filename
stringlengths
13
34
markdown
stringlengths
0
359k
10.7554_elife.88204
RESEARCH ARTICLE Transmembrane protein CD69 acts as an S1PR1 agonist Hongwen Chen1, Yu Qin1, Marissa Chou2, Jason G Cyster2,3*, Xiaochun Li1,4* 1Department of Molecular Genetics, The University of Texas Southwestern Medical Center, Dallas, United States; 2Department of Microbiology and Immunology, University of California, San Francisco, San Francisco, United States; 3Howard Hughes Medical Institute, University of California, San Francisco, San Francisco, United States; 4Department of Biophysics, The University of Texas Southwestern Medical Center, Dallas, United States Abstract The activation of Sphingosine- 1- phosphate receptor 1 (S1PR1) by S1P promotes lymphocyte egress from lymphoid organs, a process critical for immune surveillance and T cell effector activity. Multiple drugs that inhibit S1PR1 function are in use clinically for the treatment of autoimmune diseases. Cluster of Differentiation 69 (CD69) is an endogenous negative regulator of lymphocyte egress that interacts with S1PR1 in cis to facilitate internalization and degradation of the receptor. The mechanism by which CD69 causes S1PR1 internalization has been unclear. Moreover, although there are numerous class A GPCR structures determined with different small molecule agonists bound, it remains unknown whether a transmembrane protein per se can act as a class A GPCR agonist. Here, we present the cryo- EM structure of CD69- bound S1PR1 coupled to the heterotrimeric Gi complex. The transmembrane helix (TM) of one protomer of CD69 homodimer contacts the S1PR1- TM4. This interaction allosterically induces the movement of S1PR1- TMs 5–6, directly activating the receptor to engage the heterotrimeric Gi. Mutations in key residues at the interface affect the interactions between CD69 and S1PR1, as well as reduce the receptor internal- ization. Thus, our structural findings along with functional analyses demonstrate that CD69 acts in cis as a protein agonist of S1PR1, thereby promoting Gi- dependent S1PR1 internalization, loss of S1P gradient sensing, and inhibition of lymphocyte egress. Editor's evaluation This important study provides unprecedented molecular insight into the activation and internal- ization of an important cell surface receptor induced by another membrane protein. The data supporting the conclusions are compelling, which include rigorous electron microscopy analysis, and biochemical and cell- based functional assays. The findings here not only reveal important mecha- nisms of S1P GPCR regulation, but also have implications for other fields such as receptor pharma- cology and immunity. Introduction Sphingosine- 1- phosphate (S1P) plays an essential role in the immune system by promoting the egress of lymphocytes from lymphoid organs into blood and lymph via a direct interaction with one of its five cognate G protein–coupled receptors, S1PR1 (Baeyens and Schwab, 2020; Cartier and Hla, 2019; Cyster and Schwab, 2012; Pappu et al., 2007; Rosen et al., 2013; Spiegel and Milstien, 2003). After egressing from spleen, lymph nodes, or mucosal lymphoid tissues, T and B lymphocytes travel to other lymphoid organs in a cycle of continual pathogen surveillance. When an infection occurs, *For correspondence: jason.cyster@ucsf.edu (JGC); xiaochun.li@utsouthwestern. edu (XL) Competing interest: The authors declare that no competing interests exist. Funding: See page 12 Preprinted: 15 February 2023 Received: 04 April 2023 Accepted: 09 April 2023 Published: 11 April 2023 Reviewing Editor: Jungsan Sohn, Johns Hopkins University School of Medicine, United States Copyright Chen et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Chen et al. eLife 2023;12:e88204. DOI: https://doi.org/10.7554/eLife.88204 1 of 16 Research article there is a temporary shutdown of lymphocyte egress from the responding lymphoid organ(s) and this enables increased accumulation of lymphocytes and enhances the immune response (Cyster and Schwab, 2012). Egress shutdown is mediated by type I interferon (IFN) inducing lymphocyte CD69 expression. CD69, a type II transmembrane C- type lectin protein, intrinsically inhibits the function of S1PR1 in T and B cells (Shiow et al., 2006). CD69 also regulates T cell egress from the thymus (Nakayama et al., 2002; Zachariah and Cyster, 2010). A disulfide- bond in the extracellular domain links CD69 as a homodimer (Ziegler et al., 1994). Biochemical studies demonstrated that CD69 may associate with S1PR1 through interactions between their transmembrane domains (TMs) to facilitate S1PR1 internalization and degradation (Bankovich et al., 2010). Unlike S1P, CD69 has been shown to bind S1PR1 but not the other S1PRs (Bankovich et al., 2010; Jenne et al., 2009; Shiow et al., 2006). However, the mechanism of CD69- induced S1PR1 internalization and thus functional inactivation has been unclear. Importantly, several S1PR1 modulators (e.g. Fingolimod, also known as FTY720, Siponimod, Ozan- imod, and Etrasimod), have been approved for treating the autoimmune diseases multiple sclerosis and ulcerative colitis (Brinkmann et al., 2010; Chun et al., 2021; Dal Buono et al., 2022; Kappos et al., 2010). These immunosuppressants are believed to act by inhibiting S1PR1 function and thereby preventing autoimmune effector lymphocytes exiting lymphoid organs, blocking the autoimmune attack. Either sphingosine or FTY720 is metabolically catalyzed by two intracellular sphingosine kinases into the phosphorylated form (S1P or FTY720- P) and then exported to the extracellular space via S1P transporters (Baeyens and Schwab, 2020; Spiegel et al., 2019). There, S1P binds to its receptors for initiation of the signal while FTY720- P activates the S1PR1 but causes a persistent internalization and degradation of S1PR1 to attenuate the signal (Brinkmann et al., 2010). Recently, cryogenic electron microscopy (cryo- EM) structures of S1PR1 complexed with different small molecule ligands have been determined (Liu et al., 2022; Xu et al., 2022; Yu et al., 2022; Yuan et al., 2021). These findings reveal a mechanism of how S1PR1 engages its endogenous ligand S1P and its modulators to adopt the active conformation for recruiting the heterotrimeric Gi protein. The previously determined crystal structure of antagonist ML056- bound S1PR1 reveals its inactive state (Hanson et  al., 2012). However, the molecular mechanism remains unknown of how CD69 binds to S1PR1 to trigger its internalization. Therefore, structural study on the S1PR1- CD69 complex will provide molecular insights into the CD69- mediated functional inhibition of S1PR1 and reveal how a class A GPCR can be regulated by a transmembrane protein modulator. In this manuscript, we deter- mined the structure of CD69- bound S1PR1 coupled to Gαiβ1γ2 heterotrimer by cryo- EM at 3.15 Å reso- lution. Our findings reveal that TM of CD69 contacts TM4 of S1PR1 to activate the receptor allowing it to engage the α5 helix of Gαi in the absence of S1P ligand, thereby disrupting the receptor’s egress- promoting function. Results Since serum contains an abundance of lipids including S1P, we expressed human S1PR1 or CD69 in HEK293 cells cultured in a medium with lipid- deficient serum. Then, we purified human S1PR1 protein alone to validate its activation in the presence of S1P using the GTPase- Glo assay (Figure 1A). We then tested the effect on S1PR1 of adding the CD69 homodimer in the absence of S1P. Remarkably, addition of CD69 alone caused a similar amount of Gi activation as addition of S1P indicating that CD69 functions as a protein agonist of S1PR1 (Figure 1A). To perform structural studies, we mixed lysates from HEK293 cells that independently expressed human CD69 and human S1PR1. The CD69- S1PR1 complex was then incubated with Gαiβ1γ2 hetero- trimer and scFv16 (Maeda et al., 2018) at 1:1.2:1.4 molar ratio. After gel- filtration purification, the resulting complex was concentrated for cryo- EM analysis (Figure  1—figure supplement 1A). We obtained over 1  million particles from  ~4000 cryo- EM images. The overall structure of the CD69- bound S1PR1 coupled to heterotrimeric Gi was determined at 3.15 Å resolution by 293,516 particles (Figure 1—figure supplement 1B–F; Table 1). The structure shows that one S1PR1 binds one CD69 homodimer and one Gi heterotrimer. It also revealed well- defined features for the canonical seven transmembrane helices (7- TMs) of S1PR1, the Gαi Ras- like domain, the Gβ and Gγ subunits and scFv16 (Figure 1B, Figure 1—figure supplement 2A). The intracellular loop 3 (ICL3) and the C- terminus of S1PR1 and the intracellular and extracellular domains of CD69 were not found in the cryo- EM map indicating their flexibility in the complex. In contrast, the TMs of the CD69 homodimer were clearly Chen et al. eLife 2023;12:e88204. DOI: https://doi.org/10.7554/eLife.88204 2 of 16 Structural Biology and Molecular Biophysics Research article A C ) % ( r e v o n r u t P T G d e z i l a m r o N 80 70 60 50 40 B **** **** CD69-a CD69-b **** **** **** **** S1PR1 no ligand 1 nM S1P 10 nM S1P 100 nM S1P 0.5 µM CD69 2.5 µM CD69 10 µM CD69 scFv16 Gβ Gγ Giα Extracellular Space N CD69-a S1PR1 90° 6 Cytosol ICL3 C ICL2 Giα Gβ CD69-b scFv16 Gγ CD69-a CD69-b 5 7 1 4 3 2 ligand-binding pocket Figure 1. Overall structure of human CD69- S1PR1- Gi complex. (A) S1PR1- induced GTP turnover for Gi1 in the presence of purified CD69 or S1P. Luminescence signals were normalized relative to the condition with Gi1 only. Data are mean ± s.e.m. of three independent experiments. One- way ANOVA with Tukey’s test; ****p<0.0001. Experiments were repeated at least three times with similar results. (B) Cryo- EM map of human CD69 bound S1PR1- Gi complex. (C) Cartoon presentation of the complex in the same view and color scheme as shown in (B). Slab view of S1PR1 from the extracellular side showing that the orthosteric binding pocket is vacant. The online version of this article includes the following source data and figure supplement(s) for figure 1: Figure supplement 1. Cryo- EM reconstruction of CD69 bound S1PR1- Gi complex. Figure supplement 1—source data 1. Original uncropped SDS- PAGE gels for data in Figure 1—figure supplement 1. Figure supplement 1—source data 2. Uncropped SDS- PAGE gels for data in Figure 1—figure supplement 1 with the relevant bands labeled. Figure supplement 2. The cryo- EM density map of CD69- bound S1PR1- Gi complex. Figure supplement 3. Structural comparison between CD69- bound S1PR1 and S1P- bound S1PR1. Figure supplement 4. Structures of homodimeric and heterodimeric GPCRs. defined in the map owing to their interactions with S1PR1 (Figure 1B, Figure 1—figure supplement 2A; the interacting TM helix is referred to as CD69- a). Because no lipid was supplemented into the protein during the expression and purification, there is no notable lipid ligand in the 7- TM bundle of S1PR1, which is different from the previous structural discoveries on S1PRs (Chen et al., 2022; Liu et al., 2022; Xu et al., 2022; Yu et al., 2022; Yuan et al., 2021; Zhao et al., 2022). Structural comparison shows that the entire complex and the S1P bound S1PR1- Gi complex share a similar conformation with a root- mean- square deviation (RMSD) of 0.82 Å (Figure 1—figure supple- ment 3A). The receptors in both complexes can be aligned well; however, the F1614.43 in TM4 presents Chen et al. eLife 2023;12:e88204. DOI: https://doi.org/10.7554/eLife.88204 3 of 16 Structural Biology and Molecular Biophysics Research article Table 1. Cryo- EM data collection, processing, and refinement statistics. Structure CD69- S1PR1- Gi- scFv16 PDB EMDB Data collection/ processing Magnification Voltage (kV) Pixel size (Å) Defocus range (μm) Electron exposure (e-/Å2) Symmetry imposed 8G94 EMD- 29861 105,000 300 0.83 1.0–2.0 60 C1 Initial particles (No.) ~1.1 million Final particles (No.) Map resolution (Å) FSC threshold Map resolution range (Å) Refinement Model Resolution (Å) FSC threshold Map sharpening B- factor (Å2) Model composition Non- hydrogen atoms Protein residues Ligand B- factors (Å2) Protein R.m.s. deviations Bond lengths (Å) Bond angles (°) Validation MolProbity score Clashscore Rotamers outliers (%) Ramachandran plot (%) Favored Allowed Outliers 293,516 3.14 0.143 25–3.0 3.3 0.5 –60 9223 1187 0 98.23 0.006 0.702 1.64 6.49 0.00 95.87 4.13 0.00 a notable shift due to CD69 binding (Figure 1— figure supplement 3B). We docked the S1PR1 bound to the other TM of the CD69 homodimer which showed that the modeled receptor would sterically clash with the Giα subunit (Figure  1— figure supplement 3C). This may explain why only one receptor binds one CD69 homodimer in the presence of the heterotrimeric G- protein. Receptor activity- modifying protein 1 (RAMP1), a type I transmembrane domain protein, binds the calcitonin receptor- like receptor (CLR) class B GPCR to form the Calcitonin gene- related peptide (CGRP) receptor which is involved in the pathology of migraine (Russell et al., 2014). The structure of Gs- protein coupled CGRP receptor uncovers that TM of RAMP1 interacts with TMs 3–5 of CLR and the extracellular domains of RAMP1 and CLR have extensive interactions (Liang et  al., 2018; Figure  1—figure supple- ment 4A). Both CLR and RAMP1 contribute to the engagement of their agonist CGRP. However, in our structure, CD69 acts as an agonist to activate S1PR1 through a direct binding to TM4 of S1PR1 in the absence of a canonical agonist (e.g. S1P or FTY720- P). The extracellular domain of CD69 is completely invisible in the complex and may not interact with the extracellular loops of S1PR1. Another type of intramembrane interaction observed for GPCRs is the formation of either homodimers or heterodimers. The metabotropic glutamate receptor 2 (mGlu2), a Class- C GPCR, employs TM4 to maintain its inactive dimeric state or TM6 to assemble as a homodimer in the pres- ence of its agonist (Du et  al., 2021; Figure  1— figure supplement 4B). The structure of inactive mGlu2–mGlu7 heterodimer shows that TM5 plays a key role in the complex assembly (Du et  al., 2021; Figure  1—figure supplement 4C). More- over, TM1 of the class D GPCR Ste2 is responsible for engaging the TM1 of another Ste2 to form a homodimer (Velazhahan et al., 2021; Figure 1— figure supplement 4D). These findings elucidate that GPCRs could employ distinct TMs to recruit their transmembrane binding partners. The TM of one protomer of CD69 homodimer interacts with the TM4 of S1PR1 (Figure 1C). The interface area between TMs is about 600 Å2. Struc- tural analysis shows that residues V41, V45, V48, V49, T52, I56, I59, A60 of CD69 mediate its exten- sive interactions with the receptor (Figure  2A, Figure  1—figure supplement 2B). Residues L1604.42, F1614.43, I1644.46, W1684.50, V1694.51, L1724.54, I1734.55, G1764.58, I1794.61 and M1804.62 of S1PR1- TM4 contribute to the interaction with CD69 (Figure 2B, Figure 1—figure supplement Chen et al. eLife 2023;12:e88204. DOI: https://doi.org/10.7554/eLife.88204 4 of 16 Structural Biology and Molecular Biophysics Research article A C E A60 I179 M180 I59 I173 I56 L172 T52 V169 W168 I164 V49 V48 V45 F161 TM4 L160 CD69-a V41 90° B I173 V169 M180 T52 I59 I56 A60 L172 V49 I179 –50 S1PR1 –37 –50 CD69 –37 –50 CD69 –37 –50 Tubulin F S1PR1 – WT V169Y M180Y CD69 WT WT WT WT Lane # 1 2 3 4 WT V48F, V49F 5 WT I56F, I59F 6 IP: Flag WB: Flag IP: Flag WB: Strep Lysates WB: Strep Lysates WB: Tubulin D % e s a e e r l F G T - P A 14 12 10 8 6 4 2 0 -2 S1PR1WT S1PR1V169Y S1PR1M180Y -12 -11 -10 -9 -7 -8 Log [S1P (M)] -6 -5 -4 ) % ( r e v o n r u t P T G d e z i l a m r o N 70 65 60 55 50 Control CD69(V48F/V49F) CD69(WT) CD69(I56F/I59F) S1PR1 S1PR1 + CD69 S1PR1 + CD69(V48F/V49F) S1PR1 + CD69(I56F/I59F) Figure 2. The binding interface between CD69 and S1PR1. (A) and (B) Detailed interactions between CD69- a and TM4 of S1PR1. Residues that contribute to complex formation are labeled. CD69 is shown in green and S1PR1 in slate. (C) S1PR1- Flag and CD69- StrepII co- immunoprecipitation assay in transfected HEK293 GnTI- cells from one experiment that is representative of three. (D) Dose- response curves of S1PR1WT, S1PR1V169Y and S1PR1M180Y for the TGFα shedding assay using S1P. Data are mean ± s.d. (n=3). (E) S1PR1- induced GTP turnover for Gi1 in the presence of purified wild- type and mutant CD69. Luminescence signals were normalized relative to the condition with Gi1 only. Data are mean ± s.e.m. of three independent experiments. One- way ANOVA with Tukey’s test; ***p<0.001, ****p<0.0001. Experiments in (C)-(E) were repeated at least twice with similar results. (F) Flow cytometric analysis of S1PR1 surface expression on WEHI231 lymphoma cells transduced with S1PR1 and CD69 wild- type and mutant constructs as indicated. From one experiment that is representative of three. The online version of this article includes the following source data and figure supplement(s) for figure 2: Source data 1. Original uncropped western blots for data in Figure 2. Source data 2. Uncropped western blots for data in Figure 2 with the relevant bands labeled. Figure supplement 1. Size exclusion column profiles of CD69 wild type and mutants. Figure supplement 1—source data 1. Original uncropped SDS- PAGE gels for data in Figure 2—figure supplement 1. Figure supplement 1—source data 2. Uncropped SDS- PAGE gels for data in Figure 2—figure supplement 1 with the relevant bands labeled. 2B). However, the TM of another CD69 does not have any interactions with the receptor and hetero- trimeric Gi protein (Figure 1C). Further structural comparison with the S1P- bound S1PR1- Gi complex indicates that the heterotrimeric Gi proteins in both complexes exhibit a similar state with a RMSD of 0.45  Å. Also, the intracellular regions of the heptahelical domain adopt a similar conformation to accommodate the Gi proteins. This finding implies that S1P and CD69 stimulate the receptor to engage the heterotrimeric Gi proteins in an analogous fashion. To validate our structural observations, we performed the co- immunoprecipitation (co- IP) assay using S1PR1 and CD69 variants. Compared to the wild- type S1PR1, two mutants (V1694.51Y and M1804.62Y) present reduced binding to CD69 (Figure  2C). The TGFα shedding assay showed that these two mutants retained normal activity in response to S1P (Figure 2D). We also tested two CD69 double mutations (V48F/V49F and I56F/I59F) for their association with S1PR1. The co- IP results show that the interaction between S1PR1 and either mutant is considerably attenuated, thus directly supporting the role of CD69- TM in the complex assembly (Figure 2C). Moreover, we have purified Chen et al. eLife 2023;12:e88204. DOI: https://doi.org/10.7554/eLife.88204 5 of 16 Structural Biology and Molecular Biophysics Research article A C ML056 TM6 inactive S1PR1 (3V2Y) B 90° TM6 TM4 CD69-a V49 V45 V41 L174 I170 F210 C167 F133 TM4 TM3 TM5 90° W269 TM6 F273 C167 F133 F210 TM5 L174 D E F273 W269 TM5 TM6 W269 F273 F133 F210 S1P bound S1PR1 (7TD3) L174 I170 Figure 3. Comparison between CD69- bound S1PR1 and ML056- or S1P- bound S1PR1. (A) Overall structures of S1PR1 binding with CD69 and ML056. The CD69- bound S1PR1 structure was aligned to ML056- bound inactive S1PR1 (PDB code: 3V2Y). ML056- bound receptor is shown in brown, CD69- bound receptor in blue, and the TM of CD69 in green. The same color scheme is used (C) and (D). (B) The movements of TM4 and TM6 of CD69- bound S1PR1 compared with ML056- bound inactive S1PR1. (C) Residues involved in the TM movements. (D) TM6 movement around W2696.48 and F2736.52. (E) Comparison between CD69- bound S1PR1 and S1P- bound S1PR1 (PDB code: 7TD3). Residues in the ligand binding pocket are shown. CD69- bound receptor in blue and S1P- bound S1PR1 in cyan. S1P is shown as balls and sticks in yellow. The online version of this article includes the following source data and figure supplement(s) for figure 3: Figure supplement 1. S1PR1 specificity for CD69 binding. Figure supplement 1—source data 1. Original uncropped western blots for data in Figure 3—figure supplement 1. Figure supplement 1—source data 2. Uncropped western blots for data in Figure 3—figure supplement 1 with the relevant bands labeled. Figure supplement 2. Structures of GPCRs with their positive allosteric modulators. CD69(V48F/V49F) and CD69(I56F/I59F) individually and mixed with S1PR1 and Gαiβ1γ2 to conduct a GTPase- Glo assay (Figure 2—figure supplement 1). Consistent with the results of our co- IP assays, the activation of Gi proteins in the presence of either variant was decreased (Figure 2E). To further validate the physiological role of the CD69- S1PR1- Gi complex, we tested the two CD69 variants, for their influence on CD69- mediated S1PR1 internalization in WEHI231 B lymphoma cells. In accord with the biochemical data, CD69(V48F/V49F), and CD69(I56F/I59F) were both reduced in their ability to downregulate S1PR1 (Figure 2F). The structures of the S1PR1 complex with its small molecule modulators (including S1P, FTY720- P, BAF312, and ML056) uncover that the TMs 3, 5, 6, and 7 contribute to accommodate the modulators in the orthosteric site (Hanson et  al., 2012; Liu et  al., 2022; Xu et  al., 2022; Yu et al., 2022; Yuan et al., 2021). In contrast, S1PR1 employs its TM4 to associate with CD69 which functions as a protein agonist for triggering receptor activation. Structural comparison with the inactive state of ML056 bound S1PR1 reveals a unique mechanism of CD69- mediated S1PR1 activation (Figure  3A). The binding of CD69 induces a 4  Å shift at the intracellular end of TM4 causing the residues C1674.49, I1704.52, and L1744.56 in TM4 to face TM3 (Figure 3B). C1674.49 and I1704.52 have hydrophobic contacts with the F1333.41 in TM3, and L1744.56 pushes the F2105.47 in TM5 towards the edge of the receptor to form the hydrophobic interactions with W2696.48 and F2736.52 in TM6 (Figure  3C, Figure  1—figure supplement 2C). These interactions trigger the notable Chen et al. eLife 2023;12:e88204. DOI: https://doi.org/10.7554/eLife.88204 6 of 16 Structural Biology and Molecular Biophysics Research article movement of TM5 and TM6 allowing the opening of intracellular regions to engage the hetero- trimeric Gi proteins (Figure 3A and D). Residues A1373.45, I1704.52, L1744.56, F2105.47, W2696.48, and F2736.52 are conserved among S1PR1, S1PR2 and S1PR3, but not F1333.41 and C1674.49. Remarkably, further comparison shows that the key residues, which are crucial for the S1P binding and receptor activation, present similar conformations in the structures of S1P- bound S1PR1 and CD69- bound S1PR1, although S1P and CD69 have different structural natures and completely distinct binding sites in the receptor (Figure 3E). To date, five S1PRs have been identified. These receptors have different tissue distributions, and they also function via distinct kinds of G proteins (including Gi, Gq, and G12/13) (Cartier and Hla, 2019). Previous work showed that CD69 specifically binds to S1PR1, and it does not associate with S1PR2, S1PR3 or S1PR5 (Bankovich et  al., 2010; Jenne et  al., 2009; Shiow et  al., 2006). To dissect the binding specificity of CD69, we carried out the co- IP assays to show a very weak interaction between S1PR2 and CD69 (Figure 3—figure supplement 1A). Although the sequence homology among five S1PRs is high, residues L1574.51 and L1684.62 in S1PR2- TM4 are not conserved with those in S1PR1 and are determinants for specific recognition of CD69 (Figure 3—figure supplement 1B). We spec- ulated that converting these two residues to those in S1PR1 may prompt the interaction between S1PR2 variant and CD69. Our co- IP result clearly shows that S1PR2(L1574.51V/L1684.62M) could interact with CD69 albeit the interactions are weaker than that between S1PR1 and CD69 (Figure 3—figure supplement 1A). This finding further demonstrates the essential role of S1PR1- TM4 in the CD69- mediated S1PR1 signaling. All the known small molecule S1PR1 agonists or antagonists bind to the orthosteric site in the heptahelical domain (Hanson et al., 2012; Liu et al., 2022; Xu et al., 2022; Yu et al., 2022; Yuan et al., 2021). Interestingly, the CD69 binding site is akin to that of the allosteric agents which attach to receptors (Draper- Joyce et al., 2021; Mao et al., 2020; Yang et al., 2022), although the nature of these agents and CD69 is quite different. The diversity of the allosteric modulator binding sites in GPCRs has been revealed by numerous structures (Figure 3—figure supplement 2). When the ortho- steric site is occupied, the positive allosteric modulator attaches to the receptor and then increases agonist affinity and/or efficacy. CD69 binds to the edge of S1PR1, but it acts as a protein agonist to directly activate the receptor in the absence of any agonists in the orthosteric site. Thus, our finding suggests CD69 is different from other S1PR1 agonists in that it functions via a direct binding to the edge of the receptor. It remains unknown whether the antagonist of S1PR1 bound to the 7- TMs will affect the CD69- mediated regulation of S1PR1. We co- transfected S1PR1- GFP and CD69- mCherry into HEK293 cells in a lipid depleted medium. After 24  hr, the fluorescence images show that substantial receptors (~80%) have been internalized with CD69. However, when we added the Ex26, a potent S1PR1 antag- onist (Cahalan et  al., 2013), into the cells 6 hr after transfection, the images show that just ~50% receptors have been internalized (Figure 4A and B). This finding indicates that the CD69- mediated S1PR1 activation could be reversed when the 7- TMs pocket is preoccupied by an antagonist. It has been known that S1P, FTY720- P and CD69, could promote the internalization of S1PR1. However, the mechanisms of S1P- and FTY720- P- mediated internalization appear to be different. While both S1P and FTY720- P activate Gi- signaling, FTY720- P is considered as a β-arrestin- biased agonist, and FTY720- P- induced S1PR1 internalization is β-arrestin- dependent (Oo et  al., 2007; Xu et  al., 2022). The pathway of S1P- mediated internalization can be β-arrestin- dependent or independent (Galvani et  al., 2015; Reeves et  al., 2016). To test the mechanism of how CD69 induces the receptor internalization, we performed a fluorescence imaging assay to check the internalization of S1PR1 in the presence of either Gi inhibitor Pertussis toxin (PTX) or β-arrestin inhibitor Barbadin. The plasmids encoding S1PR1- GFP and CD69- mCherry were co- transfected into HEK293 cells in a lipid depleted medium. After 6 hr, we added PTX or Barbadin. On day 2, we calculated the fraction of internalized S1PR1 in each group by fluorescence imaging. The results show that Barbadin does not interfere with CD69- induced receptor internalization (Figure 4C and D), but PTX could prevent half of the receptors from internalization (Figure 4E and F). Our finding also supports that Barbadin was effective in reducing FTY70- P- induced S1PR1 internalization (Figure 4—figure supplement 1). Thus, CD69 agonism of S1PR1 induces Gi- dependent internal- ization of the complex. Chen et al. eLife 2023;12:e88204. DOI: https://doi.org/10.7554/eLife.88204 7 of 16 Structural Biology and Molecular Biophysics Research article A Vehicle Ex26 C Vehicle Barbadin E Vehicle PTX S1PR1-EGFP CD69-mCherry Merge B 100 S1PR1-EGFP CD69-mCherry Merge D 1 R P 1 S r a u l l l e c a r t n I S1PR1-EGFP CD69-mCherry Merge F 1 R P 1 S r a u l l l e c a r t n I 1 R P 1 S r a u l l l e c a r t n I ) l l l e c e o h w e h t f o % ( 80 60 40 20 0 Vehicle Ex26 100 ns ) l l l e c e o h w e h t f o % ( 80 60 40 20 0 Vehicle Barbadin 100 ) l l l e c e o h w e h t f o % ( 80 60 40 20 0 Vehicle PTX Figure 4. CD69 induced S1PR1 internalization. (A) HEK293 cells were treated with 2 μM Ex26 or vehicle for 12 hr and imaged using confocal microscopy. Scale bar, 10 μm. (B) Quantification of intracellular S1PR1 of the cells in (A). (C) HEK293 cells were treated with 20 μM Barbadin for 12 hr and imaged for analysis. Scale bar, 10 μm. (D) Quantification of intracellular S1PR1 of the cells in (C). (E) HEK293 cells were treated with 200 ng/ml pertussis toxin (PTX) for 12 hr and imaged for analysis. Scale bar, 10 μm. (F) Quantification of intracellular S1PR1 of the cells in (E). Data are mean ± s.e.m. Two- sided Welch’s t- test; ns, not significant, **p<0.01, ****p<0.0001. All experiments were repeated at least three times with similar results. The online version of this article includes the following figure supplement(s) for figure 4: Figure supplement 1. Barbadin alters the FTY720- P mediated S1PR1 internalization. Figure supplement 2. S1PR1- induced GTP turnover for Gi1 in the presence of purified CD69 and S1P. Discussion Our studies provide a model for understanding how the lymphocyte activation marker CD69 controls lymphocyte egress and thus augments adaptive immunity. As an immediate early gene, CD69 is strongly transcriptionally induced in lymphocytes within an hour of exposure to type I IFN, toll- like receptor (TLR) ligands, or antigen receptor engagement (Grigorova et al., 2010; Shiow et al., 2006; Ziegler et al., 1994). Following induction, CD69 protein engages S1PR1 as an agonist, causing S1PR1 internalization and loss of the ability to sense S1P gradients. We speculate that even prior to internal- ization, CD69 disrupts S1PR1’s egress promoting function by acting as a high concentration agonist Chen et al. eLife 2023;12:e88204. DOI: https://doi.org/10.7554/eLife.88204 8 of 16 Structural Biology and Molecular Biophysics Research article and thus making the receptor ‘blind’ to S1P distribution. Consistently, our functional analysis reveals that CD69 could not synergize with S1P to trigger S1PR1 activation (Figure 4—figure supplement 2). Previous work has shown the critical importance of correctly distributed S1P and thus correctly localized S1PR1 activation for effective lymphocyte egress (Schwab et al., 2005). As well as promoting egress, S1PR1, transmits signals needed for maintaining T cell survival (Mendoza et al., 2017) and CD69 has been implicated in transmitting signals that influence T cell differentiation (Cibrián and Sánchez- Madrid, 2017; Kimura et al., 2017). Whether the CD69- S1PR1 complex contributes to these signals before undergoing degradation merits further study. GRK2 (Arnon et  al., 2011; Oo et  al., 2007; Watterson et al., 2002) and dynamin (Willinger et al., 2014) participate in S1PR1 internaliza- tion in response to S1P. In accord with these factors possibly having a role in CD69- mediated S1PR1 internalization, they have been shown to promote internalization of some receptors independently of β-arrestins (Moo et al., 2021). The selectivity of CD69 for S1PR1 is important for allowing activated CD69+ lymphocytes and natural killer cells to employ other S1PRs, such as S1PR2 and S1PR5, to carry out functions without interruption by CD69 (Jenne et  al., 2009; Laidlaw et  al., 2019; Moriyama et al., 2014). The lack of conservation of key residues that mediate the S1PR1- CD69 interaction in TM4 of S1PR2, S1PR5 and the other S1PRs provides an explanation for this selectivity (Figure 3— figure supplement 1B). In summary, we provide the first example of GPCR activation by interaction in cis with a transmembrane ligand and thereby explain the mechanism of lymphocyte egress shutdown. The structure also offers insights that may enable introduction of transcriptionally inducible GPCR switches into CAR- T cells and other engineered cell types. Methods Constructs For expression and purification, the wild- type human S1PR1 (a.a.1–347, UniProt: P21453) and CD69 (full- length, UniProt: Q07108) were separately cloned into pEZT- BM vector (Morales- Perez et  al., 2016) with a C- terminal Flag tag and StrepII tag, respectively. Plasmids of Gαi1, Gβ1/Gγ2 and scFv16 are kind gifts from Brian Kobilka (Stanford University). For co- immunoprecipitation assay, the full- length wild- type human S1PR1 fused with a C- terminal Flag tag and CD69 fused with a C- terminal StrepII tag, were separately cloned into pCAGGS vector (Niwa et al., 1991) with modified multiple cloning sites. For fluorescence microscopy, the plasmids pCAGGS- S1PR1- Flag- GFP and pCAGGS- CD69- StrepII- mCherry were constructed. Protein expression and purification S1PR1- Flag and CD69- StrepII were separately expressed using baculovirus- mediated transduction of mammalian HEK293S GnTI− cells (ATCC CRL- 3022) in a medium containing FreeStyle 293 (Gibco Cat# 12338018) supplemented with 2% charcoal- dextran stripped fetal bovine serum (Gibco Cat# 12676029), penicillin (100 U/mL), and streptomycin (100 μg/mL) (Corning Cat# 30–002 CI). Baculo- viruses were generated in Sf9 cells, and P2 virus was used to infect HEK293S GnTI− cells at 37 °C. At 8 hr after infection, sodium butyrate at a final concentration of 10 mM was added to the culture. After further incubation for 64 hr at 30 °C, cells expressing S1PR1- Flag and CD69- StrepII were mixed together and resuspended in buffer A (20  mM HEPES, pH 7.5, 150  mM NaCl) supplemented with protease inhibitors and then homogenized by sonication. The protein was solubilized with 1% LMNG (lauryl maltose neopentyl glycol) /0.1% CHS (cholesteryl hemisuccinate) for 1  hr at 4  °C. Insoluble material was removed by centrifugation at 40,000 g, 4 °C for 30 min, and the supernatant was incu- bated with Strep- Tactin XT resin (IBA Cat# 2- 5030- 025) for batch binding. The resin was washed with 20 column volumes (CV) of buffer A containing 0.01% LMNG/0.001% CHS. The protein complex was eluted with 6 CVs of buffer A containing 0.01% LMNG/0.001% CHS and 50 mM biotin, followed by a second affinity purification by anti- Flag M2 resin (Sigma- Aldrich). The excessive CD69- StrepII was washed off with 20 CVs of buffer A containing 0.01% LMNG/0.001% CHS, and the complex was eluted with 5 CVs of 3×Flag peptide (0.1 mg/ml; ApexBio). The eluted protein was further purified by gel filtration using a Superose 6 Increase 10/300 GL column (Cytiva) with 20 mM HEPES, pH 7.5, 150 mM NaCl, 0.001% L- MNG/0.0001% CHS, and 0.0025% glyco- diosgenin (GDN). The peak frac- tions were collected for complex assembly. Chen et al. eLife 2023;12:e88204. DOI: https://doi.org/10.7554/eLife.88204 9 of 16 Structural Biology and Molecular Biophysics Research article To assemble the CD69- S1PR1- Gi- scFv16 complex, purified CD69- S1PR1 was mixed with the Gi heterotrimer at a 1:1.2 molar ratio. This mixture was incubated on ice for 1 hr, followed by the addi- tion of apyrase to catalyze the hydrolysis of unbound GDP on ice for 1 hr. Then, scFv16 was added at a 1.4:1 molar ratio (scFv16: CD69- S1PR1) followed by 30 min incubation on ice. The mixture was diluted 10- fold by gel filtration column buffer. To remove excess Gi and scFv16 proteins, the mixture was purified by anti- Flag M2 affinity chromatography. The complex was eluted and concentrated using an Amicon Ultra Centrifugal Filter (molecular weight cutoff 100 kDa). The complex was further purified by gel filtration (Superose 6 Increase 10/300 GL) with buffer 20 mM HEPES, pH 7.5, 150 mM NaCl, 0.001% L- MNG/0.0001% CHS, and 0.0025% GDN. Peak fractions consisting of CD69- S1PR1- Gi complex were concentrated to ~10–12 mg/ml for cryo- EM studies. Cryo-EM sample preparation and data acquisition The freshly purified CD69- S1PR1- Gi- scFv16 complex was added to Quantifoil R1.2/1.3 400- mesh Au holey carbon grids (Quantifoil), blotted using a Vitrobot Mark IV (FEI), and vitrified in liquid ethane. The grids were imaged in a 300- kV Titan Krios (FEI) with a Gatan K3 Summit direct electron detector. Data were collected in super- resolution mode at a pixel size of 0.415 Å with a dose rate of 23 electrons per physical pixel per second. Images were recorded for 5 s exposures in 50 subframes with a total dose of 60 electrons per Å2. Imaging processing and 3D reconstruction A total of 4,239 dose- fractionated image stacks of CD69- S1PR1- Gi complex were collected and subjected to single particle analysis using RELION- 3.1 (Zivanov et  al., 2018) and cryoSPARC v3.3 (Punjani et  al., 2017). MotionCor2 (Zheng et  al., 2017) was used for motion correction and dose weighting, CTFFIND- 4.1 Rohou and Grigorieff, 2015 for contrast transfer function (CTF) estimation, and crYOLO Wagner et al., 2019 for particle picking with a general model. A total of 1,113,446 parti- cles were extracted with a pixel size of 1.66 Å in RELION and imported to cryoSPARC. The imported particles were subjected to ab initio model reconstruction and several rounds of alternating 2D classi- fication and heterogeneous refinement. Then 336,669 particles from the best class were re- extracted at full pixel size (0.83 Å) in RELION and imported to cryoSPARC again. Two heterogeneous refine- ments were performed in parallel and the resulting particles from the two best classes were combined with duplicates removed. These 293,516 particles were subjected to CTF refinement and Bayesian polishing followed by masked 3D auto refinement. RELION postprocessing was used for sharpening of the final map. Model construction and refinement The cryo- EM structure of the S1PR1- Gi bound to S1P (PDB: 7TD3) (Liu et  al., 2022) was used as initial models and manually docked into cryo- EM density map with UCSF Chimera- 1.15 (Pettersen et  al., 2004). The transmembrane helix of CD69 was manually built using Coot- 0.9.6 (Emsley and Cowtan, 2004). Due to the limited local resolution, the TM of CD69- b was built as polyalanine. The resulting model was subjected to iterative rounds of manual adjustment and rebuilding in Coot and real- space refinement in Phenix- 1.16 (Adams et  al., 2010). MolProbity (Williams et  al., 2018) was used to validate the geometries of the model. Structural figures were generated using UCSF Chime- ra- 1.15, ChimeraX- 1.5 (Pettersen et al., 2021), and PyMOL- 2.3 (https://pymol.org/2/). GTP turnover assay GTP turnover was analyzed using GTPase- Glo Assay kit (Promega Cat# V7681). Briefly, the purified S1PR1 was first incubated with purified CD69 and/or S1P followed by mixing with isolated Gi protein in an assay buffer containing 20 mM HEPES, pH7.5, 150 mM NaCl, 0.01% LMNG/0.001% CHS, 10 mM MgCl2, 100  μM TCEP, 10  μM GDP and 5  μM GTP. After incubation for 60  min, the reconstituted GTPase- Glo reagent was added to the sample and incubated for 30 min at room temperature. The amount of remaining GTP was assessed by measuring luminescence after adding and incubation with the detection reagent for 10 min at room temperature. The luminescence signal was normalized in each case to that of G- protein alone. Data were analyzed using GraphPad Prism 9. Co-immunoprecipitation and immunoblotting assay HEK293 GnTI- cells were transfected with plasmids encoding CD69- StrepII and S1PR1- Flag using FuGene 6 transfection reagent in 60 mm dishes. Forty- eight hr post transfection, cells were harvested Chen et al. eLife 2023;12:e88204. DOI: https://doi.org/10.7554/eLife.88204 10 of 16 Structural Biology and Molecular Biophysics Research article and whole cell lysates were prepared using IP lysis buffer (Thermo Scientific) supplemented with protease inhibitor cocktail (Roche). Lysates were cleared by centrifuging at 20,000 g for 15 min at 4 °C. Supernatants were incubated with anti- Flag M2 affinity beads (MilliporeSigma) with end- over- end rotation for 2 h at 4 °C. Beads were washed three times with lysis buffer for 5  min per wash with end- over- end rotation at 4 °C. Proteins were eluted from beads with lysis buffer supplemented with 0.4  mg/ml 3×Flag peptide. Protein samples were loaded Bolt 4–12% Bis- Tris plus gels (Invitrogen) and transferred to TransBlot Turbo nitrocellulose membranes (Bio- Rad). Membranes were blocked for 1 hr at room temperature with 5% milk in PBS with 0.05% Tween 20 (PBST) followed by primary antibody incubation, three- times wash, secondary antibody incubation, and three- times wash again. Membranes were developed for 2 min at room temperature using SuperSignal West Pico PLUS Chemi- luminescent Substrate (Thermo Scientific) and then imaged using the LI- COR Odyssey Fc imaging system. The following primary antibodies were used: Tubulin (D3U1W), Cell Signaling Cat# 86298 (1:3000 dilution); Flag tag (FLA- 1), MBL International Cat# M185- 3L (1:3000 dilution); StrepII tag, IBA GmbH Cat# 2- 1507- 001 (1:2000 dilution). Anti- mouse IgG HRP- linked secondary antibody (Cell Signaling Cat# 7076) was used for chemiluminescent detection (1:3000 dilution). TGFα shedding assay The agonist activity of S1P for the mutant S1PR1s was determined by the TGFα shedding assays (Inoue et al., 2012). Briefly, three pCAGGS plasmids encoding the human full- length S1PR1 variant (empty vector as negative control), the chimeric Gαq/i1 subunit and alkaline phosphatase- fused TGFα (AP- TGFα) were co- transfected into HEK293 cells using FuGene 6 transfection reagent in a 12- well plate. After 24  hr, the transfected cells were collected by trypsinization, washed with phosphate- buffered saline (PBS), and resuspended in Hanks’ balanced salt solution (HBSS) with 5 mM HEPES (pH 7.4). Then, the cells were seeded into a 96- well culture plate and treated with S1P, which was serially diluted in HEPES- containing HBSS with 0.01% fatty acid–free bovine serum albumin. After incubation with S1P, the cell plate was spun, and conditioned media was transferred to an empty 96- well plate. AP reaction solution (120 mM Tris- HCl, pH 9.5, 40 mM NaCl, 10 mM MgCl2, and 10 mM p- nitrophenyl phosphate) was added into the cell plates and the conditioned media plates. The absorbance at 405 nm was measured using a microplate reader (Synergy Neo2, BioTek) before and after 2 hr incuba- tion at 37 °C. Ligand- induced AP- TGFα release was calculated as described previously (Inoue et al., 2012). AP- TGFα release signal of empty vector- transfected cells were subtracted from that of S1PR1 cells at the corresponding S1P concentration points. Then, the vehicle- treated AP- TGFα release signal was set as a baseline and ligand- induced AP- TGFα release signals were fitted to a four- parameter sigmoidal concentration–response curve using GraphPad Prism 9 software. Fluorescence microscopy HEK293 cells were plated in 35 mm glass bottom dishes (Cellvis Cat# D35141.5N) followed by trans- fection with S1PR1- GFP and/or CD69- mCherry using FuGene 6 reagent on the next day. Twenty- four  hr post transfection, the cells were stained with Hoechst 33342 reagent (Thermo Fisher Cat# R37605) and fluorescence images were acquired using a Zeiss LSM 800 microscope system with ZEN imaging software (Zeiss). For fluorescence quantification of intracellular S1PR1 and CD69, outside and inside of plasma membrane were circled manually in Fiji software (Schindelin et al., 2012). The fluorescence intensi- ties in each circle were measured and regarded as whole- cell and intracellular fluorescence intensity, respectively. The intracellular fluorescence intensity was normalized to its corresponding whole- cell fluorescence intensity. For each data point, ~30 cells were randomly selected for quantification. The data shown in the figures are representative of two or more independent experiments. WEHI231 cell retroviral transduction WEHI231 B lymphoma cells were co- transduced with retroviral constructs encoding OX56 N- ter- minal tagged human S1PR1 containing an IRES- hCD4 reporter and either empty vector or constructs encoding wildtype, V48F/V49F or I56F/I59F human CD69 and an IRES- GFP reporter using methods previously described (Lu et al., 2019). After 3–5 days, the cells were harvested and rested for 20 min at 37 °C in PBS without serum, then stained to detect OX56, CD69, and hCD4. OX56 (S1PR1) staining on hCD4 +GFP + CD69 + cells were plotted. Chen et al. eLife 2023;12:e88204. DOI: https://doi.org/10.7554/eLife.88204 11 of 16 Structural Biology and Molecular Biophysics Research article Acknowledgements The data were collected at the UT Southwestern Medical Center Cryo- EM Facility (funded in part by the CPRIT Core Facility Support Award RP170644). We thank L Beatty, L Esparza, and Y Xu for technical support. This work was supported by NIH P01 HL160487, R01 GM135343, and Welch Foundation (I- 1957) (to XL) and R01 AI040098 (to JGC). JGC is an investigator of Howard Hughes Medical Institute. This article is subject to HHMI’s Open Access to Publications policy. HHMI lab heads have previously granted a nonexclusive CC BY 4.0 license to the public and a sublicensable license to HHMI in their research articles. Pursuant to those licenses, the author- accepted manu- script of this article can be made freely available under a CC BY 4.0 license immediately upon publication. Additional information Funding Funder National Institutes of Health National Institutes of Health Grant reference number Author P01 HL160487 Xiaochun Li R01 GM135343 Xiaochun Li Welch Foundation I-1957 Howard Hughes Medical Institute National Institutes of Health Xiaochun Li Jason G Cyster R01 AI040098 Jason G Cyster The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Author contributions Hongwen Chen, Conceptualization, Investigation, Writing – review and editing; Yu Qin, Marissa Chou, Investigation, Writing – review and editing; Jason G Cyster, Xiaochun Li, Conceptualization, Supervi- sion, Funding acquisition, Writing – original draft, Writing – review and editing Author ORCIDs Hongwen Chen Jason G Cyster Xiaochun Li http://orcid.org/0000-0002-1065-9808 http://orcid.org/0000-0002-4735-9745 http://orcid.org/0000-0002-0177-0803 Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.88204.sa1 Author response https://doi.org/10.7554/eLife.88204.sa2 Additional files Supplementary files • MDAR checklist Data availability The 3D cryo- EM density maps have been deposited in the Electron Microscopy Data Bank under the accession number EMD- 29861. Atomic coordinates for the atomic model have been deposited in the Protein Data Bank under the accession number 8G94. All other data needed to evaluate the conclu- sions in the paper are present in the paper and/or the supplementary materials. Chen et al. eLife 2023;12:e88204. DOI: https://doi.org/10.7554/eLife.88204 12 of 16 Structural Biology and Molecular Biophysics Research article The following datasets were generated: Author(s) Chen H, Li X Year 2023 Chen H, Li X 2023 Dataset title Dataset URL Database and Identifier Structure of CD69- bound S1PR1 coupled to heterotrimeric Gi Structure of CD69- bound S1PR1 coupled to heterotrimeric Gi https://www. rcsb. org/ structure/ 8G94 RCSB Protein Data Bank, 8G94 https://www. ebi. ac. uk/ emdb/ EMD- 29861 EMBD, EMD- 29861 References Adams PD, Afonine PV, Bunkóczi G, Chen VB, Davis IW, Echols N, Headd JJ, Hung L- W, Kapral GJ, Grosse- Kunstleve RW, McCoy AJ, Moriarty NW, Oeffner R, Read RJ, Richardson DC, Richardson JS, Terwilliger TC, Zwart PH. 2010. PHENIX: a comprehensive python- based system for macromolecular structure solution. Acta Crystallographica. Section D, Biological Crystallography 66:213–221. DOI: https://doi.org/10. 1107/S0907444909052925, PMID: 20124702 Arnon TI, Xu Y, Lo C, Pham T, An J, Coughlin S, Dorn GW, Cyster JG. 2011. GRK2- dependent S1PR1 desensitization is required for lymphocytes to overcome their attraction to blood. Science 333:1898–1903. DOI: https://doi.org/10.1126/science.1208248, PMID: 21960637 Baeyens AAL, Schwab SR. 2020. Finding a way out: S1P signaling and immune cell migration. Annual Review of Immunology 38:759–784. DOI: https://doi.org/10.1146/annurev-immunol-081519-083952, PMID: 32340572 Bankovich AJ, Shiow LR, Cyster JG. 2010. CD69 suppresses sphingosine 1- phosophate receptor- 1 (S1P1) function through interaction with membrane helix 4. The Journal of Biological Chemistry 285:22328–22337. DOI: https://doi.org/10.1074/jbc.M110.123299, PMID: 20463015 Brinkmann V, Billich A, Baumruker T, Heining P, Schmouder R, Francis G, Aradhye S, Burtin P. 2010. Fingolimod (FTY720): discovery and development of an oral drug to treat multiple sclerosis. Nature Reviews. Drug Discovery 9:883–897. DOI: https://doi.org/10.1038/nrd3248, PMID: 21031003 Cahalan SM, Gonzalez- Cabrera PJ, Nguyen N, Guerrero M, Cisar EAG, Leaf NB, Brown SJ, Roberts E, Rosen H. 2013. Sphingosine 1- phosphate receptor 1 (S1P(1)) upregulation and amelioration of experimental autoimmune encephalomyelitis by an S1P(1) antagonist. Molecular Pharmacology 83:316–321. DOI: https://doi.org/10. 1124/mol.112.082958, PMID: 23204443 Cartier A, Hla T. 2019. Sphingosine 1- phosphate: lipid signaling in pathology and therapy. Science 366:eaar5551. DOI: https://doi.org/10.1126/science.aar5551, PMID: 31624181 Chen H, Chen K, Huang W, Staudt LM, Cyster JG, Li X. 2022. Structure of S1PR2- heterotrimeric G(13) signaling complex. Science Advances 8:eabn0067. DOI: https://doi.org/10.1126/sciadv.abn0067 Chun J, Giovannoni G, Hunter SF. 2021. Sphingosine 1- phosphate receptor modulator therapy for multiple sclerosis: differential downstream receptor signalling and clinical profile effects. Drugs 81:207–231. DOI: https://doi.org/10.1007/s40265-020-01431-8, PMID: 33289881 Cibrián D, Sánchez- Madrid F. 2017. CD69: from activation marker to metabolic gatekeeper. European Journal of Immunology 47:946–953. DOI: https://doi.org/10.1002/eji.201646837, PMID: 28475283 Cyster JG, Schwab SR. 2012. Sphingosine- 1- phosphate and lymphocyte egress from lymphoid organs. Annual Review of Immunology 30:69–94. DOI: https://doi.org/10.1146/annurev-immunol-020711-075011, PMID: 22149932 Dal Buono A, Gabbiadini R, Alfarone L, Solitano V, Repici A, Vetrano S, Spinelli A, Armuzzi A. 2022. Sphingosine 1- phosphate modulation in inflammatory bowel diseases: keeping lymphocytes out of the intestine. Biomedicines 10:1735. DOI: https://doi.org/10.3390/biomedicines10071735, PMID: 35885040 Draper- Joyce CJ, Bhola R, Wang J, Bhattarai A, Nguyen ATN, Cowie- Kent I, O’Sullivan K, Chia LY, Venugopal H, Valant C, Thal DM, Wootten D, Panel N, Carlsson J, Christie MJ, White PJ, Scammells P, May LT, Sexton PM, Danev R, et al. 2021. Positive allosteric mechanisms of adenosine A1 receptor- mediated analgesia. Nature 597:571–576. DOI: https://doi.org/10.1038/s41586-021-03897-2, PMID: 34497422 Du J, Wang D, Fan H, Xu C, Tai L, Lin S, Han S, Tan Q, Wang X, Xu T, Zhang H, Chu X, Yi C, Liu P, Wang X, Zhou Y, Pin J- P, Rondard P, Liu H, Liu J, et al. 2021. Structures of human mGlu2 and mGlu7 homo- and heterodimers. Nature 594:589–593. DOI: https://doi.org/10.1038/s41586-021-03641-w Emsley P, Cowtan K. 2004. Coot: model- building tools for molecular graphics. Acta Crystallographica. Section D, Biological Crystallography 60:2126–2132. DOI: https://doi.org/10.1107/S0907444904019158, PMID: 15572765 Galvani S, Sanson M, Blaho VA, Swendeman SL, Obinata H, Conger H, Dahlbäck B, Kono M, Proia RL, Smith JD, Hla T. 2015. HDL- bound sphingosine 1- phosphate acts as a biased agonist for the endothelial cell receptor S1P1 to limit vascular inflammation. Science Signaling 8:ra79. DOI: https://doi.org/10.1126/scisignal.aaa2581, PMID: 26268607 Grigorova IL, Panteleev M, Cyster JG. 2010. Lymph node cortical sinus organization and relationship to lymphocyte egress dynamics and antigen exposure. PNAS 107:20447–20452. DOI: https://doi.org/10.1073/ pnas.1009968107, PMID: 21059923 Chen et al. eLife 2023;12:e88204. DOI: https://doi.org/10.7554/eLife.88204 13 of 16 Structural Biology and Molecular Biophysics Research article Hanson MA, Roth CB, Jo E, Griffith MT, Scott FL, Reinhart G, Desale H, Clemons B, Cahalan SM, Schuerer SC, Sanna MG, Han GW, Kuhn P, Rosen H, Stevens RC. 2012. Crystal structure of a lipid G protein- coupled receptor. Science 335:851–855. DOI: https://doi.org/10.1126/science.1215904, PMID: 22344443 Inoue A, Ishiguro J, Kitamura H, Arima N, Okutani M, Shuto A, Higashiyama S, Ohwada T, Arai H, Makide K, Aoki J. 2012. TGFα shedding assay: an accurate and versatile method for detecting GPCR activation. Nature Methods 9:1021–1029. DOI: https://doi.org/10.1038/nmeth.2172, PMID: 22983457 Jenne CN, Enders A, Rivera R, Watson SR, Bankovich AJ, Pereira JP, Xu Y, Roots CM, Beilke JN, Banerjee A, Reiner SL, Miller SA, Weinmann AS, Goodnow CC, Lanier LL, Cyster JG, Chun J. 2009. T- bet- dependent S1P5 expression in NK cells promotes egress from lymph nodes and bone marrow. The Journal of Experimental Medicine 206:2469–2481. DOI: https://doi.org/10.1084/jem.20090525, PMID: 19808259 Kappos L, Radue E- W, O’Connor P, Polman C, Hohlfeld R, Calabresi P, Selmaj K, Agoropoulou C, Leyk M, Zhang- Auberson L, Burtin P. 2010. A placebo- controlled trial of oral fingolimod in relapsing multiple sclerosis. New England Journal of Medicine 362:387–401. DOI: https://doi.org/10.1056/NEJMoa0909494 Kimura MY, Hayashizaki K, Tokoyoda K, Takamura S, Motohashi S, Nakayama T. 2017. Crucial role for CD69 in allergic inflammatory responses: CD69- myl9 system in the pathogenesis of airway inflammation. Immunological Reviews 278:87–100. DOI: https://doi.org/10.1111/imr.12559, PMID: 28658550 Laidlaw BJ, Gray EE, Zhang Y, Ramírez- Valle F, Cyster JG. 2019. Sphingosine- 1- Phosphate receptor 2 restrains egress of γδ T cells from the skin. The Journal of Experimental Medicine 216:1487–1496. DOI: https://doi.org/ 10.1084/jem.20190114, PMID: 31160320 Liang Y- L, Khoshouei M, Deganutti G, Glukhova A, Koole C, Peat TS, Radjainia M, Plitzko JM, Baumeister W, Miller LJ, Hay DL, Christopoulos A, Reynolds CA, Wootten D, Sexton PM. 2018. Cryo- Em structure of the active, Gs- protein complexed, human CGRP receptor. Nature 561:492–497. DOI: https://doi.org/10.1038/ s41586-018-0535-y Liu S, Paknejad N, Zhu L, Kihara Y, Ray M, Chun J, Liu W, Hite RK, Huang XY. 2022. Differential activation mechanisms of lipid GPCRs by lysophosphatidic acid and sphingosine 1- phosphate. Nature Communications 13:731. DOI: https://doi.org/10.1038/s41467-022-28417-2 Lu E, Wolfreys FD, Muppidi JR, Xu Y, Cyster JG. 2019. S- geranylgeranyl- L- glutathione is a ligand for human B cell- confinement receptor P2RY8. Nature 567:244–248. DOI: https://doi.org/10.1038/s41586-019-1003-z, PMID: 30842656 Maeda S, Koehl A, Matile H, Hu H, Hilger D, Schertler GFX, Manglik A, Skiniotis G, Dawson RJP, Kobilka BK. 2018. Development of an antibody fragment that stabilizes GPCR/G- protein complexes. Nature Communications 9:3712. DOI: https://doi.org/10.1038/s41467-018-06002-w, PMID: 30213947 Mao C, Shen C, Li C, Shen D- D, Xu C, Zhang S, Zhou R, Shen Q, Chen L- N, Jiang Z, Liu J, Zhang Y. 2020. Cryo- Em structures of inactive and active GABAB receptor. Cell Research 30:564–573. DOI: https://doi.org/10. 1038/s41422-020-0350-5 Mendoza A, Fang V, Chen C, Serasinghe M, Verma A, Muller J, Chaluvadi VS, Dustin ML, Hla T, Elemento O, Chipuk JE, Schwab SR. 2017. Lymphatic endothelial S1P promotes mitochondrial function and survival in naive T cells. Nature 546:158–161. DOI: https://doi.org/10.1038/nature22352, PMID: 28538737 Moo EV, van Senten JR, Bräuner- Osborne H, Møller TC. 2021. Arrestin- Dependent and -independent internalization of G protein- coupled receptors: methods, mechanisms, and implications on cell signaling. Molecular Pharmacology 99:242–255. DOI: https://doi.org/10.1124/molpharm.120.000192, PMID: 33472843 Morales- Perez CL, Noviello CM, Hibbs RE. 2016. Manipulation of subunit stoichiometry in heteromeric membrane proteins. Structure 24:797–805. DOI: https://doi.org/10.1016/j.str.2016.03.004, PMID: 27041595 Moriyama S, Takahashi N, Green JA, Hori S, Kubo M, Cyster JG, Okada T. 2014. Sphingosine- 1- phosphate receptor 2 is critical for follicular helper T cell retention in germinal centers. The Journal of Experimental Medicine 211:1297–1305. DOI: https://doi.org/10.1084/jem.20131666, PMID: 24913235 Nakayama T, Kasprowicz DJ, Yamashita M, Schubert LA, Gillard G, Kimura M, Didierlaurent A, Koseki H, Ziegler SF. 2002. The generation of mature, single- positive thymocytes in vivo is dysregulated by CD69 blockade or overexpression. Journal of Immunology 168:87–94. DOI: https://doi.org/10.4049/jimmunol.168.1. 87, PMID: 11751950 Niwa H, Yamamura K, Miyazaki J. 1991. Efficient selection for high- expression transfectants with a novel eukaryotic vector. Gene 108:193–199. DOI: https://doi.org/10.1016/0378-1119(91)90434-d, PMID: 1660837 Oo ML, Thangada S, Wu MT, Liu CH, Macdonald TL, Lynch KR, Lin CY, Hla T. 2007. Immunosuppressive and anti- angiogenic sphingosine 1- phosphate receptor- 1 agonists induce ubiquitinylation and proteasomal degradation of the receptor. The Journal of Biological Chemistry 282:9082–9089. DOI: https://doi.org/10. 1074/jbc.M610318200, PMID: 17237497 Pappu R, Schwab SR, Cornelissen I, Pereira JP, Regard JB, Xu Y, Camerer E, Zheng Y- W, Huang Y, Cyster JG, Coughlin SR. 2007. Promotion of lymphocyte egress into blood and lymph by distinct sources of sphingosine- 1- phosphate. Science 316:295–298. DOI: https://doi.org/10.1126/science.1139221, PMID: 17363629 Pettersen EF, Goddard TD, Huang CC, Couch GS, Greenblatt DM, Meng EC, Ferrin TE. 2004. UCSF chimera--a visualization system for exploratory research and analysis. Journal of Computational Chemistry 25:1605–1612. DOI: https://doi.org/10.1002/jcc.20084, PMID: 15264254 Pettersen EF, Goddard TD, Huang CC, Meng EC, Couch GS, Croll TI, Morris JH, Ferrin TE. 2021. UCSF chimerax: structure visualization for researchers, educators, and developers. Protein Science 30:70–82. DOI: https://doi.org/10.1002/pro.3943, PMID: 32881101 Chen et al. eLife 2023;12:e88204. DOI: https://doi.org/10.7554/eLife.88204 14 of 16 Structural Biology and Molecular Biophysics Research article Punjani A, Rubinstein JL, Fleet DJ, Brubaker MA. 2017. CryoSPARC: algorithms for rapid unsupervised cryo- EM structure determination. Nature Methods 14:290–296. DOI: https://doi.org/10.1038/nmeth.4169, PMID: 28165473 Reeves PM, Kang YL, Kirchhausen T. 2016. Endocytosis of ligand- activated sphingosine 1- phosphate receptor 1 mediated by the clathrin- pathway. Traffic 17:40–52. DOI: https://doi.org/10.1111/tra.12343, PMID: 26481905 Rohou A, Grigorieff N. 2015. CTFFIND4: fast and accurate defocus estimation from electron micrographs. Journal of Structural Biology 192:216–221. DOI: https://doi.org/10.1016/j.jsb.2015.08.008, PMID: 26278980 Rosen H, Stevens RC, Hanson M, Roberts E, Oldstone MBA. 2013. Sphingosine- 1- Phosphate and its receptors: structure, signaling, and influence. Annual Review of Biochemistry 82:637–662. DOI: https://doi.org/10.1146/ annurev-biochem-062411-130916, PMID: 23527695 Russell FA, King R, Smillie SJ, Kodji X, Brain SD. 2014. Calcitonin gene- related peptide: physiology and pathophysiology. Physiological Reviews 94:1099–1142. DOI: https://doi.org/10.1152/physrev.00034.2013, PMID: 25287861 Schindelin J, Arganda- Carreras I, Frise E, Kaynig V, Longair M, Pietzsch T, Preibisch S, Rueden C, Saalfeld S, Schmid B, Tinevez J- Y, White DJ, Hartenstein V, Eliceiri K, Tomancak P, Cardona A. 2012. Fiji: an open- source platform for biological- image analysis. Nature Methods 9:676–682. DOI: https://doi.org/10.1038/nmeth.2019, PMID: 22743772 Schwab SR, Pereira JP, Matloubian M, Xu Y, Huang Y, Cyster JG. 2005. Lymphocyte sequestration through S1P lyase inhibition and disruption of S1P gradients. Science 309:1735–1739. DOI: https://doi.org/10.1126/science. 1113640, PMID: 16151014 Shiow LR, Rosen DB, Brdicková N, Xu Y, An J, Lanier LL, Cyster JG, Matloubian M. 2006. Cd69 acts downstream of interferon- alpha/beta to inhibit S1P1 and lymphocyte egress from lymphoid organs. Nature 440:540–544. DOI: https://doi.org/10.1038/nature04606, PMID: 16525420 Spiegel S, Milstien S. 2003. Sphingosine- 1- Phosphate: an enigmatic signalling lipid. Nature Reviews. Molecular Cell Biology 4:397–407. DOI: https://doi.org/10.1038/nrm1103, PMID: 12728273 Spiegel S, Maczis MA, Maceyka M, Milstien S. 2019. New insights into functions of the sphingosine- 1- phosphate transporter SPNS2. Journal of Lipid Research 60:484–489. DOI: https://doi.org/10.1194/jlr.S091959, PMID: 30655317 Velazhahan V, Ma N, Pándy- Szekeres G, Kooistra AJ, Lee Y, Gloriam DE, Vaidehi N, Tate CG. 2021. Structure of the class D GPCR Ste2 dimer coupled to two G proteins. Nature 589:148–153. DOI: https://doi.org/10.1038/ s41586-020-2994-1, PMID: 33268889 Wagner T, Merino F, Stabrin M, Moriya T, Antoni C, Apelbaum A, Hagel P, Sitsel O, Raisch T, Prumbaum D, Quentin D, Roderer D, Tacke S, Siebolds B, Schubert E, Shaikh TR, Lill P, Gatsogiannis C, Raunser S. 2019. SPHIRE- cryolo is a fast and accurate fully automated particle picker for cryo- EM. Communications Biology 2:218. DOI: https://doi.org/10.1038/s42003-019-0437-z, PMID: 31240256 Watterson KR, Johnston E, Chalmers C, Pronin A, Cook SJ, Benovic JL, Palmer TM. 2002. Dual regulation of EDG1/S1P(1) receptor phosphorylation and internalization by protein kinase C and G- protein- coupled receptor kinase 2. The Journal of Biological Chemistry 277:5767–5777. DOI: https://doi.org/10.1074/jbc.M110647200, PMID: 11741892 Williams CJ, Headd JJ, Moriarty NW, Prisant MG, Videau LL, Deis LN, Verma V, Keedy DA, Hintze BJ, Chen VB, Jain S, Lewis SM, Arendall WB, Snoeyink J, Adams PD, Lovell SC, Richardson JS, Richardson DC. 2018. MolProbity: more and better reference data for improved all- atom structure validation. Protein Science 27:293– 315. DOI: https://doi.org/10.1002/pro.3330, PMID: 29067766 Willinger T, Ferguson SM, Pereira JP, De Camilli P, Flavell RA. 2014. Dynamin 2- dependent endocytosis is required for sustained S1PR1 signaling. The Journal of Experimental Medicine 211:685–700. DOI: https://doi. org/10.1084/jem.20131343, PMID: 24638168 Xu Z, Ikuta T, Kawakami K, Kise R, Qian Y, Xia R, Sun M- X, Zhang A, Guo C, Cai X- H, Huang Z, Inoue A, He Y. 2022. Structural basis of sphingosine- 1- phosphate receptor 1 activation and biased agonism. Nature Chemical Biology 18:281–288. DOI: https://doi.org/10.1038/s41589-021-00930-3 Yang X, Wang X, Xu Z, Wu C, Zhou Y, Wang Y, Lin G, Li K, Wu M, Xia A, Liu J, Cheng L, Zou J, Yan W, Shao Z, Yang S. 2022. Molecular mechanism of allosteric modulation for the cannabinoid receptor CB1. Nature Chemical Biology 18:831–840. DOI: https://doi.org/10.1038/s41589-022-01038-y Yu L, He L, Gan B, Ti R, Xiao Q, Hu H, Zhu L, Wang S, Ren R. 2022. Structural insights into sphingosine- 1- phosphate receptor activation. PNAS 119:e2117716119. DOI: https://doi.org/10.1073/pnas.2117716119, PMID: 35412894 Yuan Y, Jia G, Wu C, Wang W, Cheng L, Li Q, Li Z, Luo K, Yang S, Yan W, Su Z, Shao Z. 2021. Structures of signaling complexes of lipid receptors s1pr1 and s1pr5 reveal mechanisms of activation and drug recognition. Cell Research 31:1263–1274. DOI: https://doi.org/10.1038/s41422-021-00566-x, PMID: 34526663 Zachariah MA, Cyster JG. 2010. Neural crest- derived pericytes promote egress of mature thymocytes at the corticomedullary junction. Science 328:1129–1135. DOI: https://doi.org/10.1126/science.1188222, PMID: 20413455 Zhao C, Cheng L, Wang W, Wang H, Luo Y, Feng Y, Wang X, Fu H, Cai Y, Yang S, Fu P, Yan W, Shao Z. 2022. Structural insights into sphingosine- 1- phosphate recognition and ligand selectivity of s1pr3- gi signaling complexes. Cell Research 32:218–221. DOI: https://doi.org/10.1038/s41422-021-00567-w, PMID: 34545189 Zheng SQ, Palovcak E, Armache JP, Verba KA, Cheng Y, Agard DA. 2017. MotionCor2: anisotropic correction of beam- induced motion for improved cryo- electron microscopy. Nature Methods 14:331–332. DOI: https://doi. org/10.1038/nmeth.4193, PMID: 28250466 Chen et al. eLife 2023;12:e88204. DOI: https://doi.org/10.7554/eLife.88204 15 of 16 Structural Biology and Molecular Biophysics Research article Ziegler SF, Ramsdell F, Alderson MR. 1994. The activation antigen CD69. Stem Cells 12:456–465. DOI: https:// doi.org/10.1002/stem.5530120502, PMID: 7804122 Zivanov J, Nakane T, Forsberg BO, Kimanius D, Hagen WJ, Lindahl E, Scheres SH. 2018. New tools for automated high- resolution cryo- EM structure determination in RELION- 3. eLife 7:e42166. DOI: https://doi.org/ 10.7554/eLife.42166, PMID: 30412051 Chen et al. eLife 2023;12:e88204. DOI: https://doi.org/10.7554/eLife.88204 16 of 16 Structural Biology and Molecular Biophysics
10.7554_elife.85878
RESEARCH ARTICLE Two RNA- binding proteins mediate the sorting of miR223 from mitochondria into exosomes Liang Ma, Jasleen Singh, Randy Schekman* Department of Molecular and Cell Biology, Howard Hughes Medical Institute, University of California, Berkeley, United States Abstract Fusion of multivesicular bodies (MVBs) with the plasma membrane results in the secre- tion of intraluminal vesicles (ILVs), or exosomes. The sorting of one exosomal cargo RNA, miR223, is facilitated by the RNA- binding protein, YBX1 (Shurtleff et al., 2016). We found that miR223 specif- ically binds a ‘cold shock’ domain (CSD) of YBX1 through a 5’ proximal sequence motif UCAGU that may represent a binding site or structural feature required for sorting. Prior to sorting into exosomes, most of the cytoplasmic miR223 resides in mitochondria. An RNA- binding protein local- ized to the mitochondrial matrix, YBAP1, appears to serve as a negative regulator of miR223 enrich- ment into exosomes. miR223 levels decreased in the mitochondria and increased in exosomes after loss of YBAP1. We observed YBX1 shuttle between mitochondria and endosomes in live cells. YBX1 also partitions into P body granules in the cytoplasm (Liu et al., 2021). We propose a model in which miR223 and likely other miRNAs are stored in mitochondria and are then mobilized by YBX1 to cyto- plasmic phase condensate granules for capture into invaginations in the endosome that give rise to exosomes. Editor's evaluation This important study presents a novel mechanism of miRNA223 sorting into exosomes involving its storage within mitochondria, specifically by a mitochondrially localized protein YBAP1. The evidence supporting the findings is convincing and opens avenues for future studies on molecular mecha- nisms. This paper is a valuable addition to the cellular sorting of miRNA involving interplay with and between the organelles, interesting for miRNAs researchers, as well as cell biologists. Introduction Extracellular vesicles (EVs) bud from the plasma membrane or are secreted when multivesicular bodies (MVB) fuse with the plasma membrane to release a population of vesicles called exosomes. EVs and their cargos are highly dependent on their membrane source. Microvesicles released by budding from the plasma membrane are a heterogeneous population of EVs ranging in size from 30 nm to 1000 nm (Cocucci et al., 2009). Exosomes are 30 nm to 150 nm in size and originate as vesicles invaginated into the interior of an MVB to form intraluminal vesicles (ILVs; Harding et al., 1983). Many RNAs are selectively sorted into EVs, especially small RNAs. Several studies have indicated that RNA binding proteins (RNPs) may be involved in the enrichment of RNAs into EVs (Mukherjee et al., 2016; Santangelo et al., 2016; Teng et al., 2017; Villarroya- Beltri et al., 2013). However, many of these studies used sedimentation at  ~100,000  g to collect EVs, which may also collect RNP particles not enclosed within membranes which complicates the interpretation of these data. To address this question, we previously developed buoyant density- based methods to separate EVs *For correspondence: schekman@berkeley.edu Competing interest: The authors declare that no competing interests exist. Funding: See page 20 Received: 30 December 2022 Preprinted: 11 January 2023 Accepted: 24 July 2023 Published: 25 July 2023 Reviewing Editor: Agnieszka Chacinska, IMol Polish Academy of Sciences, Poland Copyright Ma et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 1 of 23 Research article from non- vesicular aggregates and found that EVs form two distinct populations of high and low buoyant density (Shurtleff et al., 2016; Temoche- Diaz et al., 2020). We found that some miRNAs are selectively enriched in a high buoyant density vesicle fraction characterized by an enrichment in the exosomal marker protein CD63, whereas the low buoyant density EVs are fairly non- selective in the capture of miRNAs (Temoche- Diaz et  al., 2019). We developed a cell- free reaction to identify YBX1 as required for miR223 sorting into exosomes and demonstrated that it plays an important role in the enrichment of miR223 into exosomes in HEK293T cells (Shurtleff et al., 2016). We subse- quently found that phase separated YBX1 condensates selectively recruit miR223 in vitro and sort it into exosomes in cells (Liu et al., 2021). In this study, we report that YBX1 directly and specifically binds miR223 by its ‘cold shock’ domain (CSD). We have identified a sequence motif, UCAGU, that facilitates the sorting of miR223 into exosomes. We also found a significant fraction of cytoplasmic miR223 localized within mitochondria, tightly associated with the mitochondrial envelope and that a mitochondrial RNA- binding protein, YBAP1, may control the transfer of miR223 from mitochondria to exosomes. Results YBX1 directly and specifically binds miR223 We previously documented that YBX1 facilitates miR223 sorting into exosomes (Shurtleff et  al., 2016) and that exosomal miR223 is decreased in YBX1 knockout cells (Liu et al., 2021). We reexam- ined the enrichment and confirmed that exosomal miR223 was decreased in exosomes purified from YBX1 KO cells (Figure 1a). We used a Nanosight particle tracking device to quantify buoyant density purified vesicles and found that knockout of YBX1 did not affect exosome secretion (Figure 1—figure supplement 1). Whereas the importance of YBX1 for miR223 sorting has been established, the mechanism of their interaction was not known. To examine the direct interaction of YBX1 and miR223, we used an electrophoretic mobility shift assay (EMSA) with purified recombinant YBX1, expressed in insect cells (Figure 1—figure supplement 2a), and chemically synthetic miR223 and miR190, a cytoplasmic miRNA that is not enriched in exosomes. Purified YBX1 was titrated and incubated with 5’ fluorescently labeled miR223 at 30 ℃ for 30 min. miR223- YBX1 complexes were separated by electrophoresis and detected by in- gel fluorescence. The EMSA data showed that YBX1 directly and specifically bound to miR223, but ~140 fold less well with miR190 (Figure 1b–c). The measured Kd for YBX1:miR223 was 4.2 nM (Figure 1d). YBX1 has three major domains including an N- terminal alanine/proline- rich (A/P) domain, a central cold shock domain (CSD) and a C- terminal domain (CTD) (Figure 1e). To explore which specific domain of YBX1 binds miR223, we constructed a series of fragments: the A/P domain, CSD and CTD. The YBX1 fragments were expressed in and purified from insect cells (Figure 1—figure supplement 2b). EMSA data showed that the A/P domain and CTD had little or no affinity for miR223, whereas the CSD domain bound miR223 but with an affinity much reduced compared to full length YBX1 (Figure 1f). We then constructed two combined fragments of the A/P and CSD and CSD and CTD domains (Figure 1—figure supplement 2c). The EMSA data showed that the A/P domain was dispensable, whereas binding of miR223 to CSD plus CTD was comparable to full- length YBX1 (Figure 1g). YBX1- F85A in the CSD domain was reported to block the YBX1- specific binding of mRNA (Lyons et al., 2016). Purified YBX1- F85A protein failed to bind miR223 (Figure 1g, Figure 1—figure supple- ment 2c). These data suggest that YBX1 directly and specifically binds miR223 via the CSD. The CTD of YBX1 did not appear to bind miR223 but may somehow facilitate a higher affinity interaction of the CSD with miR223. A binding or structural motif on miR223 that promotes interaction with YBX1 and enrichment into exosomes We next sought to determine the miR223 sequence motif responsible for interaction with YBX1 and enrichment into exosomes. We used an EMSA competition assay with a series of miR223 mutants. Purified YBX1 and 5’ fluorescently labeled miR223 were incubated with miR223 mutant constructs titrated in a range from 1 nM to 1 μM. miR223 variants in a binding domain should not compete for interaction of YBX1 with 5’ fluorescently tagged miR223 whereas variations in sequences irrelevant to Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 2 of 23 Cell Biology Research article a. e g n a h c d o F l / ) T W O K 1 X B Y ( 8 4 2 1 0.5 0.25 d. miR223-3p miR190a-5p Cell Exosome 100 80 60 40 20 ) % ( d n u o B n o i t c a r F 0 10 0 M μ 1 [miRNA] = 1nM Kd of miR223 = 4.2nM Kd of miR190 = 574.5nM 10 1 10 2 YBX1 [ nM ] 10 3 c. M μ 0 YBX1 M μ 1 Bound miR190 Free miR190 miR223 miR190 b. M μ 0 YBX1 Bound miR223 Free miR223 YBX1 N 1 51 129 A/P CSD CTD 324 C M μ μ 0 0 YBX1(1-51) M μ 1 M μ μ 0 0 YBX1(52-129) M μ 1 YBX1(130-324) M μ μ 0 0 M μ 1 1 51 A/P 52 129 CSD 130 324 CTD YBX1(1-129) M μ μ 0 0 M μ 1 M μ μ 0 0 YBX1(52-324) M μ 1 M μ μ 0 0 YBX1(F85A) M μ 1 e. f. Bound miR223 Free miR223 g. Bound miR223 Free miR223 1 129 A/P CSD 52 324 1 51 324 CSD CTD A/P CSD CTD F85A Figure 1. YBX1 directly and specifically binds miR223. (a) RT- qPCR analysis of fold change of miR- 223 and miR- 190 in cells and purified exosomes from 293 T WT cells and YBX1 knockout cells. Data are plotted from three independent experiments, each independent experiment with triplicate qPCR reactions; error bars represent standard deviations. (b–c) EMSA assays using 1 nM 5’ fluorescently labeled miR223 or miR190 and purified YBX1. Figure 1 continued on next page Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 3 of 23 Cell Biology Research article Figure 1 continued Purified YBX1 was titrated from 500pM to 1 μM. In gel fluorescence was detected. Quantification of (d) shows the calculated Kd. (e) Schematic diagrams of the different domains of YBX1. (f) EMSA assay using 1 nM 5’ fluorescently labeled miR223 and purified YBX1 truncations. [YBX1(1–51) or YBX1(52–129) or YBX1(130–324).] (g) EMSA assay using 1 nM 5’ fluorescently labeled miR223 and purified YBX1 truncations [YBX1(1–129) or YBX1(52–324)] or YBX1(F85A) mutant. The online version of this article includes the following source data and figure supplement(s) for figure 1: Source data 1. Uncropped gel images corresponding to Figure 1. Figure supplement 1. Knockout of YBX1 did not change exosome secretion. Figure supplement 2. Purified YBX1 full length protein and different truncations and mutation. Figure supplement 2—source data 1. Uncropped gel images corresponding to Figure 1—figure supplement 2. interaction would compete. Using this EMSA competition assay to screen the miR223 binding motif, we found that the competitive binding of miR223mut (3- 6) and miR223mut (4- 7) were decreased (Figure  2—figure supplement 1). This suggested that the sequence UCAGU was critical for inter- action with YBX1. To test this directly, we employed a variant sequence, termed miR223mut, where the UCAGU was substituted with AGACA. As a positive control, we employed a variant of miR190, miR190sort, where the sequence AUAUG was substituted with UCAGU (Figure  2a). EMSA data showed a~27- fold reduced YBX1 interaction of with miR223mut, whereas the affinity of miR190sort with YBX1 was increased ~eightfold compared to wt sequences. To test whether this motif is critical for miR223 enrichment into exosomes, we purified exosomes from 293T cells transiently transfected to overexpress one of the four miRNA constructs (Figure 2d). RT- qPCR data showed that the level of miR223 in exosomes was ~fourfold dependent on the puta- tive exosomal sorting motif (Figure  3e) and the enrichment of miR190sort into exosomes was increased ~fivefold compared to miR190 WT (Figure 2f). In previous work, we developed a cell- free reaction to test the biochemical requirement for YBX1 in the sorting of miR223 into vesicles formed with membranes and cytosol isolated from broken HEK293 cells (Shurtleff et  al., 2016). In this work, we showed that the sorting of miR223 and of a CD63- luciferase fusion protein into an enclosed membrane were coincidentally inhibited by GW4869 an inhibitor of neutral sphingomyelinase (NS2) known to interfere with exosome biogenesis and secre- tion. On this basis, we concluded that the cell- free reaction recapitulated the sorting event leading to the packaging of miR223 into exosomes. We refined this assay to measure the incorporation of 32P- 5’ end- labeled wt and mutant miR223 into vesicles formed in vitro. Isolated membranes and cytosol were incubated with 32P- labled wt or mutant miR223 at 30 °C for 20 min, after which RNase I was added to digest any unpackaged miRNA. Controls including 1% Triton X- 100 were used to measure background RNase resistant radiolabel. Samples were resolved on a gel for visual and quantitative evaluation of membrane sequestered RNA (Figure 2g and h). The results suggested that the UCAGU motif is critical for miR223 packaging into vesicles in the cell- free reaction. Taken together the results in Figure  2 show that the miR223 sequence UCAGA promotes the binding of YBX1 in order to sort the miRNA into vesicles formed in cells and in a cell- free reaction. We suggest this sorting facilitates the export of miR223 in exosomes secreted from HEK293 cells. Mitochondria contribute to miR223 enrichment into exosomes In a recent study, we showed that YBX1 is sorted into P- bodies in cells and that these biomolecular condensates may initiate the sorting of miR223 into vesicles budding into the interior of endosomes (Liu et al., 2021; Shurtleff et al., 2016). Mitochondria represent another apparent intracellular loca- tion of miR223 (Wang et al., 2020). We used cell fractionation of homogenates of HEK293 cells to evaluate the subcellular distribution of endogenous miR223. Fractionation was evaluated by immu- noblot using marker proteins characteristic of various cell organelles (Figure  3a). Analysis of RNA extracted from isolated membranous organelles confirmed that miR223 but not miR190 was signifi- cantly enriched in mitochondria but not in ER or cytosol (Figure 3b, Figure 3—figure supplement 1a; Wang et al., 2020). Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 4 of 23 Cell Biology Research article a. miR223-3p UGUCAGUUUGUCAAAUACCCCA miR223mut UGAGACAUUGUCAAAUACCCCA miR190-5p UGAUAUGUUUGAUAUAUUAGGU miR190sort UGUCAGUUUUGAUAUAUUAGGU c. 100 ) % ( d n u o B n o i t c a r F 75 50 25 b. M μ 0 YBX1 M μ 1 M μ 0 YBX1 0 0.1 M μ 1 1 M μ 0 10 100 YBX1 [nM] YBX1 miR223 miR190sort miR190a miR223mut [miRNA] = 1nM Kd of miR223mut = 112.4nM Kd of miR190sort = 77.5nM 1000 10000 M μ 1 M μ 0 YBX1 M μ 1 d. miR223 Transfect plasmid which express miRNAs or mutants miR223mut miR190 miR190sort medium 1,500 g Supernatant100,000 g 10% 40% 150,000 g Exosomes 10,000 g 120,000 g 60% Sucrose cushion EV 60% e. f. p=0.0013 1.00 0.27 1.2 1.0 0.8 0.6 0.4 0.2 0.0 g. Temp (℃) RNase I Membranes Cytosol Triton(1%) miR223 3030 30 30 4 4 _ + + + + + _ + + + + + + _ + + + + _ _ _ _ + _ miR223mut 30 + + _ + + _ _ _ _ 3030 304 + + + + + + + + + _ + 4 _ + + _ l e g n a h C d o F s A N R m i l a m o s o x E ) l l e c o t e v i t a e r ( l miR223 miR223mut p=0.0194 5.33 h. 1.00 miR190 miR190sort 8 6 4 2 0 i 3 2 2 R m d e t c e t o r p % 25 20 15 10 5 0 C M l e g n a h C d o F s A N R m i l a m o s o x E ) l l e c o t e v i t a e r ( l miR223 miR223mut degree) C+M+Triton C+M C+M(4 Figure 2. miR223 sequence motif UCAGU binds YBX1. (a) RNA oligonucleotides corresponding to miR223, miR190 and versions with mutated sorting motif (miR223mut) or mutation to introduce the sorting motif (miR190sort). (b) EMSA assays using 1 nM 5’ fluorescently labeled miR223 WT or miR223mut or miR190 WT or miR190sort and purified YBX1. Purified YBX1 was titrated from 500pM to 1 µM. In gel fluorescence was detected. (c) Binding affinity curves as calculated by EMSA data from (b) (d) Schematic shows exosome purification with buoyant density flotation in a sucrose Figure 2 continued on next page Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 5 of 23 Cell Biology Research article Figure 2 continued step gradient from 293T cells overexpressing miR223 WT or mutant or miR190 WT or miR190sort. (e) RT- qPCR analysis of relative abundance of miR223 or miR223mut detected in exosomes compared to cellular level in 293T cells overexpressing miR223 WT or miR223mut. Data are plotted from three independent experiments and error bars represent standard deviations. (f) RT- qPCR analysis of relative abundance of miR190 or miR190sort detected in exosomes compared to cellular level in 293T cells overexpressing miR190 WT or miR190sort. Data are plotted from three independent experiments and error bars represent standard derivations. (g) In vitro packaging assay using 32P 5’end- labeled miR223 and miR223mut. Cell- free packaging of miR223 and miR223mut measured as protected radioactive signal from 32P labeled miR223 and miR223mut. Reactions with or without membrane, cytosol, and 1% Triton X- 100, and incubated at 4 or 30 °C are indicated. For the samples containing only cytosol plus membrane at 4 °C, only one- third of the samples were loaded. Each sample was supplemented with 300 mM urea to reduce the background signal. (h) Data quantification showed protected fraction of miR223 and miR223mut as calculated from in vitro packaging data shown in (g). The online version of this article includes the following source data and figure supplement(s) for figure 2: Source data 1. Uncropped gel images corresponding to Figure 2. Figure supplement 1. Screening of exosomal sorting motif of miR223. Figure supplement 1—source data 1. Uncropped gel images corresponding to Figure 2—figure supplement 1. To determine the localization of miR223 on or within mitochondria, we prepared mitoplasts using digitonin to strip away the mitochondrial outer membrane followed by fractionation on a Percoll density gradient. Immunoblots of the enriched mitochondria and isolated mitoplasts showed that the outer membrane, marked by Tom20, was largely removed with retention of the inner membrane marker Tim23 (Figure 3c). RNA was extracted from the purified mitoplast and RT- qPCR data indicated that miR223 was enriched along with mRNA for COX1, but not with nuclear U6 snRNA (Figure 3d). As an independent means to assess the localization of cytoplasmic miR223, we used immunopre- cipitation to purify mitochondria. Isolated mitochondria were then converted to mitoplasts by osmotic shock and treated with proteinase K and RNase. Immunoblots of the immunoprecipitated mitochon- dria and isolated mitoplasts showed that the outer membrane, marked by Tom20, and intermembrane space, marked by AIF, were largely removed with retention of the mitochondrial matrix marker citrate synthase (Figure 3e). RNA was extracted from the immunoprecipitated mitochondria and mitoplasts and RT- qPCR data indicated that miR223 was enriched along with mRNA for COX1, but not with miR190 or nuclear U6 snRNA (Figure 3f). We also used immunoprecipitated mitochondria (Figure  3—figure supplement 1b) and either Triton X- 100 to solubilize the membrane or freeze- thaw to allow the matrix and envelope fractions to be separated by centrifugation. Mitochondrial membrane proteins, such as Tom20 and COX IV, were solubilized and retained in the supernatant fraction (Figure  3—figure supplement 1c). The freeze- thaw regimen released citrate synthase to a supernatant fraction whereas Tom20 and COX IV sedimented in the pellet fraction. RNA was extracted from the detergent supernatant and pellet fractions where we found similar distributions of COX1 and miR223, neither of which were as readily solubilized as the inner and outer membrane proteins (Figure 3—figure supplement 1d). RT- qPCR quantification of fractions from the freeze- thaw regimen showed that both COX1 mRNA and miR223 remained largely associated with the sedimentable membrane fraction (Figure 3—figure supplement 1e). We conclude that miR223 is enclosed within mitochondria, possibly in association with the inner membrane. We sought a test of the role of mitochondria in the secretion of miR223 in exosomes. For this purpose, we generated cells depleted of mitochondria (Correia- Melo et  al., 2017). U- 2 OS cells expressing GFP- parkin were treated with CCCP for 48 hr, conditions that cause mitochondria to be removed by mitophagy. We confirmed mitochondrial depletion after CCCP treatment by RT- qPCR of mitochondrial COX1 mRNA (Figure  3g) and immunoblot of the mitochondrial inner membrane marker Tim23 (Figure 3h). We then compared the levels of both miR223 and miR190 from GFP- parkin expressing U- 2 OS cells with and without CCCP treatment. RT- qPCR data showed that miR223, but not miR190, increased threefold in cells treated with CCCP (Figure  3j). To test the possibility that miR223 accumulated in cells as a result of a failure of mobilization into exosomes, we compared the miR223 levels in exosomes purified from untreated and CCCP treated cells (Figure  3j). Although exosome secretion, as measured with a CD63- luciferase marker, did not change after CCCP treatment (Figure 3—figure supplement 2a), we found that CCCP treatment lowered the amount of miR223 in EVs fourfold (Figure 3j). The increase in cellular at the expense of exosomal miR223 may reflect a critical role for mitochondria in the mobilization of this RNA to exosomes. Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 6 of 23 Cell Biology c. d. l l e C T M P M ⍺-Tim23 ⍺-Tom20 ⍺-Tubulin ⍺-PDI t n e m h c i r n E A N R 10 8 6 4 2 0 COX1 miR223 U6 snRNA cell mitoplast cell mitoplast cell mitoplast COX1 miR223 miR190 U6 snRNA IP-MT Cell IP-MP IP-MT Cell IP-MP IP-MT IP-MP Cell IP-MT Cell IP-MP + CCCP Exosome Free Medium + Uridine CCCP removed exosome Cell Exosome Research article a. l l e C o t y C o t i M R E b. t n e m h c i r n E 3 2 2 R m i ⍺-Tim23 ⍺-Calnexin ⍺-Tubulin ) l l e c o t e v i t a e r ( l 40 30 20 10 0 Cyto ER Mito * f. t n e m h c i r n E A N R 20 15 10 5 0 ⍺-Citrate Synthase ⍺-LAMP1 ⍺-Tom20 ⍺-AIF ⍺-GRP78 ⍺-GAPDH h. P=0.0004 i. ⍺-Tim23 ⍺-GAPDH Parkin-GFP e. g. l l e c n i A N R m 1 X O C f o % 120 100 80 60 40 20 0 NC 24uM CCCP j. e g n a h c d o F l ) / C N P C C C ( 8 4 2 1 0.5 0.25 0.125 miR223 miR190 Figure 3. Mitochondria contribute to miR223 enrichment into exosomes. (a) Immunoblot analysis of protein markers for different subcellular fractions isolated from 293T cells. (b) RT- qPCR analysis of miR223 fold changes of different subcellular fractions isolated from 293T cells relative to cell lysate. (c) Immunoblot analysis of protein markers for mitoplasts purified from 293T cells by Percoll gradient fractionation (MT: mitochondria; MP: mitoplast). (d) RT- analysis of COX1 mRNA, miR223 and U6 snRNA fold changes for mitoplasts purified from 293T cells relative to cell lysate. (e) Immunoblot analysis Figure 3 continued on next page Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 7 of 23 Cell Biology Research article Figure 3 continued of protein markers for immunoprecipitated mitochondria and osmotic shock generated mitoplasts. Mitochondria were purified from a 293T 3xHA- EGFP- OMP25 overexpressing cell line using anti- HA magnetic beads. Mitoplasts were purified following mitochondrial immunoprecipitation by osmotic shock, proteinase K and RNase treatment (IP- MT: immunoprecipitated mitochondria; IP- MP: immunoprecipitated mitoplasts). (f) RT- analysis of COX1 mRNA, miR223, miR190 and U6 snRNA fold changes for immunoprecipitaed mitochondria and mitoplasts purified from the 293T 3xHA- EGFP- OMP25 overexpressing cell line. Data are plotted from three independent experiments and error bars represent standard deviations. (g) RT- qPCR analysis of mitochondrial mRNA COX1 in U2OS cells expressing GFP- Parkin treated with or without CCCP. Data are plotted from three independent experiments and error bars represent standard deviations. (h) Immunoblot analysis of mitochondrial marker Tim23 in U2OS cells expressing GFP- Parkin treated with or without CCCP. (i) Schematic of exosome purification from mitochondria depleted GFP- Parkin expressing U2OS cells. (j) RT- qPCR analysis of fold change of miR- 223 and miR- 190 in cells and purified exosomes from U2OS cells expressing GFP- Parkin which were treated with or without CCCP. Data are plotted from three independent experiments and error bars represent standard deviations. The online version of this article includes the following source data and figure supplement(s) for figure 3: Source data 1. Uncropped immunoblot images corresponding to Figure 3. Figure supplement 1. miR223, but not miR190, enriched in mitochondria. Figure supplement 1—source data 1. Uncropped immunoblot images corresponding to Figure 3—figure supplement 1. Figure supplement 2. Mitochondrial depletion did not change exosome secretion. YBAP1 binds miR223 in the mitochondria and in vitro In the course of purifying a tagged version of YBX1 from 293T cells, we observed another protein that copurified and found that it corresponded to YBAP1 (Figure 4a and b). Such a complex of YBX1 and YBAP1 has previously been reported (Matsumoto et al., 2005). We confirmed that purified YBX1 and YBAP1 bind each other by coexpression and affinity purification from insect cells (Figure 4—figure supplement 1b). YBAP1 is a mitochondrial matrix protein with a standard N- terminal transit peptide sequence (Muta et al., 1997). We confirmed this mitochondrial localization in U- 2 OS cells expressing Tom22- mCherry transiently transfected with a YBAP1- GFP construct (Figure 4c–d). We also showed that YBAP1 is localized within mitochondria by performing a proteinase K protection assay on purified mitochondria. Mitochondria were isolated from non- transfected cells and exposed to proteinase K in the presence or absence of Triton X- 100 and the degradation of YBAP1 was evaluated by immuno- blot. YBAP1 was resistant to proteinase K digestion as was the mitochondrial inner membrane marker Tim23. Both were degraded by proteinase treatment in the presence of Triton X- 100 (Figure  4e), consistent with the localization of YBAP1 within mitochondria. To test whether YBAP1 was bound to miR223 in mitochondria, we used YBAP1 immunoprecipita- tion with mitochondria purified by fractionation on a Percoll density gradient (Figure 4f). RT qPCR data showed that mitochondrial miR223 was immunoprecipitated by YBAP1 antibody but not by a control antibody (Figure 4g). To determine whether the YBAP1 and miR223 interaction was direct, we used the EMSA assay and found that purified YBAP1 bound miR223, but not miR190. The YBAP1 interaction with miR223 was not dependent on the RNA sequence motif responsible for YBX1 binding (Figure 4—figure supplement 2a–b). Taken together, these data suggest that YBAP1 binds miR223 in mitochondria and in vitro. YBAP1 may control the transit of miR223 from mitochondria to exosomes To investigate the function of YBAP1 in the transit of miR223 into exosomes, we generated a 293T YBAP1 KO cell line and compared the level of miR223 enrichment in exosomes and mitochondria isolated from WT and mutant cells (Figure 5a and b). Although knockout of YBAP1 did not change exosome secretion (Figure  5—figure supplement 1a), RT- qPCR analysis showed that miR223 decreased twofold in mitochondria but increased eightfold in exosomes purified from mutant and WT cells, respectively (Figure 5c). This apparent inverse relationship is consistent with a role for YBAP1 protein in the retention of miR223 in mitochondria. YBX1 puncta shuttled from mitochondria to endosomes In previous work we reported the localization of YBX1 to P- bodies and suggested this may repre- sent an intermediate stage in the concentrative sorting of miRNAs for secretion in exosomes (Liu et al., 2021). In other earlier work, P- bodies were seen in association with mitochondria (Huang et al., Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 8 of 23 Cell Biology Research article a. b. d. e u a v l YBX1 Copurified Band ⍺-YBAP1 c. y a r G d e z i l a m r o N 1.5 1.0 0.5 0.0 0 Tom22 YBAP1 100 50 Distance (pixels) 150 Tom22-mC e. e. Mitochondria Protease K Triton X-100 + - - + + - + + + 10um YBAP1-GFP f. Input IP ⍺-Tim23 ⍺-Tom20 ⍺-YBAP1 ⍺-YBAP1 ⍺-Tom20 h. Bound miR223 Free miR223 i. M μ μ 0 0 YBAP1 M μ 1 M μ μ 0 0 YBAP1 M μ 1 Bound miR190 Free miR190 merge g. 3 2 2 R m i f o t u p n I % 60 40 20 0 j. 100 ) % ( d n o B n o i t c a r F 75 50 25 P<0.0001 mIgGIP YBAP1IP miR223-3p miR190a-5p 0 0.1 1 10 100 YBAP1 [nM] 1000 10000 [miRNA] = 1nM Kd of miR223 = 173.2nM Figure 4. YBAP1 directly and specifically binds miR223. (a) Strep II- YBX1 was overexpressed in HEK293T cells. Coomassie blue detection of unknown band copurified with YBX1 from 293T cells. (b) Immunoblot identified unknown band was YBAP1. (c) Tom22- mCherry expressing U2OS was transfected with a YBAP1- GFP- expressing plasmid, cultured for 12 hr and observed by confocal microscopy. Scale bar, 10 μm. (d) Quantification of the fluorescence intensity of the different channels indicated by the solid white line of (c). (e) YBAP1 resides in mitochondria. Proteinase K protection assay for YBAP1 Figure 4 continued on next page Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 9 of 23 Cell Biology Research article Figure 4 continued using purified mitochondria from 293T cells. Samples were treated with or without proteinase K (10 μg/ml) and or Triton X- 100 (0.5%). Immunoblots for Tim23, Tom20, and YBAP1 are shown. (f) Mitochondria were purified for immunoprecipitation with YBAP1 antibody. Immunoblot detection of YBAP1 and Tom20. (g) RT- qPCR analysis of miR223 fold changes of YBAP1 IP samples. Data are plotted from three independent experiments and error bars represent standard deviations. (h–i) EMSA assays using 1 nM 5’ fluorescently labeled miR223 or miR190. Purified YBAP1 was titrated from 500pM to 1 μM. In gel fluorescence was detected. Quantification of (j) shown the calculated Kd. The online version of this article includes the following source data and figure supplement(s) for figure 4: Source data 1. Uncropped immunoblot and gel images corresponding to Figure 4. Figure supplement 1. YBX1 and YBAP1 copurify as a complex from transfected SF9 cells. Figure supplement 1—source data 1. Uncropped gel images corresponding to Figure 4—figure supplement 1. Figure supplement 2. YBAP1 does not share the same miR223- binding motif as YBX1. Figure supplement 2—source data 1. Uncropped gel images corresponding to Figure 4—figure supplement 2. 2011). To explore this possibility, we visualized endogenous YBX1 and YBAP1 by IF and observed YBX1 puncta colocalized with mitochondria (Figure  6a and b). In order to detect the proximity of endosomes to this point of contact between YBX1 puncta and mitochondria, we used U- 2 OS cells transfected with Rab5(Q79L)- mCherry, which we employed previously to enlarge and detect the inter- nalization of YBX1 into endosomes (Liu et al., 2021). We then used three color visualization of the U- 2 OS cells also transfected with YFP- YBX1 and mito- BFP. Time- lapse imaging showed YBX1 puncta in close proximity to mitochondria or endosomes, followed quickly by transfer between them (Figure 6c and d). Taken together, these data suggest a mechanism whereby miR223 stored in mitochondria, possibly sequestered by YBAP1, may be captured in a tighter interaction with YBX1 in P- bodies and delivered to endosomes for sorting and secretion in exosomes. Discussion Selected miRNAs are sorted, some with very high fidelity, into invaginations in the endosome that give rise to exosomes secreted from cultured human cells and likely from many if not all cells in metazoan organisms. The means by which these miRNAs are sorted and the possible extracellular functions they serve is a subject of interest in normal and disease physiology. Here we report the role of the RNA- binding protein YBX1 and a sorting or structural signal on one target RNA, miR223, and the indirect path miR223 may take from storage in mitochondria into exosomes. We have identified a sequence motif on miR223, UCAGU, responsible for high- affinity interaction with YBX and for sorting into vesicles formed in a cell- free reaction as well as for secretion in exosomes by HEK293 cells. Previously we performed this in vitro packaging assay in the presence of an inhibitor (GW4869) of neutral sphingomyelinase (NS2). This inhibitor has been shown to reduce the secretion of exosomes and exosome- associated miRNAs in other studies (Li et al., 2013; Trajkovic et al., 2008; Yuyama et  al., 2012). In our cell- free assay, GW4869 inhibited the protection of CD63- luciferase and miR- 223 at concentrations known to inhibit the activity of NS2 in partially purified enzyme frac- tions (Shurtleff et al., 2016). We concluded that our cell- free reaction provides a model that mimics aspects of exosome biogenesis. The YBX1 protein has three distinct domains, one of which, the cold- shock domain (CSD) appears to be the principal site for RNA binding, including at least one critical residue, F85, required for binding miR223 as well as other RNAs (Lyons et al., 2016). The C- terminal domain (CTD) includes an intrinsically disordered domain (IDR) that promotes the formation of a liquid- liquid phase separation likely responsible for the organization of YBX1 in P- bodies (Liu et al., 2021). This domain does not itself interact with RNA, but it appears to facilitate the folding or stabilization of the CSD to promote high affinity binding to miR223. In other work using a similar approach, we identified two separate sorting signals, a 5’UGGA and a 3’UUU, on miR122 to which the RNA- binding protein La binds en route to secretion in exosomes by the breast cancer cell line MDA- MB- 231 (Temoche- Diaz et al., 2019). Other distinct sorting signals and their cognate RNA- binding proteins have been documented in different cell lines. miRNAs with a GGAG sorting motif recognized by a sumolyated form of hnRNPA2B1 was shown to be enriched in exosomes (Villarroya- Beltri et al., 2013). Another sequence, AAUGC, was found to be enriched in Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 10 of 23 Cell Biology Research article a. Mitochondria Cell medium RNA Extraction and RT-qPCR WT or YBAP1-KO Exosome b. c. e g n a h c d o F l / ) T W O K 1 P A B Y ( 16 8 4 2 1 0.5 0.25 ⍺-YBAP1 ⍺-Citrate Synthase ⍺-COX IV ⍺-Tim23 ⍺-AIF ⍺-Tom20 ⍺-YBX1 ⍺-Tubulin Cell Exosome mitochondria miR223 miR190 Figure 5. YBAP1 sequesters miR223 which is released and secreted in YBAP1 KO cells. (a) Schematic shows exosome and mitochondria purification from 293 T WT cells and YBAP1 knock out cells for RT- qPCR analysis. (b) Analysis of 293 T WT and CRISPR/Cas9 genome- edited cells by immunoblot for YBAP1, YBX1 and mitochondrial markers (c) RT- qPCR analysis of miR223 enrichment in mitochondria purified from 293 T WT cells and YBAP1 KO cells relative to cell lysate. Data are plotted from three independent experiments and error bars represent standard deviations. (d) RT- qPCR analysis of miR223 and miR190 fold change in cells, purified mitochondria and exosomes from 293 T WT cells and YBAP1 KO cells. Data are plotted from three independent experiments and error bars represent standard deviations. The online version of this article includes the following source data and figure supplement(s) for figure 5: Source data 1. Uncropped immunoblot images corresponding to Figure 5. Figure supplement 1. Knockout of YBAP1 did not change exosome secretion. exosomal miRNA and dependent on the RNA- binding protein FMR1 for miRNA secretion (Wozniak et al., 2020). Diverse cell lines and likely tissues appear to invoke distinct sorting signals decoded by different RNA- binding proteins. Many of the proteins may engage in biomolecular condensates such as P- bodies as a mechanism to sort RNAs for secretion (Liu et al., 2021). Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 11 of 23 Cell Biology Research article a. b. d e h c a t t a a t c n u p 1 X B Y % a i r d n o h c o t i m e h t n o 80 60 40 20 0 d. s t n e v e e l t t u h s a t c n u p 1 X B Y l l e c r e p 40 30 20 10 0 YBX1 YBAP1 5μm c. YFP-YBX1 Rab5(Q79L)-mC mito-BFP 0s 3m48s 6m39s 2μm 9m30s 10m27s 28m28s Figure 6. YBX1 puncta relocalize from mitochondria to endosomes. (a) YBX1 puncta on the mitochondria. U2OS cells were stained with anti- YBX1 and anti- YBAP1 antibodies and observed by confocal microscopy. The right panel shows enlarged regions of interest from the left panel. Scale bar, 5 μm. (b) The statistics are of the percentage of YBX1 puncta detected in proximity to mitochondria. N=30 cells. (c) YBX1 puncta relocalize from mitochondria to endosomes. U2OS cells overexpressed YFP- YBX1, Rab5(Q79L)- mCherry and mito- BFP. Time- lapse images were acquired with a Zeiss LSM900 confocal microscope. Scale bar, 2 μm. (d) The statistics are of YBX1 puncta shuttle events per cell. The data was represented as violin plots. N=34 cells. Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 12 of 23 Cell Biology Research article miR223 appears to be an example of a number of small nuclear- encoded RNAs localized to mito- chondria (Jeandard et al., 2019). In some cases these RNAs, such as tRNAs, serve an essential func- tion such as in mitochondrial protein synthesis, however, for miRNAs with no obvious mitochondrial genome target, the function remains unclear (Jeandard et  al., 2019). Nonetheless, others have documented the localization of these miRNAs enclosed within the mitochondrion and in the case of miR223, it appears to be tightly associated with the inner membrane. The exact organization and function of miR223 in this location remains to be investigated but in the context of exosomal secre- tion, the mitochondrial localization appears to serve as a reservoir. In relation to the mitochondrial localization of miR223, we found a mitochondrial RNA- binding protein, YBAP1, that copurified with a tagged form of YBX1 expressed in HEK293 cells. YBAP1 has previously been reported to interact with YBX1 and independently found associated with mitochon- dria where its localization is dependent on an N- terminal transit peptide sequence (Muta et  al., 1997). The association of mitochondrial YBAP1 and cytoplasmic YBX1 was reproduced by coexpres- sion of recombinant forms of the two proteins in baculovirus- infected SF9 cells (Figure  4—figure supplement 1b). YBAP1 binds miR223 selectively but with an affinity significantly below that of YBX1 (Figure 4h–j). Although YBX1 does not localize to the mitochondrion, the stable interaction of the complex may suggest a transient relationship, perhaps during the biogenesis of YBAP1 as it transits from the cytoplasm into the mitochondrion. A functional relationship between YBAP1 and YBX1 is suggested by the reduction in miR223 in mito- chondria and increase in secretion of miR223 in exosomes secreted from YBAP1 KO cells (Figure 5). In contrast, removal of mitochondria by treatment of cells with CCCP resulted in an increase in cyto- plasmic miR223 at the expense of secretion in exosomes (Figure 3). Although YBAP1 may facilitate the retention of miR223 within mitochondria, mitochondrial RNA import and export may serve an MVB Exosomes YBX1 YBAP1 miR223 Figure 7. Diagram representing a model of miR223 sorting from mitochondria into exosomes. Stages in the transfer of miR223 from mitochondria. Cytosolic miR223 is enriched in mitochondria where it may be sequestered by a weak interaction with YBAP1. Cytoplasmic YBX1 interacts more tightly with miR223 which may drive the removal of miR223 from mitochondria. YBX1 in RNA granules may accumulate miR223 removed from mitochondria. YBX1 puncta may give rise to small particles carrying miR223 for uptake into endosomes and secretion in exosomes. Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 13 of 23 Cell Biology Research article independent role in the selective capture of miRNAs by YBX1 in cytoplasmic P- body condensates. YBX1 puncta appear to shuttle between mitochondria and endosomes at which point miR223 bound to YBX1 may be further sorted into invaginations budding into the interior of endosomes. The highly selective nature of miRNA sorting and secretion in exosomes suggests an important role in the trafficking of miRNAs between cells. Numerous studies have suggested a role for secreted miRNAs in recipient cells (Cha et  al., 2015; Mittelbrunn et  al., 2011; Pegtel et  al., 2010; Valadi et al., 2007). Nonetheless, as miRNAs ordinarily act stoichiometrically on target mRNAs, the extremely low abundance and copy number of miRNAs/vesicle is hard to reconcile with such a functional role of secreted miRNA (Chevillet et al., 2014; Shurtleff et al., 2017). Our observation that the bulk of cellular miR223 is held within mitochondria suggests an alternative role in some structural or regula- tory process, perhaps essential for mitochondrial homeostasis, controlled by the selective extraction of unwanted miRNA into RNA granules and further by secretion in exosomes (Figure 7). Key resources table Materials and methods Reagent type (species) or resource Designation Source or reference Identifiers Additional information Cell line (Spodoptera frugiperda) Sf9 Cell line (Homo sapiens) HEK 293T cells Cell line (Homo sapiens) HEK 293T- YBX1 KO cells Other Other Other Cell culture facility at UC Berkeley Cell culture facility at UC Berkeley Obtained by CRISPR- Cas9 in Schekman Lab Cell line (Homo sapiens) HEK 293T- YBAP1 KO This study Obtained by CRISPR- Cas9 in Schekman Lab Cell line (Homo sapiens) HEK 293T- 3xHA- EGFP- OMP25 This study Obtained by overexpression of pLJM1- 3XHA- EGFP- OMP25 in Schekman lab Cell line (Homo sapiens) U- 2OS cells Other Cell culture facility at UC Berkeley Cell line (Homo sapiens) U- 2OS Parkin- GFP cells This study Obtained by overexpression of Parkin- GFP in Schekman lab Recombinant DNA reagent pFastBac His6 MBP N10 TEV LIC cloning vector (4 C) Addgene RRID: Addgene_30116 N/A Recombinant DNA reagent Tom22- mCherry (plasmid) This study Gift of Dr Li Yu lab Recombinant DNA reagent His- MBP- YBX1 (plasmid) This study Recombinant DNA reagent His- MBP- YBX1(1–51) (plasmid) This study Recombinant DNA reagent His- MBP- YBX1(52–129) (plasmid) This study Recombinant DNA reagent His- MBP- YBX1(130–324) (plasmid) This study Recombinant DNA reagent His- MBP- YBX1(1–129) (plasmid) This study Recombinant DNA reagent His- MBP- YBX1(F85A) (plasmid) This study Recombinant DNA reagent His- MBP- YBAP1 (plasmid) This study To express YBX1 in insect cells. Plasmid maintained in Schekman lab To express YBX1(1–51) in insect cells. Plasmid maintained in Schekman lab To express YBX1(52–129) in insect cells. Plasmid maintained in Schekman lab To express YBX1(1–51) in insect cells. Plasmid maintained in Schekman lab To express YBX1(1–129) in insect cells. Plasmid maintained in Schekman lab To express YBX1(F85A) in insect cells. Plasmid maintained in Schekman lab To express YBAP1 in insect cells. Plasmid maintained in Schekman lab Recombinant DNA reagent Mito- BFP This study Gift of Dr. Samantha Lewis lab Recombinant DNA reagent mCherry- Rab5(Q79L) (plasmid) Addgene RRID: Addgene_35138 Recombinant DNA reagent pLJM1- 3XHA- EGFP- OMP25 This study Continued on next page To express 3xHA- EGFP- OMP25 in HEK293T cells. Plasmid maintained in Schekman lab Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 14 of 23 Cell Biology Research article Continued Reagent type (species) or resource Designation Source or reference Identifiers Additional information Anti- YBX1 (Rabbit polyclonal) Anti- YBAP1 (Mouse monoclonal) Anti- YBAP1(Rabbit polyclonal) Abcam RRID: AB_1950384 WB 1:1000 Santa Cruz Thermo Fisher Scientific RRID: AB_10611471 WB 1:1000 RRID: AB_2638956 WB 1:1000 Anti- Tim23 (Mouse monoclonal) BD Biosciences RRID: AB_398754 WB 1:1000 Anti- Tom20 (Mouse monoclonal) Abcam RRID: AB_945896 WB 1:1000 Anti- Calnexin (Rabbit polyclonal) Abcam RRID: AB_2069006 WB 1:2000 Anti- HA (Rabbit monoclonal) Cell Signaling RRID: AB_1549585 WB 1:1000 Anti- COX IV (Rabbit Monoclonal) Cell signaling RRID: AB_2085424 WB 1:1000 Anti- Citrate Synthase (Rabbit monoclonal) Cell signaling RRID: AB_2665545 WB 1:1000 Anti- Rab5 (Rabbit monoclonal) Cell signaling RRID: AB_2300649 WB 1:1000 Anti- LAMP1 (Rabbit monoclonal) Cell signaling RRID: AB_2687579 WB 1:1000 Anti- GRP78 (Rabbit polyclonal) Abcam RRID: AB_2119834 WB 1:3000 Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Anti- GADPH (Rabbit monoclonal) Anti- alpha Tubulin (Mouse monoclonal) Anti- beta Actin (Mouse monoclonal) Cell signaling RRID: AB_561053 WB 1:5000 Abcam RRID: AB_2241126 WB 1:5000 Abcam LICOR NIH RRID: AB_449644 WB 1:5000 https://www.licor.com/bio/image-studio-lite/ RRID: SCR_002285 https://fiji.sc/ Software, algorithm Image Studio Lite Software, algorithm FIJI Software, algorithm Prism 9 GraphPad RRID: SCR_002798 https://www.graphpad.com Cell lines and cell culture All immortalized cell lines were obtained from the UC- Berkeley Cell Culture Facility and were confirmed by short tandem repeat (STR) profiling and tested negative for mycoplasma contamina- tion. HEK 293T cells were cultured in DMEM with 10% FBS(VWR), NEAA (Gibco, Cat No: 11140050) and 1  mM Sodium Pyruvate (Gibco, Cat No: 11360070). For exosome production, we seeded cells at 10~20% confluency in 150 mm tissue culture dishes (Fisher Scientific, Cat No: 12- 565- 100) containing 30  ml of exosome- free medium. Exosomes were collected from 80% confluent cells (~48 hr). Exosome purification Conditioned medium was harvested from 80% to 90%  confluent HEK 293T cultured cells. All procedures were performed at 4 ℃. Cells and large debris were removed by centrifugation in a Sorvall R6  +centrifuge at 1000xg for 15  min followed by 10,000xg for 15  min using a FIBERlite F14−6x500 y rotor. The supernatant fraction was then centrifuged onto a 60% sucrose cushion in a buffer with 10 mM HEPES (pH 7.4) and 0.85% w/v NaCl at ~100,000 x g (28,000 RPM) for 1.5 h in a SW32Ti rotor. The interface over the sucrose cushion was collected and pooled for an additional centrifugation onto a 2 ml 60% sucrose cushion at ~120,000 x g (31,500 RPM) for 15 h using an SW41Ti rotor. The first collected interface was measured by refractometry and adjusted a sucrose concentration not exceeding 21%. For bulk purification, the EVs collected from the interface over the sucrose cushion after the first SW41Ti centrifugation were mixed with 60% sucrose to a final volume of 10  ml (the concentration of sucrose  ~50%). One ml of 40% and 1  ml of 10% sucrose Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 15 of 23 Cell Biology Research article were sequentially overlaid and the samples were centrifuged at ~150,000 x g (36,500 rpm) for 15 h in an SW41Ti rotor. The exosomes were located at the 10%/40% interface and collected for RNA extraction or immunoblot. Density gradient isolation of mitochondria and mitoplasts Mitochondria were isolated according to a well- established published protocol (Wang et al., 2020). HEK293T Cells were harvested at 80% confluency and were homogenized in 6  vol of HB buffer (225 mM mannitol, 25 mM sucrose, 0.5% BSA, 0.5 mM EGTA, 30 mM Tris–HCl, pH 7.4, and protease inhibitors) in a prechilled Dounce homogenizer (Kontes). The lysate was centrifuged and the postnu- clear supernatant was collected. Crude mitochondria were centrifuged at 6300 x g for 8 min, washed once with MRB buffer (250 mM mannitol, 0.5 mM EGTA, and 5 mM HEPES, pH 7.4), resuspended in 1 ml MRB buffer, laid over a 30% Percoll solution (9 ml) and centrifuged at 95,000 g for 45 min. The buoyant, purified fraction of mitochondria was collected for further analysis. For mitoplast purification, crude mitochondria were resuspended into 10 vol MRB buffer with 0.2 mg/ml digitonin and incubated on ice for 15 min. Digitonin- treated crude mitochondria were laid over a 30% Percoll solution (9 ml) and centrifuged at 95,000 g for 45 min. A buoyant, purified fraction of mitoplasts was collected for further analysis. YBAP1 immunoprecipitation from mitochondria After the Percoll gradient purification, the enriched mitochondria were diluted 2 x into MRB buffer and centrifuged at 12,000  g for 10  min. The mitochondrial pellet was lysed in 0.5  ml RIPA buffer (50 mM Tris- HCl, pH 7.5, 150 mM NaCl, 2 mM EDTA, 1% DDM, 1 mM PMSF) containing protease inhibitors (1  mM 4- aminobenzamidine dihydrochloride, 1  µg/ml antipain dihydrochloride, 1  µg/ml aprotinin, 1  µg/ml leupeptin, 1  µg/ml chymostatin, 1  mM phenylmethylsulphonyl fluoride, 50  µM N- tosyl- L- phenylalanine chloromethyl ketone and 1 µg/ml pepstatin) and RNase inhibitors, followed by centrifugation at 12,000  g for 10  min. Supernatant fractions were incubated with 10  µl washed protein A Dynabeads (ThermoFisher Scientific, Catalog number: 10001D) and 0.5 µg mouse mono- clonal IgG antibody and rotated at 4 ℃ for 1 h. A magnetic rack was used to remove protein A beads and the resulting supernatant fractions were incubated with 40 µl washed protein A Dynabeads and 4 µg YBAP1 antibody or mouse IgG antibody and rotated at 4 ℃ overnight. The beads were collected using a magnetic rack, washed 3 x with 1 ml of RIPA buffer, and collected for immunoblot and RNA extraction. Mitochondria immunoprecipitation Mito- IP was performed as previously described with slight modifications (Chen et  al., 2016). The mito- IP cell- line was grown to ~90% confluency in 15 cm dishes. All the subsequent steps of mito- IP were performed using ice- cold buffers either in a cold- room or on ice. Cells (2x107) were washed twice with 10  ml of PBS and then harvested in 10  ml of mito- IP buffer (10  mM KH2PO4, 137  mM KCl) containing protease inhibitors (1  mM 4- aminobenzamidine dihydrochloride, 1  µg/ml antipain dihydrochloride, 1  µg/ml aprotinin, 1  µg/ml leupeptin, 1  µg/ml chymostatin, 1  mM phenylmethyl- sulphonyl fluoride, 50 µM N- tosyl- L- phenylalanine chloromethyl ketone and 1 µg/ml pepstatin) and TCEP 0.5 mM. The final mito IP buffer also contained 6 ml of OptiPrep (Sigma) per 100 ml. Cells were collected at 700xg for 5 min and resuspended in 1 ml of mito- IP buffer per 15 cm plate and then lysed using 5–10 passes through a 22 G needle. A post- nuclear supernatant (PNS) fraction was obtained after centrifuging the lysate at 1500xg for 10  min to remove unbroken cells and nuclei. Whenever necessary, a fraction of PNS was saved for immunoblot analysis. The resulting PNS was incubated with 100 µl of anti- HA magnetic beads (Sigma) pre- equilibrated in the mito- IP buffer in 1.5 ml microcentri- fuge tubes and then gently rotated on a mixer for 15 min. The beads were collected using a magnetic rack and washed 3 x for 5 min with 1 ml of mito- IP buffer. For mitoplast purification by osmotic shock, the supernatant was discarded after the final wash of the mito- IP sample, and the beads were gently resuspended in 200 µl of hypotonic osmotic shock buffer (OSB) containing 20  mM HEPES at pH 7.4. The resuspended sample was incubated on ice for 30 min and then the beads were centrifuged at 15,000 g for 15 min to sediment mitochondria/ mitoplasts. Beads were then resuspended in 100 µl of KPBS and proteinase K was added to achieve a final concentration of 10 µg/ml and samples were incubated on a rotating mixer at 4 °C for 15 min. Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 16 of 23 Cell Biology Research article Subsequently, PMSF was added to a final concentration of 1  mM, along with a protease inhibitor cocktail, and the sample was incubated on ice for 5 min. Next, 2.5 units of RNase ONE (Promega) was added to the sample, which was further incubated on a rotating mixer at room temperature for 15 min. For protein analysis, the sample was eluted directly in the SDS loading buffer. Alternatively, Trizol was added to the sample to stop the reaction for RNA purification. To assess their quality, we assayed the immunoprecipitated mitochondria for the following protein markers by immunoblotting using the rabbit primary antibodies- anti- HA (1:1000), anti- COX IV (1:1000), anti- TOM20 (1:1000), anti- Citrate Synthase (1:1000), anti- RAB5 (1:1000), anti- LAMP1 (1:2000), anti- GAPDH (1:5000), and anti- GRP78 (Abcam) (1:3000). All the above antibodies were sourced from Cell Signaling Technology, unless stated otherwise. As a negative control, non- transduced HEK293T cells were used in these experiments to assess the non- specific capture of the marker proteins. Mitochondrial fractionation One mito- IP was performed per sample as described above. To ensure an even distribution of mito- chondria across the samples, we pooled washed beads from all the IPs and equally distributed aliquots for subsequent treatments. Mitochondria were lysed in a 50 μl final volume using either 1% vol/vol Triton X- 100 (final concentration) or by three sequential rounds of freeze/thaw using liquid- nitrogen, as indicated. Urea was added to a final concentration of 3 M. After a 10 min incubation on ice, samples were centrifuged at 15000xg for 15 min and supernatant and pellet fractions were collected as indi- cated. The total fractionated mitochondria were analyzed by immunoblotting for various mitochon- drial markers. To analyze the specific RNA content of total or fractionated mitochondria, we extracted RNA using Trizol (Invitrogen) as per manufacturer’s recommendations followed by q- PCR. In vitro packaging of miR223 and miR223mut Preparation of membranes and cytosol The membrane and cytosol fractions were prepared from HEK293T cells as previously described with slight modifications (Shurtleff et al., 2016; Temoche- Diaz et al., 2020). All steps were carried out in either the cold- room or on ice using ice- cold buffers and pre- chilled equipment. Briefly, HEK293T cells (80% confluency) were washed twice with PBS and then harvested in the homoge- nization buffer (HB) (250 mM sorbitol, 20 mM HEPES- KOH pH 7.4) containing protease inhibitor cocktail (1  mM 4- aminobenzamidine dihydrochloride, 1  µg/ml antipain dihydrochloride, 1  µg/ml aprotinin, 1 µg/ml leupeptin, 1 µg/ml chymostatin, 1 mM phenylmethylsulphonyl fluoride, 50 µM N- tosyl- L- phenylalanine chloromethyl ketone and 1 µg/ml pepstatin). The cell pellet was obtained by centrifuging the cells at 500xg for 5 min. After discarding the supernatant, cells were weighed and resuspended in two volumes of HB followed by lysis with 5–10 passages through a 22  G needle. The lysate was centrifuged at 1500xg for 10  min to remove unbroken cells and nuclei to obtain a PNS which was then centrifuged at 20,000xg for 30 min to obtain a membrane frac- tion. The supernatant from above was centrifuged at 150,000xg for 30 min using a TLA- 55 rotor (Beckman Coulter ) and the resulting supernatant was used as the cytosol fraction (~6 mg protein/ ml). Membranes from the first 20,000xg sedimentation were resuspended in 1 ml of HB and centri- fuged again at 20,000xg for 30 min. The pellet fraction was resuspended in one volume of HB and rested on an ice block for a minimum of 10 min until the insoluble components and debris settled at the bottom of the tube. The finely resuspended material in the resulting supernatant fraction was then transferred to a new microcentrifuge tube (to avoid the settled debris) and was used as the membrane fraction. Preparation of radiolabeled miR223 and miR223mut substrates HPLC purified miR223 and miR223mut oligos were obtained from IDT. A stock solution of these oligos (1  μl of a 10  μM) was 5’-end- labeled using T4PNK (NEB) and 5  μl of ATP, [γ–32P]- 6000  Ci/mmol 10mCi/ml EasyTide (PerkinElmer BLU502Z250UC) as per manufacturer’s recommendations in a 50 μl reaction volume. T4PNK was heat- inactivated at 70 °C for 15 min. Unincorporated radionucleotides were removed by passing through PerformaTM spin columns (EdgeBio). The flow- through (radiola- beled substrate) was collected and stored at –20 °C until further use. Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 17 of 23 Cell Biology Research article In vitro miR223 packaging reaction Wherever indicated, 10 μl of cytosol (~5 mg/ml), 17 μl of membranes, 2 μl of radiolabeled substrate, 9 μl of 5 x incorporation buffer (400 mM KCl, 100 mM CaCl2, 60 mM HEPES- NaOH, pH 7.4, 6 mM, MgOAc), 4.5 μl of 10 x ATP regeneration system (400 mM creatine phosphate, 2 mg/ml creatine phos- phokinase, 10 mM ATP, 20 mM HEPES pH 7.2, 250 mM sorbitol, 150 mM KOAc, 5 mM MgOAc), 1 μl of ATP (100 mM, Promega), 0.5 μl of GTP (100 mM, Promega), 1 μl of Ribolock (40 U/μl, Invitrogen) were mixed to setup a 45 μl in vitro packaging reaction. In samples without the cytosol or membranes, the final reaction volumes were adjusted to 45 μl using HB. The reactions were incubated at either 30 °C or on ice, as indicated, for 15 min. Following the incubation, the indicated samples were subjected to RNAse ONE(Promega) using 10 U of the enzyme in the presence of urea (300 mM final concentration) in a total reaction volume of 60 μl. Wherever indicated, TritonX- 100 was added to a final concentra- tion of 1%. The RNAse treatments were carried out for 20 min at 30 °C followed by RNA extraction using DirectZol (Zymo Research) kits as per manufacturer’s protocol. RNA was precipitated overnight at –20 °C by the addition of 3 vol of ethanol, 1/10th volume of 3 M sodium acetate (pH5.2) and 30 μg Glycoblue reagent (Invitrogen). Precipitated RNA was sedimented at 16,000xg for 30 min followed by washing with ice- cold 70% ethanol. The RNA pellet was resuspended in 2  X RNA loading dye (NEB) and heated for 5 min at 70 °C. RNA was separated using a 15% denaturing polyacrylamide gel, followed by gel drying using a vacuum gel dryer (Model 583, Biorad). Radioactive bands were visual- ized by phosphorimaging using a Kodak storage phosphor screen and the Pharos FX Plus Molecular Imager (Biorad). Immunoblots Cell lysates and other samples were prepared by adding 2% SDS and heated at 95 ℃ for 10 min. Protein was quantified using a BCA Protein Assay Kit (Thermo Fisher Scientific) and appropriate amounts were mixed with 5 x SDS loading buffer. Samples were heated at 95℃ for 10 min and sepa- rated on 4–20% acrylamide Tris- glycine gradient gels (Life Technologies). Proteins were transferred to PVDF membranes (EMD Millipore, Darmstadt, Germany) and the membrane was blocked with 5% fat- free milk powder in TBST and incubated for 1 h at room temperature or overnight at 4 °C with primary antibodies. Blots were then washed in three washes of TBST for 10 min each. Membranes were incubated with anti- rabbit or anti- mouse secondary antibodies (GE Healthcare Life Sciences, Pittsburgh, PA) for 1 hr at room temperature and rinsed in three washes of TBST for 10 min each. Blots were developed with ECL- 2 reagent (Thermo Fisher Scientific). Primary antibodies used in this study were as follows: anti- Tim23 (BD, 611222), Calnexin (Abcam, ab22595), Actin (Abcam, ab8224), Tubulin (Abcam, ab7291), YBAP1 (Santa Cruz Biotechnology, sc- 271200). Immunofluorescence Cells were cultured on 12 mm round coverslips (corning) and were fixed with 4% EM- grade parafor- maldehyde (Electron Microscopy Science, Hatfield, PA) in PBS pH7.4 for 10 min at room temperature. Cells were then washed 3 x with PBS for 10 min each, treated with permeabilizing buffer (10% FBS in PBS) containing 0.1% saponin for 20 min and treated in blocking buffer for 30 min. Subsequently, cells were incubated with primary antibodies in permeabilizing buffer for 1 hr at room temperature, washed 3 x with PBS for 10 min each and incubated with secondary antibodies in permeabilizing buffer for 1 hr at room temperature and finally washed 3 x with PBS for 10 min each. Cells were mounted on slides with Prolong Gold with DAPI (Thermo Fisher Scientific, P36931). Primary antibodies used in the immu- nofluorescence studies were as follows: anti- YBX1 (Abcam, ab12148), YBAP1 (Santa Cruz Biotech- nology, sc- 271200). Images were acquired with Zeiss LSM900 confocal microscope and analyzed with the Fiji software (http://fiji.sc/Fiji). Quantitative real-time PCR Cellular and EV RNAs were extracted using a mirVana miRNA isolation kit (Thermo Fisher Scien- tific, AM1560) or Direct- zol RNA Miniprep kits (Zymo Research). Taqman miRNA assays for miRNA detection were purchased from Life Technologies. Assay numbers were: hsa- miR- 223–3 p, 002295; hsa- mir- 190–5  p, 000489; U6 snRNA, 001973. Total RNAs were quantified using RNA bioanalyzer (Agilent). Taqman qPCR master mix with no Amperase UNG was obtained from Life Technologies for reverse transcription. For mRNA, RevertAid First Strand cDNA Synthesis Kit (Thermo Scientific, K1621) Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 18 of 23 Cell Biology Research article was used for reverse transcription. COX1 qPCR primers: Forward- 5’- TCTC AGGC TACA CCCT AGAC CA-3’, Reverse- 5’- ATCG GGGT AGTC CGAG TAAC GT-3’. GAPDH qPCR primers: Forward- 5’- CTGA CTTC AACA GCGA CACC -3’, Reverse- 5’- TAGC CAAA TTCG TTGT CATA CC-3’. Quantitative real- time PCR was performed using an QuantStudio 5 Real- Time PCR System (Applied Biosystems). Protein purification Twin Strep tag hybrid YBX1 was expressed and the protein was isolated 48  hr after PEI- mediated transfection of 293T cells. Cells were resuspended in PBS and collected by centrifugation for 5 min at 600  g. Pellet fractions were resuspended in 35  ml lysis buffer (50  mM Tris- HCl (pH 8),150mM NaCl,1mM EDTA, 2  mM DTT, 1  mM PMSF and 1  x protease inhibitor cocktail). After sonication of the cell suspension the crude lysate was centrifuged for 60 min at 20,000 rpm at 4 °C. The resulting supernatant fraction was incubated with 2 ml Strep- Tactin Sepharose resin (IBA, 2- 1201- 010) for 1 h. Strep- Tactin Sepharose resin samples were transferred to columns (18 ml) and protein- bound beads were washed with 60 ml wash buffer (50 mM Tris- HCl (pH 8), 500 mM NaCl, 1 mM EDTA, 2 mM DTT) until no protein was eluted as monitored by the Bio- Rad protein assay (Bio- Rad, Catalog #5000006). Proteins were eluted with 10 ml elution buffer (50 mM Tris- HCl (PH = 8),150mM NaCl, 10 mM desthi- obiotin, 1  mM EDTA, 2  mM DTT) and concentrated using an Amicon Ultra Centrifugal Filter Unit (50 kDa, 4 ml) (Fisher Scientific, EMD Millipore). Proteins were further purified by gel filtration chroma- tography (Superdex- 200, GE Healthcare) with columns equilibrated in storage buffer (50 mM Tris- HCl 7.4, 500 mM KCl, 5% glycerol, 1 mM DTT). Peak fractions corresponding to the appropriate fusion protein were pooled, concentrated, and distributed in 10  µl aliquots in PCR tubes, flash- frozen in liquid nitrogen and stored at –80 °C. Protein concentration was determined by known concentrations of BSA assessed by Coomassie Blue staining. Tagged (6xHis) and maltose- binding protein hybrid genes were expressed in baculovirus- infected SF9 insect cells (Lemaitre et al., 2019). Insect cell cultures (1 l, 1x106 cells/ml) were harvested 48 h after viral infection and collected by centrifugation for 20 min at 2000 rpm. The pellet fractions were resuspended in 35  ml lysis buffer (50  mM Tris- HCl 7.4, 0.5  M KCl, 5% glycerol, 10  mM imidazole, 0.5 µl/ml Benzonase nuclease (Sigma, 70746–3), 1 mM DTT, 1 mM PMSF and 1 x protease inhibitor cocktail). Cells were lysed by sonication and the crude lysate was centrifuged for 60 min at 20,000 rpm at 4 °C. After centrifugation, the supernatant fraction was incubated with 2 ml Ni- NTA His- Pur resin (Thermo Fisher, PI88222) for 1  hr. Ni- NTA resin samples were transferred to columns (18  ml) and protein- bound beads were washed with 60 ml lysis buffer until no protein was eluted as monitored by the Bio- Rad protein assay (Bio- Rad, Catalog #5000006). Proteins were eluted with 10 ml elution buffer (50 mM Tris- HCl 7.4, 0.5 M KCl, 5% glycerol, 500 mM imidazole). The eluted sample was incubated with 2 ml amylose resin (New England Biolabs, E8021L) for 1 hr at 4 °C. Amylose resin samples were transferred to columns and protein- bound beads were washed with 60 ml lysis buffer until no protein was eluted as monitored by the Bio- Rad protein assay. Proteins were eluted with 10 ml elution buffer (50  mM Tris- HCl 7.4, 500  mM KCl, 5% glycerol, 50  mM maltose) and were concentrated using an Amicon Ultra Centrifugal Filter Unit (50 kDa, 4 ml) (Thermo Fisher Scientific, EMD Millipore). Proteins were further purified by gel filtration chromatography (Superdex- 200, GE Healthcare) with columns equilibrated in storage buffer (50 mM Tris- HCl 7.4, 500 mM KCl, 5% glycerol, 1 mM DTT). Peak frac- tions corresponding to the appropriate fusion protein were pooled, concentrated, and distributed in 10 µl aliquots in PCR tubes, flash- frozen in liquid nitrogen and stored at –80 °C. Protein concentration was determined by known concentrations of BSA based on Coomassie blue staining. CRISPR/Cas9 genome editing A pX330- based plasmid expressing Venus fluorescent protein (Shurtleff et  al., 2016) was used to clone the gRNAs targeting YBAP1. A CRISPR guide RNA targeting the first exon of the YBAP1 open reading frame was designed following the CRISPR design website (http://crispor.tefor.net/crispor.py): CGCT GCGT GCCC CGTG TGCT . Oligonucleotides encoding gRNAs were annealed and cloned into pX330- Venus as described (Cong et  al., 2013). HEK293T cells were transfected by Lipofectamine 2000 for 48 hr at low passage number, trypsinized and sorted for single, Venus positive cells in 96- well plates by a BD Influx cell sorter. YBAP1 knockout candidates were confirmed by immunoblot. HEK 293T YBX1 knockout cells were generously provided by Dr. Xiaoman Liu (Liu et al., 2021). Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 19 of 23 Cell Biology Research article Electrophoretic mobility shift assay Fluorescently labeled RNAs (5’-IRD800CWN) for detecting free and protein- bound RNA were ordered from Integrated DNA Technologies (IDT, Coralville, IA). EMSA was performed as described with some modification (Rio, 2014). Briefly, 1  nM of IRD800CWN- labeled RNA was incubated with increasing amounts of purified proteins, ranging from 500 pM - 1  μM. Buffer E was used in this incubation (25 mM Tris pH8.0, 100 mM KCl, 1.5 mM MgCl2, 0.2 mM EGTA, 0.05% Nonidet P- 40, 1 mM DTT, 5% glycerol, 50 µg/ml heparin). Reactions were incubated at 30 ℃ for 30 min then chilled on ice for 10 min. Samples were mixed with 6 x loading buffer (60 mM KCl, 10 mM Tris pH 7,6, 50% glycerol, 0.03% (w/v) xylene cyanol). Mixtures (5  µl) were loaded onto a 6% native polyacrylamide gel and electrophoresed at 200 V for 45 min in a cold room. The fluorescence signal was detected using an Odyssey CLx Imaging System (LI- COR Biosciences, Lincoln, NE). The software of the Odyssey CLx Imaging System was used to quantify fluorescence. To calculate Kds, we fitted used Hill equations with quantified data points. CD63-Nluc exosome secretion assay The CD63- Nluc exosome secretion assay was carried out as described (Williams et al., 2023). Briefly, cells stably expressing CD63- Nluc were cultured in 24- well plates until reaching approximately 80% confluence. All subsequent procedures were performed at 4 °C. Conditioned medium (200 µl) was collected from the appropriate wells and transferred to microcentrifuge tubes. The tubes were subjected to centrifugation at 1000×g for 15 min to remove intact cells, followed by an additional centrifugation at 10,000×g for 15 min to eliminate cellular debris. Supernatant fractions (50 µl) were used for measuring CD63- Nluc exosome luminescence. Cells were kept on ice and washed once with cold PBS, and then lysed in 200 µl of PBS containing 1% TX- 100 and protease inhibitor cocktail. For the measurement of CD63- Nluc exosome secretion, a master mix was prepared by diluting the Extracellular NanoLuc Inhibitor at a 1:1000 ratio and the NanoBRET Nano- Glo Substrate at a 1:333 ratio in PBS (Promega, Madison, WI, USA). Aliquots of the Nluc substrate/inhibitor master mix (100 µl) were added to 50  µl of the supernatant fraction obtained from the medium- speed centrifugation. The mixture was briefly vortexed, and luminescence was measured using a Promega GlowMax 20/20 Luminometer (Promega, Madison, WI, USA). Following luminescence measurements, 1.5 µl of 10% TX- 100 was added to each reaction tube to achieve a final concentration of 0.1% TX- 100. Samples were vortexed briefly, and luminescence was measured again. For intracellular normalization, the lumi- nescence of 50  µl of cell lysate was measured using the Nano- Glo Luciferase Assay kit (Promega, Madison, WI, USA) following the manufacturer’s instructions. The exosome production index (EPI) for each sample was calculated using the formula: EPI = ([medium] - [medium +0.1% TX- 100])/cell lysate. Acknowledgements We thank Dr. Samantha Lewis for advice about localization to mitochondria and for sharing a plasmid; thanks Matthew J Shurtleff, David Melville, Shenjie Wu, Jordan Ngo, Congyan Zhang, Justin Williams, Morayma M Temoche- Diaz for suggestions and reading and editing the manuscript. We also thank staff at the UC Berkeley shared facilities, the Cell Culture Facility, the Flow Cytometry Facility and QB3- Berkeley (The California Institute for Quantitative Biosciences at UC Berkeley). LM and JS are supported as Research Associates of the HHMI. RS is an Investigator of the HHMI, a Senior fellow of the UC Berkeley Miller Institute of Science and Scientific Director of Aligning Science Across Parkin- son’s Disease. Additional information Funding Funder Howard Hughes Medical Institute Grant reference number Author Randy Schekman Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 20 of 23 Cell Biology Research article Funder Grant reference number Author The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Author contributions Liang Ma, Conceptualization, Data curation, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing – original draft, Writing – review and editing; Jasleen Singh, Data curation, Formal analysis, Methodology, Writing – original draft, Writing – review and editing; Randy Schekman, Conceptualization, Resources, Supervision, Funding acquisition, Validation, Investigation, Method- ology, Writing – original draft, Project administration, Writing – review and editing Author ORCIDs Liang Ma Randy Schekman http://orcid.org/0000-0003-3227-5917 http://orcid.org/0000-0001-8615-6409 Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.85878.sa1 Author response https://doi.org/10.7554/eLife.85878.sa2 Additional files Supplementary files • MDAR checklist Data availability All data generated or analyzed in this study are included in the manuscript and supporting files. Source data files have been provided for Figure 1b, Figure 1c, Figure 1f, Figure 1g, Figure 1- figure supplement 2a- c, Figure 2b, Figure 2g, Figure 2- figure supplement 1a- f, Figure 3a, Figure 3c, Figure 3e, Figure 3h, Figure 3 supplement 1b- c, Figure 4a, Figure 4b, Figure 4e, Figure 4f, Figure 4h, Figure 4i, Figure 4- figure supplement 1a- b, Figure 4- figure supplement 2a, Figure 5b. References Cha DJ, Franklin JL, Dou Y, Liu Q, Higginbotham JN, Demory Beckler M, Weaver AM, Vickers K, Prasad N, Levy S, Zhang B, Coffey RJ, Patton JG. 2015. KRAS- dependent sorting of miRNA to exosomes. eLife 4:e07197. DOI: https://doi.org/10.7554/eLife.07197, PMID: 26132860 Chen WW, Freinkman E, Wang T, Birsoy K, Sabatini DM. 2016. Absolute quantification of matrix metabolites reveals the dynamics of mitochondrial metabolism. Cell 166:1324–1337.. DOI: https://doi.org/10.1016/j.cell. 2016.07.040, PMID: 27565352 Chevillet JR, Kang Q, Ruf IK, Briggs HA, Vojtech LN, Hughes SM, Cheng HH, Arroyo JD, Meredith EK, Gallichotte EN, Pogosova- Agadjanyan EL, Morrissey C, Stirewalt DL, Hladik F, Yu EY, Higano CS, Tewari M. 2014. Quantitative and stoichiometric analysis of the microRNA content of exosomes. PNAS 111:14888–14893. DOI: https://doi.org/10.1073/pnas.1408301111, PMID: 25267620 Cocucci E, Racchetti G, Meldolesi J. 2009. Shedding microvesicles: artefacts no more. Trends in Cell Biology 19:43–51. DOI: https://doi.org/10.1016/j.tcb.2008.11.003, PMID: 19144520 Cong L, Ran FA, Cox D, Lin S, Barretto R, Habib N, Hsu PD, Wu X, Jiang W, Marraffini LA, Zhang F. 2013. Multiplex genome engineering using CRISPR/Cas systems. Science 339:819–823. DOI: https://doi.org/10. 1126/science.1231143, PMID: 23287718 Correia- Melo C, Ichim G, Tait SWG, Passos JF. 2017. Depletion of mitochondria in mammalian cells through enforced mitophagy. Nature Protocols 12:183–194. DOI: https://doi.org/10.1038/nprot.2016.159, PMID: 28005069 Harding C, Heuser J, Stahl P. 1983. Receptor- mediated endocytosis of transferrin and recycling of the transferrin receptor in rat reticulocytes. The Journal of Cell Biology 97:329–339. DOI: https://doi.org/10.1083/jcb.97.2. 329, PMID: 6309857 Huang L, Mollet S, Souquere S, Le Roy F, Ernoult- Lange M, Pierron G, Dautry F, Weil D. 2011. Mitochondria associate with P- bodies and modulate microRNA- mediated RNA interference. The Journal of Biological Chemistry 286:24219–24230. DOI: https://doi.org/10.1074/jbc.M111.240259, PMID: 21576251 Jeandard D, Smirnova A, Tarassov I, Barrey E, Smirnov A, Entelis N. 2019. Import of Non- Coding RNAs into Human Mitochondria: A. Critical Review and Emerging Approaches. Cells 8:286. DOI: https://doi.org/10.3390/ cells8030286 Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 21 of 23 Cell Biology Research article Lemaitre RP, Bogdanova A, Borgonovo B, Woodruff JB, Drechsel DN. 2019. FlexiBAC: a versatile, open- source baculovirus vector system for protein expression, secretion, and proteolytic processing. BMC Biotechnology 19:20. DOI: https://doi.org/10.1186/s12896-019-0512-z, PMID: 30925874 Li J, Liu K, Liu Y, Xu Y, Zhang F, Yang H, Liu J, Pan T, Chen J, Wu M, Zhou X, Yuan Z. 2013. Exosomes mediate the cell- to- cell transmission of IFN-α-induced antiviral activity. Nature Immunology 14:793–803. DOI: https://doi. org/10.1038/ni.2647, PMID: 23832071 Liu XM, Ma L, Schekman R. 2021. Selective sorting of microRNAs into exosomes by phase- separated YBX1 condensates. eLife 10:e71982. DOI: https://doi.org/10.7554/eLife.71982, PMID: 34766549 Lyons SM, Achorn C, Kedersha NL, Anderson PJ, Ivanov P. 2016. YB- 1 regulates tiRNA- induced Stress Granule formation but not translational repression. Nucleic Acids Research 44:6949–6960. DOI: https://doi.org/10. 1093/nar/gkw418, PMID: 27174937 Matsumoto K, Tanaka KJ, Tsujimoto M. 2005. An acidic protein, YBAP1, mediates the release of YB- 1 from mRNA and relieves the translational repression activity of YB- 1. Molecular and Cellular Biology 25:1779–1792. DOI: https://doi.org/10.1128/MCB.25.5.1779-1792.2005, PMID: 15713634 Mittelbrunn M, Gutiérrez- Vázquez C, Villarroya- Beltri C, González S, Sánchez- Cabo F, González MÁ, Bernad A, Sánchez- Madrid F. 2011. Unidirectional transfer of microRNA- loaded exosomes from T cells to antigen- presenting cells. Nature Communications 2:282. DOI: https://doi.org/10.1038/ncomms1285, PMID: 21505438 Mukherjee K, Ghoshal B, Ghosh S, Chakrabarty Y, Shwetha S, Das S, Bhattacharyya SN. 2016. Reversible HuR- microRNA binding controls extracellular export of miR- 122 and augments stress response. EMBO Reports 17:1184–1203. DOI: https://doi.org/10.15252/embr.201541930, PMID: 27402548 Muta T, Kang D, Kitajima S, Fujiwara T, Hamasaki N. 1997. p32 protein, a splicing factor 2- associated protein, is localized in mitochondrial matrix and is functionally important in maintaining oxidative phosphorylation. The Journal of Biological Chemistry 272:24363–24370. DOI: https://doi.org/10.1074/jbc.272.39.24363, PMID: 9305894 Pegtel DM, Cosmopoulos K, Thorley- Lawson DA, van Eijndhoven MAJ, Hopmans ES, Lindenberg JL, de Gruijl TD, Würdinger T, Middeldorp JM. 2010. Functional delivery of viral miRNAs via exosomes. PNAS 107:6328–6333. DOI: https://doi.org/10.1073/pnas.0914843107, PMID: 20304794 Rio DC. 2014. Electrophoretic Mobility Shift Assays for RNA–Protein Complexes. Cold Spring Harbor Protocols 2014:435–440. DOI: https://doi.org/10.1101/pdb.prot080721 Santangelo L, Giurato G, Cicchini C, Montaldo C, Mancone C, Tarallo R, Battistelli C, Alonzi T, Weisz A, Tripodi M. 2016. The RNA- Binding Protein SYNCRIP Is a Component of the Hepatocyte Exosomal Machinery Controlling MicroRNA Sorting. Cell Reports 17:799–808. DOI: https://doi.org/10.1016/j.celrep.2016.09.031, PMID: 27732855 Shurtleff MJ, Temoche- Diaz MM, Karfilis KV, Ri S, Schekman R. 2016. Y- box protein 1 is required to sort microRNAs into exosomes in cells and in a cell- free reaction. eLife 5:e19276. DOI: https://doi.org/10.7554/ eLife.19276, PMID: 27559612 Shurtleff MJ, Yao J, Qin Y, Nottingham RM, Temoche- Diaz MM, Schekman R, Lambowitz AM. 2017. Broad role for YBX1 in defining the small noncoding RNA composition of exosomes. PNAS 114:E8987–E8995. DOI: https://doi.org/10.1073/pnas.1712108114, PMID: 29073095 Temoche- Diaz MM, Shurtleff MJ, Nottingham RM, Yao J, Fadadu RP, Lambowitz AM, Schekman R. 2019. Distinct mechanisms of microRNA sorting into cancer cell- derived extracellular vesicle subtypes. eLife 8:e47544. DOI: https://doi.org/10.7554/eLife.47544, PMID: 31436530 Temoche- Diaz MM, Shurtleff MJ, Schekman R. 2020. Buoyant density fractionation of small extracellular vesicle sub- populations derived from mammalian cells. Bio- Protocol 10:e3706. DOI: https://doi.org/10.21769/ BioProtoc.3706, PMID: 33659370 Teng Y, Ren Y, Hu X, Mu J, Samykutty A, Zhuang X, Deng Z, Kumar A, Zhang L, Merchant ML, Yan J, Miller DM, Zhang H- G. 2017. MVP- mediated exosomal sorting of miR- 193a promotes colon cancer progression. Nature Communications 8:14448. DOI: https://doi.org/10.1038/ncomms14448, PMID: 28211508 Trajkovic K, Hsu C, Chiantia S, Rajendran L, Wenzel D, Wieland F, Schwille P, Brügger B, Simons M. 2008. Ceramide triggers budding of exosome vesicles into multivesicular endosomes. Science 319:1244–1247. DOI: https://doi.org/10.1126/science.1153124, PMID: 18309083 Valadi H, Ekström K, Bossios A, Sjöstrand M, Lee JJ, Lötvall JO. 2007. Exosome- mediated transfer of mRNAs and microRNAs is a novel mechanism of genetic exchange between cells. Nature Cell Biology 9:654–659. DOI: https://doi.org/10.1038/ncb1596, PMID: 17486113 Villarroya- Beltri C, Gutiérrez- Vázquez C, Sánchez- Cabo F, Pérez- Hernández D, Vázquez J, Martin- Cofreces N, Martinez- Herrera DJ, Pascual- Montano A, Mittelbrunn M, Sánchez- Madrid F. 2013. Sumoylated hnRNPA2B1 controls the sorting of miRNAs into exosomes through binding to specific motifs. Nature Communications 4:2980. DOI: https://doi.org/10.1038/ncomms3980, PMID: 24356509 Wang WX, Prajapati P, Nelson PT, Springer JE. 2020. The mitochondria- associated er membranes are novel subcellular locations enriched for inflammatory- responsive microRNAs. Molecular Neurobiology 57:2996–3013. DOI: https://doi.org/10.1007/s12035-020-01937-y, PMID: 32451872 Williams JK, Ngo JM, Lehman IM, Schekman R. 2023. Annexin A6 mediates calcium- dependent exosome secretion during plasma membrane repair. eLife 12:e86556. DOI: https://doi.org/10.7554/eLife.86556, PMID: 37204294 Wozniak AL, Adams A, King KE, Dunn W, Christenson LK, Hung WT, Weinman SA. 2020. The RNA binding protein FMR1 controls selective exosomal miRNA cargo loading during inflammation. The Journal of Cell Biology 219:e201912074. DOI: https://doi.org/10.1083/jcb.201912074, PMID: 32970791 Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 22 of 23 Cell Biology Research article Yuyama K, Sun H, Mitsutake S, Igarashi Y. 2012. Sphingolipid- modulated exosome secretion promotes clearance of amyloid-β by microglia. The Journal of Biological Chemistry 287:10977–10989. DOI: https://doi.org/10. 1074/jbc.M111.324616, PMID: 22303002 Ma et al. eLife 2023;12:e85878. DOI: https://doi.org/10.7554/eLife.85878 23 of 23 Cell Biology
10.1016_j.isci.2020.100959
Article Loss of Asb2 Impairs Cardiomyocyte Differentiation and Leads to Congenital Double Outlet Right Ventricle Substrate adapter polyubiquitinated Substrate E3 Cullin5 Asb2 E2 X Heart Failure Embryonic Lethality DORV Flna Smad2 E7.5 FHF AHF Cardiac Crescent E9.0-9.5 OFT LV RV Looped Heart Negative regulation Positive regulation Abir Yamak, Dongjian Hu, Nikhil Mittal, ..., Christel Moog- Lutz, Patrick T. Ellinor, Ibrahim J. Domian fyamak@mgh.harvard.edu (A.Y.) idomian@mgh.harvard.edu (I.J.D.) HIGHLIGHTS Flna removal partially rescues embryonic lethality of Asb2-heart- specific knockout AHF-Asb2 knockouts harboring one Flna allele have double outlet right ventricle Asb2-Flna regulate TGFb-Smad2 signaling in the heart Conserved role of Asb2 in heart morphogenesis between mice and humans DATA AND CODE AVAILABILITY GSE145495 Yamak et al., iScience 23, 100959 March 27, 2020 ª 2020 The Author(s). https://doi.org/10.1016/ j.isci.2020.100959 Article Loss of Asb2 Impairs Cardiomyocyte Differentiation and Leads to Congenital Double Outlet Right Ventricle Abir Yamak,1,2,3,9,* Dongjian Hu,2,4 Nikhil Mittal,1,2 Jan W. Buikema,2,5 Sheraz Ditta,2,6 Pierre G. Lutz,7 Christel Moog-Lutz,7 Patrick T. Ellinor,1,2,3 and Ibrahim J. Domian1,2,8,* SUMMARY Defining the pathways that control cardiac development facilitates understanding the pathogenesis of congenital heart disease. Herein, we identify enrichment of a Cullin5 Ub ligase key subunit, Asb2, in myocardial progenitors and differentiated cardiomyocytes. Using two conditional murine knockouts, Nkx+/Cre.Asb2fl/fl and AHF-Cre.Asb2fl/fl, and tissue clarifying technique, we reveal Asb2 requirement for embryonic survival and complete heart looping. Deletion of Asb2 results in upregu- lation of its target Filamin A (Flna), and concurrent Flna deletion partially rescues embryonic lethality. Conditional AHF-Cre.Asb2 knockouts harboring one Flna allele have double outlet right ventricle (DORV), which is rescued by biallelic Flna excision. Transcriptomic and immunofluorescence analyses identify Tgfb/Smad as downstream targets of Asb2/Flna. Finally, using CRISPR/Cas9 genome editing, we demonstrate Asb2 requirement for human cardiomyocyte differentiation suggesting a conserved mechanism between mice and humans. Collectively, our study provides deeper mechanistic under- standing of the role of the ubiquitin proteasome system in cardiac development and suggests a pre- viously unidentified murine model for DORV. INTRODUCTION Congenital heart diseases (CHDs) are prenatal defects that affect the heart’s structure and/or function and are the leading cause of infant mortality under 1 year of age. Approximately 1%–2% of human babies are born with cardiac malformations that pose as major risk factors for adult cardiovascular problems (Bruneau, 2008; Nemer, 2008). The heart, the first functional organ in the developing embryo, starts to form early on during development, before the end of gastrulation. The first and second heart fields (FHF and SHF, respectively) as well as the proepicardial organ and the cardiac neural crest are the major contributors to the forming heart (Martinsen and Lohr, 2015). The FHF gives rise primarily to the left ventricle and most of the atria; the SHF contributes to the right ventricle, outflow tract, and parts of the atria (Srivastava, 2006; Yamak and Nemer, 2015). Induction of the cardiac fate and the proper morphogenesis of the verte- brate heart are controlled by a well-characterized and highly conserved combinatorial network of transcrip- tion factors and signaling molecules that act together to orchestrate the embryonic development of the four-chambered mammalian heart and the subsequent post-natal maturation. Of important note, the adult heart has minimal intrinsic regenerative capacity (Mercola et al., 2011). As a result, significant stressors on the heart can result in loss of viable or functional myocardial tissue and ultimately heart failure. This renders cardiovascular disease a leading cause of death worldwide and highlights an unmet clinical need for novel approaches for heart regeneration. One major approach is the use of stem cells that can be induced to give rise to the different cell types that constitute the heart. Understanding the cellular processes and signaling pathways that govern in vivo heart formation and maturation is necessary for the generation of functional mature cardiac tissue for clinical and preclinical applications (Hu et al., 2018). Targeted protein degradation by the ubiquitin proteasome system (UPS) is important for the regulation of cellular physiology and is required for normal organ formation (Glickman and Ciechanover, 2002). The UPS consists of three enzymes: Ubiquitin (Ub) activating enzyme, E1, which transfers activated Ub to the Ub conjugating enzyme, E2. This then interacts with the E3 Ub ligase that covalently links the Ub or Ub chain to a lysine residue in the substrate thus targeting it for degradation by the proteasome. The E3 Ub ligase is responsible for substrate specificity (Jung et al., 2009). Recent evidence points to a role of the UPS in heart disease, particularly in myocardial remodeling, familial cardiomyopathies, chronic heart failure, and 1Harvard Medical School, Boston, MA 02115, USA 2Cardiovascular Research Center, Massachusetts General Hospital, 185 Cambridge Street, CPZN3200, Boston, MA 02114, USA 3Cardiovascular Disease Initiative, Broad Institute of MIT and Harvard, Cambridge, MA 02142, USA 4Department of Biomedical Engineering, Boston University, Boston, MA 02215, USA 5University Medical Center Utrecht, 3584 CX Utrecht, Netherlands 6Department of Pharmaceutical Sciences, Utrecht University, 3512 JE Utrecht, Netherlands 7Institut de Pharmacologie et de Biologie Structurale, IPBS, Universite´ de Toulouse, CNRS, UPS, Toulouse, France 8Harvard Stem Cell Institute, Cambridge, MA 02138, USA 9Lead Contact *Correspondence: fyamak@mgh.harvard.edu (A.Y.), idomian@mgh.harvard.edu (I.J.D.) https://doi.org/10.1016/j.isci. 2020.100959 iScience 23, 100959, March 27, 2020 ª 2020 The Author(s). This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). 1 B D A C E Figure 1. Asb2 Is Expressed in the Developing and Adult Heart and Undergoes Isoform Switching during Differentiation (A) qPCR analysis of embryonic cardiomyocytes reveals predominant Asb2a expression in the R-G+ and the R+G+ populations. R-G+: Mef2c-.Nkx2-5+; R+G+: Mef2c+.Nkx2-5+; R+G-: Mef2c+.Nkx2-5-; NEG: Mef2c-.Nkx2-5-. (B) qPCR analysis of Asb2a and Asb2b on RNA from murine hearts of different embryonic stages as well as neonates and postnatal day 8–9. Note that the a isoform is equally expressed at all stages, whereas the b isoform expression increases with development. (C) Western blot analysis on whole tissue extracts from embryonic and adult heart, spleen, and skeletal muscle using Asb2-specific antibody. Note that Asb2 corresponding band in the embryonic heart co-migrates with that in the spleen where only the a isoform is expressed, whereas that in the adult heart co-migrates with that in the skeletal muscle that is known to express on the b isoform. These data are consistent with the qPCR data in (B). 2 iScience 23, 100959, March 27, 2020 Figure 1. Continued (D) In situ hybridization on E9.5 mouse embryo showing robust Asb2 expression in the LV, RV, and OFT and to a lesser extent the IFT. LV, left ventricle; RV, right ventricle; OFT, outflow tract; IFT, inflow tract. (E) Immunohistochemistry on E10.5 and E12.5 mouse embryos using Asb2-specific antibody (green) and Troponin T (red). DAPI (blue) marks nuclei. Note Asb2 expression colocalizes with Troponin T in the myocardium (arrows) and no expression is seen in the endocardial cells (arrow heads). Scale bar is equivalent to 250 mm in the first three heart images left to right at E10.5 (top) and E12.5 (bottom), 25 mm in the E10.5 heart image top far right, and 50 mm in the E12.5 heart image bottom far right as indicated in the figure. ischemia-reperfusion injury (Pagan et al., 2013). Pharmacological inhibition of the proteasome is a new and promising means for cardioprotection (Pagan et al., 2013). Paradoxically, enhancing UPS activity has in some cases also provided protection against heart disease (Bulteau et al., 2001; Li et al., 2011; Powell et al., 2007) highlighting the importance of defining the role of the UPS as a therapeutic target in cardiac disease. In addition to its role as a protein quality control, the UPS has also been shown to regulate the turnover of sarcomere proteins, including the myofibrillar proteins myosin, actin, and troponin. Examples of this include the E3 Ub ligases MuRF1 (muscle-specific RING finger 1, targeting troponin I) and F box pro- tein Fbx122 (targeting a-actinin and filamin C) (Kedar et al., 2004; Spaich et al., 2012). E3 ligases have also been shown to regulate important signaling pathways in the heart, such as the JNK (c-Jun N terminal kinase) (Laine and Ronai, 2005), calcineurin (Fan et al., 2008), and the VEGF (vascular endothelial growth factor) signaling pathways (Murdaca et al., 2004). This important role of the UPS in the heart and its poten- tial for a therapeutic target in cardiac disease brings about a need to understand its specific function in heart development and disease. Using our previously described transgenic reporter system (Domian et al., 2009), we identified Asb2 (ankyrin repeat-containing protein with a suppressor of cytokine signaling box [SOCS box] 2) as being enriched in FHF and SHF cardiac progenitors. Asb2, which encodes specificity subunit of Cullin 5 RING E3 Ub ligase, exists in two isoforms: Asb2a and Asb2b in mouse, corresponding to variants 2 and 1 in humans, respectively (Bello et al., 2009). It has previously been shown to regulate differ- entiation of myeloid leukemia cells and skeletal myogenesis through proteasomal degradation of filamin proteins (Bello et al., 2009; Guibal et al., 2002; Heuze´ et al., 2008). Filamins (Flna, Flnb, and Flnc in mice) are actin-binding proteins important for the stabilization of the actin-cytoskeleton (van der Flier and Son- nenberg, 2001). Flnc is the only isoform expressed in the heart muscle where it is required for normal contractility (Fujita et al., 2012). Flna expression in the heart is restricted to endocardial and mesenchymal cells of the cardiac cushions during development and to valve leaflets in the adult heart (Norris et al., 2010). FLNA and FLNC mutations have been linked to cardiac defects in humans (de Wit et al., 2011, 2009; Kyndt et al., 2007; Valde´ s-Mas et al., 2014). In zebrafish, an additional Asb2 target TCF3 was very recently identi- fied where TCF3 was negatively regulated by Asb2 during cardiogenesis (Fukuda et al., 2017). Asb2 down- regulation was also shown to be a mediator of follistatin-induced muscle hypertrophy and SMAD2/3 regu- lation of skeletal muscle mass in young adults in mice. The repression of Asb2 was, however, ameliorated in aging mice, some of which also displayed increasing Asb2 baseline levels (Davey et al., 2016). Asb2 over- expression was also shown to drive skeletal muscle atrophy in mice (Davey et al., 2016). A recent study also showed that Asb2 knockout is embryonic lethal and that Asb2a targets Flna for proteasomal degradation during early cardiomyocyte differentiation (Me´ tais et al., 2018). The embryonic lethality of Asb2 mutants was shown to be primarily due to heart defects (Me´ tais et al., 2018). Herein, we show that Asb2 knockout in the FHF and SHF are both embryonic lethal by E10.5 and E12.5, respectively. Using tissue clearing combined with immunofluorescence technique, we show that Asb2 mutant hearts have incomplete looping. Moreover, Asb2 regulates cardiac morphogenesis partly through Flna turnover, and we hereby propose a model where Asb2-Flna controls TGFb-SMAD signaling to drive early cardiac formation. Additionally, Asb2 lethality in the anterior heart field (AHF) is partially rescued by Flna removal from these hearts. We also show that Asb2 ablation in the AHF leads to double outlet right ventricle (DORV), which is corrected upon further deletion of Flna from these hearts. Finally, we reveal that Asb2 role in cardiomyocyte differentiation is conserved in human cardiomyocytes as well. Collectively, our results shed light on the UPS regulation of heart development and its role as a cardio-therapeutic target and provide evidence for the first time for the role of the UPS in the rare congenital heart defect, DORV. RESULTS Asb2 Is Highly Enriched in the Embryonic Heart We have previously characterized a transgenic reporter system for the isolation of three distinct mouse cardiac progenitor cells from developing embryos: FHF population, marked by Nkx2.5+.Mef2c- expression, and two SHF population subsets: Nkx2.5-.Mef2c+ and Nkx2.5+.Mef2c+ (Domian et al., 2009). Genome-wide iScience 23, 100959, March 27, 2020 3 E A B C D Figure 2. Asb2 Is Essential for Early Cardiac Development (A) Nkx2-5+/Cre.Asb2 E9.5 and E11.5 knockout (KO) embryos (fl/fl) versus wild-type littermates (Wt). Note the resorbing KO embryo at E11.5. Scale bar is equivalent to 0.4 mm for E9.5 and 0.5 mm for E11.5 as indicated. (B) AHF-Cre.Asb2 E10.5 and E12.5 knockout (KO) embryos (fl/fl) versus wild-type littermates (Wt). Note the resorbing KO embryo at E12.5. Scale bar is equivalent to 0.5 mm for E10.5 and E12.5 as indicated in the figure. (C) 3D reconstruction of CUBIC-cleared, Troponin-T-stained E9.5 whole control and Asb2 mutant embryos showing both ventral and dorsal views. Note the bulging in the right ventricle of the control heart that is lacking in the mutant (indicated by the red arrow heads). Scale bar is equivalent to 200 mm as indicated. 4 iScience 23, 100959, March 27, 2020 Figure 2. Continued (D) Measurement of the heart tube of control and Asb2 mutant hearts. Note the statistically significant shorter heart tubes of the mutants. N = 5 per group. Data are represented as mean G SEM. * = p < 0.005. Unpaired t test was used using GraphPad Prism; p < 0.05 is considered statistically significant. (E) Heatmap analysis of a subset of cardiac looping differentially expressed genes in RNA-seq data from control (Nkx2-5+/Cre.Asb2fl/+) versus Nkx2-5+/Cre.Asb2 knockout E9.5 murine hearts. N = 3 in each group (each sample is in itself a combination of three to four hearts to account for heterogeneity among different litters). transcriptional profiling and real-time PCR (qPCR) reveal Asb2 transcripts enrichment in the three popula- tions (Figure 1A) (Domian et al., 2009). To investigate the temporal expression of Asb2 in the developing heart, we performed qPCR analysis on RNA from mouse hearts at different stages of embryonic develop- ment. Our data show that Asb2a is expressed similarly throughout heart development, whereas Asb2b expression increases with development (Figure 1B). This is further confirmed by western blot analysis, which shows that the Asb2 band in the embryonic heart co-migrates with that in the spleen (which expresses Asb2a [Spinner et al., 2015]), whereas the Asb2 band in the adult heart co-migrates with that in the skeletal muscle (known to express Asb2b [Bello et al., 2009]) (Figure 1C). To further investigate in vivo spatial cardiac expres- sion of Abs2, we performed in situ hybridization on E9.5 embryos. Our data reveal robust expression of Asb2 transcripts predominantly in the left (LV) and right ventricles (RV) and to a lower extent in inflow (IFT) and outflow tracts (OFT) (Figure 1D). Furthermore, immunostaining of E10.5 and E11.5 (Figure 1E, upper and lower panels, respectively) embryonic sections using Asb2-specific antibody shows that, in the heart, Asb2 expression (green) is restricted to the myocardium overlapping with cardiac Troponin T (red). White arrows in the zoomed merged image at E10.5 (right panel) indicate overlap of Asb2 and Troponin T in the myocardial layer, but no expression is seen in the endocardial layer indicated by arrow heads. Asb2 Is Required for Early Cardiac Formation To investigate the role of Asb2 during cardiac development, we generated two conditional knockout lines (KO): Nkx2-5+/Cre (a mouse line with the Cre recombinase knocked into the Nkx2-5 locus) and AHF-Cre (a mouse line with a transgene placing Cre under the transcriptional control of the AHF enhancer of the Mef2c gene). These mouse lines allow for the targeted removal of (Lombardi et al., 2009) Asb2 from the whole heart and the SHF, respectively (Lombardi et al., 2009). The floxed alleles are in common region and inactivate both Asb2 isoforms. Both conditional KOs have pericardial edema and are embryonic lethal: Nkx2-5+/Cre.Asb2fl/fl mice die at E10.5–11 and AHF-Cre.Asb2fl/fl die at E11.5–12 (Figures 2A and 2B, respec- tively). AHF-Cre.Asb2fl/fl mice analyzed at E10.5 also have shorter OFT compared with their control litter- mates (Figure S1D). For Nkx2-5+/Cre.Asb2fl/fl, mice were analyzed at E8.5 (3 litters), E9.5 (23 litters), E10.5 (3 litters), and E11.5 (2 litters); for AHF-Cre.Asb2fl/fl, mice were analyzed at E9.5 (3 litters), E10.5 (4 litters), and E12.5 (2 litters). Each litter consists of 8–11 embryos in total. All embryos were genotyped. Figure S1A shows the reduced level of Asb2 in the heterozygotes (Nkx2-5+/Cre.Asb2fl/+) and the complete loss of Asb2 in the knockouts (Nkx2-5+/Cre.Asb2fl/fl). In order to perform a phenotypic analysis of the Nkx2-5+/Cre-Asb2fl/fl mutant embryos, we used state-of- the-art tissue clearing technique CUBIC combined with immunostaining. CUBIC can effectively clear mice embryos and embryonic hearts while preserving immunolabels (Kolesova´ et al., 2016; Tainaka et al., 2014). Nkx2-5+/Cre-Asb2fl/fl and control littermates e9.5 mice embryos were cleared with CUBIC and stained for Troponin T to mark cardiomyocytes as well as DAPI for nuclei. Confocal microscopy with optical sectioning followed by 3D-reconstruction allowed the precise visualization of the developing hearts without disruption of underlying anatomy. During cardiac morphogenesis, the straight heart tube undergoes sequential looping steps to get to the fully looped heart. The fully looped heart acquires a he- lical shape in mice that is also referred to as the mature S-loop in chicks (Le Garrec et al., 2017; Ma¨ nner, 2009). In Le Garrec et al. paper, they used computer modeling to simulate the biological process of mouse cardiac looping, incorporating in their model the left-right asymmetry and mechanical constraints seen in the looping heart. Their findings suggest that the lack of any of these parameters would lead to a C-shaped heart loop rather than the helical structure. In the chick, the heart is first transformed into a C-shaped heart, a process known as dextral looping. The C-loop is then converted into an immature S-loop that then trans- forms into a mature S-looped heart where the ventricular segments are curved outward to generate the left and right chambers (Ma¨ nner, 2009). As shown in Figure 2C, the mutant embryos do not form the full helical structure seen in the control littermates. Instead, they have partially looped hearts that resemble the C-shaped hearts in the chick (Ma¨ nner, 2009). Moreover, measurement of the heart tube length in mutant versus control hearts reveals statistically significant shorter tubes in the mutant hearts (Figure 2D). Video S1 is a z stack of stained control and mutant e9.5 embryos showing the incomplete looping in the mutant embryo. Figure S1C represents four images from the z stack at different depth in the embryo. Four to five iScience 23, 100959, March 27, 2020 5 A B C 6 iScience 23, 100959, March 27, 2020 Figure 3. Asb2 Targets Flna for Proteasomal Degradation in the Developing Heart and Asb2-Mutant Hearts Have an Altered Gene Expression Profile (A) Immunohistochemistry on E9.5 Abs2 heterozygote (Nkx2-5+/Cre.Asb2fl/+, middle pane) and mutant hearts (Nkx2-5+/Cre.Asb2fl/fl, lower panel) as well as Wt controls (top panel) using Flna (red) and Troponin-T (green)-specific antibodies. Note that FlnA expression is restricted to the endocardial layer (white arrow heads) in the Wt heart, whereas it is abnormally expressed in the myocardial layer in the Asb2-mutant hearts co-localizing with Troponin-T expression there (white arrows). Moreover, some cardiomyocytes in the outflow tract of the Asb2-heterozygous hearts also express Flna (yellow arrows) suggesting a dose- dependent regulation. Scale bar is equivalent to 250 mm in the first column (left), 100 mm in the second, third, and fourth columns, and 25 mm in the fifth (far right) column as indicated in the figure. (B) Heatmap analysis of RNA-seq data from control (Group1: Nkx2-5+/Cre.Asb2fl/+), Asb2 mutant (Group2: Nkx2-5+/Cre.Asb2fl/fl), Flna mutant (Group3: Nkx2-5+/Cre.Flnafl/y), and Asb2-Flna double mutant (Group4: Nkx2-5+/Cre.Asb2fl/fl.Flnafl/y) E9.5 murine hearts. Note the high level of differentially expressed genes in the Asb2-mutant and Asb2-Flna double mutant versus the control groups. A small subset of genes (indicated by arrows) that are perturbed in the Asb2-mutant hearts are restored to normal in the Asb2-Flna double mutants. N = 3 in each group (each sample is in itself a combination of three to four hearts to account for heterogeneity among different litters). (C) Heatmap analysis of a subset of genes from the RNA-seq data in (B) that are part of the Tgfb/Smad signaling pathway. Note that the Foxa genes expression levels (indicated with a yellow line) that are downstream of the Tgfb/Smad are restored to normal in the Asb2-Flna double mutants versus the Asb2-mutant hearts. embryos were analyzed for each condition. The efficiency of the CUBIC/immunostaining technique on e9.5 mouse embryo is evidenced by the clearly visible striations of the cardiac muscle fibers (Figure S1B). In order to identify Asb2 downstream targets in the heart, RNA sequencing (RNA-seq) analysis was per- formed on Nkx2-5+/Cre.Asb2fl/fl and control littermates (Figure 3B) (Figure S2C shows reduced levels of Asb2 transcripts in the Nkx2-5+/Cre.Asb2fl/fl knockout compared with the Nkx2-5+/Cre.Asb2fl/+ heterozygote control). The gene expression profile was greatly altered in the Asb2 mutant hearts compared with their control littermates (Group 2 versus Group 1) (Figure 3B). Of note, a number of genes that are mis-expressed in the Asb2 cardiac mutant hearts have been previously linked to abnormal cardiac looping in mice (Figure 2E) (Azhar et al., 2003; Bardot et al., 2017; Chen et al., 1997; Le Garrec et al., 2017; Mine et al., 2008; Ribeiro et al., 2007; Vincentz et al., 2011). Ingenuity Pathway Analysis also shows that ‘‘cardiovascular system devel- opment and function’’ as well as ‘‘cardiovascular disease’’ are among the top pathways altered in the Asb2 mutant hearts (Table S1, yellow highlights). Table S2 is an upstream analysis with the ones with a positive activation Z score > 1.5 highlighted in yellow. This list shows the pathways whose downstream targets are altered (upregulated or downregulated) in our knockouts versus controls. Targets with a positive Z score suggest upregulation pathways in the Asb2 mutant hearts. Asb2 Controls Cardiac Morphogenesis Partly through Regulating Filamin A Since Asb2 targets filamin proteins for degradation (Me´ tais et al., 2018) and Flna perturbations lead to car- diac defects and embryonic lethality (Feng et al., 2006), we investigated cardiac Flna expression in the Nkx2-5+/Cre.Asb2fl/fl. Flna expression in the control heart (Figure 3A, top panel) is restricted to endocardial and pericardial layers (red staining, white arrow heads). In the knockout embryos (Figure 3A, third panel), Flna’s expression domain is abnormally expanded to include the myocardial layer (white arrows), co-local- izing with Troponin T expression (green for Troponin and yellow for the co-localization). Moreover, in Nkx2-5+/Cre.Asb2fl/+ heterozygous hearts (Figure 3A, second panel), Flna is abnormally expressed in some cardiomyocytes of the OFT myocardium (yellow arrows) suggesting that Asb2 regulation of Flna turn- over is dose dependent. We then hypothesized that, if Asb2 cardiac mutant phenotype is due to overexpression of Flna, then concurrently deleting Flna along with Asb2 should suppress the Asb2 phenotype. (Please note that Flna is an x-linked gene so a knockout is denoted by fl/fl for female or fl/y for male, whereas a heterozygous is denoted by fl/x or fl/+.) To examine this hypothesis, we developed Nkx2-5+/Cre.Asb2fl/fl.Flnafl/y double mutants. Removal of Flna from the hearts of Nkx2-5+/Cre.Asb2fl/fl did not rescue lethality (Figure S2A). Approximately 16 litters were analyzed at E9.5 and 3 litters at E10.5. As expected, Nkx2-5+/Cre.Asb2fl/fl.Flnafl/fl double knockouts no longer harbor ectopic Flna expression in the myocardium (Figure S2B) as was previously seen with the Nkx2-5+/Cre.Asb2fl/fl single knockouts (Figure 3A). Instead, the double knockouts have normal endocardial expression of Flna similar to their control litter- mates (Figure S2B). RNA-seq analysis on e9.5 hearts from these mice show that their gene expression pro- file is closely related to the Nkx2-5+/Cre.Asb2fl/fl group (Group 2 versus Group 4) (Figure S2C shows reduced levels of Asb2 transcripts in the Nkx2-5+/Cre.Asb2fl/fl.Flnafl/y double knockout compared with the Nkx2-5+/Cre.Asb2fl/+ heterozygote control). However, some genes whose expression was altered in the Nkx2-5+/Cre.Asb2fl/fl group are restored to normal in the Nkx2-5+/Cre.Asb2fl/fl.Flnafl/y hearts (indicated by arrows and shown in Table S3). These results suggest that Flna concurrent deletion can restore the normal expression level of a subset of genes in the Asb2 mutant hearts. Among these genes are the iScience 23, 100959, March 27, 2020 7 A B C 8 iScience 23, 100959, March 27, 2020 Figure 4. Tgfb/Smad Signaling Activity Is Downstream Asb2-Flna in the Developing Heart (A) Schematic representation of Asb2-Flna-Smad2 interaction network using MetaCore Clarivate Analytics software. Note that Asb2 ubiquitinates and negatively regulates Flna, whereas Flna binds directly to and positively regulates Smad2. (B) Immunohistochemistry on Nkx2-5+/Cre.Asb2 mutant (middle panel) and Nkx2-5+/Cre.Asb2-Flna double mutant (last panel) murine hearts as well as wild- type controls (top panel) using pSmad2-specific antibody (green) and Troponin T (red). Note the nuclear localization of pSmad2 as a sign of Tgfb/Smad2 cycle activation. Examples of positive (purple arrowhead) and negative (yellow arrowheads) nuclei are indicated in the Wt sample (red box). DAPI marks all nuclei (AV, atrioventricular canal; V, primitive ventricle; OFT, outflow tract; Myo, myocytes; Endo, endocardial cells). Myocardial cells are marked by white arrows; endocardial cells are marked by white arrowheads. Scale bar is equivalent to 75 mm in the first two columns from the left and 25 mm in the third, fourth, and fifth columns as indicated in the figure. (C) Quantification of the immunostaining in (B) of percentage of pSmad2-positive nuclei in cardiomyocytes ((AV+V) Myo and OFT Myo) as well as endocardial cells (AV endo). Note the increased level of pSmad2-positive nuclei in Asb2-mutant myocytes that are restored to normal in the Asb2-FlnA double mutants. This regulation is not seen in the endocardial cells that do not express Asb2 (Figure 1E). n = 7 for Wt; n = 4 for Nkx2-5+/Cre.Asb2fl/fl; n = 3 for Nkx2-5+/Cre.Asb2fl/fl.Flnafl/fl. Data are represented as mean G SEM. * = p < 0.05 Control versus Nkx2-5+/Cre.Asb2fl/fl; # = p < 0.05 Nkx2-5+/Cre.Asb2fl/fl versus Nkx2-5+/Cre.Asb2fl/fl.Flnafl/fl; NS = not significant. Two-way ANOVA was used for analysis using GraphPad Prism. p < 0.05 is considered statistically significant. Foxa genes, which are downstream of the Tgfb/Smad signaling (Figure 3C, yellow line) (Tang et al., 2011). Other genes in the Tgfb/Smad pathway are also altered in both Asb2-mutant and Asb2.Flna double mutant hearts (Figure 3C). Figure S2D is a qPCR analysis confirming some of these altered genes. Tgfbr1 and InhA (which encodes a member of the Tgfb superfamily) are also among the positively regulated targets in the upstream analysis of the RNA-seq data of Asb2-mutant hearts versus control (Table S2). Both genes are no longer positively regulated in the upstream analysis of the list of genes corrected in the Asb2-Flna double mutant hearts (Table S3, yellow highlights). These data prompted further analysis of the Asb2/Flna regula- tion of TGFb/Smad signaling in the heart of these mice. Asb2 Regulates TGFb/Smad Signaling through Regulating Filamin A Protein TGFb signaling is initiated upon ligand-stimulated activation of serine/threonine receptor kinases that in turn lead to phosphorylation and activation of Smad proteins. Activated Smads interact with common signaling transducer Smad4, translocate to the nucleus, and activate downstream targets (Shi and Massague´ , 2003). Flna directly associates with Smad2 and Smad2 phosphorylation, and TGFb/Smad2 signaling is impaired in Fln-null human melanoma cells (Sasaki et al., 2001; Zhou et al., 2011). Moreover, FLNA mutations were linked to x-linked myxomatous valvular dystrophy, a multivalve degeneration disor- der, and disrupted TGFb/Smad2/3 signaling was implicated in the disease pathogenesis (Geirsson et al., 2012; Norris et al., 2010). Using the ‘‘Build Network’’ module in MetaCore Clarivate Analytics software, we investigated the Asb2-Flna-Smad2 interaction. As shown in Figure 4A, Asb2 negatively regulates Flna through ubiquitination and Flna positively regulates Smad2 through direct binding. Asb2, Flna, and Smad2 are shown in red for visualization. In order to investigate further Asb2/Flna regulation of TGFb/ Smad2 signaling in cardiac development, we immunostained E9.5 Asb2-mutant hearts with antisera directed against pSmad2 (Figure 4B) and then quantified the pSmad2-positive nuclei. Figure 4C shows significant increase in the percentage of pSmad2-positive nuclei in the Nkx2-5+/Cre.Asb2fl/fl myocytes (pre- viously shown to have overexpression of Flna [Figure 3A]) compared with their littermate controls. This increase was not seen in the endocardial cells where Flna expression is normal (Figure 3A). Interestingly, pSmad2 levels were restored to normal in the Nkx2-5+/Cre.Asb2fl/fl.Flnafl/fl (double mutant) myocytes further confirming that Asb2 regulates pSmad2 in the heart through the regulated turnover of Flna. Flna Removal from AHF-Cre.Asb2 Mutant Hearts Partially Rescues Embryonic Lethality To examine Flna expression in the AHF-Cre.Asb2fl/fl hearts (where Asb2 is knocked out in the RV and OFT only), Flna immunostaining was performed. As shown in Figure 5A, Flna expression (red) is restricted to the endocardial and epicardial layers in the control hearts (top panel, white arrow heads), whereas it is aber- rantly expressed in the myocardial layer of the OFT and RV only (red staining lower panel, white arrows), co-localizing with TroponinT expression (yellow staining lower panel) there. Flna expression was normal in the myocardial layer of the primitive left ventricle (PV) that harbors normal Asb2 expression and acts as an internal control in these mice. We then sought to examine the effect of further knocking out Flna from the AHF-Cre.Asb2-mutant hearts. To do this, we crossed Asb2fl/fl.Flnafl/fl with AHF-Cre.Asb2fl/+ mice. Our results show that AHF-Cre.Asb2fl/fl.Flnafl/y are born with the expected Mendelian ratios (Figure 5B); however, newborn pups die between P0.5 and P1.5. These results show that Flna deletion partially rescues Asb2 lethality. The AHF-Cre.Asb2fl/fl.Flnafl/+ also survive to birth albeit at a lower percent- age from what is expected by Mendelian ratios; these mice also die right after birth at P0.5. iScience 23, 100959, March 27, 2020 9 Troponin FlnA OFT DAPI Merge Ε9.5 250μm Ε9.5 100μm 100μm 100μm 25μm PV Troponin FlnA OFT PV DAPI Merge Ε9.5 250μm Ε9.5 100μm 100μm 100μm 25μm + / l f 2 b s A . e r C - F H A l f / l f 2 b s A . e r C - F H A A B AHF-Cre.Asb2fl/+ X Asb2fl/fl Expected Observed at P0.5 AHFCreAsb2fl/fl AHFCreAsb2fl/+ Control 25% 25% 50% 0/29 (0%) 13/29 (44.8%) 16/29 (55.1%) AHF-Cre.Asb2fl/+ X Asb2fl/fl.FlnAfl/fl e r C F H A Asb2fl/fl.FlnAfl/+ Asb2fl/fl.FlnAfl/y Asb2fl/+.FlnAfl/+ Asb2fl/+.FlnAfl/y Expected Observed at P0.5 1/31 (3.2%) 4/31 (12.9%) 6/31 (19.3%) 6/31 (19.3%) 12.5% 12.5% 12.5% 12.5% Control 50% 14/31 (45.2%) Dead at P0.5-1.5 Figure 5. Flna Removal from AHFCre.Asb2 Mutant Hearts Partially Rescues Their Lethality (A) Immunohistochemistry on E9.5 AHF-Cre.Abs2 mutant hearts (AHF-Cre.Asb2fl/fl, lower panel) and littermate controls (AHF-Cre.Asb2fl/+, top panel) using Flna (red)- and Troponin-T (green)-specific antibodies. Note overexpression of Flna in the OFT (white arrows) of the AHF-Cre.Asb2 mutant hearts but not the primitive left ventricle (PV) that harbors normal Asb2 levels thus serving as an internal control. Flna expression in the control hearts is restricted to the endocardial layer (white arrow heads). Scale bar is equivalent to 250 mm in the first column (left); 100 mm in the second, third, and fourth columns; 25 mm in the fifth column (far right) as indicated. (B) Table showing the survival of AHF-Cre.Asb2 mutant (top) and AHF-Cre.Asb2-Flna double mutant (bottom) mice. Note that no AHF-Cre.Abs2 mutant mice are observed at P0.5. However, the AHF-Cre.Asb2-Flna double mutant (Asb2fl/fl.Flnafl/y) mice are born at the expected Mendelian ratios. AHF-Cre.Asb2- mutant mice harboring one copy of Flna (Asb2fl/fl.Flnafl/+) are also born yet at lower percentage than what is expected by Mendelian genetics. These mice die, however, right after birth. P0.5, postnatal day 0.5. Asb2 Removal from the Anterior Heart Field Leads to Double Outlet Right Ventricle) in Mice To determine the cardiac defects of the AHF-Cre.Asb2fl/fl.Flnafl/+, we examined these mice at e16.5–e17.5 after the completion of cardiac morphogenesis but prior to the perinatal mortality associated with this genotype. Five litters were analyzed. Figure 6A shows the survival rate of these mice at E16.5. Gross examination of these hearts revealed that both the aorta and the pulmonary artery originate in the RV (Figure 6B, middle panel, yellow circle). In contrast, both the control hearts (Figure 6B, left panel) and those with AHF-Cre.Asb2-Flna double mutant (Fig- ure 6B, right panel) were grossly normal with the pulmonary artery originating in the RV and the aorta originating in 10 iScience 23, 100959, March 27, 2020 C * A B D Figure 6. AHF-Asb2 Mutant Hearts Have Double Outlet Right Ventricle and Ventricular Septal Defect (A) Table showing the survival of AHF-Cre.Asb2-Flna double mutant mice at E16.5. (B) E16.5 whole hearts of AHF-Cre.Asb2-mutant embryos with one copy of Flna (AHF-Cre.Asb2fl/fl.Flnafl/x), AHF- Cre.Asb2.Flna double mutants (AHF-Cre.Asb2fl/fl.Flnafl/fl), and wild-type control. Note that both the pulmonary artery (PA) and the aorta (Ao) are open in the right ventricle (RV) of the AHF-Cre.Asb2fl/fl.Flnafl/x hearts (yellow circle). Scale bar is equivalent to 0.02 mm as indicated. (C) Masson trichrome staining of E16.5 heart sections of control (Wt) (top), AHF-Cre.Asb2fl/fl.Flnafl/x (middle), and AHF- Cre.Asb2fl/fl.Flnafl/fl (bottom) embryos. Note that both the pulmonary artery and the aorta open in the right ventricle of the AHF-Cre.Asb2fl/fl.Flnafl/x hearts (yellow circle, middle panel) but not the Wt or the AHF-Cre.Asb2fl/fl.Flnafl/fl hearts. The AHF-Cre.Asb2fl/fl.Flnafl/x also have a VSD indicated by asterisk (middle panel, right). Ao, aorta; PA, pulmonary artery; RV, right ventricle; LV, left ventricle; IVS, interventricular septum. Scale bar is equivalent to 250 mm. (D) Number of E16.5 hearts with DORV in Wt, AHF-Cre.Asb2fl/fl.Flnafl/x, and AHF-Cre.Asb2fl/fl.Flnafl/fl embryos. Note that 5/5 Asb2fl/fl.Flnafl/x have DORV accompanied by a VSD suggesting 100% disease penetrance in these mice. the LV. Serial sections of mutant and control hearts (Figure 6C) further confirm that the AHF-Cre.Asb2fl/fl.Flnafl/x mice have DORV (Figure 6C middle panel, yellow oval). This is also accompanied by a ventricular septal defect (Figure 6C middle panel right, indicated by asterisk), a feature commonly associated with DORV in patients with congenital heart disease (Obler et al., 2008). As shown in Figure 6D, the DORV phenotype appeared to be fully penetrant in the AHF-Cre.Asb2fl/fl.Flnafl/x hearts. Notably, the DORV phenotype is corrected in the AHF-Cre.Asb2-Flna double mutant hearts (AHF-Cre.Asb2fl/fl.Flnafl/y, Figure 6C lower panel). Asb2 Is Required for Human Embryonic Stem Cell-Derived Cardiomyocyte Differentiation To further investigate if the requirement for Asb2 for cardiac development is conserved during human cardiomyo- (hESC)-derived cardiomyocyte in vitro cyte differentiation, we turned to human embryonic stem cell differentiation. Both ASB2 variants 1 (Asb2b in mice) and 2 (Asb2a in mice) are expressed at different stages of cardiomyocyte differentiation (Figure 7A, top and bottom graphs, respectively). Using CRISPR/Cas9 genome ed- iting technology, we then generated ASB2-null hESCs. The guides were designed in exon 2 (targeting variant 1 specifically) or exon 4 (targeting variants 1 and 2) (Figure S3A). Four wild-type (Wt) (received the CRISPR/Cas9 con- structs but failed to generate an in/del) and four knockout (KO) lines were generated. The genotype of all lines was confirmed by sequencing (refer to Transparent Methods), and the knockouts were confirmed by qPCR (Figure 7B, right panel). Wt clones were able to differentiate into beating cardiomyocytes, whereas all four KO lines failed to do so (Video S2, top panels for Wt clones and bottom panels for KO clones). Calcium cycling was also impaired in the Asb2-null derived hESCs (Video S3, left panel for Wt and right panel for KO, and Figure S3B). Two Wt and two KO lines were used for further investigation. qPCR analysis on RNA from cardiomyocytes derived from these cells iScience 23, 100959, March 27, 2020 11 A C B D Figure 7. ASB2 Is Essential for Human Embryonic Stem Cell (hESC)-derived Cardiomyocytes Differentiation (A) qPCR analysis of RNA from hESC-derived cardiomyocytes at different stages of differentiation. Note that both ASB2 variants are expressed at the different stages. N = 4 for D0, n = 5 for all other stages. Data are represented as mean G SEM. (B) qPCR analysis of RNA extracted from two different wild-type (Wt) clones and two ASB2 mutant (KO) clones d7 (left) and d15 (right) differentiated hESCs. Note reduced Troponin T (differentiation marker) transcripts in the mutants at d15 and no difference in MESP1 and NKX2-5 (cardiac progenitor markers) levels at d7. N = 3 per sample for d7 and N = 4 per sample for d14. Data are represented as mean G SEM. * = p < 0.05. Two-way ANOVA was used for analysis using GraphPad Prism. p < 0.05 is considered statistically significant. (C and D) Immunostaining of d8 (C) and d15 (D) differentiated Wt and ASB2 mutant cells (KO). Note reduced Troponin T (red)-positive mutant cells at d15 but no difference in NKX2-5 (green) levels between Wt and mutant cells at both stages. DAPI (blue) marks all nuclei. Scale bar is equivalent to 8 mm in (C) and 10 mm in (D) as indicated. shows that cardiac Troponin T transcript levels (TNNT2, marker of cardiomyocyte differentiation) are greatly reduced in the KO lines at d15 of differentiation (Figure 7B, right). On the other hand, both NKX2-5 and MESP1 (markers of cardiac progenitors) are normally expressed at d7 of differentiation (Figure 7B, left). This was further confirmed at the protein level by immunostaining that shows great reduction in cardiac Troponin T (red) expression at d15 (Figure 7D) and normal NKX2-5 levels (green) at days 15 and 8 (Figures 7C and 7D, respec- tively). These data suggest that Asb2-null hES cells can commit to the cardiac lineage but arrest in differentiation prior to the generation of functional cardiomyocytes. We then examined if ASB2 regulation of the TGFb/SMAD signaling seen in mice hearts is conserved in the human cells. As discussed above, upon TGFb/Smad activation, the signaling transducer Smad4 is 12 iScience 23, 100959, March 27, 2020 A B C Figure 8. ASB2 Is an Upstream Regulator of TGFb/SMAD Pathway in hESC-derived Cardiomyocytes (A) Immunostaining of d15 Wt and ASB2 mutant (KO) hESC-derived cardiomyocytes using SMAD4-specific antibody (green). Note reduced level of nuclear but not total mean fluorescence intensity of SMAD4-positive cells in the KOs (quantification graphs on the right). DAPI (blue) marks all nuclei. Scale bar is equivalent to 75 mm in the first column and 25 mm in the second and third columns as indicated. *: p < 0.005 significant versus Wt. Unpaired t test was used for analysis using GraphPad Prism. (B and C) (B) Western blot analysis of Wt and ASB2 mutant (KO) hESC-derived cardiomyocytes using SMAD4, SMAD2, and pSMAD2 antibodies. Note increase of SMAD4 and pSMAD2 in the mutant clones. Data are representative of three separate experiments (C) Quantification of the western blot analysis in (B). Data are average of quantification from three separate experiments. *: significant versus Wt1 and Wt2. p < 0.05 is considered statistically significant. One-way ANOVA was used for analysis using GraphPad Prism. translocated to the nucleus to activate downstream targets. Figure 8A shows an increase in nuclear SMAD4 (green) in the ASB2-null hES-derived cardiomyocytes. The nuclear versus total SMAD4 was quantified (Fig- ure 8A, right graph) showing that nuclear SMAD4 signal is doubled in the ASB2-null cells, whereas total Smad4 levels remain the same. Western blot analysis on total protein extracts from these cells also confirms significant increase in both SMAD4 and pSMAD2 protein levels (Figures 8B and 8C). This further confirms that the TGFb/SMAD signaling pathway is activated in Asb2-null cardiomyocytes. DISCUSSION In this study, we provide strong evidence for the role of Asb2 in controlling heart morphogenesis partly through its regulation of the actin-binding protein, Filamin A (Flna), and Tgfb/Smad signaling. We further show that this regulation is part of the DORV disease pathogenesis. Using CUBIC clearing technique combined with immunofluorescence and confocal microscopy, we show that the Asb2-mutant hearts have shorter heart tubes and do not form the fully looped helical structure. RNA-seq analysis also reveals that a number of genes that have been linked to cardiac looping defects iScience 23, 100959, March 27, 2020 13 are altered in the Asb2-mutant hearts. Recent morphological analysis of Asb2 null embryos suggested that cardiac looping in the total body null is largely intact (Me´ tais et al., 2018). To examine this more carefully, we exploited recent advances in tissue clearing coupled to optical sectioning and 3D reconstruction. This anal- ysis of the intact embryos, however, allows us to refine these findings and to examine the Asb2 mutant hearts more thoroughly and at a slightly later point in development. Although the hearts do start to loop, they arrest early on before making it to the helical fully looped heart. Measurement of the heart tube length reveals shorter heart tubes in the mutant hearts, which could explain the inability of the heart to fully form the helical structure. These data further reveal the important role of Asb2 regulation of cardi- omyocyte differentiation on the normal growth of the heart tube. We show that the CUBIC technique combined with immunofluorescence/confocal microscopy has distinct advantages over traditional morphological analysis for the phenotypic analysis of mouse embryos and allows for the detection of subtle phenotypes and morphological abnormalities. Our data further reveal that Filamin A (Flna) is aberrantly overexpressed in the Asb2-mutant cardiomyo- cytes that normally do not express Flna protein. This is consistent with the data that Metais et al. reported. We also show that this regulation is dose dependent. We further show that Asb2-Flna regulate Tgfb-Smad signaling. Nuclear pSmad2 is overexpressed in the Asb2-mutant hearts consistent with the upregulation of this signaling pathways. Its levels are restored to normal in the Asb2.Flna double mutants further showing that Asb2 regulates SMAD signaling through the Flna pathway. RNA-seq analysis also reveals that regula- tion of the Tgfb-Smad pathway in the Asb2-mutant hearts and the Foxa genes, which are downstream effectors of the Tgfb/Smad signaling (Tang et al., 2011), is in fact restored to normal in the Asb2-Flna dou- ble mutants. Flna has been previously shown to associate with Smad2 signaling (Sasaki et al., 2001). Moreover, Tgfb/Smad2/3 signaling is impaired in the multivalve degeneration disorder, X-linked myxoma- tous valvular dystrophy, in which FLNA mutations were reported (Geirsson et al., 2012; Norris et al., 2010). Our data provide further evidence for regulation of the Tgfb/Smad cycle by Flna and show that Asb2 is upstream of this regulatory pathway in the developing heart. Using human embryonic stem cell (hESC)-derived cardiomyocytes, we further show that the Asb2 role in embryonic heart differentiation is conserved in humans. Although ASB2-null hESCs are able to form cardiac progenitor cells (marked by expression of MESP1 and NKX2-5), they have an impaired ability to differen- tiate into beating TNNT+ cardiomyocytes. It is important to note here that the difference between Troponin T levels in the Asb2-null hESCs and the Asb2 mutants in vivo could be due to the total knockout in the cells that is more severe than the conditional in vivo knockout. Additionally, the cell system lacks signaling coming from the endocardium, which could also explain this difference. These results demon- strate that, in human PSCs differentiating in vitro, ASB2-mediated targeted degradation is required for the differentiation from NKX2-5+ progenitors to beating TNNT+ cardiomyocytes and that deletion of ASB2 results in a differentiation arrest at the progenitor stage. Moreover, the finding that these cells have increased levels of SMAD4 and pSMAD2, markers of TGFb/SMAD pathway activation, provides further evidence that ASB2 is an upstream regulator of the TGFb/SMAD pathway during the differentiation of human cardiomyocytes. These considerations become increasingly important given the potential of pluripotent stem cell-derived CMs to serve as a renewable cell source for cardiac regeneration in the injured heart. Given that Flna is a direct target of Asb2 that is aberrantly upregulated in Asb2-mutant cardiomyocytes, we then investigated whether Asb2 cardiac mutant embryonic lethality can be rescued by the concurrent deletion of Flna. Accordingly, we generated AHF-Cre.Asb2fl/fl.Flnafl/y double mutants. Our data show that, as opposed to the AHF-Cre.Asb2fl/fl single mutants that die by E11.5, the AHF-Cre.Asb2fl/fl.Flnafl/y double mutants are born with the expected Mendelian ratios (Figure 5B) but die shortly after birth. This suggests partial rescue of lethality seen in the AHF-Cre.Asb2fl/fl mutant hearts. Of note, we also generated Nkx2-5+/Cre.Asb2fl/fl.Flnafl/Y double mutants that did not rescue the Asb2 lethality (Figure S2A), suggesting a greater Asb2 dependency or a more complex phenotype in these mice. These findings are not surprising owing to earlier and broader expression domain of the Nxk2-5Cre compared with the AHF-Cre line. Moreover, AHF-Asb2-mutant mice with one Flna allele (AHF-Cre.Asb2fl/fl.Flnafl/x) sometimes also survive to P0, albeit at significantly lower than expected ratios (Fig- ure 5B). Not only do these results suggest a dose dependency of Asb2 and its target Flna but they also allow us to identify DORV associated with ventricular septal defect (VSD) as a penetrant cardiac phenotype. More inter- estingly, this phenotype was rescued when Flna was abrogated by the concurrent deletion of Flna showing that both Asb2 and Flna play a functional role in the pathogenesis of DORV. 14 iScience 23, 100959, March 27, 2020 During cardiac development, the heart first forms as a primitive heart tube that then elongates and starts to loop by addition of cells from the anterior, posterior, and second heart field at both the venous and arterial poles. At the onset of looping, left-right asymmetry in the heart becomes morphologically evident and any defects in this process can lead to complex congenital heart problems, including DORV and VSDs (Rams- dell, 2005). The heart is the first organ to break the left-right symmetry in the developing embryo, and it has been shown that the actin-cytoskeleton is fundamental for laterality and modulation associated with heart looping. It was shown to provide the built-in mechanism required for cells to acquire left-right asymmetry (Linask and Vanauker, 2007; Tee et al., 2015). Abnormalities in the control of construction of the cytoskel- eton has been previously shown to result in looping defects and ultimately lead to congenital heart prob- lems (Langdon et al., 2012; Linask and Vanauker, 2007). Our data and the data from Metais et al. provide solid evidence for the Asb2-Flna regulation of the actin cytoskeleton during heart morphogenesis (Me´ tais et al., 2018). Our data extend this regulation to show that it is important for normal heart tube and OFT development and, if perturbed, leads to DORV and VSD in the developing mammalian heart. Additionally, we suggest a mechanism where Asb2 downregulation leads to abnormal overexpression of Flna that ulti- mately leads to increased activity of the Tgfb/Smad2 signaling in the myocardium thus causing growth/ elongation defects and DORV in the mammalian heart. Indeed, prior reports have implicated the Tgfb superfamily and Smad2/3 in left-right asymmetry, and Tgfb2 mutant mice have been shown to develop DORV and die right after birth (Azhar et al., 2003; Sanford et al., 1997; Whitman and Mercola, 2001). In humans, a missense mutation in Flna (c.5290G>A (p.A1764T) has been reported in a patient with DORV (de Wit et al., 2011). Since missense mutations can result in both loss and gain of function, future studies will be required to determine the effect of this mutation on Flna expression and function. Thus, our data demonstrate a link between targeted protein turnover and the development of DORV and highlights the potential of the ASB2/FLNA axis as a diagnostic, prognostic, and/or therapeutic target for patients with DORV. Limitations of the Study Although our data show that the role of Asb2 in heart morphogenesis is conserved between mice and human, a limitation is that the in vivo murine system is a conditional knockout compared with the total knockout in the human cells system. Additionally, as we know, a cross talk between the endocardium and the myocardium occurs during heart morphogenesis, and this again is lacking in our human cell system. METHODS All methods can be found in the accompanying Transparent Methods supplemental file. DATA AND CODE AVAILABILITY RNA-seq data have been deposited in NCBI’s Gene Expression Omnibus (GEO). The accession number for the RNA-seq data reported in this paper is GEO: GSE145495. SUPPLEMENTAL INFORMATION Supplemental Information can be found online at https://doi.org/10.1016/j.isci.2020.100959. ACKNOWLEDGMENTS We thank all members of the Domian Lab for valuable insight and suggestions. We also thank the his- tology core at Dana Farber Cancer Institute/Harvard Medical School and the NextGen Sequencing core at Massachusetts General Hospital for technical support. This work was supported by the American Heart Association (17GRNT33630170), the Centre National de la Recherche Scientifique, and the Univer- sity of Toulouse. A.Y. is a recipient of the Fund for Medical Discovery (FMD) Award from the Massachu- setts General Hospital/Harvard Medical School. AUTHOR CONTRIBUTIONS Conceptualization, A.Y. and I.J.D.; Methodology, A.Y. and I.J.D.; Investigation A.Y., D.H., N.M., J.W.B., and S.D.; Writing – Original Draft, A.Y.; Writing – Review & Editing, A.Y., P.G.L., C.M.-L, P.T.L., and I.J.D.; Visu- alization, A.Y.; Supervision, A.Y. and I.J.D.; Funding Acquisition, A.Y. and I.J.D. iScience 23, 100959, March 27, 2020 15 DECLARATION OF INTERESTS The authors declare no competing interests. Received: August 16, 2019 Revised: December 17, 2019 Accepted: February 26, 2020 Published: March 27, 2020 REFERENCES Azhar, M., Schultz, J.E.J., Grupp, I., Dorn, G.W., Meneton, P., Molin, D.G.M., Gittenberger-de Groot, A.C., and Doetschman, T. (2003). Transforming growth factor beta in cardiovascular development and function. Cytokine Growth Factor Rev. 14, 391–407. Bardot, E., Calderon, D., Santoriello, F., Han, S., Cheung, K., Jadhav, B., Burtscher, I., Artap, S., Jain, R., Epstein, J., et al. (2017). Foxa2 identifies a cardiac progenitor population with ventricular differentiation potential. Nat. Commun. 8, 14428. Bello, N.F., Lamsoul, I., Heuze´ , M.L., Me´ tais, A., Moreaux, G., Calderwood, D.A., Duprez, D., Moog-Lutz, C., and Lutz, P.G. (2009). The E3 ubiquitin ligase specificity subunit ASB2beta is a novel regulator of muscle differentiation that targets filamin B to proteasomal degradation. Cell Death Differ. 16, 921–932. Bruneau, B.G. (2008). The developmental genetics of congenital heart disease. Nature 451, 943–948. Bulteau, A.L., Lundberg, K.C., Humphries, K.M., Sadek, H.A., Szweda, P.A., Friguet, B., and Szweda, L.I. (2001). Oxidative modification and inactivation of the proteasome during coronary occlusion/reperfusion. J. Biol. Chem. 276, 30057–30063. Chen, J.N., van Eeden, F.J., Warren, K.S., Chin, A., Nu¨ sslein-Volhard, C., Haffter, P., and Fishman, M.C. (1997). Left-right pattern of cardiac BMP4 may drive asymmetry of the heart in zebrafish. Development 124, 4373–4382. Davey, J.R., Watt, K.I., Parker, B.L., Chaudhuri, R., Ryall, J.G., Cunningham, L., Qian, H., Sartorelli, V., Sandri, M., Chamberlain, J., et al. (2016). Integrated expression analysis of muscle hypertrophy identifies Asb2 as a negative regulator of muscle mass. JCI Insight 1, https:// doi.org/10.1172/jci.insight.85477. de Wit, M.C.Y., de Coo, I.F.M., Lequin, M.H., Halley, D.J.J., Roos-Hesselink, J.W., and Mancini, G.M.S. (2011). Combined cardiological and neurological abnormalities due to filamin A gene mutation. Clin. Res. Cardiol. 100, 45–50. de Wit, M.C.Y., Kros, J.M., Halley, D.J.J., de Coo, I.F.M., Verdijk, R., Jacobs, B.C., and Mancini, G.M.S. (2009). Filamin A mutation, a common cause for periventricular heterotopia, aneurysms and cardiac defects. J. Neurol. Neurosurg. Psychiatry 80, 426–428. Domian, I.J., Chiravuri, M., van der Meer, P., Feinberg, A.W., Shi, X., Shao, Y., Wu, S.M., Parker, K.K., and Chien, K.R. (2009). Generation of functional ventricular heart muscle from mouse ventricular progenitor cells. Science 326, 426–429. 16 iScience 23, 100959, March 27, 2020 Fan, Y., Xie, P., Zhang, T., Zhang, H., Gu, D., She, M., and Li, H. (2008). Regulation of the stability and transcriptional activity of NFATc4 by ubiquitination. FEBS Lett. 582, 4008–4014. Feng, Y., Chen, M.H., Moskowitz, I.P., Mendonza, A.M., Vidali, L., Nakamura, F., Kwiatkowski, D.J., and Walsh, C.A. (2006). Filamin A (FLNA) is required for cell-cell contact in vascular development and cardiac morphogenesis. Proc. Natl. Acad. Sci. U S A 103, 19836–19841. Fujita, M., Mitsuhashi, H., Isogai, S., Nakata, T., Kawakami, A., Nonaka, I., Noguchi, S., Hayashi, Y.K., Nishino, I., and Kudo, A. (2012). Filamin C plays an essential role in the maintenance of the structural integrity of cardiac and skeletal muscles, revealed by the medaka mutant zacro. Dev. Biol. 361, 79–89. Fukuda, R., Gunawan, F., Beisaw, A., Jimenez- Amilburu, V., Maischein, H.-M., Kostin, S., Kawakami, K., and Stainier, D.Y.R. (2017). Proteolysis regulates cardiomyocyte maturation and tissue integration. Nat. Commun. 8, 14495. Geirsson, A., Singh, M., Ali, R., Abbas, H., Li, W., Sanchez, J.A., Hashim, S., and Tellides, G. (2012). Modulation of transforming growth factor-b signaling and extracellular matrix production in myxomatous mitral valves by angiotensin II receptor blockers. Circulation 126, S189–S197. Glickman, M.H., and Ciechanover, A. (2002). The ubiquitin-proteasome proteolytic pathway: destruction for the sake of construction. Physiol. Rev. 82, 373–428. Guibal, F.C., Moog-Lutz, C., Smolewski, P., Di Gioia, Y., Darzynkiewicz, Z., Lutz, P.G., and Cayre, Y.E. (2002). ASB-2 inhibits growth and promotes commitment in myeloid leukemia cells. J. Biol. Chem. 277, 218–224. Heuze´ , M.L., Lamsoul, I., Baldassarre, M., Lad, Y., Le´ veˆ que, S., Razinia, Z., Moog-Lutz, C., Calderwood, D.A., and Lutz, P.G. (2008). ASB2 targets filamins A and B to proteasomal degradation. Blood 112, 5130–5140. Hu, D., Linders, A., Yamak, A., Correia, C., Kijlstra, J.D., Garakani, A., Xiao, L., Milan, D.J., van der Meer, P., Serra, M., et al. (2018). Metabolic maturation of human pluripotent stem cell- derived cardiomyocytes by inhibition of HIF1a and LDHA. Circ. Res. 123, 1066–1079. Jung, T., Catalgol, B., and Grune, T. (2009). The proteasomal system. Mol. Aspects Med. 30, 191–296. Kedar, V., McDonough, H., Arya, R., Li, H.-H., Rockman, H.A., and Patterson, C. (2004). Muscle- specific RING finger 1 is a bona fide ubiquitin ligase that degrades cardiac troponin I. Proc. Natl. Acad. Sci. U S A 101, 18135–18140. (cid:1) Capek, M., Radochova´ , B., Jana´ (cid:1)cek, Kolesova´ , H., J., and Sedmera, D. (2016). Comparison of different tissue clearing methods and 3D imaging techniques for visualization of GFP-expressing mouse embryos and embryonic hearts. Histochem. Cell Biol. 146, 141–152. Kyndt, F., Gueffet, J.-P., Probst, V., Jaafar, P., Legendre, A., Le Bouffant, F., Toquet, C., Roy, E., McGregor, L., Lynch, S.A., et al. (2007). Mutations in the gene encoding filamin A as a cause for familial cardiac valvular dystrophy. Circulation 115, 40–49. Laine, A., and Ronai, Z. (2005). Ubiquitin chains in the ladder of MAPK signaling. Sci. STKE 2005, re5. Langdon, Y., Tandon, P., Paden, E., Duddy, J., Taylor, J.M., and Conlon, F.L. (2012). SHP-2 acts via ROCK to regulate the cardiac actin cytoskeleton. Development 139, 948–957. Le Garrec, J.-F., Domı´nguez, J.N., Desgrange, A., Ivanovitch, K.D., Raphae¨ l, E., Bangham, J.A., Torres, M., Coen, E., Mohun, T.J., and Meilhac, S.M. (2017). A predictive model of asymmetric morphogenesis from 3D reconstructions of mouse heart looping dynamics. Elife 6, https:// doi.org/10.7554/eLife.28951. Li, J., Horak, K.M., Su, H., Sanbe, A., Robbins, J., and Wang, X. (2011). Enhancement of proteasomal function protects against cardiac proteinopathy and ischemia/reperfusion injury in mice. J. Clin. Invest. 121, 3689–3700. Linask, K.K., and Vanauker, M. (2007). A role for the cytoskeleton in heart looping. ScientificWorldJournal 7, 280–298. Lombardi, R., Dong, J., Rodriguez, G., Bell, A., Leung, T.K., Schwartz, R.J., Willerson, J.T., Brugada, R., and Marian, A.J. (2009). Genetic fate mapping identifies second heart field progenitor cells as a source of adipocytes in arrhythmogenic right ventricular cardiomyopathy. Circ. Res. 104, 1076–1084. Ma¨ nner, J. (2009). The anatomy of cardiac looping: a step towards the understanding of the morphogenesis of several forms of congenital cardiac malformations. Clin. Anat. 22, 21–35. Martinsen, B.J., and Lohr, J.L. (2015). Cardiac development. In Handbook of Cardiac Anatomy, Physiology, and Devices, P.A. Laizzo, ed. (Springer), pp. 23–33. Mercola, M., Ruiz-Lozano, P., and Schneider, M.D. (2011). Cardiac muscle regeneration: lessons from development. Genes Dev. 25, 299–309. Me´ tais, A., Lamsoul, I., Melet, A., Uttenweiler- Joseph, S., Poincloux, R., Stefanovic, S., Valie` re, A., Gonzalez de Peredo, A., Stella, A., Burlet- Schiltz, O., et al. (2018). Asb2a-filamin A axis is essential for actin cytoskeleton remodeling during heart development. Circ. Res. 122, e34–e48. Mine, N., Anderson, R.M., and Klingensmith, J. (2008). BMP antagonism is required in both the node and lateral plate mesoderm for mammalian left-right axis establishment. Development 135, 2425–2434. Murdaca, J., Treins, C., Monthoue¨ l-Kartmann, M.-N., Pontier-Bres, R., Kumar, S., Van Obberghen, E., and Giorgetti-Peraldi, S. (2004). Grb10 prevents Nedd4-mediated vascular endothelial growth factor receptor-2 degradation. J. Biol. Chem. 279, 26754–26761. Nemer, M. (2008). Genetic insights into normal and abnormal heart development. Cardiovasc. Pathol. 17, 48–54. Norris, R.A., Moreno-Rodriguez, R., Wessels, A., Merot, J., Bruneval, P., Chester, A.H., Yacoub, M.H., Hage` ge, A., Slaugenhaupt, S.A., Aikawa, E., et al. (2010). Expression of the familial cardiac valvular dystrophy gene, filamin-A, during heart morphogenesis. Dev. Dyn. 239, 2118–2127. Obler, D., Juraszek, A.L., Smoot, L.B., and Natowicz, M.R. (2008). Double outlet right ventricle: aetiologies and associations. J. Med. Genet. 45, 481–497. Pagan, J., Seto, T., Pagano, M., and Cittadini, A. (2013). Role of the ubiquitin proteasome system in the heart. Circ. Res. 112, 1046–1058. Powell, S.R., Davies, K.J.A., and Divald, A. (2007). Optimal determination of heart tissue 26S- proteasome activity requires maximal stimulating ATP concentrations. J. Mol. Cell. Cardiol. 42, 265–269. Ramsdell, A.F. (2005). Left-right asymmetry and congenital cardiac defects: getting to the heart of the matter in vertebrate left-right axis determination. Dev. Biol. 288, 1–20. single-cell resolution by tissue decolorization. Cell 159, 911–924. Ribeiro, I., Kawakami, Y., Bu¨ scher, D., Raya, A., Rodrı´guez-Leo´ n, J., Morita, M., Rodrı´guez Esteban, C., and Izpisu´ a Belmonte, J.C. (2007). Tbx2 and Tbx3 regulate the dynamics of cell proliferation during heart remodeling. PLoS One 2, e398. Sanford, L.P., Ormsby, I., Gittenberger-de Groot, A.C., Sariola, H., Friedman, R., Boivin, G.P., Cardell, E.L., and Doetschman, T. (1997). TGFbeta2 knockout mice have multiple developmental defects that are non-overlapping with other TGFbeta knockout phenotypes. Development 124, 2659–2670. Sasaki, A., Masuda, Y., Ohta, Y., Ikeda, K., and Watanabe, K. (2001). Filamin associates with Smads and regulates transforming growth factor- beta signaling. J. Biol. Chem. 276, 17871–17877. Shi, Y., and Massague´ , J. (2003). Mechanisms of TGF-beta signaling from cell membrane to the nucleus. Cell 113, 685–700. Spaich, S., Will, R.D., Just, S., Spaich, S., Kuhn, C., Frank, D., Berger, I.M., Wiemann, S., Korn, B., Koegl, M., et al. (2012). F-box and leucine-rich repeat protein 22 is a cardiac-enriched F-box protein that regulates sarcomeric protein turnover and is essential for maintenance of contractile function in vivo. Circ. Res. 111, 1504–1516. Spinner, C.A., Uttenweiler-Joseph, S., Metais, A., Stella, A., Burlet-Schiltz, O., Moog-Lutz, C., Lamsoul, I., and Lutz, P.G. (2015). Substrates of the ASB2a E3 ubiquitin ligase in dendritic cells. Sci. Rep. 5, 16269. Srivastava, D. (2006). Making or breaking the heart: from lineage determination to morphogenesis. Cell 126, 1037–1048. Tainaka, K., Kubota, S.I., Suyama, T.Q., Susaki, E.A., Perrin, D., Ukai-Tadenuma, M., Ukai, H., and Ueda, H.R. (2014). Whole-body imaging with Tang, Y., Shu, G., Yuan, X., Jing, N., and Song, J. (2011). FOXA2 functions as a suppressor of tumor metastasis by inhibition of epithelial-to- mesenchymal transition in human lung cancers. Cell Res. 21, 316–326. Tee, Y.H., Shemesh, T., Thiagarajan, V., Hariadi, R.F., Anderson, K.L., Page, C., Volkmann, N., Hanein, D., Sivaramakrishnan, S., Kozlov, M.M., and Bershadsky, A.D. (2015). Cellular chirality arising from the self-organization of the actin cytoskeleton. Nat. Cell Biol. 17, 445–457. Valde´ s-Mas, R., Gutie´ rrez-Ferna´ ndez, A., Go´ mez, J., Coto, E., Astudillo, A., Puente, D.A., Reguero, J.R., A´ lvarez, V., Morı´s, C., Leo´ n, D., et al. (2014). Mutations in filamin C cause a new form of familial hypertrophic cardiomyopathy. Nat. Commun. 5, 5326. van der Flier, A., and Sonnenberg, A. (2001). Structural and functional aspects of filamins. Biochim. Biophys. Acta 1538, 99–117. Vincentz, J.W., Barnes, R.M., and Firulli, A.B. (2011). Hand factors as regulators of cardiac morphogenesis and implications for congenital heart defects. Birth Defects Res. A Clin. Mol. Teratol. 91, 485–494. Whitman, M., and Mercola, M. (2001). TGF-beta superfamily signaling and left-right asymmetry. Sci. STKE 2001, re1. Yamak, A., and Nemer, M. (2015). Role of embryonic and differentiated cells in cardiac development. In Biomaterials for Cardiac Regeneration, E.J. Suuronen and M. Ruel, eds. (Springer), pp. 37–70. Zhou, A.-X., Toylu, A., Nallapalli, R.K., Nilsson, G., Atabey, N., Heldin, C.-H., Bore´ n, J., Bergo, M.O., and Akyu¨ rek, L.M. (2011). Filamin a mediates HGF/c-MET signaling in tumor cell migration. Int. J. Cancer 128, 839–846. iScience 23, 100959, March 27, 2020 17 iScience, Volume 23 Supplemental Information Loss of Asb2 Impairs Cardiomyocyte Differentiation and Leads to Congenital Double Outlet Right Ventricle Abir Yamak, Dongjian Hu, Nikhil Mittal, Jan W. Buikema, Sheraz Ditta, Pierre G. Lutz, Christel Moog-Lutz, Patrick T. Ellinor, and Ibrahim J. Domian Supplementary Data Transparent Methods Animals. All animal experimentations were carried out in accordance with institutional guidelines for animal care. Experiments were approved by the Massachusetts General Hospital’s Subcommittee on Research Animal Care (SRAC), which serves as the Institutional Animal Care and Use Committee (IACUC) as required by the Public Health Service (PHS) Policy on Humane Welfare Regulations. The program and facilities have been fully accredited by the American Association for the Accreditation of Laboratory Animal Care (AAALAC) since July 30, 1993. The institutional assurance number with the Office for Protection from Research Risks at the N.I.H. is DI6-00361. All mice lines were kept on a C57BL/6 background. Approximately, 20 AHF-Cre, 20 Nkx2-5+/Cre, 100 Asb2fl/fl and 100 Asb2fl/fl.Flnafl/fl mice were used. To isolate embryos from pregnant females, cervical dislocation was used for euthanasia which is required for embryo collection in mice. Sex of the embryos was not an influence in this study due to the very early developmental stage. Embryos were analyzed at E8.5, E9.5, E10.5, and E11.5 as indicated in the results section where applicable. For the double outlet right ventricle analysis, hearts of E16.5 embryos were used. Generation of Asb2 and Flna knockout embryos. Asb2fl/fl or Asb2fl/fl.Flnafl/fl females were mated with Nkx2-5+/Cre or AHF-Cre male mice and plugs were checked on a daily basis. The day a plug is seen is considered embryonic day e0.5. Asb2fl/fl, Flnafl/fl, Nkx2-5+/Cre and AHF-Cre mice are previously described (Lamsoul et al., 2013; Lombardi et al., 2009; Pinto et al., 2014). Mice genotypes (adult and embryos) were determine by PCR genotyping. Genotyping oligos used are: Flna flox: 5’ TCT TCC TCT TTC AGC TGG 3’and 5’ ACA ACT GCT GCT CCA GAG 3’; Asb2 flox: 5’ CAGTGTCTGCTCTGAGGTCTCTC 3’ and 5’ CAATCTCTCCCTGGTAGAAACAGTTTGG 3’; Nkx2-5 Cre: 5’ GATTAGCTTAAGCGGAGCTGGGTGTCC 3’ and 5’ GCCGCATAACCAGTGAAACAGCATTGC 3’; AHF-Cre: 5’ CCAGGCAAAGGCAAGAATAA 3’ and 5’ ATGTTTAGCTGGCCCAAATG 3’. Immunohistochemistry. Immunofluorescence was done as previously described (Domian et al., 2009). Tissues were permeabilized with 0.3% Triton and antigen retrieval was done using citrate buffer. Tissues were blocked with goat or donkey serum and primary antibodies were incubated overnight at 4oC. Secondary antibodies linked to appropriate alexa fluor were incubated for 1 hour at room temperature. Excess antibodies were washed with Phosphate buffer saline with 0.2% tween-20. Tissues were mounted with prolong gold anti-fade mounting media. Antibodies used were: Asb2 (Abcam, ab13710); Filamin a (Abcam, ab76289); Nkx2-5 (Invitrogen, PA5-49431); pSmad2(Millipore, AB3849); Troponin T (Thermo- Scientific, MA5-12960; SMAD4 (Proteintech, 10231-1-AP). Masson Trichrome Staining was done on paraffin heart sections using the American Mastertech Scientific kit (Item No. KTMTR) according to the manufacturer’s protocol. Paraffin sections were deparaffinized with 3 rounds of xylene followed by rehydration with serial dilutions of ethanol baths prior to staining. Outflow tract measurements were done on 2D images using ImageJ. The landmarks used for measurement are as shown in supplementary figure 1D. CUBIC clearing and Immunostaining. Embryos were immersed in CUBIC-1 solution (25% urea, 15% TritonX-100, 25% N,N,N,N-tetrakis(2-hydroxypropyl)ethyl-enediamine) at 37oC with gentle shaking till efficiently cleared (2-5 days depending on developmental stage). Following clearing, embryos were washed thoroughly with PBS and stained with Troponin T and/or Filamin A antibodies for 4-5 days (at 4oC), washed with PBS and then incubated with the corresponding secondary antibodies coupled to Alexa Fluor 488 or 546 for additional 3-4 days (at 4oC). DAPI was added to CUBIC-1 solution and the following PBS washes to mark nuclei. Following staining, embryos were then cleared with CUBIC-2 solution (50% sucrose, 25% urea and 10% 2,2’,2’-nitrilotriethanol) for 1-2 days at 37oC with gentle shaking and then immediately transferred to immersion oil and imaged with laser confocal microscopy (Leica TCS SP8). 90-120 z-stacks were taken for each embryo that were then used to generate the 3D reconstructions using either the Leica software or image J. The 3D images were then further analyzed for phenotypic defects. At least 5 embryos were analyzed for each condition. The clearing/staining technique was adapted from the established protocol by Kolesova et al (Kolesová et al., 2016). Heart tube measurements were done on 3D images using ImageJ. The landmarks used for measuring the tube’s length are as described in Le Garrec et al paper (Le Garrec et al., 2017). RNA Extraction and qPCR. RNA extraction was done using the Qiagen RNeasy Micro Kit (Cat No. 74004) according to the manufacturer’s protocol. qPCR analysis was done using the Applied Biosystems PowerUp SYBR Green Master mix (Cat No. A25742) according to manufacturer’s protocol. Oligos used were: msAsb2a: 5’ GCTCTGTTTCACTCTGGCTCT 3’ and 5’ CTTCAGCACGGGGTCCATAG 3’; msAsb2b: 5’ AACCACCAGCCAGGACATTT 3’ and 5’ ACTTCTGCATGACCCCTTGG 3’; huASB2V1: 5’ ATTGGGCAGGAGGAGTACAG 3’ and 5’ AACTCTCAGGAGGTGCAGT 3’; huASB2V2: 5’ ATGACCCGCTTCTCCTATGC 3’ and 5’ CGAACTCTCAGGAGGTGCAG 3’. huTNNT2: 5’ ACTTGGAGGCAGAGAAGTTCG 3’ and 5’ CCCGGTGACTTTAGCCTTCC 3’; huNKX2-5: 5’ CGCACAGCTCTTTCTTTTCGG 3’ and 5’ CGCCTTCTATCCACGTGCC 3’; huMESP1: 5’ CTTTTTGGCCTCAGCACCTTC 3’ and 5’ AGTGTCTAGCCCTATGGGTCC 3’. RNA Sequencing. RNA was extracted from e9.5 embryo hearts using the Qiagen RNeasy Micro Kit (Cat No. 74004) and sent to the MGH Next Generation sequencing core. The libraries were sequenced using illumina HiSeq platform. Splice-aware alignment program STAR was used to map the sample sequencing reads to the Mus musculus mm10 reference genome. Gene expression counts were calculated using HTSeq based on current Ensembl annotation for mm10. The R package “edgeR” was then employed to make differential gene expression calls. Pathway analyses were done using “MetaCore-Clarivate” and “Ingenuity Pathway Analysis-Qiagen” softwares. Human Pluripotent Stem Cell Culture and Differentiation. HUES9 hESC line (NIH Human Embryonic Stem Cell Registry Number 0022, generated by HSCI iPS Core at Harvard University) was used in generating CRISPR KO cell line. hESC culture, differentiation and dissociation protocols were based on previously published works (Hu et al., 2018). Briefly, hESCs were cultured in Essential 8 Medium (Thermo Fisher Scientific, MA) in Matrigel (BD Biosciences) coated cell culture plates. hESCs were differentiated in RPMI GlutaMAX (Thermo Fisher Scientific, MA) plus Gem21 NeuroPlex Serum-Free Supplement without insulin (Gemini Bio Products, CA) for the first 5 days. Small molecules CHIR99021 (STEMCELL Technologies, Vancouver, Canada) and IWP-4 (STEMCELL Technologies, Vancouver, Canada) were added on day 1 and 3, respectively. Differentiation media was then switched to RPMI GlutaMAX plus Gem21 NeuroPlex Serum-Free Supplement from day 7 to 10. Differentiating hESCs then underwent glucose starvation for 6 days, which resulted in highly pure populations of beating CMs. hESC-CMs were re-plated onto Matrigel coated PDMS plates for confocal imaging. Imaging procedure and analysis were done based on previously published methods (Kijlstra et al., 2015). Briefly, Fluo-4, AM (Thermo Fisher Scientific, MA) calcium indicator were incubated with hESC-CMs prior to imaging. Movies of CMs at randomly selected regions were acquired in both DIC and GFP channels at 50 frames per second for 10 seconds. Calcium transients were analyzed using ImageJ software. In vitro differentiation of the SHF-dsRed/Nkx2.5-eGFP cells was done as previously described (Domian et al., 2009). Generation of ASB2-null hESCs. We used CRISPR/Cas9 genome editing technology to generate the ASB2-null hESCs according to the described protocol (Ran et al., 2013). Guide RNAs (gRNAs) specific for hASB2 variant 1 (equivalent to Asb2β in mouse) and those common for both variants 1 & 2 (mouse Asb2β & α respectively) were designed using CRISPR design online tool, cloned into CRISPR/Cas9-GFP plasmid backbone (pSpCas9 from Addgene) and sequenced. Plasmids with the efficient gRNAs were delivered by electroporation to hESCs. Single cell CRISPR clone selection, expansion and sequencing protocols were adapted from Peters et al (Peters et al., 2008). Following FACS selection, GFP+ hESCs were plated sparsely onto Matrigel coated dishes for growing single cell clones. After 10 days, individual clones were picked, plated into 96-well plates, and sequenced. Four clones harboring ASB2 gene locus modification along with four wild type (Wt) clones were expanded and differentiated into CMs for further analysis. 6 guides were tested individually (sequences below). Guides 1 and 6 were successful in inducing the knockouts. 1: used were: Guide 5' CACCGGTTGGTACATGCAGACGCGG 5' Guides AAACCCGCGTCTGCATGTACCAACC 3’; Guide 2: 5’ CACCGGTCCGCTAGGCTCTGCTCGA and 5’ AAACTCGAGCAGAGCCTAGCGGACC 3’; Guide 3: 5' CACCGGGCCCCTTGTCTTGTCCGCT 3’ and 5’ AAACAGCGGACAAGACAAGGGGCCC 3’; Guide 4: 5' CACCGGCCCGGGCCGGCGAACTCTC 3’ and 5' AAACGAGAGTTCGCCGGCCCGGGCC 3’; Guide 5: 5' CACCGCTCCTGAGAGTTCGCCGGCC 3’ and 5' AAACGGCCGGCGAACTCTCAGGAGC 3’; Guide 6: 5' CACCGCTGCACGAGGCCGCATACTA 3’ and 5' AAACTAGTATGCGGCCTCGTGCAGC 3’ and 3’ Western blot analysis. Total protein extracts were prepared using RIPA buffer. Proteins were run on 10% TGX pre-cast gels from biorad and transferred to PVDF membranes using Trans-blot turbo transfer kit (Biorad). Membranes were blocked with 5% non-fat milk or BSA (in case of pSmad2) and primary antibodies were incubated overnight at 4oC. Secondary antibodies linked to HRP (horseradish peroxidase) were incubated for 1 hour at room temperature and signal was revealed using super signal west femto or pico ECL substrates (Thermo-scientific). Antibodies used were Smad2 (5339, Cell Signaling), pSmad2 (3108, cell signaling) and Smad4 (ab40759, ABCAM). Western blots were then quantified using the Image Lab software. Statistical Analysis: Standard t-test was used for the QPCR analysis and the heart tube measurements. One-way ANOVA was used for the western blot quantification as well as the percentages of Smad4 and pSmad2 positive cells where. p<0.05 is considered statistically significant. Data and Software availability: RNAseq data have been deposited in NCBI’s Gene Expression Omnibus (GEO) and can be accessed through GEO Series accession number GSE145495. B. DAPI Troponin T Control Nkx2-5+/Cre.Asb2fl/+ Nkx2-5+/Cre.Asb2fl/fl 25μm + / fl 2 b s A . e r C / + 5 - 2 x k N fl / fl 2 b s A . e r C / + 5 - 2 x k N l o r t n o C H D P A G / 2 b s A l o r t n o C o t e v i t a e r l 1 0.8 0.6 0.4 0.2 0 Asb2 GAPDH E9.5 100μm A. C. 75 50 37 25 l o r t n o C t n a t u M E9.5 100μm D. Control AHF-Cre.Asb2fl/fl DAPI Troponin OFT OFT Ε10.5 100μm 100μm ) m μ ( h t g n e l T F O 600 400 200 0 * Control AHF-Cre.Asb2fl/fl Supplementary figure 1, related to figure 2. A. Western Blot analysis on hearts of e9.5 Nkx2- 5+/Cre.Asb2fl/+, Nkx2-5+/Cre.Asb2fl/fl, and their control littermates using Asb2 antibody and GAPDH for loading control. Notice reduced Asb2 protein levels in the heterozygous mice (fl/+) and the complete loss of Asb2 in the knockout mice (fl/fl) (quantification analysis on the rights) (5-6 hearts were uses per condition). B & C. CUBIC/Immunofluorescence of E9.5 mice embryos. B. High magnification showing the cardiac myocardial region of an E9.5 mouse embryo cleared with CUBIC and immuno-stained for Troponin T (green). Blue marls DAPI. Note the visible striations (yellow arrows). Scale bar is equivalent to 25µm. C. Serial sections of Control (top) and Mutant (bottom) E9.5 cleared/stained mice embryos, showing the heart region. Troponin T (green) was used to mark the myocardium. Note the bulging in the Control heart (right arrow) which is missing in the Mutant. D. Immunohistochemistry on E10.5 AHF-Cre.Asb2 hearts (AHF- Cre.Asb2fl/fl, right panel) and littermate control (left panel) using Troponin-T (gree)-specific antibody. Note the shorter outflow tract (OFT) of the AHF-Cre.Asb2fl/fl heart. Scale bar is equivalent to 100µm. 4 Control and 3 knockout hearts were analyzed. *: p<0.005 significant vs. control. Unpaired t-test was used for analysis using Graphpad Prism. Supplementary Figure 1, related for figure 2. A. Western Blot analysis on hearts of e9.5 Nkx+/Cre.Asb2fl/+, Nkx+/Cre.Asb2fl/fl and their control littermates using Asb2 antibody and GAPDH for loading control. Notice reduced Asb2 protein levels in the heterozygous mice (fl/+) and the complete loss of Asb2 in the knock out mice (fl/fl) (quantification analysis on the right) (5-6 hearts were used per condition). B. & C. CUBIC/Immunofluorescence in e9.5 mice embryos. B&C. CUBIC/Immunofluorescence in e9.5 mice embryos. B. High magnification showing the cardiac myocardial region of an E9.5 mouse embryo cleared with CUBIC and immuno-stained for TroponinT (green). Blue marks DAPI. Note the visible striations (yellow arrows). Scale bar is equivalent to 25μm. C. Serial sections of Control (top) and Mutant (bottom) E9.5 cleared/stained mice embryos, showing the heart region. TroponinT (green) was used to mark the myocar- dium. Note the bulging in the Control heart (right arrow) which is missing in the Mutant. D. Immunohistochem- istry on E10.5 AHF-Cre.Abs2 hearts (AHF-Cre.Asb2fl/fl, right panel) and littermate control (left panel) using Troponin-T (green)-specific antibody. Note the shorter outflow tract (OFT) of the AHF-Cre.Asb2fl/fl heart. Scale bar is equivalent to 100μm. 4 Control and 3 knockout hearts were analyzed. *: p<0.005 significant vs control. Unpaired t-test was used for analysis using Graphpad prism. A.A.A.A. B. Control Nkx2-5+/CreAsb2fl/fl.Flnafl/+ Nkx2-5+/CreAsb2fl/fl.Flnafl/y Troponin FlnA DAPI E9.5 0.2mm 0.2mm 0.2mm E10.5 0.4mm 0.4mm 0.4mm C. D. l o r t n o C fl / fl a n F l . fl / fl 2 b s A / . e r C + 5 - 2 x k N Ε9.5 Troponin 75μm 75μm 75μm FlnA DAPI Ε9.5 75μm 75μm 75μm + / fl 2 b s A / . e r C + 5 - 2 x k N o t e v i t a e R l Asb2 * * 1.5 1.0 0.5 0.0 Nkx2-5+/Cre.Asb2fl/+ Nkx2-5+/Cre.Asb2fl/fl Nkx2-5+/Cre.Asb2fl/fl.Flnafl/fl s l e v e l A N R m e v i t a l e R 6.0 4.5 3.0 3.0 2.5 2.0 1.5 1.0 0.5 0.0 * # @ * # * * # * Shh Hand2 Foxa2 Foxa3 *: significant vs Nkx2-5+/Cre.Asb2fl/+ Nkx2-5+/Cre.Asb2fl/+ Nkx2-5+/Cre.Asb2fl/fl #: significant vs Nkx2-5+/Cre.Asb2fl/+.Flnafl/y Nkx2-5+/Cre.Asb2fl/+.Flnafl/y @: significant vs Nkx2-5+/Cre.Asb2fl/fl.Flnafl/y Nkx2-5+/Cre.Asb2fl/fl.Flnafl/y Supplementary Figure 2, related to figure 3. A. Nkx2-5+/Cre.Asb2.Flna E9.5 and E10.5 embryos. Note Supplementary Figure 2, related to figure 3. A. Nkx2-5+/Cre.Asb2.Flna E9.5 and E10.5 embryos. Note the the smaller Nkx2-5+/Cre.Asb2fl/fl.Flnafl/+ and Nkx2-5+/Cre.Asb2fl/fl.Flnafl/y at E9.5 and E10.5. Nkx2- smaller Nkx2-5+/CreAsb2fl/fl.Flnafl/+ and Nkx2-5+/CreAsb2fl/fl.Flnafl/y at E9.5 and E10.5. Nkx2-5+/CreAsb2fl/fl.Flnafl/+ and Nkx2- 5+/Cre.Asb2fl/fl.Flnafl/+ and Nkx2-5+/Cre.Asb2fl/fl.Flnafl/y often presented with pericardial edema at both stages. 5+/CreAsb2fl/+.Flnafl/y often presented with pericardial edema at both stages. 16 litters were analyzed at E9.5 and 16 litters were analyzed at E9.5 and 3 litters at E10.5. Scale bar is equivalent to 0.2mm at E9.5 and 0.4mm 3 litters at E10.5. Scale bar is equivalent to 0.2mm in E9.5 embryos and 0.4mm in E10.5 embryos as at E10.5 embryos as indicated. B. Immunofluorescence on Nkx2-5+/Cre.Asb2fl/fl.Flnafl/y double knockouts and indicated. B. Immunofluorescence on Nkx2-5+/Cre.Asb2fl/fl.Flnafl/fl double knockouts and controls using Flna controls using Flna (red) and Troponin T (green) antibodies. Note absence of Flna expression in the (red) and Troponin T (green) antibodies. Note absence of Flna expression in the myocardium of the double myocardium of the double knockouts as opposed to its expression in the myocardium of the single knockouts as opposed to its expression in the myocardium of the single knockouts in figure 3A. Scale bar is knockouts in figure 3A. Scale bar is equivalent to 75µm. C. Asb2 transcipt levels from RNAseq data showing equivalent to 75μm. C. Asb2 transcript levels from RNAseq data showing reduced Asb2 levels in the single reduced Asb2 levels in the single and double knockouts compared to the Asb2 heterozygote control. *: and double knockouts compared to the Asb2 heterozygote control. * p<0.05. One-way ANOVA was used for p<0.05. One-way ANOVA was used for analysis using Graphpad Prism. D. QPCR analysis of hearts from analysis using Graphpad prism. D. QPCR analysis of hearts from Nkx2-5+/Cre.Asb2fl/+, Nkx2-5+/Cre.Asb2fl/fl, Nkx2-5+/Cre.Asb2fl/+, Nkx2-5+/Cre.Asb2fl/fl, Nkx2-5+/Cre.Asb2fl/+.Flnafl/y, Nkx2-5+/Cre.Asb2fl/fl.Flnafl/y E9.5 mice. Nkx2-5+/Cre.Asb2fl/+.Flnafl/y, and Nkx2-5+/Cre.Asb2fl/fl.Flnafl/y E9.5 mice. n=5-6 for Nkx2-5+/Cre.Asb2fl/+; n=5 for N=5-6 for Nkx2-5+/Cre.Asb2fl/+; n=5 for Nkx2-5+/Cre.Asb2fl/fl; n=5-6 for Nkx2-5+/Cre.Asb2fl/+.Flnafl/y; n=3 for Nkx2- Nkx2-5+/Cre.Asb2fl/fl; n=5-6 for Nkx2-5+/Cre.Asb2fl/+.Flnafl/y; n=3 for Nkx2-5+/Cre.Asb2fl/fl.Flnafl/y (each sample was a 5+/Cre.Asb2fl/fl.Flnafl/y (each sample is a combination of 2-3 hears to account for littermate variability). The combination of 2-3 hearts to account for littermate variability). The selected genes are among those identified selected genes are among those identified in RNAseq analysis in fugures 2 and 3. P<0.05 is considered in RNAseq analysis in figures 2 and 3. p<0.05 is considered statistically significant. T-test was used for anly- statistically significant. T-test was used for analysis using Graphpad Prism. sis using Graphpad Prism. A. 1 e d u G i Variant 1 1 2 3a 5’UTR 6 e d u G i 4 5 6 7 8 9 10 3’UTR Variant 2 3b 4 5 6 7 8 9 10 B. 3.0 2.5 0 F F / 2.0 1.5 1.0 0 5’UTR WT KO 3’UTR 2 4 6 Time (s) 8 10 Supplementary Figure 3, related to figure 7. A. Schematic representation of the two Asb2 isoforms Supplementary Figure 3, related to figure 7. A. Schematic representation of the two ASB2 isoforms showing the location of the guides used for CRISPR/Cas9 genome editing. B. Representative calcium showing the location of the guides used for CRISPR/Cas9 genome editing. B. Representative calcium transients of hiPSC-CMs (WT: blue, KO: red). transients of hiPSC-CMs (WT: blue, KO: red).
10.1073_pnas.2221103120
RESEARCH ARTICLE | PHYSIOLOGY OPEN ACCESS Structural insights into plasmalemma vesicle-associated protein (PLVAP): Implications for vascular endothelial diaphragms and fenestrae Tao-Hsin Changa,b,1 , Fu-Lien Hsieha,b, Xiaowu Gua,b,2, Philip M. Smallwooda,b, Jennifer M. Kavranc,d, Sandra B. Gabellic,e,f,3, and Jeremy Nathansa,b,g,h,1 Contributed by Jeremy Nathans; received December 12, 2022; accepted February 20, 2023; reviewed by Daniel J. Leahy and Qun Liu In many organs, small openings across capillary endothelial cells (ECs) allow the diffusion of low–molecular weight compounds and small proteins between the blood and tissue spaces. These openings contain a diaphragm composed of radially arranged fibers, and current evidence suggests that a single-span type II transmem- brane protein, plasmalemma vesicle-associated protein-1 (PLVAP), constitutes these fibers. Here, we present the three-dimensional crystal structure of an 89-amino acid segment of the PLVAP extracellular domain (ECD) and show that it adopts a parallel dimeric alpha-helical coiled-coil configuration with five interchain disulfide bonds. The structure was solved using single-wavelength anomalous diffraction from sulfur-containing residues (sulfur SAD) to generate phase information. Biochemical and circular dichroism (CD) experiments show that a second PLVAP ECD segment also has a parallel dimeric alpha-helical configuration—presumably a coiled coil— held together with interchain disulfide bonds. Overall, ~2/3 of the ~390 amino acids within the PLVAP ECD adopt a helical configuration, as determined by CD. We also determined the sequence and epitope of MECA-32, an anti-PLVAP antibody. Taken together, these data lend strong support to the model of capillary diaphragms formulated by Tse and Stan in which approximately ten PLVAP dimers are arranged within each 60- to 80-nm-diameter opening like the spokes of a bicycle wheel. Passage of molecules through the wedge-shaped pores is presumably determined both by the length of PLVAP—i.e., the long dimension of the pore—and by the chemical properties of amino acid side chains and N-linked glycans on the solvent-accessible faces of PLVAP. vasculature | coiled-coil | single-wavelength anomalous dispersion of sulfur atoms | permeability | alpha helix The efficient movement of molecules between the intravascular space and the surrounding tissues is central to vascular function (1). Movement of molecules across capillary endothelial cells (ECs) can occur via a paracellular (between cells) pathway or via any of several tran- scellular pathways (1). The transcellular pathways include i) transporter- or channel-mediated uptake or release of small molecules (amino acids, glucose, ions, xenobiotics, etc.), ii) receptor-mediated endocytosis, as seen for iron uptake by the transferrin receptor, and iii) diffusion of small- and moderately sized molecules, such as peptide hormones, across 60- to 80-nm circular openings at the endothelial cell plasma membrane (1–3). These circular openings can either span the width of the cell where the EC is extremely thin (fenestrae) or they can serve as the mouth of transendothelial channels or transcytotic vesicular carriers (caveolae), where the EC is thicker (4). The relative prominence of these different pathways differs among vascular beds in different tissues (5, 6). In the electron microscope, the 60- to 80-nm openings are seen to possess a delicate diaphragm with a thickness of tens of nanometers (Fig. 1A). In en face views (Fig. 1 B and C), the diaphragm is observed to consist of 8 to 10 radial strands joined together at a central hub, like the spokes of a bicycle wheel (7). Current evidence suggests that a single transmembrane protein, plasmalemma vesicle-associated protein (PLVAP or PV1), con- stitutes the diaphragm’s sole structural protein (8). PLVAP is a type II transmembrane protein of ~60 kDa with an ~390-amino acid C-terminal extracellular domain (ECD), a single transmembrane domain, and a 26-amino acid N-terminal cytoplasmic domain (9). PLVAP localizes to diaphragms by immunoelectron microscopy, and targeted disruption of the mouse Plvap gene leads to loss of diaphragms from fenestrae, transendothelial channels, and caveolae (10, 11). Current models envision a diffusional pathway consisting of i) the spaces between PLVAP strands within the diaphragm and ii) a meshwork of N-linked glycans on the PLVAP extracellular domain (4). Additionally, PLVAP functions as a gatekeeper for lymphocyte transmigration through ECs (12) and for the entry of lymphocytes into lymph nodes (13). In gut ECs, Salmonella typhimurium has been found Significance The microscopic openings across the walls of many capillaries contain a protein mesh that is formed by the extracellular domain (ECD) of plasmalemma vesicle-associated protein (PLVAP). The three-dimensional structure of one segment of the PLVAP ECD together with biochemical and spectroscopic experiments on this and other segments show that the majority of the PLVAP ECD consists of a dimeric alpha-helical coiled coil, consistent with electron micrographs showing thin fibrous strands arranged across the capillary openings like the spokes of a bicycle wheel. Protein structure determination benefited from the availability of a microfocused long-wavelength X-ray beam, a helical data collection strategy, and an ultrasensitive detector, which together allowed rapid rotation and translation of the crystal during collection of anomalous scattering data from sulfur atoms. Author contributions: T.-H.C., F.-L.H., X.G., and J.N. designed research; T.-H.C., F.-L.H., X.G., and P.M.S. performed research; T.-H.C. contributed new reagents/analytic tools; T.-H.C., F.-L.H., X.G., J.M.K., S.B.G., and J.N. analyzed data; and T.-H.C., F.-L.H., and J.N. wrote the paper. Reviewers: D.J.L., University of Texas at Austin; and Q.L., Brookhaven National Laboratory. The authors declare no competing interest. Copyright © 2023 the Author(s). Published by PNAS. This open access article is distributed under Creative Commons Attribution License 4.0 (CC BY). 1To whom correspondence may be addressed. Email: taohsin.chang@gmail.com or jnathans@jhmi.edu. 2Present address: Genentech, South San Francisco, CA 94080. 3Present address: Merck & Co., Inc., West Point, PA 19486. This article contains supporting information online at https://www.pnas.org/lookup/suppl/doi:10.1073/pnas. 2221103120/-/DCSupplemental. Published March 30, 2023. PNAS  2023  Vol. 120  No. 14  e2221103120 https://doi.org/10.1073/pnas.2221103120   1 of 12 A B C Fig.  1. Fenestral diaphragms visualized by electron microscopy and a schematic of PLVAP within a single diaphragm. (A) Transmission electron micrograph of a part of a choriocapillaris EC in cross-section showing several fenestrae, each with a single diaphragm (dark line). Two fenestrae are marked with arrows. Reproduced with permission from  Hernnberger et  al (10). (B) Scanning electron micrograph of a part of the luminal face of a rat kidney EC showing an en face view with seven fenestrae. Radial strands that constitute the diaphragm are seen within each fenestra. Reproduced with permission from Bearer and Orci (7). (Scale bars for (A) and (B), 100 nm.) (C) Diagram illustrating a model for the arrangement of PLVAP homodimers within an EC diaphragm. Each black line represents a single PLVAP monomer. Ten dimers are arranged within a circle formed by the plasma membrane (in red). to induce the expression of PLVAP, promoting vascular dissemi- nation of gut antigens (14). PLVAP self-associates based on coimmunoprecipitation exper- iments with epitope-tagged PLVAP produced in transfected cells, and that association appears to represent a disulfide-linked dimer based on the electrophoretic mobility of PLVAP in the presence vs. absence of reducing agents (15, 16). These data, together with computational analyses of the secondary structure that predict high alpha-helical content for the PLVAP ECD, have led to a model in which this domain forms an extended and largely helical homodimer that spans the ~30- to 40-nm distance from the plasma membrane to the central hub (4). As the ECD contains nine cysteines, the dimer may be held together by multiple inter- chain disulfide bonds. To understand how endothelial diaphragms determine vascular permeability, it is essential to define the three-dimensional structure of PLVAP at atomic resolution. Prior to the present study, there were no biophysical studies of purified PLVAP or its fragments, and there was no high-resolution structural information for PLVAP. Here, we present circular dichroism (CD) analyses of the PLVAP ECD and its fragments showing that these polypeptides are almost entirely alpha-helical, electrophoretic analyses of secreted PLVAP fragments with distinct tags showing that they form parallel dimers, and an atomic resolution X-ray structure of an 89-amino acid frag- ment of the PLVAP ECD showing that it consists of an extended alpha-helical homodimer, held together by five disulfide bonds. The structure was solved using single-wavelength anomalous diffraction from sulfur-containing amino acids (sulfur SAD). We also demon- strate the utility of combining sulfur SAD with molecular replace- ment and protein modeling for coiled-coil structure determination. Finally, we map the epitope of MECA-32, an anti-PLVAP mono- clonal antibody (mAb) (15), and determine the sequences of the MECA-32 light- and heavy-chain variable domains. Results PLVAP ECD Adopts a Predominantly Alpha-Helical Configuration. The secondary structure predictions for the PLVAP ECD using the Quick2D program (17) which contains five algorithms (PsiPred, SPIDER3, PSSpred, DeepCNF, and NetSurfP-2.0) show high alpha-helical propensity [~91% helicity (SI Appendix, Fig. S1A)]. Two regions spanning amino acids 141-224 and 272- 389 are predicted to be in a coiled-coil (CC) configuration by MARCOIL (17) and DeepCoil (18), and we, therefore, refer to these as CC1 and CC2, respectively (SI Appendix, Fig. S1). To express biophysical quantities of the full-length PLVAP ECD and the individual CC1 and CC2 subregions in a soluble and correctly folded and disulfide-bonded form, PLVAP coding region sequences were inserted downstream of DNA sequences coding for E. coli thioredoxin, an 8xHis tag, and the cleavage site for 3C protease (SI Appendix, Figs. S2 A and B and S3 A and B). Protein production in E. coli SHuffle cells, in which the intracellular redox potential favors disulfide bond formation (19), led to the accumulation of the predicted fusion proteins as the major soluble polypeptides (SI Appendix, Figs. S2 C and S3 C). Following immobilized metal affinity chromatography (IMAC) purification, cleavage with 6xHis-tagged 3C protease, and removal of the His- tagged thioredoxin and 3C protease, the PLVAP ECD and its subregions appeared to be present as disulfide-linked dimers based on their electrophoretic mobilities in reducing and nonreducing gels (SI Appendix, Figs. S2 C and D and S3 C and D). To assess the secondary structure content of the PLVAP ECD, CD spectra were obtained from IMAC-purified soluble dimeric ECD segments encompassing residues 52 to 438 (for the full-length ECD), 141 to 229 (for CC1), and 270 to 395 (for CC2) (Fig. 2). A derivative of CC2, in which both cysteines were mutated to alanine, was also analyzed. The K2D3 algorithm (20) was used to analyze the CD spectra (Fig. 2 B–G), revealing ~90% helicity for CC1 (pPMS1170), ~74 to 78% helicity for CC2 (irrespective of the presence or absence of the two cysteines, pXG52 and pXG53, respectively), and ~64% helicity for the full-length PLVAP ECD (pXG61), lower than the predicted ~91% helicity (SI Appendix, Fig. S1A). PLVAP ECD Forms Parallel Disulfide-Linked Dimers. To explore the assembly of the PLVAP ECD in a more native context, segments corresponding to its N- and C-terminal halves were expressed as secreted fusions to human growth hormone (hGH) in transiently transfected HEK/293T cells and harvested from serum-free conditioned medium (SFCM). A set of PLVAP ECD N-terminal regions, encompassing amino acids 53-222, 53-235, and 53-257 (constructs A-C, respectively, in Fig. 3A), all of which include CC1, were efficiently secreted and found to migrate in nonreducing SDS-PAGE at a mobility corresponding to 2 to 3 times the molecular weight at which they migrate in reducing SDS-PAGE (Fig. 3B). hGH alone (construct O) migrates with a predicted molecular weight of ~20 kDa under both reducing and nonreducing conditions, consistent with its monomeric structure. The C-terminal half of the PLVAP ECD, amino acids 271-438, encompasses CC2 and contains two cysteines. In addition to the wild-type (WT) version of this region, site-directed mutants with the first cysteine changed to alanine (C337A) and the second cysteine changed to alanine (C386A), or both cysteines changed to alanine were produced as hGH fusions (constructs D-G in Fig. 3A). In reducing SDS-PAGE, all four hGH fusion proteins migrated at the same mobility, consistent with their monomeric molecular weights (~40 kDa; Fig. 3B). However, under nonre- ducing conditions, only the cysteine-to-alanine double mutant migrated at the monomeric molecular weight, whereas the three mutants that contain either one or two cysteines exhibited lower mobility. Under nonreducing conditions (Fig. 3B), the C337A mutant (construct E) shows a higher mobility than the other two cysteine-containing fusion proteins (constructs D and F), presum- ably reflecting a distinct conformation of its unfolded state. In sum, the electrophoretic mobilities of hGH–PLVAP fusion 2 of 12   https://doi.org/10.1073/pnas.2221103120 pnas.org A (cid:31) (cid:31) (cid:31) (cid:31) (cid:31) (cid:31) (cid:31) (cid:31) (cid:31) (cid:31) B D F C E G Fig.  2. Circular dichroism (CD) spectra of the PLVAP ECD show predominantly alpha-helical content. (A) Map of full-length PLVAP (Top) and the four soluble ECD fragments analyzed by CD (Below). The 10 cysteines are numbered and represented by vertical black lines in the full-length schematic. (B–E), CD spectra of the indicated ECD segments. (F) Superposition of the four CD spectra. (G) Reference CD spectra for alpha-helix, beta-sheet, and random coil. proteins A-G under reducing vs. nonreducing conditions imply that both the N-terminal and C-terminal halves of the PLVAP ECD form disulfide-linked dimers. To more comprehensively probe the dimeric nature of the C-terminal half of the PLVAP ECD, we prepared constructs H–K, a set of myc epitope–tagged versions of constructs D–G, and constructs L–N, a set of rim epitope-tagged and fibronectin type 3 (Fn3) domain-containing versions of constructs D–F. Based on their distinct epitope tags, members of these two sets of constructs can be independently visualized in fluorescent immunoblots. Additionally, the higher molecular weight of constructs L–N low- ers their mobility in nonreducing gels. Heterodimers formed between individual members of constructs H–K and constructs L–N would be predicted to exhibit a mobility distinct from and intermediate to the mobilities of the corresponding homodimers. In the immunoblots of nonreducing SDS-PAGE gels shown on the left side of Fig. 3C, SFCM harvested from cells cotransfected with construct L together with constructs H, I, J, or K reveals L+L homodimers (red; low mobility), H+H, I+I, and J+J homodimers (green; high mobility), and L+H, L+I, and L+J heterodimers (yellow; intermediate electrophoretic mobility). Consistent with the behavior of construct G (Fig. 3B), cotransfection of constructs L and K shows that L+K heterodimers and K+K homodimers are unstable to heating in SDS. To further explore the disulfide bonding arrangement within PLVAP CC2, we tested the stabilities of heterodimers formed between all pairwise combinations of constructs I and J cotrans- fected with constructs M and N. In each of these constructs, one of the two cysteines has been mutated to alanine. If the presump- tive dimers are aligned in a parallel fashion and if disulfide bonding is intermolecular, we would predict that only the pairs of con- structs in which the same cysteine is preserved would form disulfide-linked heterodimers, i.e., I+M and J+N. By contrast, if the presumptive dimers are aligned in an anti-parallel fashion and if disulfide bonding is intermolecular, we would predict that only the pairs of constructs in which different cysteines are preserved would form disulfide-linked heterodimers, i.e., I+N and J+M. As seen in the upper immunoblot on the right side of Fig. 3C, only the SFCM harvested from cells cotransfected with constructs I+M and J+N shows the heterodimeric yellow band at intermediate PNAS  2023  Vol. 120  No. 14  e2221103120 https://doi.org/10.1073/pnas.2221103120   3 of 12 A B C Fig. 3. Biochemical evidence that PLVAP ECD fragments form parallel homodimers linked by disulfide bonds between corresponding cysteines. (A) Map of full-length PLVAP (Top) and secreted human growth hormone (hGH; blue rectangle) fusions with the indicated ECD fragments (Below). The CC2 region of PLVAP has two cysteines at positions 337 and 386, and various derivatives have either one or both mutated to alanine, as indicated for the fusion proteins labeled D to N. The tenth fibronectin type 3 (Fn3) domain from human fibronectin (amino acids 1,538 to 1,630) is shown as a green rectangle in fusion proteins L, M, and N. (B) SDS-PAGE and anti-hGH immunoblots of serum-free conditioned medium from transiently transfected HEK/293T cells showing the electrophoretic mobilities of the indicated fusion proteins under either reducing [+beta-mercaptoethanol (BME)] or nonreducing [-BME] conditions. Fusion proteins D, E, and F have identical lengths but exhibit distinct electrophoretic mobilities under nonreducing conditions based on the location of the disulfide bonds. Fusion protein G, which lacks interchain disulfide bonds, migrates as a monomer after heating in SDS in the absence of BME. D, dimer. M, monomer. (C) SDS-PAGE and anti-epitope tag immunoblots of serum-free conditioned medium (SFCM) showing that coexpression and cosecretion of pairs of PLVAP ECD fusion proteins with different electrophoretic mobilities and different epitope tags lead to both homo- and heterodimer formation. The dimers are stable to heating in SDS in the absence of BME if there is at least one interchain disulfide bond. Heterodimers do not form if SFCM containing the individual homodimeric fusion proteins are mixed together prior to gel electrophoresis. Immunoblotting was performed with mouse anti-rim mAb and rabbit anti-myc, and the primary antibodies were visualized with fluorescent secondary antibodies. mobility. As a further control for specificity, the lower immunoblot on the right side of Fig. 3C shows that pairwise mixing of SFCM from cells transfected with individual constructs in the same com- binations as shown for the cotransfection experiment does not generate heterodimer species. This experiment implies that dimers of the C-terminal half of the PLVAP ECD are likely formed intra- cellularly and do not rapidly exchange their subunits at room temperature. Taken together, these experiments show that CC2 forms a parallel dimer with two intermolecular disulfide bonds. Protein Production and Crystallization of PLVAP CC1 and CC2. For structure determination, we focused on PLVAP CC1 and CC2. Both protein fragments can be expressed as soluble disulfide-linked dimers as judged by their mobilities on reducing vs. nonreducing SDS-PAGE gels, and both are monodisperse as judged by their behavior in size-exclusion chromatography (SEC; SI  Appendix, Figs. S2 and S3). PLVAP CC2 crystallized in several conditions (e.g., 0.1 M MgCl2, 0.1 M MES, pH 6.0, and 8% polyethylene glycol 6K), and the crystals diffracted to a resolution of 3.65 Å, revealing an I212121 space group with unit-cell parameters 82.6 Å, 98.2 Å, and 413.6 Å. Of note is that PLVAP CC2 crystals revealed a severely anisotropic diffraction pattern, with resolution limits along to a*, b*, and c* axes of 6.6 Å, 5.3 Å, and 3.65 Å. Moreover, each asymmetric unit appears to contain 4 to 5 dimers of PLVAP CC2 based on an analysis of the Matthews coefficient (21). Without an accurate model for molecular replacement, and with low-resolution and severe anisotropic diffraction data (22), we were unable to solve the structure of this crystal form. PLVAP CC1 crystallized in the I212121 space group with unit-cell parameters 33.8 Å, 89.0 Å, and 180.9 Å, referred to hereafter as crystal form II. X-ray diffraction data were collected and processed to 1.95 Å (SI Appendix, Fig. S4 A–C and Table S1). Because PLVAP CC1 contains only one methionine, we initially produced PLVAP CC1 with selenium-labeled methionine (SeMet) for multi- or single-wavelength anomalous diffraction (MAD or SAD). However, the resulting anomalous signals were too weak to be useful for structure determination. Structure Determination of PLVAP CC1 by Sulfur SAD Phasing. PLVAP CC1 contains five cysteines and one methionine, and previous studies have demonstrated that SAD from sulfur- containing residues—sulfur SAD—can be used for experimental phasing and structure determination (23–26). The recent development and installation of a microfocused X-ray beam, fast pixel array detector, and stable vector/helical data collection system at the National Synchrotron Light Source II (NSLS II) inspired us to ask whether it might be possible to collect accurate sulfur SAD data from a single PLVAP CC1 crystal for de novo phasing. For this experiment, we first screened additional crystallization conditions and obtained PLVAP CC1 crystals in the P212121 space group with unit-cell parameters 35.6 Å, 86.2 Å, and 181.6 Å, referred to hereafter as crystal form I. We then collected data 4 of 12   https://doi.org/10.1073/pnas.2221103120 pnas.org at a wavelength of 1.77 Å, using a low-noise and fast readout EIGER X detector, with an exposure time of 0.015 s per 0.2° oscillation. The X-ray beam was 90% attenuated to give ~5 × 1011 photons s−1 with a beam cross-section of 5 μm × 7 μm. The data collection encompassed a total rotation of 1,800° while continuously advancing the crystal 300 μm (Fig. 4A). Constraints related to the detector-to-crystal distance and the long wavelength of the X-ray beam limited the resolution of the sulfur SAD data to 2.4 Å (SI Appendix, Fig. S4 D–F). To determine how much data are required for sulfur SAD phas- ing of PLVAP CC1, we indexed and integrated the complete 1,800° dataset and then divided it into five separately scaled data- sets, each comprising one 360° rotation (Fig. 4 A and B). We found that integrating all five sulfur SAD datasets provided anom- alous signals with a signal-to-noise ratio sufficient for structure determination of PLVAP CC1, whereas integrating only the first three sulfur SAD datasets was insufficient (Fig. 4C; compare SI Appendix, Figs. S5 and S6). Phases derived from the full sulfur SAD dataset revealed electron densities corresponding to ten disulfide bonds and four methionines (Fig. 5A). The Bijvoet dif- fraction ratio ( < |ΔF ±h| >∕< |F | > ) was ~1.4% at 1.77 Å (23). Subsequent structure refinement revealed two PLVAP CC1 dimers in the asymmetric unit (Fig. 5B and SI Appendix, Table S1). Integrating Sulfur SAD with Structure Prediction and Molecular Replacement. The past several years have witnessed dramatic advances in protein structure prediction (27, 28), which makes structure determination using molecular replacement (MR) based on predicted structures an increasingly attractive strategy. Although predicted structures can capture the overall secondary and tertiary structural features of the protein of interest, they often exhibit multiple differences from the actual structure on a scale of one to several angstroms. As a result, phases generated by MR based on predicted structures can be of variable utility. One approach to improving MR phasing is to combine it with SAD phasing (MR–SAD) (29–32). This strategy has not yet been applied to coiled-coil structures. Therefore, we have explored the utility of this approach with PLVAP CC1 as a test case. For this test, we first combined the X-ray diffraction data from the first three crystal rotations corresponding to 3 × 360 o (datasets 1 + 2 + 3) = 1,080°, as described in the preceding section (Fig. 4 A and B). As noted above, phases calculated from this 1,080° sulfur SAD dataset failed to solve the structure of PLVAP CC1 (SI Appendix, Fig. S6). Next, we used the CCFold algorithm (33), a protein structure prediction algorithm for coiled-coil proteins, to generate a computational model of PLVAP CC1. This model has a rmsd of 1.5 Å relative to the crystal structure of PLVAP CC1, with the N-terminal region showing the largest deviation (SI Appendix, Fig. S7). We used this computational model for MR with the 1,080° diffraction dataset to generate an electron density map that exhibited a coiled-coil-like pattern (Fig. 4D and SI Appendix, Fig. S8A). However, this approach appeared to retain an inherent MR model bias, with the result that the MR-based phases were of insufficient quality to permit structure refinement beyond values of Rwork = 46.8% and Rfree = 50.4% (Fig. S8B). To test the MR–SAD strategy, we used the MR-based electron density map and its protein model in combination with the sulfur SAD phases from the 1,080° diffraction dataset to improve the anomalous difference Fourier analyses. Subsequent automated model building and structure refinement from this starting point Fig. 4. Data collection and phasing methods to determine the structure of PLVAP CC1. (A) The vector module established on the AMX beamline (NSLS II) was used to collect a continuous 1,800° oscillation range (9,000 images) along 300 μm of a single PLVAP CC1 crystal (P212121 space group). For the phase determination and analysis, every 360° rotation was scaled as a single dataset (1 to 5) using AIMLESS. (B) The plot shows the data quality for each image. The Rmerge (red lines) are not visible on the plot because they are completely overlapping with Rmeas (blue lines). Each 360° rotation is denoted by double-headed arrows. (C) The initial density-modified map (blue meshes) from PHENIX RESOLVE was calculated with experimental sulfur SAD phases from five datasets and is contoured at 1.3 σ. The anomalous difference map (purple meshes) for sulfur atoms is contoured at 3 σ with sulfur-containing residues labeled. (D) The initial electron density map (cyan meshes) from PHASER contoured at 1.3 σ, was obtained by MR using the computational model of PLVAP CC1 generated by the CCFold algorithm as the search model. (E) The initial density-modified map (yellow meshes) from PHENIX RESOLVE was calculated with the phases obtained by combining MR with sulfur SAD using only the first three datasets and is contoured at the 1.3 σ level. The anomalous difference map (red meshes) for sulfur atoms is contoured at the 3 σ level with sulfur-containing residues labeled. PNAS  2023  Vol. 120  No. 14  e2221103120 https://doi.org/10.1073/pnas.2221103120   5 of 12 Fig. 5. Structure determination of PLVAP CC1 (crystal form I) using experimental sulfur SAD phases. (A) The anomalous difference map (purple meshes) for sulfur atoms, contoured at 2.5 σ, is shown with two dimers of PLVAP CC1 in the asymmetric unit of crystal form I (P212121 space group) in a ribbon representation. Five disulfide bonds (SS-1 to SS-5) and two methionine sulfur atoms (M169) can be identified in each CC1 dimer using sulfur SAD phases. The N and C termini are labeled. (B) The sigmaA-weighted 2|FO|-|FC| electron density map is shown after structure refinement (gray meshes) and is contoured at 1.0 σ. resulted in successful structure determination with Rwork = 31.7% and Rfree = 37.3% (Fig. 4E and SI Appendix, Fig. S8C). During the course of this work, the MR–SAD pipeline in PHENIX AutoSol (34, 35) was updated to improve anomalous difference Fourier analyses by integrating the model and the electron density map obtained from MR. Taken together, these tests show that the combination of protein modeling and MR–SAD can successfully solve a coiled-coil structure in a situation where phases based on modeling and MR alone or sulfur SAD alone cannot. Structure of PLVAP CC1. PLVAP CC1 is 89 amino acids in length and contains five cysteines (Fig.  6A). The structure of PLVAP CC1, determined from crystal form I by sulfur SAD, reveals a parallel dimer with five symmetric interchain disulfide bonds (SS-1 by Cys142, SS-2 by Cys153, SS-3 by Cys178, SS-4 by Cys199, and SS-5 by Cys224) (Fig. 6 B and C). The five disulfide bonds were confirmed in the anomalous difference map (Fig. 6 B and C). The PLVAP CC1 dimer model from crystal form I was used to determine the structure of PLVAP CC1 in crystal form II by MR (Fig.  6D). The electron density for residues 141 to 145 of crystal form II chain B was not visible (Fig. 6D). Structure-based analysis by the Socket2 algorithm (36) for both crystal forms shows that most of the PLVAP CC1 dimer exhibits a classical coiled-coil configuration characterized by heptad repeats (abcdefg in Fig. 6A). In addition to the five disulfide bonds, the PLVAP CC1 dimer interface features both hydrophobic and hydrophilic interactions (Fig.  7A). More specifically, there are sixteen hydrophobic interactions illustrated in cross-section (CS) in Fig. 7B: CS-1, Leu146; CS-2, Val149; CS-4, Leu156; CS-5, Leu157; CS-6, Leu160; CS-7, Val164; CS-8, Leu167; CS-10, Leu185; CS-11, Leu186; CS-12, Lys189 (using the proximal hydrophobic region of the side chain); CS-13, Thr192; CS-14, Leu196; CS-15, Arg203 (using the proximal hydrophobic region of the side chain); CS-16, Thr213; CS-17, Leu217; and CS-18, Val220. Two hydrophilic interactions are mediated by Asn150 (CS-3) and Asp181 and Lys182 (CS-9) (Fig. 7B). The PLVAP CC1 heptad repeats present several exceptions to the classic coiled-coil pattern of hydrophobic amino acids at posi- tions one (a) and four (d) that form the basis of knobs-in-holes packing (37, 38). In particular, Asn150, Asp181, Lys189, Thr192, Arg203, and Thr213 at positions a and d are hydrophilic (Fig. 6A). Asn150 contributes a hydrophilic interaction (CS-3), and the side chains of Lys189 (CS-12), Thr192 (CS-13), Arg203 (CS-15), and Thr213 (CS-16) form hydrophobic interactions that presumably stabilize the coiled-coil configuration (Fig. 7). Glu174 (position d), Gln206 (position d), and Gln210 (position a) do not appear to contribute to the coiled-coil configuration (Fig. 6A). The desta- bilizing effects of these noncanonical amino acids are presumably 6 of 12   https://doi.org/10.1073/pnas.2221103120 pnas.org (IC), the transmembrane domain Fig. 6. Structures of PLVAP CC1. (A) Schematic diagram of mouse PLVAP primary structure showing the intracellular domain (TM), the extracellular domain (ECD), and the location and sequence of PLVAP CC1 (residues 141 to 229). The five cysteine residues in CC1 are denoted in red bold letters. The putative N-linked glycosylation site (Asn-X-Ser/Thr; X is any residue) is indicated by black underline. The heptad repeat pattern (abcdefg; a and d are generally hydrophobic residues) is labeled and was derived by the Socket2 algorithm (36) based on the structures of PLVAP CC1. The unfilled triangles indicate amino acids at positions a and d that are hydrophilic (Asn150, Asp181, Lys189, Thr192, Arg203, and Thr213) and contribute to coiled-coil interactions (Fig.  7). The filled triangles indicate amino acids at positions a and d (Glu174, Gln206, and Gln210) that do not appear to contribute to coiled-coil interactions. The green line indicates an insertion of a noncanonical heptad repeat (defg). In this region, disulfide bonds SS-3 and SS-4 likely maintain the coiled-coil configuration. (B and C) Crystal form I of PLVAP CC1 solved by sulfur SAD. The anomalous difference map (purple meshes) contoured at 3.0 σ confirms the presence of five disulfide bonds (SS-1 to SS-5) shown in stick representation. The N and C termini are labeled. The positions of Asn residues in the putative N-linked glycosylation sites are indicated by white circles. In (C), the two PLVAP CC1 dimers from crystal form I are shown in side and top views. The small vertical green line denotes the position of a noncanonical heptad repeat (defg). (D) The PLVAP CC1 dimer from crystal form II is shown in side and top views. offset by the presence of disulfide bonds SS-3, SS-4, and SS-5 (Fig. 6 C and D). Last, a noncanonical heptad repeat (defg; resides 185 to 188; Fig. 6A) was identified by the Socket2 algorithm (36). Its destabilizing effect on the coiled-coil assembly is presumably offset by the presence of disulfide bonds SS-3 and SS-4 (Fig. 6C). Among PLVAP CC1 dimers, the average rmsds are 1.15 Å for crystal form I chain A + B vs. chain C + D, 0.89 Å for crystal form I chain A + B vs. crystal form II chain A + B, and 0.84 Å for crystal form I chain C + D vs. crystal form II chain A + B (Fig. 8). Conformational differences are greatest at the N terminus. Mapping the MECA-32 Epitope and Genetically Engineering MECA- 32. MECA-32, an anti-PLVAP mAb, is widely used as a marker of high-permeability vasculature (39, 40). Mapping the MECA- 32 epitope and cloning and expressing recombinant MECA-32 could be of utility in the structural analysis of PLVAP and in the development of improved reagents for visualizing PLVAP in vivo. To map the MECA-32 epitope by immunoblotting, a series of deletion mutants of the PLVAP ECD were expressed as hGH fusion proteins (Fig. 9 A and B). Deletions from the PLVAP N terminus showed that the MECA-32 epitope is between amino acids 322 and 390, the interval defined by deletions C and D. A finer set of deletions from the PLVAP C terminus showed that the MECA-32 epitope is between amino acids 361 and 371, with an essential region between amino acids 367 and 371, the interval defined by deletions G and H. To clone the variable domains of MECA-32, we reverse-transcribed and PCR amplified light-chain (VL) and heavy-chain (VH) segments from the MECA-32 hybridoma cell line (15) and inserted the PCR products into plasmids coding for a human light-chain constant region (CL with a C-terminal rhodopsin 1D4 epitope tag) and a human heavy-chain IgG constant region (CH1, CH2, and CH3 with a human rhinovirus 3C protease cleavage site between CH1 and CH2 and a C-terminal 8xHis tag), respectively (Fig. 9C). The light and heavy chains encoded in the resulting constructs were coexpressed, assembled, and secreted from HEK/293T cells; we refer to this engineered protein as MECA-32/hIgG (Fig. 9D). The binding properties of MECA-32/hIgG were tested by immobilizing it on protein A–coated microwells, probing the wells with alkaline phos- phatase (AP) fused to PLVAP CC2 (AP–PLVAP CC2), and detect- ing AP activity with a colorimetric enzyme assay (Fig. 9E). This assay shows strong and specific binding between MECA-32/hIgG and AP–PLVAP CC2 (Fig. 9F). Discussion The three-dimensional structure of PLVAP CC1 presented here shows that it forms a parallel disulfide-bonded dimeric coiled coil. Additionally, our spectroscopic and biochemical studies of the PLVAP ECD and its subregions show that the entire PLVAP ECD adopts a predominantly alpha-helical conformation and that it assembles into a parallel disulfide-bonded, and presumably highly extended, dimer. PNAS  2023  Vol. 120  No. 14  e2221103120 https://doi.org/10.1073/pnas.2221103120   7 of 12 Fig.  7. Interactions mediating PLVAP CC1 dimer formation. (A) Interactions for PLVAP CC1 dimer formation include (i) five disulfide bridges (SS-1 to SS-5) shown in stick representation, (ii) residues for hydrophilic interactions [cross- section (CS)-3 and CS-9], and (iii) hydrophobic interactions (multiple CSs) shown as sphere represen- tations (atom coloring: carbon, marine, and cyan for each chain of the dimer, respectively; nitrogen, dark blue; oxygen, red). (B) Close-up views of the helix–helix interactions. The sigmaA-weighted 2|FO|-|FC| (golden electron density map meshes) is contoured at 1.0 σ, and residues for these interactions are shown in stick representation. Hydrophilic interactions are display- ed as red dashed lines. maintaining serum oncotic pressure (11). It seems likely that, during evolution, these sieving properties of the EC diaphragm imposed a selective pressure to i) cap the molecular weights of polypeptide hormones at ~40 kDa and ii) increase the molecular Fig.  8. Structural comparisons between PLVAP CC1 dimers. (A) The superposition of chains A + B with chains C + D of PLVAP CC1 crystal form I has a rmsd of 1.15 Å over 139 Cα atoms. (B) The superposition of chains A + B of PLVAP CC1 crystal form I with chains A + B of crystal form II has a rmsd of 0.89 Å over 125 Cα atoms. (C) The superposition of chains C + D of PLVAP CC1 crystal form I and chains A + B of crystal form II has a rmsd of 0.84 Å over 150 Cα atoms. These data lend strong support to the EC diaphragm model proposed by Tse and Stan (4) in which low–molecular weight blood and tissue substituents traverse the fenestral and vesicular diaphragms through wedge-shaped channels bounded on two sides by PLVAP dimers and on one side by the curved plasma membrane at the edge of the fenestra or vesicle (Fig. 1C). In this model, the lengths of the two longer sides of each opening are determined by the length of the PLVAP dimer, and the chemical properties of the channel walls are determined by the amino acid side chains and N-linked glycans (four conserved Asn-X-Ser/Thr are present; SI Appendix, Fig. S1A) on the solvent-accessible faces of PLVAP. The 64% estimate for the helical content of the ~390-amino acid PLVAP ECD—based on the CD spectrum in Fig. 2E—implies that ~248 amino acids are in a helical configuration. Assuming that the helices form a dimeric coiled coil with ~3.5 residues and ~5.1 Å per turn, then the helical part of the PLVAP ECD would extend ~361 Å (~36 nm), a close match to the ~40-nm radius of a diaphragm. It is interesting that, among mam- mals, the PLVAP ECD length is highly conserved (SI Appendix, Fig. S1B), consistent with the concept that coiled coils act as “molec- ular rulers” within macromolecular assemblies (37). Fenestrated capillaries are present in a wide variety of locations, including endocrine organs, intestinal villi, renal peritubular cap- illaries, the choroid plexus, and circumventricular organs (1–3). In endocrine organs such as the pituitary, polypeptide hormones are secreted into the extracellular space and traverse the capillary wall, presumably via its fenestrations. The largest polypeptide hormones—thyroid-stimulating hormone, luteinizing hormone, and follicle-stimulating hormone—are heterodimeric glycopro- teins of 28 to 35 kDa (41). Horse radish peroxidase (HRP), a 44-kDa protein, can also traverse EC diaphragms (42). In contrast, serum albumin, a nonglycosylated protein of 66 kDa, is unable to pass through EC diaphragms, as required for its role in 8 of 12   https://doi.org/10.1073/pnas.2221103120 pnas.org B PLVAP amino acids 265-400 J I H FG E anti-hGH + MECA-32 anti-hGH MECA-32 A B C D E F G H I J A B C D E F G H I J A B C D E F G H I J C D hGH-PLVAP fusion proteins A-J hGH PLVAP segment A B 62 49 38 28 Hybridoma cells C RT-PCR (MECA-32 V and V ) H L Molecular cloning LV LC Rho 1D4 160 120 80 50 3C protease cleavage site H3C HV H1C H2C 8xHis 20 D Engineered MECA-32 E AP-PLVAP CC2 F AP PLVAP CC2 Engineered MECA-32 Immobilized Target Engineered MECA-32 IgG Probe IgG AP-PLVAP CC2 Co-transfection in HEK293 cells SDS-PAGE (non-reducing) Protein A Fig. 9. Mapping the epitope for MECA-32 on the PLVAP ECD. (A) The sequence of the PLVAP ECD between amino acids 265 and 400 showing the end points of deletions that remove progressively larger regions from the N terminus (rightward arrows B–D) and the end points of deletions that remove progressively larger regions from the C terminus (leftward arrows E–J). Amino acids essential for MECA-32 binding, as defined by PLVAP deletion analysis, are highlighted in red, with the overlying red bar indicating a larger region that likely encompasses the MECA-32 epitope. (B) SDS-PAGE and immunoblotting of hGH fusions with the PLVAP ECD deletions indicated in (A). Blots were probed with rabbit anti-hGH and mouse mAb MECA-32, and the primary antibodies were visualized with fluorescent secondary antibodies. Fusion proteins A-D extend to the C terminus of PLVAP. The N terminus of fusion protein A is at PLVAP amino acid 104 (TRREME...). Fusion proteins E-J have their N termini at PLVAP amino acid 143 (QEQLKE...). MECA-32 binds to fusion proteins A-C and E-G but not to fusion proteins D and H-J. (C) Workflow for MECA-32 Ab molecular cloning, construct design, and protein expression. (D) SDS-PAGE under nonreducing conditions of the engineered MECA-32 Ab (MECA-32/hIgG). (E) Diagram of the protein–protein interaction assay. MECA-32/hIgG was immobilized on protein A–coated wells. A fusion protein comprising alkaline phosphatase (AP) fused to PLVAP CC2 (AP–PLVAP CC2) was used as a probe. (F) A colorimetric AP reaction was used to detect the interaction between MECA-32/hIgG and AP–PLVAP CC2. Bovine IgG served as the negative control. weights of nonhormone serum proteins to greater than ~60 kDa. The selective pressure exerted on hormone and serum protein molecular weights was presumably mirrored by selective pressure on the sequence and structure of PLVAP so that it would produce EC diaphragms with an ~50-kDa cutoff. Phase information is required for X-ray crystal structure deter- mination, but it is lost during the collection of diffraction data (43–45). This “phase problem” is especially challenging for pro- teins that lack structurally defined homologs or accurate com- putational models because their structures cannot be solved by MR. In such cases, phase information is generally obtained by multiple isomorphous replacement (MIR) after soaking crystals with heavy atoms (e.g., gold, platinum, and mercury derivatives) or by MAD or SAD if an appropriate anomalous scatterer is present or can be introduced (43, 44, 46, 47). However, these approaches can be problematic (43, 48). For example, heavy atoms can degrade crystal quality or alter crystal packing so that the resulting crystal is no longer isomorphous (49), and SeMet incorporation often results in lower protein yield from eukaryotic cells, which can be problematic for proteins that are difficult to produce (50). Sulfur SAD phasing is an attractive alternative for de novo phas- ing from native crystals, but this approach faces several technical challenges (23–26, 45, 51–55). First, the anomalous scattering signal from sulfur [atomic number (Z) = 16] is weak relative to scattering from selenium (Z = 34) and mercury (Z = 80). Second, the maximum anomalous signal is near the sulfur absorption edge at a wavelength of 5 Å, but such long-wavelength X-rays are asso- ciated with increased radiation damage, background absorption, and diffuse scattering and are generally unavailable at conventional synchrotron sources. Therefore, sulfur SAD data are typically col- lected at a wavelength of 1.5 to 2.5 Å with a synchrotron source (53) or at a wavelength of 2.29 Å with a chromium source (52). Third, the requirement for highly accurate sulfur SAD data requires the collection and subsequent merging of multiple datasets with minimal radiation damage during data collection (24, 25, 51, 56). To usefully merge sulfur SAD data from different crystals, the crys- tals must be isomorphous, and their point group symmetry must permit unambiguous indexing (57). Finally, the quality of sulfur SAD phasing is also dependent on the resolution of the diffraction data, with a resolution of 2.3 to 2.8 Å being generally desirable (24). Thus, collecting sulfur SAD data from poorly diffracting crys- tals has the added challenge that the higher X-ray dose required for useful diffraction leads to increased radiation damage. In response to these challenges, several strategies have been devel- oped. These include i) reducing air absorption with a vacuum or a PNAS  2023  Vol. 120  No. 14  e2221103120 https://doi.org/10.1073/pnas.2221103120   9 of 12 helium-filled cone between the crystal and the detector (24, 55), ii) collecting diffraction data with an inverse-beam algorithm to minimize any differential effects of radiation damage on the signals from Friedel mates (25, 45), iii) using a semicylindrical detector to collect high-angle diffraction (55), and iv) using a high-precision multiaxis goniometer to collect multiple orientations from each crystal (26, 54, 55). At present, the instrumentation for implement- ing these strategies is available at only a few synchrotron facilities. A protein engineering strategy to enhance the utility of sulfur SAD involves adding rationally designed disulfide bonds (58). In the present study, we combined a microfocused beam (~5 × 1011 photons s−1 with a beam cross-section of 5 μm × 7 μm) at a wavelength of 1.77 Å, a vector/helical data collection strategy, and an ultrafast shutter and high-sensitivity detector to collect 1,800° of sulfur SAD data from a single crystal in ~2.25 min. For com- parison, Weinert et al. reported using a 90 μm × 45 μm beam cross-section with ~1010 photons s−1 at a wavelength of 2.07 Å to collect 2,880° (8 × 360°) of sulfur SAD data from a single crystal (26). We note that the Weinert et al.’s data collection required a multiaxis goniometer, an instrument that is not available in most synchrotron facilities (26). In X-ray crystallography, structure determination of coiled-coil proteins has posed a long-standing challenge to the use of MR (22, 43, 59). MR is based on comparisons between the Patterson synthesis calculated from the model and the analogous synthesis generated from the diffraction data (43). Since coiled coils tend to pack in a parallel arrangement in the crystal, they generate many similar self- and cross-Patterson vectors that are difficult to dis- ambiguate (22). Although recent advances in computational algo- rithms, including AMPLE (60), ARCIMBOLDO (61, 62), CCFold (33), and CCsolve (63), have made significant progress in structure determination of coiled-coil proteins by integrating computational modeling with MR, this strategy alone failed to solve the phase problem for PLVAP CC1. As demonstrated here, the combination of protein modeling and MR with sulfur SAD is a powerful approach for structure determination (29–32). With MR–SAD, an initial phase calcu- lation from either a full or partial MR difference Fourier analysis can be used to search for anomalous scattering peaks indicative of sulfur atoms. Importantly, the independence of MR and sulfur SAD phases minimizes the model bias that is inherent to the MR method. The MR–SAD strategy, together with technical advances in X-ray data collection, represents an important advance for struc- ture determination of challenging targets. Materials and Methods Production and Purification of PLVAP ECD Fragments in E coli. For expres- sion in E. coli, a DNA segment coding for each PLVAP ECD fragment was cloned into a modified pET-11d vector (SI Appendix, Figs S2 B and S3 B). Starting from the N terminus, this vector codes for E. coli thioredoxin (TrxA), an 8xHis tag, and a 3C protease cleavage site (SI Appendix, Figs S2 B and S3 B). All constructs were confirmed by sequencing. The plasmid was transformed into E. coli SHuffle T7 cells (New England Biolabs) and induced with 0.2 mM isopropyl β-thiogalactopyrano- side (IPTG) in terrific broth containing 100 μg/mL ampicillin (MilliporeSigma) at room temperature (~25o C) overnight. For cell disruption, the cells were harvested by centrifugation and resuspended in B-PER bacterial protein extract reagent (ThermoFisher Scientific) supplemented with 50 mM HEPES, pH 7.5, 0.3 M NaCl, 30 mM imidazole, 1 mM MgCl2, 500 U benzonase (MilliporeSigma), 0.2 mg/mL lysozyme, and “cOmplete” Protease Inhibitor Cocktail (MilliporeSigma). The cell lysate was clarified by centrifugation, and the supernatant was filtered using a 0.45 μm Steritop filter (MilliporeSigma). Proteins were purified by immobi- lized metal affinity chromatography (IMAC) using Ni Sepharose 6 Fast Flow resin (Cytiva). The resin was washed with 20 mM HEPES, pH 7.5, 0.5 M NaCl, 30 mM imidazole, and10% glycerol and eluted in 20 mM HEPES, pH 7.5, 0.15 M NaCl, and 0.5 M imidazole. The eluted protein was dialyzed against 20 mM Tris, pH 7.5, 0.5 M NaCl, and 5% glycerol and treated with His-tagged 3C protease prepared as described previously (64). The 3C protease cleaved sample was further purified by IMAC and subjected to size-exclusion chromatography (SEC) using HiLoad Superdex 200 (Cytiva) in either 10 mM HEPES, pH 7.5, and 0.15 M NaCl or 10 mM Tris, pH 8.0, and 0.3 M NaCl. CD Spectra. The four PLVAP ECD fragments (pXG52, pXG53, pXG61, and pPMS1170; Fig. 2A) were produced in E. coli SHuffle T7 cells as soluble thiore- doxin fusions, released by 3C protease, and purified to apparent homogeneity as described above. Protein concentrations were determined by the Bradford assay and confirmed by Coomassie blue staining following SDS-PAGE, with a BSA dilution series as an internal standard. Stock solutions of highly concentrated protein (5 to 10 mg/mL) in 10 mM Tris, pH 8.0, 0.3 M NaCl (pXG52, pXG53, and pPMS1170) or 10 mM Tris, pH 7.0, and 0.4 M NaCl (pXG61) were dialyzed for 2.5 h against 10 mM Tris, pH 7.4, and 0.15 M NaCl, degassed for 30 min, and then diluted with dialysis buffer to a final protein concentration of 100 μM (pXG52, pXG53, and pPMS1170) or 40 μM (pXG61) for CD measurements. Far-UV CD spectra were collected on a Jasco J-810 spectrophotometer. Spectra of 60 ul samples were recorded at 20°C using a 0.2-cm path-length cuvette and a 0.2-nm step size at a rate of one second per step. Spectra from the buffer control were subtracted from each protein’s spectrum. The data were then converted to molar residue ellipticity (MRE). Production and Immunoblotting of PLVAP Fragments Secreted from HEK/293T Cells. For production of hGH fusion proteins in HEK/293T cells (ATCC CRL-11268), DNA segments coding for the PLVAP ECD or its fragments were inserted into the pSGHP1 expression vector (a derivative of pSGHV0; (65) 3′ of the hGH coding region and 8xHis tag). Adherent HEK/293T cells were transfected with polyethyleneimine (PEI) in the wells of a six-well tray, and the serum containing medium was replaced 1 d later with serum-free medium. After an additional day, the serum-free conditioned medium was collected, centrifuged 15 min at 3,000 rpm, and the supernatant stored in aliquots at −80 °C. For SDS-PAGE, serum-free conditioned medium was mixed with an equal volume of 2xSDS sample buffer either with or without beta-mercaptoethanol (BME) and heated to 90 °C for 3 min before loading. Immunoblots were probed with rat mAb MECA-32 (553849; BD Biosciences), rabbit polyclonal anti-hGH (RDI-HGHabrx1; Fitzgerald Industries International, Concord, MA), mouse mAb anti-rim (66), or rabbit polyclonal anti- myc as indicated in the figures. Blots were then probed with fluorescent goat anti-rabbit, anti-rat, or anti-mouse secondary antibodies (LI-COR Biosciences; Lincoln, Nebraska) and imaged with a LI-COR Odyssey Fc Imaging System. Crystallization and Data Collection. PLVAP fragments CC1 and CC2 were concentrated to 10 mg/mL in 10 mM HEPES, pH 7.5, and 0.15 M NaCl and 12 mg/mL in 3 mM Bis-Tris, pH 6.5, and 90 mM NaCl for crystallization trials, respec- tively. Native crystals of PLVAP CC1 (space group I212121) were grown in 0.1 M sodium acetate, pH 5.0, and 15% polyethylene glycol (PEG) 4000 by the hanging drop vapor diffusion method at 21 °C. For sulfur SAD experiments, PLVAP CC1 crystals (space group P212121) were grown in 0.2 M sodium malonate, pH 5.0, and 20% PEG3350 by the sitting drop vapor diffusion method at 21 °C. PLVAP CC2 crystallized in 0.1 M MgCl2, 0.1 M MES, pH 6.0, and 8% PEG 6K at 21 °C. Crystals were transferred into a reservoir solution supplemented with 5 to 10% propane-1,2-diol for PLVAP CC1 and 25 to 30% glycerol for PLVAP CC2 and then cryocooled in liquid nitrogen. Native diffraction data were collected at −173 °C on the 12-2 beamline with a PILATUS 6M detector (DECTRIS) at the Stanford Synchrotron Radiation Light Source at the National Accelerator Laboratory and were processed with the XIA2 system (67) with Diffraction Integration for Advanced Light Sources (DIALS) (68, 69). Two datasets were scaled and merged using AIMLESS (70). Because of severe anisotropic diffraction along the h axis of the reciprocal lattice, ellipsoidal and anisotropic scaling of diffraction data was performed using the STARANISO web server (71). Sulfur SAD datasets were collected at −173 °C on the highly automated macromolecular crystallography (AMX) beamline in the National Synchrotron Light Source II (NSLS II) at the Brookhaven National Laboratory at a wavelength of 1.77 Å using an EIGER X 9M detector (DECTRIS) with the crystal-to-detec- tor distance set to record reflections to 2.4 Å resolution at the corners of the square detector and to 2.6 Å resolution at the center of the detector’s sides. The vector module was used to collect continuous helical data over a 1,800° 10 of 12   https://doi.org/10.1073/pnas.2221103120 pnas.org rotation with an exposure time of 0.015 s per 0.2° (with beam transmission attenuated to 10% from ~5 × 1012 photons s−1 to ~5 × 1011 photons s−1, with a beam cross-section of 5 μm × 7 μm). Images were indexed, integrated, and scaled using the XIA2 system (67) coupled with DIALS, POINTLESS, and AIMLESS (68–70, 72). A randomly selected subset of 5% of the diffraction data was used as a cross-validation dataset to calculate Rfree. Structure Determination and Refinement. The structure of PLVAP CC1 was solved using sulfur SAD data from a 1,800° rotation of a single crystal (see preced- ing paragraph). Sulfur identification (substructure site) and initial anomalous signal analysis were conducted in the HKL2MAP interface (73) using SHELX C/D/E (74). Specifically, we tested various resolution cutoffs from 3.0 Å to 4.5 Å in steps of 0.5 Å with 1,000 trials per attempt using SHELX D searches. We used a resolution cutoff at 3.5 Å with 1,000 trials, Patterson search, and an Emin value of 1.5 to search for five disulfide bridges, resulting in 17 substructure sites. The substructures found in SHELX D were refined and completed for the final 24 substructure sites using PHASER (75) at PHENIX AutoSol (34, 35) with a resolution cutoff at 2.4 Å, and 4 noncrystallographic symmetry copies using THOROUGH for further substructure searches. The phases were calculated using PHASER (75) and then subjected to density modification using PHENIX RESOLVE (76) to generate an interpretable electron density map. Next, an initial model generated by PHENIX AutoBuild (35, 77) was manually rebuilt by COOT (78) and then fed into CCP4 BUCCANEER (79) and PHENIX Rosetta (35, 80) for further model building and geometry optimization. An anomalous difference Fourier map was calculated to assign the sulfur positions in the map for model building and structure validation. The native PLVAP CC1 structure with the space group I212121 was determined by MR using PHASER (75), with the structure of the PLVAP CC1 dimer obtained from the sulfur SAD data used as the template for MR. For MR–SAD, the CCFold protein modeling algorithm (33) was used to gener- ate a dimer model of PLVAP CC1, which was then used as a template for MR using PHASER (75). The resulting model and electron density map from PHASER (75) were entered into the MR–SAD pipeline in PHENIX AutoSol (34, 35) for anomalous difference Fourier analyses. An initial model generated by PHENIX AutoBuild (35, 77) was refined in PHENIX Refine (35). Manual model building in COOT and structure refinement in CCP4 REFMAC5 (81), PHENIX Refine (35), and autoBUSTER (82) were carried out iteratively for all structures. Translation–libration–screw (TLS) rotation was applied with noncrystallographic symmetry (NCS) restraints in all refinements. All models were validated with MOLPROBITY (83). The crystallo- graphic statistics are listed in SI Appendix, Table S1. Bioinformatic Analysis and Graphic Presentation. Structure-based multiple sequence alignment was performed using Clustal Omega (84) and ESPript (85). The secondary structure predictions for PLVAP were calculated using the Quick2D algorithm (17). Structure superposition was performed using SUPERPOSE (86) in the CCP4 suite (87). Structure-based coiled-coil analysis was performed using the Socket2 algorithm (36). High-quality images of the molecular structures were generated with the PyMOL Molecular Graphic System (version 2.5, Schrödinger, LLC). Schematic figures and other illustrations were prepared using CorelDRAW (Corel Corporation) and Illustrator (Adobe). Structural biology applications used in this work were compiled and configured by SBGrid (88). Sequencing MECA-32 mRNA and Generating Recombinant MECA-32/hIgG. Total RNA from approximately 2 × 105 MECA-32 hybridoma cells (15) was isolated using the RNeasy Plus Micro Kit (QIAGEN) and reverse-transcribed with random primers using SuperScript III Reverse Transcriptase (ThermoFisher Scientific). The variable domains of the heavy and light chains (VH and VL) were PCR amplified from the resulting complementary DNA using the PCR primers described in ref. 89. Sequences of the PCR products showed the following protein sequences. Mature VH (i.e., without the signal peptide): E VQL VES GGG LVQ PGR SMK LSC AAS GFT FSD YYM AWV RQA PMK GLK W VAS ISY EGN KTY YGD SVK GRF TIS RDN AKS ILY LQM NSL KSE DTA TYY CAR QSY SSYIFDYWGQGVMVTVSS. Mature VL: D IQM TQT PSS MSA SLG ERV TIS CGT SQG VNI FLN WYQ QKP DGT IKP LIF FTS HLQ SGV PSR FSG SGS GTA YSL TIS SLE PEDFAVYYCQQYDSSPPTFGGGTKLYLK. The PCR-amplified VH and VL segments were inserted into expression vectors containing a human heavy-chain IgG constant region (CH1, CH2, and CH3 with an HRV 3C protease cleavage site between CH1 and CH2 and a C-terminal 8xHis tag) and a human light-chain constant region [CL with a C-terminal rhodopsin 1D4 epitope tag (90)], respectively (Fig. 9C). The heavy- and light-chain plasmids were cotransfected into HEK/293T cells, and the secreted MECA-32/hIgG was purified from conditioned media collected 2 d post-trans- fection by the IMAC method. For the human placental alkaline phosphatase (AP) fusion protein, DNA coding for PLVAP CC2 was cloned into the plasmid pHL-N-AP- Myc-8H which codes for AP followed by a Myc tag and an 8xHis tag (91). AP–PLVAP CC2 was expressed in HEK/293T cells. MECA-32/IgG was immobilized in the wells of protein A–coated 96-well plates (ThermoFisher Scientific) at 4 °C overnight. The wells were then washed three times with wash buffer (10 mM HEPES, pH 7.5, 0.15 M NaCl, and 0.05% Tween-20) and incubated with a 10-fold dilution of bovine serum albumin (BSA) blocking buffer (ThermoFisher Scientific) in wash buffer for 1 h at 25 °C. The wells were washed with wash buffer and incubated with conditioned media containing AP–PLVAP CC2 at 25 °C for 1 h. The wells were subsequently washed three times with wash buffer and incubated with BluePhos phosphatase substrate solution (Kirkegaard and Perry Laboratories) to visualize the bound AP probe. Data, Materials, and Software Availability. 3D structure data have been deposited in [Protein Data Bank (PDB)] (8FBY and 8FCF) for PLVAP CC1 crystal forms I and II, respectively (92, 93). ACKNOWLEDGMENTS. We thank Dr. Eugene Butcher (Stanford University) for sharing the MECA-32 hybridoma cell line; John Williams for constructing several plasmids shown in Fig.  3; the Phasing@Home organizers Dr. Claudia Millán, Dr. Massimo Sammito, and Dr. Isabel Usón at Institut de Biologia Molecular de Barcelona for helpful discussions about the phase problem; Dr. Randy Read (Cambridge University) and Dr. Thomas Terwilliger (Los Alamos National Laboratory) for helpful discussions about MR–SAD; Dr. Jean Jakoncic and Dr. Alexei Soares for excellent support at the AMX beamline; and Dr. Yun- Sil Lee for helpful comments on the manuscript. This study was supported by the National Eye Institute (NIH) R01EY018637 (J.N.), the HHMI (J.N., T.-H.C., F-L.H, and P.M.S.), a Human Frontiers Postdoctoral Fellowship LT000130/2017-L (T.-H.C), the National Cancer Institute (NIH) (R01CA204345) (S.B.G.), and the National Institute of General Medical Sciences (NIH) R01GM134000 (J.M.K.). Work at the AMX (17-ID-1) and FMX (17-ID-2) beamlines was supported by the NIH, the National Institute of General Medical Sciences (P41GM111244), the US Department of Energy (DOE) Office of Biological and Environmental Research (KP1605010), and the National Synchrotron Light Source II at the Brookhaven National Laboratory, which is supported by the DOE Office of Basic Energy Sciences under contract DE-SC0012704 (KC0401040). Author affiliations: aDepartment of Molecular Biology and Genetics, Johns Hopkins University School of Medicine, Baltimore, MD 21205; bHHMI, Johns Hopkins University School of Medicine, Baltimore, MD 21205; cDepartment of Biophysics and Biophysical Chemistry, Johns Hopkins University School of Medicine, Baltimore, MD 21205; dDepartment of Biochemistry and Molecular Biology, Bloomberg School of Public Health, Johns Hopkins University, Baltimore, MD 21205; eDepartment of Oncology, Johns Hopkins University School of Medicine, Baltimore, MD 21205; fDepartment of Medicine, Johns Hopkins University School of Medicine, Baltimore, MD 21205; gDepartment of Neuroscience, Johns Hopkins University School of Medicine, Baltimore, MD 21205; and hWilmer Eye Institute, Johns Hopkins University School of Medicine, Baltimore, MD 21205 1. 2. 3. 4. J. A. Nagy, A. M. Dvorak, H. F. Dvorak, Vascular hyperpermeability, angiogenesis, and stroma generation. Cold Spring Harb. Perspect. Med. 2, a006544 (2012). R. V. Stan, Endothelial stomatal and fenestral diaphragms in normal vessels and angiogenesis. J. Cell Mol. Med. 11, 621–643 (2007). E. K. Bosma, C. J. F. van Noorden, R. O. Schlingemann, I. Klaassen, The role of plasmalemma vesicle- associated protein in pathological breakdown of blood-brain and blood-retinal barriers: potential novel therapeutic target for cerebral edema and diabetic macular edema. Fluids Barriers CNS 15, 24 (2018). D. Tse, R. V. Stan, Morphological heterogeneity of endothelium. Semin. Thromb. Hemost. 36, 236–245 (2010). 5. W. C. Aird, Phenotypic heterogeneity of the endothelium: I. Structure, function, and mechanisms. Circ. Res. 100, 158–173 (2007). 6. W. C. Aird, Phenotypic heterogeneity of the endothelium: II. Representative vascular beds. Circ. Res. 7. 8. 9. 100, 174–190 (2007). E. L. Bearer, L. Orci, Endothelial fenestral diaphragms: A quick-freeze, deep-etch study. J. Cell Biol. 100, 418–428 (1985). R. V. Stan, M. Kubitza, G. E. Palade, PV-1 is a component of the fenestral and stomatal diaphragms in fenestrated endothelia. Proc. Natl. Acad. Sci. U.S.A. 96, 13203–13207 (1999). R. V. Stan, L. Ghitescu, B. S. Jacobson, G. E. Palade, Isolation, cloning, and localization of rat PV-1, a novel endothelial caveolar protein. J. Cell Biol. 145, 1189–1198 (1999). PNAS  2023  Vol. 120  No. 14  e2221103120 https://doi.org/10.1073/pnas.2221103120   11 of 12 10. L. Herrnberger et al., Lack of endothelial diaphragms in fenestrae and caveolae of mutant Plvap- 51. M. S. Weiss, T. Sicker, R. Hilgenfeld, Soft X-rays, high redundancy, and proper scaling: a new deficient mice. Histochem. Cell Biol. 138, 709–724 (2012). 11. R. V. Stan et al., The diaphragms of fenestrated endothelia: gatekeepers of vascular permeability and blood composition. Dev. Cell 23, 1203–1218 (2012). 12. J. Keuschnigg et al., The prototype endothelial marker PAL-E is a leukocyte trafficking molecule. Blood 114, 478–484 (2009). 13. P. Rantakari et al., The endothelial protein PLVAP in lymphatics controls the entry of lymphocytes 14. and antigens into lymph nodes. Nat. Immunol. 16, 386–396 (2015). I. Spadoni et al., A gut-vascular barrier controls the systemic dissemination of bacteria. Science 350, 830–834 (2015). 15. R. Hallmann, D. N. Mayer, E. L. Berg, R. Broermann, E. C. Butcher, Novel mouse endothelial cell surface marker is suppressed during differentiation of the blood brain barrier. Dev. Dyn. 202, 325–332 (1995). procedure for automated protein structure determination via SAS. Structure 9, 771–777 (2001). 52. C. Yang, J. W. Pflugrath, D. A. Courville, C. N. Stence, J. D. Ferrara, Away from the edge: SAD phasing from the sulfur anomalous signal measured in-house with chromium radiation. Acta. Crystallogr. D Biol. Crystallogr. 59, 1943–1957 (2003). 53. C. Mueller-Dieckmann, S. Panjikar, P. A. Tucker, M. S. Weiss, On the routine use of soft X-rays in macromolecular crystallography. Part III. The optimal data-collection wavelength. Acta. Crystallogr. D. Biol. Crystallogr. 61, 1263–1272 (2005). 54. V. Olieric et al., Data-collection strategy for challenging native SAD phasing. Acta. Crystallogr. D Struct. Biol. 72, 421–429 (2016). 55. A. Wagner, R. Duman, K. Henderson, V. Mykhaylyk, In-vacuum long-wavelength macromolecular crystallography. Acta. Crystallogr. D Struct. Biol. 72, 430–439 (2016). 56. C. Nave, E. F. Garman, Towards an understanding of radiation damage in cryocooled macromolecular 16. R. V. Stan, Multiple PV1 dimers reside in the same stomatal or fenestral diaphragm. Am. J. Physiol. crystals. J. Synchrotron. Radiat. 12, 257–260 (2005). Heart. Circ. Physiol. 286, H1347–1353 (2004). 17. L. Zimmermann et al., A completely reimplemented MPI bioinformatics toolkit with a new HHpred server at its core. J. Mol. Biol. 430, 2237–2243 (2018). 57. R. J. Gildea, G. Winter, Determination of Patterson group symmetry from sparse multi-crystal data sets in the presence of an indexing ambiguity. Acta. Crystallogr. D Struct. Biol. 74, 405–410 (2018). 58. S. H. W. Kraatz, S. Bianchi, M. O. Steinmetz, Combinatorial use of disulfide bridges and native sulfur- 18. J. Ludwiczak, A. Winski, K. Szczepaniak, V. Alva, S. Dunin-Horkawicz, DeepCoil-a fast and accurate SAD phasing for rapid structure determination of coiled-coils. Biosci Rep 38 (2018). prediction of coiled-coil domains in protein sequences. Bioinformatics 35, 2790–2795 (2019). 19. J. Lobstein et al., SHuffle, a novel Escherichia coli protein expression strain capable of correctly folding disulfide bonded proteins in its cytoplasm. Microb. Cell Fact. 11, 56 (2012). 20. C. Louis-Jeune, M. A. Andrade-Navarro, C. Perez-Iratxeta, Prediction of protein secondary structure from circular dichroism using theoretically derived spectra. Proteins 80, 374–381 (2012). 21. B. W. Matthews, Solvent content of protein crystals. J. Mol. Biol. 33, 491–497 (1968). 22. Z. Dauter, Solving coiled-coil protein structures. IUCrJ 2, 164–165 (2015). 23. W. A. Hendrickson, M. M. Teeter, Structure of the hydrophobic protein crambin determined directly 59. R. J. Read, A. J. McCoy, Maximum-likelihood determination of anomalous substructures. Acta. Crystallogr. D Struct. Biol. 74, 98–105 (2018). 60. J. M. Thomas et al., Routine phasing of coiled-coil protein crystal structures with AMPLE. IUCrJ 2, 198–206 (2015). 61. I. Caballero et al., ARCIMBOLDO on coiled coils. Acta. Crystallogr. D Struct. Biol. 74, 194–204 (2018). 62. D. D. Rodriguez et al., Crystallographic ab initio protein structure solution below atomic resolution. Nat Methods 6, 651–653 (2009). 63. S. Ramisch, R. Lizatovic, I. Andre, Automated de novo phasing and model building of coiled-coil from the anomalous scattering of sulphur. Nature 290, 107–113 (1981). proteins. Acta. Crystallogr. D Biol. Crystallogr. 71, 606–614 (2015). 24. Q. Liu et al., Structures from anomalous diffraction of native biological macromolecules. Science 64. T. H. Chang et al., Structure and functional properties of Norrin mimic Wnt for signalling with 336, 1033–1037 (2012). 25. K. El Omari et al., Pushing the limits of sulfur SAD phasing: De novo structure solution of the N-terminal domain of the ectodomain of HCV E1. Acta. Crystallogr. D Biol. Crystallogr. 70, 2197–2203 (2014). 26. T. Weinert et al., Fast native-SAD phasing for routine macromolecular structure determination. Nat. Methods 12, 131–133 (2015). 27. M. Baek et al., Accurate prediction of protein structures and interactions using a three-track neural Frizzled4, Lrp5/6, and proteoglycan. Elife 4, e06554 (2015). 65. D. J. Leahy, C. E. Dann III, P. Longo, B. Perman, K. X. Ramyar, A mammalian expression vector for expression and purification of secreted proteins for structural studies. Protein. Expr. Purif. 20, 500–506 (2000). 66. M. Illing, L. L. Molday, R. S. Molday, The 220-kDa rim protein of retinal rod outer segments is a member of the ABC transporter superfamily. J. Biol. Chem. 272, 10303–10310 (1997). 67. G. Winter, xia2: an expert system for macromolecular crystallography data reduction. J. Appl. network. Science 373, 871–876 (2021). Crystallogr. 43, 186–190 (2010). 28. J. Jumper et al., Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021). 68. J. Beilsten-Edmands et al., Scaling diffraction data in the DIALS software package: Algorithms and new approaches for multi-crystal scaling. Acta. Crystallogr. D Biol. Crystallogr. 76, 385–399 (2020). 29. J. P. Schuermann, J. J. Tanner, MRSAD: Using anomalous dispersion from S atoms collected at Cu 69. G. Winter et al., DIALS: Implementation and evaluation of a new integration package. Acta. Kalpha wavelength in molecular-replacement structure determination. Acta. Crystallogr. D Biol. Crystallogr. 59, 1731–1736 (2003). Crystallogr. D. Biol. Crystallogr. 74, 85–97 (2018). 70. P. R. Evans, G. N. Murshudov, How good are my data and what is the resolution? Acta. Crystallogr. D. 30. S. Panjikar, V. Parthasarathy, V. S. Lamzin, M. S. Weiss, P. A. Tucker, On the combination of molecular replacement and single-wavelength anomalous diffraction phasing for automated structure determination. Acta. Crystallogr. D. Biol. Crystallogr. 65, 1089–1097 (2009). Biol. Crystallogr. 69, 1204–1214 (2013). I. J. F. Tickle et al., STARANISO (Global Phasing Ltd., Cambridge, United Kingdom, 2021). 71. 72. P. Evans, Scaling and assessment of data quality. Acta. Crystallogr. D Biol. Crystallogr. 62, 72–82 31. G. Bunkoczi et al., Macromolecular X-ray structure determination using weak, single-wavelength (2006). anomalous data. Nat. Methods 12, 127–130 (2015). 73. T. Pape, Thomas R. Schneider, HKL2MAP: A graphical user interface for macromolecular phasing with 32. P. Skubak et al., A new MR-SAD algorithm for the automatic building of protein models from low- SHELX programs. J. Appl. Crystallogr. 37, 843–844 (2004). resolution X-ray data and a poor starting model. IUCrJ 5, 166–171 (2018). 74. G. M. Sheldrick, Experimental phasing with SHELXC/D/E: Combining chain tracing with density 33. D. Guzenko, S. V. Strelkov, CCFold: Rapid and accurate prediction of coiled-coil structures and application to modelling intermediate filaments. Bioinformatics 34, 215–222 (2018). 34. T. C. Terwilliger et al., Decision-making in structure solution using Bayesian estimates of map quality: the PHENIX AutoSol wizard. Acta. Crystallogr. D Biol. Crystallogr. 65, 582–601 (2009). 35. D. Liebschner et al., Macromolecular structure determination using X-rays, neutrons and electrons: modification. Acta. Crystallogr. D Biol. Crystallogr. 66, 479–485 (2010). 75. A. J. McCoy et al., Phaser crystallographic software. J. Appl. Crystallogr. 40, 658–674 (2007). 76. T. C. Terwilliger, SOLVE and RESOLVE: automated structure solution and density modification. Methods Enzymol 374, 22–37 (2003). 77. T. C. Terwilliger et al., Iterative model building, structure refinement and density modification with recent developments in Phenix. Acta. Crystallogr. D. Struct. Biol. 75, 861–877 (2019). the PHENIX AutoBuild wizard. Acta. Crystallogr. D. Biol. Crystallogr. 64, 61–69 (2008). 36. P. Kumar, D. N. Woolfson, Socket2: A program for locating, visualising, and analysing coiled-coil 78. P. Emsley, B. Lohkamp, W. G. Scott, K. Cowtan, Features and development of Coot. Acta. Crystallogr. interfaces in protein structures. Bioinformatics 37, 4575–4577 (2021). D. Biol. Crystallogr. 66, 486–501 (2010). 37. L. Truebestein, T. A. Leonard, Coiled-coils: The long and short of it. Bioessays 38, 903–916 (2016). 38. A. N. Lupas, J. Bassler, Coiled coils - a model system for the 21st century. Trends Biochem. Sci. 42, 130–140 (2017). 79. K. Cowtan, The Buccaneer software for automated model building. 1. Tracing protein chains. Acta. Crystallogr. D. Biol. Crystallogr. 62, 1002–1011 (2006). 80. F. DiMaio et al., Improved low-resolution crystallographic refinement with Phenix and Rosetta. Nat. 39. F. Benz et al., Low wnt/beta-catenin signaling determines leaky vessels in the subfornical organ and Methods 10, 1102–1104 (2013). affects water homeostasis in mice. Elife 8, e43818 (2019). 81. G. N. Murshudov et al., REFMAC5 for the refinement of macromolecular crystal structures. Acta. 40. Y. Wang et al., Beta-catenin signaling regulates barrier-specific gene expression in circumventricular Crystallogr. D. Biol. Crystallogr. 67, 355–367 (2011). organ and ocular vasculatures. Elife 8, e43257 (2019). 41. V. M. Thorner, E. Horvath, K. Kovacs, “The anterior pituitary” in Endocrinology, J. D. Wilson, D. W. 82. B. E. Bricogne, BUSTER (version 2.10.3, Global Phasing Ltd., Cambridge, United Kingdom, 2017). 83. V. B. Chen et al., MolProbity: All-atom structure validation for macromolecular crystallography. Acta. Foster, eds. (WB Saunders, Philadelphia), pp. 221–310 (1992). Crystallogr. D Biol. Crystallogr. 66, 12–21 (2010). 42. A. Ben-Zvi et al., Mfsd2a is critical for the formation and function of the blood-brain barrier. Nature 509, 507–511 (2014). 84. F. Sievers, D. G. Higgins, Clustal omega. Curr. Protoc. Bioinform. 48, 3.13.1–3.13.16 (2014). 85. X. Robert, P. Gouet, Deciphering key features in protein structures with the new ENDscript server. 43. B. Rupp, Biomolecular Crystallography: Principles, Practice, and Application to Structural Biology Nucleic Acids Res. 42, W320–W324 (2014). (Garland Science, New York, 2010), p. 809. 44. J. P. Rose, B. C. Wang, SAD phasing: History, current impact and future opportunities. Arch. Biochem. Biophys. 602, 80–94 (2016). 86. E. Krissinel, K. Henrick, Secondary-structure matching (SSM), a new tool for fast protein structure alignment in three dimensions. Acta. Crystallogr. D. Biol. Crystallogr. 60, 2256–2268 (2004). 87. M. D. Winn et al., Overview of the CCP4 suite and current developments. Acta. Crystallogr. D. Biol. 45. W. A. Hendrickson, Evolution of diffraction methods for solving crystal structures. Acta. Crystallogr. A Crystallogr. 67, 235–242 (2011). 69, 51–59 (2013). 46. B. C. Wang, Resolution of phase ambiguity in macromolecular crystallography. Methods Enzymol. 115, 90–112 (1985). 88. A. Morin et al., Collaboration gets the most out of software. Elife 2, e01456 (2013). 89. L. von Boehmer et al., Sequencing and cloning of antigen-specific antibodies from mouse memory B cells. Nat. Protoc. 11, 1908–1923 (2016). 47. W. A. Hendrickson, Determination of macromolecular structures from anomalous diffraction of 90. L. L. Molday, R. S. Molday, 1D4: a versatile epitope tag for the purification and characterization of synchrotron radiation. Science 254, 51–58 (1991). expressed membrane and soluble proteins. Methods Mol. Biol. 1177, 1–15 (2014). 48. A. C. Pike, E. F. Garman, T. Krojer, F. von Delft, E. P. Carpenter, An overview of heavy-atom 91. F. L. Hsieh, T. H. Chang, Antibody display of cell surface receptor Tetraspanin12 and SARS-CoV-2 derivatization of protein crystals. Acta. Crystallogr. D Struct. Biol. 72, 303–318 (2016). 49. M. G. Joyce, S. Radaev, P. D. Sun, A rational approach to heavy-atom derivative screening. Acta. Crystallogr. D Biol. Crystallogr. 66, 358–365 (2010). 50. S. Wenzel, T. Imasaki, Y. Takagi, A practical method for efficient and optimal production of Seleno- methionine-labeled recombinant protein complexes in the insect cells. Protein. Sci. 28, 808–822 (2019). spike protein. bioRxiv [Preprint] (2021), https://doi.org/10.1101/2021.05.29.446300 (Accessed 30 May 2021). 92. T. H. Chang et al., Crystal structure of PLVAP CC1 determined by sulfur SAD. Worldwide Protein Data Bank (wwPDB). https://doi.org/10.2210/pdb8FBY/pdb. Deposited 30 November 2022. 93. T. H. Chang et al., Crystal structure of PLVAP CC1 in I212121 space group. Worldwide Protein Data Bank (wwPDB). https://doi.org/10.2210/pdb8FCF/pdb. Deposited 1 December 2022. 12 of 12   https://doi.org/10.1073/pnas.2221103120 pnas.org
10.7554_elife.79027
RESEARCH ARTICLE Rapid encoding of task regularities in the human hippocampus guides sensorimotor timing Ignacio Polti1,2*†, Matthias Nau1,2*†, Raphael Kaplan1,3, Virginie van Wassenhove4, Christian F Doeller1,2,5* 1Kavli Institute for Systems Neuroscience, Centre for Neural Computation, The Egil and Pauline Braathen and Fred Kavli Centre for Cortical Microcircuits, Jebsen Centre for Alzheimer’s Disease, Norwegian University of Science and Technology, Trondheim, Norway; 2Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany; 3Department of Basic Psychology, Clinical Psychology, and Psychobiology, Universitat Jaume I, Castellón de la Plana, Spain; 4CEA DRF/Joliot, NeuroSpin; INSERM, Cognitive Neuroimaging Unit; CNRS, Université Paris- Saclay, Gif- Sur- Yvette, France; 5Wilhelm Wundt Institute of Psychology, Leipzig University, Leipzig, Germany Abstract The brain encodes the statistical regularities of the environment in a task- specific yet flexible and generalizable format. Here, we seek to understand this process by bridging two parallel lines of research, one centered on sensorimotor timing, and the other on cognitive mapping in the hippocampal system. By combining functional magnetic resonance imaging (fMRI) with a fast- paced time- to- contact (TTC) estimation task, we found that the hippocampus signaled behav- ioral feedback received in each trial as well as performance improvements across trials along with reward- processing regions. Critically, it signaled performance improvements independent from the tested intervals, and its activity accounted for the trial- wise regression- to- the- mean biases in TTC estimation. This is in line with the idea that the hippocampus supports the rapid encoding of temporal context even on short time scales in a behavior- dependent manner. Our results emphasize the central role of the hippocampus in statistical learning and position it at the core of a brain- wide network updating sensorimotor representations in real time for flexible behavior. Editor's evaluation This important work brings ideas about hippocampal learning and involvement in temporal processing to a sensorimotor timing task, "time- to- contact estimation", that is not typically consid- ered to be hippocampus- dependent. The study found that activity in the hippocampus measured with fMRI was related to feedback received about the accuracy of timing estimation and to perfor- mance improvement across trials in a manner not tied to the specific time interval tested. The evidence presented for the nature of the involvement of the hippocampus in this task is compelling. Introduction When someone throws us a ball, we can anticipate its future trajectory, its speed and the time it will reach us. These expectations then inform the motor system to plan an appropriate action to catch it. Generating expectations and planning behavior accordingly builds on our ability to learn from past experiences and to encode the statistical regularities of the tasks we perform. At the core of this ability lies a continuous perception- action loop, initially proposed for sensorimotor systems (e.g. *For correspondence: ignacio.polti@ntnu.no (IP); matthias.nau@ntnu.no (MN); doeller@cbs.mpg.de (CFD) †These authors contributed equally to this work Competing interest: See page 17 Funding: See page 18 Preprinted: 03 August 2021 Received: 28 March 2022 Accepted: 02 October 2022 Published: 01 November 2022 Reviewing Editor: Anna C Schapiro, University of Pennsylvania, United States Copyright Polti, Nau et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 1 of 22 Research article Wolpert et al., 2011), which is now at the heart of many leading theories of brain function including active inference (Friston et al., 2016), predictive coding (Huang and Rao, 2011) and reinforcement learning (Daw and Dayan, 2014). Critically, the brain needs to balance three primary objectives to effectively guide behavior in a dynamic environment. First, it needs to capture the specific aspects of the task that inform the relevant behavior (e.g. the remaining time to catch the ball). Second, it needs to generalize from a limited set of examples to novel and noisy situations. This can be achieved by regularizing the currently encoded information based on past experiences (e.g. by inferring how fast previous balls flew on average). Third, the sensorimotor representations that guide the behavior need to be updated flexibly when- ever feedback about our actions becomes available (e.g. when we catch or miss the ball), or when task demands change (e.g. when someone throws us a frisbee instead). Herein, we refer to these objec- tives as specificity, regularization, and flexibility. While these are all fundamental principles underlying human cognition broadly, how the brain forms and continuously updates sensorimotor representa- tions that balance these three objectives remains unclear. Here, we approach this question with a new perspective by bridging two parallel lines of research centered on sensorimotor timing and hippocampal- dependent cognitive mapping. Specifically, we test how the human hippocampus, an area often implicated in episodic- memory formation (Schiller et  al., 2015; Eichenbaum, 2017), may support the flexible updating of sensorimotor representa- tions in real time and in concert with other regions. Importantly, the hippocampus is not traditionally thought to support sensorimotor functions, and its contributions to memory formation are typically discussed for longer time scales (hours, days, weeks). Here, however, we characterize in detail the rela- tionship between hippocampal activity and real- time behavioral performance in a fast- paced timing task, which is traditionally believed to be hippocampal- independent. We propose that the capacity of the hippocampus to encode statistical regularities of our environment (Doeller et al., 2005; Schapiro et al., 2017; Momennejad, 2020) situates it at the core of a brain- wide network balancing specificity vs. regularization in real time as the relevant behavior is performed. An optimal behavioral domain to study these processes is sensorimotor timing (Gershman et al., 2014; Petter et  al., 2018). This is because prior work suggested that timing estimates indeed rely on the statistics of prior experiences (Wolpert et  al., 2011; Jazayeri and Shadlen, 2010; Acerbi et al., 2012; Chang and Jazayeri, 2018). Crucially, however, timing estimates are not always accurate. Instead, they directly reflect the trade- off between specificity and regularization, which is expressed in systematic behavioral biases. Estimated intervals regress towards the mean of the distribution of tested intervals (Jazayeri and Shadlen, 2010), a well- known effect that we will refer to as the regression effect (Petzschner et al., 2015). The regression effect suggests that the brain encodes a probability distribution of possible intervals rather than the exact information obtained in each trial (Wolpert et  al., 2011). Timing estimates therefore depend not only on the interval tested in a trial, but also on the temporal context in which they were encountered (i.e. the intervals tested in all other trials). This likely helps to predict future scenarios, to adapt behavior flexibly and to generalize to novel or noisy situations (Jazayeri and Shadlen, 2010; Acerbi et al., 2012; Roach et al., 2017). Importantly, the hippocampus proper codes for time and temporal context on various scales (Howard, 2017) and it has been shown to process behavioral feedback in decision- making tasks (Shohamy and Wagner, 2008), pointing to a role in feedback learning. Moreover, the hippocampal formation has been implicated in encoding the latent structure of a task along with the individual features that were tested (Kumaran, 2012; Schlichting and Preston, 2015; Schapiro et al., 2017; Wikenheiser et al., 2017; Behrens et al., 2018; Schuck and Niv, 2019; Whittington et al., 2020; Peer et al., 2021), providing a unified account for its many proposed roles in navigation (Burgess et al., 2002), memory (Schiller et al., 2015; Eichenbaum, 2017) and decision making (Kaplan et al., 2017; Vikbladh et al., 2019). We propose that a central function of the human hippocampus is to encode the temporal context of stimuli and behaviors rapidly, and that this process manifests as the behavioral regression effect observed in time estimation and other domains (Petzschner et al., 2015). This puts the hippocampus at the core of a brain- wide network solving the trade- off between speci- ficity and regularization for flexible behavior by continuously updating sensorimotor representations in a feedback- dependent manner. Using functional magnetic resonance imaging (fMRI) and a senso- rimotor timing task, we here test this proposal empirically. Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 2 of 22 Neuroscience Research article Results In the following, we present our experiment and results in four steps. First, we introduce our task, which built on the estimation of the time- to- contact (TTC) between a moving fixation target and a visual boundary, as well as the behavioral and fMRI measurements we acquired. On a behavioral level, we show that participants’ timing estimates systematically regress towards the mean of the tested intervals. Second, we demonstrate that anterior hippocampal fMRI activity and functional connectivity tracks the behavioral feedback participants received in each trial, revealing a link between hippo- campal processing and timing- task performance. Third, we show that this hippocampal feedback modulation reflects improvements in behavioral performance over trials. We interpret this activity to signal the updating of task- relevant sensorimotor representations in real time. Fourth, we show that these updating signals in the posterior hippocampus were independent of the specific interval that was tested and activity in the anterior hippocampus reflected the magnitude of the behavioral regres- sion effect in each trial. Notably, for each of the hippocampal main analyses, we also performed whole- brain voxel- wise analyses to uncover the larger brain network at play. We found that in addition to the hippocampus, regions typically associated with timing and reward processing signaled sensorimotor updating in our task, particularly the striatum. Follow- up analyses further revealed a striking distinction in TTC- specific and TTC- independent updating signals between striatal sub- regions. We conclude by discussing the potential neural underpinnings of these results and how the hippocampus may contribute to solving the trade- off between task specificity and regularization in concert with this larger brain network. Time-to-contact (TTC) estimation task We monitored whole- brain activity using fMRI with concurrent eye tracking in 34 participants performing a TTC task. This task offered a rich behavioral read- out and required sustained attention in every single trial. During scanning, participants visually tracked a fixation target, which moved on linear trajectories within a circular boundary. The target moved at one of four possible speed levels and in one of 24 possible directions (Figure 1A, similar to Nau et al., 2018a). The sequence of tested speeds was counterbalanced across trials. Whenever the target stopped moving, participants esti- mated when the target would have hit the boundary if it had continued moving. They did so while maintaining fixation, and they indicated the estimated TTC by pressing a button. Feedback about their performance was provided foveally and instantly with a colored cue. The received feedback depended on the timing error, that is the difference between objectively true and estimated TTC (Figure  1B), and it comprised three levels reflecting high, middle, and low accuracy (Figure  1C). Because timing judgements typically follow the Weber- Fechner law (Rakitin et al., 1998), the feed- back levels were scaled relative to the ground- truth TTC of each trial. This ensured that participants were exposed to approximately the same distribution of feedback at all intervals tested (Figure 1C, Figure 1—figure supplement 1B). After a jittered inter- trial interval (ITI), the next trial began and the target moved into another direction at a given speed. The tested speeds of the fixation target were counterbalanced across trials to ensure a balanced sampling within each scanning run. Because the target always stopped moving at the same distance to the boundary, matching the boundary’s retinal eccentricity across trials, the different speeds led to four different TTCs: 0.55, 0.65, 0.86, and 1.2 s. Each participant performed a total of 768 trials. Please see Materials and methods for more details. Analyzing the behavioral responses revealed that participants performed the task well and that the 16 ). estimated and true TTCs were tightly correlated (Figure  1B; Spearman’s rho = 0.91, p = 2.2x10− However, participants’ responses were also systematically biased towards the grand mean of the TTC distribution (0.82 s), indicating that shorter durations tended to be overestimated and longer dura- tions tended to be underestimated. We confirmed this in all participants by examining the slopes of linear regression lines fit to the behavioral responses (Figure 1—figure supplement 1D). These slopes differed from 1 (veridical performance; Figure 1B, diagonal dashed line; one- tailed one- sample t test, 2.47] ) as well as from 0 (grand mean; Figure 1B, t(33) = 16, d = 3.71, CI : [2.79, 4.72] ) horizontal dashed line; one- tailed one- sample t test, t(33) = 21.62, p = 2.2x10− and clustered at 0.5. Moreover, the slopes also correlated positively with behavioral accuracy across 08 ), consistent with participants (Figure 1—figure supplement 1E; Spearman’s rho = 0.794, p = 2.1x10− previous reports (Cicchini et al., 2012). Notably, the regression effect we observed in behavior has been argued to show that timing estimates indeed rely on the latent task regularities that our brain has 19.26, p = 2.2x10− 3.30, CI : [ 16, d = 4.22, − − − − Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 3 of 22 Neuroscience Research article A) Visual tracking & time-to-contact (TTC) estimation Typical trial Fixation Visual tracking Time Task overview Gaze trajectory 15° 5° TTC judgement Feedback 10° Boundary Four speeds Durations Feedback levels 4.2 °/s 5.8 °/s 7.5 °/s 9.1 °/s 1.2 s 0.86 s 0.67 s 0.55 s Accurate Small under-/overshoot Large under-/overshoot All scaled by speed B) TTC-task performance C) Received feedback ) s ( C T T d e t a m i t s E 1.3 0.9 0.5 100 l s a i r t t n e c r e P 50 0 0.9 True TTC (s) 1.3 High Medium Low Accuracy Figure 1. Visual tracking and Time- To- Contact (TTC) estimation task. (A) Task design. In each trial during fMRI scanning, participants fixated a target (phase 1), which started moving at one of 4 possible speeds and in one of 24 possible directions for 10° visual angle (phase 2). After the target stopped moving, participants kept fixating and estimated when the fixation target would have hit a boundary 5° visual angle apart (phase 3). After pressing a button at the estimated TTC, participants received feedback (phase 4) according to their performance. Feedback was scaled relative to target TTC. (B) Task performance. True and estimated TTC were correlated, showing that participants performed the task well. However, they overestimated short TTCs and underestimated long TTCs. Their estimates regressed towards the grand- mean of the TTC distribution (horizontal dashed line), away from the line of equality (diagonal dashed line). (C) Feedback. On average, participants received high- accuracy feedback on half of the trials (also see Figure 1—figure supplement 1B, Figure 1—figure supplement 1C). (BC) We plot the mean and SEM (black dots and lines) as well as single- participant data as dots (n=34). Feedback levels are color coded. The online version of this article includes the following figure supplement(s) for figure 1: Figure supplement 1. Behavioral analyses. Figure supplement 2. Eye tracking analyses. encoded (e.g. Jazayeri and Shadlen, 2010; Roach et al., 2017). It may therefore reflect a key behavioral adaptation helping to regularize encoded intervals to optimally support both current task performance and generalization to future scenarios. In support of this, participants’ regression slopes converged over time towards the optimal value of 0.5, that is the slope value between veridical performance and the grand mean (Figure  1—figure supplement 1F; linear mixed- effects model with task segment as a predictor and participants as the error term, F(1) = 8.172, p = 0.005, ϵ2 = 0.08, CI : [0.01, 0.18] ), and participants’ slope values became more similar (Figure 1—figure supplement 1G; linear regres- sion with task segment as predictor, F(1) = 6.283, p = 0.046, ϵ2 = 0.43, CI : [0, 1] ). Consequently, this also led to an improvement in task performance over time on group level (i.e. task accuracy and Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 4 of 22 Neuroscience Research article A) Wide-spread brain activity reflects feedback received in past trial -18 -8 -42 B) ROI analysis C) Feedback-dependent hippocampal connectivity p<0.05, FWE p<0.05, FWE T=-9.4 0 8.9 * 0.3 0 -0.3 e t a m i t s e a t e B -0.6 ant. post. Hippocampus -22 14 38 p<0.05, FWE T=0 9.7 Figure 2. Feedback on the previous trial (n- 1) modulates network- wide activity and hippocampal connectivity in subsequent trials (n). (A) Voxel- wise analysis. Activity in each trial was modeled with a separate regressor as a function of feedback received in the previous trial. Insert zooming in on hippocampus added. (B) Independent regions- of- interest analysis for the anterior (ant.) and posterior (post.) hippocampus. We plot the beta estimates obtained for the contrast between low- accuracy vs. high- accuracy feedback. Negative values indicate that smaller errors, and higher- accuracy feedback, led to stronger activity. Depicted are the mean and SEM across participants (black dot and line) overlaid on single participant data (coloured dots; n=34). Activity in the anterior hippocampus is modulated by feedback received in previous trial. Statistics reflect p<0.05 at Bonferroni- corrected levels (*) obtained using a group- level two- tailed one- sample t- test against zero. (C) Feedback- dependent hippocampal connectivity. We plot results of a psychophysiological interactions (PPI) analysis conducted using the hippocampal peak effects in (A) as a seed for low vs. high- accuracy feedback. (AC) We plot thresholded t- test results at 1 mm resolution overlaid on a structural template brain. MNI coordinates added. Hippocampal activity and connectivity is modulated by feedback received in the previous trial. The online version of this article includes the following figure supplement(s) for figure 2: Figure supplement 1. Regions of interest (ROIs). Figure supplement 2. Current trial effects. Figure supplement 3. Brain activity reflects feedback received in past trial. Figure supplement 4. Remaining contrasts from Figure 2A, Figure 2B. precision increased; Figure  1—figure supplement 1I), and the relationship between accuracy and precision became stronger (Figure 1—figure supplement 1H), linear mixed- effect model results for 5 , accuracy: F(1) = 15.127, p = 1.3x10− ϵ2 = 0.32, CI : [0.13, 1] , accuracy- precision relationship: F(1) = 8.288, p = 0.036, ϵ2 = 0.56, CI : [0, 1] , see methods for model details. 4 , ϵ2 = 0.06, CI : [0.02, 0.11] , precision: F(1) = 20.189, p = 6.1x10− Behavioral feedback predicts hippocampal activity Importantly, sensorimotor updating is expected to occur right after the value of the performed action became apparent, which is when participants received feedback. As a proxy, we therefore analyzed how activity in each voxel reflected the feedback participants received in the previous trial. Using a mass- univariate general linear model (GLM), we modeled the three feedback levels with one regressor each plus additional nuisance regressors (see Materials and methods for details). The three feedback levels (high, medium, and low accuracy) corresponded to small, medium and large timing errors, respectively. We then contrasted the beta weights estimated for low- accuracy vs. high- accuracy feed- back and examined the effects on group- level averaged across runs. We performed both whole- brain Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 5 of 22 Neuroscience Research article A) Distinct networks update TTC-specific or TTC-independent task information Regressor design TTC-independent Trial 1 2 3 4 TTC: 1.2s .55s .55s 1.2s 6 21 -30 B) ROI analysis * 0.3 e t a m i t s e a t e B 0 TTC-specific Updating > no updating: t-statistic, p<0.05, FWE TTC-independent: TTC-specific: 0 0 8.7 10.3 -0.3 post. ant. Hippocampus Figure 3. Distinct cortical and subcortical networks signal the updating of TTC- specific and TTC- independent task information. (A) Left panel: Visual depiction of parametric modulator design. Two regressors per run modeled the improvement in behavioral performance since the last trial independent of the tested TTC (Regressor 1: TTC- independent) or the improvement since the last trial when the same target TTC was tested (Regressor 2: TTC- specific). Right panel: Voxel- wise analysis results for TTC- specific and TTC- independent regressors. We plot thresholded t- test results at 1 mm resolution at p<0.05 whole- brain Family- wise- error (FWE) corrected levels overlaid on a structural template brain. Insert zooming in on hippocampus and MNI coordinates added. (B) Independent regions- of- interest analysis for the anterior (ant.) and posterior (post.) hippocampus. We plot the beta estimates obtained for TTC- independent in orange and TTC- specific regressors in blue. Depicted are the mean and SEM across participants (black dot and line) overlaid on single participant data as dots (n=34). Statistics reflect p<0.05 at Bonferroni- corrected levels (*) obtained using a group- level one- tailed one- sample t- test against zero. The online version of this article includes the following figure supplement(s) for figure 3: Figure supplement 1. Distinct networks support TTC- specific and TTC- independent updating. Figure supplement 2. TTC- independent hippocampal connectivity. voxel- wise analyses as well as regions- of- interest (ROI) analysis for anterior and posterior hippocampus separately, for which prior work suggested functional differences with respect to their contributions to memory- guided behavior (Poppenk et al., 2013; Strange et al., 2014). In both our ROI analysis and a voxel- wise analysis, we found that hippocampal activity could indeed be predicted by the feedback participants received in the previous trial (Figure  2A, Figure  2B). Higher- accuracy feedback resulted in overall stronger activity in the anterior section of the hippocampus (Figure  2B, Figure  2—figure supplement 1A; two- tailed one- sample t tests: 4, pfwe = 0.001, d = 0.28] ; posterior HPC, anterior HPC, t(33) = − 0.62, 0.07] ). Moreover, the voxel- wise analysis t(33) = 0.27, CI : [ revealed feedback- related activity also in the thalamus and the striatum (Figure 2A), and in the hippo- campus when the feedback of the current trial was modeled (Figure 2—figure supplement 2A). − 1.60, p = 0.119, pfwe = 0.237, d = 3.80, p = 5.9x10− 0.65, CI : [ 1.03, − − − − − − − − − − − − − 0.63, CI : [ 3.65, p = 8.9x10− 3.59, p = 0.002, pfwe = 0.005, d = 0.99, p = 0.329, pfwe = 0.659, d = 4, pfwe = 0.002, d = 0.25, CI : [ − 1.43, p = 0.161, pfwe = 0.322, d = Note that these results were robust even when fewer nuisance regressors were included to control for model over- specification (Figure  2—figure supplement 3B; two- tailed one- sample t 0.26] ; posterior tests: anterior HPC, t(33) = 1.01, 0.59, 0.10] ), and when all three feedback HPC, t(33) = levels were modeled with one parametric regressors (Figure 2—figure supplement 3C; two- tailed 0.20] ; one- sample t tests: anterior HPC, t(33) = 0.51, 0.17] ). In addition, hippo- posterior HPC, t(33) = campal activity was higher for medium- accuracy feedback relative to low- accuracy feedback on voxel- wise and ROI level (Figure  2—figure supplement 4A; two- tailed one- sample t tests: ante- 0.37] ; posterior HPC, rior HPC, t(33) = − 0.25] ) and for high- accuracy feedback 0.62 , CI : [ t(33) = vs. medium- accuracy feedback on voxel- wise but not ROI level (Figure 2—figure supplement 4B; two- 0.36, 0.33] ; tailed one- sample t tests: anterior HPC, t(33) = 0.01, CI : [ 0.17, 0.52] ). Further, there was no posterior HPC, t(33) = t = 0.99, p = 0.327, pfwe = 0.654, d = 0.17, CI : [ systematic relationship between subsequent trials on a behavioral level (Figure  1—figure supple- 0.17, 0.52] ; see Materials ment 1A; two- tailed one- sample t test; t(33) = 1.03, p = 0.312, d = 0.18, CI : [ and methods for details) and that the direction of the effects differed across regions (Figure  2A), speaking against potential feedback- dependent biases in attention. 4 , pfwe = 2.110 − 0.08, p = 0.933, pfwe = 1, d = 4 , pfwe = 0.002, d = 4.40, p = 1.110− 3.62, p = 9.810− 0.17, CI : [ 0.76, CI : [ 0.56, CI : [ 4 , d = − 1.00, 0.93, 1.15, − − − − − − − − − − − − − − − − Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 6 of 22 Neuroscience Research article Feedback-dependent hippocampal functional connectivity Having established that hippocampal activity reflected feedback in the TTC task, we reasoned that its activity may also show systematic co- fluctuations with other task- relevant brain regions as well. To test this, we estimated the functional connectivity of a 4 mm radius sphere centered on the hippo- campal peak main effect (x=-32, y=-14, z=-14) using a seed- based psychophysiological interaction (PPI) analysis (see Materials and methods). We reasoned that larger timing errors and therefore low- accuracy feedback would result in stronger updating compared to smaller timing errors and high- accuracy feedback, a relationship that should also be reflected in the functional connectivity between the hippocampus and other regions. We specifically tested this using the PPI analysis by contrasting trials in which participants performed poorly compared to those trials in which they performed well. We found that hippocampal activity co- fluctuated with activity in the primary motor cortex, the parahippocampal gyrus and medial parietal lobe as well as the cerebellum (Figure 2C). These co- fluc- tuations were stronger when participants performed poorly in the previous trial and therefore when they received low- accuracy feedback. Combined with the previous analysis, this means that the abso- lute hippocampal activity scaled positively (Figure 2A, Figure 2B) and functional connectivity scaled negatively (Figure 2C) with feedback valence. Hippocampal activity explains accuracy and biases in task performance Two critical open questions remained. First, did the observed feedback modulation actually reflect improvements in behavioral performance over trials? Second, was the information that was learned specific to the interval that was tested in a given trial, likely serving task specificity, or was independent of the tested interval, potentially serving regularization? To answer these questions in one analysis, we used a GLM modeling activity not as a function of feedback received in the previous trial (Figure 2), but as a function of the difference in feedback between trials (Figure 3). Specifically, we modeled with two separate parametric regressors the improvements in TTC task performance across subsequent trials (regressor 1: TTC- independent updating) as well as the improvements over subsequent trials in which the same TTC interval was tested (regressor 2: TTC- specific updating). We again accounted for nuisance variance as before, and we contrasted trials in which participants had improved versus the ones in which they had not improved or got worse (see Materials and methods for details). Because stronger sensorimotor updating should lead to larger performance improvements, we predicted to find stronger activity for improvements vs. no improvements in these tests (one- tailed hypothesis). We found both TTC- specific and TTC- independent activity throughout cortical and subcortical regions. Distinct areas engaged in either one or in both of these processes (Figure 3A, Figure 3— figure supplement 1). Crucially, we found that hippocampal activity signaled behavioral improve- ments independent of the TTC intervals tested. This effect was localized to the posterior section of the hippocampus (Figure  3B, Figure  2—figure supplement 1A; one- tailed one- sample t tests; TTC- independent: anterior HPC, t(33) = 0.36, p = 0.360 , pfwe = 1, d = 0.06, CI : [ 0.28, 0.40] , poste- rior HPC, t(33) = 2.81, p = 0.004 , pfwe = 0.017, d = 0.48, CI : [0.12, 0.85] ; TTC- specific: anterior HPC, t(33) = 1.29, p = 0.103 , t(33) = 0.57, p = 0.285 , 0.12, 0.57] ). We then again estimated the functional connectivity profile of pfwe = 0.413, d = 0.22, CI : [ the hippocampal main effect using a PPI analysis (sphere with 4 mm radius centered on the peak voxel at x=-30, y=-24, z=-18), revealing co- fluctuations in multiple regions including the putamen and the thalamus that were specific to behavioral improvements (Figure 3—figure supplement 2). 0.24, 0.44] , posterior HPC, pfwe = 1, d = 0.10, CI : [ − − − We reasoned that updating TTC- independent information may support generalization perfor- mance by means of regularizing the encoded intervals based on the temporal context in which they were encoded. In our task, an efficient way of regularizing the encoded information is to bias one’s TTC estimates towards the mean of the TTC distribution, which corresponds to the regression effect that we observed on a behavioral level (Figure  1B, Figure  1—figure supplement 1D). Given the hippocampal feedback modulation and updating activity we reported above, we hypothesized that hippocampal activity should therefore also reflect the magnitude of the regression effect in behavior. To test this in a final analysis, we modeled the activity in each trial parametrically either as a function of performance (i.e. the absolute difference between estimated and true TTC) or as a function of the strength of the regression effect in each trial (i.e. the absolute difference between the estimated TTC and the mean of the tested intervals). Voxel- wise weights for these two regressors were estimated in two independent GLMs (see Materials and methods for details). Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 7 of 22 Neuroscience Research article A) Brain activity reflects TTC-task performance and the magnitude of the regression effect -2 -23 -6 Performance: F=0 p<0.05, FWE 149 Regression effect: F=0 p<0.05, FWE 142 B) ROI analysis e t a m i t s e a t e B 1 0 −1 −2 Performance * * Regression effect * ant. post. ant. post. Hippocampus Figure 4. TTC- task performance vs. behavioral regression effect. (A) Voxel- wise analysis. We plot thresholded F- test results for the task- performance regressor and the regression- to- the- mean regressor at 1 mm resolution overlaid on a structural template brain. MNI coordinates added. Distinct networks reflect task performance and the magnitude of the regression effect. (B) Independent regions- of- interest analysis for the anterior (ant.) and posterior (post.) hippocampus. We plot the beta estimates obtained for each participant for each of the two regressors. Negative values indicate a linear increase between hippocampal activity and either task performance (left, Performance) or the magnitude of the regression effect (right, Regression effect). Depicted are the mean and SEM across participants (black dot and line) overlaid on single participant data (colored dots; n=34). Statistics reflect p<0.05 at Bonferroni- corrected levels (*) obtained using a group- level two- tailed one- sample t- test against zero. 5 − − , 0.49 , pfwe = 5.8x10− 4.85, p = 2.9x10− 2.88, p = 0.007 , 5 , pfwe = 0.014, d = Our analyses showed that trial- wise hippocampal activity increased with better TTC- (Figure  4A, Figure  4B; two- tailed one- sample t tests; anterior HPC, task performance posterior HPC, t(33) = 0.83, CI : [ 0.14] ), and consistently also with t(33) = 0.86, the valence of the feedback received in the current trial (Figure  2—figure supplement 2). In addition, however, and as predicted, it also reflected the trial- wise magnitude of the behav- ioral regression effect (Figure  4A, Figure  4B; two- tailed one- sample t tests; anterior HPC, HPC, t(33) = 0.53, 0.16] ). Activity in the anterior hippocampus t(33) = was stronger in trials in which participants’ TTC estimates were more biased towards the mean of the sampled intervals (indicated by a negative beta estimate). Notably, similar effects were observed in prefrontal and posterior cingulate areas (Figure 4A). 5.55, p = 3.6x10− 1.06, p = 0.295, pfwe = 0.886, d = 5, d = 0.18, CI : [ 6, pfwe = 1.1x10− d = − CI : [ − 0.95, CI : [ posterior 0.55] ; 0.44] ; 1.24, 1.37, − − − − − − − − − − − 0.04, r = 0.32, 0.22] ; 4.17°/s vs.9.09°/s: V = 161 , p = 0.019 , pfwe = 0.112 , r = Eye tracking: no relevant biases in viewing behavior To ensure that our results could not be attributed to systematic error patterns in viewing behavior, we analyzed the co- recorded eye tracking data of our participants in detail. After data cleaning (see Materials and methods), we used Wilcoxon signed- rank tests for paired samples to control for differ- ences in fixation accuracy across speed levels (Figure  1—figure supplement 2A; 4.17°/s vs.5.81°/s: p = 0.012 , V = 171, p = 0.03, pfwe = 0.179 , 0.08, 0.06, CI : [ pfwe = 0.071 , r = 0.28, 0.26] ; 5.81°/s CI : [ − vs.9.09°/s: V = 217, p = 0.174, pfwe = 1 , r = 0.31, 0.23] ; 7.45°/s vs.9.09°/s: V = 263, 0.28, 0.26] ) and accuracy levels (Figure  1—figure supplement pfwe = 1 p = 0.566, , r = 2B; Low vs. Medium: V = 380, p = 0.163, pfwe = 0.489 , r = 0.02, CI : [ 0.25, 0.29] ; Low vs. High: V = 366 , 0.04 , p = 0.249 , pfwe = 0.747 , r = 0.03 , CI : [ 0.31, 0.23] ). Moreover, we examined the relationship of the fixation error with TTC- task performance CI : [ (Figure 1—figure supplement 2C; Spearman’s rho = 0.17 , p = 0.344 ) as well as with the behavioral regres- sion effect (Figure 1—figure supplement 2C; Spearman’s rho = 0.26 , p = 0.131 ). None of these control analyses suggested that biased patterns in viewing behavior could hinder the interpretation of our results. 0.34, 0.2] ; 5.81°/s vs.7.45°/s: V = 224, p = 0.215 , pfwe = 1 , r = 0.24, 0.3] ; Medium vs. High: V = 278 , p = 0.748 , pfwe = 1 , r = 0.31, 0.23] ; 4.17°/s vs.7.45°/s: 0.04, CI : [ 0.01, CI : [ 0.01, CI : [ V = 152 , CI : [ − − − − − − − − − − − − − − − Discussion This study investigated how the human brain flexibly updates sensorimotor representations in a feedback- dependent manner in the service of timing behavior. We specifically focused on the hippo- campus, due to its known role in temporal coding and learning, asking how hippocampal processing Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 8 of 22 Neuroscience Research article may support behavioral flexibility, specificity, and regularization. Because anterior and posterior sections of the hippocampus differ in whole- brain connectivity as well as in their contributions to memory- guided behavior (Strange et  al., 2014), we analyzed the two sections separately. More- over, we explored the larger brain- wide network involved in balancing these objectives. To do so, we monitored human brain activity with fMRI while participants estimated the time- to- contact between a moving target and a visual boundary. This allowed us to analyze brain activity as a function of task performance and as a function of the improvements in performance over time. We found that anterior hippocampal activity as well as functional connectivity reflected the feedback participants received during this task, and its activity followed the performance improvements in a temporal- context- dependent manner. Its activity reflected trial- wise behavioral biases towards the mean of the sampled intervals, and activity in the posterior hippocampus signaled sensorimotor updating independent of the specific intervals tested. In what follows, we discuss our results in the context of prior work on timing behavior and on hippocampal spatiotemporal coding. Moreover, we elaborate on the domain- general nature of hippocampal- cortical interactions and on the sensorimotor updating mechanisms that potentially underlie the effects observed in this study. Spatiotemporal coding in the hippocampus The hippocampus encompasses neurons sensitive to elapsed time (Paton and Buonomano, 2018; Eichenbaum, 2014; Umbach et al., 2020). These cells might play an important role in guiding timing behavior (Nobre and van Ede, 2018), which potentially explains why hippocampal damage or inacti- vation impairs the ability to estimate durations in rodents (Meck et al., 1984) and humans (Richards, 1973). Our results are in line with these reports, showing that hippocampal fMRI activity also reflects participants’ TTC estimation ability (Figure 4). They are also in line with other human neuroimaging studies suggesting that the hippocampus bridges temporal gaps between two stimuli during trace eyeblink conditioning (Cheng et al., 2008), and that it represents duration within event sequences (Barnett et al., 2014; Thavabalasingam et al., 2018; Thavabalasingam et al., 2019). Our results speak to the above- mentioned reports by revealing that the hippocampus is an integral part of a widespread brain network contributing to sensorimotor updating of encoded intervals in humans (Figure  2, Figure  3, Figure  4, Figure  2—figure supplement 2, Figure  3—figure supple- ment 1, Figure 3—figure supplement 2). Moreover, they demonstrate a direct link between hippo- campal activity, the feedback participants received and the behavioral improvements expressed over time (Figure 3), emphasizing its role in feedback learning. Critically, the underlying process must occur in real- time when feedback is presented, suggesting that it plays out on short time scales. Notably, the human hippocampus is neither typically linked to sensorimotor timing tasks such as ours, nor is its activity considered to reflect temporal relationships on such short time scales. Instead, human hippo- campal processing is often studied in the context of much longer time scales (Schiller et al., 2015; Eichenbaum, 2017), which showed that it may support the encoding of the progression of events into long- term episodic memories (Deuker et al., 2016; Montchal et al., 2019; Bellmund et al., 2022) or contribute to the establishment of chronological relations between events in memory (Gauthier et al., 2019; Gauthier et  al., 2020). Intriguingly, the mechanisms at play may build on similar temporal coding principles as those discussed for motor timing (Yin and Troger, 2011; Eichenbaum, 2014; Howard, 2017; Palombo and Verfaellie, 2017; Nobre and van Ede, 2018; Paton and Buonomano, 2018; Bellmund et al., 2020; Bellmund et al., 2022; Shikano et al., 2021; Shimbo et al., 2021), with differential contributions of the anterior and posterior hippocampus. Note that our observation of distinct activity modulations in the anterior and posterior hippocampus suggests that the functions and coding principles discussed here may be mediated by at least partially distinct populations of hippocampal cells. Our task can be solved by estimating temporal intervals directly, but also by extrapolating the movement of the fixation target over time, shifting the locus of attention towards the target boundary (Figure 1). The brain may therefore likely monitor the temporal and spatial task regularities in parallel. Participants’ TTC estimates were further informed exclusively by the speed of the target, which inher- ently builds on tracking kinematic information over time, which may explain why TTC tasks also engage visual motion regions in humans (de Azevedo Neto and Amaro Júnior, 2018). While future studies could tease apart spatial and temporal factors explicitly, our results are in line with both accounts. For example, the hippocampus and surrounding structures represent maps of visual space in primates, Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 9 of 22 Neuroscience Research article which potentially mediate a coordinate system for planning behavior, integrating visual information with existing knowledge and to compute vectors in space (Nau et al., 2018b; Bicanski and Burgess, 2020). These visuospatial representations are perfectly suited to guide attention and therefore the relevant behaviors in our task (Aly and Turk- Browne, 2017), which could be tested in the future akin to prior work using a similar paradigm (Nau et al., 2018a). The role of feedback in timed motor actions Importantly, our results neither imply that the hippocampus acts as an ‘internal clock’, nor do we think of it as representing action sequences or coordinating motor commands directly. Rather, its activity may indicate the feedback- dependent updating of encoded information more generally and indepen- dent of the task that was used. The hippocampal formation has been proposed as a domain- general learning system (Kumaran, 2012; Schlichting and Preston, 2015; Chersi and Burgess, 2015; Scha- piro et al., 2017; Wikenheiser et al., 2017; Behrens et al., 2018; Bellmund et al., 2018; Vikbladh et al., 2019; Geerts et al., 2020; Momennejad, 2020; Bellmund et al., 2022), which may encode the structure of a task abstracted away from our immediate experience. In contrast, the striatum was proposed to encode sensory states or actions, supporting the learning of task- specific (egocentric) information (Chersi and Burgess, 2015; Geerts et al., 2020). Together, the two regions may there- fore play an important role in decision making in general also in other non- temporal domains. Consistent with these ideas, we observed that striatal and hippocampal activity was modulated by behavioral feedback received in each trial (Figure 2, Figure 2—figure supplement 1). Similar feed- back signals have been previously linked to learning (Schönberg et al., 2007; Cohen and Ranganath, 2007; Shohamy and Wagner, 2008; Foerde and Shohamy, 2011; Wimmer et  al., 2012) and the successful formation of hippocampal- dependent long- term memories in humans (Wittmann et  al., 2005). Moreover, hippocampal activity is known to signal learning in other tasks (Doeller et al., 2008; Foerde and Shohamy, 2011; Dickerson and Delgado, 2015; Wirth et al., 2009; Schapiro et al., 2017; Kragel et al., 2021). Here, we show a direct relationship between hippocampal activity and ongoing timing behavior, and we show that receiving behavioral feedback modulates widespread brain activity (Figure 2, Figure 2—figure supplement 1), which potentially reflects the involvement of these areas in the coordination of reward behavior observed earlier (LeGates et al., 2018). These regions include those serving sensorimotor functions, but also those encoding the structure of a task or the necessary value functions associated with specific actions (Lee et al., 2012). The present study further demonstrates that activity in the hippocampus co- fluctuates with activity in other likely task- relevant regions in a task- dependent manner. We observed such co- fluctuations in the striatum and cerebellum, often associated with reward processing and action coordination (Bostan and Strick, 2018; Cox and Witten, 2019), the motor cortex, typically involved in action plan- ning and execution, as well as the parahippocampal gyrus and medial parietal lobe, often associated with visual- scene analysis (Epstein and Baker, 2019). This may indicate that behavioral feedback also affects the functional connectivity profile of the hippocampus with those domain- selective regions that are currently engaged in the ongoing task. In the present report, this included the motor cortex, the parahippocampal gyrus, the medial parietal lobe and the cerebellum. This may also relate to previous reports of the cerebellum contributing temporal signals to cortical regions during similar tasks as ours (O’Reilly et al., 2008). Interestingly, we observed that functional connectivity of the anterior hippo- campus scaled negatively (Figure 2C) with feedback valence, unlike its absolute activity, which scaled positively with feedback valence (Figure 2A, Figure 2B), suggesting that the two measures may be sensitive to related but distinct processes. What might be the neural mechanism underlying sensorimotor updating signals in our task? Prior work has shown that hippocampal, frontal and striatal temporal receptive fields scale relative to the tested intervals, and that they re- scale dynamically when those tested intervals change (MacDonald et  al., 2011; Gouvêa et  al., 2015; Mello et  al., 2015; Wang et  al., 2018). This may enable the encoding and continuous maintenance of optimal task priors, which keep our actions well- adjusted to our current needs. We speculate that such receptive- field re- scaling also underlies the continuous updating effects discussed here. Consistent with this idea and the present results, receptive- field re- s- caling can occur on a trial- by- trial basis in the hippocampus (Shikano et al., 2021; Shimbo et al., 2021) but also in other regions such as the striatum and frontal cortex (Mello et al., 2015; Gouvêa et al., 2015; Wang et al., 2018). Such network- wide receptive- field re- scaling likely builds on a re- weighting Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 10 of 22 Neuroscience Research article of functional connections between neurons and regions, which may explain why anterior hippocampal connectivity correlated negatively with feedback valence in our data. Larger errors may have led to stronger re- scaling, which may be grounded in a corresponding change in functional connectivity. A trade-off between specificity and regularization? So far, we discussed how the brain may capture the temporal structure of a task and how the hippo- campus supports this process. However, how do we encode specific task details while still forming representations that generalize well to new scenarios? In other words, how does the brain encode the probability distribution of the intervals we tested optimally without overfitting? Our behavioral and neuroimaging results suggest that this trade- off between specificity and regularization is governed by many regions, updating different types of task information in parallel (Figure 3A). For example, hippo- campal activity reflected performance improvements independent of the tested interval, whereas the caudate signaled improvements specifically over those trials in which the same TTC was tested. In the putamen, we found evidence for both processes (Figure  3—figure supplement 1B). This suggests that different regions encode distinct task regularities in parallel to form optimal sensorim- otor representations to balance specificity and regularization. This is in line with our behavioral results, showing that TTC- task performance became more optimal in the face of both of these two objec- tives. Over time, behavioral responses clustered more closely between the diagonal and the average line in the behavioral response profile (Figure 1B, Figure 1—figure supplement 1G), and the TTC error decreased over time. While different participants approached these optimal performance levels from different directions, either starting with good performance or strong regularization, the group approached overall optimal performance levels over the course of the experiment. Because hippocampal activity (Julian and Doeller, 2020) and the regression effect (Jazayeri and Shadlen, 2010) were previously linked to the encoding of context, we reasoned that hippocampal activity should also be related to the regression effect directly. This may explain why hippocampal activity reflected the magnitude of the regression effect as well as behavioral improvements inde- pendently from TTC, and why it reflected feedback, which informed the updating of the internal prior. Notably, our results make a central prediction for future research. We anticipate that partic- ipants with stronger updating activity in the hippocampus should be able to generalize better to new scenarios, for example when new intervals are tested. While we could not test this prediction directly in our study, we did test for a link to a related phenomenon, and that is the regression effect we observed on the behavioral level. We found that TTC estimates regressed towards the mean of the sampled intervals in all participants (Figure 1B, Figure 1—figure supplement 1D), an effect that is well known in the timing literature (Jazayeri and Shadlen, 2010) and other domains (Petzschner and Glasauer, 2011; Petzschner et  al., 2015). This regression effect likely reflects regularization in support of generalization (Roach et al., 2017), because time estimates are biased towards the mean of the tested intervals, and because the mean will likely be close to the mean of possible future intervals. We therefore hypothesized that this effect is grounded in the activity of the hippocampus, because it plays a central role in generalization in other non- temporal domains (Kumaran, 2012; Schlichting and Preston, 2015; Schapiro et  al., 2017; Momennejad, 2020). Our analyses revealed that this was indeed the case. We found that hippocampal activity followed the magnitude of the regression effect in each trial (Figure 4), potentially reflecting the temporal- context- dependent regularization of encoded intervals toward the grand mean of the tested inter- vals (Jazayeri and Shadlen, 2010). In addition, our voxel- wise results showed that striatal subregions only tracked how accurate partic- ipants’ responses were, not how strongly they regressed towards the mean (Figure 4A). This dovetails with literature on spatial- navigation (Doeller et al., 2008; Chersi and Burgess, 2015; Goodroe et al., 2018; Gahnstrom and Spiers, 2020; Geerts et  al., 2020; Wiener et  al., 2016), showing that the striatum supports the reinforcement- dependent encoding of locations relative to landmarks, whereas the hippocampus may help to encode the structure of the environment in a generalizable and map- like format. This matches the functional differences observed here in the time domain, where caudate activity reflects the encoding of individual details of our task such as the TTC intervals (Figure 3A, Figure 3—figure supplement 1A, Figure 3—figure supplement 1A, Figure 3—figure supplement 1B), while the hippocampus generalizes across TTCs to encode the overall task structure (Figure 3A, Figure 3B, Figure 3—figure supplement 1A). Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 11 of 22 Neuroscience Research article Conclusion In sum, we combined fMRI with time- to- contact estimations to show that the hippocampus supports the formation of task- specific yet flexible and generalizable sensorimotor representations in real time. Hippocampal activity reflected trial- wise behavioral feedback and the behavioral improve- ments across trials, suggesting that it supports sensorimotor updating even on short time scales. The observed updating signals were independent from the tested intervals, and they explained the regression- to- the- mean biases observed on a behavioral level. This is in line with the notion that the hippocampus encodes temporal context in a behavior- dependent manner, and that it supports finding an optimal trade off between specificity and regularization along with other regions. We show that it does so even in a fast- paced timing task typically considered to be hippocampal- independent. Our results show that the hippocampus supports rapid and feedback- dependent updating of sensorimotor representa- tions, suggesting that it is a central component of a brain- wide network balancing task specificity vs. regularization for flexible behavior in humans. Table 1. Target TTCs’ response windows for each feedback level. Target TTC = 0.55 s Accuracy High Medium Low Target TTC = 0.67 s Accuracy High Medium Low Target TTC = 0.86 s Accuracy High Medium Low Target TTC = 1.2 s Accuracy High Response window (s) 0.47–0.63 0.38–0.47 | 0.63–0.71 <0.38 | >0.71 Response window (s) 0.57–0.77 0.47–0.57 | 0.77–0.87 <0.47 | >0.87 Response window (s) 0.73–0.99 0.60–0.73 | 0.99–1.12 <0.60 | >1.12 Response window (s) 1.02–1.38 Medium 0.84–1.02 | 1.38–1.56 Low <0.84 | >1.56 Materials and methods Participants We recruited 39 healthy volunteers with normal to corrected- to- normal vision for this study (16 females, 19–35 years old). Five participants were excluded: one participant did not comply with the task instructions; one was excluded due to a failure of the eye- tracker calibration; three partici- pants were excluded due to technical issues during scanning. A total of 34 participants entered the analysis. The sample size was chosen to accord with previous publications using similar procedures (Nau et al., 2018a; Montchal et al., 2019; Schuck and Niv, 2019). The study was approved by the regional committee for medical and health research ethics (project number 2017/969) in Norway and participants gave written consent prior to scanning in accordance with the declaration of Helsinki (World Medical Association, 2013). Task Participants performed two tasks simultaneously: a smooth pursuit visual- tracking task and a time- to- contact estimation task. The visual tracking task entailed fixation at a fixation disc that moved on predefined linear trajectories with one of four speeds: 4.17°/s, 5.81°/s, 7.45°/s, and 9.09°/s. Upon reaching the end of such a linear trajectory, the dot stopped moving until the second task was completed. This second task was a time- to- collision (TTC) estimation task in which participants indicated when the fixation target would have hit a circular boundary if it had continued moving. This boundary was a yellow circular line surrounding the target trajectory with 10° radius. Participants gave their response by pressing a button at the anticipated moment of collision. They performed this task while still keeping fixation, and the individual linear trajectories were all of the same length (10° visual angle), leading to four target TTC durations of 1.2 s, 0.88 s, 0.67 s, and 0.55 s tested in a counterbal- anced fashion across trials. After the button press, participants received feedback for 1 s informing them about the accuracy of their response. When participants overestimated the TTC, half of the fixation disc closest to the boundary changed color (orange or red) as a function of response accu- racy (medium or low, respectively). When participants underestimated the TTC, half of the fixation Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 12 of 22 Neuroscience Research article ∗ ± ((k disc further away from the boundary changed color. When participants were accurate, two opposing quadrants of the fixation disc would turn green. This allowed us to present feedback at fixation while keeping the number of informative pixels matched across feedback levels. To calibrate performance feedback across different TTC durations, the precise response window widths of each feedback level scaled with the speed of the fixation target (Table 1). The following formula was used to scale the response window width: d d)/2) where d is the target TTC and k is a constant proportional to 0.3 and 0.6 for high and medium accuracy, respectively. This ensured that participants received approx- imately the same feedback for tested TTCs despite the known differences in absolute performance between target TTCs due to inherent scalar variability (Gibbon, 1977). When no response was given, participants received low- accuracy feedback (two opposing quadrants of the fixation dot turned red) after a 4 s timeout. After the feedback, the disc remained in its last position for a variable inter- trial interval (ITI) sampled randomly from a uniform distribution between 0.5 s and 1.5 s. Following the end of the ITI, the dot continued moving in a different direction. In the course of 768 trials, each target TTC was sampled 192 times. We sampled eye- movement directions with 15° resolution, leading to an overall trajectory that was star- shaped, similar to earlier reports (Nau et al., 2018a). The full trajectory was never explicitly shown to the participants. Behavioral analysis Participants indicated the estimated TTC in each trial via button press. In line with previous work (Jazayeri and Shadlen, 2010), participants tended to overestimate shorter durations and under- estimate longer durations (Figure 1B). In order to quantify this behavioral effect we extracted the slope value of a linear regression line fit between estimated and target TTCs separately for each participant. A slope of 1 would indicate that participants performed perfectly accurately for all inter- vals. A slope of 0 would indicate that participants always gave the same response independent of the tested interval, fully regressing to the mean of the sampled intervals. Two separate one- tailed one- sample t tests (against 1 or 0) were performed to corroborate that participants’ slope values regressed towards the mean of the sampled TTCs (Figure  1—figure supplement 1D). A Spearman’s rank- order correlation tested if slope values correlated with the percent of high accuracy trials (Figure 1—figure supplement 1E), to further demonstrate that participants relied to different degrees on both, the target TTCs and the mean of the sampled TTCs, in order to achieve an optimal performance tradeoff. As the TTC task progressed, it would be expected that participants adjusted their TTC estimates in order to find the best tradeoff. Thus, we tested if the slope converged over time towards the value of 0.5 (the slope value between veridical performance and the mean of the sampled TTCs) by using a linear mixed- effects model with task segment as a predictor, the absolute difference between the slope and the value of 0.5 as the dependent variable and participants as the error term (Figure  1—figure supplement 1F). We also corroborated this effect by measuring the dispersion of slope values between participants across task segments using a linear regression model with task segment as a predictor and the standard deviation of slope values across participants as the dependent variable (Figure 1—figure supplement 1G). As a measure of behavioral performance, we computed two variables for each target- TTC level: sensorimotor timing accuracy, defined as the absolute difference in estimated and true TTC, and sensorimotor timing precision, defined as coeffi- cient of variation (standard deviation of estimated TTCs divided by the average estimated TTC). To study the interaction between these two variables for each target TTC over time, we first normalized accuracy by the average estimated TTC in order to make both variables comparable. We then used a linear mixed- effects model with precision as the dependent variable, task segment and normalized accuracy as predictors and target TTC as the error term. In addition, we tested whether accuracy and precision increased over the course of the experiment using a linear mixed effects model with task segment as predictor and participants as the error term. Participants received feedback after each trial corresponding to the absolute TTC error of that trial. On average, 46.9% ( σ = 9.1 ) of trials were of high accuracy, 31.2% ( σ = 3.9 ) were of medium accuracy and 21.1% ( σ = 9.8 ) were of low accuracy (Figure 1C). Moreover, we found that this feedback distribution was indeed similar across target- TTC levels as planned (Figure 1—figure supplement 1B), as well as across TTC over- and underestima- tion trials (Figure 1—figure supplement 1C). To control that there was no systematic and predict- able relationship between subsequent trials on a behavioral level, we estimated the n- 1 Pearson autocorrelation between feedback values received on each trial and then performed a two- tailed Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 13 of 22 Neuroscience Research article one- sample t- test on group level against zero using the extracted correlation coefficients from each participant (Figure 1—figure supplement 1A). Imaging data acquisition and preprocessing Imaging data were acquired on a Siemens 3T MAGNETOM Skyra located at the St. Olavs Hospital in Trondheim, Norway. A T1- weighted structural scan was acquired with 1 mm isotropic voxel size. Following EPI- parameters were used: voxel size = 2 mm isotropic, TR = 1020ms, TE = 34.6ms, flip angle = 55°, multiband factor = 6. Participants performed a total of four scanning runs of 16–18 min each including a short break in the middle of each run. Functional images were corrected for head motion and co- registered to each individual’s structural scan using SPM12 (https://www.fil.ion.ucl.ac. uk/spm/). We used the FSL topup function to correct field distortions based on one image acquired with (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/topup). Functional images were then spatially normalized to the Montreal Neurological Institute (MNI) brain template and smoothed with a Gaussian kernel with full- width- at- half- maximum of 4 mm for regions- of- interest analysis or with 8 mm for whole- brain analysis. Time series were high- pass filtered with a 128 s cut- off period. The results of all voxel- wise analyses were overlaid on a structural T1- template (colin27) of SPM12 for visualization. inverted phase- encoding direction Regions of interest definition and analysis Regions- of- interest masks for different brain areas were generated for each individual participant based on the automatic parcellation derived from FreeSurfer’s structural reconstruction (https://surfer. nmr.mgh.harvard.edu/). The ROIs used in the present study include the Hippocampus as main area of interest (Figure 2—figure supplement 1A) as well as the Caudate Nucleus, Nucleus Accumbens, Thalamus, Putamen, Amygdala, and Globus Pallidum (Figure 2—figure supplement 1B). The hippo- campal ROI was manually segmented following previous reports into its anterior and posterior sections based on the location of the uncal apex in the coronal plane as a bisection point (Poppenk et al., 2013). All individual ROIs were then spatially normalized to the MNI brain template space and re- sliced to the functional imaging resolution using SPM12. All ROI analyses were conducted using 4 mm spatial smoothing. All ROI analyses described in the following were conducted using the following procedure. We extracted beta estimates estimated for the respective regressors of interest for all voxels within a region in both hemispheres, averaged them across voxels within that region and hemispheres and performed one- sample t- tests on group level against zero as implemented in the software R (https:// www.R-project.org). Brain activity as a function of feedback on the previous trial To examine how feedback modulates activity in the subsequent trial, we used a mass- univariate general linear model (GLM) analysis to model the activity of each voxel and trial as a function of feedback received in the previous trial. The GLM included three boxcar regressors modeling the trial period for each feedback level, plus one boxcar regressor for ITIs, one for button presses and one for periods of rest (inter- session interval, ISI), which were all convolved with the canonical hemodynamic response function of SPM12. The start of the trial was considered as the trial onsets for modeling (i.e. the time when the visual- tracking target started moving). The trial end was the offset of the feedback phase (i.e. the moment in which the feedback disappeared from the screen). The ITI was the time between the offset of the feedback- phase and the subsequent trial onset. In addition, the model included the six realignment parameters obtained during pre- processing as well as a constant term modeling the mean of the time series. On the group level, we then contrasted the weights obtained for the low- accuracy vs. high- accuracy feedback regressors and tested for differences using t- tests implemented in SPM12 (Figure 2A). Additionally, we again conducted ROI analyses for the anterior and posterior sections of the hippocampus (Figure 2—figure supplement 1A) following the same procedure as described earlier (section "Regions of interest definition and analysis"). Here, we tested beta estimates obtained in the first- level analysis for the feedback- in- previous- trial regressor of interest (Figure 2B). ITIs and ISIs were modeled to reduce task- unrelated noise, but to ensure that this did not lead to over- specification of the above- described GLM, we repeated the full analysis without modeling the Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 14 of 22 Neuroscience Research article two. All other regressors including the main feedback regressors of interest remained unchanged, and we repeated both the voxel- wise and ROI- wise statistical tests as described above (Figure 2—figure supplement 3B). Moreover, instead of modeling the three feedback levels with three independent regressors, we repeated the analysis modeling the three feedback levels as one parametric regressor with three levels. In addition, one boxcar regressor was added to model all trial periods independent from feed- back level. All other regressors remained unchanged, and the model included the regressors for ITIs and ISIs. We then conducted t- tests implemented in SPM12 using the beta estimates obtained for the parametric feedback regressor (Figure 2—figure supplement 3C). Compared to the initial analyses presented above, this has the advantage that medium- accuracy feedback trials are considered for the statistics as well. Hippocampal functional connectivity as a function of previous-trial feedback We conducted a psychophysiological interactions (PPI) analysis to examine whether hippocampal functional connectivity with the rest of the brain depended on the participant’s performance on the previous trial. To do so, we centered a sphere onto the group- level peak effects within the HPC using main- effect GLM described in the previous section. The sphere was 4 mm in radius and was centered on the following MNI coordinates: x=-32, y=-14, z=-14. The GLM included a PPI regressor, a nuisance regressor accounting for the main effect of past- trial performance, and a nuisance regressor explaining variance due to inherent physiological signal correlations between the HPC and the rest of the brain. The PPI regressor was an interaction term containing the element- by- element product of the task time course (effects due to past- trial performance) and the HPC spherical seed ROI time course. The PPI model was built using the same model that revealed the main effects used to define the HPC sphere. The estimated beta weight corresponding to the interaction term was then tested against zero on the group- level using a t- test implemented in SPM12 (Figure 2C). The contrast reflects the difference between low vs. high- accuracy feedback. This revealed brain areas whose activity was co- varying with the hippocampus seed ROI as a function of past- trial performance (n- 1). Brain activity as a function of current-trial performance and feedback In two independent GLMs, we analyzed the time courses of all voxels in the brain as a function of behavioral performance (i.e. TTC error) in each trial, and as a function of feedback received at the end of each trial. The models included one mean- centered parametric regressor per run, modeling either the TTC error or the three feedback levels in each trial, respectively. Note that the feedback itself was a function of TTC error in each trial (high accuracy = 0, medium accuracy = 0.5 and low accuracy = 1). In addition, we added three nuisance regressors per run modeling ITIs, button presses, and periods of rest. These regressors were convolved with the canonical hemodynamic response function of SPM12. Moreover, the realignment parameters and a constant term were again added. We estimated weights for all regressors and conducted a t- test against zero using SPM12 for our feedback and performance regressors of interest on the group level (Figure  2—figure supplement 2A). Importantly, positive t- scores indicate a positive relationship between fMRI activity and TTC error and hence with poor behavioral performance. Conversely, negative t- scores indicate a negative relation between the two variables and hence better behavioral performance. In addition to the voxel- wise whole- brain analyses described above, we conducted independent ROI analyses for the anterior and posterior sections of the hippocampus (Figure 2—figure supple- ment 1A). Here, we tested the beta estimates obtained in our first- level analysis for the feedback and performance regressors of interest (Figure 2—figure supplement 2B; two- tailed one- sample t tests: anterior HPC, t(33) = 0.6] ; posterior 5.92, p = 1.2x10− − 4, pfwe = 5.4x10− HPC, t(33) = 0.32] ). See section ‘Regions of 4.07, p = 2.7x10− interest definition and analysis’ for more details. 6, pfwe = 2.4x10− 4, d = 1.02, CI : [ 0.7, CI : [ − 1.09, 6, d = 1.45, − − − − − − Brain activity as a function of improvements in behavioral performance across trials We used a GLM to analyze activity changes associated with behavioral improvements across trials. One regressor modeled the main effect of the trial and two parametric regressors modeled the Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 15 of 22 Neuroscience Research article following contrasts: Parametric regressor 1: trials in which behavioral performance improved vs. parametric regressor 2: trials in which behavioral performance did not improve or got worse rela- tive to the previous trial. These regressors modeled the behavioral improvements either relative to the previous trial, and therefore independently of TTC (likely serving regularization), or relative to the previous trial in which the same target TTC was presented (likely serving specificity). These two regressors reflect the tests for target- TTC- independent and target- TTC- specific updating, respec- tively, and they were not orthogonalized to each other. Because we predicted to find stronger activity for improvements vs. no improvements in behavioral performance, we here performed one- tailed statistical tests, consistent with the direction of this hypothesis. Improvement in perfor- mance was defined as receiving feedback of higher valence than in the corresponding previous trial. The same nuisance regressors were added as in the other GLMs and all regressors except the realignment parameters and the constant term were convolved with the canonical hemody- namic response function of SPM12. On the group level, we tested the two parametric regres- sors of interest against zero using a t- test implemented in SPM12, effectively contrasting trials in which behavioral performance improved against trials in which behavioral performance did not improve or got worse relative to the respective previous trials (Figure 3A). All runs were modeled separately. Moreover, we again conducted ROI analyses for the anterior and posterior sections of the hippocampus (Figure 2—figure supplement 1A) following the same procedure as described earlier (see section ‘Regions of interest definition and analysis’). Here, we tested beta estimates obtained in the first- level analysis for the TTC- specific and TTC- independent updating regressors using one- tailed one- sample t- tests (Figure 3B). In addi- tion, to test which specific subcortical regions were involved in these processes, we conducted post- hoc ROI analyses for subcortical regions after the whole- brain results were known (Figure 3—figure supplement 1B; one- tailed one- sample t tests; TTC- specific: caudate: t(33) = 5.95 , p = 5.6x10− , d = 1.02 , 5 , pfwe = 3.1x10− CI : [0.61, 1.45] , nucleus accumbens: t(33) = 4.41 , p = 5.2x10− , d = 0.76 , CI : [0.38, 1.15] , 7 , d = 1.21 , CI : [0.77, 1.67] , putamen: t(33) = 8.07 , globus pallidus: t(33) = 7.05 , 2.3x10− 9 , d = 1.38 , CI : [0.92, 1.88] , amygdala: t(33) = 1.78 , p = 0.042 , pfwe = 0.255 , p = 1.3x10− 0.04, 0.66] , thalamus: t(33) = 2.61 , p = 0.007 , pfwe = 0.007 , d = 0.45 , CI : [0.09, 0.81] ; TTC- d = 0.30 , CI : [ − 0.46, 0.23] , nucleus accumbens: independent: caudate: t(33) = 8 , t(33) = 1.82 , p = 0.039 , pfwe = 0.235 , d = 0.31 , CI : [ , d = 1.06 , pfwe = 1.3x10− , d = 0.73 , CI : [0.35, 1.12] , thalamus: CI : [0.65, 1.50] , amygdala: t(33) = 4.25 , p = 8.3x10− , d = 0.69 , CI : [0.32, 1.08] ). The subcortical ROIs (Figure  2— t(33) = 4.05 , p = 1.5x10− figure supplement 1B) were based on the FreeSurfer parcellation as described in the section ‘Regions of interest definition and analysis’. 7 , d = 1.21 , CI : [0.77, 1.68] , putamen: t(33) = 6.21 , p = 2.6x10− 0.04, 0.66] , globus pallidus: t(33) = 7.06 , p = 2.2x10− 0.67 , p = 0.746 , pfwe = 1, d = 7 , pfwe = 3.4x10− 4 4 , pfwe = 8.9x10− 8 , pfwe = 1.4x10− 9 , pfwe = 7.7x10− 7 , pfwe = 1.6x10− 5 , pfwe = 4.9x10− 0.11 , CI : [ − − − − 4 4 6 6 Hippocampal functional connectivity as a function of TTC-independent updating To examine which brain regions whose activity co- fluctuated with the one of the hippocampus during TTC- independent updating, we again conducted a PPI analysis similar to the one described earlier. A spherical seed ROI with a radius of 4 mm was centered around the hippocampal group- level peak effect (x=-30, y=-24, z=-18) observed for the TTC- independent updating regressor described above. The GLM included a PPI regressor and two nuisance regressors accounting for task- related effects from our contrast of interest (Behavioral improvements vs. no behavioral improvements) as well as physiological correlations that could arise due to anatomical connections to the hippocampal seed region or shared subcortical input. On the group- level, we then tested the weights estimated for our PPI regressor of interest against zero using a t- test implemented in SPM12. This revealed areas whose activity co- fluctuated with the one of the hippocampus as a function TTC- independent updating (Figure 3—figure supplement 2A). Moreover, we conducted independent ROI analyses for subcortical regions as described in the section ‘Regions of interest definition and analysis’. Here, we tested the beta estimates obtained for the hippo- campal seed- based PPI regressor of interest (Figure 3—figure supplement 2B; one- tailed one- sample 0.16, 0.53] , putamen: t(33) = 2.79 , t tests: caudate: t(33) = 1.06, p = 0.149 , pfwe = 0.894, d = 0.18, CI : [ p = 0.004 , pfwe = 0.026 , d = 0.48 , CI : [0.12, 0.84] globus pallidus: t(33) = 2.52 , p = 0.008 , pfwe = 0.050 , d = 0.43 , CI : [0.08, 0.79] , amygdala: t(33) = 2.60, p = 0.007, pfwe = 0.041, d = 0.45, CI : [0.09, 0.81] , nucleus − Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 16 of 22 Neuroscience Research article accumbens: t(33) = − pfwe = 0.032, d = 0.46, CI : [0.11, 0.83] ). 1.14, p = 0.869 , pfwe = 1, d = − 0.20, CI : [ − 0.54, 0.15] , thalamus: t(33) = 2.71, p = 0.005 , Brain activity as a function of behavioral performance and as a function of the behavioral regression effect To examine the neural underpinnings governing specificity and regularization in timing behavior in detail, we analyzed the trial- wise activity of each voxel as a function of performance in the TTC task (i.e. the absolute difference between estimated and true TTC in each trial) and as a function of the regression effect in behavior (i.e. the absolute difference between the estimated TTC and the mean of the sampled intervals, which was 0.82 s). To avoid effects of potential co- linearity between these regressors, we estimated model weights using two independent GLMs, which modeled the time course of each trial with either one of the two regressors. In addition, we again accounted for nuisance variance as described before, and all regressors except the realignment parameters and the constant term were convolved with the canonical HRF of SPM12. After fitting the model, we used the weights estimated for the two regressors to perform voxel- wise F- tests using SPM12, revealing activity that was correlated with these two regressors independent of the sign of the correlation (Figure 4A). In addition, we again performed ROI analyses using two- tailed one- sample t- tests for the anterior and posterior hippocampus (Figure 2—figure supplement 1A, Figure 4B). Eye tracking: Fixation quality does not affect the interpretation of our results We used an MR- compatible infrared eye tracker with long- range optics (Eyelink 1000) to monitor gaze position at a rate of 500 hz during the experiment. After blink removal, the eye tracking data was linearly detrended, median centered, downsampled to the screen refresh rate of 120 hz and smoothed with a running- average kernel of 100ms. Wilcoxon signed- rank tests for paired samples were used in order to test for potential biases in fixation error across speeds (Figure 1—figure supplement 2A) or across feedback levels (Figure 1—figure supplement 2B). Moreover, we tested if differences in fixation error could either explain individual differences in the regression effect, or individual differ- ences in absolute TTC error in behavior using Spearman’s rank- order correlations (Figure 1—figure supplement 2C). Acknowledgements We thank Raymundo Machado de Azevedo Neto for helpful comments on an earlier version of this manuscript. CFD’s research is supported by the Max Planck Society, the Kavli Foundation, the Jebsen foundation, the Centre of Excellence scheme of the Research Council of Norway – Centre for Neural Computation (223262 /F50), The Egil and Pauline Braathen and Fred Kavli Centre for Cortical Micro- circuits and the National Infrastructure scheme of the Research Council of Norway – NORBRAIN (197467 /F50). MN’s research is supported by a Feodor- Lynen Research Fellowship of the Alexander von Humboldt Foundation. RK’s research is supported by a CIDEGENT grant (CIDEGENT/2021/027) from the Valencian Community’s program for the support of talented researchers and the Ministerio de Ciencia, Innovación y Universidades, which is part of the Agencia Estatal de Investigación (AEI), through the project PID2021- 12233NA- 100. Additional information Competing interests Virginie van Wassenhove: Reviewing editor, eLife. The other authors declare that no competing inter- ests exist. Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 17 of 22 Neuroscience Research article Funding Funder Grant reference number Author European Research Council ERC-CoG GEOCOG 724836 Christian F Doeller Max Planck Society Kavli Foundation Kristian Gerhard Jebsen Foundation Christian F Doeller Christian F Doeller Christian F Doeller Norges Forskningsråd 223262/F50 Christian F Doeller Egil and Pauline Braathen and Fred Kavli Centre for Cortical Microcircuits Christian F Doeller Norges Forskningsråd NORBRAIN 197467/F50 Christian F Doeller Alexander von Humboldt Foundation Feodor-Lynen Research Fellowship Matthias Nau Generalitat Valenciana CIDEGENT/2021/027 Raphael Kaplan Ministerio de Ciencia, Innovación y Universidades Commissariat à l'Énergie Atomique et aux Énergies Alternatives Institut National de la Santé et de la Recherche Médicale PID2021-122338NA-I00 Raphael Kaplan Virginie van Wassenhove Virginie van Wassenhove The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Author contributions Ignacio Polti, Conceptualization, Data curation, Software, Formal analysis, Investigation, Visualization, Methodology, Writing – original draft, Project administration, Writing – review and editing; Matthias Nau, Conceptualization, Data curation, Supervision, Visualization, Methodology, Writing – original draft, Project administration, Writing – review and editing; Raphael Kaplan, Supervision, Project administration, Writing – review and editing; Virginie van Wassenhove, Supervision, Writing – review and editing; Christian F Doeller, Conceptualization, Supervision, Funding acquisition, Project adminis- tration, Writing – review and editing Author ORCIDs Ignacio Polti Matthias Nau Raphael Kaplan Virginie van Wassenhove Christian F Doeller http://orcid.org/0000-0002-6631-4315 http://orcid.org/0000-0003-0956-7815 http://orcid.org/0000-0002-5023-1566 http://orcid.org/0000-0002-2569-5502 http://orcid.org/0000-0003-4120-4600 Ethics The study was approved by the regional committee for medical and health research ethics (project number 2017/969) in Norway and participants gave written consent prior to scanning in accordance with the declaration of Helsinki (World Medical Association, 2013). Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.79027.sa1 Author response https://doi.org/10.7554/eLife.79027.sa2 Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 18 of 22 Neuroscience Research article Additional files Supplementary files • MDAR checklist Data availability Source data and analysis code are available at the following Open Science Framework repository: https://osf.io/cs8d6/. Pre- processed eye- tracker data can be found here: https://osf.io/mrhk9/. Raw fMRI data is available at the following G- Node Infrastructure repository: https://gin.g-node.org/ipolti/ TTC_HPCF.git. The following datasets were generated: Author(s) Polti I, Nau M, Kaplan R, Wassenhove van, Doeller CF Year 2022 Dataset title Dataset URL Database and Identifier Time- To- Contact https:// gin. g- node. org/ ipolti/ TTC_ HPCF. git G- Node Infrastructure, 10.12751/g- node.pwn4qz Frey M, Nau M, Doeller CF 2021 DeepMReye https:// osf. io/ mrhk9/ Open Science Framework, 10.17605/OSF.IO/MRHK9 https:// osf. io/ cs8d6/ Open Science Framework, cs8d6 Polti I, Nau M 2022 Rapid encoding of task regularities in the human hippocampus guides sensorimotor timing References Acerbi L, Wolpert DM, Vijayakumar S. 2012. Internal representations of temporal statistics and feedback calibrate motor- sensory interval timing. PLOS Computational Biology 8:e1002771. DOI: https://doi.org/10. 1371/journal.pcbi.1002771, PMID: 23209386 Aly M, Turk- Browne NB. 2017. How hippocampal memory shapes, and is shaped by, attention. Hannula DE, Duff MC (Eds). The Hippocampus from Cells to Systems: Structure, Connectivity, and Functional Contributions to Memory and Flexible Cognition. Cham: Springer International Publishing. p. 369–403. DOI: https://doi.org/10. 1007/978-3-319-50406-3_12 Barnett AJ, O’Neil EB, Watson HC, Lee ACH. 2014. The human hippocampus is sensitive to the durations of events and intervals within a sequence. Neuropsychologia 64:1–12. DOI: https://doi.org/10.1016/j. neuropsychologia.2014.09.011, PMID: 25223466 Behrens TEJ, Muller TH, Whittington JCR, Mark S, Baram AB, Stachenfeld KL, Kurth- Nelson Z. 2018. What is a cognitive map? Organizing knowledge for flexible behavior. Neuron 100:490–509. DOI: https://doi.org/10. 1016/j.neuron.2018.10.002, PMID: 30359611 Bellmund JLS, Gärdenfors P, Moser EI, Doeller CF. 2018. Navigating cognition: spatial codes for human thinking. Science 362:eaat6766. DOI: https://doi.org/10.1126/science.aat6766, PMID: 30409861 Bellmund JLS, Polti I, Doeller CF. 2020. Sequence memory in the hippocampal- entorhinal region. Journal of Cognitive Neuroscience 32:2056–2070. DOI: https://doi.org/10.1162/jocn_a_01592, PMID: 32530378 Bellmund JLS, Deuker L, Montijn ND, Doeller CF. 2022. Mnemonic construction and representation of temporal structure in the hippocampal formation. Nature Communications 13:3395. DOI: https://doi.org/10.1038/ s41467-022-30984-3, PMID: 35739096 Bicanski A, Burgess N. 2020. Neuronal vector coding in spatial cognition. Nature Reviews. Neuroscience 21:453–470. DOI: https://doi.org/10.1038/s41583-020-0336-9, PMID: 32764728 Bostan AC, Strick PL. 2018. The basal ganglia and the cerebellum: nodes in an integrated network. Nature Reviews. Neuroscience 19:338–350. DOI: https://doi.org/10.1038/s41583-018-0002-7, PMID: 29643480 Burgess N, Maguire EA, O’Keefe J. 2002. The human hippocampus and spatial and episodic memory. Neuron 35:625–641. DOI: https://doi.org/10.1016/s0896-6273(02)00830-9, PMID: 12194864 Chang CJ, Jazayeri M. 2018. Integration of speed and time for estimating time to contact. PNAS 115:E2879– E2887. DOI: https://doi.org/10.1073/pnas.1713316115 Cheng DT, Disterhoft JF, Power JM, Ellis DA, Desmond JE. 2008. Neural substrates underlying human delay and trace eyeblink conditioning. PNAS 105:8108–8113. DOI: https://doi.org/10.1073/pnas.0800374105, PMID: 18523017 Chersi F, Burgess N. 2015. The cognitive architecture of spatial navigation: hippocampal and striatal contributions. Neuron 88:64–77. DOI: https://doi.org/10.1016/j.neuron.2015.09.021 Cicchini GM, Arrighi R, Cecchetti L, Giusti M, Burr DC. 2012. Optimal encoding of interval timing in expert percussionists. The Journal of Neuroscience 32:1056–1060. DOI: https://doi.org/10.1523/JNEUROSCI.3411-11. 2012, PMID: 22262903 Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 19 of 22 Neuroscience Research article Cohen MX, Ranganath C. 2007. Reinforcement learning signals predict future decisions. The Journal of Neuroscience 27:371–378. DOI: https://doi.org/10.1523/JNEUROSCI.4421-06.2007, PMID: 17215398 Cox J, Witten IB. 2019. Striatal circuits for reward learning and decision- making. Nature Reviews. Neuroscience 20:482–494. DOI: https://doi.org/10.1038/s41583-019-0189-2, PMID: 31171839 Daw ND, Dayan P. 2014. The algorithmic anatomy of model- based evaluation. Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences 369:1655. DOI: https://doi.org/10.1098/rstb.2013.0478, PMID: 25267820 de Azevedo Neto RM, Amaro Júnior E. 2018. Bilateral dorsal fronto- parietal areas are associated with integration of visual motion information and timed motor action. Behavioural Brain Research 337:91–98. DOI: https://doi.org/10.1016/j.bbr.2017.09.046, PMID: 28964911 Deuker L, Bellmund JL, Navarro Schröder T, Doeller CF. 2016. An event map of memory space in the hippocampus. eLife 5:e16534. DOI: https://doi.org/10.7554/eLife.16534, PMID: 27710766 Dickerson KC, Delgado MR. 2015. Contributions of the hippocampus to feedback learning. Cognitive, Affective & Behavioral Neuroscience 15:861–877. DOI: https://doi.org/10.3758/s13415-015-0364-5, PMID: 26055632 Doeller CF, Opitz B, Krick CM, Mecklinger A, Reith W. 2005. Prefrontal- hippocampal dynamics involved in learning regularities across episodes. Cerebral Cortex 15:1123–1133. DOI: https://doi.org/10.1093/cercor/ bhh211, PMID: 15563722 Doeller CF, King JA, Burgess N. 2008. Parallel striatal and hippocampal systems for landmarks and boundaries in spatial memory. PNAS 105:5915–5920. DOI: https://doi.org/10.1073/pnas.0801489105, PMID: 18408152 Eichenbaum H. 2014. Time cells in the hippocampus: a new dimension for mapping memories. Nature Reviews. Neuroscience 15:732–744. DOI: https://doi.org/10.1038/nrn3827, PMID: 25269553 Eichenbaum H. 2017. On the integration of space, time, and memory. Neuron 95:1007–1018. DOI: https://doi. org/10.1016/j.neuron.2017.06.036, PMID: 28858612 Epstein RA, Baker CI. 2019. Scene perception in the human brain. Annual Review of Vision Science 5:373–397. DOI: https://doi.org/10.1146/annurev-vision-091718-014809, PMID: 31226012 Foerde K, Shohamy D. 2011. Feedback timing modulates brain systems for learning in humans. The Journal of Neuroscience 31:13157–13167. DOI: https://doi.org/10.1523/JNEUROSCI.2701-11.2011, PMID: 21917799 Friston K, FitzGerald T, Rigoli F, Schwartenbeck P, Pezzulo G. 2016. Active inference: a process theory. Neural Computation 29:1–49. DOI: https://doi.org/10.1162/NECO_a_00912, PMID: 27870614 Gahnstrom CJ, Spiers HJ. 2020. Striatal and hippocampal contributions to flexible navigation in rats and humans. Brain and Neuroscience Advances 4:2398212820979772. DOI: https://doi.org/10.1177/ 2398212820979772, PMID: 33426302 Gauthier B, Pestke K, van Wassenhove V. 2019. Building the arrow of time… over time: a sequence of brain activity mapping imagined events in time and space. Cerebral Cortex 29:4398–4414. DOI: https://doi.org/10. 1093/cercor/bhy320, PMID: 30566689 Gauthier B, Prabhu P, Kotegar KA, van Wassenhove V. 2020. Hippocampal contribution to ordinal psychological time in the human brain. Journal of Cognitive Neuroscience 32:2071–2086. DOI: https://doi.org/10.1162/jocn_ a_01586, PMID: 32459130 Geerts JP, Chersi F, Stachenfeld KL, Burgess N. 2020. A general model of hippocampal and dorsal striatal learning and decision making. PNAS 117:31427–31437. DOI: https://doi.org/10.1073/pnas.2007981117, PMID: 33229541 Gershman SJ, Moustafa AA, Ludvig EA. 2014. Time representation in reinforcement learning models of the basal ganglia. Frontiers in Computational Neuroscience 7:194. DOI: https://doi.org/10.3389/fncom.2013.00194, PMID: 24409138 Gibbon J. 1977. Scalar expectancy theory and Weber’s law in animal timing. Psychological Review 84:279–325. DOI: https://doi.org/10.1037/0033-295X.84.3.279 Goodroe SC, Starnes J, Brown TI. 2018. The complex nature of hippocampal- striatal interactions in spatial navigation. Frontiers in Human Neuroscience 12:250. DOI: https://doi.org/10.3389/fnhum.2018.00250, PMID: 29977198 Gouvêa TS, Monteiro T, Motiwala A, Soares S, Machens C, Paton JJ. 2015. Striatal dynamics explain duration judgments. eLife 4:e11386. DOI: https://doi.org/10.7554/eLife.11386, PMID: 26641377 Howard MW. 2017. Temporal and spatial context in the mind and brain. Current Opinion in Behavioral Sciences 17:14–19. DOI: https://doi.org/10.1016/j.cobeha.2017.05.022, PMID: 28845441 Huang Y, Rao RPN. 2011. Predictive coding. Wiley Interdisciplinary Reviews. Cognitive Science 2:580–593. DOI: https://doi.org/10.1002/wcs.142, PMID: 26302308 Jazayeri M, Shadlen MN. 2010. Temporal context calibrates interval timing. Nature Neuroscience 13:1020–1026. DOI: https://doi.org/10.1038/nn.2590, PMID: 20581842 Julian J, Doeller CF. 2020. Context in spatial and episodic memory. Poeppel D, Mangun GR, Gazzaniga MS (Eds). The Cognitive Neurosciences. The MIT Press. p. 219–234. Kaplan R, Schuck NW, Doeller CF. 2017. The role of mental maps in decision- making. Trends in Neurosciences 40:256–259. DOI: https://doi.org/10.1016/j.tins.2017.03.002, PMID: 28365032 Kragel JE, Schuele S, VanHaerents S, Rosenow JM, Voss JL. 2021. Rapid coordination of effective learning by the human hippocampus. Science Advances 7:25. DOI: https://doi.org/10.1126/sciadv.abf7144, PMID: 34144985 Kumaran D. 2012. What representations and computations underpin the contribution of the hippocampus to generalization and inference? Frontiers in Human Neuroscience 6:157. DOI: https://doi.org/10.3389/fnhum. 2012.00157, PMID: 22675298 Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 20 of 22 Neuroscience Research article Lee D, Seo H, Jung MW. 2012. Neural basis of reinforcement learning and decision making. Annual Review of Neuroscience 35:287–308. DOI: https://doi.org/10.1146/annurev-neuro-062111-150512, PMID: 22462543 LeGates TA, Kvarta MD, Tooley JR, Francis TC, Lobo MK, Creed MC, Thompson SM. 2018. Reward behaviour is regulated by the strength of hippocampus- nucleus accumbens synapses. Nature 564:258–262. DOI: https:// doi.org/10.1038/s41586-018-0740-8, PMID: 30478293 MacDonald CJ, Lepage KQ, Eden UT, Eichenbaum H. 2011. Hippocampal “ time cells ” bridge the gap in memory for discontiguous events. Neuron 71:737–749. DOI: https://doi.org/10.1016/j.neuron.2011.07.012, PMID: 21867888 Meck WH, Church RM, Olton DS. 1984. Hippocampus, time, and memory. Behavioral Neuroscience 98:3–22. DOI: https://doi.org/10.1037//0735-7044.98.1.3, PMID: 6696797 Mello GBM, Soares S, Paton JJ. 2015. A scalable population code for time in the striatum. Current Biology 25:1113–1122. DOI: https://doi.org/10.1016/j.cub.2015.02.036, PMID: 25913405 Momennejad I. 2020. Learning structures: predictive representations, replay, and generalization. Current Opinion in Behavioral Sciences 32:155–166. DOI: https://doi.org/10.1016/j.cobeha.2020.02.017, PMID: 35419465 Montchal ME, Reagh ZM, Yassa MA. 2019. Precise temporal memories are supported by the lateral entorhinal cortex in humans. Nature Neuroscience 22:284–288. DOI: https://doi.org/10.1038/s41593-018-0303-1, PMID: 30643291 Nau M, Navarro Schröder T, Bellmund JLS, Doeller CF. 2018a. Hexadirectional coding of visual space in human entorhinal cortex. Nature Neuroscience 21:188–190. DOI: https://doi.org/10.1038/s41593-017-0050-8, PMID: 29311746 Nau M, Julian JB, Doeller CF. 2018b. How the brain’s navigation system shapes our visual experience. Trends in Cognitive Sciences 22:810–825. DOI: https://doi.org/10.1016/j.tics.2018.06.008, PMID: 30031670 Nobre AC, van Ede F. 2018. Anticipated moments: temporal structure in attention. Nature Reviews. Neuroscience 19:34–48. DOI: https://doi.org/10.1038/nrn.2017.141, PMID: 29213134 O’Reilly JX, Mesulam MM, Nobre AC. 2008. The cerebellum predicts the timing of perceptual events. The Journal of Neuroscience 28:2252–2260. DOI: https://doi.org/10.1523/JNEUROSCI.2742-07.2008, PMID: 18305258 Palombo DJ, Verfaellie M. 2017. Hippocampal contributions to memory for time: evidence from neuropsychological studies. Current Opinion in Behavioral Sciences 17:107–113. DOI: https://doi.org/10.1016/ j.cobeha.2017.07.015 Paton JJ, Buonomano DV. 2018. The neural basis of timing: distributed mechanisms for diverse functions. Neuron 98:687–705. DOI: https://doi.org/10.1016/j.neuron.2018.03.045, PMID: 29772201 Peer M, Brunec IK, Newcombe NS, Epstein RA. 2021. Structuring knowledge with cognitive maps and cognitive graphs. Trends in Cognitive Sciences 25:37–54. DOI: https://doi.org/10.1016/j.tics.2020.10.004, PMID: 33248898 Petter EA, Gershman SJ, Meck WH. 2018. Integrating models of interval timing and reinforcement learning. Trends in Cognitive Sciences 22:911–922. DOI: https://doi.org/10.1016/j.tics.2018.08.004, PMID: 30266150 Petzschner FH, Glasauer S. 2011. Iterative Bayesian estimation as an explanation for range and regression effects: a study on human path integration. The Journal of Neuroscience 31:17220–17229. DOI: https://doi. org/10.1523/JNEUROSCI.2028-11.2011, PMID: 22114288 Petzschner FH, Glasauer S, Stephan KE. 2015. A Bayesian perspective on magnitude estimation. Trends in Cognitive Sciences 19:285–293. DOI: https://doi.org/10.1016/j.tics.2015.03.002, PMID: 25843543 Poppenk J, Evensmoen HR, Moscovitch M, Nadel L. 2013. Long- axis specialization of the human hippocampus. Trends in Cognitive Sciences 17:230–240. DOI: https://doi.org/10.1016/j.tics.2013.03.005, PMID: 23597720 Rakitin BC, Gibbon J, Penney TB, Malapani C, Hinton SC, Meck WH. 1998. Scalar expectancy theory and peak- interval timing in humans. Journal of Experimental Psychology. Animal Behavior Processes 24:15–33. DOI: https://doi.org/10.1037//0097-7403.24.1.15, PMID: 9438963 Richards W. 1973. Time reproductions by H.M. Acta Psychologica 37:279–282. DOI: https://doi.org/10.1016/ 0001-6918(73)90020-6, PMID: 4743299 Roach NW, McGraw PV, Whitaker DJ, Heron J. 2017. Generalization of prior information for rapid bayesian time estimation. PNAS 114:412–417. DOI: https://doi.org/10.1073/pnas.1610706114, PMID: 28007982 Schapiro AC, Turk- Browne NB, Botvinick MM, Norman KA. 2017. Complementary learning systems within the hippocampus: a neural network modelling approach to reconciling episodic memory with statistical learning. Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences 372:1711. DOI: https:// doi.org/10.1098/rstb.2016.0049, PMID: 27872368 Schiller D, Eichenbaum H, Buffalo EA, Davachi L, Foster DJ, Leutgeb S, Ranganath C. 2015. Memory and space: towards an understanding of the cognitive map. The Journal of Neuroscience 35:13904–13911. DOI: https:// doi.org/10.1523/JNEUROSCI.2618-15.2015, PMID: 26468191 Schlichting ML, Preston AR. 2015. Memory integration: neural mechanisms and implications for behavior. Current Opinion in Behavioral Sciences 1:1–8. DOI: https://doi.org/10.1016/j.cobeha.2014.07.005, PMID: 25750931 Schönberg T, Daw ND, Joel D, O’Doherty JP. 2007. Reinforcement learning signals in the human striatum distinguish learners from nonlearners during reward- based decision making. The Journal of Neuroscience 27:12860–12867. DOI: https://doi.org/10.1523/JNEUROSCI.2496-07.2007, PMID: 18032658 Schuck NW, Niv Y. 2019. Sequential replay of nonspatial task states in the human hippocampus. Science 364:eaaw5181. DOI: https://doi.org/10.1126/science.aaw5181, PMID: 31249030 Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 21 of 22 Neuroscience Research article Shikano Y, Ikegaya Y, Sasaki T. 2021. Minute- encoding neurons in hippocampal- striatal circuits. Current Biology 31:1438–1449.. DOI: https://doi.org/10.1016/j.cub.2021.01.032, PMID: 33545048 Shimbo A, Izawa EI, Fujisawa S. 2021. Scalable representation of time in the hippocampus. Science Advances 7:eabd7013. DOI: https://doi.org/10.1126/sciadv.abd7013, PMID: 33536211 Shohamy D, Wagner AD. 2008. Integrating memories in the human brain: hippocampal- midbrain encoding of overlapping events. Neuron 60:378–389. DOI: https://doi.org/10.1016/j.neuron.2008.09.023, PMID: 18957228 Strange BA, Witter MP, Lein ES, Moser EI. 2014. Functional organization of the hippocampal longitudinal axis. Nature Reviews. Neuroscience 15:655–669. DOI: https://doi.org/10.1038/nrn3785, PMID: 25234264 Thavabalasingam S, O’Neil EB, Lee ACH. 2018. Multivoxel pattern similarity suggests the integration of temporal duration in hippocampal event sequence representations. NeuroImage 178:136–146. DOI: https:// doi.org/10.1016/j.neuroimage.2018.05.036, PMID: 29775662 Thavabalasingam S, O’Neil EB, Tay J, Nestor A, Lee ACH. 2019. Evidence for the incorporation of temporal duration information in human hippocampal long- term memory sequence representations. PNAS 116:6407– 6414. DOI: https://doi.org/10.1073/pnas.1819993116, PMID: 30862732 Umbach G, Kantak P, Jacobs J, Kahana M, Pfeiffer BE, Sperling M, Lega B. 2020. Time cells in the human hippocampus and entorhinal cortex support episodic memory. PNAS 117:28463–28474. DOI: https://doi.org/ 10.1073/pnas.2013250117, PMID: 33109718 Vikbladh OM, Meager MR, King J, Blackmon K, Devinsky O, Shohamy D, Burgess N, Daw ND. 2019. Hippocampal contributions to model- based planning and spatial memory. Neuron 102:683–693.. DOI: https:// doi.org/10.1016/j.neuron.2019.02.014, PMID: 30871859 Wang J, Narain D, Hosseini EA, Jazayeri M. 2018. Flexible timing by temporal scaling of cortical responses. Nature Neuroscience 21:102–110. DOI: https://doi.org/10.1038/s41593-017-0028-6, PMID: 29203897 Whittington JCR, Muller TH, Mark S, Chen G, Barry C, Burgess N, Behrens TEJ. 2020. The Tolman- Eichenbaum machine: unifying space and relational memory through generalization in the hippocampal formation. Cell 183:1249–1263.. DOI: https://doi.org/10.1016/j.cell.2020.10.024, PMID: 33181068 Wiener M, Michaelis K, Thompson JC. 2016. Functional correlates of likelihood and prior representations in a virtual distance task. Human Brain Mapping 37:3172–3187. DOI: https://doi.org/10.1002/hbm.23232, PMID: 27167875 Wikenheiser AM, Marrero- Garcia Y, Schoenbaum G. 2017. Suppression of ventral hippocampal output impairs integrated orbitofrontal encoding of task structure. Neuron 95:1197–1207.. DOI: https://doi.org/10.1016/j. neuron.2017.08.003, PMID: 28823726 Wimmer GE, Daw ND, Shohamy D. 2012. Generalization of value in reinforcement learning by humans. The European Journal of Neuroscience 35:1092–1104. DOI: https://doi.org/10.1111/j.1460-9568.2012.08017.x, PMID: 22487039 Wirth S, Avsar E, Chiu CC, Sharma V, Smith AC, Brown E, Suzuki WA. 2009. Trial outcome and associative learning signals in the monkey hippocampus. Neuron 61:930–940. DOI: https://doi.org/10.1016/j.neuron.2009. 01.012, PMID: 19324001 Wittmann BC, Schott BH, Guderian S, Frey JU, Heinze HJ, Düzel E. 2005. Reward- Related fMRI activation of dopaminergic midbrain is associated with enhanced hippocampus- dependent long- term memory formation. Neuron 45:459–467. DOI: https://doi.org/10.1016/j.neuron.2005.01.010, PMID: 15694331 Wolpert DM, Diedrichsen J, Flanagan JR. 2011. Principles of sensorimotor learning. Nature Reviews. Neuroscience 12:739–751. DOI: https://doi.org/10.1038/nrn3112, PMID: 22033537 World Medical Association. 2013. World medical association declaration of helsinki: ethical principles for medical research involving human subjects. JAMA 310:2191–2194. DOI: https://doi.org/10.1001/jama.2013. 281053, PMID: 24141714 Yin B, Troger AB. 2011. Exploring the 4th dimension: hippocampus, time, and memory revisited. Frontiers in Integrative Neuroscience 5:36. DOI: https://doi.org/10.3389/fnint.2011.00036, PMID: 21886612 Polti, Nau et al. eLife 2022;11:e79027. DOI: https:// doi. org/ 10. 7554/ eLife. 79027 22 of 22 Neuroscience
10.7554_elife.85145
ReSeaRCH aRtICLe Statistical inference reveals the role of length, GC content, and local sequence in V(D)J nucleotide trimming Magdalena L Russell1,2*, Noah Simon3, Philip Bradley1,4*†, Frederick A Matsen IV1,5,6,7*† 1Computational Biology Program, Fred Hutchinson Cancer Center, Seattle, United States; 2Molecular and Cellular Biology Program, University of Washington, Seattle, United States; 3Department of Biostatistics, University of Washington, Seattle, United States; 4Institute for Protein Design, Department of Biochemistry, University of Washington, Seattle, United States; 5Department of Genome Sciences, University of Washington, Seattle, United States; 6Department of Statistics, University of Washington, Seattle, United States; 7Howard Hughes Medical Institute, Seattle, United States Abstract To appropriately defend against a wide array of pathogens, humans somatically generate highly diverse repertoires of B cell and T cell receptors (BCRs and TCRs) through a random process called V(D)J recombination. Receptor diversity is achieved during this process through both the combinatorial assembly of V(D)J- genes and the junctional deletion and inser- tion of nucleotides. While the Artemis protein is often regarded as the main nuclease involved in V(D)J recombination, the exact mechanism of nucleotide trimming is not understood. Using a previously published TCRβ repertoire sequencing data set, we have designed a flexible proba- bilistic model of nucleotide trimming that allows us to explore various mechanistically interpre- table sequence- level features. We show that local sequence context, length, and GC nucleotide content in both directions of the wider sequence, together, can most accurately predict the trimming probabilities of a given V- gene sequence. Because GC nucleotide content is predictive of sequence- breathing, this model provides quantitative statistical evidence regarding the extent to which double- stranded DNA may need to be able to breathe for trimming to occur. We also see evidence of a sequence motif that appears to get preferentially trimmed, independent of GC- content- related effects. Further, we find that the inferred coefficients from this model provide accurate prediction for V- and J- gene sequences from other adaptive immune receptor loci. These results refine our understanding of how the Artemis nuclease may function to trim nucle- otides during V(D)J recombination and provide another step toward understanding how V(D)J recombination generates diverse receptors and supports a powerful, unique immune response in healthy humans. Editor's evaluation Russell et al. study and reveal compelling evidence for potential sequence- based factors that may drive VDJ trimming, a mechanism involved in VDJ recombination that shapes adaptive immune repertoire generation. The work is based on a rigorous statistical comparison of logistic regression models to reveal the role and function of cutting enzymes in shaping T- and B- cell receptor diversity which could provide fundamental new insights into these processes. *For correspondence: magruss@uw.edu (MLR); pbradley@fredhutch.org (PB); matsen@fredhutch.org (FaM) †These authors contributed equally to this work Competing interest: The authors declare that no competing interests exist. Funding: See page 23 Received: 24 November 2022 Preprinted: 12 December 2022 Accepted: 11 April 2023 Published: 25 May 2023 Reviewing Editor: Frederik Graw, Friedrich- Alexander- University Erlangen- Nürnberg, Germany Copyright Russell et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 1 of 40 Research article Introduction Cells throughout the body regularly present protein fragments, known as antigens, on cell- surface molecules called major histocompatibility complex (MHC). Receptors on the surface of T cells can bind to these MHC- bound antigens, recognize them, and, if necessary, initiate an immune response. For an individual to be capable of defending against a wide array of potential foreign pathogens, they somatically generate a massive diversity of T cell receptors (TCRs) through a random process called V(D)J recombination. After generation, TCRs undergo a selection process to ensure proper expres- sion, MHC recognition, and limited autoreactivity. The collection of TCRs in an individual comprises their TCR repertoire. The majority of human T cells express α-β receptors that consist of an α and a β protein chain. During the V(D)J recombination process of the β chain, a single V-, D-, and J- gene are randomly chosen from a pool of V- gene, D- gene, and J- gene segments within the germline TCRβ locus over a series of steps. To begin this process, the recombination activating gene protein complex aligns a randomly chosen D- and J- gene, removes the intervening chromosomal DNA between the two genes, and forms a hairpin loop at the end of each gene (Gellert, 1994; Fugmann et al., 2000; Schatz and Swanson, 2011). Each hairpin loop is then nicked open, typically in an asymmetrical fashion, by the Artemis:DNA- PKcs protein complex (Ma et  al., 2002; Lu et  al., 2007). This asymmetrical hairpin opening creates a single- stranded DNA overhang at the end of both genes that, depending on the location of the hairpin nick, may contain P- nucleotides (short palindromes of gene terminal sequence) (Gauss and Lieber, 1996; Nadel and Feeney, 1997; Ma et al., 2002; Jackson et al., 2004; Lu et al., 2007). The most dominant hairpin opening position leads to a single- stranded 3’ overhang that is 4 nucleotides in length (2 nucleotides of which are P- nucleotides) (Lu et al., 2007). From here, nucleo- tides may be deleted from each gene end through an incompletely understood mechanism suggested to involve Artemis (Feeney et al., 1994; Nadel and Feeney, 1995; Nadel and Feeney, 1997; Jackson et al., 2004; Gu et al., 2010; Chang et al., 2015; Chang and Lieber, 2016; Zhao et al., 2020; Russell et al., 2022b). This nucleotide trimming can remove traces of P- nucleotides (Gauss and Lieber, 1996; Srivastava and Robins, 2012). Non- template- encoded nucleotides, known as N- insertions, can also be added to each gene end by the enzyme terminal deoxynucleotidyl transferase (Kallenbach et al., 1992; Gilfillan et al., 1993; Komori et al., 1993). Once the nucleotide addition and deletion steps are completed, the gene segments are paired and ligated together (Zhao et al., 2020). From here, the process is repeated between a random V- gene and this combined D- J junction to complete the TCRβ chain. A similar TCR chain recombination then proceeds, though without a D- gene, to complete the α-β TCR. Other adaptive immune receptor loci, such as TRG, TRD, and all IG loci, also undergo V(D)J recombination during the development of γ-δ T cells and B cells, respectively. Junctional diversity created by the deletion and non- templated insertion of nucleotides during V(D) J recombination contributes substantially to the resulting diversity of the TCR repertoire. Small varia- tions in gene sequence have been shown to lead to large changes in the extent of nucleotide deletion (Nadel and Feeney, 1995; Gauss and Lieber, 1996; Nadel and Feeney, 1997; Jackson et al., 2004). For example, sequences with high AT content suffer greater nucleotide loss than sequences with high GC content (Gauss and Lieber, 1996). These findings are suggestive of a nuclease that either binds an AT- rich sequence motif or requires an AT- specific structure (e.g. a sequence that breathes, Tsai et al., 2009), however, further work is required to quantify this mechanistic preference. The Artemis protein is often regarded as the main nuclease involved in V(D)J recombination (Chang et  al., 2015; Chang and Lieber, 2016; Zhao et  al., 2020). Artemis is a member of the metallo- β - lactamase family of nucleases (Moshous et al., 2001) and is widely regarded as a structure- specific nuclease as opposed to a nuclease that binds specific DNA sequences (Ma et  al., 2005; Chang et  al., 2015; Chang and Lieber, 2016; Yosaatmadja et  al., 2021). Members of this family are characterized by their conserved metallo-β-lactamase and β-CASP domains and their ability to nick DNA or RNA in various configurations (Dominski, 2007; Pettinati et al., 2016). Alone, Artemis possesses intrinsic 5’-to- 3’ exonuclease activity on single- stranded DNA (Li et al., 2014). On double- stranded DNA, Artemis, in complex with DNA- PKcs, has endonuclease activity on 5’ and 3’ DNA over- hangs and on DNA hairpins (Ma et al., 2002; Lu et al., 2007; Lu et al., 2008). It has been proposed that the Artemis:DNA- PKcs complex binds single- stranded- to- double- stranded DNA boundaries prior to nicking (Ma et  al., 2002; Ma et  al., 2005; Lu et  al., 2007; Chang and Lieber, 2016); for blunt DNA ends, previous work has concluded that sequence- breathing is required to achieve this Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 2 of 40 Computational and Systems Biology | Immunology and Inflammation Research article structural configuration prior to Artemis action (Chang et al., 2015). Further, Artemis, in complex with XRCC4- DNA ligase IV, has additional endonuclease activity on 3’ DNA overhangs and preferentially nicks one nucleotide at a time from the single- stranded 3’ end (Chang et al., 2016; Gerodimos et al., 2017). Despite these diverse nucleolytic functions, the extent of involvement and exact mechanism of action for the Artemis protein during the nucleotide trimming step of V(D)J recombination, and how it relates to observed sequence- dependent changes in trimming (Nadel and Feeney, 1995; Gauss and Lieber, 1996; Nadel and Feeney, 1997; Jackson et al., 2004), has yet to be fully understood. While molecular experiments using model organisms have been essential for establishing the current mechanistic understanding of the nucleotide trimming process, studies in humans have been limited. Statistical inference on high- throughput repertoire sequencing data sets allows for explora- tion of the in vivo V(D)J recombination mechanism outside of model organisms. In particular, analysis of trimming in high- throughput data sets should lead to insights about the natural underlying process, in the same way that analysis of large data sets has led to insight into the process of somatic hypermu- tation. There, researchers have found quite significant connections between local sequence identity and mutation patterns, leading to a rich literature (Rogozin and Kolchanov, 1992; Dunn- Walters et al., 1998; Cohen et al., 2011; Yaari et al., 2013; Elhanati et al., 2015; Wei et al., 2015; Cui et al., 2016; Feng et al., 2019; Spisak et al., 2020). In contrast, we are only aware of one statistical analysis connecting sequence identity to trimming lengths (Murugan et al., 2012). This one existing analysis (Murugan et al., 2012) has shown that a simple position- weight- matrix style (PWM) model does a surprisingly good job of predicting the distri- bution of trimming lengths for a variety of V- genes. However, while this trimming model has good model fit and predictive accuracy, it is limited by the assumption that the trimming mechanism relies solely on a sequence motif and, as such, is not designed in a way that allows us to explore alternative hypotheses. In this paper, we explore the sequence- level determinants of nucleotide trimming during V(D) J recombination using statistical inference on high- throughput TCRβ repertoire sequencing data (Emerson et al., 2017). With the goal of informing our mechanistic understanding in a quantitative way, we have designed a flexible probabilistic model of nucleotide trimming that allows us to explore various sequence- level features. Our results show that trimming probabilities are highest for DNA positions near the end of the sequence that contain high GC content upstream, quantifying the role of sequence- breathing dynamics in the trimming process. We also see evidence of a sequence motif that appears to get preferentially trimmed, independent of possible sequence- breathing effects. As such, we can predict trimming probabilities most accurately using a model that includes features for local sequence context, length, and GC nucleotide content in both directions of the wider sequence. We show that this model has high predictive accuracy for V- and J- gene sequences from an indepen- dent TCRβ-sequencing data set, and also extends well to TCRα, TCRγ, and IGH sequences. Further, we demonstrate that genetic variations within the gene encoding the Artemis protein that were previ- ously identified as being associated with increasing the extent of trimming (Russell et al., 2022b) are also associated with changes in several model coefficients. Results Training data description We worked with TCRβ-immunosequencing data representing 666 individuals (Emerson et al., 2017). V(D)J recombination scenarios were assigned to each sequence from each individual using the IGoR software which is designed to learn unbiased V(D)J recombination statistics from immune sequence reads (Marcou et al., 2018). Using these V(D)J recombination statistics, IGoR output a list of poten- tial recombination scenarios with their corresponding likelihoods for each TCRβ-chain sequence in the training data set. We annotated each sequence with a single V(D)J recombination scenario by sampling from these potential scenarios according to the posterior probability of each scenario (see Materials and methods for further details). Annotated TCR sequences can be separated into two categories: ‘productive’ rearrangements which code for a complete, full- length protein and ‘non- productive’ rearrangements which do not. Non- productive sequences are generated when the V(D)J recombination process produces a sequence that is either out- of- frame or contains a stop codon. Each T cell contains two loci which can Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 3 of 40 Computational and Systems Biology | Immunology and Inflammation Research article undergo the V(D)J recombination process. When the first recombination fails to generate a functional receptor (creating a non- productive sequence), followed by a successful rearrangement on the T cell’s second chromosome (a productive sequence), the non- productive rearrangement can be sequenced as part of the repertoire. Non- productive sequences do not generate proteins that undergo functional selection in the thymus, and their recombination statistics should reflect only the V(D)J recombination generation process (Robins et al., 2010; Murugan et al., 2012; Sethna et al., 2019). In contrast, the recombination statistics of productive sequences should reflect both V(D)J recombination generation and functional selection. Because we are interested in nucleotide trimming during the V(D)J recom- bination generation process, prior to selection, we only include non- productive sequences in our training data set. Further, because V- gene sequences within the TRB locus contain more sequence variation than D- and/or J- genes, we only include V- gene sequences in our training data set. Replicating a previous model of nucleotide trimming The extent of nucleotide trimming varies substantially from gene to gene (Nadel and Feeney, 1995; Nadel and Feeney, 1997; Jackson et al., 2004; Murugan et al., 2012). Previous work has identified an interesting impact of sequence features, such as sequence nucleotide context, on trimming prob- abilities using a PWM model (Murugan et al., 2012). To our knowledge, this is the only model that takes nucleotide sequence identity into account when predicting trimming probabilities. Specifically, this model leverages a ‘trimming motif’ containing 2 nucleotides 5’ of the trimming site and 4 nucleo- tides 3’ of the trimming site to predict the probability of trimming at a given site. It was designed and trained using sequencing data from just nine individuals (Murugan et al., 2012), and has surprisingly good model fit and predictive accuracy across many V- genes despite its simplicity. Using a different, and much larger, repertoire sequencing data set, we have trained this PWM model and replicated previous work (Figure 2—figure supplement 1). We will refer to this model as the 2×4 motif model. It is important to note that this PWM model is not the primary model described in Murugan et al., 2012, but again is the only one that relates nucleotide identity to trimming. Model set-up overview While the 2×4 motif model has good predictive accuracy and model fit (Murugan et al., 2012), it is limited by its assumption that the trimming mechanism relies solely on a sequence motif. Here, we have generalized this PWM model to a model that allows for arbitrary sequence features, and trained each new model using conditional logistic regression (see Materials and methods). With this set- up, we were able to evaluate the relative importance of new mechanistically interpretable features for predicting trimming probabilities. Specifically, we designed features to measure the effects of DNA- shape, length, and GC nucleotide content in both directions of the wider sequence on the probability of trimming at a given position in a gene sequence. We parameterize each of these features as follows. An example of how an arbitrary V- gene sequence is transformed into features for modeling is shown in Figure 1. To parameterize DNA- shape, we used previously developed methods (Zhou et al., 2013; Chiu et al., 2016) to estimate various DNA- shape values (i.e. roll, twist, electrostatic potential, minor groove width, etc.) for each single- nucleotide posi- tion within a sequence window surrounding the trimming site. To parameterize length, we measure the sequence- independent distance from the end of the gene (i.e. the number of nucleotides from the 3’-end of the sequence) as an integer- valued variable. We parameterize GC nucleotide content using the raw counts of AT and GC nucleotides on both sides of the trimming site (the two- side base- count). By using raw nucleotide counts, this measure also serves to parameterize length. Because AT nucleotides have a greater potential for sequence- breathing compared to GC nucleotides within a sequence (Jose et al., 2009), these two- side base- count terms may be serving as a proxy for the capacity of a sequence to breathe. As such, because sequence- breathing potential is only relevant for nucleotides that are paired, we do not include the nucleotides within the 3’ single- stranded- overhang when counting 3’ AT and GC nucleotides (see Appendix 2). With these features, we designed models containing various feature combinations (Figure  1B). Collectively, these models allow us to explore other possible sequence- level determinants of nucle- otide trimming, in addition to the previously proposed (Murugan et  al., 2012) “trimming motif” hypothesis. We trained each model using the V- gene training data set described above (see Materials and methods for further model training details), and evaluated performance using a suite of different Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 4 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Figure 1. Overview of how a sequence is transformed into features for regression. (A) As described, during the early stages of V(D)J recombination between two genes, the hairpin of each gene is opened; here, we are showing this hairpin- opening step for a single arbitrary V- gene. The most common hairpin- opening position leads to a 4- nucleotide- long single- stranded overhang (2 nucleotides of which are considered P- nucleotides, as shown in purple). From here, each gene can undergo nucleotide trimming. In this example, the V- gene is trimmed back 6 nucleotides. (B) All models were trained with non- productive V- gene sequences whose trimming positions were inferred during a sequence annotation step. For our model parameterization, we only consider the top strand (5’-to- 3’) of the observed sequence. Here, the sequence features parameterized for each model type are shown for the example sequence from (A). The pink boxes surround nucleotides included in the matrix representation of motif features. The turquoise boxes surround nucleotides used to estimate and parameterize DNA- shape features (see Appendix 2 for further details). The green boxes surround nucleotides included in the counts of GC nucleotides 5’ of the trimming site; in our actual models, we count nucleotides within a 10- nucleotide window (a 5- nucleotide window is shown in the figure). Because this window size is fixed, we do not need to include an additional parameter for AT nucleotide count 5’ of the trimming site (since it is already indirectly modeled). The yellow boxes surround double- stranded nucleotides included in the Figure 1 continued on next page Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 5 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Figure 1 continued counts of AT and GC nucleotides 3’ of the trimming site. These raw 3’-nucleotide counts also indirectly parameterize length; as such, we never include both length and two- side base- count parameters in the same model. In addition to the models shown in the figure, we also evaluated a null model which does not contain any parameters. held- out data groups (Figure 2). Specifically, to evaluate model fit, we computed the expected per- sequence conditional log loss of each model using the full V- gene training data set. To evaluate model generalizability, we computed the expected per- sequence conditional log loss using the following held- out groups: • many random, held- out subsets of the V- gene training data set; • held- out subsets of the V- gene training data set containing groups of V- genes defined to be the ‘most- different’ from all other genes using either the terminal sequences (last 25 nucleotides of each sequence) or the full gene sequences; the full J- gene data set. • For each of these held- out group analyses, each model was re- trained using the full V- gene training data set with the held- out group- of- interest removed (see Materials and methods and Appendix 3 for Figure 2. Overview of analysis strategy. The T cell receptor (TCR)β V- gene training data was used to train each trimming model containing various combinations of sequence- level features (Figure 1) by minimizing the associated loss function. The loss function is given by a sum across individuals i , genes σ , and trimming lengths n of the sampling probability of each observation Ps multiplied by the gene- specific trimming probability predicted by a model with β parameters (see Materials and methods for further details). Each potential model first underwent a ‘model evaluation’ stage (shown by the dashed lines) during which the model performance was evaluated using various subsets of the training TCRβ V- gene data set. Once all models were evaluated, a subset of the potential models continued on to the ‘model validation’ stage (shown by the solid lines) during which the performance of the model coefficients from the previous TCRβ V- gene training run were validated using several independent testing data sets including TCRβ, TCRα, TCRγ, and IGH sequences. At each stage, the performance of each model was compared to a null model (containing zero parameters, see Materials and methods). The online version of this article includes the following figure supplement(s) for figure 2: Figure supplement 1. Using a different, and much larger, repertoire sequencing data set, we have closely replicated previous work (Murugan et al., 2012) which illustrated that a simple position- weight- matrix- style model has good predictive accuracy for many T cell receptor β V- genes. Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 6 of 40 Computational and Systems Biology | Immunology and InflammationTCRβ V-gene training datamodel evaluation pathsmodel validation pathsmechanistically-interpretablesequence-level featuresindependent testing datatrain trimming modelmodel evaluationmodel validationsplit dataCGATGTATCRγδIGHTCRαβloss = losstesting data set(better)(worse)nullmodelpotentialmodellosssubset of training data(better)(worse)nullmodelpotentialmodel Research article Figure 3. Expected per- sequence conditional log loss computed for various models using the full V- gene training data set, many random, held- out subsets of the V- gene training data set, a held- out subset of the V- gene training data set containing a group of V- genes defined to be the ‘most- different’ using the terminal sequences (last 25 nucleotides of each sequence), a held- out subset of the V- gene training data set containing a group of V- genes defined to be the ‘most- different’ using the full gene sequences, and the full J- gene data set. Each model was trained using the full V- gene training data set with the held- out group or ‘most- different’ group (if applicable) removed (see Materials and methods and Appendix 3). Lower expected per- sequence log loss corresponds to better a model fit. The 1×2 motif + two- side base- count beyond model has the best model fit and generalizability across all data sets. The online version of this article includes the following source data and figure supplement(s) for figure 3: Source data 1. Expected per- sequence conditional log loss reported for each model and validation data set. Figure supplement 1. For each model, there was some variation in the expected per- sequence conditional log loss values computed across the 20 random, held- out subsets of the V- gene training data set. further details) prior to computing the loss. A lower expected per- sequence conditional log loss indi- cated better model fit and/or model generalizability. Following this model evaluation, we validated a subset of the models by using the model coefficients from the previous TCRβ V- gene training run and computing the expected per- sequence conditional log loss of the model using several independent testing data sets (Figure 2). Local sequence context, length, and GC nucleotide content in both directions of the wider sequence, together, accurately predict the trimming probabilities of a given V-gene sequence In an effort to capture the complex underlying biochemistry of the deletion process, we trained models containing various combinations of sequence- level feature types (Figure 1B) and evaluated their ability to accurately predict V- gene trimming probabilities. With this approach, we found that a model containing parameterizations of local sequence context, length, and GC nucleotide content in both directions of the wider sequence (the 1×2 motif + two- side base- count beyond model) had the best model fit and generalizability across different data sets (Figure  3 and Figure  3—figure supplement 1). This model contains a 1×2 motif, including 1- nucleotide position 5’ of the trimming Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 7 of 40 Computational and Systems Biology | Immunology and Inflammation Research article site and 2- nucleotide positions 3’ of the trimming site within the trimming window, and includes only bases beyond this trimming window in the AT and GC two- side base- count terms (Figure 1). Despite containing fewer total parameters than the original 2×4 motif model (Murugan et  al., 2012) (12 parameters compared to 18 parameters), the 1×2 motif + two- side base- count beyond model had better predictive accuracy (Figure 4 and Figure 4—figure supplement 1). − We considered the significance of the inferred model coefficients using a Bonferroni- corrected significance threshold of 0.0033 (corrected for the total number of model coefficients). With this threshold, we found that many of the inferred model coefficients were significant and quan- tified mechanistic patterns. Each coefficient represents the change in log10 odds of trimming at a given site resulting from an increase in the feature value, given that all other features are held constant. Within the nucleotides immediately surrounding the trimming site, bases 5’ of the trim- ming site have a slightly stronger effect on the trimming probability than bases 3’ of the trimming site (Figure  4B). Specifically, 5’ of the trimming site, C nucleotides have a strong positive effect on the trimming probability ( log10 coefficient = 0.2388 ) whereas A and T nucleotides have a nega- tive effect ( log10 coefficientA = 0.137 ). In contrast, immediately 0.108 and log10 coefficientT = 3’ of the trimming site, G and T nucleotides have a positive effect on the trimming probability ( log10 coefficientG = 0.093 and log10 coefficientT = 0.125 ) whereas C nucleotides have a negative effect ( log10 coefficient = 0.174 ). These results suggest a different possible mechanistic pattern than previous motif- only models (Murugan et al., 2012; Figure 2—figure supplement 1B). Further, beyond the 1×2 motif sequence window, the count of GC nucleotides 5’ of the motif (within a 10- nucleotide window) has a strong positive effect on the trimming probability ( log10 coefficient = 0.164 ) (Figure  4C). The counts of both AT and GC nucleotides 3’ of the motif have a strong negative effect on the trimming probability ( log10 coefficientAT = 0.126 ). Interestingly, the magnitude of these negative effects are very similar between AT and GC counts. This suggests that the raw number of nucleotides 3’ of the motif (e.g. the length) is more important for predicting the trimming probability at a given site compared to the identity of the nucleotides. p- values for each of these 308 ). We noted minimal coefficients were reported to be smaller than machine tolerance ( 2.23 variation in the magnitude of each inferred coefficient even when changing the number of sequences included in the training data set (Figure 4—figure supplement 7). 0.123 and log10 coefficientGC = 10− − − − × − Because we were interested in parameterizing sequence- breathing effects using the two- side base- count terms, we only included nucleotides that are considered to be double- stranded after hairpin- opening within each count. In our modeling, we assume that the DNA hairpin is opened at the  +2 position, leading to a 4- nucleotide- long 3’-single- stranded- overhang (the 2 nucleotides furthest 3’ are considered P- nucleotides) (Ma et al., 2002; Lu et al., 2007). As such, the first 2 nucleotides of the gene sequence can be considered single- stranded, and we do not include them in the two- side base- count terms. When we train a model that ignores this distinction, and include all gene sequence nucleotides in the two- side base- count terms, we note very similar inferred coefficients and model fit (Figure 4— figure supplement 2). We acknowledge that other hairpin- opening positions may be possible. To explore whether the  +2- hairpin- opening- position assumption could be affecting our inferences, we trained the 1×2 motif  + two- side base- count beyond model with other possible hairpin- opening- position assumptions and noted minimal variation in model fit (Figure 4—figure supplement 3). We also evaluated the predictive accuracy of motif  + two- side base- count beyond models containing different ‘trimming motif’ sizes. We find that models containing a small motif (e.g. a 1×2 motif) achieve similar predictive accuracy and are more generalizable compared to models containing a larger motif (Figure 4—figure supplement 4). Because the trimming mechanism is thought to be consistent across V-, D-, and J- genes from both productive and non- productive sequences, we were also interested in whether the inferred coeffi- cients for the 1×2 motif + two- side base- count beyond model would be consistent between the model trained using the non- productive V- gene training data set, a model trained using a non- productive J- gene data set, and a model trained using a productive V- gene data set. As such, we trained a new 1×2 motif + two- side base- count beyond model using only non- productive J- gene sequences and a separate, new 1×2 motif + two- side base- count beyond model using only productive V- gene sequences (both sequence sets were from the same cohort of individuals as the V- gene training data set). We found that the inferred coefficients were highly similar between the three models (Figure 4— figure supplement 5 and Figure 4—figure supplement 6). Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 8 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Figure 4. Performance of the 1×2 motif + two- side base- count beyond model. (A) Inferred trimming profiles using the 1×2 motif + two- side base- count beyond model have good predictive accuracy overall; here, we show the inferred trimming profiles (in blue) for the most frequently used V- genes. Gene- specific trimming profiles for each individual in the training data set are shown in gray. The sequence context with the highest probability of trimming (3’-TTC- 5’ or 3’-TGC- 5’) from (B and C) is highlighted in orange. (B) Position- weight- matrix of the local sequence context dependence of V- gene trimming probabilities consisting of 1 nucleotide 5’ of the trimming site and 2 nucleotide 3’ of the trimming site from fitting the 1×2 motif + two- side base- count beyond model. Positions 5’ and 3’ of the trimming site have a strong effect on the probability of trimming. (C) Inferred two- side base- count beyond model coefficients from fitting the 1×2 motif + two- side base- count beyond model suggest that the count of GC bases 5’ of the motif has a strong positive effect on the trimming probability whereas the count of GC and/or AT bases 3’ of the motif has a negative effect. The count of AT nucleotides 5’ of the motif (shown in gray) was not included in this model. The black vertical line corresponds to the trimming site. Each inferred coefficient is given as the change in log10 odds of trimming at a given site resulting from an increase in the feature value, given that all other features are held constant. The online version of this article includes the following source data and figure supplement(s) for figure 4: Source data 1. Inferred and observed trimming profiles for the most frequently used V- genes in the V- gene training data set. Source data 2. Inferred 1x2 motif + two- side base- count beyond model coefficients. Figure supplement 1. Performance of the 1×2 motif + two- side base- count beyond model across all TRB V- genes, ordered by the frequency of usage in the training data set. Figure supplement 2. Including all gene sequence nucleotides in the two- side base- count terms, instead of restricting to double- stranded nucleotides, leads to very similar inferred coefficients and model fit. Figure supplement 3. The assumed position of the initial hairpin- opening nick during the early stages of V(D)J recombination has little effect on the inferred coefficients and model fit. Figure 4 continued on next page Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 9 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Figure 4 continued Figure supplement 4. Models containing a small motif (e.g.a 1×2 motif) achieve similar predictive accuracy and are more generalizable compared to models containing a larger motif. Figure supplement 5. Inferred coefficients from a 1×2 motif + two- side base- count beyond model trained using only J- gene sequences are highly similar to those from the model trained using the V- gene training data set. Figure supplement 6. Inferred coefficients from a 1×2 motif + two- side base- count beyond model trained using only productive V- gene sequences are highly similar to those from the model trained using the non- productive V- gene training data set. Figure supplement 7. The magnitudes of the inferred coefficients from the 1×2 motif + two- side base- count beyond model have minimal variance when changing the number of sequences included in the training data set. When evaluating models containing only a single feature type, we find that the two- side base- count model which parameterizes GC nucleotide content on both sides of the trimming site (and, indi- rectly, length) has the best model fit and generalizability across all held- out groups tested (Figure 3). As such, these GC- content features, which are likely parameterizing the capacity for the sequence to breathe, are more predictive of V- gene trimming probabilities than local sequence context or DNA- shape alone. This finding supports previous observations that Artemis may act as a structure- specific nuclease as opposed to a nuclease that binds specific DNA sequences (Ma et al., 2005; Chang et al., 2015; Chang and Lieber, 2016; Yosaatmadja et al., 2021). Inferred local sequence context coefficients suggest a biological trimming motif A persistent concern with the 1×2 motif  + two- side base- count beyond model was that the motif coefficients could be driven by certain genes, instead of representing an actual gene- segment- wide signal. When comparing the inferred trimming profiles from the two- side base- count model to those from the 1×2 motif + two- side base- count beyond model, we identified a group of V- genes which had drastically lower prediction error when the 1×2 motif terms were included. These V- genes had a difference in per- gene root mean squared error between the two models that was greater than –0.13 (Figure  5A). The genes included in this group were TRBV5- 3, TRBV7- 3*01, TRBV7- 3*04, TRBV7- 4, TRBV9, TRBV11, and TRBV13. To evaluate whether these genes could be driving the observed motif signal, we explored whether the prediction error for these genes changed when they were removed from the model training data set. In fact, we found that the inferred trimming profiles for these genes still had very low prediction error despite the genes not being included in the model training data set (Figure 5B and C), showing the generalizability of these features. The inferred model coefficients from this 1×2 motif + two- side base- count beyond model fit using the subsetted training data set were highly similar to those from the original model fit using the full training data set. Because genes which are highly similar sequence- wise to the group of held- out genes could still be present in the training data set and be driving these similarities, we defined a new data set that excluded this larger group of genes. When we repeated the same experiment with this new, more- restricted training data set, we observed similar results (Figure 5B and C). As such, both of these experiments provided evidence that the motif signal may actually represent a gene- segment- wide sequence motif that appears to get preferentially trimmed, independent of GC- content- related effects. Trimming-associated variation within the Artemis locus is associated with a change in model coefficients Using a subset of the V- gene training data set used here, we previously identified a set of single nucle- otide polymorphisms (SNPs) within the gene encoding the Artemis protein that are associated with increasing the extent of V- and J- gene trimming (Russell et al., 2022b). This result suggested that trimming profiles may subtly vary in the context of these SNPs. As such, we were interested in whether these SNPs could be mediating (or serving as a proxy for) a change in the trimming mechanism. To explore this, we worked with paired SNP array and TCRβ-immunosequencing data representing 611 of the original 666 individuals in the V- gene training data set used here. Our previous work Russell et al., 2022b used data from only 398 of these individuals, however, the conclusions of that paper held when using this expanded group of 611 individuals in the analysis. With these data, we asked Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 10 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Figure 5. The 1×2 motif coefficients represent a gene- segment- wide trimming motif. (A) Distribution of the difference in per- gene root mean squared error (RMSE) between the 1×2 motif + two- side base- count beyond model and the two- side base- count model. V- genes with an RMSE difference less than –0.127 (gray vertical line) were in the lowest 10% of all RMSE differences. These ‘improved genes’ showed a large RMSE improvement when including motif terms in the model. (B) Inferred trimming profiles for TRBV9, the gene which showed the largest RMSE improvement in (A). TRBV9 had an RMSE difference of –0.31. (C) Inferred trimming profiles for TRBV13, the gene which showed the second largest RMSE improvement in (A). TRBV13 had an RMSE difference of –0.15. The inferred trimming profiles for TRBV9 and TRBV13 using models which contain motif terms have very low prediction error even when the genes are not included in the model training data set. Gene- specific trimming profiles for each individual in the training data set are shown in gray. The sequence context with the highest probability of trimming (3’-TTC- 5’ or 3’-TGC- 5’ from Figure 4B) are highlighted in orange. The online version of this article includes the following source data for figure 5: Source data 1. Per- gene mean squared error difference between the 1×2 motif + two- side base- count beyond model and the two- side base- count model. Source data 2. Inferred and observed trimming profiles for the genes with largest root mean squared error (RMSE) improvement. whether the inferred coefficients from the V- gene- specific 1×2 motif + two- side base- count beyond model varied significantly in the context of the non- coding Artemis- locus SNP (rs41298872) that was found to be most strongly associated with increasing the extent of V- gene trimming in our previous work (Russell et al., 2022b). As such, we re- defined the model to include an interaction coefficient between the SNP genotype and each model parameter (see Materials and methods). We then used a Bonferroni- corrected significance threshold of 0.0033 (corrected for the total number of interaction coefficients) to evaluate the significance of each interaction coefficient. For each significant interaction coefficient, we concluded that the corresponding model coefficient varied significantly in the context of the SNP genotype. Using these methods, we found that several of the 1×2 motif  + two- side base- count beyond model coefficients varied significantly in the context of the Artemis- locus SNP rs41298872 (Figure 6). Specifically, we found that 3’ of the trimming site, the negative effect of A nucleotides on the trim- ming odds varied in the context of the SNP for the position immediately 3’ of the trimming site ( log10 interaction coefficient = 0.006 , p = 0.0006 ) and one position away ( log10 interaction coefficient = 0.007 , p = 0.0006 ). Further, we found that the negative effect of the count of AT nucleotides 3’ of the motif Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 11 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Figure 6. Inferred single nucleotide polymorphism (SNP)- parameter- interaction coefficients from fitting the 1×2 motif + two- side base- count beyond SNP- interaction model. Note that the inferred coefficients for each main parameter (as shown in Figure 4) are not displayed here; only the inferred interaction coefficients between the SNP and each parameter are shown. (A) Inferred interaction coefficients between rs41298872 SNP genotype and motif parameters for 1- nucleotide position 5’ of the trimming site and 2- nucleotide positions 3’ of the trimming site. The interaction coefficients between the SNP genotype and the presence of A nucleotides (at all positions 3’ of the motif) are significant. This figure is a different representation of the information shown in (A). (B) Inferred interaction coefficients between rs41298872 SNP genotype and two- side base- count beyond model coefficients. The interaction coefficients between the SNP genotype and the count of AT nucleotides 3’ of the motif are significant. The interaction coefficient between the SNP genotype and the count of AT nucleotides 5’ of the motif (shown in gray) was not included in this model. The black vertical line corresponds to the trimming site. Each inferred interaction coefficient is given as the change in log10 odds of trimming at a given site resulting from an increase in the feature value and a change in genotype, given that all other features are held constant. The online version of this article includes the following source data and figure supplement(s) for figure 6: Source data 1. Inferred 1×2 motif + two- side base- count beyond, single nucleotide polymorphism (SNP) interaction model coefficients. Figure supplement 1. The significance of the 3’-AT- nucleotide count single nucleotide polymorphism (SNP)- interaction coefficient appears to be related to length effects rather than nucleotide content. varied strongly in the context of the SNP ( log10 interaction coefficient = 0.010 , p = 1.47 motif or two- side base- count coefficients were found to significantly vary. × 10− 12 ). No other Because the 3’-side base- count- beyond terms parameterize both GC nucleotide content and length in their definition, we were interested in whether the significance of the 3’-AT- nucleotide count SNP variation effect was related to GC nucleotide content, length, or both. To do this, we re- defined the 3’-side base- count- beyond parameters to be a proportion instead of raw AT/GC nucleotide counts and included an additional length term in the model to remove length- related effects from the inferred 3’-side base- count- beyond coefficients. Using this new model, we repeated the analysis and found that the length coefficient varied significantly in the context of the SNP ( log10 interaction coefficient = 0.005 , 23 ), but the 3’-AT- nucleotide- proportion term did not (Figure 6—figure supplement 1). p = 6.24 This result is fully consistent with the fact that the Artemis- locus SNP is known to be associated with increasing the extent of trimming (a proxy for length). 10− × Local sequence context, length, and GC nucleotide content in both directions of the wider sequence can also accurately predict the trimming probabilities of a given sequence from other receptor loci To validate our previously trained models, we worked with TCRα- and TCRβ-immunosequencing data representing 150 individuals, TCRγ-immunosequencing data representing 23 individuals, and IGH- immunosequencing data representing 9 individuals from three independent validation cohorts. Before analyzing these data, we ‘froze’ our trained model coefficients in git commit 093610a on our reposi- tory. In contrast to the training data cohort, these validation cohorts contain different demographics and were each processed using different sequence annotation methods (see Materials and methods). To explore the potential effects of using a different sequence annotation method, we re- annotated Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 12 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Figure 7. Expected per- sequence conditional log loss computed for various models using the T cell receptor β V- gene training data set and non- productive V- and J- gene sequences from several independent testing data sets. Each model was trained using the full non- productive TCRβ V- gene training data set. Lower expected per- sequence log loss corresponds to a better model fit. The 1×2 motif + two- side base- count beyond model has the best model fit and generalizability across all testing data sets. The horizontal dashed line corresponds to the expected per- sequence log loss of the 1×2 motif + two- side base- count beyond model computed for V- gene trimming using the non- productive TCRβ V- gene training data set. The online version of this article includes the following source data and figure supplement(s) for figure 7: Source data 1. Expected per- sequence conditional log loss reported for each model and testing data set. Figure supplement 1. Differing methods of sequence annotation have little to no effect on the model fit or performance. Figure supplement 2. Model performance is similar for productive sequences compared to non- productive sequences from each testing data set. Figure supplement 3. Using the 1×2 motif + two- side base- count beyond model, the weight of the two- side base- count beyond terms are dominant relative to the 1×2 motif terms for every testing data set. Figure supplement 4. Sequence motifs appear at varying frequencies within the germline TRB and IGH genes. the TCRβ training data set using the same annotation method as the TCRα-β testing data and found that it had little to no effect on the model fit or performance (Figure 7—figure supplement 1). To evaluate the performance of the 1×2 motif + two- side base- count beyond model using these testing data, we used the model coefficients from the previous TCRβ V- gene training run and computed the expected per- sequence conditional log loss of the model using each testing data set (TCRβ V- gene sequences, TCRα V- gene sequences, TCRγ V- gene sequences, IGH V- gene sequences, TCRβ J- gene sequences, etc.). We found that the model has high predictive accuracy (i.e. low expected per- sequence conditional log loss) for both non- productive V- and J- gene sequences from the TCRβ Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 13 of 40 Computational and Systems Biology | Immunology and Inflammation Research article testing data set (Figure 7). The model also extends well to non- productive V- and J- gene sequences from the TCRα and TCRγ testing data sets and to non- productive V- gene sequences from the IGH testing data set. The model has relatively poor predictive accuracy for non- productive IGH J- gene sequences, however. We noted very similar results when validating model performance using produc- tive V- and J- gene sequences from each testing data set (Figure 7—figure supplement 2). We hypothesized that the weight of the 1×2 motif and two- side base- count beyond model terms may vary across each testing data set. To explore this for each data set, we again used the model coefficients from the previous TCRβ V- gene training run and trained a new two- parameter model containing one coefficient scaling the 1×2 motif terms and a second coefficient scaling the two- side base- count beyond terms (see Materials and methods). With this approach, we found that the two- side base- count beyond terms were dominant compared to the 1×2 motif terms for every data set (Figure  7—figure supplement 3A). The scale coefficient for the 1×2 motif terms was very small for several of the data sets, especially the IGH data set, indicating only a weak motif- related signal. The sequence motifs that lead to a large increase in trimming probabilities in the model appear at relatively low frequencies within the germline IGH genes (Figure 7—figure supplement 4), perhaps explaining the weakness of the motif- related signal. When evaluating the expected per- sequence conditional log loss of these partially re- trained models, we note a small improvement in model fit for each re- trained model compared to the original model (Figure 7—figure supplement 3B). Discussion The junctional deletion and insertion steps of the V(D)J recombination process are essential for creating diversity within the TCR repertoire. While the Artemis protein is often regarded as the main nuclease involved in V(D)J recombination, the exact mechanism of nucleotide trimming has yet to be understood in a human system. Using a previously published high- throughput TCRβ sequencing data set, we designed a flexible probabilistic model of nucleotide trimming that allowed us to explore the relative importance of various sequence- level features. While we recognize that these general model features may not capture the full complexity of the trimming mechanism and establish causation, we were primarily interested in identifying mechanistically interpretable features which could confirm and extend our current understanding of the nucleotide trimming process. With this framework, we have (1) revealed a set of sequence- level features which can be used to accurately predict trimming probabilities across various adaptive immune receptor loci, (2) shown that length and GC nucleotide content in both directions of the wider sequence are highly predictive of trimming probabilities, quan- tifying how double- stranded DNA needs to be able to breathe for trimming to occur, (3) identified a sequence motif that appears to get preferentially trimmed, independent of length- and GC- content- related effects, and (4) demonstrated that a genetic variant within the gene encoding the Artemis protein is associated with changes in several model coefficients. Specifically, we find that a model containing parameterizations of both local sequence context, length, and GC nucleotide content in both directions of the wider sequence can most accurately predict the trimming probabilities of a given TCRβ gene sequence. In addition to having fewer parameters, this model also had better predictive accuracy than a previously proposed sequence context model (Murugan et al., 2012). Models containing other sequence- level parameters such as DNA- shape and length also had relatively worse predictive accuracy. The TR and IG V(D)J recombination processes, including trimming profiles, have previously been suggested to vary substantially across individuals (Slabodkin et al., 2021; Russell et al., 2022b). Here, our results support a universal, sequence- based trimming mechanism underlying this variation across TR and IG loci in humans. Specifically, in addi- tion to TCRβ sequences, we find that local sequence context, length, and GC nucleotide content in both directions of the wider sequence can be used to accurately predict trimming probabilities across TCRα, TCRγ, and IGH sequences. For all of these loci, we find that length and GC nucleotide content are relatively more important than local sequence context terms for making accurate model predictions. The Artemis protein, in complex with DNA- PKcs, is responsible for opening the DNA hairpin during the early steps of V(D)J recombination to generate a 4- nucleotide- long 3’-single- stranded overhang at the end of each gene, and has been suggested to continue on to trim nucleotides from this resulting DNA structure (Feeney et al., 1994; Nadel and Feeney, 1995; Nadel and Feeney, 1997; Jackson et al., 2004; Gu et al., 2010; Chang et al., 2015; Chang and Lieber, 2016; Zhao et al., 2020; Russell Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 14 of 40 Computational and Systems Biology | Immunology and Inflammation Research article et  al., 2022b). The Artemis protein, with and without DNA- PKcs, has been shown to bind single- stranded- to- double- stranded DNA boundaries prior to nicking DNA (Ma et al., 2002; Ma et al., 2005; Chang et al., 2015; Chang and Lieber, 2016). While the single- stranded overhang created during hairpin- opening may create a natural single- stranded- to- double- stranded DNA substrate for Artemis binding near the end of the gene sequence, we find that many trimming events occur further into the double- stranded gene sequence. Indeed, previous in vitro DNA nuclease assays involving Artemis have shown that sequence- breathing dynamics are often required to generate a transient single- stranded- to- double- stranded DNA substrate prior to Artemis action (Chang et al., 2015). Using our model of nucleotide trimming, we have shown that trimming probabilities are highest for DNA posi- tions closer to the end of the sequence. Because these DNA positions have fewer double- stranded nucleotides on the 3’-side of the trimming site, they may have more capacity for sequence- breathing. On the 5’-side of the trimming site, we find that having a larger number of G- C nucleotides, and perhaps less sequence- breathing capacity, increases the trimming probability. Perhaps this breathing transition can create a transient single- stranded- to- double- stranded DNA substrate that is suitable for Artemis to bind and trim. As such, this finding quantifies sequence- breathing effects that were previ- ously identified through in vitro DNA nuclease assay studies involving Artemis (Chang et al., 2015). Independent of GC- content- related effects, we have also identified a gene- segment- wide sequence motif that appears to get preferentially trimmed. This motif is suggestive of sequence- specific nucle- olytic activity, however, Artemis is widely regarded as a structure- specific nuclease as opposed to a nuclease that binds specific DNA sequences (Ma et al., 2005; Chang et al., 2015; Chang and Lieber, 2016; Yosaatmadja et al., 2021). This suggests that either (1) Artemis actually does possess some ability to recognize specific nucleotides, (2) the observed sequence motif is serving as a proxy for DNA structure induced by the motif, or (3) another nuclease, in addition to Artemis, is responsible for the sequence- specific trimming we observe. However, because the strength of this sequence motif signal varied across receptor loci, further work will be required to explore its mechanistic basis and presence. We found that several model coefficients related to local sequence context, length, and GC nucle- otide content in both directions of the wider sequence varied significantly in the context of the non- coding Artemis- locus SNP rs41298872. We previously identified this Artemis- locus SNP as being associated with increasing the extent of TCRβ V- and J- gene trimming (Russell et al., 2022b). While many previous studies have reported a high consistency of TCRβ trimming profiles across individuals (Murugan et al., 2012; Marcou et al., 2018; Sethna et al., 2020), our results begin to explore how the trimming mechanism may vary across individuals in the context of Artemis genetic variation. We reported that trimming probabilities decrease as the number of double- stranded nucleotides 3’ of the trimming site increases. In the context of the SNP rs41298872, we found that as the number of double- stranded AT nucleotides 3’ of the trimming site increases, the trimming probabilities do not decrease as quickly. This suggests that individuals homozygous (or heterozygous) for rs41298872 may be more capable of trimming at positions that have a larger number of double- stranded nucleotides 3’ of the trimming site, especially if the additional nucleotides are AT bases. This may be possible if, for example, rs41298872 increases Artemis expression. If there is more Artemis available, then trimming at less optimal positions (i.e. positions further into the sequence which have less breathing) may be possible. Additional work will be required to define the relationship between rs41298872 genotype and Artemis expression. We also identified several local sequence context coefficients that varied in the context of rs41298872, however, their mechanistic interpretation remains unclear. Earlier, we noted that A nucle- otides 3’ of the trimming site have a negative effect on the trimming probability while T nucleotides have a strong positive effect. In the context of rs41298872, we found that the magnitude of the negative effect of 3’ A nucleotides on the trimming probability was reduced. This may suggest that individuals homozygous (or heterozygous) for rs41298872 may trim in a less motif- dependent fashion, and are instead more reliant on sequence openness 3’ of the trimming site. In this way, having A or T nucleotides 3’ of the trimming site would create a more open local sequence for trimming. There are several key limitations of our approach which are intrinsic to the use of adaptive immune receptor repertoire data. First, we have used trimming statistics from non- productive rearrangements as a means of studying the nucleotide trimming process in the absence of selection. Non- productive sequences can be sequenced as part of the repertoire when they are present within a cell expressing a productive rearrangement that survived the selection process. While we are not aware of a mechanism Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 15 of 40 Computational and Systems Biology | Immunology and Inflammation Research article through which non- productive and productive rearrangements within a single cell could be correlated, we also acknowledge that the repertoire of non- productive rearrangements may be an imperfect proxy for a pre- selection repertoire. However, as is common in the literature (Robins et  al., 2010; Murugan et al., 2012; Marcou et al., 2018; Sethna et al., 2019; Sethna et al., 2020), we assume that the two recombination events are independent and that the non- productive rearrangements reflect the statistics of the repertoire prior to selection. Next, because many V(D)J rearrangement scenarios can give rise to the same final nucleotide sequence, possible error related to the annota- tion of each sequence may have restricted our ability to model the actual trimming distributions of each gene. Although we cannot rule out some effect of incorrect sequence annotation on our model inferences, we found that the exact sequence annotation method used, including sampling from the posterior distribution of rearrangement events, had little to no effect on the model fit or performance. In summary, we have found that local sequence context, length, and the GC nucleotide content in both directions of the wider sequence can accurately predict the trimming probabilities of TR and IG gene sequences. These results refine our understanding of how nucleotides are trimmed during V(D)J recombination. The sequence- level features identified here lay the groundwork for further exploration into the trimming mechanism and how it may vary across individuals. Such insights will provide another step toward understanding how V(D)J recombination generates diverse receptors and supports a powerful, unique immune response in humans. Materials and methods Training data set TCRβ repertoire sequence data for 666 healthy bone marrow donor subjects was downloaded from the Adaptive Biotechnologies immuneACCESS database using the link provided in the original publi- cation (Emerson et  al., 2017). V(D)J recombination scenarios were assigned to each sequence for each individual using the IGoR software (version 1.4.0) (Marcou et  al., 2018) as follows. The IGoR software can learn unbiased V(D)J recombination statistics from immune sequence reads. Using these statistics, IGoR can output a list of potential recombination scenarios with their corresponding like- lihoods for each sequence. As such, using the default IGoR V(D)J recombination statistics, the 10 highest probability V(D)J recombination scenarios were inferred for each TCRβ-chain sequence in the training data set (Marcou et al., 2018). We then annotated each TCRβ-chain sequence with a single V(D)J recombination scenario by sampling from these 10 scenarios according to the posterior prob- ability of each scenario. We filtered these sequences for rearrangements which contained more than 1 trimmed nucleotide and less than 15 trimmed nucleotides (see the ‘Notation’ section for further details). We further subset the data to include only non- productive sequences, and used these data for all subsequent model training. After these processing and filtering steps, we used V- gene trim- ming length distributions from 21,193,153 non- productive sequences for all model training. To test each trained model, we used V- gene trimming length distributions from the remaining 107,121,841 productive sequences (as described in Appendix 3). From this same data set, we also used J- gene trimming length distributions from 107,255,406 productive sequences and 20,204,801 non- productive sequences to test each model. Testing data sets TCRα and TCRβ testing data sets Annotated TCRα and TCRβ repertoire sequence data for 150 healthy subjects was downloaded using the link provided in the original publication (Russell et al., 2022b). In contrast to the training data cohort, this cohort contains different demographics, shallower RNA- seq- based TCR sequencing, and was processed using a different sequence annotation methods (i.e. TCRdist [version 0.0.2] [Dash et  al., 2017] as described in a previous publication [Russell et  al., 2022b]). Sequences were split into non- productive and productive groups for model validation. From the TCRα data set, we used V- gene trimming length distributions from 123,496 non- productive sequences and 862,096 produc- tive sequences and J- gene trimming length distributions from 141,451 non- productive sequences and 1,101,114 productive sequences to test each model. From the TCRβ data set, we used V- gene trimming length distributions from 64,738 non- productive sequences and 1,435,153 productive Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 16 of 40 Computational and Systems Biology | Immunology and Inflammation Research article sequences and J- gene trimming length distributions from 59,608 non- productive sequences and 1,496,953 productive sequences to test each model. TCRγ testing data set Annotated TCRγ repertoire sequence data for 23 healthy bone marrow donor subjects was down- loaded from the Adaptive Biotechnologies immuneACCESS database (Robins and Pearson, 2015). Sequences were split into non- productive and productive groups for model validation. We used V- gene trimming length distributions from 2,403,293 non- productive sequences and 1,002,662 productive sequences and J- gene trimming length distributions from 568,824 non- productive sequences and 250,493 productive sequences to test each model. IGH testing data sets Annotated IgG class non- productive IGH repertoire sequence data for nine healthy subjects was obtained from the authors of a previous publication (Spisak et al., 2020). The raw sequence data is available using the link provided in the original publication (Briney et al., 2019). In contrast to the training data cohort, this cohort contains different demographics, shallower RNA- seq based IGH- sequencing, and was processed using a different sequence annotation method (i.e. a combination of Immcantation [Vander Heiden et al., 2014] and IgBlast [Ye et al., 2013] as described in a previous publication [Spisak et al., 2020]). Further, these data are restricted to rearrangements that lead to a clonal family with at least six members. Likewise, productive IGH repertoire sequence data for four healthy subjects was downloaded using the link provided in the original publication (Jaffe et al., 2022) and the sequences were annotated using partis (version 0.16.0) (Ralph and Matsen, 2016). Due to the large size of this data set, 100k sequences were randomly sampled from the original data set prior to model validation. For both IGH data sets, only a single sequence from each inferred clonal family was included in each model testing data set. From these data sets, we used V- gene trimming length distributions from 160,714 non- productive sequences and 32,245 productive sequences and J- gene trimming length distributions from 297,298 non- productive sequences and 74,884 productive sequences to test each model. Artemis-locus SNP data set Genome- wide SNP array data corresponding to 611 of the training data set individuals was down- loaded from The database of Genotypes and Phenotypes (accession number: phs001918). Details of the SNP array data set, genotype imputation, and quality control have been described previously (Martin et al., 2020). We only used SNP data corresponding to the Artemis locus (rs41298872) which we previously found to be strongly associated with increasing the extent of V- gene trimming (Russell et al., 2022b). ∈ Notation Let I be a set of individuals. For each subject i I , assume we have a TCR repertoire consisting of sequences indexed by k such that k = 1, . . . , Ki . We assume that each sequence can be unambiguously annotated with being from a specific V- gene and J- gene sequence, and having a number of deleted nucleotides from each gene. For modeling purposes, we combine TRB V- gene or J- gene alleles that have identical terminal nucleotide sequences (last 24 nucleotides of each sequence) into TRB V- gene allele groups and TRB J- gene allele groups. As such, each TCR sequence is annotated with being from a V- gene allele group and J- gene allele group. Because we are requiring that each gene allele group originates from the same TRB V- gene or J- gene, there may still be overlap in terms of sequence identity between allele groups. For simplicity, we orient all sequences in the 5’-to- 3’ direction, and use the top strand for V- gene sequences and the bottom strand for J- gene sequences. We will be introducing modeling methods as they relate to V- genes and V- gene trimming, however, with this sequence orientation, the same methods can be applied to J- genes and J- gene trimming. We will use σ to represent a gene sequence oriented in the 5’-to- 3’ direction and n to represent the number of nucleotides deleted from the 3’ end of this sequence as we describe our modeling. We are interested in modeling the probability of trimming n nucleotides from a given gene sequence σ , P(n | σ) . We can define an empirical conditional probability density function to estimate this probability. To start, we can uniformly sample from any given individual’s repertoire. Let S be a Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 17 of 40 Computational and Systems Biology | Immunology and Inflammation Research article random variable that represents the gene- allele- group sequence from such a sample. Let N be a random variable that represents the number of deleted nucleotides, which for notational convenience we assume take on a non- negative integer value (nonsensical values will have probability zero). Let Ki 0 represent the number of TCRs that have gene allele group σ and n gene nucleotides deleted. With these data, we can form the empirical conditional probability density function: Ki represent the number of TCRs that use gene allele group σ . Let 0 C(i)(n, σ) C(i)(σ) ≤ ≤ ≤ ≤ Pemp(N = n | S = σ, i) = C(i)(n, σ) C(i)(σ) . (1) Using these TCRβ repertoire data, we want to model the influence of various sequence- level parameters on P(n | σ) . With this assumption, let L and U be lower and upper bounds, respectively, on n such that N′ = {L, . . . , U} is the set of all reasonable nucleotide deletion amounts. The precise location of hairpin opening and its relationship to deletion is unclear. Hence, we have chosen to define L = 2 since smaller trimming amounts may result from an alternative, hairpin- opening- position- related (or other) trimming mechanism. Likewise, we have chosen to define U = 14 since trimming amounts greater than 14 nucleotides are uncommon and could also result from an alternative trim- ming mechanism. We will subset the training data set, after IGoR annotation (see details in a previous section), such that we will only consider TCRs that have 2 14 . Similarly, the one existing analysis (Murugan et al., 2012) exploring the relationship between sequence context and nucleotide trim- ming only considered TCRs that had 2 12 for their modeling. We summarize all of the notation n ≤ discussed in this section, as well as in the following sections, in Appendix 1—table 1. ≤ ≤ ≤ n V(D)J recombination modeling assumptions For our model, we make the following assumptions about V(D)J recombination biology: 1. During the V(D)J recombination process, the gene DNA hairpin is nicked open by a single- stranded break (Gauss and Lieber, 1996; Nadel and Feeney, 1997; Ma et al., 2002; Jackson et al., 2004; Lu et al., 2007). 2. This hairpin nick occurs at the +2 position, leading to a 4- nucleotide- long 3’-single- stranded- overhang (the 2 nucleotides furthest 3’ are considered P- nucleotides) (Ma et al., 2002; Lu et al., 2007). We will discuss a sensitivity analysis to this assumption, which showed that the assumed hairpin- nick position had little impact on our model fitting, in the appendix. 3. If any nonzero amount of the original gene sequence is deleted, all P- nucleotides will also be deleted (Gauss and Lieber, 1996; Srivastava and Robins, 2012). 4. Nucleotide trimming occurs before N- insertion. With these assumptions, we can resolve the nucleotide sequence on both sides of the trimming site and define mechanistically interpretable model features using these two sequences. Specifically, Table 1. Summary of all parameter- specific coefficients and covariate functions for a trimming site n and gene sequence σ . Here, a and b represent the number of nucleotides 5’ and 3’ of the trimming site to be included in the ‘trimming motif,’ respectively, and c represents the number of nucleotides 5’ of the trimming site to be included in the base- count. Parameter Model coefficient variables Parameter- specific covariate function Motif parameters βmotif coefficients f1(n, σ; βmotif , a, b) (Equation 14) Base- count- beyond parameters βAT and βGC coefficients f2(n, σ; βAT , βGC , a, b, c) (Equation 17) DNA- shape parameters βshape coefficients Length parameters βldist coefficients f3(n, σ; βshape , a, b) (Equation 19) f4(n, σ; βldist ) (Equation 20) Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 18 of 40 Computational and Systems Biology | Immunology and Inflammation Research article we define a ‘trimming motif’ consisting of several nucleotides on either side of the trimming site, the predicted ‘DNA- shape’ of the nucleotides and bonds in close proximity to the trimming site, the counts of GC or AT nucleotides on either side of the trimming site beyond the ‘trimming motif’ region (e.g. the ‘two- side base- count beyond’), and the sequence- independent ‘length’ from the end of the gene to the trimming site (see Appendix 2 for further details). An example of how an arbitrary V- gene sequence is transformed into features for modeling is shown in Figure 1. We will assume that observations can be drawn from a model in which these features vary across trimming lengths n for a given gene allele group σ . We can then explore the influence of these features on the probability of trimming at a certain site given a gene sequence. Defining a model covariate function With the features summarized above, we can define a model covariate function f than contains any unique combination of parameter- specific covariate functions (Table  1). This function f will be the sum of each of the desired parameter- specific covariate functions. This framework allows us to generalize the existing PWM model (Murugan et al., 2012) to a model that allows for arbitrary sequence features. For example, we replicate this PWM model using the model covariate function, f1(n, σ; βmotif , a = 2, b = 4) , where n represents the number of trimmed nucleotides, σ represents the gene- allele- group sequence, βmotif represents motif- specific parameter coefficients, and a and b are non- negative integer values that represent the number of nucleotides 5’ and 3’ of the trimming site, respectively, that are included in the ‘trimming motif’. This function is described further in (Equation 14). To extend this model to a model containing motif parameters and base- count- beyond parame- ters, the model covariate function will be f(n, σ; βmotif , βAT , βGC , a, b, c) := f1(n, σ; βmotif , a, b) +f 2(n, σ; βAT , βGC , a, b, c) (2) and βGC where f2 represents the base- count- beyond model covariate function (Equation 17), βAT represent base- count- beyond- specific parameter coefficients, and c represents the number of nucle- otides 5’ of the trimming site to be included in the base- count. We will use this motif and base- count- beyond model example to discuss the model formulation in the following sections, however, many other parameter combinations are possible. We will not define a model covariate function that contains two parameters that model the same feature. For example, length and base- count- beyond coefficients will never be included in a model covariate function together (since they both parame- terize length). Likewise, motif and DNA- shape coefficients will never both be included in a model covariate function. Predicting trimming probabilities using conditional logistic regression We will be using the motif and base- count- beyond parameters given by (Equation 2) as examples for the remainder of this section, however, we could also formulate a model with any other parameter of interest, as described in the previous section (Table 1). As such, we can fit a conditional logit model which posits that P(n | σ; βmotif , βAT , βGC , a, b, c) := ∑ exp(f(n, σ; βmotif , βAT N′ exp(f(n′, σ; βmotif , βGC , βAT , a, b, c)) , βGC , a, b, c)) n′∈ (3) where N′ is the set of all reasonable trimming lengths, a and b represent the number of nucleotides 5’ and 3’ of the trimming site to be included in the ‘trimming motif,’ respectively, c represents the number of nucleotides 5’ of the trimming site to be included in the base- count parameters, and f(n, σ; βmotif , a, b, c) is the model covariate function for the motif and base- count- beyond model given by (Equation 2). We will let P(n | σ; βmotif , βGC , a, b, c) denote the conditional proba- bility that a given gene will be trimmed by n nucleotides. , βGC , βAT , βAT Let yikσn equal 1 if a gene allele group σ is trimmed by n nucleotides for TCR k from subject i , and , βAT , a, b, c) , such that , a, b, c) , is the likelihood of the model parameters, , given that we observed a set of trimming amounts for a set of given genes. As equal 0 otherwise. With this, we can define a likelihood function, L(βmotif for a random sample of subjects, L(βmotif βmotif such, the log- likelihood function can be written as , and βGC , βAT , βGC , βAT , βGC Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 19 of 40 Computational and Systems Biology | Immunology and Inflammation Research article log L(βmotif , βAT , βGC , a, b, c) = yikσn · ∑i,k,σ,n log P(n | σ; βmotif , βAT , βGC , a, b, c) , βGC , β AT where P(n | σ; βmotif , a, b, c) is given by (Equation 3). Instead of maximizing this log- likelihood directly, we may wish to aggregate the data to reduce the number of observations and simplify model fitting. Recall that for subject i , C(i)(σ) represents the number of TCRs which use gene allele group σ and C(i)(n, σ) represents the number of TCRs which have gene allele group σ and n gene nucleotides deleted. As such, C(i)(n, σ) is the count of observations which will have the same trimming probabilities P(n | σ; βmotif , a, b, c) and will have been trimmed by n for subject i and gene allele group σ . Thus, using this aggregated data from all subjects i I , we can re- write the log- likelihood function equivalently as , β AT , βGC ∈ log L(βmotif , βAT , βGC , a, b, c) = C(i)(n, σ) · ∑i,σ,n log P(n | σ; βmotif , βAT , βGC , a, b, c). (4) As above, for a random sample of subjects, L(βmotif , and βGC , a, b, c) is the likelihood of the model , given that we observed a set of trimming amounts for a set of given , βAT , βAT , βGC parameters, βmotif genes. With this likelihood formulation, all observations in the sample get uniform treatment in the construction of the likelihood. However, subjects may differ in their repertoire size and composition for reasons other than trimming. For example, it is known that gene usage differs across subjects. ˆβGC inference, we propose a Thus, to avoid having these differences pollute our subject and gene weighting scheme. ˆβmotif ˆβAT , , and As such, we can define the expected likelihood of a process where we first draw a subject i uniformly at random, then we sample TCR sequences from their repertoire according to a given distribution, as follows. For a single TCR sequence from such a sample, let S be a random variable representing the gene of the sequence, and let N be a random variable representing the number of deleted nucleo- tides. We can sample each TCR sequence with probability Psamp(N = n, S = σ) which we will specify , and βGC later. Also, given random S and N , the log- likelihood of the model parameters, βmotif , is given by , βAT log L(βmotif , βAT , βGC , a, b, c; N, S) = log P(N | S; βmotif , βAT , βGC , a, b, c). With this, the expected log- likelihood of the model parameters, βmotif random sample is given by , βAT , and βGC given this E[log L(βmotif, βAT, βGC, a, b, c) | I = i] Psamp(N = n, S = σ) = n,σ ∑ = n,σ ∑ log P(N = n, S = σ; βmotif , βAT , βGC , a, b, c) · Psamp(N = n, S = σ) log P(N = n | S = σ; βmotif , βAT , βGC , a, b, c). · We can define a new, weighted log- likelihood function, log Lexpected(βmotif to this expected log- likelihood: , βAT , βGC , a, b, c) , equivalent log Lexpected(βmotif , βAT , βGC , a, b, c) := E[log L(βmotif , βAT , βGC , a, b, c) | I = i]. ∑i (5) For a random sample of subjects, the weighted likelihood, Lexpected(βmotif , βGC , a, b, c) , represents the likelihood of the model parameters, βmotif , given that we observed a set of trimming amounts for a given set of gene allele groups after weighting observations according to the sampling procedure Psamp(N = n, S = σ) . We can use whichever sampling procedure, Psamp(N = n, S = σ) , we , and βGC , βAT , βAT Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 20 of 40 Computational and Systems Biology | Immunology and Inflammation Research article want. For example, recall that we originally formed the empirical conditional PDFs in (Equation 1) for each subject i by uniformly sampling from each TCR repertoire to get a total repertoire size of Ki : and Pemp(N = n | S = σ, i) = C(i)(n,σ) C(i)(σ) , Pemp(S = σ | i) = C(i)(σ) Ki , Pemp(i) = 1 I . With this, we can define a sampling procedure equivalent to this empirical joint PDF as follows: Psamp(N = n, S = σ) With this sampling procedure, := Pemp(N = n, S = σ) = Pemp(n | σ, i) Pemp(σ | i) · Pemp(i) · log Lexpected(βmotif = , βAT i,σ,n Pemp(n | σ, i) , βGC , a, b, c) Pemp(σ | i) log P(n | σ; βmotif · Pemp(i) , βGC · , βAT , a, b, c). ∑ · (6) (7) As such, each subject, instead of each observation, gets uniform treatment in the construction of the weighted likelihood. While this procedure would correct for individual subjects having different repertoire sizes, it does ˆβAT not account for gene usage differences. To avoid having these differences pollute our , , ˆβGC inference, we propose a subject- independent gene- allele- group sampling scheme. While we and could use any distribution on σ , including a uniform weight by gene allele groups, we have chosen to define: ˆβmotif Pmarg(σ) = Pemp(σ | i). 1 I ∑i We can reformulate the sampling procedure which is an empirical average per- gene- allele- group frequency such that: Psamp(N = n, S = σ) := Pemp(n | σ, i) Pmarg(σ) · Pemp(i). · (8) With this subject- independent gene sampling procedure, we can define a weighted likelihood LW(βmotif , a, b, c) such that , βGC , βAT log LW(βmotif , βGC , βAT , a, b, c) i,σ,n Pemp(n | σ, i) := log P(n | σ; βmotif · Pmarg(σ) Pemp(i) · , βGC , βAT , a, b, c) ∑ · (9) As such, each gene and each subject get uniform treatment in the construction of the weighted likelihood. From here, we can maximize this weighted log- likelihood, log LW(βmotif , where βmotif the log- probabilities βmotif To estimate each coefficient, we can solve the weighted maximum likelihood estimation problem: , and βGC , βAT , βGC , a, b, c) , to estimate is equivalent to a (log) position- weight- matrix. , βAT ˆβmotif ˆβAT , , ˆβGC ( ) = argmaxβmotif,βAT,βGC log LW(βmotif , βAT , βGC , a, b, c) (10) using the mclogit package in R. We can formulate a weighted maximum likelihood problem in a similar way for any model covariate function f containing a unique combination of parameter- specific covariate functions (Table 1). Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 21 of 40 Computational and Systems Biology | Immunology and Inflammation Research article We compare our inferred coefficients to the existing PWM model which was designed and trained using least squares (Murugan et al., 2012). When replicating this model using our methods described above (i.e. the 2×4 motif model), we note highly similar results (Figure 2—figure supplement 1). Evaluating model fit and generalizability across genes In order to evaluate the model fit and generalizability of each model, we use a variety of training and testing data sets to train each model and calculate the log loss. We will describe our general model evaluation procedure here. We describe variations of this general model evaluation procedure in Appendix 3. Let T represent a training data set and H represent a held- out testing data set. With the training set T , we can train each model of interest as described above in (Equation 10). After this model fitting, we can calculate the expected per- sequence conditional log loss of the model with given coefficients, , for a given held- out testing set, H , such that M ℓ( M | H) := = − − i,σ,n PempH (n, σ, i) i,σ,n PempH (n | σ, i) ∑ · · ∑ log P(n | σ; PempH (σ | i) M · ) PempH (i) ) M log P(n | σ; · (11) where i represents a subject, n represents a trimming length, and σ represents a gene allele group. Because we are incorporating the empirically observed frequency of each subject, trimming length, and gene allele group within each ‘held- out testing set,’ PempH (n, σ, i) , in this formulation, the expected per- sequence conditional log loss values are guaranteed to be directly comparable between held- out testing sets with varying compositions. Models that have lower expected per- sequence conditional log loss will indicate that the model has a better fit. , ∈ ˆβGC ˆβAT ˆβAT , ˆβmotif ˆβmotif { Assessing significance of model coefficients ˆβGC by maximizing the During model fitting, we estimated the model coefficients weighted likelihood function given by (Equation 9). To measure the significance of each of these model coefficients ˆβ } we want to test whether each coefficient ˆβ = 0 . To do this, we can first estimate the standard error of each inferred coefficient using a clustered bootstrap (with subject- gene pairs as the sampling unit). As such, for each bootstrap iterate, we sampled subject- gene pairs from the full V- gene training data set with replacement. Using this re- sampled data, we maximized the weighted likelihood function given by (Equation 9) to re- estimate each coefficient. We repeated this bootstrap process 1000 times and used the resulting 1000 coefficient estimates to esti- mate a standard error for each model coefficient. With this estimated standard error of each inferred model coefficient ˆβ } , we test whether ˆβ = 0 by calculating the test statistic ˆβmotif { , and ˆβAT ˆβGC , , , ∈ T( ˆβ) = ˆβ se( ˆβ) (12) and comparing T( ˆβ) to a N(0, 1) distribution to obtain each p- value. We consider the significance of each model coefficient using a Bonferroni- corrected threshold. To establish the threshold, we corrected for the total number of model coefficients being evaluated in the given model. Evaluating model coefficient variation in the context of SNPs With the motif and base- count- beyond model, we are interested in quantifying variation in model coefficients in the context of genetic variations within the gene encoding the Artemis protein that were previously identified as being associated with increasing the extent of trimming (Russell et al., 2022b). Recall that we trained this model using the model covariate function given by (Equation 2). ˆβGC by maximizing the During model fitting, we estimated the model coefficients weighted likelihood function given by (Equation 9). ˆβmotif ˆβAT , , and We have previously identified a set X of SNPs within the gene encoding the Artemis protein that are significantly associated with increasing the extent of trimming (Russell et  al., 2022b). For each SNP x {1, . . . , I} , we measure the number of minor alleles in the ∈ genotype, gix ∈ {0, 1, 2} . We are interested in whether each of the inferred model coefficients X and individual i ∈ Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 22 of 40 Computational and Systems Biology | Immunology and Inflammation Research article ˆβmotif { ˆβAT ˆβGC , , ∈ ∈ , βAT x , βGC {βmotif x ˆβ } vary in the context of genotype for each genetic variant x X . As such, for each SNP of interest, we can adapt the 1×2 motif + two- side base- count beyond model covariate function to allow for genotype- specific variation of each model coefficient by incorporating additional inter- x } to model the relationship between each model parameter action coefficients βx ∈ ˆβGC , and the SNP x genotype. We can then estimate the coefficients of this new model, , ˆβGC ˆβmotif x , as before by maximizing the weighted likelihood given by (Equation 9) using x the adapted model covariate function. We can measure the significance of each of the model coeffi- cients using the methods described in the previous section. Ultimately if a SNP- coefficient interaction ˆβGC x } is significant, we can conclude that the corresponding model coefficient term ˆβ varies significantly in the context of the genotype of SNP x X . We use this same procedure to evaluate whether each model coefficient varies in the context of each SNP of interest. ˆβAT x , and , ˆβmotif x ˆβmotif ˆβx ∈ ˆβAT , ˆβAT x , ∈ { , Acknowledgements The authors thank David Schatz and Thayer Fisher for helpful discussions regarding this paper, as well as Duncan Ralph for processing the productive IGH sequence data from Jaffe et al., 2022, and Nathaniel Spisak, Thierry Mora, and Aleksandra Walczak for sharing preprocessed data from Spisak et al., 2020. The authors would also like to thank Fred Hutch scientific computing, supported by the National Institutes of Health award S10OD028685. This work was supported by the National Institutes of Health under awards R01 AI146028, R01 AI136514, and R35 GM141457. Dr. Matsen is an Investi- gator of the Howard Hughes Medical Institute (HHMI). This article is subject to HHMI’s Open Access to Publications policy. HHMI lab heads have previously granted a nonexclusive CC BY 4.0 license to the public and a sublicensable license to HHMI in their research articles. Pursuant to those licenses, the author- accepted manuscript of this article can be made freely available under a CC BY 4.0 license immediately upon publication. Additional information Funding Funder Grant reference number Author National Institutes of Health R01 AI146028 Magdalena L Russell Philip Bradley Noah Simon Frederick A Matsen IV National Institutes of Health National Institutes of Health Howard Hughes Medical Institute R01 AI136514 Philip Bradley R35 GM141457 Philip Bradley Investigator Frederick A Matsen IV The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Author contributions Magdalena L Russell, Conceptualization, Software, Formal analysis, Validation, Investigation, Visualiza- tion, Methodology, Writing – original draft, Writing – review and editing; Noah Simon, Methodology, Writing – review and editing; Philip Bradley, Frederick A Matsen IV, Conceptualization, Supervision, Funding acquisition, Investigation, Methodology, Writing – review and editing Author ORCIDs Magdalena L Russell Philip Bradley Frederick A Matsen IV, http://orcid.org/0000-0002-1068-1968 http://orcid.org/0000-0002-0224-6464 http://orcid.org/0000-0003-0607-6025 Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 23 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.85145.sa1 Author response https://doi.org/10.7554/eLife.85145.sa2 Additional files Supplementary files • MDAR checklist Data availability The current manuscript is a computational study, so no data have been generated for this manuscript. Code is available on GitHub (copy archived at Russell et al., 2022a). Numerical data used to generate figures is available as source data for Figures 3, 4, 5, 6, and 7. The following previously published datasets were used: Year 2017 2022 Author(s) Emerson RO, DeWitt WS, Vignali M, Gravley J, Osborne EJ, Desmarais C, Klinger M, Carlson CS, Hansen JA, Rieder M, Robins HS, Hu JK Russell ML, Souquette A, Levine DM, Allen EK, Kuan G, Simon N, Balmaseda A, Gordon A, Thomas PG, Matsen FA, Bradley P Robins H, Pearson O 2015 Dataset title Dataset URL Database and Identifier https:// doi. org/ 10. 21417/ B7001Z ImmuneACCESS, 10.21417/ B7001Z Immunosequencing identifies signatures of cytomegalovirus exposure history and HLA mediated effects on the T cell repertoire Combining genotypes and T cell receptor distributions to infer genetic loci determining V(D)J recombination probabilities https://www. ncbi. nlm. nih. gov/ bioproject/ PRJNA762269 NCBI BioProject, PRJNA762269 Normal Human PBMC Deep Sequencing TCRB versus TCRG comparison https:// clients. adaptivebiotech. com/ pub/ TCRB- TCRG- comparison ImmuneACCESS, TCRB- TCRG- comparison Briney B, Inderbitzin A, Joyce C, Burton DR 2019 Commonality despite exceptional diversity in the baseline human antibody repertoire https://www. ncbi. nlm. nih. gov/ bioproject/ PRJNA406949 NCBI BioProject, PRJNA406949 2022 Functional antibodies exhibit light chain coherence https:// doi. org/ 10. 25452/ figshare. plus. 20338177 Figshare, 10.25452/ figshare. plus. 20338177 Jaffe DB, Shahi P, Adams BA, Chrisman AM, Finnegan PM, Raman N, Royall AE, Tsai F, Vollbrecht T, Reyes DS, McDonnell WJ Martin PJ, Levine DM, Storer BE, Nelson SC, Dong X, Hansen JA 2020 Recipient and donor genetic variants associated with mortality after allogeneic hematopoietic cell transplantation https://www. ncbi. nlm. nih. gov/ projects/ gap/ cgi- bin/ study. cgi? study_ id= phs001918. v1. p1 NCBI dbGaP, phs001918 References Briney B, Inderbitzin A, Joyce C, Burton DR. 2019. Commonality despite exceptional diversity in the baseline human antibody repertoire. Nature 566:393–397. DOI: https://doi.org/10.1038/s41586-019-0879-y, PMID: 30664748 Chang HHY, Watanabe G, Lieber MR. 2015. Unifying the DNA end- processing roles of the ARTEMIS nuclease. Journal of Biological Chemistry 290:24036–24050. DOI: https://doi.org/10.1074/jbc.M115.680900 Chang HHY, Lieber MR. 2016. Structure- Specific nuclease activities of ARTEMIS and the ARTEMIS: DNA- PKcs complex. Nucleic Acids Research 44:4991–4997. DOI: https://doi.org/10.1093/nar/gkw456, PMID: 27198222 Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 24 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Chang HHY, Watanabe G, Gerodimos CA, Ochi T, Blundell TL, Jackson SP, Lieber MR. 2016. Different DNA end configurations dictate which NHEJ components are most important for joining efficiency. The Journal of Biological Chemistry 291:24377–24389. DOI: https://doi.org/10.1074/jbc.M116.752329, PMID: 27703001 Chiu TP, Comoglio F, Zhou T, Yang L, Paro R, Rohs R. 2016. DNAshapeR: an R/bioconductor package for DNA shape prediction and feature encoding. Bioinformatics 32:1211–1213. DOI: https://doi.org/10.1093/ bioinformatics/btv735, PMID: 26668005 Cohen RM, Kleinstein SH, Louzoun Y. 2011. Somatic hypermutation targeting is influenced by location within the immunoglobulin V region. Molecular Immunology 48:1477–1483. DOI: https://doi.org/10.1016/j.molimm.2011. 04.002, PMID: 21592579 Cui A, Di Niro R, Vander Heiden JA, Briggs AW, Adams K, Gilbert T, O’Connor KC, Vigneault F, Shlomchik MJ, Kleinstein SH. 2016. A model of somatic hypermutation targeting in mice based on high- throughput Ig sequencing data. Journal of Immunology 197:3566–3574. DOI: https://doi.org/10.4049/jimmunol.1502263, PMID: 27707999 Dash P, Fiore- Gartland AJ, Hertz T, Wang GC, Sharma S, Souquette A, Crawford JC, Clemens EB, Nguyen THO, Kedzierska K, La Gruta NL, Bradley P, Thomas PG. 2017. Quantifiable predictive features define epitope- specific T cell receptor repertoires. Nature 547:89–93. DOI: https://doi.org/10.1038/nature22383, PMID: 28636592 Dominski Z. 2007. Nucleases of the metallo- beta- lactamase family and their role in DNA and RNA metabolism. Critical Reviews in Biochemistry and Molecular Biology 42:67–93. DOI: https://doi.org/10.1080/ 10409230701279118, PMID: 17453916 Dunn- Walters DK, Dogan A, Boursier L, MacDonald CM, Spencer J. 1998. Base- Specific sequences that bias somatic hypermutation deduced by analysis of out- of- frame human igvh genes. Journal of Immunology 160:2360–2364 PMID: 9498777. Elhanati Y, Sethna Z, Marcou Q, Callan CG, Mora T, Walczak AM. 2015. Inferring processes underlying B- cell repertoire diversity. Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences 370:20140243. DOI: https://doi.org/10.1098/rstb.2014.0243, PMID: 26194757 Emerson RO, DeWitt WS, Vignali M, Gravley J, Hu JK, Osborne EJ, Desmarais C, Klinger M, Carlson CS, Hansen JA, Rieder M, Robins HS. 2017. Immunosequencing identifies signatures of cytomegalovirus exposure history and HLA- mediated effects on the T cell repertoire. Nature Genetics 49:659–665. DOI: https://doi.org/ 10.1038/ng.3822, PMID: 28369038 Feeney AJ, Victor KD, Vu K, Nadel B, Chukwuocha RU. 1994. Influence of the V (D) J recombination mechanism on the formation of the primary T and B cell repertoires. Seminars in Immunology 6:155–163. DOI: https://doi. org/10.1006/smim.1994.1021, PMID: 7948955 Feng J, Shaw DA, Minin VN, Simon N, Matsen FA. 2019. Survival analysis of DNA mutation motifs with penalized proportional hazards. The Annals of Applied Statistics 13:1268–1294. DOI: https://doi.org/10.1214/18- aoas1233, PMID: 33214798 Fugmann SD, Lee AI, Shockett PE, Villey IJ, Schatz DG. 2000. The RAG proteins and V (D) J recombination: complexes, ends, and transposition. Annual Review of Immunology 18:495–527. DOI: https://doi.org/10.1146/ annurev.immunol.18.1.495, PMID: 10837067 Gauss GH, Lieber MR. 1996. Mechanistic constraints on diversity in human V (D) J recombination. Molecular and Cellular Biology 16:258–269. DOI: https://doi.org/10.1128/MCB.16.1.258, PMID: 8524303 Gellert M. 1994. Dna double- strand breaks and hairpins in V (D) J recombination. Seminars in Immunology 6:125–130. DOI: https://doi.org/10.1006/smim.1994.1018, PMID: 7948952 Gerodimos CA, Chang HHY, Watanabe G, Lieber MR. 2017. Effects of DNA end configuration on XRCC4- DNA ligase IV and its stimulation of Artemis activity. The Journal of Biological Chemistry 292:13914–13924. DOI: https://doi.org/10.1074/jbc.M117.798850, PMID: 28696258 Gilfillan S, Dierich A, Lemeur M, Benoist C, Mathis D. 1993. Mice lacking TdT: mature animals with an immature lymphocyte repertoire. Science 261:1175–1178. DOI: https://doi.org/10.1126/science.8356452, PMID: 8356452 Gu J, Li S, Zhang X, Wang LC, Niewolik D, Schwarz K, Legerski RJ, Zandi E, Lieber MR. 2010. Dna- Pkcs regulates a single- stranded DNA endonuclease activity of ARTEMIS. DNA Repair 9:429–437. DOI: https://doi.org/10. 1016/j.dnarep.2010.01.001, PMID: 20117966 Jackson KJL, Gaeta B, Sewell W, Collins AM. 2004. Exonuclease activity and P nucleotide addition in the generation of the expressed immunoglobulin repertoire. BMC Immunology 5:19. DOI: https://doi.org/10.1186/ 1471-2172-5-19, PMID: 15345030 Jaffe DB, Shahi P, Adams BA, Chrisman AM, Finnegan PM, Raman N, Royall AE, Tsai F, Vollbrecht T, Reyes DS, Hepler NL, McDonnell WJ. 2022. Functional antibodies exhibit light chain coherence. Nature 611:352–357. DOI: https://doi.org/10.1038/s41586-022-05371-z, PMID: 36289331 Jose D, Datta K, Johnson NP, von Hippel PH. 2009. Spectroscopic studies of position- specific DNA `` breathing'' fluctuations at replication forks and primer- template junctions. PNAS 106:4231–4236. DOI: https://doi.org/10. 1073/pnas.0900803106 Kallenbach S, Doyen N, Fanton d’Andon M, Rougeon F. 1992. Three lymphoid- specific factors account for all junctional diversity characteristic of somatic assembly of T- cell receptor and immunoglobulin genes. PNAS 89:2799–2803. DOI: https://doi.org/10.1073/pnas.89.7.2799 Komori T, Okada A, Stewart V, Alt FW. 1993. Lack of N regions in antigen receptor variable region genes of tdt- deficient lymphocytes. Science 261:1171–1175. DOI: https://doi.org/10.1126/science.8356451, PMID: 8356451 Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 25 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Li S, Chang HH, Niewolik D, Hedrick MP, Pinkerton AB, Hassig CA, Schwarz K, Lieber MR. 2014. Evidence that the DNA endonuclease ARTEMIS also has intrinsic 5′-exonuclease activity. Journal of Biological Chemistry 289:7825–7834. DOI: https://doi.org/10.1074/jbc.M113.544874 Lu H, Schwarz K, Lieber MR. 2007. Extent to which hairpin opening by the ARTEMIS: DNA- PKcs complex can contribute to junctional diversity in V (D) J recombination. Nucleic Acids Research 35:6917–6923. DOI: https:// doi.org/10.1093/nar/gkm823 Lu H, Shimazaki N, Raval P, Gu J, Watanabe G, Schwarz K, Swanson PC, Lieber MR. 2008. A biochemically defined system for coding joint formation in V (D) J recombination. Molecular Cell 31:485–497. DOI: https:// doi.org/10.1016/j.molcel.2008.05.029 Ma Y, Pannicke U, Schwarz K, Lieber MR. 2002. Hairpin opening and overhang processing by an artemis/ DNA- dependent protein kinase complex in nonhomologous end joining and V (D) J recombination. Cell 108:781–794. DOI: https://doi.org/10.1016/S0092-8674(02)00671-2 Ma Y, Schwarz K, Lieber MR. 2005. The ARTEMIS: DNA- PKcs endonuclease cleaves DNA loops, flaps, and gaps. DNA Repair 4:845–851. DOI: https://doi.org/10.1016/j.dnarep.2005.04.013 Marcou Q, Mora T, Walczak AM. 2018. High- Throughput immune repertoire analysis with igor. Nature Communications 9:561. DOI: https://doi.org/10.1038/s41467-018-02832-w, PMID: 29422654 Martin PJ, Levine DM, Storer BE, Nelson SC, Dong X, Hansen JA. 2020. Recipient and donor genetic variants associated with mortality after allogeneic hematopoietic cell transplantation. Blood Advances 4:3224–3233. DOI: https://doi.org/10.1182/bloodadvances.2020001927 Moshous D, Callebaut I, de Chasseval R, Corneo B, Cavazzana- Calvo M, Le Deist F, Tezcan I, Sanal O, Bertrand Y, Philippe N, Fischer A, de Villartay JP. 2001. Artemis, a novel DNA double- strand break repair/V (D) J recombination protein, is mutated in human severe combined immune deficiency. Cell 105:177–186. DOI: https://doi.org/10.1016/s0092-8674(01)00309-9, PMID: 11336668 Murugan A, Mora T, Walczak AM, Callan CG. 2012. Statistical inference of the generation probability of T- cell receptors from sequence repertoires. PNAS 109:16161–16166. DOI: https://doi.org/10.1073/pnas. 1212755109, PMID: 22988065 Nadel B, Feeney AJ. 1995. Influence of coding- end sequence on coding- end processing in V (D) J recombination. Journal of Immunology 155:4322–4329 PMID: 7594591. Nadel B, Feeney AJ. 1997. Nucleotide deletion and P addition in V (D) J recombination: a determinant role of the coding- end sequence. Molecular and Cellular Biology 17:3768–3778. DOI: https://doi.org/10.1128/MCB. 17.7.3768, PMID: 9199310 Pettinati I, Brem J, Lee SY, McHugh PJ, Schofield CJ. 2016. The chemical biology of human metallo-β-lactamase fold proteins. Trends in Biochemical Sciences 41:338–355. DOI: https://doi.org/10.1016/j.tibs.2015.12.007, PMID: 26805042 Ralph DK, Matsen FA. 2016. Consistency of VDJ rearrangement and substitution parameters enables accurate B cell receptor sequence annotation. PLOS Computational Biology 12:e1004409. DOI: https://doi.org/10.1371/ journal.pcbi.1004409, PMID: 26751373 Robins HS, Srivastava SK, Campregher PV, Turtle CJ, Andriesen J, Riddell SR, Carlson CS, Warren EH. 2010. Overlap and effective size of the human CD8+ T cell receptor repertoire. Science Translational Medicine 2:47. DOI: https://doi.org/10.1126/scitranslmed.3001442, PMID: 20811043 Robins H, Pearson O. 2015. Normal human PBMC, deep sequencing, TCRB vs TCRG comparison. https://clients. adaptivebiotech.com/pub/TCRB-TCRG-comparison [Accessed October 25, 2022]. Rogozin IB, Kolchanov NA. 1992. Somatic hypermutagenesis in immunoglobulin genes. II. Influence of neighbouring base sequences on mutagenesis. Biochimica et Biophysica Acta 1171:11–18. DOI: https://doi. org/10.1016/0167-4781(92)90134-l, PMID: 1420357 Russell ML, Simon N, Bradley P, Matsen FA. 2022a. Mechanistic- trimming. swh:1:rev:2ba723b3bd4a354fe78f677230b8a6dfb506422d. Software Heritage. https://archive. softwareheritage.org/swh:1:dir:85c0bff261bbd6cb74debec11141d2f4195907f1;origin=https://github.com/ magdalenarussell/mechanistic-trimming;visit=swh:1:snp:d9194a4fb82240b9842906aed109cca46c80ae4f; anchor=swh:1:rev:2ba723b3bd4a354fe78f677230b8a6dfb506422d Russell ML, Souquette A, Levine DM, Schattgen SA, Allen EK, Kuan G, Simon N, Balmaseda A, Gordon A, Thomas PG, Matsen FA, Bradley P. 2022b. Combining genotypes and T cell receptor distributions to infer genetic loci determining V (D) J recombination probabilities. eLife 11:e73475. DOI: https://doi.org/10.7554/ eLife.73475, PMID: 35315770 Schatz DG, Swanson PC. 2011. V (D) J recombination: mechanisms of initiation. Annual Review of Genetics 45:167–202. DOI: https://doi.org/10.1146/annurev-genet-110410-132552, PMID: 21854230 Sethna Z, Elhanati Y, Callan CG, Walczak AM, Mora T. 2019. OLGA: fast computation of generation probabilities of B- and T- cell receptor amino acid sequences and motifs. Bioinformatics 35:2974–2981. DOI: https://doi.org/ 10.1093/bioinformatics/btz035, PMID: 30657870 Sethna Z, Isacchini G, Dupic T, Mora T, Walczak AM, Elhanati Y. 2020. Population variability in the generation and selection of T- cell repertoires. PLOS Computational Biology 16:e1008394. DOI: https://doi.org/10.1371/ journal.pcbi.1008394, PMID: 33296360 Slabodkin A, Chernigovskaya M, Mikocziova I, Akbar R, Scheffer L, Pavlović M, Bashour H, Snapkov I, Mehta BB, Weber CR, Gutierrez- Marcos J, Sollid LM, Haff IH, Sandve GK, Robert PA, Greiff V. 2021. Individualized VDJ Recombination Predisposes the Available Ig Sequence Space. bioRxiv. DOI: https://doi.org/10.1101/2021.04. 19.440409 Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 26 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Spisak N, Walczak AM, Mora T. 2020. Learning the heterogeneous hypermutation landscape of immunoglobulins from high- throughput repertoire data. Nucleic Acids Research 48:10702–10712. DOI: https://doi.org/10.1093/ nar/gkaa825, PMID: 33035336 Srivastava SK, Robins HS. 2012. Palindromic nucleotide analysis in human T cell receptor rearrangements. PLOS ONE 7:e52250. DOI: https://doi.org/10.1371/journal.pone.0052250, PMID: 23284955 Tsai AG, Engelhart AE, Hatmal MM, Houston SI, Hud NV, Haworth IS, Lieber MR. 2009. Conformational variants of duplex DNA correlated with cytosine- rich chromosomal fragile sites. The Journal of Biological Chemistry 284:7157–7164. DOI: https://doi.org/10.1074/jbc.M806866200, PMID: 19106104 Vander Heiden JA, Yaari G, Uduman M, Stern JNH, O’Connor KC, Hafler DA, Vigneault F, Kleinstein SH. 2014. Presto: a toolkit for processing high- throughput sequencing RAW reads of lymphocyte receptor repertoires. Bioinformatics 30:1930–1932. DOI: https://doi.org/10.1093/bioinformatics/btu138 Wei L, Chahwan R, Wang S, Wang X, Pham PT, Goodman MF, Bergman A, Scharff MD, MacCarthy T. 2015. Overlapping hotspots in CDRs are critical sites for V region diversification. PNAS 112:E728–E737. DOI: https:// doi.org/10.1073/pnas.1500788112 Yaari G, Vander Heiden JA, Uduman M, Gadala- Maria D, Gupta N, Stern JNH, O’Connor KC, Hafler DA, Laserson U, Vigneault F, Kleinstein SH. 2013. Models of somatic hypermutation targeting and substitution based on synonymous mutations from high- throughput immunoglobulin sequencing data. Frontiers in Immunology 4:358. DOI: https://doi.org/10.3389/fimmu.2013.00358, PMID: 24298272 Ye J, Ma N, Madden TL, Ostell JM. 2013. IgBLAST: an immunoglobulin variable domain sequence analysis tool. Nucleic Acids Research 41:W34–W40. DOI: https://doi.org/10.1093/nar/gkt382 Yosaatmadja Y, Baddock HT, Newman JA, Bielinski M, Gavard AE, Mukhopadhyay SMM, Dannerfjord AA, Schofield CJ, McHugh PJ, Gileadi O. 2021. Structural and mechanistic insights into the ARTEMIS endonuclease and strategies for its inhibition. Nucleic Acids Research 49:9310–9326. DOI: https://doi.org/10.1093/nar/ gkab693 Zhao B, Rothenberg E, Ramsden DA, Lieber MR. 2020. The molecular basis and disease relevance of non- homologous DNA end joining. Nature Reviews. Molecular Cell Biology 21:765–781. DOI: https://doi.org/10. 1038/s41580-020-00297-8, PMID: 33077885 Zhou T, Yang L, Lu Y, Dror I, Dantas Machado AC, Ghane T, Di Felice R, Rohs R. 2013. DNAshape: a method for the high- throughput prediction of DNA structural features on a genomic scale. Nucleic Acids Research 41:W56–W62. DOI: https://doi.org/10.1093/nar/gkt437, PMID: 23703209 Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 27 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Appendix 1 Extended notation Appendix 1—table 1. Summary of all notation used in our modeling. Variable General notation I i Ki k S σ σV σJ N n L U C(i)(σ) C(i)(n, σ) N′ Pemp(N = n | S = σ, i) Motif parameter- specific notation a b {σ(n + j)} b 1 − a j= − βmotif js βmotif Description Set of all individuals Index for an individual in the set I of all individuals Total number of TCRs in the repertoire of individual i Index of a sequence in the TCR repertoire of individual i Random variable that represents the gene sequence General notation for a gene- allele- group sequence oriented 5’-to- 3’ V- gene- allele- group sequence (‘top’ strand oriented 5’-to- 3’) J- gene- allele- group sequence (‘bottom’ strand oriented 5’-to- 3’) Random variable that represents the number of deleted nucleotides Number of deleted nucleotides from the 3’-side of a gene sequence Lower bound of ‘reasonable’ trimming amounts, we have defined L = 2 Upper bound of ‘reasonable’ trimming amounts, we have defined U = 14 The number of TCRs that use gene allele group σ in the sampled repertoire of individual i The number of TCRs that have gene allele group σ and n nucleotides deleted in the sampled repertoire of individual i Set of all ‘reasonable’ trimming amounts; N′ = {2, . . .14} Empirical conditional probability density function (Equation 1) Non- negative integer value that represents the number of nucleotides 5’ of the trimming site to be included in the ‘trimming motif’ Non- negative integer value that represents the number of nucleotides 3’ of the trimming site to be included in the ‘trimming motif’ ‘Trimming motif’ sequence (Equation 13) (Log) position weight matrix coefficient for trimming motif position j 1} and nucleotide s {A, T, C, G} a, . . . , b { − ∈ − Set of all motif coefficients βmotif j − ∈ js 1} and nucleotide s a, . . . , b − { ∈ ∈ for all positions {A, T, C, G} , a, b) f(n, σ; βmotif Base- count- beyond parameter- specific notation Motif- specific covariate function (Equation 14) c CAT (x) Non- negative integer value that represents the number of nucleotides 5’ of the trimming site to be included in the 5’ base- count- beyond the ‘trimming motif’ Count of nucleotides that are A or T in an arbitrary sequence x (x) CGC Appendix 1—table 1 Continued on next page Count of nucleotides that are G or C in an arbitrary sequence x Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 28 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Appendix 1—table 1 Continued Variable Description seq5(n, σ, a, c) seq3(n, σ, b) βAT 5 and βAT 3 βAT βGC 5 and βGC 3 βGC , βGC f(n, σ; βAT DNA- shape parameter- specific notation , a, b, c) seqexpd(n, σ, a, b) E W P R H shapeu (j, seqexpd(n, σ, a, b)) shapev (d, seqexpd(n, σ, a, b)) βshape uj vd βshape βshape f(n, σ; βshape Length parameter- specific notation , a, b) βldist f(n, σ; βldist Modeling notation ) The nucleotide sequence 5’ of the trimming site, beyond the ‘trimming motif’ (Equation 15) The nucleotide sequence 3’ of the trimming site, beyond the ‘trimming motif’ (Equation 16) Base- count- beyond model coefficients for the 5’ and 3’ sequence base- counts of A and T nucleotides beyond the trimming motif Set of AT-base- count- beyond model coefficients (includes βAT βAT 3 ) Base- count- beyond model coefficients for the 5’ and 3’ sequence base- counts of G and C nucleotides beyond the trimming motif 5 and Set of GC-base- count- beyond model coefficients (includes βGC βGC 3 ) 5 and Base- count- beyond- specific covariate function (Equation 14) ‘Expanded trimming sequence window’ (Equation 18); consists of the ‘trimming motif’ sequence extended by 2 nucleotides in both the 5’ and 3’ direction Nucleotide electrostatic potential Nucleotide minor groove width Nucleotide propeller twist Di- nucleotide roll Di- nucleotide helical twist { ∈ ∈ − − {E, W, P} for the nucleotide at 1} within the ‘expanded trimming Measure of nucleotide shape u a, . . . , b position j sequence window’ seqexpd(n, σ, a, b) Measure of di- nucleotide shape v ∈ position d sequence window’ seqexpd(n, σ, a, b) DNA- shape coefficients for nucleotide shape type u {E, W, P} and ‘expanded trimming sequence window’ nucleotide position j {R, H} for the di- nucleotide at 1} within the ‘expanded trimming a + 1, . . . , b a, . . . , b 1} − − ∈ ∈ { { ∈ − − DNA- shape coefficients for di- nucleotide shape type v and ‘expanded trimming sequence window’ di- nucleotide position d a + 1, . . . , b {R, H} 1} ∈ { ∈ − − Set of all nucleotide and di- nucleotide DNA- shape coefficients DNA- shape- specific covariate function (Equation 19) Length specific model coefficient Length- specific covariate function f(n, σ; βmotif , βAT , βGC , a, b, c) Example model covariate function including motif and base- count- beyond model parameters (Equation 2) , βGC P(n | σ; βmotif Appendix 1—table 1 Continued on next page , a, b, c) , β AT Conditional logit model formulation using the motif and base- count- beyond model covariate function (Equation 3) Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 29 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Appendix 1—table 1 Continued Variable Description log L(βmotif , βAT , βGC , a, b, c) Psamp(N = n, S = σ) log Lexpected(βmotif , βAT , βGC , a, b, c) log Lemp(βmotif , βAT , βGC , a, b, c) Pmarg(σ) , βGC log LW(βmotif Model evaluation notation , βAT , a, b, c) M V J H P(H) ℓ( M | H) E[ℓ( M )] RMSE(σ, , V) M Coefficient evaluation notation T( ˆβ) X gix Aggregated log- likelihood for the conditional logit model; this likelihood function is un- weighted (Equation 4) and gives every observation uniform treatment in the likelihood Sampling procedure for the construction of the expected likelihood Expected log- likelihood for the conditional logit model; this likelihood function (Equation 5) weights each observation by its sampling probability, Psamp(N = n, S = σ) Expected log- likelihood for the conditional logit model; this likelihood function (Equation 7) weights each observation by its sampling probability from the empirical joint PDF (Equation 6) Empirical average per- gene- allele- group frequency used in formulating a subject- independent gene sampling procedure (Equation 8) Expected log- likelihood for the conditional logit model; this likelihood function (Equation 9) weights each observation using a subject- independent gene sampling procedure (Equation 8) An arbitrary model trained on a specified training data set Full V- gene data set Full J- gene data set Arbitrary held- out data set Probability of the arbitrary held- out data set (Equation 21) Expected per- sequence conditional log loss (Equation 11) of a evaluated on a data set H trained model M Expected per- sequence conditional log loss across 20 random held- out data sets (Equation 22) Per- gene mean squared error (Equation 23) for a gene σ using a trained using the V- gene training data set V model M Test statistic (Equation 12) for evaluating the significance of a single inferred coefficient ˆβ Set of SNPs within the gene encoding the Artemis protein that were previously identified to be associated with increasing the extent of trimming (Russell et al., 2022b) Number of minor alleles in the genotype of an individual i x X ∈ I for SNP ∈ {βmotif x , βAT x , βGC x } Set of interaction coefficients between each model parameter and the SNP x genotype Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 30 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Appendix 2 Extended parameter description Defining the ‘trimming motif’ and position-weight-matrix weight for a given gene and trimming site Appendix 2—figure 1. Summary of trimming motif parameters. Let a = 2 and b = 4 . The 6- nucleotide trimming motif given by (Equation 13) is shown in the orange box and the trimming site is shown by the vertical orange line. An arbitrary gene sequence is highlighted in gray and the two possible P- nucleotides are highlighted in purple. (A) For n = 5 , the 6- nucleotide trimming motif will not contain P- nucleotides. (B) For n = 3 , the 6- nucleotide trimming motif will contain one P- nucleotide. (C) For n = 1 , the trimming motif will contain two P- nucleotides and will be ‘incomplete’ (contain less than 6 nucleotides). Existing probabilistic models of nucleotide trimming using repertoire sequencing data have shown that the local nucleotide context around the trimming site, which we refer to as the ‘trimming motif,’ do a surprisingly good job of predicting the distribution of trimming lengths for a variety of genes (Murugan et  al., 2012). This simple PWM model uses a trimming motif containing 2 nucleotides 5’ of the trimming site and 4 nucleotides 3’ of the trimming site to predict the probability of trimming at that site. In practice, we can define the trimming motif to be any size. Let a and b be non- negative integer values that represent the number of nucleotides 5’ and 3’ of the trimming site, respectively. Together, these a + b nucleotides will compose the trimming motif. For a gene- allele- group sequence σ and a number of deleted nucleotides n , let σ(n + j) represent the nucleotide a, . . . , b identity at the trimming motif position j 1} where positions j < 0 represent motif { − positions 5’ of the trimming site and positions j 0 represent motif positions 3’ of the trimming site. ≥ As such, the trimming motif sequence is given by − ∈ {σ(n + j)} b 1 a. − j= − (13) Depending on n , this trimming motif may or may not include P- nucleotides. For example, for n b , the b 3’ trimming motif nucleotides will include the b deleted gene sequence nucleotides 3’ of the trimming site (and no P- nucleotides) (Appendix  2—figure 1A). Since we are assuming that the initial hairpin nick occurs at the +2 position, there will be two P- nucleotides present in the 5’-to- 3’ gene sequence. For b n < b , where the 2 represents the total P- nucleotide count in the full sequence, P- nucleotides will be included in the trimming motif sequence. Specifically, the b total 3’ trimming motif nucleotides will include b n P- nucleotides and n deleted gene sequence nucleotides (Appendix 2—figure 1B, C). Likewise, as a result of the +2 hairpin nick position assumption, TCRs that have n < b 2 will not have a full, (a + b) - length nucleotide trimming motif (Appendix 2—figure 1C). For these ‘off- the- end’ motif cases, we assign zero influence to the missing nucleotides during model fitting. − ≤ − − ≥ 2 Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 31 of 40 Computational and Systems Biology | Immunology and Inflammation Research article With this trimming motif, let βmotif − a, . . . , b motif position j − position- weight- matrix weight ∈ { js be a (log) position- weight- matrix coefficient for trimming {A, T, C, G} . We can define an un- normalized 1} and nucleotide s ∈ f(n, σ; βmotif , a, b) := b 1 − a ∑j= − βmotif jσ(n+j) (14) that will serve as a motif- specific model covariate function in subsequent modeling. As described above, since we are considering ‘off- the- end’ motif cases, σ(n + j) represent the nucleotide identity at sequence position j where positions j < 0 represent sequence positions 5’ of the trimming site and positions j 0 represent sequence positions 3’ of the trimming site. ≥ AT and GC base-count-beyond the trimming motif For an arbitrary sequence x , we can count the number of AT and GC nucleotides within the sequence as and respectively. CAT (x) = CA (x) +C T (x) CGC (x) = CG (x) +C C (x), Because the count of AT or GC nucleotides within the sequences 5’ and 3’ of the trimming site may influence the probability of trimming differently, we will calculate the counts separately. We will not include nucleotides that were already included in the motif parameterization. As above, for a gene- allele- group sequence σ and a number of deleted nucleotides n , let σ(n + j) represent the nucleotide identity at sequence position j where positions j < 0 represent sequence positions 5’ of the trimming site and positions j 0 represent sequence positions 3’ of the trimming site. Let c be a non- negative integer value that represents the number of nucleotides 5’ of the trimming site that will be included in the 5’-nucleotide counts (Appendix 2—figure 2). Recall that a is a non- negative integer value that represents the number of nucleotides 5’ of the trimming site that are included in the ‘trimming motif’ described in the previous section. As such, the nucleotide sequence 5’ of the trimming site, beyond the ‘trimming motif,’ is given by ≥ seq5(n, σ, a, c) = {σ(n + j)}(a+c) j=(a+1). (15) Within this sequence seq5(n, σ, a, c) , we can count the number of AT and GC nucleotides as CAT (seq5(n, σ, a, c)) = CA (seq5(n, σ, a, c)) + CT (seq5(n, σ, a, c)) and respectively CGC (seq5(n, σ, a, c)) = CG (seq5(n, σ, a, c)) + CC (seq5(n, σ, a, c)), To count the number of AT and GC nucleotides in the sequence 3’ of the trimming site, we will include all nucleotides located 3’ of the trimming site that are beyond the ‘trimming motif.’ However, because we are interested in using GC nucleotide content in both directions of the wider sequence as a proxy for the capacity for sequence- breathing and since sequence- breathing is only relevant for nucleotides that are paired, we will not include the nucleotides within the 3’ single- stranded- overhang when counting 3’ AT and GC nucleotides (Appendix  2—figure 2). Since we are assuming that the initial hairpin nick occurs at the +2 position leading to a 4- nucleotide- long 3’ single- stranded- overhang, for n > 2 , the nucleotide sequence 3’ of the trimming site, beyond the ‘trimming motif,’ is given by Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 32 of 40 Computational and Systems Biology | Immunology and Inflammation Research article seq3(n, σ, b) =   b {σ(n + j)}− j=3 n − {} if (n if (n − − b 3) ≥ 3) < b (16)  where b is a non- negative integer value that represents the number of nucleotides 3’ of the trimming 3) < b , all site that are included in the ‘trimming motif‘ described in the previous section. For (n nucleotides 3’ of the trimming site are considered single- stranded and, thus, no nucleotides will be included in the sequence used to calculate the AT and GC base- counts (Appendix 2—figure 2C). Within this sequence seq3(n, σ, b) , we can count the number of AT and GC nucleotides as − CAT (seq3(n, σ, b)) = CA (seq3(n, σ, b)) + CT (seq3(n, σ, b)) and CGC (seq3(n, σ, b)) = CG (seq3(n, σ, b)) + CC (seq3(n, σ, b)), respectively. As defined, these GC and AT base- counts for the 3’ sequence are dependent on sequence length and provide a parameterization of both GC nucleotide content in both directions of the wider sequence and length. With these 5’ and 3’ base counts, we can define βAT 3 to be base- count- beyond model coefficients for 5’ and 3’ sequence base- counts of AT and GC beyond the ‘trimming motif,’ respectively. With these coefficients, we can define a base- count- beyond covariate function for each trimming site n and gene σ : 5 , and βGC 5 , βAT 3 , βGC f(n, σ; βAT , βGC CAT , a, b, c) := βAT 5 · +βGC 5 · (seq5(n, σ, a, c)) + βAT 3 · (seq5(n, σ, a, c)) + βGC CGC 3 · CAT (seq3(n, σ, b)) CGC (seq3(n, σ, b)). (17) Appendix 2—figure 2. Summary of base- count parameters. Let a = 1 , b = 2 , and c = 5 . An arbitrary gene sequence is highlighted in gray and the two possible P- nucleotides are highlighted in purple. The trimming site is shown by the vertical orange line and the ‘trimming motif,’ as defined in (Equation 13), is shown by the orange box. The c nucleotides included in the count of AT and GC nucleotides 5’ of the trimming site, beyond the ‘trimming motif,’ are expressed by (Equation 15) and are shown in the green box. The nucleotides included in the count of AT and GC nucleotides 3’ of the trimming site, beyond the ‘trimming motif,’ are expressed by (Equation 16) and are shown in the yellow box. As described in the text, we are assuming that the initial hairpin nick occurs at the +2 position leading to a 4- nucleotide- long 3’ single- stranded- overhang. We exclude these single- stranded nucleotides in the 3’-base- count- beyond sequence. In this figure, the 4 nucleotides nearest to the 3’ side of each sequence (this includes the two P- nucleotides and the two 3’-most gene sequence nucleotides) are considered single- stranded and will not be included in the 3’-base- count- beyond sequence. (A) For n = 6 , 2 nucleotides 3’ of the trimming site will be used in the 3’ sequence base- counts. (B) For n = 5 , 1 nucleotide 3’ of the trimming site will be used in the 3’ sequence base- counts. (C) For n = 4 , all nucleotides 3’ of the trimming site are considered single- stranded and, thus, no nucleotides will be used for the 3’ sequence base- counts. Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 33 of 40 Computational and Systems Biology | Immunology and Inflammation Research article DNA-shape around the trimming site Methods have been previously developed to estimate DNA- shape features at a single- nucleotide position using the sequence context of 2 neighboring nucleotides on both sides of the nucleotide of interest (Zhou et al., 2013; Chiu et al., 2016). As such, these methods use a sliding- pentamer model, centered at each nucleotide of interest, to derive the structural features of nucleotides within a sequence window of any length. These structural features include estimations of electrostatic potential (E), minor groove width (W), and propeller twist (P) for each nucleotide in the sequence window and estimations of roll (R) and helical twist (H) for each di- nucleotide pair in the sequence window. For simplicity, we will use the term ‘DNA- shape parameters’ to refer to all five of these structural features. For our purposes, we can define a ‘trimming sequence window’ of size a + b , as introduced in the ‘trimming motif’ section with (Equation 13), where a and b are non- negative integer values that represent the number of nucleotides 5’ and 3’ of the trimming site, respectively. In order to estimate the DNA- shape for all nucleotides within this window, we will expand the ‘trimming sequence window’ by 2 nucleotides on both sides such that there are a + 2 nucleotides 5’ and b + 2 nucleotides 3’ of the trimming site included in an ‘expanded trimming sequence window.’ For a gene- allele- group sequence σ and a number of deleted nucleotides n , let σ(n + j) represent the nucleotide identity at the ‘expanded trimming sequence window’ position j 1} where positions j < 0 represent expanded trimming sequence window positions 5’ of the trimming site and positions j 0 represent expanded trimming sequence window positions 3’ of the trimming site. As such, the expanded trimming sequence window is given by (a + 2), . . . , (b + 2) − ≥ − ∈ { seqexpd(n, σ, a, b) := {σ(n + j)}(b+2) 1 (a+2). − j= − (18) ≥ ≤ − For each nucleotide position j Depending on n , this expanded trimming sequence window may or may not include P- nucleotides. (b + 2) , the (b + 2) 3’ expanded trimming sequence window nucleotides will For example, for n include the (b + 2) deleted gene sequence nucleotides 3’ of the trimming site (and no P- nucleotides) (Appendix  2—figure 3A). For b n < b + 2 , the (b + 2) 3’ expanded trimming sequence window nucleotides will include (b + 2) n P- nucleotides and n deleted gene sequence nucleotides (Appendix 2—figure 3B). Since we are assuming that the initial hairpin nick occurs at the +2 position, TCRs that have n < b will not have a full, (a + b + 4) - length nucleotide expanded trimming sequence window (Appendix  2—figure 3C). The sliding- pentamer model (Zhou et  al., 2013; Chiu et  al., 2016) requires a full pentamer for estimating the DNA- shape of each base of interest, and, thus, for these ‘off- the- end’ expanded trimming sequence window cases, we cannot estimate DNA- shape parameters for all nucleotides within the trimming sequence window. As such, when estimating DNA- shape parameters, we must choose b such that b n for all trimming lengths n in the data set. 1} within the expanded trimming sequence window seqexpd(n, σ, a, b) , we can estimate the nucleotide electrostatic potential, shapeE (j, seqexpd(n, σ, a, b)) , minor groove width, shapeW (j, seqexpd(n, σ, a, b)) . We then standardize the estimated values for each shape type. We can define βshape to be a nucleotide uj shape model coefficient for nucleotide shape type u {E, W, P} and trimming sequence window nucleotide position j 1} be the location of each di- nucleotide in the trimming sequence window such that d = 0 represents the location of the trimming site, d < 0 represents di- nucleotide positions 5’ of the trimming site, and d > 0 represents di- nucleotide positions 1} within the expanded trimming 3’ of the trimming site. For each di- nucleotide d sequence window seqexpd(n, σ, a, b) , we can estimate the di- nucleotide roll, shapeR (d, seqexpd(n, σ, a, b)) and helical twist, shapeH (d, seqexpd(n, σ, a, b)) . As above, we then standardize the estimated values for each di- nucleotide shape type. We can define βshape to be a di- nucleotide shape model coefficient for di- nucleotide shape type v {R, H} and trimming sequence window di- nucleotide position 1} . We use the R package DNAshapeR (Chiu et  al., 2016) to estimate these d DNA- shape parameters for each trimming sequence window. With these standardized DNA- shape estimates, we can define a DNA- shape covariate function for each trimming site n and gene σ (j, seqexpd(n, σ, a, b)) , and propeller twist, shapeP ∈ a + 1, ..., b a + 1, ..., b a + 1, ..., b 1} . Let d a, ..., b a, ..., b ≤ − − − − − − − − − − ∈ ∈ ∈ ∈ ∈ ∈ vd { { { { { Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 34 of 40 Computational and Systems Biology | Immunology and Inflammation Research article f(n, σ; βshape , a, b) := b − 1 {E,W,P} a ∑u ∑j= − ∈ b 1 − + βshape uj · shapeu (j, seqexpd(n, σ, a, b)) βshape vd · shapev (d, seqexpd(n, σ, a, b)). (19) ∑d= − a+1 ∑v {R,H} ∈ Appendix 2—figure 3. Summary of DNA- shape parameters. Let a = 1 and b = 2 . The 3- nucleotide trimming sequence window is shown in the orange box and the trimming site is shown by the vertical orange line. The 7- nucleotide expanded trimming sequence window is represented by the pink boxes in addition to the original trimming sequence window orange box. An arbitrary gene sequence is highlighted in gray and the two possible P- nucleotides are highlighted in purple. (A) For n = 5 , both the 7- nucleotide expanded trimming sequence window and the original 3- nucleotide trimming sequence window will not contain P- nucleotides. (B) For n = 3 , the 7- nucleotide expanded trimming sequence window will contain one P- nucleotide and the original 3- nucleotide trimming sequence window will not contain P- nucleotides. (C) For n = 1 , the 7- nucleotide expanded trimming sequence window will be ‘incomplete’ (contain less than 7 nucleotides), and thus, will be invalid for estimating DNA- shape for the nucleotides within the original trimming sequence window. Length We can think of the trimming amount n as a measure of the sequence- independent length from the end of the gene for each gene and trimming site, and define βldist to be a length model coefficient. As such, we can define a length covariate function for each trimming site n f(n, σ; βldist ) := βldist n. · (20) Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 35 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Appendix 3 Extended model validation methods Calculating the expected per-sequence conditional log loss across the full V-gene training data set With the full V- gene training set, we can train each model of interest as described above in (Equation 10) to obtain a trained model . After this model fitting, we can calculate the expected per- sequence conditional log loss of the model, , for the full V- gene training data set, V , using the procedure described above in (Equation 11). Here, we use the full V- gene data set as both the training data set and the testing data set. Models that have lower expected per- sequence conditional log loss on the V- gene training data set will indicate that the model has a better fit. Model evaluation using held- out testing sets, as described below, is required for evaluating model generalizability. M M Calculating the expected per-sequence conditional log loss across held-out samples Because our goal is to learn a model that is gene- agnostic, we will evaluate the performance and generalizability of each model by calculating the expected per- sequence conditional log loss using many different held- out data sets. A model that is generalizable across many genes will perform well and have a good fit across all held- out samples despite their varying gene compositions. To test this, we will create each random, held- out sample from the original training data set by cluster- sampling all observations from V- gene allele groups, σV , uniformly at random. We will refer to each random, held- out sample as the ‘held- out testing set.’ Let G be the total number of unique V- gene allele G) be an integer which represents the number groups in the original data set. Let Gtest = Round(0.3 of unique genes included in each ‘held- out testing set.’ As such, we can sample each gene σV with probability · such that the probability of each ‘held- out testing set’ H is given by Psample(S = σV) := 1 G Gtest P(H) = Psample(S = σV) ∏σV=1 Gtest ∏σV=1 = 1 G . (21) The remaining genes not sampled as part of the ‘held- out testing set’ H will compose the ‘training set’ T . Using this ‘training set,’ we can train each model of interest as described above in (Equation 10). After this model training, we can calculate the expected per- sequence conditional log loss of the model, , for the ‘held- out testing set,’ H , as described above in (Equation 11). To achieve an unbiased estimate of the model performance, we will repeat the above procedure across 20 unique held- out testing sets and calculate the expected per- sequence conditional log loss across all samples. As such, the expected per- sequence conditional log loss across these random samples is given by M E[ℓ( )] = M 20 ∑H=1 P(H) ℓ( · M | H). (22) We use the same, unique held- out testing sets to calculate the expected per- sequence conditional log loss of each model of interest, and thus, we can compare model fit and generalizability by directly comparing the expected per- sequence conditional log loss of each model. Models that have lower expected per- sequence conditional log loss will indicate that the model is a better fit and is more generalizable across genes. Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 36 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Calculating the expected per-sequence conditional log loss across held-out samples of the ‘most-different’ V-genes While the previously described procedure for evaluating the expected per- sequence conditional log loss across held- out samples of the V- gene data set provided a metric for evaluating model generalizability across different gene sets, we were interested in evaluating model performance for groups of genes which were considered ‘most- different’ sequence- wise. Many of the germline V- gene sequences are quite similar, however, there are subgroups of these sequences which share unique sequence traits. We can characterize these ‘most- different’ V- genes by either using only the ‘terminal’ V- gene sequences (e.g. that last 24 nucleotides of each sequence which is directly parameterized in the models) or using the entire V- gene sequences. To define the ‘most- different’ V- gene allele group using the ‘terminal‘ V- gene sequences, we first calculate the pairwise hamming distance between each gene- allele- group pair. We then use hierarchical clustering to cluster V- gene allele groups based on their pairwise hamming distances (Appendix  3—figure 1A). The cluster that has the smallest average pairwise hamming distance within the cluster and the largest average pairwise hamming distance outside of the cluster is defined to be the ‘most- different’ V- gene- allele- group cluster. To define the ‘most- different’ V- gene allele group using the entire V- gene sequences, we first align all gene sequences using the DECIPHER package in R. Using these aligned sequences, we can then proceed with the same procedure as described for the ‘terminal’ V- gene sequences to define the ‘most- different’ V- gene allele group (Appendix 3—figure 1B). Once we have defined a cluster of the ‘most- different’ V- gene allele groups, using either the ‘terminal’ V- gene sequences or the full sequences, we can define a held- out testing data set H containing all data observations from the V- gene allele groups within this ‘most- different’ V- gene- allele- group cluster. All data observations from the remaining gene allele groups that were not defined to be part of the ‘most- different’ cluster will compose the ‘training set’ T . Using this ‘training set,’ we can train each model of interest as described above in (Equation 10). After this model training, we can calculate the expected per- sequence conditional log loss of the model, , for the ‘held- out testing set,’ H , as described above in (Equation 11). Models that have lower expected per- sequence conditional log loss will indicate better fit and generalizability across even the ‘most- different’ genes. We can repeat this process for other V- gene- allele- group clusters (e.g. the ‘second- most- different’ V- gene- allele- group cluster) as desired. M Appendix 3—figure 1. Un- rooted trees of ‘terminal‘ V- gene sequences (A) and full- length V- gene sequences (B) derived from hierarchical clustering. Tips are colored according to cluster membership. The tips corresponding to the ‘most- different’ group within each tree are colored in orange. Calculating the expected per-sequence conditional log loss across the full J-gene data set With the full V- gene training set, we can train each model of interest as described above in (Equation 10) to obtain a trained model . After this model fitting, we can calculate the expected per- , for the full J- gene training data set, J , using the sequence conditional log loss of the model, procedure described above in (Equation 11). Here, we use the full V- gene data set as the training M M Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 37 of 40 Computational and Systems Biology | Immunology and Inflammation Research article data set and the full J- gene data set as the testing data set. Models that have lower expected per- sequence conditional log loss on the J- gene data set will indicate that the model is a better fit and is more generalizable. Evaluating TCRβ V-gene trimming models using the expected per-sequence conditional log loss across testing data sets To validate the performance of each model, we worked with TCRα- and TCRβ-immunosequencing data representing 150 individuals, TCRγ-immunosequencing data representing 23 individuals, and IGH- immunosequencing data representing 9 individuals from three independent validation cohorts (described above). With these data, we used the model coefficients from the previous TCRβ V- gene training run (‘frozen’ in git commit 093610a on our repository) and then compute the expected per- sequence conditional log loss of the model using each independent validation data set of interest. Models that have low expected per- sequence conditional log loss across all testing data sets will indicate that the model is more generalizable and less overfit to the training data. We validated each model using V- and J- gene sequences separately. Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 38 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Appendix 4 Extended experimental analyses Exploring the gene specificity of the ‘trimming motif’ To evaluate the specificity of the motif coefficients across different genes, we can compare the per- gene model predictions for the motif and base- count beyond model to a model that only contains base- count beyond parameters. To do this, we first use the entire V- gene data set V to train both the motif and base- count beyond model as before in (Equation 10) and a model that contains only base- count beyond parameters (and no motif parameters). We can then use these models to predict the probability of trimming each possible trimming amount, 2 14 , for each gene- allele- group sequence σ . For each of these models, we can then calculate the per- gene root mean squared error, RMSE , for each gene σ such that ≤ ≤ n RMSE(σ, , V) = M I 14 ∑i=1 ∑n=2 ( � � � � � � Pemp(n | σ, i) P(n | σ; ) M − I | | 2 ) (23) where is a model trained using the V- gene training data set V , I is the set of all individuals in M | is the length of the set of individuals I , Pemp(n | σ, i) is the empirical conditional PDF the data set, | I given by (Equation 1) for trimming length n , gene σ , and individual i ) is the predicted trimming probability from a specified model . We can then compare this per- gene root mean squared error for the model trained using both motif and base- count beyond parameters with a model trained using just base- count beyond parameters. I , and P(n | σ; M M ∈ Sensitivity analysis for hairpin nick position For our modeling, we assume that the initial hairpin nick occurs at the +2 position and will create two P- nucleotides at the end of the 5’-to- 3’ gene sequence. Assuming a different hairpin nick position would incorporate a different number of P- nucleotides at the end of the gene sequence (Appendix  4—figure 1). While the hairpins are assumed to be nicked at the  +2 position most frequently (Ma et  al., 2002; Lu et  al., 2007), we wanted to test the sensitivity of our models to this hairpin nick position assumption. To do this, we assumed each of the other possible hairpin opening positions (e.g. −2,–1, 0,+1,+3) one- at- a- time and appended the appropriate number of associated of P- nucleotides given the assumed hairpin nick position to the 3’-end of each V- gene- allele- group sequence in the data set. With each of these hairpin position data sets, we re- trained the motif and base- count beyond model as before in (Equation 10) and calculate the expected per- sequence conditional log loss of the model using (Equation 11). We can compare these expected per- sequence conditional log losses to evaluate the sensitivity of the model to the +2 hairpin nick assumption. Appendix 4—figure 1. An arbitrary DNA hairpin can be nicked opened at various positions near the hairpin (left figure). Hairpin nick position 0 refers to a nick at the tip of the hairpin, position –1 refers to a nick before the Appendix 4—figure 1 continued on next page Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 39 of 40 Computational and Systems Biology | Immunology and Inflammation Research article Appendix 4—figure 1 continued last nucleotide on the 5’ strand, position +1 refers to a nick before the last nucleotide on the 3’ strand, etc. The resulting 5’-to- 3’ sequences from the various nick positions for the arbitrary gene sequence are shown on the right. Nucleotides originating from the 5’ strand of the DNA hairpin are highlighted in gray and P- nucleotides (originating from the 3’ strand of the DNA hairpin) are highlighted in purple. The various hairpin nick positions lead to 5’-to- 3’ sequences that contain different amounts of P- nucleotides. Hairpin nick positions gt0 lead to 5’-to- 3’ sequences that contain P- nucleotides, nick positions equal to zero lead to 5’-to- 3’ sequences without P- nucleotides, and nick positions < 0 lead to 5’-to- 3’ sequences without P- nucleotides and with portions of the original 5’ DNA hairpin strand removed. Evaluating the weight of the 1×2 motif and two-side base-count beyond model terms across data sets For each testing data set, we can measure the weight of the 1×2 motif and two- side base- count beyond model terms within the full 1×2 motif + two- side base- count beyond model. Recall that we trained the full 1×2 motif + two- side base- count beyond model using the model covariate function given by (Equation 2) f(n, σV; βmotif , βAT , βGC , a, b, c) := f(n, σV; βmotif , a, b) + f(n, σV; βAT , βGC , a, b, c) represents motif- specific parameter coefficients, βAT where n represents the number of trimmed nucleotides, σV represents the V- gene- allele- group sequence, βmotif represent base- count- beyond- specific parameter coefficients, a and b are non- negative integer values that represent the number of nucleotides 5’ and 3’ of the trimming site, respectively, that are included in the ‘trimming motif, and c represents the number of nucleotides 5’ of the trimming site to be included in the base- ˆβGC , count. As such, for each training data set, we can use the inferred coefficients, from a previous training run and define a new two- parameter model containing a scale coefficient for the 1×2 motif terms and a second scale coefficient for the two- side base- count beyond terms. The covariate function for this new model is given by and βGC ˆβmotif ˆβAT , , and f(n, σV; ˆβmotif ˆβAT , , ˆβGC , αmotif, αcount, a, b, c) := αmotif · f(n, σV; βmotif , βGC f(n, σV; βAT , a, b) , a, b, c) +αcount · where αmotif is the scale coefficient for the 1×2 motif terms and αcount is the scale coefficient for the two- side base- count beyond terms. We can then train this new model as described previously for each data set of interest and compare the inferred scale coefficients. Russell et al. eLife 2023;12:e85145. DOI: https://doi.org/10.7554/eLife.85145 40 of 40 Computational and Systems Biology | Immunology and Inflammation
10.1073_pnas.2304319120
RESEARCH ARTICLE | IMMUNOLOGY AND INFLAMMATION OPEN ACCESS IL- 7R licenses a population of epigenetically poised memory CD8+ T cells with superior antitumor efficacy that are critical for melanoma memory Goran Micevica,b,1 Esen Sefika, Julie F. Cheunga, Noah I. Hornicka and Richard A. Flavella,e,h,2 , Koonam Parkb, Ronan Taltyc, Meaghan McGearyc, Haris Mirzaa,c, Holly N. Blackburna,d, , Marcus W. Bosenberga,b,c,e,g,i,2, , Lilach Aizenbude,f, Nikhil S. Joshia, Harriet Klugere,f,g, Akiko Iwasakia,g,h , Andrew Danielsa,c,1, Karine Flem- Karlsenb Contributed by Richard A. Flavell; received March 17, 2023; accepted June 8, 2023; reviewed by David Fisher and Taha Merghoub Recurrence of advanced melanoma after therapy is a major risk factor for reduced sur- vival, and treatment options are limited. Antitumor immune memory plays a critical role in preventing melanoma recurrence and memory T cells could be a potent cell- based therapy, but the identity, and functional properties of the required immune cells are incompletely understood. Here, we show that an IL- 7Rhi tumor- specific CD8+ popula- tion is critical for antitumor memory and can be epigenetically augmented to drive pow- erful antitumor immune responses. Using a model of functional antimelanoma memory, we found that high IL- 7R expression selectively marks a CD8+ population in lymphoid organs that plays critical roles in maintaining tumor remission after immunotherapy or surgical resection. This population has intrinsic cytotoxic activity, lacks markers of exhaustion and has superior antitumor efficacy. IL- 7Rhi cells have a functionally poised epigenetic landscape regulated by DNA methylation, which can be augmented by hypo- methylating agents to confer improved survival and complete melanoma clearance in naive mice. Importantly, greater than 95% of tumor- specific T cells in draining lymph nodes after therapy express high levels of IL- 7R. This overlap between IL- 7Rhi and antigen- specific T cells allows for enrichment of a potent functional CD8+ population without determining antigen- specificity, which we demonstrate in a melanoma model without a known antigen. We identify that IL- 7R expression in human melanoma is an independent prognostic factor of improved survival. These findings advance our basic understanding of antitumor memory and suggest a cell- based therapy using high IL- 7R expression to enrich for a lymph node population with superior antitumor activity that can be augmented by hypomethylating agents. melanoma | immunology | immunotherapy Despite significantly improved outcomes with anti- CTLA- 4 and anti- PD- 1/PD- L1 inhib- itors, approximately half of patients with advanced melanoma will not achieve a durable response and face a high risk of recurrence and death (1). Treatment options for patients with recurrent melanoma are limited, and there is a major unmet need to both understand the determinants of a durable response and develop therapies for recurrent melanoma. Cell- based immunotherapies using adoptively transferred T cells (ACT) hold great poten- tial for patients with advanced melanoma recurrence (2–4). For instance, patients with advanced melanoma refractory to anti- PD- 1 therapy who receive adoptive transfer of tumor infiltrating lymphocytes (TILs) have better survival compared with anti- CTLA- 4 blockade (1). Cases of remission have been reported in subgroups of patients; however, significant challenges limiting the benefit of ACT remain (5). These challenges include: 1) enriching for tumor- specific T cells, 2) selecting the optimal T cell population with intrinsic characteristics suitable to mediate effective antitumor responses, and 3) overcom- ing contextual signals in the tumor microenvironment (TME) and chronic TCR stimu- lation that drive T cell dysfunction/exhaustion (6). ACT effectiveness correlates with the ability to transfer tumor- specific T cells which can recognize and kill cancer cells (7–9), however such T cells make up only a small fraction of TILs. TILs also include suppressive T regulatory cells (10) and CD39- expressing T cells that limit antitumor immunity (11). Additionally, TILs lack the memory/stemness properties (12, 13) that correlate with clinical benefit upon transfer (14, 15). Tumor- specific T cells from the TME are largely terminally differentiated (16) and express markers of exhaustion, such as PD- 1, LAG- 3, and TIM- 3 (17, 18), limiting their functional antitumor activity. T cell exhaustion also occurs during chronic viral infections, where epigenetic scarring (19, 20) limits the effectiveness of viral- specific lymphocytes similar to what is observed in the TME (21–23). Despite persistent antigen encounters promoting Significance Treatment options for patients with recurrent melanoma are limited and understanding antitumor memory is critical to prevent recurrence and develop improved therapies. Previous studies found that adoptively transferring antigen- specific T cells with memory traits provides improved clinical benefit, suggesting that antitumor memory T cells could be the ideal candidate for cell- based therapies. However, this population and its markers remain elusive. We analyzed tumor- specific T cells using single- cell approaches in a model of antimelanoma memory. Our results identify a CD8+ population selectively marked by IL- 7R expression that drives antitumor memory and can be used as a potent therapy for melanoma. The antitumor function of this population can be epigenetically augmented to develop powerful adoptive cell therapies that could improve melanoma survival. Copyright © 2023 the Author(s). Published by PNAS. This open access article is distributed under Creative Commons Attribution License 4.0 (CC BY). 1G.M. and A.D. contributed equally to this work. 2To whom correspondence may be addressed. Email: marcus.bosenberg@yale.edu or richard.flavell@yale. edu. This article contains supporting information online at https://www.pnas.org/lookup/suppl/doi:10.1073/pnas. 2304319120/- /DCSupplemental. Published July 17, 2023. PNAS  2023  Vol. 120  No. 30  e2304319120 https://doi.org/10.1073/pnas.2304319120   1 of 12 exhaustion, long- lived memory CD8+ populations with intrinsic stemness features can sustain a T cell response and immunity to chronic infection (24–26). However, the identity of an analogous CD8+ T cell population that sustains antitumor memory is cur- rently elusive. Identifying and characterizing an antitumor mem- ory population is important, not only because it would advance our understanding of antitumor immunity, but would be a poten- tial source of tumor- specific cells with memory/stemness proper- ties, which have shown promising antitumor efficacy in preclinical and small clinical studies (12, 13, 27, 28). Herein, we use an experimental model of antimelanoma mem- ory to identify and characterize an endogenous population of IL- 7Rhi CD8+ T cells with superior antitumor activity that plays critical roles in antitumor memory. This tumor- specific IL- 7Rhi CD8+ population resides in lymphoid organs, has memory and intrinsic cytotoxic features, and lacks transcriptional and epige- netic markers of exhaustion. The distinct epigenetic landscape of this IL- 7Rhi CD8+ population plays central roles in its antitumor function and can be augmented by hypomethylating agents. We find that IL- 7R signaling is essential for this cell phenotype and selecting for high IL- 7R expression can be used to enrich for cells with superior antitumor function even in tumors without known antigens. Integrating our findings with studies from human mel- anoma, we find that the IL- 7Rhi CD8+ signature is an independent prognostic factor of survival. These findings extend our funda- mental understanding of antitumor memory and suggest that adoptive transfer of IL- 7Rhi memory populations in combination with epigenetic therapies could be used to develop cell- based ther- apies for melanoma. Results + T Cells Are Necessary for Functional Antitumor Memory. CD8 To identify the T cell populations necessary for a successful antitumor memory response, we used the YUMM1.7 (BrafV600ECdkn2a−/−) melanoma model (29), engineered to express the LCMV gp33- 41 and gp66- 77 antigens as well as a fluorescent label (YUMM- GFP33/66) or chicken ovalbumin (YUMM- OVA) (30) (SI Appendix, Figs. S1A and S6). These dominant antigen models uniformly form tumors in immune competent mice but are robustly immunogenic (SI Appendix, Fig. S1B) and recapitulate key properties of human melanoma in the context of immunotherapy, including response to PD- 1 and CTLA- 4 blockade (SI Appendix, Fig. S1C), presence of progenitor- exhausted and exhausted T cell populations (SI Appendix, Fig. S1D) and response to adoptive cell transfer (SI Appendix, Fig. S1E) while facilitating interrogation of antigen- specific T cells. Using these reagents, we developed a model of functional antitu- mor memory (Fig. 1A) which exhibits 100% tumor rejection upon rechallenge. To acquire functional memory, wild- type C57BL/6 (naive) mice are initially grafted (challenged) with YUMM- GFP33/66 or YUMM- OVA tumor cells before the tumor is cleared by surgical resection or combined anti- PD- 1/anti- CTLA- 4 blockade. Four to six wk after tumor clearance by immunotherapy or surgery, mice are rechallenged with 1 to 10 × 106 tumor cells, and a robust memory response characterized by rapid tumor rejection is seen in all tested mice (Fig. 1 B and C). We validated the functional memory pheno- type with a widely used syngeneic colon cancer model, MC38, as well as a nondominant antigen- expressing model, YUMMER1.7 (SI Appendix, Fig. S1F) (31). Importantly, this memory response is tumor specific, and tumor rejection does not occur upon rechallenge with a different tumor e.g., YUMM- GFP33/66- derived functional memory mice fail to reject a rechallenge with MC38. (SI Appendix, Fig. S1G). Importantly, functional memory cannot be generated in immunodeficient RAG−/− mice, due to universal spontaneous tumor recurrence after resection (Fig. 1D). Recurrence in RAG−/− mice occurs despite confirming surgically complete resections using a fluorophore- labeled tumor line (YUMMER1.7GFP) under a microscope (Fig. 1D). This observation suggests that mature lympho- cytes are necessary to prevent tumor recurrence even after resection and to lay the foundations for memory. To further investigate which immune cells are required for func- tional memory, we depleted CD8+ T cells (Fig. 1E). The functional memory phenotype was abolished by depletion, suggesting that CD8+ T cells are necessary for functional antitumor memory. Notably, we detected a significant expansion of dominant antigen- specific CD8+ T cells in both the tumor- draining lymph node (tdLN) as well as the spleens of mice after YUMM- GFP33/66 challenge (SI Appendix, Fig. S2A). Additionally, a continued pro- portional expansion of these CD8+ T cells 6 wk after tumor resection is seen only in the tdLN (SI Appendix, Fig. S1H). These observations together led us to hypothesize that influx of T cell populations from lymphoid organs, and particularly the tdLN, is necessary for the antitumor memory response. To test this hypothesis, we treated functional memory mice with the sphingosine 1- phosphate receptor- 1 (S1PR1) inhibitor FTY- 720 (32), which blocks lympho- cyte egress from lymphoid organs. Majority of the memory mice treated with FTY- 720 failed to mount an effective antitumor mem- ory response. Thus, blocking lymphocyte egress from lymphoid organs significantly blunted the memory phenotype (Fig. 1F), and suggests that the CD8+ T cell population(s) necessary for functional antitumor memory reside in lymphoid organs. We next isolated CD8+ T cells from lymphoid organs of func- tional memory mice and analyzed their transcriptome using bulk- RNA Seq. We used lymph nodes from mice with outgrowing tumors (exhausted) as control. We performed gene set enrichment analysis and found that CD8+ T cells in lymph nodes of functional memory mice had a T cell memory–like transcriptional profile (Fig. 1 G and H), including high expression of Il7r. IL- 7R is the interleukin- 7 receptor (also known as CD127), which is critical for T cell development, survival, and memory differentiation (33–36). Whether this high IL- 7R expression was marking a tumor- specific CD8+ population became important to address. hi CD8 + Population Plays a Critical Role in  Antitumor An IL- 7R Memory. IL- 7R can mark multiple T cell states (37). IL- 7R is expressed by naive T cells, is rapidly lost upon T cell activation and is absent from most effector T cell populations but reexpressed by several types of memory CD8+ T cells (38, 39). We next sought to investigate whether IL- 7R was expressed on tumor antigen- specific T cells. We designed a flow- cytometry panel equipped to detect antigen- specific T- cells (SI Appendix, Fig. S2A), canonical memory markers (SI Appendix, Fig. S2C) and markers of T cell exhaustion (SI  Appendix, Fig.  S2B). We used this panel to interrogate the tumor- specific and polyclonal CD8+ T cells in the tdLN and spleen of functional memory and exhausted mice (Fig.  2A). Unsurprisingly, tumor- specific (tetramer positive) CD8+ T cells from exhausted conditions (outgrowing tumors) showed increased expression of exhaustion/activation marker PD- 1 when compared with functional memory (Fig. 2B). Conversely, functional memory mice had a significant increase in the proportion of tumor- specific T cells marked by IL- 7R compared with exhaustion (Fig. 2C and SI Appendix, Fig. S2D). Strikingly, we found that IL- 7R was highly expressed on the surface of tumor- specific CD8+ T cells (IL- 7Rhi) in functional memory, but not on tetramer- negative CD8+ T cells (IL- 7Rlo) in this setting, making it unique in this aspect from other canonical markers of memory that were evaluated (Fig. 2D). Only 3% of tetramer- negative CD8+ T cells were IL- 7R+ (Fig. 2D). However, 2 of 12   https://doi.org/10.1073/pnas.2304319120 pnas.org Fig. 1. CD8+ T cells are necessary for functional antitumor memory.  (A) Diagram summarizing generation of functional memory mice. Wild- type C57BL/6 mice were injected with 1 to 2 × 106 YUMM- OVA or YUMM- GFP33/66 cells on day zero. Mice were treated with either surgical resection (day 10) or immunotherapy (anti- CTLA- 4 and anti- PD1, days 8 to 18). Tumor rechallenge with 2 × 106 YUMM- OVA or YUMM- GFP33/66 melanoma cells was done on day 43. (B, Left) Tumor growth upon rechallenge of functional memory mice (red) generated by tumor resection (B, Left) or control naive mice (black) with 3 × 106 YUMM- GFP33/66 melanoma cells. (B, Right) Kaplan–Meier curve comparing survival of age- matched functional memory or control (naive) mice upon tumor rechallenge. (C, Left) Tumor growth upon rechallenge of functional memory mice (red) generated by complete response to immunotherapy or control naive mice (black) with 3  × 106 YUMM- OVA or YUMM- GFP33/66 melanoma cells. (C, Right) Kaplan– Meier curve comparing survival of age- matched functional memory or control (naive) mice upon tumor rechallenge. (D, Left) Tumor growth upon resection of RAG WT (black) and RAG KO (red) mice. (Middle) Kaplan–Meier curve comparing survival of age- matched Rag KO (red line) and WT mice (black line) after resection of YUMM- OVA or YUMM- GFP33/66 tumors. RAG KO mice fail to develop functional memory and tumors recur at the site of resection. (D, Right) fluorescent images taken preresection (D,  Top) and postresection (D,  Bottom) of GFP- expressing YUMMER1.7GPF tumors under fluorescent dissecting microscope. Normal autofluorescence seen in lower image but no sign of tumor cells. (E, Left) Comparison of CD8+ percent in spleen and circulation of CD8 depleted (red) and sham- treated functional memory mice (black).(E, Right) Kaplan–Meier curve comparing survival of functional memory mice treated with isotype control (control, black line) or CD8 depletion (red line). CD8 depletion was done for 3 wk beginning on day 40, 3 d prior to tumor rechallenge with YUMM- OVA or YUMM- GFP33/66 melanoma cells. (F, Left) Comparison of CD3+ percent in circulation of FTY- 720- treated (red) and sham- treated functional memory mice (black). (F, Right) Kaplan–Meier curve comparing survival of the control (black) and FTY720 (red)- treated groups. Functional memory mice were treated with FTY720 (red line) or phosphate buffered saline (black line) beginning on day forty, 3 d prior to YUMM- OVA or YUMM- GFP33/66 rechallenge. FTY720 administration leads to failure of functional memory. (G) Heatmap summarizing scaled expression values of a subset of variable genes between CD8+ T cells from lymph nodes of functional memory and exhausted mice (RNA- Seq data). (H) Volcano plot of top gene- sets enriched in CD8+ T–cells from tumor- draining lymph nodes of functional memory mice. Tmem (P = 8.53 × 10−29) and T–cells (5.3 × 10−25) are the top overrepresented cell types/gene signatures (red). Student’s t test, ANOVA or Log- rank test; *P < 0.05, **P < 0.01, ***P < 0.001. Bar graphs shown as mean +/− SEM, based on three biological replicates. the population of IL- 7R+ CD8+ T cells accounted for over 90% of tumor- specific CD8+ T cells (Fig. 2D) in lymph nodes in con- ditions associated with functional memory. The proportion of tumor- specific IL- 7Rhi CD8+ T cells was highest in the tdLN and expanded over time from approximately 85% after initial tumor clearance to over 95% after 2 wk (SI Appendix, Fig. S2D). This population did not express exhaustion- associated markers PD- 1 or TIM- 3 (SI Appendix, Fig. S2F). To determine if this expanding IL- 7Rhi tumor- specific CD8+ population is functionally important for antitumor memory, we blocked IL- 7R signaling both after initial tumor challenge (priming) or during functional memory rechallenge (Fig. 2E). Mice treated with IL- 7R blocking antibody after priming, but not during rechallenge, failed to form functional antitumor memory and exhibited tumor formation and growth identical to naive mice upon rechallenge (Fig. 2F). Administration of IL- 7R blocking antibody immediately prior to functional mem- ory rechallenge had no effect on the effective memory pheno- type (Fig. 2F). CD8+ T cell levels were unaffected at the time of rechallenge (SI Appendix, Fig. S2E ). Mice receiving IL- 7R blockade during priming (Fig. 2F) had a median survival of 21 d. The median survival in functional memory mice (control) was undefined, since all tumors were rejected. Thus, IL- 7R blockade after priming prevented formation of functional anti- tumor memory. Taken together, these findings suggest that an IL- 7Rhi tumor–specific CD8+ population in lymphoid organs is critical for antitumor memory. hi CD8 + A Tumor- Specific IL- 7R T Cell Population Has Central Memory- Like Features. To further characterize the IL- 7Rhi CD8+ cell population important for antitumor memory, we used scRNA- seq to characterize the immune cell populations in tumor- draining lymph nodes from mice with functional memory, including both tumor- specific and nonspecific (polyclonal) populations. After dimensionality reduction, graph- based clustering and mapping gene expression in two dimensions via t- distributed stochastic neighbor embedding (tSNE), we identified 13 distinct cell clusters (Fig. 3A) from the tdLN of functional memory mice. Top marker genes were identified for each cluster (Fig. 3B and SI Appendix, Fig. S3 A and B). Clusters 0 to 3 were marked by canonical B cell markers, included naive and pre- B cells. Cluster 4 was marked by Cd8a expression, while cluster 5 consisted of CD4 cells. Both clusters 4 and 5 expressed high levels of Il7r (Fig. 3C). Overall, CD8 expression was found in clusters 4, 12, and 13 of the datasets, and those clusters were selected for further analysis (SI Appendix, Fig. S3F). The remaining clusters in the dataset included natural killer cells (cluster 8), γδ- T cells (cluster 9), neutrophils (cluster 6), and monocytes (cluster 10) (SI Appendix, Fig. S3C). We characterized the clusters more closely with a panel of common markers of memory and exhaustion (Fig. 3 C–F). PNAS  2023  Vol. 120  No. 30  e2304319120 https://doi.org/10.1073/pnas.2304319120   3 of 12 Fig. 2. An IL- 7Rhi CD8+ population plays a critical role in antitumor memory. (A) Diagram illustrating generation of exhausted and functional memory mice. Wild- type C57BL/6 mice were injected with 1 to 2 × 106 YUMM- OVA or YUMM- GFP33/66 tumors on day zero. For functional memory generation, mice started receiving immunotherapy on day eight (five treatments of combined anti- CTLA4/anti- PD- 1) or surgical resection on day ten. Functional memory mice were used for experiments 2 wk after tumor clearance. For exhausted mice, no therapy was administered. Exhausted mice were used for experiments 2 wk after initial tumor injection. (B) Percent PD- 1 expression in exhausted (black) versus functional memory (red) tumor- specific (Tet+) CD8+ T cells in tumor- draining lymph nodes (tdLN) and spleen. (C) Percent IL- 7R positivity of tumor- specific CD8+ T cells (C, Left) in exhausted (black) versus functional memory (red) tdLN and histogram overlay (C, Right) of IL- 7R expression in CD8+ T cells of exhausted (gray), recently resected (blue) and functional memory (red) mice. (D) Percent positivity of key memory and stemness markers in antigen- specific (tetramer positive, red) versus polyclonal (tetramer negative, black) CD8+ T cells in the tumor- draining lymph node of functional memory mice. (E) Diagram summarizing IL- 7R blocking antibody administration. Mice were treated for 21 d, starting either 3 d prior to resection (periresection, red bar days 7 to 28) or 3 d prior to rechallenge (rechallenge, black bar days 40 to 61). See also Fig. 2F. (F) Tumor volume (F, Left) and Kaplan–Meier curves (F, Right) showing tumor growth after rechallenge of functional memory mice, which received no additional therapy (solid black line), IL- 7R blockade at time of rechallenge (dashed line), or periresection IL- 7R blockade (red line). Student’s t test, ANOVA or Log- rank test; *P < 0.05, **P < 0.01, ***P < 0.001. Bar graphs shown as mean +/− SEM, based on three biological replicates. Cluster 4 showed high expression of Il7r, Ccr7 (C- C chemokine receptor type 7), Sell (also known as CD62L), Cd27 and Cd28. Ccr7 is expressed by naive, regulatory, and central memory T cells, and plays important roles in regulating T cell migration and homing to lymph nodes (40). As opposed to Ccr7, which mediates ‘hard stop’ signaling and transmigration into the lymph node, Sell is an L- selectin that mediates the initial rolling adhe- sion and deceleration. It is expressed on naive and central mem- ory T cells and is generally absent from effector T cell populations (41). Tcf7 is a Wnt signaling target essential for T cell memory differentiation and self- renewal and is essential for formation of central memory T cells (42–44). Cluster 4 also exhibited high expression of Cd28 (Fig. 3D), which is required for proliferation of stem- like cells (45, 46) as well as Eomes, a transcription factor with important roles in generation and persistence of memory CD8+ T cells. High levels of Bcl- 2, which can eschew apoptotic death of lymphocytes and promotes long- term survival of mem- ory populations (47–49) was notable in cluster 4 (SI Appendix, Fig. S3E). Entpd1 (CD39), which has recently been shown to play immunosuppressive functions in the TME (11) was not expressed by cluster 4 (Fig. 3C). Markers of exhaustion, such as Pdcd1, Havcr2, Lag3, Tigit, and Id2 were absent from cluster 4 (Fig. 3E). Clusters 12 and 13 were also CD8+; however, they did not express memory markers (Fig. 3D). In contrast, they expressed markers of exhaustion (Fig. 3F). We also profiled the TME from mice at day 14 after tumor grafting (5,192 cells). This condition represents an exhausted phenotype generated by grafting YUMM- GFP33/66 cells subcu- taneously and not providing any therapy. A failed immune response and uniform tumor outgrowth occurs in this setting. In contrast to the functional memory signature (Fig. 3E), nearly all CD8+ cells in the exhausted TME expressed high levels of exhaustion markers Pdcd1, Havcr2 and Lag3 (SI Appendix, Fig. S3G). Based on our findings, cluster 4 likely represents the IL- 7Rhi CD8+ population that is critical for functional antitumor memory. It recapitulated the key features detected by flow cytometry, including high expression of IL- 7R (Fig. 2D), absence of exhaus- tion markers (Fig. 2B) and tumor- specificity. Interestingly, gene set expression analysis of tumor- specific CD8+ T cells in cluster 4 showed enrichment of central- memory (TCM)- like and cytotoxic- like T cell phenotypes (Fig. 3G). hi Cells Constitutively Express Intrinsic Tumor- Specific IL- 7R Cytotoxicity Mediators, Have Superior Antitumor Function. We next sought to further investigate the potential cytotoxic properties of the IL- 7Rhi CD8+ T cells in cluster 4. We used the composite expression of 12 cytotoxicity- associated genes to calculate an overall cytotoxicity score for all cell clusters. IL- 7Rhi CD8+ T cells (cluster 4) and NK cells (cluster 8) had the highest cytotoxicity score (Fig. 4A). Tumor- specific IL- 7Rhi CD8+ T cells (cluster 4) expressed intrinsic cytotoxic mediators including natural killer cell granule protein 7 (Nkg7), killer cell lectin receptor D1 (Klrd1), cathepsin W (Ctsw), cystatin F (Cst7), and Ccl5 (Fig. 4B), suggesting a capacity for cytotoxicity. A high level of Nkg7 expression was reported on antigen- specific CD8+ T cells, on cytotoxic lymphocytes (CTLs) infiltrating tumors in patients treated with immunotherapy (50) and regulates cytotoxic granule exocytosis in effector lymphocytes (51, 52). Ccl5 plays important roles in the onset of a proliferative burst in the setting of antigen reencounter (53–55). Ctsw and Cystatin F (Cst7) are membrane- associated cysteine protease whose expression is restricted to cytotoxic cells (56), are secreted 4 of 12   https://doi.org/10.1073/pnas.2304319120 pnas.org Fig.  3. A tumor- specific IL- 7Rhi CD8+ T cell population has central memory- like features. (A) tSNE (t- distributed stochastic neighbor embedding) plot of 13 distinct cell clusters from tumor- draining lymph nodes of functional memory mice (n = 13,963 cells). Cells are colored based on clusters identified by the Louvain  algorithm.  See also SI  Appendix, Fig.  S3 C and D for automated cell type identification and their connectedness. (B) Heatmap summarizing scaled expression values of marker genes for each individual cluster, as shown in A. See also SI Appendix, Fig. S3 A and B. (C) Dot plot showing relative mean expression level of markers of T cell memory–associated genes. (D) Violin plots of a subset of memory marker genes in CD8+ clusters 4 (IL- 7Rhi), 12 and 13. (E) Dot plot showing relative mean expression level of markers of exhaustion- associated genes. (F) Violin plots of a subset of exhaustion marker genes in CD8+ clusters 4 (IL- 7Rhi), 12 and 13. (G) Volcano plot of top gene- sets enriched among the tumor- specific IL- 7Rhi CD8+ cluster. Central Memory T cells and Cytotoxic T cells are the top overrepresented gene sets (red). Student’s t test or ANOVA; *P < 0.05, **P < 0.01, ***P < 0.001. Bar graphs shown as mean +/− SEM, based on three biological replicates. Violin plots show the median (dashed line) and quartiles (dotted lines). Cytokines measured by ELISA using recombinant protein standard curve, based on three biological replicates. during target cell killing (57) and regulate cytotoxicity in NK cells and effector- memory T cells (Tem) (58). Genes canonically associated with a CD8+ T cell effector signature (59), such as Gzmb, Prf1, Ifng were not highly expressed by IL- 7Rhi CD8+ T cells (cluster 4) (Fig. 4B). The other CD8+ clusters, 12 and 13, had no appreciable or significantly lower expression of these cytotoxicity- associated genes compared with the IL- 7Rhi CD8+ cluster (Fig. 4C). Antigen- specific memory T cells are long- lived, respond rapidly to antigen rechallenge, and migrate efficiently to sites of rechallenge, making them particularly well for cell- based antitumor therapies (60). As noted previously, identifying a cell surface marker to indi- rectly enrich for tumor- specific T cells without the need to explicitly determine tumor antigens could increase the clinical benefit of ACT (61). As we have shown above, >95% of tumor- specific T cells expressed high levels of IL- 7R (Fig. 2D). Additionally, based on the central memory- like and cytotoxic traits we found in IL- 7Rhi CD8+ cells, we hypothesized they could be used for adoptive cell transfer without the need to identify tumor antigens, which is usually not feasible in clinical practice. To test this possibility, we adoptively transferred IL- 7Rhi and IL- 7Rlo CD8+ cells from tdLN of functional memory mice into wild- type mice (Fig. 4D). We subsequently injected the recipient naive mice with tumor cells and monitored tumor growth and survival. Mice receiving IL- 7Rlo CD8+ T cells showed similar tumor growth kinetics as negative controls (no ACT), with no significant difference in median overall survival between the IL- 7Rlo and negative control group (Fig. 4E). By contrast, transfer of IL- 7Rhi cells conferred a twofold decrease in tumor growth rate (Fig. 4D), significantly prolonged overall survival by 50%, and cured a subset of mice (Fig. 4E). Collectively, these experiments suggest that IL- 7R is a func- tional marker of a tumor- specific memory/cytotoxic CD8+ T cell population with superior antitumor function. PNAS  2023  Vol. 120  No. 30  e2304319120 https://doi.org/10.1073/pnas.2304319120   5 of 12 Fig.  4. Tumor- specific IL- 7Rhi cells constitutively express intrinsic cytotoxicity mediators, have superior antitumor function. (A) Single- cell transcription level and distribution of expression for indicated genes in the tSNE plot from Fig. 3A. Level of transcription displayed as gradient from gray (no expression) to purple (expression). Cytotoxicity score (A, Left) based on expression of Ctsw, Cst7, Gzmb, Prf1, Nkg7, Klrd1, Ccl5, Gzmm, Runx3, Ifng, Il2 and Tnf. (B) Dot plot showing relative mean expression level of markers of cytotoxicity across the CD8+ T cell clusters 4 (IL- 7Rhi), 12 and 13. (C) Violin plots of a subset of cytotoxicity marker genes in CD8+ clusters 4 (IL- 7Rhi), 12 and 13. (D, Left) Tumor volume curves for C57BL/6 mice grafted with melanoma cells after receiving an IL- 7Rhi (red) or IL- 7Rlo (black)  population by adoptive transfer from tdLN of functional memory mice. (D, Right) Comparison of tumor volume at day 21 post tumor cell injection in the IL- 7Rhi (red) and IL- 7Rlo (black) recipient group. (E) Kaplan–Meier survival curve for tumor- bearing mice from indicated groups described in D. Student’s t test or ANOVA; *P < 0.05, **P < 0.01, ***P < 0.001. Bar graphs shown as mean +/– SEM, based on three biological replicates. Violin plots show the median (dashed line) and quartiles (dotted lines). Five mice were assigned to each experimental group. hi + Cells Have a Distinct Poised Epigenetic Landscape CD8 IL- 7R Regulated by DNA Methylation That Imparts Superior Antitumor Function. To better understand the epigenetic landscape of IL- 7Rhi CD8+ cells, we explored the epigenetic regulation of IL- 7Rhi CD8+- related genes, such as Il7r and Tcf7. DNA methylation regulates Il7r expression in human CD8+ T cells (62). Similarly, Tcf7 methylation can regulate Tcf7 expression during early effector T cell differentiation (63, 64). We hypothesized that Tcf7 and Il7r expression may also be regulated by DNA methylation in the context of functional memory. We sorted IL- 7Rhi and IL- 7Rlo CD8+ T cell populations and analyzed Tcf7 promoter methylation by methylation- specific PCR. IL- 7Rlo cells had significantly higher Tcf7 DNA methylation compared with IL- 7Rhi cells (Fig. 5 A and B), suggesting a transcriptionally repressive role. To functionally test the effect of DNA methylation on Tcf7 and Il7r expression, we treated TILs with the hypomethylating agent RG108 ex vivo and measured Tcf7 expression by quantitative RT- PCR (Fig. 5 D and E). Treatment with RG108 increased the expression of Tcf7 twofold (Fig. 5E), and similar findings were seen with Il7r (Fig. 5D). Promoter hypermethylation is generally associated with stable gene repression (65, 66), while promoter hypomethylation, as seen in IL- 7Rhi CD8+ cells, is typically considered permissive of tran- scription (67) and can be associated with accessible chromatin (68). To investigate the chromatin accessibility of the Il7r and Tcf7 promoter in functional memory, we quantified open chromatin accessible regions using ATAC- seq (Fig. 5F and SI Appendix, Fig. S4A). Il7r, Tcf7, and other memory- associated loci showed an open chromatin configuration near their transcriptional start site (TSS) (Fig. 5 F, Upper Left), while genes associated with exhaustion, such as Pdcd1 and Havcr2 (Fig. 5 F, Upper Right), had a closed chromatin configuration, in agreement with our single- cell RNA- seq expression data. Interestingly, canonical effector loci associated with cytotoxicity and production of IFN- g, IL- 2 and TNF- a (Fig. 5F bottom row) as well as Gzma, Gzmb and Prf1 retained an open chromatin accessibility region, but were not appreciably expressed. An open chromatin configuration allows binding of transcription factors and gene expression (69). The open chromatin configuration at these loci suggested to us that genes encoding effector functions are poised for transcription and that IL- 7Rhi cells are functionally primed to respond to tumor rechallenge. Memory- permissive poised epigenetic states have been previously reported in cytokine regulatory elements of antigen- specific (70) and memory populations (71). To investigate if this open chromatin state has functional signifi- cance, we rechallenged functional memory mice with YUMM- OVA or YUMM- GFP33/66 melanoma cells (day 43) and sorted CD8+ T cells one day after tumor rechallenge to measure cytokine pro- duction (Fig. 5G and SI Appendix, Fig. S4B). Following rechallenge, there was a significant production of effector cytokines, including IFN- g, IL- 2, IL- 12, and IL- 1ß (Fig. 5H and SI Appendix, Fig. S4C). To delineate the contributions of the IL- 7Rhi population to this poised effector state, we sorted IL- 7Rhi and IL- 7Rlo CD8+ T cells from tumor- draining lymph nodes of functional memory mice and measured their ability to produce effector cytokines. Strikingly, after stimulation, the IL- 7Rhi population showed a significantly higher increase in IFN- g, TNF- a, and IL- 2 production compared with the IL- 7Rlo population (Fig. 5I). Given the ability of RG108 to epigenetically induce memory- related transcripts (Fig. 5 D and E) and based on the above functional studies (Fig. 5I), we hypothesized that the epigenome plays critical roles in the ability of IL- 7Rhi CD8+ to produce antitumor responses (Fig. 4E), and that pharmacologic epigenetic agents could further improve their antitumor activity. To test this possibility, we used the YUMMER1.7 cell line (31), which was generated by exposing the parental YUMM 1.7 cell line to UV radiation to mimic the neoantigen load of human tumors but without a dominant antigen. We generated YUMMER1.7 6 of 12   https://doi.org/10.1073/pnas.2304319120 pnas.org Fig. 5. IL- 7Rhi CD8+ cells have a distinct poised epigenetic landscape regulated by DNA methylation that imparts superior antitumor immunity. (A) Methylation- specific PCR targeting CpG loci in Tcf7 promoter region in nonmethylated DNA standard (Ctrl, leftmost lane), IL- 7Rlo and IL- 7Rhi CD8+ T cells from tumor- draining lymph nodes of functional memory mice. PCR amplification of Tcf7 locus in IL- 7Rlo and IL- 7Rhi genomic DNA (loading control, regular primers). (B) Relative quantification of Tcf7 methylation IL- 7Rlo and IL- 7Rhi CD8+ T cells (C) Relative quantification of genomic DNA/loading control in the IL- 7Rlo and IL- 7Rhi conditions. (D and E) Quantification of Il7r (D) and Tcf7 (E) expression by RT- PCR in murine T cells upon treatment with vehicle control (black) or hypomethylating agent RG- 108 (red). (F) Visualization of chromatin accessibility at or near transcriptional start sites (TSS) of indicated genes. Peaks from exhausted phenotype shown in black, functional memory shown in red. See also SI Appendix, Fig. S4A. (G) Diagram of tumor rechallenge experiment in functional memory mice shown in H and I. (H) Bar graphs showing comparison of IFN- g, IL- 2, IL- 12, and IL- 1ß cytokine levels in functional memory (red) and tumor rechallenged conditions (purple). See also SI Appendix, Fig. S4 B and C. (I) Quantification of IFN- g, TNF- a and IL- 2 production upon stimulation of IL- 7Rlo (black) and IL- 7Rhi (red) CD8+ T cells from tumor- draining lymph nodes of functional memory mice. (J) Kaplan–Meier survival curves (J, Left) for C57BL/6 mice grafted with melanoma cells after receiving untreated IL- 7Rlo (dark blue) population, untreated IL- 7Rhi population (crimson), RG- 108- treated IL- 7Rlo population (light blue), or RG- 108- treated IL- 7Rhi population (red) of CD8 T cells from tdLN of functional memory mice by adoptive transfer (red). Pie charts (J, Right) showing percent survival in these same groups. (K) Scatter plot of TCF7 methylation (beta value) and gene expression (RSEM log2) in cohort of human melanoma samples (72). Each dot represents an individual sample (Pearson’s r = 0.67, P = 2.51 × 10−62). See also SI Appendix, Fig. S4D. (L) Bar graph of Tcf7 promoter methylation in cohort of human melanoma samples stratified into an IL- 7Rhi (red) and IL- 7Rlow (black) group. (M) Heatmap showing DNA methylation of IL7Rhi- defining genes in a large cohort of melanoma patients stratified by IL- 7R expression. Hypomethylation shown in blue, hypermethylation shown in red. Student’s t test, Log- rank or ANOVA; *P < 0.05, **P < 0.01, ***P < 0.001. Bar graphs shown as mean +/− SEM, based on three biological replicates. Violin plots show the median (dashed line) and quartiles (dotted lines). Cytokines measured by ELISA using recombinant protein standard curve. Cell count based on luminescence of CellTiterGLO v2 assay. Results based on three biological replicates. functional memory mice and adoptively transferred lymph node IL- 7Rhi or IL- 7Rlo CD8+ T cells treated ex vivo with hypometh- ylating agent RG- 108 to naive mice. Strikingly, transfer of RG- 108- treated IL- 7Rhi CD8+ T cells conferred protective immu- nity against YUMMER1.7 to naive mice (Fig. 5J) (P = 0.001, Log- rank test), extending the median survival from 30 d to unde- fined (>80 d) and led to tumor clearance in 75% of the mice (Fig. 5J). Equally important, RG- 108 treatment led to increased IL- 7R expression (Fig. 5D) and abrogated the difference in anti- tumor efficacy seen between IL- 7Rlo and IL- 7Rhi we previously found (Fig. 4D), leading to tumor clearance in 60% of the RG- 108- treated IL- 7Rlo T cell recipients (Fig. 5J). These findings suggest that the epigenetic state is a key determinant of antitumor function in these cell populations. To determine the relevance of these findings in human melanoma, we analyzed the pattern of IL- 7R expression in the TME of a large cohort of melanoma patients (72). High Tcf7 promoter methylation in human melanoma samples is significantly associated with low Tcf7 expression level, suggesting a suppressive role for DNA methylation (Fig. 5K). We next stratified the melanoma cohort based on IL- 7R expression into an IL- 7Rhi and IL- 7Rlo group. Consistent with our experimental findings, the IL- 7Rhi group exhibited hypomethylation of the Tcf7 promoter and significantly higher Tcf7 expression com- pared to the IL- 7Rlo group (Fig. 5L). Importantly, we found that the IL- 7Rhi CD8+ memory signature we identified experimentally in our mouse model, was hypomethylated in the IL- 7Rhi melanoma group (Fig. 5M and SI Appendix, Fig. S4D), suggesting that DNA methyl- ation likely regulates this memory signature. PNAS  2023  Vol. 120  No. 30  e2304319120 https://doi.org/10.1073/pnas.2304319120   7 of 12 Our findings suggest that IL- 7Rhi CD8+ cells have a distinct, func- tionally poised epigenetic landscape regulated by DNA methylation, which confers superior antitumor cytotoxicity and can be potentiated by hypomethylating agents to improve cell- based therapies. IL- 7R Is a Marker of a Memory- Like Population in the TME and a Prognostic Factor for Melanoma Survival. Based on our observations, we hypothesized that high IL- 7R expression may be associated with improved melanoma survival and durable response to therapy. We first investigated whether the IL- 7Rhi memory transcriptional signature we identified experimentally in our functional memory model is present in the human melanoma TME. We stratified a large cohort of patients with advanced melanoma (72) into an IL- 7Rhi and IL- 7Rlo group (Fig. 6A) and compared expression of the signature we found experimentally between the IL- 7Rhi and IL- 7Rlo groups. We found that the human IL- 7Rhi melanoma TME group recapitulated the transcriptional signature we experimentally detected in functional memory mice (Fig. 6B and SI Appendix, Fig. S5A). Gene set enrichment analysis of the IL- 7Rhi group identified T cell signaling and immune response path- ways, including immune response regulation (P = 5.78 × 10−23), antigen receptor–mediated signaling (P = 2.34 × 10−21), cytokine- mediated signaling (P = 1.48 × 10−19), regulation of T cell activation (P = 1.79 × 10−19), and regulation of interferon- g produc- tion (P = 7.56 × 10−15) among the top enriched pathways (Fig. 6C and SI Appendix, Fig. S5B). Overall, the pathways enriched in the IL- 7Rhi melanoma TME cohort showed significant overlap with the experimentally identified IL- 7Rhi cells, suggesting that a memory- like transcriptional profile is present in the melanoma TME. Importantly, the IL- 7Rhi TME group (Fig. 6 D, Left) had longer overall survival (2,421 d vs. 875 d, P < 0.0001, Log- rank test) and progression- free survival (755 vs. 523 d, P = 0.03, Log- rank test) Fig. 6. IL- 7R is a marker of a memory- like population in the TME and a prognostic factor for melanoma survival. (A) Melanoma patient cohort [72] was stratified based on IL- 7R expression into an IL- 7Rhi (red) and IL- 7Rlo group (black). (B) Violin plots comparing expression of indicated genes in the IL- 7Rhi (red) and IL- 7Rlo groups (black) in the melanoma patient cohort. See also SI Appendix, Fig. S5A. (C) Top ten gene sets enriched in the IL- 7Rhi melanoma cohort. See also SI Appendix, Fig. S5B for detailed statistics. (D, Left) Kaplan–Meier curve comparing overall survival of the IL- 7Rhi and IL- 7Rlo melanoma groups (n = 96 patients per group). (D, Right) Kaplan–Meier curve comparing progression- free survival of the IL- 7Rhi and IL- 7Rlo melanoma groups in the same cohort (n = 96 patients per group). (E) Kaplan–Meier curve comparing overall survival of the IL- 7Rhi and IL- 7Rlo groups in a Yale cohort of 60 patients with melanoma who received immune checkpoint inhibitors. (F) Immunohistochemical staining for IL- 7R of pretreatment whole- tissue sections in the Yale melanoma cohort. (F, Left) Representative image of IL- 7Rhi staining and accompanying 40× magnification insets. (F, Right) Representative image of IL- 7Rlo staining and accompanying 40× magnification insets. (G) Kaplan–Meier curve comparing overall survival of IL- 7Rhi and IL- 7Rlo lung cancer groups (n = 140 patients per group). (H) Experimental approach to IL- 7R blockade. Wild- type C57BL/6 mice received immunotherapy alone (anti- CTLA4 and anti- PD1, shown in red) or immunotherapy with concurrent IL- 7R blockade (shown in black) 7 d after melanoma cell injection. (I) Tumor volume curves for mice receiving immunotherapy alone (black) or immunotherapy with concurrent IL- 7R blockade (red) from indicated groups described in H. Statistical tests used: Student’s t test or ANOVA and Log- rank test; *P < 0.05, **P < 0.01, ***P < 0.001. Bar graphs shown as mean +/− SEM, based on three experiments (except human data). Violin plots show the median (dashed line) and quartiles (dotted lines). n = 5 mice per group. 8 of 12   https://doi.org/10.1073/pnas.2304319120 pnas.org (Fig. 6 D, Right) compared with the IL- 7Rlo TME group. Expression of IL7R did not vary significantly with ulceration, clinical substage or Breslow thickness (SI Appendix, Fig. S5G). Interestingly, TCF7 expression was not associated with differences in survival, suggesting that the effect of IL- 7R is not simply a reflection of CD8+ infiltration or a stemness program orchestrated by TCF7 (SI Appendix, Fig. S5 C and D). We next performed Cox- proportional hazard analysis of IL- 7R expression in relation to overall survival in the melanoma cohort. Common prognostic clinicopathologic variables, including age, gender, Breslow thick- ness, clinical stage, and ulceration were used as covariates. In uni- variate and multivariate analyses, IL- 7R expression was an independent prognostic factor of overall survival (HR = 0.91, 95%CR 0.85 to 0.97, P = 0.01) (SI Appendix, Fig. S5E). To investigate these findings further, we evaluated IL- 7R expres- sion by immunohistochemistry in a cohort of 60 patients with mel- anoma treated with checkpoint inhibitors at Yale. In this independent cohort, high IL- 7R expression was associated with significantly longer tumor- specific survival (undefined vs. 678 d, P = 0.02, Log- rank test) compared with the IL- 7Rlo group (Fig. 6 E and F). TCF7hi was not associated with significantly different survival com- pared with TCF7lo samples (SI Appendix, Fig. S5F), in agreement with our prior observations. Beyond melanoma, IL- 7Rhi is also asso- ciated with significantly longer overall survival in a large cohort of patients with lung cancer (Fig. 6G). To better understand the association of IL- 7R with survival, we next investigated how IL- 7R signaling impacts the acute immune response to anti- CTLA4 and anti- PD1 therapy. We grafted wild- type mice with YUMM- GFP33/66 cells, and 7 d later the mice were divided into a group receiving immunotherapy alone (combined anti- CTLA4 and anti- PD1) or immunotherapy with IL- 7R block- ade (Fig. 6H). Blocking IL- 7R concurrently with immunotherapy did not affect the remission induced by anti- PD1 and anti- CTLA4 therapy (Fig. 6I). Strikingly however, IL- 7R blockade enhanced recurrence (Fig. 6I), suggesting a critical role for IL- 7R signaling in maintaining a durable antitumor immune response. Collectively, our work identifies IL- 7R as a functional marker of a memory/cytotoxic CD8+ T cell population, which is critical for functional antitumor immunity after checkpoint therapy or surgery. This tumor- specific population has a distinct epigenome, superior antitumor function, and can be enriched for adoptive cell therapy based on high IL- 7R expression, without knowledge of tumor antigens, and may thus help improve melanoma therapy in the postsurgical or postimmunotherapy setting. Discussion Despite significantly improved outcomes with anti- CTLA- 4 and anti- PD- 1/PD- L1 therapies, approximately half of patients with advanced melanoma do not achieve a durable response and face a high risk of recurrence, for which treatment options are limited. Adoptive transfer is a promising treatment modality for patients with advanced melanoma recurrence (1), but the identity of an optimal T cell population functionally poised to deliver effective antitumor immunity has been poorly defined (73). Our work, using a model for the induction of antimelanoma memory with a dominant tumor antigen, identifies a population of tumor- specific IL- 7Rhi memory CD8+ T cells that resides in lymphoid organs, plays critical roles in antitumor memory, and has superior antitumor functional capacity. These tumor- specific IL- 7Rhi cells have a distinct transcriptional program, including TCM- like memory markers, cytotoxic properties reminiscent of TEM cells, and notably lack exhaustion markers. The formation of these long- lived cells is necessary to maintain antitumor immunity in the context of checkpoint blockade or surgical excision. The IL- 7Rhi memory/cytotoxic population is found in the tumor- draining lymph node and spleen early after priming and is functionally dependent on IL- 7R signaling. Blocking the IL- 7 receptor abolishes the establishment of the population and results in prevention of antitumor memory. Our studies use clinically relevant immuno- genic melanoma lines that have the capacity to respond to check- point inhibitor therapies. Whether the T cell population we describe also plays important roles in less immunogenic melanomas is subject to future research. Enriching for tumor- specific T cells for adoptive transfer in the clinical setting remains challenging (74). TILs commonly used for adoptive transfer include only a small fraction of tumor- specific cells, also include suppressive T cell populations (10), largely lack memory/stemness properties (12, 13) and commonly express markers of exhaustion, (17, 18), limiting their functional antitu- mor activity. The tumor- draining lymph node is a potential source of T cells with a memory/stemness program that lack exhaustion and has been proposed in other cancer types (75). Importantly, we show that >95% of tumor- specific CD8+ T cells in the postim- munotherapy tumor- draining lymph node express high levels of IL- 7R, which is otherwise sparsely expressed in the lymph node by nonantigen- specific cells. Selection for IL- 7Rhi could be a strat- egy to indirectly enrich for a tumor- specific CD8+ population without known tumor antigens. Tumor- draining lymph node tissue is available for a subset of advanced melanoma patients (usu- ally intermediate thickness melanomas T2/T3 corresponding to Stage III substages) as part of melanoma clinical management (76). Functionally, IL- 7Rhi CD8+ T cells have superior antitumor activ- ity compared to their IL- 7Rlo counterparts. Transfer of the IL- 7Rhi CD8+ population significantly decreases tumor growth, prolongs survival, and leads to tumor clearance in a subset of naive mice. In addition to constitutive intrinsic cytotoxic properties, these cells have a functionally poised chromatin landscape without epigenetic “scars” that allows them to rapidly recall effector function, such as production of IFN- g and IL- 2. DNA methylation plays a prom- inent role in their epigenetic identity and hypomethylating agents can induce and IL- 7Rhi state ex vivo and significantly potentiate antitumor function. Alterations in DNA methylation, through TET2 and DNMT3A, have been implicated in regulating CAR- T cell stemness/exhaustion and antitumor activity (77, 78), but are incompletely understood. Integrating and contextualizing our findings with studies from human melanoma, we found that an IL- 7Rhi memory signature is present in the melanoma TME. Our findings agree with data reported by Sade- Feldman et al. who identified a CD8+ cluster with increased expression of genes linked to memory and associated with improved immune response (79). Interestingly, our data suggest that blocking IL- 7R does not preclude response to checkpoint inhibition acutely (Fig. 6I), but IL- 7R signaling is required to maintain a response to therapy. A recent study reported that anti- PD- 1 response and melanoma patient survival is associated with a late T cell mem- ory transcriptional profile (80). They identified a responder- associated single- cell cluster with increased IL- 7R expression, corresponding to long- lived memory T cell programming, in agreement with our findings. We revealed that IL- 7R is an independent prognostic factor of survival in melanoma and other malignancies. These findings advance our basic understanding of antitumor memory in the context of checkpoint inhibition or surgical resection, and suggest a strategy of using high IL- 7R expression to enrich for memory T cells with superior antitumor activity from the tumor- draining lymph node, which can be augmented by epigenetic therapies for adoptive cell transfer. Adoptive cell transfer classically relies on TILs, which are largely terminally differentiated, exhausted, PNAS  2023  Vol. 120  No. 30  e2304319120 https://doi.org/10.1073/pnas.2304319120   9 of 12 and include immunosuppressive populations. The lymph node IL- 7Rhi cells with cytotoxic/memory properties and a permissive “unscarred” epigenome are prime candidates for adoptive T cell trans- fer therapies and can help improve T cell–based immunotherapy. Materials and Methods Animal Experiments. Mice were maintained at Yale University in accordance with Institutional Animal Care and Use Committee  guidelines. Mouse strains were purchased from Jackson Labs, including WT C57BL/6 (#000664), P14 TCR transgenic mice (#037394), RAG KO mice (#002216). 1 to 20 × 105. Tumors were injected subcutaneously into the flanks of 8- wk- old, age- matched C57BL/6J mice in 100 μL of PBS (GIBCO). Tumor size was measured using an electronic caliper. FTY720 (#S5002, Selleck Chemicals) was diluted in PBS and mice were injected with 3 mg/kg three times weekly for the duration of the experiment. For CD8+ depletion experiments, mice were treated twice weekly with 200 μg of anti- CD8 antibody (SI Appendix). Cell Lines and Plasmids. Cells were grown at 37 °C and 5% CO2, in DMEM F- 12 or Opti- Mem (GIBCO) media supplemented with heat- inactivated Fetal Bovine Serum (Sigma) and Pen/Strep (GIBCO). YUMM- OVA was a kind gift from Dr. Ping- Chih Ho (University of Lausanne). YUMM- GFP33/66 was generated by Gibson cloning into a lentiviral backbone, transfecting 293FS* cells harvesting virus titer and transduc- ing YUMM1.7 an MOI of two followed by selection for GFP expression by FACS (Sony SH800). MC38 was purchased from Kerafast Biotech. Sequences available in SI Appendix, Fig. S6. Adoptive T Cell Transfer. Murine melanoma tumors, lymph nodes, or spleens were processed into single- cell suspensions, stained with antibodies against mouse CD45, CD3, CD8, gp33- tetramer, CD127, and sorted using a BD Aria Cell Sorter (BD Biosciences). Sorted T cells (10,000 to 100,000) were reconstituted in PBS (Sigma) and a final volume of 100 μL was injected retroorbitally. Unless otherwise specified in figure legend, 1 × 104 CD127+ and 1 to 6 × 104 CD127- cells were transferred. ELISA. IFN- gamma, IL- 2 and TNF- a ELISAs were performed according to the manufacturer’s instructions using ELISA kits (R&D systems, DY402- 05, DY410- 05; Biolegend 430801) and the Mouse Cytokine Array Discovery Assay (EVE Technologies). Bisulfite Modification, qPCR and Methylation- Specific PCR. Of note, 1 μg of genomic DNA per sample was bisulfite converted using the Zymo EZ DNA Methylation Kit (Zymo Research). Methylation- specific PCR primers were designed using BiSearch and conversion performed on a C1000 Touch Thermal Cycler (BioRad). with ZymoTaq PreMix and HotStart Polymerase on a C1000 Touch Thermal Cycler (BioRad). For qRT- PCR, total RNA was extracted using TRIzol (Qiagen) and the RNeasy Plus mini kit (Qiagen). cDNA was made using SuperScript IV Reverse Transcriptase (Thermo Fisher Scientific) on a CFX96 Real- Time PCR System (Bio- Rad) and iTaq Universal Probes Supermix (Bio- Rad) and sequence- specific oligonucleotide primers purchased from Sigma- Aldrich. Expression values were calculated using the standard curve method. Flow Cytometry Analysis. Fluorescence spectra were acquired using a BD LSRII or BD Symphony (BD Biosciences) flow cytometers and analyzed by FlowJo (Version 10, BD). For flow cytometry analysis, splenocytes and/or fluorescent minus one staining was used for gating. Sorting was performed at the Yale Flow Cytometry Core on a BD FACSAria instrument. CD8+ T Cell Enrichment. CD8+ T cells were purified from single- cell suspension using the CD8a+ T Cell Isolation Kit (Miltenyi Biotech) according to the manu- facturer’s instructions. Murine melanoma lymph were processed into single- cell suspensions as described above, and 103 to 108 cells were incubated with an antibody cocktail (#130- 095- 236) and conjugated to magnetic beads (Miltenyi Biotech). The CD8+ fraction was allowed to elute by gravity, and the column was washed with 6 mL of wash buffer. Single- Cell RNA- Seq. Tumors or lymph nodes were dissociated and processed into single- cell suspensions and sort- purified: P1: Tetramer+CD45+CD3+ (antigen- specific T cells), P2: CD45+CD3+Tetramer− (polyclonal T cells), P3: CD45+CD3− (Non- T cell immune cells), and P4: CD45−. P1, P2, P3, and P4 were mixed at a 2:1:1:1 ratio and cells were encapsulated into droplets using 10x Chromium GEM and libraries prepared using the Single Cell 5′ Reagent Kit version 2.0 (10× Genomics) prior to sequencing using a NovaSeq instrument, as previously described (81). ATAC- Seq. Mouse lymph node tissue was processed into single- cell suspensions, and 50,000 cells were tagmented and processed following the manufacturer’s protocol (ActiveMotif). Indexed libraries were prepared and sequenced to at least 30 × 107 reads on a NovaSeq instrument, as previously described (82). Bulk RNA- Seq. RNA was isolated using the Qiagen RNeasy Plus mini kit and QC was performed on an Agilent 2,200 TapeStation. RNA with satisfactory RIN values were used for stranded library preparation and sequenced with the target of at least 3 × 107 reads per sample. Reads were quality filtered and trimmed of Illumina adapters using FastQC and Cutadapt. Filtered reads were aligned to referen genome mm10 using STAR aligner and quantified using featureCounts. Differential expression analysis was performed with DESeq2. Immunohistochemistry. Tissue microarrays were a gift from H.K. The Yale mel- anoma cohort sample collection was approved by the Yale Human Investigation Committee protocol in accordance with the Declaration of Helsinki. Antigen retrieval was performed in Target Retrieval Solution (Dako), and slides were incu- bated with primary antibodies against IL- 7R (LS- B2830- 50 LSBio), Tcf7 (2203S Rabbit mAb), or CD3 (Cell Signaling Technology 99940S). After washing, slides were incubated with biotinylated secondary antibody and Vectastain ABC kit (Vector Labs), developed using a di- amino- benzidine- peroxidase substrate kit (Vector Labs). Scoring was performed by a board- certified pathologist (HM). Statistics and Reproducibility. Statistical analyses were conducted using R v4.0.2 and Prism 7 (GraphPad). IHC staining and qPCR analyses have been repeated at least twice. At least five mice were used per experimental group. Log- rank (Mantel–Cox) tests were used for tumor survival curve statistical anal- yses. P- values for all qPCR were calculated with two- sided Student’s t test. Benjamini–Hochberg or Bonferroni correction was used for multiple statistical comparisons. Data, Materials, and Software Availability. All data supporting the findings of this study are available within the Article, the SI Appendix Data and the NCBI Sequence Read Archive (SRA) repository (PRJNA902911) (83). All study data are included in the article and/or SI Appendix. Previously published data were used for this work (72). ACKNOWLEDGMENTS. We would like to thank Dr. Qin Yan (Yale University) for constructive discussion and feedback on the manuscript. This article is sub- ject to Howard Hughes Medical Institute’s Open Access to Publications policy. Howard Hughes Medical Institute lab heads have previously granted a nonex- clusive CC BY 4.0 license to the public and a sublicensable license to Howard Hughes Medical Institute in their research articles. Pursuant to those licenses, the author- accepted manuscript of this article can be made freely available under a CC BY 4.0 license immediately upon publication. The results shown here are in part based upon data generated by the TCGA Research Network. G.M. and A.D. are supported by an NIAID- funded fellowship T32AR007016- 47 to Yale Department of Dermatology. G.M. has been supported by the Dermatology Foundation Career Development Award and American Skin Association Research Grant. R.T. is supported by an NCI- funded fellowship F30CA254246. M.W.B. is supported by NIH grants P50CA121974, U01CA233096, U01238728, P30CA016359, and a Melanoma Research Alliance Team Science Award. N.S.J. was supported by a Melanoma Research Alliance Young Investigator Award. H.K. is supported by K12CA215119, the Yale SPORE in Skin Cancer P50 CA121974, R01 CA216846. This work was also supported, in part, by the Howard Hughes Medical Institute (A.I. and R.A.F.). Author affiliations: aDepartment of Immunobiology, Yale School of Medicine, New Haven, CT 06520; bDepartment of Dermatology, Yale School of Medicine, New Haven, CT 06520; cDepartment of Pathology, Yale School of Medicine, New Haven, CT 06520; dDepartment of Surgery, Yale School of Medicine, New Haven, CT 06520; eYale Cancer Center, Yale School of Medicine, New Haven, CT 06520; fDepartment of Medicine (Medical Oncology), Yale School of Medicine, New Haven, CT 06520; gYale Stem Cell Center, Yale School of Medicine, New Haven, CT 06520; hHHMI, Chevy Chase, MD 20815; and iYale Center for Immuno- Oncology, Yale School of Medicine, New Haven, CT 06520 10 of 12   https://doi.org/10.1073/pnas.2304319120 pnas.org Author contributions: G.M., A.D., M.W.B., and R.A.F. designed research; G.M., A.D., K.F.- K., K.P., R.T., and M.M. performed research; G.M., K.F.- K., R.T., M.M., H.N.B., E.S., J.F.C., N.I.H., L.A., N.S.J., H.K., and A.I. contributed new reagents/analytic tools; G.M., A.D., and H.M. analyzed data; and G.M., A.D., M.W.B., and R.A.F. wrote the paper. Reviewers: D.F., Massachusetts General Hospital Harvard Medical School; and T.M., Weill Cornell Medicine. Competing interest statement: R.A.F. is an advisor to GlaxoSmithKline, Zai Lab and Ventus Therapeutics. H.K. has received research support from Merck, Apexigen and Bristol- Myers Squibb, and has served as a consultant for Iovance, Celldex, Merck, Bristol- Myers Squibb, Clinigen, Shionogi, Chemocentryx, Calithera, Signatero, Gigagen, GI reviewers, Seranova, Pilant Therapeutics. None of the disclosures are directly related to the work presented herein. 1. M. W. Rohaan et al., Tumor- infiltrating lymphocyte therapy or ipilimumab in advanced melanoma. 38. T. J. Fry, C. L. Mackall, The many faces of IL- 7: From lymphopoiesis to peripheral T cell maintenance. 2. N. Engl. J. Med. 387, 2113–2125 (2022). S. A. Rosenberg, N. P. Restifo, Adoptive cell transfer as personalized immunotherapy for human cancer. Science 348, 62–68 (2015). S. A. Rosenberg et al., Durable complete responses in heavily pretreated patients with metastatic melanoma using T- cell transfer immunotherapy. Clin. Cancer Res. 17, 4550–4557 (2011). 4. M. J. Besser et al., Clinical responses in a phase II study using adoptive transfer of short- term 3. cultured tumor infiltration lymphocytes in metastatic melanoma patients. Clin. Cancer Res. 16, 2646–2655 (2010). J. Immunol. 174, 6571–6576 (2005). 39. A. Ma, R. Koka, P. Burkett, Diverse functions of IL- 2, IL- 15, and IL- 7 in lymphoid homeostasis. Annu. Rev. Immunol. 24, 657–679 (2006). 40. R. Forster, A. C. Davalos- Misslitz, A. Rot, CCR7 and its ligands: Balancing immunity and tolerance. Nat. Rev. Immunol. 8, 362–371 (2008). 41. A. Bai, H. Hu, M. Yeung, J. Chen, Kruppel- like factor 2 controls T cell trafficking by activating L- selectin (CD62L) and sphingosine- 1- phosphate receptor 1 transcription. J. Immunol. 178, 7632–7639 (2007). 5. M. Philip, A. Schietinger, CD8(+) T cell differentiation and dysfunction in cancer. Nat. Rev. Immunol. 42. S. Kurtulus et al., Checkpoint blockade immunotherapy induces dynamic changes in PD- 1(- )CD8(+) 22, 209–223 (2022). 6. A. Schietinger, Turbocharging the T cell to fight cancer. N. Engl. J. Med. 386, 2334–2336 (2022). 7. N. P. Kristensen et al., Neoantigen- reactive CD8+ T cells affect clinical outcome of adoptive cell therapy with tumor- infiltrating lymphocytes in melanoma. J. Clin. Invest. 132, e150535 (2022). 8. N. Zacharakis et al., Immune recognition of somatic mutations leading to complete durable 9. regression in metastatic breast cancer. Nat. Med. 24, 724–730 (2018). E. Tran et al., T- cell transfer therapy targeting mutant KRAS in cancer. N. Engl. J. Med. 375, 2255–2262 (2016). tumor- infiltrating T cells. Immunity 50, 181–194.e186 (2019). 43. G. Jeannet et al., Essential role of the Wnt pathway effector Tcf- 1 for the establishment of functional CD8 T cell memory. Proc. Natl. Acad. Sci. U.S.A. 107, 9777–9782 (2010). 44. D. Pais Ferreira et al., Central memory CD8(+) T cells derive from stem- like Tcf7(hi) effector cells in the absence of cytotoxic differentiation. Immunity 53, 985–1000.e1011 (2020). 45. S. J. Im et al., Defining CD8+ T cells that provide the proliferative burst after PD- 1 therapy. Nature 537, 417–421 (2016). 46. A. B. Borowski et al., Memory CD8+ T cells require CD28 costimulation. J. Immunol. 179, 10. Y. Simoni et al., Bystander CD8(+) T cells are abundant and phenotypically distinct in human 6494–6503 (2007). tumour infiltrates. Nature 557, 575–579 (2018). 47. J. M. Grayson, A. J. Zajac, J. D. Altman, R. Ahmed, Cutting edge: Increased expression of Bcl- 2 in 11. P. D. A. Vignali et al., Hypoxia drives CD39- dependent suppressor function in exhausted T cells to antigen- specific memory CD8+ T cells J. Immunol. 164, 3950–3954 (2000). limit antitumor immunity. Nat. Immunol. 24, 267–279 (2023). 48. S. Wojciechowski et al., Bim/Bcl- 2 balance is critical for maintaining naive and memory T cell 12. S. Krishna et al., Stem- like CD8 T cells mediate response of adoptive cell immunotherapy against homeostasis. J. Exp. Med. 204, 1665–1675 (2007). 13. human cancer. Science 370, 1328–1334 (2020). I. Siddiqui et al., Intratumoral Tcf1(+)PD- 1(+)CD8(+) T cells with stem- like properties promote tumor control in response to vaccination and checkpoint blockade immunotherapy. Immunity 50, 195–211.e110 (2019). 49. S. Kurtulus et al., Bcl- 2 allows effector and memory CD8+ T cells to tolerate higher expression of Bim. J. Immunol. 186, 5729–5737 (2011). 50. T. Wen et al., NKG7 Is a T- cell- intrinsic therapeutic target for improving antitumor cytotoxicity and cancer immunotherapy. Cancer Immunol. Res. 10, 162–181 (2022). 14. J. A. Fraietta et al., Determinants of response and resistance to CD19 chimeric antigen receptor 51. S. S. Ng et al., The NK cell granule protein NKG7 regulates cytotoxic granule exocytosis and (CAR) T cell therapy of chronic lymphocytic leukemia. Nat. Med. 24, 563–571 (2018). inflammation. Nat. Immunol. 21, 1205–1218 (2020). 15. J. Huang et al., Survival, persistence, and progressive differentiation of adoptively transferred 52. X. Y. Li et al., NKG7 is required for optimal antitumor T- cell Immunity. Cancer Immunol. Res. 10, tumor- reactive T cells associated with tumor regression. J. Immunother. 28, 258–267 (2005). 16. L. Zhang et al., Lineage tracking reveals dynamic relationships of T cells in colorectal cancer. Nature 564, 268–272 (2018). 154–161 (2022). 53. A. Marcais et al., Cell- autonomous CCL5 transcription by memory CD8 T cells is regulated by IL- 4. J. Immunol. 177, 4451–4457 (2006). 17. A. Gros et al., PD- 1 identifies the patient- specific CD8(+) tumor- reactive repertoire infiltrating 54. A. Crawford, J. M. Angelosanto, K. L. Nadwodny, S. D. Blackburn, E. J. Wherry, A role for the human tumors. J. Clin. Invest. 124, 2246–2259 (2014). 18. G. A. Fenton, D. A. Mitchell, Cellular cancer immunotherapy development and manufacturing in the clinic. Clin. Cancer Res. 29, 843–857 (2023). chemokine RANTES in regulating CD8 T cell responses during chronic viral infection. PLoS Pathog. 7, e1002098 (2011). 55. T. J. Schall, K. Bacon, K. J. Toy, D. V. Goeddel, Selective attraction of monocytes and T lymphocytes of 19. K. B. Yates et al., Epigenetic scars of CD8(+) T cell exhaustion persist after cure of chronic infection in the memory phenotype by cytokine RANTES. Nature 347, 669–671 (1990). humans. Nat. Immunol. 22, 1020–1029 (2021). 20. M. S. Abdel- Hakeem et al., Epigenetic scarring of exhausted T cells hinders memory differentiation 56. T. Wex et al., Human cathepsin W, a cysteine protease predominantly expressed in NK cells, is mainly localized in the endoplasmic reticulum. J. Immunol. 167, 2172–2178 (2001). upon eliminating chronic antigenic stimulation. Nat. Immunol. 22, 1008–1019 (2021). 57. C. Stoeckle et al., Cathepsin W expressed exclusively in CD8+ T cells and NK cells, is secreted during 21. K. E. Pauken et al., Epigenetic stability of exhausted T cells limits durability of reinvigoration by PD- 1 target cell killing but is not essential for cytotoxicity in human CTLs. Exp. Hematol. 37, 266–275 (2009). blockade. Science 354, 1160–1165 (2016). 22. B. C. Miller et al., Subsets of exhausted CD8(+) T cells differentially mediate tumor control and respond to checkpoint blockade. Nat. Immunol. 20, 326–336 (2019). 58. G. Pizzolato et al., Single- cell RNA sequencing unveils the shared and the distinct cytotoxic hallmarks of human TCRVdelta1 and TCRVdelta2 gammadelta T lymphocytes. Proc. Natl. Acad. Sci. U.S.A. 116, 11906–11915 (2019). 23. J. M. Angelosanto, S. D. Blackburn, A. Crawford, E. J. Wherry, Progressive loss of memory T cell potential 59. S. M. Kaech, E. J. Wherry, R. Ahmed, Effector and memory T- cell differentiation: Implications for and commitment to exhaustion during chronic viral infection. J. Virol. 86, 8161–8170 (2012). 24. D. T. Utzschneider et al., T cell factor 1- expressing memory- like CD8(+) T cells sustain the immune vaccine development. Nat. Rev. Immunol. 2, 251–262 (2002). 60. W. Leung, H. E. Heslop, Adoptive immunotherapy with antigen- specific T cells expressing a native response to chronic viral infections. Immunity 45, 415–427 (2016). TCR. Cancer Immunol. Res. 7, 528–533 (2019). 25. T. Wu et al., The TCF1- Bcl6 axis counteracts type I interferon to repress exhaustion and maintain T cell stemness. Sci. Immunol. 1, eaai8593 (2016). 26. W. H. Hudson et al., Proliferating transitory T cells with an effector- like transcriptional signature emerge from PD- 1(+) stem- like CD8(+) T cells during chronic infection. Immunity 51, 1043–1058. e1044 (2019). 27. C. A. Klebanoff et al., Central memory self/tumor- reactive CD8+ T cells confer superior antitumor immunity compared with effector memory T cells. Proc. Natl. Acad. Sci. U.S.A. 102, 9571–9576 (2005). 28. K. A. Connolly et al., A reservoir of stem- like CD8(+) T cells in the tumor- draining lymph node preserves the ongoing antitumor immune response. Sci. Immunol. 6, eabg7836 (2021). 29. K. Meeth, J. X. Wang, G. Micevic, W. Damsky, M. W. Bosenberg, The YUMM lines: A series of congenic mouse melanoma cell lines with defined genetic alterations. Pigment Cell Melanoma Res. 29, 590–597 (2016). 61. C. Yee et al., Adoptive T cell therapy using antigen- specific CD8+ T cell clones for the treatment of patients with metastatic melanoma: in vivo persistence, migration, and antitumor effect of transferred T cells. Proc. Natl. Acad. Sci. U.S.A. 99, 16168–16173 (2002). 62. H. R. Kim, K. A. Hwang, K. C. Kim, I. Kang, Down- regulation of IL- 7Ralpha expression in human T cells via DNA methylation. J. Immunol. 178, 5473–5479 (2007). 63. B. H. Ladle et al., De novo DNA methylation by DNA methyltransferase 3a controls early effector CD8+ T- cell fate decisions following activation. Proc. Natl. Acad. Sci. U.S.A. 113, 10631–10636 (2016). 64. C. D. Scharer, B. G. Barwick, B. A. Youngblood, R. Ahmed, J. M. Boss, Global DNA methylation remodeling accompanies CD8 T cell effector function. J. Immunol. 191, 3419–3429 (2013). 65. D. Sutter, W. Doerfler, Methylation of integrated adenovirus type 12 DNA sequences in transformed cells is inversely correlated with viral gene expression. Proc. Natl. Acad. Sci. U.S.A. 77, 253–256 (1980). 30. Y. R. Yu et al., Disturbed mitochondrial dynamics in CD8(+) TILs reinforce T cell exhaustion. Nat. 66. M. Busslinger, J. Hurst, R. A. Flavell, DNA methylation and the regulation of globin gene expression. Immunol. 21, 1540–1551 (2020). Cell 34, 197–206 (1983). 31. J. Wang et al., UV- induced somatic mutations elicit a functional T cell response in the YUMMER1.7 67. A. P. Bird, A. P. Wolffe, Methylation- induced repression–belts, braces, and chromatin. Cell 99, mouse melanoma model. Pigment Cell Melanoma Res. 30, 428–435 (2017). 451–454 (1999). 32. S. Mandala et al., Alteration of lymphocyte trafficking by sphingosine- 1- phosphate receptor 68. C. Luo, P. Hajkova, J. R. Ecker, Dynamic DNA methylation: In the right place at the right time. Science agonists. Science 296, 346–349 (2002). 361, 1336–1340 (2018). 33. S. M. Kaech et al., Selective expression of the interleukin 7 receptor identifies effector CD8 T cells 69. S. Henikoff, J. G. Henikoff, H. S. Kaya- Okur, K. Ahmad, Efficient chromatin accessibility mapping that give rise to long- lived memory cells. Nat. Immunol. 4, 1191–1198 (2003). in situ by nucleosome- tethered tagmentation. Elife 9, e63274 (2020). 34. J. C. Rathmell, E. A. Farkash, W. Gao, C. B. Thompson, IL- 7 enhances the survival and maintains the size of naive T cells. J. Immunol. 167, 6869–6876 (2001). 35. L. M. Bradley, L. Haynes, S. L. Swain, IL- 7: Maintaining T- cell memory and achieving homeostasis. Trends Immunol. 26, 172–176 (2005). 70. R. R. Jadhav et al., Epigenetic signature of PD- 1+ TCF1+ CD8 T cells that act as resource cells during chronic viral infection and respond to PD- 1 blockade. Proc. Natl. Acad. Sci. U.S.A. 116, 14113–14118 (2019). 71. A. Barski et al., Rapid recall ability of memory T cells is encoded in their epigenome. Sci. Rep. 7, 36. S. L. Colpitts, N. M. Dalton, P. Scott, IL- 7 receptor expression provides the potential for long- term 39785 (2017). survival of both CD62Lhigh central memory T cells and Th1 effector cells during leishmania major infection. J. Immunol. 182, 5702–5711 (2009). 72. Cancer Genome Atlas, Genomic classification of cutaneous melanoma. Cell 161, 1681–1696 (2015). 37. J. T. Barata, S. K. Durum, B. Seddon, Flip the coin: IL- 7 and IL- 7R in health and disease. Nat. 73. J. Reiser, A. Banerjee, Effector, memory, and dysfunctional CD8(+) T cell fates in the antitumor Immunol. 20, 1584–1593 (2019). immune response. J. Immunol. Res. 2016, 8941260 (2016). PNAS  2023  Vol. 120  No. 30  e2304319120 https://doi.org/10.1073/pnas.2304319120   11 of 12 74. M. Morotti et al., Promises and challenges of adoptive T- cell therapies for solid tumours. Br. J. Cancer 79. M. Sade- Feldman et al., Defining T cell states associated with response to checkpoint 124, 1759–1776 (2021). immunotherapy in melanoma. Cell 175, 998–1013.e1020 (2018). 75. K. Okamura et al., Lymphocytes in tumor- draining lymph nodes co- cultured with autologous tumor 80. A. Jaiswal et al., An activation to memory differentiation trajectory of tumor- infiltrating cells for adoptive cell therapy. J. Transl. Med. 20, 241 (2022). 76. S. L. Wong et al., Sentinel lymph node biopsy and management of regional lymph nodes in melanoma: American society of clinical oncology and society of surgical oncology clinical practice guideline update. J. Clin. Oncol. 36, 399–413 (2018). 77. B. Prinzing et al., Deleting DNMT3A in CAR T cells prevents exhaustion and enhances antitumor activity. Sci. Transl. Med. 13, eabh0272 (2021). 78. N. Jain et al., TET2 guards against unchecked BATF3- induced CAR T cell expansion. Nature 615, 315–322 (2023), 10.1038/s41586- 022- 05692- z. lymphocytes informs metastatic melanoma outcomes. Cancer Cell 40, 524–544.e525 (2022). 81. T. Zhou et al., IL- 18BP is a secreted immune checkpoint and barrier to IL- 18 immunotherapy. Nature 583, 609–614 (2020). 82. S. M. Zhang et al., KDM5B promotes immune evasion by recruiting SETDB1 to silence retroelements. Nature 598, 682–687 (2021). 83. G. Micevic et al., IL- 7R licenses a population of epigenetically poised memory CD8+ T- cells with superior antitumor efficacy that are critical for melanoma memory. NCBI Sequence Read Archive. https://www.ncbi.nlm.nih.gov/bioproject/PRJNA9029115. Deposited 17 November 2022. 12 of 12   https://doi.org/10.1073/pnas.2304319120 pnas.org
10.7554_elife.83884
RESEARCH ARTICLE Lipid homeostasis is essential for a maximal ER stress response Gilberto Garcia1,2, Hanlin Zhang1, Sophia Moreno1, C Kimberly Tsui1, Brant Michael Webster1, Ryo Higuchi- Sanabria2, Andrew Dillin1* 1Department of Molecular & Cellular Biology, Howard Hughes Medical Institute, University of California, Berkeley, Berkeley, United States; 2Leonard Davis School of Gerontology, University of Southern California, Los Angeles, United States Abstract Changes in lipid metabolism are associated with aging and age- related diseases, including proteopathies. The endoplasmic reticulum (ER) is uniquely a major hub for protein and lipid synthesis, making its function essential for both protein and lipid homeostasis. However, it is less clear how lipid metabolism and protein quality may impact each other. Here, we identified let- 767, a putative hydroxysteroid dehydrogenase in Caenorhabditis elegans, as an essential gene for both lipid and ER protein homeostasis. Knockdown of let- 767 reduces lipid stores, alters ER morphology in a lipid- dependent manner, and blocks induction of the Unfolded Protein Response of the ER (UPRER). Interestingly, a global reduction in lipogenic pathways restores UPRER induction in animals with reduced let- 767. Specifically, we find that supplementation of 3- oxoacyl, the predicted metabolite directly upstream of let- 767, is sufficient to block induction of the UPRER. This study high- lights a novel interaction through which changes in lipid metabolism can alter a cell’s response to protein- induced stress. Editor's evaluation In this work, the often- surmised but still- poorly understood connection between lipid metabolism and ER stress was explored. The work using genetic techniques and a variety of parallel approaches, implicates key lipid synthetic pathways in the rise and strength of an ER stress signal. The studies as submitted were strong, and create a compelling case for these new connections between lipid metabolism and cellular stress response. Introduction The cell must monitor protein and lipid quality to maintain cellular homeostasis. Disruptions in protein folding have been implicated in numerous neurodegenerative diseases, while lipid imbalances result in an increased risk for cardiovascular disease, diabetes, and various cancers (Hartl, 2017; Sletten et  al., 2018; Chaurasia and Summers, 2015; Avgerinos et  al., 2019). These two distinct meta- bolic pathways have been extensively studied independently, yet increasing evidence suggests that examining their relationship to one another may provide novel insights into human health. Individuals with Alzheimer’s disease (AD) and Parkinson’s disease (PD), diseases known to be associated with increased development of abnormal protein aggregates, also show dysregulation of lipid metabolism (Yerbury et al., 2016; Alecu and Bennett, 2019; Kao et al., 2020). As such, changes in lipid metab- olism are now suspected to contribute to the development of these proteopathies through mecha- nisms that are not fully understood (Chang et al., 2017; Chew et al., 2020; Xicoy et al., 2019). The endoplasmic reticulum (ER) is a major metabolic hub, responsible for the synthesis of secreted and integral proteins, as well as a major portion of a cell’s lipids (Schwarz and Blower, 2016). The close *For correspondence: dillin@berkeley.edu Competing interest: The authors declare that no competing interests exist. Funding: See page 17 Received: 01 October 2022 Preprinted: 27 October 2022 Accepted: 08 May 2023 Published: 25 July 2023 Reviewing Editor: Randy Y Hampton, University of California, San Diego, United States Copyright Garcia et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 1 of 21 Research article relationship between protein and lipid quality control within the ER is highlighted by the Unfolded Protein Response of the ER (UPRER), a system capable of sensing and responding to both protein misfolding and membrane lipid disequilibrium (Metcalf et al., 2020; Xu and Taubert, 2021). In higher eukaryotes, the UPRER contains three unfolded protein sensors, each with their own inde- pendent signaling pathways. While all three sensors are ER- localized transmembrane proteins with a luminal unfolded protein sensing domain, the most conserved of these pathways is the inositol- requiring enzyme- 1 (ire- 1) branch. The IRE- 1 luminal domain is bound by the resident ER HSP70 chap- erone (C. elegans HSP- 4) under basal conditions. Upon protein folding stress, the HSP70 chaperone is titrated away to allow for the interaction of the luminal domain with misfolded proteins, leading to oligomerization and activation of IRE- 1’s cytosolic RNase domain. Activated IRE- 1 then initiates noncanonical splicing of xbp- 1 mRNA from its xbp- 1u form to its effective xbp- 1s variant at the ER membrane (Gómez- Puerta et al., 2022). The XBP- 1s transcription factor is then able to upregulate the cell’s UPRER target genes to increase protein folding, protein turnover, and lipid metabolism to help ameliorate the stress. (Frakes and Dillin, 2017; Adams et al., 2019). All three of the UPRER sensors are also able to sense membrane lipid disequilibrium through domains adjacent to their transmembrane helices. These domains activate the UPRER independent of proteotoxic stress and the luminal sensing domains (Volmer and Ron, 2015; Halbleib et al., 2017; Tam et al., 2018). This is unsurprising considering the ER’s critical role in the synthesis of major lipids including membrane lipids, cholesterol, and neutral lipids (Lodhi and Semenkovich, 2014; Fagone and Jackowski, 2009). Accordingly, the ER actively regulates the cell’s lipid status through changes to lipid synthesis enzymes, transcriptions factors, and interactions with lipid droplets (LDs), organ- elles tasked with the storage and gatekeeping of surplus lipid stores (Shimano and Sato, 2017; Jacquemyn et al., 2017; Olzmann and Carvalho, 2019). Utilizing the same stress response sensors for protein and membrane lipid stress suggests an inter- dependent link between lipid and protein homeostasis within the ER. Ectopic activation of the UPRER results in changes to lipid metabolism, while changes in sphingolipid, lipid saturation, ceramides, and loss of various lipid enzymes activate the UPRER (Volmer and Ron, 2015; Tam et al., 2018; Imanikia et al., 2019; Daniele et al., 2020; Contreras et al., 2014; Promlek et al., 2011). Interestingly, both lipids and LDs have been shown to contribute to proteostasis. A sterol pathway localized to LDs is required for clearing inclusion bodies and LDs themselves are necessary for transport of damaged proteins (Moldavski et al., 2015; Vevea et al., 2015). Whether other lipid pathways in the ER or on LDs are required for an effective UPRER response to protein stress has not been fully investigated. Here, we performed a genetic screen of LD- associated genes to identify genes whose knockdown affected UPRER activation in the nematode C. elegans. We identified the hydroxysteroid dehydro- genase, let- 767, as required for both ER lipid and protein homeostasis. Loss of let-767 resulted in severe defects in UPRER activation, lipid homeostasis, and ER morphology. Defects in ER morphology could be rescued through lipid supplementation; however, the deficiencies in UPRER activation were not. Instead, knockdown of the upstream regulator of the let- 767 pathway resulted in a significant recovery of the UPRER activation. Thus, we propose that loss of let- 767 results in accumulation of fatty acid metabolites which leads to detrimental remodeling of the ER membrane and disruption of ER functions in lipid synthesis and UPRER induction. Our work highlights a unique cellular interaction in which lipid metabolism negatively impacts ER protein homeostasis. Further understanding of how these two systems influence one another may bring new insights into the mechanisms of protein and lipid disorders. Results LET-767 is a regulator of the UPRER To identify novel lipid genes that influence the UPRER, we carried out an RNAi screen under condi- tions of proteotoxic stress. We utilized the hsp- 4 (C. elegans Hsp70/BiP) transcriptional GFP reporter (hsp- 4p::GFP) to assess the UPRER induction, and sec- 11 (an ER serine- peptidase) RNAi to generate ER stress and induce the GFP reporter (Calfon et al., 2002). In combination with the sec- 11 RNAi, we individually knocked down each candidate gene to perform a double RNAi screen (Figure 1A). We focused on proteins associated with LDs instead of general lipid synthesis genes due to the LD’s central role in lipid regulation and their contribution to ER proteostasis (Moldavski et  al., 2015; Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 2 of 21 Cell Biology | Genetics and Genomics Research article A hsp-4p::GFP ER Stress B P F G : : p 4 - p s h sec-11 RNAi (ER stress) + Candidate RNAi (Lipid gene) EV sec-11 tag-335 EV let-767 EV let-767 EV let-767 Fluorescence Score -3 0 3 l i d e (cid:31) h g i r B Untreated EV let-767 Heatshocked EV let-767 D P F G : : p 2 . 6 1 - p s h i l d e (cid:31) h g i r B E ) 0 0 0 0 1 x ( . U . F . A d e z i l a m r o N 4 3 2 1 0 EV cco-1 EV let-767 EV let-767 n. s. F P F G : : p 6 - p s h i l d e (cid:31) h g i r B let-767 EV Untreated EV let-767 Heatshocked C G ) 0 0 0 0 1 x ( . U . F . A d e z i l a m r o N 4 3 2 1 0 ) 0 0 0 0 1 x ( . U . F . A d e z i l a m r o N 4 3 2 1 0 **** **** EV let-767 EV EV let-767 sec-11 let-767 EV tag-335 **** EV let-767 EV EV let-767 cco-1 Figure 1. Knockdown of let- 767 specifically suppresses the UPRER. (A) Schematic for screening method used to identify UPRER modulators from candidate genes. Animals expressing hsp- 4p::GFP were grown from L1 on candidate RNAi mixed in a 1:1 Ratio with ER stress inducing sec- 11 RNAi. Animals were then screened at day 1 of adulthood and scored for changes in fluorescence compared to the sec- 11/Empty Vector (EV) control. (B) Fluorescent micrographs of day 1 adult transgenic animals expressing hsp- 4p::GFP grown from L1 on EV, sec- 11, or tag- 335 RNAi combined in a 1:1 ratio with either EV or let- 767 RNAi to assay effects on UPRER induction. (C) Quantification of (B) normalized to size using a BioSorter. Lines represent mean and standard deviation. n=500. Mann- Whitney test p- value ****<0.0001. Representative data shown is one of three biological replicates. (D) Fluorescent micrographs of day 1 adult transgenic animals expressing hsp- 16.2p::GFP grown from L1 on EV or let- 767 RNAi with or without 2 hr 34 °C heat- shock treatment to assay heat shock response. Animals imaged 2 hr after recovery at 20 °C. (E) Quantification of (D) normalized to size using a BioSorter. Lines represent mean and standard deviation. n=400. Mann- Whitney test n.s.=not significant. Representative data shown is 1 of 3 biological replicates. (F) Fluorescent micrographs of day 1 adult transgenic animals expressing hsp- 6p::GFP, grown from L1 on EV or cco- 1 RNAi combined in a 1:1 ratio with either EV or let- 767 RNAi to assay effects on UPRmt induction. (G) Quantification of (F) normalized to size using a BioSorter. Lines represent mean and standard deviation. n=431. Mann- Whitney test p- value ****<0.0001. Representative data shown is one of three biological replicates. The online version of this article includes the following figure supplement(s) for figure 1: Figure supplement 1. Knockdown of let- 767 suppresses the UPRER more severely than other lipid related genes. Vevea et  al., 2015). By cross- referencing two independent C. elegans LD proteomic datasets, we identified 163 high- confidence LD proteins (Zhang et al., 2012; Na et al., 2015; Vrablik et al., 2015). We found that RNA interference of 49 genes led to reduced induction of the UPRER reporter (Supple- mentary file 1). While a large portion of these genes are annotated as functioning in general transla- tion (e.g. ribosomal subunits), potentially affecting global gene expression, a subset of 11 belonged to other functional groups and were therefore more likely to specifically affect the UPRER (Figure 1— figure supplement 1A). From this subset of genes only let- 767, an acyl- CoA reductase, was noted as having lipid enzymatic activity, while the other genes function as ATP synthases, GTPases, heat shock proteins, ATPases, methyltransferases, and tubulin. We focused on let- 767 of the screen subset due to its direct role in lipid metabolism, which has been previously characterized (Entchev et al., 2008; Desnoyers et al., 2007; Kuervers et al., 2003). Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 3 of 21 Cell Biology | Genetics and Genomics Research article let- 767 has been implicated in long- chain fatty acid (LCFA) and monomethyl branched- chain fatty acid (mmBCFA) synthesis as an acyl- CoA reductase, and in steroid metabolism as a 17- beta hydroxysteroid dehydrogenase that directly acts on steroid molecules. Within the fatty acid elongation pathway, after addition of a malonyl- CoA, acyl- CoA reductase is responsible for reduction of the 3- ketoacyl- CoA into a 3- hydroxyacyl- CoA which is then further processed into a fatty acid with 2 more carbon molecules. Knockdown of let- 767 suppressed the UPRER induction caused by RNAi of two different ER genes, sec- 11, which would impact ER signal peptide cleavage, and tag- 335, a GDP- mannose pyrophosphor- ylase, which would reduce protein glycosylation (Figure  1B–C; Bar- Ziv et  al., 2020). Furthermore, qPCR analysis confirmed that let- 767 RNAi also resulted in reduced splicing of xbp- 1u to xpb- 1s while under stress, suggesting that loss of let- 767 suppresses the UPRER induction by preventing IRE- 1’s splicing of xbp- 1 (Figure 1—figure supplement 1B). Next, we utilized the transcriptional reporters of the mitochondrial unfolded protein response (UPRmt, hsp- 6p::GFP) and the heat shock response (HSR, hsp- 16.2::GFP) to determine the effect of let- 767 knockdown on other stress responses. We observed that let- 767 RNAi does not suppress the heat shock response and only has minor effects on the mitochondrial stress response in comparison to the highly suppressed UPRER (Figure 1D–G). Together, these results show that the function of let- 767 is specifically important for ER function and homeostasis. Both fatty acid elongation and steroid processing is carried out by steroid hydrogenases within mammals (Sakurai et al., 2006). To determine whether reduced UPRER induction was a general pheno- type of steroid dehydrogenase knockdown, we performed our double RNAi protocol on the four most closely related (steroid dehydrogenase family) genes found in C. elegans, stdh- 1,2,3 and 4. Knockdown of any of the stdh genes had only mild effects on the UPRER induction, showing that the diminished UPRER induction phenotype was more specific to let- 767 and not a general phenomenon of decreased steroid dehydrogenase function (Figure 1—figure supplement 1C–D). We then tested whether RNAi of other genes implicated in the LCFA and mmBCFA pathways also affected the UPRER induction (Zhang et al., 2011; Kniazeva et al., 2004; Kniazeva et al., 2012). Of these genes, acs- 1 had the most similar phenotype to let- 767, while elo- 5 and hpo- 8 had a significant, but less severe effects on the UPRER induction without stalling development, like pod- 2 (Figure  1—figure supple- ment 1E–F). acs- 1 functions in the first step of fatty acid processing by attaching a CoA to the fatty acid, particularly for the mmBCFA isoC17 (Kniazeva et al., 2012; Zhang et al., 2021). While elo- 5 and hpo- 8 function as a fatty acid elongase and a 3- hydroxyacyl- CoA dehydratase, the steps before and after let- 767 in the fatty elongation pathway, respectively. These results indicate that perturbation of the LCFA/mmBCFA pathways negatively impacts the UPRER, with let- 767 and acs- 1 being the most critical. Loss of LET-767 function impacts the UPRER independent of lipid depletion Phenotypes caused by a reduction in the LET- 767 enzyme could likely result from insufficient produc- tion of key metabolites, i.e., mmBCFAs or LCFAs. The UPRER induction might then be restored to wild- type levels by supplementation of these lipids. To this end, we supplemented animals with a crude lysate composed of homogenized N2 (wild type) adult animals to provide a complete panel of lipids. We observed that supplementation of lysate was sufficient to rescue the phenotypes of acs- 1 RNAi, including suppression of the UPRER upon ER stress (Figure 2—figure supplement 1A–B). Phenotypes of acs- 1 RNAi have been shown to be a results of insufficient lipid species, suggesting that the lysate is sufficient to rescue deficiencies in lipids (Zhang et al., 2021). However, lysate supplementation of let- 767 RNAi treated animals resulted in a significant improvement in organismal size, but only a slight improvement in the UPRER induction (Figure 2A–B). The ability of lysate supplementation to rescue a known essential lipid phenotype, but not suppression of the UPRER from let- 767 RNAi, would suggest a potential mechanistic difference between the UPRER and size phenotypes. Although the reduced animal size is likely the result of insufficient lipids, the suppression of the UPRER is potentially due to other complications caused by knockdown of the let- 767 pathway. To investigate whether lysate supplementation had an impact on ER and lipid subcellular pheno- types of let- 767 RNAi, we further characterized the impact of let- 767 RNAi and lysate supplemen- tation on the LD and ER morphologies. We found that knockdown of let- 767 caused an extreme reduction in LD size, from large spheres to small points (Figure 2C). let- 767 knockdown also resulted Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 4 of 21 Cell Biology | Genetics and Genomics Research article Control EV sec-11 Lysate Supplement EV sec-11 EV let-767 EV let-767 EV let-767 EV let-767 A P F G : : p 4 - p s h B ) 0 0 0 0 1 x ( . U . F . A d e z i l a m r o N 4 3 2 1 0 D **** **** **** EV let-767 EV EV let-767 sec-11 Control EV let-767 EV EV let-767 sec-11 Lysate Supplement EV let-767 EV let-767 L E D H : : y b u R m l o r t n o C l t n e m e p p u S e t a s y L l i d e (cid:31) h g i r B C P F G : : 3 - S H D l o r t n o C l t n e m e p p u S e t a s y L Figure 2. Supplementation of lysate does not restore the UPRER suppressed by let- 767 RNAi. (A) Fluorescent micrographs of transgenic animals expressing hsp- 4p::GFP grown on Empty Vector (EV) or let- 767 RNAi combined in a 1:1 ratio with either EV or sec- 11 RNAi supplemented with vehicle or N2 lysate to assay effects on UPRER induction. (B) Quantification of (A) normalized to size using a BioSorter. Lines represent mean and standard deviation. n=189. Mann- Whitney test p- value ****<0.0001. Representative data shown is one of three biological replicates. (C) Representative fluorescent micrograph projections of day 1 adult transgenic animal expressing LD- localized dhs- 3::GFP, grown on EV or let- 767 RNAi with or without N2 lysate supplementation to assay LD quality. Yellow arrowheads point to example lipid droplets. Scale bar, 5 μm. (D) Representative fluorescent micrograph projections of day 1 adult transgenic animal expressing ER lumen- localized mRuby::HDEL, grown on EV or let- 767 RNAi with or without N2 lysate supplementation to assay ER quality. Orange arrowheads point to wide ER structures. Scale bar, 5 μm. Images for organelle markers individually contrasted for clarity. The online version of this article includes the following figure supplement(s) for figure 2: Figure supplement 1. Supplementation of mmBCFA or oleic acid does not restore the UPRER suppressed by let-RNAi. Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 5 of 21 Cell Biology | Genetics and Genomics Research article in substantial changes to the ER morphology, practically eliminating the wider structures of the ER. However, unlike the UPRER induction, supplementation of lysate improved the organelle morphology appearance by restoring the presence of wide ER structures and lipid droplets, albeit at considerably lower abundance and size (Figure 2C–D). These observations mark a distinction between the ability of supplementation to rescue morphological and functional phenotypes of the ER caused by let- 767 RNAi. The rescue of animal size and ER morphology phenotypes by lysate supplementation hint that they are a result of insufficient lipids, while the abilities of the ER to induce the UPRER and expand lipid droplets are potentially not due to the depleted levels of lipids. One potential reason for the ineffective rescue of the UPRER induction by lysate could be insuffi- cient levels of lipids within the lysate. We sought to determine whether excess supplementation of two lipids associated with let- 767, mmBCFAs or LCFAs (Entchev et al., 2008), could rescue the UPRER induction of animals with reduced let- 767. We supplemented animals grown on let- 767 and sec- 11 RNAi with exogenous isoC17, an essential mmBCFA, or oleic acid, a LCFA known to be a signifi- cant component of the C. elegans fatty acid content and to increase lifespan (Imanikia et al., 2019; Kniazeva et al., 2004; Henry et al., 2016). mmBCFA supplementation was not sufficient to rescue induction of the UPRER in let- 767 knockdown animals, showing only a slight improvement in the stress response induction (Figure 2—figure supplement 1C–D). However, mmBCFA supplementation was sufficient to rescue phenotypes of acs- 1 RNAi, providing evidence that our supplementation effec- tively delivered the essential mmBCFA and was also sufficient to rescue functional phenotypes of mmBCFA insufficiency (Figure 2—figure supplement 1E–F). Similar to isoC17, supplementation of oleic acid showed a very minor improvement in the ER stress response of let- 767 knockdown animals (Figure 2—figure supplement 1G–H). These data show that while the UPRER dysfunction caused by let- 767 knockdown is likely not due to the specific loss of the essential mmBCFA or LCFA (oleic acid), isoC17 is indeed essential to the UPRER in the absence of acs- 1. Knockdown of lipid biosynthesis pathways restores the UPRER signaling under let-767 RNAi Since supplementation of a WT mixture of lipids was not sufficient to recover the UPRER induction, but was able to rescue ER morphology, we considered whether let- 767 knockdown could be compro- mising the ER membrane through unbound LET- 767 partners or accumulation of upstream meta- bolic intermediates. A disrupted membrane could hinder ER membrane protein function for ER stress signaling as well as for lipid droplet and lipid synthesis enzymes, explaining both phenotypes and why the restored ER morphology did not coincide with restored UPRER function. Indeed, a disrupted ER membrane affecting ER function was suggested as the mechanistic cause of acs- 1 phenotypes, where insufficient isoC17 disrupted the ER membrane quality and hindered lipid droplet produc- tion (Zhang et al., 2021). Furthermore, ER stress through sec- 11 RNAi alone caused a reduction in let- 767 transcript levels (Figure  1—figure supplement 1B), suggesting that a reduction of let- 767 itself is not directly suppressing the UPRER induction, but possibly dependent on the levels of other factors. Instead, we hypothesized that disequilibrium within the let- 767 pathway could be the cause of membrane disruption and ultimately the loss of UPRER induction. To identify a lipid that could disrupt the ER membrane, we performed untargeted complex lipidomic analysis of worms treated with let- 767 RNAi (Supplementary file 3, Source data 1). We observed a reduction in most lipids, including nearly all triglycerides, which would be expected from the reduction in lipid droplets. However, we were not able to identify a potentially accumulated lipid, with many unknown features unable to be accurately identified. Therefore, we tested whether reducing potentially accumulated upstream metabolites of let- 767 would improve the UPRER activation. Since the complete enzymatic pathway for let- 767 has yet to be definitively identified and previous works have implicated the enzyme in multiple pathways, we accom- plished this by knocking down the ortholog of human SREBP, sbp- 1, a major transcriptional regulator of numerous lipogenic enzymes and pathways (Nomura et al., 2010). Analysis of a published RNAseq dataset of sbp- 1 RNAi treated nematodes showed downregulation of numerous lipid synthesis genes including the fatty elongation pathway genes, such as hpo- 8, elo- 1/2/4/5/6/9, and let- 767, the latter of which we confirmed through qPCR (Figure 3A; Lee et al., 2015). We found that animals grown on let- 767 and sbp- 1 RNAi appeared larger and healthier than on let- 767 RNAi alone. More impor- tantly, knockdown of sbp- 1 was able to significantly improve the UPRER reporter induction in let- 767 Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 6 of 21 Cell Biology | Genetics and Genomics Research article EV sec-11/EV/EV sec-11/let-767/EV sbp-1/EV/EV sec-11/let-767/sbp-1 B P F G : : p 4 - p s h EV sbp-1 EV sec-11 EV sec-11 EV let-767 EV let-767 EV let-767 EV let-767 i l d e (cid:31) h g i r B EV let-767/EV sbp-1/EV let-767/sbp-1 25 ng/uL Tunicamycin 0 5 10 15 20 25 Day E 100 l a v i v r u S t n e c r e P 50 EV let-767/EV sbp-1/EV let-767/sbp-1 DMSO 0 0 10 20 Day 30 40 xbp-1s xbp-1 Total let-767 **** **** D 100 l a v i v r u S t n e c r e P 50 0 EV let-767 EV EV let-767 sec-11 EV let-767 EV EV let-767 sec-11 EV EV sbp-1 EV let-767 EV let-767 sbp-1 EV + Lysate let-767 + Lysate + Lysate + Lysate A C F e g n a h C d o F l 4 3 2 1 0 ) 0 0 0 0 1 x ( . U . F . A d e z i l a m r o N 4 3 2 1 0 L E D H : : y b u R m P F G : : 3 - S H D Figure 3. Reduced global lipid synthesis rescues the UPRER suppression caused by let- 767 RNAi. (A) Quantitative RT- PCR transcript levels of xbp- 1 total, xbp- 1s, and let- 767 from day 1 adult N2 animals grown from L1 on EV, sec- 11 and EV, let- 767 and sec- 11 RNAi combined with either EV or sbp- 1 RNAi in a 1:1:1 ratio. Fold- change compared to EV treated N2 animals. Lines represent standard deviation across three biological replicates, each averaged from two technical replicates. (B) Fluorescent micrographs of day 1 adult transgenic animals expressing hsp- 4p::GFP grown on EV, let- 767, sec- 11, and/ or sbp- 1 RNAi mixed in a 1:1:1 ratio to assay effects on the UPRER induction. (C) Quantification of (D) normalized to size using a BioSorter. Lines represent mean and standard deviation. n=426. Mann- Whitney test p- value ****<0.0001. Representative data shown is one of three biological replicates. (D– E) Concurrent survival assays of N2 animals transferred to ER stress conditions of 25 ng/uL Tunicamycin (D) or control DMSO (E) conditions at day 1 of adulthood. Animals continuously grown on EV or let- 767 RNAi combined in 1:1 ratio with EV or sbp- 1 RNAi from L1 synchronization. (F) Representative fluorescent micrograph projections of day 1 adult transgenic animal expressing ER lumen- localized mRuby::HDEL or LD- localized dhs- 3::GFP, grown on EV or let- 767 RNAi mixed in 1:1 ratio with sbp- 1 RNAi with or without N2 lysate supplementation to assay ER and LD quality. Yellow arrowheads point to wide ER structures. Orange arrowheads point to lipid droplets. Scale bar, 5 μm. Images for organelle markers individually contrasted for clarity. knockdown animals under ER stress from sec- 11 RNAi and effectively restore splicing of xbp- 1 to xbp- 1s (Figure 3A–C). To test whether the improved UPRER signaling was indeed a functional restoration of the ER stress response, we performed an ER stress survival assay. As expected, animals treated with let- 767 RNAi have a severe defect in survival when exposed to the proteotoxic stress of Tunicamycin, an inhibitor of N- linked glycosylation. Correlating with our observations of the UPRER reporter, sbp- 1 knockdown rescued the survival of let- 767 RNAi treated animals to the level of sbp- 1 RNAi alone (Figure 3D–E, Supplementary file 2). Interestingly, sbp- 1 RNAi itself reduced the induction level of Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 7 of 21 Cell Biology | Genetics and Genomics Research article the UPRER reporter and the survival rate of animals on Tunicamycin, suggesting that while a reduction in the entire let- 767 pathway could indeed rescue the UPRER suppression caused by let- 767 RNAi, a global reduction in lipid synthesis enzymes limits the maximal induction of the UPRER. Finally, we determined whether sbp- 1 RNAi would also rescue the defects in LD and ER morphology caused by let- 767 knockdown. Congruent with SREBP’s central role in promoting lipogenesis, we observed a depletion of lipid droplets and a slight perturbation of the ER morphology with sbp- 1 RNAi (Figure  3F). In combination with let- 767 RNAi, sbp- 1 knockdown did not rescue the ER and lipid droplet morphology to WT conditions, highlighting that restoring the UPRER signaling is not completely dependent on restoring ER lipid levels and morphology. However, in the presence of lysate supplementation, sbp- 1 RNAi alone or combined with let- 767 RNAi resulted in ER and lipid droplet morphology resembling that of WT animals with the presence of wide sheet- like structures and larger lipid droplets. Therefore, a more complete rescue of the phenotypes exhibited by let- 767 knockdown animals could be achieved by a combination of (1) the reduction of the let- 767 pathway through knockdown of sbp- 1 to reduce global lipid synthesis, and (2) exogenous supplementation of the lipids to restore lipids lost by sbp- 1 and let- 767 RNAi. let-767 knockdown impacts xbp-1 splicing and xbp-1s activity Proper UPRER signaling is dependent on dimerization of IRE- 1 at the ER membrane to splice xbp- 1 mRNA to its active xbp- 1s isoform. Therefore, one possible mechanism by which loss of let- 767 can impact the UPRER is by impeding IRE- 1 activity through altered membrane dynamics. Indeed, knock- down of let- 767 results in altered ER membrane dynamics where the mobile fraction of an ER trans- membrane protein (SPCS- 1) is significantly reduced when measured by Fluorescence Recovery After Photobleaching (FRAP) (Figure  4A–B). The percent mobile being the calculated fraction of fluoro- phore that can move within the bleached area during the FRAP process. Supplementation of lysate improved the ER membrane mobility to 62% of WT levels. Comparatively, ER luminal mRuby dynamics, which would likely be indirectly affected by the global changes in ER 3D structure, were less severely impacted by let- 767 RNAi. These effects were rescued to a greater extent by lysate supplementation, potentially due to the recovered ER structure, restoring the mobile fraction of let- 767 RNAi treated animals to 84% of WT levels (Figure 4C–D). While specific changes in dynamics are likely dependent on the individual protein being observed, our results provide evidence that membrane dynamics have been altered when let- 767 is knocked down. To determine whether the UPRER was indeed being affected at the ER membrane, we exam- ined the impact of let- 767 knockdown on ectopic UPRER activation at two different points along the mechanistic pathway: (1) overexpression of ire- 1, which would ectopically activate the UPRER at the membrane by constitutively splicing xbp- 1 and (2) overexpression of the active xbp- 1s isoform, which would bypass the requirement of IRE- 1 splicing at the ER membrane (Li et al., 2010). In creating the necessary strains, we found that intestinal over- expression of the full- length ire- 1a isoform proved to be lethal. However, overexpression of the ire- 1b isoform lacking the luminal domain, fused with the mRuby fluorophore (mRuby::ire- 1b), was viable and had constitutive activation of the UPRER. To ensure that our fusion protein was responsible for activating the UPRER, we knocked down ire- 1 through an RNAi targeting the N- terminal region only found in ire- 1a. The ire- 1a RNAi was able to suppress the UPRER signaling in WT animals with and without ER stress but did not eliminate the basal UPRER signal in animals expressing mRuby::ire- 1b. Conversely, RNAi targeting common sequences of both ire- 1a and ire- 1b suppressed the UPRER induction in both WT and animals expressing our mRuby::ire- 1b construct, even when ER stress was induced by sec- 11 RNAi. (Figure 5—figure supplement 1A–D). Therefore, our mRuby::ire- 1b construct was sufficient to induce the intestinal UPRER without the endog- enous ire- 1a. Likewise, we confirmed that overexpression of our intestinal mRuby::xbp- 1s construct was able to induce the UPRER in an xbp- 1 (zc12) null- mutant background (Figure 5—figure supple- ment 1E–F). Beyond observing that our construct was sufficient to induce the UPRER, we also found that loss of endogenous xbp- 1 resulted in an even higher level of basal UPRER induction, potentially due to a negative regulatory role for unspliced xbp- 1, that has been previously suggested (Yoshida et al., 2006). Next, we tested UPRER induction of our ire- 1 and xbp- 1s overexpression strains on let- 767 RNAi. let- 767 knockdown reduced the induction of the UPRER reporter in mRuby::ire- 1b animals (Figure 5A–B). However, let- 767 RNAi also reduced the UPRER reporter induction in the mRuby::xbp- 1s, xbp- 1(zc12) Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 8 of 21 Cell Biology | Genetics and Genomics Research article A 1.0 y t i s n e t n I e v (cid:31) a e R l 0.5 0.0 0 C 1.0 0.5 y t i s n e t n I e v (cid:31) a e R l ER Membrane-bound GFP EV let-767/EV EV + lysate let-767 /EV + lysate 50 100 150 200 250 Time (s) ER Luminal mRuby EV Control let-767/EV Control EV + Lysate let-767 + Lysate B 80 60 e l i 40 b o M % 20 0 D 100 80 **** **** **** EV let-767 EV Control EV let-767 EV Lysate Supp. * **** **** e l i b o M % 60 40 0.0 0 20 40 60 Time (s) 80 100 20 0 EV let-767 EV EV Control let-767 EV Lysate Supp. Figure 4. Lysate supplementation rescues luminal ER protein dynamics but not ER membrane protein dynamics. (A) FRAP curve of intestinal ER- transmembrane protein, SPCS- 1::GFP, from day 1 adult animals grown on Empty Vector (EV) or let- 767 RNAi mixed with EV in a 1:1 ratio supplemented with vehicle or N2 lysate (geometric mean for n=20 pooled from two biological replicates). Lines represent geometric standard deviation. (B) Calculated percent mobile SPCS- 1::GFP of (A). Lines represent geometric mean and geometric standard deviation. Mann- Whitney test p- value ****<0.0001. (C) FRAP curve of intestinal ER- lumen protein, mRuby::HDEL, from day 1 adult animals grown on EV or let- 767 RNAi mixed with EV in a 1:1 ration supplemented with vehicle or N2 lysate (geometric mean for n=20 pooled from two biological replicates). Lines represent geometric standard deviation. (D) Calculated percent mobile mRuby::HDEL of (A). Lines represent geometric mean and geometric standard deviation. Mann- Whitney test p- value ****<0.0001 and *<0.05. strain (Figure 5C–D). By utilizing the mRuby::xbp- 1s, xbp- 1 (zc12) null- mutant we eliminated the possi- bility that the reduced induction was due to the negative regulatory effects of unspliced xbp- 1s. To ensure that the reduced UPRER induction was not due to altered expression of our mRuby::xbp- 1s construct, we performed qPCR for xbp- 1 in animals grown on let- 767 RNAi. We observed a significant drop in the percent of xbp- 1 splicing in animals expressing mRuby::ire- 1b when they were grown on let- 767 RNAi, but not in animals expressing mRuby::xbp- 1s (Figure 5E–F). However, the reduced UPRER induction despite the similar levels of xbp- 1s transcripts between EV and let- 767 RNAi suggests that in addition to affecting xbp- 1 splicing, let- 767 RNAi is likely affecting xbp- 1s activity downstream of its splicing. Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 9 of 21 Cell Biology | Genetics and Genomics Research article A B EV EV let-767 , E O b 1 - e r i : : y b u R m P F G : : p 4 - p s h ) 0 0 0 0 1 x ( . U . F . A d e z i l a m r o N 4 3 2 1 0 l i d e (cid:31) h g i r B E e g n a h C d o F l 2.0 1.5 1.0 0.5 0.0 EV let-767 EV *** n.s. EV let-767 /EV xbp-1s xbp-1 Total C D EV EV let-767 **** , E O s 1 - p b x : : y b u R m ) 2 1 c z ( 1 - p b x P F G : : p 4 - p s h i l d e (cid:31) h g i r B ) 0 0 0 0 1 x ( . U . F . A d e z i l a m r o N 4 3 2 1 0 **** EV let-767 EV F e g n a h C d o F l 1.5 1.0 0.5 0.0 ** * EV let-767/EV xbp-1s xbp-1 Total Figure 5. let- 767 knockdown impacts UPRER induction independent of xbp- 1 splicing. (A) Fluorescent micrographs of day 1 adult transgenic animals expressing hsp- 4p::GFP and intestinal mRuby::ire- 1b grown on Empty Vector (EV) or let- 767 RNAi mixed with EV in a 1:1 ratio to assay the UPRER induction. (B) Quantification of (A) normalized to size using a BioSorter. Lines represent mean and standard deviation. n=290. Mann- Whitney test P- value ****<0.0001. Representative data shown is 1 of 3 biological replicates. (C) Fluorescent micrographs of day 1 adult xbp- 1(zc12) transgenic animals expressing hsp- 4p::GFP and intestinal mRuby::xbp- 1s grown on EV or let- 767 RNAi mixed with EV in a 1:1 ratio to assay the UPRER induction. (D) Quantification of (C) normalized to size using a BioSorter. Lines represent mean and standard deviation. n=398. Mann- Whitney test p- value ****<0.0001. Representative data shown is one of three biological replicates. (E) Quantitative RT- PCR transcript levels of xbp- 1s and total xbp- 1 from day 1 adult mRuby::ire- 1b animals grown from L1 on let- 767 RNAi mixed 1:1 with EV. Fold- change compared to EV treated animals.Unpaired t- test p- value ***<0.0005. Error bars indicate ± standard deviation across three biological replicates, each averaged from two technical replicates. (F) Quantitative RT- PCR transcript levels of xbp- 1s and total xbp- 1 from day 1 adult mRuby::xbp- 1s animals grown from L1 on let- 767 RNAi mixed 1:1 with EV. Fold- change compared to EV treated animals. Unpaired t- test p- value **<0.005 and *<0.05. Error bars indicate ± standard deviation across three biological replicates, each averaged from two technical replicates. Dots indicate averaged biological replicate values. The online version of this article includes the following figure supplement(s) for figure 5: Figure supplement 1. Overexpression of ire- 1b or xbp- 1s induces the UPRER independently of endogenous ire- 1a or xbp- 1, respectively. The let-767 upstream metabolite, 3-oxostearic acid, reduces induction of the UPRER To identify a node within the let- 767 pathway that might be responsible for the disruptive metabolite or protein, we utilized mammalian cell culture to probe the effect of specific metabolites in the let- 767 pathway on ER homeostasis. This would allow us to saturate cellular exposure to lipid intermediates Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 10 of 21 Cell Biology | Genetics and Genomics Research article Figure 6. Fatty acid intermediate, 3- oxoacyl, is sufficient to reduce UPRER induction. (A) Fatty acid elongation pathway displaying intermediate lipid metabolites and annotated C. elegans genes. (B) Flow cytometry measurement histogram of huh7 UPRE reporter fluorescence normalized to EIF2A promoter driving mCherry. Cells were treated for 18 hr with Tunicamycin and vehicle, 50 μM Cerulenin, 50 μM 3- oxostearic acid, 50 μM stearic acid, or 50 μM 3- hydroxystearic acid. Data is representative of three biological replicates. (C) Median bar graph of (B). Unpaired t- test p- value **=0.004 and ****<0.0001. Error bars indicate ± standard deviation across three technical replicates. and avoid the possibility of bacteria processing the intermediates prior to digestion by C. elegans animals. Previous works have provided evidence that the human ortholog of let- 767 are HSD17B12 and HSD17B3 (Entchev et al., 2008). Interestingly, LET- 767 shares the most protein sequence percent identity (39.87%) with HSD17B12, the 3- ketoacyl reductase, which caused similar phenotypes in mice when knocked out. HSD17B12 was found to be essential for development in mice and knockout of the gene in adult mice resulted in reduced body weight, reduced lipid content, and caused liver toxicity hypothesized to be from accumulation of toxic intermediates (Rantakari et al., 2010; Heikelä et al., 2020). 3- Ketoacyl reductases perform the second step in fatty acid synthesis/elongation, metabolizing 3- oxoacyl- CoA to 3- hydroxyacyl- CoA (Figure  6A; Moon and Horton, 2003). While the fatty acid elongation pathway has not been shown to affect the UPRER, use of the fatty acid synthase inhibitor, Cerulenin, has been shown to increase levels of XBP1s with reduced transcriptional activity due to changes in palmitoylation (Chen et al., 2020). From these studies, we hypothesized that the accumu- lation of the metabolites upstream of LET- 767 could be reducing the transcriptional activity of XBP- 1S. To this end, we tested whether an upstream metabolite of LET- 767 was sufficient to reduce UPRER induction in huh7 cells containing a 5 x unfolded protein response element (UPRE) GFP reporter and EIF2A promoter driving mCherry. As fatty acid elongation extends fatty acids beyond the 16 carbons synthesized by fatty acid synthase, we supplemented our reporter line with the 18 carbon fatty acid metabolites upstream and downstream of the 3- ketoacyl reductase in combination with Tunicamycin to induce ER stress (Figure  6B–C). We observed that the upstream metabolite, 3- oxostearic acid, reduced the normalized 5xUPRE reporter induction. In comparison, the downstream metabolites, 3- hydroxystearic acid and stearic acid, did not reduce the normalized 5xUPRE reporter signal. Instead, stearic acid caused a slight induction of the reporter signal, in agreement with studies showing that saturated lipids induce the UPRER (Volmer et al., 2013). This indicated that the 5xUPRE was specifically Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 11 of 21 Cell Biology | Genetics and Genomics Research article Figure 7. Model for let- 767 RNAi blocking UPRER induction. (A) Under WT condition, acyl- CoA metabolites are elongated by the let- 767/HSD17B12 pathway and utilized at the membrane to synthesize other lipids such as neutral lipids stored in lipid droplets. IRE- 1 responds to ER stress by splicing xbp- 1u to xbp- 1s, which is then translated and able to induce expression of the UPRER target genes (B) Knockdown of let- 767 results in disequilibrium of let- 767/HS17B12 pathway, leading to accumulation of intermediates such as 3- oxoacyl- CoA and reduced lipid production. Intermediate metabolites disrupt membrane quality (dashed line) and negatively affect induction of the UPRER by reducing splicing of xbp- 1s (smaller arrows) by IRE- 1 and reducing the function of xbp- 1s post- splicing (smaller arrows). sensitive to the 3- oxo metabolite upstream of the let- 767 pathway and that this interaction between lipid metabolism and the UPRER is likely conserved in mammalian cells. Discussion Cells must monitor and maintain both lipid and protein homeostasis to preserve cellular function. The ER is uniquely a major site of both protein and lipid synthesis. Here, we performed a genetic screen to identify lipid related genes that impact the ER response to protein stress and found let- 767 to be necessary for proper animal size, organelle morphology, neutral lipid accumulation, and induction of the UPRER. Under WT conditions, the let- 767/HSD17B12 pathway elongates fatty acids that are then used to produce other lipids, such as triglycerides stored in lipid droplets (Figure 7A). The UPRER is also able to respond to ER stress by inducing XBP- 1s activity. When the let- 767 is knocked down, the pathway is unable to elongate fatty acids and accumulates the upstream metabolite, 3- oxoacyl- CoA, and possibly other metabolic intermediates that disrupt the ER membrane quality (Figure 7B). Cells are unable to synthesize downstream lipids for lipid droplets and either through disruption of the membrane or direct interaction with 3- oxoacyl- CoA, xbp- 1 splicing and the activity of spliced xbp- 1s is significantly reduced in the presence of protein induced ER stress. While unfolded proteins and lipid disequilibrium are known to independently induce the UPRER, our novel finding demonstrates that changes in lipid pathways could significantly impact a cell’s ability to respond to protein stress and brings to light a potential mechanism through which lipid disequilibrium might facilitate the progres- sion of proteopathic diseases. LET- 767 has been characterized as a hydroxysteroid dehydrogenase with evidence as both a steroid modifying enzyme and a 3- ketoacyl reductase located on the ER, consistent with mammalian HSD17B3 and HSD17B12 (Entchev et  al., 2008; Desnoyers et  al., 2007). LET- 767 has also been implicated as a requirement for branched- chain and long- chain fatty acid production, however, its linear metabolic pathway has not been thoroughly investigated so its exact lipid products or protein interactors are yet to be identified. Through our depletion of mmBCFA and LCFA pathway genes, we find that both pathways are essential to having a maximal ER unfolded protein response to protein stress, with acs- 1 RNAi having the most similar phenotypes to let- 767 knockdown. While individual supplementation of mmBCFAs or LCFAs did not rescue the let- 767 RNAi phenotypes, supplementa- tion of a more complex lipid mixture, crude worm lysate, was able to significantly rescue organelle Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 12 of 21 Cell Biology | Genetics and Genomics Research article morphology, size, and reproduction phenotypes. This suggests that loss of let- 767 has a more global effect on lipid production rather than affecting a single lipid species. A straightforward explanation for its requirement in global lipid equilibrium and the UPRER could be that the let- 767 pathway is crucial for an elemental component of the ER, such as production of fundamental lipids required for ER membrane quality as has been shown for the acs- 1 pathway (Zhang et al., 2021). The compromised ER integrity caused by acs- 1 RNAi demonstrates that a in addition to affecting lipid droplet produc- tion, a disrupted ER membrane can impact UPRER induction and that these phenotypes can be rescued through correction of lipid deficiencies such as lipid supplementation. We would then expect that let- 767 levels would be critical for ER functions, including UPRER signaling. However, the inability of supplementation to rescue let- 767 RNAi phenotypes would suggest a different mechanism for the compromised ER membrane integrity. Interestingly, ER stress results in reduced let- 767 transcript levels. Furthermore, the reduced UPRER signaling caused by let- 767 RNAi was significantly rescued by knockdown of the upstream transcription factor, sbp- 1, which also reduced let- 767 transcript levels and lipid stores. A possible interpretation of these results is that an alternative pathway to produce the essential lipids is upregulated under ER stress or sbp- 1 RNAi, but with the existence of an alternative pathway, the UPRER would not be suppressed by knockdown of let- 767. Instead, we propose that the let- 767 RNAi phenotypes are caused by disequilibrium of the fatty elongation pathway at the let- 767 node. By reducing this specific node of the fatty elongation pathway, intermediate metabolites upstream of let- 767 could interfere with ER membrane dynamics, interactions, and functions. A disrupted ER membrane would also explain the similarities in phenotypes with acs- 1 RNAi, which has been shown to affect ER integrity and which we show here can impact UPRER induction (Zhang et  al., 2021). As the UPRER sensors and numerous lipogenic enzymes reside on the ER membrane, a disrupted ER membrane would explain why knockdown of let- 767 and other lipid genes affects the capacity of the UPRER induction. While knockdown of sbp- 1 also reduced the level of UPRER induction, the reduction of numerous lipogenic pathways, including the entire let- 767 pathway, was able to improve the UPRER function of animals on let- 767 RNAi compared to let- 767 RNAi alone. Mechanistically, this could function by preventing the lipid disequilibrium caused by intermediates of the fatty acid elongation pathway by reducing the entire pathway instead of just one node. While our untargeted lipidomic analysis did not reveal an increase in any specific lipid, it did confirm that let- 767 RNAi disrupts lipid levels globally. With multiple types of fatty acids of different lengths being elongated by the same pathway, identifying a single lipid species may prove difficult since numerous 3- oxoacyls may be contributing to the change in ER membrane quality. To test whether the ER membrane disorganization was the source of UPRER dysfunction, we aimed to bypass the splicing of xbp- 1 by IRE- 1 at the ER membrane by overexpressing the already spliced isoform, xbp- 1s. However, let- 767 RNAi still caused a reduction in the ER stress response. The reduced UPRER induction in animals overexpressing xbp- 1s points to an additional mechanism downstream of splicing for the muted UPRER. The regulation of XBP1 mRNA has been proven to be more complex than simple splicing upon ER stress. XBP1u mRNA is required to localize to the ER membrane to facili- tate splicing to XBP1s upon ER stress through a hydrophobic region and translational stalling (Gómez- Puerta et al., 2022; Yanagitani et al., 2009; Yanagitani et al., 2011). Without proper localization, xbp- 1u would be translated and negatively regulate any xbp- 1s (Yoshida et al., 2006). Additionally, there are potentially other factors residing on the membrane that are required for effective translation of xbp- 1s, such as specific ribosomal complexes or mRNA stabilizing factors. An altered ER membrane would impact every interaction of xbp- 1 mRNA with the ER membrane, including transient interac- tions that have yet to be discovered. While splicing is a major point of regulation for xbp- 1, previous studies have found other potential nodes of regulation for xbp- 1 through SUMOylation, deacetylation, and palmitoylation (Chen et al., 2020; Bang et al., 2019; Jiang et al., 2012). Disruption of the ER membrane by let- 767 RNAi could impact these other pathways affecting fatty acid synthase activity or protein palmitoylation. Through supplementation of fatty acid elongation intermediates to human huh7 hepatocytes, we provide evidence that increased levels of the 3- oxoacyl metabolite that is upstream of the let- 767 ortholog, HSD17B12, a 3- ketoacyl reductase, is sufficient to reduce the ER stress response to a similar level as the fatty acid synthase inhibitor, Cerulenin. Mechanistically, Cerulenin has been proposed to prevent palmitoylation of XBP1S which reduced its transcriptional activity without impacting its protein levels Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 13 of 21 Cell Biology | Genetics and Genomics Research article (Chen et al., 2020). Whether the 3- ketoacyl metabolite impacts fatty acid synthase activity or protein palmitoylation requires further investigation. Fatty acids exist in numerous configurations of varying lengths and combination of branches and double bonds. These lipids are one of the basic building blocks that compose the membranes of the cell and their qualities can have drastic effects on membrane thickness, fluidity, and curva- ture (Harayama and Riezman, 2018). Therefore, it is not surprising that alterations in lipids such as mmBCFA levels can impact ER function (Zhang et al., 2021). Our work demonstrates that incomplete processing of lipids can also impact ER function, in our case by knock down a key enzyme within the fatty elongation pathway. Further work into how different nodes within lipid pathways and how levels of lipid intermediates change with age would greatly contribute to growing amount research on the relationship between lipid changes and aging, including age- dependent neurological diseases (Gille et al., 2021; Yoon et al., 2022). However, this work may prove to be labor intensive due to the complexity of lipidomics and the possible disconnect between transcript levels, protein levels, and enzymatic activity with age. [hsp- 4p:GFP] V), SJ4100 Materials and methods Nematode strains N2 Bristol, LIU1 (ldrIs[dhs- 3p::dhs- 3::GFP]), SJ4005 (zcIs4[hsp- 4p::GFP]), SJ17 (xbp- 1(zc12) III; zcIs4 (dvIs70[hsp- 16.2p::GFP]), VS25 (zcIs13[hsp- 6p::GFP]), CL2070 (hjIs[vha- 6p::GFP::C34B2.10(SP12)  +unc- 119(+)]),EG6703 (unc- 119(ed3); cxTi10816; oxEx1582[eft- 3p::GFP  +Cbr- unc- 119]) strains were obtained from the Caenorhabditis Genetics Center (CGC). AGD2192 (uthSi60[vha- 6p::ER- signal- sequence::mRuby::HDEL::unc- 54 UTR, cb- unc- 119(+)] I; unc- 119(ed3) III) (Daniele et  al., 2020). Transgenic strains created for this study were generated from EG6703 via the MosSCI method (Yanagitani et al., 2011) or through crossing strains. Transgenic strains created: AGD2424 (unc- 119(ed3) III; uthSi65[vha- 6p::ERss::mRuby::ire- 1a (344- 967aa)::unc- 54 3'UTR cb- unc- 119(+)] IV) AGD2425 (uthSi65[vha- 6p::ERss::mRuby::ire- 1a (344- 967aa)::unc- 54 3'UTR cb- unc- 119(+)] IV; zcls4[hsp- 4p::GFP] V) AGD2012 (unc- 119(ed3) III; uthSi71[vha- 6p::mRuby::xbp- 1s::unc- 54 3'UTR cb- unc- 119(+)] IV) AGD2735 (uthSi71[vha- 6p::mRuby::xbp- 1s::unc- 54 3'UTR cb- unc- 119(+)] IV; zcls4[hsp- 4p::GFP] V) AGD2996 (xbp- 1(zc12) III; uthSi71[vha- 6p::mRuby::xbp- 1s::unc- 54 3'UTR cb- unc- 119(+)] IV; zcls4[hsp- 4p::GFP] V) Worm growth and maintenance All worms were maintained at 20 °C on NGM agar plates seeded with OP50 E. coli bacteria. Prior to experiments, worms were bleach synchronized as described in Higuchi- Sanabria et  al., 2018, followed by overnight L1 arrest in M9 buffer (22  mM KH2PO4 monobasic, 42.3  mM Na2HPO4, 85.6 mM NaCl, 1 mM MgSO4) at 20 °C. For RNAi experiments, arrested L1s were plated on 1 µm IPTG, 100 µg/mL Carbenicillin NGM agar plates maintained at 20 °C and seeded with RNAi bacteria grown in LB +100 µg/mL Carbenicillin. Fluorescent microscopy Transcriptional reporter strains were imaged using a Leica DFC3000 G camera mounted on a Leica M205 FA microscope. Worms were grown to day 1 of adulthood at 20 °C, hand- picked, and immo- bilized with 100 mM Sodium Azide M9 buffer on NGM agar plates. Raw images were cropped, and contrast matched using ImageJ software. For the initial screen of hsp- 4p::GFP reporter animals, fluorescence was scored by eye using the following criteria: 2=increased fluorescence, 1=possible increase in fluorescence, 0=no change, –0.5=small regions of dimmer fluorescence, –1=small regions of complete loss of fluores- cence, –1.5=globally dimmer fluorescence and some regions of no fluorescence, –2=global  loss Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 14 of 21 Cell Biology | Genetics and Genomics Research article of fluorescence and small regions of dim fluorescence, –2.5=global loss of fluorescence except for regions within spermatheca, –3=complete loss of fluorescence. Confocal images were acquired using a Leica Stellaris 5 confocal platform with a ×63 objective. Day 1 worms were picked onto 100 mM Sodium Azide M9 buffer on slides and imaged within 30 min. Raw images were cropped and independently contrast optimized for clarity using ImageJ software. Biosorter analysis Transcriptional reporter strains were grown to day 1 of adulthood at 20 °C. Animals were collected into a 15 mL conical tube with M9 buffer and allowed to settle at the bottom before supernatant was aspirated. Animals were resuspended in M9 buffer and analyzed using a Union Biometrica COPAS Biosorter (P/N: 350- 5000- 000) as described in Bar- Ziv et al., 2020. Animals which saturated the signal capacity of 65532 or were outliers in both animal size parameters (i.e., Extinction and Time Of Flight) were censored. Mann- Whitney statistical tests were performed on fluorescence normalized to animal extinction using GraphPad Prism software. Integrated fluorescence is normalized to integrated extinc- tion (as a proxy for size) during quantification. Biosorter plots shown are of populations from a single experiment. qPCR Animals grown on EV or RNAi bacteria were collected at day 1 of adulthood using M9 and washed 3 x. M9 was aspirated and trizol added before 3 cycles of freeze/thaw in liquid nitrogen. Chloroform was then added at a ratio of 1:5 (chloroform:trizol). Separation of RNA was performed through centrif- ugation in gel phase- lock tubes. The RNA aqueous phase was transferred into new tubes containing isopropanol. RNAi was purified using the QIAGEN RNeasy Mini Kit (74106) according to the manufac- turer’s instructions. cDNA was synthesized using 2 µg of RNA and the QIAGEN QuantiTect Reverse Transcriptase kit (205314) according to the manufacturer’s instructions. qPCR was performed using SYBR- green. Analysis was performed for each biological replicate using the Delta- Delta CT method with pmp- 3, cdc- 42, and Y45F10D.4 as housekeeping genes (Hoogewijs et al., 2008). Lysate supplementation Crude lysate was obtained from N2 worms grown on 40  x concentrated EV bacteria at 20 °C.~120,000 day 1 adult worms were collected with M9 and washed 6 x with M9 buffer. Animals were then homogenized 20 x in 3 mL of M9 buffer using an ice- cold 15 mL Dura- Grind Stainless Steel Dounce Tissue Grinder (VWR, 62400–686). Crude lysate was transferred to 1.5 mL tubes and frozen in liquid N2. Supplementation experiments were prepared by mixing 4 x concentrated RNAi bacteria with crude lysate at a 2:1 ratio, respectively. The lysate mixture was plated on RNAi plates and allowed to dry. Dried plates were then UV irradiated without lids for 9 min in an ultraviolet crosslinker (UVP, CL- 1000) before plating L1 arrested worms on the plate. Lipid supplementation Lipid supplementation experiments were prepared by inoculating cultures (LB +100 µg/mL Carbeni- cillin) with RNAi bacteria or empty vector and then adding the lipids to the specified concentration or equal volumes of ethanol. The cultures were allowed to grow overnight to saturation and then concentrated to 4 x before being plated on RNAi plates. The bacteria was allowed to dry overnight and then UV irradiated without lids for 9 min in an ultraviolet crosslinker (UVP, CL- 1000) before plating L1 arrested worms on the plate. Tunicamycin survival assay Tunicamycin survival assays were conducted on NGM agar plates containing 25 ng/µL of tunicamycin in DMSO, or equal volume of DMSO with specified RNAi bacteria at 20 °C. Animals were moved daily for 4–7 days to new RNAi plates until progeny were no longer observed. Worms with protruding intes- tines, bagging phenotypes, or other forms of injury were scored as censored and not counted as part of the analysis. For combined RNAi lifespans, saturated cultures were mix 1:1 by volume. Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 15 of 21 Cell Biology | Genetics and Genomics Research article FRAP analysis FRAP 4D images were acquired using a 3i Marianas spinning- disc confocal platform. Photobleaching of a 10 µm x 10 µm region within a 4 µm Z- stack was performed using a 488 nm laser for 1–2ms. Raw images were processed with FIJI (Schindelin et al., 2012) into sum Z- projections and aligned using the ‘RigidBody’ setting of the StackReg ImageJ plugin (StackReg, 2021). FRAP analysis was then performed with the FRAP Profiler ImageJ plugin (FRAP profiler plugin, 2021) by selecting the 10 µm x 10 µm photobleached region as region 1 and the entire fluorescent area as region 2. Briefly, the plugin calculates percent mobile by fitting an exponential curve to the plotted normalized recovered fluorescence values to define the recovery curve and determine the fraction/percent that is able to move within the ROI. Lipidomic analysis N2 animals were grown on 40  x concentrated EV or let- 767 RNAi bacteria at 20  °C, respectively. Animals were collected with M9 and frozen in liquid nitrogen to be submitted to the UC Davis West Coast Metabolomics Center (WCMC) for untargeted complex lipid analysis. Peak values of lipids identified by the WCMC were first corrected by subtracting corresponding peaks of blanks and then their abundance estimated relative to internal standards of their corresponding lipid class before calculating average fold change from 6 biological replicates. Cell Culture Cells were grown in DMEM media (11995, Thermo Fisher) supplemented with 2  mM GlutaMAX (35050, Thermo Fisher), 10% FBS (VWR), Non- Essential Amino Acids (100 X, 11140, Thermo Fisher), and Penicillin- Streptomycin (100 X, 15070, Thermo Fisher) in 5% CO2 at 37 °C. Huh7 were obtained from UC Berkeley Cell Culture Facility (https://bds.berkeley.edu/facilities/cell-culture#cells) and their identity confirmed through their STR profling (CELL LINE AUTHENTICATION (ucberkeleydnaseq uencing.com)). Mycoplasma negative status confirmed by PCR Detection Kit. Generation of Huh7 cell lines To generate a UPRER transcriptional reporter, we designed a lentiviral vector that encodes sequence for 5 x UPR response element upstream of a minimal cFos promoter, driving sfGFP (Adamson et al., 2016). The sfGFP is fused to a PEST sequence for tighter regulation of the reporter (Corish and Tyler- Smith, 1999; Loetscher et  al., 1991). The vector also allows for constitutive mCherry expression for normalization of the sfGFP signal, and neomycin resistant gene for selection. The transcriptional reporter was then transduced through lentivirus into Cas9- expressing Huh7 cells and selected for using G418 at 800 µg/mL. Cell culture supplementation experiments Cells were plated onto 10  cm plates and allowed to grow overnight. Lipids, tunicamycin, and/or vehicle were directly added to media. Cells were harvested for flow cytometry 16–18  hours after addition of supplements. Flow Cytometry Cells were trypsinized on ice and resuspended in cold FACS buffer (PBS with 0.1% BSA and 2 mM EDTA). Samples were filtered through a 50 µm nylon filter mesh to remove clumps prior to analysis with a five- laser LSR Fortessa (BD Bioscience). Acquired data were analyzed using FlowJo 10.7.2. Acknowledgements We are grateful to all the members of the Dillin lab for intellectual and technical support. We are grateful to the Dernburg lab for use of their Marianas spinning- disc confocal platform. We are thankful to Rebecca A Kohnz for her help with analysis of the lipidomics data. This work was supported by the following grants: GG is supported by T32AG052374, HZ is supported by 2020- A- 018- FEL through the Larry L Hillblom Foundation, CKT is supported by F32AG069388 from the National Institute on Aging, RHS is supported by R00AG065200 from the National Institute on Aging, and AD is supported by R01AG059566 from the National Institute on Aging and the Howard Hughes Medical Institute. Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 16 of 21 Cell Biology | Genetics and Genomics Research article Some strains were provided by the CGC, which is funded by the NIH Office of Research Infrastructure Programs (P40 OD010440). Additional information Funding Funder Grant reference number Author National Institute on Aging T32AG052374 Gilberto Garcia Larry L. Hillblom Foundation 2020-A-018-FEL Hanlin Zhang National Institute on Aging F32AG069388 C Kimberly Tsui National Institute on Aging R00AG065200 Ryo Higuchi-Sanabria National Institute on Aging R01AG059566 Andrew Dillin The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Author contributions Gilberto Garcia, Conceptualization, Data curation, Formal analysis, Validation, Investigation, Visualiza- tion, Methodology, Writing - original draft, Writing - review and editing; Hanlin Zhang, Formal analysis, Investigation, Visualization, Methodology, Writing - review and editing; Sophia Moreno, Investigation; C Kimberly Tsui, Resources, Writing - review and editing; Brant Michael Webster, Resources; Ryo Higuchi- Sanabria, Supervision, Investigation, Writing - review and editing; Andrew Dillin, Resources, Supervision, Investigation, Project administration, Writing - review and editing Author ORCIDs Gilberto Garcia Hanlin Zhang C Kimberly Tsui Andrew Dillin http://orcid.org/0000-0001-9959-650X http://orcid.org/0000-0001-9353-6071 http://orcid.org/0000-0002-3807-5329 http://orcid.org/0000-0002-7427-2629 Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.83884.sa1 Author response https://doi.org/10.7554/eLife.83884.sa2 Additional files Supplementary files • Supplementary file 1. Candidate LD proteins identified by proteome meta- analysis and screen score. Proteins identified in meta- analysis of published LD isolation proteomes (Zhang et al., 2012; Na et al., 2015; Vrablik et al., 2015). Gene description, screen score (scored ±3 in 0.5 increments in comparison to Empty Vector/sec- 11 RNAi control), and approximated developmental stage at time of screen noted. N/A corresponds to genes not screened due to RNAi availability. • Supplementary file 2. Statistical analysis of tunicamycin survival assay data. Median lifespan, death events counted, and statistics for tunicamycin survival assay of worms grown on Empty Vector (EV) or let- 767 RNAi combined with EV or sbp- 1 RNAi. • Supplementary file 3. Untargeted lipidomic analysis of let- 767 RNAi treated animals. Normalized values and fold change of identified lipids from let- 767 RNAi treated animals compared to EV controls at day 1 of adulthood. Standard deviation and averages calculated from 6 biological replicates. • Transparent reporting form • Source data 1. Unprocessed lipidomic source data of day 1 adult animals. Unprocessed lipidomic analysis data from animals grown on Empty Vector control or let- 767 RNAi and collected at day 1 of adulthood. Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 17 of 21 Cell Biology | Genetics and Genomics Research article Data availability All data generated or analyzed during this study are included in the manuscript and supporting files; Unprocessed data for Supplementary File 3 has been provided in Source data 1. References Adams CJ, Kopp MC, Larburu N, Nowak PR, Ali MMU. 2019. Structure and molecular mechanism of ER stress signaling by the unfolded protein response signal activator IRE1. Frontiers in Molecular Biosciences 6:11. DOI: https://doi.org/10.3389/fmolb.2019.00011, PMID: 30931312 Adamson B, Norman TM, Jost M, Cho MY, Nuñez JK, Chen Y, Villalta JE, Gilbert LA, Horlbeck MA, Hein MY, Pak RA, Gray AN, Gross CA, Dixit A, Parnas O, Regev A, Weissman JS. 2016. A multiplexed single- cell CRISPR screening platform enables systematic dissection of the unfolded protein response. Cell 167:1867–1882.. DOI: https://doi.org/10.1016/j.cell.2016.11.048, PMID: 27984733 Alecu I, Bennett SAL. 2019. Dysregulated lipid metabolism and its role in α-synucleinopathy in Parkinson’s disease. Frontiers in Neuroscience 13:328. DOI: https://doi.org/10.3389/fnins.2019.00328, PMID: 31031582 Avgerinos KI, Spyrou N, Mantzoros CS, Dalamaga M. 2019. Obesity and cancer risk: emerging biological mechanisms and perspectives. Metabolism 92:121–135. DOI: https://doi.org/10.1016/j.metabol.2018.11.001, PMID: 30445141 Bang IH, Kwon OK, Hao L, Park D, Chung MJ, Oh BC, Lee S, Bae EJ, Park BH. 2019. Deacetylation of XBP1s by sirtuin 6 confers resistance to ER stress- induced hepatic steatosis. Experimental & Molecular Medicine 51:1–11. DOI: https://doi.org/10.1038/s12276-019-0309-0, PMID: 31541078 Bar- Ziv R, Frakes AE, Higuchi- Sanabria R, Bolas T, Frankino PA, Gildea HK, Metcalf MG, Dillin A. 2020. Measurements of physiological stress responses in C. Journal of Visualized Experiments 159:21. DOI: https:// doi.org/10.3791/61001, PMID: 32510480 Calfon M, Zeng H, Urano F, Till JH, Hubbard SR, Harding HP, Clark SG, Ron D. 2002. IRE1 couples endoplasmic reticulum load to secretory capacity by processing the XBP- 1 mRNA. Nature 415:92–96. DOI: https://doi.org/ 10.1038/415092a, PMID: 11780124 Chang TY, Yamauchi Y, Hasan MT, Chang C. 2017. Cellular cholesterol homeostasis and Alzheimer’s disease. Journal of Lipid Research 58:2239–2254. DOI: https://doi.org/10.1194/jlr.R075630 Chaurasia B, Summers SA. 2015. Ceramides - lipotoxic inducers of metabolic disorders. Trends in Endocrinology and Metabolism 26:538–550. DOI: https://doi.org/10.1016/j.tem.2015.07.006, PMID: 26412155 Chen X, Li H, Fan X, Zhao C, Ye K, Zhao Z, Hu L, Ma H, Wang H, Fang Z. 2020. Protein palmitoylation regulates cell survival by modulating XBP1 activity in glioblastoma multiforme. Molecular Therapy Oncolytics 17:518– 530. DOI: https://doi.org/10.1016/j.omto.2020.05.007, PMID: 33024813 Chew H, Solomon VA, Fonteh AN. 2020. Involvement of lipids in Alzheimer’s disease pathology and potential therapies. Frontiers in Physiology 11:598. DOI: https://doi.org/10.3389/fphys.2020.00598, PMID: 32581851 Contreras C, González- García I, Martínez- Sánchez N, Seoane- Collazo P, Jacas J, Morgan DA, Serra D, Gallego R, Gonzalez F, Casals N, Nogueiras R, Rahmouni K, Diéguez C, López M. 2014. Central ceramide- induced hypothalamic lipotoxicity and ER stress regulate energy balance. Cell Reports 9:366–377. DOI: https://doi.org/ 10.1016/j.celrep.2014.08.057, PMID: 25284795 Corish P, Tyler- Smith C. 1999. Attenuation of green fluorescent protein half- life in mammalian cells. Protein Engineering, Design and Selection 12:1035–1040. DOI: https://doi.org/10.1093/protein/12.12.1035 Daniele JR, Higuchi- Sanabria R, Durieux J, Monshietehadi S, Ramachandran V, Tronnes SU, Kelet N, Sanchez M, Metcalf MG, Garcia G, Frankino PA, Benitez C, Zeng M, Esping DJ, Joe L, Dillin A. 2020. UPRER promotes lipophagy independent of chaperones to extend life span. Science Advances 6:eaaz1441. DOI: https://doi.org/ 10.1126/sciadv.aaz1441, PMID: 31911951 Desnoyers S, Blanchard PG, St- Laurent JF, Gagnon SN, Baillie DL, Luu- The V. 2007. Caenorhabditis elegans LET- 767 is able to metabolize androgens and estrogens and likely shares common ancestor with human types 3 and 12 17beta- hydroxysteroid dehydrogenases. The Journal of Endocrinology 195:271–279. DOI: https://doi. org/10.1677/JOE-07-0248, PMID: 17951538 Entchev EV, Schwudke D, Zagoriy V, Matyash V, Bogdanova A, Habermann B, Zhu L, Shevchenko A, Kurzchalia TV. 2008. LET- 767 is required for the production of branched chain and long chain fatty acids in Caenorhabditis elegans. The Journal of Biological Chemistry 283:17550–17560. DOI: https://doi.org/10.1074/ jbc.M800965200, PMID: 18390550 Fagone P, Jackowski S. 2009. Membrane phospholipid synthesis and endoplasmic reticulum function. Journal of Lipid Research 50 Suppl:S311–S316. DOI: https://doi.org/10.1194/jlr.R800049-JLR200, PMID: 18952570 Frakes AE, Dillin A. 2017. The UPR ER: Sensor and coordinator of organismal homeostasis. Molecular Cell 66:761–771. DOI: https://doi.org/10.1016/j.molcel.2017.05.031 FRAP profiler plugin. 2021. The Hardin Lab. http://worms.zoology.wisc.edu/Library/WebServer/Documents/ research/4d/4d.html [Accessed July 20, 2021]. Gille B, Galuska CE, Fuchs B, Peleg S. 2021. Recent advances in studying age- associated lipids alterations and dietary interventions in mammals. Frontiers in Aging 2:773795. DOI: https://doi.org/10.3389/fragi.2021. 773795, PMID: 35822042 Gómez- Puerta S, Ferrero R, Hochstoeger T, Zubiri I, Chao J, Aragón T, Voigt F. 2022. Live imaging of the co- translational recruitment of XBP1 mRNA to the ER and its processing by diffuse, non- polarized IRE1α. eLife 11:e75580. DOI: https://doi.org/10.7554/eLife.75580, PMID: 35730412 Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 18 of 21 Cell Biology | Genetics and Genomics Research article Halbleib K, Pesek K, Covino R, Hofbauer HF, Wunnicke D, Hänelt I, Hummer G, Ernst R. 2017. Activation of the unfolded protein response by lipid bilayer stress. Molecular Cell 67:673–684. DOI: https://doi.org/10.1016/j. molcel.2017.06.012, PMID: 28689662 Harayama T, Riezman H. 2018. Understanding the diversity of membrane lipid composition. Nature Reviews. Molecular Cell Biology 19:281–296. DOI: https://doi.org/10.1038/nrm.2017.138, PMID: 29410529 Hartl FU. 2017. Protein misfolding diseases. Annual Review of Biochemistry 86:21–26. DOI: https://doi.org/10. 1146/annurev-biochem-061516-044518, PMID: 28441058 Heikelä H, Ruohonen ST, Adam M, Viitanen R, Liljenbäck H, Eskola O, Gabriel M, Mairinoja L, Pessia A, Velagapudi V, Roivainen A, Zhang F- P, Strauss L, Poutanen M. 2020. Hydroxysteroid (17Β) dehydrogenase 12 is essential for metabolic homeostasis in adult mice. American Journal of Physiology. Endocrinology and Metabolism 319:E494–E508. DOI: https://doi.org/10.1152/ajpendo.00042.2020, PMID: 32691632 Henry P, Owopetu O, Adisa D, Nguyen T, Anthony K, Ijoni- Animadu D, Jamadar S, Abdel- Rahman F, Saleh MA. 2016. Fatty acids composition of Caenorhabditis elegans using accurate mass GCMS- QTOF. Journal of Environmental Science and Health. Part. B, Pesticides, Food Contaminants, and Agricultural Wastes 51:546– 552. DOI: https://doi.org/10.1080/03601234.2016.1170555, PMID: 27166662 Higuchi- Sanabria R, Paul JW, Durieux J, Benitez C, Frankino PA, Tronnes SU, Garcia G, Daniele JR, Monshietehadi S, Dillin A. 2018. Spatial regulation of the actin cytoskeleton by HSF- 1 during aging. Molecular Biology of the Cell 29:2522–2527. DOI: https://doi.org/10.1091/mbc.E18-06-0362, PMID: 30133343 Hoogewijs D, Houthoofd K, Matthijssens F, Vandesompele J, Vanfleteren JR. 2008. Selection and validation of a set of reliable reference genes for quantitative sod gene expression analysis in C. elegans. BMC Molecular Biology 9:9. DOI: https://doi.org/10.1186/1471-2199-9-9, PMID: 18211699 Imanikia S, Sheng M, Castro C, Griffin JL, Taylor RC. 2019. XBP- 1 remodels lipid metabolism to extend longevity. Cell Reports 28:581–589. DOI: https://doi.org/10.1016/j.celrep.2019.06.057, PMID: 31315038 Jacquemyn J, Cascalho A, Goodchild RE. 2017. The ins and outs of endoplasmic reticulum- controlled lipid biosynthesis. EMBO Reports 18:1905–1921. DOI: https://doi.org/10.15252/embr.201643426, PMID: 29074503 Jiang Z, Fan Q, Zhang Z, Zou Y, Cai R, Wang Q, Zuo Y, Cheng J. 2012. Senp1 deficiency promotes ER stress- induced apoptosis by increasing Xbp1 Sumoylation. Cell Cycle 11:1118–1122. DOI: https://doi.org/10.4161/cc. 11.6.19529 Kao YC, Ho PC, Tu YK, Jou IM, Tsai KJ. 2020. Lipids and Alzheimer’s disease. International Journal of Molecular Sciences 21:1505. DOI: https://doi.org/10.3390/ijms21041505, PMID: 32098382 Kniazeva M, Crawford QT, Seiber M, Wang CY, Han M. 2004. Monomethyl branched- chain fatty acids play an essential role in Caenorhabditis elegans development. PLOS Biology 2:E257. DOI: https://doi.org/10.1371/ journal.pbio.0020257, PMID: 15340492 Kniazeva M, Shen H, Euler T, Wang C, Han M. 2012. Regulation of maternal phospholipid composition and IP(3)- dependent embryonic membrane dynamics by a specific fatty acid metabolic event in C. elegans. Genes & Development 26:554–566. DOI: https://doi.org/10.1101/gad.187054.112, PMID: 22426533 Kuervers LM, Jones CL, O’Neil NJ, Baillie DL. 2003. The Sterol modifying enzyme LET- 767 is essential for growth, reproduction and development in Caenorhabditis elegans. Molecular Genetics and Genomics 270:121–131. DOI: https://doi.org/10.1007/s00438-003-0900-9 Lee D, Jeong D- E, Son HG, Yamaoka Y, Kim H, Seo K, Khan AA, Roh T- Y, Moon DW, Lee Y, Lee S- JV. 2015. SREBP and MDT- 15 protect C. elegans from glucose- induced accelerated aging by preventing accumulation of saturated fat. Genes & Development 29:2490–2503. DOI: https://doi.org/10.1101/gad.266304.115, PMID: 26637528 Li H, Korennykh AV, Behrman SL, Walter P. 2010. Mammalian endoplasmic reticulum stress sensor IRE1 signals by dynamic clustering. PNAS 107:16113–16118. DOI: https://doi.org/10.1073/pnas.1010580107, PMID: 20798350 Lodhi IJ, Semenkovich CF. 2014. Peroxisomes: a nexus for lipid metabolism and cellular signaling. Cell Metabolism 19:380–392. DOI: https://doi.org/10.1016/j.cmet.2014.01.002, PMID: 24508507 Loetscher P, Pratt G, Rechsteiner M. 1991. The C terminus of mouse ornithine decarboxylase confers rapid degradation on dihydrofolate reductase. Support for the pest hypothesis. The Journal of Biological Chemistry 266:11213–11220 PMID: 2040628. Metcalf MG, Higuchi- Sanabria R, Garcia G, Tsui CK, Dillin A. 2020. Beyond the cell factory: Homeostatic regulation of and by the UPRER. Science Advances 6:eabb9614. DOI: https://doi.org/10.1126/sciadv.abb9614, PMID: 32832649 Moldavski O, Amen T, Levin- Zaidman S, Eisenstein M, Rogachev I, Brandis A, Kaganovich D, Schuldiner M. 2015. Lipid droplets are essential for efficient clearance of cytosolic inclusion bodies. Developmental Cell 33:603– 610. DOI: https://doi.org/10.1016/j.devcel.2015.04.015, PMID: 26004510 Moon YA, Horton JD. 2003. Identification of two mammalian reductases involved in the two- carbon fatty acyl elongation cascade. The Journal of Biological Chemistry 278:7335–7343. DOI: https://doi.org/10.1074/jbc. M211684200, PMID: 12482854 Na H, Zhang P, Chen Y, Zhu X, Liu Y, Liu Y, Xie K, Xu N, Yang F, Yu Y, Cichello S, Mak HY, Wang MC, Zhang H, Liu P. 2015. Identification of lipid droplet structure- like/resident proteins in Caenorhabditis elegans. Biochimica et Biophysica Acta (BBA) - Molecular Cell Research 1853:2481–2491. DOI: https://doi.org/10.1016/j.bbamcr. 2015.05.020 Nomura T, Horikawa M, Shimamura S, Hashimoto T, Sakamoto K. 2010. Fat accumulation in Caenorhabditis elegans is mediated by SREBP homolog SBP- 1. Genes & Nutrition 5:17–27. DOI: https://doi.org/10.1007/ s12263-009-0157-y, PMID: 19936816 Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 19 of 21 Cell Biology | Genetics and Genomics Research article Olzmann JA, Carvalho P. 2019. Dynamics and functions of lipid droplets. Nature Reviews. Molecular Cell Biology 20:137–155. DOI: https://doi.org/10.1038/s41580-018-0085-z, PMID: 30523332 Promlek T, Ishiwata- Kimata Y, Shido M, Sakuramoto M, Kohno K, Kimata Y. 2011. Membrane aberrancy and unfolded proteins activate the endoplasmic reticulum stress sensor Ire1 in different ways. Molecular Biology of the Cell 22:3520–3532. DOI: https://doi.org/10.1091/mbc.E11-04-0295, PMID: 21775630 Rantakari P, Lagerbohm H, Kaimainen M, Suomela J- P, Strauss L, Sainio K, Pakarinen P, Poutanen M. 2010. Hydroxysteroid (17Β) dehydrogenase 12 is essential for mouse organogenesis and embryonic survival. Endocrinology 151:1893–1901. DOI: https://doi.org/10.1210/en.2009-0929 Sakurai N, Miki Y, Suzuki T, Watanabe K, Narita T, Ando K, Yung TMC, Aoki D, Sasano H, Handa H. 2006. Systemic distribution and tissue Localizations of human 17Beta- Hydroxysteroid dehydrogenase type 12. The Journal of Steroid Biochemistry and Molecular Biology 99:174–181. DOI: https://doi.org/10.1016/j.jsbmb.2006. 01.010, PMID: 16621523 Schindelin J, Arganda- Carreras I, Frise E, Kaynig V, Longair M, Pietzsch T, Preibisch S, Rueden C, Saalfeld S, Schmid B, Tinevez J- Y, White DJ, Hartenstein V, Eliceiri K, Tomancak P, Cardona A. 2012. Fiji: an open- source platform for biological- image analysis. Nature Methods 9:676–682. DOI: https://doi.org/10.1038/nmeth.2019, PMID: 22743772 Schwarz DS, Blower MD. 2016. The endoplasmic reticulum: structure, function and response to cellular signaling. Cellular and Molecular Life Sciences 73:79–94. DOI: https://doi.org/10.1007/s00018-015-2052-6, PMID: 26433683 Shimano H, Sato R. 2017. SREBP- regulated lipid metabolism: Convergent physiology — divergent pathophysiology. Nature Reviews. Endocrinology 13:710–730. DOI: https://doi.org/10.1038/nrendo.2017.91, PMID: 28849786 Sletten AC, Peterson LR, Schaffer JE. 2018. Manifestations and mechanisms of myocardial lipotoxicity in obesity. Journal of Internal Medicine 284:478–491. DOI: https://doi.org/10.1111/joim.12728, PMID: 29331057 StackReg. 2021. StackReg. http://bigwww.epfl.ch/thevenaz/stackreg/ [Accessed July 20, 2021]. Tam AB, Roberts LS, Chandra V, Rivera IG, Nomura DK, Forbes DJ, Niwa M. 2018. The UPR activator Atf6 responds to proteotoxic and lipotoxic stress by distinct mechanisms. Developmental Cell 46:327–343. DOI: https://doi.org/10.1016/j.devcel.2018.04.023, PMID: 30086303 Vevea JD, Garcia EJ, Chan RB, Zhou B, Schultz M, Di Paolo G, McCaffery JM, Pon LA. 2015. Role for lipid droplet biogenesis and microlipophagy in adaptation to lipid imbalance in yeast. Developmental Cell 35:584– 599. DOI: https://doi.org/10.1016/j.devcel.2015.11.010, PMID: 26651293 Volmer R, van der Ploeg K, Ron D. 2013. Membrane lipid saturation activates endoplasmic reticulum unfolded protein response transducers through their transmembrane domains. PNAS 110:4628–4633. DOI: https://doi. org/10.1073/pnas.1217611110, PMID: 23487760 Volmer R, Ron D. 2015. Lipid- dependent regulation of the unfolded protein response. Current Opinion in Cell Biology 33:67–73. DOI: https://doi.org/10.1016/j.ceb.2014.12.002, PMID: 25543896 Vrablik TL, Petyuk VA, Larson EM, Smith RD, Watts JL. 2015. Lipidomic and proteomic analysis of Caenorhabditis elegans lipid droplets and identification of ACS- 4 as a lipid droplet- associated protein. Biochimica et Biophysica Acta (BBA) - Molecular and Cell Biology of Lipids 1851:1337–1345. DOI: https://doi.org/10.1016/j. bbalip.2015.06.004 Xicoy H, Wieringa B, Martens GJM. 2019. The role of lipids in Parkinson’s disease. Cells 8:27. DOI: https://doi. org/10.3390/cells8010027 Xu J, Taubert S. 2021. Beyond proteostasis: Lipid metabolism as a new player in ER homeostasis. Metabolites 11:52. DOI: https://doi.org/10.3390/metabo11010052, PMID: 33466824 Yanagitani Kota, Imagawa Y, Iwawaki T, Hosoda A, Saito M, Kimata Y, Kohno K. 2009. Cotranslational targeting of Xbp1 protein to the membrane promotes cytoplasmic splicing of its own mRNA. Molecular Cell 34:191–200. DOI: https://doi.org/10.1016/j.molcel.2009.02.033, PMID: 19394296 Yanagitani K, Kimata Y, Kadokura H, Kohno K. 2011. Translational pausing ensures membrane targeting and cytoplasmic splicing of XBP1u mRNA. Science 331:586–589. DOI: https://doi.org/10.1126/science.1197142, PMID: 21233347 Yerbury JJ, Ooi L, Dillin A, Saunders DN, Hatters DM, Beart PM, Cashman NR, Wilson MR, Ecroyd H. 2016. Walking the tightrope: proteostasis and neurodegenerative disease. Journal of Neurochemistry 137:489–505. DOI: https://doi.org/10.1111/jnc.13575, PMID: 26872075 Yoon JH, Seo Y, Jo YS, Lee S, Cho E, Cazenave- Gassiot A, Shin YS, Moon MH, An HJ, Wenk MR, Suh PG. 2022. Brain lipidomics: From functional landscape to clinical significance. Science Advances 8:eadc9317. DOI: https:// doi.org/10.1126/sciadv.adc9317, PMID: 36112688 Yoshida H, Oku M, Suzuki M, Mori K. 2006. Pxbp1(U) encoded in Xbp1 pre- mRNA negatively regulates unfolded protein response activator Pxbp1(S) in mammalian ER stress response. The Journal of Cell Biology 172:565– 575. DOI: https://doi.org/10.1083/jcb.200508145, PMID: 16461360 Zhang H, Abraham N, Khan LA, Hall DH, Fleming JT, Göbel V. 2011. Apicobasal domain identities of expanding tubular membranes depend on glycosphingolipid biosynthesis. Nature Cell Biology 13:1189–1201. DOI: https://doi.org/10.1038/ncb2328, PMID: 21926990 Zhang P, Na H, Liu Z, Zhang S, Xue P, Chen Y, Pu J, Peng G, Huang X, Yang F, Xie Z, Xu T, Xu P, Ou G, Zhang SO, Liu P. 2012. Proteomic study and marker protein identification of Caenorhabditis elegans lipid droplets. Molecular & Cellular Proteomics 11:317–328. DOI: https://doi.org/10.1074/mcp.M111.016345, PMID: 22493183 Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 20 of 21 Cell Biology | Genetics and Genomics Research article Zhang J, Hu Y, Wang Y, Fu L, Xu X, Li C, Xu J, Li C, Zhang L, Yang R, Jiang X, Wu Y, Liu P, Zou X, Liang B. 2021. mmBCFA C17iso ensures endoplasmic reticulum integrity for lipid droplet growth. The Journal of Cell Biology 220:e202102122. DOI: https://doi.org/10.1083/jcb.202102122, PMID: 34623380 Garcia et al. eLife 2023;12:e83884. DOI: https://doi.org/10.7554/eLife.83884 21 of 21 Cell Biology | Genetics and Genomics
10.2196_42866
JMIR MENTAL HEALTH Original Paper de Angel et al The Feasibility of Implementing Remote Measurement Technologies in Psychological Treatment for Depression: Mixed Methods Study on Engagement Valeria de Angel1,2, BSc, MSc; Fadekemi Adeleye3, BSc; Yuezhou Zhang4, MSc; Nicholas Cummins4, PhD; Sara Munir5, BSc; Serena Lewis1,6, BSc; Estela Laporta Puyal7,8, MSc; Faith Matcham1,9, CPsychol, PhD; Shaoxiong Sun4, PhD; Amos A Folarin2,4,10,11,12, PhD; Yatharth Ranjan13, MSc; Pauline Conde13, BSc; Zulqarnain Rashid13, PhD; Richard Dobson1,2, PhD; Matthew Hotopf1,2, PhD 1Department of Psychological Medicine, Institute of Psychiatry, Psychology and Neuroscience, King's College London, London, United Kingdom 2NIHR Maudsley Biomedical Research Centre,  South London and Maudsley NHS Foundation Trust, London, United Kingdom 3Department of Psychology, King's College London, London, United Kingdom 4Department of Biostatistics & Health Informatics, Institute of Psychiatry, Psychology and Neuroscience, King's College London, London, United Kingdom 5Lewisham Talking Therapies, South London and Maudsley NHS Foundation Trust, London, United Kingdom 6Department of Psychology, University of Bath, Bath, United Kingdom 7Biomedical Signal Interpretation and Computational Simulation Group, Aragón Institute of Engineering Research (I3A), IIS Aragón, University of Zaragoza, Zaragoza, Spain 8Centro de Investigación Biomédica en Red of Bioengineering, Biomaterials and Nanomedicine (CIBER-BBN), Madrid, Spain 9School of Psychology, University of Sussex, Brighton, United Kingdom 10Institute of Health Informatics, University College London, London, United Kingdom 11Health Data Research UK London, University College London, London, United Kingdom 12NIHR Biomedical Research Centre at University College London Hospitals, University College London Hospitals NHS Foundation Trust, London, United Kingdom 13Institute of Psychiatry, Psychology and Neuroscience, King's College London, London, United Kingdom Corresponding Author: Valeria de Angel, BSc, MSc Department of Psychological Medicine Institute of Psychiatry, Psychology and Neuroscience King's College London E3.08, 3rd Floor East Wing de Crespigny park London, SE5 8AF United Kingdom Phone: 44 20 7848 0002 Email: valeria.de_angel@kcl.ac.uk Abstract Background: Remote measurement technologies (RMTs) such as smartphones and wearables can help improve treatment for depression by providing objective, continuous, and ecologically valid insights into mood and behavior. Engagement with RMTs is varied and highly context dependent; however, few studies have investigated their feasibility in the context of treatment. Objective: A mixed methods design was used to evaluate engagement with active and passive data collection via RMT in people with depression undergoing psychotherapy. We evaluated the effects of treatment on 2 different types of engagement: study attrition (engagement with study protocol) and patterns of missing data (engagement with digital devices), which we termed data availability. Qualitative interviews were conducted to help interpret the differences in engagement. Methods: A total of 66 people undergoing psychological therapy for depression were followed up for 7 months. Active data were gathered from weekly questionnaires and speech and cognitive tasks, and passive data were gathered from smartphone sensors and a Fitbit (Fitbit Inc) wearable device. https://mental.jmir.org/2023/1/e42866 XSL•FO RenderX JMIR Ment Health 2023 | vol. 10 | e42866 | p. 1 (page number not for citation purposes) JMIR MENTAL HEALTH de Angel et al Results: The overall retention rate was 60%. Higher-intensity treatment (χ2 1=4.6; P=.03) and higher baseline anxiety (t56.28=−2.80, 2-tailed; P=.007) were associated with attrition, but depression severity was not (t50.4=−0.18; P=.86). A trend toward significance was found for the association between longer treatments and increased attrition (U=339.5; P=.05). Data availability was higher for active data than for passive data initially but declined at a sharper rate (90%-30% drop in 7 months). As for passive data, wearable data availability fell from a maximum of 80% to 45% at 7 months but showed higher overall data availability than smartphone-based data, which remained stable at the range of 20%-40% throughout. Missing data were more prevalent among GPS location data, followed by among Bluetooth data, then among accelerometry data. As for active data, speech and cognitive tasks had lower completion rates than clinical questionnaires. The participants in treatment provided less Fitbit data but more active data than those on the waiting list. Conclusions: Different data streams showed varied patterns of missing data, despite being gathered from the same device. Longer and more complex treatments and clinical characteristics such as higher baseline anxiety may reduce long-term engagement with RMTs, and different devices may show opposite patterns of missingness during treatment. This has implications for the scalability and uptake of RMTs in health care settings, the generalizability and accuracy of the data collected by these methods, feature construction, and the appropriateness of RMT use in the long term. (JMIR Ment Health 2023;10:e42866) doi: 10.2196/42866 KEYWORDS depression; anxiety; digital health; wearable devices; smartphone; passive sensing; mobile health; mHealth; digital phenotyping; mobile phone Introduction Background Depression is a leading cause of disability, with associated physical comorbidities and increased health care costs [1,2]. Psychological therapy is a recommended first-line treatment for mild to moderate depression [3,4]; however, approximately 50% of people do not recover following intervention [5,6]. Remote measurement technologies (RMTs) such as smartphones and wearables may assist in the treatment of depression to improve patient outcomes by detecting changes in its key behavioral aspects. RMTs, by generating unobtrusive, continuous, and objective measures of behavior and physiology, could overcome the pitfalls of the current clinical outcome measurements, which rely on patient recall and infrequent symptom scales. Furthermore, they could help establish clinical objectives for treatment, such as a target amount of physical activity or regular sleep-time schedule and serve as indicators of whether treatments targeting particular behaviors have been effective. RMTs may also uncover digital phenotypes to identify people who are more or less responsive to certain treatments, paving the way for increased personalization of mental health care [7]. Finally, RMTs could improve patient and clinician experience of psychotherapy by strengthening communication, helping support the emotional and cognitive needs of patients and enhancing self-awareness [8]. RMTs generally apply 2 types of data collection methods: active and passive. Active data collection requires conscious user engagement, such as responding to mood scales, questionnaires, or speech tasks delivered to a participant’s phone. Passive data collection refers to the automatic capture of information via device-embedded sensors that require minimal input from users [9]; for example, accelerometers on a fitness tracker automatically detect physical activity. Used in combination, active and passive monitoring provide a way to capture https://mental.jmir.org/2023/1/e42866 XSL•FO RenderX continuous, ecologically valid, and high-resolution measures of signs and symptoms related to depression. The extent to which these methods can be successfully implemented in health care and used in treatment depends on their feasibility and acceptability as tools for collecting longitudinal data in clinical populations. The feasibility of using RMTs is generally evaluated by measuring 2 broad parameters of engagement: attrition from longitudinal studies and data availability, which is the amount of usable data contributed by individuals through task completion or device use and, therefore, the opposite of missing data [10]. The measurement and reporting of attrition is relevant not only because attrition threatens the generalizability of longitudinal studies but also because it informs implementation efforts by mirroring the potential uptake and engagement within clinical settings. Much of the current research on attrition focuses on active data collection, with passive sensing being underreported. In general, studies have short follow-up periods, with systematic reviews finding a median follow-up period of 7 days for active data [11] and between 7 and 14 days for passive data [12], limiting their ability to be generalized to psychotherapy contexts, which usually span weeks. In addition, the context in which data collection occurs is key to understanding the difference in attrition rates between RMT studies. For example, a review of self-referral studies found, on average, 50% attrition in the first 15 days and varied retention rates depending on factors such as the presence and type of illness studied [13], whereas clinical trials on digital-based psychotherapy found similar attrition but at a much slower pace [14]. By contrast, large studies with dedicated recruitment resources have achieved attrition rates as low as 20% even if follow-up sessions were conducted after 2 years [15]. Therefore, if implementation of RMTs within health care is the aim, research on long-term attrition in active and passive data collection in the context of psychotherapy is critical. Work on data availability has generally focused on active approaches [16-18], leaving passive sensing underresearched JMIR Ment Health 2023 | vol. 10 | e42866 | p. 2 (page number not for citation purposes) JMIR MENTAL HEALTH de Angel et al and underreported [12]. Given that both approaches require varying amounts of input and commitment from the user, missingness is likely to vary in the extent to which it occurs at random and may be differentially affected by individual differences [19]. This, in turn, has implications for the integrity of the constructed variables and for understanding the potential sources of biases in the data. Sparse active data points on mood questionnaires can affect how ground truth is determined, whereas less passive data can result in inaccuracies in how features are derived and the resulting data analysis (refer to Currey and Torous [20] for an example of this). Objective We sought to explore the feasibility of using RMTs in a clinical setting to help uncover potential implementation and scaling issues, the resolution of which is crucial for widespread adoption. This study used a mixed methods design to evaluate the long-term engagement with active and passive approaches to the remote monitoring of mood and behavior in people with depression undergoing psychotherapy. Applying the framework developed by White et al [10], we focused on 2 forms of engagement as feasibility aspects of interest. The aims were (1) to measure engagement with the research protocol through recruitment and attrition rates, (2) measure engagement with RMTs through passive and active data availability rates and identify data streams that are more vulnerable to missing data, (3) assess the possible effect of treatment on both types of engagement, and (4) use the information gathered from qualitative interviews to aid in the explanation of the quantitative engagement data. Methods Study Design This study was a fully remote, mixed methods, prospective cohort study with repeated measures over a 7-month period, designed to evaluate the feasibility and acceptability of using remote data collection methods in people undergoing treatment; the full protocol has been reported elsewhere [21]. The quantitative measures included recurrent clinical questionnaires and continuous digital sensor data. Qualitative measures comprised semistructured interviews that adopted an inductive approach to thematic analysis. from Improving Access Recruitment and Setting Participants were drawn to Psychological Therapies (IAPT) services in South London and Maudsley National Health Service Foundation Trust, United Kingdom, a publicly funded outpatient program providing psychological treatments for adults with mild to moderate mental health disorders. IAPT services provide treatment at both high and low intensities (refer to Table S1 in Multimedia Appendix 1 for details), the allocation of which is based on several factors, including patient needs, preferences, and diagnosis. High-intensity therapy comprises approximately 10 to 12 comprises low-intensity sessions, whereas approximately 6 to 8 sessions. These are usually delivered 1 week apart and can be web-based or face-to-face, depending on clinician availability and patient preference. therapy Sample A total of 66 treatment-seeking adults with depression were recruited from their local IAPT services’ waiting list, which provided the study information, and screened for eligibility either over the phone by a researcher or through a web-based self-screening tool. The sample size was determined by the primary aims and followed the general recommendation for samples of 50 to 60 participants to assess feasibility outcomes [22]. Recruitment and data collection were conducted between June 2020 and March 2022. We included adults with a current episode of depression, as measured by the Mini International Neuropsychiatric Interview [23], who owned and did not extensively share an Android (Google LLC) smartphone and were able and willing to use a wrist-worn device for the duration of the study. The exclusion criteria included a lifetime diagnosis of bipolar disorder, schizophrenia, or schizoaffective disorders, as the digital patterns of these conditions are different from those of depression, and people who were working regular night shifts or were pregnant, as these external factors can cause changes in sleep patterns. Researchers discussed health anxieties with potential participants through unstructured questions. On the basis of these discussions, those who believed that their health anxieties may worsen with continuous behavioral monitoring were excluded. Ethics Approval This study was reviewed and given favorable opinion by the London Westminster Research Ethics Committee and received approval from the Health Research Authority (reference number 20/LO/0091). Procedures Overview Details of the measures, technology, and procedures have been covered in depth in a previous publication [21]. The methods described in this section refer to the primary aims and outcomes of the original study protocol. Therefore, the measures presented herein are relevant to this analysis. Overall, the participants were enrolled in the study at least a week before their first therapy session. The researchers had no control over the treatment provided. Consequently, the enrolled participants had different waiting list times, treatment lengths, and treatment intensities. They were followed up throughout treatment and up to 3 months after treatment using smartphone apps and a wrist-worn device (Fitbit Charge 3 or 4 [Fitbit Inc]). Therefore, active and passive data were collected for approximately 7 months, but this depended on the treatment length, which varied from person to person. The study procedures are depicted by Figure 1. https://mental.jmir.org/2023/1/e42866 XSL•FO RenderX JMIR Ment Health 2023 | vol. 10 | e42866 | p. 3 (page number not for citation purposes) JMIR MENTAL HEALTH de Angel et al Figure 1. The study timeline for participants, from screening to the end of the study. Baseline Session After providing signed informed consent, the participants provided in-depth baseline sociodemographic and clinical data related to their current and previous physical and mental health conditions, family history, treatment status, phone use, and social and physical activity levels. Clinical measures included the Patient Health Questionnaire 9-item [24], a widely validated depression questionnaire, and the Generalized Anxiety Disorder 7-item scale questionnaire [25]. Participants were then guided through the installation and setup of the 4 apps used in this study, which have been detailed in the next section. Data Collection Overview Active and passive data collection began from the baseline session, and data were acquired from a variety of sources. The underlying infrastructure for data collection and storage was the Remote Assessment of Disease and Relapse (RADAR)–base platform, developed by the RADAR-Central Nervous System Consortium [26]. Passive measures were gathered from (1) Fitbit wearable device sensors and (2) smartphone sensors, and active measures were gathered from (3) web-based surveys and (4) smartphone apps. Passive Measures Passive measures were collected from 2 devices. First, the participants were provided with a Fitbit and downloaded the Fitbit app, which provided a user interface where they could track their own activity. The data extracted through the Fitbit Application Programming Interface for use in this study were related to sleep, physical activity, heart rate, and step count. The participants used their own Android smartphone and were asked to download the RADAR-base passive RMT app, a purpose-built app that collects smartphone sensor data. Only data streams with a fixed sampling rate allow for the calculation of missing data, as it provides the total number of expected data points in a period, which serves as the denominator for the total number of observed data points. These data streams were acceleration, nearby Bluetooth device detection, and GPS. GPS coordinates were obfuscated by adding a participant-specific random number as a reference point, and the relative change in location was calculated from there; therefore, an individual’s home address or precise geographic location could not be gathered. Active Measures Overall, active data were collected through 2 methods (Figure 1): web-based surveys and smartphone-based tasks sent via apps. Web-based surveys were clinical measures delivered by email via the REDCap (Research Electronic Data Capture; Vanderbilt University) software, a web-based platform for research that is conducted through a browser [27]. The smartphone-based active data were collected through (1) clinical questionnaires, (2) a series of speech tasks delivered directly to the participant’s phone via a custom-built app (the RADAR active RMT app), and (3) validated cognitive assessments in gamified format requiring a separate app, the THINC-it app [28] Cognitive tasks were completed monthly, whereas the speech task, which required the participants to record themselves reading a short text [29] and answering a question aloud, was delivered fortnightly. All active measures were rotated weekly such that the tasks took an average of 10 minutes per week to complete, except for 1 week in a month, when the THINC-it task increased the completion time by approximately 15 minutes. The participants were notified when it was time to carry out the tasks. Details of all active measures can be found elsewhere [21]. https://mental.jmir.org/2023/1/e42866 XSL•FO RenderX JMIR Ment Health 2023 | vol. 10 | e42866 | p. 4 (page number not for citation purposes) JMIR MENTAL HEALTH Posttreatment Interview The first 20 participants who completed the therapy and agreed to participate in an optional posttreatment interview were included in the qualitative analysis. This was a 30-minute semistructured interview conducted on the web examining the participants’ experiences of using RMTs during psychotherapy for depression. To reduce potential social desirability bias, interviews were conducted by researchers who had little to no previous contact with their interviewee. Statistical Analysis Quantitative Data Quantitative data were collected regarding the following parameters of engagement: to or loss of withdrawal 1. Study engagement: the main outcome of attrition is defined as the division of study participants into those who completed the study (“completers”) and those who did not because follow-up (“non-completers”). To determine whether symptom severity at baseline was associated with attrition, 2-tailed t tests were performed to compare the mean severity of clinical measures taken at baseline, namely the Patient Health Questionnaire 9-item and Generalized Anxiety Disorder 7-item, across the study completion groups. The Shapiro-Wilk test was used in all cases to test for normality distributions in variables, and if this assumption was violated, nonparametric tests were used. All other assumptions for 2-tailed t test calculations were met. To test the effect of treatment characteristics on attrition, completers were compared with noncompleters in terms of treatment length and treatment intensity. The Mann-Whitney U test was used for the continuous variable treatment length given the violation of parametric assumptions, and the chi-square test was used to compare frequencies across low- and high-intensity therapy. To examine the role of overall de Angel et al time in the study as a confounder (given its association with treatment length and treatment intensity), we tested its potential association with attrition by conducting a 2-tailed t test on the mean study length across the completion groups. 2. Engagement with RMTs: engagement with RMTs was measured as the total number of data points available out of the total number of data points expected. In terms of active data, this was calculated as the number of active tasks completed out of the total number of tasks delivered. In terms of passive data, this was calculated as the number of hours in which there was at least one data point divided by the total number of hours in a day. This was then averaged to a weekly statistic. Logistic regression analyses were performed to assess whether being in treatment influenced the magnitude of data availability. The 3 treatment conditions were before treatment, treatment, or after treatment. Given the expected reduction in data availability over time owing to study fatigue, we adjusted these analyses for time in weeks, age, and gender. The weeks selected for analysis had to have at least 10 participants in each treatment condition; therefore, before treatment versus treatment status comparisons involved weeks 3 to 8, and after treatment versus treatment status comparisons comprised weeks 8 to 24. Missing Data Thresholds We established missing data thresholds as follows. Passive data required at least one data point per hour for at least 8 hours per day to be considered available. The total number of available hours per week were calculated, and weeks with at least 50% of available hours were deemed available. Active data were sampled weekly; therefore, active data availability for each participant was defined as the completion of at least one active task that week. The proportion of participants in the study with available data each week has been presented by the dotted line in Figure 2A. https://mental.jmir.org/2023/1/e42866 XSL•FO RenderX JMIR Ment Health 2023 | vol. 10 | e42866 | p. 5 (page number not for citation purposes) JMIR MENTAL HEALTH de Angel et al Figure 2. (A) Data availability by type of data. Y-axis 1 shows the percentage of people who contributed data out of the total number of available participants. Participant numbers are plotted against the secondary y-axis. (B) Data availability by passive data stream. This shows the proportion of participants with available data, averaged per week. Data were deemed available if there was at least one data point available per hour on at least 8 hours a day. (C) Data availability by active data. This shows the proportion of participants with at least one active data task completed, averaged per week. IQR: Interquartile range, ACC: accelerometer; BT: Bluetooth. Qualitative Data Transcriptions of the recordings of the semistructured interviews were checked for accuracy by a second researcher and analyzed using a deductive approach to thematic analysis, with the iterative categorization technique [30]. The deductive approach was used in favor of an inductive approach, as certain core themes in this field have been previously reported [8,31]. These were used as initial frameworks from which to organize the initial coding, as we anticipated that these concepts would also emerge from the current data, but flexibility was given to reorganize these codes as they applied to the current data. https://mental.jmir.org/2023/1/e42866 XSL•FO RenderX JMIR Ment Health 2023 | vol. 10 | e42866 | p. 6 (page number not for citation purposes) JMIR MENTAL HEALTH de Angel et al Overarching themes, such as device engagement and the impact of treatment, were preestablished according to the quantitative objectives of the study. was significant: N=66, χ2 1=4.6; P=.03. The participants who received low-intensity treatment were more likely to complete the study than those who received high-intensity treatment. All quantitative data processing and analyses were performed using R (version 4.0.2, R Core Team), and qualitative data were analyzed using NVivo (released in March 2020, QSR International). Results Study Engagement: Recruitment and Attrition Over 900 people were contacted, and of these, 66 (7.3%) were finally enrolled (Figure 3). Of the 66 enrolled individuals, 40 (61%) completed the study. Sample characteristics are presented in Table 1 and show that our sample was similar in demographic proportions to the total IAPT population in South London in terms of age, gender, ethnicity structures [32], and employment status [33]. Table 2 shows the means, medians, and proportions for those who completed the study versus those who did not on treatment-related variables. A chi-square test of independence was performed to examine the relationship between treatment intensity and attrition. The relationship between these variables Figure 3. Recruitment flowchart. No significant differences were found between the attrition groups across the sample characteristics of age, gender, ethnicity, educational level, employment status, and previous experience with digital health tools. A Mann-Whitney U test was conducted to determine whether there was a difference in treatment length between the attrition groups. The results indicated a trend toward significance in terms of the difference in treatment length between the groups (W=339.5; P=.05). A significance threshold of P<.05 would not regard this observation as evidence for a significant difference in treatment length between completers and noncompleters, where longer treatments would be associated with attrition. These associations cannot be accounted for by symptom severity or overall time in the study, given that study length was not associated with attrition, and the severity of depression or anxiety was associated with neither treatment length nor treatment intensity. t tests (2-tailed) revealed that the severity of anxiety (t56=−2.80; P=.007), but not depression (t50=−0.18; P=.86), was associated with attrition such that higher anxiety at baseline was associated with higher attrition levels (Table 2). These associations are mapped in Figure S1 in Multimedia Appendix 1. https://mental.jmir.org/2023/1/e42866 XSL•FO RenderX JMIR Ment Health 2023 | vol. 10 | e42866 | p. 7 (page number not for citation purposes) JMIR MENTAL HEALTH Table 1. Study sample characteristics (N=66). Age (years), mean (SD) Gender, n (%) Woman Man Nonbinary Ethnicity, n (%) Asian or Asian British Black, African, Caribbean, or Black British Middle Eastern Mixed or multiple ethnic groups White British White (other) Education level, n (%) Secondary education Degree-level education or diploma (eg, BSca and BAb) Postgraduate degree (eg, MScc, MAd, and PhDe) Employment status, n (%) Paid employment Unpaid employment Unemployment Furlough Student Retired Previous experience with digital health tools, n (%) Psychiatric comorbidities, n (%) 0 (single psychiatric diagnosis) 1 ≥2 Therapy intensity, n (%) Low intensity High intensity Treatment data (time in weeks), mean (SD) Treatment start lag Treatment lengthf Posttreatment follow-up Total study PHQ-9g GAD-7h aBSc: Bachelor of Science. bBA: Bachelor of Arts. cMSc: Master of Science. https://mental.jmir.org/2023/1/e42866 XSL•FO RenderX Values 34.6 (11.1) 40 (61) 23 (35) 3 (5) 2 (3) 11 (17) 1 (2) 6 (9) 37 (56) 9 (14) 26 (39) 26 (39) 14 (21) 42 (64) 4 (6) 12 (18) 3 (5) 4 (6) 1 (2) 55 (83) 17 (26) 12 (18) 37 (56) 33 (50) 32 (48) 4.7 (4.7) 11.6 (6.5) 14.5 (6.4) 29.6 (6.6) 16.7 (5.1) 13.3 (4.7) de Angel et al JMIR Ment Health 2023 | vol. 10 | e42866 | p. 8 (page number not for citation purposes) JMIR MENTAL HEALTH de Angel et al dMA: Master of Arts. ePhD: Doctor of Philosophy. fTreatment length is the number of weeks between the first and last sessions, and not the total number of sessions. gPHQ-9: Patient Health Questionnaire 9-item. hGAD-7: Generalized Anxiety Disorder 7-item. Table 2. Summary statistics and statistical analyses of associations among completion groups, explanatory variables, and covariates. Attrition Analysis Completers (n=40) Noncompleters (n=26) Test df Test statistic P value Treatment intensity (%) Low intensity High intensity 75.76 46.88 24.24 53.13 Chi-square test —a 1 — 4.6 — Treatment length (weeks), median (IQR) Length 7 (5-12.25) 12 (6-22.25) Mann-Whitney U test N/Ab 339.5 Covariates, mean (SD) Anxiety severityc Depression severityd 12.33 (4.66) 15.4 (4.06) 16.65 (5.08) 16.88 (5.17) Study length (weeks) 29.38 (6.02) 29.92 (7.56) t test t test t test 56.28 -2.80 50.40 -0.18 45.02 -0.31 .03 — .05 .007 .86 .76 aNot available. bN/A: not applicable. cAnxiety was measured with the Generalized Anxiety Disorder 7-item scale questionnaire. dDepression was measured using the Patient Health Questionnaire 9-item. Engagement With RMTs: Data Availability The four main types of data collected were (1) wearable passive data, (2) smartphone-based passive data, (3) smartphone-based active data, and (4) web-based active data. Figure 2 shows how these data types vary in terms of their availability in the study, where 100% data completion would mean that the participants supplied, on average, 100% of the data that week. The data availability for smartphone-based passive data was between 20% and 40% for the duration of the study. Fitbit-based passive data and both active data streams had a similar proportion of data availability, but the rate of decline was lower for wearable passive data than for active data. To describe the missing data patterns across the passive data streams, the proportion of participants who provided sensor data for at least 8 hours is plotted in Figure 2B. There was no established threshold for the minimum acceptable quantity of passive data necessary to perform an analysis of these data. Therefore, we established a limit of 8 hours per day as an acceptable threshold for missing data, as the main daily activities, such as work and sleep, can be broken down into 8-hour cycles. GPS location was the passive data stream most vulnerable to missing data, followed by Bluetooth (Figure 2B). The active data stream most vulnerable to missing data for the first 10 weeks was speech and the THINC-it cognitive task thereafter (Figure 2C). To study whether being in treatment affected data availability, the amount of data contributed by the participants who were actively receiving treatment was compared with that contributed by those who were in either pretreatment (on the waiting list) or posttreatment. Logistic regression, adjusted for time in weeks, revealed minor effects of treatment status on data availability. Significant differences in data availability were found for active smartphone and passive wearable data. When comparing those in treatment with those in pretreatment (Table 3), we found that the odds of those in treatment having active app smartphone were 2.54 times that of those on the waiting list having active app data, regardless of the time in the study. Conversely, there was a 54% decrease in the odds of contributing Fitbit data for those in treatment compared with those in pretreatment. In summary, more clinical questionnaires were completed while in treatment than while on the waiting list, but more Fitbit data were available while on the waiting list than during treatment. https://mental.jmir.org/2023/1/e42866 XSL•FO RenderX JMIR Ment Health 2023 | vol. 10 | e42866 | p. 9 (page number not for citation purposes) JMIR MENTAL HEALTH de Angel et al Table 3. Odds ratios (ORs) for data availability in treatment as compared with that in pretreatment for overlapping weeks, adjusting for time (in weeks), age, and gender. A positive absolute difference in ORs show that those in treatment had more data availability. Data streams Treatment versus pretreatment (weeks 3-8)a Absolute difference (unadjusted), % OR (95% CI) P value Active Web based Smartphone Passive Wearable Smartphone −4.40 +b9.24 −9.94 −6.31 0.89 (0.44-1.79) 2.34 (1.16-4.78) 0.46 (0.21-0.95) 0.72 (0.44-1.16) .76 .01 .04 .17 aWeeks 3 to 8 when n for all groups was <10. bThe absolute difference in available data (as a %) between treatment and pretreatment conditions. That is, there was 9.24% more smartphone data during treatment versus pretreatment. Qualitative Evaluation A total of 4 major themes related to the study aims were developed from the 20 semistructured interviews, as shown in Figure 4. Quotes associated with each subtheme can be found in Table S2 in Multimedia Appendix 1, and quantified participant responses are described in Figure S2. The first major theme was the general participation experience. Protocol-related subthemes revolve around the idea that having a good relationship with the study team improves their general experience. A strong motivator of engagement was knowing that they were contributing to research, but experience was dampened by tedious study procedures in the form of repetitive and high-frequency questionnaires. Given the differences in preferences over when to receive feedback, how to receive feedback, and how much feedback to receive regarding the participants’ measured mood and behavior, it was thought that the flexibility to control these would have improved the experience. The level of engagement with the apps and devices was affected by physical discomfort of wearing a Fitbit; technology-related issues, which relate to any technical challenges, such as battery issues and measurement accuracy; and the tasks themselves, specifically their complexity and enjoyability, which added burden or ease to their engagement. Figure 4. Four major themes were developed from the interviews; each subpanel shows the numbered minor themes and subthemes. https://mental.jmir.org/2023/1/e42866 XSL•FO RenderX JMIR Ment Health 2023 | vol. 10 | e42866 | p. 10 (page number not for citation purposes) JMIR MENTAL HEALTH de Angel et al The interplay between mental health and engagement was found to be bidirectional. For example, when participants were unwell, some reported avoiding the self-reflection required by the questionnaires, whereas others experienced this only when feeling well. On the one hand, some reported an improvement through the encouragement of health-promoting behaviors, whereas others experienced guilt or anxiety from obsessing over data, especially if no improvement was apparent. As for the interplay between treatment and device engagement, the participants felt that the study had to integrate well with their treatment schedules. Treatment milestones were found to affect engagement with the study, with some people losing interest after therapy and others finding it harder to remain engaged with the increased burden of treatment. RMTs were seen increasing accountability with the therapist, providing targets to complete homework, and helping therapeutic conversations. The following section draws upon these results to aid the interpretation of engagement patterns. treatment effectiveness by to promote Discussion Principal Findings We evaluated the feasibility of using both active and passive data collection methods in the psychological treatment for depression. We examined recruitment, engagement with the study protocol through attrition, and engagement with the technology through patterns of missing data. We then used qualitative interviews to help interpret the feasibility data and gain insight into such data patterns. Study Engagement: Recruitment and Attrition Recruitment rates were low in proportion to the number of people initially contacted, which is in line with previous depression studies that found recruitment challenging, with large variations in success rates [34]. Digital health studies present potential participants with additional concerns, including unfamiliarity with technology and privacy misgivings [31], which may have contributed to low uptake]. Remote sensing studies tend to show a higher recruitment uptake than this study, for example, the study by Matcham et al [15], a discrepancy that could be largely explained by the exclusion of iPhone (Apple Inc) users and requirement to time recruitment with the start of treatment. Importantly, the similarity of our sample to the target population in key demographic aspects provides some reassurance that uptake was equitable across the main sample characteristics. A retention of 60% after 7 months is markedly lower than the 94% retention after a 9-month follow-up in a sample at different stages of recurrent major depressive disorder [15], yet it is higher than the 50% retention after 15 days in a self-referral study with little researcher contact [13] and the 53% retention in studies on psychotherapy treatments [35]. Some of the key differences between these studies that can help explain the differences in engagement rates stem from the context in which the studies were carried out, participant burden, and questionnaire frequency, the latter 2 being subthemes emerging from our interviews. https://mental.jmir.org/2023/1/e42866 XSL•FO RenderX We also found that treatment length and intensity affected attrition, as did higher baseline anxiety, with longer and more intense treatment being associated with early disengagement. In line with the finding from the participant interviews that participation burden increases during treatment, our quantitative analysis found that treatment characteristics and symptom severity influence retention. This could be partly because of the competing cognitive resources between engagement with RMT and treatment tasks, as described in the interviews. For example, other studies have found that constant feedback from health devices may worsen health anxieties [8], which may disproportionately impact those with higher anxiety compared with those with depression. The main implications for engagement with the study protocol relate to scalability, generalizability, and digital divide. On the one hand, slow recruitment may reflect a low readiness among patients to sign up for remote monitoring within health care services, with implications for increased scale-up costs and staff training. Higher attrition in people with clinical and treatment-related complexities may result in them deriving fewer benefits from the implementation of RMTs than their counterparts. In addition, studies using these samples may have a higher risk of attrition bias than those on less complex treatments. This limitation to the generalizability of research findings owing to bias would, therefore, disproportionally affect those with more complex needs, widening the digital divide. Engagement With RMTs: Data Availability The availability of data from the wearables and active questionnaires showed a similar pattern of decline over time. By contrast, smartphone-based passive data showed a low but stable data pattern. Some data streams, such as GPS and Bluetooth for passive data and speech and cognitive tests for active data, are more likely to be incomplete. Some forms of data collection place a greater burden on the user than others. Therefore, it was expected that active data collection forms that have the highest participant burden would have a faster pace of decline as a function of time and cause study fatigue [36]. Therefore, it was unsurprising that the active data streams that contributed the fewest data points were speech and cognitive tasks, which were lengthier and, according to our interviews, more cognitively demanding. Although wearables require very little engagement, they still involve some level of action: they must be worn, charged, and synchronized. According to the participant interviews, 50% of those interviewed chose not to wear the Fitbit because of comfort and privacy issues, among other reasons. By contrast, passive apps are unobtrusive in their data collection, as they do not require a regular smartphone user to deviate from their usual behavior, and, therefore, produced a more stable pattern of data availability, which seemed to be less affected by study burden. The passive data streams most vulnerable to missing data were the GPS and Bluetooth sensors. However, other passive sensing studies on mental health have found the opposite pattern, with more data being available from GPS than from accelerometers [15,20,37]. Sensor noncollection can occur for multiple reasons, including participants turning off the data permissions or the sensor itself. GPS and Bluetooth are sensors that can be easily JMIR Ment Health 2023 | vol. 10 | e42866 | p. 11 (page number not for citation purposes) JMIR MENTAL HEALTH de Angel et al switched off from a smartphone’s main setting page and may be seen as more intrusive forms of monitoring. This was supported by the finding that 35% of those interviewed felt “monitored” by the apps and the recurrence of “privacy” as a theme in RMT research for health care [38]. When comparing the participants in treatment with those on the waiting list, there was an increased completion of active tasks during therapy. In the current sample, the participants expressed the benefits of having a cohesive experience with RMTs and treatment such that completing active tasks during treatment helped with homework, promoted working on their mental health, and sparked conversations with their therapist. The literature on self-management in digital health shows that, despite the potential for added burden, there is a disposition for symptom tracking during treatment [39]. Conversely, people in treatment had less Fitbit wear time than those on the waiting list. The increased self-awareness that comes from tracking health with the Fitbit can be demotivating if there are no evident improvements in health outcomes such as sleep and physical activity [8], which might increase the likelihood of participants removing the device to avoid feelings of guilt and internal pressure [40]. The implications for engagement with RMTs relate to the integrity of the data collected and the differing acceptability thresholds for different devices. Low data availability means that the features derived from passive data may lack accuracy and could lead to false interpretations; however, even from the same device, different sensors contribute different amounts of data. Given that multiple sensor combinations are used to infer different aspects of behavior, accurate feature construction may require longer data collection windows for certain sensors depending on their target behavior. Although data imputation methods may help address some of these issues, increasing data availability using engagement strategies is likely to yield more accurate results. Several suggestions have been proposed by Currey and Torous [20], including overcoming the tendency of smartphones to halt data collection when apps are idle by including active components in passive apps. Additionally, if different devices (eg, smartphones vs wearables) produce different levels of data availability, this may have implications for the appropriateness of their use, depending on the purpose and length of data collection. Smartphone-based data collection for long-term monitoring is only appropriate if the resources are available to support increased data availability strategies; otherwise, the amount of data may be too scarce to be informative. If such strategies are in place, smartphone-based data, despite having a lesser overall amount, may be a more suitable option than wearables for long-term monitoring since wearables initially provide more data, but that amount gradually decreases over time. This study demonstrated wearables to be a feasible method for collecting activity and sleep data, 2 core items in psychotherapy for depression, for at least 32 consecutive weeks, before data availability falls below the 40% to mark. Therefore, smartphone-based data in a naturalistic context involving relatively longer-term treatments. However, it is important to consider strategies to increase user engagement with technology this method may be superior https://mental.jmir.org/2023/1/e42866 XSL•FO RenderX that take a patient-centered approach, including selecting measures that are meaningful to patients [41]. Limitations and Future Directions This was a longitudinal cohort study, so the comparison groups of treatment versus nontreatment differed in more ways than only the exposure to treatment and different treatment intensities. The participants in the treatment group were compared with those in the same week of the study but who had yet to start treatment; delayed treatment start was related to treatment intensity, clinical risk, catchment area for the health care center, and symptom severity. Despite our efforts to account for these variables in the analysis, there is a possibility of residual confounders. Future studies could quantify the components of treatment that are related to poorer engagement. Engagement with RMTs is broadly defined as data availability, which assumes that the occurrence of missing data is because of the participants deliberately disengaging. However, despite presenting some evidence of personal and clinical characteristics related to data availability, missing data can also be completely missing at random because of software errors. These factors may affect data streams differently based on technical factors. Future research could determine the nature of missingness by mapping technical issues to missing data. There is no standard method to establish a threshold for “missing data.” In this study, we justify a minimum of 8 hours of passive data and at least one active task completed; however, other studies (eg, the study by Matcham et al [15]) considered data availability as a single point of data per hour. It is critical to understand how much missing data are admissible before the integrity of the data is affected so that there can be an accurate characterization of the behavior. A single point of active data may describe a symptom experience for the previous 2 weeks, whereas a single passive data point covers a second’s worth of activity. Therefore, rates of missingness need to be interpreted with this relativity in mind, and future studies should work toward establishing acceptable thresholds for data availability for each behavioral feature under study. This study has shown engagement differences between data collection types and a difference in engagement between those in treatment and those on the waiting list; however, despite the statistical significance, future work should attempt to establish whether these differences are clinically meaningful. Finally, the COVID-19 pandemic has given rise to a rapid adoption of technology, especially in the health care sector [42]. This is likely to have had an impact on people’s attitudes toward digital tools for health monitoring and, consequently, engagement with RMTs. As a result, this study may have picked up on higher technology acceptance, or conversely, technology fatigue, as a factor of time. Conclusions We investigated the feasibility of remote collection methods in the psychological treatment for depression and reported the extent to which it was feasible to collect active and passive data via RMTs in a population with depression within a health care setting. Uptake was low but equal across the main demographic JMIR Ment Health 2023 | vol. 10 | e42866 | p. 12 (page number not for citation purposes) JMIR MENTAL HEALTH de Angel et al categories and, therefore, broadly representative of the target population. Retention in our study was low, but comparable with retention rates in psychological therapy [35]. Treatment characteristics such as length and intensity were associated with attrition, as was higher baseline anxiety, suggesting that patients undergoing more complex treatment may perceive fewer benefits from long-term remote monitoring. In addition, different data streams showed different levels of missing data despite being gathered from the same device, implying that different sensors may require different data collection protocols to ensure sufficient data for accurate feature construction. Being in treatment also affected RMT engagement in different ways, depending on the device, with Fitbit contributing less data during treatment but active tasks being completed more often. Future work should establish acceptable thresholds for data availability for different sensors and devices to ensure a minimum requirement for the integrity of RMT data and investigate which aspects of treatment are related to poorer engagement. Finally, successful implementation of RMTs requires more than the user engagement measures presented in this study; however, adopting these user engagement measures is a key first step. Acknowledgments This research was reviewed by a team with experience of mental health problems and their caregivers, trained to advise on research through the Feasibility and Acceptability Support Team for Researchers: a free, confidential service in England provided by the National Institute for Health Research (NIHR) Maudsley Biomedical Research Centre. This study represents independent research funded by the National Institute for Health and Care Research Biomedical Research Centre in South London and Maudsley National Health Service (NHS) Foundation Trust and King’s College London. The views expressed are those of the authors and not necessarily those of the NHS, National Institute for Health and Care Research, or Department of Health and Social Care. The funders had no role in the study design, data collection, data analysis, data interpretation, or writing of the manuscript. The authors would like to acknowledge the staff and students of King’s College London and the South London and Maudsley Improving Access to Psychological Therapies services in Lewisham, Croydon, and Lambeth who contributed to and supported this project. The authors convey special thanks to Katherine Tallent, Areej Elgaziari, Rebecca Chapman, and Alice Pace at Improving Access to Psychological Therapies for their help in recruiting and coordinating the link between their services and our project and to Nor Jamil, Lucy Hyam, and Alice Nynabb at King’s College London for their hard work on recruitment and the overall running of the study. RD was supported by the following: (1) NIHR Biomedical Research Centre at South London and Maudsley NHS Foundation Trust and King’s College London, United Kingdom; (2) Health Data Research United Kingdom, which is funded by the UK Medical Research Council, Engineering and Physical Sciences Research Council, Economic and Social Research Council, Department of Health and Social Care (England), Chief Scientist Office of the Scottish Government Health and Social Care Directorates, Health and Social Care Research and Development Division (Welsh Government), Public Health Agency (Northern Ireland), British Heart Foundation, and Wellcome Trust; (3) The BigData@Heart consortium, funded by the Innovative Medicines Initiative-2 Joint Undertaking under grant agreement number 116074 (this joint undertaking receives support from the European Union’s Horizon 2020 research and innovation program and European Federation of Pharmaceutical Industries and Associations and is chaired by DE Grobbee and SD Anker, partnering with 20 academic and industry partners and European Society of Cardiology) (4) the NIHR University College London Hospitals Biomedical Research Centre; (5) the NIHR Biomedical Research Centre at South London and Maudsley NHS Foundation Trust and King’s College London; (6) the UK Research and Innovation London Medical Imaging and Artificial Intelligence Centre for Value Based Healthcare; and (7) the NIHR Applied Research Collaboration South London (NIHR Applied Research Collaboration South London) at King’s College Hospital NHS Foundation Trust. Conflicts of Interest MH is the principal investigator of the Remote Assessment of Disease and Relapse–Central Nervous System program, a precompetitive public-private partnership funded by the Innovative Medicines Initiative and the European Federation of Pharmaceutical Industries and Associations. Multimedia Appendix 1 Supplementary information to the paper, including information on treatments and quantitative and qualitative analyses. [DOCX File , 137 KB-Multimedia Appendix 1] References 1. GBD 2019 Diseases and Injuries Collaborators. Global burden of 369 diseases and injuries in 204 countries and territories, 1990-2019: a systematic analysis for the Global Burden of Disease Study 2019. Lancet 2020 Oct 17;396(10258):1204-1222 [FREE Full text] [doi: 10.1016/S0140-6736(20)30925-9] [Medline: 33069326] https://mental.jmir.org/2023/1/e42866 XSL•FO RenderX JMIR Ment Health 2023 | vol. 10 | e42866 | p. 13 (page number not for citation purposes) JMIR MENTAL HEALTH de Angel et al 5. 4. 3. 2. Leone M, Kuja-Halkola R, Leval A, D'Onofrio BM, Larsson H, Lichtenstein P, et al. Association of youth depression with subsequent somatic diseases and premature death. JAMA Psychiatry 2021 Mar 01;78(3):302-310 [FREE Full text] [doi: 10.1001/jamapsychiatry.2020.3786] [Medline: 33295952] Gelenberg AJ, Freeman MP, Markowitz JC, Rosenbaum JF, Thase ME, Trivedi MH, et al. Practice Guideline for the Treatment of Patients with Major Depressive Disorder. 3rd edition. Washington, DC, USA: American Psychiatric Association; 2010. Depression in adults: treatment and management - NICE Guideline. National Institute for Health and Care Excellence. 2022 Jun 29. URL: https://www.nice.org.uk/guidance/ng222/chapter/Recommendations#treatment-for-a-new-episode- of-less-severe-depression [accessed 2022-07-13] Clark DM. Realizing the mass public benefit of evidence-based psychological therapies: the IAPT program. Annu Rev Clin Psychol 2018 May 07;14:159-183 [FREE Full text] [doi: 10.1146/annurev-clinpsy-050817-084833] [Medline: 29350997] 6. More than half of patients who finished psychological therapy recovered in 2017-18. National Health Service Digital. 2018 Nov 20. URL: https://digital.nhs.uk/news/2018/more-than-half-of-patients-who-finished-psychological-therapy-recovered- in-2017-18 [accessed 2022-02-23] Passos IC, Ballester P, Rabelo-da-Ponte F, Kapczinski F. Precision psychiatry: the future is now. Can J Psychiatry 2022 Jan;67(1):21-25 [FREE Full text] [doi: 10.1177/0706743721998044] [Medline: 33757313] de Angel V, Lewis S, White KM, Matcham F, Hotopf M. Clinical targets and attitudes toward implementing digital health tools for remote measurement in treatment for depression: focus groups with patients and clinicians. JMIR Ment Health 2022 Aug 15;9(8):e38934 [FREE Full text] [doi: 10.2196/38934] [Medline: 35969448] Insel TR. Digital phenotyping: technology for a new science of behavior. JAMA 2017 Oct 03;318(13):1215-1216. [doi: 10.1001/jama.2017.11295] [Medline: 28973224] 8. 7. 9. 10. White KM, Williamson C, Bergou N, Oetzmann C, de Angel V, Matcham F, et al. A systematic review of engagement 11. reporting in remote measurement studies for health symptom tracking. NPJ Digit Med 2022 Jun 29;5(1):82 [FREE Full text] [doi: 10.1038/s41746-022-00624-7] [Medline: 35768544] de Girolamo G, Barattieri di San Pietro C, Bulgari V, Dagani J, Ferrari C, Hotopf M, et al. The acceptability of real‐time health monitoring among community participants with depression: a systematic review and meta‐analysis of the literature. Depress Anxiety 2020 Apr 27;37(9):885-897 [FREE Full text] [doi: 10.1002/da.23023] 12. De Angel V, Lewis S, White K, Oetzmann C, Leightley D, Oprea E, et al. Digital health tools for the passive monitoring 13. of depression: a systematic review of methods. NPJ Digit Med 2022 Jan 11;5(1):3 [FREE Full text] [doi: 10.1038/s41746-021-00548-8] [Medline: 35017634] Pratap A, Neto EC, Snyder P, Stepnowsky C, Elhadad N, Grant D, et al. Indicators of retention in remote digital health studies: a cross-study evaluation of 100,000 participants. NPJ Digit Med 2020 Feb 17;3:21 [FREE Full text] [doi: 10.1038/s41746-020-0224-8] [Medline: 32128451] 14. Torous J, Lipschitz J, Ng M, Firth J. Dropout rates in clinical trials of smartphone apps for depressive symptoms: a systematic review and meta-analysis. J Affect Disord 2020 Mar 15;263:413-419. [doi: 10.1016/j.jad.2019.11.167] [Medline: 31969272] 15. Matcham F, Leightley D, Siddi S, Lamers F, White KM, Annas P, RADAR-CNS consortium. Remote Assessment of Disease and Relapse in Major Depressive Disorder (RADAR-MDD): recruitment, retention, and data availability in a longitudinal remote measurement study. BMC Psychiatry 2022 Mar 21;22(1):136 [FREE Full text] [doi: 10.1186/s12888-022-03753-1] [Medline: 35189842] 16. Goldberg SB, Buck B, Raphaely S, Fortney JC. Measuring psychiatric symptoms remotely: a systematic review of remote measurement-based care. Curr Psychiatry Rep 2018 Aug 28;20(10):81. [doi: 10.1007/s11920-018-0958-z] [Medline: 30155749] 17. Gual-Montolio P, Martínez-Borba V, Bretón-López JM, Osma J, Suso-Ribera C. How are information and communication technologies supporting routine outcome monitoring and measurement-based care in psychotherapy? A systematic review. Int J Environ Res Public Health 2020 May 02;17(9):3170 [FREE Full text] [doi: 10.3390/ijerph17093170] [Medline: 32370140] 18. Ng MM, Firth J, Minen M, Torous J. User engagement in mental health apps: a review of measurement, reporting, and validity. Psychiatr Serv 2019 Jul 01;70(7):538-544 [FREE Full text] [doi: 10.1176/appi.ps.201800519] [Medline: 30914003] 19. Matcham F, Carr E, White KM, Leightley D, Lamers F, Siddi S, RADAR-CNS consortium. Predictors of engagement with remote sensing technologies for symptom measurement in Major Depressive Disorder. J Affect Disord 2022 Aug 01;310:106-115. [doi: 10.1016/j.jad.2022.05.005] [Medline: 35525507] 20. Currey D, Torous J. Increasing the value of digital phenotyping through reducing missingness: a retrospective analysis. 21. 22. medRxiv 2022 May 17 [FREE Full text] [doi: 10.1101/2022.05.17.22275182] de Angel V, Lewis S, Munir S, Matcham F, Dobson R, Hotopf M. Using digital health tools for the Remote Assessment of Treatment Prognosis in Depression (RAPID): a study protocol for a feasibility study. BMJ Open 2022 May 06;12(5):e059258 [FREE Full text] [doi: 10.1136/bmjopen-2021-059258] [Medline: 35523486] Sim J, Lewis M. The size of a pilot study for a clinical trial should be calculated in relation to considerations of precision and efficiency. J Clin Epidemiol 2012 Mar;65(3):301-308. [doi: 10.1016/j.jclinepi.2011.07.011] [Medline: 22169081] https://mental.jmir.org/2023/1/e42866 XSL•FO RenderX JMIR Ment Health 2023 | vol. 10 | e42866 | p. 14 (page number not for citation purposes) JMIR MENTAL HEALTH de Angel et al 23. Sheehan DV, Lecrubier Y, Sheehan KH, Amorim P, Janavs J, Weiller E, et al. The Mini-International Neuropsychiatric Interview (M.I.N.I.): the development and validation of a structured diagnostic psychiatric interview for DSM-IV and ICD-10. J Clin Psychiatry 1998;59 Suppl 20:22-57. [Medline: 9881538] 24. Kroenke K, Spitzer RL, Williams JB. The PHQ-9: validity of a brief depression severity measure. J Gen Intern Med 2001 25. Sep;16(9):606-613 [FREE Full text] [doi: 10.1046/j.1525-1497.2001.016009606.x] [Medline: 11556941] Spitzer RL, Kroenke K, Williams JB, Löwe B. A brief measure for assessing generalized anxiety disorder: the GAD-7. Arch Intern Med 2006 May 22;166(10):1092-1097. [doi: 10.1001/archinte.166.10.1092] [Medline: 16717171] 26. Ranjan Y, Rashid Z, Stewart C, Conde P, Begale M, Verbeeck D, RADAR-CNS Consortium. RADAR-base: open source mobile health platform for collecting, monitoring, and analyzing data using sensors, wearables, and mobile devices. JMIR Mhealth Uhealth 2019 Aug 01;7(8):e11734 [FREE Full text] [doi: 10.2196/11734] [Medline: 31373275] 27. Harris PA, Taylor R, Thielke R, Payne J, Gonzalez N, Conde JG. Research electronic data capture (REDCap)--a metadata-driven methodology and workflow process for providing translational research informatics support. J Biomed Inform 2009 Apr;42(2):377-381 [FREE Full text] [doi: 10.1016/j.jbi.2008.08.010] [Medline: 18929686] 28. McIntyre RS, Best MW, Bowie CR, Carmona NE, Cha DS, Lee Y, et al. The THINC-integrated tool (THINC-it) screening assessment for cognitive dysfunction: validation in patients with major depressive disorder. J Clin Psychiatry 2017 Jul;78(7):873-881. [doi: 10.4088/JCP.16m11329] [Medline: 28858441] International Phonetic Association. Handbook of the International Phonetic Association: A Guide to the Use of the International Phonetic Alphabet. Cambridge, UK: Cambridge University Press; 1999. 29. 30. Neale J. Iterative categorization (IC): a systematic technique for analysing qualitative data. Addiction 2016 31. Jun;111(6):1096-1106 [FREE Full text] [doi: 10.1111/add.13314] [Medline: 26806155] Simblett S, Matcham F, Siddi S, Bulgari V, Barattieri di San Pietro C, Hortas López J, RADAR-CNS Consortium. Barriers to and facilitators of engagement with mHealth technology for remote measurement and management of depression: qualitative analysis. JMIR Mhealth Uhealth 2019 Jan 30;7(1):e11325 [FREE Full text] [doi: 10.2196/11325] [Medline: 30698535] 32. Harwood H, Rhead R, Chui Z, Bakolis I, Connor L, Gazard B, et al. Variations by ethnicity in referral and treatment 33. pathways for IAPT service users in South London. Psychol Med (forthcoming) 2021 Aug 02:1-12. [doi: 10.1017/S0033291721002518] [Medline: 34334151] Psychological Therapies, Annual Reports on the use of IAPT services. National Health Service Digital. 2021. URL: https:/ /digital.nhs.uk/data-and-information/publications/statistical/psychological-therapies-annual-reports-on-the-use-of-iapt-services [accessed 2022-07-08] 34. Liu Y, Pencheon E, Hunter RM, Moncrieff J, Freemantle N. Recruitment and retention strategies in mental health trials - a systematic review. PLoS One 2018 Aug 29;13(8):e0203127 [FREE Full text] [doi: 10.1371/journal.pone.0203127] [Medline: 30157250] 35. Wierzbicki M, Pekarik G. A meta-analysis of psychotherapy dropout. Prof Psychol Res Pr 1993 May;24(2):190-195 [FREE Full text] [doi: 10.1037/0735-7028.24.2.190] 36. Dineley J, Lavelle G, Leightley D, Matcham F, Siddi S, Peñarrubia-María MT, et al. Remote smartphone-based speech collection: acceptance and barriers in individuals with major depressive disorder. In: Proceedings of the 2021 Interspeech. 2021 Presented at: Interspeech '21; August 30-September 3, 2021; Brno, Czechia p. 631-635 URL: https://www. isca-speech.org/archive/interspeech_2021/dineley21_interspeech.html [doi: 10.21437/interspeech.2021-1240] 37. Kiang MV, Chen JT, Krieger N, Buckee CO, Alexander MJ, Baker JT, et al. Sociodemographic characteristics of missing data in digital phenotyping. Sci Rep 2021 Jul 29;11(1):15408 [FREE Full text] [doi: 10.1038/s41598-021-94516-7] [Medline: 34326370] 38. Zhang H, Ibrahim A, Parsia B, Poliakoff E, Harper S. Passive social sensing with smartphones: a systematic review. Computing (forthcoming) 2022 Aug 12 [FREE Full text] [doi: 10.1007/s00607-022-01112-2] 39. Niendam TA, Tully LM, Iosif AM, Kumar D, Nye KE, Denton JC, et al. Enhancing early psychosis treatment using smartphone technology: a longitudinal feasibility and validity study. J Psychiatr Res 2018 Jan;96:239-246. [doi: 10.1016/j.jpsychires.2017.10.017] [Medline: 29126059] 40. Kerner C, Goodyear V. The motivational impact of wearable healthy lifestyle technologies: a self-determination perspective on Fitbits with adolescents. Am J Health Educ 2017 Jul 28;48(5):287-297 [FREE Full text] [doi: 10.1080/19325037.2017.1343161] 41. Manta C, Patrick-Lake B, Goldsack J. Digital measures that matter to patients: a framework to guide the selection and development of digital measures of health. Digit Biomark 2020 Sep 15;4(3):69-77 [FREE Full text] [doi: 10.1159/000509725] [Medline: 33083687] 42. Hutchings R. The impact of Covid-19 on the use of digital technology in the NHS. Nuffield Trust. 2020 Aug. URL: https:/ /www.nuffieldtrust.org.uk/files/2020-08/the-impact-of-covid-19-on-the-use-of-digital-technology-in-the-nhs-web-2.pdf [accessed 2022-09-09] https://mental.jmir.org/2023/1/e42866 XSL•FO RenderX JMIR Ment Health 2023 | vol. 10 | e42866 | p. 15 (page number not for citation purposes) JMIR MENTAL HEALTH Abbreviations IAPT: Improving Access to Psychological Therapies RADAR: Remote Assessment of Disease and Relapse REDCap: Research Electronic Data Capture RMT: remote measurement technology de Angel et al Edited by J Torous; submitted 21.09.22; peer-reviewed by M Kapsetaki, S Difrancesco; comments to author 05.11.22; revised version received 10.11.22; accepted 26.11.22; published 24.01.23 Please cite as: de Angel V, Adeleye F, Zhang Y, Cummins N, Munir S, Lewis S, Laporta Puyal E, Matcham F, Sun S, Folarin AA, Ranjan Y, Conde P, Rashid Z, Dobson R, Hotopf M The Feasibility of Implementing Remote Measurement Technologies in Psychological Treatment for Depression: Mixed Methods Study on Engagement JMIR Ment Health 2023;10:e42866 URL: https://mental.jmir.org/2023/1/e42866 doi: 10.2196/42866 PMID: ©Valeria de Angel, Fadekemi Adeleye, Yuezhou Zhang, Nicholas Cummins, Sara Munir, Serena Lewis, Estela Laporta Puyal, Faith Matcham, Shaoxiong Sun, Amos A Folarin, Yatharth Ranjan, Pauline Conde, Zulqarnain Rashid, Richard Dobson, Matthew Hotopf. Originally published in JMIR Mental Health (https://mental.jmir.org), 24.01.2023. This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Mental Health, is properly cited. The complete bibliographic information, a link to the original publication on https://mental.jmir.org/, as well as this copyright and license information must be included. https://mental.jmir.org/2023/1/e42866 XSL•FO RenderX JMIR Ment Health 2023 | vol. 10 | e42866 | p. 16 (page number not for citation purposes)
10.1021_acschembio.3c00092
pubs.acs.org/acschemicalbiology Articles A Fluorescence Polarization Assay for Macrodomains Facilitates the Identification of Potent Inhibitors of the SARS-CoV‑2 Macrodomain Ananya Anmangandla,# Sadhan Jana,# Kewen Peng,# Shamar D. Wallace,# Saket R. Bagde, Bryon S. Drown, Jiashu Xu, Paul J. Hergenrother, J. Christopher Fromme,* and Hening Lin* Cite This: ACS Chem. Biol. 2023, 18, 1200−1207 Read Online ACCESS Metrics & More Article Recommendations *sı Supporting Information ABSTRACT: Viral macrodomains, which can bind to and/or hydrolyze adenine diphosphate ribose (ADP-ribose or ADPr) from proteins, have been suggested to counteract host immune response and be viable targets for the development of antiviral drugs. Therefore, developing high-throughput screening (HTS) techni- ques for macrodomain inhibitors is of great interest. Herein, using a novel tracer TAMRA-ADPr, an ADP-ribose compound conjugated with tetramethylrhodamine, we developed a robust fluorescence polarization assay for various viral and human macrodomains including SARS-CoV-2 Macro1, VEEV Macro, CHIKV Macro, human MacroD1, MacroD2, and PARP9 Macro2. Using this assay, we validated Z8539 (IC50 6.4 μM) and GS441524 (IC50 15.2 μM), two literature-reported small-molecule inhibitors of SARS-CoV-2 Macro1. Our data suggest that GS441524 is highly selective for SARS-CoV-2 Macro1 over other human and viral macrodomains. Furthermore, using this assay, we identified pNP-ADPr (ADP-ribosylated p-nitrophenol, IC50 370 nM) and TFMU-ADPr (ADP-ribosylated trifluoromethyl umbelliferone, IC50 590 nM) as the most potent SARS-CoV-2 Macro1 binders reported to date. An X-ray crystal structure of SARS-CoV-2 Macro1 in complex with TFMU-ADPr revealed how the TFMU moiety contributes to the binding affinity. Our data demonstrate that this fluorescence polarization assay is a useful addition to the HTS methods for the identification of macrodomain inhibitors. ■ INTRODUCTION COVID-19 is an ongoing global pandemic caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) that has led to more than 6.8 million deaths and over 759 million cases worldwide.1 As one of the major host defense mechanisms against viral infections, interferon (IFN) signaling is activated when host cells detect viral invasions.2 A central effector of IFN activation is the ADP-ribosylation of host cell proteins, and these ADP-ribose (ADPr) tags play important roles in regulating protein activities and are thus vital for a successful defense against viral infection.3 However, SARS-CoV-2 can counter IFN-induced mono- ADP-ribosylation (MARylation) in host cells through its first macrodomain (Macro1) encoded within the non-structural protein 3 (nsp3).4 Macrodomains are ancient and well- conserved structural modules found in a wide range of proteins with diverse biological functions. Macrodomains are known to bind, and in some cases, hydrolyze ADP-ribosylated proteins, thus functioning as either “readers” or “erasers” of ADPr modifications. Viral macrodomains, including those of corona- viruses (CoVs), the Venezuelan equine encephalitis virus (VEEV), and the Chikungunya virus (CHIKV), are reported to hydrolyze MARylated host proteins and are responsible for attenuating host immune responses against viral infection.4−6 Given the central roles that viral macrodomains play in host cell immune responses, macrodomain inhibitors are potential antiviral agents. However, several macrodomains are also encoded by human proteins, including MacroD1, MacroD2, PARP9, and TARG1, which may have important physiological functions.7 Therefore, selective viral macrodomain inhibitors with minimal off-target effects are highly desirable. With the emergence of the COVID-19 global pandemic, multiple research efforts have been directed toward developing high-throughput screening (HTS) methods for the identifica- tion of SARS-CoV-2 Macro1 inhibitors. For instance, Dasovich et al.8 reported a luminescence-based assay termed ADPr-Glo, which utilizes an ADP-ribosylated peptide that can be hydrolyzed by SARS-CoV-2 Macro1. The hydrolysis product ADPr can be further hydrolyzed into AMP by the Received: February 10, 2023 Accepted: April 10, 2023 Published: May 1, 2023 © 2023 The Authors. Published by American Chemical Society 1200 https://doi.org/10.1021/acschembio.3c00092 ACS Chem. Biol. 2023, 18, 1200−1207 ACS Chemical Biology pubs.acs.org/acschemicalbiology Articles phosphodiesterase NudF and is subsequently converted into luminance by the commercially available AMP-Glo and quantified. Schuller et al.9 screened over 200 crystallographic and virtual screening hits using a homogeneous time-resolved fluorescence (HTRF) assay and differential scanning fluorim- etry (DSF) assay. These HTS assays for SARS-CoV-2 Macro1 have several limitations. For the ADPr-Glo assay, an extra enzyme NudF is used, which can complicate the result since compounds may also affect NudF activity. HTRF utilizes three expensive reagents (ADPr-conjugated biotin peptide, FRET donors and acceptors), rendering it less cost-effective for large-scale screening. Finally, the DSF assay is not suitable for high-throughput screening. More facile HTS methods have been developed for other macrodomains such as PARG,10,11 but they cannot be used for other macrodomains. The fluorescence polarization (FP) assay, which exploits the polarization of a fluorophore being inversely related to its freedom of motion,12 provides a useful addition to the screening methods mentioned above. The FP assay is a simple and high- throughput assay that can be performed in ambient conditions, and the only reagent it requires other than the protein of interest is a fluorophore-conjugated ligand (so-called “tracer”). Very recently, Roy et al.13 developed an FP assay for the screening of SARS-CoV-2 Macro1 using fluorescein-labeled and ADP- ribosylated peptide as the tracer. However, they could only achieve an assay window of less than 60 milipolarization (mP) using a high protein concentration of 15 μM, which suggests that the tracer may not be a high-affinity binder of SARS-CoV-2 Macro1 and necessitates the use of large quantities of protein, limiting its use as a high-throughput screening method. Herein, we designed and synthesized a novel FP tracer, TAMRA-ADPr. Using this tracer, we established a robust binding assay with a wider mP shift window and successfully applied this assay to a variety of macrodomains, including SARS- CoV-2 Macro1, VEEV Macro, CHIKV Macro, human MacroD1, MacroD2, and PARP9 Macro2. Using this assay, we were able to validate two small-molecule SARS-CoV-2 Macro1 inhibitors reported in the literature. Furthermore, we tested several ADPr derivatives and identified two compounds to be the most potent binders of SARS-CoV-2 Macro1 known to date. ■ RESULTS AND DISCUSSION To develop an FP assay suitable for the high-throughput screening of SARS-CoV-2 Macro1 inhibitors, we first designed a tracer molecule TAMRA-ADPr (Figure 1), inspired by the presumed structure of macrodomain substrates.7 For the synthesis of TAMRA-ADPr, ADPr-N3 was first synthesized and then coupled to alkyne-TAMRA at the C1″ position via click chemistry (see the SI for details). Gratifyingly, TAMRA-ADPr showed relatively strong binding to SARS-CoV-2 Macro1 in a titration assay where the mP shift reached over 110 when 10 μM Macro1 was used (Figure 2A). Encouraged by this result, we tested five additional macro domains: human MacroD1, MacroD2, PARP9 Macro2, VEEV Macro, and CHIKV Macro. TAMRA-ADPr could bind to each of these macro domains, albeit with different affinities. Based on the mP shift data, MacroD1 and MacroD2 are the most potent binders of TAMRA-ADPr, reaching mP shifts of more than 100 at a low concentration of 0.38 μM (Figure 2B). This is consistent with the previous finding that ADPr is a strong binder of both MacroD1 (KD = 0.72 μM)14 and MacroD2 (KD = 0.15 μM).15 Figure 1. Design and mechanism of a fluorescence polarization (FP) assay for ADPr-binding macrodomains. (A) Structure of TAMRA- ADPr. The TAMRA fluorophore is coupled to ADPr at C1″ through a long triazole-alkane linker. (B) In the absence of inhibitors, the majority of tracers is bound to protein. Thus, the free rotation of the fluorophore is hindered and a high fluorescence polarization is observed. Upon addition of inhibitor, there is competition for binding and the tracer is released from the macrodomain. The unbound tracer molecules are now free to rotate, leading to a lower observed polarization. On the other hand, only an ∼70 mP shift could be achieved by VEEV Macro and CHIKV Macro at 6 μM (Figure 2C), suggesting that they are weaker binders of TAMRA-ADPr. Having obtained a satisfactory tracer, we next designed a convenient “mix and read” FP assay where different concen- trations of compounds to be tested were incubated with SARS- CoV-2 Macro1 and the tracer for 30 min before the mP shifts were read on a plate reader. The percent binding of the tracer relative to the negative control (protein and tracer only) was calculated and fitted to an IC50 curve. It should be noted that the protein concentrations were chosen to give an mP shift window of at least 50 to yield data with acceptable errors and therefore differ for each macrodomain (see Materials and Methods). We first tested ADPr, a well-characterized ligand for SARS-CoV-2 Macro1 as well as many other macro domains, to see whether our FP assay could quantitatively capture the binding affinity of macrodomain ligands. The IC50 of ADPr against the tracer binding to SARS-CoV-2 Macro1 was determined to be 15.5 μM (Figure 2D,F), which is comparable to the reported KD value of 11.6 μM.16 We further determined the IC50 values of ADPr against other macrodomains (Figure 2D,F) and were pleased to find the IC50 values were all consistent with the reported KD values of ADPr for different macrodomains.6,14,15 As a negative control, we also showed that iso-ADPr, the smallest internal structural unit containing the characteristic ribose−ribose glycosidic bond for poly-ADPr (PAR),17,18 did not compete with the tracer in macrodomain binding (Figure 2E). Therefore, we concluded that this competitive FP assay is a reliable screening method for potential inhibitors of SARS-CoV-2 Macro1 and other ADPr-binding macro domains. Since TAMRA-ADPr is a good binder of SARS-CoV-2 Macro1, we thought it would be interesting to see whether ADPr-N3, the precursor of TAMRA-ADPr, could also bind SARS-CoV-2 Macro1. As shown in Figure 3B, ADPr-N3 was identified to be a more potent binder of SARS-CoV-2 Macro1 than ADPr with a nearly 2-fold smaller IC50. The binding activity of ADPr-N3 is not unexpected since the azido group is similar to the hydroxyl group in size and thus unlikely to cause steric 1201 https://doi.org/10.1021/acschembio.3c00092 ACS Chem. Biol. 2023, 18, 1200−1207 ACS Chemical Biology pubs.acs.org/acschemicalbiology Articles Figure 2. Validation of the FP assay. (A−C) mP shift values measured after 30 min incubation of 20 nM tracer with varying concentrations of macrodomains. (D) KD curves of ADPr for different macrodomains. (E) iso-ADPr does not compete with the tracer for all the macrodomains tested. (F) IC50 values for ADPr with different macrodomains. For all the data presented, error bars indicate SEM and IC50 values are reported as mean ± SEM, n = 2 or 3. Figure 3. IC50 determination of ADPr-N3, Z8539, and GS-441524 on SARS-CoV-2 Macro1. (A) Chemical structures of ADPr-N3, Z8539, and GS- 441524. (B) IC50 curve of ADP-N3 for SARS-CoV-2 Macro1. (C) IC50 curve of Z8539 for SARS-CoV-2 Macro1. (D) IC50 curves and values of GS- 441524 for different macrodomains. For all the data presented, error bars indicate SEM and IC50 values are reported as mean ± SEM, n = 2 or n = 3. clashes with the protein. Given that SARS-CoV-2 Macro1 can accommodate much bulkier groups at the C1″ position of ADPr, as shown by TAMRA-ADPr, ADPr-N3 may be a useful precursor for the development of ADPr-based inhibitors of SARS-CoV-2 Macro1 through click chemistry. We then tested two recently reported SARS-CoV-2 Macro1 inhibitors using the FP assay. Z8539 (Figure 3A) is a potent small-molecule inhibitor of SARS-CoV-2 Macro1 discovered very recently by Gahbauer et al.19 through a combined approach of virtual screening and fragment linking. Z8539 was found to be 1202 https://doi.org/10.1021/acschembio.3c00092 ACS Chem. Biol. 2023, 18, 1200−1207 ACS Chemical Biology pubs.acs.org/acschemicalbiology Articles Figure 4. pNP-ADPr and TFMU-ADPr are potent macrodomain binders. (A) Chemical structures of pNP-ADPr and TFMU-ADPr. (B) IC50 curves of pNP-ADPr and TFMU-ADPr for SARS-CoV-2 Macro1. IC50 curve of ADPr is also shown as a reference. (C) IC50 values of pNP-ADPr and TFMU- ADPr for SARS-CoV-2 Macro1. (D) Fold decrease in the IC50 values of pNP-ADPr and TFMU-ADPr relative to those of ADPr for each macrodomain tested. (E) Electron density of TFMU-ADPr in the complex with SARS-CoV-2 Macro1. (F) Structure of Macro1 in complex with TFMU-ADPr (cyan) is superimposed with that of Macro1 in complex with ADPr (gray, PDB 6YWL). TFMU-ADPr and ADPr are shown in stick representation. (G) Aromatic ring of TFMU interacts with the side chain of Ile131 and on the other side, Gly46 and Gly47. For all the data presented, error bars indicate SEM and IC50 values are reported as mean ± SEM, n = 2 or 3. a slightly better SARS-CoV-2 Macro1 binder than ADPr in a homogeneous time-resolved fluorescence (HTRF) assay. This result was validated in our FP assay, where the IC50 of Z8539 is 2-fold smaller than ADPr against SARS-CoV-2 Macro1 (Figure 3C). Z8539 is an encouraging example, showing that small- molecule inhibitors of SARS-CoV-2 Macro1 with structures unrelated to ADPr are possible. However, its binding affinity for SARS-CoV-2 Macro1 is only ∼6.4 μM. Another reported small-molecule inhibitor of SARS-CoV-2 Macro1 is GS-441524 (Figure 3A), the active metabolite of targets the viral RNA- Remdesivir, an antiviral drug that dependent RNA polymerase (RdRp).20 Remdesivir was shown to be effective against SARS-CoV-221 and is the first COVID-19 therapy approved by the FDA. Recently, Ni et al.22 found that GS-441524 can bind SARS-CoV-2 Macro1 and solved the crystal structure of SARS-CoV-2 Macro1 bound with GS- 441524. Using isothermal titration calorimetry (ITC), they determined that the KD of GS-441524 for SARS-CoV-2 Macro1 is 10.8 μM, similar to that of ADPr. Intrigued by this finding, we also tested GS-441524 in our FP assay. Consistent with the reported data, the IC50 of GS-441524 for SARS-CoV-2 Macro1 was determined to be 15.2 μM. Additionally, we tested whether GS-441524 could inhibit other macrodomains and were surprised to find that GS-441524 is a selective SARS-CoV-2 Macro1 inhibitor with no significant binding to all other macrodomains tested (Figure 3D). This result coincides with another paper published very recently,23 which showed that GS- 441524 is selective for SARS-CoV-2 Macro1 over other macrodomains including MERS-CoV Mac, CHIKV Macro, PARP14 Macro2, and PARP15 Macro2 in ITC experiments. Taken together, GS-441524 is a promising lead compound against SARS-CoV-2 Macro1 with high selectivity and ligand efficiency. Although the actual physiological substrates/binding partners for SARS-CoV-2 Macro1 are still unknown, it has been proposed that the most likely substrates are ADPr C1″-esters coupled to glutamic or aspartic acid protein residues.24 Taken together with our finding that TAMRA-ADPr with a C1″-triazole linkage can bind SARS-CoV-2 Macro1 with high affinity, it seems that bulky groups at the C1″ position would not disrupt binding and instead may confer a higher affinity. We therefore tested several other ADPr compounds. TFMU-ADPr and pNP-ADPr (Figure 1203 https://doi.org/10.1021/acschembio.3c00092 ACS Chem. Biol. 2023, 18, 1200−1207 ACS Chemical Biology pubs.acs.org/acschemicalbiology Articles 4A) are previously developed assay substrates for poly(ADP- ribosyl)glycohydrolase (PARG).25 Based on our proposal that bulkier substituents at the C1″ position may boost macro- domain binding, TFMU-ADPr and pNP-ADPr may be potential binders for macrodomains since the aromatic rings are introduced at C1″ similar to TAMRA-ADPr. Therefore, we tested these two compounds in the FP assay. We were surprised to find that pNP-ADPr is 40-fold more potent than ADPr for SARS-CoV-2 Macro1 with an IC50 of only 0.37 μM (Figure 4B− E), which is the strongest binder of SARS-CoV-2 Macro1 reported so far. Similarly, the IC50 of TFMU-ADPr against SARS-CoV-2 Macro1 is 0.59 μM, 25-fold smaller than ADPr (Figure 4B−E). We also tested these two compounds with other macrodomains and found that their IC50 values are similar to those of ADPr with the exception of MacroD2, for which both compounds showed a more than 10-fold increase in activity over ADPr (Figure 4C,D). Therefore, pNP-ADPr and TFMU-ADPr are both potent binders of macrodomains with strong preferences for SARS-CoV-2 Macro1. We also found that pNP-ADPr and TFMU-ADPr could inhibit the hydrolysis of ADPr by SARS-CoV-2 Marco1 and human MacroD1 in the cell lysate (Figure S1, Supporting Information). To understand why TFMU-ADPr binds strongly to Macro1, we determined the X-ray crystal structure of the Macro1- TFMU-ADPr complex using diffraction data that extended to 1.9 Å resolution. The TFMU-ADPr electron density is well resolved (Figure 4E). As expected, the inhibitor binds within the known ADPr binding site of Macro1 (Figure 4F, the structure is superimposed to that of the Macro1-ADPr complex PDB 6YWL). The TFMU moiety extends from the binding site along a narrow hydrophobic groove, bracketed on one side by the Ile131 side chain and on the other side by Gly46 and Gly47 (Figure 4G). ■ CONCLUSIONS In summary, we have developed TAMRA-ADPr, an ADPr- based tracer, and devised an FP-based binding assay for the screening of ADPr-binding macrodomain inhibitors. The reliability of the FP assay was confirmed by testing the IC50 values of ADPr against different macrodomains and comparing them to the reported KD values. Using this assay, we tested and validated Z8539 and GS-441524, two recently reported small- molecule inhibitors of SARS-CoV-2 Macro1. An interesting finding of this work is that pNP-ADPr and TFMU-ADPr are strong binders of SARS-CoV-2 Macro1 and several other macrodomains. Their structures may provide clues for the future design of more potent ADPr-based probe molecules of SARS- CoV-2 Macro1. We believe that the FP assay described herein is a convenient and robust screening method that can facilitate future drug discovery efforts for macrodomain inhibitors. ■ MATERIALS AND METHODS Reagents. pNP-ADPr and TFMU-ADPr are synthesized as previously described.25 Z8539 was obtained from Enamine (Z4718398539). GS-441524 was obtained from MedChemExpress (HY-103586). Expression and Purification of Macrodomains. Macrodomain plasmids were purchased from Twist Biosciences or Genscript in pET28 vectors (full sequences available in the SI). The plasmids were transformed into BL21(DE3) chemically competent E. coli. 4 L of LB broth with 50 μg/mL kanamycin was inoculated with an overnight starter grown at 37 °C. Cultures were grown at 200 rpm and 37 °C for ∼4 h until the OD600 reached 0.8. Then, IPTG was added to 0.5 mM and the cells were incubated at 16 ° C overnight to allow protein expression. Cells were harvested by centrifugation at 6000g. Cell pellets were frozen at −80 °C or immediately used for purification. Pellets were resuspended in lysis buffer (50 mM Tris pH 8.0, 500 mM NaCl, 0.5 mg/ mL lysozyme, 1 mM PMSF, and Pierce universal nuclease). Following a 30 min incubation, cells were sonicated on ice for 4 min total at 60% amplitude. The lysate was clarified at 4 °C and 30,000g for 35 min. The clarified lysate was loaded onto Ni-NTA resin, washed with 50 mL wash buffer (50 mM Tris pH 8.0, 500 mM NaCl, 20 mM imidazole), and eluted with elution buffer (50 mM Tris pH 8, 500 mM NaCl, 200 mM imidazole). Crude macrodomains were concentrated using a 10 kDa MWCO Amicon filter and loaded onto a HiLoad 16/600 Superdex 75 gel filtration column equilibrated with storage buffer (25 mM Tris pH 8.0, 150 mM NaCl, 10% glycerol) on an Ä KTA FPLC system. Fractions containing macrodomains were pooled, concentrated, flash-frozen in liquid nitrogen, and stored at −80 °C for future use. For SARS-CoV-2, the sample was supplemented with DTT (2 mM) and tobacco-etch protease and incubated at 4 °C overnight. The reaction mixture was then subjected to subtractive nickel chelate chromatography, and the eluate was injected into a HiLoad 16/600 Superdex75 gel filtration column equilibrated with protein storage buffer (5 mM HEPES and 150 mM NaCl, pH 7.5). Fractions containing purified the SARS-CoV-2 macrodomain were combined and concentrated. Then, samples were aliquoted, flash frozen using liquid nitrogen, and stored at −80 °C. Fluorescence Polarization Assay. The stock solution of purified macrodomain proteins was diluted with the assay buffer (25 mM Tris pH 8.0, 150 mM NaCl, and 0.01% Tween-20) to 2× final concentration. Final concentrations for each macrodomain were as follows: 0.5 μM for MacroD1 and MacroD2, 1.5 μM for SARS-CoV-2 Macro1 and PARP9 Macro2, and 5 μM for VEEV Macro and CHIKV Macro. TAMRA-ADPr (40 nM, 2× final concentration) was then added to the protein solution to give the assay solution. To each well of a 96-well black plate (Corning, #3915) was added 50 μL of the assay solution followed by the addition of 50 μL of the compound solution (2× final concentration) in the assay buffer. The plate was wrapped with aluminum foil and left at room temperature for 30 min. The plate was then scanned on a Cytation5 using a FP filter cube (Agilent, part number: 8040562, Ex: 530/25, Em: 590/35). The parallel and perpendicular fluorescence intensities of each well were recorded, and the mP values were then calculated based on the blank-subtracted data. Control wells include tracer-only wells where only 20 nM tracers were present and negative-control wells where only an appropriate concentration of macrodomain protein and 20 nM tracers were present. The percent binding of tracer relative to the control wells was calculated as follows: where mPtest, mPtracer, and mPneg are mP values of the test wells, tracer- only wells, and negative-control wells, respectively. The obtained data were then fitted into an IC50 curve using the sigmoidal four-parameter logistic model (bottom and top were constrained to be 0 and 100, respectively) implemented in GraphPad Prism 9.4.1 (GraphPad Software, Inc.). Co-crystallization of the SARS-CoV-2 Macro1-TFMU-ADPr Complex. SARS-CoV-2 Macro1 was mixed with TFMU-ADPr to final concentrations of 0.4 and 2 mM. The Macro1-inhibitor complex was crystallized by the hanging-drop method at 20 °C by mixing 1 μL of the Macro1-TFMU-ADPr solution with 1 μL well solution (200 mM sodium acetate, 100 mM Tris−HCl pH 8, and 30% PEG-4000). Crystals were observed after 5 days. Prior to freezing with liquid nitrogen, crystals were cryo-protected in well solution containing 10% ethylene glycol. Diffraction Data Collection, Structure Solution, Model Building, and Refinement. Diffraction data was collected on Northeastern Collaborative Access Team (NE-CAT) beamline 24- ID-E at Advanced Photon Source (APS). Initial data processing was performed by the NE-CAT ‘RAPD’ pipeline, which uses XDS for scaling and merging.26 The structure was solved by molecular replacement using Phaser27 in Phenix28 using a previously published 1204 https://doi.org/10.1021/acschembio.3c00092 ACS Chem. Biol. 2023, 18, 1200−1207 =relative%bindingoftracermPmPmPmPtesttracernegtracer ACS Chemical Biology pubs.acs.org/acschemicalbiology Articles structure of SARS-CoV-2 Macro122 (PDB:6YWL) as the search model. Coot29 was used for model building, and refinement and validation were performed in Phenix.30 There are two copies of the Macro1- inhibitor complex in the asymmetric unit, so non-crystallographic symmetry restraints were used during refinement. The final structures of the two copies are nearly identical, with no obvious differences in the inhibitor or inhibitor binding sites between the two copies, although the density for the inhibitor was stronger in one copy than the other. Data and refinement statistics are presented in Table 1. Table 1. Data Collection and Refinement Statisticsa PDB code resolution range space group unit cell total reflections unique reflections multiplicity completeness (%) mean I/sigma(I) Wilson B-factor R-merge R-meas R-pim CC1/2 CC* reflections used in refinement reflections used for R-free R-work R-free CC(work) CC(free) number of non-hydrogen atoms macromolecules ligands solvent protein residues RMS (bonds) RMS (angles) Ramachandran favored (%) Ramachandran allowed (%) Ramachandran outliers (%) Rotamer outliers (%) Clashscore average B-factor macromolecules ligands solvent number of TLS groups 8GIA 68.92−1.86 (1.926−1.86) C 1 2 1 140.517 Å 36.668 Å 65.056 Å 90° 101.211° 90° 186,998 (18900) 27,584 (2723) 6.8 (6.9) 99.15 (99.02) 6.94 (1.36) 28.01 0.1788 (1.245) 0.1943 (1.347) 0.07503 (0.5084) 0.987 (0.704) 0.997 (0.909) 27,525 (2717) 1353 (146) 0.2083 (0.3189) 0.2586 (0.3661) 0.938 (0.828) 0.918 (0.743) 2813 2535 150 176 335 0.005 0.60 97.89 2.11 0.00 0.36 7.90 35.84 35.81 30.80 39.19 12 aStatistics for the highest-resolution shell are shown in parentheses. ■ ASSOCIATED CONTENT *sı Supporting Information The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acschembio.3c00092. TFMU-ADPr and pNP-ADPr inhibition of macrodomain enzymatic activity, supplementary methods, and NMR spectra of compounds (PDF) ■ AUTHOR INFORMATION Corresponding Authors J. Christopher Fromme − Department of Molecular Biology and Genetics, Weill Institute for Cell and Molecular Biology, Cornell University, Ithaca, New York 14853, United States; Email: jcf14@cornell.edu Hening Lin − Department of Chemistry and Chemical Biology and Howard Hughes Medical Institute; Department of Chemistry and Chemical Biology, Cornell University, Ithaca, New York 14853, United States; orcid.org/0000-0002- 0255-2701; Email: hl379@cornell.edu Authors Ananya Anmangandla − Department of Chemistry and Chemical Biology, Cornell University, Ithaca, New York 14853, United States; orcid.org/0000-0002-2999-4067 Sadhan Jana − Department of Chemistry and Chemical Biology, Cornell University, Ithaca, New York 14853, United States Kewen Peng − Department of Chemistry and Chemical Biology, Cornell University, Ithaca, New York 14853, United States Shamar D. Wallace − Department of Molecular Biology and Genetics, Weill Institute for Cell and Molecular Biology, Cornell University, Ithaca, New York 14853, United States Saket R. Bagde − Department of Molecular Biology and Genetics, Weill Institute for Cell and Molecular Biology, Cornell University, Ithaca, New York 14853, United States; orcid.org/0000-0001-9800-9326 Bryon S. Drown − Department of Chemistry, Institute for Genomic Biology, and Cancer Center at Illinois, University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, United States; Present Address: Current address: Department of Chemistry, Purdue University, West Lafayette, Indiana 47906, United States Jiashu Xu − Department of Chemistry and Chemical Biology, Cornell University, Ithaca, New York 14853, United States Paul J. Hergenrother − Department of Chemistry, Institute for Genomic Biology, and Cancer Center at Illinois, University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, United States Complete contact information is available at: https://pubs.acs.org/10.1021/acschembio.3c00092 Author Contributions #Equal contribution. Funding This work is supported in part by NIH/NIAMS grant R01AR078555 and NIH/NIGMS training grants T32GM008500 and T32GM138826. J.C.F., S.D.W., and S.R.B. were supported by NIH/NIGMS grant R35GM136258 to J.C.F. We are grateful for the assistance of David Neau at NE- CAT. This work is based upon research conducted at the NE- CAT beamlines, which are funded by the National Institute of General Medical Sciences from the National Institutes of Health (P30 GM124165). The Eiger 16 M detector on the 24-ID-E funded by a NIH-ORIP HEI grant beam line is (S10OD021527). This research used resources of the APS, a U.S. Department of Energy (DOE) Office of Science User Facility operated for the DOE Office of Science by Argonne National Laboratory under Contract No. DE-AC02- 06CH11357. 1205 https://doi.org/10.1021/acschembio.3c00092 ACS Chem. Biol. 2023, 18, 1200−1207 ACS Chemical Biology pubs.acs.org/acschemicalbiology Articles Notes The authors declare the following competing financial for Sedec interest(s): H.L. Therapeutics. is a founder and consultant ■ REFERENCES (1) World Health Orgnization WHO Coronavirus (COVID-19) Dashboard (https://covid19.who.int/, accessed on March 29, 2023). (2) Ivashkiv, L. B.; Donlin, L. T. Regulation of type I interferon responses. Nat Rev Immunol 2014, 14, 36−49. (3) Fehr, A. R.; Singh, S. A.; Kerr, C. M.; Mukai, S.; Higashi, H.; Aikawa, M. The impact of PARPs and ADP-ribosylation on inflammation and host-pathogen interactions. Genes Dev. 2020, 34, 341−359. (4) Russo, L. C.; Tomasin, R.; Matos, I. A.; Manucci, A. C.; Sowa, S. T.; Dale, K.; Caldecott, K. W.; Lehtio, L.; Schechtman, D.; Meotti, F. C.; Bruni-Cardoso, A.; Hoch, N. C. The SARS-CoV-2 Nsp3 macrodomain reverses PARP9/DTX3L-dependent ADP-ribosylation induced by interferon signaling. J. Biol. Chem. 2021, 297, No. 101041. (5) Li, C.; Debing, Y.; Jankevicius, G.; Neyts, J.; Ahel, I.; Coutard, B.; Canard, B. Viral Macro Domains Reverse Protein ADP-Ribosylation. J. Virol. 2016, 90, 8478−8486. (6) Malet, H.; Coutard, B.; Jamal, S.; Dutartre, H.; Papageorgiou, N.; Neuvonen, M.; Ahola, T.; Forrester, N.; Gould, E. A.; Lafitte, D.; Ferron, F.; Lescar, J.; Gorbalenya, A. E.; de Lamballerie, X.; Canard, B. The crystal structures of Chikungunya and Venezuelan equine encephalitis virus nsP3 macro domains define a conserved adenosine binding pocket. J. Virol. 2009, 83, 6534−6545. (7) Rack, J. G. M.; Perina, D.; Ahel, I. Macrodomains: Structure, Function, Evolution, and Catalytic Activities. Annu. Rev. Biochem. 2016, 85, 431−454. (8) Dasovich, M.; Zhuo, J.; Goodman, J. A.; Thomas, A.; McPherson, R. L.; Jayabalan, A. K.; Busa, V. F.; Cheng, S. J.; Murphy, B. A.; Redinger, K. R.; Alhammad, Y. M. O.; Fehr, A. R.; Tsukamoto, T.; Slusher, B. S.; Bosch, J.; Wei, H.; Leung, A. K. L. High-Throughput Activity Assay for Screening Inhibitors of the SARS-CoV-2 Mac1 Macrodomain. ACS Chem. Biol. 2022, 17, 17−23. (9) Schuller, M.; Correy, G. J.; Gahbauer, S.; Fearon, D.; Wu, T.; Diaz, R. E.; Young, I. D.; Carvalho Martins, L.; Smith, D. H.; Schulze- Gahmen, U.; Owens, T. W.; Deshpande, I.; Merz, G. E.; Thwin, A. C.; Biel, J. T.; Peters, J. K.; Moritz, M.; Herrera, N.; Kratochvil, H. T.; Consortium, Q. S. B.; Aimon, A.; Bennett, J. M.; Brandao Neto, J.; Cohen, A. E.; Dias, A.; Douangamath, A.; Dunnett, L.; Fedorov, O.; Ferla, M. P.; Fuchs, M. R.; Gorrie-Stone, T. J.; Holton, J. M.; Johnson, M. G.; Krojer, T.; Meigs, G.; Powell, A. J.; Rack, J. G. M.; Rangel, V. L.; Russi, S.; Skyner, R. E.; Smith, C. A.; Soares, A. S.; Wierman, J. L.; Zhu, K.; O’Brien, P.; Jura, N.; Ashworth, A.; Irwin, J. J.; Thompson, M. C.; Gestwicki, J. E.; von Delft, F.; Shoichet, B. K.; Fraser, J. S.; Ahel, I. Fragment binding to the Nsp3 macrodomain of SARS-CoV-2 identified through crystallographic screening and computational docking. Sci. Adv. 2021, 7, eabf8711. (10) Stowell, A. I. J.; James, D. I.; Waddell, I. D.; Bennett, N.; Truman, C.; Hardern, I. M.; Ogilvie, D. J. A high-throughput screening- compatible homogeneous time-resolved fluorescence assay measuring the glycohydrolase activity of human poly(ADP-ribose) glycohydro- lase. Anal. Biochem. 2016, 503, 58−64. (11) Kim, I.-K.; Stegeman, R. A.; Brosey, C. A.; Ellenberger, T. A Quantitative Assay Reveals Ligand Specificity of the DNA Scaffold Repair Protein XRCC1 and Efficient Disassembly of Complexes of XRCC1 and the Poly(ADP-ribose) Polymerase 1 by Poly(ADP-ribose) Glycohydrolase*. J. Biol. Chem. 2015, 290, 3775−3783. (12) Lea, W. A.; Simeonov, A. Fluorescence polarization assays in small molecule screening. Expert Opin. Drug Discovery 2011, 6, 17−32. (13) Roy, A.; Alhammad, Y. M.; McDonald, P.; Johnson, D. K.; Zhuo, J.; Wazir, S.; Ferraris, D.; Lehtio, L.; Leung, A. K. L.; Fehr, A. R. Discovery of compounds that inhibit SARS-CoV-2 Mac1-ADP-ribose binding by high-throughput screening. Antiviral Res. 2022, 203, No. 105344. (14) Yang, X.; Ma, Y.; Li, Y.; Dong, Y.; Yu, L. L.; Wang, H.; Guo, L.; Wu, C.; Yu, X.; Liu, X. Molecular basis for the MacroD1-mediated hydrolysis of ADP-ribosylation. DNA Repair (Amst) 2020, 94, No. 102899. (15) Neuvonen, M.; Ahola, T. Differential activities of cellular and viral macro domain proteins in binding of ADP-ribose metabolites. J. Mol. Biol. 2009, 385, 212−225. (16) Brosey, C. A.; Houl, J. H.; Katsonis, P.; Balapiti-Modarage, L. P. F.; Bommagani, S.; Arvai, A.; Moiani, D.; Bacolla, A.; Link, T.; Warden, L. S.; Lichtarge, O.; Jones, D. E.; Ahmed, Z.; Tainer, J. A. Targeting SARS-CoV-2 Nsp3 macrodomain structure with insights from human poly(ADP-ribose) glycohydrolase (PARG) structures with inhibitors. Prog Biophys Mol Biol 2021, 163, 171−186. (17) Wang, Z.; Xu, W., Biochemical and Biophysical Assays of PAR- WWE Domain Interactions and Production of iso-ADPr for PAR- Binding Analysis. In ADP-ribosylation and NAD+ Utilizing Enzymes: Methods and Protocols, Chang, P., Ed. Springer New York: New York, NY, 2018; pp. 65−73, DOI: 10.1007/978-1-4939-8588-3_5. (18) Wang, Z.; Michaud, G. A.; Cheng, Z.; Zhang, Y.; Hinds, T. R.; Fan, E.; Cong, F.; Xu, W. Recognition of the iso-ADP-ribose moiety in poly(ADP-ribose) by WWE domains suggests a general mechanism for poly(ADP-ribosyl)ation-dependent ubiquitination. Genes Dev. 2012, 26, 235−240. (19) Gahbauer, S.; Correy, G. J.; Schuller, M.; Ferla, M. P.; Doruk, Y. U.; Rachman, M.; Wu, T.; Diolaiti, M.; Wang, S.; Neitz, R. J.; Fearon, D.; Radchenko, D.; Moroz, Y.; Irwin, J. J.; Renslo, A. R.; Taylor, J. C.; Gestwicki, J. E.; von Delft, F.; Ashworth, A.; Ahel, I.; Shoichet, B. K.; Fraser, J. S. Iterative computational design and crystallographic screening identifies potent inhibitors targeting the Nsp3 Macrodomain of SARS-CoV-2. bioRxiv 2022. (20) Siegel, D.; Hui, H. C.; Doerffler, E.; Clarke, M. O.; Chun, K.; Zhang, L.; Neville, S.; Carra, E.; Lew, W.; Ross, B.; Wang, Q.; Wolfe, L.; Jordan, R.; Soloveva, V.; Knox, J.; Perry, J.; Perron, M.; Stray, K. M.; Barauskas, O.; Feng, J. Y.; Xu, Y.; Lee, G.; Rheingold, A. L.; Ray, A. S.; Bannister, R.; Strickley, R.; Swaminathan, S.; Lee, W. A.; Bavari, S.; Cihlar, T.; Lo, M. K.; Warren, T. K.; Mackman, R. L. Discovery and Synthesis of a Phosphoramidate Prodrug of a Pyrrolo[2,1-f][triazin-4- amino] Adenine C-Nucleoside (GS-5734) for the Treatment of Ebola and Emerging Viruses. J. Med. Chem. 2017, 60, 1648−1661. (21) Beigel, J. H.; Tomashek, K. M.; Dodd, L. E.; Mehta, A. K.; Zingman, B. S.; Kalil, A. C.; Hohmann, E.; Chu, H. Y.; Luetkemeyer, A.; Kline, S.; Lopez de Castilla, D.; Finberg, R. W.; Dierberg, K.; Tapson, V.; Hsieh, L.; Patterson, T. F.; Paredes, R.; Sweeney, D. A.; Short, W. R.; Touloumi, G.; Lye, D. C.; Ohmagari, N.; Oh, M. D.; Ruiz-Palacios, G. M.; Benfield, T.; Fatkenheuer, G.; Kortepeter, M. G.; Atmar, R. L.; Creech, C. B.; Lundgren, J.; Babiker, A. G.; Pett, S.; Neaton, J. D.; Burgess, T. H.; Bonnett, T.; Green, M.; Makowski, M.; Osinusi, A.; Nayak, S.; Lane, H. C.; Members, A.-S. G. Remdesivir for the Treatment of Covid-19 - Final Report. N Engl J Med 2020, 383, 1813− 1826. (22) Ni, X.; Schroder, M.; Olieric, V.; Sharpe, M. E.; Hernandez- Olmos, V.; Proschak, E.; Merk, D.; Knapp, S.; Chaikuad, A. Structural Insights into Plasticity and Discovery of Remdesivir Metabolite GS- 441524 Binding in SARS-CoV-2 Macrodomain. ACS Med. Chem. Lett. 2021, 12, 603−609. (23) Tsika, A. C.; Gallo, A.; Fourkiotis, N. K.; Argyriou, A. I.; Sreeramulu, S.; Lohr, F.; Rogov, V. V.; Richter, C.; Linhard, V.; Gande, S. L.; Altincekic, N.; Krishnathas, R.; Elamri, I.; Schwalbe, H.; Wollenhaupt, J.; Weiss, M. S.; Spyroulias, G. A. Binding Adaptation of GS-441524 Diversifies Macro Domains and Downregulates SARS- CoV-2 de-MARylation Capacity. J. Mol. Biol. 2022, 434, No. 167720. (24) Correy, G. J.; Kneller, D. W.; Phillips, G.; Pant, S.; Russi, S.; Cohen, A. E.; Meigs, G.; Holton, J. M.; Gahbauer, S.; Thompson, M. C.; Ashworth, A.; Coates, L.; Kovalevsky, A.; Meilleur, F.; Fraser, J. S. The mechanisms of catalysis and ligand binding for the SARS-CoV-2 NSP3 macrodomain from neutron and x-ray diffraction at room temperature. Sci. Adv. 2022, 8, eabo5083. (25) Drown, B. S.; Shirai, T.; Rack, J. G. M.; Ahel, I.; Hergenrother, P. J. Monitoring Poly(ADP-ribosyl)glycohydrolase Activity with a 1206 https://doi.org/10.1021/acschembio.3c00092 ACS Chem. Biol. 2023, 18, 1200−1207 ACS Chemical Biology pubs.acs.org/acschemicalbiology Articles Continuous Fluorescent Substrate. Cell. Chem. Biol. 2018, 25, 1562− 1570e19. (26) Kabsch, W. Xds. Acta Crystallogr D Biol Crystallogr 2010, 66, 125−132. (27) McCoy, A. J.; Grosse-Kunstleve, R. W.; Adams, P. D.; Winn, M. D.; Storoni, L. C.; Read, R. J. Phaser crystallographic software. J. Appl. Crystallogr. 2007, 40, 658−674. (28) Liebschner, D.; Afonine, P. V.; Baker, M. L.; Bunkoczi, G.; Chen, V. B.; Croll, T. I.; Hintze, B.; Hung, L. W.; Jain, S.; McCoy, A. J.; Moriarty, N. W.; Oeffner, R. D.; Poon, B. K.; Prisant, M. G.; Read, R. J.; Richardson, J. S.; Richardson, D. C.; Sammito, M. D.; Sobolev, O. V.; Stockwell, D. H.; Terwilliger, T. C.; Urzhumtsev, A. G.; Videau, L. L.; Williams, C. J.; Adams, P. D. Macromolecular structure determination using X-rays, neutrons and electrons: recent developments in Phenix. Acta Crystallogr D Struct Biol 2019, 75, 861−877. (29) Emsley, P.; Lohkamp, B.; Scott, W. G.; Cowtan, K. Features and development of Coot. Acta Crystallogr., Sect. D: Biol. Crystallogr. 2010, 66, 486−501. (30) Williams, C. J.; Headd, J. J.; Moriarty, N. W.; Prisant, M. G.; Videau, L. L.; Deis, L. N.; Verma, V.; Keedy, D. A.; Hintze, B. J.; Chen, V. B.; Jain, S.; Lewis, S. M.; Arendall, W. B., 3rd; Snoeyink, J.; Adams, P. D.; Lovell, S. C.; Richardson, J. S.; Richardson, D. C. MolProbity: More and better reference data for improved all-atom structure validation. Protein Sci. 2018, 27, 293−315. 1207 https://doi.org/10.1021/acschembio.3c00092 ACS Chem. Biol. 2023, 18, 1200−1207
10.7554_elife.85492
RESEaRCH aRTICLE On the limits of fitting complex models of population history to f- statistics Robert Maier1*†, Pavel Flegontov1,2*†, Olga Flegontova2, Ulaş Işıldak2, Piya Changmai2, David Reich1,3,4,5* 1Department of Human Evolutionary Biology, Harvard University, Cambridge, United States; 2Department of Biology and Ecology, Faculty of Science, University of Ostrava, Ostrava, Czech Republic; 3Broad Institute of Harvard and MIT, Cambridge, United States; 4Howard Hughes Medical Institute, Harvard Medical School, Boston, United States; 5Department of Genetics, Harvard Medical School, Boston, United States Abstract Our understanding of population history in deep time has been assisted by fitting admixture graphs (AGs) to data: models that specify the ordering of population splits and mixtures, which along with the amount of genetic drift and the proportions of mixture, is the only informa- tion needed to predict the patterns of allele frequency correlation among populations. The space of possible AGs relating populations is vast, and thus most published studies have identified fitting AGs through a manual process driven by prior hypotheses, leaving the majority of alternative models unexplored. Here, we develop a method for systematically searching the space of all AGs that can incorporate non- genetic information in the form of topology constraints. We implement this findGraphs tool within a software package, ADMIXTOOLS 2, which is a reimplementation of the ADMIXTOOLS software with new features and large performance gains. We apply this method- ology to identify alternative models to AGs that played key roles in eight publications and find that in nearly all cases many alternative models fit nominally or significantly better than the published one. Our results suggest that strong claims about population history from AGs should only be made when all well- fitting and temporally plausible models share common topological features. Our re- evaluation of published data also provides insight into the population histories of humans, dogs, and horses, identifying features that are stable across the models we explored, as well as scenarios of populations relationships that differ in important ways from models that have been highlighted in the literature. Editor's evaluation This is a rigorous and critical analysis of the performance of a popular suite of methods for inferring population history, accompanied by improvements. Should be of broad interest to anyone inter- ested in human history. *For correspondence: robertmaier@gmx.net (RM); Pavel_Flegontov@hms.harvard. edu (PF); reich@genetics.med.harvard. edu (DR) †These authors contributed equally to this work Competing interest: The authors declare that no competing interests exist. Funding: See page 25 Preprinted: 08 May 2022 Received: 10 December 2022 Accepted: 05 April 2023 Published: 29 June 2023 Reviewing Editor: Magnus Nordborg, Gregor Mendel Institute, Austria Copyright Maier, Flegontov et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Introduction Admixture graph models provide a powerful intellectual framework for describing the relationships among populations that allows not only branching of populations from a common ancestor but also mixture events. An admixture graph (abbreviated below as AG), as fit in the widely used software packages ADMIXTOOLS (Patterson et al., 2012) and TreeMix (Pickrell and Pritchard, 2012; Molloy et  al., 2021), is a directed acyclic bifurcating graph with two types of edges: those representing genetic drift, and those representing gene flow. Each admixture event is represented as a confluence Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 1 of 62 Research article of two gene flow edges. Nodes of such a graph represent unsampled intermediate populations, and terminal nodes (leaves) represent sampled present- day or ancient groups (see a mathematical defini- tion in Soraggi and Wiuf, 2019). An attractive feature of AGs is that they can summarize important features of population history without requiring specification of all parameters such as population sizes, split times, mixture times, and distinguishing between sudden splits or drawn- out separations. All these parameters describe important features of demographic history and are fit by many methods for fitting demographic models (Gutenkunst et al., 2009; Gronau et al., 2011; Schiffels et al., 2016; Flegontov et al., 2019; Kamm et al., 2020, Rogers, 2019; Hubisz et al., 2020). However, the fact that it is possible to factor this difficult problem by first inferring important aspects of the topology (AGs fitted to allele frequency correlation statistics), and then fitting additional demographic parame- ters to data such as site frequency spectra, simplifies demographic inference (Patterson et al., 2012; Pickrell and Pritchard, 2012; Lipson et al., 2013; Leppälä et al., 2017; Lipson et al., 2020b; Molloy et al., 2021; Yan et al., 2021). AGs thus serve both as conceptual frameworks that allow us to think about the relationships of populations deep in time, and as mathematical models we can fit to genetic data. ∗ AGs are fitted to f- statistics (Reich et al., 2009; Patterson et al., 2012; Peter, 2016; Soraggi and Wiuf, 2019). For convenience, below we use a concise definition of f- statistics by Lipson, 2020a: ‘The most general definition is that of the f4- statistic f4(A, B; C, D), which measures the average correla- tion in allele frequency differences between (1) populations A and B and (2) populations C and D that is, (pA – pB) (pC – pD), for allele frequencies p, typically averaged over many biallelic single- nucleotide polymorphisms. This f4- statistic is the same as the D- statistic up to a normalization factor.’ The other f- statistics (f2 and f3) can be defined as special cases of f4- statistics: f2(A, B) = f4(A, B; A, B) and f3(A; B, C) = f4(A, B; A, C). f4- Statistics can be written as linear combinations of f3- or f2- statistics, and f3- statistics can be written as linear combinations of f4- and f2- statistics. f2-, f3-, and f4- statistics have straightforward interpretations in terms of drift edges along the tree, see Figure 2 in Patterson et al., 2012 and Appendix 1—figure 1b. A challenge for fitting AG models is that they are often not uniquely constrained by the data, with many providing equally good fits to the f2-, f3-, and f4- statistics used to constrain them within the limits of statistical resolution. Previously published methods for finding fitting AGs (mainly qpGraph, Patterson et  al., 2012 and TreeMix, Pickrell and Pritchard, 2012; Molloy et al., 2021) were not well equipped to handle the large range of equally well- fitting models for three reasons: (1) They did not reliably provide information on whether there is a uniquely fitting parsimonious model or alternatively whether there are many models that fit equally well to the limits of statistical resolution, (2) they did not provide formal goodness- of- fit tests, and related to this, (3) they did not provide tests for whether the difference between the fits of any two models is statistically significant. As a consequence, and as we demonstrate in what follows, many published AG models have been interpreted as providing more confidence than is merited about the extent to which genetic data allows us to disentangle ancestral relationships. To appreciate these problems, we first need to consider two main approaches that were utilized to study demographic history with AGs. The first approach is to identify AGs automatically, either without human intervention or with guid- ance. It is possible in theory to exhaustively test all possible graphs for a given set of populations and pre- specified number of admixture events, as implemented, for example, in the admixturegraph R package (Leppälä et al., 2017). An exhaustive approach can provide a complete view of the range of models that are consistent with the data for a specified level of parsimony (total number of admix- ture events allowed in the graph), which is not biased by the algorithm used to explore the space of possible AGs. However, this approach is limited to small graphs (typically up to six groups, two admix- ture events) due to the rapid increase in the number of possible AGs as the number of populations and admixture events grows. As we show in our discussion of case studies, the simple models explored with an exhaustive approach can lead to misleading conclusions about population history because not including additional populations can blind users to additional mixture events that occurred (and whose existence is revealed by examining data from additional populations). Furthermore, models with additional admixture events that are qualitatively different to the best- fitting parsimonious graph and that capture the true history, will sometimes be completely missed when constraining the number of gene flows. Alternatively, the programs TreeMix (Pickrell and Pritchard, 2012; Molloy et  al., 2021), MixMapper (Lipson et al., 2013), miqoGraph (Yan et al., 2021), and AdmixtureBayes (Nielsen Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 2 of 62 Evolutionary Biology | Genetics and Genomics Research article et al., 2023 preprint) all address the problem of how to rapidly explore the vast space of AGs relating a set of populations by applying algorithmic ideas or heuristics; all of these methods speed up model search by orders of magnitude. The second approach to fitting AGs is to manually build them up by grafting additional populations onto simpler smaller graphs that fit the data. This approach involves stepwise addition of populations in an order that is chosen based on the best judgment of the user, and for each newly added population involves adding admixture events or tweaks in the graph until a fit is obtained; the user then moves on to adding the next population (see Reich et al., 2009; Reich et al., 2011; Reich et al., 2012; Lazaridis et al., 2014; Seguin- Orlando et al., 2014; Fu et al., 2016; Skoglund et al., 2016; Yang et al., 2017; McColl et al., 2018; Moreno- Mayar et al., 2018; Tambets et al., 2018; van de Loosdrecht et al., 2018; Flegontov et al., 2019; Sikora et al., 2019; Wang et al., 2019; Lipson et al., 2020b; Shinde et al., 2019; Yang et al., 2020; Hajdinjak et al., 2021; Wang et al., 2021; Bergström et al., 2022 for examples). The program qpGraph in the ADMIXTOOLS package (Patterson et al., 2012) has been the most common computational method used for testing fits of individual AGs. Most AGs in the liter- ature have been constructed manually in this way, often acknowledging the existence of alternative models by presenting plausible models side- by- side, and this approach has been the basis for many claims about population history (Lazaridis et al., 2014; Yang et al., 2017; Posth et al., 2018; Sikora et al., 2019; Shinde et al., 2019; Bergström et al., 2020; Lipson et al., 2020b; Hajdinjak et al., 2021; Wang et al., 2021; Bergström et al., 2022). A strength of this approach is that it takes advan- tage of human judgment and outside knowledge about what graphs best fit the history of the human or animal populations being analyzed. This external information is powerful as it can incorporate non- genetic evidence such as geographic plausibility and temporal ordering of populations or linguistic similarity, or other genetic data such as estimates of population split times, or shared Y chromosomes, or rejection of proposed scenarios based on joint analysis of much larger numbers of populations than can reasonably be analyzed within a single AG. Thus, while manual approaches explore many orders of magnitude fewer topologies than automatic approaches often do, they still may provide inferences about population history that are more useful than those provided by automatic approaches. These methods’ strength is also their weakness: by relying on intuition, following a manual approach has the potential to validate the biases users have as to what types of histories are most plausible (these may be the only types of histories that will be carefully explored). This can blind users to surprises: to profoundly different topologies that may correspond more closely to the true history, and we discuss examples of this in the Results section. In this study, we introduce a new method, findGraphs, that belongs to the first class of algorithms (those for automated AG topology inference). Algorithmic innovations and speedups in findGraphs enable us to explore a much larger proportion of plausible AG space than many other methods reported to date. The findGraphs method combines the advantages of automated and manual topology exploration by allowing users to encode various sources of information as constraints on the space of AGs, which is then explored automatically. However, the main innovations in findGraphs are not computational, but instead conceptual. Instead of finding one or a few AGs fitting the data well, we use findGraphs for exploring AG spaces and assessing if any reliable information on popu- lation history can be extracted from a given AG space (defined by a population set and parsimony constraints) in the first place. Results Regardless of the approach used to search through the space of possibly fitting AGs, a challenge in the effort to find a uniquely well- fitting AG (or group of topologically similar AGs) is that it has been difficult to quantify the absolute goodness of fit of a model to date. We have not been entirely successful with this and are not aware of other work that has been successful. It is also difficult to assess the relative fits of multiple models, especially if they differ in complexity. Performance gains relative to the original implementation of qpGraph allow us to address this problem by obtaining bootstrap confidence intervals and p- values for estimated parameters of single models, as well as for the difference in fit quality of two models (see Appendix 1, Sections 1.B.3 and 2.E). In combination with the approach to automating the search of well- fitting AGs, this leads to a situation where we are able to find and test a large number of models, many of which fit equally well despite often having very different topological features. Published approaches to comparing the fits of AG models based Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 3 of 62 Evolutionary Biology | Genetics and Genomics Research article on Akaike information criterion (AIC) or Bayesian information criterion (BIC), see Flegontov et  al., 2019; Shinde et al., 2019 have the problem that it is often not clear what the effective number of degrees of freedom is in the two models being compared since in the case of AGs it depends not only on the number of graph edges, but also on graph topology. The methods for automated graph topology inference and model comparison relying on bootstrap resampling are implemented in ADMIXTOOLS 2, a comprehensive platform for learning about popu- lation history from f- statistics. It is built to provide a stand- alone workspace for research in this area and is implemented as an R package. For all computations, ADMIXTOOLS 2 exhibits large speedups rela- tive to previously published platforms for f- statistic analysis (e.g., popstats and ADMIXTOOLS version 6.0 which we call ‘Classic ADMIXTOOLS’ in what follows to distinguish it from updated ADMIXTOOLS version 7.0.2 which implements some of the speedup ideas also implemented in ADMIXTOOLS 2). This is achieved by deploying a series of algorithmic improvements, most notably storage of precom- puted f- statistics in random access memory, which avoids having to rely on reading in extremely large genotype matrices to perform most computations. In addition to the new algorithmic ideas allowing efficient searching through the space of AGs and comparing the fits of two AGs, ADMIXTOOLS 2 also provides a solution to the question of which parameters of an AG are identifiable in the limit of infinite data. Methodological details are presented in Appendix 1, and below we focus on documenting problems of AG inference on simulated data and revisiting AGs from the literature to understand the extent to which methodological challenges with AG fitting biased previous studies. Topological diversity of well-fitting models and effects of parsimony constraints on simulated data First, we explored the performance of the findGraphs method for automated topology inference on simulated AGs of random topology, focusing on the following questions: (1) among findGraphs results, how common are AGs fitting nominally or significantly better than the true one but different topo- logically; (2) what is the degree of topological diversity among these models fitting the data better than the true one? For this purpose, we simulated AGs of four complexity classes using msprime v.1.1.1: eight or nine non- outgroup populations, and four or five admixture events. Only simulations where pairwise FST for groups were in the range characteristic for anatomically modern and archaic humans were selected for further analysis, resulting in 20 random topologies per complexity class, each including a distant outgroup that facilitates automated exploration of the topology space. We ran findGraphs on each simulated dataset starting from random graphs and pre- specifying the true number of admixture events (n), or n − 1, or n + 1 events. For each of these graph complexity levels, we performed 100 independent findGraphs runs and recorded 5 AGs from each run having the best log- likelihood (LL) scores. Topologically redundant AGs were discarded, and for the remaining AGs we calculated worst f4- statistic residuals (WR) and tested if the newly found models fit significantly better than the true model, using the bootstrap model comparison method developed in this study (see Appendix 1, Sections 1.B.3 and 2.E). In Figure 1a–c, we show the following statistics for each simulated AG, summarized across simulated complexity classes and parsimony levels allowed at the stage of topology exploration: fraction of topologies found with findGraphs that fit better than the true AG (according to LL score), or that fit significantly better than the true AG, or those with plausible absolute fits (WR < 3 SE). It is clear that for the great majority of simulated datasets, even a shallow exploration of the topology space with findGraph (100 independent runs) uncovers AGs that fit nomi- nally better than the true topology (Figure 1a and d) and are topologically diverse (see Figure 1e for examples). When allowing for n admixture events, at least one AG fitting significantly better than the true one was found for 60% of simulated datasets. When n + 1 admixture events were allowed, this grew to 100% (all 80 datasets). It should be noted that some admixture events are indistinguish- able with f- statistics; for instance, successive gene flows between two lineages, with no other edges branching off between the gene flows. If such gene flows were included in the random topologies we simulated, AGs with n events were overly complex for representing the true history. Thus, if we are dealing with random histories, choosing an optimal complexity class for topology search is not straightforward. These results on simulated data raise concerns about the extent to which fitting AG topologies provide reliable information about population history. Even for histories including eight or nine groups, an outgroup, and four or five pulse- like admixture events, perfect diploid data, and groups as Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 4 of 62 Evolutionary Biology | Genetics and Genomics Research article Figure 1. Computer simulations show that when the true admixture graph (AG) topology is complex, findGraphs frequently finds AGs fitting the data better than the true AG. (a) Fractions of distinct AGs found with findGraphs that fit the data nominally better than the true AG (according to log- likelihood [LL] scores). The simulated datasets are grouped by complexity class (eight or nine leaves, four or five admixture events) and by the number of admixture events allowed at the topology search stage (n − 1 on the left, n in the middle, and n + 1 on the right, where n is the true number of simulated admixture events). Each dot represents a simulated random history, and 20 such histories were simulated for each complexity class. (b) Fractions of distinct AGs found with findGraphs that fit the data significantly better than the true AG (two- tailed empirical p- value of the bootstrap model comparison method <0.05). (c) Fractions of distinct AGs found with findGraphs that fit the data well in absolute terms (WR < 3 SE). (d) Distinct AGs found for a particular simulated history (eight groups and four admixture events) in the LL and WR coordinates. Only best- fitting graphs with WR < 3 SE are shown. The fit of the true topology is shown in yellow, and topologies that fit the data significantly better than the true one are in purple. The true topology was not recovered by our findGraphs searches. (e) The true model from panel (d) and two alternative models found with findGraphs, both fitting significantly better than the true one (based on the bootstrap p- value) and very different topologically. This is presented as an example of very high topological diversity seen among well- fitting models. Model parameters (graph edges) that were inferred to be unidentifiable (see Appendix 1, Section 2.F) are plotted in red. differentiated as Neanderthals and anatomically modern humans—a complexity class that is simpler than many models fitted to real genetic data in published papers—models fitting the data as well as or better than the true one are common, and their topological diversity is in most cases so high that it precludes consensus inference of topology by analysis of multiple topologies. As we demonstrate on another set of simulated data in Appendix 1 (Section 1.B.1), the probability of finding a ‘wrong’ model that fits better than the true one grows with increasing graph complexity, and that effect is reproduced with both findGraphs and TreeMix (Appendix 1—figure 2). We expect this problem to be even more acute when researchers are dealing with realistic complex histories. However, geneticists often rely on external constraints on AG topologies (such as temporal plausibility of a topology, results from qpAdm modelling, f4- statistics, PCA, ADMIXTURE, geographical, archaeological, and linguistic considerations) that were not used for filtering the results of our topology searches on simulated data. Thus, it is possible in principle that published AG models are more robust than our results on simu- lated data, and we explore this issue in depth in the next section and in Appendix 2. Revisiting published AGs We studied AGs from eight publications (Lazaridis et al., 2014; Shinde et al., 2019; Sikora et al., 2019; Bergström et  al., 2020; Lipson et  al., 2020b; Hajdinjak et  al., 2021Librado et  al., 2021; Wang et al., 2021) with the goal of comparing published models to models identified by our algo- rithm for automatically inferring optimal (best- fitting) AGs (Table 1). In all but one study, qpGraph or its automated reimplementation (admixturegraph, Leppälä et al., 2017) was used for fitting topologies to genetic data, while Librado et al., 2021 relied on the automated OrientAGraph method (Molloy Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 5 of 62 Evolutionary Biology | Genetics and GenomicsNum. of adm. events:4Num. of adm. events: 5Num. of leaves: 8Num. of leaves: 9n−1nn+1n−1nn+10.000.250.500.751.000.000.250.500.751.00Proportion of newly found AG thatfit better than the true onea.Num. of adm. events: 4Num. of adm. events: 5Num. of leaves: 8Num. of leaves: 9n−1nn+1n−1nn+10.000.250.500.751.000.000.250.500.751.00Proportion of newly found AG thatfit significantly better than the true oneb.Num. of adm. events: 4Num. of adm. events: 5Num. of leaves: 8Num. of leaves: 9n−1nn+1n−1nn+10.000.250.500.751.000.000.250.500.751.00Proportion of newly found AGwith WR < 3c.Num. of adm. eventsn−1nn+1n−1nn+15101551015510151.01.52.02.53.0LLWRModelTrue AGNewly found AGNewly found AG(significant)d.e.NewlyfoundAGLLscore=2.78p-value=0.008OUT51%C12%88%93%D49%52%G7%E48%HABFNewlyfoundAGLLscore=7.57p-value=0.00218%82%52%40%48%60%F91%9%GOUTCEDHABTrueAGLLscore=7.9375%88%25%12%28%C72%88%12%FGOUTDEHBA Research article % l i , s e g o o p o t g n i t t fi e s r o w y l t n a c fi n g S i i i y l t n a c fi n g i s - n o N i y l t n a c fi n g i s - n o N y l t n a c fi n g S i i g n i t t fi e s r o w g n i t t fi r e t t e b g n i t t fi r e t t e b t c n i t s i D e v i t a n r e t l a : l e d o m . l b u P i s e g o o p o t l , l a u d i s e r t s r o w e r u t x m d A i s p u o r G % i , s e g o o p o t l % i , s e g o o p o t l % i , s e g o o p o t l d n u o f E S d e s u s P N S s t n e v e ) s n o i t a u p o p l ( 5 . 0 8 2 . 6 7 . 3 9 4 . 3 5 5 . 5 9 3 . 1 7 0 . 2 2 4 . 0 1 0 . 0 0 . 8 4 2 . 9 8 7 . 6 1 7 . 0 8 5 . 3 1 . 4 2 5 . 4 4 . 8 2 6 . 6 1 . 7 7 1 . 3 6 . 4 3 8 . 9 3 . 2 1 . 2 1 8 . 2 7 . 5 1 0 . 0 3 . 0 7 . 5 5 9 . 1 1 3 . 4 8 1 . 7 1 9 . 0 5 0 . 0 1 . 0 0 . 8 6 . 0 0 . 0 0 . 1 2 2 6 0 3 3 4 1 4 2 3 5 3 5 4 8 7 . 7 5 1 8 8 9 1 0 0 . 0 0 0 2 . 6 2 1 3 0 . 1 0 . 8 7 7 1 4 9 8 5 8 7 2 1 2 . 2 2 . 6 2 . . 9 3 2 . 1 4 1 9 6 . 8 4 . 3 2 . 8 3 . 8 3 . 2 4 . 2 8 2 2 1 3 , 7 4 2 2 4 6 , , 9 0 0 9 4 2 , , 9 1 4 7 6 7 1 , 8 9 6 3 6 2 , 3 4 3 3 4 5 8 8 3 7 1 1 2 , † 1 1 3 5 7 3 0 2 , 3 0 9 4 4 3 , 9 0 5 3 1 6 , 8 † 6 † 6 7 7 * 8 * 0 1 2 1 2 1 2 1 3 1 4 1 l i a n g i r o e h t n o i t a c i l b u p n i e r u g F i e 1 3 3 b 3 d 5 t x E e 5 t x E 4 t x E d 2 ) t f e l ( f 3 6 t x E ) t h g i r ( f 3 , . l a t e s i d i r a z a L 4 1 0 2 , . l a t e e d n h S i 9 1 0 2 n o i t a c i l b u P m ö r t s g r e B 0 2 0 2 , . l a t e , . l a t e o d a r b L i 1 2 0 2 , . l a t e i j k a n d a H j , . l a t e n o s p L i b 0 2 0 2 , . l a t e g n a W 1 2 0 2 1 2 0 2 , . l a t e a r o k S i 9 1 0 2 . ) s l i a t e d r o f 1 l e fi y r a t n e m e p p u S e e s ( l a t a d r a l i m i s y r e v r o e m a s e h t n o d e r r e f n i s h p a r g e v i t a n r e t l a o t s n o i t a c i l b u p t h g e m o r f i s h p a r g d e r a p m o c e W . s h p a r g d n u o f y l l a c i t a m o t u a f o t x e t n o c e h t n i s h p a r g d e h s i l b u P . 1 e b a T l , s P N S ( a t a d l i a n g i r o e h t d e t s e t e w i , s e d u t s e s a c l l a r o F . s G A e h t g n i t t fi r o f d e s u ) l e v e l p u o r g e h t t a a t a d g n i s s i m o n h t i w ; s P N S ( s m s i h p r o m y l o p e d i t o e c u n - e g n i s l l f o r e b m u n e h T : d e s u s P N S . d e t n e s e r p s i G A e h t e r e h w r e p a p l i a n g i r o e h t n i r e b m u n e r u g F i : n o i t a c i l b u p l i a n g i r o e h t n i e r u g F i . n o i t a c i l b u p t n a v e e r e h t l f o r a e y d n a r o h t u a t s r fi e h t f o e m a n t s a L : n o i t a c i l b u P . h p a r g h c a e n i s t n e v e e r u t x m d a i f o r e b m u n e h T : s t n e v e e r u t x m d a i . h p a r g h c a e n i s n o i t a u p o p l f o r e b m u n e h T : ) s n o i t a u p o p l ( s p u o r G e m o s n i e w , h c r a e s l y g o o p o t i t n e c fi f e f o e s o p r u p e h t r o f , r e v e w o H . s e n o d e h s i l b u p e h t o t r a l i m i s y r e v s t fi l e d o m d e n a t b o d n a i l ) y g o o p o t h p a r g d e h s i l b u p e h t d n a , n o i t i s o p m o c n o i t a u p o p l . t x e t e h t n i d e s s u c s i d d n a , l 1 e fi y r a t n e m e p p u S n l i , s e t o n t o o f e h t n i d e t o n s a l y t i x e p m o c h p a r g r o , n o i t i s o p m o c n o i t a u p o p l , l n o i t a u c a c l c i t s i t a t s - 3 f r o f s g n i t t e s d e t s u d a j s e s a c l g n o a d e t u b i r t s i d y l n e v e s l e d o m f o 3 / 1 o t 0 2 / 1 ( s l e d o m f l o e p m a s e v i t a t n e s e r p e r a , e g r a l y r e v i s a w s e g o o p o t l t c n i t s i d f o r e b m u n e h t f I . ) 5 0 . 0 < e u a v - p l l a c i r i p m e d e l i a t - o w t ( t s e t n o s i r a p m o c l e d o m p a r t s t o o b e h t o t g n d r o c c a h p a r g d e h s i l i b u p e h t n a h t r e t t e b y l t n a c fi n g i s i t fi t a h t i l s e g o o p o t e v i t a n r e t l a t c n i t s i d f o e g a t n e c r e p e h T : % i l , s e g o o p o t g n i t t fi r e t t e b y l t n a c fi n g S i i . ) E S ( s r o r r e d r a d n a t s n i d e r u s a e m , n m u o c l ’ d e s u s P N S ‘ e h t n i n w o h s t e s P N S e h t o t d e t t fi h p a r g d e h s i l b u p e h t f o l a u d i s e r c i t s i t a t s - f t s r o w e h T : E S , l a u d i s e r t s r o W : l e d o m . l b u P . e n o d e h s i l b u p e h t m o r f g n i r e f f i i l d s e g o o p o t d n u o f y l w e n t c n i t s i d f o r e b m u n e h T : d n u o f i l s e g o o p o t e v i t a n r e t l a t c n i t s i D . l e p m a s l l s i h t n o d e t a u c a c e r e w s n m u o c g n w o l i l l o f d n a s i h t n i s e g a t n e c r e p e h t d n a , d a e t s n i e n o d e h s i l b u p e h t o t d e r a p m o c s a w ) m u r t c e p s d o o h i l e k i l - g o l e h t l e d o m p a r t s t o o b e h t o t g n d r o c c a h p a r g d e h s i l i b u p e h t n a h t e s r o w ) y l l i a n m o n ( i y l t n a c fi n g i s - n o n t fi t a h t i s e g o o p o t l t c n i t s i d f o e g a t n e c r e p e h T : % i l , s e g o o p o t g n i t t fi e s r o w y l t n a c fi n g i s - n o N i . ) 5 0 . 0 ≥ e u a v - p l l a c i r i p m e d e l i a t - o w t ( t s e t n o s i r a p m o c t s e t n o s i r a p m o c l e d o m p a r t s t o o b e h t o t g n d r o c c a h p a r g d e h s i l i b u p e h t n a h t e s r o w y l t n a c fi n g i s i t fi t a h t i s e g o o p o t l t c n i t s i d f o e g a t n e c r e p e h T : % l i , s e g o o p o t g n i t t fi e s r o w y l t n a c fi n g S i i . t x e t e h t d n a 1 l e fi y r a t n e m e p p u S e e s l , y t i c i l p m i s r o f l e d o m d e h s i l b u p e h t m o r f d e v o m e r e r e w s w o fl e n e g n a t r e C i † . t x e t e h t d n a 1 l e fi y r a t n e m e p p u S e e s l , i d e fi d o m s a w n o i t i s o p m o c n o i t a u p o p e h T * l . ) 5 0 . 0 < e u a v - p l l a c i r i p m e d e l i a t - o w t ( i p a r t s t o o b e h t o t g n d r o c c a h p a r g d e h s i l b u p e h t n a h t r e t t e b ) y l l i a n m o n ( i y l t n a c fi n g i s - n o n t fi t a h t i s e g o o p o t l t c n i t s i d f o e g a t n e c r e p e h T : % i l , s e g o o p o t g n i t t fi r e t t e b y l t n a c fi n g i s - n o N i . ) . 5 0 0 ≥ e u a v - p l l a c i r i p m e d e l i a t - o w t ( t s e t n o s i r a p m o c l e d o m Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 6 of 62 Evolutionary Biology | Genetics and Genomics Research article et al., 2021). The main question we were interested in is whether we can find alternative models which (1) fit as well as, or better than the published graph, (2) differ in important ways from the published graph, and (3) cannot immediately be rejected based on other evidence such as temporal plausibility. The studies were selected according to the criterion that an AG model inferred in the study is used as primary evidence for at least one statement about population history in the main text of the study. In other words, the AG method was used in the original studies to support new conclusions about population history, and not simply to show that there is a model that exists that does not contradict results of other genetic analyses, an approach that is a valid use of AGs and has been taken in some studies (e.g., Seguin- Orlando et al., 2014; Narasimhan et al., 2019; Wang et al., 2019). There are many published studies that could have been included in our re- evaluation exercise as they meet our key criterion (e.g., Yang et al., 2017; McColl et al., 2018; Posth et al., 2018; Flegontov et al., 2019; Carlhoff et  al., 2021, Kutanan et  al., 2021; Bergström et  al., 2022; Lipson et  al., 2022; Vallini et al., 2022). However, critical re- evaluation of each published graph is an intensive process, and the sample of studies we revisited is diverse enough to identify some general patterns. Here we present a high- level summary of these analyses. Discussion of individual case studies follows below, and for details see the exposition in Appendix 2. For 19 out of 22 published graphs we examined, we were able to find at least one, but usually many, graphs of the same complexity (number of groups and admixture events), with an LL score that was nominally better than that of the published graph (see results for 11 selected graphs in Table 1 and full results for all 22 graphs in Supplementary file 1). The 22 graphs were drawn from the 8 publications as there were multiple final graphs presented in some of the publications (Shinde et al., 2019; Sikora et al., 2019; Librado et al., 2021), or we examined selected intermediates in the model construction process (Bergström et al., 2020; Lazaridis et al., 2014; Lipson et al., 2020b; Wang et al., 2021), or we introduced an outgroup not used in the original study (Hajdinjak et al., 2021; Sikora et al., 2019), or we tested additional graph complexity classes dropping ‘unnecessary’ admix- ture events (Lipson et al., 2020b; Sikora et al., 2019). These alternative graphs often fit not significantly better than the published one after taking into account variability across single- nucleotide polymorphisms (SNPs) via bootstrapping. In the following cases, at least one model that fits significantly better than the published one according to our bootstrap model comparison method was found: the Bergström et al. and Lazaridis et al. 7- population graphs; the Librado et al., 2021 graph with 3 admixture events; the Hajdinjak et al. graphs with or without adding a chimpanzee outgroup; the Lipson et  al., 2020b. intermediate graphs with 7 groups and 4 admixture events and with 10 groups and 8 admixture events; the Wang et al. 12- population graph; and the Sikora et al. graphs for West Eurasians and for East Eurasians with 10 or 6 admixture events (Supplementary file 1). In nearly all cases (except for the Lazaridis et al. six- population graph, Shinde et al. graph with eight populations and three admixture events, and the Librado et al., 2021. graph with four admixture events), we also identified many additional graphs that fit the data not significantly worse than the published ones. In every example, some of these graphs have topologies that are qualitatively different in important ways from those of the published graphs. Features such as which populations are admixed or unadmixed, direction of gene flow, or the order of split events, if not constrained a priori, are generally not the same between alternative fitting models for the same populations. This result agrees with the expectation from our exploration of simulated AGs (Figure 1). While some of these graphs can be rejected since their topologies appear highly unlikely because of non- genetic or unrelated genetic evidence, for all of the publications except one (Shinde et al., 2019), there are alternative equally- well- or- better- fitting graphs we identified and examined manually that differ in qualitatively important ways with regard to the implications about history, are temporally plausible (for instance, very ancient populations do not receive gene flows from sources closely related to much less ancient groups), and not obviously wrong based on other lines of evidence. These findings and the results on simulated AGs suggest that complex AG models, even with a very good fit to the data, often differ in important ways from true population histories. The previous statements are valid if the original parsimony constraints are applied, that is, if the graph complexity (the number of admixture events) is not altered. Below in selected case studies (Shinde et al., Librado et al., 2021.) we also explore the effect of relaxing the parsimony constraint. Table 1 and Figure 2 summarize these results for one or a few graphs from each publication, while Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 7 of 62 Evolutionary Biology | Genetics and Genomics Research article Figure 2. Log- likelihood (LL) scores of published graphs (those shown in Table 1) and automatically inferred graphs. Each dot represents the LL score of a best- fitting graph from one findGraphs iteration (low values of the score indicate a better fit); only topologically distinct graphs are shown. LL scores for the published models and best- fitting alternative models found are shown by blue and pink x’s, respectively. Bootstrap distributions of LL scores for these models (vertical lines, 90% CI) and their medians (solid dots) are also shown. Lower scores of the fits obtained using all single- nucleotide polymorphisms (SNPs), relative to the bootstrap distribution, indicate overfitting. Green and red horizontal lines show the approximate locations where newly found models consistently have fits significantly better or worse, respectively, than those of the published model. In the case of the Bergström et al., Lazaridis et al., and Hajdinjak et al. studies, one or more worst- fitting models were removed for improving the visualization. The setups shown here (population composition, number of groups and admixture events, topology search constraints) match those shown in Table 1. Supplementary file 1 contains the full results for all studied graphs and setups. Table 2 summarizes our assessment of inferences in the original publications that were supported by the published graphs. To identify alternative models, we ran many iterations of findGraphs for each set of input popula- tions, constraints, and the number of admixture events being fit to the data, and we selected the best- fitting graph in each iteration, that is, a graph with the lowest LL score. Each iteration was initiated from a random graph. The algorithm is non- deterministic so that in each iteration it takes a different trajectory through graph space, possibly terminating in a different final best graph. The number of admixture events in the initial random graphs and in the output graphs was always kept equal to that of the published graph. For each example, we counted how many distinct topologies were found with significantly or non- significantly better or worse LL scores than that of the published graph (Table 1, Supplementary file 1). To obtain a formally correct comparison of model fit, the published graph and each alternative model were fitted to resampled replicates of the dataset and the resulting LL score distributions were compared (see Appendix 1, Sections 1.B.3 and 2.E). As shown in Figure 2, for four of the eight publications we re- analyzed, the LL score of the published graph run on the full data is better than almost all the bootstrap replicates on the same data (it falls below the fifth percentile), which is a sign of overfitting, and underscores the importance of applying bootstrap to assess the robustness of fitted models and conclusions drawn from them. Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 8 of 62 Evolutionary Biology | Genetics and Genomics Research article e r e w e n o d e h s i l b u p e h t n a h t e s r o w y l t n a c fi n g i s - n o n d n a i r e t t e b g n i t t fi s l e d o m l l a ) s k s i r e t s a y b d e k r a m ( s e s a c o w t n i l y n o , y l l a u n a m d e m r o f r e p e b o t d a h t n e m s s e s s a s i h t e c n S i . s h p a r G d n fi y b d e t a r e n e g s l e d o m e b i s u a p y l l l l a r o p m e t d n a g n i t t fi - l l e w e v i t a n r e t l a f o s t e s n i d e s s e s s a e w t r o p p u s e s o h w s e r u t a e f y e k s t s i l l e b a t e h T l . s i s y a n a - e r r u o n i t r o p p u s r i e h t f o l e v e l e h t d n a i s e d u t s l i a n g i r o e h t n i s e c n e r e f n i t r o p p u s t a h t ) s G A ( s h p a r g e r u t x m d a d e h s i l i b u p e h t f o s e r u t a e F . 2 e b a T l . ) s l i a t e d r o f 2 i x d n e p p A n i s n o i t c e s e v i t c e p s e r e h t e e s ( i d e n m a x e s a w s l e d o m g n i t t fi - t s e b f o t e s b u s a l y n o s e s a c r e h t o n i ; d e z i n i t u r c s y l l a r o p m e t g n o m a t r o p p u s l i a s r e v n u g n k c a i l l e d o m d e h s i l b u p e h t f o s e r u t a e F l e b i s u a p y l l l a r o p m e t l l a y b d e t r o p p u s l e d o m d e h s i l b u p e h t f o s e r u t a e F s h p a r G d n fi y b d e t a r e n e g s l e d o m e v i t a n r e t l a l e b i s u a p l d e z i n i t u r c s e w t a h t s h p a r G d n fi y b d e t a r e n e g s l e d o m e v i t a n r e t l a e r u t x m d a i / s p u o r G s t n e v e y d u t S d n a , i d e x m d a n u e r a s e g a e n i l g o d n a e n a r r e t i d e M t s a E d n a , n a c i r e m A , ) l a k a B i ( n a i r e b S i f o s g o d d n a , ) a i l e r a K ( n a e p o r u E t s a E i , ) c h t i l o e N y l r a E y n a m r e G ( n a e p o r u E t s e W e h t . i d e x m d a e r a ) g o d g n g n i s i i a e n u G w e N ( i n g i r o n a i s A t s a e h t u o S n a i l e r a K e h t f o e t a d e h t o t r o i r p ( s e g a e n i l g o d d e t a c i t s e m o d f o e c n e g r e v i d y l r a E . ) a y 0 0 9 , 0 1 , g o d * 3 / 7 0 2 0 2 , . l a t e m ö r t s g r e B e h t m o r f d e v i r e d t o n s i p u o r g y r e h p i r e P s u d n I e h t n i l y r t s e c n a d e t a e r - r e m r a f n a n a r I i ) 1 ( . s p u o r g c h t i l i l o c a h C r a s s i H e p e T r o c h t i l i o e N z u r i F i j j a H . p u o r g y r e h p i r e P s u d n I e h t n i y r t s e c n a n a i s A s i e r e h T ) 2 ( 4 / 8 9 1 0 2 , . l a t e e d n h S i m o r f d e v i r e d t o n s i p u o r g y r e h p i r e P s u d n I e h t n i l y r t s e c n a d e t a e r - r e m r a f n a n a r I i ) 1 ( A / N . p u o r g y r e h p i r e P s u d n I e h t n i y r t s e c n a d e t a e r - n a i s A s i l e r e h T ) 2 ( * 3 / 8 . s p u o r g c h t i l i l o c a h C r a s s i H e p e T r o c h t i l i o e N z u r i F i j j a H e h t o t d e t a e r l s e c r u o s l a r t s e c n a e e r h t f o e r u t x m a i t n e s e r p e r s n a e p o r u E y a d - t n e s e r P A / N y l r a e d n a , s r e r e h t a g – r e t n u h n a e p o r u E t s e W , ) 1 A M ( a t ’ l a M : s p u o r g g n w o i l l o f e h t . s r e m r a f n a e p o r u E 4 / 7 4 1 0 2 , . l a t e s i d i r a z a L . e g a e n i l l d e t a e r - 2 M O D a d n a d e t a e r - l C W C a f o e r u t x m a i s i n a p r a T ) 5 ( h t i w d e t a c o s s a i l r e t s u c e h t d n a p u o r g C W C e h t g n i t c e n n o c w o fl e n e g o n s i e r e h T . p u o r g 2 M O D e h t n i t n e s b a s i i e r u t x m d a d e t a e r - l - A N A O E N ) 1 ( ) 3 ( . ) e d a c l a m r o f y e h t ( s p u o r g r e t s i s e r a T N O P - C d n a 2 M O D d e z i i m x a m s i y r t s e c n a e s o h w e r u t l u c a t h s a t n S r e t a i l e h t f o s e s r o h d n a s e s r o h a y a n m a Y . ) G R U T , T N O P - C , 2 M O D ( e p p e t S n r e t s e W e h t n i . - p u o r g A N A O E N e h t o t p u o r g t s o h g g n h c n a r b - p e e d a m o r f i w o fl e n e g s i e r e h T ) 2 ( ) 4 ( 9 r o 8 / 0 1 1 2 0 2 , . l a t e o d a r b L i l n o i t a u p o p a o t d e g n o e b 3 5 6 1 K B l l a u d i v i d n i l e v a C o r i K o h c a B d o - r a e y - 0 0 0 , 5 3 ~ e h T ) 2 ( , i m h s I - ’ t s U e h t o t e g a e n i l i d e t a c o s s a - ) P U I ( i c h t i l l o e a P r e p p U l a i t i n I o r i K o h c a B d o l . s e g a e n i l 1 - 6 1 1 Q t e y o G d n a , n a u y n a T i - r a e y - 0 0 0 , 3 4 o t - 0 0 0 , 5 4 ~ e h t n i d n u o f e g a e n i l e h t m o r f s w o fl e n e g e r a e r e h T ) 1 ( l d e t a e r - 3 5 6 1 K B a d n a d e t a e r - r i h g n u S a l f o e r u t x m a i s i e g a e n i l i 6 1 e c n o t s e V e h T ) 3 ( . e g a e n i l d n a , e d n e M , e d n a m e L ( s p u o r g n a c i r f A t s e W y a d - t n e s e r p n i d e z i i m x a m e g a e n i l A ) 1 ( d e z i i m x a m e g a e n i l i g n h c n a r b - p e e d a s i a k a L m u h S n i t n e n o p m o c y r t s e c n a r e h t o n A ) 2 ( o t d n a l a u d i v i d n i a k a L m u h S t n e c n a e h t o t i y r t s e c n a e m o s d e t u b i r t n o c o s l a ) a b u r o Y . i t u b M d n a i a k a B y a d - t n e s e r p . i t u b M d n a i a k a B s r e r e h t a g – r e t n u h t s e r o f n a r e h t n i i t i l p s l a h t r e d n a e N / n a m u h n r e d o m e h t i t a g n g r e v i d , . e . i ( y r t s e c n a i ’ c a h c r a - r e p u S ‘ ) 3 ( d n a , e d n e M , e d n a m e L , a k a L m u h S , i t u b M i , a k a B o t d e t u b i r t n o c ) r e p e e d r o t n o p i . a b u r o Y , a k a B i , a t o M , w a g A o t d e t u b i r t n o c ) s e g a e n i l r o ( e g a e n i l n a m u h n r e d o m t s o h g A ) 4 ( . a b u r o Y d n a , e d n e M , e d n a m e L , a k a L m u h S , i t u b M l a s r e v i n u t s o m a l s i s r e r e h t a g – r e t n u h e s e n a m a d n A o t d e t a e r e c r u o s l a m o r f e r u t x m d A i i , c h t i l o e N e t a L r e v i R w o l l e Y r e p p U , n a t e b T i , n o m o J e h t n i g n i r r u c c o , s n a i s A t s a E n i i c h t i l o e N y l r a E d n a l s I i a n h C d n a , e g A n o r I n a w a T i i , c h t i l o e N e t a L r e v i R o a L i t s e W . s p u o r g ) o a d g n a L ( i r e r e h t a g – r e t n u h s u s a c u a C e h t m o r f w o fl e n e g d e v i e c e r e g a e n i l ) E N A _ 1 A M ( a t ’ l a M e h T . e g a e n i l ) G H C r o P L _ G H s u s a c u a C ( A / N A / N A / N e g a p t x e n n o d e u n i t n o c 2 e b a T l 8 / 2 1 6 / 3 1 1 2 0 2 , . l a t e g n a W 9 1 0 2 , . l a t e a r o k S i ’ t s e W ‘ 1 1 / 2 1 b 0 2 0 2 , . l a t e n o s p L i . l a u d i v i d n i 1 - 6 1 1 Q t e y o G e h t f o t a h t o t , l a c i t n e d i t o n t u b , d e t a e r l s a w t a h t 8 / 2 1 1 2 0 2 , . l a t e j i k a n d a H j Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 9 of 62 Evolutionary Biology | Genetics and Genomics Research article y l l a r o p m e t g n o m a t r o p p u s l i a s r e v n u g n k c a i l l e d o m d e h s i l b u p e h t f o s e r u t a e F s h p a r G d n fi y b d e t a r e n e g s l e d o m e v i t a n r e t l a l e b i s u a p l a m o r f w o fl e n e g d e v i e c e r s e g a e n i l ) P U _ a n a Y ( a n a Y d n a ) E N A _ 1 A M ( a t ’ l a M e h T ) 1 ( o t g n i t u b i r t n o c i s e n o e h t e r o f e b g n g r e v i d e c r u o s d e t a c o s s a - n a i s A i t s a E n o m m o c l s i v o C d n a , ) P L _ a k s a A l ( 1 R S U , ) M _ a m y l o K ( a m y l o K , ) N _ e v a C s l i v e D ( e v a C s ’ l i v e D e h t . s e g a e n i l ) P L _ s i v o C l ( s a h a m y l o K d n a l , s w o fl e n e g d e t a e r - n a e p o r u E o n d e v i e c e r e g a e n i l e v a C s ’ l i v e D e h T ) 3 ( l . ) s i v o C d n a 1 R S U ( s n a c i r e m A i t n e c n a n a h t l y r t s e c n a d e t a e r - n a e p o r u E s s e l o t r e s o c l s i s e g a e n i l l s i v o C d n a , 1 R S U , a m y l o K e h t n i y r t s e c n a d e t a e r - n a e p o r u E l ) 2 ( . a n a Y o t n a h t a t ’ l a M d e z i n i t u r c s e w t a h t s h p a r G d n fi y b d e t a r e n e g s l e d o m e v i t a n r e t l a l e b i s u a p y l l l a r o p m e t l l a y b d e t r o p p u s l e d o m d e h s i l b u p e h t f o s e r u t a e F e r u t x m d a i / s p u o r G s t n e v e y d u t S d e u n i t n o c 2 e b a T l 6 / 4 1 9 1 0 2 , . l a t e a r o k S i ’ t s a E ‘ Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 10 of 62 Evolutionary Biology | Genetics and Genomics Research article The fraction of graphs with scores better than the score of the published graph should not be overinterpreted, as it is influenced by the findGraphs algorithm, which does not guarantee ergodic sampling from the space of well- fitting AGs. In particular, it is possible that despite findGraph’s strat- egies for efficiently identifying classes of well- fitting AGs (see Appendix 1, Sections 1.B.1 and 2.C), it has a bias toward missing particular classes of graph topologies. However, even one alternative graph which is not significantly worse- fitting than the published graph suggests that we are not able to iden- tify a single best- fitting model. Many of these alternatives, despite providing a good fit to the data, appear unlikely, for example, because they suggest that Paleolithic- era humans are mixed between different lineages closely related to present- day humans. We were mainly interested in alternative models which are also plausible, and so we constrained the space of allowed topologies in findGraphs to those we considered plausible a priori, in cases where this was necessary for reducing the search space size. Constraints were either integrated into the topology search itself, or were applied to outcomes of unconstrained searches, as detailed below. Below we summarize our key findings and the methodological implications from our re- analysis of the eight published datasets. For more detailed discussions see Appendix 2. 1. Bergström et al., 2020 The AG for ancient and present- day dogs in Figure 1e of Bergström et al., 2020 includes an outgroup, six other groups and three admixture events (Figure  3a, Figure 3—source data 1). A best- fitting newly found graph fits the data nominally better than the published one (two- tailed empirical p- value = 0.332), and it bears a closer resemblance to the human population history (Figure 3—source data 2). In this new seven- population model (Figure 3a), both American and Siberian dog lineages repre- sent a mixture between groups related to the Asian and East European dog lineages, and robust genetic results suggest that in the time horizon investigated in the original publication (after ca. 10,900 years ago) nearly all Siberian (Jeong et al., 2019; Sikora et al., 2019) and all American (Raghavan et al., 2014; Raghavan et al., 2015; Moreno- Mayar et al., 2018) human populations were admixed between groups most closely related to Europeans and East Asians. According to this model, East Mediterranean dogs are modeled as a mixture of a basal branch (splitting deeper than the divergence of the Asian and European dogs) and West European dogs, again in agreement with current models of genetic history of West Asian human populations who are modeled as a mixture of ‘basal Eurasians’ and West European hunter–gatherers (Lazaridis et al., 2016; Lipson et al., 2017). Although greater congruence with human history increases the plausibility of findGraph’s newly identified model rela- tive to the published model, to make unbiased comparisons between the history of the two species, model selection should be done strictly independently for each species, and so the genetic data alone does not favor one model more than another. To explain why the original paper on the population history of dogs missed the model that find- Graphs identified, we observe that the Bergström et al., 2020 AG search was exhaustive under the parsimony constraint (no more than two admixture events for six populations, with the seventh group added at a later stage without an exhaustive topology search), and thus missed the potentially true topology including three admixture events for these six populations. This case study also illustrates that even in a relatively low complexity context (seven groups and three admixture events) applying manual approaches for finding optimal models is risky. When any new group such as an Early Neolithic dog from Germany is added to the model, it may introduce crucial new constraints into the system, and re- exploring the whole graph space in an automated way is necessary to avoid missing the true model. In contrast, mapping a newly added group on a simple skeleton graph (even when that skel- eton is a uniquely best- fitting model like in Bergström et al., 2020) may yield a topology that is at odds with the true history. 2. Lazaridis et al., 2014 The graph in Figure 3 (in Lazaridis et al., 2014) suggested that present- day Europeans are derived from at least three populations that are very much differentiated genetically: West European hunter– gatherers (WHG), early European farmers (EEF), and Siberian hunter–gatherers from the same lineage as that of the Mal’ta boy who lived about 24,000 years ago (MA1). For seven- population graphs with four admixture events, we found 40 out of 306 distinct graphs with a score better than that of the published graph (10 of those graphs are shown in Figure 3—source data 3). The best- fitting newly Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 11 of 62 Evolutionary Biology | Genetics and Genomics Research article Figure 3. Published graphs and selected alternative models from three studies for which we explored alternative admixture graph (AG) fits. In all cases, we selected a temporally plausible alternative model that fits nominally or significantly better than the published model and has important qualitative differences compared to the published model with respect to the interpretation about population relationships. In all but one case, the model has the Figure 3 continued on next page Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 12 of 62 Evolutionary Biology | Genetics and Genomics Research article Figure 3 continued same complexity as the published model shown on the left with respect to the number of admixture events; the exception is the re- analysis of the Librado et al., 2021 horse dataset since the published model with three admixture events is a poor fit (worst Z- score comparing the observed and expected f- statistics has an absolute value of 23.9 even when changing the composition of the population groups to increase their homogeneity and improve the fit relative to the composition used in the published study). For this case, we show an alternative model with 8 admixture events that fits well and has important qualitative differences from the point of view of population history interpretation. The existence of well- fitting AG models does not mean that the alternative models are the correct models; however, their identification is important because they prove that alternative reasonable scenarios exist that are qualitatively different from published models. Model parameters (graph edges) that were inferred to be unidentifiable (see Appendix 1, Section 2.F) are plotted in red. (a) The graph published by Bergström et al., 2020 (on the left) and a nominally better fitting graph for dogs that is more congruent to human history (on the right). For both species, Baikal and Native American groups are mixed between European- and East Asian- related lineages, and a ‘Basal Eurasian’ lineage contributes to West Asian groups; these features are all characteristic of human history but absent in the published dog graph. (b) The graph published by Librado et al., 2021 (modified population composition, on the left) and a significantly better fitting AG that is temporally and geographically plausible (on the right). In contrast to the published graph, in this graph with eight mixture events (the minimum necessary to obtain an acceptable statistical fit to the data), a lineage maximized in horses associated with Yamnaya steppe pastoralists or their Sintashta descendants (C- PONT, TURG, or DOM2) contributes a substantial proportion of ancestry to the horses from the Corded Ware Complex (CWC). Thus, in this model both CWC humans and horses are mixtures of Yamnaya and European farmer- associated lineages. This is qualitatively different from the suggestion that there was no Yamnaya- associated contribution to CWC horses which was a possibility raised in the paper. The AG with eight admixture events is also different from the published model in that it shows a fitting model where the Tarpan horse does not have the history claimed in the study (as an admixture of the CWC and DOM2 horses). (c) The graph published by Hajdinjak et al., 2021 (on the left) and a significantly better fitting AG, but without a specific lineage shared between the Bacho Kiro Initial Upper Paleolithic group and East Asians (on the right). In this model, all the lineages shared between Bacho Kiro IUP and East Asians contributed a large fraction of the ancestry of later European hunter–gatherers as well, and thus this graph does not imply distinctive shared ancestry between the earliest modern humans in Europe and later people in East Asia, and instead could be explained by a quite different and also archaeologically plausible scenario of a primary modern human expansion out of West Asia contributing serially to the major lineages leading to Bacho Kiro, then later East Asians, then Ust’-Ishim, then the primary ancestry in later European hunter–gatherers. The online version of this article includes the following source data for figure 3: Source data 1. The published (Bergström et al., 2020) and alternative admixture graphs for dogs found with findGraphs. Source data 2. Alternative admixture graphs for humans found with findGraphs for the dataset from Bergström et al., 2020. Source data 3. The published admixture graph from Lazaridis et al., 2014 and alternative graphs found with findGraphs (seven populations, four admixture events). Source data 4. The published admixture graph from Shinde et al., 2019 and alternative graphs found with findGraphs (8 pops., 3 adm. events) relying on the original set of SNPs and group composition, and original (incorrect) algorithm for calculating f- statistics. Source data 5. The published admixture graph from Shinde et al., 2019 and alternative graphs found with findGraphs (eight populations, three admixture events) for the modified group composition and using the updated algorithm for calculating f- statistics. Source data 6. Alternative graphs allowing for an additional admixture event found with findGraphs for the dataset from Shinde et al., 2019: 8 populations, 4 admixture events, the modified group composition, and the updated algorithm for calculating f- statistics. Source data 7. The published admixture graphs from Librado et al., 2021 and alternative graphs found with findGraphs (10 populations, 3–5 admixture events) for the modified group composition and using the updated algorithm for calculating f- statistics. Source data 8. Published admixture graph from Hajdinjak et al., 2021 and alternative graphs found with findGraphs (12 populations, 8 admixture events). Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 13 of 62 Evolutionary Biology | Genetics and Genomics Research article found model and two other models fit the data significantly better than the published model (Supple- mentary file 1), but their topology is qualitatively very similar to that of the published graph (Figure 3—source data 3). In the best- fitting newly found model, French and Karitiana share some drift to the exclusion of MA1, while in the published model the source of MA1- related ancestry in French is closer to MA1 than to Karitiana. It is important to point out that not all of the 40 alternative graphs that fit nominally or significantly better than the published one are consistent with the conclusion that modern European populations are admixed between three different ancestral populations (Figure 3—source data 3). According to the fifth alternative graph in Figure 3—source data 3 that fits nominally better than the published model (p- value = 0.464), the present- day European population was formed by admixture of an MA1- related lineage and a European Neolithic- related lineage, with no WHG contribution. Of course, other lines of evidence make it clear that LBK Stuttgart is a mixture of Anatolian farmer- and WHG Loschbour- related ancestry (e.g., Lazaridis et al., 2016; Lipson et al., 2017), thus providing external information in favor of the Lazaridis et  al., 2014 model. The use of such ancillary information in concert with graph exploration is important in order to obtain more confident inferences about popu- lation history taking advantage of AGs. We also note that a large group of newly found models (247 graphs) fits not significantly worse than the published one (Supplementary file 1), and those are topologically diverse. Thus, strictly speaking, the AG method on the given dataset cannot be used to prove that the published model is the only one fitting the data. 3. Shinde et al., 2019 The skeleton AG in the original study (Shinde et  al., 2019) was constructed manually, and subse- quently all possible branching orders (105) within the five- population Iranian farmer/herder- related clade were tested. The published model (Figure  3 in that study) included nine groups and three admixture events, but one group (Belt Cave Mesolithic) had a very high missing data rate. Following the approach of the published paper, we repeated findGraphs analysis both with and without the Belt Cave individual. Thus, we initially explored the following topology classes: 9 groups with 3 admix- ture events on ca. 19,000 polymorphic sites and 8 groups with 3 admixture events on ca. 470,000 sites (Figure 3—source data 4, Supplementary file 1). The finding that the predominant ancestry component of the Indus Periphery group was the most basal branch in the Iranian farmer clade was a prominent claim of the original study (Shinde et al., 2019). This finding if correct is important, since it implies that the Iranian- related ancestry in the Indus Valley Civilization genetic grouping (which is the same group as Indus Periphery or IP) split from the Iranian- related ancestry in the first Iranian plateau farmers before the date of the Hajji Firuz farmers, who at ~8000 years ago are among the earliest people living on the Iranian plateau known to have grown West Asian crops. The ancient DNA record combined with radiocarbon dating evidence suggests that beginning around the time of the Hajji Firuz farmers, both West Asian domesticated plants such as wheat and barley, and Anatolian farmer- related admixture, began spreading eastward across the Iranian plateau. If the Iranian- related ancestry in IP was spread eastward into the Indus Valley across the Iranian plateau as part of the same agricul- turally associated expansion—perhaps brought by people speaking Indo- European languages as well as introducing West Asian crops—then we would expect to see at least some of the Iranian- related ancestry in IP being a clade with that in Hajji Firuz relative to Ganj Dareh. The following groups were admixed by default in the graph models compared in the original study: Hajji Firuz Neolithic (labeled ‘Chalcolithic’ in that study but the dates are Neolithic) and Tepe Hissar Chalcolithic were considered as mixtures of an Anatolian farmer- related lineage and an Iranian farmer- related lineage. Indus Periphery was assumed to be a mixture of an Andamanese- related lineage representing ancient South Indians (ASI) and an Iranian farmer- related lineage. The original study differed from ours since these constraints were introduced manually, but we wanted our topology search to be automatic and to explore a wider range of parameter space. To provide power to detect negative f3- statistics useful for constraining the model search, we introduced several modifications to the original group composition (described in Appendix 2) and a new algorithm that makes it possible to compute negative f3- statistics on pseudo- haploid data, but at a cost of removing sites with only one chromosome genotyped in any non- singleton population (see Appendix 1, Section 1.A). We repeated topology search with this set of f- statistics providing additional constraints, performing 4,000 runs of the findGraphs algorithm. The Mota ancient African Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 14 of 62 Evolutionary Biology | Genetics and Genomics Research article individual was set as an outgroup and three admixture events were allowed in the eight- population graph. Among 4,000 resulting graphs (one from each findGraphs run), 144 were distinct topologi- cally, and the published model was recovered in 13 runs of 4,000 (Supplementary file 1). Only four distinct topologies fitting nominally better than the published one were found, and those had LL scores almost identical to that of the published AG. These four alternative models (Figure 3—source data 5b) shared all topologically important features of the published model (Figure 3—source data 5). Five other topologies differed in important ways from the published one and emerged as fitting the data worse, but not significantly worse, than the published one (Figure 3—source data 5c). But in fact, the AG analysis reported above may not be an adequate exploration of the problem. Although absolute fits of the best models found are good (WR = 2.5 SE), the parsimony constraint allowing only three admixture events precluded correct modeling of basal Eurasian ancestry shared by all West Asian groups (Lazaridis et  al., 2016) or of the Indus Periphery group itself, for which a more complex 3- component admixture model was proposed elsewhere (Narasimhan et  al., 2019). Concerned that this oversimplification could be causing our search to miss important classes of models, we explored qpAdm models for the Indus Periphery group, following the protocol by Narasimhan et al., 2019 (see the dataset composition in Supplementary file 4). Our qpAdm results (Supplementary files 5 and 6, Appendix 2) show that the parsimony assumption that was made when constructing the AG analysis in Shinde et al., 2019 is contradicted by f- statistic evidence since the simplest fitting qpAdm model for the IP group includes four ancestry sources, not two (Indus Periphery = Ganj Dareh Neolithic + Onge (ASI) + WSHG + Anatolia Neolithic), and indeed Narasimhan et al. themselves showed this when they presented a qpAdm model that was more complex (Ganj Dareh Neolithic + Onge (ASI) + WSHG) than the one used for constraining the AG model comparison (Ganj Dareh Neolithic + Onge (ASI)). To explore how the parsimony constraint influences results, we allowed four admixture events in the eight- population graph (Supplementary file 1). Among 4,000 resulting graphs (one from each findGraphs run), 443 were distinct topologically, and 270 had WRs between 2 and 3 SE, that is, fitted the data well. In Figure 3—source data 6b, we show four graphs with four admixture events that model the Indus Periphery group as a mixture of three or four sources, with a significant fraction of its ancestry derived from the Hajji Firuz Neolithic or Tepe Hissar Chalcolithic lineages including both Iranian and Anatolian ancestries. The fits of these models are just slightly different (e.g., LL = 11.7 vs 9.3, both WRs = 2.4 SE) from that of the best- fitting model (Figure 3—source data 6a), and are similar to that of the simpler published graph. Besides these four illustrative graphs, dozens of topol- ogies with very different models for the Indus Periphery group fit the data approximately equally well, suggesting that there is no useful signal in this type of AG analysis when the parsimony constraint is relaxed (this finding is similar to that in our re- analysis of the dog AG in Bergström et al., 2020, where relaxation of the parsimony constraint identified equally well- fitting AGs that were very different with regard to their inferences about population history). These results show that at least with regard to the AG analysis, a key historical conclusion of the study (that the predominant genetic component in the Indus Periphery lineage diverged from the Iranian clade prior to the date of the Ganj Dareh Neolithic group at ca. 10 kya and thus prior to the arrival of West Asian crops and Anatolian genetics in Iran) depends on the parsimony assumption, but the preference for three admixture events instead of four is hard to justify based on archaeological or other arguments. Why did the Shinde et al., 2019 AG analysis find support for the IP Iranian- related lineage being the first to split, while our findGraphs analysis did not? Shinde et al., 2019 study sought to carry out a systematic exploration of the AG space in the same spirit as findGraphs—one of only a few papers in the literature where there has been an attempt to do so—and thus this qualitative difference in findings is notable. We hypothesize that the inconsistency reflects the fact that the deeply diverging WSHG- related ancestry (Narasimhan et al., 2019) present in the IP genetic grouping at a level of ca. 10% was not taken into account explicitly neither in the AG analysis nor in the admixture- corrected f4- symmetry tests also reported in Shinde et al., 2019. The difference in qualitative conclusions may also reflect the fact that the Shinde et al. study was distinguishing between fitting models relying on an LL difference threshold of 4 units (based on the AIC). As discussed in Appendix 1, AIC is not appli- cable to AGs where the number of independent model parameters is topology- dependent even if the numbers of groups and admixture events are fixed, and models compared with AIC should have the same number of parameters. Thus, we believe that the analysis by Shinde et al. was over- optimistic Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 15 of 62 Evolutionary Biology | Genetics and Genomics Research article (as compared to the bootstrap model comparison method we use) about being able to reject models that were in fact plausible using its AG fitting setup. 4.Librado et al., 2021 In contrast to the other studies revisited in our work, the AG published by Librado et al., 2021 was inferred automatically using OrientAGraph. Models with three (Figure 3b in that study) and zero to five (Ext. Data Figure 5a–d in that study) admixture events were shown. The dataset included 10 popu- lations (9 horse populations and donkey as an outgroup) and was based on 7.4 million polymorphic transversion sites with no missing data at the group level. Unlike all the other AGs we re- evaluate in this study whose fits to the data were evaluated in the published studies using qpGraph, the topolo- gies published in Librado et al., 2021 were not evaluated for statistical goodness- of- fit, and in fact fit the f- statistic data so poorly that even simple statistics show they cannot be correct (Figure 3b, Figure 3—source data 7a, c, e,, Supplementary file 1). In this case, the approach of using findGraphs to identify alternative topologies with the same number of admixture events that fit the data better is meaningless, as both the published models and the alternative models do not have enough degrees of freedom to accommodate the complexity present in the real data (Figure 3—source data 7). In particular, we found that WR of the published model with three admixture events is 23.9 SE (Figure 3—source data 7a). For this reason, we moved to topology searches in more complex model spaces incorporating six to nine admixture events. Temporally plausible models with even a modest fit (WR between 3 and 4 SE) were encountered only among models with eight and nine admixture events (Figure 3—source data 7j–r). Librado et al., 2021 discussed five inferences relying fully or partially on their published AGs reported in that study (Table  2). The simplest temporally plausible and best- fitting (WR = 3.4 SE) model we found (eight admixture events, see Figure 3b and the first model in Figure 3—source data 7j) supports inferences 2 and 4, and is incompatible with inferences 1, 3, and 5 (Table 2). We consider this model to be plausible also from the geographical perspective (see Appendix 2 for an interpretation of this topology). We are not arguing here that this AG represents the true history; in fact, it is highly unlikely to be the truth, given how large the space of all possible admixture events is and how much admixture evidently occurred relating all these groups (which makes finding the true model extremely unlikely, see the results on simulated data in Figure 1 and Appendix 1—figure 2). However, our set of 16 temporally plausible and fitting (WR < 4 SE) models with eight or nine admix- ture events (Figure 3—source data 7j–r) is consistent with some features of the published graph being stable: the features (2) that DOM2 and C- PONT are sister groups, and (4) that there was a gene flow from a deep- branching ghost group to NEO- ANA (Table 2). Equally important is our finding that there are plausible models that are inconsistent with other inferences in Librado et al., 2021. (Table 2). For example, 13 of these 16 models are inconsistent with the suggestion that there was no gene flow connecting the CWC group and the cluster maxi- mized in the Western steppe (DOM2, C- PONT, and TURG) (Figure 3—source data 7j–r). In the eight- admixture- event best- fitting AG (Figure 3b, the first model in Figure 3—source data 7j), CWC actually derives appreciable ancestry from the early domestic horse lineage (DOM2) associated with the Sintashta culture to the exclusion of the more distant Yamnaya- associated TURG and C_PONT horses. This scenario presents a parallel to the one observed in humans, with individuals associated with the CWC receiving admixture from Steppe pastoralists albeit in different proportions: ~75% for humans, versus ~20% in horses. These models specifying a substantial Steppe horse contribution to CWC horses are inconsistent with the inference in Librado et al., 2021. that ‘Our results reject the commonly held association between horseback riding and the massive expansion of Yamnaya steppe pastoralists into Europe around 3000 BC’. We are not aware of other lines of evidence in the paper (apart from the fitted AG) that support the claim of no Yamnaya horse impact on CWC horses. Another example of a feature of the published graph that turned out to be unstable is the model for the Tarpan horse. Only 8 of 16 temporally plausible and fitting models (Figure 3—source data 7j–r) support the conclusion by Librado et al. that the Tarpan is a mixture of a DOM2- related and a CWC- related lineage. The other eight models suggest that Tarpan is a mixture of a deep lineage and a DOM2- related lineage (Figure 3b, the first model in Figure 3—source data 7j), echoing a hypoth- esis that Tarpan may be a hybrid with Przewalski- related horses not represented in the AG (Librado et al., 2021). Again, we are not arguing here that our alternative model is right—indeed we are nearly Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 16 of 62 Evolutionary Biology | Genetics and Genomics Research article certain it is wrong in important aspects—but we are merely pointing out that the complexity of the AG space means that qualitatively quite different conclusions are compatible with the statistics fitted in the published paper. 5. Hajdinjak et al., 2021 The AG inferred by Hajdinjak et al. was constructed manually and incorporated 11 groups and 8 admixture events (Figure 2d in the original study). Most (71.4%) models found with findGraphs fit nominally better, and 15.7% fit significantly better than the published model (Table 1, Supplementary file 1, Figure 2), which has a poor absolute fit on this set of sites and groups (WR = 4.8 SE, Figure 3c, Figure 3—source data 8). The statistics described above and the fact that LL scores on all sites lie outside of the LL distribution on resampled datasets (Figure 2) suggest that models in this complexity class are overfitted, but the published topology emerged as fitting relatively poorly. Overfitting arises naturally during manual graph construction as performed in many studies (not only in Hajdinjak et al., 2021, but also in Fu et  al., 2016; Skoglund et  al., 2016; Yang et  al., 2017; Posth et  al., 2018; McColl et al., 2018; Moreno- Mayar et al., 2018; Tambets et al., 2018; van de Loosdrecht et al., 2018; Flegontov et al., 2019; Sikora et al., 2019; Wang et al., 2019; Lipson et al., 2020b; Shinde et al., 2019; Yang et al., 2020; Wang et al., 2021). The graph grew one group at a time, and each newly added group was mapped on to the pre- existing skeleton graph as unadmixed or as a two- way mixture. Another requirement was that all intermediate graphs have good absolute fits (WR below 3 or 4 SE). When the model- building process is constrained in a particular path and fits of all intermedi- ates are required to be good, unnecessary admixture events are often added along the way, and the resulting graph belongs to a complexity class in which models are overfitted. Hajdinjak et  al., 2021’s published graph had three notable features that were interpreted by the authors and used to support some conclusions of the study (Table 2), with the following feature considered the most important: there are gene flows from the lineage found in the ~45,000- to 43,000- year- old Bacho Kiro Initial Upper Paleolithic (IUP) individuals to the Ust’-Ishim, Tianyuan, and GoyetQ116- 1 lineages. We identified 1,421 topologies fitting nominally or significantly better than the published model and moved on to inspect 50 best- fitting topologies for temporal plausibility (all of them fitting significantly better than the published model). All non- African individuals included in the model are Upper Paleolithic and their dates are not drastically different in relative terms, from ca. 45 to 30 kya (1,000 years before present). Nevertheless, we considered most gene flows from later- to earlier- attested lineages as temporally implausible (for instance, GoyetQ116- 1 (~35 kya) Ust’-Ishim (~44 kya), Sunghir III (34.5 kya) Tianyuan (40 kya), etc.) since they imply great antiquity of the later- attested lineages and of all lineages derived from them at least partially. → → Of the 50 topologies inspected, 32 were considered temporally plausible. Of those topologies, none supported the finding of gene flows from the Bacho Kiro IUP lineage specifically into all three of the Ust’-Ishim, Tianyuan, and GoyetQ116- 1 lineages. A total of 17 topologies supported features 2 and 3 but were inconsistent with feature 1; and 14 topologies supported feature 3 only (Table 2). Best- fitting representatives of each of these topology classes are shown along with the published model in Figure 3—source data 8. Considering topological diversity among models that are temporally plausible, conform to current knowledge about relationships between modern and archaic humans, and fit significantly better than the published model, we conclude that feature 3 is probably robust but other details of the fitted AG in Hajdinjak et al.—for example, gene flows to the Ust’-Ishim, Tianyuan, and Goyet Q116- 1 lineages from sources sharing drift exclusively with the Upper Paleolithic Bacho Kiro lineage—should not be interpreted as providing meaningful inferences about population history of Upper Paleolithic modern humans. A central finding of Hajdinjak et al. is that the Bacho Kiro IUP group shares more alleles with pres- ent- day East Asians than with Upper Paleolithic Holocene Europeans despite coming from Europe. Specifically, the study documents significantly positive statistics of the form D(an Asian group, Kostenki14; Bacho Kiro IUP, Mbuti). Hajdinjak et al.’s interpretation of this observation is that ‘there was at least some continuity between the earliest modern humans in Europe [Bacho Kiro IUP] and later people in Eurasia [East Asians]’. However, a significant D- or f4- statistic can have multiple expla- nations. The statistic f4 (Tianyuan, Kostenki14; Bacho Kiro IUP, Mbuti) is fitted equally well by the published 12- population AG (Z- score for the difference between the observed and fitted statistics = 0.64) and by, for example, the AG in Figure 3c (Z- score = 0.94). Under the latter model that fits the Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 17 of 62 Evolutionary Biology | Genetics and Genomics Research article data significantly better than the published model (p- value = 0.02), the Bacho Kiro IUP and Tianyuan branches are not connected by a gene flow and do not receive gene flows from a third common source, but the common ancestor of Ust’-Ishim and all European Paleolithic lineages receives an 8% gene flow from a divergent modern human lineage splitting deeper than Bacho Kiro IUP and Tianyuan (Figure 3c, Figure 3—source data 8c). This scenario or some version of it seems archaeologically and geographically plausible and is not disproven by any other line of genetic or non- genetic evidence of which we are aware. It could correspond to a scenario where a primary modern human expansion out of West Asia contributed serially to the major lineages leading to Bacho Kiro, then later East Asians, then Ust’-Ishim, and finally the primary ancestry in later European hunter–gatherers. This has a very different interpretation from the scenario of distinctive shared ancestry between the earliest modern humans in Europe such as Bacho Kiro IUP and later people in East Asia—to the exclusion of later Euro- pean hunter–gatherers—that is suggested by the Hajdinjak et al. published graph. We are not claiming that this specific alternative model is correct—indeed, it is almost certainly not the correct one given the topological complexity of the set of all AGs consistent with the data—but the existence of it and many other models that fit the data makes it clear that we do not yet have a unique historical explanation for the excess sharing of alleles that has been documented between some Upper Paleolithic European groups (Bacho Kiro IUP, Hajdinjak et al., 2021, and GoyetQ116- 1, Yang et al., 2017 and Hajdinjak et al., 2021) and all East Asians. 6. Lipson et al., 2020b The AG in the original study was constructed manually (Extended Data Figure 4 in that study) and is very complex (12 groups and 12 admixture events): it exists in a space of ~1044 topologies of this complexity. We note that one admixture event was added by Lipson et  al., 2020b to account for potential modern DNA contamination in ancient Shum Laka individuals, and removing it caused a negligible difference in the fit of the published model (Supplementary file 1). Thus, to decrease the complexity of the graph search space, we considered graphs with 12 groups and 11 admixture events. Among 2,000 newly found models, 11.9% fit nominally (but not significantly) better than the published model (Table 1, Supplementary file 1, Figure 2), and absolute fits of 36.7% of novel models are good (WR <3 SE). These metrics, along with the fact that LL scores on all sites lie outside of the LL distri- bution on resampled datasets (Figure 2), suggest that models in this complexity class, including the published model, are overfitted. Of the AGs we re- evaluate in this study, the graph from Lipson et al., 2020b shares with the graphs from Hajdinjak et al., 2021; Sikora et al., 2019; Wang et al., 2021, evidence of being overfitted (Figure 2). Below we discuss four prominent features of the AG published in the original study (that were inter- preted by the authors and used to support some conclusions of the study) and the extent to which these features consistently replicate across the large number of fitting 12- population graphs with 11 admixture events (Table 2). High topological diversity is observed among temporally plausible newly found AGs (see an example in Figure 4a and further topologies in Figure 4—source data 1). Consid- ering extreme cases, two AGs completely lacked support for three features of the published graph (Figure 4a, Figure 4—source data 1c), and one graph supported all four features of the published graph fully (Figure 4—source data 1q, the second model). There are some graphs where defining two distinct ancestral lineages maximized in West Africans and in Mbuti and Biaka (features 1 and 2, Table 2) is essentially impossible since all or nearly all Africans are modeled as a mixture of at least two deep lineages (see alternative graph no. 4 shown in Figure 4—source data 1d, the second model). In some graphs there is no single lineage specific to rainforest hunter–gatherers (Biaka, Mbuti, and Shum Laka) since the primary ancestries in these groups form independent deep branches in the African graph (see Figure 4a and graph no. 16 shown in Figure 4—source data 1j, the second model). The ghost modern and super- archaic gene flows to Africans also had no universal support in the set of alternative graphs we examined (see, for example, Figure 4a and Figure 4—source data 1c). Considering the high degree of topological diversity among models that are temporally plausible, conform to known findings about relationships between modern and archaic humans, and fit nomi- nally better than the published model, we conclude that none of the four AG features from the orig- inal study are consistently supported by our re- analysis (Table 2). This situation may be attributed to (1) overfitting and/or to (2) the lack of information in the dataset (in the combination of groups and SNP sites) and/or to (3) inherent limitations of f- statistics, when distinct topologies predict identical Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 18 of 62 Evolutionary Biology | Genetics and Genomics Research article Figure 4. Published graphs and selected alternative models from three further studies for which we explored alternative admixture graph (AG) fits. (a) The graph published by Lipson et al., 2020b (on the left) and a nominally better fitting AG (on the right). In contrast to the published graph, there is no single lineage specific to modern rainforest hunter–gatherers (Biaka and Mbuti) and Shum Laka (Cameroon_SMA). Rather, the primary ancestries Figure 4 continued on next page Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 19 of 62 Evolutionary Biology | Genetics and Genomics Research article Figure 4 continued in each group are separate deep- branching lineages (the deeper lineage they all share is also the source of the majority of ancestry in all anatomically modern humans modeled here). In contrast to the graph in the published paper, there is no West African- maximized ancestry present in mixed form in Biaka, Mbuti, and Shum Laka; archaic admixture is not limited to a subset of Africans but is present in all anatomically modern humans in various proportions; and there is no ghost modern human ancestry in Agaw, Biaka, Lemande, Mbuti, Mende, Mota, Shum Laka, and Yoruba. (b) The admixture graph published by Wang et al., 2021 (on the left) and a significantly better fitting AG meeting the constraints used to inform model building in the published paper (on the right). The finding of Onge- related admixture that is widespread in East Asia suggesting an early peopling via a coastal route is not a feature of this model. (c) The admixture graph published by Sikora et al., 2019 (simplified "Western" graph, on the left) and a nominally better fitting AG (on the right). The striking feature of the AG suggested in the paper whereby Mal’ta (MA1_ANE) derives some ancestry from a CHG- associated lineage is not a feature of this alternative model. The online version of this article includes the following source data for figure 4: Source data 1. The published admixture graph from Lipson et al., 2020b and alternative graphs found with findGraphs (12 populations, 11 admixture events) using the updated algorithm for calculating f- statistics. Source data 2. The published admixture graph from Wang et al., 2021 and alternative graphs found with findGraphs (12 populations, 8 admixture events) using the updated algorithm for calculating f- statistics. Source data 3. The simplified published admixture graph for West Eurasian groups from Sikora et al., 2019 and alternative graphs found with findGraphs (13 populations, 6 admixture events). Source data 4. The simplified published admixture graph for East Eurasian groups from Sikora et al., 2019 and alternative graphs found with findGraphs (14 populations, 6 admixture events). f- statistics. Our results highlight the mystery around the highly distinctive genetic ancestry of the Shum Laka individuals themselves, who represent the newly reported data in the Lipson et al., 2020b study, and represent a highly important set of genetic datapoints that was not available prior to the study. The ancestral relationships of these four individuals to rainforest hunter–gatherers, and to the primary lineage in present- day West Africans, remains an open question, one whose resolution prom- ises meaningful new insights into modern human population history. 7. Wang et al., 2021 The AG inferred by Wang et  al., 2021 was constructed manually, and the final graph (Extended Data Figure 6 in the original study) included 12 groups and 8 admixture events. We applied several constraints on the graph space exploration process all of which were shared with the Wang et al. graphs (Supplementary file 1). An important feature of the published graphs in Wang et al., 2021 that was remarked upon in the study is admixture from a source related to Andamanese hunter– gatherers that is almost universal in East Asians (Table 2). For example, the abstract states ‘Hunter- gatherers from Japan, the Amur River Basin, and people of Neolithic and Iron Age Taiwan and the Tibetan Plateau are linked by a deeply splitting lineage that probably reflects a coastal migration during the Late Pleistocene epoch.’ We performed 2,000 findGraphs iterations and obtained 1,778 distinct topologies satisfying all the constraints, nearly all of them (1,724) fitting nominally better than the published model, and 12.6% fitting significantly better (Table  1, Supplementary file 1). The models were ranked by LL scores, and 56 highest- ranking topologies, all of them fitting signifi- cantly better than the published one, were assessed for temporal plausibility, and 20 topologies were considered temporally plausible (all of them are shown in Figure 4—source data 2). According to these topologies, 0–2 East Asian groups had a fraction of their ancestry derived from a source specif- ically related to Onge, and 19 topologies included gene flows from the European (Loschbour)- related branch to all 8 East Asian groups (Figure 4—source data 2). The inferred topological relationships among East Asians are variable in this group of AGs, and we decided to apply further constraints that guided model ranking and elimination by Wang et al., based on considerations from archaeological evidence, Y chromosome haplogroup divergence patterns, and population split time estimation (see Appendix 2 for details). Applying these three additional constraints, we identified two models (among the 56 subjected to manual inspection) that satisfied all of them. The highestranking of those models is shown in Figure  4b and Figure 4—source data 2c (the second model), and it includes a 13% (deeply) European- related gene flow to the common ancestor of all East Asians, and gene flows from Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 20 of 62 Evolutionary Biology | Genetics and Genomics Research article the Onge- related branch to just two East Asian groups: Nepal Chokhopani and China WLR LN. This model fits the data significantly better than the published model (p- value = 0.028). We do not claim that this is the correct model (indeed we are almost certain that it is not given the high topological diversity of fitting models), but it is not obviously wrong and differs in qualitatively important ways from the published one. The Wang et  al., 2021 AG provides an illuminating example that helps us to understand the value added by AG construction. The AG construction process in Wang et al. followed a philosophy of not relying entirely on the allele frequency correlation data (not treating the genetic data as inde- pendent to explore how much new insight could come from genetic data alone). Instead, the study integrated other lines of genetic evidence as well as linguistic and archaeological insights explicitly into the AG construction process, with the goal of identifying models consistent with multiple lines of evidence. The fact that after this procedure a fitting graph was obtained is not of great interest, as it is essentially always possible to obtain a fit to allele frequency correlation data when enough admixture events are added. The important question is whether any of the emergent features of the graph that were not applied as constraints in the construction process—for example the evidence of ubiquitous Andamanese- related gene flow throughout East Asia suggesting a coastal route expansion that admixed with an interior route expansion proxied by Tianyuan—were stably inferred. Our analysis does not come to this finding consistently among well- fitting and plausible AGs. We conclude that this important feature of the published graph is not supported by f- statistic analysis alone (Table 2), and indeed we are not aware of a single feature of the Wang et al., 2021 AG that is stably inferred beyond the constraints applied to build it. 8. Sikora et al., 2019 Two AGs inferred by Sikora et al., 2019 were constructed manually based on an SNP set derived from whole- genome shotgun data and incorporated 12 or 13 groups and 10 admixture events (Extended Data Figure 3f in the original study). One graph was focused on West Eurasians, and the other one on East Eurasians, and both included a Neanderthal, a Denisovan, and an African group (Dinka). Although the chimpanzee outgroup was not included in the original graphs, we added it as it dras- tically constrains the topology search space. In contrast to most other published graphs discussed above, gene flows in the graphs inferred by Sikora et al. do not have equal standing: four low- level gene flows (0–1%) connect the Neanderthal lineage to Upper Paleolithic lineages. We repeated each topology search under two alternative settings: either keeping the number of admixture events at 10 to match the published graphs, or at 6 to match simplified versions of the published graphs lacking these low- level Neanderthal gene flows. We performed that modification to simplify the search space and to alleviate the overfitting problem which becomes severe if 10 gene flows across the graph are allowed (Supplementary file 1). In the case of the "Western" graphs with 6 admixture events, 1,000 topology search iterations were performed, 894 distinct topologies were found, 4 models fit significantly better, and 151 models fit nominally better than the published one (Table 1, Supplementary file 1). We inspected those 155 topologies and identified 29 topologies (Figure 4—source data 3) that are temporally plausible. Sikora et al. came to the following striking conclusion relying on the "Western" AG (Table  2): the Mal’ta (MA1_ANE) lineage received a gene flow from the Caucasus hunter–gatherer (CHG) lineage. However, in our findGraphs exploration this direction of gene flow (CHG Mal’ta) was supported by two of the 29 topologies, and the opposite gene flow direction (from the Mal’ta and East European hunter–gatherer lineages to CHG) was supported by the remaining 27 plausible topologies (Figure 4—source data 3). The highest- ranking plausible topology (Figure 4c) has a fit that is not significantly different from that of the simplified published model with six admixture events (p- value = 0.392). We note that the gene flow direction contradicting the graph by Sikora et al. was supported by published qpAdm analyses (Lazaridis et al., 2016; Narasimhan et al., 2019), and qpAdm is not affected by the same model degeneracy issues that are the focus of this study. Considering the topological diversity among models that are temporally plausible, conform to robust findings about relationships between modern and archaic humans, and fit nominally better than the published model, we conclude that the direction of the Mal’ta- CHG gene flow cannot be resolved by AG analysis (Table 2). → Some important conclusions based on the "Eastern" graph also do not replicate across all plausible AGs (Table 2). In the case of the "Eastern" graphs with 6 admixture events, 4,446 topology search Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 21 of 62 Evolutionary Biology | Genetics and Genomics Research article iterations were performed, and 2,785 distinct topologies were found. Only 3 topologies fit signifi- cantly and 13 nominally better than the published one, and 9.8% of topologies fit not significantly worse than the published one (Table 1, Supplementary file 1). Of the AGs belonging to these groups, we inspected 116 best- fitting ones and identified 97 AGs that are temporally plausible. The Sikora et al. "Eastern" AG had three distinctive features that were used to support some conclusions of the study (Table 2). Only feature 2 was universally supported by all the 97 plausible alternative models fitting significantly better, nominally better, or not significantly worse than the simplified published model, while feature 3 was supported by 83 of 97 plausible models, and feature 1 was supported by 28 of 97 plausible models (Table 2). We plotted 14 plausible graphs as examples of topologies supporting all three features, two features, or one feature of the published graph (Figure 4—source data 4). We note that all the "Eastern" graphs discussed here, both the published and alternative ones, have relatively poor absolute fits with WR above 4 or 5 SE. Increasing the number of gene flows to 10 allowed us to reach much better absolute fits (with WR as low as 2.42 SE), but that resulted in high topological diversity (on a par with some other case studies discussed above). Discussion A proposed protocol for using AG fitting in genetic studies AGs represent a conceptually powerful framework for thinking about demographic history, but, as we demonstrate in this study (see also Appendix 2), the practice of manually constructing a small number of complex models without exploring AG space in an automated way can lead to overconfidence in the validity of these models. An ideal outcome of an AG model exploration exercise would be the identification of a model or a group of topologically very similar models which fit the data well and significantly better than all alternative models with the same number of admixture events; however, this is almost never achieved for graphs with more than eight populations and three admixture events in our experience (Appendix 2), and even this approach can lead to potentially unstable results as relaxing the assumption of parsimony (that fewer admixture events is more likely) can lead to quali- tatively quite different equally well- fitting topologies as in our re- analysis of the Bergström et al. and Shinde et al. datasets. Most of the examples of AGs in eight recently published studies we revisited do not fit this ideal pattern, as we were able to identify many topologically different alternative models that could not easily be rejected based on temporal plausibility or other constraints (Figure 3—source data 4, Figure 3—source data 5, Figure 3—source data 6, Figure 3—source data 7, Figure 3— source data 8, Figure 4—source data 1, Figure 4—source data 2, Figure 4—source data 3, Figure 4—source data 4). In particular, for all studies except Shinde et al., 2019 (under a strict parsimony assumption however), we identified AGs that were not significantly worse fitting than the published ones, and with topological features that were different in qualitatively important ways. There were also some more encouraging findings of the exercise we performed to re- evaluate published models. For example, at least one of the key inferences about population history relying on AG modeling were stable for all analyzed models for the Librado et  al., 2021., Hajdinjak et al., Shinde et al. (under the parsimony assumption), and Sikora et al. (simplified "Eastern" graph) studies (Appendix 2). The existence of some stable features in these graphs helps to point the way toward a protocol that we believe should be applied in all future studies that use AG fitting exercises to support claims about population history. We propose the following tentative protocol to identify features of fitting AGs that are stable enough to be used to make inferences about population history. 1. For a given combination of populations, carry out an initial scan using findGraphs to identify reasonable parameter values for the number of allowed admixture events (graph complexity class). For example, run findGraphs allowing between zero and eight admixture events (100 algorithm iterations per graph complexity class), saving one or a few best- fitting AGs after each iteration. The smallest number of admixture events that yields models where the (negative) LL score or the worst f- statistic residual is lower than some threshold can then be explored more deeply by running more iterations of findGraphs. 2. Run findGraphs on the chosen complexity class, where some of the resulting graphs should be inspected manually to determine whether they could in principle be historically plausible models. Implausible models (e.g., models where a very ancient population appears to be admixed between two modern populations) can be filtered out by imposing topological constraints. If no Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 22 of 62 Evolutionary Biology | Genetics and Genomics Research article or only a few graphs remain, findGraphs can be run again under these constraints. This can be repeated until one or more graphs with an acceptable LL score or worst residual has been iden- tified. At this stage, apply the bootstrap method to determine whether the best- fitting graph is significantly better than the next best- fitting graph. If it is not, identify a set of graphs which are not clearly worse than the best- fitting graph by performing bootstrap model comparison for many model pairs. 3. Researchers should compare the resulting graphs to each other with the goal of identifying common features. Although ADMIXTOOLS 2 includes automated tools for cataloguing common topological features (Appendix 1, Sections 1.B.5 and 2.G), we found a manual approach to be valuable as the fitted parameters (especially admixture proportions) are as important for this task as graph topology. 4. Once a set of fitting graphs and stable topological features shared between them is identified, researchers should carry out a findGraphs exploration of the space of graphs with one addi- tional admixture event. If inferences are stable even when fitting graphs with one more level of complexity than the graphs with the minimal number of admixture events needed to fit the data, this increases confidence in the inferences. Furthermore, addition of a new population may introduce crucial information to an existing set of populations, which can change the space of fitting topologies in a profound way, as in our re- analysis of the data from Bergström et al., 2020 (Figure 3a, Figure 3—source data 1). Thus, it is advisable that the topology optimiza- tion procedure is repeated on several alternative population sets, in addition to considering models that allow an additional admixture event beyond the minimum required for parsimony, to explore if inferences about topology change qualitatively. 5. AGs fitted with f- statistics do not distinguish between time and population size as the two factors affecting genetic drift. Moreover, many different complex genetic histories for a set of populations can result in the exact same expected f- statistics. This provides an opportunity to further constrain a model fitting procedure. Methods that take advantage of information from the site frequency spectra (momi2, fastsimcoal, Kamm et  al., 2020, Excoffier et  al., 2013) or derived site patterns, a special case of site frequency spectra (Legofit, Rogers, 2019), can supply alternative information not captured by f- statistics (further information can come from methods that fit haplotype divergence patterns such as MSMC, Schiffels and Durbin, 2014 and SMC++, Terhorst et al., 2017, or inferences based on fitted gene trees such as RELATE, Speidel et al., 2019, and ARGweaver, Hubisz et al., 2020; Hubisz and Siepel, 2020). These tools are too computationally intensive to explore a large number of models, but the advan- tages of the different approaches can be combined by first identifying a set of candidate models using findGraphs, and then testing these candidate models with other methods. This approach is also expected to help address overfitting since different data types almost always include different variable site sets. We believe that researchers should only begin to make strong claims about population history with AGs once a protocol such as we propose is applied. We see the guidelines above as analogous in spirit to the protocols that were introduced in medical genetics at a time of the reproducibility crisis in the field of candidate gene association studies. Many studies looking for risk factors for common, complex diseases resulted in publications with marginally significant p- values without correcting for multiple hypothesis testing that was implicitly performed due to many candidate genes being tested and only those with significant findings being published. Unsurprisingly, most of these claims failed to replicate in follow- up studies in independent sets of samples (Ioannidis, 2005; Border et al., 2019; Collins et al., 2012; Duncan et al., 2019). The human medical genetic community addressed this challenge by coming together to support a rigorous set of commonly accepted standards for declaring genome- wide statistical significance, such as the require- ment that p- values be corrected for the effective number of independent common variants in the genome and requiring correction for the known confounders of population structure and undocu- mented relatedness among individuals (Hirschhorn and Daly, 2005). Conclusions Sampling AG space is a useful method for modeling population histories, but finding robust and accu- rate models can be challenging. As we demonstrated by revisiting a handful of published AGs and re- analyzing the datasets used to fit them, f- statistics are usually insufficient for identifying uniquely fitting AG models, making it necessary to incorporate other sources of evidence. This provides a challenge to previous approaches for automated model building. We investigated several published Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 23 of 62 Evolutionary Biology | Genetics and Genomics Research article AG models and, in nearly all cases, found many alternative models, some of which are historically and geographically plausible but contradict conclusions that were derived from the published models. To conduct these analyses, we developed a method for automated AG topology optimization that can incorporate external sources of information as topological constraints. This method is developed in the ADMIXTOOLS 2 framework, which aside from AG modeling, implements many other methods for population history inference based on f- statistics. It is important to recognize that the key concern we have highlighted in this study—the fact that there can often be thousands of different topologies that are equally good fits to the allele frequency correlation patterns relating a set of populations—does not invalidate the use of allele frequency correlation testing in many other contexts in which it has been applied to make inferences about population history. For example, negative f3- statistics (‘admixture’ f3- statistics) continue to provide unambiguous evidence for a history of mixture in tested populations, and f4- and D- symmetry statistics remain powerful ways to evaluate whether a tested pair of populations is consistent with descending from a common ancestral population since separation from the ancestors of two groups used for comparison. The qpWave methodology remains a fully valid generalization of f4- statistics, making it possible to test whether a set of populations is consistent with descending from a specified number of ancestral populations (which separated at earlier times from a comparison set of popula- tions). In addition, Haak et al., 2015 and Harney et al., 2021 the qpAdm extension of qpWave— which allows for estimating proportions of mixtures for the tested population under the assumption that we have data from the source populations for the mixture—remains a valid approach, unaffected by the concerns identified here. Instead of relying on a specific model of deep population relation- ships, qpAdm relies on an empirically measured covariance matrix of f4- statistics for the analyzed populations, which is highly constraining with respect to estimation of mixture proportions but can be consistent with a wide range of deep history models. All these methods are implemented in ADMIXTOOLS 2. Finally, approaches that use AGs to adjust for the covariance structure relating a set of populations without insisting that the particular AG model that is proposed is true with can be useful, for example for the purpose of analyzing shared genetic drift patterns of a group of populations that derive from similar mixtures. One example was a study that attempted to test for different source populations for Neolithic migrations into the Balkans after controlling for different proportions of hunter–gatherer admixture (Mathieson et al., 2018). Another example was a study that attempted to study shared ancestry between different East African forager populations after controlling for different proportions of deeply divergent source populations (Lipson et al., 2022). However, with respect to the inferences about deep history produced by AGs themselves, our results highlight the importance of caution in proposing specific models of population history that relate a set of groups. Acknowledgements We thank Anders Bergström, Esther Brielle, Mateja Hajdinjak, Iosif Lazaridis, Pablo Librado, Mark Lipson, Vagheesh Narasimhan, Ludovic Orlando, Nick Patterson, Mary Prendergast, Jakob Sedig, Kendra Sirak, Pontus Skoglund, and Chuanchao Wang, for suggestions for how to improve specific analyses, and for conversations and critical comments. We thank Matthew Mah, Shop Mallick, Adam Micco, Nadin Rohland, Ron Pinhasi for help in generating additional data from an ancient DNA library from individual I8726 for which 1.24 million SNP capture data was generated and published in Narasimhan et al., 2019 and for which we report 2.6- fold shotgun data here (Supplementary file 2). P.F., P.C., O.F., and U.I. were supported by the Czech Ministry of Education, Youth and Sports (program ERC CZ, project no. LL2103). P.F., P.C., and O.F. were supported by the Czech Science Foundation (project no. 21- 27624S). P.F. and P.C. were also supported by the Czech Ministry of Education, Youth and Sports: the "Large Infrastructures for Research, Experimental Develop- ment and Innovations" program (project "IT4Innovations National Supercomputing Center" no. LM2015070) and the Inter- Excellence program (project no. LTAUSA18153). R.M. and D.R. were supported by grants from the National Institutes of Health (GM100233 and HG012287), the John Templeton Foundation (grant 61220), and the Allen Discovery Center program, a Paul G Allen Fron- tiers Group advised program of the Paul G Allen Family Foundation. D.R. and P.F. were supported by private gifts from Jean- Francois Clin. D.R. is an Investigator of the Howard Hughes Medical Institute. Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 24 of 62 Evolutionary Biology | Genetics and Genomics Research article Additional information Funding Funder Grant reference number Author Czech Ministry of Education, Youth and Sports LL2103 Czech Ministry of Education, Youth and Sports Czech Ministry of Education, Youth and Sports National Institutes of Health National Institutes of Health John Templeton Foundation The Czech Science Foundation LM2015070 LTAUSA18153 GM100233 HG012287 grant 61220 21-27624S Pavel Flegontov Piya Changmai Olga Flegontova Ulaş Işıldak Pavel Flegontov Piya Changmai Pavel Flegontov Piya Changmai Robert Maier David Reich Robert Maier David Reich Robert Maier David Reich Pavel Flegontov Piya Changmai Olga Flegontova The funders had no role in study design, data collection, and interpretation, or the decision to submit the work for publication. Author contributions Robert Maier, Conceptualization, Resources, Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing – original draft, Writing – review and editing; Pavel Flegontov, Conceptualization, Resources, Data curation, Formal analysis, Supervision, Funding acqui- sition, Validation, Investigation, Visualization, Writing – original draft, Project administration, Writing – review and editing, Methodology; Olga Flegontova, Formal analysis; Ulaş Işıldak, Formal analysis, Visualization; Piya Changmai, Formal analysis, Funding acquisition; David Reich, Conceptualization, Resources, Supervision, Funding acquisition, Project administration, Writing – review and editing Author ORCIDs Robert Maier Pavel Flegontov Ulaş Işıldak David Reich http://orcid.org/0000-0002-3044-090X http://orcid.org/0000-0001-9759-4981 http://orcid.org/0000-0001-6497-6254 http://orcid.org/0000-0002-7037-5292 Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.85492.sa1 Author response https://doi.org/10.7554/eLife.85492.sa2 Additional files Supplementary files •  Supplementary file 1. Published graphs in the context of automatically found graphs. We compared 22 different graphs from 8 publications to alternative graphs inferred on the same or very similar data; these findGraphs runs are highlighted in blue in the ‘Iterations’ column. In total, 51 findGraphs runs are summarized here since in some cases models more complex or less complex than the published one were explored and/or different population compositions were tested (see the ‘Topology search constraints and population modifications’ column and the footnotes for details). The columns with names in blue show various information on the published graphs or their modified versions and some properties of the published population sets. The columns with names in Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 25 of 62 Evolutionary Biology | Genetics and Genomics Research article magenta show settings used for calculating f- statistics and for exploring the AG space, and the number of SNPs used that depends on them. The columns with names in black summarize the outcomes of findGraphs runs, that is, the properties of alternative model sets found. Publication: Last name of the first author and year of the relevant publication. Figure in the original publication: Figure number in the original paper where the AG is presented. Groups (populations): The number of populations in each graph. Singleton pseudo- haploid populations: The number of populations in the graph composed of a single pseudo- haploid individual. Calculation of negative ‘admixture’ f3- statistics is impossible for such populations since their heterozygosity cannot be estimated (see the text for details). No. of negative f3- stats (allsnps: YES): The number of negative f3- statistics among all possible f3- statistics for a given set of populations when all available sites are used for each statistic. If no negative f3- statistics exist for a set of populations, AG fits are not affected by the ‘minac2=2’ setting intended for accurate calculation of f- statistics for non- singleton pseudo- haploid groups. Admixture events: The number of admixture events in each graph. Publ. model: log- likelihood (LL): Log- likelihood score of the published graph fitted to the SNP set shown in the ‘SNPs used’ column. Publ. model: LL, median of bootstrap distr.: Median of the log- likelihood scores of 100 or 500 fits of the published graph using bootstrap resampled SNPs. Publ. model: worst residual (WR), SE: The worst f- statistic residual of the published graph fitted to the SNP set shown in the ‘SNPs used’ column, measured in standard errors (SE). SNPs used: The number of SNPs (with no missing data at the group level) used for fitting the AG. For all case studies, we tested the original data (SNPs, population composition, and the published graph topology) and obtained model fits very similar to the published ones. However, for the purpose of efficient topology search we adjusted settings for f3- statistic calculation, population composition, or graph complexity as shown here and discussed in the text. Settings for calculating f2- statistics: Arguments of the extract_f2 function used for calculating all possible f2- statistics for a set of groups, which were then used by findGraphs for calculating f3- statistics needed for fitting AG models. See Appendix 1, Section 2.A for descriptions of each argument. Topology search constraints and population modifications: Constraints applied when generating random starting graphs and/or when searching the topology space. Modifications of the original population composition are also described in this column, where applicable. Iterations: The number of findGraphs iterations (runs), each started from a random graph of a certain complexity. For each case study, findGraphs setups that were considered optimal are highlighted in blue in this column. Iterations confirming published graph: The number of iterations (runs) in which the resulting graph was topologically identical to the published graph. In the cases, when the published model was irrelevant since more complex graphs were explored, ‘N/A’ appears in this and subsequent columns. If less complex models were explored, the published model was still relevant since its version without selected admixture edges was tested. Distinct alternative topologies found: The number of distinct newly found topologies. If graph complexity was equal to (or less than) that of the published graph, the published topology (or its simplified version) is not counted here. If graph complexity exceeded that of the published graph, all newly found topologies are counted. If the published topology was recovered by findGraphs, the numbers in this column are shown in bold. Significantly better fitting topologies: The number of distinct topologies that fit significantly better than the published graph according to the bootstrap model comparison test (two- tailed empirical p- value <0.05). If the number of distinct topologies was very large, a representative sample of models (1/20 to 1/3 of models evenly distributed along the log- likelihood spectrum) was compared to the published one instead. These cases are marked as ‘a fraction of models tested’ in this column. If model complexity was higher than that of the published model, model comparison was irrelevant and was not performed. Non- significantly better fitting topologies: The number of distinct topologies that fit non- significantly (nominally) better than the published graph according to the bootstrap model comparison test (two- tailed empirical p- value ≥0.05). Non- significantly worse fitting topologies: The number of distinct topologies that fit non- significantly (nominally) worse than the published graph according to the bootstrap model comparison test (two- tailed empirical p- value ≥0.05). Significantly worse fitting topologies: The number of distinct topologies that fit significantly worse than the published graph according to the bootstrap model comparison test (two- tailed empirical p- value <0.05). Significantly better fitting topologies, %: The percentage of distinct topologies that fit significantly better than the published graph according to the bootstrap model comparison test (two- tailed empirical p- value <0.05). If the number of distinct topologies was very large, a representative sample of models (1/20 to 1/3 of models evenly distributed along the log- likelihood spectrum) was compared to the published one instead, and the percentages in this and following columns were calculated on this sample. Non- significantly better fitting topologies, %: The percentage of distinct topologies that fit non- significantly (nominally) better than the published Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 26 of 62 Evolutionary Biology | Genetics and Genomics Research article graph according to the bootstrap model comparison test (two- tailed empirical p- value ≥0.05). Non- significantly worse fitting topologies, %: The percentage of distinct topologies that fit non- significantly (nominally) worse than the published graph according to the bootstrap model comparison test (two- tailed empirical p- value ≥0.05). Significantly worse fitting topologies, %: The percentage of distinct topologies that fit significantly worse than the published graph according to the bootstrap model comparison test (two- tailed empirical p- value <0.05). p- value best alternative vs. publ.: An empirical two- tailed p- value of a test comparing log- likelihood distributions across bootstrap replicates for two topologies, the highest- ranking newly found topology and the published topology. In some cases, the highest- ranking newly found topology (according to LL) has a fit that is not significantly better than that of the published model, but other newly found models fit significantly better despite having higher LL. P- values below 0.05 are highlighted in green. Used in Table 1: Here the findGraphs runs featured in Table 1 are marked. •  Supplementary file 2. Statistics for shotgun sequencing of individual I8726. •  Supplementary file 3. Group labels, archaeological and geographic meta- data, and sequencing statistics for the individuals used for admixture graph fitting in Shinde et al., 2019. See the ‘Comment’ column for a list of updates to the dataset composition we performed prior to admixture graph fitting in our study. •  Supplementary file 4. Group labels, archaeological and geographic meta- data, and sequencing statistics for the individuals used for qpAdm modelling that revisits qpAdm results by Narasimhan et al., 2019. •  Supplementary file 5. A summary of selected qpAdm results. This is a re- analysis of the data from Narasimhan et al., 2019, with a modified group composition that is described in Appendix 2. •  Supplementary file 6. qpAdm results for separate individuals from selected target groups (Indus Periphery and others). qpAdm p- values are shown for each individual for models of varying complexity, from one- to four- way models. •  Supplementary file 7. Comparison of the original group composition used for admixture graph fitting in Librado et al., 2021 and the modified groups used in our study. •  MDAR checklist Data availability The new software presented in this manuscript (the ADMIXTOOLS 2 R package) is freely avail- able at https://github.com/uqrmaie1/admixtools (copy archived at Maier et  al., 2022), along with a detailed manual at https://uqrmaie1.github.io/admixtools/. The ancient human genome newly reported in this manuscript (Supplementary file 2) is freely available at the European Nucleotide Archive in the form of an alignment of reads to the hg19 human reference genome (project acces- sion number PRJEB58199). Published software packages re- used in this manuscript are available at: https://bitbucket.org/nygcresearch/treemix/src/master/ (TreeMix, Pickrell and Pritchard, 2012) and at https://github.com/DReichLab/AdmixTools (David Reich Lab, 2023, ADMIXTOOLS, Patterson et al., 2012). Published archaeogenetic datasets re- analyzed in this manuscript were kindly shared by the corresponding authors of the following publications upon our requests: Bergström et al., 2020; Lazaridis et al., 2014; Librado et al., 2021; Lipson et al., 2020b; Shinde et al., 2019; Sikora et al., 2019; Wang et  al., 2021; Hajdinjak et  al., 2021. Various statistics for these re- used datasets are summarized in Supplementary file 1. The following dataset was generated: Author(s) Maier R, Flegontov P, Flegontova O, Işıldak U, Changmai P, Reich D Year 2022 Dataset title Dataset URL Database and Identifier On the limits of fitting complex models of population history to f- statistics https://www. ebi. ac. uk/ ena/ browser/ view/ PRJEB58199 European Nucleotide Archive, PRJEB58199 Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 27 of 62 Evolutionary Biology | Genetics and Genomics Research article References Baumdicker F, Bisschop G, Goldstein D, Gower G, Ragsdale AP, Tsambos G, Zhu S, Eldon B, Ellerman EC, Galloway JG, Gladstein AL, Gorjanc G, Guo B, Jeffery B, Kretzschumar WW, Lohse K, Matschiner M, Nelson D, Pope NS, Quinto- Cortés CD, et al. 2022. Efficient ancestry and mutation simulation with msprime 1.0. Genetics 220:iyab229. DOI: https://doi.org/10.1093/genetics/iyab229, PMID: 34897427 Bellwood P. 2011. The checkered prehistory of rice movement southwards as a domesticated cereal—from the yangzi to the equator. Rice 4:93–103. DOI: https://doi.org/10.1007/s12284-011-9068-9 Bergström A, Frantz L, Schmidt R, Ersmark E, Lebrasseur O, Girdland- Flink L, Lin AT, Storå J, Sjögren KG, Anthony D, Antipina E, Amiri S, Bar- Oz G, Bazaliiskii VI, Bulatović J, Brown D, Carmagnini A, Davy T, Fedorov S, Fiore I, et al. 2020. Origins and genetic legacy of prehistoric dogs. Science 370:557–564. DOI: https://doi.org/ 10.1126/science.aba9572, PMID: 33122379 Bergström A, Stanton DWG, Taron UH, Frantz L, Sinding M- HS, Ersmark E, Pfrengle S, Cassatt- Johnstone M, Lebrasseur O, Girdland- Flink L, Fernandes DM, Ollivier M, Speidel L, Gopalakrishnan S, Westbury MV, Ramos- Madrigal J, Feuerborn TR, Reiter E, Gretzinger J, Münzel SC, et al. 2022. Grey wolf genomic history reveals a dual ancestry of dogs. Nature 607:313–320. DOI: https://doi.org/10.1038/s41586-022-04824-9 Boos DD. 2003. Introduction to the bootstrap world. Statistical Science 18:168–174. DOI: https://doi.org/10. 1214/ss/1063994971 Border R, Johnson EC, Evans LM, Smolen A, Berley N, Sullivan PF, Keller MC. 2019. No support for historical candidate gene or candidate gene- by- interaction hypotheses for major depression across multiple large samples. The American Journal of Psychiatry 176:376–387. DOI: https://doi.org/10.1176/appi.ajp.2018. 18070881, PMID: 30845820 Campbell MC, Tishkoff SA. 2008. African genetic diversity: implications for human demographic history, modern human origins, and complex disease mapping. Annual Review of Genomics and Human Genetics 9:403–433. DOI: https://doi.org/10.1146/annurev.genom.9.081307.164258 Carlhoff S, Duli A, Nägele K, Nur M, Skov L, Sumantri I, Oktaviana AA, Hakim B, Burhan B, Syahdar FA, McGahan DP, Bulbeck D, Perston YL, Newman K, Saiful AM, Ririmasse M, Chia S, Pulubuhu DAT, Jeong C, Peter BM, et al. 2021. Genome of a middle Holocene Hunter- Gatherer from Wallacea. Nature 596:543–547. DOI: https://doi.org/10.1038/s41586-021-03823-6, PMID: 34433944 Chen FH, Dong GH, Zhang DJ, Liu XY, Jia X, An CB, Ma MM, Xie YW, Barton L, Ren XY, Zhao ZJ, Wu XH, Jones MK. 2015. Agriculture facilitated permanent human occupation of the Tibetan Plateau after 3600 B.P. Science 347:248–250. DOI: https://doi.org/10.1126/science.1259172, PMID: 25593179 Collins AL, Kim Y, Sklar P, O’Donovan MC, Sullivan PF, International Schizophrenia Consortium. 2012. Hypothesis- driven candidate genes for schizophrenia compared to genome- wide association results. Psychological Medicine 42:607–616. DOI: https://doi.org/10.1017/S0033291711001607, PMID: 21854684 David Reich Lab. 2023. Admixtools. ed1d857. GitHub. https://github.com/DReichLab/AdmixTools Duncan LE, Ostacher M, Ballon J. 2019. How genome- wide association studies (GWAS) made traditional candidate gene studies obsolete. Neuropsychopharmacology 44:1518–1523. DOI: https://doi.org/10.1038/ s41386-019-0389-5, PMID: 30982060 Excoffier L, Dupanloup I, Huerta- Sánchez E, Sousa VC, Foll M. 2013. Robust demographic inference from genomic and SNP data. PLOS Genetics 9:e1003905. DOI: https://doi.org/10.1371/journal.pgen.1003905, PMID: 24204310 Flegontov P, Altınışık NE, Changmai P, Rohland N, Mallick S, Adamski N, Bolnick DA, Broomandkhoshbacht N, Candilio F, Culleton BJ, Flegontova O, Friesen TM, Jeong C, Harper TK, Keating D, Kennett DJ, Kim AM, Lamnidis TC, Lawson AM, Olalde I, et al. 2019. Palaeo- eskimo genetic ancestry and the peopling of chukotka and North America. Nature 570:236–240. DOI: https://doi.org/10. 1038/s41586-019-1251-y, PMID: 31168094 Fu Q, Posth C, Hajdinjak M, Petr M, Mallick S, Fernandes D, Furtwängler A, Haak W, Meyer M, Mittnik A, Nickel B, Peltzer A, Rohland N, Slon V, Talamo S, Lazaridis I, Lipson M, Mathieson I, Schiffels S, Skoglund P, et al. 2016. The genetic history of ice age Europe. Nature 534:200–205. DOI: https://doi.org/10.1038/ nature17993, PMID: 27135931 Gray RD, Jordan FM. 2000. Language trees support the express- train sequence of austronesian expansion. Nature 405:1052–1055. DOI: https://doi.org/10.1038/35016575, PMID: 10890445 Green RE, Krause J, Briggs AW, Maricic T, Stenzel U, Kircher M, Patterson N, Li H, Zhai W, Fritz MH- Y, Hansen NF, Durand EY, Malaspinas A- S, Jensen JD, Marques- Bonet T, Alkan C, Prüfer K, Meyer M, Burbano HA, Good JM, et al. 2010. A draft sequence of the Neandertal genome. Science 328:710–722. DOI: https://doi.org/10.1126/science.1188021 Gronau I, Hubisz MJ, Gulko B, Danko CG, Siepel A. 2011. Bayesian inference of ancient human demography from individual genome sequences. Nature Genetics 43:1031–1034. DOI: https://doi.org/10.1038/ng.937, PMID: 21926973 Gutenkunst RN, Hernandez RD, Williamson SH, Bustamante CD. 2009. Inferring the joint demographic history of multiple populations from multidimensional SNP frequency data. PLOS Genetics 5:e1000695. DOI: https://doi. org/10.1371/journal.pgen.1000695, PMID: 19851460 Haak W, Lazaridis I, Patterson N, Rohland N, Mallick S, Llamas B, Brandt G, Nordenfelt S, Harney E, Stewardson K, Fu Q, Mittnik A, Bánffy E, Economou C, Francken M, Friederich S, Pena RG, Hallgren F, Khartanovich V, Khokhlov A, et al. 2015. Massive migration from the steppe was a source for Indo- European languages in Europe. Nature 522:207–211. DOI: https://doi.org/10.1038/nature14317, PMID: 25731166 Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 28 of 62 Evolutionary Biology | Genetics and Genomics Research article Hajdinjak M, Mafessoni F, Skov L, Vernot B, Hübner A, Fu Q, Essel E, Nagel S, Nickel B, Richter J, Moldovan OT, Constantin S, Endarova E, Zahariev N, Spasov R, Welker F, Smith GM, Sinet- Mathiot V, Paskulin L, Fewlass H, et al. 2021. Initial upper Palaeolithic humans in Europe had recent neanderthal ancestry. Nature 592:253–257. DOI: https://doi.org/10.1038/s41586-021-03335-3 Hammer MF, Woerner AE, Mendez FL, Watkins JC, Wall JD. 2011. Genetic evidence for archaic admixture in Africa. PNAS 108:15123–15128. DOI: https://doi.org/10.1073/pnas.1109300108, PMID: 21896735 Harney É, Patterson N, Reich D, Wakeley J. 2021. Assessing the performance of qpAdm: a statistical tool for studying population Admixture. Genetics 217:iyaa045. DOI: https://doi.org/10.1093/genetics/iyaa045, PMID: 33772284 Hirschhorn JN, Daly MJ. 2005. Genome- Wide association studies for common diseases and complex traits. Nature Reviews. Genetics 6:95–108. DOI: https://doi.org/10.1038/nrg1521, PMID: 15716906 Hubisz M, Siepel A. 2020. Inference of ancestral recombination graphs using argweaver. Methods in Molecular Biology 2090:231–266. DOI: https://doi.org/10.1007/978-1-0716-0199-0_10, PMID: 31975170 Hubisz MJ, Williams AL, Siepel A. 2020. Mapping gene flow between ancient hominins through demography- aware inference of the ancestral recombination graph. PLOS Genetics 16:e1008895. DOI: https://doi.org/10. 1371/journal.pgen.1008895, PMID: 32760067 Ioannidis JPA. 2005. Why most published research findings are false. PLOS Medicine 2:e124. DOI: https://doi. org/10.1371/journal.pmed.0020124, PMID: 16060722 Jeong C, Balanovsky O, Lukianova E, Kahbatkyzy N, Flegontov P, Zaporozhchenko V, Immel A, Wang CC, Ixan O, Khussainova E, Bekmanov B, Zaibert V, Lavryashina M, Pocheshkhova E, Yusupov Y, Agdzhoyan A, Koshel S, Bukin A, Nymadawa P, Turdikulova S, et al. 2019. The genetic history of admixture across inner Eurasia. Nature Ecology & Evolution 3:966–976. DOI: https://doi.org/10.1038/s41559-019-0878-2, PMID: 31036896 Kamm J, Terhorst J, Durbin R, Song YS. 2020. Efficiently inferring the demographic history of many populations with allele count data. Journal of the American Statistical Association 115:1472–1487. DOI: https://doi.org/10. 1080/01621459.2019.1635482, PMID: 33012903 Ko AM- S, Chen C- Y, Fu Q, Delfin F, Li M, Chiu H- L, Stoneking M, Ko Y- C. 2014. Early Austronesians: into and out of Taiwan. American Journal of Human Genetics 94:426–436. DOI: https://doi.org/10.1016/j.ajhg.2014.02.003, PMID: 24607387 Kutanan W, Liu D, Kampuansai J, Srikummool M, Srithawong S, Shoocongdej R, Sangkhano S, Ruangchai S, Pittayaporn P, Arias L, Stoneking M. 2021. Reconstructing the human genetic history of mainland Southeast Asia: insights from genome- wide data from Thailand and laos. Molecular Biology and Evolution 38:3459–3477. DOI: https://doi.org/10.1093/molbev/msab124, PMID: 33905512 Lachance J, Vernot B, Elbers CC, Ferwerda B, Froment A, Bodo JM, Lema G, Fu W, Nyambo TB, Rebbeck TR, Zhang K, Akey JM, Tishkoff SA. 2012. Evolutionary history and adaptation from high- coverage whole- genome sequences of diverse African hunter- gatherers. Cell 150:457–469. DOI: https://doi.org/10.1016/j.cell.2012.07. 009, PMID: 22840920 Lazaridis I, Patterson N, Mittnik A, Renaud G, Mallick S, Kirsanow K, Sudmant PH, Schraiber JG, Castellano S, Lipson M, Berger B, Economou C, Bollongino R, Fu Q, Bos KI, Nordenfelt S, Li H, de Filippo C, Prüfer K, Sawyer S, et al. 2014. Ancient human Genomes suggest three ancestral populations for present- day Europeans. Nature 513:409–413. DOI: https://doi.org/10.1038/nature13673, PMID: 25230663 Lazaridis I, Nadel D, Rollefson G, Merrett DC, Rohland N, Mallick S, Fernandes D, Novak M, Gamarra B, Sirak K, Connell S, Stewardson K, Harney E, Fu Q, Gonzalez- Fortes G, Jones ER, Roodenberg SA, Lengyel G, Bocquentin F, Gasparian B, et al. 2016. Genomic insights into the origin of farming in the ancient near East. Nature 536:419–424. DOI: https://doi.org/10.1038/nature19310, PMID: 27459054 Leppälä K, Nielsen SV, Mailund T. 2017. Admixturegraph: an R package for admixture graph manipulation and fitting. Bioinformatics 33:1738–1740. DOI: https://doi.org/10.1093/bioinformatics/btx048, PMID: 28158333 Librado P, Khan N, Fages A, Kusliy MA, Suchan T, Tonasso- Calvière L, Schiavinato S, Alioglu D, Fromentier A, Perdereau A, Aury J- M, Gaunitz C, Chauvey L, Seguin- Orlando A, Der Sarkissian C, Southon J, Shapiro B, Tishkin AA, Kovalev AA, Alquraishi S, et al. 2021. The origins and spread of domestic horses from the Western Eurasian steppes. Nature 598:634–640. DOI: https://doi.org/10.1038/s41586-021-04018-9, PMID: 34671162 Lipson M, Loh PR, Levin A, Reich D, Patterson N, Berger B. 2013. Efficient moment- based inference of admixture parameters and sources of gene flow. Molecular Biology and Evolution 30:1788–1802. DOI: https://doi.org/10. 1093/molbev/mst099, PMID: 23709261 Lipson M, Szécsényi- Nagy A, Mallick S, Pósa A, Stégmár B, Keerl V, Rohland N, Stewardson K, Ferry M, Michel M, Oppenheimer J, Broomandkhoshbacht N, Harney E, Nordenfelt S, Llamas B, Gusztáv Mende B, Köhler K, Oross K, Bondár M, Marton T, et al. 2017. Parallel palaeogenomic transects reveal complex genetic history of early European farmers. Nature 551:368–372. DOI: https://doi.org/10.1038/nature24476, PMID: 29144465 Lipson M. 2020a. Applying f4- statistics and admixture graphs: theory and examples. Molecular Ecology Resources 20:1658–1667. DOI: https://doi.org/10.1111/1755-0998.13230 Lipson M, Ribot I, Mallick S, Rohland N, Olalde I, Adamski N, Broomandkhoshbacht N, Lawson AM, López S, Oppenheimer J, Stewardson K, Asombang RN, Bocherens H, Bradman N, Culleton BJ, Cornelissen E, Crevecoeur I, de Maret P, Fomine FLM, Lavachery P, et al. 2020b. Ancient West African foragers in the context of African population history. Nature 577:665–670. DOI: https://doi.org/10.1038/s41586-020-1929-1, PMID: 31969706 Lipson M, Sawchuk EA, Thompson JC, Oppenheimer J, Tryon CA, Ranhorn KL, de Luna KM, Sirak KA, Olalde I, Ambrose SH, Arthur JW, Arthur KJW, Ayodo G, Bertacchi A, Cerezo- Román JI, Culleton BJ, Curtis MC, Davis J, Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 29 of 62 Evolutionary Biology | Genetics and Genomics Research article Gidna AO, Hanson A, et al. 2022. Ancient DNA and deep population structure in sub- Saharan African foragers. Nature 603:290–296. DOI: https://doi.org/10.1038/s41586-022-04430-9, PMID: 35197631 Lu D, Lou H, Yuan K, Wang X, Wang Y, Zhang C, Lu Y, Yang X, Deng L, Zhou Y, Feng Q, Hu Y, Ding Q, Yang Y, Li S, Jin L, Guan Y, Su B, Kang L, Xu S. 2016. Ancestral origins and genetic history of Tibetan highlanders. American Journal of Human Genetics 99:580–594. DOI: https://doi.org/10.1016/j.ajhg.2016.07.002, PMID: 27569548 Maier R, Isildak U, rmmaier. 2022. Admixtools. swh:1:rev:45fcc41768d88559863d6dfae9ccaceaf247f81e. Software Heritage. https://archive.softwareheritage.org/swh:1:dir:38fceb428451354c86a1c41fa7291ea1 17a15d2b;origin=https://github.com/uqrmaie1/admixtools;visit=swh:1:snp:ec5ab16d631d19a1943b1f874db5 a4df0cee6ced;anchor=swh:1:rev:45fcc41768d88559863d6dfae9ccaceaf247f81e Mallick S, Li H, Lipson M, Mathieson I, Gymrek M, Racimo F, Zhao M, Chennagiri N, Nordenfelt S, Tandon A, Skoglund P, Lazaridis I, Sankararaman S, Fu Q, Rohland N, Renaud G, Erlich Y, Willems T, Gallo C, Spence JP, et al. 2016. The simons genome diversity project: 300 genomes from 142 diverse populations. Nature 538:201–206. DOI: https://doi.org/10.1038/nature18964 Mathieson I, Alpaslan- Roodenberg S, Posth C, Szécsényi- Nagy A, Rohland N, Mallick S, Olalde I, Broomandkhoshbacht N, Candilio F, Cheronet O, Fernandes D, Ferry M, Gamarra B, Fortes GG, Haak W, Harney E, Jones E, Keating D, Krause- Kyora B, Kucukkalipci I, et al. 2018. The genomic history of southeastern Europe. Nature 555:197–203. DOI: https://doi.org/10.1038/nature25778, PMID: 29466330 McColl H, Racimo F, Vinner L, Demeter F, Gakuhari T, Moreno- Mayar JV, van Driem G, Gram Wilken U, Seguin- Orlando A, de la Fuente Castro C, Wasef S, Shoocongdej R, Souksavatdy V, Sayavongkhamdy T, Saidin MM, Allentoft ME, Sato T, Malaspinas A- S, Aghakhanian FA, Korneliussen T, et al. 2018. The prehistoric peopling of Southeast Asia. Science 361:88–92. DOI: https://doi.org/10.1126/science.aat3628, PMID: 29976827 Molloy EK, Durvasula A, Sankararaman S. 2021. Advancing admixture graph estimation via maximum likelihood network orientation. Bioinformatics 37:i142–i150. DOI: https://doi.org/10.1093/bioinformatics/btab267, PMID: 34252951 Moreno- Mayar JV, Vinner L, de Barros Damgaard P, de la Fuente C, Chan J, Spence JP, Allentoft ME, Vimala T, Racimo F, Pinotti T, Rasmussen S, Margaryan A, Iraeta Orbegozo M, Mylopotamitaki D, Wooller M, Bataille C, Becerra- Valdivia L, Chivall D, Comeskey D, Devièse T, et al. 2018. Early human dispersals within the Americas. Science 362:2621. DOI: https://doi.org/10.1126/science.aav2621 Narasimhan VM, Patterson N, Moorjani P, Rohland N, Bernardos R, Mallick S, Lazaridis I, Nakatsuka N, Olalde I, Lipson M, Kim AM, Olivieri LM, Coppa A, Vidale M, Mallory J, Moiseyev V, Kitov E, Monge J, Adamski N, Alex N, et al. 2019. The formation of human populations in South and central Asia. Science 365:7487. DOI: https://doi.org/10.1126/science.aat7487 Nelson D, Kelleher J, Ragsdale AP, Moreau C, McVean G, Gravel S. 2020. Accounting for long- range correlations in genome- wide simulations of large cohorts. PLOS Genetics 16:e1008619. DOI: https://doi.org/10.1371/ journal.pgen.1008619, PMID: 32369493 Nielsen SV, Vaughn AH, Leppälä K, Landis MJ, Mailund T, Nielsen R. 2023. Bayesian inference of Admixture graphs on native American and Arctic populations. PLOS Genetics 1:1010410. DOI: https://doi.org/10.1371/ journal.pgen.1010410 Patterson N, Moorjani P, Luo Y, Mallick S, Rohland N, Zhan Y, Genschoreck T, Webster T, Reich D. 2012. Ancient admixture in human history. Genetics 192:1065–1093. DOI: https://doi.org/10.1534/genetics.112.145037 Peter BM. 2016. Admixture, population structure, and F- statistics. Genetics 202:1485–1501. DOI: https://doi. org/10.1534/genetics.115.183913, PMID: 26857625 Pickrell JK, Pritchard JK. 2012. Inference of population splits and mixtures from genome- wide allele frequency data. PLOS Genetics 8:e1002967. DOI: https://doi.org/10.1371/journal.pgen.1002967, PMID: 23166502 Posth C, Nakatsuka N, Lazaridis I, Skoglund P, Mallick S, Lamnidis TC, Rohland N, Nägele K, Adamski N, Bertolini E, Broomandkhoshbacht N, Cooper A, Culleton BJ, Ferraz T, Ferry M, Furtwängler A, Haak W, Harkins K, Harper TK, Hünemeier T, et al. 2018. Reconstructing the deep population history of central and South America. Cell 175:1185–1197.. DOI: https://doi.org/10.1016/j.cell.2018.10.027, PMID: 30415837 Prüfer K, Racimo F, Patterson N, Jay F, Sankararaman S, Sawyer S, Heinze A, Renaud G, Sudmant PH, de Filippo C, Li H, Mallick S, Dannemann M, Fu Q, Kircher M, Kuhlwilm M, Lachmann M, Meyer M, Ongyerth M, Siebauer M, et al. 2014. The complete genome sequence of a Neanderthal from the altai mountains. Nature 505:43–49. DOI: https://doi.org/10.1038/nature12886 Raghavan M, Skoglund P, Graf KE, Metspalu M, Albrechtsen A, Moltke I, Rasmussen S, Stafford TW, Orlando L, Metspalu E, Karmin M, Tambets K, Rootsi S, Mägi R, Campos PF, Balanovska E, Balanovsky O, Khusnutdinova E, Litvinov S, Osipova LP, et al. 2014. Upper Palaeolithic Siberian genome reveals dual ancestry of native Americans. Nature 505:87–91. DOI: https://doi.org/10.1038/nature12736, PMID: 24256729 Raghavan M, Steinrücken M, Harris K, Schiffels S, Rasmussen S, DeGiorgio M, Albrechtsen A, Valdiosera C, Ávila- Arcos MC, Malaspinas A- S, Eriksson A, Moltke I, Metspalu M, Homburger JR, Wall J, Cornejo OE, Moreno- Mayar JV, Korneliussen TS, Pierre T, Rasmussen M, et al. 2015. Population genetics: genomic evidence for the Pleistocene and recent population history of native Americans. Science 349:aab3884. DOI: https://doi. org/10.1126/science.aab3884, PMID: 26198033 Reich D, Thangaraj K, Patterson N, Price AL, Singh L. 2009. Reconstructing Indian population history. Nature 461:489–494. DOI: https://doi.org/10.1038/nature08365, PMID: 19779445 Reich D, Patterson N, Kircher M, Delfin F, Nandineni MR, Pugach I, Ko AM- S, Ko Y- C, Jinam TA, Phipps ME, Saitou N, Wollstein A, Kayser M, Pääbo S, Stoneking M. 2011. Denisova admixture and the first modern human Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 30 of 62 Evolutionary Biology | Genetics and Genomics Research article dispersals into Southeast Asia and Oceania. The American Journal of Human Genetics 89:516–528. DOI: https://doi.org/10.1016/j.ajhg.2011.09.005 Reich D, Patterson N, Campbell D, Tandon A, Mazieres S, Ray N, Parra MV, Rojas W, Duque C, Mesa N, García LF, Triana O, Blair S, Maestre A, Dib JC, Bravi CM, Bailliet G, Corach D, Hünemeier T, Bortolini MC, et al. 2012. Ruiz- linares a. Reconstructing Native American Population History. Nature 488:370–374. DOI: https:// doi.org/10.1038/nature11258 Rogers AR. 2019. Legofit: estimating population history from genetic data. BMC Bioinformatics 20:526. DOI: https://doi.org/10.1186/s12859-019-3154-1, PMID: 31660852 Scally A, Durbin R. 2012. Revising the human mutation rate: implications for understanding human evolution. Nature Reviews. Genetics 13:745–753. DOI: https://doi.org/10.1038/nrg3295, PMID: 22965354 Schiffels S, Durbin R. 2014. Inferring human population size and separation history from multiple genome sequences. Nature Genetics 46:919–925. DOI: https://doi.org/10.1038/ng.3015, PMID: 24952747 Schiffels S, Haak W, Paajanen P, Llamas B, Popescu E, Loe L, Clarke R, Lyons A, Mortimer R, Sayer D, Tyler- Smith C, Cooper A, Durbin R. 2016. Iron age and anglo- saxon genomes from East England reveal British migration history. Nature Communications 7:10408. DOI: https://doi.org/10.1038/ncomms10408, PMID: 26783965 Seguin- Orlando A, Korneliussen TS, Sikora M, Malaspinas A- S, Manica A, Moltke I, Albrechtsen A, Ko A, Margaryan A, Moiseyev V, Goebel T, Westaway M, Lambert D, Khartanovich V, Wall JD, Nigst PR, Foley RA, Lahr MM, Nielsen R, Orlando L, et al. 2014. Paleogenomics: genomic structure in Europeans dating back at least 36,200 years. Science 346:1113–1118. DOI: https://doi.org/10.1126/science.aaa0114, PMID: 25378462 Shinde V, Narasimhan VM, Rohland N, Mallick S, Mah M, Lipson M, Nakatsuka N, Adamski N, Broomandkhoshbacht N, Ferry M, Lawson AM, Michel M, Oppenheimer J, Stewardson K, Jadhav N, Kim YJ, Chatterjee M, Munshi A, Panyam A, Waghmare P, et al. 2019. An ancient harappan genome lacks ancestry from steppe pastoralists or Iranian farmers. Cell 179:729–735.. DOI: https://doi.org/10.1016/j.cell.2019.08.048, PMID: 31495572 Sikora M, Pitulko VV, Sousa VC, Allentoft ME, Vinner L, Rasmussen S, Margaryan A, de Barros Damgaard P, de la Fuente C, Renaud G, Yang MA, Fu Q, Dupanloup I, Giampoudakis K, Nogués- Bravo D, Rahbek C, Kroonen G, Peyrot M, McColl H, Vasilyev SV, et al. 2019. The population history of northeastern Siberia since the Pleistocene. Nature 570:182–188. DOI: https://doi.org/10.1038/s41586-019-1279-z Skoglund P, Posth C, Sirak K, Spriggs M, Valentin F, Bedford S, Clark GR, Reepmeyer C, Petchey F, Fernandes D, Fu Q, Harney E, Lipson M, Mallick S, Novak M, Rohland N, Stewardson K, Abdullah S, Cox MP, Friedlaender FR, et al. 2016. Genomic insights into the peopling of the southwest Pacific. Nature 538:510–513. DOI: https://doi. org/10.1038/nature19844, PMID: 27698418 Soraggi S, Wiuf C. 2019. General theory for stochastic admixture graphs and F- statistics. Theoretical Population Biology 125:56–66. DOI: https://doi.org/10.1016/j.tpb.2018.12.002, PMID: 30562538 Speidel L, Forest M, Shi S, Myers SR. 2019. A method for genome- wide genealogy estimation for thousands of samples. Nature Genetics 51:1321–1329. DOI: https://doi.org/10.1038/s41588-019-0484-x, PMID: 31477933 Tambets K, Yunusbayev B, Hudjashov G, Ilumäe AM, Rootsi S, Honkola T, Vesakoski O, Atkinson Q, Skoglund P, Kushniarevich A, Litvinov S, Reidla M, Metspalu E, Saag L, Rantanen T, Karmin M, Parik J, Zhadanov SI, Gubina M, Damba LD, et al. 2018. Genes reveal traces of common recent demographic history for most of the uralic- speaking populations. Genome Biology 19:139. DOI: https://doi.org/10.1186/s13059-018-1522-1, PMID: 30241495 Terhorst J, Kamm JA, Song YS. 2017. Robust and scalable inference of population history from hundreds of unphased whole genomes. Nature Genetics 49:303–309. DOI: https://doi.org/10.1038/ng.3748, PMID: 28024154 Vallini L, Marciani G, Aneli S, Bortolini E, Benazzi S, Pievani T, Pagani L. 2022. Genetics and material culture support repeated expansions into paleolithic Eurasia from a population hub out of Africa. Genome Biology and Evolution 14:evac045. DOI: https://doi.org/10.1093/gbe/evac045, PMID: 35445261 van de Loosdrecht M, Bouzouggar A, Humphrey L, Posth C, Barton N, Aximu- Petri A, Nickel B, Nagel S, Talbi EH, El Hajraoui MA, Amzazi S, Hublin J- J, Pääbo S, Schiffels S, Meyer M, Haak W, Jeong C, Krause J. 2018. Pleistocene North African genomes link near eastern and sub- Saharan African human populations. Science 360:548–552. DOI: https://doi.org/10.1126/science.aar8380, PMID: 29545507 Wang CC, Reinhold S, Kalmykov A, Wissgott A, Brandt G, Jeong C, Cheronet O, Ferry M, Harney E, Keating D, Mallick S, Rohland N, Stewardson K, Kantorovich AR, Maslov VE, Petrenko VG, Erlikh VR, Atabiev BC, Magomedov RG, Kohl PL, et al. 2019. Ancient human genome- wide data from a 3000- year interval in the Caucasus corresponds with eco- geographic regions. Nature Communications 10:590. DOI: https://doi.org/10. 1038/s41467-018-08220-8, PMID: 30713341 Wang C- C, Yeh H- Y, Popov AN, Zhang H- Q, Matsumura H, Sirak K, Cheronet O, Kovalev A, Rohland N, Kim AM, Mallick S, Bernardos R, Tumen D, Zhao J, Liu Y- C, Liu J- Y, Mah M, Wang K, Zhang Z, Adamski N, et al. 2021. Genomic insights into the formation of human populations in East Asia. Nature 591:413–419. DOI: https://doi. org/10.1038/s41586-021-03336-2 Yan J, Patterson N, Narasimhan VM. 2021. MiqoGraph: fitting admixture graphs using mixed- integer quadratic optimization. Bioinformatics 37:2488–2490. DOI: https://doi.org/10.1093/bioinformatics/btaa988, PMID: 33247708 Yang M.A, Gao X, Theunert C, Tong H, Aximu- Petri A, Nickel B, Slatkin M, Meyer M, Pääbo S, Kelso J, Fu Q. 2017. 40,000- year- old individual from Asia provides insight into early population structure in Eurasia. Current Biology 27:3202–3208.. DOI: https://doi.org/10.1016/j.cub.2017.09.030, PMID: 29033327 Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 31 of 62 Evolutionary Biology | Genetics and Genomics Research article Yang MA, Fan X, Sun B, Chen C, Lang J, Ko Y- C, Tsang C, Chiu H, Wang T, Bao Q, Wu X, Hajdinjak M, Ko AM- S, Ding M, Cao P, Yang R, Liu F, Nickel B, Dai Q, Feng X, et al. 2020. Ancient DNA indicates human population shifts and admixture in northern and southern China. Science 369:282–288. DOI: https://doi.org/10.1126/ science.aba0909 Zhang M, Yan S, Pan W, Jin L. 2019. Phylogenetic evidence for sino- tibetan origin in northern China in the late neolithic. Nature 569:112–115. DOI: https://doi.org/10.1038/s41586-019-1153-z Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 32 of 62 Evolutionary Biology | Genetics and Genomics Research article Appendix 1 1. ADMIXTOOLS 2. A new software package and set of algorithmic ideas for fitting allele frequency correlation statistics to genetic data relating populations We present a new implementation of the popular ADMIXTOOLS software (called ‘Classic ADMIXTOOLS’) (Patterson et al., 2012; Haak et al., 2015). Our implementation (ADMIXTOOLS 2) enhances performance by greatly reducing runtime and memory requirements across a wide range of different methods, relative to Classic ADMIXTOOLS (Appendix  1—figure 1a). Some of these improvements have now been implemented in version 7.0.2 of ADMIXTOOLS (https://github.com/ DReichLab/AdmixTools/). The present study focuses not on the performance differences between Classic ADMIXTOOLS and ADMIXTOOLS 2, but on the description of new ideas implemented in one or both of these tools. 1.A. Computation and use of f-statistics A key idea that facilitates the performance increases shared by ADMIXTOOLS 2 and ADMIXTOOLS v. 7.0.2 is that any f- statistic (which form the basis of almost all ADMIXTOOLS programs as well as other toolkits for studying population history such as popstats) can be computed from a small number of f2- statistics. For most f- statistic- based analyses (for example qpWave, qpAdm, and qpGraph; Appendix  1—figure 1c), the time required to compute f3- and f4- statistics algebraically from f2- statistics is trivial compared to the time required to load genotype data and compute f2- statistics. These f2- statistics can be stored and re- used to compute f3- and f4- statistics, thus reducing the size of the input data, runtime, and memory requirements by orders of magnitude (Appendix 1—figure 1a, d). Using precomputed f2- statistics is not always the best solution. In datasets with large amounts of missing data, computing f3- and f4- statistics from precomputed f2- statistics may introduce bias. In this case, it is necessary to compute f3- and f4- statistics directly, using different SNPs for each f- statistic (all available SNPs in each population triplet or quadruplet). However, even without the use of precomputed f2- statistics, ADMIXTOOLS 2 often achieves large performance gains (Appendix 1— figure 1a). The program qpfstats in Classic ADMIXTOOLS implements an idea which strikes a balance between these two extremes. It increases the accuracy of estimation of f- statistics by using a regression approach to jointly estimate the values of all f2-, f3-, and f4- statistics relating a set of populations. Specifically, qpfstats searches for values of these statistics that are not only consistent with information from the SNPs that have data in the groups used to compute each particular f- statistic, but also satisfy the algebraic relationships expected with other f- statistics (thus incorporating information from data at many additional SNPs). See further details on this algorithm at https:// github.com/DReichLab/AdmixTools/blob/master/qpfs.pdf, (Patterson et al., 2012). This feature is available in ADMIXTOOLS 2 through the qpfstats option in the extract_f2 function. Another improvement introduced in ADMIXTOOLS 2 relates to accurate evaluation of the match between observed and expected f3- statistics when fitting AGs where at least one population is represented by a single individual with genotypes derived by randomly selecting one sequencing read at each variable position (‘pseudo- haploid’ data). f- Statistic computations need to be modified when analyzing pseudo- haploid data, because heterozygosity cannot be computed using comparisons of sequences within the same individual; however, computation of heterozygosity is essential to calculate ‘admixture’ f3- statistics f3(target; A, B), where negative values provide proof of the mixed nature of the target population. When a target population is represented by multiple individuals, unbiased estimation of admixture f3- statistics can be carried out even for pseudo- haploid data by analyzing positions covered by sequences from at least two individuals and only computing variation rates across individuals. This approach is implemented in Classic ADMIXTOOLS with the ‘inbreed: YES’ option. However, no admixture f3- statistic can be computed with this algorithm if the target population is represented by a single individual (as no variation across individuals within a population can be detected in this case). Classic ADMIXTOOLS deals with this case by failing to run if any population in an analysis is represented by a single individual and the ‘inbreed: YES’ option is turned on. Because the datasets from all the AGs revisited here included at least one population represented by a single individual (Supplementary file 1), the ‘inbreed: YES’ option could not be Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 33 of 62 Evolutionary Biology | Genetics and Genomics Research article used in the original studies (the program failed with this option, by design). Thus, AG fitting in those studies relied on the incorrect algorithm for calculating f3- statistics (except for Librado et al., 2021, which used TreeMix instead of qpGraph) and, as a result, some f3- statistics that are negative and could provide important constraints for AG fitting were evaluated as positive. These concerns are relevant for the Shinde et al., 2019, Lipson et al., 2020b, and Wang et al., 2021 studies we revisit below (see Supplementary file 1 for a list of datasets where negative f3- statistics were encountered). To be able to detect negative f3- statistics and thus take advantage of their power for constraining the space of possibly fitting historical models, in ADMIXTOOLS 2 we introduced a similar algorithm which makes it possible to compute negative f3- statistics on pseudo- haploid data, at a cost of removing sites with only one chromosome genotyped in any population that is represented by at least two individuals (so that it is possible in theory to compute heterozygosity in these populations). Admixture f3- statistics continue to be incorrectly computed using ADMIXTOOLS 2 for targets that are singleton populations represented by pseudo- haploid data, as there is no avoiding this particular problem. Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 34 of 62 Evolutionary Biology | Genetics and Genomics Research article Appendix 1—figure 1. Performance comparison of f- statistic computation and AG fitting in Classic ADMIXTOOLS and ADMIXTOOLS 2 and an overview of the major ADMIXTOOLS programs. (a) Performance comparison of f- statistic computation and AG fitting. Top: Memory usage and runtime for computing f- statistics using (1) the qpDstat program in ADMIXTOOLS v7.0.2 released in 06/2021, (2) the f4 function in ADMIXTOOLS 2 without precomputing f2- statistics, and (3) the f4 function in ADMIXTOOLS 2 with precomputed f2- statistics. (1) and (2) Appendix 1—figure 1 continued on next page Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 35 of 62 Evolutionary Biology | Genetics and Genomics Research article Appendix 1—figure 1 continued give identical results, whereas (3) only gives identical results in the absence of missing data, which limits its usefulness beyond a moderate number of populations. Bottom: Runtime comparison of qpGraph with and without precomputed f- statistics. (b) Illustration of f2- and f4- statistics. f2 measures the amount of drift separating any two populations, while f4 measures the amount of drift shared between two population pairs. Every f4- statistic is a linear combination of four f2- statistics. (c) Overview of the major ADMIXTOOLS programs, their primary use cases, and their associated f- statistics. (d) Schematic representation of the computations behind the ADMIXTOOLS programs qpGraph, qpWave, and qpAdm. ADMIXTOOLS 2 separates the computation of f2- statistics from the later steps in the pipeline. Shown below are the number of data points for N individuals, M SNPs, and k populations. The exact number of all possible non- redundant f2-, f3-, and f4- statistics for k populations are . A small number of f2- statistics can be used to obtain a much larger number of f3- and f4- statistics and require much less ) ( storage space than the raw genotype data. 1 , and 3 ) k 1 , 2 2 ) k 3 ( k 4 ( 1.B. AG fitting, model comparison, and interpretation There are several challenges that arise when modeling the ancestral relationships among populations with AGs, and ADMIXTOOLS 2 implements solutions to several key problems that were not adequately addressed with previous approaches: (1) Automated AG inference; (2) Estimating confidence intervals for AG parameters; (3) Comparing fits of different AGs; (4) Determining identifiability of AG parameters; and (5) Drawing conclusions from a large number of fitting graphs. Here, we describe these challenges, how we address them, and how our approaches compare to other approaches, while the section ‘Technical presentation of ADMIXTOOLS 2 in the context of methods based on f- statistics’ gives detailed descriptions. 1.B.1. Automated AG inference Constructing AGs manually runs the risk of overlooking models that challenge conventional hypotheses. On the other hand, current methods for inferring AGs automatically (Leppälä et al., 2017; Molloy et al., 2021; Pickrell and Pritchard, 2012; Yan et al., 2021) do not allow external information to be integrated into the analysis, and often result in models that may fit the genetic data but can be rejected on other grounds. In addition, TreeMix (Pickrell and Pritchard, 2012), as well as OrientAGraph (Molloy et  al., 2021), an improved version of TreeMix, can miss AG topologies that exist on parts of the non- convex likelihood surface that are bypassed by these algorithms for exploring AGs (e.g., topology M7 in Figure 4 of Molloy et al., 2021). MixMapper (Lipson et  al., 2013) and miqoGraph Yan et  al., 2021 have a different limitation: exploring topologies with more than one admixture event in the history of any group is not possible. Due to these limitations, many published findings are based on manual proposal of topologies and evaluation of fit, and the great majority of studies using this manual approach (see, e.g., Reich et al., 2011; Reich et al., 2012; Lazaridis et al., 2014; Fu et al., 2016; Skoglund et al., 2016; Yang et al., 2017; McColl et al., 2018; Moreno- Mayar et al., 2018; Tambets et al., 2018; van de Loosdrecht et  al., 2018; Flegontov et  al., 2019; Sikora et  al., 2019; Wang et  al., 2019; Lipson et al., 2020b; Shinde et al., 2019; Yang et al., 2020; Hajdinjak et al., 2021; Wang et al., 2021; Bergström et  al., 2022) rely on the software qpGraph. We introduce an approach for finding well- fitting AGs automatically that can integrate external information, and that recovers graph topologies more accurately than TreeMix (Appendix  1—figure 2). External information can be integrated by specifying a set of constraints that AGs must satisfy. This not only ensures that resulting models are temporally plausible, but also cleanly separates prior assumptions from the independent constraints provided by genetic data. Our strategy implemented in the function ‘findGraphs’, differs from TreeMix/OrientAGraph in several deep ways, most notably in that it optimizes graphs directly, rather than optimizing trees first and adding admixture events later. This makes it less prone to getting stuck in local optima: our simulation results show that findGraphs is more accurate for random graphs (Appendix  1—figure 2), and that it can recover specific topologies that pose problems for TreeMix and OrientAGraph. 1.B.2. Estimating confidence intervals for AG parameters Since our new implementation of qpGraph can evaluate models much more rapidly, it becomes feasible to evaluate the same model multiple times on different SNP sets. This allows us to derive bootstrap confidence intervals (Boos, 2003) for all parameters estimated by qpGraph, including drift lengths, admixture weights, LL scores, and f4- statistic residuals. It should also be noted that the estimated confidence intervals do not take into account uncertainty about the graph topology. Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 36 of 62 Evolutionary Biology | Genetics and Genomics Research article 1.B.3. Comparing the fits of different AGs Using the bootstrap method for evaluating a graph multiple times on different SNP sets not only allows us to obtain confidence intervals for single graphs, but also allows us to test whether the fit of one graph is significantly better than the fit of another graph, by obtaining confidence intervals for the LL score difference or difference in the largest f4- statistic residuals (worst residuals, WR) of two graphs. When we apply this approach to a range of datasets, we find that models with modest LL differences are often not distinguishable after accounting for the variability across SNPs, even if one might expect them to be distinguishable based on the magnitude of the likelihood difference (Appendix 1—figure 3a, b). Thus, previous methods relying on AIC or BIC (such as Shinde et al., 2019; Flegontov et al., 2019) that used specified likelihood difference thresholds to reject some models over others, were over- aggressive. These methods are problematic since they generally assume that the models compared have the same effective number of degrees of freedom, but the number of independent parameters estimated in an AG is not simply determined by the number of groups, drift edges, or admixture events, as it also depends on the graph topology in a complex way. A second challenge in comparing different AG models arises when comparing models of different complexity (i.e., with a different number of admixture events). Established methods such as AIC and BIC can also account for different model complexity, but if the number of independent parameters in a model is known. We implement a method to compare AG models of different complexity by using a new scoring function, which uses different blocks of SNPs for deriving fitted and estimated f- statistics. This ensures that our model comparison test does not favor more complex models by allowing them to overfit the data. This cross- validation approach can also be used to rank alternative models of the same complexity and deal with overfitting. We note that the calculation of cross- validated LL scores is not turned on in findGraphs by default, and to make our results more comparable to those of the published studies we revisited, we relied on standard LL scores in this study. To test if our method is well calibrated, we simulated 100,000 SNPs under the same graph in 1,000 replicates. We then created two new topologies by removing one out of two symmetric edges from the first graph (Appendix 1—figure 3c). These new incorrect models are symmetrically related to the first graph and can be used to test whether the true difference in LL scores of these two graphs is zero. The uniform distribution of p- values confirms that our method is well calibrated (Appendix 1—figure 3d). A caveat is that only one symmetric topology was explored in this way. 1.B.4. Determining identifiability of AG parameters Fitting AGs results in an estimate of the overall model fit, as well as in estimates of branch lengths and admixture weights. However, even with infinite data some of these parameters cannot be estimated, as they are not identifiable from the system of equations that corresponds to the AG. Issues like this have been well described for simple topological features of a graph. For example, the lengths of the two branches connected to the root node cannot be estimated independently. Furthermore, in a graph with n populations and a admixture events, at least one parameter will not be identifiable unless n 2   the inequality  >= 2n + 2a − is satisfied (Lipson, 2020a). However, even in graphs that meet 3  this criterion, some parameters are not identifiable, and until the development of ADMIXTOOLS 2, there was no method for testing whether any given parameter in an AG is identifiable. We introduce a method for testing which parameters in an AG are identifiable, and which are not, based on the Jacobi matrix of the graph’s system of f2 equations. Like our method for deriving confidence intervals for AG parameters, this can improve the interpretability of AG analyses. Our methods for automated topology inference, for bootstrapping LL scores or worst residuals for comparing model fits, for cross- validation of AGs of different complexities, for estimating confidence intervals and determining identifiability of AG parameters can greatly improve the interpretability of AG analyses. We implement all these methods in ADMIXTOOLS 2 to assist the user in accurately testing a series of admixture models for ancient populations. 1.B.5. Drawing conclusions from a large number of fitted models As discussed in the Results section, when we apply our methods for finding optimal graphs and comparing AGs to a number of previously published models, we find that there often exists a much larger number of fitting models than has previously been appreciated. In these cases, we are unable to prioritize a single model, or even a small number of models, based on the evidence we Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 37 of 62 Evolutionary Biology | Genetics and Genomics Research article have. However, we are still able to reject the vast majority of all tested models. This suggests that insight can be gained by identifying common features among the well- fitting models. We therefore introduce methods for summarizing collections of well- fitting AGs to determine which features they share. In practice, we find that these methods can aid manual inspection of findGraphs results, but the high diversity of well- fitting topologies we see in most case studies and the importance of fitted parameters (especially admixture proportions) for historical interpretation of topologies makes it difficult to reliably automatize the process of interpreting fitted AG models. 2. Technical presentation of ADMIXTOOLS 2 in the context of methods based on f-statistics Much of the content that follows recapitulates theory presented in previous work, notably Reich et al., 2009, Green et al., 2010, and Patterson et al., 2012, but we summarize it here for coherence. 2.A. f-Statistics All ADMIXTOOLS programs are based on the statistics f2, f3, and f4, for population pairs, triplets, and quadruples, respectively. j 2 bj = 1 M f2 quantifies the genetic drift separating two populations A and B . For a single SNP, it is given by , where aj and bj are the allele frequencies for SNP j in populations A and f2 A, B B. When allele frequencies are estimated using a small number of samples, this estimator of f2 will be biased upwards. An unbiased estimator of f2 is given by aj aj 1 − nA,j− 1 − ( ) , where nA,j and nB,j are the observed allele counts in bj bj 1 − nB,j− 1 ( ) aj − ( f2 = 1 M ∑ − ) ( ) 2 j ( bj f3 aj − is the covariance of the allele frequency differences between populations A and B, and the allele frequency differences between populations A and C (assuming ) that alleles are coded randomly, so that a – b and a – c are both 0 in expectation). Significantly negative values of f3(A; B, C) suggest that A is a mixture of sources related to B and C (although the converse does not hold: A might be admixed between B and C even if f3 is positive). aj − ) ( ∑ cj ) ( bj aj − populations A and B. ∑ A; B, C ( = 1 M ) j = 1 M j ( bj f4 aj − A, B; C, D is the covariance of the allele frequency differences between A and B, and the allele frequency differences between C and D. Significantly positive values ) of f4(A, B; C, D) (or equivalently significantly negative values of f4(A, B; D, C)) reveal that A and B do not form a clade with respect to C and D, and that some of the drift separating A from C is shared with the drift separating B from D. cj − ) ( ∑ dj ( ) f3 and f4 can be written as linear combinations of f2- statistics: f3 A; B, C = 1 2 f2 A, B + f2 A, C f2 B, C − ( f4 A, B; C, D ) = 1 2 ( ( A, D ) + f2 ( B, C f2 ) f2 ( A, C )) f2 − − B, D (1) (2) ( This implies that all f3- and f4- statistics can be computed from f2- statistics as long as they are )) ( ( ( ( ( ) ) ) ) defined on the same SNPs. For revisiting published studies, we used the ‘extract_f2’ function with the ‘maxmiss’ argument set at 0, which corresponds to the ‘useallsnps: NO’ setting in classic ADMIXTOOLS. It means that no missing data are allowed (at the level of populations) in the specified set of populations for which pairwise f2- statistics are calculated. For the values of the ‘blgsize’, ‘adjust_pseudohaploid’, and ‘minac2’ arguments we use in our analyses, see Supplementary file 1. The ‘blgsize’ argument sets the SNP block size in Morgans, and we used either the default value of 0.05 (5 cM), or 4,000,000 bp when a genetic map was not available. Genotypes of pseudo- haploid samples are usually coded as 0 or 2 (i.e., they are, strictly speaking, pseudo- diploid), even though only one allele is observed. The ‘adjust_pseudohaploid’ argument ensures that the observed allele count increases only by 1 for each pseudo- haploid sample. If ‘TRUE’ (default), samples that do not have any genotypes coded as 1 among the first 1,000 SNPs are automatically identified as pseudo- haploid. This leads to slightly more accurate estimates of f- statistics. Setting this parameter to ‘FALSE’ treats all samples as diploid. Another important argument (‘minac2=2’) of the ‘extract_f2’ function removes sites with only one chromosome genotyped in any non- singleton population and is needed for unbiased estimation of negative f3- statistics in non- singleton pseudo- haploid populations. In the absence of negative Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 38 of 62 Evolutionary Biology | Genetics and Genomics Research article f3- statistics or pseudo- haploid populations, this argument has no influence on AG LL scores. This algorithm for calculation of f- statistics triggered by the ‘minac2=2’ argument is described below. For f3(A; B, C), we compute the uncorrected numerator for each SNP, (a – b) × (a – c). We then subtract a bias correction factor at each SNP, p(1 – p)/(ac – 1), which we only need for population a (because the other factors cancel out); p is the allele frequency, and ac is the observed allele count. In pseudo- haploid samples, (ac – 1) would be zero and produce an error in any sites with only one observed allele. With the ‘inbreed: NO’ setting in Classic ADMIXTOOLS, the smallest non- zero value for ac is 2, so the division by 0 problem is avoided, but the correction factor is slightly smaller than it should be. ADMIXTOOLS 2 adds only an allele count of 1 for each site in a pseudo- haploid sample (with the default option ‘adjust_pseudohaploid = TRUE’), so there can be cases where ac = 1. To imitate what the setting ‘inbreed: NO’ in Classic ADMIXTOOLS is doing, ac is set to 2 at those sites (or the denominator is set to 1). There is still a small difference between Classic ADMIXTOOLS and ADMIXTOOLS 2 at other sites because each observed site adds two alleles in ADMIXTOOLS with the default setting ‘inbreed: NO’, but only 1 allele in ADMIXTOOLS 2 with the default setting ‘adjust_pseudohaploid = TRUE’, but for AG fitting that does not matter. One solution to avoid biased correction factors is to only consider sites with ac of at least two, which is what the ‘inbreed: YES’ setting in Classic ADMIXTOOLS does. The problem with this is that we cannot use populations with a single pseudo- haploid sample, which is often useful, and would only give misleading results if that population is admixed. The new option ‘minac2=2’ in ADMIXTOOLS 2 is different from the ‘inbreed: YES’ setting in Classic ADMIXTOOLS since it makes an exception for populations consisting of a single pseudo- haploid sample in that it sets ac to 2 at each site (denominator is set to 1) when computing the correction factor for those populations. 2.B. Fitting AGs An AG is a directed acyclic graph specifying the topology of the ancestral relationships among a set of populations. Each node in this graph represents a (present- day or ancient) population. Terminal nodes (also called leaf nodes) represent observed populations, while internal nodes represent unobserved ancestral populations. Modeling all observed populations as leaf nodes confers some robustness to drift specific to single populations and to genotyping errors. The edges connecting the populations are weighted and correspond either to the magnitude of genetic drift that has occurred along that branch (drift edges), or to the admixture proportions (admixture edges, where two edges point to the same node). The goal of qpGraph is to test how well a given graph topology fits the observed f- statistics. This is achieved by varying the edge weights until the maximum likelihood fit is obtained. The following section describes the graph fitting in more detail. k+1 2 k First, for k populations, all f3- statistics of the form f3(O; X1, X2) are computed, where O is one of the k populations (typically an outgroup), and X1 and X2 are all pairs formed from the other populations (including pairs where X1 = X2). These f3- statistics can then be used to fit the graph and to compute the likelihood. The likelihood score of a graph is the dot product of the differences between the expected and observed f3- statistics, weighted by the inverse covariance matrix of f3- statistics: ( ) L g = 1 2 − f3, obs − f3, fit ′ Q− 1 f3, obs − f3, fit ( ) ( ) ( ) (3) k k+1 2 Here, f3, obs are the observed f3- statistics and f3, fit are the fitted f3- statistics. Both are vectors of for k populations excluding the outgroup. Q is the q × q covariance matrix of f3- length q = statistics, where the diagonal entries are the f3- statistic variances, and the off- diagonal entries are the covariances for all pairs of f3- statistics. Just like the variances (the squared standard errors), the covariances are estimated from the jackknife leave- one- block- out f3- statistics. ) ( Finding the edge weights which maximize the likelihood score involves two nested optimization steps. The inner optimization finds the drift weights which maximize the likelihood score while fixing the admixture weights. The outer optimization finds the admixture weights which maximize the likelihood score, while optimizing the drift weights for each set of admixture weights. The inner optimization uses a quadratic programming solver to find the optimal drift weights, while the outer optimization uses a general purpose optimization algorithm to find the optimal admixture weights. Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 39 of 62 Evolutionary Biology | Genetics and Genomics Research article While the gradient function in the outer optimization adjusts the admixture weights, the objective function iterates over the following steps: 1. Optimization of drift weights conditional on admixture weights 2. Estimation of fitted f3- statistics 3. Calculation of the graph likelihood using observed and fitted f3- statistics These steps are repeated until convergence is reached and the likelihood score can no longer be improved by adjusting the admixture weights. Step 1 optimizes the drift edge weights, while holding the admixture weights constant. All drift edge weights are required to be non- negative, which makes this a constrained quadratic programming problem (hence qpGraph). Additional upper and lower bounds can be specified for individual graph edges. Step 2 turns the edge weights into fitted f3- statistics. To see how edge weights in an AG translate to f3- statistics, it helps to first consider how they translate into f2- statistics for a pair of populations. Without any admixture events, there is exactly one path p connecting any two populations. The fitted f2- statistic (f2, fit) is the sum of edge weights we along this path p connecting two populations. The fitted f2- statistic is the sum of edge weights we along this path: f2, fit = we e p ∈ ∑ In the presence of admixture events, two populations may be connected via multiple paths. Each admixture node that lies between the two populations increases the number of possible paths. The fitted f2- statistic for the two populations now becomes the weighted sum of all these paths, where the weight of each path is given by the product of all estimated admixture proportions wa along this path ( ∏ p wa ): ∈ a f2, fit = p P ∈ ∑ a p ∈ ∏ wa we e p ∈ ∑ (4) The fitted f2- statistics are then used to obtain fitted f3- statistics using Equation 1. Step 3 uses the fitted and observed f3- statistics to estimate the likelihood score using Equation 3. Prior to these three steps, initial admixture weights are drawn randomly. To ensure that the end results do not depend on the random initialization, the whole optimization process is repeated multiple times with different random initial values. The original ADMIXTOOLS implementation retains only the results from the initial values resulting in the lowest absolute likelihood score. The new ADMIXTOOLS 2 implementation provides an option to retrieve the results for all random initializations. This can be useful, as large fluctuations between different random initializations can be an indicator of an overparameterized or otherwise poorly fitting model. 2.C. Automated AG inference To find graph topologies that could conceivably have given rise to the observed f- statistics, we start with a randomly generated graph with a fixed number of admixture events, apply a number of modifications to this graph, and evaluate each of the resulting graphs. We then pick the best- fitting graph and repeat this procedure until graph modifications no longer lead to improved scores. We use a number of random graph modifications, as well as targeted modifications which are informed by parameters obtained during the fitting of the current graph. For the targeted modifications, we change the optimization of a single graph from a constrained optimization problem, in which drift edges are constrained to be positive and admixture weights are constrained to be between zero and one, to an unconstrained optimization problem in which both types of parameters can take any real values. Rearranging the nodes adjacent to edges which were estimated to be negative results in an improved fit at a much higher rate than random graph adjustments. The random modifications include (1) pruning and randomly re- grafting leaf nodes, (2) pruning and randomly re- grafting a set of connected nodes in the graph, (3) swapping the orientation of admixture edges, (4) shifting admixture edges, (5) re- rooting the graph, and (6) combinations of two or more of any of these modifications. Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 40 of 62 Evolutionary Biology | Genetics and Genomics Research article The number of admixture events is not affected by the graph modifications described so far. A significant score improvement can often be achieved by adding a single admixture edge to several random positions in a graph. This is unsurprising since it increases the degrees of freedom of the original graph. However, picking the best fitting graph with one admixture edge added, and testing all graphs that result from removing a single admixture edge from that graph, often results in a graph with the same number of admixture events and a better fit than the original graph. We employ this strategy whenever the regular graph modifications described above do not lead to any further improvements. We keep track of the search tree of all previously evaluated graphs and their scores in order to not evaluate any graph more than once, and so that backtracking in the search space is possible in cases where no more local improvements can be identified. Nevertheless, multiple iterations with different random starting graphs are usually necessary to find graphs with good fits. The number of iterations needed to approach a global optimum depends on the size of the search space, but the optimal number of iterations is hard to estimate in practice. For revisiting published studies, we used the following settings of the findGraphs algorithm: •  mutfuns = namedList(spr_leaves, spr_all, swap_leaves, move_admixedge_once, flipadmix_ random, place_root_random, mutate_n), a list of functions used to modify graphs. •  numgraphs = 10, number of alternative graphs produced by randomly applying the mutation functions at the start of each generation. •  stop_gen = 10,000, total number of generations after which to stop. •  stop_gen2=30, number of generations without LL score improvement after which to stop. •  plusminus_generations = 10. If the best score does not improve after plusminus_generations generations, another approach to improving the score is attempted: A number of graphs with an additional admixture event is generated and evaluated. The resulting graph with the best score is picked, and new graphs are created by removing any one admixture event (bringing the number back to what it was originally). The graph with the lowest score is then selected. This approach often makes it possible to break out of local optima. •  opt_worst_residual = FALSE. Optimize for lowest worst residual instead of best score. ‘FALSE’ by default, because the LL score is generally a better indicator of the quality of the model fit, and because optimizing for the lowest worst residual is much slower since f4- statistics need to be computed for each graph. •  reject_f4z=0. If this is a number greater than zero, all f4- statistics with |Z- score|>reject_f4z will be used to constrain the search space of AGs: Any graphs in which Z- scores greater than reject_f4z are expected to be zero will not be evaluated. •  diag = 1e−04. This argument is passed to the qpgraph function and determines the regulari- zation term added to the diagonal elements of the covariance matrix of fitted branch lengths (after scaling by the matrix trace). Default is 0.0001. •  numstart = 10. This argument is passed to the qpgraph function and determines the number of random initializations of starting weights (defaults to 10). Increasing this number will make the optimization slower but reduce the risk of not finding the optimal weights. lsqmode = FALSE. This argument is passed to the qpgraph function. If set to ‘FALSE’, the inverse f3- statistic covariance matrix is not discarded by the algorithm. • The arguments ‘admix_constraints’ (constraints on the number of admixture events in the history of a given population), ‘event_constraints’ (constraints on the branching order of specified lineages), and ‘outpop’ (the population assigned as an outgroup) were set according to Supplementary file 1. Each findGraphs run was initiated by a random graph with a specified number of admixture events. Usually, the same topology constraints were applied at the stage of random graph generation and the topology search stage, for exceptions see Supplementary file 1. 2.D. Evaluating automated AG inference through simulations We evaluated the performance of findGraphs by simulating genetic data under a large number of different AG models, applying findGraphs to each simulated dataset in three independent iterations, and comparing the resulting best graph across three iterations to the simulated graph. We applied TreeMix to the same simulated data for comparison. We simulated between 8 and 16 populations per graph, and between 0 and 10 admixture events. For each parameter combination, we simulated 20 different AGs generated by the random_admixturegraph function. We counted both the fraction Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 41 of 62 Evolutionary Biology | Genetics and Genomics Research article of random simulated graphs where the best inferred graph was identical to the simulated graph, as well as the fraction of random simulated graphs where the best inferred graph was either identical to the simulated graph or had a better score than the simulated graph. For models with a large number of admixture events the number of possible models is so large that it becomes increasingly likely that there will be some alternative models which fit the data better than the model under which the data were simulated. We used msprime v.0.7.4 and the msprime_sim wrapper function in ADMIXTOOLS 2 to simulate data for 100,000 unlinked SNPs and 100 diploid samples per population for each AG. The simulation parameters we chose were aimed at facilitating fast simulations of large numbers of informative SNPs rather than at being as realistic as possible. We therefore expect that the simulation results allow us to make comparisons across groups, but not that they are informative about the rate at which ‘true’ models can be recovered in empirical data. We simulated under a constant mutation rate of 0.001 per site per generation, a constant haploid effective population size of 1,000, with neighboring nodes in the graph separated by 1,000 or more generations, and all admixture events occurring in discrete pulses of 50%/50% proportions. To allow for a fair comparison between findGraphs and TreeMix, we made sure that small differences in the way AGs are modeled in findGraphs and in TreeMix were accounted for before testing graphs for identical topology. For example, TreeMix AGs can have lineages terminating at an admixture node, whereas in findGraphs lineages always end at a ‘leaf’ node with a single ancestor. More realistic simulations were performed with msprime v.1.1.1 which allows accurate simulation of recombination and of multi- chromosome diploid genomes relying on the Wright–Fisher model (Nelson et al., 2020, Baumdicker et al., 2022). We simulated three chromosomes (each 100 Mbp long) in a diploid genome by specifying a flat recombination rate (2 × 10−8 per bp per generation) along the chromosome and a much higher rate at the chromosome boundaries (loge2 or ~0.693 per bp per generation, see https://tskit.dev/msprime/docs/stable/ancestry.html#multiple-chromosomes). A flat mutation rate, 1.25 × 10−8 per bp per generation (Scally and Durbin, 2012), and the binary mutation model were used. To maintain the correct correlation between chromosomes, the discrete time Wright–Fischer model was used for 25 generations into the past, and deeper in the past the standard coalescent simulation algorithm was used (as recommended by Nelson et al., 2020). We simulated AGs of four complexity classes: eight or nine groups sampled at leaves, four or five pulse- like admixture events. All group sizes were identical: 10 diploid individuals with no missing data. Demographic events were separated by date intervals ranging randomly between 1,500 and 8,000 generations, with an upper bound on the tree time depth at 40,000 generations. Effective population sizes were constant along each edge, and were picked randomly from the range of 2,000–40,000 diploid individuals. Admixture proportions for all admixture events varied randomly between 10% and 40%. For subsequent analyses we selected only simulations where pairwise FST for groups were in the range characteristic for anatomically modern and archaic humans (there was at least one FST value below 0.15). In this way, 20 random topologies were simulated per graph complexity class, each including also a distant outgroup that facilitates exploration of the topology space. The outgroup diverged at 40,000 generations ago and had a constant diploid population size of 100,000 individuals. Since there was no missing data and all individuals were diploid, we first calculated all possible f2- statistics for 4 Mbp- sized genome blocks (with the ‘maxmiss = 0’, ‘adjust_pseudohaploid = FALSE’, and ‘minac2=FALSE’ settings) and then used them for calculating f4- statistics as linear combinations of f2- statistics or for fitting AGs (with the ‘numstart = 100’ and ‘diag = 0.0001’ settings). 2.E. Comparing the fits of different AGs We are interested in determining whether one AG fits the data significantly better than another AG, or whether an observed score difference Δ = S1 – S2 can be attributed to variability across independent SNPs. We first consider two AGs with the same number of admixture events, where we can ignore the problem of comparing two models with different complexity. As in other bootstrap standard error calculations, we divide the genome into n blocks indexed by i, and we draw b sets of n blocks with replacement, indexed by j. We fit both graphs b times—once for each bootstrap set of SNP blocks. This results in a set of b score differences Δj. The bootstrap confidence interval for the difference in scores is given by the quantiles of the distribution of Δj. We also compute an empirical bootstrap p- value, testing the null hypothesis that two different graphs fit the data Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 42 of 62 Evolutionary Biology | Genetics and Genomics Research article 1 b , 2δ p = max equally well. It is computed as (Boos, 2003), where δ is either the fraction of Δj > 0, or the fraction of Δj < 0, whichever is smaller. The reason for applying bootstrap resampling, as ) opposed to jackknife resampling in this case, is that the distribution of score differences tends to have a high kurtosis, which can make jackknife estimates inaccurate. Simulating data under the null hypothesis is not straightforward in this case, because it involves finding two non- identical graphs which in expectation fit the data equally well. We decided to simulate under one graph and compare two graphs which are symmetrically related to the simulated graph (Appendix  1—figure 3). This confirmed that the p- value follows a uniform distribution under the null hypothesis. ( g L Next, we consider comparisons of two graphs of different complexity. The problem here is that more complex graphs have more degrees of freedom which allow them to overfit the data better, without necessarily being any closer to the truth. To solve this problem, we introduce an out- of- sample likelihood score. The regular likelihood score is given by: ′ Q− , with f3, obs and f3, fit defined on the same set of SNPs. The out- of- sample likelihood score is defined in the same way, except that f3, obs and f3, fit are ) defined on mutually exclusive sets of SNP blocks, thereby preventing any overfitting. The covariance matrix Q is defined on the same set of SNP blocks as f3, fit. As described earlier, we use block- bootstrap to fit both graphs multiple times on different SNP blocks. In each bootstrap iteration, we use all SNP blocks which are not used in fitting the graph for estimating f3, obs. f3, obs − f3, obs − f3, fit f3, fit − = 1 2 ( ) ) ( ( 1 2.F. AG identifiability An edge in an AG is unidentifiable, if small changes to the weight of this edge (admixture proportions in the case of an admixture edge, drift length in the case of a drift edge) do not necessarily lead to changes in expected f- statistics. This is the case if the small change in weight can be offset by small changes in other graph edges, leading to a situation where observed f- statistics can be explained by more than one weight estimate for that edge. To find unidentifiable edges, we derive the Jacobi matrix of the graph’s system of f2 equations (Equation 4 applied to each population pair). In principle, whether a parameter is identifiable can depend on the values of all other parameters. However, in practice this is rarely the case, and so we draw values for all parameters from a uniform distribution, which gives us a Jacobi matrix with numeric values. We then determine the rank of the Jacobi matrix, along with the rank of all matrices that result from dropping a single column (a parameter corresponding to a graph edge). For identifiable edges, the rank of the full matrix will be greater than the rank of the reduced matrix, and for unidentifiable edges, the ranks will be the same. 2.G. Drawing conclusions from a large number of fitting models We developed several methods that aim to summarize a collection of graphs which all fit the data similarly well. By highlighting features which are observed repeatedly across graphs, it becomes possible to extract interpretable conclusions from an otherwise hard to interpret collection of possible models. These graph summaries identify features in each graph that can be compared to different graphs describing the same populations. We summarize each graph in several ways: 1. Admixture status of each population For each population, we count the total number of admixture events that are encountered along all paths from the chosen leaf to the root. 2. Order of population split events For each pair of population pairs, we determine if the most recent split of the first pair has occurred before or after the most recent split of the second pair, or whether the graph does not specify the order in which those splits occurred. 3. Proxy populations For each admixed population in a graph, we attempt to identify proxy sources: populations closest to the admixing populations. In contrast to the other approaches to summarizing graphs which are based only on the topology of each graph, this can also rely on information about the estimated graph parameters. 4. Cladality For each group of four populations, we test whether the graph implies that any f4- statistic describing the relationship between the four populations is expected to be zero. 5. Node descendants Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 43 of 62 Evolutionary Biology | Genetics and Genomics Research article Each internal node in an AG is an ancestor to a specific set of leaf populations. An AG can be characterized by the sets of leaf populations formed by the internal nodes. Multiple AGs may be compared by counting the number of overlapping sets. This also makes it possible to quantify for each internal node in a single graph, how often a matching internal node can be found across a collection of alternative graphs, which is conceptually similar to bootstrap support values in phylogenetic trees. While these methods provide some help in comparing features across many graphs, they are not able to reliably answer the question whether the fitting graphs are relatively similar or dissimilar from each other, and whether they are similar to any particular graph. This is in part due to the fact that small topological changes involving populations of interest may be more relevant than similar topological changes involving only populations that are not the focus of the study. Appendix 1—figure 2. Comparison of accuracy of automated search for optimal topology in the findGraphs function of ADMIXTOOLS 2 and in TreeMix using simulated graphs with 8, 10, 12, and 16 populations, and 0–10 admixture events. Error bars show standard errors calculated as SE2 = p (1 – p) / n, where p is the fraction on the y- axis and n is the number of simulations in each group (typically 20). In the case of ADMIXTOOLS 2, we applied findGraphs three times on each simulated dataset and picked a result with the best fit score. More details are provided in Methods. (a) Fraction of simulations where the simulated graph is recovered exactly. (b) Fraction of simulations where the simulated graph is either recovered exactly, or the score is at least as good as the score of the simulated graph, when both graphs are evaluated by ADMIXTOOLS 2. More admixture edges greatly increase the search space and make it more difficult to recover the simulated graph, but they do make it easier to find alternative graphs with good fits. Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 44 of 62 Evolutionary Biology | Genetics and Genomics Research article Appendix 1—figure 3. Calibrating the bootstrap model comparison approach. (a) Bootstrap sampling distributions of the log- likelihood scores for two AGs (shown in Appendix 1—figure 3—Figure supplement 1) for the same populations fitted using real data. Vertical lines show the log- likelihood scores computed on all SNP blocks. (b) Distribution of differences of the bootstrap log- likelihood scores for both graphs (same data as in a). The purple area shows the proportion of resamplings in which the first graph has a higher score than the second graph. The two- sided p- value for the hypothesis of no difference is equivalent to twice that area (or one over the number of bootstrap iterations if all values fall on one side of zero). In this case it is 0.078. (c) The AG which was used to evaluate our method for testing the significance of the difference of two graph fits on simulated data. We simulated under the full graph and fitted two graphs that result from deleting either the red admixture edge or the blue admixture edge. These two graphs have the same expected fit score but can have different scores in any one simulation iteration. (d) QQ plot of p- values testing for a score difference between the two graphs (on simulated data) under the hypothesis of no difference, confirming that the method is well calibrated. The online version of this article includes the following figure supplement(s) for appendix 1—figure 3: Appendix 1—Figure 3 supplement 1. The admixture graphs compared in (Appendix 1—figure 3). Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 45 of 62 Evolutionary Biology | Genetics and Genomics Research article Appendix 2 1. Bergström et al., 2020 The AG for dogs in Figure 1e of Bergström et al., 2020 was inferred by exhaustively evaluating all graphs with two admixture events and outgroup ‘Andean fox’ for the six populations that remain in the graph after excluding an Early Neolithic dog from Germany. The only six- population graph with a worst residual (WR) below 3 standard errors (SE) was then chosen as a scaffold onto which the Early Neolithic dog genome from Germany was mapped, allowing for one more admixture event, and a seven- population graph with the lowest LL score was shown as the final model in the paper. Alternative six- population scaffolds were not explored in the original publication, although two six- population graphs with fits very similar to the best one were found. LL was not used as a ranking metric for alternative models in the original study; instead, the number of f4- statistics having residuals above 3 SE was considered. Since no f3- statistics were negative when all sites available for each population triplet were used (the ‘useallsnps: YES’ option), we did not use the upgraded algorithm for calculating f3- statistics on pseudo- haploid data. Our findGraphs results confirm the published six- population graph in that no graph with lower LL score is identified, but 3 of 14 unique alternative graphs found fit not significantly worse than the published graph (the published graph was also recovered by findGraphs) (Table 1, Supplementary file 1). When we used findGraphs to infer seven- population graphs with three admixture events (again fixing Andean fox as the outgroup), we identified five graphs with LL score nominally better than that of the published graph and one with a score that is slightly lower than that of the published graph but actually significantly better according to our model comparison methodology (this model is very similar to the published graph, Figure 3—source data 1). In the newly found seven- population graph with the best LL score (Figure 3—source data 1), the Siberian (Baikal), American, and East Mediterranean dogs are admixed, and the West European, East European (Karelia), and dogs of Southeast Asian origin (New Guinea singing dog) are unadmixed, while the opposite pattern is found in the published graph (Table 2). The best- fitting graph does not fit the data significantly better than the published graph (two- tailed empirical p- value = 0.332), but it bears a closer resemblance to the human population history (see the third- best graph found by findGraphs on human data from Bergström et  al., 2020 in Figure 3—source data 2) than the published seven- population graph (Figure 3a, Figure 3—source data 1). In this new seven- population model (Figure  3a), both American and Siberian dog lineages represent a mixture between groups related to the Asian and East European dog lineages, and robust genetic results suggest that in the time horizon investigated in the original publication (after ca. 10,900 years ago) nearly all Siberian (Jeong et al., 2019; Sikora et al., 2019) and all American (Raghavan et al., 2014; Raghavan et al., 2015; Moreno- Mayar et al., 2018) human populations were admixed between groups most closely related to Europeans and Asians. According to this model, East Mediterranean dogs are modeled as a mixture of a basal branch (splitting deeper than the divergence of the Asian and European dogs) and West European dogs, again in agreement with current models of genetic history of West Asian human populations who are modeled as a mixture of ‘basal Eurasians’ and WHG (Lazaridis et al., 2016; Lipson et al., 2017). Although greater congruence with human history increases the plausibility of findGraph’s newly identified model relative to the published model, to make unbiased comparisons between the history of the two species, model selection should be done strictly independently for each species, and so the genetic data alone does not favor one model more than another. Our results provide a specific alternative hypothesis that differs in qualitatively important ways from the published model and can be tested against new genetic data as it becomes available as well as other lines of genetic analysis of existing data. To explain why the original paper on the population history of dogs missed the model that findGraphs identified that is plausibly a closer match to the true history, we observe that the Bergström et al., 2020 AG search was exhaustive under the parsimony constraint (no more than twp admixture events for six populations), and thus missed the potentially true topology including three admixture events for these six populations. This case study also illustrates that even in a relatively low complexity context (seven groups and three admixture events) applying manual approaches for finding optimal models is risky. When any new group such as an Early Neolithic dog from Germany is added to the model, it may introduce crucial information into the system, and re- exploring the Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 46 of 62 Evolutionary Biology | Genetics and Genomics Research article whole graph space in an automated way is advisable. In contrast, mapping a newly added group on a simple skeleton graph (even when that skeleton is a uniquely best- fitting model) may yield a topology that is at odds with the true history. As the original Bergström et al. paper noted (Figure 3C of that study), no congruent six- population graph models were found for humans and dogs under the parsimony assumption: the three most congruent graphs for dogs resulted in poorly fitting models for the corresponding human populations (WR above 10 SE), and the three most congruent graphs for humans resulted in poorly fitting models for the corresponding dog populations (WR between 5 and 10 SE). We added a WHG group to the set of human groups from the original publication and using findGraphs on the original set of 77 K transversion SNPs we found that the third best- fitting model for humans (Figure 3—source data 2) (which is not significantly different in fit from the first one) is nearly identical topologically to the newly found dog graph. Even though findGraphs identified an AG topology that fits the data as well as the seven- population graph in Bergström et al. and is qualitatively quite different with respect to which populations were admixed, the new topology continues to support another of the key inferences of that study: that many of the early divergences among domesticated dog lineages occurred prior to the date of the Karelian dog (~10,900 ya). Thus, both graphs concur in providing strong evidence that the radiation of domesticated dog lineages occurred by the early Holocene, prior to the domestication of other animals. We further emphasize that the Bergström et  al., 2020 graph is the best- case scenario (along with Lazaridis et al., 2014 discussed below) for published AGs. Most published graphs are far less stable even than this. 2. Lazaridis et al., 2014 The graph in Figure 3 (from Lazaridis et  al., 2014) was inferred in the following manner. First, a phylogenetic tree without admixture was constructed which was the best fit for all f4- statistics among the populations ‘Mbuti’, ‘WHG Loschbour’ (Lazaridis et al., 2014), ‘LBK Stuttgart’ (Lazaridis et al., 2014), ‘Onge’, and ‘Karitiana’, with ‘Mbuti’ fixed as an outgroup. Next, all possible AGs were considered that result from adding a single admixture edge to this tree. After it was found that each of them had a WR > 3 SE, several graphs with two admixture events were considered, and some of them had WR < 3 SE. The ‘MA1’ genome (Raghavan et al., 2014) was added to these graphs in several different ways, and only one of these configurations was found to have WR < 3 SE. This was then used as a skeleton graph onto which a European population (represented by different present- day groups) was added. No fitting graph was found in which present- day Europeans could be modeled as a two- way mixture (adding one admixture event to the graph). After inspecting the non- fitting f- statistics of one of these graphs, it was found that modeling modern Europeans as a three- way mixture (adding two admixture events to the graph) is consistent with all f- statistics. Six f3- statistics were negative when all sites available for each population triplet were used (the ‘useallsnps: YES’ option, Supplementary file 1), but the upgraded algorithm for calculating f3- statistics on pseudo- haploid data had no effect since the only pseudo- haploid group in the dataset (MA1) was a singleton population, and the algorithm removes sites with only one chromosome genotyped in any non- singleton population. Thus, below we show results generated using the standard algorithm for calculating f- statistics. First, we considered the published skeleton graph onto which a European population was later added (Lazaridis et al., 2014). As in the Bergström et al., 2020 example, the best six- population graph with two admixture events found by findGraphs is identical to the published six- population graph, which has an LL score of 3.0 (Supplementary file 1). The second- best graph found has an LL score of 31.8. When computing the bootstrap p- value for the difference between these two graphs, we find that in 1.6% of all SNP resamplings the second- best graph has a better score than the published graph, resulting in a two- tailed empirical p- value of 0.032 for a difference in fits between these two graphs. All 14 alternative graphs found by our algorithm fit significantly worse than the published graph (Supplementary file 1). When we add the European population (French) and consider seven- population graphs with four admixture events, we find 40 out of 306 distinct graphs with a score better than that of the published graph (10 of those graphs are shown in Figure 3—source data 3). The best- fitting newly found model and two other models fit the data significantly better than the published model (Supplementary file 1), but their topology is qualitatively very similar to that of the published graph (Figure 3—source data 3). In the best- fitting newly found model, French and Karitiana share some drift to the exclusion of MA1, while in the published model the source of MA1- related ancestry in French is closer to MA1 than to Karitiana. Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 47 of 62 Evolutionary Biology | Genetics and Genomics Research article It is important to point out that not all of the 40 alternative graphs that fit nominally or significantly better than the published one are consistent with the conclusion that modern European populations are admixed between three different ancestral populations (Figure 3—source data 3). For example, the fifth alternative graph in Figure 3—source data 3 that is fitting nominally better than the published model (p- value = 0.464) includes no basal Eurasian ancestry in EEF (LBK Stuttgart), and instead models Onge as having ~50% West Eurasian- related ancestry and MA1 as having ~25% Asian ancestry. According to that graph, the present- day European population was formed by admixture of an MA1- related lineage and a European Neolithic- related lineage, with no West European hunter– gatherer (WHG) contribution. Of course, other lines of evidence make it clear that LBK Stuttgart is a mixture of Anatolian farmer- related ancestry and WHG Loschbour- related ancestry (Lazaridis et al., 2016; Lipson et al., 2017), thus providing external information in favor of the Lazaridis et al., 2014 model, and the use of such ancillary information in concert with graph exploration is important in order to obtain more confident inferences about population history taking advantage of AGs. The second alternative graph in Figure 3—source data 3 that fits just negligibly worse than the highest- ranking model has another distinctive feature: LBK Stuttgart is modeled as a mixture of a WHG- related and a basal Eurasian lineage, but modern Europeans receive a gene flow not from the LBK- related lineage, but from its basal Eurasian source. Although temporally plausible, this model is much less plausible from the archaeological point of view than the published model, and thus in this case too we can reject it as unlikely based on non- genetic evidence. We note, however, that a large group of newly found models (247 graphs) fits not significantly worse than the published one (Supplementary file 1), and those are topologically diverse. Thus, strictly speaking, the AG method on the given dataset cannot be used to prove that the published model is the only one fitting the data. 3. Shinde et al., 2019 The skeleton AG in the original study (Shinde et al., 2019) was constructed manually on the basis of an SNP set derived from the 1240K enrichment panel, and subsequently all possible branching orders (105) within the five- population Iranian farmer- related clade were tested. The published model (Figure 3 in that study) included 9 groups and 3 admixture events, but one group (Belt Cave Mesolithic) had a very high missing data rate, and as a result model fitting relied not just on the merged dataset which included 19,000 polymorphic sites without missing data across groups, but also on a dataset with approximately 470,000 sites that excluded the Belt Cave individual. The topological inferences were consistent for both analyses (Table S3 of that study). Following the approach of the published paper, we repeated findGraphs analysis both with and without the Belt Cave individual. Thus, we initially explored the following topology classes: 9 groups with 3 admixture events on ca. 19,000 polymorphic sites and 8 groups with 3 admixture events on ca. 470,000 sites (Supplementary file 1). The sample composition of the groups and the SNP dataset matched that in the original study. We summarize results across 4,000 independent iterations of the findGraphs algorithm for each topology class. For the nine- population graph we found 89 models with LL nominally better than that of the published model (Supplementary file 1). For the eight- population graph, we found 61 nominally and 4 significantly better fitting models (Supplementary file 1), and their topological diversity was high (Figure 3—source data 4). We note that the following groups were admixed by default in the graph models compared in the original study: Hajji Firuz Neolithic (labeled ‘Chalcolithic’ in that study but the dates are Neolithic) and Tepe Hissar Chalcolithic were considered as mixtures of an Anatolian farmer- related lineage and an Iranian farmer- related lineage; Indus Periphery was considered as a mixture of an Andamanese- related lineage representing ancient South Indians (ASI) and an Iranian farmer- related lineage. However, calculation of negative ‘admixture’ f3- statistics for these target groups is impossible using the original dataset and the original model fitting algorithm for several reasons. First, the Indus Periphery group was represented by a single pseudo- haploid individual (I8726) from the ‘Indus Valley cline’ for whom the best- quality data were available. But direct calculation of ‘admixture’ f3- statistics for such a group as a target is impossible since its heterozygosity cannot be estimated. Second, as discussed above, in Classic ADMIXTOOLS it is impossible to apply a correction intended for accurate calculation of f3- statistics on pseudo- haploid data (the ‘inbreed: YES’ option) if there is at least one population composed of one individual only (a singleton population). Third, the original Hajji Firuz Neolithic group composed of five individuals Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 48 of 62 Evolutionary Biology | Genetics and Genomics Research article included a family of three second- or third- degree relatives, and that artificially inflated the drift on the Hajji Firuz branch and made detecting a negative statistic f3(Hajji Firuz Neolithic; X, Y), even if present, highly unlikely. Indeed, no f3- statistic turned out to be nominally negative for the groups on the eight- population graph when statistics were calculated according to the original settings (we used settings equivalent to ‘useallsnps: NO’ and ‘inbreed: NO’ in classic ADMIXTOOLS, 470,389 polymorphic sites were available). Considering this fact, it is not surprising that our automated topology space search is not well constrained. The original study differed from ours since the constraints were introduced manually, but we wanted our topology search to be automatic and to explore a wider range of the parameter space. In order to provide power to detect negative f3- statistics useful for constraining the model search, we (1) removed two members of the family from the Hajji Firuz Neolithic group, (2) extended the number of individuals and sites available for the Indus Periphery group by generating new shotgun- sequencing data for a previously published library (Narasimhan et al., 2019) derived from individual I8726 (see Supplementary file 2) and by adding published data for three other individuals from the Indus Valley cline (from Gonur in Turkmenistan and Shahr- i- Sokhta in Iran; Narasimhan et al., 2019; Shinde et  al., 2019), (3) removed from other groups two individuals based on second to third degree relatedness, and (4) removed two individuals from other groups based on evidence of contamination with modern human DNA. All the changes to the dataset are shown in Supplementary file 3. In addition to these dataset adjustments, the new algorithm for calculating f- statistics makes it possible to compute negative f3- statistics on pseudo- haploid data, but at a cost of removing sites with only one chromosome genotyped in any non- singleton population (see Appendix 1). We eventually detected significantly negative ‘admixture’ f3- statistics f3(Tepe Hissar Chalcolithic; Ganj Dareh Neolithic, Anatolia Neolithic), f3(Indus Periphery; Ganj Dareh Neolithic, Onge), and other similar statistics for the same target groups. We also observed a nominally negative (Z- score = −0.6) statistic f3(Hajji Firuz Neolithic; Ganj Dareh Neolithic, Anatolia Neolithic), which is suggestive but does not by itself prove admixture in the recent history of the Hajji Firuz Neolithic group. For this updated analysis, 249,009 variable sites without missing data at the group level were available for the eight populations. We repeated topology search with this set of f- statistics providing additional constraints, performing 4,000 runs of the findGraphs algorithm. The Mota ancient African individual was set as an outgroup and three admixture events were allowed in the eight- population graph. Among 4,000 resulting graphs (one from each findGraphs run), 144 were distinct topologically, and the published model was recovered in 13 runs of 4,000 (Supplementary file 1). Only four distinct topologies fitting nominally better than the published one were found, and those had LL scores almost identical to that of the published eight- population model (16.97 and 17.66 vs. 17.85). These four alternative models (Figure 3—source data 5b) shared all topologically important features of the published model (Figure 3—source data 5a). Five other topologies differed in important ways from the published one and emerged as fitting the data worse, but not significantly worse, than the published one (Figure 3—source data 5c): two- tailed empirical p- values reported by our bootstrap model comparison method ranged between 0.060 and 0.112. Three of these topologies included a trifurcation of Iranian farmer- related lineages leading to the Indus Periphery, Hajji Firuz Neolithic, and Ganj Dareh Neolithic groups. The other two topologies included Hajji Firuz Neolithic as an unadmixed Anatolian- related lineage. In both cases, the Indus Periphery group was modeled as receiving a gene flow from either the Onge lineage (a proxy for ASI) or a deep Asian lineage. The finding that the predominant ancestry component of the Indus Periphery group was the most basal branch in the Iranian farmer clade was a prominent claim of the original study Shinde et al., 2019; for example, the abstract stated: ‘The Iranian- related ancestry in the IVC derives from a lineage leading to early Iranian farmers, herders, and hunter gatherers before their ancestors separated.’ Our finding that the Hajji Firuz Neolithic lineage may be as deep within the Iranian clade as the Indus Periphery lineage or may even diverge from the Anatolian branch shows that this statement cannot be confidently made based on AG analysis alone. However, the findings we have described up to this point do not invalidate the broader conclusion that the AG modeling in Shinde et al. was used to support; namely (using the phrasing from the abstract) that the genetic data ‘contradict… the hypothesis that the shared ancestry between early Iranians and South Asians reflects a large- scale spread of western Iranian farmers east.’ This Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 49 of 62 Evolutionary Biology | Genetics and Genomics Research article finding if correct is important, since it implies that the Iranian- related ancestry in the IVC (Indus Valley Civilization genetic grouping, which is the same group as IP), split from the Iranian- related ancestry in the first Iranian plateau farmers before the date of the Hajji Firuz farmers, who at ~8000 years ago are among the earliest people living on the Iranian plateau known to have grown West Asian crops. The ancient DNA record combined with radiocarbon dating evidence suggests that beginning around the time of the Hajji Firuz farmers, both West Asian domesticated plants such as wheat and barley, and Anatolian farmer- related admixture, began spreading eastward across the Iranian plateau. If the Iranian- related ancestry in IP was spread eastward into the Indus Valley across the Iranian plateau as part of the same agriculturally associated expansion—perhaps brought by people speaking Indo- European languages as well as introducing West Asian crops—then we would expect to see at least some of the Iranian- related ancestry in IP being a clade with that in Hajji Firuz relative to Ganj Dareh. The fact that we do not find any models compatible with this scenario is thus a potentially important finding. In summary, there are two reasons the genetic analyses we have reported up to this point continue to support the finding that the Iranian- related ancestry in IP is not a clade with the Iranian- related ancestry in Hajji Firuz (and Tepe Hissar) and thus is unlikely to reflect the same eastward movement of agriculturalists. First, in findGraphs analysis, all models specifying IP and Tepe Hissar and/or Hajji Firuz as a clade relative to Ganj Dareh were significantly worse- fitting that the published one. Instead, either the Iranian- related ancestry in IP definitively splits off first (the topology from Shinde et al.), or the branching order of the IP, Ganj Dareh, and the Hajji Firuz/ Tepe Hissar lineages cannot be determined, or IP, Ganj Dareh, and Tepe Hissar are a clade relative to Hajji Firuz. In all these fitting topologies, the ~10,000- year- old radiocarbon date of the Ganj Dareh individuals sets a lower bound on the split time between IP and Hajji Firuz/Tepe Hissar, which is pre- agriculturalist. This suggests that the Iranian- related ancestry in IP is not due to an eastward agriculturalist expansion. But in fact, the AG analysis reported above is not an adequate exploration of the problem. Although absolute fits of the best models found are good (WR = 2.5 SE), the parsimony constraint allowing only three admixture events precluded correct modeling of basal Eurasian ancestry shared by all West Asian groups (Lazaridis et al., 2016) or of the Indus Periphery group itself, for which a more complex 3- component admixture model was proposed (Narasimhan et al., 2019). Concerned that this oversimplification could be causing our search to miss important classes of models, we explored qpAdm models for the Indus Periphery group further, following the ‘distal’ protocol with ‘rotating’ outgroups outlined by Narasimhan et  al., 2019 and using the dataset and outgroups (‘right’ populations) from that study. All sites available for analyses were used, following Narasimhan et al. (the ‘useallsnps: YES’ option). The combined Indus Periphery group we analyzed included seven individuals from Shahr- i- Sokhta and three individuals from Gonur (three individuals were removed from the Narasimhan et al., 2019 dataset due to potential contamination with modern human DNA and low coverage). We removed one individual from the Ganj Dareh Neolithic group as potentially contaminated, and one second- or third- degree relative was removed from the Anatolia Neolithic group, see the dataset composition in Supplementary file 4. We note that no ‘distal’ qpAdm models were tested for the combined Indus Periphery group by Narasimhan et al., 2019, and individuals from this group were modeled one by one (Table S82 from Narasimhan et al., 2019), which potentially reduced the sensitivity of the method. A model ‘Indus Periphery = Ganj Dareh Neolithic + Onge (ASI)’ was strongly rejected for the Indus Periphery group of 10 individuals with a p- value = 2 × 10−15, and a model that was shown to be fitting for all Indus Periphery individuals modeled one by one by Narasimhan et al. (Ganj Dareh Neolithic + Onge (ASI) + West Siberian hunter–gatherers (WSHG)) was rejected for the grouped individuals with a p- value = 0.0044. In contrast, a model ‘Indus Periphery = Ganj Dareh Neolithic + Onge (ASI) + WSHG + Anatolia Neolithic’ was not rejected based on the p > 0.01 threshold used in Narasimhan et al. (the p- value was marginal but passing at 0.03) and produced plausible admixture proportions for all four sources that are confidently above zero: 53.2 ± 5.3%, 28.7 ± 2.1%, 10.5 ± 1.3%, and 7.7 ± 2.9%, respectively (Supplementary file 5). The same ‘distal’ model albeit with Anatolian Neolithic always in higher proportion was found as one of the simplest models (or the only simplest model) fitting the data for many other groups from Iran and Central Asia explored by Narasimhan et  al., 2019: Aligrama2_IA (13% Anatolia Neolithic), Barikot_H (21%), BMAC (26%), Bustan_BA_o2 (15%), Butkara_H (24%), Saidu_Sharif_H_o (12%), Shahr_I_Sokhta_BA1 (19%), and Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 50 of 62 Evolutionary Biology | Genetics and Genomics Research article SPGT (23%) (Supplementary file 5 gives a compendium of ‘distal’ modeling results by Narasimhan et al.). When we modeled Indus Periphery individuals separately, as in Narasimhan et  al., 2019, the simplest two- component model ‘Ganj Dareh Neolithic + Onge (ASI)’ was rejected for 5 of 10 individuals (at least ~315,000 sites were genotyped per individual), including the individual I8726 used for the AG analysis in Shinde et al. (Supplementary file 6). The model ‘Ganj Dareh Neolithic + Onge (ASI)’ was not rejected only for individuals with fewer than 141,000 sites genotyped, suggesting that this result is attributed not to population heterogeneity, but to lack of power. These qpAdm results show that the parsimony assumption that was made when constructing the AG analysis in Shinde et al., 2019 is contradicted by f- statistic evidence, and indeed Narasimhan et al. themselves showed this when they presented a distal qpAdm model that was more complex (Ganj Dareh Neolithic + Onge (ASI) + WSHG) than the one used for constraining the AG model comparison (Ganj Dareh Neolithic + Onge (ASI)). Another line of evidence used to support the principal historical conclusion by Shinde et al. was a series of f4- statistic cladality tests following correction of allele frequencies using an admixture model ‘target group = Iranian farmer + Anatolia Neolithic + Onge (ASI)’, with a great majority of tests supporting the deepest position of the Iranian farmer ancestry component in the Indus Periphery group within the Iranian farmer clade (Shinde et al., 2019). However, the model used for allele frequency correction (Ganj Dareh Neolithic + Onge (ASI) + Anatolia Neolithic) was simpler than the 4- component model we found and different from the 3- component model for the Indus Periphery group suggested by Narasimhan et al., 2019, which is a weakness of that analysis. A valuable direction for future work would be to repeat this analysis with a 4- component allele frequency correction model (Ganj Dareh Neolithic + Onge (ASI) + WSHG + Anatolia Neolithic), although that is beyond the scope of the present study, which simply aims to revisit the reported analyses and test if they fully support their inferences by ruling out alternative explanations. To explore how the parsimony constraint influences results, we allowed four admixture events in the eight- population graph (Supplementary file 1). Among 4,000 resulting graphs (one from each findGraphs run), 443 were distinct topologically, and 270 had WRs between 2 and 3 SE, that is, fitted the data well. We explored 35 topologies with LL scores in a narrow range between 9.3 (the best value) and 13.3. In Figure 3—source data 6, we show four graphs with four admixture events that model the Indus Periphery group as a mixture of three or four sources, with a significant fraction of its ancestry derived from the Hajji Firuz Neolithic or Tepe Hissar Chalcolithic lineages including both Iranian and Anatolian ancestries. The fits of these models are just slightly different (e.g., LL = 11.7 vs 9.3, both WRs = 2.4 SE) from that of the best- fitting model (Figure 3—source data 6), and similar to that of the published graph. Besides these four illustrative graphs, dozens of topologies with very different models for the Indus Periphery group fit the data approximately equally well, suggesting that there is no useful signal in this type of AG analysis when the parsimony constraint is relaxed (this finding is similar to that in our re- analysis of the dog AG in Bergström et al., 2020, where relaxation of the parsimony constraint identified equally well- fitting AGs that were very different with regard to their inferences about population history). These results show that at least with regard to the AG analysis, a key historical conclusion of the study (that the predominant genetic component in the Indus Periphery lineage diverged from the Iranian clade prior to the date of the Ganj Dareh Neolithic group at ca. 10 kya and thus prior to the arrival of West Asian crops and Anatolian genetics in Iran) depends on the parsimony assumption, but the preference for three admixture events instead of four is hard to justify based on archaeological or other arguments. Why did the Shinde et al., 2019 AG analysis find support for the IP Iranian- related lineage being the first to split, while our findGraphs analysis did not? Shinde et al., 2019 study sought to carry out a systematic exploration of the AG space in the same spirit as findGraphs—one of only a few papers in the literature where there has been an attempt to do so—and thus this qualitative difference in findings is notable. We hypothesize that the inconsistency reflects the fact that the deeply diverging WSHG- related ancestry (Narasimhan et  al., 2019) present in the IVC (Indus Valley Civilization genetic grouping, which is the same group as Indus Periphery) at a level of ca. 10% was not taken into account explicitly neither in the AG analysis nor in the admixture- corrected f4- symmetry tests also reported in Shinde et al., 2019. The difference in qualitative conclusions may also reflect the fact that the Shinde et al. study was distinguishing between fitting models relying on a LL difference threshold of 4 units (based on the AIC). As discussed in Appendix 1, AIC is not applicable to AGs Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 51 of 62 Evolutionary Biology | Genetics and Genomics Research article where the number of independent model parameters is topology dependent even if the numbers of groups and admixture events are fixed, and models compared with AIC should have the same number of parameters. Thus, the analysis by Shinde et al. was over- optimistic about being able to reject models that were in fact plausible using its AG fitting setup. The archaeological and linguistic implications of the Shinde et al. study are important, and there are several avenues available for further attempting to distinguish historical scenarios using f- statistics that are outside the scope of a methodological study like this one. Some of our observations that are most challenging for the conclusions of Shinde et al. are those related to the graphs with four admixture events in Figure 3—source data 6b that fit the Iranian farmer- related ancestry in the Indus Periphery group as deriving partially from the Hajji Firuz Neolithic or Tepe Hissar Chalcolithic- related lineages. The qpAdm method (Haak et al., 2015, Harney et al., 2021) is able to use information from distal outgroups (such as WSHG) not included in the AG modeling exercise revisited here. Leveraging this information might be able to obtain constraints that would further test the key historical conclusions from Shinde et al. Non- f- statistic- based methods could also be informative. Finally, we emphasize that the f4- statistic cladality tests correcting for the Anatolian farmer- and Onge- related admixture in the Indus Periphery grouping do continue to provide support for the historical conclusion of Shinde et al. (these analyses reject models where the Tepe Hissar or Hajji Firuz groups share genetic drift with the Indus Periphery individuals), with the caveat that they do not correct for the WSHG admixture. 4.Librado et al., 2021 In contrast to the other studies revisited in our work, the AG published by Librado et al., 2021was inferred automatically using OrientAGraph. Models with three (Figure 3b in that study) and zero to five (Ext. Data Fig. 5a- d in that study) admixture events were shown. The dataset included 10 populations (nine horse populations and donkey as an outgroup) and was based on 7.4  million polymorphic transversion sites with no missing data at the group level. We observed that some groups used for the OrientAGraph and qpAdm analyses were very broad geographically and temporally (see Table S1 in the original study), and thus we tested two alternative group compositions: the original one and a streamlined one. In the latter case we included individuals from one archaeological site and one archaeological period per group: the Botai, C- PONT, DOM2, ELEN, and NEO- ANA groups were modified in this way, and the CWC, LP- SFR, Tarpan, and TURG groups were left with the composition used in the original paper (Supplementary file 7). In addition, seven individuals with missing data proportion exceeding 80% were removed from the analysis, affecting the donkey outgroup, DOM2, and NEO- ANA groups (Supplementary file 7). Since among all possible f3- statistics for the 10 populations three were negative (using all sites available for each population triplet, ‘useallsnps: YES’), we applied the upgraded algorithm for calculating f- statistics, which removed sites with only one chromosome genotyped in any non- singleton population, resulting in the following site counts for the original and modified population compositions: 11,092 and 1,767,419 sites, respectively. The very low number of sites available in the former case is due to the fact that all individuals are pseudo- haploid, and that two groups (the donkey outgroup and NEO_ANA) are composed of two individuals, a high- coverage one and a low- coverage one. Thus, just sites genotyped in both donkey individuals and in both NEO_ANA individuals were kept. Considering this problem, we focused on the modified group composition only. We tested a range of model complexities (from 3 to 9 gene flows) and performed 1,000 findGraphs topology search runs per model complexity class. Unlike all the other AGs we re- evaluate in this study whose fits to the data were evaluated in the published studies using qpGraph, the topologies published in Librado et al., 2021 (with three to five admixture events) were not evaluated for statistical goodness- of- fit, and in fact fit the f- statistic data so poorly that even simple statistics show they cannot be correct (Figure 3b, Figure 3—source data 7a, c, e, Supplementary file 1). In this case, the approach of using findGraphs to identify alternative topologies with the same number of admixture events that fit the data better is meaningless, as both the published models and the alternative models do not have enough degrees of freedom to accommodate the complexity present in the real data; all models are guaranteed to be wrong. In particular, we found that WR of the published model with three admixture events is 23.9 SE (Figure 3—source data 7a). In this complexity class findGraphs found 22 topologically diverse models that fit significantly better than the published one (Table 1, Supplementary file 1), but nevertheless have extremely poor absolute fits (from 16.2 to 21.3 SE, see a temporally plausible example in Figure 3— source data 7b). In the complexity class with four admixture events, no model fitting better than the Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 52 of 62 Evolutionary Biology | Genetics and Genomics Research article published one was found; however, five alternative models fitting not significantly worse than the published one had lower WR (10 or 12 SE vs. 14.1 SE, Figure 3—source data 7d). The WR of the published model with five admixture events was 6.9 SE (Figure 3—source data 7e); just two models fitting nominally better and 223 models fitting non- significantly worse than the published model and having similar or higher WR were found (Table 1, Supplementary file 1). These results suggest that while OrientAGraph was often (but not always) able to find the same tentative global likelihood optimum as findGraphs, neither three nor five admixture events are enough to explain the data since nearly all the groups are probably admixed. For this reason, we moved to topology searches in more complex model spaces incorporating six to nine admixture events. Temporally plausible models with even a modest fit (WR between 3 and 4 SE) were encountered only among models with eight and nine admixture events (Figure 3—source data 7j- r). In the complexity class with eight admixture events, five such temporally plausible fitting models were found, with WRs ranging from 3.4 to 3.9 SE (all these models are shown in Figure 3— source data 7j- l). In the complexity class with 9 admixture events, 11 such models were found, with WRs ranging from 3.4 to 4.0 SE (all these models are shown in Figure 3—source data 7m- r). Librado et al., 2021 discussed the following inferences relying fully or partially on their published AGs reported in that study (Table 2): (1) NEO- ANA- related admixture is absent in DOM2; (2) DOM2 and C- PONT are sister groups (they form a clade); (3) there is no gene flow connecting the CWC group and the cluster associated with Yamnaya horses and horses of the later Sintashta culture whose ancestry is maximized in the Western Steppe (DOM2, C- PONT, TURG); (4) there was gene flow from a deep- branching ghost group to NEO- ANA; and (5) Tarpan is a mixture of a CWC- related and a DOM2- related lineage. The simplest temporally plausible and best- fitting (WR = 3.4 SE) model we found (modified group composition, eight admixture events; see Figure  3b and the second model in Figure 3—source data 7j) supports inferences 2 and 4, and is incompatible with inferences 1, 3, and 5 (Table 2). This newly found model can be interpreted as follows. There is a trifurcation of three deep lineages: a lineage maximized in Western and Central Europe (up to 100% of ancestry in a Late Paleolithic group from France, LP_SFR), a Western- Steppe- specific lineage (up to 55% in TURG), and a Tarpan- specific lineage (22% in Tarpan). Western and Central European horses, represented by LP- SFR, by the majority ancestry in horses found in the Corded Ware culture context (CWC), and by the majority ancestry in wild Neolithic Anatolian horses (NEO_ANA), contributed about half of the ancestry in the Western Steppe groups TURG, C- PONT, and DOM2. The other half of ancestry in the Western Steppe groups is represented by the Western Steppe- specific lineage. That lineage also contributed about 50% of ancestry in wild horses from the Yana Upper Paleolithic site (ELEN), and the other half of ELEN’s ancestry is derived from an even deeper lineage. The Botai group is modeled as a mixture of European horses (69%) and Siberian horses (31% ELEN- related ancestry). In contrast to Librado et al., 2021, Tarpan is modeled as a mixture of its specific lineage (22%) and a DOM2- related group (78%), and CWC also received ancestry (21%) from a DOM2- related group. All the populations included in the model except for LP_SFR are admixed, and there is evidence of substantial genetic influence from a lineage that was eventually maximized in the Western Steppe (although it did not necessarily originate there) in the ELEN and Botai groups. We consider this model to be plausible from both temporal and geographical perspectives. We are not arguing here that our eight- admixture- event model represents the true history; in fact, it is highly unlikely to be entirely true, given how large the space of all possible admixture events is and how much admixture evidently occurred relating all these groups (which makes finding the unique truly fitting model extremely unlikely based on f- statistic fitting, see the results on simulated data in Figure 1 and Appendix 1—figure 2b). We have also not attempted in any way to replicate the AG exploration procedure performed in the Librado et al., 2021. study; the graph fitting procedure was quite different from ours, based on OrientAGraph optimization rather than findGraphs optimization, and a Block Jackknife procedure with a different genome block size for determining standard errors (4 Mbp in our protocol and ca. 500 kbp in the Librado et al., 2021. study). Regardless of how the graph was obtained, it is valuable for providing readers with guidance about which topological features of the graphs are meaningful and stable, and which are less certain, especially—as in the case of the AG presented in the paper—when some features of the presented model do not fit the data by a wide margin, as evident by the WR of 6.9 in the published model for five admixture events. Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 53 of 62 Evolutionary Biology | Genetics and Genomics Research article Our set of 16 temporally plausible and fitting (WR  < 4  SE) models with eight or nine admixture events (Figure 3—source data 7j–r) is consistent with some features of the published graph being stable: the features (2) that DOM2 and C- PONT are sister groups, and (4) that there was a gene flow from a deep- branching ghost group to NEO- ANA (Table 2). Equally important, however, is our finding that there are plausible models that are inconsistent with other inferences in Librado et al., 2021. (Table 2). For example, 13 of these 16 models are inconsistent with the suggestion that there was no gene flow connecting the CWC group and the cluster maximized in the Western steppe (DOM2, C- PONT, and TURG) (Figure 3—source data 7j–r). In the eight- admixture- event best- fitting plausible model (Figure 3b and the second model in Figure 3—source data 7j), CWC actually derives appreciable ancestry from the early domestic horse lineage (DOM2) associated with the Sintashta culture to the exclusion of the more distant Yamnaya- associated TURG and C_PONT horses. This scenario presents a parallel to the one observed in humans, with individuals associated with the CWC receiving admixture from Steppe pastoralists albeit in different proportions: ~75% for humans, versus ~20% in horses. These models specifying a substantial Steppe horse contribution to CWC horses would weaken support for the inference in Librado et  al., 2021. that ‘Our results reject the commonly held association between horseback riding and the massive expansion of Yamnaya steppe pastoralists into Europe around 3000 BC.’ We are not aware of other lines of evidence in the paper (apart from the fitted AG) that support the claim of no Yamnaya horse impact on CWC horses. Another example of a feature of the published graph that turned out to be unstable is the model for the Tarpan horse. Only 8 of 16 temporally plausible and fitting models (Figure 3—source data 7j–r) support the conclusion by Librado et  al., 2021. that the Tarpan is a mixture of a DOM2- related and a CWC- related lineage. The other 8 models suggest that Tarpan is a mixture of a deep lineage and a DOM2- related lineage (Figure 3b and the second model in Figure 3—source data 7j), echoing a hypothesis that Tarpan may be a hybrid with the Przewalski horse lineage not represented in the AG (Librado et al., 2021). Again, we are not arguing here that our fitting alternative model is right—indeed we are nearly certain it is wrong in important aspects—but we are merely pointing out that the complexity of the AG space means that qualitatively quite different conclusions are compatible with the genetic data. Other aspects of the Librado et al., 2021. study, most notably the dramatic geographic expansion of the DOM2 modern domestic horse lineage after 4000 years ago in association with the Sintashta culture which is the most extraordinary finding of Librado et al., 2021., are in no way challenged by our results. 5. Hajdinjak et al., 2021 The AG inferred by Hajdinjak et al. was constructed manually on the basis of an SNP set derived from in- solution enrichment of two SNP panels (1240K and a further million of transversion polymorphisms discovered as polymorphic within one or two sub- Saharan African individuals or among archaic humans) and incorporated 11 groups and 8 admixture events (Figure 2d in the original study). The published graph has no clear outgroup since the deepest branch (Denisovan) is admixed. This property of the graph makes automated graph space exploration difficult. We explored two topology classes: (1) 11 groups with 8 admixture events, the original SNP set, Denisovan assigned as an outgroup only at the stage of generating random starting graphs (gene flows to/from the Denisovan branch were allowed at the topology optimization step); and (2) 12 groups with 8 admixture events, chimpanzee added and the original SNP set changed due to the zero missing rate condition, and chimpanzee assigned as an outgroup at both algorithm stages (Supplementary file 1). For both graph complexity classes, two topology search settings were tested: (1) either no additional constraints were applied beyond the outgroup constraints described above, or (2) the Vindija Neanderthal and Mbuti were allowed to have no admixture events in their history, and the Denisovan lineage was allowed to have up to one admixture event in its history (these constraints were in line with the model in the original study and with literature on the genetic history of archaic humans, e.g., Prüfer et al., 2014). The composition of the groups matched that in the original study, as did the parameter settings for qpGraph, with the exception of ‘least squares mode’, which was used in the original study, but not in our analysis. ‘Least squares mode’ computes LL scores without taking into account the f- statistic covariance matrix, and we confirmed that changing this parameter does not qualitatively change our results. Since no f3- statistics were negative when all sites available for each population triplet were used (the ‘useallsnps: YES’ option), we did not use the upgraded algorithm for calculating f3- Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 54 of 62 Evolutionary Biology | Genetics and Genomics Research article statistics on pseudo- haploid data. We summarize results across 2,000–4,000 independent runs of the findGraphs algorithm (Supplementary file 1). When chimpanzee was not included into the analysis and no topology constraints were applied, nearly all newly found models turned out to be distinct (3,996 of 4,000), nearly all (96.8%) fit nominally better and 15.9% fit significantly better than the published model (Supplementary file 1), and absolute fits of 91.3% of novel models are good (WR < 3 SE). Similar results were obtained when the topology search algorithm was constrained: nearly all (89.5%) of 1,999 newly found models fit nominally better and 26% fit significantly better than the published model (Supplementary file 1). When chimpanzee was set as an outgroup and no topology constraints were applied, the picture remained similar. Nearly all newly found models turned out to be distinct (1,996 of 2,000), and a very large fraction of them (56.8%) fit significantly better than the published model (Supplementary file 1); 16.4% of novel models demonstrated WR < 3 SE. Similar results were obtained when the topology search algorithm was constrained: most (71.4%) newly found models fit nominally better and 15.7% fit significantly better than the published model (Table 1, Supplementary file 1, Figure 2), which has a poor absolute fit on this set of sites and groups (WR = 4.8 SE, Figure 3c, Figure 3—source data 8). The statistics described above and the fact that LL scores on all sites lie outside of the LL distribution on resampled datasets (Figure 2) suggest that models in this complexity class are overfitted, but the published topology emerged as fitting relatively poorly. Overfitting arises naturally during manual graph construction as performed in many studies (not only in Hajdinjak et al., 2021, but also in, e.g., Fu et al., 2016; Skoglund et al., 2016; Yang et al., 2017; Posth et al., 2018; McColl et al., 2018; Moreno- Mayar et al., 2018; Tambets et al., 2018; van de Loosdrecht et al., 2018; Flegontov et al., 2019; Sikora et al., 2019; Wang et al., 2019; Lipson et al., 2020b; Shinde et al., 2019; Yang et al., 2020; Wang et al., 2021). The graph grew one group at a time, and each newly added group was mapped on to the pre- existing skeleton graph as unadmixed or as a two- way mixture. This imposed constraints on the model- building process. Another constraint imposed was the requirement that all intermediate graphs have good absolute fits (WR below 3 or 4 SE). When the model- building process is constrained in a particular path and fits of all intermediates are required to be good, unnecessary admixture events are often added along the way, and the resulting graph belongs to a complexity class in which models are overfitted and many alternative models fit equally well. There is no single obviously correct order of adding branches to a growing graph. For example, the Kostenki and Sunghir lineages were included into the initial graph (Fig. S6.1 in the original study) as unadmixed lineages, and their admixture status was not revisited at subsequent steps (unlike that of Tianyuan and Ust’-Ishim), except for adding the archaic gene flow common for non- Africans. For that reason, the published graph differs from many alternative better- fitting and temporally plausible graphs where the Kostenki and Sunghir lineages are modeled as more complex mixtures (Figure 3—source data 8). Hajdinjak et al., 2021’s published graph had the following notable features that were interpreted by the authors and used to support some conclusions of the study (Table 2): (1) there are gene flows from the lineage found in the ~45,000- to 43,000- year- old Bacho Kiro Initial Upper Paleolithic (IUP) individuals to the Ust’-Ishim, Tianyuan, and GoyetQ116- 1 lineages; (2) the ~35,000- year- old Bacho Kiro Cave individual BK1653 belonged to a population that was related, but not identical, to that of the GoyetQ116- 1 individual; and (3) the Vestonice16 lineage is a mixture of a Sunghir- related and a BK1653- related lineage. To assess if these features are supported by our re- analysis, we focused on our most constrained findGraphs run: with chimpanzee set as an outgroup and with the topology constraints applied at the topology search step. We identified 1,421 topologies fitting nominally or significantly better than the published model and satisfying the constraints and moved on to inspect 50 best- fitting topologies for temporal plausibility (all of them fitting significantly better than the published model). All non- African individuals included in the model are Upper Paleolithic and their dates are not drastically different in relative terms: from ca. 45 kya (thousand years before present) for some Bacho Kiro IUP individuals (Hajdinjak et al., 2021) to ca. 30 kya for the Vestonice16 individual (Fu et al., 2016). Nevertheless, we considered most gene flows from later- to earlier- attested lineages as temporally implausible (for instance, GoyetQ116- 1 (~35 kya) Bacho Kiro IUP (45–43 kya), Kostenki14 (38 kya) → Tianyuan → (40 kya)) since they imply great antiquity of the later- attested lineages, for example, >40 kya for Ust’-Ishim (~44 kya), GoyetQ116- 1 (35 kya) Ust’-Ishim (44 kya), Sunghir III (34.5 kya) Tianyuan (40 kya), Vestonice16 (30 kya) → → → Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 55 of 62 Evolutionary Biology | Genetics and Genomics Research article the Vestonice16 lineage, and even greater antiquity for the related lineages such as Sunghir III and Kostenki14. → → Tianyuan and Bacho Kiro GoyetQ116- 1 gene flow, but no Bacho Kiro Of the 50 topologies inspected, 32 were considered temporally plausible. Of those topologies, none supported feature 1 of the published AG (there is no replication of the finding of gene flows from the Bacho Kiro IUP lineage specifically to all three of the Ust’-Ishim, Tianyuan, and GoyetQ116- 1 lineages). One topology supported features 2 and 3, and partially supported feature 1 (there was Ust’-Ishim Bacho Kiro gene flows). A total of 17 topologies supported features 2 and 3 but were inconsistent with feature 1; and 14 topologies supported feature 3 only (Table  2). Best- fitting representatives of each of these topology classes are shown along with the published model in Figure 3—source data 8. Considering topological diversity among models that are temporally plausible, conform to current knowledge about relationships between modern and archaic humans, and fit significantly better than the published model, we conclude that feature 3 is probably robust but other details of the fitted AG in Hajdinjak et al. (Figure 2d of that study)—for example, gene flows to the Ust’-Ishim, Tianyuan and Goyet Q116- 1 lineages from sources sharing drift exclusively with the Upper Paleolithic Bacho Kiro lineage—should not be interpreted as providing meaningful inferences about population history of Upper Paleolithic modern humans. For example, the upper right- hand alternative model plotted in Figure 3—source data 8c supports features 2 and 3 but includes no gene flows from the Bacho Kiro IUP lineage. → A central finding of Hajdinjak et al. is that the Bacho Kiro IUP group shares more alleles with present- day East Asians than with Upper Paleolithic Holocene Europeans despite coming from Europe. Specifically, the study documents significantly positive statistics of the form D(an Asian group, Kostenki14; Bacho Kiro IUP, Mbuti) (Figure 2b and Extended Data Figure 5 in the original study). For example, D(Tianyuan, Kostenki14; Bacho Kiro IUP, Mbuti) is significantly positive (D = 0.0032, SE = 0.0010, Z = 3.2) on the dataset used for testing the 12- population graphs (263,698 sites without missing data across all 12 groups). The same statistic is also significantly positive (D = 0.0029, SE = 0.0006, Z = 4.4) when all 1,312,292 non- missing sites in the population quadruplet are analyzed. Hajdinjak et al.’s interpretation of this observation, using the language from the abstract, is that ‘there was at least some continuity between the earliest modern humans in Europe [Bacho Kiro IUP] and later people in Eurasia [East Asians]’. However, a significant D- statistic can have multiple explanations. The statistic f4(Tianyuan, Kostenki14; Bacho Kiro IUP, Mbuti) is fitted equally well by the published 12- population AG (Z- score for the difference between the observed and fitted statistics = 0.64) and by, for example, the lower left- hand graph in Figure 3—source data 8c (Z- score = 0.94) reproduced in Figure 3c. Under the latter model that fits the data significantly better than the published model (p- value = 0.02), the Bacho Kiro IUP and Tianyuan branches are not connected by a gene flow and do not receive gene flows from a third common source, but the common ancestor of Ust’-Ishim and all European Paleolithic lineages receives an 8% gene flow from a divergent modern human lineage splitting deeper than Bacho Kiro IUP and Tianyuan (Figure 3c, Figure 3—source data 8c). This scenario or some version of it seems archaeologically and geographically plausible and is not disproven by any other line of genetic or non- genetic evidence of which we are aware. It could correspond to a scenario where a primary modern human expansion out of West Asia contributed serially to the major lineages leading to Bacho Kiro, then later East Asians, then Ust’-Ishim, and finally the primary ancestry in later European hunter–gatherers. This has a very different interpretation from the scenario of distinctive shared ancestry between the earliest modern humans in Europe such as Bacho Kiro IUP and later people in East Asia—to the exclusion of later European hunter–gatherers—that is suggested by the Hajdinjak et al. published graph. We are not claiming that this specific alternative model is correct—indeed, it is almost certainly not the correct one given the topological complexity of the set of all AGs consistent with the data— but the existence of it and many other models that fit the data makes it clear that we do not yet have a unique historical explanation for the excess sharing of alleles that has been documented between some Upper Paleolithic European groups (Bacho Kiro IUP, Hajdinjak et  al., 2021 GoyetQ116- 1, Yang et al., 2017 and Hajdinjak et al., 2021) and all East Asians. Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 56 of 62 Evolutionary Biology | Genetics and Genomics Research article 6. Lipson et al., 2020b The AG in the original study (Lipson et  al., 2020b) was constructed manually based on an SNP set derived from the 1240K enrichment panel, and the final model was alternatively tested on the combined Human Origins subpanels 4 and 5 (each ascertained on one African individual) or on sites ascertained as polymorphic in archaic humans. The final published model (Extended Data Figure 4 in that study) is very complex (12 groups and 12 admixture events): it exists in a space of ~1044 topologies of this complexity. We note that one admixture event was added by Lipson et al., 2020b to account for potential modern DNA contamination in ancient Shum Laka individuals, and removing it caused a negligible difference in the fit of the published model (Supplementary file 1). Thus, to decrease the complexity of the graph search space, we considered graphs with 12 groups and 11 admixture events. Twenty- two f3- statistics for these 12 groups turned out to be negative (when the ‘useallsnps: YES’ setting was used), and thus for exploring this graph complexity class we had to remove sites with only one chromosome genotyped in any non- singleton population (Supplementary file 1). The following constraints were applied during the topology search: chimpanzee was assigned as an outgroup at both stages of the process (while generating random starting graphs and while searching the topology space); Altai Neanderthal was required to be unadmixed; and non- Africans (French) were required to have at least one admixture event in their history. The composition of the groups we analyzed matched that in the original study. We summarize results across 2,000 independent iterations of the findGraphs algorithm. All newly found models turned out to be distinct (2,000), and 11.9% fit nominally (but not significantly) better than the published model (Table 1, Supplementary file 1, Figure 2). Absolute fits of 36.7% of novel models are good (WR  < 3  SE). Fits of the highest- ranking model and the published model are not significantly different according to the bootstrap model comparison method (p- value = 0.176). These metrics, along with the fact that LL scores on all sites lie outside of the LL distribution on resampled datasets (Figure 2), suggest that models in this complexity class, including the published model, are overfitted. Of the AGs we re- evaluate in this study, Lipson et al., 2020b shares with Hajdinjak et al., 2021, Sikora et al., 2019, and Wang et al., 2021 evidence of being overfitted (Figure 2). We also wanted to check if overfitting would be found in the graph complexity classes corresponding to two simpler intermediate graphs from the original study (Supplementary file 1): 7 groups and 4 admixture events (Figure S3.24 in that study) and 10 groups and 8 admixture events (Figure S3.25 in that study). The population composition of the dataset we used for this analysis was slightly different from the dataset used by Lipson et al.: the ancient South African hunter–gatherer group was replaced by a related group (present- day Juǀʼhoan North), and instead of the Shum Laka ancient group, only one high- coverage individual from the same group (I10871) was used. We summarize results across 2,000 or 10,000 independent findGraphs runs for each SNP set, for the small and large graphs, respectively. For 7 groups, we found 201 novel topologies fitting better than the published one, and for 10 groups we found nearly 9,000 such topologies (Supplementary file 1). In the latter case, 6.8% of newly found topologies fit significantly better than the published topology. For the more complex graph class with 10 groups and 8 admixture events we also found evidence of overfitting: the LL score of the published graph run on the full data is better than almost all the bootstrap replicates on the same data (it falls below the 5th percentile). Below we discuss selected prominent features of the AG published in the original study (that were interpreted by the authors and used to support some conclusions of the study) and the extent to which these features consistently replicate across the large number of fitting 12- population graphs with 11 admixture events (Table  2): (1) A lineage maximized in present- day West African groups (Lemande, Mende, and Yoruba) also contributed some ancestry to the ancient Shum Laka individual and to present- day Biaka and Mbuti; (2) another ancestry component in Shum Laka is a deep- branching lineage maximized in the rainforest hunter–gatherers Biaka and Mbuti; (3) ‘super- archaic’ ancestry (i.e., diverging at the modern human/Neanderthal split point or deeper) contributed to Biaka, Mbuti, Shum Laka, Lemande, Mende, and Yoruba; and (4) a ghost modern human lineage (or lineages) contributed to Agaw, Mota, Biaka, Mbuti, Shum Laka, Lemande, Mende, and Yoruba. We identified 232 12- population topologies that fit nominally better than the published one, 34 best- fitting topologies (of 232) were manually assessed for temporal plausibility, and we focus on 30 topologies identified as temporally plausible and including a low- level Neanderthal contribution (≤10%) in non- Africans (French). These 30 topologies are shown along with the published model in Figure 4—source data 1. Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 57 of 62 Evolutionary Biology | Genetics and Genomics Research article In this set of alternative models, high topological diversity is observed (see an example in Figure 4a and further topologies in Figure 4—source data 1). We classified the topologies as follows. If an ancestral lineage defined above (for example, a deep- branching lineage maximized in rainforest hunter–gatherers Biaka and Mbuti) exists in the graph, we compared the sets of populations where it is found in the published model and in the model examined. If there was no more than one population where the ancestry is expected according to the published model but not present, or present but not expected, we considered this feature of the published graph supported by the alternative graph. If no ancestral lineage meets the definition above, the feature of the published graph was considered not supported. In all other cases, partial support for the feature was declared (Figure 4—source data 1). Considering extreme cases, two alternative graphs completely lacked support for three features of the published graph (Figure 4a, Figure 4—source data 1c), and one graph supported all four features of the published graph fully (Figure 4—source data 1q, the second model). There are some graphs where defining two distinct ancestral lineages maximized in West Africans and in Mbuti and Biaka (features 1 and 2) is essentially impossible since all or nearly all Africans are modeled as a mixture of at least two deep lineages (see the second model in Figure 4— source data 1d). In some graphs, there is no single lineage specific to rainforest hunter–gatherers (Biaka, Mbuti, and Shum Laka) since the primary ancestries in these groups form independent deep branches in the African graph (see Figure 4a and the second model in Figure 4—source data 1j). The ghost modern and super- archaic gene flows to Africans also had no universal support in the set of alternative graphs we examined (see, e.g, Figure 4a and Figure 4—source data 1c). Considering the high degree of topological diversity among models that are temporally plausible, conform to known findings about relationships between modern and archaic humans, and fit nominally better than the published model, we conclude that all the four AG features from the original study are not supported by our re- analysis (Table 2). As in the case study above, the published manually constructed model is a representative of a large class of models that are equally well fitting to the limits of our resolution. This situation may be attributed to (1) overfitting and/or to (2) the lack of information in the dataset (in the combination of groups and SNP sites) and/or to (3) inherent limitations of f- statistics, when distinct topologies predict identical f- statistics. In reconsidering the findings of Lipson et al., 2020b it is important to keep in mind that analysis of allele frequency correlation statistics is not the only type of information that can be used to make inferences about population relationships in deep time. Other methodologies have provided important insights into deep African population history, and the model building in Lipson et  al., 2020b was guided in an informal way by these other lines of evidence. For example, unknown archaic lineages admixing into some African populations were hypothesized through identification of deeply splitting haplotypes that are too long to have been freely mixing with other haplotypes in present- day populations for all of their history (Hammer et al., 2011; Lachance et al., 2012; Speidel et al., 2019). Similarly, analysis of haplotype divergence times of pairs of populations has been used to provide evidence of an early radiation of modern human lineages maximized today in southern African hunter–gatherers, Mbuti rainforest hunter–gatherers, and the great majority of other present- day populations; and a later split of lineages related to East African hunter–gatherers, West African agriculturalists, and non- Africans, which is a feature of the Lipson et al. model (Campbell and Tishkoff, 2008; Mallick et al., 2016). Notably, some alternative models we found do not contradict the above- mentioned results and are profoundly different from the published model at the same time (see, e.g., Figure 4a). These constraints are not enough, however, to provide evidence for all the topological details of the Lipson et al., 2020b AG highlighted in this section, or for other features of the Lipson et al., 2020b AG that were not invoked in the previous literature and newly proposed in that study, such as the ‘ghost modern’ lineage splitting around the same time as the lineages leading to southern African hunter–gatherers and central African rainforest hunter–gatherers and mixing in highest proportion to Ethiopian hunter–gatherers and to a lesser proportion to West Africans, and the ‘basal West African’ lineage that contributes uniquely to Shum Laka. Many of the models that emerged as good fits in our AG- building exercise as the published one did not share some of these features (Figure 4—source data 1). The high diversity of well- fitting AG models that satisfy known constraints relating diverse African populations highlights the need for further research based on multiple lines of genetic analysis (in addition to allele frequency correlation patterns) to obtain further insights into deep African history. Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 58 of 62 Evolutionary Biology | Genetics and Genomics Research article Our results particularly highlight the mystery around the highly distinctive genetic ancestry of the Shum Laka individuals themselves, who represent the newly reported data in the Lipson et  al., 2020b study and a highly important set of genetic datapoints that was not available prior to the study. The ancestral relationships of these four individuals to both rainforest hunter–gatherers, and to the primary lineage in present- day West Africans, remains an open question, one whose resolution promises meaningful new insights into human population history. 7. Wang et al., 2021 The AG inferred by Wang et al., 2021 was constructed manually on the basis of an SNP set derived from the 1240K enrichment panel. We focused our analysis on the final graph (Extended Data Figure 6 in Wang et al., 2021, 12 groups and 8 admixture events) and on two simpler intermediates in the model- building process (Figures SI3- 9 and SI3- 10a in Wang et  al., 2021). To simplify the latter two models further, we removed a low- level gene flow (1%) from a WHG- related lineage (Loschbour) to the Mongolia Neolithic group, which resulted in negligible LL differences (0.5 and 2.4 log- units, respectively). Thus, using findGraphs we explored the following topology classes: 9 groups with 4 admixture events, 10 groups with 5 admixture events, and 12 groups with 8 admixture events (Supplementary file 1). The composition of the groups matched that in the original study. We summarized results across 2,000 independent iterations of the findGraphs algorithm for each topology class. In the case of the most extensive population set (12 groups), three f3- statistics turned out to be negative (when the ‘useallsnps: YES’ setting was used), and thus for exploring this graph complexity class we had to remove sites with only one chromosome genotyped in any non- singleton population (Supplementary file 1). For this complexity class, we also applied several constraints on the graph space exploration process all of which were shared with the Wang et al. graphs: the Denisovan genome was assigned as an outgroup in the random starting graphs, but not at the topology search stage; up to one admixture event was allowed in the history of the Denisovan group; no admixture events were allowed in the history of Mbuti, Loschbour, and Onge; and the (Denisovan, (Mbuti, (Loschbour, Onge))) branching order was required. For each topology class we found hundreds to thousands of topologically unique graphs fitting nominally better than the published models (Table  1, Supplementary file 1). For both simple topology classes, no model fitting significantly better than the published one was found (Supplementary file 1). However, the final published model fits the data significantly worse than 12.6% of newly found models of the same complexity (Table 1, Supplementary file 1). The fact that many topologically diverse models had good absolute fits (65%, 55%, and 15% of distinct newly found graphs with 9, 10, and 12 groups, respectively, had WR < 3 SE) suggests that AG models in these complexity classes are overfitted. Further evidence of overfitting comes from the poor fits of the published model on bootstrap- resampled datasets as compared to their fits on all sites (Figure 2). An important feature of the published graphs in Wang et al., 2021 that was remarked upon in the study is admixture from a source related to Andamanese hunter–gatherers that is almost universal in East Asians, occurring in the Jomon, Tibetan, Upper Yellow River Late Neolithic, West Liao River Late Neolithic, Taiwan Iron Age, and China Island Early Neolithic (Liangdao) groups (Table 2). For example, the abstract states ‘Hunter- gatherers from Japan, the Amur River Basin, and people of Neolithic and Iron Age Taiwan and the Tibetan Plateau are linked by a deeply splitting lineage that probably reflects a coastal migration during the Late Pleistocene epoch.’ We performed 2,000 findGraphs iterations and obtained 1,778 distinct topologies satisfying all the constraints, nearly all of them (1,724) fitting nominally better than the published model, and 12.6% fitting significantly better (Supplementary file 1). The models were ranked by LL, and 56 highest- ranking topologies, all of them fitting significantly better than the published one, were assessed for temporal plausibility (models with gene flows from a later group to Tianyuan dated to 40 kya were removed), and 20 topologies were considered temporally plausible (all of them are shown in Figure 4—source data 2). According to these topologies, 0–2 East Asian groups had a fraction of their ancestry derived from a source specifically related to Onge, and 19 topologies included gene flows from the European (Loschbour)- related branch to all 8 East Asian groups (Figure 4—source data 2). The inferred topological relationships among East Asians are variable in this group of 20 models, and we decided to apply further constraints that guided model ranking and elimination by Wang et al., based on considerations from archaeological evidence, Y chromosome haplogroup divergence patterns, and population split time estimation. Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 59 of 62 Evolutionary Biology | Genetics and Genomics Research article The constraints that are not based on correlation of allele frequencies across populations that Wang et al. applied and that we applied in our re- examination are as follows. First, combined evidence from archaeology, linguistics, and genetics (a closely shared Y chromosome haplogroup) suggests that the present- day Tibetan Plateau population harbors a substantial proportion of ancestry from a large- scale migration from the Neolithic farming groups from the Upper and Middle Yellow River (Chen et al., 2015; Lu et al., 2016; Zhang et al., 2019). These arguments and radiocarbon dates favor the following branching order of predominant ancestry components: (Mongolia East N, (China Upper YR LN, Nepal Chokhopani)). Second, evidence from archaeology, linguistics and genetics suggests that the expansion of Austronesian speakers and the peopling of Taiwan was from southeast coastal China to Taiwan and Southeast Asia, but not from Taiwan to mainland China (Bellwood, 2011; Gray and Jordan, 2000; Ko et al., 2014). These arguments make a China Island EN Taiwan IA gene flow direction plausible and make the opposite direction of flow less likely. Third, in the original study MSMC cross- coalescence rates were computed for a few pairs of present- day proxies for the ancient groups, and it was argued that they impose constraints on the graph topology. The inferred coalescence date for the Tibetan and Ulchi groups was slightly younger than the Tibetan- Ami and Tibetan- Atayal dates (see Fig. SI3- 1 in the original study), suggesting that the Nepal Chokhopani and Mongolia East N group may share ancestral source populations more recently than these two groups and Taiwan IA. We note that it was not clear in the original paper if the difference in coalescence dates is statistically significant, the finding was clearer in MSMC than in MSMC2 analysis, and there was no attempt to calculate expected cross- coalescence profiles using these methods for models incorporating many gene flows. Nevertheless, we applied this constraint as well in an attempt to understand whether, if we used a constraint system similar to that in Wang et al., we would obtain results that agreed with respect to the finding of Onge- related admixture ubiquitous among East Asian groups. → Applying these three additional constraints, we identified two models (among the 56 ones subjected to manual inspection) that satisfied all of them. The highest- ranking of those models is shown in Figure  4b and Figure 4—source data 2c (the second model), and it includes a 13% (deeply) European- related gene flow to the common ancestor of all East Asians, and gene flows from the Onge- related branch to just two East Asian groups: Nepal Chokhopani and China WLR LN. This model fits the data significantly better than the published model (p- value = 0.028). We do not claim that this is the correct model (indeed we are almost certain that it is not given the high topological diversity of fitting models), but it is not obviously wrong and differs in qualitatively important ways from the published one. The Wang et al., 2021 AG provides an illuminating example that helps us to understand the value added by AG construction. The AG construction process in Wang et al. followed a philosophy of not relying entirely on the allele frequency correlation data (not treating the genetic data as independent to explore how much new insight could come from genetic data alone). Instead, the study integrated other lines of genetic evidence as well as linguistic and archaeological insights explicitly into the AG construction process, with the goal of identifying models consistent with multiple lines of evidence. The fact that after this procedure a fitting graph was obtained is not of great interest, as it is essentially always possible to obtain a fit to allele frequency correlation data when enough admixture events are added. The important question is whether any of the emergent features of the graph that were not applied as constraints in the construction process—for example the evidence of ubiquitous Andamanese- related gene flow throughout East Asia suggesting a coastal route expansion that admixed with an interior route expansion proxied by Tianyuan—were stably inferred. Our analysis does not come to this finding consistently among well- fitting and plausible AGs. We conclude that an important feature of the published graph, that is variable levels of Andamanese- related ancestry found in all East Asians except for Siberians (Mongolia Neolithic) and the Upper Paleolithic Tianyuan (Figure 2 in Wang et al., 2021), is not supported by f- statistic analysis alone (Table 2), and indeed we are not aware of a single feature of the Wang et al., 2021 AG that is stably inferred beyond the constraints applied to build it. 8. Sikora et al., 2019 Two AGs inferred by Sikora et al., 2019 were constructed manually based on an SNP set derived from whole- genome shotgun data and incorporated 12 or 13 groups and 10 admixture events (Extended Data Figure 3f in the original study). One graph was focused on West Eurasians, and the Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 60 of 62 Evolutionary Biology | Genetics and Genomics Research article other one on East Eurasians, and both included a Neanderthal, a Denisovan, and an African group (Dinka). Although the chimpanzee outgroup was not included in the original graphs, we added it as it drastically constrains the topology search space. The following additional constraints were applied at the findGraphs model optimization stage: the Neanderthal and African groups were unadmixed and the Denisovan group had no more than one admixture event in its history. These three constraints match the features of the published graph. We also repeated topology searches without constraining the admixture status of the Neanderthal, Denisovan, and Dinka. Since no f3- statistics were negative when all sites available for each population triplet were used (the ‘useallsnps: YES’ option), we did not apply the algorithm that allows unbiased calculation of f3- statistics on pseudo- haploid data at the expense of loss of analyzed SNPs. In contrast to most other published graphs discussed above, gene flows in the graphs inferred by Sikora et al. do not have equal standing: four low- level gene flows (0–1%) connect the Neanderthal lineage to Upper Paleolithic lineages (Kostenki, Sunghir, Yana, Ust’-Ishim in the "Western" graph and Sunghir, Yana, Mal’ta, Ust’-Ishim in the "Eastern" graph). We repeated each topology search under two alternative settings: either keeping the number of admixture events at 10 to match the published graphs, or at 6 to match simplified versions of the published graphs lacking these low- level Neanderthal gene flows. We performed that modification to simplify the search space and to alleviate the overfitting problem which becomes severe if 10 gene flows across the graph are allowed (Supplementary file 1). Here, we compare LL and WR for the original published models and their simplified versions: the "Western" graph including chimpanzee (LL = 65.7, WR = 3.32 SE) vs. its simplified version (LL = 76.5, WR = 3.78 SE) and the "Eastern" graph including chimpanzee (LL = 85.3, WR = 3.11 SE) vs. its simplified version (LL = 102.4, WR = 4.16 SE). In both cases, we found no statistically significant differences in model fits (relying on the bootstrap model comparison method). In summary, topology search was repeated under 8 settings: for the "Western" or "Eastern" graphs, with no constrains on the admixture status or with the constraints specified above, and with 10 or 6 gene flows (Supplementary file 1). Below we focus on results for constrained models with 6 admixture events. In contrast, Figure 2 and Table 1 show results for constrained "Western" graphs with 10 admixture events. In the case of the constrained "Western" graphs with 6 admixture events, 1,000 findGraphs runs were performed, 894 distinct topologies were found, 4 models fit significantly better, and 151 models fit nominally better than the published one (Table 1, Supplementary file 1). We inspected those 155 topologies and identified 29 topologies (Figure 4—source data 3) that are temporally plausible and include no non- canonical gene flows from archaic groups such as Denisovan or a ghost archaic group to non- Africans. Sikora et al. came to the following striking conclusion relying on the "Western" AG (Table 2): the Mal’ta (MA1_ANE) lineage received a gene flow from the Caucasus hunter–gatherer (CaucasusHG_LP or CHG) lineage. However, in our findGraphs exploration this direction of gene Mal’ta) was supported by two of the 29 topologies, and the opposite gene flow flow (CHG direction (from the Mal’ta and East European hunter–gatherer lineages to CHG) was supported by the remaining 27 plausible topologies (Figure 4—source data 3). The highest- ranking plausible topology (Figure 4c) has a fit that is not significantly different from that of the simplified published model (p- value = 0.392). We note that the gene flow direction contradicting the graph by Sikora et al. was supported by a published qpAdm analyses (Lazaridis et al., 2016; Narasimhan et al., 2019), and qpAdm is not affected by the same model degeneracy issues that are the focus of this study. Considering the topological diversity among models that are temporally plausible, conform to robust findings about relationships between modern and archaic humans, and fit nominally better than the published model, we conclude that the direction of the Mal’ta- CHG gene flow cannot be resolved by AG analysis (Table 2). → Some important conclusions based on the "Eastern" graph also do not replicate across all plausible AGs (Table 2). In the case of the constrained "Eastern" graphs with 6 admixture events, 4,446 topology search iterations were performed, and 2,785 distinct topologies were found. Only 3 topologies fit significantly and 13 nominally better than the published one (p- value for the highest- ranking newly found model vs. the simplified published model = 0.112), and 9.8% of topologies fit not significantly worse than the published one (Table 1, Supplementary file 1). Of the topologies belonging to these groups, we inspected 116 best- fitting ones and identified 97 topologies that are temporally plausible and include no gene flows from archaic groups such as Denisovan or ghost archaic to non- Africans that are qualitatively different from the gene flows that are currently widely Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 61 of 62 Evolutionary Biology | Genetics and Genomics Research article accepted. The Sikora et al. "Eastern" AG had the following distinctive features that were used to support some conclusions of the study (Table  2): (1) the Mal’ta (MA1_ANE) and Yana (Yana_UP) lineages receive a gene flow from a common East Asian- associated source diverging before the ones contributing to the Devil’s Cave (DevilsCave_N), Kolyma (Kolyma_M), USR1 (Alaska_LP), and Clovis (Clovis_LP) lineages; (2) European- related ancestry in the Kolyma, USR1, and Clovis lineages is closer to Mal’ta than to Yana; (3) the Devil’s Cave lineage received no European- related gene flows, and Kolyma has less European- related ancestry than ancient Americans (USR1 and Clovis). Only feature 2 was universally supported by all the 97 plausible alternative models fitting significantly better, nominally better, or not significantly worse than the simplified published model, while feature 3 was supported by 83 of 97 plausible models, and feature 1 was supported by 28 of 97 plausible models (Table 2). We plotted 14 plausible graphs as examples of topologies supporting all three features, two features, or one feature of the published graph (Figure 4—source data 4). We note that all the "Eastern" graphs discussed here, both the published and alternative ones, have relatively poor absolute fits with WR above 4 or 5 SE. Increasing the number of gene flows to 10 allowed us to reach much better absolute fits (with WR as low as 2.42 SE), but that resulted in high topological diversity (on a par with some other case studies discussed above). In the case of the constrained "Eastern" graphs with 10 admixture events, 1,000 findGraphs runs were performed, and 1000 distinct topologies were found. Of these topologies, 13.2% fit significantly better, 30% nominally better, and 17.6% non- significantly worse than the published model (p- value for the highest- ranking newly found model vs. the published model <0.002) (Supplementary file 1). Maier, Flegontov et al. eLife 2023;12:e85492. DOI: https://doi.org/10.7554/eLife.85492 62 of 62 Evolutionary Biology | Genetics and Genomics
10.7554_elife.80854
RESEARCH ARTICLE MLL3 regulates the CDKN2A tumor suppressor locus in liver cancer Changyu Zhu1†, Yadira M Soto- Feliciano2,3*†, John P Morris1,4†, Chun- Hao Huang1, Richard P Koche5, Yu- jui Ho1, Ana Banito1, Chun- Wei Chen1, Aditya Shroff1, Sha Tian1, Geulah Livshits1, Chi- Chao Chen1, Myles Fennell1, Scott A Armstrong6, C David Allis2, Darjus F Tschaharganeh7*, Scott W Lowe1,8* 1Department of Cancer Biology and Genetics, Memorial Sloan Kettering Cancer Center, New York, United States; 2Laboratory of Chromatin Biology and Epigenetics, The Rockefeller University, New York, United States; 3Koch Institute for Integrative Cancer Research, Massachusetts Institute of Technology, Cambridge, United States; 4Department of Pharmacology, The University of North Carolina at Chapel Hill, Chapel Hill, United States; 5Center for Epigenetics Research, Memorial Sloan Kettering Cancer Center, New York, United States; 6Dana- Farber Cancer Institute, Boston, United States; 7Helmholtz- University Group "Cell Plasticity and Epigenetic Remodeling", German Cancer Research Center, Heidelberg, Germany; 8Howard Hughes Medical Institute, New York, United States *For correspondence: ysoto@mit.edu (YMS- F); d.tschaharganeh@dkfz.de (DFT); lowes@mskcc.org (SWL) †These authors contributed equally to this work Competing interest: See page 18 Funding: See page 19 Received: 07 June 2022 Preprinted: 09 June 2022 Accepted: 31 May 2023 Published: 01 June 2023 Reviewing Editor: Hao Zhu, University of Texas Southwestern Medical Center, United States Copyright Zhu, Soto- Feliciano, Morris et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Abstract Mutations in genes encoding components of chromatin modifying and remod- eling complexes are among the most frequently observed somatic events in human cancers. For example, missense and nonsense mutations targeting the mixed lineage leukemia family member 3 (MLL3, encoded by KMT2C) histone methyltransferase occur in a range of solid tumors, and heterozygous deletions encompassing KMT2C occur in a subset of aggressive leukemias. Although MLL3 loss can promote tumorigenesis in mice, the molecular targets and biological processes by which MLL3 suppresses tumorigenesis remain poorly characterized. Here, we combined genetic, epigenomic, and animal modeling approaches to demonstrate that one of the mechanisms by which MLL3 links chromatin remodeling to tumor suppression is by co- activating the Cdkn2a tumor suppressor locus. Disruption of Kmt2c cooperates with Myc overexpression in the development of murine hepatocellular carcinoma (HCC), in which MLL3 binding to the Cdkn2a locus is blunted, resulting in reduced H3K4 methylation and low expression levels of the locus- encoded tumor suppressors p16/Ink4a and p19/Arf. Conversely, elevated KMT2C expres- sion increases its binding to the CDKN2A locus and co- activates gene transcription. Endogenous Kmt2c restoration reverses these chromatin and transcriptional effects and triggers Ink4a/Arf- dependent apoptosis. Underscoring the human relevance of this epistasis, we found that genomic alterations in KMT2C and CDKN2A were associated with similar transcriptional profiles in human HCC samples. These results collectively point to a new mechanism for disrupting CDKN2A activity during cancer development and, in doing so, link MLL3 to an established tumor suppressor network. Editor's evaluation This paper convincingly shows that MLL3 regulates the CDKN2A tumor suppressor in MYC- driven liver cancers. The use of in vivo models and epigenomic analysis made the findings particularly robust. This work significantly advances our understanding of the function of MLL3 in cancer. Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 1 of 25 Research article Introduction Hepatocellular carcinoma (HCC) is a deadly primary liver cancer with a 5  year survival rate of only 18% (Jemal et al., 2017). HCC is currently the fourth most frequent cause of cancer- related mortality worldwide, and its incidence continues to grow (Llovet et al., 2021). Genomic alterations found in HCC are highly diverse and are characterized by promoter mutations in TERT (telomerase reverse transcriptase), amplifications, or chromosomal gains encompassing the MYC oncogene, activating hotspot mutations in CTNNB1 (β-catenin), and inactivating mutations and deletions in the TP53 and CDKN2A tumor suppressor genes (2017; Schulze et al., 2015). Among these alterations, genetic gain of MYC and inactivation of tumor suppressor p53 are known to cooperate to drive tumorigenesis in HCC (Molina- Sánchez et al., 2020). Mechanistically, oncogenic MYC activation triggers increased expression of the tumor suppressor ARF, one of two proteins encoded in CDKN2A in alternative reading frames. ARF binds to the E3 ubiquitin ligase MDM2 to prevent p53 degradation, leading to apoptosis to restrain MYC- driven tumorigenesis (Lowe and Sherr, 2003). However, it is unclear how the CDKN2A locus is regulated in response to MYC overexpression. Beyond these well- studied drivers, HCC frequently harbors mutations in one or more chromatin modifying enzymes, including MLL3 (encoded by KMT2C; Fujimoto et al., 2012; Kan et al., 2013). MLL3 is a component of the COMPASS- like complex that has structural and functional similarities to the developmentally essential Drosophila Trithorax- related complex (Schuettengruber et  al., 2017). This multiprotein complex controls gene expression through its histone H3 lysine 4 (H3K4) methyltransferase activity, which establishes chromatin modifications most often associated with tran- scriptional activation (Shilatifard, 2012). Most studies have shown that MLL3 and its paralog MLL4 (encoded by KMT2D) typically catalyze H3K4 monomethylation (H3K4me1) at enhancers (Herz et al., 2012; Hu et al., 2013), while the MLL1/2 complex is responsible for H3K4 trimethylation (H3K4me3) at promoters and enhancers in a locus- specific manner (Denissov et al., 2014; Rickels et al., 2016; Wang et al., 2009). While less characterized, MLL3/4 regulation of promoter activity is emerging as an additional mech- anism connecting the COMPASS- like complex to gene expression. Some publications report that H3K4me1 enrichment at promoters has been associated with gene repression (Cheng et al., 2014), and MLL3 inactivation decreases H3K4me3 levels at the promoters of metabolism- related genes in normal murine livers (Valekunja et  al., 2013) and human liver cancer cells (Ananthanarayanan et al., 2011). Furthermore, a recent study in leukemia cells demonstrated that MLL3 and MLL4, in the absence of MLL 1/2 complex, are capable of binding to promoters to activate tumor suppressor genes (Soto- Feliciano et al., 2023). These divergent results suggest that the genomic binding pattern and functions of MLL3 are highly context dependent. Notably, HCC also harbors mutations in KMT2D (Cleary et  al., 2013), while KDM6A/UTX, an H3K27 demethylase within the COMPASS- like complex, has been functionally established as a potent tumor suppressor in pancreatic and liver cancers (Revia et  al., 2022). These observations suggest that epigenetic- based mechanisms of gene regulation controlled by the MLL3 complex may constrain HCC development. However, because chromatin regulators such as ARID1A often exhibit context- specific tumor suppressive and oncogenic roles in liver cancer development (Sun et al., 2017), it is unclear whether MLL3 is a bona fide tumor suppressor in HCC. We therefore employed mouse models of HCC to investigate the molecular targets of MLL3 and the biological processes it affects. Results MLL3 is a tumor suppressor in Myc-driven liver cancer To better understand the functional significance of genes commonly inactivated in HCC, including a number of chromatin regulators, we selected 12 genes with recurrent inactivating mutations in human HCC (Cancer Genome Atlas Research Network, 2017; Ahn et al., 2014; Fujimoto et al., 2012; Figure 1—figure supplement 1A) and performed a CRISPR- based in vivo screen to determine whether they behave as tumor suppressors in HCC. Specifically, the screen tested whether loss of each of these 12 genes would drive hepatic tumorigenesis in cooperation with Myc—one of the most frequently gained and/or amplified oncogenes in HCC (Huang et  al., 2014). We applied hydrody- namic tail vein injection (HTVI) in wild- type mice to directly introduce genetic manipulations into adult Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 2 of 25 Cancer Biology Research article hepatocytes in vivo (Bell et al., 2007). We introduced both a transposon vector for stable genomic integration of oncogenic Myc cDNA and plasmids designed for transient expression of Cas9 and single guide RNAs (sgRNAs; a mix of two for each gene; Figure 1B; Largaespada, 2009; Moon et al., 2019; Tschaharganeh et al., 2014; Xue et al., 2014). 3 months after HTVI, only sgKmt2c resulted in liver tumor formation with high penetrance (Figure 1—figure supplement 1B), suggesting that MLL3 likely acts as a tumor suppressor to constrain Myc- driven liver cancer. Supporting this idea, KMT2C mutations co- occur with MYC genomic gains and amplifications in human HCC tumors (Figure 1A). To validate and extend the results from the screen, we applied the same approach to test whether the screen phenotype could be recapitulated with oncogenic Myc and single Kmt2c- targeted sgRNAs (Figure 1B). Mice injected with an Myc cDNA transposon combined with either of two independent Cas9/Kmt2c sgRNAs (Myc; sgKmt2c.1 or Myc; sgKmt2c.2) developed liver tumors, with a slightly later onset and slightly longer survival than mice receiving the Myc transposon combined with an sgRNA targeting Trp53 (Myc; sgTrp53; Figure  1C and D). In contrast, mice injected with Myc and a control sgRNA (sgChrom8) did not succumb to disease over the observation period (Figure 1C). These findings were confirmed in an independent cohort of mice (Figure  1—figure supplement 2b). Analyses of tumor- derived genomic DNA revealed insertions and deletions (indels) in either Kmt2c or Trp53 depending on the genotype of tumor- derived cells (Figure 1—figure supplement 2B). DNA sequencing of the CRISPR- targeted region from two independent Myc; sgKmt2c tumors revealed indels predicted to generate premature stop codons (Figure 1—figure supplement 2C). In one case, the indel was heterozygous, implying that even partial suppression of Kmt2c can promote tumorigenesis. In support of this, GFP- linked Kmt2c shRNAs efficiently cooperated with Myc overex- pression to drive liver cancer, producing tumors with 50–80% reduction in Kmt2c mRNA expression (Figure 1—figure supplement 2D–G). shKmt2c.2 resulted in less potent knockdown than shKmt2c.1 yet produced faster tumor formation, suggesting that, as in acute myeloid leukemia (Chen et  al., 2014), MLL3 can likely act as a haploinsufficient tumor suppressor in liver cancer (Figure 1—figure supplement 2E, G). Apart from MYC, CTNNB1 (β-catenin) is a frequently mutated oncogene in human HCC (Rebou- issou et  al., 2016), although the co- occurrence between CTNNB1 and KMT2C mutations was not statistically significant (Figure 1A). To test whether MLL3 loss can cooperate with oncogenic CTNNB1 to promote liver tumorigenesis, we performed analogous HTVI of a transposon vector expressing constitutively active β-catenin (Ctnnb1- N90; Tward et al., 2007) in combination with Kmt2c- or Trp53- targeted sgRNAs (Figure 1—figure supplement 3A). However, no tumor formation was observed in mice that received the Kmt2c- targeted sgRNAs (Figure 1—figure supplement 3B), indicating that unlike p53, the tumor- suppressive role of MLL3 is specific to the oncogene and contexts. MLL3 loss alters the chromatin landscape of liver cancer cells MLL3 and MLL4 are histone methyltransferases that can deposit the H3K4 monomethylation mark at genomic enhancers and intergenic regions during organ development (Hu et  al., 2013). However, more studies indicate that MLL3 and MLL4 are also capable of binding to promoter regions (Chen et al., 2014; Dhar et al., 2016; Wang et al., 2010), especially in the context of cancer (Soto- Feliciano et al., 2023). To determine the genomic binding patterns of MLL3 in HCC, we performed MLL3 chro- matin immunoprecipitation (ChIP)- sequencing (ChIP- Seq) analysis in Myc; sgKmt2c (sgKmt2c.1 which generates heterozygous or homozygous indels) and Myc; sgTrp53 liver cancer cell lines. Compared to the sgTrp53 cells, sgKmt2c cells had a marked reduction in MLL3 chromatin binding at a subset of genomic loci (Figure  2A). Approximately 40% of the peaks that were selectively lost in Kmt2c- deficient cells occurred at promoter regions, whereas unchanged MLL3 peaks between the two geno- types were more likely to be within intergenic regions (Figure 2B, Figure 2—figure supplement 1). Therefore, our data suggest that, beyond the canonical action of MLL3 at gene enhancers, MLL3 can also occupy promoter regions in Myc- induced liver cancer. Of note, the residual ChIP- seq signal observed in the sgKmt2c cells most likely reflects the binding of MLL4 and/or remnant MLL3 since the antibody used in these experiments can recognize both MLL3 and MLL4 proteins (Dorighi et al., 2017). Nonetheless, the downregulated peak signals in Myc; sgKmt2c cells were specifically due to MLL3 disruption. Similar to the Drosophila Trithorax- related complex (Schuettengruber et al., 2017), the mamma- lian MLL3 and MLL4 complexes facilitate gene transcription by establishing permissive modifications Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 3 of 25 Cancer Biology Research article Figure 1. MLL3 constrains Myc- driven liver tumorigenesis. (A) Oncoprints displaying genomic mutations and deletions of KMT2C and TP53, gains and amplifications of MYC, and activating CTNNB1 mutations in merged publicly available datasets (TCGA, MSK, INSERM, RIKEN, AMC, and MERCi) of 1280 sequenced hepatocellular carcinomas, and the table showing their relationships. p- Values were calculated by Fisher exact tests. (B) Schematic for hydrodynamic tail vein injection (HTVI) of gene delivery into murine livers. Vectors permitting stable expression of Myc transposon (top) and transient expression of Cas9 and single guide RNAs (sgRNAs) targeting putative tumor suppressors (bottom) via sleeping beauty transposase were introduced into hepatocytes by HTVI. (C) Survival curves of mice injected with Myc transposon and pX330 expressing two independent sgRNAs targeting Kmt2c after HTVI (Myc; sgKmt2c.1, n=5; Myc; sgKmt2c.2, n=5). Myc; sgTrp53 (n=5), and Myc; sgChrom8 (n=5) serve as controls. Survival curves were compared using log- rank tests. (D) Representative images (left, liver macro- dissection, scale bar: 0.5 cm; right, H&E staining, scale bar: 100 μm) of mouse liver tumors generated by HTVI delivery of Myc transposon and in vivo gene editing. The dashed lines indicate the boundaries between liver tumors and non- tumor liver tissues. The online version of this article includes the following source data and figure supplement(s) for figure 1: Figure supplement 1. In vivo screen identifies MLL3 as a tumor suppressor in Myc- driven liver cancer. Figure supplement 2. Suppression of Kmt2c by CRISPR or RNAi promotes Myc- driven liver cancer. Figure supplement 2—source data 1. Original gel for surveyor assays in Figure 1—figure supplement 2B. Figure supplement 3. MLL3 loss does not cooperate with CTNNB1 oncogene to drive liver cancer. Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 4 of 25 Cancer Biology Research article Figure 2. MLL3 disruption alters the chromatin and transcriptional landscape of liver cancer cells. (A) Tornado plots showing MLL3 chromatin immunoprecipitation- sequencing (ChIP- Seq) signal (peaks) that were down or remained unchanged in Myc; sgKmt2c cells relative to Myc; sgTrp53 cells. Center: transcriptional start site (TSS). (B) Alluvial plot showing the percentages of MLL3 ChIP- Seq peaks in different genomic elements in Myc; sgKmt2c vs Myc; sgTrp53 cells, including 16,999 peaks down, 48,815 peaks unchanged, and 265 peaks up in sgKmt2c cells. Promoter regions were defined as Figure 2 continued on next page Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 5 of 25 Cancer Biology Research article Figure 2 continued TSS±2 kb. (C) Heatmaps of histone modification ChIP- Seq signals (H3K4me3, H3K4me1, H3K27ac, left panel) and MLL3 ChIP- Seq signal (right panel) at promoter or intergenic regions in three independent Myc; sgKmt2c and Myc; sgTrp53 liver tumor- derived cell lines. Cluster 1: loss of promoter and enhancer activity (loss of H3K4me3, H3K4me1, and H3K27ac); cluster 2: gain of enhancer activity (gain of H3K4me1 and H3K27ac); and cluster 3: gain of promoter activity (increase of H3K4me3). Representative top five loci for each cluster were listed on the right. (D) Volcano plot of differentially expressed genes revealed by RNA- sequencing of three independent Myc; sgKmt2c and Myc; sgTrp53 hepatocellular carcinoma (HCC) cell lines. Genes in sgKmt2c cells with more than twofold expression change and exceeding adjusted p- value<10–5 are color- labeled (orange: upregulated; green: downregulated). Some differentially expressed genes are labeled with gene symbols, and p53 targets are bolded. The online version of this article includes the following figure supplement(s) for figure 2: Figure supplement 1. MLL3 deficiency disrupts its binding at promoters in liver cancer cells. Figure supplement 2. MLL3 disruption impacts transcriptional and histone modification profiles in liver tumors. on histone H3K4 via the MLL3 and MLL4 methyltransferase (Shilatifard, 2012). To determine whether MLL3 disruption impacts the local or global chromatin landscape of HCC cells, we performed ChIP- Seq analyses for H3K4 methylation and H3K27 acetylation in six independently derived tumor cell lines: three each for Myc; sgKmt2c and Myc; sgTrp53 (Figure 2C). Cluster analysis on genomic areas revealed three clusters of genomic loci that showed enrichment or depletion between Myc; sgKmt2c and Myc; sgTrp53 tumor cells for each tested histone modification (Figure 2C, Figure 2— figure supplement 2A–C). Loci in cluster 1 (reduced H3K4me3, H3K4me1, and H3K27ac in sgKmt2c cells) showed the most pronounced differences in chromatin modifications between the two liver tumor genotypes. In contrast, the loci in cluster 2 showed increased H3K4me1 and H3K27ac marks, most of which mapped to intergenic regions. The loci in cluster 3 showed increased H3K4me3 and included some p53 target genes such as Cdkn1a and Eda2r. To determine whether these drastic changes in the chromatin landscape were associated with changes in MLL3 binding, we integrated the chromatin modifications results with our MLL3 ChIP- Seq results (Figure 2C). Interestingly, the loci in cluster 1, which displayed the most substantial changes in histone modifications, involved genes that showed enriched MLL3 binding in Myc; sgTrp53 cells compared to the Myc; sgKmt2c genotype. These data support a model whereby MLL3 binding to these loci facilitates the acquisition of a chromatin environment conducive for active gene transcription. MLL3 regulates specific tumor suppression programs in liver cancer cells Transcriptional profiling helped hone in on potentially critical targets of MLL3. Specifically, we deter- mined the output of these chromatin landscape changes by transcriptional profiling of the same set of Myc; sgTrp53 and Myc; sgKmt2c liver cancer cell lines described above. Despite the broad binding of MLL3 across the genome, we found only 248 differentially expressed genes (DEGs): 132 significantly upregulated (p<0.05, log2 fold- change  >2) and 116 significantly downregulated (p<0.05, log2 fold- change <−2) in Myc; sgKmt2c liver tumor cells compared to Myc; sgTrp53 controls. As predicted, transcripts encoding p53 and p53 target genes such as Ccng1, Cdkn1a, and Zmat3 (Bieging- Rolett et  al., 2020) were upregulated in Myc; sgKmt2c cells, consistent with nonsense- mediated decay of truncated p53 transcripts and a concomitant reduction in p53 effector genes. Strikingly, some of the downregulated genes in Myc; sgKmt2c lines mapped to loci enriched in cluster 1, including Cdkn2a, Bmp6, and Lrp2 (Figure 2C–D, Figure 2—figure supplement 2D, E). Of note, sgKmt2c did not lead to compensatory changes in the transcript levels of other major components of the COMPASS- like complexes, including Mll4 (Kmt2d), Utx (Kdm6a), Mll1 (Kmt2a), and Mll2 (Kmt2b; Figure 3—figure supplement 1A), suggesting that the alterations in MLL3 binding, histone modifica- tion, and transcription were specifically attributed to MLL3 disruption. We reason that the mediators of MLL3 actions in tumor suppression should be within cluster 1 with reduced transcription and MLL3 binding in Kmt2c- deficient cells. To further characterize the gene repertoire directly regulated by MLL3 genomic binding, we integrated the results of MLL3 ChIP- seq and RNA- seq. Specifically, we selected downregulated DEGs that show concordant decreased binding of MLL3 in Myc; sgKmt2c lines and subjected them to gene ontology analysis. Apart from the cluster 1 genes noted above, the integrative analysis revealed multiple MLL3- regulated tumor suppressive programs (Figure 3A, Figure 3—figure supplement 1B), including both cell- autonomous Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 6 of 25 Cancer Biology Research article A B C Mouse HCC sgKmt2c vs sgTrp53 up-regulated genes 0.5 - 0.4 - 0.3 - 0.2 - 0.1 - 0.0 - ) S E ( e r o c s t n e m h c i r n E CDKN2A mutated TP53 mutated Human HCC KMT2C vs TP53 mutated up-regulated genes 0.5 - 0.4 - 0.3 - 0.2 - 0.1 - 0.0 - ) S E ( e r o c s t n e m h c i r n E Mouse HCC sgKmt2c vs sgTrp53 down-regulated genes 0.00 - -0.05 - -0.10 - -0.15 - -0.20 - -0.25 - -0.30 - -0.35 - CDKN2A mutated TP53 mutated Human HCC KMT2C vs TP53 mutated down-regulated genes 0.0 - -0.1 - -0.2 - -0.3 - -0.4 - -0.5 - ) S E ( e r o c s t n e m h c i r n E ) S E ( e r o c s t n e m h c i r n E CDKN2A mutated TP53 mutated CDKN2A mutated TP53 mutated Figure 3. MLL3 regulates specific transcription programs including tumor suppressor CDKN2A. (A) Network plot showing the major biological processes and related genes directly regulated by MLL3 binding. p- Values and cluster sizes were calculated by the integrative analyses of RNA- seq and MLL3 chromatin immunoprecipitation- sequencing (ChIP- Seq), as detailed in the Materials and methods. (B) Gene set enrichment analysis (GSEA) plots of transcriptional signatures derived from mouse hepatocellular carcinomas (HCCs; Myc; sgKmt2c vs Myc; sgTrp53) against transcriptomics of HCCs Figure 3 continued on next page Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 7 of 25 Cancer Biology Research article Figure 3 continued with CDKN2A vs TP53 mutations. (C) GSEA plots of transcriptional signatures derived from KMT2C mutated/deleted human HCCs against the ones with CDKN2A mutations or homozygous deletions. HCCs with TP53 mutations were used as the controls for both comparisons. Normalized enrichment scores (NES) and false discovery rate (FDR) q- values were calculated by GSEA. The online version of this article includes the following figure supplement(s) for figure 3: Figure supplement 1. MLL3 loss impacts specific transcription programs. Figure supplement 2. CDKN2A and KMT2C mutations cause similar transcriptional changes in human hepatocellular carcinoma (HCC). mechanisms (cellular metabolism) and non- autonomous mechanisms (interaction with extracellular matrix and immune system). KMT2C and CDKN2A mutations result in similar transcriptomes in human HCC One genomic locus that stood out in our integrative analysis was Cdkn2a, which encompasses both the p16/Ink4a and p19/Arf (p14 in human) tumor suppressors (Gil and Peters, 2006). CDKN2A is located on the human chromosome 9p and is deleted or epigenetically silenced in many cancer types (Sherr, 2012), including HCC (2017). While MLL3 likely regulates a plethora of genes that contribute to its tumor- suppressive potential, the well- defined and potent antitumor functions of Cdkn2a- encoded proteins make them attractive candidates as functionally relevant MLL3 effectors. Furthermore, in our analysis publicly available genomic data on 1280 HCC samples, we found that CDKN2A alterations, like KMT2C alterations, showed significant co- occurrence with MYC gains and amplifications. However, we were unable to conduct a meaningful test of mutual exclusivity between CDKN2A and KMT2C alterations (Figure 3—figure supplement 2A), given the constraints of sample size and the modest frequencies of alteration in each gene. Further dissection of transcrip- tional profiling datasets from human and mouse HCCs harboring known gene alterations using gene set enrichment analysis (GSEA) revealed that human tumors with CDKN2A deletions transcriptionally resembled both mouse and human HCC harboring KMT2C alterations (Figure  3B and C) but not those harboring RB1 loss (Figure 3—figure supplement 2B), even though the tumor suppressor RB1 is regulated by CDKN2A/p16INK4A and their genomic alterations exhibit mutual exclusivity in multiple cancer types (Knudsen et al., 2020). While we cannot rule out the possibility that other factors drive these associations, our results support a biologically meaningful relationship between MLL3 and CDKN2A. Cdkn2a locus is a genomic and transcriptional target of MLL3 in liver cancer To explore the relationship between MLL3 and Cdkn2a locus in more detail, we tested whether genes encoded by Cdkn2a were direct targets of MLL3- regulated transcription. Indeed, Cdkn2a is a cluster 1 locus that, in Myc; sgKmt2c cancer cells, displays significant reduction in (1) expression, (2) H3K4me1/3 and H3K27ac levels, and (3) MLL3 binding at the Cdkn2a promoter compared with Myc; sgTrp53 cells (Figure 2C–D, Figure 4A). Of note, MLL3 binding peaks were also observed within the gene body of Cdkn2a. The differential expression of Ink4a and Arf was confirmed by qPCR, immu- noblotting, and ChIP- qPCR analyses on multiple Myc; sgKmt2c and Myc; sgTrp53 liver cancer lines (Figure 4B, Figure 4—figure supplement 1A, B). These results imply that Cdkn2a locus is a genomic and transcriptional target of MLL3 in liver cancer cells. Since the Myc; sgTrp53 and Myc; sgKmt2c cells we studied above are not isogenic, we performed a series of additional experiments to demonstrate a direct transcriptional effect of MLL3 on the Cdkn2a locus. Because p53 inactivation can lead to compensatory increases in Ink4a and Arf expression (Stott et al., 1998), representing an alternative possibility accounting for the observed difference of Cdkn2a expression in sgTrp53 vs sgKmt2c cells. However, p53 suppression in sgKmt2c cells produced only a subtle and inconsistent effect on the expression of Ink4a and Arf, whereas Kmt2c suppression in sgTrp53 cells consistently attenuated p16Ink4a and p19Arf protein levels (Figure  4—figure supple- ment 1C, D). As another means of ruling out the p53 pathway as an explanation for altered Cdkn2a expression in Myc; sgKmt2c cells, we tested the ability of MLL3 to regulate Cdkn2a transcripts in an orthogonal liver cancer model driven by Myc and inactivation of Axin1, which is a well- defined Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 8 of 25 Cancer Biology Research article Figure 4. CDKN2A locus is a genomic and transcriptional target of MLL3 in liver cancer. (A) Genome browser tracks for MLL3 and H3K4me1 chromatin immunoprecipitation- sequencing (ChIP- Seq) in Myc; sgTrp53 (red) and Myc; sgKmt2c (blue) hepatocellular carcinoma (HCC) cell lines at the Cdkn2a locus. (B) qPCR analysis for mRNA expression of Ink4a and Arf from three independent Myc; sgKmt2c and Myc; sgTrp53 HCC lines (n=3 cell lines each genotype). Values are shown as mean ± SD. ***=p<0.001 (unpaired two- tailed t- test). (C) Schematic for CRISPR activation (CRISPRa) system of nuclease- Figure 4 continued on next page Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 9 of 25 Cancer Biology Research article Figure 4 continued dead Cas9 (dCas9) and VP64- p65- Rta (VPR) guided by sgKMT2C to activate KMT2C expression in human HLE HCC cell line. (D) qPCR analysis for mRNA expression of KMT2C in HLE cells with sgGFP (control) or two different CRISPRa single guide RNAs (sgRNAs) targeting KMT2C (n=4 cell lines each genotype). Each data point represents the average of technical duplicates. Data are shown as mean ± SEM. ***=p<0.001 (one- way ANOVA followed by post- hoc t- tests). (E) Genome browser tracks for MLL3 ChIP- Seq at the CDKN2A locus in HLE cells with sgGFP (control, black) or sgKmt2c.1 (blue). (F and G) qPCR analysis for mRNA expression of (F) INK4A and (G) ARF in HLE cells with sgGFP (control) or sgKMT2C (n=4). Each data point represents the average of technical duplicates. Data are shown as mean ± SEM. ***=p<0.001 (one- way ANOVA followed by post- hoc t- tests). The online version of this article includes the following source data and figure supplement(s) for figure 4: Figure supplement 1. MLL3 directly regulates Cdkn2a expression in liver cancer cells. Figure supplement 1—source data 1. Original western blots for Figure 4—figure supplement 1A,C,D,E. tumor suppressor that negatively regulates β-catenin activity in HCC (Satoh et al., 2000). Liver cancer cells produced by hydrodynamic delivery of the Myc transposon vector and Axin1 sgRNAs displayed reduced Ink4a and Arf expression upon Kmt2c knockdown without targeting p53 (Figure 4—figure supplement 1E,F). Importantly, MLL3 binding peaks at the Cdkn2a locus were also detected in Myc; sgAxin1 liver cancer cells (Figure 4—figure supplement 1G), suggesting that Cdkn2a transcription is directly regulated by MLL3 rather than an indirect outcome of p53 loss. These data imply that MLL3 supports a chromatin environment at the Cdkn2a locus that facilitates the transcription of both Ink4a and Arf and raises the possibility that these factors contribute to the tumor suppressor activity of MLL3 in liver cancer. We next set out to determine whether MLL3 binding is sufficient to induce transcriptional activa- tion of the CDKN2A locus and, in doing so, extend our analysis to human liver cancer cells. As the KMT2C transcript is too large (14,733 bp) for cDNA transduction, we turned to the CRISPR activation (CRISPRa) system (Chavez et al., 2015) in a human hepatocellular carcinoma cell line (HLE). Following stable integration of the nuclease dead Cas9 fused to the VP64- p65- Rta (VPR) transcriptional acti- vator, cells were transduced with two orthogonal sgRNAs targeting the human KMT2C promoter (or, as control, transduced with sgRNA against GFP; Figure  4C). Cells expressing the KMT2C sgRNAs showed a marked and specific increase in the expression of endogenous KMT2C, but not of KMT2D or TP53 (Figure 4D, Figure 5—figure supplement 1A, B), which was accompanied by an increase in MLL3 binding to the CDKN2A locus (Figure 4E) and transcriptional upregulation of both CDKN2A transcripts (Figure  4F and G). Therefore, MLL3 directly binds and co- activates transcription of the CDKN2A locus in human liver cancer cells. MLL3 mediates oncogene-induced apoptosis in a Cdkn2a-dependent manner The above results raise the possibility that the Cdkn2a products, INK4A and ARF, may contribute to the tumor suppressive activity of MLL3. In this regard, Myc overexpression in primary cells (mouse embryonic fibroblasts; MEFs) often triggers apoptosis (Evan et  al., 1992), and this in turn limits tumorigenesis in a manner that is dependent on Cdkn2a (Zindy et  al., 1998). This pathway also suppresses liver tumorigenesis since concomitant disruption of Ink4a and Arf using CRISPR, or germ- line deletion of Arf alone, cooperated with Myc overexpression to rapidly promote tumor develop- ment (Figure 5—figure supplement 1C). Similarly, Kmt2c suppression also attenuated MYC- induced apoptosis, as shown by tumor histology and apoptosis by TUNEL assay (Negoescu et  al., 1997), 5  days after hydrodynamic delivery of transposon vectors encoding Myc together with GFP- linked shRNAs targeting Kmt2c (or Renilla luciferase as a control; Figure 5A and B). This difference in apop- tosis correlated with an increase in retention of GFP- shKmt2c expressing cells 10 days after injection (Figure 5—figure supplement 1D, E). Altogether, these results show that Kmt2c suppression impairs Myc- induced apoptosis in vivo in a manner that is reminiscent of the anti- apoptotic effects of Cdkn2a loss in the context of aberrant Myc activation (Eischen et  al., 1999; Jacobs et  al., 1999; Schmitt et al., 1999). To model the interaction between Myc overexpression, MLL3 function, and Cdkn2a regulation, we transduced liver progenitor cells (LPCs) with retroviral vectors encoding Myc linked to a reverse tetracycline transactivator (rtTA3), together with doxycycline (dox)- inducible Kmt2c shRNAs to enable reversible Kmt2c silencing (Figure  5—figure supplement 2A). Infection of LPCs with Myc in the Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 10 of 25 Cancer Biology Research article Figure 5. MLL3 mediates oncogene- induced apoptosis in a Cdkn2a- dependent manner. (A) Representative images of TUNEL- positive nuclei (red staining) in murine livers 5 days after hydrodynamic injection of the indicated vector combinations. DAPI(4′,6- diamidino- 2- phenylindole) was used to visualize nuclei. (B) Quantification of TUNEL- positive nuclei in mouse livers 5 days after hydrodynamic tail vein injection (HTVI) of the indicated vector combinations. Data points represent the number of TUNEL- positive cells in five different high- power fields in three independent murine livers per group. ***=p<0.001 (one- way ANOVA followed by post hoc t- tests). (C) Chromatin immunoprecipitation (ChIP)- qPCR analysis for H3K4me3 signals at Arf Figure 5 continued on next page Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 11 of 25 Cancer Biology Research article Figure 5 continued and Ink4a promoters 4 days after doxycycline (dox) withdrawal in Myc- rtTA3; TRE- shKmt2c cells. Values are mean ± SD from technical replicates (n=3), and the experiments were conducted in two independent liver progenitor cell (LPC) lines with different shKmt2c. (D) qPCR analysis for mRNA expression of Arf and Ink4a 4 days after dox withdrawal in two independent LPC lines with different shKmt2c. Values are mean ± SD from technical replicates (n=3). ***=p<0.001 and **=p<0.01 (unpaired two- tailed t- test). (E) Representative images of colony formation assay of the indicated cell lines 5 days after dox withdrawal. (F) Quantification of colony formation assay. Values are mean ± SD of three independent experiments with two independent LPC lines. *=p<0.05 (unpaired two- tailed t- test). (G) Time course analysis of Draq7- positive (dead or permeabilized) cells as a fraction of Venus- positive, Myc- rtTA3; TRE- shKmt2c cells expressing constitutive shRNAs targeting both Ink4a and Arf (shCdkn2a) or Renilla luciferase (shRen) off and on dox. Values represent mean ± SEM of triplicate wells of each genotype at each timepoint of two independently derived LPC lines, infected with either shRen or shCdkn2a. *=p<0.05 (unpaired two- tailed t- test of final average percentage Draq7+/GFP+). NS, not significant (p>0.05). The online version of this article includes the following source data and figure supplement(s) for figure 5: Figure supplement 1. Kmt2c suppression reduces cell clearance upon enforced Myc expression in vivo. Figure supplement 2. Endogenous Kmt2c restoration triggers apoptosis and is accompanied by increased Cdkn2a expression. Figure supplement 2—source data 1. Original western blots for Figure 5—figure supplement 2B,F,G. presence of MLL3 (i.e. cells infected with Myc- rtTA3 and a dox- inducible shRNA targeting Renilla luciferase) acutely activated INK4A and ARF expression (Figure 5—figure supplement 2B), and these cells could not be maintained in culture. Phenocopying the ability of Myc and Kmt2c suppression to transform liver cells in vivo, combined Myc and shKmt2c expression facilitated the persistent growth of cells maintained on Dox (Figure 5—figure supplement 2C,D). By contrast, dox withdrawal induced Kmt2c mRNA expression and H3K4me3 deposition at the Cdkn2a promoters, ultimately leading to elevations in Arf and Ink4a mRNA and protein (Figure  5C and D), reduced colony formation, and increased apoptosis (Figure  5—figure supplement 2D–F). Furthermore, constitutive shRNA- mediated knockdown of Arf and Ink4a through targeting of the shared exon 2 (shCdkn2a) significantly rescued colony- forming capacity and prevented cell death following Kmt2c restoration as determined by time- lapse microscopy of cells cultured with a fluorescent dye that stains dead cells (Figure 5E–G, Figure 5—figure supplement 2G). These data support a model whereby a prominent tumor suppres- sive output of MLL3 in liver cancer involves direct upregulation of Cdkn2a that, when impaired, atten- uates the MYC- induced apoptotic program and permits tumor progression. Figure 6. Model of MLL3 as a tumor suppressor in liver cancer. MLL3 restricts MYC- induced liver tumorigenesis by directly activating the Cdkn2a locus to mediate tumor cell apoptosis. Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 12 of 25 Cancer Biology Research article Discussion Our study combined genetic, epigenomic, and animal modeling approaches to identify Cdkn2a as an important regulatory target of MLL3 in both mouse and human liver cancers. Our results support a model whereby oncogenic stress, herein produced by MYC, leads to an increase in the binding of MLL3 to the CDKN2A locus, an event that is associated with the accumulation of histone marks linked to the biochemical activity of MLL3- containing complexes and conducive to gene activation (Figure 6). Accordingly, these events are accompanied by transcriptional upregulation of two key Cdkn2a gene products, Ink4a and Arf. Moreover, suppression of Kmt2c phenocopies the effects of Cdkn2a inac- tivation in abrogating MYC- induced apoptosis. Conversely, suppression of Cdkn2a diminishes the anti- proliferative effects of Kmt2c restoration. As such, our results establish a conserved epistatic rela- tionship between the chromatin modifier MLL3 and a well- characterized tumor suppressor network. The epistatic relationship described above might be expected to lead to mutual exclusivity of KMT2C and CDKN2A alterations; however, we did not observe significant mutual exclusivity in liver cancer, which is likely due to insufficient samples sizes needed to obtain statistical power. Alterna- tively, other functionally important components linked to the CDKN2A locus could produce CDKN2A- independent forces that drive selection for chromosome 9p deletions, including type I interferon genes, CDKN2B, and MTAP (Barriga et al., 2022). Alternatively, mutual exclusivity between KMT2C and CDKN2A alterations would be expected only under circumstances where CDKN2A action is the most dominant MLL3 effector. Indeed, it seems likely that multiple downstream genes, including factors involved in interactions with stromal and immune populations, contribute to MLL3- driven tumor suppression, and their relative importance may vary between cell and tissue types. Such a vari- able output in cancer- relevant gene regulation has been noted for other chromatin regulators that, at the extreme, serve as pro- oncogenic factors in some contexts and tumor suppressors in others (Fountain et al., 1992; Schmid et al., 2000; Sun et al., 2017; Xia et al., 2021). Furthermore, our observation that Kmt2c deficiency cooperated with MYC but not CTNNB1 to drive HCC highlights such context specificity and is in line with recent findings that chromatin context could favor particular oncogenic alterations over others (Weiss et al., 2022). UTX (KDM6A), MLL3 (KMT2C), and MLL4 (KMT2D), the core catalytic components of the COMPASS- like complex, are all considered tumor suppressors, with frequent loss- of- function genomic alterations found in a broad spectrum of human cancers (Revia et al., 2022; Sze and Shilatifard, 2016). While each of these components regulates redundant sets of genes (Hu et  al., 2013; Lee et  al., 2009), they may exert their tumor suppressive functions through different mechanisms. In liver and pancreas cancer models, UTX can control the expression of negative regulators of mTOR such as DEPTOR, and its disruption prevents their transcription and facilitates tumorigenesis through increased mTORC1 activity (Revia et al., 2022). Additionally, while the mechanisms of MLL4 activity have not been exam- ined in liver cancer, studies suggest that MLL4 suppresses skin carcinogenesis by promoting lineage stability and ferroptosis independently of MLL3 (Egolf et  al., 2021). Our study demonstrates that MLL3 is both necessary and sufficient for efficient transcriptional activation of the CDKN2A locus that drives oncogene- induced apoptosis. The molecular basis for this heterogeneity in effector output remains to be determined, but it seems likely that different subsets of target genes are preferen- tially disabled by haploinsufficiency of individual components and/or subject to compensation by the remaining COMPASS complex activities. Systematic studies comparing the binding, histone modifi- cations, and transcriptional output of cells across a spectrum of allelic configurations of COMPASS complex factors will be needed to achieve a more holistic understanding of their functions and inter- actions in different contexts. The most well- established role for MLL3/4- UTX- containing complexes is the control of H3K4 monomethylation at enhancers during development (Herz et al., 2010; Hu et al., 2013). While our ChIP- Seq studies also revealed binding of MLL3/4 to enhancers in liver tumor cells, an even larger fraction of genes—including Cdkn2a—showed MLL3/4 chromatin enrichment at gene promoters, and indeed, transcription of this class of genes was most affected by Kmt2c disruption. Interestingly, Kmt2c suppression preferentially limited the MLL3/4 enrichment at promoters and shifted residual complex binding toward intergenic regions. Such dynamic regulation of distinct cis- acting elements by the MLL3/4 complex has also been observed in other contexts (Cheng et al., 2014; Soto- Feliciano et al., 2023), where the non- canonical binding of MLL3/4 at promoters is a recurrent tumor suppressive mechanism in cancer cells. MLL3/4 has also been observed to bind within the exons and introns, which Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 13 of 25 Cancer Biology Research article may enable chromatin looping of enhancers to activate gene expression (Panigrahi and O’Malley, 2021). Further studies into the action and regulation of MLL3/4 complexes at promoters and gene bodies will be informative and may yield new insights into the actions of the COMPASS- like complex in cancer. While CDKN2A showed a surprisingly dominant role in mediating the tumor- suppressive effects of the broadly acting MLL3 enzyme, there are precedents for a predominant contribution of a single gene to the functional output of chromatin- complex disruption. Indeed, polycomb repressive complexes (PRCs) broadly repress gene expression in different cell types through the coordinated action of PRC1 and PRC2 complexes that deposit and maintain repressive H3K27me3 marks on the enhancers of target genes, including CDKN2A (Bracken et al., 2007; Kotake et al., 2007). Despite these similarly broad effects, CDKN2A is often the most functionally relevant target of PRC- mediated repression, as genetic deletion of either the PRC1 component Bmi1 or the PRC2 component Ezh2, or treatment with small molecule inhibitors of EZH2, can facilitate Cdkn2a induction in normal and tumor cells. This, in turn, triggers anti- proliferative responses that can be rescued by Cdkn2a deletion (Jacobs et al., 1999; Richly et al., 2011). Notably, the COMPASS- like complexes are biochemically and functionally similar to Trithorax complexes in Drosophila, which have an evolutionarily conserved antagonistic relationship with PRC1 and PRC2 that controls epigenetic memory and cell fate during development (Mills, 2010; Piunti and Shilatifard, 2016). Our findings suggest such antagonism extends to tumor suppression in mammalian cells, likely via regulation of Cdkn2a and other tumor suppressor genes (Soto- Feliciano et al., 2023). Materials availability statement Source files of all original gels and western blots were provided for the following figures: Figure 1—figure supplement 2B; Figure 4—figure supplement 1A, C, D, E; Figure 5—figure supplement 2B, F, G. RNA sequencing and ChIP- Seq data files that support the findings of this study have been depos- ited in the Gene Expression Omnibus under the accession code GSE85055, as well as in the Dryad digital repository (doi:10.5061/dryad.7pvmcvdwm; doi:10.5061/dryad.f1vhhmh0h). Sequences of sgRNAs, shRNAs, and primers used in this manuscript are included in the Supplementary file 1. Materials and methods Animal experiments 8- to 10- week- old female C57BL/6 animals were purchased from Envigo (formerly Harlan). Each exper- iment was performed in mice from the same order. Arf- null animals (C57BL/6 background), originally provided by Dr. Charles Sherr, St. Jude Children’s Research Hospital, were maintained in our breeding colony. For HTVI, a sterile 0.9% NaCl solution/plasmid mix was prepared containing oncogene trans- posons (5 µg DNA of pT3- Myc or 10 µg pT3-Ctnnb1 N90) with either 20 µg of pX330 expressing the indicated sgRNAs or 20 µg of pT3- EF1a- GFP- miRE plasmid together with CMV- SB13 Transposase (1:5 ratio). Mice were randomly assigned to experimental groups and injected with the 0.9% NaCl solu- tion/plasmid mix into the lateral tail vein with a total volume corresponding to 10% of body weight in 5–7 s as described before (Largaespada, 2009; Moon et al., 2019; Tschaharganeh et al., 2014; Xue et  al., 2014). Injected mice were monitored for tumor formation by abdominal palpation. All animal experiments were approved by the Memorial Sloan Kettering Cancer Center (MSK) Institu- tional Animal Care and Use Committee (protocol 11- 06- 011). Animals were monitored for signs of ill health by veterinary staff at the Research Animal Resource Center at MSK, and efforts were made to minimize suffering. Vector constructs The pT3- Myc vector Addgene (#92046) and pT3- EF1a- GFP- miRE plasmid were described before (Huang et al., 2014). The pT3-Ctnnb1 N90 vector (Tward et al., 2007) was obtained from Addgene (#31785). For CRISPR/Cas9- mediated genome editing, sgRNAs were subcloned into pX330 (Addgene, #42230; Hsu et al., 2013). All shRNA and sgRNA sequences are listed in Supplementary file 1. Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 14 of 25 Cancer Biology Research article Derivation of primary liver tumor cell lines Liver tumors were resected with sterile instruments, and 10–50 mg of tumor tissue was minced and washed in sterile PBS, incubated in a mix of 1 mg/mL collagenase IV and 3 mg/mL dispase (dissolved in sterile, serum- free DMEM(Dulbecco's Modified Eagle Medium)) with gentle shaking, washed with PBS, incubated for 5  min in 0.05% (w/v) trypsin, and washed and plated in complete DMEM (10% FBS(fetal bovine serum), 1× penicillin/streptomycin) on collagen- coated plates (PurCol, Advanced Biomatrix). Primary cultures were passaged until visibly free from fibroblasts. Cell lines were authen- ticated on a routine basis using short tandem repeat profiling, as well as tested for mycoplasma contamination and immediately discarded upon a positive test. Analysis of CRISPR-directed mutations CRISPR- mediated insertions and deletions were detected by surveyor assay as directed by the manu- facturer (Transgenomic/IDT). Briefly, after overnight lysis of primary tumors and cell lines at 37°C in buffer containing 0.4 mg/mL proteinase K, 10 mM Tris, 100 mM NaCl, 10 mM EDTA, and 0.5% SDS, pH 8.0, genomic DNA was extracted by isopropanol precipitation.  ~250–500  bp regions flanking predicted CRISPR cleavage sites were PCR amplified with Herculase II taq polymerase, column puri- fied (Qiagen), heated to 95°C, and slowly cooled to promote annealing of heteroduplexes. Following treatment with Surveyor nuclease, products were analyzed by electrophoresis on a 2% polyacrylamide gel. Primers used for surveyor assay are listed in Supplementary file 1. Amplified PCR products were separately gel purified and ligated into blunt- end digested pBlueScript (Stratagene). DNA from 48 transformed colonies was analyzed by Sanger sequencing using a T7 primer. CRISPR activation Human HCC cell line HLE, purchased from JCRB Cell Bank (JCRB0404), was transduced by the lenti- virus expressing nuclease- dead Cas9 (dCas9) fused with VPR (Chavez et  al., 2015) and sgRNAs against KMT2C (sequence in Supplementary file 1) to generate stable MLL3 CRISPRa HLE line by puromycin selection. Generation and modification of primary cells LPCs from E13.5–15.5 C57BL/6 embryos were isolated and grown in hepatocyte growth media (HGM) as previously described (Zender et al., 2005). To simultaneously overexpress Myc and condi- tionally suppress Kmt2c, LPCs were co- infected with a retroviral construct constitutively expressing both Myc and a reverse tet- transactivator (rtTA) (MSCV- Myc- IRES- rtTA) along with retroviral TRMPV vectors (MSCV- TRE- dsRed- miR30/shRNA- PGK- Venus- IRES- NeoR) (Zuber et  al., 2011) expressing ds- Red linked, teint- responsive shRNAs targeting Kmt2c cloned into an optimized mir- 30 context (‘mir- E,’ TRPMVe; Zuber et al., 2011). For selection of infected cells and sustained shKmt2c expres- sion, cells were maintained in HGM with neomycin (1 mg/mL) and dox (1 µg/mL) starting 2 days after infection. To introduce constitutively expressed shRNAs in the setting of inducible shKmt2c, retro- viral MLPe vectors (MSCV- LTR- mir- E- PGK- Puro- IRES- GFP; Dickins et al., 2005). GFP- linked shRNAs targeting either Cdkn2a or Renilla luciferase (as control) were co- infected with MSCV- Myc- IRES- rtTa and TRMPVe- shKmt2c. Triple- infected cells were maintained in media with neomycin, puromycin (2 µg/mL), and dox 2 days post infection. Infected continuously proliferating cells were transitioned to growth in complete DMEM and maintained on collagen- coated plates. Colony assays For measurement of cell proliferation, 5000 transduced and selected LPCs or MEFs were plated in triplicate in 6- well plates. Tetracycline- inducible shKmt2c–expressing LPCs were grown in the pres- ence or absence of dox, and after 5 days, cells were fixed with formalin and methanol and stained with 0.05% crystal violet. MEFs were fixed after 6 days with formalin and methanol and stained with 0.05% crystal violet. Apoptosis assays Apoptosis was measured in LPCs via Annexin V staining according to the manufacturer’s instructions (eBiosciences, Annexin- V APC). 25,000 cells were grown with and without dox for 3 days, trypsinized, Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 15 of 25 Cancer Biology Research article washed with Annexin- V binding buffer, and ~100,000 cells were incubated with Annexin- V APC and analyzed on an LSRII flow cytometer (BD). Live imaging Imaging was performed on LPCs immortalized by linked overexpression of Myc and two independent, inducible Kmt2c shRNAs constitutively expressing shRNAs targeting Renilla luciferase or Cdkn2a (generated as detailed above). 1000  cells were plated on collagen- coated, 96 well, clear bottom imaging plates in media supplemented with 300 nM Draq7 (Invitrogen) with and without dox, in trip- licate by genotype. 18 hr after plating cells, Venus (marking all plated cells) and Draq7 fluorescence was collected in two, 10× fields of each well every 15 min for 41 hr using an automated, high content microscope (InCell 6000, General Electric). Chromatin immunoprecipitation Histone ChIP was performed as previously described (Lee et  al., 2006). Briefly, cell samples were cross- linked in 1% formaldehyde for 10  min, and the reaction was stopped by addition of glycine to 125  mM final concentration. Fixed cells were lysed in SDS lysis buffer, and the chromatin was fragmented by sonication (Covaris). Sheared chromatin was incubated with antibodies (final concen- tration 10 µg/mL) against H3K4me3 (Abcam, ab8580, Lot:GR164706- 1), H3K27ac (Abcam, ab4729, Lot:GR200563- 1), or H3K4me1 (Abcam; ab8895, Lot:GR114265- 2) or with normal rabbit IgG (Abcam, ab46540) at 4°C for overnight. Antibodies were recovered by binding to protein A/G agarose (Milli- pore), and the eluted DNA fragments were used directly for qPCR or subjected to high- throughput sequencing (ChIP- Seq) using a HiSeq 2000 platform (Illumina). High- throughput reads were aligned to mouse genome assembly NCBI37/mm9 as previously described (Barradas et al., 2009). Reads that aligned to multiple loci in the mouse genome were discarded. The ChIP- Seq signal for each gene was quantified as total number of reads per million in the region 2 kb upstream to 2 kb downstream of the transcription start site (TSS). Primers used for ChIP- qPCR of mouse Cdkn2a promoter (Barradas et al., 2009) are listed in Table S1. The complete dataset is available at NCBI Gene Expression Omnibus (GSE85055), as well as the Dryad digital repository (doi:10.5061/dryad.7pvmcvdwm). For the MLL3 ChIP- Seq, the following protocol was used. Cross- linking ChIP in mouse and human HCC cells was performed with 10–20×107 cells per immunoprecipitation. Cells were collected, washed once with ice- cold PBS, and flash- frozen. Cells were resuspended in ice- cold PBS and cross- linked using 1% paraformaldehyde (PFA; Electron Microscopy Sciences) for 5 min at room temperature with gentle rotation. Unreacted PFA was quenched with glycine (final concentration 125 mM) for 5 min at room temperature with gentle rotation. Cells were washed once with ice- cold PBS and pelleted by centrifugation (800 g for 5 min). To obtain a soluble chromatin extract, cells were resuspended in 1 mL of LB1 (50 mM HEPES pH 7.5, 140 mM NaCl, 1 mM EDTA, 10% glycerol, 0.5% NP- 40, 0.25% Triton X- 100, and 1× complete protease inhibitor cocktail) and incubated at 4°C for 10  min while rotating. Samples were centrifuged (1400 g for 5 min), resuspended in 1 mL of LB2 (10 mM Tris- HCl pH 8.0, 200  mM NaCl, 1  mM EDTA, 0.5  mM EGTA, and 1× complete protease inhibitor cocktail), and incubated at 4°C for 10 min while rotating. Finally, samples were centrifuged (1400 g for 5 min) and resuspended in 1 mL of LB3 (10 mM Tris- HCl pH 8.0, 100 mM NaCl, 1 mM EDTA, 0.5 mM EGTA, 0.1% sodium deoxycholate, 0.5% N- lauroylsarcosine, and 1× complete protease inhibitor cocktail). Samples were homogenized by passing seven to eight times through a 28- gauge needle, then Triton X- 100 was added to a final concentration of 1%. Chromatin extracts were sonicated for 14 min using a Covaris E220 focused ultrasonicator. Lysates were centrifuged at 20,000 g for 10 min at 4°C, and 5% of the supernatant was saved as input DNA. Beads were prepared by incubating them in 0.5% BSA in PBS and antibodies overnight (100 μL of Dynabeads Protein A or Protein G [Invitrogen] plus 20 μL of antibody). The antibody was anti- MLL3/4, which was kindly provided by the Wysocka laboratory (Dorighi et al., 2017). Antibody- beads mixes were washed with 0.5% BSA in PBS and then added to the lysates overnight while rotating at 4°C. Beads were then washed six times with RIPA buffer (50 mM HEPES pH 7.5, 500 mM LiCl, 1 mM EDTA, 0.7% sodium- deoxycholate, and 1% NP- 40) and once with TE- NaCl Buffer (10 mM Tris- HCl pH 8.0, 50 mM NaCl, and 1 mM EDTA). Chromatin was eluted from beads in Elution buffer (50 mM Tris- HCl pH 8.0, 10 mM EDTA, and 1% SDS) by incubating at 65°C for 30 min while shaking, supernatant was removed by centrifugation, and crosslinking was reversed by Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 16 of 25 Cancer Biology Research article further incubating chromatin overnight at 65°C. The eluted chromatin was then treated with RNase A (10 mg/mL) for 1 hr at 37°C and with proteinase K (Roche) for 2 hr at 55°C. DNA was purified by using phenol- chloroform extraction followed with ethanol precipitation. The NEBNext Ultra II DNA Library Prep kit was used to prepare samples for sequencing on an Illumina NextSeq500 (75 bp read length, single- end, or 37 bp read length, and paired- end). The complete dataset for MLL3 ChIP- Seq is available at the Dryad digital repository (doi:10.5061/ dryad.f1vhhmh0h). Immunoblotting Cell pellets were lysed in Laemmli buffer (100  mM Tris- HCl pH 6.8, 5% glycerol, 2% SDS, and 5% 2- mercaptoethanol). Equal amounts of protein were separated on 12% SDS–polyacrylamide gels and transferred to PVDF(polyvinylidene difluoride) membranes (90 V, 75 min). β-actin was used as a control to ensure equal loading, and images were analyzed using the AlphaView software (ProteinSimple). Immunoblotting was performed using antibodies for MYC (1:1000, Abcam, ab32072), p53 (1:500, Leica Biosystems, NCL- p53- 505), p19 (1:250, Santa Cruz Biotechnology, sc- 32748), p16 (1:250, Santa Cruz Biotechnology, sc- 1207), Axin1 (1:1000, Cell Signaling, #2074), and β-actin (1:10000, Sigma- Aldrich, clone AC- 15). Source files of all western blots were provided for Figure 1—figure supple- ment 2, Figure 4—figure supplement 1, Figure 5—figure supplement 2. Quantitative RT-PCR Total RNA was isolated using RNeasy Mini Kit, QIAshredder Columns, and RNase- Free DNase Set (Qiagen). cDNA synthesis was performed using TaqMan Reverse Transcription Reagents (Thermo Fisher Scientific). Real- time PCR was carried out using Power SYBR Green Master Mix (Thermo Fisher Scientific) and the Life Technologies ViiA 7 machine. Transcript levels were normalized to the levels of mouse or human Actb mRNA expression and calculated using the ΔΔCt method. Each qRT- PCR was performed in triplicate using gene- specific primers (sequences listed in Table S1). RNA sequencing and differential expression analysis For RNA sequencing, total RNA from three independent tumor- derived cell lines (Myc; sgTrp53 and Myc; sgKmt2c) was isolated using RNeasy Mini Kit, QIAshredder Columns and RNase- Free DNase Set (Qiagen). RNA- Seq library construction and sequencing were performed according to proto- cols used by the integrated genomics operation Core at MSK. 5–10  million reads were acquired per replicate sample. After removing adaptor sequences with Trimmomatic, RNA- seq reads were aligned to GRCm38.91(mm10) with STAR (Dobin et al., 2013). Genome- wide transcript counting was performed by HTSeq to generate an FPKM(Fragments Per Kilobase per Million mapped fragments) matrix (Anders et al., 2015). DEGs were identified by DESeq2 (v.1.8.2, package in R) and plotted in the volcano plot. The complete dataset is available at NCBI Gene Expression Omnibus (GSE85055). Integrative analyses of RNA-seq and MLL3 ChIP-seq Differential peaks from ChIP- Seq data were annotated by assigning all intragenic peaks to that gene while intergenic peaks were assigned using linear genomic distance to the TSS. Genes that were coor- dinately regulated (fold change >1.5 and adjusted p- value<0.1) in MLL3 ChIP- seq and RNA- seq data were selected for the integrated analysis. Enriched pathways were scored using the enrichGO function with ‘biological process’ in the clusterProfiler R package. Redundant pathways were collapsed using the ‘simplify’ function with a cutoff of 0.7 with the p.adjust metric. Network analysis was performed using differential peaks and genes by running enrichplot::cnetplot in R with default parameters. Human cancer analyses RNA sequencing data of selected samples with somatic mutations or homozygous deletions of KMT2C, CDKN2A, TP53, or RB1 in the TCGA HCC dataset were downloaded from Broad Institute TCGA Genome Data Analysis Center. To obtain transcriptional signatures of HCC with genomic muta- tions and deletions of either KMT2C, CDKN2A, and RB1, differential gene expression analyses were performed by DESeq2 (with TP53- mutated HCCs as controls). The oncoprints of homozygous dele- tions and somatic mutations of KMT2C, CDKN2A, and TP53, as well as MYC gains and amplifications from human HCC datasets (Cancer Genome Atlas Research Network, 2017, MSK [Harding et al., Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 17 of 25 Cancer Biology Research article 2019; Zheng et al., 2018], INSERM [Schulze et al., 2015], RIKEN [Fujimoto et al., 2012], AMC [Ahn et al., 2014], and MERCi [Ng et al., 2022]) were generated by cBioPortal (https://www.cbioportal.org ; Cerami et al., 2012; Gao et al., 2013). Gene set enrichment analysis GSEA was performed using the GSEAPreranked tool for conducting GSEA of data derived from RNA- seq experiments (version 2.07) against other signatures. The metric scores (normalized enrichment scores and false discovery rate q- values) were calculated using the sign of the fold change multiplied by the inverse of the p- value (Subramanian et al., 2005). Specifically, transcriptional signatures were derived based on significantly changed genes (p- adjusted<0.05, absolute fold change >2) from RNA- seq of mouse HCC cell lines (Myc; sgKmt2c vs Myc; sgTrp53, n=3  each genotype), and in human HCCs with mutations in KMT2C vs TP53 (p- adjusted<0.05, absolute fold change >2). These signatures were compared to the transcriptional comparison of TCGA human HCCs with genomic inactivation of CDKN2A vs TP53. Statistical analyses Data are presented as mean ± D or SEM as specified. The statistical comparison between two groups was accomplished with the two- tailed student’s t- test or one- way ANOVA followed by post hoc t- tests among three or more groups. The analyses for co- occurrence or mutual exclusivity of mutations were performed using Fisher Exact test. Comparisons of survival curves were performed by log- rank tests. All statistical tests were performed using the Prism 8 software. All data presented in the manuscript have been replicated in independent cohorts of mice or in at least three biological replicates for in vitro experiments. On the basis of predicted effects of oncogene- tumor suppressor interaction introduced by HTVI in mice, with a power of 0.8 and p<0.05, we calculated a minimum sample size of 5 mice per group. Animals within the same cage were randomly allocated into control and exper- imental groups, with the group assignment recorded in a master spreadsheet and unmasked only when all samples of the respective experiments were analyzed. Data collection of each experiment was detailed in the respective figures, figure legends, and methods. No data were excluded from studies in this manuscript. Acknowledgements We thank Charles Sherr and Janet Novak for constructive guidance and advice on all aspects of this study. We thank Ali Shilatifard, Lu Wang, and all members of the Lowe lab for helpful and stimu- lating discussions. We gratefully thank A Chramiec for excellent technical assistance. We thank Joanna Wysocka (Stanford University) for kindly sharing the anti- MLL3/4 antibody used in our ChIP- Seq experiments. This work was supported by grants to SWL (P01 CA013106 and R01 CA233944) from the NIH/NCI, as well as by the National Center for Tumor Disease, Heidelberg, and grants of the German Research Foundation to DFG (SFB/TRR77). This work was also supported by the NIH/NCI Cancer Center Support Grant to Memorial Sloan Kettering Cancer Center (P30 CA008748). YMSF is supported by a MOSAIC K99/R00 Award from the NIH/NIGMS (1K99GM140265- 01). CZ is supported by an F32 Postdoctoral Fellowship (1F32CA257103) from the NIH/NCI. JPM was a recipient of a Postdoctoral Fellowship (PF- 14- 066- 01- TBE) from the American Cancer Society. DFT is supported by a Young Investigator Group (VH- NG- 1114) by the Helmholtz foundation. SWL is the Geoffrey Beene Chair for Cancer Biology and an investigator of the Howard Hughes Medical Institute. Additional information Competing interests C David Allis: is a co founder of Chroma Therapeutics and Constellation Pharmaceuticals and a Scien- tific Advisory Board member of EpiCypher. Scott W Lowe: is an advisor for and has equity in the following biotechnology companies: ORIC Pharmaceuticals, Faeth Therapeutics, Blueprint Medicines, Geras Bio, Mirimus Inc, Senescea, and PMV Pharmaceuticals. S.W.L. also acknowledges receiving funding and research support from Agilent Technologies and Calico, for the purposes of massively Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 18 of 25 Cancer Biology Research article parallel oligo synthesis and single- cell analytics, respectively. The other authors declare that no competing interests exist. Funding Funder Grant reference number Author National Cancer Institute P01 CA013106 Scott W Lowe National Cancer Institute R01 CA233944 Scott W Lowe National Institute of General Medical Sciences 1K99GM140265-01 Yadira M Soto-Feliciano National Cancer Institute 1F32CA257103 Changyu Zhu American Cancer Society PF-14-066-01-TBE John P Morris Helmholtz foundation VH-NG-1114 Darjus F Tschaharganeh National Cancer Institute P30 CA008748 Scott W Lowe The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Author contributions Changyu Zhu, Yadira M Soto- Feliciano, Conceptualization, Data curation, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing – original draft, Writing – review and editing; John P Morris, Conceptualization, Data curation, Formal analysis, Validation, Investigation, Methodology, Writing – original draft, Writing – review and editing; Chun- Hao Huang, Conceptualization, Data curation, Formal analysis, Validation, Investigation, Methodology, Writing – original draft; Richard P Koche, Yu- jui Ho, Software, Formal analysis, Visualization, Methodology; Ana Banito, Data curation, Methodology; Chun- Wei Chen, Formal analysis; Aditya Shroff, Sha Tian, Geulah Livshits, Chi- Chao Chen, Myles Fennell, Scott A Armstrong, Methodology; C David Allis, Supervision, Writing – review and editing; Darjus F Tschaharganeh, Conceptualization, Data curation, Formal analysis, Supervision, Validation, Investigation, Methodology, Writing – original draft, Writing – review and editing; Scott W Lowe, Conceptualization, Resources, Supervision, Funding acquisition, Investigation, Writing – orig- inal draft, Project administration, Writing – review and editing Author ORCIDs Changyu Zhu Yadira M Soto- Feliciano Richard P Koche Ana Banito Chun- Wei Chen Scott A Armstrong Scott W Lowe http://orcid.org/0000-0003-3583-3638 http://orcid.org/0000-0002-8523-7917 http://orcid.org/0000-0002-6820-5083 http://orcid.org/0000-0003-2188-0003 http://orcid.org/0000-0002-8737-6830 http://orcid.org/0000-0002-9099-4728 http://orcid.org/0000-0002-5284-9650 Ethics All animal experiments were approved by the MSKCC Institutional Animal Care and Use Committee (protocol 11- 06- 011). Animals were monitored for signs of ill- health by veterinary staff at the Research Animal Resource Center (RARC) at MSKCC and efforts were made to minimize suffering. Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.80854.sa1 Author response https://doi.org/10.7554/eLife.80854.sa2 Additional files Supplementary files • Supplementary file 1. Tables displaying the sequences of single guide RNA (sgRNA), shRNA, qPCR primers, and chromatin immunoprecipitation (ChIP)- qPCR primers used in the studies of this manuscript. Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 19 of 25 Cancer Biology Research article • MDAR checklist Data availability Source files of all original gels and Western Blots were provided for the following figures: Figure 1— figure supplement 2B; Figure 4—figure supplement 1A, C, D, E; Figure 5—figure supplement 2B, F, G. RNA sequencing and ChIP sequencing data files that support the findings of this study have been deposited in the Gene Expression Omnibus under the accession code GSE85055, as well as in the Dryad digital repository (doi:10.5061/dryad.7pvmcvdwm; doi:10.5061/dryad.f1vhhmh0h). Sequences of sgRNAs, shRNAs, and primers used in this manuscript are included in the Supplementary File 1. The following datasets were generated: Year 2022 Dataset title Dataset URL Database and Identifier Mll3 suppresses tumorigenesis by activating the Ink4a/Arf locus https:// doi. org/ 10. 5061/ dryad. 7pvmcvdwm Dryad Digital Repository, 10.5061/dryad.7pvmcvdwm 2022 MLL3 ChIP sequencing in murine and human HCC cells https:// doi. org/ 10. 5061/ dryad. f1vhhmh0h Dryad Digital Repository, 10.5061/dryad.f1vhhmh0h Author(s) Soto- Feliciano MY, Zhu C, Morris JP, Huang C- H, Koche RP, Y- J Ho, Banito A, Chen C- W, Shroff A, Tian S, Livshits G, Chen C- C, Fennell M, Armstrong SA, Allis CD, Tschaharganeh DF, Lowe SW Soto- Feliciano MY, Zhu C, Morris JP, Huang C- H, Roche RP, Y- J Ho, Banito A, Chen C- W, Shroff A, Tian S, Livshits G, Chen C- C, Fennell M, Armstrong SA, Allis CD, Tschaharganeh DF, Lowe SW Lowe SW 2017 Mll3 suppresses tumorigenesis by activating the Ink4a/Arf locus https://www. ncbi. nlm. nih. gov/ geo/ query/ acc. cgi? acc= GSE85055 NCBI Gene Expression Omnibus, GSE85055 References Ahn SM, Jang SJ, Shim JH, Kim D, Hong SM, Sung CO, Baek D, Haq F, Ansari AA, Lee SY, Chun SM, Choi S, Choi HJ, Kim J, Kim S, Hwang S, Lee YJ, Lee JE, Jung WR, Jang HY, et al. 2014. Genomic portrait of Resectable hepatocellular Carcinomas: implications of Rb1 and Fgf19 aberrations for patient stratification. Hepatology 60:1972–1982. DOI: https://doi.org/10.1002/hep.27198, PMID: 24798001 Ananthanarayanan M, Li Y, Surapureddi S, Balasubramaniyan N, Ahn J, Goldstein JA, Suchy FJ. 2011. Histone H3K4 Trimethylation by Mll3 as part of ASCOM complex is critical for NR activation of bile acid transporter genes and is downregulated in cholestasis. American Journal of Physiology. Gastrointestinal and Liver Physiology 300:G771–G781. DOI: https://doi.org/10.1152/ajpgi.00499.2010, PMID: 21330447 Anders S, Pyl PT, Huber W. 2015. Htseq--a python framework to work with high- throughput sequencing data. Bioinformatics 31:166–169. DOI: https://doi.org/10.1093/bioinformatics/btu638, PMID: 25260700 Barradas M, Anderton E, Acosta JC, Li S, Banito A, Rodriguez- Niedenführ M, Maertens G, Banck M, Zhou MM, Walsh MJ, Peters G, Gil J. 2009. Histone demethylase Jmjd3 contributes to epigenetic control of Ink4A/ARF by Oncogenic RAS. Genes & Development 23:1177–1182. DOI: https://doi.org/10.1101/gad.511109, PMID: 19451218 Barriga FM, Tsanov KM, Ho YJ, Sohail N, Zhang A, Baslan T, Wuest AN, Del Priore I, Meškauskaitė B, Livshits G, Alonso- Curbelo D, Simon J, Chaves- Perez A, Bar- Sagi D, Iacobuzio- Donahue CA, Notta F, Chaligne R, Sharma R, Pe’er D, Lowe SW. 2022. MACHETE identifies interferon- encompassing Chromosome 9P21.3 deletions as mediators of immune evasion and metastasis. Nature Cancer 3:1367–1385. DOI: https://doi.org/ 10.1038/s43018-022-00443-5, PMID: 36344707 Bell JB, Podetz- Pedersen KM, Aronovich EL, Belur LR, McIvor RS, Hackett PB. 2007. Preferential delivery of the sleeping beauty Transposon system to livers of mice by Hydrodynamic injection. Nature Protocols 2:3153– 3165. DOI: https://doi.org/10.1038/nprot.2007.471, PMID: 18079715 Bieging- Rolett KT, Kaiser AM, Morgens DW, Boutelle AM, Seoane JA, Van Nostrand EL, Zhu C, Houlihan SL, Mello SS, Yee BA, McClendon J, Pierce SE, Winters IP, Wang M, Connolly AJ, Lowe SW, Curtis C, Yeo GW, Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 20 of 25 Cancer Biology Research article Winslow MM, Bassik MC, et al. 2020. Zmat3 is a key splicing regulator in the P53 tumor suppression program. Molecular Cell 80:452–469. DOI: https://doi.org/10.1016/j.molcel.2020.10.022, PMID: 33157015 Bracken AP, Kleine- Kohlbrecher D, Dietrich N, Pasini D, Gargiulo G, Beekman C, Theilgaard- Mönch K, Minucci S, Porse BT, Marine JC, Hansen KH, Helin K. 2007. The Polycomb group proteins bind throughout the Ink4A- ARF locus and are disassociated in Senescent cells. Genes & Development 21:525–530. DOI: https://doi.org/10. 1101/gad.415507, PMID: 17344414 Cancer Genome Atlas Research Network. 2017. Comprehensive and integrative Genomic characterization of hepatocellular carcinoma. Cell 169:1327–1341. DOI: https://doi.org/10.1016/j.cell.2017.05.046 Cerami E, Gao J, Dogrusoz U, Gross BE, Sumer SO, Aksoy BA, Jacobsen A, Byrne CJ, Heuer ML, Larsson E, Antipin Y, Reva B, Goldberg AP, Sander C, Schultz N. 2012. The cBio cancer Genomics portal: an open platform for exploring multidimensional cancer Genomics data. Cancer Discovery 2:401–404. DOI: https://doi.org/10. 1158/2159-8290.CD-12-0095, PMID: 22588877 Chavez A, Scheiman J, Vora S, Pruitt BW, Tuttle M, P R Iyer E, Lin S, Kiani S, Guzman CD, Wiegand DJ, Ter- Ovanesyan D, Braff JL, Davidsohn N, Housden BE, Perrimon N, Weiss R, Aach J, Collins JJ, Church GM. 2015. Highly efficient Cas9- mediated transcriptional programming. Nature Methods 12:326–328. DOI: https:// doi.org/10.1038/nmeth.3312, PMID: 25730490 Chen C, Liu Y, Rappaport AR, Kitzing T, Schultz N, Zhao Z, Shroff AS, Dickins RA, Vakoc CR, Bradner JE, Stock W, LeBeau MM, Shannon KM, Kogan S, Zuber J, Lowe SW. 2014. Mll3 is a Haploinsufficient 7Q tumor Suppressor in acute myeloid leukemia. Cancer Cell 25:652–665. DOI: https://doi.org/10.1016/j.ccr.2014.03.016, PMID: 24794707 Cheng J, Blum R, Bowman C, Hu D, Shilatifard A, Shen S, Dynlacht BD. 2014. A role for H3K4 Monomethylation in Gene repression and partitioning of Chromatin readers. Molecular Cell 53:979–992. DOI: https://doi.org/10. 1016/j.molcel.2014.02.032, PMID: 24656132 Cleary SP, Jeck WR, Zhao X, Chen K, Selitsky SR, Savich GL, Tan TX, Wu MC, Getz G, Lawrence MS, Parker JS, Li J, Powers S, Kim H, Fischer S, Guindi M, Ghanekar A, Chiang DY. 2013. Identification of driver genes in hepatocellular carcinoma by Exome sequencing. Hepatology 58:1693–1702. DOI: https://doi.org/10.1002/hep. 26540, PMID: 23728943 Denissov S, Hofemeister H, Marks H, Kranz A, Ciotta G, Singh S, Anastassiadis K, Stunnenberg HG, Stewart AF. 2014. Mll2 is required for H3K4 Trimethylation on Bivalent promoters in embryonic stem cells, whereas Mll1 is redundant. Development 141:526–537. DOI: https://doi.org/10.1242/dev.102681, PMID: 24423662 Dhar SS, Lee SH, Chen K, Zhu G, Oh W, Allton K, Gafni O, Kim YZ, Tomoiga AS, Barton MC, Hanna JH, Wang Z, Li W, Lee MG. 2016. An essential role for UTX in resolution and activation of Bivalent promoters. Nucleic Acids Research 44:3659–3674. DOI: https://doi.org/10.1093/nar/gkv1516, PMID: 26762983 Dickins RA, Hemann MT, Zilfou JT, Simpson DR, Ibarra I, Hannon GJ, Lowe SW. 2005. Probing tumor phenotypes using stable and regulated synthetic microRNA precursors. Nature Genetics 37:1289–1295. DOI: https://doi. org/10.1038/ng1651, PMID: 16200064 Dobin A, Davis CA, Schlesinger F, Drenkow J, Zaleski C, Jha S, Batut P, Chaisson M, Gingeras TR. 2013. STAR: Ultrafast universal RNA- seq aligner. Bioinformatics 29:15–21. DOI: https://doi.org/10.1093/bioinformatics/ bts635, PMID: 23104886 Dorighi KM, Swigut T, Henriques T, Bhanu NV, Scruggs BS, Nady N, Still CD, Garcia BA, Adelman K, Wysocka J. 2017. Mll3 and Mll4 facilitate enhancer RNA synthesis and transcription from promoters independently of H3K4 Monomethylation. Molecular Cell 66:568–576. DOI: https://doi.org/10.1016/j.molcel.2017.04.018, PMID: 28483418 Egolf S, Zou J, Anderson A, Simpson CL, Aubert Y, Prouty S, Ge K, Seykora JT, Capell BC. 2021. Mll4 mediates differentiation and tumor suppression through Ferroptosis. Science Advances 7:eabj9141. DOI: https://doi.org/ 10.1126/sciadv.abj9141, PMID: 34890228 Eischen CM, Weber JD, Roussel MF, Sherr CJ, Cleveland JL. 1999. Disruption of the ARF- Mdm2- P53 tumor Suppressor pathway in Myc- induced Lymphomagenesis. Genes & Development 13:2658–2669. DOI: https:// doi.org/10.1101/gad.13.20.2658, PMID: 10541552 Evan GI, Wyllie AH, Gilbert CS, Littlewood TD, Land H, Brooks M, Waters CM, Penn LZ, Hancock DC. 1992. Induction of apoptosis in fibroblasts by C- Myc protein. Cell 69:119–128. DOI: https://doi.org/10.1016/0092- 8674(92)90123-t, PMID: 1555236 Fountain JW, Karayiorgou M, Taruscio D, Graw SL, Buckler AJ, Ward DC, Dracopoli NC, Housman DE. 1992. Genetic and physical map of the interferon region on Chromosome 9p. Genomics 14:105–112. DOI: https:// doi.org/10.1016/s0888-7543(05)80290-3, PMID: 1385297 Fujimoto A, Totoki Y, Abe T, Boroevich KA, Hosoda F, Nguyen HH, Aoki M, Hosono N, Kubo M, Miya F, Arai Y, Takahashi H, Shirakihara T, Nagasaki M, Shibuya T, Nakano K, Watanabe- Makino K, Tanaka H, Nakamura H, Kusuda J, et al. 2012. Whole- genome sequencing of liver cancers identifies Etiological influences on Mutation patterns and recurrent mutations in Chromatin regulators. Nature Genetics 44:760–764. DOI: https://doi.org/ 10.1038/ng.2291, PMID: 22634756 Gao J, Aksoy BA, Dogrusoz U, Dresdner G, Gross B, Sumer SO, Sun Y, Jacobsen A, Sinha R, Larsson E, Cerami E, Sander C, Schultz N. 2013. Integrative analysis of complex cancer Genomics and clinical profiles using the cBioPortal. Science Signaling 6:pl1. DOI: https://doi.org/10.1126/scisignal.2004088, PMID: 23550210 Gil J, Peters G. 2006. Regulation of the Ink4B- ARF- Ink4A tumour Suppressor locus: all for one or one for all. Nature Reviews. Molecular Cell Biology 7:667–677. DOI: https://doi.org/10.1038/nrm1987, PMID: 16921403 Harding JJ, Nandakumar S, Armenia J, Khalil DN, Albano M, Ly M, Shia J, Hechtman JF, Kundra R, El Dika I, Do RK, Sun Y, Kingham TP, D’Angelica MI, Berger MF, Hyman DM, Jarnagin W, Klimstra DS, Janjigian YY, Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 21 of 25 Cancer Biology Research article Solit DB, et al. 2019. Prospective Genotyping of hepatocellular carcinoma: clinical implications of next- generation sequencing for matching patients to targeted and immune therapies. Clinical Cancer Research 25:2116–2126. DOI: https://doi.org/10.1158/1078-0432.CCR-18-2293, PMID: 30373752 Herz HM, Madden LD, Chen Z, Bolduc C, Buff E, Gupta R, Davuluri R, Shilatifard A, Hariharan IK, Bergmann A. 2010. The H3K27Me3 demethylase dUTX is a Suppressor of Notch- and RB- dependent tumors in Drosophila. Molecular and Cellular Biology 30:2485–2497. DOI: https://doi.org/10.1128/MCB.01633-09, PMID: 20212086 Herz HM, Mohan M, Garruss AS, Liang K, Takahashi YH, Mickey K, Voets O, Verrijzer CP, Shilatifard A. 2012. Enhancer- associated H3K4 monomethylation by Trithorax- related, the Drosophila homolog of mammalian Mll3/ Mll4. Genes & Development 26:2604–2620. DOI: https://doi.org/10.1101/gad.201327.112, PMID: 23166019 Hsu PD, Scott DA, Weinstein JA, Ran FA, Konermann S, Agarwala V, Li Y, Fine EJ, Wu X, Shalem O, Cradick TJ, Marraffini LA, Bao G, Zhang F. 2013. DNA targeting specificity of RNA- guided Cas9 nucleases. Nature Biotechnology 31:827–832. DOI: https://doi.org/10.1038/nbt.2647, PMID: 23873081 Hu D, Gao X, Morgan MA, Herz HM, Smith ER, Shilatifard A. 2013. The Mll3/Mll4 branches of the COMPASS family function as major Histone H3K4 Monomethylases at enhancers. Molecular and Cellular Biology 33:4745– 4754. DOI: https://doi.org/10.1128/MCB.01181-13, PMID: 24081332 Huang CH, Lujambio A, Zuber J, Tschaharganeh DF, Doran MG, Evans MJ, Kitzing T, Zhu N, de Stanchina E, Sawyers CL, Armstrong SA, Lewis JS, Sherr CJ, Lowe SW. 2014. Cdk9- mediated transcription elongation is required for MYC addiction in hepatocellular carcinoma. Genes & Development 28:1800–1814. DOI: https:// doi.org/10.1101/gad.244368.114, PMID: 25128497 Jacobs JJ, Scheijen B, Voncken JW, Kieboom K, Berns A, van Lohuizen M. 1999. Bmi- 1 Collaborates with C- Myc in tumorigenesis by inhibiting C- Myc- induced apoptosis via Ink4A/ARF. Genes & Development 13:2678–2690. DOI: https://doi.org/10.1101/gad.13.20.2678, PMID: 10541554 Jemal A, Ward EM, Johnson CJ, Cronin KA, Ma J, Ryerson AB, Mariotto A, Lake AJ, Wilson R, Sherman RL, Anderson RN, Henley SJ, Kohler BA, Penberthy L, Feuer EJ, Weir HK. 2017. Annual report to the nation on the status of cancer, 1975- 2014. Journal of the National Cancer Institute 109:djx030. DOI: https://doi.org/10.1093/ jnci/djx030, PMID: 28376154 Kan Z, Zheng H, Liu X, Li S, Barber TD, Gong Z, Gao H, Hao K, Willard MD, Xu J, Hauptschein R, Rejto PA, Fernandez J, Wang G, Zhang Q, Wang B, Chen R, Wang J, Lee NP, Zhou W, et al. 2013. Whole- genome sequencing identifies recurrent mutations in hepatocellular carcinoma. Genome Research 23:1422–1433. DOI: https://doi.org/10.1101/gr.154492.113, PMID: 23788652 Knudsen ES, Nambiar R, Rosario SR, Smiraglia DJ, Goodrich DW, Witkiewicz AK. 2020. Pan- cancer molecular analysis of the RB tumor Suppressor pathway. Communications Biology 3:158. DOI: https://doi.org/10.1038/ s42003-020-0873-9, PMID: 32242058 Kotake Y, Cao R, Viatour P, Sage J, Zhang Y, Xiong Y. 2007. pRB family proteins are required for H3K27 Trimethylation and Polycomb repression complexes binding to and silencing P16Ink4Alpha tumor Suppressor Gene. Genes & Development 21:49–54. DOI: https://doi.org/10.1101/gad.1499407, PMID: 17210787 Largaespada DA. 2009. Transposon- mediated Mutagenesis of somatic cells in the mouse for cancer Gene identification. Methods 49:282–286. DOI: https://doi.org/10.1016/j.ymeth.2009.07.002, PMID: 19607923 Lee TI, Johnstone SE, Young RA. 2006. Chromatin Immunoprecipitation and Microarray- based analysis of protein location. Nature Protocols 1:729–748. DOI: https://doi.org/10.1038/nprot.2006.98, PMID: 17406303 Lee J, Kim DH, Lee S, Yang QH, Lee DK, Lee SK, Roeder RG, Lee JW. 2009. A tumor suppressive coactivator complex of P53 containing ASC- 2 and Histone H3- Lysine- 4 Methyltransferase Mll3 or its Paralogue Mll4. PNAS 106:8513–8518. DOI: https://doi.org/10.1073/pnas.0902873106, PMID: 19433796 Llovet JM, Kelley RK, Villanueva A, Singal AG, Pikarsky E, Roayaie S, Lencioni R, Koike K, Zucman- Rossi J, Finn RS. 2021. Hepatocellular carcinoma. Nature Reviews. Disease Primers 7:6. DOI: https://doi.org/10.1038/ s41572-020-00240-3, PMID: 33479224 Lowe SW, Sherr CJ. 2003. Tumor suppression by Ink4A- ARF: progress and puzzles. Current Opinion in Genetics & Development 13:77–83. DOI: https://doi.org/10.1016/s0959-437x(02)00013-8, PMID: 12573439 Mills AA. 2010. Throwing the cancer switch: reciprocal roles of Polycomb and Trithorax proteins. Nature Reviews. Cancer 10:669–682. DOI: https://doi.org/10.1038/nrc2931, PMID: 20865010 Molina- Sánchez P, Ruiz de Galarreta M, Yao MA, Lindblad KE, Bresnahan E, Bitterman E, Martin TC, Rubenstein T, Nie K, Golas J, Choudhary S, Bárcena- Varela M, Elmas A, Miguela V, Ding Y, Kan Z, Grinspan LT, Huang KL, Parsons RE, Shields DJ, et al. 2020. Cooperation between distinct cancer driver genes underlies Intertumor heterogeneity in hepatocellular carcinoma. Gastroenterology 159:2203–2220. DOI: https://doi.org/ 10.1053/j.gastro.2020.08.015, PMID: 32814112 Moon SH, Huang CH, Houlihan SL, Regunath K, Freed- Pastor WA, Morris JP, Tschaharganeh DF, Kastenhuber ER, Barsotti AM, Culp- Hill R, Xue W, Ho YJ, Baslan T, Li X, Mayle A, de Stanchina E, Zender L, Tong DR, D’Alessandro A, Lowe SW, et al. 2019. P53 represses the Mevalonate pathway to mediate tumor suppression. Cell 176:564–580. DOI: https://doi.org/10.1016/j.cell.2018.11.011, PMID: 30580964 Negoescu A, Lorimier P, Labat- Moleur F, Azoti L, Robert C, Guillermet C, Brambilla C, Brambilla E. 1997. TUNEL: Improvement and evaluation of the method for in situ apoptotic cell identification. Biochemica 2:12–17. Ng CKY, Dazert E, Boldanova T, Coto- Llerena M, Nuciforo S, Ercan C, Suslov A, Meier MA, Bock T, Schmidt A, Ketterer S, Wang X, Wieland S, Matter MS, Colombi M, Piscuoglio S, Terracciano LM, Hall MN, Heim MH. 2022. Integrative Proteogenomic characterization of hepatocellular carcinoma across Etiologies and stages. Nature Communications 13:2436. DOI: https://doi.org/10.1038/s41467-022-29960-8, PMID: 35508466 Panigrahi A, O’Malley BW. 2021. Mechanisms of enhancer action: the known and the unknown. Genome Biology 22:108. DOI: https://doi.org/10.1186/s13059-021-02322-1, PMID: 33858480 Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 22 of 25 Cancer Biology Research article Piunti A, Shilatifard A. 2016. Epigenetic balance of gene expression by Polycomb and COMPASS families. Science 352:aad9780. DOI: https://doi.org/10.1126/science.aad9780, PMID: 27257261 Rebouissou S, Franconi A, Calderaro J, Letouzé E, Imbeaud S, Pilati C, Nault JC, Couchy G, Laurent A, Balabaud C, Bioulac- Sage P, Zucman- Rossi J. 2016. Genotype- phenotype correlation of Ctnnb1 mutations reveals different SS- Catenin activity associated with liver tumor progression. Hepatology 64:2047–2061. DOI: https://doi.org/10.1002/hep.28638, PMID: 27177928 Revia S, Seretny A, Wendler L, Banito A, Eckert C, Breuer K, Mayakonda A, Lutsik P, Evert M, Ribback S, Gallage S, Chikh Bakri I, Breuhahn K, Schirmacher P, Heinrich S, Gaida MM, Heikenwälder M, Calvisi DF, Plass C, Lowe SW, et al. 2022. Histone H3K27 demethylase Kdm6A is an epigenetic Gatekeeper of Mtorc1 signalling in cancer. Gut 71:1613–1628. DOI: https://doi.org/10.1136/gutjnl-2021-325405, PMID: 34509979 Richly H, Aloia L, Di Croce L. 2011. Roles of the Polycomb group proteins in stem cells and cancer. Cell Death & Disease 2:e204. DOI: https://doi.org/10.1038/cddis.2011.84, PMID: 21881606 Rickels R, Hu D, Collings CK, Woodfin AR, Piunti A, Mohan M, Herz HM, Kvon E, Shilatifard A. 2016. An evolutionary conserved epigenetic mark of Polycomb response elements implemented by TRX/MLL/ COMPASS. Molecular Cell 63:318–328. DOI: https://doi.org/10.1016/j.molcel.2016.06.018, PMID: 27447986 Satoh S, Daigo Y, Furukawa Y, Kato T, Miwa N, Nishiwaki T, Kawasoe T, Ishiguro H, Fujita M, Tokino T, Sasaki Y, Imaoka S, Murata M, Shimano T, Yamaoka Y, Nakamura Y. 2000. Axin1 mutations in hepatocellular Carcinomas, and growth suppression in cancer cells by virus- mediated transfer of Axin1. Nature Genetics 24:245–250. DOI: https://doi.org/10.1038/73448, PMID: 10700176 Schmid M, Sen M, Rosenbach MD, Carrera CJ, Friedman H, Carson DA. 2000. A Methylthioadenosine Phosphorylase (MTAP) fusion transcript identifies a new Gene on Chromosome 9P21 that is frequently deleted in cancer. Oncogene 19:5747–5754. DOI: https://doi.org/10.1038/sj.onc.1203942, PMID: 11126361 Schmitt CA, McCurrach ME, de Stanchina E, Wallace- Brodeur RR, Lowe SW. 1999. Ink4A/ARF mutations accelerate Lymphomagenesis and promote Chemoresistance by disabling P53. Genes & Development 13:2670–2677. DOI: https://doi.org/10.1101/gad.13.20.2670, PMID: 10541553 Schuettengruber B, Bourbon HM, Di Croce L, Cavalli G. 2017. Genome regulation by Polycomb and Trithorax: 70 years and counting. Cell 171:34–57. DOI: https://doi.org/10.1016/j.cell.2017.08.002, PMID: 28938122 Schulze K, Imbeaud S, Letouzé E, Alexandrov LB, Calderaro J, Rebouissou S, Couchy G, Meiller C, Shinde J, Soysouvanh F, Calatayud AL, Pinyol R, Pelletier L, Balabaud C, Laurent A, Blanc JF, Mazzaferro V, Calvo F, Villanueva A, Nault JC, et al. 2015. Exome sequencing of hepatocellular Carcinomas identifies new mutational signatures and potential therapeutic targets. Nature Genetics 47:505–511. DOI: https://doi.org/10.1038/ng. 3252, PMID: 25822088 Sherr CJ. 2012. Ink4- ARF locus in cancer and aging. Wiley Interdisciplinary Reviews. Developmental Biology 1:731–741. DOI: https://doi.org/10.1002/wdev.40, PMID: 22960768 Shilatifard A. 2012. The COMPASS family of Histone H3K4 Methylases: mechanisms of regulation in development and disease pathogenesis. Annual Review of Biochemistry 81:65–95. DOI: https://doi.org/10. 1146/annurev-biochem-051710-134100, PMID: 22663077 Soto- Feliciano YM, Sánchez- Rivera FJ, Perner F, Barrows DW, Kastenhuber ER, Ho Y- J, Carroll T, Xiong Y, Anand D, Soshnev AA, Gates L, Beytagh MC, Cheon D, Gu S, Liu XS, Krivtsov AV, Meneses M, de Stanchina E, Stone RM, Armstrong SA, et al. 2023. A molecular switch between mammalian MLL complexes dictates response to Menin- MLL inhibition. Cancer Discovery 13:146–169. DOI: https://doi.org/10.1158/2159-8290. CD-22-0416, PMID: 36264143 Stott FJ, Bates S, James MC, McConnell BB, Starborg M, Brookes S, Palmero I, Ryan K, Hara E, Vousden KH, Peters G. 1998. The alternative product from the human Cdkn2A locus, P14(ARF), participates in a regulatory feedback loop with P53 and Mdm2. The EMBO Journal 17:5001–5014. DOI: https://doi.org/10.1093/emboj/ 17.17.5001, PMID: 9724636 Subramanian A, Tamayo P, Mootha VK, Mukherjee S, Ebert BL, Gillette MA, Paulovich A, Pomeroy SL, Golub TR, Lander ES, Mesirov JP. 2005. Gene set enrichment analysis: a knowledge- based approach for interpreting genome- wide expression profiles. PNAS 102:15545–15550. DOI: https://doi.org/10.1073/pnas.0506580102, PMID: 16199517 Sun X, Wang SC, Wei Y, Luo X, Jia Y, Li L, Gopal P, Zhu M, Nassour I, Chuang JC, Maples T, Celen C, Nguyen LH, Wu L, Fu S, Li W, Hui L, Tian F, Ji Y, Zhang S, et al. 2017. Arid1A has context- dependent Oncogenic and tumor Suppressor functions in liver cancer. Cancer Cell 32:574–589. DOI: https://doi.org/10.1016/j.ccell.2017.10.007, PMID: 29136504 Sze CC, Shilatifard A. 2016. Mll3/Mll4/COMPASS family on epigenetic regulation of enhancer function and cancer. Cold Spring Harbor Perspectives in Medicine 6:a026427. DOI: https://doi.org/10.1101/cshperspect. a026427, PMID: 27638352 Tschaharganeh DF, Xue W, Calvisi DF, Evert M, Michurina TV, Dow LE, Banito A, Katz SF, Kastenhuber ER, Weissmueller S, Huang CH, Lechel A, Andersen JB, Capper D, Zender L, Longerich T, Enikolopov G, Lowe SW. 2014. P53- dependent Nestin regulation links tumor suppression to cellular plasticity in liver cancer. Cell 158:579–592. DOI: https://doi.org/10.1016/j.cell.2014.05.051, PMID: 25083869 Tward AD, Jones KD, Yant S, Cheung ST, Fan ST, Chen X, Kay MA, Wang R, Bishop JM. 2007. Distinct pathways of Genomic progression to benign and malignant tumors of the liver. PNAS 104:14771–14776. DOI: https:// doi.org/10.1073/pnas.0706578104, PMID: 17785413 Valekunja UK, Edgar RS, Oklejewicz M, van der Horst GTJ, O’Neill JS, Tamanini F, Turner DJ, Reddy AB. 2013. Histone Methyltransferase Mll3 contributes to genome- scale circadian transcription. PNAS 110:1554–1559. DOI: https://doi.org/10.1073/pnas.1214168110, PMID: 23297224 Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 23 of 25 Cancer Biology Research article Wang P, Lin C, Smith ER, Guo H, Sanderson BW, Wu M, Gogol M, Alexander T, Seidel C, Wiedemann LM, Ge K, Krumlauf R, Shilatifard A. 2009. Global analysis of H3K4 methylation defines MLL family member targets and points to a role for Mll1- mediated H3K4 methylation in the regulation of transcriptional initiation by RNA polymerase II. Molecular and Cellular Biology 29:6074–6085. DOI: https://doi.org/10.1128/MCB.00924-09, PMID: 19703992 Wang JK, Tsai MC, Poulin G, Adler AS, Chen S, Liu H, Shi Y, Chang HY. 2010. The Histone demethylase UTX enables RB- dependent cell fate control. Genes & Development 24:327–332. DOI: https://doi.org/10.1101/gad. 1882610, PMID: 20123895 Weiss JM, Hunter MV, Cruz NM, Baggiolini A, Tagore M, Ma Y, Misale S, Marasco M, Simon- Vermot T, Campbell NR, Newell F, Wilmott JS, Johansson PA, Thompson JF, Long GV, Pearson JV, Mann GJ, Scolyer RA, Waddell N, Montal ED, et al. 2022. Anatomic position determines Oncogenic specificity in Melanoma. Nature 604:354–361. DOI: https://doi.org/10.1038/s41586-022-04584-6, PMID: 35355015 Xia Y, Liu Y, Yang C, Simeone DM, Sun T- T, DeGraff DJ, Tang M- S, Zhang Y, Wu X- R. 2021. Dominant role of Cdkn2B/P15Ink4B of 9P21.3 tumor Suppressor Hub in inhibition of cell- cycle and Glycolysis. Nature Communications 12:2047. DOI: https://doi.org/10.1038/s41467-021-22327-5, PMID: 33824349 Xue W, Chen S, Yin H, Tammela T, Papagiannakopoulos T, Joshi NS, Cai W, Yang G, Bronson R, Crowley DG, Zhang F, Anderson DG, Sharp PA, Jacks T. 2014. CRISPR- mediated direct Mutation of cancer genes in the mouse liver. Nature 514:380–384. DOI: https://doi.org/10.1038/nature13589, PMID: 25119044 Zender L, Xue W, Cordón- Cardo C, Hannon GJ, Lucito R, Powers S, Flemming P, Spector MS, Lowe SW. 2005. Generation and analysis of genetically defined liver Carcinomas derived from Bipotential liver Progenitors. Cold Spring Harbor Symposia on Quantitative Biology 70:251–261. DOI: https://doi.org/10.1101/sqb.2005.70.059, PMID: 16869761 Zheng J, Sadot E, Vigidal JA, Klimstra DS, Balachandran VP, Kingham TP, Allen PJ, D’Angelica MI, DeMatteo RP, Jarnagin WR, Ventura A. 2018. Characterization of hepatocellular adenoma and carcinoma using microRNA profiling and targeted Gene sequencing. PLOS ONE 13:e0200776. DOI: https://doi.org/10.1371/journal.pone. 0200776, PMID: 30052636 Zindy F, Eischen CM, Randle DH, Kamijo T, Cleveland JL, Sherr CJ, Roussel MF. 1998. Myc signaling via the ARF tumor Suppressor regulates P53- dependent apoptosis and Immortalization. Genes & Development 12:2424– 2433. DOI: https://doi.org/10.1101/gad.12.15.2424, PMID: 9694806 Zuber J, McJunkin K, Fellmann C, Dow LE, Taylor MJ, Hannon GJ, Lowe SW. 2011. Toolkit for evaluating genes required for proliferation and survival using Tetracycline- regulated Rnai. Nature Biotechnology 29:79–83. DOI: https://doi.org/10.1038/nbt.1720, PMID: 21131983 Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 24 of 25 Cancer Biology Research article Appendix 1 Appendix 1—key resources table Reagent type (species) or resource Designation Source or reference Identifiers Additional information Strain and strain background (M. musculus) Wild- type C57BL/6 J The Jackson Laboratory Stock #000664 Cell line (Homo- sapiens) HLE HCC cell line JCRB Cell Bank JCRB0404 Cell line (M. musculus) Myc; sgTrp53 HCC cell lines This paper Cell line (M. musculus) Myc; sgKmt2c HCC cell lines This paper Cell line (M. musculus) Myc; sgAxin1 HCC cell lines This paper Cell line (M. musculus) Cell line (M. musculus) TRE- shKmt2c.1 liver progenitor line TRE- shKmt2c.2 liver progenitor line This paper This paper NA NA NA NA NA Three independent cell lines derived from different mice were used as biological replicates Three independent cell lines derived from different mice were used as biological replicates Three independent cell lines derived from different mice were used as biological replicates Antibody Recombinant DNA reagent Recombinant DNA reagent Recombinant DNA reagent Recombinant DNA reagent Recombinant DNA reagent Anti- MLL3/4 (Rabbit polyclonal) Dorighi et al., 2017; PMID:28483418 NA ChIP- seq (1:500) pT3- Myc Addgene #92046 pT3-Ctnnb1 N90 Addgene #31785 PX330- Cas9- U6- sgRNA Addgene #42230 CMV- SB13 pT3- EF1a- GFP- miRE Huang et al., 2014; PMID:25128497 Huang et al., 2014; PMID:25128497 NA NA Zhu, Soto- Feliciano, Morris et al. eLife 2023;12:e80854. DOI: https://doi.org/10.7554/eLife.80854 25 of 25 Cancer Biology
10.7554_elife.80923
RESEARCH ARTICLE Flexible specificity of memory in Drosophila depends on a comparison between choices Mehrab N Modi1, Adithya E Rajagopalan1,2, Hervé Rouault1,3, Yoshinori Aso1, Glenn C Turner1* 1HHMI Janelia Research Campus, Ashburn, United States; 2The Solomon H. Snyder Department of Neuroscience, Johns Hopkins University School of Medicine, Baltimore, United States; 3Aix- Marseille Univ, Université de Toulon, CNRS, CPT (UMR 7332), Turing Centre for Living Systems, Marseille, France Abstract: Memory guides behavior across widely varying environments and must therefore be both sufficiently specific and general. A memory too specific will be useless in even a slightly different environment, while an overly general memory may lead to suboptimal choices. Animals successfully learn to both distinguish between very similar stimuli and generalize across cues. Rather than forming memories that strike a balance between specificity and generality, Drosophila can flexibly categorize a given stimulus into different groups depending on the options available. We asked how this flexibility manifests itself in the well- characterized learning and memory pathways of the fruit fly. We show that flexible categorization in neuronal activity as well as behavior depends on the order and identity of the perceived stimuli. Our results identify the neural correlates of flexible stimulus- categorization in the fruit fly. Editor's evaluation Memory recall is more precise when discrimination is required. This work in Drosophila shows that two related odors trigger near identical Kenyon cell responses when tested in isolation, but trigger different responses to the second odor if these are experienced in sequence within a small temporal window. The authors argue that this template comparison requires some activity downstream of Kenyon cells, that is recruited by MBONs. Overall, the experiments, building on a clever method to build "miminal memories" via optogenetically restricting the formation of memory traces in selec- tive output compartments of the Kenyon cell (KC) axon terminals, provide very nice physiological evidence for a neural mechanism that underlies a contextual basis for the precision of memory recall. *For correspondence: turnerg@janelia.hhmi.org Competing interest: The authors declare that no competing interests exist. Funding: See page 22 Preprinted: 26 May 2022 Received: 09 June 2022 Accepted: 14 June 2023 Published: 15 June 2023 Reviewing Editor: Mani Ramaswami, Trinity College Dublin, Ireland Copyright Modi et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Introduction Animals routinely encounter competing options and must select between them to survive in complex environments. Making such a choice requires assigning options with subjective values that can be updated through learned experience (Hare et  al., 2011; Hunt et  al., 2012; Glimcher and Fehr, 2013). In such a framework, storing and updating values for every potential option separately would be computationally taxing (Seger and Miller, 2010). Instead, it is beneficial for the brain to main- tain overlapping sensory representations, which would allow options to be grouped downstream into categories based on sensory similarity and assigned with common values (Seger, 2008). Such a coding- scheme would allow animals to distinguish between options in different categories and perform appropriate behavioral responses (Kudryavitskaya et  al., 2021). For example, the most Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 1 of 26 Research article adaptive response when faced with a choice is to pick the highest value stimulus, taking into account the values of available alternatives (Glimcher and Fehr, 2013; Hayden, 2018; Padoa- Schioppa and Conen, 2017). However, this scheme would not readily allow for distinguishing between two options from the same category. It is therefore essential in such a coding- scheme that category boundaries be flexible. One prominent hypothesis suggests that animals directly compare the values of the available stimuli and make use of this relative value signal to guide flexible categorization (Itti and Koch, 2001; Carello and Krauzlis, 2004; Mysore and Knudsen, 2011; Mysore et al., 2011). In this study, we ask how this flexibility arises and identify neural correlates of stimulus comparison in the relatively simple learning and memory circuitry of the Drosophila mushroom body. This flexibility can be studied by examining how animals use an associative memory in two different tasks: discrimination and generalization (Mackintosh, 1974). In a discrimination task, the animal has to choose between a cue associated with reward and a second cue that could either be similar to (hard discrimination) or distinct (easy discrimination) from the original cue. In the generalization task, the flies have to choose between a cue that is perceptually similar to the trained cue and a cue that is very different. The correct choice in this task depends on the animal generalizing its learned response to the similar cue. So the response to the perceptually similar cue differs between the two tasks – the animal chooses it when generalizing and chooses against it when discriminating. Despite the need to switch choices, performance can be extremely high on both these types of tasks (Campbell et al., 2013; Xu and Südhof, 2013; Chen and Gerber, 2014), suggesting that comparisons between avail- able alternatives have a strong impact on animals’ behavioral responses. We employed these paradigms using aversive olfactory conditioning in Drosophila, as its well- studied memory circuit provides a strong framework to understand the neural basis of flexible catego- rization. Olfactory learning takes place in the mushroom body (MB), where odors are represented by sparse activity patterns of the 2000 intrinsic neurons termed Kenyon cells (KCs) (Turner et al., 2008; Murthy et al., 2008; Honegger et al., 2011). Although different odor response patterns are largely uncorrelated, chemically similar odors can elicit partly overlapping patterns of activity (Campbell et al., 2013; Lin et al., 2014). These sensory representations are converted into value- representing memory traces that guide behavioral outputs downstream of the KCs, at the synapses that they form with MB output neurons (MBONs) (Aso et al., 2014b; Hige et al., 2015a; Owald et al., 2015; Villar et  al., 2022). These  ~30 distinct MBONs form compartments that integrate input from different subsets of KCs (Ito et al., 1998; Strausfeld et al., 2003; Lin et al., 2007; Tanaka et al., 2008; Aso et al., 2009; Aso et al., 2014a; Takemura et al., 2017; Li et al., 2020). Furthermore, these synapses are plastic and primarily undergo depression as flies learn an olfactory association (Berry et al., 2018; Cohn et  al., 2015; Hige et  al., 2015a; Perisse et  al., 2016; Séjourné et  al., 2011; but see also Plaçais et al., 2013; Stahl et al., 2022). This plasticity is mediated by dopaminergic neurons (DANs) that convey information about reward or punishment and arborize in corresponding compartments as the MBONs forming a series of DAN- MBON modules (Aso et al., 2014a). The degree to which KC response patterns overlap has been shown to drive the specificity of learning and resulting behavior (Campbell et al., 2013; Lin et al., 2014). During learning, synapses from odor- activated KCs to specific MBONs are depressed. A similar odor with an overlapping KC response pattern thus also exhibits a reduced synaptic drive onto the same MBONs (Hige et al., 2015a; Perisse et al., 2016; Berry et al., 2018). The greater the overlap, the more extensive the depression of this other odor’s activation of the MBONs, and the greater the generalization. In contrast, when overlap is low, downstream MBON activity is minimally affected and the animal discriminates between the two cues. Although this model has a lot of explanatory power, it does not include a means for explicitly comparing between available options. A comparison would allow for the estimation of a relative value between the available options and explain the high performance of flies in both gener- alization and discrimination tasks. The explanatory framework must therefore move beyond the idea of overlap in sensory representations. In this study, we combined neural activity measurements with behavioral experiments, to expand our understanding of flexible categorization. We observed that flies could achieve high levels of performance for both discrimination and generalization tasks and identified a single MB compart- ment capable of supporting both. Surprisingly, MBON responses in this compartment showed no measurable stimulus- specificity to simple pulses of the two similar odors we used, despite being able to distinguish them behaviorally. However, when we presented odors in sequence, one transitioning Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 2 of 26 Neuroscience Research article immediately into the other, similar to what flies experience in the behavioral task, we found that MBON responses to these odors were clearly distinct. These findings show that MBON activity is modulated by a temporal comparison of the alternatives presented to the fly, allowing for switches in the categorization of odor stimuli. Importantly, KC representations did not show categorization switching to either simple stimuli or transitions suggesting the involvement of downstream mecha- nisms. Moreover, behavioral experiments showed that these comparisons are made when stimuli are experienced close together in time. Both imaging and behavior provide complementary evidence that comparing available alternatives ‘side- by- side’ in time is important for flexible categorization.These results show that the MB circuit implements a comparison, augmenting small differences between overlapping sensory representations to guide flexible stimulus categorization and choice behavior. Results Precision of memory recall depends on MB compartment Previous work has shown that flies are capable of high levels of performance on both hard discrimina- tion and generalization tasks (Campbell et al., 2013). This study identified a trio of odors to use for experiments on the specificity of memory, based on the degree of overlap of KC response patterns: pentyl acetate (PA) butyl acetate (BA) and ethyl lactate (EL) (Figure 1A, left). PA and BA are chemically similar and elicit highly overlapping response patterns in the KC population (Campbell et al., 2013). EL is distinct, both chemically and in terms of KC response patterns. Choices between different combi- nations of these cues can be used to test flies’ ability to flexibly classify odors and measure memory specificity. Take, for example, an experiment where flies are trained to form an association with PA. We can present flies with a difficult discrimination task by giving them a choice between the similar odors (PA and BA), or an easy discrimination with a choice between the paired odor (PA) and the dissimilar odor (EL) (Figure 1A, right). We can also test whether the association with PA generalizes to the similar odor BA, by giving flies a choice between BA and EL. Since we use these odors in many different combinations for different task structures, with and without reciprocal design, here we will use A to refer to the paired odor (PA or BA) and A’ to refer to the other similar odor, which is unpaired, while B always refers to the dissimilar odor, EL. With this nomenclature, hard discrimination involves an A versus A’ choice, easy discrimination is A versus B and generalization is A’ versus B (Figure 1A). Although previous work showed flies can flexibly categorize odors and learn both generalization and discrimination tasks using these odors, electric shock was used as the reinforcement (Campbell et al., 2013). Consequently the synaptic changes responsible were likely distributed across many areas of the mushroom body, and possibly elsewhere. To confine plasticity to a more restricted region of the brain, we used optogenetic reinforcement, pairing the activation of specific DANs with odor presen- tation (Figure  1B; Claridge- Chang et  al., 2009; Schroll et  al., 2006). We used drivers to express CSChrimson in specific DANs from the PPL1 cluster that target different compartments involved in aversion learning: α3 (MB630B) and γ2α’1 (MB296B) (Aso et al., 2014a; Aso and Rubin, 2016). Since compartments have different time courses for memory acquisition and recall (Aso and Rubin, 2016), the number of repetitions of odor- reinforcement pairing and the time between training and testing differed depending on the compartment tested (see Methods). We found that these two compartments exhibited contrasting properties in the easy and hard discrimination tasks (Figure 1C). Flies that received reinforcement from DAN PPL1-α3 were poor at the hard discrimination, although they performed significantly better on the easy task (Figure  1D, p=0.007, n=12). On the other hand, flies that received optogenetic reinforcement via DAN PPL1-γ2α’1 performed the hard discrimination as effectively as the easy discrimination (Figure  1E, p=0.08, n=12). Empty driver controls performed no better than chance at either easy or hard discrimina- tion (Figure 1—figure supplement 1A, p=0.052, p=0.38, n=12). These results show that these two compartments have different capacities for discrimination, with α3 weakly discriminating and γ2α’1 stronger. The difference in ability to support fine discrimination between these two compartments raises the question of whether and how they differ in a generalization task. In a simple model where perfor- mance reflects overlap between the test stimulus and the trained odor, the harder the discrimina- tion, the easier the generalization. Does the weakly discriminating α3 compartment support strong Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 3 of 26 Neuroscience Research article A C butyl acetate O O pentyl acetate O O ethyl lactate O O OH odor A' odor B train A B B A y s a e test / A B / A B 0 2 min train A A' d r a h A' A test / A A' / A A' 0 2 min train A A -2 0 2 train A A B B B B test / A' B / A' B min test / A B / A B n o i t a n m i i r c s d i F n o i t a z i l a r e n e g n o i t a n m i i r c s d i y s a e -2 0 2 min training easy discrimination B odor + reinforcement KCs DANs MBONs hard discrimination generalization α3 * E γ2α'1 n.s. easy hard easy hard α3 n.s. H γ2α'1 n.s. γ2α'1 * I x e d n i e c n a m r o f r e p 0.6 0.4 0.2 0 -0.2 discr. gen. discr. gen. -0.4 DA rescue knockout hard discrimination D x e d n i e c n a m r o f r e p 0.8 0.6 0.4 0.2 0 -0.2 G x e d n i e c n a m r o f r e p 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 Figure 1. A single set of changed synapses can result in generalization or discrimination. (A) Left: Chemical structures of the three odors used in the study, the similar odors butyl acetate (BA) and pentyl acetate (PA) and the dissimilar odor ethyl lactate (EL). Middle: During training the similar odors are interchangeably used as the odors that are paired (A) or unpaired (A’) with optogenetic reinforcement (LED). Right: Trained flies are then given one of three different choices between odors in opposing arena quadrants. These choices represent the three kinds of tasks used here to study memory specificity. Performance index measures the bias in the distribution of flies across the different quadrants (see Methods). The circles depict fly population behavior in our arenas and the vertical bars depict stimulus choices. The dashed, red line depicts the discrimination boundary in each choice. This boundary shifts relative to the light- green stimulus, depending on the options. (B) Mushroom body learning schematic. KCs activated by an odor (blue) form synapses on MBONs in two compartments (red and gray shading). Reinforcement stimulates the DAN projecting to one compartment (red) leading to synaptic depression. (C) Behavior protocols for discrimination tasks at two levels of difficulty. Colored bars represent odor delivery periods, red dashes indicate LED stimulation for optogenetic reinforcement. A represents the paired odor, A’ the similar odor and B the dissimilar odor. (D) Significantly lower performance on the hard discrimination task with reinforcement to α3 (p=0.007, n=12). Flies received 10 cycles of training and were Figure 1 continued on next page Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 4 of 26 Neuroscience Research article Figure 1 continued tested for memory 24 hours later. CsChrimson- mVenus driven in DAN PPL1-α3 by MB630B- Gal4. (E) No significant difference in performance on easy versus hard discrimination with reinforcement to γ2α’1 (p=0.08, n=12 reciprocal experiments). Flies received three cycles of training and were tested for memory immediately after. CsChrimson- mVenus driven in DAN PPL1 γ2α’1 by MB296B- Gal4. (F) Behavior protocol for generalization. Scores here are compared to a control protocol where light stimulation is not paired with odor presentation in time. (G) No significant difference in performance on generalization and easy discrimination with reinforcement to α3 (p=0.84, n=12). Flies received 10 cycles of training and were tested 24 hr later. (H) No significant difference in performance on generalization and easy discrimination with reinforcement to γ2α’1 (p=0.89, n=12 unpaired control performance scores). Flies received three cycles of training and were tested immediately after. (I) Rescue of the dopamine biosynthesis pathway in DAN PPL1-γ2α’1 is sufficient for performance on the hard discrimination task (p=0.04, n=8). Black circles and error bars are mean and SEM. Statistical comparisons made with an independent sample Wilcoxon rank sum test. The online version of this article includes the following figure supplement(s) for figure 1: Figure supplement 1. Control behavior experiments. performance on generalization, while the strongly discriminating γ2α’1 compartment does not? Or does the γ2α’1 compartment somehow have the flexibility to support strong performance on both tasks? We tested this by examining relative performance on generalization and easy discrimination tasks in these two compartments. We kept a parallel structure between the two types of tasks by quanti- fying performance against control experiments where optogenetic stimulation was delivered unpaired to odor delivery (Figure 1F; see Methods; note that since these experiments did not have reciprocal controls the performance scores in Figure 1G–H are computed differently than in Figure 1D–E). As expected, training flies using DAN PPL1-α3 yielded similarly high performance on both generaliza- tion and easy discrimination tasks (Figure 1G, p=0.84, n=12), while empty driver controls performed no better than chance (Figure 1—figure supplement 1A and B, p=0.052, p=0.91, n=12). However, performance on the generalization task was also high in the strongly discriminating compartment γ2α’1, with a performance level indistinguishable from that in the easy discrimination task (Figure 1H p=0.89, n=12). Although the experiments above target optogenetic punishment to specific sites within the MB, there is the possibility that there are secondary sites of plasticity that contribute to the behavioral performance we observe, via indirect connections between MB compartments. To more rigorously confine plasticity to γ2α’1, we performed an experiment where dopamine production is restricted solely to DAN PPL1-γ2α’1 within the fly. Dopamine is necessary for flies to show any measurable aversive learning (Aso et al., 2019; Kim et al., 2007; Qin et al., 2012), and its production requires the Drosophila tyrosine hydroxylase enzyme, DTH (Cichewicz et al., 2017; Neckameyer and White, 1993; Riemensperger et al., 2011). So we examined performance of flies lacking DTH throughout the nervous system (Cichewicz et al., 2017), but with production rescued specifically in PPL1-γ2α’1 by driving expression of UAS- DTH using the split hemidrivers TH- DBD and 73F07- AD (Aso et al., 2019). Performance was significantly higher for the DTH- rescue flies than for the mutants in hard discrimina- tion (Figure 1I, p=0.038, n=8) and generalization tasks (Figure 1—figure supplement 1D, p=0.041, n=6), indicating that plasticity in this set of synapses is sufficient for both behaviors (For control exper- iments with easy discrimination, see Figure 1—figure supplement 1C). These results show that a single memory trace formed via plasticity confined to γ2α’1 supports strong performance on the hard discrimination and generalization tasks. We note that the choice outcomes of these paradigms are opposite: in the generalization experiments flies distribute away from odor A’, while in the hard discrimination task, flies accumulate in the A’ quadrant. We next sought to understand how plasticity in this one compartment can result in this flexible categorization of A’. KC inputs to both MB compartments contain enough information for discrimination We started by evaluating whether the odor inputs to the γ2α’1 and α3 compartments carry enough information to discriminate between the two similar odors used in our behavior experiments. Previous measurements of KC responses to these odors showed that they exhibit overlapping response patterns, but did not determine whether that overlap was differentially distributed across different KC subtypes (Campbell et al., 2013). We used two- photon calcium imaging to measure cell population responses in the KC subtypes that send axons to γ2α’1 (γ and α’/β’ KCs) and α3 (α/β KCs) (Figure 2A Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 5 of 26 Neuroscience PA BA EL C γ KCs 20 60 r e b m u n input for γ2α'1 l l e c 100 140 15 30 15 30 time (s) 15 30 15 30 15 30 time (s) 15 30 Research article A odor B d 1 2 10 m 1 / F F Δ 4 4 s 2 10.0 7.5 5.0 2.5 0.0 F/FΔ 20 60 r e b m u n l l e c 100 α'/β' KCs input for γ2α'1 r e b m u n l l e c 20 40 60 α/β KCs input for α3 D 10 2 C P 5 0 40 30 2 C P 20 10 1.5 1.0 0.5 0.0 ΔF/F 6.0 4.0 2.0 0.0 ΔF/F 3.0 2.0 1.0 0.0 ΔF/F n.s. n.s. E 100 r e d o c e d 0 5 PC1 10 50 0 PA EL BA n.s. n.s. 100 0 -20 20 0 PC1 2 C P 6 4 2 0 50 0 PA BA EL n.s. n.s. 100 50 0 PA BA EL y c a r u c c a y c a r u c c a r e d o c e d y c a r u c c a r e d o c e d 15 30 30 15 time (s) 15 30 0 20 10 PC1 Figure 2. KC responses to single odor pulses contain enough information to discriminate similar odors. (A) Schematic of in vivo imaging preparation. (B) Example single- trial odor response patterns in α/β KCs. ΔF/F responses (color bar) are shown overlaid on baseline fluorescence (grayscale). Numbered circles indicate cells for which ΔF/F traces are plotted below. Black bar indicates odor delivery. (C) ΔF/F responses of different KC subtypes to the three odors used in this study. Rows show responses of individual KCs, averaged across trials, sorted by responses to PA. GCaMP6f was driven in γ KCs by d5HT1b- Gal4, in α’/β’ KCs by c305a- Gal4 and in α/β KCs by c739- Gal4. Colored bars above plots indicate the odor delivery period. (D) Odor response patterns for the same example flies as in C, projected onto the first two principal component axes to show relative distances between representations for the different odors. (E) Decoder prediction accuracies, plotted across flies. Each gray circle is the accuracy of the decoder for one fly for a given odor, averaged across all trials. Black circles and error bars are means and SEM. For γ KCs (top), decoder accuracies for PA (n=7 flies, p=0.06) and BA (p=0.08) were not significantly different from EL accuracy. This was also true for α’/β’ KCs (middle, n=5 flies, p=0.13 for the PA- EL comparison and p=0.13 for BA- EL) and α/β KCs (bottom, n=6 flies, p=0.63 for PA- EL and p=0.73 for BA- EL). All statistical testing was done with a paired- sample, Wilcoxon signed rank test with a Bonferroni- Holm correction for multiple comparisons. (C,D) In this figure, each shade of green denotes one of the two similar odor chemicals. But in subsequent figures, the darker shade represents the odor paired with reinforcement and the lighter shade, the unpaired, similar odor. In the reciprocal design we use, each of the odor chemicals is the paired odor in half the experimental repeats. The online version of this article includes the following figure supplement(s) for figure 2: Figure supplement 1. Similar odors have similar KC response patterns. and B). In separate sets of flies, GCaMP6f (Chen et al., 2013) was expressed in γ KCs (d5HT1b Yuan et  al., 2006), α’/β’ KCs (c305a Armstrong et  al., 2006; Krashes et  al., 2007) and α/β KCs (c739 McGuire et  al., 2001). γ and α’/β’ KCs had to be imaged separately since there is no driver that exclusively labels both subtypes. The trial- averaged response traces of individual KCs for each of the three subtypes showed that many of the same cells respond to the two similar odors (PA and BA), but representations did not completely overlap (Figure  2C). Responses were very different for the dissimilar odor, EL. KC population response vectors from single trials, plotted as projections along the first two principal component axes (Figure  2D), also show the similarity in KC representations between the chemically similar odors. Finally, we examined the similarity of responses for individual KCs to the different pairs of odors. Pooling cells across all imaged flies, we found that similar odors elicited similar response strengths in individual KCs (Figure 2—figure supplement 1A, γ KCs: r=0.74, p<0.001; α’/β’ KCs: r=0.76, p<0.001  and α/β KCs: r=0.63, p<0.001). Correlation coefficients were Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 6 of 26 Neuroscience Research article lower and were not significant for the dissimilar odors (Figure  2—figure supplement 1B, γ KCs: r=0.04, p=0.60; α’/β’ KCs: r=0.06, p=0.55 and α/β KCs: r=–0.04, p=0.77). To quantify how effectively KC activity patterns could distinguish between odors, we used logistic regression models to determine the probability a particular odor evoked the KC activity pattern observed on a given trial. We trained logistic regression decoders to recognize KC response patterns using leave- one- out cross- validation. We computed the average decoder accuracy for the 8 odor presentation trials of each odor for each fly. Decoder accuracies for the two similar odors were as high as they were for the dissimilar odor, across all KC subtypes (Figure 2E)(γ KCs: comparing accuracies for PA and EL p=0.06, BA- EL p=0.08, n=7 flies; α’/β’ KCs: PA- EL p=0.13, BA- EL p=0.13, n=5 flies; α/β KCs: PA- EL p=0.63, BA- EL p=0.73, n=6 flies). Even though compartments γ2α’1 and α3 receive olfactory input from totally distinct subsets of KCs, input activity patterns appear capable of supporting fine discrimination in all three KC subtypes. Is this information retained one synapse downstream, when hundreds of KCs converge onto the MBONs in these two compartments? Plasticity in MBON γ2α’1 is not sufficiently odor-specific for discrimination We next examined plasticity in the downstream MBONs, to test whether activity of MBON-γ2α’1 could potentially support fine discrimination after training. We carried out on- rig optogenetic rein- forcement, and imaged MBON-γ2α’1 odor responses pre- and post- pairing (Figure 3A). γ2α’1 spans parts of both the γ and α’ MB lobes, but receives reinforcement from the single DAN, PPL1-γ2α’1 (Aso et al., 2014a). Two MBONs send dendrites to the same region of neuropil; here we treat them as a single cell type, MBON-γ2α’1, and we imaged from their overlapping dendritic projections (Figure 3B and C). We expressed Chrimson88.tdTomato (Strother et al., 2017) in the DAN PPL1-γ2α’1 (driven by 82C10- LexA which also drives weak expression in compartments α2 and α3 Pfeiffer et al., 2013) and opGCaMP6f selectively in MBON-γ2α’1 (MB077B Aso et al., 2014a). We imaged MBON-γ2α’1 responses to pulses of all three odors, before and after pairing one of the similar odors with opto- genetic reinforcement (Figure 3A–C). We delivered two presentations of each odor stimulus before and after pairing, and imaged only one, to minimize adaptation effects (Berry et al., 2018). Based on previous studies, we expected to see depression of the MBON-γ2α’1 response specifically (or at least preferentially) for the reinforced odor (Berry et al., 2018; Cohn et al., 2015; Hige et al., 2015a; Owald et al., 2015; Perisse et al., 2016; Séjourné et al., 2011). However, after pairing, MBON-γ2α’1 responses to A and A’ were both strongly depressed (Figure 3D and E, p=0.001 for A and p=0.001 for A’, n=11 flies). In fact we could not detect a difference in response size between the two, even though only one (A) had been paired with reinforcement (Figure 3F, p=0.77). As expected, responses to the dissimilar odor (B) were not affected (p=0.18). This strong depression of MBON-γ2α’1 responses to both similar odors suggests that downstream of the KCs, A and A’ are grouped into the same category. Such a grouping should elicit the same behavior response to both odors and allow for generalization. How then do flies discriminate between them after learning in our hard discrimination task? We postulated that the apparent discrepancy between our behavioral observations and measurements of MBON activity might be because we did not adequately reproduce the fly’s sensory experience when it is presented as a choice between two odors. MBON responses to odor transitions reflect discrimination behavior When flies make a choice between two odors in the behavioral arena, they encounter an odor boundary, where the concentration of one odor rapidly drops off and the other rises. To mimic this experience while imaging neural activity on the microscope, we designed an odor delivery system to deliver rapid transitions between odors. We characterized the performance of this odor delivery system with a photo- ionization detector to measure odor concentration changes and an anemometer to measure airflow (Figure 4—figure supplement 1). We then examined how plasticity affects MBON-γ2α’1 responses to these odor transitions. As above, we used single odor pulses for training, to match how flies are trained behaviorally. However, we examined MBON responses to odor transitions pre- and post- pairing, to match how flies expe- rience the choice between odors. These results showed a sharp contrast to our observations with Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 7 of 26 Neuroscience Research article A A A' B pre-pairing 5 s 45 s 45 s 45 s odor-LED pairing 2 s post-pairing DAN γ2α'1 Chrimson B C / F F Δ 4 4 s d MBON-γ2α'1 GCaMP6f 5.0 2.5 10 m 0.0 ΔF/F 8s mean * * n.s. E 5 4 3 2 1 0 / ) F F d ( e z s i e s n o p s e r -1 pre-pairing post-pairing 10 time (s) 20 8s mean n.s. n.s. D / ) F F Δ ( e s n o p s e r r o d o 4 2 0 4 2 0 4 2 0 0 F / ) F F d ( e z s i e s n o p s e r 4 3 2 1 0 pre post pre post pre post paired unpaired similar unpaired dissimilar -1 paired unpaired paired unpaired pre post Figure 3. Plasticity in MBON γ2α’1 is not sufficiently odor specific for hard discrimination. (A) Stimulus protocol for on- rig, in vivo training. Plasticity in MBON γ2α’1 was assessed by imaging pre- and post- pairing with optogenetic reinforcement via DAN PPL1-γ2α’1. There was no imaging during the pairing itself. Colored bars represent 5 s odor delivery (dark green: odor A paired; light green: odor A’ unpaired similar; purple: odor B unpaired dissimilar). PA and BA were used as the paired odor for every alternate fly. (B) Schematic of experimental design. Expression of GCaMP6f in MBON-γ2α’1 driven by MB077B- Gal4 and Chrimson88- tdTomato in DAN PPL1-γ2α’1 by 82C10- LexA. Imaging plane in the γ lobe as indicated. (C) Example MBON-γ2α’1 single- trial odor response. ΔF/F responses (color bar) are shown overlaid on baseline fluorescence (grayscale). White ROI indicates neuropil region for which a single- trial ΔF/F trace is plotted below. Black bar indicates odor delivery. (D) MBON γ2α’1 ΔF/F response traces pre- (grey) and post- (black) pairing (mean +- SEM, n=11 flies). Bars indicate 5 s odor delivery period; colors Figure 3 continued on next page Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 8 of 26 Neuroscience Research article Figure 3 continued correspond to odor identities in a. (E) Response sizes pre- and post- pairing show a reduction for both paired (dark green, p=0.001), and similar unpaired (light green, p=0.001) odors but not the dissimilar odor (purple, p=0.18). Response amplitude calculated as mean ΔF/F over an 8 s window starting at odor onset (inset). Connected circles indicate data from individual flies. (F) Data as in E, re- plotted to compare responses to the paired odor with responses to the unpaired, similar odor before and after training. Response sizes were not significantly different pre- (p=0.77) or post- pairing (p=0.77). Statistical comparisons made with the paired sample Wilcoxon signed rank test, with a Bonferroni Holm correction for multiple comparisons. single odor pulses. Surprisingly, the depression of responses to A’ seen in single pulses was not readily apparent in an A to A’ transition (Figure 4A). Responses were similar to pre- pairing levels when A’ was preceded by A, but not when the order was reversed (Figure 4B). Indeed, quantifying the size of the MBON response to the second pulse showed A’ responses were not significantly different pre- and post- pairing (Figure 4C; p=0.376, n=13). As expected, responses to A as the second pulse were significantly lower after pairing (Figure 4C; p<0.001, n=13). The contrast with the single odor pulse results is clearest when comparing responses to A versus A’ as the second pulse in a transi- tion, where A’ responses were now significantly larger (Figure 4—figure supplement 2C p=0.004). Control experiments where LED stimulation was omitted showed no significant differences pre- and post- mock pairing (Figure 4—figure supplement 3A–C). Additionally, when we examined responses to A- B and A’-B transitions before and after training, we saw no effect on responses to odor B, indi- cating that transitions selectively enhance the otherwise depressed responses to the similar odor A’ (Figure 4—figure supplement 3D–F). These results indicate that the way the fly encounters the odor has a profound effect on MBON responses after learning. Isolated pulses of A and A’ elicit similar strongly depressed responses, while transitioning from one odor to the next, as at an odor boundary, responses were clearly distinct, with the A’ response now much stronger. To quantify how effectively MBON-γ2α’1 activity captures an odor boundary, we computed a contrast score reflecting the change in MBON activity at the transition. This score was the difference between the minimum ΔF/F value during the first pulse and the maximum during the second pulse. After learning, contrast around the odor transition was significantly higher for A to A’ transitions than the reverse (Figure 4—figure supplement 2D; p=0.001). These results show that the way the animal experiences the odors has a significant effect on how differentially the downstream MBON-γ2α’1 responds to them. To further evaluate whether this effect contributes to hard discrimination, we next examined the effects of plasticity on odor transition responses in MBON-α3. Reinforcement in this compartment does not support fine discrimination, so if the transition effect is important for discrimination, it should be absent here. α3 is a slow- learning compartment; when an odor is paired with reinforcement via DAN PPL1-α3, behavioral performance gradually rises until it peaks 24 hr after training (Aso and Rubin, 2016). To examine MBON responses when behavioral performance is at this peak, we could not use on- rig optogenetic reinforcement. Instead, we trained flies in a behavior chamber, by pairing an odor with a shock reinforcement. We retrieved flies from the arena and imaged MBON-α3 responses 20–28 hr after training (Figure 4D, detailed protocol in Figure 4—figure supplement 2B). This experimental approach did not permit us to measure pre- and post- training responses in the same fly. So in these experiments, we compared responses observed in trained flies with those in a mock- trained cohort, where shock was delivered at a different time than odor (Figure 4—figure supplement 2B). We note that optogenetic reinforcement was not possible for experiments targeting the α3 compartment; despite extensive efforts we were unable to identify LexA driver lines either with sufficient strength to image MBON-α3 activity, or to get effective reinforcement selectively via DAN PPL1-α3 (Figure 4— figure supplement 4). We found that in response to odor transitions, there was no modulation in MBON-α3 responses (Figure  4E). Responses to the second odor in the transition were depressed for both transition orders (Figure 4F; A’-A, p<0.001, A- A’, p<0.001, pooled n=14 PA- paired and n=12 BA- paired flies). As expected, responses to the dissimilar odor (B) showed no significant depression (p=0.48). We also evaluated odor boundary detection post- training by computing a contrast score as we did for Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 9 of 26 Neuroscience Research article A B imaging pre-pairing odor-LED pairing imaging post-pairing DAN γ2α'1 Chrimson MBON-γ2α'1 GCaMP6f pre-pairing post-pairing 3 2 1 0 3 2 1 0 6 4 2 0 / ) F F Δ ( e s n o p s e r r o d o 24 hrs odor-shock pairing imaging post-pairing MBON-α3 GCaMP6f control post-pairing D E / ) F F Δ ( e s n o p s e r r o d o 3 2 1 0 3 2 1 0 4 3 2 1 0 0 20 10 time (s) 8s mean n.s. n.s. * 0 20 10 time (s) * * n.s. F / ) F F d ( e z s i e s n o p s e r 8 6 4 2 0 pre post paired pre post unpaired similar pre post unpaired dissimilar control paired unpaired similar odors control unpaired dissimilar odor C / ) F F d ( e z s i e s n o p s e r 6 5 4 3 2 1 0 Figure 4. Odor transitions enable discrimination by MBON-γ2α’1 but not MBON-α3. (A) Schematic of experimental design to assess plasticity in MBON-γ2α’1. Protocol was identical to Figure 3A, except odor transitions were used as pre- and post- pairing test stimuli to mimic odor boundaries from the behavioral arena. See Figure 4—figure supplement 2A for the detailed protocol. (B) MBON-γ2α’1 ΔF/F response time courses pre- Figure 4 continued on next page Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 10 of 26 Neuroscience Research article Figure 4 continued (gray) and post- (black) pairing to different odor transitions (mean +- SEM, n=13 flies). Colored bars indicate timing of odor delivery, dark green: paired odor, light green: unpaired, similar odor, purple: dissimilar odor. Schematic of the corresponding fly movement in the behavior arena at left. (C) Response sizes in MBON-γ2α’1 pre- and post- pairing for the second odor in the transition. Responses calculated as mean over an 8 s window starting at the onset of the second odor (inset). Responses when the paired odor is second are significantly reduced after pairing (dark green n=13 flies, p<0.001). By contrast there is no significant reduction when the unpaired similar odor comes second (light green p=0.376). The dissimilar odor control showed no significant change (purple, p=1). Connected circles indicate data from individual flies. (D) Schematic of experimental design for MBON-α3. Flies were trained by pairing odor with shock in a conditioning apparatus and odor responses were imaged 24 hr later. Plasticity was assessed by comparing responses against those from a control group of flies exposed to odor and shock but separated by 7 min. See Figure 4—figure supplement 2B for the detailed protocol. GCaMP6f was driven in MBON-α3 by MB082C- Gal4 (bottom). (E) MBON-α3 ΔF/F response traces as in B. Light gray traces are from control shock- exposed flies (n=12 flies), dark gray traces are from odor- shock paired flies (pooled n=14 PA- paired and n=12 BA- paired flies). Averaged control traces were pooled across trials where either PA or BA was the second odor in a transition. (F) Response sizes in MBON-α3 in control (gray) and trained flies (colored) for the second odor in the transition (computed as in C). Responses are significantly reduced in trained flies both when the paired odor is second (dark green, n=14 PA- paired and n=12 BA- paired flies pooled, p<0.001) and when the unpaired similar odor is second (light green, p<0.001). Responses to the dissimilar odor were not significantly different (purple, p=0.48). Plotted control responses were pooled across trials where PA or BA was the second odor in a transition. (C,F) Statistical comparisons for MBON-γ2α’1 made with the paired sample, Wilcoxon signed- rank test. For MBON-α3, where responses were compared across different flies, we used the independent samples Wilcoxon’s rank- sum test. p- values were Bonferroni- Holm corrected for multiple comparisons. The online version of this article includes the following figure supplement(s) for figure 4: Figure supplement 1. Time courses of odor delivery and air flow for odor pulses and odor transitions. Figure supplement 2. Contrast around odor transitions is high for MBON-γ2α’1 but not MBON-α3. Figure supplement 3. Observations from no- LED control training and odor transitions to the dissimilar odor. Figure supplement 4. PPL1-α3 split- LexA lines drive expression poorly. MBON-γ2α’1. With MBON-α3, we saw very little contrast at the transition point, and contrast was similarly low for either order of the transition (Figure 4—figure supplement 2E, P=0.86). Mimicking the fly’s experience in the behavioral arena by presenting odor transitions revealed a strong concordance between neural activity and behavior. In MBON-γ2α’1, when odors are presented in isolation, responses to A and A’ were not measurably different (Figure 3). This would allow flies to generalize learning between these odors in most circumstances. But when the two similar odors are juxtaposed in time, matching what they experience when making a choice, MBON-γ2α’1 responses were clearly distinct and could support fine discrimination (Figure 4). The ability of MBON-γ2α’1 to respond differently in these two conditions likely reflects the flexible categorization that enables flies to perform both generalization and discrimination. In agreement with this hypothesis, the effect of odor transitions is absent in the α3 compartment, which does not support fine discrimination. Odor transition effects on MBONs are not present in the KCs We have shown that MBON-γ2α’1 responses show a stimulus- history dependent modulation at odor transitions, but MBON-α3 does not. To determine whether this arises upstream of the MBONs, we examined KC responses to odor transitions. Early sensory processing in the antennal lobe could alter odor representations when delivered as transitions, as seen in locusts (Nizampatnam et al., 2018; Saha et  al., 2013). So we examined responses to odor transitions in the input KC populations for MBON-γ2α’1 (γ and α’/β’ KCs) and MBON-α3 (α/β KCs) (Figure 5A). We attempted to use the KC activity patterns we measured to reproduce our observations of MBON activity. Specifically, we used logistic regression models, adjusting the weights of KC inputs so that model outputs were low for A and A’ and high for B. To match the training procedure the flies experienced, we first trained the models using isolated odor pulses, and tested predictions for odor transition stimuli. To ensure we did not penalize cells that responded uniquely to transitions, weights were initialized at 1 and we trained models without any weight regularization. Trained weights were negatively correlated with responses to A, as expected (Figure 5B, top, Pearson correlation coefficient for γ KCs = - 0.39, α’/β’ = - 0.28 and α/β = - 0.22). These weights were then used to calculate model output for A’-A and A- A’ transitions. In Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 11 of 26 Neuroscience Research article γ KCs α'/β' KCs α/β KCs input for γ2α'1 input for γ2α'1 input for α3 BA PA PA BA BA PA PA BA BA PA PA BA A B train = 0 = 0 = 1 1 t u p t u o . r g e r . g o l 0.8 0.6 0.4 0.2 0 C train = 0 = 0 = 1 = 0 = 1 1 t u p t u o . r g e r . g o l 0.8 0.6 0.4 0.2 0 20 60 r e b m u n l l e c 100 140 15 30 15 30 time (s) 20 0 i t h g e w C K 1.5 1.0 0.5 0.0 ΔF/F 20 60 100 6.0 4.0 2.0 0.0 ΔF/F 20 40 60 15 30 15 30 time (s) 5 0 3.0 2.0 1.0 0.0 ΔF/F 15 30 15 30 time (s) 10 0 -10 -20 0 0.5 1 -20 0 1 0.5 norm. resp. to A -5 0 0.5 1 n.s. n.s. n.s. n.s. n.s. n.s. 1 0.8 0.6 0.4 0.2 0 1 0.8 0.6 0.4 0.2 0 A A' B A'-A A-A' A A' B A'-A A-A' A A' B A'-A A-A' odor odor odor 20 i t h g e w C K 0 -20 0 1 0.5 norm. resp. to A n.s. * A A' B A'-A A-A' odor 1 0.8 0.6 0.4 0.2 0 5 0 -5 0 0.5 1 * * A A' B A'-A A-A' odor 20 0 -20 0 n.s. 0.5 1 * AAAA A A' B A'-A A-A' odor 1 0.8 0.6 0.4 0.2 0 Figure 5. KC responses to odor transitions are not sufficient for hard discrimination. (A) ΔF/F responses of different KC subtypes to odor transitions. Rows show responses of individual KCs, averaged across trials, sorted by responses to BA- PA. GCaMP6f was driven in γ KCs by d5HT1b- Gal4 (left), in α’/β’ KCs by c305a- Gal4 (middle) and in α/β KCs by c739- Gal4 (right). The odor delivery periods are indicated by colored bars at the top. (B) We fitted KC weights with logistic regression to give high or low outputs to odors consistent with measured MBON outputs (synaptic weight plots, black circles Figure 5 continued on next page Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 12 of 26 Neuroscience Research article Figure 5 continued are individual fitted weights, pooled across flies). Individual logistic regression model outputs for held out test data for all types of odor stimuli are plotted in black. The gray background indicates that odor transition data was not part of the training set (n=96 models for γ, n=80 for α’/β’ and n=96 for α/β KCs, respectively), red circles and error bars are mean +/-SEM. The dashed, gray line at 0.5 indicates the logistic regression output threshold. Mean model outputs were below the decision threshold for A and A’ and were not significantly different (p=1 for γ, p=1 for α’/β’ and p=1 for α/β KCs, respectively), as was the case for A’-A and A- A’ (p=1 for γ, p=0.95 for α’/β’ and p=1 for α/β KCs, respectively). (C) Fitted weights and outputs for logistic regression models as in B, except that these were trained on single pulse as well as odor transitions. Average model outputs for A and A’ were below the decision threshold and were significantly different only for one KC sub- type (p=0.026 for γ, p<0.001 for α’/β’ and p=0.062 for α/β KCs, respectively). Mean outputs for A’-A were below the decision threshold, but outputs for A- A’ were above it and significantly different for all KC subtypes (p<0.001 for γ, p<0.001 for α’/β’ and p<0.001 for α/β KCs, respectively). All statistical comparisons were made with the Wilcoxon signed- rank test with a Bonferroni- Holm correction for multiple comparisons. The online version of this article includes the following figure supplement(s) for figure 5: Figure supplement 1. KC response patterns are similar for both isolated odor pulses and transitions. contrast to our observations of MBON activity, model outputs were not significantly different between the two transitions, and were low for both (Figure 5B), indicating that transition- evoked changes in the KC odor representations do not underlie the effects on MBON-γ2α’1 activity. In fact, KC responses to single odor pulses were coarsely similar to responses when those odors came second in a transition; these decoders could also effectively discriminate odors in a transition, although accuracy was slightly lower than with isolated pulses (Figure 5—figure supplement 1). To directly evaluate how distinctively the KC population responds to odor transitions, we re- trained the models, adding the requirement that they respond differentially to A- A’ versus A’-A transitions. We found that all three KC subtypes could distinguish odor transitions when trained to do so (Figure 5C, p<0.001 for all KC subtypes). These results show that it is possible for the model to discriminate tran- sitions, but only if trained using transition- evoked KC activity. By contrast flies learn to discriminate when trained solely with the isolated odor pulses. Overall, these results show that MBON activity is modulated by a temporal comparison of the alternatives presented to the fly. These observations lead us to the prediction that even if learning is restricted to the γ2α’1 compartment, flies would only be able to discriminate odors if they experi- enced odor transitions. We tested this prediction with behavioral experiments using odor sequences. Odor sequences show that a temporal comparison contributes to odor discrimination We have shown that MBON-γ2α’1 responses to the similar odors only became distinguishable when presented as transitions. We predicted that flies’ behavioral response to these odors should also be indistinguishable, unless they are encountered as transitions. Further, since MBON-γ2α’1 signals posi- tive valence (Aso et al., 2014b), our activity measurements predict that flies might be attracted to A’ if they encounter an A to A’ transition. To test these predictions, we examined behavioral responses to temporal sequences of odor, converting the spatial odor border flies encountered in our earlier behav- ioral experiments, into an odor transition in time. Flies were trained in the circular arena, and then tested by flooding the entire arena with a sequence of odor pulses. We then compared their behav- ioral response to direct odor transitions to their response when we interrupted the transition with 25 s of clean air. We determined the timing of odor pulse transitions using photo- ionization detector measurements at the exhaust outlet of the arena (Figure 4—figure supplement 1), and analyzed fly behavior around these timepoints. Attraction to an odor was quantified by how much the flies move upwind; in the arena odors flow inwards from the periphery so we measured displacement away from the center of the arena. We examined the time course of upwind displacement for direct and interrupted transitions (Figure 6A–E). We observed strong upwind displacement during the second pulse of an A- A’ transition, which was significantly larger than during the reverse A’-A sequence (Figure 6B, C and F, p=0.004, n=12). This contrasted with results observed with a 25  s gap in between the two odor pulses. In these inter- rupted transitions, responses to the second pulse were not significantly different depending on tran- sition order (Figure  6 D, E and G, p=0.85, n=12 for A- gap- A’, n=13 for A’-gap- A,), and showed a similar degree of upwind displacement to that evoked during the first pulse, as expected. Note that starting locations at the onset of the second odor pulse were not significantly different in any Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 13 of 26 Neuroscience Research article A B ) m m ( . l p s d i i d n w p u training first pulse second pulse no gap training first pulse air gap second pulse 30 s 30 s 30 s 25 s 30 s 5 0 -5 5 0 -5 D 5 0 -5 5 0 -5 0 10 20 0 10 20 0 10 20 0 10 20 C ) m m ( . l p s d i i d n w p u 5 0 -5 5 0 -5 E 5 0 -5 5 0 -5 0 20 10 time (s) 0 20 10 time (s) 0 20 10 time (s) 0 20 10 time (s) n.s. n.s. 10 5 0 -5 F ) m m ( . l p s d i i d n w p u n a e m -10 no gap 25s gap first pulse G ) m m ( . l p s d i i d n w p u * n.s. 10 5 0 -5 n a e m -10 no gap 25s gap second pulse Figure 6. Flies are attracted to the unpaired odor only in transitions. (A) Experimental strategy for measuring behavioral responses to odor transitions. Flies were trained by pairing one of the similar odors with optogenetic activation of DAN PPL1-γ2α’1. They were then tested with 30 s odor pulses presented either as direct transitions (left) or interrupted by a 25 s air period (right). Schematics illustrate A- A’ transitions but both sequences were tested, as indicated by the bars on top of panels B- E. (B) Upwind displacement during the first and second pulses of an A- A’ odor transition, as indicated by the green bars up top. This was computed as the increase in each fly’s distance from the arena center over the odor delivery period, then averaged across all flies in an arena (approximately 15 flies per arena). Traces in dark and light green are responses to A and A’ respectively. Plots are mean +/-SEM (n=12 arena runs for all stimulus types). (C) Upwind displacement for the reverse odor transition i.e. A’-A. (D) Upwind displacement for A- gap- A’ interrupted transition. (E) Upwind displacement for the reverse A’-gap- A interrupted transition. (F) Upwind displacement in response to the first odor pulse, averaged across flies in each arena experiment. Mean displacement was not significantly different between unpaired and paired odors for experiments with no gap (n=12, 12 experiments for paired and unpaired odors, p=0.70) and a 25 s gap (n=12, Figure 6 continued on next page Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 14 of 26 Neuroscience Research article Figure 6 continued 13, p=0.40). (G) As in F except for responses to the second odor pulse. Displacement was significantly different for transitions with no gap between pulses (n=12, 12 experiments for paired and unpaired odors, p=0.004) but not different when transitions were interrupted by a 25 s gap (n=12, 13, p=0.85). Statistical comparisons in F and G were made with the independent- sample Wilcoxon rank sum test with a Bonferroni- Holm correction for multiple comparisons. The online version of this article includes the following figure supplement(s) for figure 6: Figure supplement 1. Transition dependent attraction is not a result of linearly summed, single- pulse responses. Figure supplement 2. Flies reinforced via DAN PPL1-α3 do not respond to transitions between A and A’. condition, ruling out the possibility that flies go more upwind with the A- A’ transition because they start from further downwind in the arena (Figure 6—figure supplement 1B, n=12 for A- A’, n=12 for A’-A, p=0.08 for direct; n=12 for A- gap- A’, n=13 for A’-gap- A, p=0.39 for interrupted). Additionally, we ruled out the possibility that the increased upwind displacement during such a transition comes from a linear combination of the response to the end of the first odor pulse and the beginning of the second (Figure 6—figure supplement 1C- G). These results show that behavioral responses to A’ are distinct only when it immediately follows the paired odor A, matching the odor transition responses we observed in MBON-γ2α’1. The upwind displacement during the A- A’ transition is consistent with our observation that MBON-γ2α’1, a positive valence MBON that drives upwind behavior, is highly active during these transitions. In fact, the mean upwind displacement after an A- A’ transition was similar to that caused by optogenetic activation of MBON-γ2α’1 in the arena (unpublished communication - Y. Aso). Overall, these results show that flies compare available alternatives ‘side- by- side’ in time and that stimulus history is important for flexible categorization and behavior. When the two odors are encountered separately, both MBON-γ2α’1 output and fly behavior are indistinguishable for the two similar odors. However, when they are closely apposed in time, MBON-γ2α’1 activity is enhanced and flies are attracted to A’. Discussion Using a pair of perceptually similar odors (A and A’) and one distinct odor (B), we identified a site in the MB circuit that switches the neuronal and behavioral categorization of A’ depending on whether flies are presented an A’ vs A or an A’ vs B choice. Learning- related synaptic plasticity resulted in depressed neuronal responses to both similar odors when presented in isolation, consistent with strong behav- ioral generalization. However, when the odors were presented sequentially, as at an odor boundary, neuronal responses to A and A’ were distinct. Moreover, behavioral experiments with carefully timed odor delivery showed that flies’ response to A’ was distinct from A only if they were delivered in a transition. These results demonstrate how presenting cues as a choice can influence behavioral responses. An odor boundary presents an opportunity to compare stimuli, and this comparison modu- lates memory traces by amplifying small differences between stimuli to change categorization and behavior. Memory specificity is determined by more than overlap of KC somatic activity patterns KC representations are conventionally thought to be the key player in determining whether to discriminate or generalize. KC activity patterns in response to distinct odors have little overlap (Perez- Orive et  al., 2002; Murthy et  al., 2008; Turner et  al., 2008; Honegger et  al., 2011; Campbell et  al., 2013), and this allows synaptic changes to be highly stimulus specific (Hige et  al., 2015a). However, activity patterns are not so sparse that pattern separation is complete (Campbell et  al., 2013; Dasgupta et al., 2017; Endo et al., 2020; Hige et al., 2015b). Overlap between different odor response patterns exists, and correlates with both the strength of generalization and the specificity of plastic changes in MBONs. Importantly, prior work did not examine whether the extent of overlap is distributed differentially across different KC subtypes. This could serve as the basis for differences in discrimination we observe between compartments, with optogenetic training in γ2α’1 but not α3 Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 15 of 26 Neuroscience Research article capable of supporting hard discrimination. However, we found that all three major subtypes of KCs exhibit similar levels of overlap across our odor test set. Furthermore, we found that odor representa- tions in all three KC subtypes contain enough information for A- A’ discrimination. Despite this, when we examined KC responses to odor transitions, we found that they could not account for the odor- transition responses we observed in MBON-γ2α’1. Another factor that could potentially contribute to the different specificity in these compartments are the numbers of KC inputs. Theoretical work suggests that – holding the absolute number of responding KCs constant – the larger the total population of KCs, the less overlap there will be between different odor response patterns (Marr, 1969; Albus, 1971; Babadi and Sompolinsky, 2014; Cayco- Gajic and Silver, 2019). However, contrary to theoretical expectations, of the two MBON types we studied MBON-α3 received more synaptic inputs from a greater number of KCs, however it was poorer at hard discrimination (MBON γ2α’1=2,959 synapses from 336 α’/β’ KCs and 3773 synapses from 683 γ KCs per hemisphere; MBON α3=11,360 synapses from 888 α/β KCs per hemisphere; Clements et al., 2020; Li et al., 2020), again indicating that compartments’ different capabilities for hard discrimination were not a result of differences in sensory representations. We note that our KC activity measurements are all from cell bodies, while KC output synapses are the site of plasticity (Bilz et al., 2020), and we cannot formally exclude that activity at KC synapses may differ based on type- specific integrative properties (DasGupta et al., 2014; Groschner et al., 2018; Vrontou et al., 2021), or axo- axonic connections between KCs (Bielopolski et al., 2019; Manoim et al., 2022). Adapting memory usage through stimulus comparisons Our work instead highlights the importance of comparisons in determining how a stimulus is cate- gorized. In particular we observe that flies make a temporal comparison of inputs and identify the underlying neural implementation of comparison in the MB. Prior work has shown history- dependent effects at multiple layers of the olfactory circuit. For example, in the rodent olfactory bulb, responses to odor sequences are linear combinations of the responses to individual pulses (Gupta et al., 2015). On the other hand, in locusts, presenting odors singly or in transitions altered odor representations non- linearly in KCs (Broome et al., 2006). The extent of response alteration correlated with the accu- racy of behavioral recall (Saha et al., 2013). In Drosophila, odor transitions can cause similar changes in PN representations that result in altered innate odor preference (Badel et al., 2016). In another locust study (Nizampatnam et al., 2018), presenting an odor in a transition altered its representation to enhance contrast in the locust antennal lobe glomeruli. These observations suggest that changes in response to transitioning stimuli are mediated by a mechanism that takes place early on in the olfac- tory circuit. However, our observations of KC activity indicate that, although KCs response patterns to an odor presented as a single pulse versus in a transition are distinct, they are not sufficiently so to generate the sequence- specific transition effect we observe here. We suggest instead that flexible categorization in Drosophila involves a mechanism at or down- stream of the KC- MBON synapses modified during learning. Examples of such downstream modu- lation of memory have already been observed in the MB. In Drosophila, recalling food reward associations from one MB compartment is gated by another MB compartment depending on whether or not the fly is hungry (Perisse et  al., 2016). In this case, recall is regulated by the addition of a layer of contextual modulation through neuropeptide signaling that couples neural activity to satiety state. Ongoing motor activity can also affect dopaminergic inputs along the MB lobes (Cohn et al., 2015). These could modulate memory- traces on relatively short timescales based on the behavioral state of the animal. Switching between independently stored short and long- term memories provides another solution (Trannoy et al., 2011; Huetteroth et al., 2015; Yamagata et al., 2015). Experiments have shown that long- term memory allows for more generalization than short- term memory (Ichinose et  al., 2015; König et  al., 2017). However, all these mechanisms rely on an internal state signal (satiety or locomotion) rather than comparisons between external stimuli. Here we establish a few important constraints on a possible mechanism for flexible categorization in Drosophila: (i) it manifests at or downstream of the sites of learning, the KC >MBON synapses and (ii) it modifies responses to stimuli asymmetrically - in A to A’ transitions and not the reverse. One mechanism that could satisfy these criteria, would involve an explicit comparison of MBON activities in time, much like the delay- lines in the auditory pathways of owls and crickets (Schöneich et al., 2015; Sullivan and Konishi, 1986). This could be implemented via a downstream neuron that receives a Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 16 of 26 Neuroscience Research article real- time and a delayed copy of MBON-γ2α’1 activity, which then provides a positive feed- back signal to the MBON to amplify small increases in activity. Both of these motifs have been observed in the EM connectome (Li et al., 2020). As the extent of depression of KC >MBON synapses is inevitably slightly weaker for any odor that overlaps imperfectly with the learned odor, this mechanism would sensitize the circuit to small differences in MBON activity that arise around an odor transition. Another class of mechanisms centers on the observation that the KC population exhibits a distinct pattern of responses to odor offset (Tanaka et al., 2008; Lüdke et al., 2018). Offset responses in other MBONs can be potentiated (Vrontou et  al., 2021), presumably due to the timing of reinforcement (Cohn et al., 2015; Handler et al., 2019), suggesting a similar mechanism might operate in MBON-γ2α’1 to augment responses to the second odor in a transition. An additional candidate mechanism is plas- ticity of inhibitory input to KCs from the APL neuron. Activity of this inhibitory neuron is reduced by training (Zhou et al., 2019; Liu and Davis, 2009), so when a non- overlapping set of KCs is activated at an odor transition, that excitation may more effectively drive the downstream MBON. Inhibition at odor offset is a particularly prominent feature of the α’/β’ KCs that are input to MBON-γ2α’1 (Inada et al., 2017), so this effect could act in combination with potentiation of KC offset responses to create pronounced changes in KC output at an odor transition. Future work will be needed to resolve these different possibilities. Many animals use a stored memory to support different behaviors based on the choices available to them. We have shown that in Drosophila, this response flexibility relies on comparing cues side by side in time. Making fine- grained distinctions is easier when a temporal comparison is possible, but when it is not, more generalized categorizations can be an adaptive default. Methods Fly strains Drosophila melanogaster were raised on standard cornmeal food at 21 °C at 60% relative humidity on standard cornmeal food on a 12–12 hr light- dark cycle. For optogenetics behavior experiments, crosses were set on food supplemented with 0.2  mM all- trans- retinal and moved to 0.4  mM after eclosion and kept in the dark throughout. Transgene Expression target/reporter description Bloomington stock number, reference MB296B split Gal4 DAN PPL1-γ2α’1 MB630B split Gal4 d5HT1b- Gal4 c305a- Gal4 c739- Gal4 MB077B split Gal4 MB082C split Gal4 DAN PPL1-α3 γ KCs α’/β’ KCs α/β KCs MBONs γ2α’1 MBONs α3 R82C10- LexA DANs PPL1-γ2α’1, α2, α3 BDSC:68253 Aso and Rubin, 2016 BDSC:68290 Aso and Rubin, 2016 BDSC:27637 Yuan et al., 2006 BDSC:30829 Krashes et al., 2007 BDSC:7362 McGuire et al., 2001 BDSC:68283 Aso et al., 2014a BDSC:68286 Aso et al., 2014a BDSC:54981 Pfeiffer et al., 2013 20XUAS- CsChrimson- mVenus attp18 Optogenetic activation for behavior BDSC:55134 Klapoetke et al., 2014 13XLexAop2- IVS- Syn21- Chrimson88- tdT- 3.1- P10 Optogenetic activation for imaging BDSC: n.a. Strother et al., 2017 20XUAS- IVS- Syn21- opGCaMP6f- P10 Codon- optimized Ca2+ reporter BDSC: n.a. Chen et al., 2013 Expression patterns of split- GAL4 lines produced by Janelia FlyLight (Jenett et al., 2012) can be viewed online (http://splitgal4.janelia.org/cgi-bin/splitgal4.cgi). Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 17 of 26 Neuroscience Research article Behavior DAN driver split Gal4 crossed with 20XUAS- CsChrimson- mVenus attp18 TH- rescue experiment (genetic strategy as in Aso et al., 2019) knockout w, 20XUAS- CSChrimson- mVenus attP18; +; ple2, DTHFS ±BAC attP2, TH- ZpGAL4DBD VK00027 / TM6 B crossed with w; R73F07- p65ADZp attP40 /CyO; ple2, DTHFS ±BAC attP2 /TM6B knockout and rescue in DAN PPL1-γ2α’1 w, 20XUAS- CSChrimson- mVenus attP18; UAS- DTH1m; ple2, DTHFS ±BAC attP2, TH- ZpGAL4DBD VK00027 /TM6 B crossed with w; R73F07- p65ADZp attP40 /CyO; ple2, DTHFS ±BAC attP2 /TM6B KC imaging γ KCs: w; +/+; d5HT1b- Gal4/20XUAS- IVS- Syn21- opGCaMP6f- P10 VK00005 α’/β’ KCs: w; c305a- Gal4/+; 20XUAS- IVS- Syn21- opGCaMP6f- P10 VK00005/+ α/β KCs: w; c739- Gal4/+; 20XUAS- IVS- Syn21- opGCaMP6f- P10 VK00005/+ MBON γ2α’1 imaging 20XUAS- IVS- Syn21- opGCaMP6f- P10 Su(Hw)attP8 /w; R25D01- ZpGAL4DBD attP40 /82C10- LexAp65 attP40; R19F09- p65ADZp attP2 /13XLexAop2- IVS- Syn21- Chrimson88::tdT- 3.1- p10 in VK00005 R25D01 and R19F09 are components of the MB077B stable split- GAL4 driver (BDSC: 68283) MBON α3 imaging w; +/+; 20XUAS- IVS- Syn21- opGCaMP6f- P10 VK00005  /R23C06- ZpGAL4DBD in attP2, R40B08- p65ADZp VK00027 R23C06 and R40B08 are components of the MB082C stable split- GAL4 driver (BDSC: 68286) Behavior experiments Odor quadrant choice assay Groups of approximately 20 females, aged 4–10 d post- eclosion were anaesthetized on a cold plate and collected at least two day prior to experiments. After a day of recovery on 0.4 mM all- trans- retinal food, they were transferred to starvation vials containing nutrient- free agarose. Starved females were trained and tested at 25 °C at 50% relative humidity in a dark circular arena described in Aso and Rubin, 2016. The arena consisted of a circular chamber surrounded by four odor delivery ports that divide the chamber into quadrants. The input flow rate through each port was 100 mL/min, which was actively vented out a central exhaust at 400 mL/min. Odors were pentyl acetate, butyl acetate and ethyl lactate (Sigma- Aldrich product numbers 109584, 287725, and W244015 respectively). Except for the TH- rescue experiments shown in Figure 1I, these odors were diluted 1:10000 in paraffin oil (Sigma- Aldrich product number 18512). For the experiments in Figure 1I, we used a different odor delivery system which utilizes air dilution of saturated odorant vapor, and delivered odors at a 1:16 dilution of saturated vapor. Flies were aspirated into the arena via a small port, and allowed 60 s to acclimatize before training commenced. Training consisted of exposing the flies to one of the odors while providing optogenetic stimulation via a square array of red LEDs (617 nm peak emission, Red- Orange LUXEON Rebel LED, 122 lm at 700mA) which shone through an acrylic diffuser to illuminate flies from below. LED activation consisted of 30 pulses of 1 s duration with a 1 s inter- flash interval, commencing 5 s after switching on the odor valves and terminating 5 s after valve shut- off. To optimize learning scores, we used different training regimes depending on the compartments receiving optogenetic reinforcement, according to Aso and Rubin, 2016. A single training session was used for MB296B, TH- mutant, TH- rescue, while 3 training sessions, separated by 60 s, were used for some MB296B experiments, as indicated in the text. For MB630B we used 10 training sessions separated by 15 min. Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 18 of 26 Neuroscience Research article Following training, testing was carried out with the appropriate odors for each task. In the test configuration, the two different odor choices are presented in opposing quadrants for 60 s. Videos of fly behavior were captured at 30 frames per second using MATLAB (Mathworks, USA) and BIAS (http://archive.iorodeo.com/content/basic-image-acquisition-software-bias.html) and analyzed using custom- written code in MATLAB. Odor attraction assay For the odor attraction assay, the outputs of odor machines were re- configured to inject the output of a single odor machine into all four quadrants. We switched output from one machine to the other to deliver rapid odor transitions in time. About 15 flies were introduced into the arena for each experi- ment. The rest of the behavioral procedures were identical to those used in the quadrant choice assay. Optogenetic MBON-activation assay For this assay, a clean air stream was delivered into all four arena quadrants throughout the experi- ment. Flies expressed CSChrimson in MBON γ2α’1. Flies received six 10 s long LED flashes, separated by 60 s of darkness. The rest of the behavioral procedures were identical to those used in the quadrant choice assay. Calcium imaging Flies were imaged on a resonant- scanning, Janelia, jET MIMMS2.0 custom- designed two- photon microscope, with a Chameleon Ultra II, Titanium- sapphire laser (Coherent, USA) tuned to emit 920 nm. Images were acquired using a 20 x, NA 1.0, water- immersion objective lens XLUMPLFLN (Olympus, Japan) and a GaAsP PMT H11706P- 40 SEL (Hamamatsu, Japan). Power after the objective ranged from 4 to 5  mW for MBON imaging and 4–7  mW for KC imaging, depending on the preparation. Microscope control and data acquisition ran on the Scanimage platform (Vidrio, USA). Frames were acquired at 30 Hz, but three frames at a time were averaged during acquisition, for a final frame rate of 10 Hz. For KC imaging, pixels were sampled at 0.22 μm/pixel and for MBON imaging, at 0.18 μm/ pixel. For photostimulation, flies were fully illuminated from beneath with 617  nm light through a liquid light- guide (LLG- 03- 59- 340- 0800- 2, Mightex, USA) butt- coupled to an LED light source (GCS- 0617–04 A0510, Mightex, USA). Intensity at the fly was 1 mW/mm2. LED pulses were delivered at a frequency of 1 Hz, with a duty- cycle of 50%, for 5 s, starting 2 s after paired- odor onset. For optogenetics imaging experiments, crosses were set on food supplemented with 0.4  mM all- trans- retinal, and maintained on the same food at 25  ° C until flies were used for experiments. Flies were prepared as described previously (Campbell et al., 2013; Honegger et al., 2011). Three- to 8- day- old female flies were immobilized in a 0.25  mm thick stainless- steel sheet with a photo- chemically etched tear- drop shaped hole (PhotoFab, UK) and glued into place with two- component epoxy (Devcon, USA). For imaging in the KC somata and the MBON dendrites, head angle was adjusted differently to give best optical access to the target region, taking care to keep the antennae dry beneath the metal plate. For KC and MBON α3 imaging, the back of the head was submerged in Ringer’s bath solution consisting in mM: NaCl, 103; KCl, 3; CaCl2, 1.5; MgCl2, 4; NaHCO3, 26; N- tris(hydroxymethyl) methyl- 2- aminoethane- sulfonic acid, 5; NaH2PO4, 1; trehalose, 10; glucose, 10 (pH 7.3, 275 mOsm). For γ2α’1 MBON experiments, the flies were starved (24 hours in nutrient- free, distilled- water agarose vials). Previous studies have shown that hemolymph sugar is halved in flies starved for 24 hrs (Dus et al., 2011). So we used bath Ringer’s where glucose and trehalose were halved to 5 mM each and the non- metabolizable sugar arabinose (10 mM) was substituted to maintain osmolarity. For KC imaging, once the brain was exposed, bath solution was momentarily aspirated away and the preparation was covered in a drop of 5% (w/v) agarose (Cambrex Nusieve, catalog #50080) in Ringer’s, cooled to 36 ° C, which was then flattened with a 5 mm diameter circular coverslip that was then removed just prior to imaging. For KC imaging, 8 repeats of single odor pulses and each kind of odor transition were deliv- ered with an inter trial interval of 45 s. Stimulus types were randomly interleaved. For MBON γ2α’1 imaging, we delivered two repeats of either single odor pulses or transitions before and after odor- reinforcement pairing (Figure 4—figure supplement 4A), adapted from Berry et al., 2018. Only one repeat was imaged before and after pairing. For MBON α3 imaging, only the second presentation of each transition stimulus type was imaged (Figure 4—figure supplement 4B). Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 19 of 26 Neuroscience Research article Odor delivery for imaging experiments To deliver rapid odor transitions, we set up two separate odor delivery machines (Honegger et al., 2011) and joined their outputs upstream of the final tube delivering odor to the fly. These systems use saturated odor vapor which is then serially diluted in clean air to a final dilution of 0.8% (v/v). This was delivered to the fly at a flow rate of 400 mL/min from a tube with an inner diameter of 3 mm. We measured relative odor concentrations with a photoionization detector (200B miniPID, Aurora Scientific, Canada). Different chemical vapors at the same concentration generate different PID signal amplitudes. Thus, the PID signal is linearly related to concentration only for a given odor chemical. The PID probe was used to measure and tune odor pulse shapes and to measure and account for the time taken for an odor pulse to reach the fly. The short period of overlap between the fall of the first odor pulse and the rise of the second occurred for both kinds of transition stimuli, paired to unpaired transitions and unpaired to paired transitions (Figure 4—figure supplement 1C). Hence, this overlap would not affect our measures of discriminability between the two similar odors. A hot- wire anemom- eter S490 (Kurz, USA) was used to measure air- velocity at a sampling rate of 10 KHz while mock- odor pulses were being delivered though empty odor- vials. This was to rule out any mechanical transients at the time of odor- transitions being an external cue to the flies. To minimize transients, we combined the steps of the second serial dilution of the second odor pulse and mixing the outputs of the two odor machines. The second pulse in all transition stimuli was introduced into the final air stream at one- tenth the flow- rate. Any pressure- transients due to valve switches during the transition to the second pulse were too small to be measured by the anemometer in the final output. We saw a small valve- switching transient at the beginning of the first pulse in any transition (8% the size of the steady- state flow, Figure 4—figure supplement 1D). Since this transient was always at the onset of the first pulse, and not during transitions, again, it did not affect discriminability. Final air flow rate and odor concentration were adjusted to best match the odor flux a fly would experience in the arena. Since odor flows inward from the circumference of the arena, odor flux increases with distance from the periphery. We computed the odor flux on the circle that covers half the arena area and matched the flux delivered on the rig to it. Data analysis Behavior Videos recorded during the test phase were analyzed using custom- written MATLAB code. The centroid of each fly was identified and the number of centroids in each quadrant computed for every frame of the experiment. For discrimination experiments, a Performance Index (PI) was calculated as the number of flies in the quadrants containing the paired odor minus the number in the quadrants with the unpaired odor, divided by the total number of flies (Tully and Quinn, 1985). This value was calculated for every frame of the movie, and the values over the final 30 s of the test period averaged to compute a single PI. Discrimination experiments employed a reciprocal design where the identity of the paired and unpaired odors was swapped and a single data point represents the averaged PI from two reciprocally trained groups of flies. Generalization experiments could not employ a reciprocal design, so instead we compared scores against control experiments where flies were exposed to LED stimulation that was not paired with odor delivery; instead stimulation preceded odor by 2  min. In this case the PI score reported as a single data point is the PI observed from the generalization experiment minus the PI observed in the unpaired control, after both PIs were corrected for biases in initial quadrant occupancies by subtracting away the pre- odor baseline. Statistical testing was done as described in figure legends. We used the non- parametric, inde- pendent sample, Wilcoxon rank sum test to compare performance indices across treatment groups. Statistical testing was performed with custom code written in Matlab (Mathworks, USA). The appro- priate sample size was estimated based on the standard deviation of performance indices in previous studies using the same assay (Aso and Rubin, 2016). For the odor attraction and the MBON- activation assays, computing upwind displacement required us to track each fly’s trajectory in time. We used the Caltech Fly Tracker (Eyjolfsdottir et al., 2014) to automatically extract fly trajectories from videos. Odor stimulus onset time in the arena was determined from PID measurements of odor concentration at the arena exhaust port. For the Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 20 of 26 Neuroscience Research article MBON- activation assay, stimulus onset was set as the moment the LED turned on. Upwind displace- ment was computed as the increase in the distance from the center for each fly, relative to its location at stimulus onset, for each time- point over the entire stimulus window. The displacement for all flies in an arena experiment were then averaged before plotting and statistical testing. Calcium imaging For KC data, fluorescence time- series images were first analyzed with the Suite2P analysis pipeline (Pachitariu et al., 2016) running in Matlab to register data and identify active single- cell regions of interest (ROIs). For MBON imaging data, ROIs were manually drawn using a custom Matlab script. For both types of experiments, average, raw fluorescence intensity for each ROI was then extracted by a separate, custom script. A background region with no labeling in each imaging field was manually defined, and background fluorescence (this consisted of the PMT offset and autofluorescence) was subtracted from all measured fluorescence values for that field. ΔF/F was computed according to the following equation ΔF/Fi = (Fi - F0)/F0. where Fi is the fluorescence of a given cell ROI at a given time- point in a trial, and F0 is the same ROI’s fluorescence in an 8 s window during the baseline period on that trial, prior to odor delivery. For all plotted fluorescence traces, ΔF/F time- series data was boxcar filtered with a window- width of 0.2 s. All other analysis was done with un- filtered ΔF/F data. For making statistical comparisons, ΔF/F values during stimulus presentation were averaged over time windows as indicated in each figure. KC population activity decoders The objective of this analysis was to determine whether a linear classifier can discriminate trials of a particular odor based on the KC responses. We fitted logistic regression models to predict whether or not KC activity on a given trial was evoked by a particular odor. For example, an odorA classifier received KC population activity vectors as input and then made a prediction whether the input activity was evoked by odorA/not odorA. Separate classifiers were fitted for each fly, for each odor. We used leave- one- out cross validation (LOOCV): of the 8 repeats acquired for each odor, one was left out as a test trial and the remaining trials were used to fit the model. In this way, we systematically fitted models for each combination of training and test trail sets. All plotted accuracy scores are for model predictions on test trials not used for fitting. Model weights were initialized by sampling from a distribution of weights obtained from EM connectome synapse counts (Clements et  al., 2020). Synapse counts from a KC to an MBON were assumed to be linearly related to KC- MBON weight, and normalized to the maximum weight observed. Initial model weights were uniformly sampled from this biological distribution and then fitted without regularization. The cost function used to estimate goodness of fit was the binary cross- entropy with a quadratic regularization, defined as cost = 1 m − m ∑i=1 yi × log hi yi 1 − − × log 1 − hi + ( ) ( ) ( ) λ 2m n θ2 j ∑j=1 where m is the number of training trials, yi is the correct odor label for a given trial (0 or 1), hi is the model’s prediction (or probability that the input activity vector was in response to a given odor) for the same trial, λ is the regularization constant (we used λ = 1, but this was not a sensitive parameter), n is the number of neurons in a given dataset and θj are the weights of the neurons. For logistic regression models fitted without regularization (shown in Figure 5B), λ was set to 0. The model’s prediction, h was computed according to the equation h = 1 1 + e− θ X × ( ) here X is the m × n activity matrix for m training trials and θ is the n 1 vector of weights. × Code availability statement The custom Matlab code used for analysis in this manuscript is publicly available at https://github. com/mehrabmodi1/Drosophila_flexible_recall (Modi, 2023a; copy archived at Modi, 2023b). Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 21 of 26 Neuroscience Research article Acknowledgements This work was supported by the Howard Hughes Medical Institute and the National Institutes of Health (2R01DC010403- 06). We thank Florin Albeanu and the Albeanu group for hosting MM for part of the time that this work was carried out, and also for engaging in helpful discussions and feed- back. Robert Eifert at Cold Spring Harbor Laboratory and Steven Sawtelle, Igor Negrashov, Vasily Goncharov and others at jET, Janelia Research Campus provided vital technical support. Todd Laverty and others in the Janelia Drosophila resources team and Karen Hibbard from Janelia provided vital support with fly lines and media. Gudrun Ihrke and others in Project Technical Resources provided support with expression characterization. We also thank all members of the Turner and Aso groups and Eyal Gruntman, Vivek Jayaraman, Ann Hermundstad, Florin Albeanu and Priyanka Gupta for support, discussions and feedback. Additional information Funding Funder Howard Hughes Medical Institute National Institutes of Health Grant reference number Author Yoshinori Aso Glenn C Turner 2R01DC010403-06 Glenn C Turner The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Author contributions Mehrab N Modi, Conceptualization, Resources, Data curation, Software, Formal analysis, Investiga- tion, Visualization, Writing - original draft, Writing – review and editing; Adithya E Rajagopalan, Inves- tigation, Writing – review and editing; Hervé Rouault, Conceptualization, Formal analysis, Supervision, Investigation, Writing – review and editing; Yoshinori Aso, Conceptualization, Supervision, Funding acquisition, Investigation, Writing - original draft, Project administration, Writing – review and editing; Glenn C Turner, Conceptualization, Resources, Supervision, Funding acquisition, Writing - original draft, Writing – review and editing Author ORCIDs Yoshinori Aso Glenn C Turner http://orcid.org/0000-0002-2939-1688 http://orcid.org/0000-0002-5341-2784 Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.80923.sa1 Author response https://doi.org/10.7554/eLife.80923.sa2 Additional files Supplementary files • MDAR checklist Data availability All data have been uploaded to Dryad. Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 22 of 26 Neuroscience Research article The following dataset was generated: Author(s) Turner GC Year 2023 Dataset title Dataset URL Database and Identifier Flexible specificity of memory in Drosophila depends on a comparison between choices https:// doi. org/ 10. 5061/ dryad. 8931zcrtc Dryad Digital Repository, 10.5061/dryad.8931zcrtc References Albus JS. 1971. Communicated by Donald H. Perkel. Mathematical Biosciences 10:25–61. DOI: https://doi.org/ 10.1016/0025-5564(71)90051-4 Armstrong JD, Texada MJ, Munjaal R, Baker DA, Beckingham KM. 2006. Gravitaxis in Drosophila melanogaster: a forward genetic screen. Genes, Brain, and Behavior 5:222–239. DOI: https://doi.org/10.1111/j.1601-183X. 2005.00154.x, PMID: 16594976 Aso Y, Grübel K, Busch S, Friedrich AB, Siwanowicz I, Tanimoto H. 2009. The mushroom body of adult Drosophila characterized by Gal4 drivers. Journal of Neurogenetics 23:156–172. DOI: https://doi.org/10.1080/ 01677060802471718, PMID: 19140035 Aso Y, Hattori D, Yu Y, Johnston RM, Iyer NA, Ngo TTB, Dionne H, Abbott LF, Axel R, Tanimoto H, Rubin GM. 2014a. The neuronal architecture of the mushroom body provides a logic for associative learning. eLife 3:e04577. DOI: https://doi.org/10.7554/eLife.04577, PMID: 25535793 Aso Y, Sitaraman D, Ichinose T, Kaun KR, Vogt K, Belliart- Guérin G, Plaçais PY, Robie AA, Yamagata N, Schnaitmann C, Rowell WJ, Johnston RM, Ngo TTB, Chen N, Korff W, Nitabach MN, Heberlein U, Preat T, Branson KM, Tanimoto H, et al. 2014b. Mushroom body output neurons encode valence and guide memory- based action selection in Drosophila. eLife 3:e04580. DOI: https://doi.org/10.7554/eLife.04580, PMID: 25535794 Aso Y, Rubin GM. 2016. Dopaminergic neurons write and update memories with cell- type- specific rules. eLife 5:e16135. DOI: https://doi.org/10.7554/eLife.16135, PMID: 27441388 Aso Y, Ray RP, Long X, Bushey D, Cichewicz K, Ngo TT, Sharp B, Christoforou C, Hu A, Lemire AL, Tillberg P, Hirsh J, Litwin- Kumar A, Rubin GM. 2019. Nitric oxide acts as a cotransmitter in a subset of dopaminergic neurons to diversify memory dynamics. eLife 8:e49257. DOI: https://doi.org/10.7554/eLife.49257, PMID: 31724947 Babadi B, Sompolinsky H. 2014. Sparseness and expansion in sensory representations. Neuron 83:1213–1226. DOI: https://doi.org/10.1016/j.neuron.2014.07.035, PMID: 25155954 Badel L, Ohta K, Tsuchimoto Y, Kazama H. 2016. Decoding of context- dependent olfactory behavior in Drosophila. Neuron 91:155–167. DOI: https://doi.org/10.1016/j.neuron.2016.05.022, PMID: 27321924 Berry JA, Phan A, Davis RL. 2018. Dopamine neurons mediate learning and forgetting through bidirectional modulation of a memory trace. Cell Reports 25:651–662.. DOI: https://doi.org/10.1016/j.celrep.2018.09.051, PMID: 30332645 Bielopolski N, Amin H, Apostolopoulou AA, Rozenfeld E, Lerner H, Huetteroth W, Lin AC, Parnas M. 2019. Inhibitory muscarinic acetylcholine receptors enhance aversive olfactory learning in adult Drosophila. eLife 8:e48264. DOI: https://doi.org/10.7554/eLife.48264, PMID: 31215865 Bilz F, Geurten BRH, Hancock CE, Widmann A, Fiala A. 2020. Visualization of a distributed synaptic memory code in the Drosophila brain. Neuron 106:963–976. DOI: https://doi.org/10.1016/j.neuron.2020.03.010, PMID: 32268119 Broome BM, Jayaraman V, Laurent G. 2006. Encoding and decoding of overlapping odor sequences. Neuron 51:467–482. DOI: https://doi.org/10.1016/j.neuron.2006.07.018, PMID: 16908412 Campbell RAA, Honegger KS, Qin H, Li W, Demir E, Turner GC. 2013. Imaging a population code for odor identity in the Drosophila mushroom body. The Journal of Neuroscience 33:10568–10581. DOI: https://doi. org/10.1523/JNEUROSCI.0682-12.2013, PMID: 23785169 Carello CD, Krauzlis RJ. 2004. Manipulating intent: evidence for a causal role of the superior colliculus in target selection. Neuron 43:575–583. DOI: https://doi.org/10.1016/j.neuron.2004.07.026, PMID: 15312655 Cayco- Gajic NA, Silver RA. 2019. Re- evaluating circuit mechanisms underlying pattern separation. Neuron 101:584–602. DOI: https://doi.org/10.1016/j.neuron.2019.01.044, PMID: 30790539 Chen T- W, Wardill TJ, Sun Y, Pulver SR, Renninger SL, Baohan A, Schreiter ER, Kerr RA, Orger MB, Jayaraman V, Looger LL, Svoboda K, Kim DS. 2013. Ultrasensitive fluorescent proteins for imaging neuronal activity. Nature 499:295–300. DOI: https://doi.org/10.1038/nature12354 Chen Y, Gerber B. 2014. Generalization and discrimination tasks yield concordant measures of perceived distance between odours and their binary mixtures in larval Drosophila. The Journal of Experimental Biology 217:2071–2077. DOI: https://doi.org/10.1242/jeb.100966, PMID: 24920835 Cichewicz K, Garren EJ, Adiele C, Aso Y, Wang Z, Wu M, Birman S, Rubin GM, Hirsh J. 2017. A new brain dopamine- deficient Drosophila and its pharmacological and genetic rescue. Genes, Brain, and Behavior 16:394–403. DOI: https://doi.org/10.1111/gbb.12353, PMID: 27762066 Claridge- Chang A, Roorda RD, Vrontou E, Sjulson L, Li H, Hirsh J, Miesenböck G. 2009. Writing memories with light- addressable reinforcement circuitry. Cell 139:405–415. DOI: https://doi.org/10.1016/j.cell.2009.08.034 Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 23 of 26 Neuroscience Research article Clements J, Dolafi T, Umayam L, Neubarth NL, Berg S, Scheffer LK, Plaza SM. 2020. Neu Print: Analysis tools for EM connectomics. bioRxiv. DOI: https://doi.org/10.1101/2020.01.16.909465 Cohn R, Morantte I, Ruta V. 2015. Coordinated and compartmentalized neuromodulation shapes sensory processing in Drosophila. Cell 163:1742–1755. DOI: https://doi.org/10.1016/j.cell.2015.11.019, PMID: 26687359 DasGupta S, Ferreira CH, Miesenböck G. 2014. FoxP influences the speed and accuracy of a perceptual decision in Drosophila. Science 344:901–904. DOI: https://doi.org/10.1126/science.1252114, PMID: 24855268 Dasgupta S, Stevens CF, Navlakha S. 2017. A neural algorithm for a fundamental computing problem. Science 358:793–796. DOI: https://doi.org/10.1126/science.aam9868, PMID: 29123069 Dus M, Min S, Keene AC, Lee GY, Suh GSB. 2011. Taste- independent detection of the caloric content of sugar in Drosophila. PNAS 108:11644–11649. DOI: https://doi.org/10.1073/pnas.1017096108, PMID: 21709242 Endo K, Tsuchimoto Y, Kazama H. 2020. Synthesis of conserved odor object representations in a random, divergent- convergent network. Neuron 108:367–381.. DOI: https://doi.org/10.1016/j.neuron.2020.07.029, PMID: 32814018 Eyjolfsdottir E, Fleet D, Pajdla T, Schiele B, Tuytelaars T. 2014. Detecting Social Actions of Fruit Flies. Computer Vision – ECCV 2014. ECCV 2014. Lecture Notes in Computer Science. 834–849. DOI: https://doi.org/10.1007/ 978-3-319-10605-2 Glimcher PW, Fehr E. 2013. Neuroeconomics: Decision Making and the Brain Academic Press. GroschnerLN, ChanLH, Bogacz B, DasGupta S. 2018. Dendritic integration of sensory evidence in perceptual decision- making. Cell 173:894–905. DOI: https://doi.org/10.1016/j.cell.2018.03.075 Gupta P, Albeanu DF, Bhalla US. 2015. Olfactory bulb coding of odors, mixtures and sniffs is a linear sum of odor time profiles. Nature Neuroscience 18:272–281. DOI: https://doi.org/10.1038/nn.3913, PMID: 25581362 Handler A, Graham TGW, Cohn R, Morantte I, Siliciano AF, Zeng J, Li Y, Ruta V. 2019. Distinct dopamine receptor pathways underlie the temporal sensitivity of associative learning. Cell 178:60–75. DOI: https://doi.org/10. 1016/j.cell.2019.05.040, PMID: 31230716 Hare TA, Schultz W, Camerer CF, O’Doherty JP, Rangel A. 2011. Transformation of stimulus value signals into motor commands during simple choice. PNAS 108:18120–18125. DOI: https://doi.org/10.1073/pnas. 1109322108, PMID: 22006321 Hayden BY. 2018. Economic choice: the foraging perspective. Current Opinion in Behavioral Sciences 24:1–6. DOI: https://doi.org/10.1016/j.cobeha.2017.12.002 Hige T, Aso Y, Modi MN, Rubin GM, Turner GC. 2015a. Heterosynaptic plasticity underlies aversive olfactory learning in Drosophila. Neuron 88:985–998. DOI: https://doi.org/10.1016/j.neuron.2015.11.003, PMID: 26637800 Hige T, Aso Y, Rubin GM, Turner GC. 2015b. Plasticity- driven Individualization of olfactory coding in mushroom body output neurons. Nature 526:258–262. DOI: https://doi.org/10.1038/nature15396, PMID: 26416731 Honegger KS, Campbell RAA, Turner GC. 2011. Cellular- resolution population imaging reveals robust sparse coding in the Drosophila mushroom body. The Journal of Neuroscience 31:11772–11785. DOI: https://doi.org/ 10.1523/JNEUROSCI.1099-11.2011, PMID: 21849538 Huetteroth W, Perisse E, Lin S, Klappenbach M, Burke C, Waddell S. 2015. Sweet taste and nutrient value subdivide rewarding dopaminergic neurons in Drosophila. Current Biology 25:751–758. DOI: https://doi.org/ 10.1016/j.cub.2015.01.036, PMID: 25728694 Hunt LT, Kolling N, Soltani A, Woolrich MW, Rushworth MFS, Behrens TEJ. 2012. Mechanisms underlying cortical activity during value- guided choice. Nature Neuroscience 15:470–476, . DOI: https://doi.org/10.1038/nn.3017, PMID: 22231429 Ichinose T, Aso Y, Yamagata N, Abe A, Rubin GM, Tanimoto H. 2015. Reward signal in a recurrent circuit drives appetitive long- term memory formation. eLife 4:e10719. DOI: https://doi.org/10.7554/eLife.10719, PMID: 26573957 Inada K, Tsuchimoto Y, Kazama H. 2017. Origins of cell- type- specific olfactory processing in the Drosophila mushroom body circuit. Neuron 95:357–367. DOI: https://doi.org/10.1016/j.neuron.2017.06.039, PMID: 28728024 Ito K, Suzuki K, Estes P, Ramaswami M, Yamamoto D, Strausfeld NJ. 1998. The organization of extrinsic neurons and their implications in the functional roles of the mushroom bodies in Drosophila melanogaster Meigen. Learning & Memory 5:52–77 PMID: 10454372. Itti L, Koch C. 2001. Computational modelling of visual attention. Nature Reviews. Neuroscience 2:194–203. DOI: https://doi.org/10.1038/35058500, PMID: 11256080 Jenett A, Rubin GM, Ngo TTB, Shepherd D, Murphy C, Dionne H, Pfeiffer BD, Cavallaro A, Hall D, Jeter J, Iyer N, Fetter D, Hausenfluck JH, Peng H, Trautman ET, Svirskas RR, Myers EW, Iwinski ZR, Aso Y, DePasquale GM, et al. 2012. A Gal4- driver line resource for Drosophila Neurobiology. Cell Reports 2:991– 1001. DOI: https://doi.org/10.1016/j.celrep.2012.09.011, PMID: 23063364 Kim YC, Lee HG, Han KA. 2007. D 1 dopamine receptor Dda1 is required in the mushroom body neurons for aversive and appetitive learning in Drosophila. The Journal of Neuroscience 27:7640–7647. DOI: https://doi. org/10.1523/JNEUROSCI.1167-07.2007, PMID: 17634358 Klapoetke NC, Murata Y, Kim SS, Pulver SR, Birdsey- Benson A, Cho YK, Morimoto TK, Chuong AS, Carpenter EJ, Tian Z, Wang J, Xie Y, Yan Z, Zhang Y, Chow BY, Surek B, Melkonian M, Jayaraman V, Constantine- Paton M, Wong GK- S, et al. 2014. Independent optical excitation of distinct neural populations. Nature Methods 11:338–346. DOI: https://doi.org/10.1038/nmeth.2836 Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 24 of 26 Neuroscience Research article König C, Antwi- Adjei E, Ganesan M, Kilonzo K, Viswanathan V, Durairaja A, Voigt A, Yarali A. 2017. Aversive olfactory associative memory loses odor specificity over time. The Journal of Experimental Biology 220:1548– 1553. DOI: https://doi.org/10.1242/jeb.155317, PMID: 28468811 Krashes MJ, Keene AC, Leung B, Armstrong JD, Waddell S. 2007. Sequential use of mushroom body neuron subsets during Drosophila odor memory processing. Neuron 53:103–115. DOI: https://doi.org/10.1016/j. neuron.2006.11.021, PMID: 17196534 Kudryavitskaya E, Marom E, Shani- Narkiss H, Pash D, Mizrahi A. 2021. Flexible categorization in the mouse olfactory bulb. Current Biology 31:1616–1631.. DOI: https://doi.org/10.1016/j.cub.2021.01.063, PMID: 33571434 Li F, Lindsey JW, Marin EC, Otto N, Dreher M, Dempsey G, Stark I, Bates AS, Pleijzier MW, Schlegel P, Nern A, Takemura SY, Eckstein N, Yang T, Francis A, Braun A, Parekh R, Costa M, Scheffer LK, Aso Y, et al. 2020. The connectome of the adult Drosophila mushroom body provides insights into function. eLife 9:e62576. DOI: https://doi.org/10.7554/eLife.62576, PMID: 33315010 Lin HH, Lai JSY, Chin AL, Chen YC, Chiang AS. 2007. A map of olfactory representation in the Drosophila mushroom body. Cell 128:1205–1217. DOI: https://doi.org/10.1016/j.cell.2007.03.006, PMID: 17382887 Lin AC, Bygrave AM, de Calignon A, Lee T, Miesenböck G. 2014. Sparse, decorrelated odor coding in the mushroom body enhances learned odor discrimination. Nature Neuroscience 17:559–568. DOI: https://doi. org/10.1038/nn.3660, PMID: 24561998 Liu X, Davis RL. 2009. The GABAergic anterior paired lateral neuron suppresses and is suppressed by olfactory learning. Nature Neuroscience 12:53–59. DOI: https://doi.org/10.1038/nn.2235, PMID: 19043409 Lüdke A, Raiser G, Nehrkorn J, Herz AVM, Galizia CG, Szyszka P. 2018. Calcium in Kenyon cell somata as a substrate for an olfactory sensory memory in Drosophila. Frontiers in Cellular Neuroscience 12:128. DOI: https://doi.org/10.3389/fncel.2018.00128, PMID: 29867361 Mackintosh NJ. 1974. The Psychology of Animal Learning Academic Press. Manoim JE, Davidson AM, Weiss S, Hige T, Parnas M. 2022. Lateral axonal modulation is required for stimulus- specific olfactory conditioning in Drosophila. Current Biology 32:4438–4450.. DOI: https://doi.org/10.1016/j. cub.2022.09.007, PMID: 36130601 Marr D. 1969. A theory of cerebellar cortex. The Journal of Physiology 202:437–470. DOI: https://doi.org/10. 1113/jphysiol.1969.sp008820, PMID: 5784296 McGuire SE, Le PT, Davis RL. 2001. The role of Drosophila mushroom body signaling in olfactory memory. Science 293:1330–1333. DOI: https://doi.org/10.1126/science.1062622, PMID: 11397912 Modi MN. 2023a. Drosophila_Flexible_Recall. 82c3e32. Github. https://github.com/mehrabmodi1/Drosophila_ flexible_recall Modi MN. 2023b. Drosophila_Flexible_Recall. swh:1:rev:82c3e3247205cb08ccbc4389cde9df13c46be457. Software Heritage. https://archive.softwareheritage.org/swh:1:dir:99cf8811b1e48564ff2d5aa974c7a6d5 705d17fb;origin=https://github.com/mehrabmodi1/Drosophila_flexible_recall;visit=swh:1:snp:f70253ea56f7 3419c1b6f11e5932457d93c7ea53;anchor=swh:1:rev:82c3e3247205cb08ccbc4389cde9df13c46be457 Murthy M, Fiete I, Laurent G. 2008. Testing odor response stereotypy in the Drosophila mushroom body. Neuron 59:1009–1023. DOI: https://doi.org/10.1016/j.neuron.2008.07.040, PMID: 18817738 Mysore SP, Asadollahi A, Knudsen EI. 2011. Signaling of the strongest stimulus in the owl optic Tectum. The Journal of Neuroscience 31:5186–5196. DOI: https://doi.org/10.1523/JNEUROSCI.4592-10.2011, PMID: 21471353 Mysore SP, Knudsen EI. 2011. Flexible Categorization of relative stimulus strength by the optic Tectum. The Journal of Neuroscience 31:7745–7752. DOI: https://doi.org/10.1523/JNEUROSCI.5425-10.2011, PMID: 21613487 Neckameyer WS, White K. 1993. Drosophila tyrosine hydroxylase is encoded by the pale locus. Journal of Neurogenetics 8:189–199. DOI: https://doi.org/10.3109/01677069309083448, PMID: 8100577 Nizampatnam S, Saha D, Chandak R, Raman B. 2018. Dynamic contrast enhancement and flexible odor codes. Nature Communications 9:3062. DOI: https://doi.org/10.1038/s41467-018-05533-6, PMID: 30076307 Owald D, Felsenberg J, Talbot CB, Das G, Perisse E, Huetteroth W, Waddell S. 2015. Activity of defined mushroom body output neurons underlies learned olfactory behavior in Drosophila. Neuron 86:417–427. DOI: https://doi.org/10.1016/j.neuron.2015.03.025, PMID: 25864636 Pachitariu M, Stringer C, Dipoppa M, Schröder S, Rossi LF, Dalgleish H, Carandini M, Harris KD. 2016. Suite2p: Beyond 10,000 Neurons with standard two- photon microscopy. bioRxiv. DOI: https://doi.org/10.1101/061507 Padoa- Schioppa C, Conen KE. 2017. Orbitofrontal cortex: A neural circuit for economic decisions. Neuron 96:736–754. DOI: https://doi.org/10.1016/j.neuron.2017.09.031, PMID: 29144973 Perez- Orive J, Mazor O, Turner GC, Cassenaer S, Wilson RI, Laurent G. 2002. Oscillations and sparsening of odor representations in the mushroom body. Science 297:359–365. DOI: https://doi.org/10.1126/science. 1070502, PMID: 12130775 Perisse E, Owald D, Barnstedt O, Talbot CB, Huetteroth W, Waddell S. 2016. Aversive learning and appetitive motivation toggle feed- forward inhibition in the Drosophila mushroom body. Neuron 90:1086–1099. DOI: https://doi.org/10.1016/j.neuron.2016.04.034, PMID: 27210550 Pfeiffer BD, Ngo TB, Hibbard KL, Murphy C, Jenett A, Truman JW. 2013. LexA Driver Collection of Rubin Laboratory at Janelia Farm. Personal Communication to FlyBase. Plaçais PY, Trannoy S, Friedrich AB, Tanimoto H, Preat T. 2013. Two pairs of mushroom body efferent neurons are required for appetitive long- term memory retrieval in Drosophila. Cell Reports 5:769–780. DOI: https://doi. org/10.1016/j.celrep.2013.09.032, PMID: 24209748 Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 25 of 26 Neuroscience Research article Qin H, Cressy M, Li W, Coravos JS, Izzi SA, Dubnau J. 2012. Gamma neurons mediate dopaminergic input during aversive olfactory memory formation in Drosophila. Current Biology 22:608–614. DOI: https://doi.org/10.1016/ j.cub.2012.02.014, PMID: 22425153 Riemensperger T, Isabel G, Coulom H, Neuser K, Seugnet L, Kume K, Iché-Torres M, Cassar M, Strauss R, Preat T, Hirsh J, Birman S. 2011. Behavioral consequences of dopamine deficiency in the Drosophila central nervous system. PNAS 108:834–839. DOI: https://doi.org/10.1073/pnas.1010930108, PMID: 21187381 Saha D, Leong K, Li C, Peterson S, Siegel G, Raman B. 2013. A spatiotemporal coding mechanism for background- invariant odor recognition. Nature Neuroscience 16:1830–1839. DOI: https://doi.org/10.1038/nn. 3570, PMID: 24185426 Schöneich S, Kostarakos K, Hedwig B. 2015. An auditory feature detection circuit for sound pattern recognition. Science Advances 1:e1500325. DOI: https://doi.org/10.1126/sciadv.1500325, PMID: 26601259 Schroll C, Riemensperger T, Bucher D, Ehmer J, Völler T, Erbguth K, Gerber B, Hendel T, Nagel G, Buchner E, Fiala A. 2006. Light- induced activation of distinct modulatory neurons triggers appetitive or aversive learning in Drosophila larvae. Current Biology 16:1741–1747. DOI: https://doi.org/10.1016/j.cub.2006.07.023, PMID: 16950113 Seger CA. 2008. How do the basal ganglia contribute to categorization? Their roles in generalization, response selection, and learning via feedback. Neuroscience and Biobehavioral Reviews 32:265–278. DOI: https://doi. org/10.1016/j.neubiorev.2007.07.010, PMID: 17919725 Seger CA, Miller EK. 2010. Category learning in the brain. Annual Review of Neuroscience 33:203–219. DOI: https://doi.org/10.1146/annurev.neuro.051508.135546, PMID: 20572771 Séjourné J, Plaçais P- Y, Aso Y, Siwanowicz I, Trannoy S, Thoma V, Tedjakumala SR, Rubin GM, Tchénio P, Ito K, Isabel G, Tanimoto H, Preat T. 2011. Mushroom body efferent neurons responsible for aversive olfactory memory retrieval in Drosophila. Nature Neuroscience 14:903–910. DOI: https://doi.org/10.1038/nn.2846 Stahl A, Noyes NC, Boto T, Botero V, Broyles CN, Jing M, Zeng J, King LB, Li Y, Davis RL, Tomchik SM. 2022. Associative learning drives longitudinally graded presynaptic plasticity of neurotransmitter release along axonal compartments. eLife 11:e76712. DOI: https://doi.org/10.7554/eLife.76712, PMID: 35285796 Strausfeld NJ, Sinakevitch I, Vilinsky I. 2003. The mushroom bodies of Drosophila melanogaster: an Immunocytological and golgi study of Kenyon cell organization in the calyces and lobes. Microscopy Research and Technique 62:151–169. DOI: https://doi.org/10.1002/jemt.10368, PMID: 12966500 Strother JA, Wu ST, Wong AM, Nern A, Rogers EM, Le JQ, Rubin GM, Reiser MB. 2017. The emergence of directional selectivity in the visual motion pathway of Drosophila. Neuron 94:168–182.. DOI: https://doi.org/10. 1016/j.neuron.2017.03.010, PMID: 28384470 Sullivan WE, Konishi M. 1986. Neural map of Interaural phase difference in the owl’s brainstem. PNAS 83:8400– 8404. DOI: https://doi.org/10.1073/pnas.83.21.8400, PMID: 3022292 Takemura SY, Aso Y, Hige T, Wong A, Lu Z, Xu CS, Rivlin PK, Hess H, Zhao T, Parag T, Berg S, Huang G, Katz W, Olbris DJ, Plaza S, Umayam L, Aniceto R, Chang LA, Lauchie S, Ogundeyi O, et al. 2017. A connectome of a learning and memory center in the adult Drosophila brain. eLife 6:e26975. DOI: https://doi.org/10.7554/eLife. 26975, PMID: 28718765 Tanaka NK, Tanimoto H, Ito K. 2008. Neuronal assemblies of the Drosophila mushroom body. The Journal of Comparative Neurology 508:711–755. DOI: https://doi.org/10.1002/cne.21692, PMID: 18395827 Trannoy S, Redt- Clouet C, Dura JM, Preat T. 2011. Parallel processing of appetitive short- and long- term memories in Drosophila. Current Biology 21:1647–1653. DOI: https://doi.org/10.1016/j.cub.2011.08.032, PMID: 21962716 Tully T, Quinn WG. 1985. Classical conditioning and retention in normal and mutant Drosophila melanogaster. Journal of Comparative Physiology. A, Sensory, Neural, and Behavioral Physiology 157:263–277. DOI: https:// doi.org/10.1007/BF01350033, PMID: 3939242 Turner GC, Bazhenov M, Laurent G. 2008. Olfactory representations by Drosophila mushroom body neurons. Journal of Neurophysiology 99:734–746. DOI: https://doi.org/10.1152/jn.01283.2007, PMID: 18094099 Villar ME, Pavão- Delgado M, Amigo M, Jacob PF, Merabet N, Pinot A, Perry SA, Waddell S, Perisse E. 2022. Differential coding of absolute and relative aversive value in the Drosophila brain. Current Biology 32:4576– 4592.. DOI: https://doi.org/10.1016/j.cub.2022.08.058, PMID: 36103878 Vrontou E, Groschner LN, Szydlowski S, Brain R, Krebbers A, Miesenböck G. 2021. Response competition between neurons and antineurons in the mushroom body. Current Biology 31:4911–4922.. DOI: https://doi. org/10.1016/j.cub.2021.09.008 Xu W, Südhof TC. 2013. A neural circuit for memory specificity and generalization. Science 339:1290–1295. DOI: https://doi.org/10.1126/science.1229534, PMID: 23493706 Yamagata N, Ichinose T, Aso Y, Plaçais PY, Friedrich AB, Sima RJ, Preat T, Rubin GM, Tanimoto H. 2015. Distinct dopamine neurons mediate reward signals for short- and long- term memories. PNAS 112:578–583. DOI: https://doi.org/10.1073/pnas.1421930112 Yuan Q, Joiner WJ, Sehgal A. 2006. A sleep- promoting role for the Drosophila serotonin receptor 1A. Current Biology 16:1051–1062. DOI: https://doi.org/10.1016/j.cub.2006.04.032, PMID: 16753559 Zhou M, Chen N, Tian J, Zeng J, Zhang Y, Zhang X, Guo J, Sun J, Li Y, Guo A, Li Y. 2019. Suppression of GABAergic neurons through D2- like receptor secures efficient conditioning in Drosophila aversive olfactory learning. PNAS 116:5118–5125. DOI: https://doi.org/10.1073/pnas.1812342116, PMID: 30796183 Modi et al. eLife 2023;12:e80923. DOI: https://doi.org/10.7554/eLife.80923 26 of 26 Neuroscience
10.7554_elife.83529
ReSeaRCH aRTICLe Context- dependent requirement of G protein coupling for Latrophilin- 2 in target selection of hippocampal axons Daniel T Pederick1†, Nicole A Perry- Hauser2,3†, Huyan Meng4, Zhigang He4, Jonathan A Javitch2,3*, Liqun Luo1* 1Department of Biology, Howard Hughes Medical Institute, Stanford University, Stanford, United States; 2Departments of Psychiatry and Molecular Pharmacology and Therapeutics, Columbia University Vagelos College of Physicians and Surgeons, New York, United States; 3Division of Molecular Therapeutics, New York State Psychiatric Institute, New York, United States; 4F.M. Kirby Neurobiology Center, Department of Neurology, Boston Children’s Hospital, Harvard Medical School, Boston, United States Abstract The formation of neural circuits requires extensive interactions of cell- surface proteins to guide axons to their correct target neurons. Trans- cellular interactions of the adhesion G protein- coupled receptor latrophilin- 2 (Lphn2) with its partner teneurin- 3 instruct the precise assembly of hippocampal networks by reciprocal repulsion. Lphn2 acts as a repulsive receptor in distal CA1 neurons to direct their axons to the proximal subiculum, and as a repulsive ligand in the prox- imal subiculum to direct proximal CA1 axons to the distal subiculum. It remains unclear if Lphn2- mediated intracellular signaling is required for its role in either context. Here, we show that Lphn2 couples to Gα12/13 in heterologous cells; this coupling is increased by constitutive exposure of the tethered agonist. Specific mutations of Lphn2’s tethered agonist region disrupt its G protein coupling and autoproteolytic cleavage, whereas mutating the autoproteolytic cleavage site alone prevents cleavage but preserves a functional tethered agonist. Using an in vivo misexpression assay, we demonstrate that wild- type Lphn2 misdirects proximal CA1 axons to the proximal subiculum and that Lphn2 tethered agonist activity is required for its role as a repulsive receptor in axons. By contrast, neither tethered agonist activity nor autoproteolysis were necessary for Lphn2’s role as a repulsive ligand in the subiculum target neurons. Thus, tethered agonist activity is required for Lphn2- mediated neural circuit assembly in a context- dependent manner. Editor's evaluation This is an intriguing study investigating the molecular mechanisms of neural circuit developmental organization. Using a defined hippocampal circuit, the authors find that ectopic expression of an adhesion G protein receptor leads to axon mistargeting. This work defines new mechanisms of axon target specificity. Introduction Latrophilins (Lphn1–3) are highly expressed in the brain and were originally identified as responders to ɑ-latrotoxin, a neurotoxin from black widow spider venom that causes the profound release of neurotransmitters from nerve terminals (Davletov et al., 1996). They belong to the family of adhesion G protein- coupled receptors (aGPCRs), capable of eliciting intracellular effects through coupling with *For correspondence: jaj2@cumc.columbia.edu (JaJ); lluo@stanford.edu (LL) †These authors contributed equally to this work Competing interest: The authors declare that no competing interests exist. Funding: See page 20 Received: 17 September 2022 Preprinted: 27 September 2022 Accepted: 16 March 2023 Published: 20 March 2023 Reviewing Editor: Kelly Monk, Vollum Institute, Oregon Health & Science University, United States Copyright Pederick, Perry- Hauser et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 1 of 23 Research article eLife digest The complex brain circuits that allow animals to sense and interact with their envi- ronment start to form early during development. Throughout this period, neurons extend fiber- like projections to establish precise wiring patterns. Various types of proteins at the surface of both incoming fibers and target cells ensure that only the right partners will connect together. Latrophilin- 2, for example, is a neuronal surface protein essential for the formation of accu- rate connections in the hippocampus, a brain region important for memory. Studded through the membrane of certain neurons, it acts as a signal- sending ligand to direct incoming fibers, with neurons that carry Latrophilin- 2 repelling projections from cells that display certain protein partners. At the same time, Latrophilin- 2 also allows neurons to receive chemical signals by working with intracellular signaling proteins known as G proteins, which help to relay information between cells. It remained unclear how this role as a signalling receptor participates in the wiring of the hippocampus during development. To explore this question, Pederick, Perry- Hauser et al. examined the impact of Latrophilin- 2 on the connection patterns of mouse hippocampal neurons that do not normally carry this protein. Intro- ducing Latrophilin- 2 into these ‘proximal CA1 cells’ misdirected them away from their usual partners – unless Latrophilin- 2 was altered so that it could not interact with G proteins. In contrast, forcing the connecting partners of CA1 cells to display normal or altered versions of Latrophilin- 2 did not inter- fere with the protein acting as a repulsive ligand. Taken together, these results suggest that the ability of Latrophilin- 2 to signal through G proteins is important for neurons that are attempting to project their fibers onto other cells, but not important when Latrophilin- 2 acts in targets to direct incoming fibers from other neurons. These results show that a single protein can shape neural circuits by acting both as a signal- receiving receptor and a signal- sending ligand depending on the context. In the future, Pederick, Perry- Hauser et al. hope that their findings will shed new light on how the wiring of the brain is disrupted in neuro- developmental disorders. heterotrimeric G proteins (Lelianova et al., 1997). Additionally, as cell adhesion molecules, latrophi- lins interact via their N- terminal extracellular domain with four different families of interacting partners including neurexins (Boucard et  al., 2012), teneurins (Silva et  al., 2011), fibronectin leucine- rich transmembrane proteins (FLRTs) (O’Sullivan et al., 2012), and contactins (Zuko et al., 2016). In the central nervous system, latrophilins have been implicated in neuronal migration, circuit assembly, and synapse formation (Anderson et al., 2017; Del Toro et al., 2020; Donohue et al., 2021; Pederick et al., 2021; Sando et al., 2019; Sando and Südhof, 2021). In humans, polymorphisms in LPHN3 are associated with an increased risk of attention- deficit/hyperactivity disorder, and a missense variant in the LPHN2 gene is responsible for extreme microcephaly (Arcos- Burgos et al., 2010; Domené et al., 2011; Vezain et al., 2018). We recently showed in mice that one of the three latrophilins, Lphn2, displays expression patterns inverse to teneurin- 3 (Ten3) in two parallel hippocampal networks (Pederick et  al., 2021). While hippocampal Lphn2 is preferentially expressed in the distal CA1 and the proximal subiculum, Ten3 is enriched in the proximal CA1 and the distal subiculum. These expression patterns and reciprocal repul- sions mediated by Ten3- Lphn2 interactions instruct proximal CA1 axons to target the distal subiculum, and more distal CA1 axons to target more proximal subiculum (Figure 1A). Specifically, Lphn2 acts as a ‘receptor’ in more distal CA1 axons that is repelled by Ten3 expressed from the distal subiculum (Figure 1B). At the same time, Lphn2 acts as a repulsive ‘ligand’ in the proximal subiculum to repel Ten3- expressing (Ten3+) proximal CA1 axons; this action requires Lphn2’s teneurin- binding domain but not its FLRT- binding activity (Figure 1C; Pederick et al., 2021). Therefore, Lphn2 is required cell autonomously as a receptor in more distal CA1 axons for their precise target selection, and non- autonomously in target neurons as a ligand for precise target selection of proximal CA1 axons. While Ten3 additionally mediates homophilic attraction (Berns et al., 2018; Pederick et al., 2021), Lphn2 does not mediate homophilic binding in trans (Boucard et al., 2014; Pederick et al., 2021). Structurally, the N- terminal extracellular domain of latrophilins comprises a rhamnose- binding lectin (RBL) domain, an olfactomedin- like (OLF) ligand- binding domain, a serine/threonine- rich region and Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 2 of 23 Developmental Biology | Neuroscience Research article Figure 1. Misexpression of latrophilin- 2 (Lphn2) in proximal CA1 axons causes axon mistargeting to the proximal subiculum. (A) Cartoon depicting the topographic connections from proximal CA1 (pCA1) to distal subiculum (dSub) and distal CA1 (dCA1) to proximal subiculum (pSub). Ten3+ proximal CA1 axons are repelled from Lphn2 expressing (Lphn2+) proximal subiculum and Lphn2+ axons are repelled from Ten3+ distal subiculum. Red symbols indicate the repulsive cues experienced by CA1 axons, previously described in Pederick et al., 2021. (B) Deletion of Lphn2 from CA1 leads to distal Figure 1 continued on next page Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 3 of 23 Developmental Biology | Neuroscience Research article Figure 1 continued CA1 axons mistargeting to distal subiculum, suggesting that Lphn2 acts cell- autonomously as a repulsive receptor. (C) Deletion of Lphn2 from proximal subiculum results in proximal CA1 axon mistargeting to proximal subiculum, suggesting Lphn2 acts cell- non- autonomously as a repulsive ligand. Figures (A–C) are based on Pederick et al., 2021. (D) Experimental design of Lphn2 misexpression assay in proximal CA1. At postnatal day (P) 0, lentivirus expressing Cre or Cre and Lphn2 was injected into CA1. This was followed by injection at P42 of Cre- dependent membrane bound mCherry (mCh) into proximal CA1 as an axon tracer. (E and F) Representative images of AAV- DIO- mCh (magenta; mCh expression in a Cre- dependent manner) injections in proximal CA1 (top) and corresponding projections in the subiculum (bottom). (G) A representative image of the subiculum with proximal subiculum (pSub), mid subiculum (mSub), and distal subiculum (dSub) regions highlighted. (H) The fraction of total axon intensity within proximal, mid, and distal subiculum. Cre: N=5 and Cre- Lphn2: N=5. Means ± SEM; two- way ANOVA with Sidak’s multiple comparisons test. Injection sites of all subjects are shown in Figure 1—figure supplement 3. Scale bars represent 200 μm. The online version of this article includes the following source data and figure supplement(s) for figure 1: Source data 1. Misexpression of latrophilin- 2 (Lphn2) in proximal CA1 axons causes axon mistargeting to the proximal subiculum. Figure supplement 1. In vivo expression of lentivirus used in Figures 1, 3 and 4. Figure supplement 2. Quantification of latrophilin- 2 (Lphn2), Lphn2_F831A/M835A, and Lphn2_T829G expression in CA1. Figure supplement 2—source data 1. Quantification of latrophilin- 2 (Lphn2), Lphn2_F831A/M835A, and Lphn2_T829G expression in CA1. Figure supplement 3. Mean injection site positions for proximal CA1 axon tracing in Figures 1, 3 and 4. hormone receptor motif (HRM), and a conserved GPCR autoproteolysis- inducing (GAIN) domain that encompasses the GPCR proteolysis site (GPS) (Araç et al., 2012; Moreno- Salinas et al., 2019; Vizur- raga et al., 2020; Figure 2A). aGPCRs undergo autoproteolytic cleavage at the HL/T consensus site within the GPS. This self- cleavage divides the receptor into an extracellular N- terminal fragment (NTF) and a membrane- bound C- terminal fragment (CTF) that remain noncovalently associated throughout biosynthesis and membrane trafficking (Vizurraga et  al., 2020). The seven residues immediately C- terminal to the GPS constitute the tethered agonist peptide (also known as the Stachel or stalk peptide), which upon exposure binds within the transmembrane domain to activate heterotrimeric G proteins (Liebscher and Schöneberg, 2016). While our previous in vivo work established that interaction between Ten3 and Lphn2 was required for precise circuit assembly (Pederick et  al., 2021), it did not examine how this might depend on Lphn2- mediated signaling mechanisms. Here, we modified our previous hippocampal model to develop an Lphn2 misexpression assay (Figure  1D). We misexpressed Lphn2 in either CA1 axons distal subiculum axon or the subiculum target and assessed the impact on normal proximal CA1 targeting. We found that ectopically expressing wild- type Lphn2 in proximal CA1 axons causes their mistargeting to the proximal subiculum. This provided us with a robust platform to interrogate whether tethered agonist activity or autoproteolytic cleavage is required for axon mistargeting in this Lphn2 ectopic expression system. When misexpressed in CA1, Lphn2 tethered- agonist activity was required for Lphn2- mediated axon mistargeting. By contrast, when we misexpressed Lphn2 in subiculum target neurons, both tethered agonist activity and autoproteolysis were dispensable for Lphn2- mediated axon repulsion. Thus, our data support that Lphn2 G- protein coupling is required in axons but not target neurons during precise circuit assembly. → Results Misexpression of wild-type Lphn2 in proximal CA1 leads to axon mistargeting in the subiculum To investigate the role of Lphn2- mediated G protein activity in hippocampal axon targeting, we first designed a gain- of- function assay in which we misexpressed Lphn2 in proximal CA1 neurons. We hypothesized that this ectopic expression would cause proximal CA1 axons to avoid the Ten3+ distal subiculum and incorrectly target the proximal subiculum. If so, this platform could provide us with an assay to test Lphn2 mutants with defects in various functions to determine whether wild- type Lphn2 mistargeting is compromised. To test our hypothesis, we used a dual injection strategy to ectopically express Lphn2 in proximal CA1 and trace its axons into the subiculum (Figure 1—figure supplement 1). At postnatal day 0 (P0), lentivirus expressing Cre (LV- Cre) (control) or Cre- Lphn2 (LV- Cre- P2A- Lphn2) was injected into prox- imal CA1, followed by injection of a Cre- dependent membrane- bound mCherry (AAV- DIO- mCherry) Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 4 of 23 Developmental Biology | Neuroscience Research article Figure 2. Exposure of the latrophilin- 2 (Lphn2) tethered agonist (TA) promotes intracellular signaling through Gα12/13. (A) Cartoon representations of full- length and tethered agonist- exposed (CTF) Lphn2 with detailed amino acid sequences for the TA. The extracellular domain of Lphn2 comprises an N- terminal rhamnose- binding lectin domain (RBL), an olfactomedin- like domain (OLF), a serine/threonine- rich region, and a HormR domain (HRM). It also contains the GPCR autoproteolysis- inducing (GAIN) domain necessary for autoproteolytic cleavage. This cleavage divides the aGPCR into two Figure 2 continued on next page Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 5 of 23 Developmental Biology | Neuroscience Research article Figure 2 continued polypeptide chains: an N- terminal fragment (NTF) and a C- terminal fragment (CTF). The peptide stretch directly following the proteolytic cleavage site is known as the ‘Stachel’ or tethered agonist. Exposure of the tethered agonist results in aGPCR activation and downstream signaling. (B) Representative immunoblot analysis (N=3) of wild- type Lphn2 and Lphn2- CTF expression in HEK293T cells using a primary antibody against FLAG (1:500, ThermoFisher, PA1- 984B). Expected bands for full- length Lphn2- FLAG and Lphn2- CTF- FLAG are 164 kDa and 72 kDa, respectively. (C) Serum response element (SRE) luciferase reporter assay for Lphn2 and Lphn2- CTF shows that removing the entire NTF up to the GPS cleavage site constitutively enhances SRE signaling (N=3 biological replicates, 9 technical replicates). (D) Schematic outlining the Gβγ-release bioluminescence resonance energy transfer (BRET) assay. The Lphn2 tethered agonist is capped with an enterokinase cleavage site (EK) preceded by a hemagglutinin signal peptide (SP), the P2Y12 N- terminal extracellular sequence, and a flexible linker (Lizano et al., 2021). Addition of 10 nM enterokinase generates a tethered agonist neoepitope identical to activated endogenous Lphn2. Lphn2 activation results in G protein dissociation, allowing Gβγ-Venus to associate with the C- terminus of GPCR kinase 3 (GRK3- ct) (Hollins et al., 2009). (E) Gβγ-release BRET assay testing SP- P2Y12- EK- Lphn2- CTF activation of Gαs, Gαi, Gαq, Gα12, and Gα13 in HEKΔ7 cells (N=3–4 biological replicates, 9–12 technical replicates). β2- adrenergic receptor (β2AR) with 1 µM isoproteronal, dopamine receptor D2 (D2R) with 10 µM quinpirole, and endothelin receptor (ETA) with 100 nM ligand ET- 1 were used as positive controls, for Gαs, Gαi and Gαq/12/13 ,respectively. Means ± SEM; Multiple unpaired t tests between no G protein and G protein conditions; **p<0.01; ****p<0.0001. The online version of this article includes the following source data and figure supplement(s) for figure 2: Source data 1. Full raw unedited blot of latrophilin- 2 (Lphn2) expression in HEK293T cells. Source data 2. Uncropped immunoblot analysis of latrophilin- 2 (Lphn2) expression in HEK293T cells. Source data 3. Replicates of the immunoblot assay. Figure supplement 1. Gαq- inhibitor YM- 254890 does not impair serum response element (SRE) luciferase response of Lphn2- CTF. Figure supplement 2. Gβγ-release bioluminescence resonance energy transfer (BRET) assay shows latrophilin- 2 (Lphn2) couples to Gα12 and Gα13. into proximal CA1 in the same mice at approximately P42 (Figure 1D). We confirmed the expression of Lphn2 in CA1 axons by the presence of FLAG immunostaining in the subiculum and that ectopic expression levels were higher than that of endogenous Lphn2 (Figure 1—figure supplement 2A, C and D). As expected, in control animals (Cre), Cre expressing (Cre+) proximal CA1 axons targeted the most distal parts of the subiculum (Figure 1E). By contrast, when Lphn2 was misexpressed in prox- imal CA1 (Cre- Lphn2), Cre+ proximal CA1 axons targeted the most proximal parts of the subiculum (Figure 1F). To analyze the location of proximal CA1 axons in the subiculum, we calculated the fraction of axon intensity within thirds of the subiculum across the proximal/distal axis (Figure 1G). Proximal CA1 axons misexpressing Lphn2 are located significantly more in the proximal third of the subiculum and significantly less in the distal third of the subiculum when compared to control axons (Figure 1H). These data supported our hypothesis that ectopic expression of Lphn2 in proximal CA1 axons causes mistargeting to the proximal subiculum. Importantly, the phenotype observed when overex- pressing Lphn2 in pCA1 axons is more severe than that observed when Ten3 is deleted (Berns et al., 2018), suggesting that mistargeting is not caused by disruption of Ten3 expression alone. Having established the effect of wild- type Lphn2 misexpression in proximal CA1 axons, we next sought to characterize G protein coupling of wild- type Lphn2 and generate Lphn2 mutants to test the require- ment of G protein signaling in Lphn2 mediated neural circuit assembly. Lphn2 signals through Gα12/13 The G protein interaction partners for Lphn2 have not been previously established. We recently showed that Lphn3, another member of the latrophilin family of aGPCRs, couples principally to Gα12/13, and also more weakly to Gαq, using a combination of gene expression assays and an activation strategy that permitted acute exposure of the tethered agonist in a live- cell system (Mathiasen et al., 2020). Thus, we began our signaling characterization of Lphn2 similarly using a wild- type full- length Lphn2 construct, and a constitutively active construct termed Lphn2- CTF (Figure 2A). The wild- type Lphn2 construct comprises all extracellular elements including the RBL, OLF, HRM, and GAIN domains, in addition to the seven transmembrane helix domain. The Lphn2- CTF lacks the entire NTF up to the GPS and instead has only a methionine residue before the tethered agonist. We tested the expression of these constructs in mammalian cells using immunoblotting and showed that both full- length Lphn2 and Lphn2- CTF ran at the expected truncated position (~72 kDa) suggesting that full- length Lphn2 undergoes normal proteolytic cleavage (Figure 2B). This result for full- length Lphn2 is similar to our work characterizing autoproteolysis of Lphn3 (Perry- Hauser et al., 2022). To infer the activity of these constructs in G protein signaling pathways, we used a luminescence- based gene expression assay for serum response element (SRE), which produced a robust response Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 6 of 23 Developmental Biology | Neuroscience Research article in our previous studies of Lphn3 (Mathiasen et al., 2020). In our assay design, SRE action is coupled to the transcription and translation of firefly luciferase; this readout is then normalized to the control reporter, Renilla luciferase, expressed from the same plasmid under a constitutive promoter. We found that Lphn2- CTF significantly enhanced signaling over wild- type Lphn2 for SRE gene expression at varying levels of cDNA transfection (Figure 2C). Since the SRE assay reports on signaling by Gα12/13 as well as Gαq we tested whether Gα12/13 or Gαq was the primary contributor to this response using a selective Gαq inhibitor, YM- 254890 (Figure 2—figure supplement 1). We did not observe a significant effect upon the addition of the inhibitor, suggesting that Lphn2 signals through Gα12/13. To verify our result in the context of acute G protein activation, we next tested how tethered agonist exposure affects G protein activation in a bioluminescence resonance energy transfer (BRET) assay (Figure 2D). We designed a synthetically- activatable Lphn2 construct based on a recent publi- cation that took advantage of the protease enterokinase (Lizano et al., 2021). Enterokinase selec- tively recognizes the trypsinogen substrate sequence DDDDK and cleaves after the lysine residue, thereby exposing the native tethered agonist. Thus, we cloned an Lphn2 construct that included a modified hemagglutinin signal peptide, the P2Y12 N- terminal extracellular sequence (amino acids 1–24), a flexible linker (GGSGGSGGS), the enterokinase recognition site (DYKDDDDK), and the trun- cated Lphn2- CTF sequence. We tested this construct in a Gβγ-release assay where energy transfer was monitored between the membrane- anchored luminescent donor, GRK3- ct- Rluc8, and the fluo- rescent acceptor, Gγ-Venus (Hollins et al., 2009). This assay was performed in a HEKΔ7 cell line with targeted deletion of Gα12 and Gα13, as well as Gαs/olf, Gαq/11, and Gαz (Alvarez- Curto et al., 2016) to enable systematic re- introduction of the Gα subunits. As expected, in the absence of Gα subunits no BRET signal was observed; however, when Gα12 or Gα13 was re- introduced to cells expressing the Lphn2 construct there was a significant increase in the BRET signal upon treatment with enterokinase (Figure 2E). This increase was not observed upon co- expression of the receptor with Gαs, Gαi1, or Gαq (Figure 2E and Figure 2—figure supplement 2). This suggests that the increase in cAMP reported previously for the Lphn2 CTF (Sando and Südhof, 2021) may not result from the direct activation of Gαs, but rather from some other form of signaling crosstalk. Alternatively, it is possible that our Lphn2, which was isolated from the P8 hippocampus and lacks exons 19 and 20, may represent a different transcript variant in the brain that activates distinct signaling pathways. In fact, alternative splicing has been shown to affect G protein coupling specificity for several GPCRs, including Lphn3 (Markovic and Challiss, 2009; Röthe et al., 2019). Taken together, these data demonstrate that Lphn2 signals through the G proteins Gα12 and Gα13 in heterologous cells. Having established that these in- cell methods were sufficient to characterize G protein signaling pathways for Lphn2, we next characterized how different mutations in the tethered agonist region affect intracellular signaling. Mutating conserved residues F831A and M835A in the tethered agonist impairs G protein coupling activity Previous studies suggest that the third and seventh residues of aGPCRs are required for tethered agonist- mediated G protein activation (Stoveken et al., 2015). We hypothesized that mutating these residues in Lphn2, phenylalanine (F831), and methionine (M835), to alanine (F831A/M835A) would impair G protein signaling mediated by the tethered agonist (Figure 3A). Like our work with wild- type Lphn2 and Lphn2- CTF, we mutated the tethered agonist residues in both full- length and truncated constructs (Lphn2_F831A/M835A and Lphn2- CTF_F831A/M835A, respectively). Immunoblotting against the C- terminal FLAG- tag confirmed expression in HEK293T cells but showed that Lphn2_ F831A/M835A is largely uncleaved (Figure  3B). This is consistent with previous work with Lphn1 showing that mutating the third phenylalanine to an alanine disrupts autoproteolytic cleavage (Araç et al., 2012) and shows that the double mutation (F831A/M835A) in Lphn2 also inhibits cleavage. We also validated that Lphn2_F831A/M835A is expressed on the cell surface at a comparable level as Lphn2 wild- type (Figure 3—figure supplement 1). We then proceeded to test these constructs in our SRE gene expression system (Figure 3A). As hypothesized, both the full- length and truncated Lphn2 had dramatically impaired responses to SRE across varying levels of cDNA transfection. To confirm that the reduced SRE response was due to impaired G protein coupling and not simply to impaired proteolysis, we cloned the CTF of our Lphn2_F831A/M835A mutant into our enterokinase- activatable construct. We then tested our construct in the Gβγ-release assay and compared the BRET Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 7 of 23 Developmental Biology | Neuroscience Research article Figure 3. Lphn2_F831A/M835A has impaired G protein coupling activity and autoproteolytic cleavage and fails to misdirect proximal CA1 (pCA1) axons to the proximal subiculum (pSub) when misexpressed. (A) Schematic of the mutated tethered agonist for Lphn2_F381A/M835A. The serum response element (SRE) luciferase reporter assay shows that both the full- length Lphn2_F831A/M835A and the Lphn2- CTF_F831A/M835A have impaired signaling (N=3 biological replicates, 9 technical replicates). Means ± SEM; Multiple unpaired t- tests between full- length latrophilin- 2 (Lphn2) and Lphn2_F831A/ M835A and Lphn2- CTF and Lphn2- CTF_F831A/M835A constructs; ****p<0.0001. (B) Representative immunoblot analysis (N=3) of TA- dead Lphn2 and TA- dead Lphn2- CTF expression in HEK293T cells using a primary antibody against FLAG (1:500, ThermoFisher, PA1- 984B). Expected bands for full- length Lphn2_F831A/M835A- FLAG and Lphn2_F831A/M835A- CTF- FLAG are 164 kDa and 72 kDa, respectively. (C) Gβγ-release BRET assay testing SP- P2Y12- EK- Lphn2- CTF_F831A/M835A activation of Gα12 and Gα13 in HEKΔ7 cells (N=3–4 biological replicates, 9–12 technical replicates). SP- P2Y12- EK- Lphn2- CTF signaling is shown for comparison. Means ± SEM; Multiple unpaired t tests between no G protein and G protein conditions; *p<0.05, **p<0.01; ****p<0.0001. (D) Representative images of AAV- DIO- mCh (magenta; mCh expression in a Cre- dependent manner) injections in proximal CA1 (top) and corresponding projections in the subiculum (bottom). (E) Fraction of total axon intensity within proximal, mid, and distal subiculum. Cre: N=5, Cre- Lphn2: N=5 and Cre- Lphn2_F831A/M835A: N=6. Means ± SEM; two- way ANOVA with Sidak’s multiple comparisons test. Injection sites of all subjects are shown in Figure 1—figure supplement 3. Scale bars represent 200 μm. The online version of this article includes the following source data and figure supplement(s) for figure 3: Source data 1. Uncropped immunoblot analysis of latrophilin- 2 (Lphn2) expression in HEK293T cells. Source data 2. Lphn2_F831A/M835A has impaired G protein coupling activity and autoproteolytic cleavage and fails to misdirect proximal CA1 (pCA1) axons to the proximal subiculum (pSub) when misexpressed. Figure 3 continued on next page Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 8 of 23 Developmental Biology | Neuroscience Research article Figure 3 continued Figure supplement 1. Latrophilin- 2 (Lphn2), Lphn2_F831A/M835A, and Lphn2_T829G are expressed at the cell surface at comparable levels. Figure supplement 2. Comparison of fraction of total axon intensity across the subiculum within the same experimental condition. response to wild- type Lphn2- CTF. Unlike the wild- type receptor, Lphn2- CTF_F831A/M835A did not yield a BRET signal after the re- introduction of any of the G proteins in question (Gαs, Gαi1, Gαq, Gα12, or Gα13) (Figure 3C, Figure 2—figure supplement 2). Taken together, our findings demonstrate that the F831A/M835A mutations in Lphn2 impair tethered agonist- mediated G protein coupling. Tethered agonist activity or autoproteolysis of Lphn2 is required for its cell-autonomous effect in causing proximal CA1 axon mistargeting We next misexpressed Lphn2- F831A/M835A in proximal CA1 to determine if Lphn2 tethered agonist activity or autoproteolysis is required in vivo to direct mistargeting of proximal CA1 axons. Lphn2_ F831A/M835A was ectopically expressed at levels similar to that of wild- type Lphn2 and was detected in CA1 axons (Figure  1—figure supplement 1A and B). We injected LV- Cre- P2A- Lphn2_F831A/ M835A- FLAG into CA1 of P0 mice, followed by AAV- DIO- mCherry into proximal CA1 of the same mice as adults. The majority of Lphn2_F831A/M835A- expressing proximal CA1 axons targeted the most distal third of the subiculum, like negative control Cre animals (Figure 3D). The fraction of axon intensity in Cre- Lphn2_F831A/M835A animals was significantly lower in the proximal subiculum and significantly higher in the distal subiculum when compared to Cre- P2A- Lphn2 animals (Figure  3E). Proximal CA1 axons in Cre- P2A- Lphn2_F831A/M835A animals showed a similar pattern of targeting to negative control Cre animals (Figure 3—figure supplement 2), although the total fraction of axon intensity was significantly lower in the distal subiculum (Figure 3E). Collectively, these findings suggest that tethered agonist activity or autoproteolysis is required for Lphn2- mediated miswiring of proximal CA1 axons. Mutating residue T829G in the tethered agonist renders Lphn2 cleavage deficient but preserves the ability of the tethered agonist to activate G protein While misexpressing Lphn2_F831A/M835A failed to cause proximal CA1 axons to mistarget to the proximal subiculum, we could not definitively link this result to impaired tethered agonist activity since the Lphn2_F831A/M835A mutant was also resistant to autoproteolytic cleavage (Figure  3B). Since our initial efforts to find a tethered agonist mutant with impaired G protein signaling that retained normal autoproteolytic cleavage were unsuccessful, we designed a construct that rendered Lphn2 resistant to autoproteolytic cleavage but preserved tethered agonist activity. Previous studies showed that replacing threonine- 838 in the tethered agonist of Lphn1 or threonine- 923 in the tethered agonist of Lphn3 to glycine inhibited autoproteolysis while maintaining proper folding of the receptor (Araç et al., 2012; Kordon et al., 2023). Thus, we mutated the analogous threonine- 829 in Lphn2 (Lphn2_ T829G) and confirmed that Lphn2_T829G was cleavage resistant using immunoblotting (Figure 4A). We also validated that Lphn2_T829G is expressed on the cell surface at a comparable level as Lphn2 wild- type (Figure 3—figure supplement 1). We next assessed G protein signaling for Lphn2_T829G using our SRE gene expression system with full- length and truncated receptors (Lphn2_T829G and Lphn2- CTF_T829G, respectively) (Figure 4B). Full- length Lphn2_T829G had significantly impaired SRE response compared to wild- type Lphn2, consistent with diminished exposure of the tethered agonist in the absence of cleavage; however, when we tested Lphn2- CTF_T829G, which lacked the entire NTF up to the GPS cleavage site, we observed SRE levels comparable to Lphn2- CTF suggesting that the mutated tethered agonist is fully active if exposed. We, therefore, cloned the CTF of the T829G mutant into our enterokinase- activatable construct and tested BRET signaling following the re- introduction of Gα proteins (Figure  4C). The T829G- CTF retained BRET signaling comparable to wild- type Lphn2- CTF for Gα12 and Gα13, with no discernable BRET response for Gαs, Gαi1, or Gαq (Figure 2—figure supplement 2). Taken together, these data supported that Lphn2_T829G is resistant to autoproteolytic cleavage but maintains a func- tional tethered agonist. Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 9 of 23 Developmental Biology | Neuroscience Research article Figure 4. Lphn2_T829G impairs autoproteolytic cleavage, retains G protein activity in the truncated receptor, and misdirects axons to the proximal subiculum (pSub) when misexpressed. (A) Representative immunoblot analysis (N=3) of Lphn2_T829G and Lphn2- CTF_T829G expression in HEK293T cells using a primary antibody against FLAG (1:500, ThermoFisher, PA1- 984B). Expected bands for full- length Lphn2_T829G- FLAG and Lphn2- CTF_ T829G- FLAG are 164 kDa and 72 kDa, respectively. (B) Schematic of the mutated tethered agonist for Lphn2_T829G. The serum response element (SRE) luciferase reporter assay shows that the full- length Lphn2_T829G has impaired SRE levels while the Lphn2_T829G truncated up to the GPS cleavage site has SRE levels comparable to Lphn2_CTF (N=3 biological replicates, nine technical replicates). Means ± SEM; Multiple unpaired t tests between full- length Lphn2 and Lphn2_T829G and Lphn2- CTF and Lphn2- CTF_T829G constructs; *p<0.05; ***p<0.001; ****p<0.0001. (C) Gβγ-release BRET assay testing SP- P2Y12- EK- Lphn2- CTF_T829G activation of Gα12 and Gα13 in HEKΔ7 cells (N=3–4 biological replicates, 9–12 technical replicates). SP- P2Y12- EK- Lphn2- CTF signaling is shown for comparison. Means ± SEM; Multiple unpaired t- tests between no G protein and G protein conditions; ***p<0.001; ****p<0.0001. (D) Representative images of AAV- DIO- mCh (magenta; mCh expression in a Cre- dependent manner) injections in proximal CA1 (top) and corresponding projections in the subiculum (bottom). (E) Fraction of total axon intensity within proximal, mid, and distal subiculum. Cre: n=5, Cre- Lphn2: n=5 and Cre- Lphn2_T829G: n=5. Means ± SEM; two- way analysis of variance (ANOVA) with Sidak’s multiple comparisons test. Injection sites of all subjects are shown in Figure 1—figure supplement 3. Scale bars represent 200 μm. The online version of this article includes the following source data for figure 4: Source data 1. Uncropped immunoblot analysis of latrophilin- 2 (Lphn2) expression in HEK293T cells. Source data 2. Lphn2_T829G impairs autoproteolytic cleavage, retains G protein activity in the truncated receptor, and misdirects axons to the proximal subiculum (pSub) when misexpressed. Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 10 of 23 Developmental Biology | Neuroscience Research article Autocleavage-deficient Lphn2 retains moderate activity in directing proximal CA1 axon mistargeting To assess if autoproteolytic cleavage is required in vivo for Lphn2- mediated proximal CA1 axon mistar- geting, we injected Cre- Lphn2_T829G into proximal CA1 of P0 mice, followed by AAV- DIO- mCherry into pCA1 of the same mice as adults. Lphn2_T829G was ectopically expressed at levels similar to wild- type Lphn2 and was detected in CA1 axons (Figure 1—figure supplement 1A and B). Overall, Lphn2_T829G- expressing proximal CA1 axons did not show highly enriched targeting to a specific region of the subiculum, as observed for control proximal CA1 axons or Lphn2 misexpressing prox- imal CA1 axons, which preferentially target distal and proximal subiculum, respectively (Figure 4E). However, compared to the control there was a significant increase in the fraction of axon intensity in the proximal subiculum in Cre- Lphn2_T829G animals, even though this mistargeting was not as pronounced as seen with wild- type Cre- Lphn2 (Figure 3—figure supplement 2.). The intermediate gain- of- function phenotypes of misexpressing Lphn2_T829G compared to misexpressing wild- type Lphn2 or tethered agonist- deficient (and non- cleavable) Lphn2 suggest that autoproteolysis is not absolutely required for Lphn2 misexpression- induced miswiring of proximal CA1 axons. The weaker than wild- type overexpression phenotype is likely caused by the decreased G protein signaling of the full- length construct given the more limited exposure of the tethered agonist in the absence of cleavage. The preservation of some signaling activity of Lphn2_T829G is consistent with the ability of its tethered agonist to signal. Neither tethered agonist activity nor autoproteolysis is required for Lphn2’s action as a repulsive ligand We previously showed that misexpression of Lphn2 in distal subiculum target neurons causes prox- imal CA1 axons to avoid this area, suggesting that Lphn2 acts cell non- autonomously as a repulsive ligand in directing target selection of proximal CA1 axons (Pederick et al., 2021). In the context of repulsive axon guidance, proteolysis has been proposed as a mechanism to disassemble the extracel- lular binding complex after repulsive signaling, which is necessary for repulsion (Hattori et al., 2000). Are tethered agonist activity and/or autoproteolysis required for Lphn2’s cell non- autonomous role in neural circuit assembly? To test this, we used a strategy to ectopically express Lphn2 in the distal subiculum and trace proximal CA1 axons into the subiculum (Figure 5A; Figure 5—figure supple- ment 1) identical to the one we previously reported for comparing lentiviruses expressing GFP alone (LV- GFP) or GFP and Lphn2 (LV- GFP- P2A- Lphn2- FLAG; Pederick et  al., 2021). At postnatal day 0 (P0), LV- GFP- P2A- Lphn2_F831A/M835A- FLAG and LV- GFP- P2A- Lphn2_T829G- FLAG were injected into distal subiculum, followed by injection of membrane- bound mCherry (AAV- mCherry) into prox- imal CA1 in the same mice at approximately P42. The P0 lentivirus injection only covers a small frac- tion of the entire proximal CA1 axon projection, enabling us to assess whether proximal CA1 axons target lentivirus- expressing regions differently from adjacent regions that do not express lentivirus. To observe the relationship between proximal CA1 axon projections and lentivirus- induced regions of the subiculum, we plotted axon signal intensity (mCh) and lentivirus injection site (GFP) from the same animal as height and color, respectively. We previously reported that GFP alone does not affect the intensity of proximal CA1 axons, whereas GFP- Lphn2 regions have significantly reduced proximal CA1 axon intensity in GFP- Lphn2 positive regions (Pederick et al., 2021; Figure 5B and C; Figure 5—figure supplement 2A and B). When either GFP- Lphn2_F831A/M835A or GFP- Lphn2_T829G were expressed in the distal subic- ulum, we also observed a significant decrease in axon intensity in GFP positive regions compared to GFP (Figure 5D, E and F and Figure 5—figure supplement 2C and D). This decrease was not signifi- cantly different from GFP- Lphn2 animals (Figure  5F). These findings suggest that neither tethered agonist activity nor autoproteolysis is required for Lphn2’s cell- non- autonomous role as a ligand in the neural circuit assembly. Discussion In this study, we utilized a combination of in vivo axon target selection and in vitro cell signaling assays to determine if Lphn2 G protein signaling is required for its role as a neural wiring molecule. First, we showed that Lphn2 misexpression can cell- autonomously misdirect proximal CA1 axons to the Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 11 of 23 Developmental Biology | Neuroscience Research article Figure 5. Misexpression of latrophilin- 2 (Lphn2) mutants in target distal subiculum neurons does not cause mistargeting of proximal CA1 axons. (A) Experimental design of Lphn2 misexpression assay in distal subiculum. (B to E). Representative mountain plots showing normalized mCh fluorescence as height (proximal CA1 axon projections in subiculum) and normalized GFP fluorescence as color (lentivirus expression). P, proximal; D, distal; M, medial; L, lateral. (F) Ratio of mCh fluorescence intensity (from proximal CA1 axons) in GFP+ versus GFP– regions of the subiculum. GFP: N=5, GFP- Lphn2: N=5, GFP- Lphn2_F831A/M835A: N=5 and GFP- Lphn2_T829G: N=6. Means ± SEM. One- way analysis of variance (ANOVA) with Tukey’s multiple comparisons test. © 2021, AAAS. Data from panels B, C and F (left two columns) are respectively reproduced from Figure S8B (top left panel), Figure S8E (bottom right panel) and Figure 2G (left two columns) of Pederick et al., 2021, reprinted with permission from AAAS. The panels from B, C and F are therefore not covered by the CC- BY 4.0 license and further reproduction would need permission from the copyright holder The online version of this article includes the following source data and figure supplement(s) for figure 5: Source data 1. Misexpression of Lphn2 mutants in target distal subiculum neurons does not cause mistargeting of proximal CA1 axons. Figure supplement 1. In vivo expression of lentivirus used in Figure 5. Figure supplement 2. Representative images corresponding to Figure 5B–E. proximal subiculum, establishing an assay to test the requirements of Lphn2 G protein signaling when it acts as a receptor (Figure  1). Second, we identified the G protein interaction partners of Lphn2 (Figure  2) and validated point mutations that disrupt tethered agonist activity and/or autoproteol- ysis of the GPS region (Figures 3 and 4). Third, we showed that when Lphn2 is misexpressed in CA1 axons, tethered agonist activity is required for Lphn2’s ability to misdirect axon targeting (Figures 3 Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 12 of 23 Developmental Biology | Neuroscience Research article and 4). Finally, when Lphn2 acts as a repulsive ligand in subiculum target neurons, we demonstrated that neither tethered agonist activity nor autoproteolytic cleavage is required for the receptor’s ability to repel Ten3+ proximal CA1 axons (Figure 5). Taken together, these findings highlight the impor- tance of Lphn2 G protein signaling during precise circuit assembly in a context- specific manner. Our results also support that while aGPCR GPS cleavage is dispensable for Lphn2’s role as a receptor to direct axon targeting, an intact tethered agonist is essential. The role of autoproteolytic cleavage and tethered agonism in aGPCR activation Upon aGPCR biosynthesis, the conserved GAIN domain undergoes autoproteolytic cleavage at the GPS to generate N- and C- terminal fragments that remain non- covalently bound during trafficking to the cell surface (Araç et al., 2012). Crystal structures of GAIN domains from Lphn1 and ADGRB3 (BAI3) (Araç et al., 2012), ADGRG1 (GPR56) (Salzman et al., 2016), and ADGRG6 (GPR126) (Leon et al., 2020), revealed that in an intact aGPCR, the tethered agonist is buried as a β- strand in the GAIN domain, forming an extensive network of conserved hydrogen bonds and hydrophobic side chains. This suggests that when the complex between aGPCR’s NTF and CTF remains intact the teth- ered agonist is inaccessible for engagement with its binding site in the 7TM domain. However, several studies have reported that naturally cleavage- resistant aGPCRs can still function (Liebscher et  al., 2014; Wilde et al., 2016). This suggests that the tethered agonist can in fact interact with the 7TM independent of cleavage, although the efficacy of this interaction is likely to be diminished, as we see in the present study. Thus, the relative contributions and/or necessity of autoproteolytic cleavage and the tethered agonist to aGPCR activity remain an area of active study. Recent efforts reported the structures of eight aGPCRs, seven of which were truncated up to the GPS (Barros-Álvarez et al., 2022; Ping et al., 2022; Qu et al., 2022; Xiao et al., 2022). While most discussions from the structural aGPCR studies argued that NTF dissociation is required for tethered agonist interaction with the receptor, the structure of autoproteolysis- deficient ADGRF1 supported the possibility of the cleavage- independent manner of receptor activation (Qu et  al., 2022). The density for the tethered agonist of ADGRF1 was well- resolved and bound in an α-helical structure within the orthosteric site of the 7TM bundle. This interaction was like that of the cleaved structures, and ADGRF1 was also observed to be bound to a miniGi1, supporting that receptor cleavage and tethered agonist exposure are not absolutely required for G protein coupling. One caveat, however, is that there is little density for the NTF in this structure, suggesting that the structure obtained may result from a fraction of receptor where cleavage has still occurred and the NTF has dissociated. As mentioned above, not all aGPCRs are auto- proteolytically cleaved; therefore, activation cannot be fully dependent on tethered agonist exposure through the removal of the NTF (Kishore et al., 2016; Liebscher et al., 2022). In this regard, it is possible that full- length aGPCRs exist in multiple conformational states that include receptor molecules in which the tethered agonist is unmasked from the GAIN domain. In fact, molecular dynamics (MD) simulations of spontaneous tethered agonist exposure were recently reported for five intact aGPCR homologs (ADGRB3, ADGRE2, ADGRE5, ADGRG1, and Lphn1) (Beliu et al., 2021). Here, the authors show that tethered agonist exposure occurs due to the high intrinsic flexibility of the GAIN domain. They also used biorthogonal labeling of conserved positions within the tethered agonist to show that large portions (+6 residues) of the tethered agonist can become solvent accessible in the context of the GAIN domain. They argue that tethered agonist exposure likely occurs in a stepwise mechanism where the tethered agonist is uncovered along its N C axis. Thus, it is possible that an intact complex of aGPCR’s NTF and CTF could unmask the tethered agonist sufficiently for interaction with the 7TM, resulting in receptor activation. → The ability of tethered agonist exposure to occur in intact aGPCRs could provide an explanation for why our Lphn2_T829G mutant displays a partial axon mistargeting phenotype (Figure  4). Even though the Lphn2_T829G mutant cannot undergo autoproteolytic cleavage, it still retains a functional tethered agonist that is able to initiate G protein signaling if transiently unmasked. This is likely why activation of full- length Lphn2_T829G is less robust than the wild- type Lphn2, which can more readily unmask the TA. We also cannot rule out the possibility that a small amount of cleavage, although undetectable in the HEK cells immunoblotting, nonetheless contributes to the partial mistargeting phenotype in vivo. Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 13 of 23 Developmental Biology | Neuroscience Research article Like the Lphn2_T829G mutant, Lphn2_F831A/M835A cannot undergo autoproteolytic cleavage. However, this mutant also has impaired G protein coupling activity even with full exposure of the TA. This explains why we observed close- to- normal axon targeting when we overexpressed Lphn2_ F831A/M835A in proximal CA1 (Figure  3). Even if the tethered agonist of Lphn2_F831A/M835A becomes unmasked it still cannot initiate tethered agonist- mediated receptor. However, given the statistically significant difference in total axon intensity in dSub between control and Lphn2_F831A/ M835A (Figure  3E), we cannot rule out some residual tethered agonist- independent, G protein- mediated signaling or an Lphn2 function independent of G protein signaling, at least in the context of overexpression. Implication of Lphn2 signaling in neural circuit assembly How could Lphn2- mediated G protein signaling in the CA1 axons lead to axon repulsion? We show here that Lphn2 primarily signals through Gα12 and Gα13 in heterologous cells (Figure  2E). Gα12/ Gα13 are known to regulate Rho GTPase; for example, Gα13 binds to and activates p115RhoGEF, an exchange factor for and activator of the small GTPase RhoA (Kozasa et  al., 2011). RhoA acti- vation is known to cause growth cone collapse and neuronal process retraction via its regulation of the actin- myosin contractility (Luo, 2002; Spillane and Gallo, 2014). Thus, if Lphn2 coupling to Gα12/Gα13 also applies to CA1 neurons, as suggested by our results in heterologous cells, Gα12/ RhoA may be a plausible pathway for Lphn2 to mediate its function as a receptor Gα13 → for axon repulsion. RhoGEF → Interestingly, neither autoproteolytic cleavage nor tethered agonist activity is required for Lphn2 to act cell non- autonomously as a repulsive ligand in subiculum target neurons (Figure 5). This suggests that cleavage of Lphn2 is not required for repulsion in this context and implies that another mecha- nism mediates the disassembly of the extracellular binding complex, which is required for retracting axons to pull away from the targets. Other potential mechanisms to disassemble the extracellular binding complex include Ten3 cleavage (teneurins are known to also undergo proteolytic cleavage at its extracellular domain; Sita et al., 2019), endocytosis of the adhesion complex as in the case of ephrin/Eph receptor (Egea and Klein, 2007), or forces produced by actin- myosin contractility in axon terminals induced by repulsive signaling. Indeed, since Lphn2 also acts as a repulsive ligand and Ten3 as a repulsive receptor, extracellular binding of Lphn2 to Ten3 should also trigger a repulsive response in Ten3+ axon terminals, but the signaling mechanism is completely unknown. Future studies on the mechanisms that disassemble the extracellular complex and intracellular signaling in the axon down- stream of Ten3 will increase our understanding of how the interaction of these two molecules can lead to repulsive outcomes. Materials and methods Key resources table Reagent type (species) or resource Designation Source or reference Identifiers Additional information Cell line (Homo sapiens) HEK293T (epithelial, kidney) ATCC RRID:CVCL_0063 Cell line (Homo sapiens) HEKΔ7 Alvarez- Curto et al., 2016, PMCID:PMC5207144 JBC HEK293 cells with targeted deletion via CRISPR- Cas9 of GNAS, GNAL, GNAQ, GNA11, GNA12, GNA13, and GNAZ Antibody Antibody anti- FLAG (rabbit polyclonal) ThermoFisher, PA1- 984B RRID:AB_347227 IB: 1:500 anti- rabbit HRP (donkey polyclonal) ThermoFisher, Cat #31458 RRID:AB_228213 IB: 1:10,000 Recombinant DNA reagent SRE- luciferase/glo (plasmid) Nazarko et al., 2018, PMCID:PMC6137404 iScience Recombinant DNA reagent Lphn2- Flag (plasmid) This paper pCDNA3.1 with Kozak (GCC) and C- terminal Flag tag Recombinant DNA reagent Lphn2- CTF- Flag (plasmid) This paper Recombinant DNA reagent SP- P2Y12- EK- Lphn2- CTF (plasmid) This paper Recombinant DNA reagent Lphn2- F831A/M835A- Flag (plasmid) This paper Continued on next page pCDNA3.1 with Kozak (GCC), Lphn2 C- terminal fragment and C- terminal Flag Enterokinase cleavage site based on Lizano et al., 2021, Lphn2- CTF pCDNA3.1 with kozak (GCC), Lphn2, F831A and M835A, and C- terminal Flag Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 14 of 23 Developmental Biology | Neuroscience Research article Continued Reagent type (species) or resource Designation Source or reference Identifiers Additional information pCDNA3.1 with kozak (GCC), Lphn2, F831A and M835A C- terminal fragment, and C- terminal Flag Enterokinase cleavage site based on Lizano et al., 2021, Lphn2- F831A/M835A- CTF pCDNA3.1 with kozak (GCC), Lphn2, T829G, and C- terminal Flag pCDNA3.1 with kozak (GCC), Lphn2, T829G C- terminal fragment, and C- terminal Flag Enterokinase cleavage site based on Lizano et al., 2021, Lphn2- T829G- CTF Recombinant DNA reagent Lphn2- F831A/M835A- CTF- Flag (plasmid) This paper Recombinant DNA reagent SP- P2Y12- EK- Lphn2- F831A/M835A- CTF (plasmid) This paper Recombinant DNA reagent Lphn2- T829G- Flag (plasmid) This paper Recombinant DNA reagent Lphn2- T829G- CTF- Flag (plasmid) This paper Recombinant DNA reagent SP- P2Y12- EK- Lphn2- T829G- CTF (plasmid) This paper Recombinant DNA reagent Gαs/Gαi/ Gαq/ Gα12/Gα13 (plasmid) Recombinant DNA reagent Gβ1 Recombinant DNA reagent Gγ2- Venus Recombinant DNA reagent GRK3ct- Rluc8 Hollins et al., 2009, PMCID:PMC2668204 Hollins et al., 2009, PMCID:PMC2668204 Hollins et al., 2009, PMCID:PMC2668204 Hollins et al., 2009, PMCID:PMC2668204 Chemical compound, drug Firefly D- luciferin NanoLight Technology Chemical compound, drug Coelenterazine- h NanoLight Technology Cell Signal Cell Signal Cell Signal Cell Signal Cat #306 Cat #301 Chemical compound, drug YM- 254890 AdipoGene Life Sciences CAS 568580- 02- 9 Chemical compound, drug Janelia Fluor 646 Genetic reagent (Mus musculus) CD- 1 Lavis Lab, Howard Hughes Medical Institute Janelia Research Campus Charles River Laboratory CD- 1 mice were used for all animal experiments Recombinant DNA reagent Lentiviral UbC- Cre This paper Lentivirus co- expressing Cre for in vivo experiments Recombinant DNA reagent Lentiviral UbC- Cre- P2A- Lphn2- FLAG This paper Recombinant DNA reagent Lentiviral UbC- Cre- P2A Lphn2_F831A/ M835A- FLAG This paper Recombinant DNA reagent Lentiviral UbC- Cre- P2A- Lphn2_T828G- FLAG This paper Recombinant DNA reagent Lentiviral UbC- GFP Pederick et al., 2021, PMCID:PMC8830376 Pederick et al., 2021, PMCID:PMC8830376 Recombinant DNA reagent Lentiviral UbC- GFP- P2A- Lphn2- FLAG Science Recombinant DNA reagent Lentiviral UbC- GFP- P2A- Lphn2_F831A/ M835A- FLAG This paper Recombinant DNA reagent Lentiviral UbC- GFP- P2A- Lphn2_T828G- FLAG This paper Recombinant DNA reagent AAV8- EF1a- DIO- ChR2- mCh Addgene plasmid 20297 Lentivirus co- expressing Cre- P2A- Lphn2- FLAG for in vivo experiments Lentivirus co- expressing Cre- P2A- Lphn2_F831A/M835A- FLAG for in vivo experiments Lentivirus co- expressing Cre- P2A- Lphn2_T829G- FLAG for in vivo experiments Science Lentivirus co- expressing GFP for in vivo experiments Lentivirus co- expressing GFP- P2A- Lphn2- FLAG for in vivo experiments Lentivirus co- expressing GFP- P2A- Lphn2_F831A/M835A- FLAG for in vivo experiments Lentivirus co- expressing GFP- P2A- Lphn2_F831A/M835A- FLAG for in vivo experiments AAV used to label axons of Cre expressing neurons with mCherry Recombinant DNA reagent AAV8- CaMKIIa- ChR2- mCh Addgene plasmid26975 AAV used to label axons of neurons with mCherry Chemical compound, drug DAPI ThermoFisher D1306 1:10,000 Antibody anti- mCherry (rat monoclonal) ThermoFisher, M11217 RRID:AB_2536611 Antibody anti- Cre (Rabbit polyclonal) Synaptic Systems, 257 003, RRID:AB_2619968 Immunohistochemistry 1:1,000 Immunohistochemistry 1:500 Immunohistochemistry 1:2,500 Antibody Antibody anti- GFP (chicken polyclonal) Aves Labs, GFP- 1020 RRID:AB_10000240 anti- FLAG (goat polyclonal) Abcam, ab95045 RRID:AB_10676074 Immunohistochemistry 1:3000 Antibody anti- Lphn2 (Rabbit polyclonal) Novus Biologicals, nbp2- 58704 Immunohistochemistry 1:500 Continued on next page Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 15 of 23 Developmental Biology | Neuroscience Research article Continued Reagent type (species) or resource Designation Source or reference Identifiers Additional information Software, algorithm Zen Zeiss Software, algorithm ImageJ National Institutes of Health Software, algorithm Adobe Photoshop Software, algorithm Adobe Illustrator Software, algorithm Matlab Adobe Adobe MathWorks Software, algorithm GraphPad Prism 9 GraphPad Software Previously existing Previously existing Previously existing Previously existing Previously existing Previously existing Materials for cell culture experiments Dulbecco’s Modified Eagle Medium (DMEM), high glucose, and penicillin- streptomycin (P/S) (10,000  U/mL) were purchased from Gibco (ThermoFisher Scientific, Waltham, MA). Fetal bovine serum (FBS), 0.5% trypsin, and Dulbecco’s phosphate- buffered saline (DPBS) were purchased from Corning (Fisher Scientific, Waltham, MA). Opti- MEM reduced serum medium, no phenol red, and Lipofectamine 2000 transfection reagent was purchased from Invitrogen (ThermoFisher Scientific). FuGENE transfection reagent was purchased from Promega (Madison, WI). RIPA buffer was purchased from Sigma- Aldrich (St. Louis, MO). Triton lysis buffer consisted of 0.11 M Tris- HCl powder, 0.04 M Tris- base powder, 75  mM NaCl, 3  mM MgCl2, and 0.25% Triton X- 100 pure liquid. The 3  X Firefly Assay Buffer was freshly prepared in Triton lysis buffer and contained 15 mM DTT, 0.6 mM coenzyme A (MedChemExpress, Monmouth Junction, NJ), 0.45 mM ATP (MedChemExpress, Monmouth Junc- tion, NJ), and 0.42 mg/mL firefly D- luciferin (NanoLight Technology). Renilla Salts buffer consisted of 45 mM Na2EDTA, 30 mM Na Pyrophosphate, and 1.425 M NaCl. The 3 X Renilla Assay Buffer was freshly prepared in Renilla Salts and contained 0.06 mM PTC124 in DMSO (MedChemExpress) and 0.01 mM coelenterazine- h (NanoLight Technologies, Pinetop, AZ). For the BRET assays, enterokinase, light chain, was obtained from New England Biolabs (Ipswich, MA), isoproterenol and quinpirole from Sigma Aldrich, and endothelin 1 (ET- 1) from Tocris Bioscience (Bristol, United Kingdom). YM- 254890 was purchased from AdipoGen Life Sciences (San Diego, CA). Impermeant Janelia Fluor 646 conju- gated to benzyl guanine was a kind gift from Dr. Luke Lavis (Howard Hughes Medical Institute Janelia Research Campus). Plasmid DNA constructs Lphn2 was amplified from cDNA isolated from the P8 mouse hippocampus. Sanger sequencing confirmed that exons 19 and 20 were excluded from the amplified Lphn2 (Refer to NCBI Reference Sequence: NM_001081298.2 for exon annotation). This cDNA was used as a polymerase chain reac- tion (PCR) template to make the Lphn2 constructs used in this study. All cDNA constructs were assem- bled in a pCDNA3.1+ vector by Gibson assembly using NEBuilder HiFi DNA Assembly Master Mix (New England Biolabs). Sequences were confirmed with the Genewiz sequencing service (South Plain- field, NJ). Plasmid DNA constructs are available upon request. Cell culture HEK293T cells (American Type Culture Collection, Manassas, VA; RRID:CVCL_0063) and HEK293 cells with targeted deletion via CRISPR- Cas9 of GNAS, GNAL, GNAQ, GNA11, GNA12, GNA13, and GNAZ (HEKΔ7) (Alvarez- Curto et al., 2016) were maintained in high- glucose DMEM supplemented with 10% FBS and 1% P/S at 37 °C in a 5% CO2 humidified incubator. Cell authentication was not performed as the cells were obtained directly from the supplier. PCR- based mycoplasma testing was performed routinely using ATCC mycoplasma testing services. Cell viability was assessed for each passage using the Countess II automated cell counter (ThermoFisher Scientific). Immunoblot analysis HEK293T cells were detached for 2–3  min using 0.5% trypsin and then plated at a density of 350,000 cells/mL in a six- well culture plate. After 24 hr, the cells were transfected using FuGENE trans- fection reagent (8 μL/2 μg cDNA) and Opti- Mem with receptor cDNA (2 μg). After 24 hr, cells were placed on ice and incubated in 500 μL RIPA buffer for 30 min. Following this incubation, cells were Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 16 of 23 Developmental Biology | Neuroscience Research article scraped from the culture plate and moved into 1.5 mL microcentrifuge tubes. Cells were then spun at 15,000 × g in a 4 °C benchtop centrifuge to pellet debris. After centrifugation, 50 μL of the super- natant was transferred into a fresh microcentrifuge tube and combined with 50 μL 2 X SDS Laemmli sample buffer (Sigma- Aldrich). In preparation for immunoblot analysis, a 20 μL sample was run on an SDS- PAGE gel (Mini- PROTEAN TGX, 4–15%, Bio- Rad Laboratories, Inc, Hercules, CA) prior to transfer to a PDVF membrane (Immobilon- P Membrane, Merck Millipore Ltd., Burlington, MA). The membrane was then incubated in a 5% milk tris- buffered saline with 0.1% tween- 20 (TBS- T) solution for 1 hr at RT with gentle rotation. The membrane was washed five times 5 min in TBS- T prior to overnight incuba- tion at 4 °C with 1° anti- FLAG antibody (1:500, ThermoFisher, PA1- 984B; RRID:AB_347227). The next morning, the membrane was washed five times 5 min in TBS- T. The membrane was then incubated for 1 hr at RT with 2° anti- rabbit HRP antibody (1:10,000, ThermoFisher, Cat #31458;RRID:AB_228213). The membrane was washed five times 5  min in TBS- T prior to visualization with SuperSignal West Pico Chemiluminescent Substrate (Fisher Scientific) using the Azure c600 Gel Imaging System (Azure Biosystems, Dublin, CA). Gene expression assays In preparation for transfection, HEK293T cells were detached for 2–3 min using 0.5% trypsin and then seeded at a density of 400,000 cells/mL in a 12- well culture plate. After 24 hr, the cells were co- trans- fected using Lipofectamine 2000 (2.5 μL/1 μg cDNA) and Opti- Mem with receptor cDNA (10–600 ng), gene reporter cDNA (600 ng), and empty vector pCDNA5/FRT to balance the total amount of cDNA up to 1200 ng. After 6 hr with the transfection reagent, the media was volume exchanged to serum- free DMEM supplemented with 1% P/S (~18 hr serum starvation). After 24 hr, the media was aspirated from the cells and each well was gently rinsed with DPBS. Cells were then mechanically detached using 275 μL DPBS and 80 μL of the resuspension was distributed in triplicate to a 96- well black/white isoplate (Perkin Elmer Life Sciences). Next, 40 μL of 3 X Firefly Assay Buffer was added to each well. The emission was then read at 535 nm after 10 min incubation using a PHERAstar FS microplate reader (BMG LABTECH, Ortenberg, Germany). Next, 60 μL 3 X Renilla Assay Buffer was added to each well. The emission was then read at 475 nm after 10 min incubation using a PHERAstar FS microplate reader. For assays using the Gαq- inhibitor YM- 254890, the cell media was exchanged to DMEM containing 1 μM YM- 254890 approximately 6 hr after transfection. Bioluminescence resonance energy transfer assays In preparation for transfection, HEKΔ7 cells were detached for 2–3 min using 0.5% trypsin and then seeded at a density of 400,000 cells/mL in a 12- well culture plate. After 24 hr, the cells were co- trans- fected using Lipofectamine 2000 (2.5 μL/1 μg cDNA) and Opti- Mem with receptor cDNA (200 ng), Gα (720 ng), Gβ1 (250 ng), Gγ2- Venus (250 ng), membrane- anchored GRK3ct- Rluc8 (50 ng), and empty vector pCDNA5/FRT to balance the total amount of cDNA up to 1470 ng. After 24 hr transfection, cells were washed with DPBS before being re- suspended in 400 μL BRET buffer (DPBS containing 5 mM glucose). Next, 45 μL of the resuspension was distributed to six wells of a 96- well OptiPlate black- white plate (Perkin Elmer Life Sciences, Waltham, MA). Cells were then incubated for 10 mins with 10 μL coelenterazine- h (final concentration 5 μM) before ligand addition to reach a final well volume of 100 μL. Donor (Rluc8) and acceptor (mVenus) emission was read using a PHERAstar FS microplate reader at 485 nm and 525 nm, respectively. The BRET ratio was then measured by dividing the 525 emissions by the 485 emissions. The drug- induced BRET ratio was then calculated by subtracting the buffer BRET for each condition. Surface expression measurements using SNAPfast-tag In preparation for transfection, HEK293T cells were detached for 2–3 min using 0.5% trypsin and then seeded at a density of 900,000 cells/well in a six- well culture plate. After 24 hr, the cells were trans- fected using FuGENE transfection reagent (8 μL/2 μg cDNA) and Opti- Mem with SNAPfast- tagged receptor cDNA (2  μg). After 24  hr, cells were incubated for 30  min with 500  µL 1  µM impermeant Janelia Fluor 646 conjugated to benzyl guanine was dissolved in DMEM containing 10% FBS and 1%  P/S. Cells were then washed three times with complete DMEM and once with DPBS prior to resuspension in 500 µL DPBS. Next, 100 μL of resuspension was added to three wells of a 96- well OptiPlate black plate (Perkin Elmer Life Sciences, Waltham, MA). The emission was then read using Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 17 of 23 Developmental Biology | Neuroscience Research article the filter 640/680 at a gain of 1000 using a PHERAstar FS microplate reader (BMG LABTECH, Orten- berg, Germany). Mice All procedures followed animal care and biosafety guidelines approved by Stanford University’s Admin- istrative Panel on Laboratory Animal Care (APLAC 14007) and Administrative Panel on Biosafety (APB- 3669- LL120) in accordance with NIH guidelines. Both male and female mice were used, and mice were group housed on a 12 hr light/dark cycle with access to food and water ad libitum. CD- 1 mice from Charles River Laboratories were used for all experiments. The total number of mice injected and screened for each experiment is as follows: Figure 1: LV- Cre, 101, LV- Cre- P2A- Lphn2- FLAG, 60; Figure  3: LV- Cre- P2A- Lphn2_F831A/M835A- FLAG, 86; Figure  4: LV- Cre- P2A- Lphn2_T829G- FLAG, 102; and Figure  5: LV- GFP- P2A- Lphn2_F831A/M835A- FLAG, 79 and LV- GFP- P2A- Lphn2_T829G- FLAG, 56. Lentivirus generation All lentivirus constructs expressing Cre, GFP, Lphn2, Lphn2_F831A/M835A, or Lphn2_T829G were made by inserting corresponding cDNA into the LV- UbC plasmid (Pederick et al., 2021) with a P2A sequence between the two ORFs. Cre and GFP were amplified from LV- UbC- GFP- Cre and full- length Lphn2 cDNA was isolated from a cDNA library made from mRNAs from the P8 mouse hippocampus. GFP and Lphn2 were inserted into LV- UbC with a Gibson assembly cloning kit (NEB E5510S). The Lphn2_F831A/M835A and Lphn2_T829G mutations were made using Q5 mutagenesis (NEB, E0552S). All plasmids were sequenced and verified before the virus was produced. All custom lentiviruses were generated by transfecting 36 10 cm plates (HEK293T) with four plasmids (4.1 R, RTR2, VSVg, and transfer vector containing gene of interest). Medium was collected 48 hr later and centrifuged at 8400 relative centrifugal force (rcf) for 18 hr at 4 °C. Viral pellets were dissolved with PBS and further purified with a 20% sucrose gradient centrifugation at 80,000 rcf for 2 hr. Stereotactic injections in neonatal mice P0 mice were anesthetized using hypothermia. CA1 injections were 1.0 mm lateral, 0.85 mm anterior, and 0.8  mm ventral from lambda and subiculum injections were 1.3  mm lateral, 0.45  mm anterior, and 0.8  mm ventral from lambda. 100  nl of lentivirus was injected at 100  nl/min at the following titers: LV- Cre (7 × 1012 copies per ml), LV- Cre- P2A- Lphn2- FLAG (2.4 × 1012 copies per ml), LV- Cre- P2A- Lphn2_F831A/M831A- FLAG (2.24 × 1012 copies per ml) and LV- Cre- P2A- Lphn2_T829G- FLAG (1.2 × 1013 copies per ml), LV- GFP (6 × 1012 copies per ml), LV- GFP- P2A- Lphn2- FLAG (5 × 1012 copies per ml), LV- GFP- P2A- Lphn2_F831A/M831A- FLAG (3.6 × 1012 copies per ml), LV- GFP- P2A- Lphn2_T829G- FLAG (9 × 1012 copies per ml). Stereotactic injection in adult mice Injections of AAV8- EF1a- DIO- ChR2- mCh (2 × 1012 copies per ml, Neuroscience Gene Vector and Virus core, Stanford University) and AAV8- CaMKIIa- ChR2- mCh (2 × 1012 copies per ml, Neuroscience Gene Vector and Virus core, Stanford University) were performed at about P42. Mice were anesthetized using isoflurane and mounted in stereotactic apparatus (Kopf). Coordinates for proximal CA1 were 1.4 mm lateral and 1.25 mm posterior from bregma, and 1.12 mm ventral from the brain surface. Virus was iontophoretically injected with current parameters 5 µA, 7 s on, 7 s off, for 2 min, using pipette tips with an outside perimeter of 10–15 μm. Mice were perfused about 2 weeks later and processed for immunostaining as described below. Immunostaining Mice were injected with 2.5% Avertin and were transcardially perfused with PBS followed by 4% paraformaldehyde (PFA). Brains were dissected and post- fixed in 4% PFA overnight, and cryopro- tected for about 24 hr in 30% sucrose. Brains were embedded in Optimum Cutting Temperature (OCT, Tissue- Tek), frozen in dry ice- cooled isopentane bath, and stored at –80 °C until sectioned. 60 μm thick floating sections were collected in PBS +0.02% sodium azide and stored at 4 °C. Sections were incu- bated in the following solutions at room temperature unless indicated: 1 hr in 0.3% PBS/Triton X- 100 and 10% normal donkey serum, two nights in the primary antibody at 4 °C in 0.3% PBS/Triton X- 100 Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 18 of 23 Developmental Biology | Neuroscience Research article and 10% normal donkey serum, 3 × 15 min in 0.3% PBS/Triton X- 100, overnight in secondary anti- body +DAPI (1:10,000 of 5 mg/ml, Sigma- Aldrich) in 0.3% PBS/Triton X- 100 and 10% normal donkey serum, 2 × 15 min in 0.3% PBS/Triton X- 100, and 15 min in PBS. Sections were mounted with Fluoro- mount- G (SouthernBiotech). Primary antibodies used were rat anti- mCherry (1:1000, ThermoFisher, M11217, RRID:AB_2536611), rabbit anti- Cre (1:500, Synaptic Systems, 257 003, RRID:AB_2619968), chicken anti- GFP (1:2500, Aves Labs, GFP- 1020, RRID:AB_10000240) goat anti- FLAG (1:3000, Abcam, ab95045, RRID:AB_10676074) and rabbit anti- Lphn2 (1:500, Novus Biologicals, NBP2- 58704). Secondary antibodies conjugated to Alexa 488, Alexa 568, or Cy3 (Jackson ImmunoResearch) were used at 1:500 from 50% glycerol stocks. Image and data analysis for CA1 axon tracing Mice were only included if they passed the following criteria: (1) AAV injection site must be in prox- imal CA1 (most proximal 30%), (2) lentivirus injections sites must be in CA1 and not in the subiculum (Figures 1, 3 and 4) or lentivirus injections must be in the distal subiculum (Figure 5), (3) proximal CA1 axons must overlap with lentivirus injection site in the subiculum (Figure 5). All mice that fulfilled these criteria are reported in Figures 1 and 3–5 and were included in quantifications. Images of injection sites (5x magnification) and projections (10x magnification) were acquired for every other 60 μm sagittal section using a Zeiss epifluorescence scope. Due to variations in injection sites within each mouse, exposure was adjusted for each mouse to avoid saturation. Fluorescence intensity measurements on unprocessed images were taken using FIJI and data processing was performed using MATLAB. For injection site quantification, a 30- pixel- wide segmented line was drawn from proximal CA1 to distal CA1 using the DAPI signal as a guide. For projection quantification in the subiculum, a 200- pixel- wide segmented line was drawn from the proximal subiculum to the distal subiculum through the cell body layer using only DAPI as a guide. From this point, injection site and projection images were processed the same. Segmented lines were straightened using the ‘Straighten’ func- tion, background subtraction was performed using the ‘Subtract’ function and intensity values were measured using the ‘Plot Profile’ command (FIJI). For injections that labeled both CA2 and proximal CA1, CA2 axons were present near the distal border of CA1 and spilled into the proximal subiculum. These axons had their intensity set to zero by using area selection and the clear function (FIJI). The intensity plots were resampled into 100 equal bins using a custom MATLAB code. For trace quantification in Figures 1, 3 and 4 the axon intensity was combined for all sections by summing all intensity values at each binned position. To calculate the fraction of axon intensity across proximal, mid, and distal subiculum the total axon intensity in bins 1–33, 34–66, and 67–100 were summed, respectively. The summed value from each of these regions was then divided by the total sum of axons from bins 1–100 to obtain the fraction of axons intensity within the proximal, mid, and distal subiculum. Fractions of axon intensities were compared using a two- way ANOVA with Sidak’s multiple comparisons tests using Prism 9 (GraphPad). The mean position of the injection sites was calculated by generating a summed intensity trace as above and then multiplying the intensity value by the bin position, summing across the entire axis, and dividing by the sum of the intensity values. Representative images (Figures 1, 3 and 4) were taken using a Zeiss LSM 780 confocal microscope (20×magnification, tile scan, max projection). In Figure  5, the experimental and data analysis procedures were identical to Pederick et  al., 2021 and, therefore, we used LV- GFP and LV- GFP- P2A- Lphn2- FLAG from that study to compare with LV- GFP- P2A- Lphn2_F831A/M831A- FLAG and LV- GFP- P2A- Lphn2_T828G- FLAG data generated from this study. To quantify average axon intensity in GFP+ and adjacent GFP– regions in subiculum targets (Figure  5F), we restricted the analysis to the most distal 20% of the subiculum. To determine the GFP+ region we identified the intensity- weighted central row using the summed fluorescence of each row and determined the minimal symmetric window of rows around the central row that encompassed at least 50% of the total intensity in the restricted GFP image. This defined a rectangle in the original image that we designated as the GFP+ region. We then computed the mean fluorescence intensity in this region for the mCh channel. We used the two rows above and below (lateral and medial) the designated GFP+ region as the adjacent GFP– region and computed the mean mCh fluorescence across these four rows. To determine mCh fluorescence differences in GFP+ versus GFP– regions, we divided the mCh intensity in the GFP+ region by the mCh intensity in the GFP– region for each mouse Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 19 of 23 Developmental Biology | Neuroscience Research article (i.e. GFP+/GFP–). mCh fluorescence intensity GFP+/GFP– was compared across groups using a one- way ANOVA with Tukey’s multiple comparisons tests using Prism 9 (GraphPad). Three- dimensional mountain plots were generated using the ‘surf’ function. Quantification of overexpressed Lphn2 in CA1 P0 neonatal injections into CA1 and immunostaining were performed as stated above. 16- bit images were acquired with a Zeiss Axio Imager Z2 confocal microscope. All conditions were imaged with equal settings. For mean intensity quantification of FLAG and GFP a 50- and 20- pixel wide segmented line was drawn through the molecular layer and cell body layer of CA1, respectively, using the DAPI signal as a guide. The mean intensity of FLAG and GFP was calculated using the ‘Measure’ function. From this point, only the section with the highest mean GFP expression was used from each animal. The FLAG mean intensity value was divided by the GFP mean intensity value to determine the FLAG/ GFP mean intensity. FLAG/GFP mean intensities across conditions were compared using a one- way analysis of variance (ANOVA) with Tukey’s multiple comparisons test in Prism 9 (GraphPad). Replicates Technical replicates are defined here as repeated measurements of the same sample while biological replicates are defined as measurements of biologically distinct samples. Acknowledgements We thank Z Li, T Li, C McLaughlin, D Wang, and Y Wu for critiques of the manuscript. We are also grateful to A Inoue (Tohoku University, Japan) for the generous gift of the HEKΔ7 cell line and Dr. Luke Lavis (Janelia Research Campus) for supplying the JF646 dye. This work was supported by NIH grants T32- MH015144 (NAPH), R01- NS050835 (LL), R01- MH54137 (JAJ), the Hope for Depression Research Foundation (JAJ), and P30EY012196 (ZH). LL is an investigator at the Howard Hughes Medical Institute. Additional information Funding Funder National Institutes of Health National Institutes of Health National Institutes of Health Hope for Depression Research Foundation National Institutes of Health Grant reference number Author T32-MH015144 Nicole A Perry-Hauser R01-NS050835 Liqun Luo R01-MH54137 Jonathan A Javitch Jonathan A Javitch P30EY012196 Zhigang He The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Author contributions Daniel T Pederick, Conceptualization, Resources, Data curation, Formal analysis, Validation, Inves- tigation, Visualization, Writing – original draft, Writing – review and editing; Nicole A Perry- Hauser, Conceptualization, Resources, Data curation, Formal analysis, Funding acquisition, Validation, Inves- tigation, Visualization, Writing – original draft, Writing – review and editing; Huyan Meng, Resources, Writing – review and editing; Zhigang He, Resources, Funding acquisition, Writing – review and editing; Jonathan A Javitch, Conceptualization, Resources, Data curation, Supervision, Funding acqui- sition, Writing – original draft, Writing – review and editing; Liqun Luo, Conceptualization, Resources, Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 20 of 23 Developmental Biology | Neuroscience Research article Data curation, Supervision, Funding acquisition, Validation, Investigation, Visualization, Writing – orig- inal draft, Project administration, Writing – review and editing Author ORCIDs Daniel T Pederick Nicole A Perry- Hauser Huyan Meng Jonathan A Javitch Liqun Luo http://orcid.org/0000-0003-1870-9475 http://orcid.org/0000-0003-3130-3023 http://orcid.org/0000-0003-1511-6156 http://orcid.org/0000-0001-7395-2967 http://orcid.org/0000-0001-5467-9264 Ethics All procedures followed animal care and biosafety guidelines approved by Stanford University's Administrative Panel on Laboratory Animal Care (APLAC 14007) and Administrative Panel on Biosafety (APB- 3669- LL120) in accordance with NIH guidelines. Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.83529.sa1 Author response https://doi.org/10.7554/eLife.83529.sa2 Additional files Supplementary files •  MDAR checklist Data availability All materials are available through requests to the corresponding authors. All custom code was iden- tical to that reported in Pederick et  al., 2021 and can be accessed at https://github.com/dped- erick/Reciprocal-repulsions-instruct-the-precise-assembly-of-parallel-hippocampal-networks/tree/1 (Pederick, 2021). All data generated or analyzed during this study are included in the manuscript and supporting file. Source data files have been provided for all figures. References Alvarez- Curto E, Inoue A, Jenkins L, Raihan SZ, Prihandoko R, Tobin AB, Milligan G. 2016. Targeted elimination of G proteins and arrestins defines their specific contributions to both intensity and duration of G protein- coupled receptor signaling. The Journal of Biological Chemistry 291:27147–27159. DOI: https://doi.org/10. 1074/jbc.M116.754887, PMID: 27852822 Anderson GR, Maxeiner S, Sando R, Tsetsenis T, Malenka RC, Südhof TC. 2017. Postsynaptic adhesion GPCR latrophilin- 2 mediates target recognition in entorhinal- hippocampal synapse assembly. The Journal of Cell Biology 216:3831–3846. DOI: https://doi.org/10.1083/jcb.201703042, PMID: 28972101 Araç D, Boucard AA, Bolliger MF, Nguyen J, Soltis SM, Südhof TC, Brunger AT. 2012. A novel evolutionarily conserved domain of cell- adhesion GPCRs mediates autoproteolysis. The EMBO Journal 31:1364–1378. DOI: https://doi.org/10.1038/emboj.2012.26, PMID: 22333914 Arcos- Burgos M, Jain M, Acosta MT, Shively S, Stanescu H, Wallis D, Domené S, Vélez JI, Karkera JD, Balog J, Berg K, Kleta R, Gahl WA, Roessler E, Long R, Lie J, Pineda D, Londoño AC, Palacio JD, Arbelaez A, et al. 2010. A common variant of the latrophilin 3 gene, LPHN3, confers susceptibility to ADHD and predicts effectiveness of stimulant medication. Molecular Psychiatry 15:1053–1066. DOI: https://doi.org/10.1038/mp. 2010.6, PMID: 20157310 Barros-Álvarez X, Nwokonko RM, Vizurraga A, Matzov D, He F, Papasergi- Scott MM, Robertson MJ, Panova O, Yardeni EH, Seven AB, Kwarcinski FE, Su H, Peroto MC, Meyerowitz JG, Shalev- Benami M, Tall GG, Skiniotis G. 2022. The tethered peptide activation mechanism of adhesion GPCRs. Nature 604:757–762. DOI: https://doi. org/10.1038/s41586-022-04575-7, PMID: 35418682 Beliu G, Altrichter S, Guixà-González R, Hemberger M, Brauer I, Dahse AK, Scholz N, Wieduwild R, Kuhlemann A, Batebi H, Seufert F, Pérez- Hernández G, Hildebrand PW, Sauer M, Langenhan T. 2021. Tethered agonist exposure in intact adhesion/class B2 gpcrs through intrinsic structural flexibility of the gain domain. Molecular Cell 81:905–921. DOI: https://doi.org/10.1016/j.molcel.2020.12.042, PMID: 33497605 Berns DS, DeNardo LA, Pederick DT, Luo L. 2018. Teneurin- 3 controls topographic circuit assembly in the hippocampus. Nature 554:328–333. DOI: https://doi.org/10.1038/nature25463, PMID: 29414938 Boucard AA, Ko J, Südhof TC. 2012. High affinity neurexin binding to cell adhesion G- protein- coupled receptor CIRL1/latrophilin- 1 produces an intercellular adhesion complex. The Journal of Biological Chemistry 287:9399– 9413. DOI: https://doi.org/10.1074/jbc.M111.318659, PMID: 22262843 Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 21 of 23 Developmental Biology | Neuroscience Research article Boucard AA, Maxeiner S, Südhof TC. 2014. Latrophilins function as heterophilic cell- adhesion molecules by binding to teneurins: regulation by alternative splicing. The Journal of Biological Chemistry 289:387–402. DOI: https://doi.org/10.1074/jbc.M113.504779, PMID: 24273166 Davletov BA, Shamotienko OG, Lelianova VG, Grishin EV, Ushkaryov YA. 1996. Isolation and biochemical characterization of a Ca2+-independent alpha- latrotoxin- binding protein. The Journal of Biological Chemistry 271:23239–23245. DOI: https://doi.org/10.1074/jbc.271.38.23239, PMID: 8798521 Del Toro D, Carrasquero- Ordaz MA, Chu A, Ruff T, Shahin M, Jackson VA, Chavent M, Berbeira- Santana M, Seyit- Bremer G, Brignani S, Kaufmann R, Lowe E, Klein R, Seiradake E. 2020. Structural basis of teneurin- latrophilin interaction in repulsive guidance of migrating neurons. Cell 180:323–339.. DOI: https://doi.org/10. 1016/j.cell.2019.12.014, PMID: 31928845 Domené S, Stanescu H, Wallis D, Tinloy B, Pineda DE, Kleta R, Arcos- Burgos M, Roessler E, Muenke M. 2011. Screening of human LPHN3 for variants with a potential impact on ADHD susceptibility. American Journal of Medical Genetics. Part B, Neuropsychiatric Genetics 156B:11–18. DOI: https://doi.org/10.1002/ajmg.b.31141, PMID: 21184580 Donohue JD, Amidon RF, Murphy TR, Wong AJ, Liu ED, Saab L, King AJ, Pae H, Ajayi MT, Anderson GR. 2021. Parahippocampal latrophilin- 2 (ADGRL2) expression controls topographical presubiculum to entorhinal cortex circuit connectivity. Cell Reports 37:110031. DOI: https://doi.org/10.1016/j.celrep.2021.110031, PMID: 34818557 Egea J, Klein R. 2007. Bidirectional Eph- ephrin signaling during axon guidance. Trends in Cell Biology 17:230– 238. DOI: https://doi.org/10.1016/j.tcb.2007.03.004, PMID: 17420126 Hattori M, Osterfield M, Flanagan JG. 2000. Regulated cleavage of a contact- mediated axon repellent. Science 289:1360–1365. DOI: https://doi.org/10.1126/science.289.5483.1360, PMID: 10958785 Hollins B, Kuravi S, Digby GJ, Lambert NA. 2009. The C- terminus of GRK3 indicates rapid dissociation of G protein heterotrimers. Cellular Signalling 21:1015–1021. DOI: https://doi.org/10.1016/j.cellsig.2009.02.017, PMID: 19258039 Kishore A, Purcell RH, Nassiri- Toosi Z, Hall RA. 2016. Stalk- dependent and stalk- independent signaling by the adhesion G protein- coupled receptors GPR56 (ADGRG1) and BAI1 (ADGRB1). The Journal of Biological Chemistry 291:3385–3394. DOI: https://doi.org/10.1074/jbc.M115.689349, PMID: 26710850 Kordon SP, Dutka P, Adamska JM, Bandekar SJ, Leon K, Erramilli SK, Adams B, Li J, Kossiakoff AA, Araç D. 2023. Isoform- and ligand- specific modulation of the adhesion GPCR ADGRL3/latrophilin3 by a synthetic binder. Nature Communications 14:635. DOI: https://doi.org/10.1038/s41467-023-36312-7, PMID: 36746957 Kozasa T, Hajicek N, Chow CR, Suzuki N. 2011. Signalling mechanisms of RhoGTPase regulation by the heterotrimeric G proteins G12 and G13. Journal of Biochemistry 150:357–369. DOI: https://doi.org/10.1093/ jb/mvr105, PMID: 21873336 Lelianova VG, Davletov BA, Sterling A, Rahman MA, Grishin EV, Totty NF, Ushkaryov YA. 1997. Alpha- Latrotoxin receptor, latrophilin, is a novel member of the secretin family of G protein- coupled receptors. The Journal of Biological Chemistry 272:21504–21508. DOI: https://doi.org/10.1074/jbc.272.34.21504, PMID: 9261169 Leon K, Cunningham RL, Riback JA, Feldman E, Li J, Sosnick TR, Zhao M, Monk KR, Araç D. 2020. Structural basis for adhesion G protein- coupled receptor gpr126 function. Nature Communications 11:194. DOI: https:// doi.org/10.1038/s41467-019-14040-1, PMID: 31924782 Liebscher I, Schön J, Petersen SC, Fischer L, Auerbach N, Demberg LM, Mogha A, Cöster M, Simon KU, Rothemund S, Monk KR, Schöneberg T. 2014. A tethered agonist within the ectodomain activates the adhesion G protein- coupled receptors GPR126 and GPR133. Cell Reports 9:2018–2026. DOI: https://doi.org/10.1016/j. celrep.2014.11.036, PMID: 25533341 Liebscher I, Schöneberg T. 2016. Tethered agonism: A common activation mechanism of adhesion gpcrs. Experimental Pharmacology 234:111–125. DOI: https://doi.org/10.1007/978-3-319-41523-9_6, PMID: 27832486 Liebscher I, Schöneberg T, Thor D. 2022. Stachel- mediated activation of adhesion G protein- coupled receptors: insights from cryo- EM studies. Signal Transduction and Targeted Therapy 7:227. DOI: https://doi.org/10.1038/ s41392-022-01083-y, PMID: 35810167 Lizano E, Hayes JL, Willard FS. 2021. A synthetic method to assay adhesion- family G- protein coupled receptors. Determination of the G- protein coupling profile of ADGRG6(GPR126). Biochemical and Biophysical Research Communications 534:317–322. DOI: https://doi.org/10.1016/j.bbrc.2020.11.086, PMID: 33248691 Luo L. 2002. Actin cytoskeleton regulation in neuronal morphogenesis and structural plasticity. Annual Review of Cell and Developmental Biology 18:601–635. DOI: https://doi.org/10.1146/annurev.cellbio.18.031802.150501, PMID: 12142283 Markovic D, Challiss RAJ. 2009. Alternative splicing of G protein- coupled receptors: physiology and pathophysiology. Cellular and Molecular Life Sciences 66:3337–3352. DOI: https://doi.org/10.1007/s00018- 009-0093-4, PMID: 19629391 Mathiasen S, Palmisano T, Perry NA, Stoveken HM, Vizurraga A, McEwen DP, Okashah N, Langenhan T, Inoue A, Lambert NA, Tall GG, Javitch JA. 2020. G12/13 is activated by acute tethered agonist exposure in the adhesion GPCR ADGRL3. Nature Chemical Biology 16:1343–1350. DOI: https://doi.org/10.1038/s41589-020-0617-7, PMID: 32778842 Moreno- Salinas AL, Avila- Zozaya M, Ugalde- Silva P, Hernández- Guzmán DA, Missirlis F, Boucard AA. 2019. Latrophilins: a neuro- centric view of an evolutionary conserved adhesion G protein- coupled receptor subfamily. Frontiers in Neuroscience 13:700. DOI: https://doi.org/10.3389/fnins.2019.00700, PMID: 31354411 Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 22 of 23 Developmental Biology | Neuroscience Research article Nazarko O, Kibrom A, Winkler J, Leon K, Stoveken H, Salzman G, Merdas K, Lu Y, Narkhede P, Tall G, Prömel S, Araç D. 2018. A comprehensive mutagenesis screen of the adhesion GPCR latrophilin- 1/ADGRL1. IScience 3:264–278. DOI: https://doi.org/10.1016/j.isci.2018.04.019, PMID: 30428326 O’Sullivan ML, de Wit J, Savas JN, Comoletti D, Otto- Hitt S, Yates JR, Ghosh A. 2012. FLRT proteins are endogenous latrophilin ligands and regulate excitatory synapse development. Neuron 73:903–910. DOI: https://doi.org/10.1016/j.neuron.2012.01.018, PMID: 22405201 Pederick DT. 2021. Reciprocal- repulsions- instruct- the- precise- assembly- of- parallel- hippocampal- networks. 1f862bd. GitHub. https://github.com/dpederick/Reciprocal-repulsions-instruct-the-precise-assembly-of- parallel-hippocampal-networks/tree/1 Pederick DT, Lui JH, Gingrich EC, Xu C, Wagner MJ, Liu Y, He Z, Quake SR, Luo L. 2021. Reciprocal repulsions instruct the precise assembly of parallel hippocampal networks. Science 372:1068–1073. DOI: https://doi.org/ 10.1126/science.abg1774, PMID: 34083484 Perry- Hauser NA, VanDyck MW, Lee KH, Shi L, Javitch JA. 2022. Disentangling autoproteolytic cleavage from tethered agonist- dependent activation of the adhesion receptor ADGRL3. The Journal of Biological Chemistry 298:102594. DOI: https://doi.org/10.1016/j.jbc.2022.102594, PMID: 36244455 Ping Y- Q, Xiao P, Yang F, Zhao R- J, Guo S- C, Yan X, Wu X, Zhang C, Lu Y, Zhao F, Zhou F, Xi Y- T, Yin W, Liu F- Z, He D- F, Zhang D- L, Zhu Z- L, Jiang Y, Du L, Feng S- Q, et al. 2022. Structural basis for the tethered peptide activation of adhesion GPCRs. Nature 604:763–770. DOI: https://doi.org/10.1038/s41586-022-04619-y, PMID: 35418678 Qu X, Qiu N, Wang M, Zhang B, Du J, Zhong Z, Xu W, Chu X, Ma L, Yi C, Han S, Shui W, Zhao Q, Wu B. 2022. Structural basis of tethered agonism of the adhesion GPCRs ADGRD1 and ADGRF1. Nature 604:779–785. DOI: https://doi.org/10.1038/s41586-022-04580-w, PMID: 35418679 Röthe J, Thor D, Winkler J, Knierim AB, Binder C, Huth S, Kraft R, Rothemund S, Schöneberg T, Prömel S. 2019. Involvement of the adhesion gpcrs latrophilins in the regulation of insulin release. Cell Reports 26:1573–1584. DOI: https://doi.org/10.1016/j.celrep.2019.01.040, PMID: 30726739 Salzman GS, Ackerman SD, Ding C, Koide A, Leon K, Luo R, Stoveken HM, Fernandez CG, Tall GG, Piao X, Monk KR, Koide S, Araç D. 2016. Structural basis for regulation of GPR56/ADGRG1 by its alternatively spliced extracellular domains. Neuron 91:1292–1304. DOI: https://doi.org/10.1016/j.neuron.2016.08.022, PMID: 27657451 Sando R, Jiang X, Südhof TC. 2019. Latrophilin gpcrs direct synapse specificity by coincident binding of flrts and teneurins. Science 363:eaav7969. DOI: https://doi.org/10.1126/science.aav7969, PMID: 30792275 Sando R, Südhof TC. 2021. Latrophilin GPCR signaling mediates synapse formation. eLife 10:e65717. DOI: https://doi.org/10.7554/eLife.65717, PMID: 33646123 Silva JP, Lelianova VG, Ermolyuk YS, Vysokov N, Hitchen PG, Berninghausen O, Rahman MA, Zangrandi A, Fidalgo S, Tonevitsky AG, Dell A, Volynski KE, Ushkaryov YA. 2011. Latrophilin 1 and its endogenous ligand lasso/teneurin- 2 form a high- affinity transsynaptic receptor pair with signaling capabilities. PNAS 108:12113– 12118. DOI: https://doi.org/10.1073/pnas.1019434108, PMID: 21724987 Sita LV, Diniz GB, Horta- Junior JAC, Casatti CA, Bittencourt JC. 2019. Nomenclature and comparative morphology of the teneurin/tcap/adgrl protein families. Frontiers in Neuroscience 13:425. DOI: https://doi.org/ 10.3389/fnins.2019.00425, PMID: 31130838 Spillane M, Gallo G. 2014. Involvement of Rho- family GTPases in axon branching. Small GTPases 5:e27974. DOI: https://doi.org/10.4161/sgtp.27974, PMID: 24936971 Stoveken HM, Hajduczok AG, Xu L, Tall GG. 2015. Adhesion G protein- coupled receptors are activated by exposure of a cryptic tethered agonist. PNAS 112:6194–6199. DOI: https://doi.org/10.1073/pnas.1421785112, PMID: 25918380 Vezain M, Lecuyer M, Rubio M, Dupé V, Ratié L, David V, Pasquier L, Odent S, Coutant S, Tournier I, Trestard L, Adle- Biassette H, Vivien D, Frébourg T, Gonzalez BJ, Laquerrière A, Saugier- Veber P. 2018. A de novo variant in ADGRL2 suggests A novel mechanism underlying the previously undescribed association of extreme microcephaly with severely reduced sulcation and rhombencephalosynapsis. Acta Neuropathologica Communications 6:109. DOI: https://doi.org/10.1186/s40478-018-0610-5, PMID: 30340542 Vizurraga A, Adhikari R, Yeung J, Yu M, Tall GG. 2020. Mechanisms of adhesion G protein- coupled receptor activation. The Journal of Biological Chemistry 295:14065–14083. DOI: https://doi.org/10.1074/jbc.REV120. 007423, PMID: 32763969 Wilde C, Fischer L, Lede V, Kirchberger J, Rothemund S, Schöneberg T, Liebscher I. 2016. The constitutive activity of the adhesion GPCR GPR114/ADGRG5 is mediated by its tethered agonist. FASEB Journal 30:666– 673. DOI: https://doi.org/10.1096/fj.15-276220, PMID: 26499266 Xiao P, Guo S, Wen X, He Q- T, Lin H, Huang S- M, Gou L, Zhang C, Yang Z, Zhong Y- N, Yang C- C, Li Y, Gong Z, Tao X- N, Yang Z- S, Lu Y, Li S- L, He J- Y, Wang C, Zhang L, et al. 2022. Tethered peptide activation mechanism of the adhesion GPCRs ADGRG2 and ADGRG4. Nature 604:771–778. DOI: https://doi.org/10.1038/s41586-022- 04590-8, PMID: 35418677 Zuko A, Oguro- Ando A, Post H, Taggenbrock R, van Dijk RE, Altelaar AFM, Heck AJR, Petrenko AG, van der Zwaag B, Shimoda Y, Pasterkamp RJ, Burbach JPH. 2016. Association of cell adhesion molecules contactin- 6 and latrophilin- 1 regulates neuronal apoptosis. Frontiers in Molecular Neuroscience 9:143. DOI: https://doi.org/10.3389/fnmol.2016.00143, PMID: 28018171 Pederick, Perry- Hauser et al. eLife 2023;12:e83529. DOI: https://doi.org/10.7554/eLife.83529 23 of 23 Developmental Biology | Neuroscience
10.7554_elife.86130
TOOLS aND ReSOURceS Bacterial meningitis in the early postnatal mouse studied at single- cell resolution Jie Wang1,2, Amir Rattner1, Jeremy Nathans1,2,3,4* 1Department of Molecular Biology and Genetics, Johns Hopkins University School of Medicine, Baltimore, United States; 2Howard Hughes Medical Institute, Johns Hopkins University School of Medicine, Baltimore, United States; 3Department of Neuroscience, Johns Hopkins University School of Medicine, Baltimore, United States; 4Department of Ophthalmology, Johns Hopkins University School of Medicine, Baltimore, United States Abstract Bacterial meningitis is a major cause of morbidity and mortality, especially among infants and the elderly. Here, we study mice to assess the response of each of the major meningeal cell types to early postnatal E. coli infection using single nucleus RNA sequencing (snRNAseq), immunostaining, and genetic and pharamacologic perturbations of immune cells and immune signaling. Flatmounts of the dissected leptomeninges and dura were used to facilitiate high- quality confocal imaging and quantification of cell abundances and morphologies. Upon infection, the major meningeal cell types – including endothelial cells (ECs), macrophages, and fibroblasts – exhibit distinctive changes in their transcriptomes. Additionally, ECs in the leptomeninges redistribute CLDN5 and PECAM1, and leptomeningeal capillaries exhibit foci with reduced blood- brain barrier integrity. The vascular response to infection appears to be largely driven by TLR4 signaling, as determined by the nearly identical responses induced by infection and LPS administration and by the blunted response to infection in Tlr4-/- mice. Interestingly, knocking out Ccr2, encoding a major chemoattractant for monocytes, or acute depletion of leptomeningeal macrophages, following intracebroventricular injection of liposomal clodronate, had little or no effect on the response of leptomeningeal ECs to E. coli infection. Taken together, these data imply that EC responses to infec- tion are largely driven by the intrinsic EC response to LPS. Editor's evaluation This study presents valuable findings on the changes in immune cell populations and stromal cells occurring at the CNS borders in a neonatal bacterial meningitis model, focusing on fibroblasts, macrophages, and endothelial cells. The study provides a solid snRNA- seq dataset and high- quality immune fluorescence images of dissected brain border regions, that will be useful for the commu- nity. These observations and datasets will be of interest to the neuro- immunology community. Introduction The brain and spinal cord are protected, both physically and immunologically, by the meninges, a multi- layered tissue that occupies the space between the CNS parenchyma and the surrounding bone (Figure 1A; Coles et  al., 2017a; Weller et  al., 2018). Starting from the surface of the brain and moving toward the skin, the meninges consists of: (1) the pia, a thin and semi- permeable layer of cells that allows passage of small molecules and proteins between the CNS parenchyma and the cerebro- spinal fluid (CSF); (2) the sub- arachnoid space, a highly vascularized region containing fibroblasts and immune cells that is filled with CSF and supported by a web of trabeculae; (3) the arachnoid, including *For correspondence: jnathans@jhmi.edu Competing interest: The authors declare that no competing interests exist. Funding: See page 24 Preprinted: 11 January 2023 Received: 11 January 2023 Accepted: 22 May 2023 Published: 15 June 2023 Reviewing Editor: Florent Ginhoux, Agency for Science Technology and Research, Singapore Copyright Wang et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 1 of 27 Tools and resources an outer epithelial barrier layer that serves as the outer boundary of the CSF- accessible space; and (4) the dura, a fibrous layer containing draining sinuses (veins), lymphatics, fibroblasts, and immune cells. Layers 1–3 together comprise the leptomeninges. The meninges hosts a diverse collection of immune cells, including macrophages [referred to as barrier- associated macrophages (BAMs) or CNS- associated macrophages (CAMs)], monocytes, innate lymphoid cells (ILCs), T- cells, B- cells, dendritic cells, and mast cells. Recent single- cell RNA sequencing (scRNAseq) and immuno- phenotyping have revealed distinctive layer- specific types and abundances of these immune cells (Rua and McGavern, 2018; Alves de Lima et  al., 2020). Macrophages are especially abundant, and they can be divided into several molecularly distinct classes that are char- acterized by different half- lives and capacities for self- renewal (Mrdjen et al., 2018; Kierdorf et al., 2019; Van Hove et al., 2019; Masuda et al., 2022). Layer specific diversity is also observed among meningeal fibroblasts, with molecularly distinctive pial, arachnoid, and perivascular fibroblasts, as well as two types of dural fibroblasts (DeSisto et al., 2020; Derk et al., 2021). A wide variety of CNS injuries and disease processes involve the meninges, including stroke, traumatic brain injury, neuroinflammatory conditions such as multiple sclerosis, neurodegenerative diseases, and infections (Derk et al., 2021; Alves de Lima et al., 2020). Disease and injury responses have been observed among meningeal immune cells, fibroblasts, and vasculature (Rua and McGavern, 2018; Derk et al., 2021). Recent experiments with animal models imply that some of these changes play a causal role in injury or disease pathology. For example, in mouse models of stroke, genetic ablation of meningeal mast cells reduces infarct size and brain swelling (Arac et al., 2014). The oldest and best- established pathophysiologic role for the meninges is as a site of bacterial, fungal, or viral infection (Uiterwijk and Koehler, 2012; Williamson et al., 2017; Kohil et al., 2021). Bacterial meningitis is most common among young children and the elderly, and it is generally initi- ated by a blood- borne infection (Ku et  al., 2015; McGill et  al., 2016). The annual incidence of bacterial meningitis ranges from  ~2 per 100,000 people in Western Europe and North America to 100–1000 per 100,000 people in the Sahel region of Africa (Ku et al., 2015; GBD 2016 Meningitis Collaborators, 2018). In Western Europe and North America, mortality from bacterial meningitis is 10–20%, with >30% of survivors experiencing residual neurologic defects (McGill et al., 2016; Eisen et al., 2022). The molecules and mechanisms that mediate bacterial adhesion to and invasion of the meningeal vasculature are objects of active investigation (Coureuil et al., 2017). Meningitis in the neonatal period, when the immune system is immature, typically results from bacterial infection during or immediately prior to delivery. Bacteria, most commonly Group B Strep- tococci and E. coli, are introduced into the bloodstream through breaks in the infant’s skin or via intra- amniotic infection (Gaschignard et  al., 2011; Shane et  al., 2017). The incidence of bacterial meningitis in neonates is ~0.3 per 1000 live births in developed countries and ~4 per 1000 live births in less developed countries (Ku et al., 2015). Among survivors of neonatal meningitis, 20–70% (the number depending on the country) are left with long- lasting neurologic sequelae, including hearing loss, epilepsy, and learning and/or behavioral disabilities (Peltola et al., 2021). In the present work, we describe the molecular and cellular responses of meningeal cells in a mouse model of neonatal E. coli meningitis. Responses to infection were observed in every major meningeal cell type, including endothelial cells (ECs), macrophages, and fibroblasts. We have also used genetic and pharmacologic perturbations of immune cells and pathways to explore communication networks responsible for this complex multi- cellular response. Results Single nucleus sequencing and flatmount imaging of the mouse meninges Our point of departure in studying the murine meninges was to utilize a simple protocol for dissecting the leptomeninges and the dura free from adjacent tissues. When the skull and brain are sepa- rated in the absence of fixation, the natural cleavage plane is between the leptomeninges and dura (Figure  1A). The leptomeninges can then be peeled from the brain surface and the dura can be peeled from the inner surface of the skull. In our hands, the separation of the unfixed leptomeninges from the brain parenchyma works best with tissue from young mice: it is efficient with tissue harvested at postnatal day (P)6 but, as noted by Van Hove et al., 2019 and confirmed by us, it fails with adult Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 2 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources A Skin Meninges Bone Dura Arachnoid Pia Dural venous sinus Leptomeninges C 1 M A C E P I P A D 6 0 2 D C brain with leptomeninges attached F 2 P A M U Fb-a Fb-d2 Fb-p Fb-d1 MCs Fb-m U2 Arachnoid barrier Fb-d3 Macrophages U1 Osteoclasts Osteoblasts Neurons Glia ECs UMAP 1 Fibroblasts dura 3 (Fb-d3) Endothelial cells (ECs) Fibroblasts pia (Fb-p) Fibroblasts dura 1 (Fb-d1) Macrophages Fibroblasts arachnoid (Fb-a) Fibroblasts dura 2 (Fb-d2) Osteoblasts Osteoclasts Arachnoid barrier Glia Unkown cluster 1 (U1) Fibroblasts mitotic (Fb-m) Mural cells (MCs) Neurons Unknown cluster 2 (U2) I Macrophages ECs Mural cells Neurons Glia Osteoclasts Osteoblasts Fibroblasts pia Arachnoid barrier Fibroblasts arachnoid Fibroblasts dura 3 Fibroblasts dura 2 Fibroblasts dura 1 dissected dura dissected leptomeninges B 1 M A C E P I P A D 6 0 2 D C D I P A D 6 0 2 D C 1 M A C E P dissected dura E dissected leptomeninges brain after leptomeninges removal I P A D D A C E I P A D D A C E 1 M A C E P 1 M A C E P 1 M A C E P D A C E D A C E 1 T U L G G control infected H Fb-d1 Cdh18 Slc1a2 Kdr 2 P A M U UMAP 1 Cdh1 Arachnoid barrier 4 3 2 1 0 3 2 1 0 4 3 2 1 0 5 4 3 2 1 Glia 5 4 3 2 4 3 2 1 0 Slc47a1 Fb-d3 ECs Satb2 Osteoblasts St18 Igsf8 4.0 3.5 3.0 2.5 2.0 Fb-a Osteoclasts Tmem132c Notch3 Fb-p 4 3 2 1 0 MCs UMAP 1 3 2 1 0 5 4 3 2 1 0 3 2 1 0 4 3 2 1 0 2 P A M U Epha3 Fb-d2 F13a1 Macrophages 1 f b E 2 v a N 8 1 h d C 3 m n e T 3 2 m a d A 3 a h p E 1 a 1 1 o C l 2 f f e m T 3 1 t n a G l 1 a 7 4 c S 1 a 6 1 c S l l l 3 s t m a d A t a n N 8 f s g I l 3 p b n a R a 0 2 a 6 c S l 1 h d C 3 m p r T d 3 a m e S 1 a 5 2 o C l r d a x C 1 s b r o S 6 1 t n W i 2 c s A 2 x n u R 6 a 4 o C l l l p A 3 t a F p a F 8 1 t S 2 b t a S 4 n r z d P 2 b 9 c S l 2 d 0 v 6 p t A 1 1 a 7 c S l c 2 3 1 m e m T 5 p c A k s t C p a f G 4 p q A 2 a 1 c S l d 6 a m e S 3 g r N 3 n x r N 5 n t n C 2 m n e T 3 h c t o N 9 c c b A b r f g d P 3 c p r T 1 m a c e P r d K 1 t l F k e T g r E 1 c r M j p b R s s t C c r p t P 1 a 3 1 F 1 e r g d A 2 1 0 Percent Expressed 0 25 50 75 Average Expression Figure 1. Dissection and single nucleus (sn) RNAseq of mouse meninges. (A) Diagram of the tissue layers between brain and skin, corresponding to the red rectangle in the coronal section through the brain and skull (upper right). (B) Cross- section of dissected dura stained for PECAM1 (endothelial cells) and CD206 (macrophages). (C) Coronal section through cortex (lower) and overlying leptomeninges (upper) stained for PECAM1 and CD206. (D) Cross- scetion of the isolated leptomeninges (upper panel) and the denuded brain (lower) stained for PECAM1 and CD206. (E) Dissected dura and leptomeninges stained for PECAM1 and ECAD (arachnoid barrier epithelium); the leptomeninges is also stained for GLUT1 (a BBB marker; bottom panel). (F) UMAP plot of combined control and infected meninges snRNAseq datasets with cell clusters differentially colored and labeled. The macrophage cluster includes a small upward extension that represents monocytes and monocyte- derived cells. (G) UMAP plots of separated control and infected meninges snRNAseq datasets. (H) UMAP plots, as in panel (F) showing transcripts that are highly enriched in each of 12 cell clusters (labels in each panel). (I) Dot plot showing some of the transcript abundances that most clearly discriminate among major meningeal cell types, as well as contaminating neurons and glia. Scale bars: B- E, 100 um. All tissue and data in this and other figures are from P6 mice. The immunostaining and histochemical probes in this and other figures are indicated adjacent to the corresponding panel(s), with lettering color- coded to match the corresponding fluorescent color. The online version of this article includes the following figure supplement(s) for figure 1: Figure 1 continued on next page Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 3 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources Figure 1 continued Figure supplement 1. Pairwise Pearson correlations among snRNAseq datasets. Figure supplement 2. Flatmount images of control (i.e. uninfected) leptomeninges and dura. Figure supplement 3. Characteristics of the P5- P6 model of E. coli meningitis. tissue. Separation of the leptomeninges is also more readily accomplished if the brain is first chilled in ice- cold PBS for several minutes. Immunostaining of the isolated leptomeninges, the isolated dura, and the denuded brain reveals ECs (expressing PECAM- 1/CD31) and macrophages (expressing CD206/MRC1) in both leptomeninges and dura (Figure 1B–D). The arachnoid epithelium expresses E- cadherin (ECAD). Leptomeningeal ECs, but not dura ECs, express GLUT1, a blood- brain barrier (BBB) marker (Figure 1E). These and all other analyses in this study were conducted at P6. For transcriptome analyses at cellular resolution, we sought to obtain as representative a sampling of cell types as possible and to minimize RNA synthesis or degradation after dissecting the dura and leptomeninges. Therefore, we avoided enzymatic tissue dissociation followed by single cell analysis and, instead, purified nuclei following tissue homogenization. The meninges were obtained from P6 mice, either without infection (“control”) or following a single subcutaneous injection of E. coli K1 at P5 (‘infected’), as described more fully in the next section. Each sequencing library, two control and three infected, was produced from a single mouse and consisted of the pooled leptomeninges plus dura tissues. All experiments, except for those shown in Figure 6C- F and 7, were conducted with FVB/NJ mice. Single nucleus (sn)RNAseq data were obtained from 14,356 and 34,585 nuclei from control and infected mice, respectively, with a mean of 1,320 transcripts sequenced per nucleus, using the 10 X Genomics Chromium platform (Supplementary file 1). Pairwise Pearson correlations among the five snRNAseq datasets (two control and three infected) shows correlations of 0.98–1.00 within the same group and 0.90–0.93 between infected vs. control groups (Figure 1—figure supple- ment 1). Fourteen principal cell clusters were identified with Seurat, and their identities were assigned by immunostaining and with reference to published data, as seen in the Uniform Manifold Approximation and Projection (UMAP) plots in Figure 1F and H (Supplementary file 2). One cluster, labeled ‘macro- phages’, consists largely of macrophages but also encompasses other immune cells, as described in detail below. Two clusters represent neurons and glia, presumably brain contaminants. One small cluster derives from mitotic fibroblasts, and two small clusters (U1 and U2) are unidentified. Strikingly, five large clusters represent fibroblasts – one pial, one arachnoid, and three dural, as determined by a comparison with published data on meningeal fibroblasts (DeSisto et al., 2020; Derk et al., 2021). A dot plot of 61 transcripts that exhibit cell- type- specific patterns of enrichment supports these cell cluster assignments and also illustrates a pattern of partial overlap in gene expression among the five fibroblast clusters (Figure 1I). Comparing UMAP plots of control and infected datasets reveals shifts in the positions of the immune cell and fibroblast clusters with infection (Figure 1G). Flatmounts of the isolated leptomeninges and dura permit confocal imaging across the full depth of each of these tissues (Figure 1—figure supplement 2). Additionally, the dura can be imaged as a flatmount while it remains attached to the inner surface of the skull, although this arrangement reduces tissue access to antibody and washing solutions. Leptomeninges flatmounts show a high density of macrophages marked by (1) co- expression of macrophage markers CD206 and LYVE1, which localize to distinct cytoplasmic/surface compartments (Figure 1—figure supplement 2A–C). These cells, as well as other non- macrophage immune cells, also express SPI1/PU.1, which localizes to the nucleus, and CD45/PTPRC, a general marker for hematopoietic cells (Figure 1—figure supplement 2A–C). Flatmounts of the peripheral dura (i.e. outside the sinuses) show elongated perivascular macrophages expressing CD206 and LYVE1, as well as additional immune cells expressing CD45 (Figure 1—figure supplement 2D). The vasculature can be visualized in leptomeninges and dura flatmounts by immunostaining for PECAM- 1, and, in leptomeninges flatmounts, by immunostaining for tight junction markers Occludin (OCLN), Zonula occludens- 1 (ZO- 1), and Claudin- 5 (CLDN5) (Figure 1—figure supplement 2C and below). In dura flatmounts with the bone attached, perivascular fibroblasts express FOXP2, and osteo- blasts express SATB2 (Figure 1—figure supplement 2D). Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 4 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources The bacterial meningitis model To model human neonatal meningitis, we inoculated P5 mice with E. coli strain RFP- RS218 (O18:K1:H7). This strain is a clinical isolate from the cerebrospinal fluid (CSF) of a neonate with meningitis and it has been derivatized by the addition of a plasmid expressing red fluorescent protein (RFP) to facil- itate visualization of E. coli cells in tissue. P5 was chosen as the age of inoculation to approximate the degree of maturity of the neonatal human immune system (Holsapple et al., 2003; Kuper et al., 2016; Park et  al., 2020). Presumably, the 9- month human gestation compared to the 19–20  day mouse gestation explains the relatively greater developmental maturity of the human immune system compared to the mouse immune system at birth. In each experiment, a litter of P5 mice were randomly divided into two groups that were either subcutaneously injected in the back with 1.2×105  CFU of E. coli RFP- RS218 in 20  µL PBS or not injected. Twenty- two hours later (at P6), the mice were sacrificed and tissues analyzed. One day after E. coli inoculation, the mice appeared lethargic and they stopped gaining weight (Figure 1—figure supplement 3A). Any inoculated mice that were not sacrificed died within two days of the inoculation (Figure 1—figure supplement 3B). At sacrifice, 22  hr after subcutaneous inoculation, E. coli cells were typically observed in a patchy distribution in the leptomeninges and dura, and at much sparser density within the brain (Figure 2A–D). The number of E. coli per unit area in flatmounts of the dura sinus region was, on average, ~20- fold greater than the number of E. coli in the same area in flatmounts of the leptome- ninges, reflecting, in part, the several fold greater depth of the dura sinus (Figure 1—figure supple- ment 3C, left plot). The spatial heterogeneity in E. coli accumulation within the meninges, together with animal- to- animal variation in the severity of infection, likely accounts for some degree of vari- ability in the cellular alterations associated with infection. To address this issue, representative images were chosen for the figures and for quantification. All images and data in the present study were obtained at P6, and all quantifications of flatmount images used Z- stacks that span the full thickness of the tissue. Overview of the responses of meningeal cells to infection In comparing control vs. infected datasets for each of the principal cell clusters, scatter plots encom- passing all transcripts reveal relatively few changes in the contaminating neuron and glia clusters and many more changes in each of the principal meningeal clusters, with roughly equal numbers of up- and down- regulated transcripts (Figure 2—figure supplement 1 and Supplementary file 3). A comparison of non- immune meningeal cells (excluding contaminating non- meningeal cells [neurons, glia, osteoblasts, and osteoclasts]) that was limited to transcripts with a log2- fold change equal to or greater than 2.5 in control vs. infection conditions in any one or more of these cell types shows that dural and leptomeningeal ECs form one cluster and dural and leptomeningeal fibroblasts and arach- noid barrier cells form a second cluster (Figure 2—figure supplement 2). The following transcripts are increased broadly across cell types: (1) general stress response genes Metallothionien- 1 (Mt1), Metallothionien- 2 (Mt2), lipocalins Apolipoprotein D (Apod) and Lipocalin2 (Lcn2), (2) Lipopolysac- charide (LPS) binding protein (Lbp, which presents LPS to TLR4), (3) Serum amyloid A3 (Saa3, an acute phase protein induced by inflammation), and (4) Ceruloplasmin (Cp, a secreted copper- binding protein that is also induced by inflammation). Dot plots that include all of the cell types in the snRNAseq dataset were generated to visualize representative examples of the various patterns of altered transcript abundance with infection (Figure 2—figure supplement 3). Several transcripts – for example, Kiz, Slc39a14, Ap4e1, and Cp – are up- regulated across nearly all clusters (Figure 2—figure supplement 3A). However, for most tran- scripts that exhibit abundances changes with infection, those changes were limited to smaller subsets of cell types, for example, the down- regulation of Col14a1, Col8a1, and Slc4a10 in dural fibroblasts, the down- regulation of Igsf8, Tmtc4, and Zfp536 in arachnoid and pial fibroblasts and arachnoid barrier cells, and the up- regulation of Cxcl2 in macrophages (Figure 2—figure supplement 3A and B). Hierarchical clustering in a gene set enrichment analysis (GSEA) with control vs. infected samples shows prominent induction of inflammatory responses (‘complement’, ‘interferon gamma response’, ‘IL6 JAK STAT3 signaling’, ‘TNFA signaling via NFKB’) across all meningeal cell types. ECs and arach- noid barrier cells show a reduction in apical barrier (i.e. tight junction) transcripts, but this reduction did not reach statistical significance (Figure 2—figure supplement 4). Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 5 of 27 Microbiology and Infectious Disease | Neuroscience Control Infected n i y t i s n e d e g a h p o r c a M t n u o m a t l f i s e g n n e m o p e t Leptomeninges (flatmount) l Control Infected ) 2 m m / s l l e c . + 1 U P ( I 4000 3000 2000 1000 0 * Control Control Infected Infected E 6 0 2 D C 1 . U P F 1 E V Y L 6 0 2 D C G 0 8 1 D C 6 0 2 D C 0 8 1 D C H 6 L I 6 0 2 D C L e p t o m e n n g e s i ( f l a t m o u n t ) Tools and resources B B C C A i l o c . E ) P F G - 2 e T ( i A’ P F G i l o c . E B i l o c . E ) P F G - 2 e T ( i P F G i l o c . E 1 M A C E P 6 0 2 D C C FD J 8 A 0 0 1 S 1 M A C E P 8 A 0 0 1 S 6 0 2 D C 8 A 0 0 1 S 1 M A C E P Control Infected Control Infected Control Infected Leptomeminges (flatmount) **** 5 * y t i s n e n t i e v i t l a e r 6 0 2 D C 4 3 2 1 0 Dura sinus (flatmount) **** Leptomeminges (flatmount) *** p=0.63 Dura sinus (flatmount) **** y t i s n e t n i l e v i t a e r 6 0 2 D C y t i s n e t n i l e v i t a e r 6 L I 3 2 1 0 3 2 1 0 y t i s n e t n i e v i t l a e r 1 E V Y L y t i s n e t n i e v i t a e r l 1 E V Y L 4 3 2 1 0 3 2 1 0 y t i s n e t n i l e v i t a e r 0 8 1 D C 5 4 3 2 1 0 Control Control Infected Infected * * K e h t n i y t i s n e d l l e c + 8 A 0 0 1 S ) 2 m m / s l l e c ( s u c u s l l a r t n e c e h t n i y t i s n e d l l e c + 8 A 0 0 1 S ) 2 m m / s l l e c ( i s e g n n e m o t p e l 2000 1500 1000 500 0 500 400 300 200 100 0 Control Control Infected Infected s u r f a c e C o r t i c a l ( f l a t m o u n t ) C e n t r a l i s n u s L e p t o m e n n g e s i D u r a ( f l a t m o u n t ) L e p t o m e n n g e s i ( f l a t m o u n t ) C e n t r a l i s n u s D u r a ( f l a t m o u n t ) D u r a ( f l a t m o u n t ) - C e n t r a l i s n u s ( f l a t m o u n t ) L e p t o m e n n g e s i Figure 2. The E. coli meningitis model and some immune cell responses. (A–C) Coronal sections of P6 brain with leptomeninges, 1 day after a subcutaneous injection of 1.2×105 RFP- expressing E. coli K1. Regions within the rectangles labeled (B) and (C) are enlarged below. The Tie2- GFP transgene is expressed in ECs. (D) Leptomeninges flatmount showing scattered E. coli (RFP; false colored magenta). (E) Leptomeningeal macrophages, visualized with nuclear immunostaining for transcription factor PU.1 and cytoplasmic staining for CD206, show cytoplasmic enlargement upon Figure 2 continued on next page Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 6 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources Figure 2 continued infection (left), but only a small increase in cell number (right). (F) Infection leads to a selective reduction in LYVE1 immunostaining, and little or no change in CD206 immunostaining, in macrophages in the leptomeninges and dura. (G) Infection leads to little or no change in CD180 and CD206 immunostaining in macrophages in the leptomeninges. (H) IL6 increases in dural fibroblasts in the central sinus. (I) Quantification of images in F- H, in arbitrary units. Each point in this and other quantifications of immunofluorescent data represents the analysis of a single Z- stacked confocal image that encompasses the full depth of the tissue (leptomeninges or dura), unless noted otherwise. (J and K) Increase in cells immunostained for S100A8 in the leptomeninges and dura of infected mice. All infected tissue and data in this and other figures are from P6 mice that had been infected 22 hr earlier. Scale bars: A and A’, 500 µm; B- I, 100 µm. In this and all other figures showing quantification: (1) unless stated otherwise, each symbol in the immunofluorescent quantifications represent a single confocal image, (2) the bars show mean ± SD; (3) the number of mice used for each sample are listed in Supplementary file 4; (4) the Wilcoxon rank sum test was used to measure statistical significance, except for Figure 5D and G, in which the sample size is too small and the student t- test was used instead; and (5) abbreviations are: n.s., not significant (i.e. p>0.05); *, p<0.05; **, p<0.01; ***, p<0.001; ****, p<0.0001. The online version of this article includes the following figure supplement(s) for figure 2: Figure supplement 1. Scatterplots for the major meningeal cell types comparing snRNAseq transcript abundances in control vs. infected mice. Figure supplement 2. Heatmap of non- immune meningeal cells showing all transcripts with log2- fold change greater than 2.5 in control vs. infection conditions in any one or more of the listed cell types. Figure supplement 3. Dot plot showing some of the transcript abundance patterns that distinguish control vs. infected meninges, plotted by cell type. Figure supplement 4. Gene set enrichment analysis (GSEA) for individual cell types in a comparison of control vs. infected meninges. Responses of meningeal immune cells to infection Despite the presence of E. coli, the density of macrophages in the leptomeninges, which are normally present at 1500–2000 cells per mm2, increased by only ~20% (Figure 2E; for all image quantifications, the number of mice in each sample are listed in Supplementary file 4). Whether this increase reflects in situ proliferation, ingress from other tissue compartments [e.g. blood and/or skull bone marrow (Herisson et al., 2018; Cugurra et al., 2021)], or a combination of the two, remains to be determined. However, macrophage morphology changed dramatically with infection, from small and rounded to large and irregularly shaped (Figure 2E). Infection also led to a reduction in LYVE1 abundance, but little or no change in CD180 or CD206 abundance, in macrophages (Figure 2F, G and I). In the dura of infected mice, IL6 was induced in fibroblasts, with the cell type assignment determined by the observed increase in fibroblast- specific Il6 transcript abundance in the snRNAseq datasets (Figure 2H and I, and Figure 2—figure supplement 3A). In both the dura and leptomeninges, the number of cells expressing S100A8, a marker for monocytes and immature macrophages, increased  ~10- fold (Figure 2J and K). For a more comprehensive assessment of the immune response to infection, we further parsed the immune and immune- related clusters into microglia (a brain contaminant), innate lymphoid cells/T cells (ILC/T), osteoclasts (a contaminant from the skull), monocytes (MCs) and monocyte- derived cells, resident macrophages (MPs; subdivided by CCL2 expression), and inflammatory macrophages (subdi- vided by IL1 receptor type 1 [IL1R1] expression; Figure 3A–C , and D lower panel). Cell clusters were assigned with reference to published data, as summarized in Supplementary file 2. The designation of macrophage clusters as ‘resident’ or ‘inflammatory’ reflects their locations in the UMAP plots (i.e. their gene expression profiles), the former corresponding to UMAP locations occupied by macro- phages in the control meninges and the latter corresponding to UMAP locations occupied by macro- phages that are specific for the infected meninges (compare Figure  3A and F). Importantly, these designations refer only to patterns of gene expression and are agnostic as to the origins and migra- tory histories of the macrophage clusters. The division of resident macrophages into CCL2- and CCL2+ subtypes is based on the differential expression of multiple genes, six of which are included in the dot plot in Figure 3B, with these six plus an additional 24 also included in the dot plot in Figure 3—figure supplement 1A. Although the relative abundances of the principal meningeal cell types did not change with infec- tion, as judged by counting nuclei in the five snRNAseq libraries (Figure  3D upper panel), parsing the individual immune cell types revealed an ~twofold decrease in the abundance of CCL2- resident macrophages, a > 10- fold increase in inflammatory macrophages (IL1R1+ and IL1R1-), and a several- fold increase in the abundance of monocytes or monocyte- derived cells (Figure  3D, lower panel). While these experiments do not distinguish between changes in gene expression patterns among resident immune cells versus the ingress of circulating immune cells, the changes in macrophage Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 7 of 27 Microbiology and Infectious Disease | Neuroscience A 2 P A M U Tools and resources Resident MPs CCL2+ Resident MPs CCL2- B Microglia ILC/T cells Osteoclasts Mitotic MPs Inflammatory MPs IL1R1- Inflammatory MPs IL1R1+ Microglia MCs or MC- derived cells ILC/T cells Osteoclasts UMAP 1 MCs/MC−derived cells Resident MPs, CCL2+ Resident MPs, CCL2− Inflammatory MPs, IL1R1+ Inflammatory MPs, IL1R1− p C 3 a a S l 2 c x C 3 f s C 1 r 1 l I 8 3 d C r p e L 1 s b r o S d 1 a n c a C b 9 6 1 m a F 6 s a G 4 2 f c T 0 6 x d D k e P l l 2 c C d g p H p H c s t C p i t y C n a c V 1 b c P l 4 7 d C k s t C i k g D 2 p a y N d 1 o y M 2 b 9 c S l 4 s s a C 2 2 5 6 2 m G x o T 3 f z k I 1 p a k S l 1 3 6 1 d C 3 v a N 5 r c C 3 1 f g F 2 c d x P l 4 2 p a g h r A l 3 c a n a g 6 t S C 2 P A M U D 15 y r a r b 10 i l f o t n e c r e P 5 ILC/T cells Resident MPs CCL2+ Resident MPs CCL2- Inflammatory MPs IL1R1- Osteoclasts Mitotic MPs MCs or MC-derived cells Microglia Inflammatory MPs IL1R1+ Spi1 (PU.1) 3 2 1 0 Ptprc (Cd45) 4 3 2 1 0 Mrc1 (Cd206) 6 4 2 0 All cell types c = control i = infected − − − − − − − ) c ( a i l G − −− − − − 0 ) c ( l s t s a c o e t s O ) i ( s t s a c o e l t s O ) c ( l s t s a b o e t s O ) i ( s t s a b o e l t s O ) c ( s n o r u e N ) i ( s n o r u e N ) c ( s l l e c l a r u M ) i ( s l l e c l a r u M ) c ( s e g a h p o r c a M ) i ( s e g a h p o r c a M − − −− − − −− − − i ) i ( a p s t s a b o r b F i l ) c ( c i t o t i m s t s a b o r b F i l ) i ( c i t o t i m s t s a b o r b F i l ) c ( 3 a r u d ) i ( 3 a r u d ) c ( 2 a r u d ) i ( 2 a r u d ) c ( 1 a r u d ) i ( 1 a r u d l s t s a b o r b F i l s t s a b o r b F i l s t s a b o r b F i l s t s a b o r b F i l t s a b o r b F i l t s a b o r b F i ) c ( ) i ( i d o n h c a r a s t s a b o r b F l i i d o n h c a r a s t s a b o r b F l i − ) i ( a i l G ) c ( a p i l s t s a b o r b F i − − ) c ( s C E ) i ( s C E − − ) c ( i r e i r r a b d o n h c a r A ) i ( i r e i r r a b d o n h c a r A y r a r b i l f o t n e c r e P 10 5 0 c = control i = infected − ) c ( - 2 L C C s P M t n e d s e R i − ) i ( - 2 L C C s P M t n e d s e R i − ) c ( + 2 L C C s P M t n e d s e R i − ) i ( + 2 L C C s P M t n e d s e R i Immune cells Library JW19 JW20 JW21 control JW22 infected JW23 − − ) c ( s P M c i t o t i M ) i ( s P M c i t o t i M − ) i ( a i l g o r c M i − ) c ( a i l g o r c M i − ) i ( l s t s a c o e t s O − ) c ( l s t s a c o e t s O − ) i ( s l l e c d e v i r e d - C M r o s C M − ) c ( s l l e c d e v i r e d - C M r o s C M − ) i ( - 1 R 1 L I s P M y r o t a m m a l f n I − ) i ( + 1 R 1 L I s P M y r o t a m m a l f n I − ) c ( - 1 R 1 L I s P M y r o t a m m a l f n I − ) c ( + 1 R 1 L I s P M y r o t a m m a l f n I − ) c ( s l l e c T C L I / − ) i ( s l l e c T C L I / Percent Expressed 0 25 50 75 Average Expression −1012 Monocytes S100a8 Resident MPs, _ + CCL2 > CCL2 Resident MPs, CCL2 and CCL2 + _ Inflammatory MPs, IL1R1+ Inflammatory MPs IL1R1 _ Cass4 Ddx60 Il1r1 Ebi3 4 3 2 1 0 4 3 2 1 0 5 4 3 2 1 0 3 2 1 0 6 4 2 0 UMAP 1 E j ) d e t s u d a , e u a v - P l Resident macrophages, CCL2 Mef2c Tbc1d4 Snx29 Rnf150 Mctp1 Dse 200 Arsb Tcf24 Ddx60 100 Fmnl2 Cacna1d Lcn2 Apod Mt1 Mt2 Fkbp5 Ccl9 Cd38 _ Saa3 Cxcl2 ( 0 1 g o l - Samd4 Bank1 Cd180 0 -3 −2 -1 0 1 2 3 decrease with infection -log (0.05) 10 Resident macrophages, CCL2 + 60 40 20 0 Mef2c Ctsc Arsb Hpgds Cd180 Rnf150 Hpgd Samd4 Cd36 Tcf24 Zfp710 Saa3 Cxcl2 Mt1 Mt2 Birc3 Msr1 Fth1 Pde4b N4bp1 Csf3 Cd163 -log (0.05) 10 6 4 5 -3.75 -1 log (Average fold change) increase with infection −2.5 2 0 1 2.5 3.75 5.0 6.25 no significant change F 2 P A M U Ccl9 Ccl9 Cd163 Cd163 Control Infected Control Infected Lyve1 Lyve1 Mef2c Mef2c Control Infected Control Infected Cd180 Cd180 Hpgd Hpgd 5 4 3 2 1 0 Control Infected Control Infected UMAP 1 Figure 3. Immune subtypes and their responses to infection. (A) snRNAseq UMAP plot for immune cells from combined control and infected meninges. (B) Dot plot showing some of the transcript abundances that most clearly distinguish among meningeal immune cells. (C) UMAP plots, as in panel (A) showing eight transcripts that are expressed by all (left three panels) or by distinct subsets (right five panels) of macrophage subtypes and macrophage- like cells. Red arrows highlight regions within the UMAP clusters that correspond to distinct cell types, as defined in panel Figure 3 continued on next page Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 8 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources Figure 3 continued (A). (D) Comparing the number of nuclei in control (c) vs. infected (i) snRNAseq datasets across all meningeal cell types (upper panel) and across immune cells (bottom panel). The fraction of cells comprising the general category ‘macrophages’ shows no change with infection (black arrow in upper panel). However, the lower panel shows that several macrophage subsets decrease (green arrow) or increase (red arrows) in abundance with infection. (E) Volcano plots for CCL2- and CCL2+ macrophages showing control vs. infected snRNAseq transcript abundances (see Supplementary file 5). (F) UMAP plots, as in (A), comparing control vs. infected snRNAseq for six genes. The online version of this article includes the following figure supplement(s) for figure 3: Figure supplement 1. Comparisons between CCL2- and CCL2+ macrophages. Figure supplement 2. Leptomeningeal and dura fibroblast markers and fibroblast responses to infection. Figure supplement 3. Collagen transcripts in individual meningeal cell types from control vs. infected mice. Figure supplement 4. SLC transporter transcripts in individual meningeal cell types from control vs. infected mice. abundances could be largely explained if ~50% of the CCL2- resident macrophages present prior to infection converted to inflammatory macrophages in response to infection. Among genes that are either up- or down- regulated in CCL2- or CCL2+ resident macrophages, several dozen show greater than 5- fold changes in abundance (Figure 3E and Supplementary file 5). As seen in Figure 3—figure supplement 1B, CCL2- and CCL2+ macrophages show nearly identical changes among the transcripts with the greatest increases in abundance with infection, and somewhat greater variability among tran- scripts with the greatest decreases in abundance with infection. The UMAP plots in Figure 3F compare the expression levels of six genes in control vs. infected immune cells, and they illustrate the appearance of the inflammatory macrophage pattern of gene expression specifically in the infected meninges, represented by the lower ~50% of the macrophage cluster (see also Figure 3A). Figure 3F also illustrates the diversity of macrophage transcript changes with infection, with dramatic up- regulation of Ccl9, modest up- regulation of Cd163, modest down- regulation of Lyve1 and Mef2c, and dramatic down- regulation of Cd180 and Hpgd. Interestingly, by immunostaining, CD180 levels showed little change at this time point (22 hr post- infection; Figure 2G and I), suggestive of a long protein half- life, whereas LYVE1 levels showed a large reduction (Figure 2F and I), suggestive of post- transcriptional as well as transcriptional down- regulation. These data are consistent with a model in which infection promotes the appearance of new macrophages ‘states’, as defined by novel patterns of gene expression that are distinct from those of resting macrophages. Responses of meningeal fibroblasts to infection Fibroblasts constitute the most abundant cell type in the leptomeninges and dura (Figure 1F, Supple- mentary file 1, and Figure  3—figure supplement 2), and each of the five meningeal fibroblast subtypes shows numerous changes in transcript abundances in response to infection (Figure  2— figure supplements 1–4 and Supplementary file 3). Here, we highlight two gene families, colla- gens and SLC transporters, in which multiple family members show expression changes in control vs. infected meningeal fibroblasts. Transcripts coding for multiple collagen subtypes are down- regulated by infection: among the 50 members of the collagen gene family, 25 show detectable expression in the meninges by snRNAseq and 19/25 are down- regulated but only 2/25 are up- regulated (Figure 3— figure supplement 3). Two examples are shown in the UMAP plots in Figure 3—figure supplement 2: Col14a1 is down- regulated in type 1 and 2 dural fibroblasts, and Col25a1 is down- regulated in arachnoid barrier cells and type 3 dural fibroblasts. Immunostaining of the dura for COL14A1 did not reveal significant changes one day after infection, likely reflecting the slow turnover of mature extra- cellular matrix collagen (Jackson and Heininger, 1975; Last et al., 1989). Similarly, multiple transcripts coding for SLC transporters are down- regulated in meningeal fibro- blasts. Of the ~350 members of the mouse Slc gene family with detectable expression in meningeal cells, 37 show infection- dependent changes in transcript abundance in one or more meningeal cell types by snRNAseq, with 21/37 down- regulated and 6/37 up- regulated by log2- fold>0.25 following infection (Figure  3—figure supplement 4). The UMAP plots in Figure  3—figure supplement 1B show down- regulation of Slc16a1 in dura fibroblasts. Slc39a14, which codes for a divalent metal trans- porter, is unusual in its substantial up- regulation with infection (Figure 3—figure supplement 4B). In contrast to the pattern of down- regulation among the majority of Col and Slc transcripts, across the full transcriptome similar numbers of transcripts are up- and down- regulated by infection in each Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 9 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources of the major meningeal cell types (Figure  3—figure supplement 1). Among fibroblasts, examples of transcripts that are up- regulated include Alk in arachnoid fibroblasts, Scara5 in pial fibroblasts, Nrg3 in type 1 dura fibroblasts, Camk4 in type 2 dura fibroblasts, and Lbp in type 3 dura fibroblasts (Figure 3—figure supplement 2B). As noted above, infection up- regulates IL6 protein and Il6 tran- scripts in dura fibroblasts (Figure 2H and I, and Figure 3—figure supplement 2B). Responses of meningeal vasculature to infection In the context of bacterial meningitis, meningeal ECs serve as both a portal of entry for bacteria and immune cells and as a site of pathologically increased vascular permeability (Kim et al., 1997; Barichello et al., 2013; Coureuil et al., 2017). To explore the meningeal EC response to infection, we first parsed the EC UMAP into clusters derived from leptomeninges, dura, and arterial ECs based on the following markers: BBB markers Cldn5, Lef1, Slc2a1, and Slc7a1 for leptomeningeal ECs vs. non- BBB marker Plvap for dural ECs; and arterial markers Bmx and Fbln5 (Figure 4A and C; Supple- mentary files 2 and 6). Dural ECs were further divided into Von Willebrand Factor (VWF) expressing and non- expressing subclasses (Figure 4A and C). Immunostaining of leptomeninges and dura flat- mounts for CLDN5, SLC2A1/GLUT1, and PLVAP confirmed the cluster assignment of leptomeningeal ECs vs. dura ECs, and immunostaining for smooth muscle actin (SMA) was used to distinguish arteries vs. veins histologically. Infection leads to an expansion in the size of the leptomeninges and dura EC clusters (Figure 4A and E), indicative of increased heterogeneity in transcriptome content, with little or no change in the proportions of the different EC subtypes (Figure 4B). Multiple transcriptome changes distinguish the infection responses of different EC clusters, including down- regulation of transcripts coding for amino acid transporters SLC7A1 and SLC7A5 in leptomeningeal ECs and down- regulation of transcripts coding for plasmalemma vesicle associated protein (PLVAP; a marker of high permeability vasculature) and the mechanosensory channel PIEZO2 in dural ECs (Figure 4C–E). Transcripts coding for VEGFR2/ KDR are down- regulated in leptomeningeal and dural ECs (Figure 4C–E). In leptomeninges flatmounts, infection produces mislocalization and clustering of CLDN5 and PECAM1, disorganized capillary morphology, and an expansion of the area covered by capillaries (Figure 5A–C). Immunoblotting of leptomeningeal proteins shows a modest reduction in the mean level of CLDN5, but this trend did not reach statistical significance due to the relatively high animal- to- animal variability in the infected group (Figure  5D). Functionally, there are patchy deficiencies in vascular barrier integrity following infection, with a spatial distribution that closely matches the clustering of CLDN5 and PECAM1, as revealed by extravasation of Sulfo- NHS- biotin, a low molecular weight intravascular tracer (Figure 5A–C). The same infection- associated vascular phenotypes were also seen one day after an intraperitoneal (IP) injection of 10 mg/kg lipopolysaccharide (LPS), a potent activator of the innate immune response to gram- negative bacteria such as E. coli (Figure 5A–C). Two well- studied signaling pathways are known to control CNS vascular permeability: VEGF and WNT. Consistent with the down- regulation of Vegfr2/Kdr transcripts seen in infected ECs by snRNAseq (Figure  4C–E), VEGFR2 immunostaining is also greatly reduced in the leptomeninges (Figure 5E and H). As VEGFR2 is the principal receptor for VEGF signaling in ECs, its down- regulation suggests that the enhanced vascular permeability associated with bacterial infection is not caused by increased VEGF signaling, a known mechanism for increasing vascular permeability (Senger et  al., 1983; Roberts and Palade, 1995). WNT signaling in CNS vasculature maintains the blood- brain barrier (BBB) and it is both mediated by and up- regulates the transcription factor LEF1, which binds to target genes in combination with beta- catenin (Sabbagh et al., 2018). In control leptomeninges flatmounts, LEF1 accumulates in vein and capillary EC nuclei with minimal accumulation in arterial ECs (Figure 5F, left panels). Strikingly, in infected leptomeninges flatmounts, LEF1 immunostaining in ECs is greatly reduced, but it persists in many non- ECs (Figure 5F, right panels; and Figure 5H). By immunoblotting, total leptomeningeal LEF1 levels are reduced by ~25%, a change that was not statistically significant (Figure 5G). The right panels of Figure 5F shows reductions in CLDN5 and PECAM1 specifically in leptomeningeal veins, which appear as ‘shadows’ in the flatmount images. With infection, a reduction is also seen for ERG, a pan EC transcription factor (Figure 5F and H). These data are consistent with a model in which infec- tion leads to reduced BBB integrity in the leptomeningeal vasculature, at least in part, by reducing WNT signaling in ECs. Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 10 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources A 2 P A M U EC leptomeninges EC dura VWF+ EC dura VWF- EC artery EC mitotic control infected UMAP 1 D 200 Hmcn1 Slc7a1 Xist Tsix Slc6a6 B t n e c r e P y r a r b i l f o 50 40 30 20 10 0 − − ) c ( ) i ( i s e g n n e m o t p e l i s e g n n e m o t p e l C E C E Library JW19 JW20 JW21 JW22 JW23 − ) c ( y r e t r a C E − ) i ( y r e t r a C E − − − ) c ( + F W V a r u d C E ) i ( + F W V a r u d C E ) c ( - F W V a r u d C E − ) i ( - F W V a r u d C E C EC leptomeninges (c) EC leptomeninges (i) EC dura VWF+ (c) EC dura VWF+ (i) EC dura VWF- (c) EC dura VWF- (i) EC artery (c) EC artery (i) 1 f e L 8 p r L 1 a 7 c S l 5 a 7 c S l 5 n d C l l n c O x m B l 5 n b F 1 t a t S f w V p a v P l 2 o z e P i 3 n x r N 1 m a c e P 2 r f g e V / r d K Percent Expressed 20 40 60 80 Average Expression −1012 EC leptomeninges EC artery Mt1 Lcn2 30 Cp Cfh Pde4b Mt2 Il1r1 Csf3 Ttr Saa3 Kiz Selp -log (0.05) 10 Gm4951 Iigp1 20 Gm12216 Mllt3 Stat1 Cadps2 Tsix 10 Dkk2 Specc1 Cd74 Tanc2 -log (0.05) 10 0 Mt1 Lcn2 Ttr Saa3 Fmnl2 Cxcl1 Edn1 Csf3 Mctp1 Mt2 Apod −2 -1 0 1 2 4 EC dura VWF+ decrease with infection increase with infection no significant change −2 -1 0 1 2 4 EC dura VWF - Mt1 Lcn2 120 90 Xist Tsix Dach2 Iigp1 60 Slc6a6 Esrrg Eln 30 Slit3 Tmem108 0 Gm20663 Fmnl2 Mt2 Tspan18 Cp Kdr/Vegfr2 Pde4b Csf3 Ttr Saa3 Selp Cxcl1 Serpina3n -log (0.05) 10 Xist Tsix 80 Airn Iigp1 Igf2r Cd74 40 Rora Gab1 Klf12 Slc6a6 0 Gm20663 Piezo2 Kdr/Vegfr2 Fth1 Fkbp5 Slc39a14 Mt1 Lcn2 Mt2 Ttr Saa3 Csf3 Cxcl1 Selp Cp Sema3a -log (0.05) 10 Kiz −2.5 -1 0.0 1 2.5 5.0 -1 log (Average fold change) −2 2 0 1 2 4 Iqsec1 Kdr/Vegfr2 150 100 Dach2 Lrp8 Ptprm Rnf220 Tspan18 Spock2 50 0 ) d e t s u d a j l , e u a v - P ( 0 1 g o l - Kdr/Vegfr2 Kdr/Vegfr2 Stat1 Stat1 Cldn5 Cldn5 E 2 P A M U Control Lrp8 Control Slc7a1 Control Slc7a5 Control Nrxn3 Infected Lrp8 Infected Slc7a1 Infected Slc7a5 Infected Nrxn3 Control Cadm1 Control Piezo2 Control Plvap Infected Cadm1 Infected Piezo2 Infected Plvap Control Infected Ocln Ocln Control Infected Tjp1 (ZO-1) Tjp1 (ZO-1) Control Infected Lef1 Lef1 3 2 1 0 Control Infected Rela Rela Control Pecam1 Infected Pecam1 Control Infected Control Infected Control Infected UMAP 1 Figure 4. Changes in EC gene expression with infection in the leptomeninges and dura. (A) snRNAseq UMAP plots for ECs from combined control and infected meninges. (B) The number of nuclei from different EC subtypes is consistent across the five snRNAseq libraries (control: JW19, JW20; infected: JW21- JW23). (C) Dot plot showing changes in transcript abundances in EC subtypes in control (c) vs. infected (i) snRNAseq datasets. (D) Volcano Figure 4 continued on next page Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 11 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources Figure 4 continued plots for four EC subtypes showing control vs. infected snRNAseq transcript abundances. Xist and Tsix transcripts, referable to sex differences among embryos, are marked in red. Ttr transcripts, also marked in red, likely represent contamination from choroid plexus RNA and are present in 2/3 infected snRNAseq samples (see Supplementary file 6). (E) UMAP plots, as in (A), comparing control vs. infected snRNAseq for 15 genes. To assess the effects of infection on the organization of dura ECs, we took advantage of the large and unusually straight veins that occupy the dural sinuses adjacent to the skull’s sutures. Venous ECs throughout the body typically exhibit elongated nuclei that are aligned with the long axis of the vein, and therefore also with the direction of blood flow. With infection, this alignment is diminished in dural veins (Figure 6A and B), suggesting a general effect of infection on cytoskeletal organization within venous ECs. No assessment of permeability was made for the dura vasculature because the dura resides outside of the BBB territory (delimited by the arachnoid epithelial barrier) and, therefore, it exhibits high permeability in the control state. Microbial products, such as LPS, and any of a wide variety of cytokines could directly or indirectly alter EC structure and gene expression. As multiple immunologic stimuli are known to converge on the NF kappaB pathway and as the GSEA analysis implied that this pathway was induced upon infec- tion (Figure 2—figure supplement 4), we assessed NF kappaB signaling by generating reporter mice in which five tandem repeats of a canonical NF kappaB response element were inserted upstream of a minimal promoter to drive expression of a nuclear localized and 3xHA epitope- tagged tandem dimer Tomato from the Rosa26 locus (nls- tdT- 3xHA; Figure 6C). The parental version of this mouse line has the additional feature that a loxP- transcription stop- loxP cassette separates the promoter and the nls- tdT- 3xHA coding region, permitting cell- type specific read- out of NF kappaB signaling when crossed to a cell- type specific Cre transgene or knock- in allele. For the present analyses, we have used germline Cre- recombination to generate an allele that is predicted to be permissive for reporter expression in any cell type. In both dura and leptomeninges, NF kappaB reporter expression was induced by infection almost exclusively in ECs, with reporter positive EC nuclei increasing from  ~1%  to~20% in the dura and from  ~1%  to~5% in the leptomeninges. EC nuclei were identified based on ERG immunostaining (Figure 6D–F). As the the Rosa26 locus is generally permissive for expression in most, if not all, cell types, it was surprising that NF kappaB reporter expression was largely restricted to ECs. A second surprising feature was the cell- to- cell heterogeneity in EC expression, with reporter expressing ECs adjacent to non- expressing ECs in an apparently random pattern. The latter observation suggests substantial cell- to- cell heterogeneity in meningeal EC responses to infection, consistent with the observed broadening of infected EC clusters in the UMAP plots in Figure 4A and E. Genetic and pharmacologic perturbations of the immune response To explore the role of specific immune pathways in the vascular changes associated with infection, we applied the E. coli infection paradigm to mice with null mutations in (1) Tlr4, the gene coding for the innate immune system’s LPS receptor, or (2) Ccr2, the gene coding for one of the receptors for monocyte chemoattractant protein- 1 (MCP1/CCL2), a chemokine that recruits immune cells to sites of infection (Figure 7). Tlr4 is expressed widely among cell types within the meninges, whereas Ccr2 is expressed predominantly in monocytes and monocyte- derived cells (Figure 7—figure supplement 1; in the dot plot of all meningeal cells shown in Figure 7—figure supplement 1B, monocytes are included in the macrophage cluster). For this experiment, the WT comparator strain is C57BL/6  J, which matches the background of the Tlr4-/- and Ccr2-/- mice. The EC response to infection is milder in C57BL/6  J mice compared to FVB/NJ mice. At one day post- infection, the number of E. coli in leptomeninges flatmounts did not differ significantly between C57 WT control and Ccr2-/- mice, whereas the number of E. coli in leptomeninges flatmounts was increased  ~twofold in Tlr4-/- mice, albeit with substantial scatter in the data (Figure 1—figure supplement 3C, center plot). In response to infection, the leptomeningeal vasculature of Tlr4-/- mice showed minimal changes in the distribution of CLDN5 and very few regions of Sulfo- NHS- biotin leakage. In contrast, Ccr2-/- mice showed a clumpy redistribution of CLDN5 and localized regions of Sulfo- NHS- biotin leakage, much like the C57BL/6  J WT control (Figure  7A). Quantifying the area occupied by the leptomeningeal vasculature and the extent of Sulfo- NHS- biotin leakage confirmed this visual impression (Figure 7B). Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 12 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources Control Infected LPS (10 mg/kg) A 5 N D L C i n i t o b - S H N - o f l u S C l d o o b s e g n n e m o t p e L i ) % a e r a ( e g a r e v o c l e s s e v CLDN5 * Control Infected 80 70 60 50 40 30 80 70 60 50 40 PECAM1 CLDN5 * 100 Control Infected ** Control LPS 80 60 40 20 0 B 1 M A C E P i n i t o b - S H N - o f l u S L e p t o m e n n g e s i ( f l a t m o u n t ) ) % a e r a ( l a n g s i 80 60 i 40 n i t o b - S H N - o f l u 0S 20 Control Infected LPS (10 mg/kg) L e p t o m e n n g e s i ( f l a t m o u n t ) Control Infected LPS **** **** D Con. Inf. Con. Inf. CLDN5 GAPDH y t i s n e t n i t o b l n r e t s e W e v i t a e R l 1.5 1.0 0.5 0 CLDN5 n.s. Control Infected E 2 R F G E V 1 M A C E P F G R E 1 F E L G R E 1 F E L H Control Infected Control PECAM1 VEGFR2 PECAM1 VEGFR2 PECAM1 VEGFR2 PECAM1 VEGFR2 Infected A C o r t i c a l S u r f a c e Control Infected V V V A A A y t i s n e n t i e v i t l a e r 2 R F G E V s t n u o m t a l f i s e g n n e m o t p e l n i ) 1 M A C E P o t d e z i l a m r o n ( 2.5 2.0 1.5 1.0 0.5 0 A A A V V V **** n i y t i s n e t n i i g n n a i t s 1 F E L V V V V V V A A A A V A V A V ( f l t a m o u n t ) L e p t o m e n n g e s i 1 F E L 5 N D L C 1 M A C E P 1 M A C E P 1 F E L 5 N D L C 1 F E L L e p t o m e n n g e s i ( f l t a m o u n t ) Control Infected A V A V A V L e p t o m e n n g e s i ( f l t a m o u n t ) A A A V V V Con. Inf. Con. Inf. G LEF1 GAPDH ZO-1 n r e t s e W e v i t l a e R y t i s n e t n i t o b l LEF1 p=0.06 ZO-1 n.s. 1.5 1.0 0.5 Control Infected s t n u o m a l f t i s e g n n e m o t p e l ) s t i n u y r a r t i b r a ( l l e c r e p 5 4 3 2 1 0 **** **** n i y t i s n e t n i *** Artery Vein Capillary i i g n n a t s G R E s t n u o m a l f t i s e g n n e m o t p e l ) s t i n u y r a r t i b r a ( l l e c r e p 4 3 2 1 0 ** **** *** Control Infected Artery Vein Capillary Figure 5. Changes in EC morphology and EC protein abundance and localization with infection. (A) CLDN5 localization and Sulfo- NHS- biotin leakage in leptomeningeal vasculature following infection or LPS administration. (B) PECAM1 localization and Sulfo- NHS- biotin leakage in leptomeningeal vasculature following infection or LPS administration. (C) Infection or 10 mg/kg LPS treatment increases the area covered by vasculature in flatmounts of leptomeninges. (D) Immunoblotting shows a modest, but not statistically significant, reduction in CLDN5 level relative to GAPDH level in the Figure 5 continued on next page Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 13 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources Figure 5 continued leptomeninges with infection (n=4 independent experiments). (E) Reduced KDR (VEGFR2) immunstaining in leptomeninges ECs with infection. (F) Reduced EC nuclear LEF1 immunostaining in capillaries and veins, and reduced PECAM1 and CLDN5 staining in veins in the leptomeninges with infection. (G) Immunoblotting shows a modest reduction in LEF1 level and no change in ZO- 1 level relative to GAPDH level in the leptomeninges with infection (n=3 independent experiments for LEF1 and n=4 independent experiments for ZO- 1). (H) Quantification of images in E and F. For (D) and (G), the statistical signficance was calculated using the student’s t- test because the Wilcoxon rank sum test cannot be used on such small sample sizes. Scale bars: A and B, 100 µm; E and F, 100 µm. In control (i.e., uninfected) Tlr4-/- and Ccr2-/- mice, the density of leptomeningeal macrophages (as quantified by PU.1 immunostaining) was, respectively,~65%  and~90% of the WT value, and this density rose by ~50% in infected Tlr4-/- mice, but showed little or no change in infected Ccr2-/- or C57 WT mice (Figure 7C). Similar results were obtained by quantifying CD206 immunostaining, with Ccr2- /- leptomeninges showing ~80% of the C57 WT value (Figure 7—figure supplement 2 compares PU.1 and CD206 quantification). The lower baseline number of macrophages in the Tlr4-/- leptomeninges compared to WT could reflect a modestly higher level of macrophage proliferation in the WT driven by a basal level of constitutive TLR4 signaling, perhaps in response to LPS from the developing gut and/or skin microbiomes or from adult feces in the cage. The lower baseline number of macrophages in the Ccr2-/- leptomeninges compared to WT could reflect reduced monocyte ingress. As a complementary approach to gene inactivation, we used a single intra- cerebroventricular (ICV) injection of clodronate- containing liposomes to acutely and selectively eliminate leptomeningeal macrophages 2 days before E. coli infection (Figure 8A and B). The number of E. coli in the leptome- ninges did not differ significantly between mice that received control vs. clodronate liposomes (Figure 1—figure supplement 3C, right plot). Surprisingly, eliminating leptomeningeal macrophages had little or no effect on the infection- dependent redistribution of CLDN5 in leptomeningeal ECs, leakage of Sulfo- NHS- biotin, the increase in the area occupied by leptomeningeal vasculature, or the fractional increase in the number of ECs showing induction of the NF kappaB reporter (Figure 8B- D). The principal differences between responses of mice receiving control liposomes vs. clodronate liposomes were the modestly higher overall levels (i.e. both baseline and infected) following clodro- nate treatment of (1) the leptomeningeal blood vessel area (Figure  8D, second plot) and (2) the number of leptomeningeal ECs with NF kappaB reporter activation (expressed as HA+/ERG +ECs; Figure 8D third plot). These modest effects of clodronate treatment might represent inflammatory/ stress responses that arise from the death of large numbers of leptomeningeal macrophages and the accompanying release of bioactive substances. In support of the earlier inference that relatively few macrophages migrate into the leptomeninges 1 day after infection, the data in Figure 8C and D show that after clodronate depletion of CNS macrophages at P3, the number of macrophages in the leptomeninges at P6 increases only modestly following infection. Taken together, these experiments, together with the LPS treatment experiment (Figure 5A–C), show that (1) LPS stimulation of TLR4 signaling plays a central role in the response of the leptomenin- geal vasculature to infection (CLDN5 and PECAM1 redistribution, vessel swelling, and leakage), and (2) this vascular response is largely independent of leptomeningeal macrophages, by far the most abundant immune cells in the leptomeninges. Discussion The present study defines the responses of cells in the mouse leptomeninges and dura to bacterial meningitis in the early postnatal period. At this age, the immaturity of the adaptive immune system and the rapidity of infection imply that the host response depends largely, and perhaps exclusively, on the innate immune system. In response to bacterial infection, all of the major meningeal cell types – including ECs, macrophages, and fibroblasts – exhibit large and distinctive changes in their tran- scriptomes. In addition, ECs in leptomeningeal capillaries redistribute CLDN5 and PECAM1, leptome- ningeal capillaries become enlarged and disorganized and they exhibit foci of reduced BBB integrity, and ECs in leptomeningeal capillaries and veins lose nuclear LEF1. These capillary responses to infec- tion appear to be largely driven by TLR4 signaling, as determined by the response to LPS administra- tion and by the blunting of these responses to bacterial infection in the absence of TLR4. This simple and robust model of bacterial meningitis in infancy should prove useful in dissecting mechanisms of Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 14 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources A Control Infected B G R E 1 M A C E P D u r a ( f l a t m o u n t ) - C e n t r a l i s n u s **** a r u d e h t n i e c n e r e f f i d l r a u g n A e h t f o s x a i g n o l ( s u n s i l a r t n e c 100 ) n o i t c e r i d 80 i n e v l a r t n e c . s v s u e c u n l 60 40 20 0 Control Infected C Delta crystallin minimal promoter and intron loxP loxP Frt-Neo-3xpoly(A)-Frt (nls)tdT R26 locus Cre 3xHA Delta crystallin minimal promoter and intron loxP Frt-Neo-3xpoly(A)-Frt (nls)tdT 3xHA 5xNFKB response element 3xpoly(A) = “3xstop” bGH 3’ UTR 5xNFKB response element bGH 3’ UTR R26-NFKB-LSL-(nls)tdT-3xHA D G R E A H - i t n a A H - i t n a G R E E ) A H x 3 - T d t - s n ( l A H - i t n a 1 M A C E P Dura (flatmount) - Central sinus Leptomeninges (flatmount) Control Infected Control Infected Control A H - i t n a 1 M A C E P Infected 0.3F s C E + G R E / s C E + A H 0.2 0.1 0.0 ** Control Control Infected Infected ** Leptomeninges Dura (central sinus) Control Infected C e n t r a l i s n u s P e r i p h e r a l d u r a D u r a ( f l t a m o u n t ) Figure 6. Infection causes disorganization of EC nuclear orientation in the dural venous sinus, and an increase in NF- kappa B signaling in ECs in the leptomeninges and dura. (A) EC nuclei, visualized with ERG immunostaining, in the large vein of the central sinus. (B) Quantifying the orientation of the long axis of EC nuclei in the large vein of the central sinus, as shown in (A). Each data point represents one nucleus. (C) Structure of the NF- kappa B reporter before (left) and after (right) Cre- mediated recombination that removes a loxP- transcription stop- loxP (LSL) cassette. NF- kappa B Figure 6 continued on next page Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 15 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources Figure 6 continued reporter activation leads to expression of nls- tdT- 3xHA. (D) Infection increases expression of the NF- kappa B reporter in a subset of ECs in the dura and leptomeninges, as determined by immunostaining for HA. (E) In the dura, NF- kappa B reporter activation is observed in both ECs and non- ECs in the central sinus, but the NF- kappa B reporter is not activated in the peripheral dura. (F) Quantification of NF- kappa B reporter activation in ECs in the central sinus of the dura and in the leptomeninges. Scale bars: A, 50 µm; D and E, 100 µm. pathophysiology. In the future, it would be interesting to modify the mouse model by including antibi- otic treatment at different times after infection to parallel the clinical course of treated meningitis and to explore the neurologic sequelae that are commonly seen in human survivors of bacterial meningitis. This study has several limitations. First, the analyses focused on the response at 22 hr post- infection and earlier and later events have not been studied. Second, the movements, if any, of immune cells between compartments have not been studied, and therefore it is unclear whether some of the changes in transcriptome profiles should be ascribed to changes in cell state – that is the same cells as were present prior to infection, but with an altered transcriptional program – or the ingress of immune cells from non- meningeal pools. Interestingly, previous cell tracing analyses have shown that the pool of myeloid cells that supplies the meninges includes cells that are locally sequestered in the skull bones (Herisson et al., 2018; Cugurra et al., 2021). The modest increases in macrophage density in the leptomeninges following infection, with or without clodronate treatment (Figures 2E, 8C and D), imply a correspondingly modest influx of cells from non- meningeal reservoirs. Third, most of the tran- scriptomic changes are, at present, of unknown functional significance, and, more specifically, the clin- ical significance of the observed cellular and molecular changes is also not clear. Future work will be aimed at defining the physiologic significance of the observed transcriptome changes. For example, increases or decreases in the abundance of particular SLC transcripts (Figure 3—figure supplement 4) imply corresponding changes in the transmembrane movement of a defined set of small molecules. Comparisons with previously published single cell (sc) RNAseq data from mouse meninges empha- size the challenges of correlating data from tissues harvested at different ages, prepared by different methods, and subject to undefined batch effects. For example, in the present study we have divided dura fibroblasts into three clusters (Figure 1F, H and I), with clusters Fb- d1 and/or Fb- d3 likely derived from embryonic day (E)14 dura fibroblast clusters M4- 1 and M4- 2, as defined by DeSisto et al., 2020. Based on a comparison to scRNAseq of mouse coronal sutures dissected at E15.5 and E17.5 (Farmer et al., 2021), cluster Fb- d2 likely corresponds to Farmer et al.’s MG2 (Matn4 +and Nppc+, and occu- pying the outer dura) and Fb- d3 likely corresponds to Farmer et al.’s MG3 (Matn4- and Nppc+, and occupying the inner dura). Additional challenges attend comparisons between human and mouse meninges datasets, as scRNAseq of adult human dura reveals an even more complex landscape, with subdivision of dural fibroblasts into either six or 14 clusters, depending on the analysis method (Wang et al., 2022). Comparisons between P6 and adult macrophage subtypes in the mouse meninges, the latter defined by Van Hove et al., 2019, are challenging as adult meningeal macrophages were subdi- vided based on MHC class II (e.g. H2- Aa) transcript levels, which are uniformly low in P6 meningeal macrophages. The role of LPS in mediating BBB breakdown has been intensively studied, primarily in the context of the vasculature within the brain parenchyma (Wispelwey et  al., 1988; Banks et  al., 2015). In the brain, LPS treatment leads to a reduction in EC tight junctions secondary to a decrease in tight junction protein abundance and changes in tight junction protein localization (Peng et  al., 2021). In multiple cell types, TLR4 signaling (i.e. LPS- induced signaling) activates the NF- kappa B pathway, and that connection presumably accounts for the NF- kappa B reporter activation in meningeal ECs observed here. Although the infection- induced down- regulation of transcripts coding for cell- cell junction proteins in ECs, arachnoid barrier cells, and other cell types did not reach statistical signifi- cance (Figure 2—figure supplement 4), the altered localization of CLDN5 and the spatial correlation between this mis- localization and BBB disruption (e.g. Figure  5A) suggests that increased plasma membrane protein internalization and/or degradation could play a role in EC barrier defects. The NF- kappa B and WNT pathways can exhibit either positive and negative cross- regulation, depending on cell type and developmental context (Ma and Hottiger, 2016). By comparing BETA- CATENIN level, localization, and signaling in WT mouse embryo fibroblasts (MEFs) and in MEFs homozygous for inactivating mutations in IKKalpha or IKKbeta (the inflammation- activated kinases that phosphorylate the inhibitory binding partner of NF kappaB, leading to NF kappaB activation), Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 16 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources A CD206 CLDN5 Control CLDN5 CD206 CD206 CLDN5 Infected CLDN5 CD206 ) 7 5 C ( T W - / - 4 r l T - / - 2 r c C ) 7 5 C ( T W - / - 4 r l T - / - 2 r c C Control Infected CLDN5 Sulfo-NHS-biotin CLDN5 Sulfo-NHS-biotin CLDN5 Sulfo-NHS-biotin CLDN5 Sulfo-NHS-biotin B d o o b l i s e g n n e m o t p e L ) 0.8 % a e r a ( 0.6 e g a r e v o c l e s s e v 0.4 0.2 0.0 n.s. *** **** *** n.s. WT (C57) Tlr4 -/- Ccr2 -/- y t i s n e d e g a h p o r c a M ) 2 m m / s l l e c + 1 . U P ( 4000 3000 2000 1000 0 p=0.165 **** n.s. n.s. *** WT (C57) Tlr4 -/- Ccr2 -/- i n i t o b - S H N - o f l u S ) % a e r a ( l a n g s i 40 30 20 10 0 n.s. **** **** **** Control Infected n.s. WT (C57) Tlr4 -/- Ccr2 -/- Figure 7. Effects of Tlr4 KO and Ccr2 KO on leptomeningeal EC responses to infection. (A) Leptomeninges flatmounts of control vs. infected mice showing, in the upper panels, macrophage density (CD206) and vascular architecture (CLDN5) and, in the lower panels, vascular leakage (sulfo- NHS biotin). (B) Quantification of (left) vascular architecture based on CLDN5 immunostaining, (center) macrophage density, and (right) Sulfo- NHS- biotin leakage in WT, Tlr4-/-, and Ccr2-/- leptomeninges flatmounts in control vs. infected mice. Scale bar: A, 100 µm. Figure 7 continued on next page Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 17 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources Figure 7 continued The online version of this article includes the following figure supplement(s) for figure 7: Figure supplement 1. Expression of Tlr4 and Ccr2 in the major meningeal cell classes and in meningeal immune cells. Figure supplement 2. Comparison of leptomeningeal macrophage quantification by counting PU.1+vs. CD206 + cells, using as a test case the experiments presented in Figure 7. Lamberti et al., 2001 found that IKKalpha and IKKbeta phosphorylate BETA- CATENIN on different sites and with opposite effects. IKKalpha phosphorylation leads to BETA- CATENIN stabilization and increased WNT signaling, whereas IKKbeta phosphorylation leads to BETA- CATENIN destabilization and decreased WNT signaling. These precedents in other cell types suggest that reduced LEF1 – and presumably reduced canonical WNT signaling – in ECs in the infected leptomeninges could reflect BETA- CATENIN down- regulation via the activated NF kappaB pathway. Reduced canonical WNT signaling in ECs would be predicted to reduce BBB integrity (Rattner et al., 2022). The flat geometry of the meninges and its superficial location between the brain and the skull present unusually favorable opportunities for both in vitro and in vivo microscopy. For in vitro anal- yses, the relative thinness of the dissected mouse leptomeninges and dura provide excellent access to antibodies and also facilitate high quality confocal imaging without the need for chemical clearing agents, as described here. In vivo, thinned skull preparations that permit two- photon imaging of the mouse meninges have been described, and this approach can be applied to mice with fluorescent immune cells that have been inoculated with fluorescent bacteria to allow single- cell resolution in vivo imaging of bacterial meningitis in a native context (Kjos et  al., 2015; Coles et  al., 2017b; Mang- lani and McGavern, 2018). A largely unexplored opportunity also exists for ex vivo culture and live imaging of the dissected leptomeninges and dura (Glimcher et al., 2008). Such preparations could permit high- resolution imaging of (1) bacterial movement across the vascular wall, (2) interactions between bacteria and immune cells, and (3) interactions between host cells. The use of fluorescent reporters of signaling pathway activity would further increase the value of such analyses (Kudo et al., 2018; Clark et al., 2021). Despite decades of research, numerous gaps remain in our understanding of the pathophysiology of bacterial meningitis. These include: (1) the ways in which the imature immune system differs from the more mature immune system in its response to infection, (2) the relative importance and the precise roles of different immune cells and immune modulators, and (3) the roles played by changes in gene expression and cell behavior among non- immune cell types. In each of these areas, mouse models of meningitis, together with new technologies for interrogating these models, can provide insights that inform the understanding of human meningitis. Materials and methods Mice The following mouse lines were used: FVB/NJ (JAX#001800); C57BL/6  J (JAX#000664); Tie2- GFP (JAX#003658); B6(Cg)- Tlr4tm1.2Karp/J (JAX#029015); B6.129S4- Ccr2tm1Ifc/J (JAX#004999); and Rosa26- NF- kappaB reporter mice (described below). All mice were housed and handled according to the approved Institutional Animal Care and Use Committee protocol of the Johns Hopkins Medical Institutions. Meninges snRNA- seq experiments and histological studies used postnatal day 6 (P6) mice with age- matched controls. Construction and genotyping of the NF-kappa B reporter To construct the Cre- dependent reporter for NF- kappa B signaling at the Rosa26 locus, the following elements were inserted (from 5’ to 3’ in the order listed) into a standard Rosa26 targeting vector: an frt- phosphoglycerate kinase (Pgk)- neomycin (Neo)- frt (FNF) cassette, which includes a strong poly- adenylation signal; five tandem repeats of a canonical NF kappaB response element (GGGA CTTT CC); a minimal Delta Crystallin promoter followed by an intron; a loxP- transcription stop- loxP (LSL) cassette; an open reading frame coding for a nuclear- localization signal- tdTomato- 3xHA protein (nls- tdT- 3xHA); and a bovine growth hormone 3’UTR. The Rosa26- NF- kappaB- LSL- tdT targeting construct with a 3’ flanking Diphteria toxin- A coding sequence was electroporated into R1 ES cells, which were Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 18 of 27 Microbiology and Infectious Disease | Neuroscience l o r t n o C t e a n o r d o C l C l o r t n o C d e t c e f n I D y t i s n e d e g a h p o r c a M ) 2 m m / s l l e c + 6 0 2 D C ( 4000 3000 2000 1000 0 Tools and resources GS-lectin PECAM1 CD206 CD206 A ) V C I ( s e m o s o p i l ) V C I ( s e m o s o p i l B y t i s n e d e g a h p o r c a M ) 2 m m / s l l e c + 6 0 2 D C ( 4000 3000 2000 1000 0 Control liposomes (ICV) Clodronate (ICV) Clodronate (CM) **** **** * n.s. Dura (central sinus) Dura (no sinus) Leptomeninges Control liposomes (ICV) Clodronate liposomes (ICV) CLDN5 Sulfo-NHS-biotin CD206 VEGFR2 CLDN5 Sulfo-NHS-biotin CD206 VEGFR2 L e p t o m e n n g e s i ( f l a t m o u n t ) l o r t n o C d e t c e f n I Leptomeninges (flatmounts) Control Infected L e p t o m e n n g e s i ( f l a t m o u n t ) n.s. **** **** 100 80 60 40 20 0 ** ** Control liposomes Clodronate liposomes ) % a e r a ( e g a r e v o c l e s s e v d o o B l 100 80 60 40 20 0 ** *** ** *** 0.15 s l l e c l i l a + e h t o d n e + G R E 0.10 * 0.05 * * ** ) % a e r a ( l i a n g s n i t o b - S H N - o i / A H 0.00 Control liposomes Clodronate liposomes Control liposomes Clodronate liposomes f l u S Control liposomes Clodronate liposomes Figure 8. EC response to eliminating leptomeningeal macrophages with liposomal clodronate. (A) Intracerebroventricular (ICV) injection at P3 of empty liposomes vs. clodronate liposomes shows that clodronate almost completely eliminates leptomeningeal macrophages at P6, as visualized with CD206 immunostaining. (B) Quantification of leptomeningeal and dural macrophage abundance following ICV or cisterna magna (CM) injection of clodronate liposomes vs. control liposomes. (C) Leptomeninges flatmounts show greatly reduced numbers of macrophages in mice with or without infection following P3 treatment with clodronate liposomes, and there is little or no effect of macrophage depletion on CLDN5 relocalization and on Sulfo- NHS- biotin leakage in leptomeningeal ECs in response to infection. (D) Quantification in leptomeninges flatmounts from control vs. infected mice (from left to right): (1) macrophage density (2) vascular density, (3) NF- kappa B reporter activation in ECs (as shown in Figure 6F), and (4) Sulfo- NHS- biotin leakage. Mice received an ICV injection at P3 of empty liposomes or clodronate liposomes. CD206 immunostaining was used for macrophage quantification. Figure 7—figure supplement 2 shows that counting PU.1 or CD206 immunstained cells gives closely similar results in leptomeninges flatmounts. Scale bars: A, 500 µm (low magnification) and 50 µm (inset); B, 100 µm. Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 19 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources then subjected to G418 selection. Clones harboring the targeted Rosa26 locus were identified by Southern blot hybridization, karyotyped, and injected into blastocysts from Sv129 mice. Germline transmission to the progeny of founder males was determined by PCR. PCR primers for the parental allele: TGTC GGCC TGCA GCCA AAGC TTAT CGA (sense, at the 3’ end of the Neo casette) and TGAA GTTC TCAG GATC GGTC GCTA (antisense, in the intron). PCR primers for the Cre- recombined allele: CCCC TCTG CTAA CCAT GTTC ATGC CTT (sense, in the intron) and GGCA ACCT TCCT CTTC TTCT TAGG CATG GTGG (antisense, at the 5’ end of the nls- tdT- 3xHA open reading frame). Antibodies and other reagents The following antibodies were used for tissue immunohistochemistry and immunoblotting: goat anti- CD45 (R&D Systems AF114- SP); rat anti- CD206/MRC1 (Bio- Rad MCA2235T); goat anti- CD206 (R&D Systems AP2535); rat anti- LYVE1 (Thermo Fisher/eBioscience 14- 0443- 82); rat anti- PU.1/Spi- 1 (Novus Biologicals MAB7124); rat anti- CD180/RP105 antibody, PE (eBioscience 12- 1801- 81); goat anti- S100A8 (R&D Systems AF3059); rabbit anti- ERG (Cell Signaling Technologies 97249); rat anti- PECAM1/CD31 (BD Biosciences 553370); mouse anti- CLDN5, Alexa Fluor 488 conjugate (Invitrogen 352588); mouse anti- CLDN5 (Invitrogen 35–2500); rabbit anti- ZO1 (Invitrogen 40–2200); rabbit anti- LEF1 rabbit (Cell Signaling Technologies 2230); rabbit anti- LEF1, Alexa Fluor 647 conjugate (Cell Signaling Tech- nologies 14022); rabbit anti- Occludin (Invitrogen 406100); goat anti- VEGFR2/KDR (R&D Systems AF644- SP); sheep anti- FOXP2 (R&D Systems AF5647- SP); rabbit anti- SATB2 (Abcam ab92446); rat anti- IL- 6 (Biolegend 504501); rabbit anti- COL14A1 (Novus Biologicals NBP2- 15940); chicken anti- GFP (Abcam Ab13970), rat anti- HA (Proteintech 7c9); rabbit anti- HA (homemade); goat anti- E- cadherin (R&D Systems AF748); rabbit anti- E- cadherin (Cell Signaling Technologies 3195); rabbit anti- AIFM3 (Novus Biologicals NBP1- 76889); rabbit anti- COL25A1 (G- Biosciences ITT1021); goat anti- IGSF8 (R&D systems AF3117- SP); rabbit anti- NNAT (Abcam ab27266); mouse anti- GAPDH (Cell Signaling Technol- ogies 97166 S); rabbit anti- GAPDH (Cell Signaling Technologies 5174 S); Streptavidin, Alexa Fluor 488 conjugate (Invitrogen S11223); Streptavidin, Alexa Fluor 647 conjugate (Invitrogen S32357). Alexa- Fluor- conjugated secondary antibodies were from Invitrogen. Infrared immunoblotting secondary antibodies were from LI- COR. Other reagents used: Sulfo- NHS- biotin (Thermo Fisher Scientific #21217); LPS O111:B4 (Sigma- Aldrich L2630); Benzonase Nuclease, ultrapure (Sigma- Aldrich, E8263- 5KU); clodronate and control liposomes: Mannosylated Macrophage Depletion kit (Encapsula NanoScience SKU#CLD- 8914); Micro BCA Protein Assay kit (ThermoFisher Scientific 23235). E. coli infection and LPS injection E. coli strain RFP- RS218 (O18:K1:H7) with K1 capsule is a clinical isolate from the CSF of a neonate with meningitis (Zhu et al., 2020; a generous gift from the late Dr. Kwang Sik Kim, Johns Hopkins Medical School, Baltimore, MD). E. coli were grown overnight in Luria broth containing 100  µg/ml ampicillin at 37 °C. The following day, 400 µL of the E. coli culture was added to 20 mL of fresh Luria broth for an additional 2 hr of culture at 37 °C. The bacteria were washed one time in PBS, the OD at 620 nm was measured, and the concentration of the samples was adjusted prior to inoculation. For most experiments, E. coli meningitis was induced in FVB/NJ mice. For experiments with Tlr4- /- and Ccr2-/- mice, C57BL/6 J was used as the control to match the strain background of the KO lines. Briefly, a litter of P5 mice were randomly divided into two groups that were either subcutaneously injected in the back with 1.2×105 CFU of E. coli RFP- RS218 in 20 µL PBS or not injected. Twenty- two hours later, the mice were sacrificed as described. Tlr4-/- and Ccr2-/- on a C57BL/6 J background and Rosa26- NF- kappaB reporter mice on a mixed Sv129 x C57BL/6 J background were subcutaneously injected with 8×104  CFU of E. coli RFP- RS218. For LPS- injection, P5 mice were intraperitoneally injected with a single dose of LPS O111:B4 (10 mg/kg) or, for control mice, the same volume of PBS. Twenty- two hours later (at P6), the mice were sacrificed as described. All analyses (snRNAseq and histology) were conducted on mice that were sacrificed at P6. Meningeal macrophage depletion with clodronate liposomes CNS macrophages were depleted as described in Polfliet et al., 2001. Each mouse was injected with control or clodronate liposomes (3 µl of a 5 mg/ml stock solution) two days before being infected with E. coli or before receiving an LPS injection. Liposomes were allowed to warm from refrigerator Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 20 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources temperature to room temperature for 1 hr prior to injection. P3 mice were deeply anesthetized on ice and then slowly injected with liposomes using a Hamilton syringe (Hamilton Bonaduz, AG Switzerland) either in a lateral ventricle (intracerebroventricular administration) or in the cisterna magna. Tissue processing To prepare isolated dura and leptomeninges tissues for snRNAseq, immunoblots, and whole mount immunostaining, mice were deeply anesthetized on ice and then perfused via the cardiac route with PBS. The skullcaps (with dura attached) and brain (with leptomeninges attached) were dissected in PBS. The brain was then chilled in ice- cold PBS for several minutes. Using a fine tweezers, the leptomeninges was gently peeled from the surface of the brain and either used for protein extraction, or immersion fixed in 2% paraformaldehyde (PFA)/PBS at room temperature for 1 hr for subsequent immunostaining. The dura was gently peeled from the skullcap and immersion fixed in 2% parafor- maldehyde (PFA)/PBS at room temperature for 1 hr for subsequent immunostaining. Alternately, the skullcap and attached dura were immersion fixed overnight in 2% PFA/PBS at 4 °C without further dissection. Following immersion fixation, each sample was washed three times in PBS. For vibratome sections, the brain and attached leptomeninges was fixed overnight in 2% PFA/PBS at 4 °C, washed the following day in PBS at 4 °C for at least 3 hr, embedded in 3% agarose, and sectioned at 150 µm thickness using a vibratome (Leica). For analysis of vascular leakage, P6 mice were injected intraperitoneally with Sulfo- NHS- biotin (30 µl of 20 mg/ml Sulfo- NHS- biotin in PBS per mouse) 10–15 min before sacrifice. After IP injection, the tracer rapidly equilibrates into the systemic circulation. Mice were deeply anesthetized on ice and perfused via the cardiac route with PBS followed by leptomeninges dissection and PFA fixation, as described above. For cross- sections of isolated dura and leptomeninges, fixed tissues were embedded in optimal cutting temperature compound (OCT, Tissue- Tek), rapidly frozen in dry ice, and stored at –80  °C. Thirty µm sections were cut on a cryostat and thaw- mounted onto Superfrost plus slides. Slides were stored at –80 °C until further processing. Immunohistochemistry Whole mount leptomeninges, dura, or brain sections were incubated overnight at 4 °C with primary antibodies diluted with PBSTC (PBS with 1% Triton X- 100, 0.1 mM CaCl2) plus 10% normal goat serum (NGS). Tissues were washed four times with PBSTC over the course of 6–8 hours, and then incubated overnight at 4 °C with secondary antibodies diluted in PBSTC plus 10% NGS. If a primary rat anti- body was used, secondary antibodies were additionally incubated with 0.5% normal mouse serum as a blocking agent. The following day, tissues were washed at least four times with PBSTC over the course of 6–8 hr, flat- mounted on Superfrost Plus glass slides (Fisher Scientific), and coverslipped with Fluoromount G (EM Sciences 17984–25). For leptomeninges or dura cross- sections, sections on slides were covered with 2% PFA/PBS at room temperature for 15 min, washed three times in PBS, and incubated overnight with primary anti- bodies diluted in PBSTC plus 10% NGS at 4 °C. The following day, sections were washed at least four times with PBSTC and incubated with secondary antibodies for 2 hr at room temperature. Sections were then washed four times with PBSTC and coverslipped with Fluoromount G. For each immunos- taining analysis, whole mounts and sections were stained from at least two independent experiments. Confocal microscopy Confocal images were captured with a Zeiss LSM700 confocal microscope (20x and 63x objectives) using Zen Black 2012 software, and processed with Fiji- ImageJ, Adobe Photoshop, and Adobe Illus- trator. The depths of the Z- stacked flatmount images from the confocal series were chosen to capture the full thickness of the tissue: (1) 40–60 µm for the dura sinus region, (2) 15–20 µm for the dura periph- eral region, and (3) 20–30 µm leptomeninges. Each point in the quantification of immunofluorescent data represents the analysis of a single Z- stacked confocal image from a flatmount that encompasses the full depth of the tissue (leptomeninges or dura), unless noted otherwise. For Sulfo- NHS- biotin detection with streptavidin, ~twofold animal- to- animal variability is typically seen in the overall inten- sity of the strepatvidin signal, most likely due to variable uptake of Sulfo- NHS- biotin from the site of IP Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 21 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources injection. To permit a clearer comparison between images of Sulfo- NHS- biotin leakage, this intensity variation has been minimized by manually adjusting the brightness of the Sulfo- NHS- biotin channel. Leptomeninges tissue lysate To prepare leptomeninges proteins for western blotting, anesthetized mice were perfused with PBS and their brains were dissected in PBS. The leptomeninges was detached from the surface of the brain tissue with tweezers and then transferred into a 1.5 mL eppendorf tube and stored at –80 °C for further processing. Frozen leptomeninges from two mice were pooled, lysed in lysate buffer (50 mM Tris- HCl (pH, 7.4), 150  mM NaCl, 2  mM MgCl2, 1% Triton X- 100, 0.25  U/µL Benzonase), and then homogenized using a plastic pestle fitted for eppendorf tubes. The homogenates were incubated for 10 min at room temperature to digest the nuclear DNA, and then SDS was added to a final concen- tration of 0.5%. The lysate in SDS was incubated for 20 min at 4 °C and then centrifuged at 14,000 x g for 15 min at 4 °C. The supernatant was recovered and its protein concentration was determined using the BCA protein assay kit. Immunoblotting Protein samples (6–8 µg per sample) were loaded onto a 4–12% NuPAGE Bis- Tris protein gel, which was run at 130 V for 1.5 h and then blotted onto a nitrocellulose membrane (Millipore). Membranes were blocked with Intercept blocking buffer (LI- COR 927–60001) at room temperature for 1 hr and then probed overnight with primary antibodies diluted in blocking buffer at 4 °C. The following day, the membranes were washed four times with TBST and then probed at room temperature for 2 h with the corresponding infrared secondary antibodies diluted in blocking buffer. Bands were visualized with an Odyssey Fc Imager (LI- COR) and band intensities were quantified with Fiji- Image J software. snRNAseq Two and three independent biological replicate libraries were prepared for the control and E. coli- infection groups, respectively, with one P6 mouse used per library. For each sample, the dura and leptomeninges were rapidly dissected in ice- cold DPBS (Gibco 14287072). The combined dura and leptomeninges were minced with a razor blade and Dounce homogenized using a loose- fitting pestle in 5 mL homogenization buffer (0.25 M sucrose, 25 mM KCl, 5 mM MgCl2, 20 mM Tricine- KOH,pH 7.8) supplemented with 1  mM DTT, 0.15  mM spermine, 0.5  mM spermidine, EDTA- free protease inhibitor (Roche 11836 170 001), and 60 U/mL RNasin- Plus RNase Inhibitor (Promega N2611). A 5% IGEPAL- 630 solution was added to bring the homogenate to 0.3% IGEPAL CA- 630, and the sample was further homogenized with ten strokes of a tight- fitting pestle. The sample was filtered through a 50 μm filter (CellTrix, Sysmex, 04- 004- 2327), underlayed with solutions of 30% and 40% iodixanol (Sigma D1556) in homogenization buffer, and centrifuged at 10,000×g for 18 min in a swinging bucket centrifuge at 4 °C. Nuclei were collected at the 30–40% interface, diluted with two volumes of homog- enization buffer, and concentrated by centrifugation for 10 min at 500xg at 4 °C. snRNAseq libraries were constructed using the 10 x Genomics Chromium single- cell 3’ v3 kit following the manufacturer’s protocol (https:// support. 10xgenomics. com/ single- cell- gene- expression/ library- prep/ doc/ user- guide- chromium- single- cell- 3- reagent- kits- user- guide- v31- chemistry). Libraries were sequenced on an Illumina NovaSeq 6000. Analysis of snRNAseq data Reads were aligned to the mm10 pre- mRNA index using the Cell Ranger count program, version 3.1.0. The data for the different libraries was merged using the Cell Ranger merge command. Data analysis was performed using the Seurat R package, version 4.0.1 in RStudio. After filtering out nuclei with >1% mitochondrial transcripts or with <500 or>6000 genes, 48,941 nuclei were retained, 14,357 nuclei from the two control samples and 34,585 nuclei from the three infected samples. The data were normalized using a regularized negative binomial regression algorithm implemented in the SCTransform function as described in Hafemeister and Satija, 2019. UMAP dimensional reduction was performed using the R uwot package (https://github.com/jlmelville/uwot) integrated into the Seurat R package (Melville, 2022). To compare cell types across treatments, the data was integrated using the strategy described in Stuart and Satija, 2019. This pipeline involves split- ting the dataset by treatment using the Seurat SplitObject function and integrating the subset Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 22 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources objects using FindIntegrationAnchors and IntegrateData functions. Data for the various scatter plots were extracted using the Seurat AverageExpression function, and differential gene expression was analyzed using the Seurat FindMarkers function. The Wilcoxon Rank Sum test was used to calculate p- values. The p- values were adjusted with a Bonferroni correction using all genes in the dataset. Data exploration, analysis, and plotting were performed using RStudio (RStudio Team, 2020), the tidyverse collection of R packages (Wickham, 2017), and ggplot2 (Wickham, 2009). Dotplots were generated with the default settings, including default normalization. For Gene set enrichment analysis (GSEA; Subramanian et al., 2005), genes were ranked by the fold expression change between control and infected datasets. The ranked gene list was used to detect enriches gene sets within the Broad Insti- tute Hallmark Gene Sets using the fgsea R package (https://github.com/ctlab/fgsea; Korotkevich et al., 2019). Macrophage quantification in the meninges The density of macrophages was quantified using Fiji- ImageJ (https:// imagej. net/ software/ fiji/) from captured Z- stacked leptomeninges or dura wholemounts. The numbers of PU.1+or  CD206+ cells were quantified in representative regions and then normalized to the area analyzed. Cell orientation analysis Using Fiji- ImageJ, the axis of blood flow of the midline vein located in the superior sagittal sinus was defined as 0°, and then the long axis of each EC nucleus, identified by ERG immunostaining, was scored and its angle calculated relative to 0°. Nuclei were analyzed from three independent control vs. infection experiments. Fraction of area covered by blood vessels The relative area covered by leptomeningeal blood vessels was determined from Z- stacked flatmount images that had been immunostained with CLDN5 or PECAM1. More specifically, flatmount images of representative regions that were populated by capillaries, but not by large veins or arteries, were overlayed in Adobe Illustrator with arrays of 10 parallel and evenly- spaced straight white lines. The length of each line corresponds to 166 µm on the image and the first and tenth lines in each array are separated by a distance that corresponds to 150 µm on the image. Thus, the 10 lines define a 166 µm x 150 µm rectangle. Along each white line, the widths of all of the regions in the image that were not covered by blood vessels were manually scored by drawing (in Adobe Illustrator) a straight line across the vessel- free region. When all of the vessel- free line segments had been drawn for a given square array, the sum of their lengths was calculated with Fiji- ImageJ and divided by the sum of the lengths of the ten white lines to generate an estimate for the fraction of the area not covered by blood vessels. The fraction of the area covered by blood vessels equals one minus the fraction of the area not covered by blood vessels. Each data point in a blood vessel area plot represents the area estimate from one set of white lines, that is the estimate obtained from sampling a length of 10x166 µm=1.66 mm. Statistical analysis All statistical values are presented as mean ± SD. The number of mice used for each sample are listed in Supplementary file 4. The Wilcoxon rank sum test was used to measure statistical significance, except for Figure 5D and G, in which the sample size is too small and the student t- test was used instead. Statistical tests were carried out using the following web sites: https://www.socscistatistics. com/tests/signedranks/default2.aspx and https://www.omnicalculator.com/statistics/wilcoxon-rank- sum-test#how-do-i-calculate-wilcoxon-rank-sum-test. The statistical significance is represented graph- ically as n.s., not significant (i.e. p>0.05); *, p<0.05; **, p<0.01; ***, p<0.001; ****, p<0.0001. Acknowledgements Supported by the Howard Hughes Medical Institute. The authors thank Dr. Latika Nagpal for helpful comments on the manuscript. The authors thank the reviewers and the editors for excellent comments that improved the manuscript. Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 23 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources Additional information Funding Funder Howard Hughes Medical Institute Grant reference number Author Jeremy Nathans Jie Wang Amir Rattner National Eye Institute R01EY018637 Jeremy Nathans The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Author contributions Jie Wang, Conceptualization, Formal analysis, Investigation, Methodology, Writing - original draft, Writing - review and editing; Amir Rattner, Formal analysis, Investigation, Writing - review and editing; Jeremy Nathans, Conceptualization, Formal analysis, Supervision, Funding acquisition, Writing - orig- inal draft, Project administration, Writing - review and editing Author ORCIDs Amir Rattner Jeremy Nathans http://orcid.org/0000-0001-9542-6212 http://orcid.org/0000-0001-8106-5460 Ethics All mice were housed and handled strictly according to the approved Institutional Animal Care and Use Committee protocol of the Johns Hopkins Medical Institutions (Protocol MO19M429). Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.86130.sa1 Author response https://doi.org/10.7554/eLife.86130.sa2 Additional files Supplementary files •  Supplementary file 1. snRNAseq library statistics. •  Supplementary file 2. Criteria for assigning cell type clusters based on snRNAseq transcript profiles. •  Supplementary file 3. Differential transcript abundances among the major meningeal cell types in infected vs. control snRNAseq datasets. FC, fold change. •  Supplementary file 4. Number of mice in each group in the image quantifications. •  Supplementary file 5. Differential transcript abundances among CCL2- and CCL2+ resident macrophages in infected vs. control snRNAseq datasets. FC, fold change. •  Supplementary file 6. Differential transcript abundances among EC cell types in infected vs. control snRNAseq datasets. FC, fold change. •  MDAR checklist Data availability Sequencing data have been deposited in GEO. The following datasets were generated: Author(s) Wang J, Rattner A, Nathans J Year 2022 Continued on next page Dataset title Dataset URL Database and Identifier Bacterial meningitis in the early postnatal mouse studied at single- cell resolution https://www. ncbi. nlm. nih. gov/ geo/ query/ acc. cgi? acc= GSE221678 NCBI Gene Expression Omnibus, GSE221678 Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 24 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources Continued Author(s) Wang J, Rattner A, Nathans J Year 2022 Dataset title Dataset URL Database and Identifier snRNAseq_JW19_ meninges_control_RP1 Wang J, Rattner A, Nathans J 2022 snRNAseq_JW20_ meninges_control_RP2 Wang J, Rattner A, Nathans J 2022 snRNAseq_JW21_ meninges_infected_RP1 Wang J, Rattner A, Nathans J 2022 snRNAseq_JW22_ meninges_infected_RP2 Wang J, Rattner A, Nathans J 2022 snRNAseq_JW23_ meninges_infected_RP3 https://www. ncbi. nlm. nih. gov/ geo/ query/ acc. cgi? acc= GSM6892910 https://www. ncbi. nlm. nih. gov/ geo/ query/ acc. cgi? acc= GSM6892911 https://www. ncbi. nlm. nih. gov/ geo/ query/ acc. cgi? acc= GSM6892912 https://www. ncbi. nlm. nih. gov/ geo/ query/ acc. cgi? acc= GSM6892913 https://www. ncbi. nlm. nih. gov/ geo/ query/ acc. cgi? acc= GSM6892914 NCBI Gene Expression Omnibus, GSM6892910 NCBI Gene Expression Omnibus, GSM6892911 NCBI Gene Expression Omnibus, GSM6892912 NCBI Gene Expression Omnibus, GSM6892913 NCBI Gene Expression Omnibus, GSM6892914 References Alves de Lima K, Rustenhoven J, Kipnis J. 2020. Meningeal immunity and its function in maintenance of the central nervous system in health and disease. Annual Review of Immunology 38:597–620. DOI: https://doi.org/ 10.1146/annurev-immunol-102319-103410, PMID: 32340575 Arac A, Grimbaldeston MA, Nepomuceno ARB, Olayiwola O, Pereira MP, Nishiyama Y, Tsykin A, Goodall GJ, Schlecht U, Vogel H, Tsai M, Galli SJ, Bliss TM, Steinberg GK. 2014. Evidence that meningeal mast cells can worsen stroke pathology in mice. The American Journal of Pathology 184:2493–2504. DOI: https://doi.org/10. 1016/j.ajpath.2014.06.003, PMID: 25134760 Banks WA, Gray AM, Erickson MA, Salameh TS, Damodarasamy M, Sheibani N, Meabon JS, Wing EE, Morofuji Y, Cook DG, Reed MJ. 2015. Lipopolysaccharide- induced blood- brain barrier disruption: roles of cyclooxygenase, oxidative stress, neuroinflammation, and elements of the neurovascular unit. Journal of Neuroinflammation 12:223. DOI: https://doi.org/10.1186/s12974-015-0434-1, PMID: 26608623 Barichello T, Fagundes GD, Generoso JS, Elias SG, Simões LR, Teixeira AL. 2013. Pathophysiology of neonatal acute bacterial meningitis. Journal of Medical Microbiology 62:1781–1789. DOI: https://doi.org/10.1099/jmm. 0.059840-0, PMID: 23946474 Clark HR, McKenney C, Livingston NM, Gershman A, Sajjan S, Chan IS, Ewald AJ, Timp W, Wu B, Singh A, Regot S. 2021. Epigenetically regulated digital signaling defines epithelial innate immunity at the tissue level. Nature Communications 12:1836. DOI: https://doi.org/10.1038/s41467-021-22070-x, PMID: 33758175 Coles JA, Myburgh E, Brewer JM, McMenamin PG. 2017a. Where are we? The anatomy of the murine cortical meninges revisited for Intravital imaging, Immunology, and clearance of waste from the brain. Progress in Neurobiology 156:107–148. DOI: https://doi.org/10.1016/j.pneurobio.2017.05.002, PMID: 28552391 Coles JA, Stewart- Hutchinson PJ, Myburgh E, Brewer JM. 2017b. The mouse cortical meninges are the site of immune responses to many different pathogens, and are accessible to Intravital imaging. Methods 127:53–61. DOI: https://doi.org/10.1016/j.ymeth.2017.03.020, PMID: 28351758 Coureuil M, Lécuyer H, Bourdoulous S, Nassif X. 2017. A journey into the brain: insight into how bacterial pathogens cross blood- brain barriers. Nature Reviews. Microbiology 15:149–159. DOI: https://doi.org/10. 1038/nrmicro.2016.178, PMID: 28090076 Cugurra A, Mamuladze T, Rustenhoven J, Dykstra T, Beroshvili G, Greenberg ZJ, Baker W, Papadopoulos Z, Drieu A, Blackburn S, Kanamori M, Brioschi S, Herz J, Schuettpelz LG, Colonna M, Smirnov I, Kipnis J. 2021. Skull and vertebral bone marrow are myeloid cell reservoirs for the meninges and CNS parenchyma. Science 373:eabf7844. DOI: https://doi.org/10.1126/science.abf7844, PMID: 34083447 Derk J, Jones HE, Como C, Pawlikowski B, Siegenthaler JA. 2021. Corrigendum: living on the edge of the CNS: meninges cell diversity in health and disease. Frontiers in Cellular Neuroscience 15:761506. DOI: https://doi. org/10.3389/fncel.2021.761506, PMID: 34690706 DeSisto J, O’Rourke R, Jones HE, Pawlikowski B, Malek AD, Bonney S, Guimiot F, Jones KL, Siegenthaler JA. 2020. Single- cell transcriptomic analyses of the developing meninges reveal meningeal fibroblast diversity and function. Developmental Cell 54:43–59.. DOI: https://doi.org/10.1016/j.devcel.2020.06.009, PMID: 32634398 Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 25 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources Eisen DP, Hamilton E, Bodilsen J, Køster- Rasmussen R, Stockdale AJ, Miner J, Nielsen H, Dzupova O, Sethi V, Copson RK, Harings M, Adegboye OA. 2022. Longer than 2 hours to antibiotics is associated with doubling of mortality in a multinational community- acquired bacterial meningitis cohort. Scientific Reports 12:672. DOI: https://doi.org/10.1038/s41598-021-04349-7, PMID: 35027606 Farmer DT, Mlcochova H, Zhou Y, Koelling N, Wang G, Ashley N, Bugacov H, Chen HJ, Parvez R, Tseng KC, Merrill AE, Maxson RE, Wilkie AOM, Crump JG, Twigg SRF. 2021. The developing mouse coronal suture at single- cell resolution. Nature Communications 12:4797. DOI: https://doi.org/10.1038/s41467-021-24917-9, PMID: 34376651 Gaschignard J, Levy C, Romain O, Cohen R, Bingen E, Aujard Y, Boileau P. 2011. Neonatal bacterial meningitis: 444 cases in 7 years. The Pediatric Infectious Disease Journal 30:212–217. DOI: https://doi.org/10.1097/inf. 0b013e3181fab1e7, PMID: 21416693 GBD 2016 Meningitis Collaborators. 2018. Global, regional, and national burden of meningitis, 1990- 2016: a systematic analysis for the global burden of disease study 2016. The Lancet Neurology 17:1061–1082. DOI: https://doi.org/10.1016/S1474-4422(18)30387-9 Glimcher SA, Holman DW, Lubow M, Grzybowski DM. 2008. Ex vivo model of cerebrospinal fluid outflow across human arachnoid granulations. Investigative Ophthalmology & Visual Science 49:4721–4728. DOI: https://doi. org/10.1167/iovs.08-2238, PMID: 18614809 Hafemeister C, Satija R. 2019. Normalization and variance stabilization of single- cell RNA- seq data using regularized negative binomial regression. Genome Biology 20:296. DOI: https://doi.org/10.1186/s13059-019- 1874-1, PMID: 31870423 Herisson F, Frodermann V, Courties G, Rohde D, Sun Y, Vandoorne K, Wojtkiewicz GR, Masson GS, Vinegoni C, Kim J, Kim DE, Weissleder R, Swirski FK, Moskowitz MA, Nahrendorf M. 2018. Direct vascular channels connect skull bone marrow and the brain surface enabling myeloid cell migration. Nature Neuroscience 21:1209–1217. DOI: https://doi.org/10.1038/s41593-018-0213-2, PMID: 30150661 Holsapple MP, West LJ, Landreth KS. 2003. Species comparison of anatomical and functional immune system development. Birth Defects Research. Part B, Developmental and Reproductive Toxicology 68:321–334. DOI: https://doi.org/10.1002/bdrb.10035, PMID: 14666995 Jackson SH, Heininger JA. 1975. Proline recycling during collagen metabolism as determined by concurrent 18O2- and 3H- labeling. Biochimica et Biophysica Acta 381:359–367. DOI: https://doi.org/10.1016/0304-4165( 75)90241-x, PMID: 1111595 Kierdorf K, Masuda T, Jordão MJC, Prinz M. 2019. Macrophages at CNS interfaces: ontogeny and function in health and disease. Nature Reviews. Neuroscience 20:547–562. DOI: https://doi.org/10.1038/s41583-019- 0201-x, PMID: 31358892 Kim KS, Wass CA, Cross AS. 1997. Blood- brain barrier permeability during the development of experimental bacterial meningitis in the rat. Experimental Neurology 145:253–257. DOI: https://doi.org/10.1006/exnr.1997. 6458, PMID: 9184127 Kjos M, Aprianto R, Fernandes VE, Andrew PW, van Strijp JAG, Nijland R, Veening J- W. 2015. Bright fluorescent Streptococcus pneumoniae for live- cell imaging of host- pathogen interactions. Journal of Bacteriology 197:807–818. DOI: https://doi.org/10.1128/JB.02221-14, PMID: 25512311 Kohil A, Jemmieh S, Smatti MK, Yassine HM. 2021. Viral meningitis: an overview. Archives of Virology 166:335– 345. DOI: https://doi.org/10.1007/s00705-020-04891-1, PMID: 33392820 Korotkevich G, Sukhov V, Sergushichev A. 2019. Fast gene set enrichment analysis. bioRxiv. DOI: https://doi. org/10.1101/060012 Ku LC, Boggess KA, Cohen- Wolkowiez M. 2015. Bacterial meningitis in infants. Clinics in Perinatology 42:29–45, . DOI: https://doi.org/10.1016/j.clp.2014.10.004, PMID: 25677995 Kudo T, Jeknić S, Macklin DN, Akhter S, Hughey JJ, Regot S, Covert MW. 2018. Live- cell measurements of kinase activity in single cells using translocation reporters. Nature Protocols 13:155–169. DOI: https://doi.org/ 10.1038/nprot.2017.128, PMID: 29266096 Kuper CF, van Bilsen J, Cnossen H, Houben G, Garthoff J, Wolterbeek A. 2016. Development of immune organs and functioning in humans and test animals: implications for immune intervention studies. Reproductive Toxicology 64:180–190. DOI: https://doi.org/10.1016/j.reprotox.2016.06.002, PMID: 27282947 Lamberti C, Lin KM, Yamamoto Y, Verma U, Verma IM, Byers S, Gaynor RB. 2001. Regulation of beta- catenin function by the IkappaB Kinases. The Journal of Biological Chemistry 276:42276–42286. DOI: https://doi.org/ 10.1074/jbc.M104227200, PMID: 11527961 Last JA, Summers P, Reiser KM. 1989. Biosynthesis of collagen Crosslinks II. In vivo labelling and stability of lung collagen in rats. Biochimica et Biophysica Acta 990:182–189. DOI: https://doi.org/10.1016/s0304-4165(89) 80032-7, PMID: 2492830 Ma B, Hottiger MO. 2016. Crosstalk between Wnt/β-catenin and NF- B signaling pathway during inflammation. Frontiers in Immunology 7:378. DOI: https://doi.org/10.3389/fimmu.2016.00378, PMID: 27713747 κ Manglani M, McGavern DB. 2018. Intravital imaging of neuroimmune interactions through a thinned skull. Current Protocols in Immunology 120:24. DOI: https://doi.org/10.1002/cpim.46, PMID: 29512146 Masuda T, Amann L, Monaco G, Sankowski R, Staszewski O, Krueger M, Del Gaudio F, He L, Paterson N, Nent E, Fernández- Klett F, Yamasaki A, Frosch M, Fliegauf M, Bosch LFP, Ulupinar H, Hagemeyer N, Schreiner D, Dorrier C, Tsuda M, et al. 2022. Specification of CNS macrophage subsets occurs postnatally in defined niches. Nature 604:740–748. DOI: https://doi.org/10.1038/s41586-022-04596-2, PMID: 35444273 McGill F, Heyderman RS, Panagiotou S, Tunkel AR, Solomon T. 2016. Acute bacterial meningitis in adults. Lancet 388:3036–3047. DOI: https://doi.org/10.1016/S0140-6736(16)30654-7, PMID: 27265346 Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 26 of 27 Microbiology and Infectious Disease | Neuroscience Tools and resources Melville J. 2022. Uwot: the uniform manifold approximation and projection (UMAP) method for Dimensionality reduction. 0.1.14. R Package. https://cran.r-project.org/web/packages/uwot/index.html Mrdjen D, Pavlovic A, Hartmann FJ, Schreiner B, Utz SG, Leung BP, Lelios I, Heppner FL, Kipnis J, Merkler D, Greter M, Becher B. 2018. High- dimensional single- cell mapping of central nervous system immune cells reveals distinct myeloid subsets in health, aging, and disease. Immunity 48:380–395. DOI: https://doi.org/10. 1016/j.immuni.2018.01.011, PMID: 29426702 Park JE, Jardine L, Gottgens B, Teichmann SA, Haniffa M. 2020. Prenatal development of human immunity. Science 368:600–603. DOI: https://doi.org/10.1126/science.aaz9330, PMID: 32381715 Peltola H, Roine I, Kallio M, Pelkonen T. 2021. Outcome of childhood bacterial meningitis on three continents. Scientific Reports 11:21593. DOI: https://doi.org/10.1038/s41598-021-01085-w, PMID: 34732790 Peng X, Luo Z, He S, Zhang L, Li Y. 2021. Blood- brain barrier disruption by lipopolysaccharide and sepsis- associated encephalopathy. Frontiers in Cellular and Infection Microbiology 11:768108. DOI: https://doi.org/ 10.3389/fcimb.2021.768108, PMID: 34804998 Polfliet MM, Goede PH, van Kesteren- Hendrikx EM, van Rooijen N, Dijkstra CD, van den Berg TK. 2001. A method for the selective depletion of perivascular and meningeal macrophages in the central nervous system. Journal of Neuroimmunology 116:188–195. DOI: https://doi.org/10.1016/s0165-5728(01)00282-x, PMID: 11438173 Rattner A, Wang Y, Nathans J. 2022. Signaling pathways in neurovascular development. Annual Review of Neuroscience 45:87–108. DOI: https://doi.org/10.1146/annurev-neuro-111020-102127, PMID: 35803586 Roberts WG, Palade GE. 1995. Increased microvascular permeability and endothelial fenestration induced by vascular endothelial growth factor. Journal of Cell Science 108:2369–2379. DOI: https://doi.org/10.1242/jcs. 108.6.2369, PMID: 7673356 Rua R, McGavern DB. 2018. Advances in meningeal immunity. Trends in Molecular Medicine 24:542–559. DOI: https://doi.org/10.1016/j.molmed.2018.04.003, PMID: 29731353 Sabbagh MF, Heng JS, Luo C, Castanon RG, Nery JR, Rattner A, Goff LA, Ecker JR, Nathans J. 2018. Transcriptional and epigenomic landscapes of CNS and non- CNS vascular endothelial cells. eLife 7:e36187. DOI: https://doi.org/10.7554/eLife.36187, PMID: 30188322 Senger DR, Galli SJ, Dvorak AM, Perruzzi CA, Harvey VS, Dvorak HF. 1983. Tumor cells secrete a vascular permeability factor that promotes accumulation of ascites fluid. Science 219:983–985. DOI: https://doi.org/10. 1126/science.6823562, PMID: 6823562 Shane AL, Sánchez PJ, Stoll BJ. 2017. Neonatal sepsis. Lancet 390:1770–1780. DOI: https://doi.org/10.1016/ S0140-6736(17)31002-4, PMID: 28434651 Stuart T, Satija R. 2019. Integrative single- cell analysis. Nature Reviews. Genetics 20:257–272. DOI: https://doi. org/10.1038/s41576-019-0093-7, PMID: 30696980 Subramanian A, Tamayo P, Mootha VK, Mukherjee S, Ebert BL, Gillette MA, Paulovich A, Pomeroy SL, Golub TR, Lander ES, Mesirov JP. 2005. Gene set enrichment analysis: a knowledge- based approach for interpreting genome- wide expression profiles. PNAS 102:15545–15550. DOI: https://doi.org/10.1073/pnas.0506580102, PMID: 16199517 Uiterwijk A, Koehler PJ. 2012. A history of acute bacterial meningitis. Journal of the History of the Neurosciences 21:293–313. DOI: https://doi.org/10.1080/0964704X.2011.595653, PMID: 22724490 Van Hove H, Martens L, Scheyltjens I, De Vlaminck K, Pombo Antunes AR, De Prijck S, Vandamme N, De Schepper S, Van Isterdael G, Scott CL, Aerts J, Berx G, Boeckxstaens GE, Vandenbroucke RE, Vereecke L, Moechars D, Guilliams M, Van Ginderachter JA, Saeys Y, Movahedi K. 2019. A single- cell atlas of mouse brain macrophages reveals unique transcriptional identities shaped by ontogeny and tissue environment. Nature Neuroscience 22:1021–1035. DOI: https://doi.org/10.1038/s41593-019-0393-4, PMID: 31061494 Wang AZ, Bowman- Kirigin JA, Desai R, Kang LI, Patel PR, Patel B, Khan SM, Bender D, Marlin MC, Liu J, Osbun JW, Leuthardt EC, Chicoine MR, Dacey RG, Zipfel GJ, Kim AH, DeNardo DG, Petti AA, Dunn GP. 2022. Single- cell profiling of human dura and meningioma reveals cellular meningeal landscape and insights into meningioma immune response. Genome Medicine 14:49. DOI: https://doi.org/10.1186/s13073-022-01051-9, PMID: 35534852 Weller RO, Sharp MM, Christodoulides M, Carare RO, Møllgård K. 2018. The meninges as barriers and facilitators for the movement of fluid, cells and pathogens related to the rodent and human CNS. Acta Neuropathologica 135:363–385. DOI: https://doi.org/10.1007/s00401-018-1809-z, PMID: 29368214 Wickham H. 2009. Ggplot2: Elegant Graphics for Data Analysis Springer- Verlag. DOI: https://doi.org/10.1007/ 978-0-387-98141-3 Wickham H. 2017. Tidyverse: easily install and load the ‘Tidyverse’.’ 1.2.1. R Package. https://CRAN.R-project. org/package=tidyverse Williamson PR, Jarvis JN, Panackal AA, Fisher MC, Molloy SF, Loyse A, Harrison TS. 2017. Cryptococcal meningitis: epidemiology, Immunology, diagnosis and therapy. Nature Reviews. Neurology 13:13–24. DOI: https://doi.org/10.1038/nrneurol.2016.167, PMID: 27886201 Wispelwey B, Lesse AJ, Hansen EJ, Scheld WM. 1988. Haemophilus influenzae lipopolysaccharide- induced blood brain barrier permeability during experimental meningitis in the rat. The Journal of Clinical Investigation 82:1339–1346. DOI: https://doi.org/10.1172/JCI113736, PMID: 3262627 Zhu N, Liu W, Prakash A, Zhang C, Kim KS. 2020. Targeting E. coli invasion of the blood- brain barrier for investigating the pathogenesis and therapeutic development of E. coli meningitis. Cellular Microbiology 22:e13231. DOI: https://doi.org/10.1111/cmi.13231, PMID: 32447809 Wang et al. eLife 2023;12:e86130. DOI: https:// doi. org/ 10. 7554/ eLife. 86130 27 of 27 Microbiology and Infectious Disease | Neuroscience
10.1021_acs.jmedchem.3c00912
pubs.acs.org/jmc This article is licensed under CC-BY 4.0 Article Discovery of a First-in-Class Degrader for the Lipid Kinase PIKfyve Chungen Li,◆ Yuanyuan Qiao,◆ Xia Jiang, Lianchao Liu, Yang Zheng, Yudi Qiu, Caleb Cheng, Fengtao Zhou, Yang Zhou, Weixue Huang, Xiaomei Ren, Yuzhuo Wang, Zhen Wang,* Arul M. Chinnaiyan,* and Ke Ding* Cite This: J. Med. Chem. 2023, 66, 12432−12445 Read Online ACCESS Metrics & More Article Recommendations *sı Supporting Information ABSTRACT: The phosphoinositide kinase PIKfyve has emerged as a new potential therapeutic target in various cancers. However, limited clinical progress has been achieved with PIKfyve inhibitors. Here, we report the discovery of a first-in-class PIKfyve degrader 12d (PIK5-12d) by employing the proteolysis-targeting chimera approach. PIK5-12d potently degraded PIKfyve protein with a DC50 value of 1.48 nM and a Dmax value of 97.7% in prostate cancer VCaP cells. Mechanistic studies revealed that it selectively induced PIKfyve degradation in a VHL- and proteasome-dependent manner. PIKfyve degradation by PIK5-12d caused massive cytoplasmic vacuolization and blocked autophagic flux in multiple prostate cancer cell lines. Importantly, PIK5-12d was more effective in suppressing the growth of prostate cancer cells than the parent inhibitor and exerted prolonged inhibition of downstream signaling. Further, intraperitoneal administration of PIK5-12d exhibited potent PIKfyve degradation and suppressed tumor proliferation in vivo. Overall, PIK5-12d is a valuable chemical tool for exploring PIKfyve-based targeted therapy. 1. INTRODUCTION PIKfyve is a phosphoinositide kinase that is characterized by the presence of a FYVE finger-containing domain structure.1 As a lipid kinase, PIKfyve phosphorylates phosphatidylinositol- 3-phosphate (PI(3)P) to produce phosphatidylinositol-3,5- bisphosphate (PI(3,5)P2), which is crucial to maintain endomembrane homeostasis.2 PIKfyve plays a critical role in regulating membrane homeostasis, endosomal trafficking, and system.3−6 autophagy in the endosomal and lysosomal Accumulating evidence suggests that PIKfyve is a potential therapeutic target for various human cancers.7−9 For example, shRNA knockdown of PIKfyve induced cytoplasmic vacuoliza- tion in dividing cells and suppressed cell proliferation.10 Apilimod (1, Figure 1), a potent and highly selective PIKfyve inhibitor, effectively inhibited the proliferation of B-cell non- Hodgkin lymphoma cells.11 Our previous work also demon- strated that the inhibition of PIKfyve could suppress autophagy and potentiate response to immune checkpoint blockade in prostate cancer.12 several other Several selective small-molecule inhibitors of PIKfyve have been disclosed and YM201636 (2) and 1 were the most well- characterized examples (Figure 1).11,13 Compound 2, an initially discovered PIKfyve inhibitor, potently inhibited the kinase activity with an IC50 value of 33 nM and was selective lipid kinase family members.13−15 over Compound 1, as mentioned above, was another small- molecule PIKfyve inhibitor with both good kinase inhibitory activity (IC50 = 14 nM) and high kinome selectivity.11 Compound 1 has been advanced into clinical trials for the treatment of lymphoma, autoimmune diseases, neurodegener- ative diseases, and COVID-19 disease.11,16−19 However, due to compound stability issues in vivo, the effects in clinical trials have been limited. Therefore, the discovery of new PIKfyve modulators is highly desirable and thus we explored the potential of PIKfyve-based targeted therapy. Proteolysis targeting chimera (PROTAC), a heterobifunc- tional molecule recruiting protein-of-interest (POI) to the E3 ligase and inducing POI degradation by the ubiquitin- proteasome system (UPS), has become a novel paradigm for drug discovery.20 The POI degradation mediated by PROTACs is a catalytic and event-driven process, which is Received: May 22, 2023 Published: August 21, 2023 Figure 1. Chemical structures of selected PIKfyve inhibitors. © 2023 The Authors. Published by American Chemical Society 12432 https://doi.org/10.1021/acs.jmedchem.3c00912 J. Med. Chem. 2023, 66, 12432−12445 Journal of Medicinal Chemistry pubs.acs.org/jmc Article Figure 2. Design of PIKfyve PROTACs. (A) Docking model of compound 1 with PIKfyve protein (PDB: 7K2V); (B) chemical structures of compound 1 and the designed PIKfyve PROTACs. Scheme 1. Synthesis of Compounds 7a−7ja aReagents and conditions: (a) tert-butyl-N-(2-hydroxyethyl)carbamate, NaH, N, N-dimethylformamide (DMF), 0 °C, and 8 h; (b) hydrazine hydrate, 1,4-dioxane, 90 °C, and 12 h; (c) m-tolualdehyde, acetic acid (AcOH), ethyl alcohol (EtOH), reflux, and 6 h; (d) (i) trifluoroacetic acid (TFA), dichloromethane (CH2Cl2), rt., and 3 h; (ii) 2-(7-azabenzotriazol-1-yl)-N′, N′, N′-tetramethyluronium hexafluorophosphate (HATU), triethylamine (Et3N), DMF, rt., and 5 h; (iii) TFA, CH2Cl2, rt., and 3 h; (e) HATU, Et3N, DMF, rt., and 5 h. Scheme 2. Synthesis of Compounds 12a−12j, 13a, and 12dNa aReagents and conditions: (a) tert-butyl-(4-(2-hydroxyethyl)phenyl)carbamate, NaH, DMF, 0 °C, and 6 h; (b) hydrazine hydrate, 1,4-dioxane, 90 °C, and 12 h; (c) m-tolualdehyde, AcOH, EtOH, reflux, and 6 h; (d) (i) TFA, CH2Cl2, rt., and 3 h; (ii) HATU, Et3N, DMF, rt., and 5 h; (iii) TFA, CH2Cl2, rt., and 3 h; (e) HATU, Et3N, DMF, rt., and 5 h. usually more efficient than the inhibitor occupancy. Mean- while, PROTACs can deplete both the catalytic and non- catalytic functions of the kinase, potentially outperforming kinase inhibitors.21 Herein, we report the discovery of the first series of PIKfyve PROTAC degraders. The optimal compound 12d (PIK5-12d) showed potent degradative activity against PIKfyve with a DC50 value of 1.48 nM and a Dmax value of in prostate cancer VCaP cells. It also 97.7%, respectively, 12433 https://doi.org/10.1021/acs.jmedchem.3c00912 J. Med. Chem. 2023, 66, 12432−12445 Journal of Medicinal Chemistry pubs.acs.org/jmc Article exhibited promising PIKfyve degradative effects in vivo. Importantly, PIK5-12d exerted prolonged inhibition of PIKfyve downstream signaling and outperformed the parent PIKfyve inhibitor in the suppression of the growth of prostate cancer cells. 2. RESULTS AND DISCUSSION 2.1. Design of PIKfyve PROTAC Degraders. We chose to utilize compound 1, a PIKfyve inhibitor in Phase II clinical trials, as the “warhead” to develop PIKfyve PROTAC degraders. The computational modeling studies revealed that the pyridyl moiety of compound 1 extended to the solvent- exposed region (Figure 2A). The reported structure−activity relationship (SAR) studies on compound 1 also indicated that the pyridyl group was well tolerated to various substituents.22 Based on these observations, a series of PIKfyve degraders were designed by tethering compound 1 to a ligand for the E3 ligase von Hippel−Lindau (VHL) via a diverse set of linkers (Figure 2B). 2.2. Chemical Synthesis. The synthesis of compounds 7a−7j is depicted in Scheme 1. The starting material 4-(2,6- dichloropyrimidin-4-yl)morpholine went through a regioselec- tive nucleophilic substitution with tert-butyl-N-(2- hydroxyethyl)carbamate to produce compound 3, which reacted with hydrazine hydrate to give compound 4. The subsequent condensation reaction of compound 4 with m- tolualdehyde afforded compound 5, which was acylated by a series of linkers after the deprotection of N-Boc to yield key intermediates 6a−6j. Subsequent deprotection of tert-butyl ester of compounds 6a−6j and a further amide coupling reaction with the classical VHL ligand produced final compounds 7a−7j. Compounds 12a−12j, 13a, and 12dN were synthesized by the same reactions and conditions used in the synthesis of compounds 7a−7j (Scheme 2). The major difference was that intermediate 8 was constructed from tert-butyl-(4-(2- hydroxyethyl)phenyl) carbamate instead of tert-butyl-N-(2- hydroxyethyl)carbamate. 2.3. Structure−Degradation Relationship Study of PIKfyve Degraders. Based on the design strategy noted above, we first obtained a set of PIKfyve degraders by connecting the VHL ligand directly to the central pyrimidine ring via linkers of different lengths (Table 1). The degradation efficiency of these degraders was assessed by immunoblotting assays in VCaP cells with high expression levels of PIKfyve.12 The results showed that only compounds with longer linkers displayed good PIKfyve degradative activities. For example, compounds 7i and 7j achieved degradation rates of 72 and 67% at 0.1 μM, for PIKfyve, while other compounds with shorter linkers were much less active. Interestingly, compounds 7i and 7j displayed decreased degradative effect on PIKfyve at 1.0 μM, which may be due to the hook effects.23 respectively, We reasoned that although the pyridyl group of compound 1 was exposed to the solvent area, its aromatic ring structure may still contribute to its binding with PIKfyve protein. Thus, the second series of degraders were designed by connecting the VHL ligand to the phenyl ring that retained the aromatic ring character of the pyridyl group (Table 2). Investigation of the linker length revealed that compound 12d (PIK5-12d) with 4 −CH2− in the middle of the linker showed the best PIKfyve degradative activity with the degradation rates of 97 and 91% at 0.1 and 1 μM, respectively. Interestingly, the substitution of Table 1. Degradation Efficiency of Compounds 7a−7ja compds linker (n) 0.1 μM 1.0 μM % degradation (VCaP) 7a 7b 7c 7d 7e 7f 7g 7h 7i 7j 0 1 2 3 4 5 6 7 8 9 0 0 0 0 3 0 30 32 72 67 0 0 0 10 0 51 17 31 32 21 aDegradation efficiency was determined by immunoblotting after treatment with compounds in VCaP cells for 24 h. Table 2. Degradation Efficiency of PIKfyve Degraders 12a− 12j, 13a, and 12dNa compds R1 R2 % degradation (VCaP) linker (n) 0.1 μM 1.0 μM 12a 12b 12c 12d (PIK5-12d) 12e 12f 12g 12h 12i 12j 13a 12dN (PIK5-12dN) H H H H H H H H H H S-methyl H R-hydroxy R-hydroxy R-hydroxy R-hydroxy R-hydroxy R-hydroxy R-hydroxy R-hydroxy R-hydroxy R-hydroxy R-hydroxy S-hydroxy 0 1 2 3 4 5 6 7 8 9 3 3 0 0 0 97 90 26 84 81 64 48 47 0 0 8 0 91 75 19 92 45 41 3 44 0 aDegradation efficiency was determined by immunoblotting after treatment with compounds in VCaP cells for 24 h. the VHL ligand in PIK5-12d with a more potent version resulted in compound 13a with decreased activity.24 We also synthesized compound 12dN (PIK5-12dN) as a negative control by using the inactive isomer of the VHL ligand,25 and it turned out to have no degradative activity against PIKfyve as expected (Table 2). 2.4. Compound PIK5-12d Selectively Induced the Degradation of PIKfyve Protein in a Concentration-, Time-, VHL-, and Proteasome-Dependent Manner. PIK5-12d displayed the most potent degradative effects on 12434 https://doi.org/10.1021/acs.jmedchem.3c00912 J. Med. Chem. 2023, 66, 12432−12445 Journal of Medicinal Chemistry pubs.acs.org/jmc Article Figure 3. PIK5-12d in a concentration-dependent fashion reduced PIKfyve protein in multiple human prostate cancer cells in vitro. (A) Immunoblotting of PIKfyve and GAPDH in VCaP cells treated with increasing concentrations of PIK5-12d for 24 h (left), percent remaining PIKfyve protein was plotted for DC50 and Dmax determination (right); (B) immunoblotting of PIKfyve and GAPDH in VCaP with 100 nM PIK5- 12d for various timepoints (left), percent remaining PIKfyve protein was quantified (right); (C) global proteomic analysis of PIK5-12d in VCaP cells after 4 h treatment of DMSO or 300 nM PIK5-12d (left), mass-spec quantification of PIKfyve protein (middle), lipid kinase changes (right); and (D) immunoblotting of PIKfyve and GAPDH in multiple human prostate cancer cell lines with increasing concentrations of PIK5-12d for 24 h. Figure 4. PIK5-12d mediated-PIKfyve degradation was VHL and proteasome-dependent. (A) Immunoblotting of PIKfyve and GAPDH in VCaP cells treated with increasing concentrations of 12d (PIK5-12d), 12f, or 12h in the conditions of with or without 1 μM of proteasome inhibitor bortezomib for 24 h; (B) immunoblotting of PIKfyve and GAPDH in VCaP cells treated with increasing concentration of PIK5-12d with or without warhead 1 or VHL ligand VL285 for 24 h. PIKfyve in the preliminary screening. To further characterize this compound, we treated VCaP cells with PIK5-12d with different concentrations for 24 h. The results showed that compound PIK5-12d dose-dependently induced PIKfyve degradation with a DC50 value of 1.48 nM and a Dmax value of 97.9% (Figure 3A). The kinetics experiments revealed that PIK5-12d at a concentration of 100 nM caused fast degradation of PIKfyve protein with a t1/2 value of 1.5 h (Figure 3B). We further performed global proteomic analysis using tandem mass tags (TMT) labeled mass-spectrometry to unbiasedly quantify the protein change upon PIK5-12d treatment in VCaP cells. The results indicated that PIK5- 12d is a very specific PIKfyve degrader with only 3 proteins significantly downregulated including PIKfyve, which accounts for the off-target rate at 2 out of 7573 detectable proteins. It is worth noting that PIK5-12d is selective for PIKfyve over other In addition, PIK5-12d also lipid kinases (Figure 3C). effectively reduced PIKfyve in other prostate cancer PC3, LNCaP, and 22RV1 cells (Figure 3D). We further investigated the mechanism of PIKfyve degradation by PIK5-12d in VCaP cells. Western blot results showed that PIK5-12d and its analogues (12f and 12h) at 0.1 and 1 μM can degrade PIKfyve protein to different extents, in VCaP cells after treatment for 24 h, while pretreatment with the proteasome inhibitor bortezomib completely rescued the level of PIKfyve protein (Figure 4A). In addition, both the warhead compound 1 and VHL ligand VL285 competitively blocked the PIKfyve degradation by PIK5-12d (Figure 4B).26 These results demonstrated that PIK5-12d induced PIKfyve degradation in a VHL- and proteasome-dependent manner. 2.5. PIK5-12d Induced Massive Cytoplasmic Vacuoli- zation and Blocked Autophagy in Prostate Cancer Cells. Previous work showed that the PIKfyve inhibition could induce cytoplasmic vacuolization and block autophagy.12 Thus, 12435 https://doi.org/10.1021/acs.jmedchem.3c00912 J. Med. Chem. 2023, 66, 12432−12445 Journal of Medicinal Chemistry pubs.acs.org/jmc Article we investigated the effects of PIK5-12d on these phenotypes in prostate cancer DU145 cells. The results showed that PIK5- 12d has comparable vacuolization induction ability to inhibitor 1 in a dose-dependent manner in RFP-labeled DU145 cells (Figure 5A). In addition, PIK5-12d also concentration- dependently increased autophagy markers LC3A/B and p62 in different prostate cancer cell lines (Figure 5B). Figure 5. PIK5-12d induced massive cytoplasmic vacuolization and blocked autophagy. (A) Representative images of DU145-RFP cells with treatments of DMSO, 300 nM PIK5-12d, or 1 for 24 h. The is quantified in indicated conditions; (B) vacuole area per cell immunoblotting of PIKfyve and autophagy markers in indicated human prostate cancer cells with increasing concentrations of PIK5- 12d for 24 h. 2.6. PIK5-12d Decreased Prostate Cancer Cell Proliferation and Exerted Prolonged Suppression of PIKfyve Downstream Signaling. PROTACs usually have prolonged therapeutic effects due to the irreversible depletion of the target protein.27 Thus, we conducted an in vitro washout experiment where VCaP cells were incubated with PIK5-12d or 1 for 24 h before compounds were removed. Cells were then cultured in compound-free media for 2 weeks. The results showed that PIK5-12d inhibited VCaP cell proliferation with an IC50 of 522.3 nM, which is over twofold lower than that of compound 1 (Figure 6A). The anti-proliferative effects were significantly enhanced when VCaP cells were incubated with the compounds for 2 weeks, and PIK5-12d was found to be ∼5 times more potent than compound 1 in this experiment. PIK5-12d also outperformed the negative control PIK5-12dN in long-term anti-proliferation assay after 4 h treatment and followed by culturing in a drug-free medium for 2 weeks the (Figure 6B). We further compounds on PIKfyve and its downstream signaling using the washout approach. In the experiment, VCaP cells were incubated with PIK5-12d, PIK5-12dN, or 1 for 4 h, and compounds were then removed. Cells were further cultured in compound-free media for 72 h. It was shown that PIK5-12d significantly reduced PIKfyve and increased LC3A/B at 300 nM, while the negative control PIK5-12dN and inhibitor 1 did not show any effect on these proteins (Figure 6C). It is worth noting that the vacuolization ability triggered by PIK5-12d, negative control PIK5-12dN, and inhibitor 1 at 4 h has no investigated the effect of significant difference (Figure 6D). These results collectively suggested that PIK5-12d exerted prolonged suppression of prostate cancer cell proliferation and PIKfyve downstream signaling. 2.7. PIK5-12d Depleted PIKfyve Protein and Sup- pressed Tumor Proliferation In Vivo. We further performed a pharmacodynamic assessment of PIK5-12d in an LTL-331R human prostate cancer patient-derived xenograft (PDX) model. PIK5-12d was administrated by intraperitoneal (IP) injection with two doses of 4 and 10 mg/kg, respectively, for 3 days. The tumor tissues were harvested on day 4 and subjected to western blot analysis (Figure 7A). As shown in Figure 7B, PIK5-12d almost completely depleted PIKfyve protein at both doses compared to the vehicle control group, indicating its strong PIKfyve degradation efficiency in vivo. In addition, the depletion of PIKfyve protein by PIK5-12d also triggered tumor cell death (Figure 7C). Then, long-term tumor efficacy of the LTL-331R model was determined by once-daily administration of PIK5-12d at 5 days on and 2 days off regimen for 17 days. It was shown that PIK5-12d significantly suppressed tumor proliferation in vivo (Figure 7D). These results collectively suggested the promising therapeutic potential of PIKfyve degradation for prostate cancer treatment. 3. CONCLUSIONS for multiple types of cancer The lipid kinase PIKfyve has been increasingly recognized as a therapeutic target including multiple myeloma, prostate cancer, non-Hodgkin lymphoma, and other diseases such as neurodegenerative disorders and SARS-CoV-2 infection. Although a number of small-molecule inhibitors have been developed, only a few of them advanced in clinical development such as compound 1. Published data suggested that compound 1 has low plasma stability which limited its in vivo efficacy.1,11 Thus, there is an urgent need for the development of other modulators to target PIKfyve. PROTAC, as a new modulator type, has become a novel paradigm for kinase drug discovery. PROTAC often out- performed kinase inhibitors because it not only functions in a catalytic and event-driven manner but can disrupt both the enzymatic and scaffolding functions of the kinase. Here, we report the discovery of a first-in-class PIKfyve PROTAC degrader PIK5-12d. PIK5-12d effectively degraded PIKfyve with a DC50 value of 1.48 nM and a Dmax value of 97.7% in prostate cancer VCaP cells. Mechanistic studies showed that PIK5-12d induced PIKfyve degradation through the VHL- and proteasome-dependent manner. PIK5-12d also induced massive cytoplasmic vacuolization and blocked autophagy in prostate cancer cells. Importantly, PIK5-12d exerted pro- longed suppression of prostate cancer cell proliferation and PIKfyve downstream signaling compared to the kinase inhibitor 1. In addition, PIK5-12d exhibited potent PIKfyve degradation effects, triggered tumor cell death, and suppressed this study tumor proliferation in vivo. Taken together, discovered a first-in-class PIKfyve degrader as the valuable tool compound for the research community and provided the proof-of-concept for the degradation of PIKfyve as a promising therapeutic approach for prostate cancer as well as other cancers. 4. EXPERIMENTAL SECTION 4.1. General Methods for Chemistry. The reagents and solvents used in chemical synthesis were obtained from 12436 https://doi.org/10.1021/acs.jmedchem.3c00912 J. Med. Chem. 2023, 66, 12432−12445 Journal of Medicinal Chemistry pubs.acs.org/jmc Article Figure 6. PIK5-12d decreased prostate cancer cell proliferation and exerted prolonged suppression of PIKfyve downstream signaling. (A) Long- term cell viability is visualized by crystal violet staining in VCaP either treated with PIK5-12d or 1 for 24 h and chased for 2 weeks, or continuous treatment for 2 weeks (top), IC50s were calculated for indicated conditions (bottom); (B) long-term cell viability is determined in VCaP cells with washout in drug-free medium after 4 h treatment of PIK5-12d or negative control PIK5-12dN; and (C) immunoblotting of PIKfyve, LC3A/B, and GAPDH in whole cell lysate of VCaP cells treated with PIK5-12d, negative control PIK5-12dN, or inhibitor 1 for 4 h and chased in drug-free medium for 72 h. (D) Quantification of vacuole per cell in DU145-RFP cells for 4 h of treatment of PIK5-12d, negative control PIK5-12dN, or inhibitor 1. commercial agents without further purification. All reactions were monitored by using thin-layer chromatography (TLC). All final compounds were purified by a column chromatog- raphy on silica gel (300−400 mesh). The NMR spectra were recorded on Agilent DD2 500 spectrometer (Agilent Technologies Inc., USA) or Bruker AVANCE 600 spectrom- eter (Bruker Company, Germany) in CDCl3 or DMSO-d6. The spectra of high-resolution mass (HRMS) were monitored by Bruker MaXis 4G TOF mass spectrometer. The purities of all final compounds were identified by HPLC analysis with the Agilent 1200 system and were proved to be >95%. HPLC condition: Triart C18 reversed-phase column, 5 μm, 4.6 mm × 250 mm, and flow rate 1.0 mL/min, starting with a 15 min- gradient from 0.1% TFA in water and acetonitrile 1:9 mixture to 0.1% TFA in acetonitrile, then ending with 0.1% TFA in acetonitrile for 5 min. 4.1.1. tert-Butyl (2-((4-Chloro-6-morpholinopyrimidin-2- yl)oxy)ethyl)carbamate (3). NaH (0.2 g, 7.8 mmol, 60% dispersion in mineral oil) was added into a solution of tert- butyl-N-(2-hydroxyethyl)carbamate (0.5 g, 3.1 mmol) in anhydrous DMF (10 mL) at 0 °C. The resulting mixture was stirred at 0 °C for 30 min. Then, 4-(2,6-dichloropyrimidin- 4-yl)morpholine (0.7 g, 3.1 mmol) dissolved in anhydrous DMF (5 mL) was added dropwise. The reaction solution was stirred at the same temperature for a further 8 h, then was diluted with 50 mL of water to precipitate a white solid, which was filtered and washed with water three times (50 mL per wash). The resulting solid was dried to give a mixture of isomers tert-butyl (2-((4-chloro-6-morpholinopyrimidin-2-yl)- oxy)ethyl)carbamate and tert-butyl (2-((2-chloro-6-morpholi- nopyrimidin-4-yl)oxy)ethyl)carbamate and was purified by column chromatography [petroleum ether (PE)/ethyl acetate (EA)] to give specific intermediate 3 (600 mg, 54%). 1H NMR (500 MHz, DMSO-d6) δ 6.97 (t, J = 5.8 Hz, 1H), 6.60 (s, 1H), 4.16 (t, J = 5.8 Hz, 2H), 3.66−3.55 (m, 8H), 3.22 (q, J = 5.8 Hz, 2H), 1.35 (s, 9H). 4.1.2. tert-Butyl (2-((4-Hydrazineyl-6-morpholinopyrimi- din-2-yl)oxy)ethyl)carbamate (4). A solution of intermediate 3 (0.5 g, 1.4 mmol) and 2.5 mL of 50% hydrazine hydrate solution dissolved in dioxane (15 mL) was stirred at 90 °C for 12437 https://doi.org/10.1021/acs.jmedchem.3c00912 J. Med. Chem. 2023, 66, 12432−12445 Journal of Medicinal Chemistry pubs.acs.org/jmc Article Figure 7. PIK5-12d depleted PIKfyve protein in vivo. (A) Study design of pharmacodynamic assessment on target engagement of PIKfyve PROTAC degrader in LTL-331R human prostate cancer patient-derived xenograft model; (B) immunoblotting of PIKfyve, LC3A/B, and GAPDH in the whole cell lysate of human LTL-331R tumors treated with vehicle or various concentrations of PIK5-12d for 4 days; (C) representative images of TUNEL signal of LTL-331R tumors treated with vehicle or 10 mg/kg PIK5-12d for 4 days; and (D) tumor proliferation of LTL-331R tumors treated with vehicle or 15 mg/kg PIK5-12d at 5 days on and 2 days off regimen for 17 days, the p value was determined by two-tailed unpaired t test between vehicle and PIK5-12d groups. 12 h. Intermediate 3 was exhausted by a monitoring of TLC. reduced The reaction solution was concentrated under pressure to obtain a white solid, which was washed with 50 mL of water and dried to give the intermediate 4 (470 mg, 95%). 1H NMR (500 MHz, DMSO-d6) δ 8.96 (brs, 2H), 7.65 (s, 1H), 6.94 (t, J = 5.7 Hz, 1H), 4.05 (t, J = 6.0 Hz, 2H), 3.64−3.58 (m, 4H), 3.38 (t, J = 4.8 Hz, 4H), 3.18 (q, J = 6.0 Hz, 2H), 1.72 (s, 3H), 1.35 (s, 9H). 4.1.3. tert-Butyl (E)-(2-((4-(2-(3-Methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)- carbamate (5). Intermediate 4 (400 mg, 1.1 mmol) and 3- methylbenzaldehyde (142 mg, 1.2 mmol) were dissolved in anhydrous ethanol (10 mL). Then, a catalytic amount of acetic acid was added. The resulting solution was refluxed for 6 h, cooled to room temperature, and concentrated under reduced pressure to obtain a white solid, which was resuspended into a mixture solution of PE (5 mL) and CH2Cl2 (5 mL), then was filtered to give intermediate 5 (450 mg, 90%). 1H NMR (500 MHz, DMSO-d6) δ 10.82 (s, 1H), 7.99 (s, 1H), 7.48 (d, J = 7.7 Hz, 1H), 7.46 (s, 1H), 7.27 (t, J = 7.6 Hz, 1H), 7.15 (d, J = 7.5 Hz, 1H), 6.96 (t, J = 5.6 Hz, 1H), 6.05 (s, 1H), 4.13 (t, J = 5.9 Hz, 2H), 3.65 (t, J = 4.9 Hz, 4H), 3.51 (t, J = 4.8 Hz, 4H), 3.23 (q, J = 5.9 Hz, 2H), 2.32 (s, 3H), 1.36 (s, 9H). 4.1.4. General Procedures for the Synthesis of Inter- mediates 6a−6j. A solution of intermediate 5 (100 mg, 0.2 mmol) dissolved in CH2Cl2 (5 mL) was added 1 mL of TFA. The reaction solution was stirred at room temperature for 3 h and was concentrated under reduced pressure to obtain the deprotected intermediate residue, which was redissolved in anhydrous DMF (3.5 mL). Then, 3-(tert-butoxy)-3-oxopropa- noic acid (48 mg, 0.3 mmol), Et3N (202 mg, 2.0 mmol), and HATU (190 mg, 0.5 mmol) were added in order. The resulting solution was stirred at room temperature for 5 h, diluted with 30 mL of water, and extracted with EtOAc for three times. The organic phase was dried with anhydrous Na2SO4, filtered, and concentrated. The resulting residue was redissolved in 5 mL of CH2Cl2, which was added 1 mL of TFA. The resulting mixture was stirred at room temperature for another 3 h to deprotect the tert-butyl ester. The reaction solution was concentrated to obtain a crude residue, which was purified by chromatography on a silica gel column with CH2Cl2/MeOH to give white intermediate 6a (65 mg, 63%). 1H NMR (500 MHz, DMSO-d6) δ 10.97 (brs, 1H), 8.30 (t, J = 5.6 Hz, 1H), 8.02 (s, 1H), 7.59−7.46 (m, 2H), 7.28 (t, J = 7.6 Hz, 1H), 7.17 (d, J = 7.6 Hz, 1H), 6.03 (s, 1H), 4.22 (t, J = 5.6 Hz, 2H), 3.66 (t, J = 4.7 Hz, 4H), 3.60−3.50 (m, 4H), 3.41 (q, J = 5.6 Hz, 2H), 3.14 (s, 2H), 2.33 (s, 3H). Intermediates 6b− 6j were obtained according to the procedure of 6a. 4.1.5. General Procedures for the Synthesis of Inter- mediates 7a−7j. A solution of intermediate 6a (50 mg, 0.1 mmol) dissolved in anhydrous DMF (3.5 mL) was added the classical VHL ligand (2S,4R)-1-((S)-2-amino-3,3-dimethylbu- tanoyl)-4-hydroxy-N-(4-(4-methylthiazol-5-yl)benzyl)- pyrrolidine-2-carboxamide (42 mg, 0.1 mmol), Et3N (30 mg, 0.3 mmol), and HATU (76 mg, 0.2 mmol) in order. The resulting mixture was stirred at room temperature for 5 h, diluted with 15 mL of water, and extracted with EtOAc for two times. The organic phase was dried with anhydrous Na2SO4, filtered, and concentrated. The resulting residue was purified by chromatography on a silica gel column with CH2Cl2/ MeOH to obtain the final compound 7a (40 mg, 48%). Compounds 7b−7j were obtained according to the procedure of 7a. 4.1.6. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthiazol- 5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1-oxo- butan-2-yl)-N 3-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)- malonamide (7a). 1H NMR (600 MHz, DMSO-d6) δ 10.85 (s, 1H), 8.98 (s, 1H), 8.59 (t, J = 6.1 Hz, 1H), 8.32 (t, J = 5.6 Hz, 1H), 8.21 (d, J = 9.4 Hz, 1H), 8.02 (s, 1H), 7.51 (d, J = 12438 https://doi.org/10.1021/acs.jmedchem.3c00912 J. Med. Chem. 2023, 66, 12432−12445 Journal of Medicinal Chemistry pubs.acs.org/jmc Article 7.6 Hz, 1H), 7.48 (s, 1H), 7.43−7.38 (m, 4H), 7.29 (t, J = 7.6 Hz, 1H), 7.17 (d, J = 7.6 Hz, 1H), 6.08 (s, 1H), 5.15 (brs, 1H), 4.54 (d, J = 9.4 Hz, 1H), 4.46−4.40 (m, 2H), 4.37−4.34 (m, 1H), 4.23 (dd, J = 15.9, 5.6 Hz, 1H), 4.20 (t, J = 5.8 Hz, 2H), 3.69−3.61 (m, 6H), 3.56−3.51 (m, 4H), 3.44−3.39 (m, 2H), 3.25 (d, J = 15.0 Hz, 1H), 3.16 (d, J = 15.0 Hz, 1H), 2.45 (s, 3H), 2.34 (s, 3H), 2.08−2.01 (m, 1H), 1.94−1.87 (m, 1H), 0.94 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 171.25, 168.64, 166.85, 165.70, 163.71, 163.27, 162.81, 150.82, 147.11, 140.74, 138.86, 137.31, 134.12, 130.54, 129.17, 129.04, 128.04, 127.99, 126.81, 126.42, 122.95, 75.14, 68.26, 65.30, 63.90, 58.09, 55.80, 55.75, 43.61, 42.03, 41.04, 37.70, 37.32, 34.95, 25.62, 20.32, 15.32. HRMS [electrospray ionization (ESI)] calcd for C43H54N10O7S [M + H]+ 855.3976, found 855.3977. HPLC purity 98.63%. 4.1.7. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthiazol- 5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1-oxo- butan-2-yl)-N 4-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)- succinimide (7b). 1H NMR (600 MHz, DMSO-d6) δ 10.84 (s, 1H), 8.98 (s, 1H), 8.57 (t, J = 6.0 Hz, 1H), 8.07 (t, J = 5.5 Hz, 1H), 8.02 (s, 1H), 7.91 (d, J = 9.3 Hz, 1H), 7.50 (d, J = 7.8 Hz, 1H), 7.48 (s, 1H), 7.42 (d, J = 8.4 Hz, 2H), 7.38 (d, J = 8.3 Hz, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.17 (d, J = 7.5 Hz, 1H), 6.07 (s, 1H), 5.13 (brs, 1H), 4.52 (d, J = 9.4 Hz, 1H), 4.45− 4.39 (m, 2H), 4.35 (s, 1H), 4.22 (dd, J = 15.9, 5.5 Hz, 1H), 4.17 (t, J = 5.8 Hz, 2H), 3.69−3.60 (m, 6H), 3.57−3.50 (m, 4H), 3.41−3.35 (m, 3H), 2.44 (s, 3H), 2.40−2.28 (m, 6H), 2.07−2.01 (m, 1H), 1.93−1.86 (m, 1H), 0.93 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 171.31, 171.04, 170.64, 168.96, 163.71, 163.31, 162.82, 150.83, 147.09, 140.72, 138.88, 137.30, 134.13, 130.54, 129.16, 129.01, 128.02, 127.99, 126.80, 126.42, 122.94, 75.11, 68.26, 65.30, 63.93, 58.08, 55.80, 55.69, 43.60, 41.02, 37.60, 37.31, 34.71, 30.24, 29.87, 25.73, 20.31, 15.32. HRMS (ESI) calcd for C44H56N10O7S [M + Na]+ 891.3952, found 891.3949. HPLC purity 98.73%. 4.1.8. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthiazol- 5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1-oxo- butan-2-yl)-N 5-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)- glutaramide (7c). 1H NMR (600 MHz, DMSO-d6) δ 10.85 (s, 1H), 8.98 (s, 1H), 8.57 (t, J = 6.1 Hz, 1H), 8.01 (s, 1H), 8.00 (t, J = 5.7 Hz, 1H), 7.93 (d, J = 9.3 Hz, 1H), 7.50 (d, J = 7.8 Hz, 1H), 7.48 (s, 1H), 7.41 (d, J = 8.3 Hz, 2H), 7.37 (d, J = 8.3 Hz, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.17 (d, J = 7.5 Hz, 1H), 6.07 (s, 1H), 5.14 (brs, 1H), 4.58−4.51 (m, 1H), 4.47−4.39 (m, 2H), 4.37−4.34 (m, 1H), 4.25−4.15 (m, 3H), 3.70−3.64 (m, 6H), 3.54 (t, J = 4.9 Hz, 4H), 3.42−3.36 (m, 2H), 2.44 (s, 3H), 2.34 (s, 3H), 2.28−2.21 (m, 1H), 2.18−2.11 (m, 1H), 2.08 (t, J = 7.6 Hz, 2H), 2.06−2.01 (m, 1H), 1.94−1.87 (m, 1H), 1.75−1.67 (m, 2H), 0.94 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 171.45, 171.31, 171.14, 169.21, 163.67, 163.25, 162.81, 150.85, 150.82, 147.08, 140.76, 138.85, 137.28, 134.12, 130.53, 129.15, 129.01, 128.00, 127.97, 126.78, 126.43, 122.95, 75.10, 68.26, 65.30, 63.87, 58.11, 55.82, 55.77, 43.60, 41.03, 37.46, 37.33, 34.55, 34.22, 33.66, 25.78, 21.15, 20.31, 15.31. HRMS (ESI) calcd for C45H58N10O7S [M + H]+ 883.4289, found 883.4285. HPLC purity 98.43%. 4.1.9. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthiazol- 5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1-oxo- butan-2-yl)-N 6-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)- adipamide (7d). 1H NMR (600 MHz, DMSO-d6) δ 10.85 (s, 1H), 8.98 (s, 1H), 8.57 (t, J = 6.1 Hz, 1H), 8.04−7.98 (m, 2H), 7.88 (d, J = 9.3 Hz, 1H), 7.51 (d, J = 7.8 Hz, 1H), 7.48 (s, 1H), 7.41 (d, J = 8.3 Hz, 2H), 7.38 (d, J = 8.3 Hz, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.17 (d, J = 7.5 Hz, 1H), 6.07 (s, 1H), 5.14 (brs, 1H), 4.54 (d, J = 9.4 Hz, 1H), 4.46−4.39 (m, 2H), 4.35 (s, 1H), 4.24−4.13 (m, 3H), 3.70−3.63 (m, 6H), 3.57− 3.51 (m, 4H), 3.44−3.32 (m, 2H), 2.44 (s, 3H), 2.34 (s, 3H), 2.29−2.23 (m, 1H), 2.13−2.01 (m, 4H), 1.94−1.87 (m, 1H), 1.45 (dt, J = 11.7, 5.3 Hz, 4H), 0.93 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 171.67, 171.33, 169.13, 163.66, 163.23, 162.76, 150.83, 147.09, 140.79, 138.87, 137.29, 134.11, 130.54, 129.17, 129.01, 128.01, 127.98, 126.79, 126.44, 122.96, 75.09, 68.25, 65.30, 64.03, 58.08, 55.75, 55.71, 43.61, 41.02, 37.46, 37.32, 34.57, 34.45, 34.02, 25.76, 24.53, 24.30, 20.31, 15.32. HRMS (ESI) calcd for C46H60N10O7S [M + Na]+ 919.4265, found 919.4255. HPLC purity 99.08%. 4.1.10. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthia- zol-5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1- oxobutan-2-yl)-N7-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)- heptanediamide (7e). 1H NMR (600 MHz, DMSO-d6) δ 10.84 (s, 1H), 8.98 (s, 1H), 8.56 (t, J = 6.1 Hz, 1H), 8.02 (s, 1H), 8.00 (t, J = 5.6 Hz, 1H), 7.86 (d, J = 9.4 Hz, 1H), 7.50 (d, J = 7.9 Hz, 1H), 7.48 (s, 1H), 7.42 (d, J = 8.3 Hz, 2H), 7.38 (d, J = 8.3 Hz, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.17 (d, J = 7.5 Hz, 1H), 6.07 (s, 1H), 5.13 (brs, 1H), 4.54 (d, J = 9.4 Hz, 1H), 4.46−4.39 (m, 2H), 4.38−4.33 (m, 1H), 4.24−4.15 (m, 3H), 3.70−3.64 (m, 6H), 3.53 (t, J = 4.9 Hz, 4H), 3.44−3.38 (m, 2H), 2.44 (s, 3H), 2.34 (s, 3H), 2.27−2.21 (m, 1H), 2.14−2.02 (m, 4H), 1.95−1.86 (m, 1H), 1.52−1.42 (m, 4H), 1.24−1.19 (m, 2H), 0.93 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 171.72, 171.42, 171.31, 169.10, 163.65, 163.25, 162.75, 150.81, 147.07, 140.77, 138.86, 137.28, 134.09, 130.53, 129.16, 128.99, 127.99, 127.96, 126.77, 126.41, 122.94, 75.08, 68.23, 65.28, 63.96, 58.06, 55.72, 55.66, 43.59, 41.00, 37.49, 37.31, 34.56, 34.18, 27.73, 25.75, 24.61, 24.40, 20.30, 15.30. HRMS (ESI) calcd for C47H62N10O7S [M + Na]+ 933.4422, found 933.4414. HPLC purity 98.54%. 4.1.11. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthia- zol-5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1- oxobutan-2-yl)-N8-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)- octanediamide (7f). 1H NMR (600 MHz, DMSO-d6) δ 10.82 (s, 1H), 8.98 (s, 1H), 8.56 (t, J = 6.1 Hz, 1H), 8.01 (s, 1H), 7.99 (t, J = 5.5 Hz, 1H), 7.85 (d, J = 9.4 Hz, 1H), 7.53−7.45 (m, 2H), 7.42 (d, J = 8.3 Hz, 2H), 7.39−7.36 (m, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.17 (d, J = 7.5 Hz, 1H), 6.07 (s, 1H), 5.12 (brs, 1H), 4.54 (d, J = 9.4 Hz, 1H), 4.46−4.39 (m, 2H), 4.35 (s, 1H), 4.22 (dd, J = 15.9, 5.5 Hz, 1H), 4.18 (t, J = 5.7 Hz, 2H), 3.70−3.62 (m, 6H), 3.56−3.49 (m, 4H), 3.38−3.35 (m, 2H), 2.44 (s, 3H), 2.34 (s, 3H), 2.27−2.20 (m, 1H), 2.13− 2.08 (m, 1H), 2.08−2.05 (m, 2H), 2.05−2.00 (m, 1H), 1.94− 1.86 (m, 1H), 1.51−1.41 (m, 4H), 1.25−1.20 (m, 4H), 0.93 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 171.77, 171.47, 171.33, 169.11, 163.71, 163.33, 162.82, 150.83, 147.09, 140.71, 138.88, 137.30, 134.13, 130.54, 129.15, 129.01, 128.01, 127.98, 126.79, 126.42, 122.93, 75.11, 68.24, 65.30, 64.29, 63.94, 58.07, 55.73, 55.66, 43.59, 41.02, 37.51, 37.32, 34.67, 34.57, 34.23, 27.86, 27.84, 25.76, 24.71, 24.57, 20.31, 15.32. HRMS (ESI) calcd for C48H64N10O7S [M + Na]+ 947.4578, found 947.4577. HPLC purity 95.55%. 4.1.12. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthia- zol-5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1- 12439 https://doi.org/10.1021/acs.jmedchem.3c00912 J. Med. Chem. 2023, 66, 12432−12445 Journal of Medicinal Chemistry pubs.acs.org/jmc Article oxobutan-2-yl)-N9-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)- nonanediamide (7g). 1H NMR (600 MHz, DMSO-d6) δ 10.82 (s, 1H), 8.98 (s, 1H), 8.56 (t, J = 6.1 Hz, 1H), 8.01 (s, 1H), 8.00 (t, J = 5.7 Hz, 1H), 7.83 (d, J = 9.4 Hz, 1H), 7.50 (d, J = 7.7 Hz, 1H), 7.48 (s, 1H), 7.42 (d, J = 8.3 Hz, 2H), 7.38 (d, J = 8.3 Hz, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.17 (d, J = 7.6 Hz, 1H), 6.07 (s, 1H), 5.12 (brs, 1H), 4.54 (d, J = 9.4 Hz, 1H), 4.46−4.40 (m, 2H), 4.37−4.33 (m, 1H), 4.22 (dd, J = 15.8, 5.5 Hz, 1H), 4.18 (t, J = 5.7 Hz, 2H), 3.70−3.63 (m, 6H), 3.53 (t, J = 4.9 Hz, 4H), 3.38−3.35 (m, 2H), 2.44 (s, 3H), 2.34 (s, 3H), 2.28−2.21 (m, 1H), 2.13−2.01 (m, 4H), 1.94−1.87 (m, 1H), 1.53−1.41 (m, 4H), 1.25−1.18 (m, 6H), 0.93 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 170.57, 170.25, 170.10, 167.87, 162.49, 162.11, 161.59, 149.61, 145.87, 139.48, 137.66, 136.08, 132.91, 129.32, 127.93, 127.79, 126.79, 126.76, 125.57, 125.20, 121.71, 73.89, 67.01, 64.08, 63.07, 62.73, 56.84, 54.50, 54.42, 42.38, 39.80, 36.29, 36.11, 33.46, 33.35, 33.01, 26.76, 26.70, 24.53, 23.58, 23.40, 19.09, 14.10. HRMS (ESI) calcd for C49H66N10O7S [M + Na]+ 961.4735, found 961.4713. HPLC purity 97.86%. 4.1.13. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthia- zol-5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1- oxobutan-2-yl)-N10-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)- decanediamide (7h). 1H NMR (600 MHz, DMSO-d6) δ 10.82 (s, 1H), 8.98 (s, 1H), 8.56 (t, J = 6.1 Hz, 1H), 8.01 (s, 1H), 8.00 (t, J = 5.6 Hz, 1H), 7.83 (d, J = 9.4 Hz, 1H), 7.50 (d, J = 7.7 Hz, 1H), 7.48 (s, 1H), 7.42 (d, J = 8.2 Hz, 2H), 7.38 (d, J = 8.3 Hz, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.17 (d, J = 7.5 Hz, 1H), 6.07 (s, 1H), 5.12 (brs, 1H), 4.54 (d, J = 9.4 Hz, 1H), 4.46−4.40 (m, 2H), 4.35 (s, 1H), 4.21 (dd, J = 15.9, 5.5 Hz, 1H), 4.18 (t, J = 5.7 Hz, 2H), 3.69−3.62 (m, 6H), 3.56− 3.50 (m, 4H), 3.37−3.34 (m, 2H), 2.44 (s, 3H), 2.34 (s, 3H), 2.28−2.20 (m, 1H), 2.13−2.00 (m, 4H), 1.93−1.86 (m, 1H), 1.52−1.40 (m, 4H), 1.27−1.17 (m, 8H), 0.93 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 171.77, 171.45, 171.32, 169.08, 163.70, 163.34, 162.80, 150.82, 147.08, 140.69, 138.88, 137.29, 134.12, 130.53, 129.15, 129.00, 128.00, 127.98, 126.79, 126.41, 122.92, 75.10, 68.23, 65.29, 64.29, 63.93, 58.05, 55.71, 55.63, 43.59, 41.01, 37.52, 37.32, 34.68, 34.57, 34.23, 28.13, 28.04, 25.74, 24.79, 24.64, 20.31, 15.31. HRMS (ESI) calcd for C50H68N10O7S [M + Na]+ 975.4891, found 975.4878. HPLC purity 98.82%. 4.1.14. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthia- zol-5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1- oxobutan-2-yl)-N11-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)- undecanediamide (7i). 1H NMR (600 MHz, DMSO-d6) δ 10.82 (s, 1H), 8.98 (s, 1H), 8.56 (t, J = 6.1 Hz, 1H), 8.01 (s, 1H), 8.00 (t, J = 5.6 Hz, 1H), 7.83 (d, J = 9.3 Hz, 1H), 7.50 (d, J = 7.8 Hz, 1H), 7.48 (s, 1H), 7.42 (d, J = 8.3 Hz, 2H), 7.38 (d, J = 8.3 Hz, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.17 (d, J = 7.5 Hz, 1H), 6.07 (s, 1H), 5.12 (brs, 1H), 4.54 (d, J = 9.4 Hz, 1H), 4.46−4.40 (m, 2H), 4.37−4.33 (m, 1H), 4.22 (dd, J = 15.8, 5.5 Hz, 1H), 4.18 (t, J = 5.7 Hz, 2H), 3.71−3.63 (m, 6H), 3.53 (t, J = 4.9 Hz, 4H), 3.38−3.35 (m, 2H), 2.44 (s, 3H), 2.34 (s, 3H), 2.28−2.22 (m, 1H), 2.12−2.01 (m, 4H), 1.93−1.88 (m, 1H), 1.52−1.40 (m, 4H), 1.24−1.18 (m, 10H), 0.93 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 171.79, 171.47, 171.32, 169.09, 163.71, 163.35, 162.81, 150.83, 147.09, 140.70, 138.89, 137.30, 134.13, 130.54, 129.15, 129.01, 128.01, 127.98, 126.79, 126.41, 122.93, 75.11, 68.24, 65.30, 63.95, 58.06, 55.72, 55.64, 43.60, 41.02, 37.53, 37.33, 34.69, 34.59, 34.25, 28.25, 28.18, 28.14, 28.06, 28.04, 25.75, 24.82, 24.66, 20.31, 15.32. HRMS (ESI) calcd for C51H70N10O7S [M + Na]+ 989.5048, found 989.5040. HPLC purity 98.36%. 4.1.15. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthia- zol-5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1- oxobutan-2-yl)-N12-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)- dodecanediamide (7j). 1H NMR (600 MHz, DMSO-d6) δ 10.82 (s, 1H), 8.98 (s, 1H), 8.56 (t, J = 6.1 Hz, 1H), 8.01 (s, 1H), 7.99 (t, J = 5.7 Hz, 1H), 7.83 (d, J = 9.4 Hz, 1H), 7.50 (d, J = 7.7 Hz, 1H), 7.47 (s, 1H), 7.42 (d, J = 8.3 Hz, 2H), 7.38 (d, J = 8.3 Hz, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.17 (d, J = 7.4 Hz, 1H), 6.07 (s, 1H), 5.12 (brs, 1H), 4.54 (d, J = 9.4 Hz, 1H), 4.47−4.40 (m, 2H), 4.35 (s, 1H), 4.22 (dd, J = 15.8, 5.5 Hz, 1H), 4.18 (t, J = 5.7 Hz, 2H), 3.70−3.63 (m, 6H), 3.53 (t, J = 4.9 Hz, 4H), 3.38−3.35 (m, 2H), 2.44 (s, 3H), 2.34 (s, 3H), 2.27−2.21 (m, 1H), 2.12−2.00 (m, 4H), 1.94−1.87 (m, 1H), 1.52−1.40 (m, 4H), 1.25−1.18 (m, 12H), 0.93 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 171.78, 171.46, 171.32, 169.09, 163.71, 163.35, 162.81, 150.83, 147.09, 140.69, 138.89, 137.29, 134.13, 130.54, 129.15, 129.01, 128.01, 127.98, 126.80, 126.41, 122.92, 75.11, 68.23, 65.30, 63.94, 58.06, 55.72, 55.63, 43.60, 41.02, 37.53, 37.33, 34.69, 34.59, 34.24, 28.35, 28.30, 28.20, 28.14, 28.06, 25.75, 24.82, 24.66, 20.31, 15.32. HRMS (ESI) calcd for C52H72N10O7S [M + H]+ 981.5384, found 981.5383. HPLC purity 99.64%. tert-Butyl (4-(2-((4-Chloro-6-morpholinopyrimi- din-2-yl)oxy)ethyl)phenyl)carbamate (8). Intermediate 8 was obtained according to the procedure of 3. 1H NMR (500 MHz, CDCl3) δ 7.28 (d, J = 8.2 Hz, 2H), 7.19 (d, J = 8.2 Hz, 2H), 6.43 (s, 1H), 6.15 (s, 1H), 4.42 (t, J = 7.4 Hz, 2H), 3.86−3.68 (m, 4H), 3.59 (brs, 4H), 3.03 (t, J = 7.4 Hz, 2H), 1.51 (s, 9H). 4.1.16. 4.1.17. tert-Butyl (4-(2-((4-Hydrazineyl-6-morpholinopyr- imidin-2-yl)oxy)ethyl)phenyl)carbamate (9). Intermediate 9 was obtained according to the procedure of 4.1H NMR (400 MHz, DMSO-d6) δ 9.27 (s, 1H), 7.67 (s, 1H), 7.37 (d, J = 8.2 Hz, 2H), 7.14 (d, J = 8.2 Hz, 2H), 5.61 (s, 1H), 4.63 (brs, 3H), 4.25 (t, J = 7.1 Hz, 2H), 3.68−3.59 (m, 4H), 3.43−3.36 (m, 4H), 2.86 (t, J = 7.1 Hz, 2H), 1.46 (s, 9H). 4.1.18. tert-Butyl (E)-(4-(2-((4-(2-(3-Methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)phenyl)- carbamate (10). Intermediate 10 was obtained according to the procedure of 5.1H NMR (400 MHz, DMSO-d6) δ 10.87 (s, 1H), 9.26 (s, 1H), 7.99 (s, 1H), 7.50 (d, J = 8.1 Hz, 2H), 7.37 (d, J = 8.1 Hz, 2H), 7.29 (t, J = 7.5 Hz, 1H), 7.20−7.10 (m, 3H), 6.06 (s, 1H), 4.32 (t, J = 7.0 Hz, 2H), 3.74−3.62 (m, 4H), 3.57−3.50 (m, 4H), 2.90 (t, J = 7.0 Hz, 2H), 2.34 (s, 3H), 1.47 (s, 9H). for 4.1.19. General Procedures the Synthesis of Intermediates 11a−11j. Intermediates 11a−11j were ob- tained according to the procedure of 6a−6j. (E)-3-((4-(2-((4- (2-(3-methylbenzylidene)hydrazineyl)-6-morpholinopyrimi- din-2-yl)oxy)ethyl)phenyl)amino)-3-oxopropanoic acid (11a) 1H NMR (500 MHz, DMSO-d6) δ 10.99 (brs, 1H), 10.07 (s, 1H), 8.00 (s, 1H), 7.55−7.46 (m, 4H), 7.28 (t, J = 7.6 Hz, 1H), 7.22 (d, J = 8.4 Hz, 2H), 7.16 (s, 1H), 6.01 (s, 1H), 4.38 (s, 2H), 3.66 (t, J = 4.8 Hz, 4H), 3.54 (s, 4H), 2.95 (t, J = 7.0 Hz, 2H), 2.33 (s, 3H). 4.1.20. General Procedures the Synthesis of Intermediates 12a−12j, 13a, and 12dN. Compounds 12a− for 12440 https://doi.org/10.1021/acs.jmedchem.3c00912 J. Med. Chem. 2023, 66, 12432−12445 Journal of Medicinal Chemistry pubs.acs.org/jmc Article 12j, 13a, and 12dN were obtained according to the procedure of 7a−7j. 4.1.21. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthia- zol-5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1- oxobutan-2-yl)-N3-(4-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)phenyl)- malonamide (12a). 1H NMR (600 MHz, DMSO-d6) δ 10.87 (s, 1H), 10.05 (s, 1H), 8.98 (s, 1H), 8.60 (t, J = 6.1 Hz, 1H), 8.23 (d, J = 9.3 Hz, 1H), 7.99 (s, 1H), 7.53−7.47 (m, 4H), 7.43 (d, J = 8.4 Hz, 2H), 7.39 (d, J = 8.3 Hz, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.23 (d, J = 8.5 Hz, 2H), 7.17 (d, J = 7.4 Hz, 1H), 6.07 (s, 1H), 5.15 (brs, 1H), 4.57 (d, J = 9.4 Hz, 1H), 4.49− 4.41 (m, 2H), 4.38−4.32 (m, 3H), 4.23 (dd, J = 15.8, 5.5 Hz, 1H), 3.70−3.62 (m, 6H), 3.53 (t, J = 4.9 Hz, 4H), 3.43 (d, J = 14.9 Hz, 1H), 3.35−3.31 (m, 2H), 2.94 (t, J = 7.0 Hz, 2H), 2.45 (s, 3H), 2.34 (s, 3H), 2.09−2.02 (m, 1H), 1.98−1.87 (m, 1H), 0.96 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 169.87, 167.29, 164.17, 163.75, 162.34, 161.91, 161.48, 149.44, 145.72, 139.22, 137.48, 135.90, 135.15, 132.74, 131.43, 129.15, 127.75, 127.65, 127.18, 126.64, 126.59, 125.42, 125.01, 121.54, 117.15, 73.64, 66.88, 64.49, 63.89, 56.73, 54.54, 54.47, 42.21, 42.12, 39.65, 35.93, 33.57, 32.25, 24.32, 24.27, 18.92, 13.93. HRMS (ESI) calcd for C49H58N10O7S [M + H]+ 931.4289, found 931.4286. HPLC purity 98.68%. 4.1.22. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthia- zol-5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1- oxobutan-2-yl)-N4-(4-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)phenyl)- succinimide (12b). 1H NMR (600 MHz, DMSO-d6) δ 10.86 (s, 1H), 9.88 (s, 1H), 8.98 (s, 1H), 8.57 (t, J = 6.0 Hz, 1H), 7.99 (s, 1H), 7.96 (d, J = 9.3 Hz, 1H), 7.53−7.46 (m, 4H), 7.42 (d, J = 8.3 Hz, 2H), 7.39 (d, J = 8.3 Hz, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.20 (d, J = 8.5 Hz, 2H), 7.17 (d, J = 7.5 Hz, 1H), 6.06 (s, 1H), 5.12 (d, J = 3.5 Hz, 1H), 4.55 (d, J = 9.4 Hz, 1H), 4.47−4.40 (m, 2H), 4.33 (t, J = 7.1 Hz, 3H), 4.22 (dd, J = 15.8, 5.4 Hz, 1H), 3.72−3.60 (m, 6H), 3.55−3.49 (m, 4H), 2.92 (t, J = 7.0 Hz, 2H), 2.63−2.56 (m, 1H), 2.56−2.52 (m, 2H), 2.49−2.42 (m, 4H), 2.34 (s, 3H), 2.06−2.01 (m, 1H), 1.93−1.87 (m, 1H), 0.94 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 172.41, 171.65, 170.78, 170.05, 164.85, 164.43, 163.99, 151.93, 148.19, 141.67, 139.98, 138.40, 138.07, 135.23, 133.37, 131.64, 130.23, 130.11, 129.56, 129.12, 129.08, 127.90, 127.49, 124.01, 119.48, 76.13, 69.36, 67.01, 66.38, 59.19, 56.93, 56.82, 44.69, 42.12, 38.40, 35.86, 34.73, 32.32, 30.59, 26.84, 21.41, 16.42. HRMS (ESI) calcd for C50H60N10O7S [M + H]+ 945.4445, found 945.4438. HPLC purity 99.67%. 4.1.23. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthia- zol-5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1- oxobutan-2-yl)-N5-(4-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)phenyl)- glutaramide (12c). 1H NMR (600 MHz, DMSO-d6) δ 10.85 (s, 1H), 9.81 (s, 1H), 8.98 (s, 1H), 8.56 (t, J = 6.0 Hz, 1H), 7.99 (s, 1H), 7.92 (d, J = 9.2 Hz, 1H), 7.54−7.45 (m, 4H), 7.42 (d, J = 8.3 Hz, 2H), 7.40−7.36 (m, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.20 (d, J = 8.5 Hz, 2H), 7.17 (d, J = 7.5 Hz, 1H), 6.06 (s, 1H), 5.14 (d, J = 3.3 Hz, 1H), 4.55 (d, J = 9.3 Hz, 1H), 4.46−4.40 (m, 2H), 4.38−4.32 (m, 3H), 4.22 (dd, J = 15.8, 5.4 Hz, 1H), 3.70−3.62 (m, 6H), 3.55−3.50 (m, 4H), 2.92 (t, J = 6.9 Hz, 2H), 2.44 (s, 3H), 2.34 (s, 3H), 2.32−2.25 (m, 3H), 2.24−2.18 (m, 1H), 2.07−2.00 (m, 1H), 1.95−1.87 (m, 1H), 1.84−1.75 (m, 2H), 0.95 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 172.41, 172.18, 171.20, 170.20, 164.84, 164.43, 163.99, 151.92, 148.19, 141.68, 139.98, 138.39, 138.05, 135.23, 133.45, 131.64, 130.23, 130.11, 129.53, 129.11, 129.08, 127.89, 127.49, 124.02, 119.63, 76.13, 69.37, 67.01, 66.38, 59.18, 56.91, 56.84, 44.68, 42.12, 38.42, 36.32, 35.66, 34.74, 34.68, 26.88, 21.99, 21.41, 16.42. HRMS (ESI) calcd for C51H62N10O7S [M + H]+ 959.4602, found 959.4584. HPLC purity 99.28%. 4.1.24. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthia- zol-5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1- oxobutan-2-yl)-N6-(4-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)phenyl)- adipamide (12d). 1H NMR (600 MHz, DMSO-d6) δ 10.86 (s, 1H), 9.81 (s, 1H), 8.98 (s, 1H), 8.57 (t, J = 6.1 Hz, 1H), 7.99 (s, 1H), 7.88 (d, J = 9.3 Hz, 1H), 7.53−7.47 (m, 4H), 7.42 (d, J = 8.3 Hz, 2H), 7.40−7.36 (m, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.20 (d, J = 8.5 Hz, 2H), 7.17 (d, J = 7.5 Hz, 1H), 6.07 (s, 1H), 5.13 (d, J = 3.6 Hz, 1H), 4.55 (d, J = 9.4 Hz, 1H), 4.47− 4.41 (m, 2H), 4.33 (t, J = 7.1 Hz, 3H), 4.22 (dd, J = 15.9, 5.5 Hz, 1H), 3.71−3.62 (m, 6H), 3.56−3.50 (m, 4H), 2.92 (t, J = 7.0 Hz, 2H), 2.45 (s, 3H), 2.34 (s, 3H), 2.32−2.24 (m, 3H), 2.20−2.13 (m, 1H), 2.06−2.00 (m, 1H), 1.94−1.87 (m, 1H), 1.61−1.48 (m, 4H), 0.94 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 172.42, 172.41, 171.45, 170.19, 164.85, 164.44, 163.99, 151.92, 148.19, 141.68, 139.98, 138.39, 138.04, 135.24, 133.46, 131.64, 130.23, 130.11, 129.55, 129.11, 129.08, 127.89, 127.49, 124.02, 119.60, 76.14, 69.35, 67.01, 66.38, 59.17, 56.85, 56.81, 44.69, 42.12, 38.42, 36.69, 35.69, 35.21, 34.74, 26.93, 26.87, 25.67, 25.44, 21.41, 16.42. HRMS (ESI) calcd for C52H64N10O7S [M + H]+ 973.4758, found 973.4757. HPLC purity 99.00%. 4.1.25. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthia- zol-5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1- oxobutan-2-yl)-N7-(4-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)phenyl)- heptanediamide (12e). 1H NMR (600 MHz, DMSO-d6) δ 10.85 (s, 1H), 9.79 (s, 1H), 8.98 (d, J = 5.4 Hz, 1H), 8.56 (t, J = 6.1 Hz, 1H), 7.99 (s, 1H), 7.86 (d, J = 9.3 Hz, 1H), 7.53− 7.46 (m, 4H), 7.42 (d, J = 8.3 Hz, 2H), 7.38 (d, J = 8.3 Hz, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.20 (d, J = 8.5 Hz, 2H), 7.17 (d, J = 7.5 Hz, 1H), 6.06 (s, 1H), 5.13 (d, J = 3.6 Hz, 1H), 4.54 (d, J = 9.4 Hz, 1H), 4.45−4.40 (m, 2H), 4.37−4.30 (m, 3H), 4.22 (dd, J = 15.8, 5.5 Hz, 1H), 3.70−3.62 (m, 6H), 3.55−3.50 (m, 4H), 2.92 (t, J = 7.0 Hz, 2H), 2.44 (s, 3H), 2.34 (s, 3H), 2.30−2.22 (m, 3H), 2.16−2.10 (m, 1H), 2.07−2.00 (m, 1H), 1.95−1.86 (m, 1H), 1.62−1.43 (m, 4H), 1.32−1.23 (m, 2H), 0.93 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 172.51, 172.42, 171.51, 170.19, 164.84, 164.43, 163.99, 151.93, 148.19, 141.68, 139.98, 138.39, 138.08, 135.23, 133.41, 131.64, 130.23, 130.11, 129.54, 129.11, 129.08, 127.89, 127.49, 124.01, 119.58, 76.13, 69.34, 67.01, 66.38, 59.16, 56.84, 56.76, 44.68, 42.12, 38.43, 36.75, 35.68, 35.26, 34.74, 28.84, 26.86, 25.73, 25.41, 21.41, 16.42. HRMS (ESI) calcd for C53H66N10O7S [M + H]+ 987.4915, found 987.4912. HPLC purity 99.85%. 4.1.26. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthia- zol-5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1- oxobutan-2-yl)-N8-(4-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)phenyl)- octanediamide (12f). 1H NMR (600 MHz, DMSO-d6) δ 10.85 (s, 1H), 9.80 (s, 1H), 8.98 (s, 1H), 8.56 (t, J = 6.1 Hz, 1H), 7.99 (s, 1H), 7.85 (d, J = 9.4 Hz, 1H), 7.53−7.46 (m, 4H), 7.42 (d, J = 8.2 Hz, 2H), 7.40−7.36 (m, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.20 (d, J = 8.5 Hz, 2H), 7.17 (d, J = 7.5 Hz, 1H), 6.06 (s, 1H), 5.13 (d, J = 3.6 Hz, 1H), 4.55 (d, J = 9.4 Hz, 1H), 4.46−4.40 (m, 2H), 4.37−4.31 (m, 3H), 4.22 (dd, J = 12441 https://doi.org/10.1021/acs.jmedchem.3c00912 J. Med. Chem. 2023, 66, 12432−12445 Journal of Medicinal Chemistry pubs.acs.org/jmc Article 15.9, 5.5 Hz, 1H), 3.70−3.63 (m, 6H), 3.55−3.50 (m, 4H), 2.92 (t, J = 7.0 Hz, 2H), 2.44 (s, 3H), 2.34 (s, 3H), 2.29−2.22 (m, 3H), 2.15−2.10 (m, 1H), 2.06−2.01 (m, 1H), 1.93−1.88 (m, 1H), 1.60−1.42 (m, 4H), 1.32−1.24 (m, 4H), 0.94 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 172.56, 172.43, 171.55, 170.20, 164.84, 164.43, 163.99, 151.92, 148.19, 141.68, 139.98, 138.39, 138.07, 135.23, 133.43, 131.64, 130.23, 130.11, 129.54, 129.11, 129.08, 127.89, 127.49, 124.02, 119.59, 76.13, 69.34, 67.01, 66.38, 59.16, 56.84, 56.76, 44.68, 42.12, 38.42, 36.84, 35.68, 35.34, 34.74, 28.98, 28.94, 26.86, 25.83, 25.56, 21.41, 16.42. HRMS (ESI) calcd for C54H68N10O7S [M + H]+ 1001.5071, found 1001.5063. HPLC purity 99.74%. 4.1.27. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthia- zol-5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1- oxobutan-2-yl)-N9-(4-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)phenyl)- nonanediamide (12g). 1H NMR (600 MHz, DMSO-d6) δ 10.85 (s, 1H), 9.79 (s, 1H), 8.98 (s, 1H), 8.56 (t, J = 6.1 Hz, 1H), 7.99 (s, 1H), 7.84 (d, J = 9.4 Hz, 1H), 7.53−7.46 (m, 4H), 7.42 (d, J = 8.3 Hz, 2H), 7.40−7.36 (m, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.20 (d, J = 8.5 Hz, 2H), 7.17 (d, J = 7.5 Hz, 1H), 6.06 (s, 1H), 5.12 (d, J = 3.6 Hz, 1H), 4.54 (d, J = 9.4 Hz, 1H), 4.46−4.39 (m, 2H), 4.33 (t, J = 7.1 Hz, 3H), 4.21 (dd, J = 15.8, 5.5 Hz, 1H), 3.72−3.61 (m, 6H), 3.56−3.48 (m, 4H), 2.92 (t, J = 7.0 Hz, 2H), 2.44 (s, 3H), 2.34 (s, 3H), 2.30−2.20 (m, 3H), 2.15−2.08 (m, 1H), 2.05−1.98 (m, 1H), 1.95−1.85 (m, 1H), 1.62−1.40 (m, 4H), 1.30−1.20 (m, 6H), 0.93 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 170.70, 170.56, 169.71, 168.32, 162.98, 162.57, 162.13, 150.06, 146.33, 139.81, 138.12, 136.53, 136.21, 133.37, 131.56, 129.77, 128.37, 128.24, 127.68, 127.25, 127.22, 126.03, 125.63, 122.15, 117.72, 74.27, 67.47, 65.14, 64.52, 57.29, 54.96, 54.87, 42.82, 38.67, 36.56, 34.99, 33.81, 33.46, 32.87, 27.21, 27.16, 24.98, 24.03, 23.76, 19.55, 14.56. HRMS (ESI) calcd for C55H70N10O7S [M + H]+ 1015.5228, found 1015.5218. HPLC purity 98.87%. 4.1.28. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthia- zol-5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1- oxobutan-2-yl)-N10-(4-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)phenyl)- decanediamide (12h). 1H NMR (600 MHz, DMSO-d6) δ 10.85 (s, 1H), 9.79 (s, 1H), 8.98 (s, 1H), 8.56 (t, J = 6.0 Hz, 1H), 7.99 (s, 1H), 7.84 (d, J = 9.4 Hz, 1H), 7.54−7.46 (m, 4H), 7.42 (d, J = 8.3 Hz, 2H), 7.38 (d, J = 8.3 Hz, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.20 (d, J = 8.5 Hz, 2H), 7.17 (d, J = 7.5 Hz, 1H), 6.06 (s, 1H), 5.12 (d, J = 3.6 Hz, 1H), 4.54 (d, J = 9.4 Hz, 1H), 4.46−4.40 (m, 2H), 4.37−4.31 (m, 3H), 4.22 (dd, J = 15.8, 5.4 Hz, 1H), 3.70−3.62 (m, 6H), 3.56−3.50 (m, 4H), 2.92 (t, J = 7.0 Hz, 2H), 2.44 (s, 3H), 2.34 (s, 3H), 2.29−2.23 (m, 3H), 2.14−2.08 (m, 1H), 2.05−1.99 (m, 1H), 1.94−1.87 (m, 1H), 1.60−1.42 (m, 4H), 1.32−1.17 (m, 8H), 0.93 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 172.56, 172.42, 171.56, 170.19, 164.84, 164.43, 163.99, 151.92, 148.19, 141.67, 139.98, 138.39, 138.08, 135.24, 133.41, 131.64, 130.23, 130.11, 129.54, 129.11, 129.08, 127.89, 127.49, 124.01, 119.57, 76.13, 69.33, 67.01, 66.38, 59.16, 56.82, 56.74, 44.68, 42.11, 38.43, 36.86, 35.68, 35.32, 34.74, 29.24, 29.15, 29.13, 26.85, 25.89, 25.63, 21.41, 16.42. HRMS (ESI) calcd for C56H72N10O7S [M + H]+ 1029.5384, found 1029.5378. HPLC purity 99.35%. 4.1.29. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthia- zol-5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1- oxobutan-2-yl)-N11-(4-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)phenyl)- undecanediamide (12i). 1H NMR (600 MHz, DMSO-d6) δ 10.85 (s, 1H), 9.79 (s, 1H), 8.98 (s, 1H), 8.55 (t, J = 6.1 Hz, 1H), 7.99 (s, 1H), 7.84 (d, J = 9.4 Hz, 1H), 7.52−7.47 (m, 4H), 7.42 (d, J = 8.3 Hz, 2H), 7.39−7.37 (m, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.20 (d, J = 8.5 Hz, 2H), 7.17 (d, J = 7.5 Hz, 1H), 6.06 (s, 1H), 5.11 (d, J = 3.2 Hz, 1H), 4.54 (d, J = 9.4 Hz, 1H), 4.45−4.40 (m, 2H), 4.33 (t, J = 7.1 Hz, 3H), 4.21 (dd, J = 15.9, 5.4 Hz, 1H), 3.68−3.62 (m, 6H), 3.54−3.50 (m, 4H), 2.92 (t, J = 7.0 Hz, 2H), 2.44 (s, 3H), 2.34 (s, 3H), 2.29−2.22 (m, 3H), 2.14−2.07 (m, 1H), 2.05−1.97 (m, 1H), 1.93−1.86 (m, 1H), 1.62−1.52 (m, 2H), 1.53−1.42 (m, 2H), 1.30−1.20 (m, 12H), 0.93 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 172.56, 172.42, 171.56, 170.18, 164.84, 164.43, 163.99, 151.93, 148.19, 141.67, 139.98, 138.39, 138.08, 135.23, 133.42, 131.64, 130.23, 130.11, 129.54, 129.11, 129.08, 127.89, 127.49, 124.01, 119.58, 76.13, 69.33, 67.01, 66.38, 59.15, 56.82, 56.73, 44.68, 42.11, 38.43, 36.86, 35.68, 35.33, 34.74, 29.34, 29.27, 29.22, 29.16, 29.13, 26.85, 25.91, 25.64, 21.41, 16.42. HRMS (ESI) calcd for C57H74N10O7S [M + H]+ 1043.5541, found 1043.5527. HPLC purity 96.81%. 4.1.30. N1-((S)-1-((2S,4R)-4-Hydroxy-2-((4-(4-methylthia- zol-5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1- oxobutan-2-yl)-N12-(4-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)phenyl)- dodecanediamide (12j). 1H NMR (600 MHz, DMSO-d6) δ 10.86 (s, 1H), 9.79 (s, 1H), 8.98 (s, 1H), 8.56 (t, J = 6.1 Hz, 1H), 7.99 (d, J = 1.0 Hz, 1H), 7.84 (d, J = 9.4 Hz, 1H), 7.52− 7.47 (m, 4H), 7.42 (d, J = 8.3 Hz, 2H), 7.38 (d, J = 8.3 Hz, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.20 (d, J = 8.5 Hz, 2H), 7.17 (d, J = 7.6 Hz, 1H), 6.06 (s, 1H), 5.12 (brs, 1H), 4.54 (d, J = 9.4 Hz, 1H), 4.46−4.40 (m, 2H), 4.38−4.30 (m, 3H), 4.22 (dd, J = 15.8, 5.5 Hz, 1H), 3.68−3.63 (m, 6H), 3.53 (t, J = 4.9 Hz, 4H), 2.92 (t, J = 7.1 Hz, 2H), 2.44 (s, 3H), 2.34 (s, 3H), 2.29− 2.22 (m, 3H), 2.13−2.07 (m, 1H), 2.06−2.00 (m, 1H), 1.94− 1.87 (m, 1H), 1.57 (t, J = 7.2 Hz, 2H), 1.53−1.40 (m, 2H), 1.30−1.22 (m, 12H), 0.93 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 171.47, 171.32, 170.47, 169.09, 163.73, 163.31, 162.87, 150.82, 147.09, 140.60, 138.89, 137.29, 136.98, 134.13, 132.31, 130.54, 129.14, 129.01, 128.44, 128.01, 127.98, 126.79, 126.40, 122.92, 118.48, 75.03, 68.24, 65.93, 65.28, 58.06, 55.72, 55.64, 43.59, 41.02, 37.33, 35.75, 34.59, 34.24, 33.64, 28.33, 28.29, 28.18, 28.12, 28.08, 28.04, 25.75, 24.82, 24.53, 20.31, 15.32. HRMS (ESI) calcd for C58H76N10O7S [M + H]+ 1057.5697, found 1057.5696. HPLC purity 98.53%. 4.1.31. N1-((S)-1-((2S,4R)-4-Hydroxy-2-(((S)-1-(4-(4-meth- ylthiazol-5-yl)phenyl)ethyl)carbamoyl)pyrrolidin-1-yl)-3,3- dimethyl-1-oxobutan-2-yl)-N 6 -(4-(2-((4-(2-((E)-3- methylbenzylidene)hydrazineyl)-6-morpholinopyrimidin-2- yl)oxy)ethyl)phenyl)adipamide (13a). 1H NMR (600 MHz, DMSO-d6) δ 10.86 (s, 1H), 9.81 (s, 1H), 8.98 (s, 1H), 8.37 (d, J = 7.8 Hz, 1H), 7.99 (s, 1H), 7.81 (d, J = 9.2 Hz, 1H), 7.53− 7.46 (m, 4H), 7.43 (d, J = 8.2 Hz, 2H), 7.37 (d, J = 8.2 Hz, 2H), 7.29 (t, J = 7.6 Hz, 1H), 7.20 (d, J = 8.5 Hz, 2H), 7.17 (d, J = 8.0 Hz, 1H), 6.06 (s, 1H), 5.10 (s, 1H), 4.96−4.88 (m, 1H), 4.51 (d, J = 9.3 Hz, 1H), 4.43 (t, J = 8.0 Hz, 1H), 4.33 (t, J = 7.0 Hz, 2H), 4.28 (s, 1H), 3.69−3.64 (m, 4H), 3.64−3.58 (m, 2H), 3.56−3.49 (m, 4H), 2.93 (t, J = 7.0 Hz, 2H), 2.45 (s, 3H), 2.34 (s, 3H), 2.31−2.25 (m, 3H), 2.19−2.12 (m, 1H), 2.04−1.97 (m, 1H), 1.83−1.75 (m, 1H), 1.60−1.48 (m, 4H), 1.37 (d, J = 7.0 Hz, 3H), 0.94 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 171.26, 170.35, 169.99, 168.98, 163.73, 163.31, 162.87, 150.85, 147.12, 144.03, 140.59, 137.29, 136.94, 134.13, 132.35, 130.48, 129.13, 129.06, 128.46, 128.19, 127.98, 126.40, 125.74, 125.61, 122.92, 118.50, 75.02, 68.14, 65.92, 65.28, 12442 https://doi.org/10.1021/acs.jmedchem.3c00912 J. Med. Chem. 2023, 66, 12432−12445 Journal of Medicinal Chemistry pubs.acs.org/jmc Article 57.92, 55.78, 55.64, 47.06, 43.59, 37.08, 35.59, 34.56, 34.13, 33.64, 25.83, 24.55, 24.32, 21.81, 20.31, 15.36. HRMS (ESI) calcd for C53H66N10O7S [M + Na]+ 1009.4735, found 1009.4711. HPLC purity 97.91%. 4.1.32. N1-((S)-1-((2S,4S)-4-Hydroxy-2-((4-(4-methylthia- zol-5-yl)benzyl)carbamoyl)pyrrolidin-1-yl)-3,3-dimethyl-1- oxobutan-2-yl)-N6-(4-(2-((4-(2-((E)-3-methylbenzylidene)- hydrazineyl)-6-morpholinopyrimidin-2-yl)oxy)ethyl)phenyl)- adipamide (12dN). 1H NMR (600 MHz, DMSO-d6) δ 10.85 (s, 1H), 9.80 (s, 1H), 8.98 (s, 1H), 8.63 (t, J = 6.1 Hz, 1H), 7.99 (s, 1H), 7.87 (d, J = 8.8 Hz, 1H), 7.50 (t, J = 8.2 Hz, 3H), 7.47 (s, 1H), 7.42−7.37 (m, 4H), 7.29 (t, J = 7.6 Hz, 1H), 7.20 (d, J = 8.5 Hz, 2H), 7.17 (d, J = 7.4 Hz, 1H), 6.06 (s, 1H), 5.43 (d, J = 7.3 Hz, 1H), 4.49−4.41 (m, 2H), 4.39−4.31 (m, 3H), 4.26 (dd, J = 15.8, 5.5 Hz, 1H), 4.23−4.18 (m, 1H), 3.94 (dd, J = 10.0, 5.7 Hz, 1H), 3.69−3.64 (m, 4H), 3.55−3.51 (m, 4H), 3.44 (dd, J = 10.0, 5.3 Hz, 1H), 2.92 (t, J = 7.0 Hz, 2H), 2.44 (s, 3H), 2.34 (s, 4H), 2.33−2.22 (m, 4H), 2.19−2.11 (m, 1H), 1.77−1.71 (m, 1H), 1.58−1.47 (m, 4H), 0.95 (s, 9H). 13C NMR (150 MHz, DMSO-d6) δ 172.96, 172.74, 171.44, 170.46, 164.85, 164.44, 164.00, 148.22, 141.73, 139.68, 138.40, 138.04, 135.24, 133.47, 131.61, 130.20, 129.64, 129.49, 129.24, 129.16, 129.04, 128.02, 127.92, 127.82, 119.64, 76.14, 69.57, 67.01, 66.39, 65.40, 58.99, 57.16, 56.09, 44.69, 42.26, 37.40, 36.70, 35.13, 35.10, 34.74, 26.86, 25.64, 25.40, 21.42, 16.42. HRMS (ESI) calcd for C52H64N10O7S [M + Na]+ 995.4578, found 995.4571. HPLC purity 95.44%. 4.2. Molecular Docking. The only available protein structure of PIKfyve was the crystal complex of PIKfyve, Figure 4, and Vac14 (PDB:7K2V) to date. Figure 4 and Vac14 were removed before protein preparation, which was operated by assigning bond orders and adding hydrogens in the Protein Preparation Wizard section of Maestro Version 11.9 (Schrodinger, LLC, New York, 2019). The chemical structure of apilimod was constructed and prepared by using the LigPrep section with the default settings and the OPLS3e force field. The grid box of the receptor was generated as the center of residue Leu119. Molecular docking was operated in extra precision by using the Glide section of Maestro. The final images were prepared by Pymol (http://pymol.org). 4.3. Cell Line. Human prostate cancer cell lines VCaP, PC3, 22RV1, and LNCaP were purchased from American Type Culture Collection (ATCC) and maintained under 5% CO2 at 37 °C in a medium according to ATCC’s instruction. All cell lines were tested negative for mycoplasma and authenticated by genotyping. 4.4. Western Blot Analysis. The whole cell lysate was harvested in Pierce radioimmunoprecipitation assay (RIPA) buffer (ThermoScientific) containing protease and phospha- tase inhibitor cocktails. Protein concentration was measured using the detergent compatible (DC) protein assay (Bio-Rad). Denatured lysates were separated on NuPage 4−12% Bis-Tris Midi Protein gels (Novex) and transferred to 0.45 μm polyvinylidene difluoride membrane (Immobilon) using a TransBlot Turbo dry transfer machine (Bio-Rad). The membrane was incubated in blocking buffer (5% non-fat dry milk, Tris-buffered saline with 0.1% Tween-20) for 1 h at room temperature. The membrane was then incubated with primary antibody for 1 h at room temperature, followed by overnight incubation at 4 °C. Chemiluminescent detection using ECL Prime (Amersham) and signal were visualized by an Odyssey imaging system (Li-Cor). Primary antibodies were PIKfyve (R&D, AF7885), LC3A/B (CST, 12741S), GAPDH (CST, 3683S), and vinculin (CST, 18799S). All antibodies were used at dilutions suggested by the manufacturers. 4.5. TMT-Labeled Quantitative Proteomic Analysis. VCaP cells were plated at 3 × 106 cells per well in a 6-well plate overnight prior to treatment with DMSO or 300 nM PIK5-12d for 4 h. Whole cell lysates were collected in RIPA buffer (Thermo Fisher Scientific) without protease inhibitor. Total protein (75 μg) per condition was labeled with TMT isobaric Label Reagent (Thermo Fisher Scientific) according to the manufacturer’s protocol and subjected to 12 fractions of liquid chromatography−mass spectrometry (LC−MS)/MS analysis. 4.6. In Vivo Experiment. All in vivo experiments were approved by the University of Michigan Institutional Animal Care and Use Committee. LTL-331R tumor was kindly provided by Dr. Yuzhuo Wang’s group in Vancouver Prostate Centre and maintained subcutaneously on both sides of dorsal flanks of male CB17 severe combined immunodeficiency mice. PIK5-12d was freshly dissolved in the vehicle (5% DMSO and 95% of 40% hydroxypropyl-β-cyclodextrin) for once-daily IP injection. Pharmacodynamic assessment was performed by once-daily administration of either vehicle, 4 or 10 mg/kg PIK5-12d for 3 days, and tumor samples were collected 24 h post last dose on day 4 for protein and TUNEL in situ cell death assay assessment. Long-term tumor efficacy of the LTL- 331R model was determined by once-daily administration of PIK5-12d at 5 days on and 2 days off regimen for 17 days. Tumors were measured at least twice per week using digital calipers following the formula (p/6) (L × W2), where L and W are the length and width of the tumors, respectively. 4.7. TUNEL Assay. Tumor tissue was fixed in formalin and embedded into paraffin. Formalin-fixed, paraffin-embedded tissue was sectioned into 5 μm thickness and then deparaffined and rehydrated by xylene and ethanol gradients. TUNEL signal was stained using In Situ Cell Death Detection Kit TMR (Roche) according to manufacturer’s instruction. TUNEL signal was visualized by a Zeiss fluorescence microscope. ■ ASSOCIATED CONTENT *sı Supporting Information The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jmedchem.3c00912. 1H NMR and 13C NMR spectra, and HPLC traces for all degraders (PDF) Molecular formula strings (CSV) Docking pose of compound 1 in PIKfyve (PDB) The data set of TMT proteomics (XLSX) ■ AUTHOR INFORMATION Corresponding Authors Zhen Wang − State Key Laboratory of Chemical Biology, Shanghai Institute of Organic Chemistry, Chinese Academy of Sciences, Shanghai 200032, People’s Republic of China; orcid.org/0000-0001-8762-6089; Email: wangz@ sioc.ac.cn Arul M. Chinnaiyan − Michigan Center for Translational Pathology, Department of Pathology, and Department of Urology, University of Michigan, Ann Arbor, Michigan 48109, United States; Howard Hughes Medical Institute, University of Michigan, Ann Arbor, Michigan 48109, United States; Email: arul@med.umich.edu 12443 https://doi.org/10.1021/acs.jmedchem.3c00912 J. Med. Chem. 2023, 66, 12432−12445 Journal of Medicinal Chemistry pubs.acs.org/jmc Article Ke Ding − State Key Laboratory of Chemical Biology, Shanghai Institute of Organic Chemistry, Chinese Academy of Sciences, Shanghai 200032, People’s Republic of China; Institute of Basic Medicine and Cancer (IBMC), Chinese Academy of Sciences, Hangzhou, Zhejiang 310022, People’s Republic of China; International Cooperative Laboratory of Traditional Chinese Medicine Modernization and Innovative Drug Discovery of Chinese Ministry of Education (MOE), Guangzhou City Key Laboratory of Precision Chemical Drug Development, College of Pharmacy, Jinan University, Guangzhou 511400, People’s Republic of China; orcid.org/0000-0001-9016-812X; Phone: +86-21-5492 5100; Email: dingk@sioc.ac.cn Authors Chungen Li − State Key Laboratory of Chemical Biology, Shanghai Institute of Organic Chemistry, Chinese Academy of Sciences, Shanghai 200032, People’s Republic of China Yuanyuan Qiao − Michigan Center for Translational Pathology and Department of Pathology, University of Michigan, Ann Arbor, Michigan 48109, United States Xia Jiang − Michigan Center for Translational Pathology, University of Michigan, Ann Arbor, Michigan 48109, United States Lianchao Liu − State Key Laboratory of Chemical Biology, Shanghai Institute of Organic Chemistry, Chinese Academy of Sciences, Shanghai 200032, People’s Republic of China Yang Zheng − Michigan Center for Translational Pathology, University of Michigan, Ann Arbor, Michigan 48109, United States Yudi Qiu − State Key Laboratory of Chemical Biology, Shanghai Institute of Organic Chemistry, Chinese Academy of Sciences, Shanghai 200032, People’s Republic of China Caleb Cheng − Michigan Center for Translational Pathology, University of Michigan, Ann Arbor, Michigan 48109, United States; orcid.org/0000-0003-1872-2661 Fengtao Zhou − International Cooperative Laboratory of Traditional Chinese Medicine Modernization and Innovative Drug Discovery of Chinese Ministry of Education (MOE), Guangzhou City Key Laboratory of Precision Chemical Drug Development, College of Pharmacy, Jinan University, Guangzhou 511400, People’s Republic of China; orcid.org/0000-0003-2518-7855 Yang Zhou − International Cooperative Laboratory of Traditional Chinese Medicine Modernization and Innovative Drug Discovery of Chinese Ministry of Education (MOE), Guangzhou City Key Laboratory of Precision Chemical Drug Development, College of Pharmacy, Jinan University, Guangzhou 511400, People’s Republic of China; orcid.org/0000-0003-4167-6413 Weixue Huang − State Key Laboratory of Chemical Biology, Shanghai Institute of Organic Chemistry, Chinese Academy of Sciences, Shanghai 200032, People’s Republic of China Xiaomei Ren − State Key Laboratory of Chemical Biology, Shanghai Institute of Organic Chemistry, Chinese Academy of Sciences, Shanghai 200032, People’s Republic of China Yuzhuo Wang − The Vancouver Prostate Centre, Vancouver General Hospital and Department of Urologic Sciences, The University of British Columbia, Vancouver, British Columbia V6H 3Z6, Canada Complete contact information is available at: https://pubs.acs.org/10.1021/acs.jmedchem.3c00912 Author Contributions ◆C.L. and Y.Q. contributed equally to this work. Notes The authors declare the following competing financial interest(s): A patent application was filed on the PIKfyve degraders described in this study in which A.M.C., K.D., L.C., Y.Q., and Z.W. are named as inventors. This work received partially financial support from Livzon Pharmaceutical Group Inc. ■ ACKNOWLEDGMENTS from the National We acknowledge the financial support Natural Science Foundation of China (81820108029 and 22037003), the Open Project of Shenzhen Bay Laboratory (SZBL2021080601004), State Key Laboratory of Chemical Biology, and Livzon Pharmaceutical Group Inc. This work was also supported by National Cancer Institute Outstanding Investigator Award R35 CA231996. A.M.C. is a Howard Hughes Medical Institute Investigator, A. Alfred Taubman Scholar, and American Cancer Society Professor. ■ ABBREVIATIONS AcOH, acetic acid; ATCC, American type culture collection; CH2Cl2, dichloromethane; DC50, the half-maximal degradation concentration; Dmax, the maximal degradation rate; DMF, N,N- dimethylformamide; DMSO, dimethyl sulfoxide; EtOH, ethyl alcohol; Et3N, triethylamine; HATU, 2-(7-azabenzotriazol-1- yl)-N,N,N′,N′-tetramethyluronium hexafluorophosphate; intraperitoneal; NaH, HRMS, high-resolution mass; sodium hydride; PDX, patient-derived xenograft; PI(3)P, phosphatidylinositol-3-phosphate; PI(3,5)P2, phosphatidylino- sitol-3,5-bisphosphate; POI, protein-of-interest; PROTAC, proteolysis-targeting chimera; SAR, structure−activity relation- ship; SDR, structure−degradation relationship; TFA, trifluoro- acetic acid; TLC, thin-layer chromatography; TMT, tandem mass tags; UPS, ubiquitin-proteasome system; VHL, von Hippel−Lindau IP, ■ REFERENCES (1) Ikonomov, O. C.; Sbrissa, D.; Shisheva, A. Small molecule PIKfyve inhibitors as cancer therapeutics: translational promises and limitations. Toxicol. Appl. Pharmacol. 2019, 383, No. 114771. (2) Shisheva, A. PIKfyve: partners, significance, debates and paradoxes. Cell. Biol. Int. 2008, 32, 591−604. (3) Bissig, C.; Croise, P.; Heiligenstein, X.; Hurbain, I.; Lenk, G. M.; Kaufman, E.; Sannerud, R.; Annaert, W.; Meisler, M. H.; Weisman, L. S. The PIKfyve complex regulates the early melanosome homeostasis required for physiological amyloid formation. J. Cell Sci. 2019, 132, No. jcs229500. (4) Bissig, C.; Hurbain, I.; Raposo, G.; van Niel, G. PIKfyve activity regulates reformation of terminal storage lysosomes from endolyso- somes. Traffic 2017, 18, 747−757. (5) Hessvik, N. P.; Øverbye, A.; Brech, A.; Torgersen, M. L.; Jakobsen, I. S.; Sandvig, K.; Llorente, A. PIKfyve inhibition increases exosome release and induces secretory autophagy. Cell. Mol. Life Sci. 2016, 73, 4717−4737. (6) Lartigue, J. D.; Polson, H.; Feldman, M.; Shokat, K.; Tooze, S. A.; Urbé, S.; Clague, M. J. PIKfyve regulation of endosome-linked pathways. Traffic 2009, 10, 883−893. (7) O’Connell, C. E.; Vassilev, A. Combined inhibition of p38MAPK and PIKfyve synergistically disrupts autophagy to selectively target cancer cells. Cancer Res. 2021, 81, 2903−2917. (8) Rivero-Rios, P.; Weisman, L. S. Roles of PIKfyve in multiple cellular pathways. Curr. Opin. Cell Biol. 2022, 76, No. 102086. 12444 https://doi.org/10.1021/acs.jmedchem.3c00912 J. Med. Chem. 2023, 66, 12432−12445 Journal of Medicinal Chemistry pubs.acs.org/jmc Article I.; Klionsky, D. triple negative cell (9) Oppelt, A.; Haugsten, E. M.; Zech, T.; Håvard, E. D.; Jørgen, W. PIKfyve, MTMR3 and their product PtdIns5P regulate cancer cell migration and invasion through activation of Rac1. Biochem. J. 2014, 461, 383−390. (10) Ikonomov, O. C.; Filios, C.; Sbrissa, D.; Chen, X.; Shisheva, A. The PIKfyve−ArPIKfyve−Sac3 triad in human breast cancer: link between elevated Sac3 phosphatase and enhanced functional lines. Biochem. Biophys. Res. proliferation of Commun. 2013, 440, 342−347. (11) Gayle, S.; Landrette, S.; Beeharry, N.; Conrad, C.; Hernandez, M.; Beckett, P.; Ferguson, S. M.; Mandelkern, T.; Zheng, M.; Xu, T.; Rothberg, J.; Lichenstein, H. Identification of apilimod as a first-in- class PIKfyve kinase inhibitor for treatment of B-cell non-hodgkin lymphoma. Blood 2017, 129, 1768−1778. (12) Qiao, Y.; Choi, J. E.; Tien, J. C.; Simko, S. A.; Rajendiran, T.; Vo, J. N.; Delekta, A. D.; Wang, L.; Xiao, L.; Hodge, N. B.; Desai, P.; Mendoza, S.; Juckette, K.; Xu, A.; Soni, T.; Su, F.; Wang, R.; Cao, X.; Yu, J.; Kryczek, I.; Wang, X.-M.; Wang, X.; Siddiqui, J.; Wang, Z.; Bernard, A.; Fernandez-Salas, E.; Navone, N. M.; Ellison, S. J.; Ding, K.; Eskelinen, E.-L.; Heath, E. J.; Zou, W.; Chinnaiyan, A. M. Autophagy inhibition by targeting PIKfyve potentiates response to immune checkpoint blockade in prostate cancer. Nat. Cancer 2021, 2, 978−993. (13) Jefferies, H. B. J.; Cooke, F. T.; Jat, P.; Boucheron, C.; Koizumi, T.; Hayakawa, M.; Kaizawa, H.; Ohishi, T.; Workman, P.; Waterfield, M. D.; Parker, P. J. A selective PIKfyve inhibitor blocks PtdIns(3,5)P2 production and disrupts endomembrane transport and retroviral budding. EMBO Rep. 2008, 9, 164−170. (14) Ikonomov, O. C.; Sbrissa, D.; Shisheva, A. YM201636, an inhibitor of retroviral budding and PIKfyve-catalyzed PtdIns(3,5)P2 synthesis, halts glucose entry by insulin in adipocytes. Biochem. Biophys. Res. Commun. 2009, 382, 566−570. (15) Martin, S.; Harper, C. B.; May, L. M.; Coulson, E. J.; Osborne, S. L. Inhibition of PIKfyve by YM-201636 dysregulates autophagy and leads to apoptosis-independent neuronal cell death. PLoS One 2013, 8, No. e60152. (16) Krausz, S.; Boumans, M. J. H.; Gerlag, D. M.; Lufkin, J.; Kuijk, A. W. R.; Bakker, A.; Boer, M.; Lodde, B. M.; Reedquist, K. A.; Jacobson, E. W.; Mera, M. O.; Tak, P. P. A phase IIa, randomized, double-blind, placebo-controlled trial of apilimod mesylate, an interleukin-12/interleukin-23 inhibitor, in patients with rheumatoid arthritis. Arthritis Rheum. 2012, 64, 1750−1755. (17) Sbrissa, D.; Naisan, G.; Ikonomov, O.; Shisheva, A. Apilimod, a candidate anticancer therapeutic, arrests not only PtdIns(3,5)P2 but also PtdIns5P synthesis by PIKfyve and induces bafilomycin A1- reversible aberrant endomembrane dilation. PLoS One 2018, 13, No. e0204532. (18) Wada, Y.; Lu, R.; Zhou, D.; Chu, J.; Przewloka, T.; Zhang, S.; Li, L.; Wu, Y.; Qin, J.; Balasubramanyam, V.; Barsoum, J.; Ono, M. Selective abrogation of Th1 response by STA-5326, a potent IL-12/ IL-23 inhibitor. Blood 2007, 109, 1156−1164. (19) Baranov, M. V.; Bianchi, F.; Bogaart, G. V. D. The PIKfyve inhibitor apilimod: a double-edged sword against COVID-19. Cell 2021, 10, 30. (20) Cromm, P. M.; Crews, C. M. Targeted protein degradation: from chemical biology to drug discovery. Cell Chem. Biol. 2017, 24, 1181−1190. (21) Wang, Z.; Huang, W.; Zhou, K.; Ren, X.; Ding, K. Targeting the non-catalytic functions: a new paradigm for kinase drug discovery? J. Med. Chem. 2022, 65, 1735−1748. (22) Cai, X.; Xu, Y.; Cheung, A. K.; Tomlinson, R. C.; Alcázar- Román, A.; Murphy, L.; Billich, A.; Zhang, B.; Feng, Y.; Klumpp, M.; Rondeau, J. M.; Fazal, A. N.; Wilson, C. J.; Myer, V.; Joberty, G.; Bouwmeester, T.; Labow, M. A.; Finan, P. M.; Porter, J. A.; Ploegh, H. L.; Baird, D.; De Camilli, P.; Tallarico, J. A.; Huang, Q. PIKfyve, a class III PI kinase, is the target of the small molecular IL-12/IL-23 inhibitor apilimod and a player in Toll-like receptor signaling. Chem. Biol. 2013, 20, 912−921. (23) Casement, R.; Bond, A.; Craigon, C.; Ciulli, A. Mechanistic and structural features of PROTAC ternary complexes. In Targeted protein degradation: methods and protocols; Springer US, 2021; pp 79−113. (24) Zhou, Z.; Zhou, G.; Zhou, C.; Fan, Z.; Cui, R.; Li, Y.; Li, R.; Gu, Y.; Li, H.; Ge, Z.; Cai, X.; Jiang, B.; Wang, D.; Zheng, M.; Xu, T.; Zhang, S. Discovery of a potent, cooperative, and selective SOS1 PROTAC ZZ151 with in vivo antitumor efficacy in KRAS-mutant cancers. J. Med. Chem. 2023, 66, 4197−4214. (25) Soares, P.; Gadd, M. S.; Frost, J.; Galdeano, C.; Ellis, L.; Epemolu, O.; Rocha, S.; Read, K. D.; Ciulli, A. Group-based optimization of potent and cell-active inhibitors of the von Hippel− Lindau (VHL) E3 ubiquitin ligase: structure−activity relationships leading to the chemical probe (2S,4R)-1-((S)-2-(1-cyanocyclopropa- necarboxamido)-3,3-dimethylbutanoyl)-4-hydroxy-N-(4-(4-methyl- thiazol-5-yl)benzyl)pyrrolidine-2-carboxamide (VH298). J. Med. Chem. 2018, 61, 599−618. (26) Xiao, L.; Parolia, A.; Qiao, Y.; Bawa, P.; Eyunni, S.; Mannan, R.; Carson, S. E.; Chang, Y.; Wang, X.; Zhang, Y.; Vo, J. N.; Kregel, S.; Simko, S. A.; Delekta, A. D.; Jaber, M.; Zheng, H.; Apel, I. J.; McMurry, L.; Su, F.; Wang, R.; Zelenka-Wang, S.; Sasmal, S.; Khare, L.; Mukherjee, S.; Abbineni, C.; Aithal, K.; Bhakta, M. S.; Ghurye, J.; Cao, X.; Navone, N. M.; Nesvizhskii, A. I.; Mehra, R.; Vaishampayan, U.; Blanchette, M.; Wang, Y.; Samajdar, S.; Ramachandra, M.; Chinnaiyan, A. M. Targeting SWI/SNF ATPases in enhancer- addicted prostate cancer. Nature 2022, 601, 434−439. (27) Burslem, G. M.; Crews, C. M. Proteolysis-targeting chimeras as therapeutics and tools for biological discovery. Cell 2020, 181, 102− 114. 12445 https://doi.org/10.1021/acs.jmedchem.3c00912 J. Med. Chem. 2023, 66, 12432−12445
10.3390_insects14020130
Communication Why Petals? Naïve, but Not Experienced Bees, Preferentially Visit Flowers with Larger Visual Signals Nicholas J. Balfour * and Francis L. W. Ratnieks Laboratory of Apiculture & Social Insects, School of Life Sciences, University of Sussex, Brighton BN1 9QG, UK * Correspondence: n.balfour@sussex.ac.uk; Tel.: +44-(0)1273-872954 Simple Summary: Why do plants have showy flowers? Clearly, many plant species must attract pollinators and have floral adaptations for this. However, some flowers attract insects without showy petals. This suggests that a key function of showy visual signals is to attract naïve, first-time pollinator visitors. This is similar to how a restaurant with a large sign and showy visual signals may be especially important in gaining first-time visits when competing with other establishments or plants. Customers or pollinators will know if the first visit is rewarding and use this to decide whether to revisit. First, they must visit. Most flowers and restaurants would benefit from more visitors. Restaurants usually have empty tables, and many flowers are pollen limited in their reproduction. Here, we manipulated the ray petals of inflorescences of two garden flowers to test the hypothesis that the role of showy visual signals is to attract naïve visitors. On their first inflorescence visit to both species, naïve honey bees and bumble bees were more likely to visit intact inflorescences, than those with ray petals removed. However, by the tenth consecutive inflorescence, bees showed no preference. A positive correlation was observed between the visitation of inflorescences with no petals and the inflorescence number on both study plants, for both bees. These results strongly suggest that a key function of showy petals is to attract naïve, first-time visitors. Abstract: Flower evolution includes a range of questions concerning the function of showy morpho- logical features such as petals. Despite extensive research on the role of petals in attracting pollinators, there has been little experimental testing of their importance in attracting naïve versus experienced flower-visitors. In an exploratory field study, we manipulated the ray petals of inflorescences of two garden flowers, Rudbeckia hirta and Helenium autumnale, to test the hypothesis that these showy structures primarily function to attract first-time, naïve, visitors. On their first inflorescence visit to both species, naïve honey bees and bumble bees were more likely to visit intact inflorescences, than those with ray petals removed. However, by the tenth consecutive inflorescence on the same visit to the flower patch, test insects showed no preference. A positive correlation was observed between the visitation of inflorescences with zero petals and inflorescence number on both study plants, for both bees. These results suggest that a key function of showy petals is to attract naïve, first-time visitors. Similar to how a restaurant attracts diners with a large sign, showy signals may be vital to enticing first-time visitors when competing with other establishments or plants for customers or pollinators. We hope the findings of this exploratory study will stimulate further work in this area. Keywords: bees; behaviour; flowers; floral advertising; foraging; ray petals 1. Introduction Why do plants have showy flowers? Clearly, many plant species (~85% [1]) must attract pollinators and have floral adaptations for this [2–5]. Petals or analogous flower parts are generally not green to contrast visually against background foliage [5] and ultraviolet light, a colour seen by most insects, is often part of the visual signal [6]. This suggests that a key function of showy visual signals is to attract naïve, first-time pollinator visitors. Similar to how a restaurant with a large sign, showy visual signals may be especially important in Citation: Balfour, N.J.; Ratnieks, F.L.W. Why Petals? Naïve, but Not Experienced Bees, Preferentially Visit Flowers with Larger Visual Signals. Insects 2023, 14, 130. https:// doi.org/10.3390/insects14020130 Academic Editor: Monique M. van Oers Received: 6 December 2022 Revised: 16 January 2023 Accepted: 24 January 2023 Published: 26 January 2023 Copyright: © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). Insects 2023, 14, 130. https://doi.org/10.3390/insects14020130 https://www.mdpi.com/journal/insects insects Insects 2023, 14, 130 2 of 7 gaining first-time visits when competing with other establishments or plants. Customers or pollinators will know if the first visit is rewarding and use this to decide whether to revisit. First, they must visit. Restaurants usually have empty tables [7], and many flowers are pollen limited in their reproduction [8]. As such, most flowers and restaurants would benefit from more visitors. Asteraceae contains over 25,000 species, are abundant on every continent [9], and are considered the largest, most successful, and highly evolved of all plant families [10]. One particular adaptation that this group possesses is the capitular inflorescence, which comprises many small flowers that open over one or more weeks [11]. In addition, many Asteraceae possess flowers with accessible floral rewards, meaning that they are generalists who cater to a wide variety of potential flower-visiting species [12,13]. The small tubular disc flowers of Asteraceae are, however, modest and, unless en masse, inconspicuous. Hence, many species have showy marginal or ray petals which are either sterile or female. Here, we test and provide initial support for the hypothesis that showy petals are par- ticularly attractive to first-time visitors by removing the ray petals from the inflorescences of two Asteraceae and comparing the flower choices of naïve and experienced bee visitors to control inflorescences. 2. Methods 2.1. Experimental Setup All data were collected in a private domestic garden (Hazelmere, Magham Down, East Sussex, 50.880, 0.284) between 1000 and 1600 h during July and August 2021 and in weather conditions suitable for all flower-visitor activity (generally sunny, ≥16 ◦C and light wind). We used exotic Asteraceae to minimise the likelihood that flower visitors in the area had experienced these species: Rudbeckia hirta (var. ‘Black Eyed Susan’), and Helenium autumnale (var. ‘Sahin’s Early Flowerer’). These varieties were also selected for their large and conspicuous ray petals. Asteraceaeae were ideal subjects as the capitulum inflorescence is robust and easy to manipulate by removing ray-floret petals manually. Importantly, the central disc in the study species was large so the removal of petals did not render the remainder of the flower inadequate for the study insects to land on. In many flowers, the petals form part of the landing platform, so removing them would compromise insect visitation. Each of the four study patches measured approximately 4 m2 and comprised 16–20 plants of a single variety in full bloom with, on average, 132 (standard deviation, ±33.6) inflo- rescences. The plants were in 10 l pots, 0.5–1 m in height, and were placed close together so that the distance between inflorescences was <20 cm. For the experienced flower-visitor experiment, we set out one patch of each study plant variety with inflorescences of the three treatments on 20 July 2021. For the naïve flower-visitor experiment, we set out a further patch of H. annus on 19 August 2021 and a further patch of R. hirta on 20 August 2021. Because honeybees and other flower-visitors show an innate preference for symmet- rical flower shapes [14–16](Giurfa et al., 1996; Möller and Sorci, 1998; Orban et al., 2015), we used symmetrical ray petal treatments. Each capitulum was subject to one of three treatments: (i) zero ray petals (i.e., all petals removed), (ii) four ray petals in a cross arrange- ment, and (iii) all petals (i.e., all petals left intact). By judicially selecting which treatment newly blooming inflorescences received we ensured each patch contained equal numbers of inflorescences per treatment and that the treatments were dispersed approximately evenly across each patch. Treatments were equalised throughout the experimental period as needed. Old inflorescences were removed. 2.2. Inflorescence Measurements For both study plants, we counted the number of ray petals per inflorescence and measured the diameter of ten intact inflorescences (i.e., central disc plus ray petals) and the central discs alone. These data were used to determine the relative area of visual display of the central disc or the intact flower for the two study plant species. Insects 2023, 14, 130 3 of 7 2.3. Experienced Flower-Visitors To identify foragers that had previously visited the experimental patches (i.e., experi- enced visitors), from 09 to 11 August 2021, we used acrylic paint to uniquely mark actively foraging flower-visitors. During 12–14 August, flower-visitors were followed and the number and sequence of the treatment of the inflorescences visited by marked individuals were recorded. We collected data for a total of 22 h over this period. 2.4. Naïve Flower-Visitors Data were collected during six study days, from 19 August to 6 September 2020. When data were not being collected plants were placed in commercially available fine mesh cages (GardenSkill 1.35 m pop-up cage, MPN: GPN100/125-04) to exclude all insect visitors after removing the cages. We patiently waited beside the patches for naïve insects to arrive and recorded the sequence in which they visited ten experimental inflorescences. To ensure we only studied naïve insects, we then captured and marked all insects with acrylic paint using a honey bee queen marking cage. Although we cannot be certain that the flower- visitors studied had not visited these flower species previously as their prior experience is unknown, they were naïve to our study patches and treatments. 2.5. Statistical Analysis Statistical analyses were conducted using ‘R’ software (version 3.4.3 [17]). The as- sumptions of a beta regression model were met and the ‘betareg’ package was used for analysis [18]. Beta regression was used to test the relationship between the response vari- able, the proportion of visits to experimental treatments (zero, four or all petals), and the explanatory variable, the number of inflorescences visited. Chi-squared analyses compared raw numbers (observation proportions) to expected probabilities (equal visitation across treatments). Data visualisations were created using the package ‘ggplot2’ [19]. Bees were grouped by genus for analysis. 3. Results 3.1. Inflorescence Measurements Both study plants had hemispherical to spherical central discs with hundreds of florets. H. autumnale ray petals are yellow with orange tinges. Inflorescences had 14.7 ± 2.06 (mean ± SD) ray petals. The diameter of the whole inflorescence was 64.5 ± 3.92 mm and the central disc 21.6 ± 0.70 mm. R. hirta’s ray florets were yellow with orange/brown Inflorescences had 13.0 ± 0.67 ray petals, with a diameter of tinges in some plants. 108.5 ± 19.51 mm, with a central disc of 21.6mm ± 0.70. Given that the total inflores- cence diameters were approximately 3 and 5 times that of the central disc, the ray petals provided by far the greater area of visual display, c. 8 and 24 times that of the central disc. 3.2. Experienced Visitors We recorded ten uniquely marked experienced honey bees foraging on each of the two study flowers and followed them for an average of 23 consecutive inflorescences. These bees visited significantly fewer R. hirta inflorescences with no petals compared to those with four or all petals (no petals: 68, four petals: 99, all petals: 103; χ2 = 8.156, df = 2, p = 0.017). However, there was no visitation trend on H. autumnale inflorescences (58, 65, 60; χ2 = 0.426, df = 2, p = 0.808). 3.3. Naïve Visitors We recorded 42 naïve honey bees (R. hirta: 23; H. autumnale: 19) and 28 bumble bees (Bombus terrestris/lucorum and B. pascuorum; R. hirta: 17; H. autumnale: 11) foraging on ten consecutive inflorescences. The majority of naïve honey bees (R. hirta: 0.83; H. autumnale: 0.73) and bumble bees (R. hirta: 0.83; H. autumnale: 0.72) initially visited inflorescences with all petals on both plant species, with a very low proportion visiting those with zero petals (honey bees: 0.04, Insects 2023, 14, 130 4 of 7 0.06; bumble bees: 0.00, 0.18). This difference was significant for honey bees on both plants (R. hirta: χ2 = 25.391, df = 2, p < 0.001; H. autumnale: χ2 = 13, df = 2, p = 0.002). However, by the tenth inflorescences, all three treatments were visited by honey bees in roughly equal proportions and did not differ significantly across the three treatments on either plant species (R. hirta: χ2 = 1.652, df = 2, p = 0.438; H. autumnale: χ2 = 0.333, df = 2, p = 0.847). Bumble bees followed a similar pattern, predominately visiting R. hirta inflorescences with all petals initially (χ2 = 21, df = 2, p < 0.001) and by the tenth flower, no preference was evident (χ2 = 1.33, df = 2, p = 0.513). Sample sizes were not great enough to allow the analysis of the H. autumnale bumble bee data. Nevertheless, the trends in the data followed a similar trajectory (Figure 1). Figure 1. Proportion of visits by naïve bumbles bees (a,c) and honey bees (b,d) to inflorescences with all ray petals present (red circles), with four petals (yellow triangles), and those with all their ray petals removed (blue squares). The inflorescences number in the sequence of the bees’ first ten visits is given on the x-axis. Shown are beta regressions (lines) and 95% confidence intervals (shaded areas). Overall, the data show a negative correlation between the proportion of inflorescences with all petals visited by both honey bees and bumble bees and the inflorescence number for both R. hirta and H. autumnale (Figure 1). Two of these four analyses indicated a significant correlation (Table 1). No significant trends were observed with the visitation of flowers with four petals. By contrast, a positive correlation was observed between the proportion of inflorescences with no petals visited by honey bees and bumbles bees and inflorescence Insects 2023, 14, 130 4 of 7 3.3. Naïve Visitors We recorded 42 naïve honey bees (R. hirta: 23; H. autumnale: 19) and 28 bumble bees (Bombus terrestris/lucorum and B. pascuorum; R. hirta: 17; H. autumnale: 11) foraging on ten consecutive inflorescences. The majority of naïve honey bees (R. hirta: 0.83; H. autumnale: 0.73) and bumble bees (R. hirta: 0.83; H. autumnale: 0.72) initially visited inflorescences with all petals on both plant species, with a very low proportion visiting those with zero petals (honey bees: 0.04, 0.06; bumble bees: 0.00, 0.18). This difference was significant for honey bees on both plants (R. hirta: χ2 = 25.391, df = 2, p < 0.001; H. autumnale: χ2 = 13, df = 2, p = 0.002). However, by the tenth inflorescences, all three treatments were visited by honey bees in roughly equal proportions and did not differ significantly across the three treatments on either plant species (R. hirta: χ2 = 1.652, df = 2, p = 0.438; H. autumnale: χ2 = 0.333, df = 2, p = 0.847). Bumble bees followed a similar pattern, predominately visiting R. hirta inflorescences with all petals initially (χ2 = 21, df = 2, p < 0.001) and by the tenth flower, no preference was evident (χ2 = 1.33, df = 2, p = 0.513). Sample sizes were not great enough to allow the analysis of the H. autumnale bumble bee data. Nevertheless, the trends in the data followed a similar trajectory (Figure 1). Figure 1. Proportion of visits by naïve bumbles bees (a,c) and honey bees (b,d) to inflorescences with all ray petals present (red circles), with four petals (yellow triangles), and those with all their ray petals removed (blue squares). The inflorescences number in the sequence of the bees’ first ten visits is given on the x-axis. Shown are beta regressions (lines) and 95% confidence intervals (shaded ar-eas). Insects 2023, 14, 130 5 of 7 number on both study plants. All four analyses indicated the correlation was statistically significant (Table 1). Table 1. Output from beta regression analysis of naïve bees (Pseudo R2, z-value, and p-value) for the two study plant species, two study bee genera, and treatment types. Treatments: (i) All petals (all petals left intact), (ii) Four petals (all but four petals removed), and (ii) No petals (all petals removed). Plant Species Bee Genus Treatment Rudbeckia hirta Rudbeckia hirta Rudbeckia hirta Helenium autumnale Helenium autumnale Helenium autumnale Rudbeckia hirta Rudbeckia hirta Rudbeckia hirta Helenium autumnale Helenium autumnale Helenium autumnale 4. Discussion Apis Apis Apis Apis Apis Apis Bombus Bombus Bombus Bombus Bombus Bombus All petals Four petals No petals All petals Four petals No petals All petals Four petals No petals All petals Four petals No petals R2 0.18 0.01 0.40 0.40 0.26 0.30 0.28 0.13 0.32 0.69 0.16 0.33 z-Value p-Value −1.497 −0.309 2.263 −2.521 1.636 2.897 −1.850 1.067 2.58 −4.782 2.326 2.386 0.134 0.757 0.008 0.012 0.102 0.004 0.064 0.286 0.010 <0.001 0.259 0.017 The results indicate that the additional visual signal provided by the ray petals func- tioned differently for naïve and experienced bee visitors. Initially, naïve honey bees and bumble bees preferentially visited treatment inflorescences with all petals on both plant species. However, very quickly and by the tenth inflorescence visit, foragers showed no preference among inflorescences with all, four, or zero ray petals. Flowers “advertise” themselves via showy displays [20] and in doing so incur costs [21]. The conspicuous yellow ray petals of our study plants presumably served as long-distance signals that attracted naive flower-visitors to the patches. In many plant species, larger floral displays have been documented to increase the number of long-distance approaches by pollinators [22,23]. Both plant species also possess ultraviolet floral guides [6,24,25], which serve to orient bees at close range toward the reward (i.e., nectar and/or pollen), and which also known to increase visitation rates [6]. Large and colourful floral displays are known to reduce the search times and influence the flower choice of honey bees and bumble bees due to limitations in their visual resolution [26–28]. Numerous observational and experimental studies in other species of Asteraceae have shown that capitula with ray petals receive more visitation than rayless [4,29–32]. Our results agree with this to a degree, but with the important additional result that the ray petals primarily serve to attract naïve insects, indicating that the primary function of ray petals is to attract naïve, first-time visitors. To give a human analogy, a large showy restaurant advertisement may be attractive to first-time visitors. However, once the location and the quality of the food on offer at the restaurant are known, large signage is a less important criterion in the decision to return as the value of the resource and its location are now known. Bees are known to exhibit both long-range and short-range selectivity when foraging and the criteria can differ between the two [33]. Our results show that naïve bees very quickly changed their behaviour. This agrees with previous studies which have shown that bee foraging behaviour and flower selection is extremely flexible. For example, bees have been observed to associate display size [34], flower colour [35], and patch [36] with nectar rewards within their first three flower visits. Coevolution between flowering plants and pollinators has ramifications for many elements of pollinator biology. Classically, for example, there has been much emphasis on floral traits that advantage pollinators that have particular morphological features, such as recessed nectar requiring an insect with a long tongue, as in Darwin’s Orchid (Angraecum sesquipedale) which is pollinated by a long-tongued Sphingidae moth [37]. In the case of petals, many flowers pollinated by birds have red petals, a colour that most insects cannot Insects 2023, 14, 130 6 of 7 see. However, in the case of the red field poppy, Papaver rhoeas, in its native area it and other red-petalled flowers are visited by Glaphyridae beetles that can see red. These do not occur in Europe, where following its introduction by early agriculturalists P. rhoeas has changed its petal colour from red to red+ultra violet, which can be seen by bees [38]. Our results also show that the ability of insects to rapidly learn may also have an important role in the evolution of floral traits. This is also seen in adaptive-colour change in floral parts, by which insects learn to avoid flowers that have changed colour to signify that they are no longer rewarding [39]. In short, the ability of bees and other pollinators to rapidly responds to differences in rewards is likely to be a significant selective force on the evolution of floral traits. Our research provides interesting results and a novel hypothesis on the role of petals in plant-pollinator coevolution. But of course, further work on a larger scale is needed to test this hypothesis and determine the generality of our findings. For example, presenting bees with pure- and mixed-signal patches would allow investigation into the effect of long- versus. short-range signals. Moreover, further work should exclude the possibility that removing petals either reduces nectar production or releases repellent volatiles. We hope the findings of this exploratory study will stimulate further research in this area. Author Contributions: Conceptualization, F.L.W.R.; methodology, F.L.W.R. and N.J.B.; validation, F.L.W.R. and N.J.B.; formal analysis, N.J.B.; investigation, F.L.W.R. and N.J.B.; resources, F.L.W.R.; data curation, N.J.B.; writing—original draft preparation, N.J.B.; writing—review and editing, F.L.W.R. and N.J.B.; visualization, N.J.B.; funding acquisition, F.L.W.R. All authors have read and agreed to the published version of the manuscript. Funding: N.B. was funded by Rowse Honey Ltd. Data Availability Statement: Data associated with this manuscript are accessible at Figshare https: //www.mdpi.com/1999-4893/16/2/112 (Balfour and Ratnieks, 2022). Acknowledgments: We thank to our funders Rowse Honey Ltd. Conflicts of Interest: The authors declare no conflict of interest. References Ollerton, J.; Winfree, R.; Tarrant, S. How many flowering plants are pollinated by animals? Oikos 2011, 120, 321–326. [CrossRef] 1. 2. Marshall, D.F.; Abbott, R.J. Polymorphism for outcrossing frequency at the ray floret locus in Senecio vulgaris L. II. Confirmation. 3. 4. 5. Heredity 1984, 52, 331–336. [CrossRef] Sun, M.; Ganders, F.R. Outcrossing rates and allozyme variation in rayed and rayless morphs of Bidens pilosa. Heredity 1990, 64, 139–143. [CrossRef] Andersson, S. Pollinator and non-pollinator selection on ray morphology in Leucanthemum vulgare (oxeye daisy, Asteraceae). Am. J. Bot. 2008, 95, 1072–1078. [CrossRef] [PubMed] Bukovac, Z.; Shrestha, M.; Garcia, J.E.; Burd, M.; Dorin, A.; Dyer, A.G. Why background colour matters to bees and flowers. J. Comp. Physiol. A 2017, 203, 369–380. [CrossRef] 6. Horth, L.; Campbell, L.; Bray, R. Wild bees preferentially visit Rudbeckia flower heads with exaggerated ultraviolet absorbing 7. 8. 9. floral guides. Biol. Open 2014, 3, 221–230. [CrossRef] [PubMed] Robson, S.K.; Kimes, S.E.; Becker, F.D.; Evans, G.W. Consumers’ responses to table spacing in restaurants. Cornell Hosp. Q. 2011, 52, 253–264. [CrossRef] Ashman, T.L.; Knight, T.M.; Steets, J.A.; Amarasekare, P.; Burd, M.; Campbell, D.R.; Dudash, M.R.; Johnston, M.O.; Mazer, S.J.; Mitchell, R.J.; et al. Pollen limitation of plant reproduction: Ecological and evolutionary causes and consequences. Ecology 2004, 85, 2408–2421. [CrossRef] Thorne, R.F. The classification and geography of the flowering plants: Dicotyledons of the class Angiospermae. Bot. Rev. 2000, 66, 441–647. 10. Proctor, M.; Yeo, P.; Lack, A. The Natural History of Pollination; HarperCollins: New York, NY, USA, 1996. 11. Burtt, B.L. Compositae and the study of functional evolution. Trans. Bot. Soc. Edinb. 1961, 39, 216–232. [CrossRef] 12. Balfour, N.J.; Fensome, K.A.; Samuelson, E.E.; Ratnieks, F.L. Following the dance: Ground survey of flowers and flower-visiting insects in a summer foraging hotspot identified via honey bee waggle dance decoding. Agric. Ecosyst. Environ. 2015, 213, 265–271. [CrossRef] 13. Balfour, N.J.; Ratnieks, F.L. The disproportionate value of ‘weeds’ to pollinators and biodiversity. J. Appl. Ecol. 2022, 59, 1209–1218. [CrossRef] 14. Giurfa, M.; Eichmann, B.; Menzel, R. Symmetry perception in an insect. Nature 1996, 382, 458–461. [CrossRef] Insects 2023, 14, 130 7 of 7 15. Møller, A.P.; Sorci, G. Insect preference for symmetrical artificial flowers. Oecologia 1998, 114, 37–42. [CrossRef] 16. Orbán, L.L.; Chartier, S. Unsupervised neural network quantifies the cost of visual information processing. PLoS ONE 2015, 10, e0132218. [CrossRef] 17. R-Project. R: A Language and Environment for Statistical Computing; R Foundation: Vienna, Austria, 2019; Available online: http://www.r-project.org/ (accessed on 21 January 2022). 18. Zeileis, A.; Cribari-Neto, F.; Gruen, B.; Kosmidis, I.; Simas, A.B.; Rocha, A.V.; Zeileis, M.A. Package ‘Betareg’. Available online: https://cran.r-project.org/web/packages/betareg/index.html (accessed on 25 January 2018). 19. Wickham, H. ggplot2: Elegant Graphics for Data Analysis; Springer: London, UK, 2016. 20. Dafni, A.; Lehrer, M.; Kevan, P.G. Spatial flower parameters and insect spatial vision. Biol. Rev. 1997, 72, 239–282. [CrossRef] 21. Primack, R.B.; Hall, P. Costs of reproduction in the pink lady’s slipper orchid: A four-year experimental study. Am. Nat. 1990, 136, 638–656. [CrossRef] Sih, A.; Baltus, M.S. Patch size, pollinator behavior, and pollinator limitation in catnip. Ecology 1987, 68, 1679–1690. [CrossRef] 22. 23. Grindeland, J.M.; Sletvold, N.; Ims, R.A. Effects of floral display size and plant density on pollinator visitation rate in a natural population of Digitalis purpurea. Funct. Ecol. 2005, 19, 383–390. [CrossRef] 24. Watt, W.B.; Hoch, P.C.; Mills, S.G. Nectar resource use by Colias butterflies. Oecologia 1974, 14, 353–374. [CrossRef] 25. Martin, N.H. Flower size preferences of the honeybee (Apis mellifera) foraging on Mimulus guttatus (Scrophulariaceae). Evol. Ecol. Res. 2004, 6, 777–782. 26. Lunau, K. Innate flower recognition in bumblebees (Bombus terrestris, B. lucorum; Apidae): Optical signals from stamens as 27. landing reaction releasers. Ethology 1991, 88, 203–214. [CrossRef] Spaethe, J.; Tautz, J.; Chittka, L. Visual constraints in foraging bumblebees: Flower size and color affect search time and flight behavior. Proc. Natl. Acad. Sci. USA 2001, 98, 3898–3903. [CrossRef] 28. Dyer, A.G.; Spaethe, J.; Prack, S. Comparative psychophysics of bumblebee and honeybee colour discrimination and object detection. J. Comp. Physiol. A 2008, 194, 617–627. [CrossRef] 29. Lack, A.J. Competition for pollinators in the ecology of Centaurea scabiosa L. and Centaurea nigra L. III. Insect visits and the number of successful pollinations. New Phytol. 1982, 91, 321–339. [CrossRef] 30. Andersson, S. Floral display and pollination success in Senecio jacobaea (Asteraceae): Interactive effects of head and corymb size. Am. J. Bot. 1996, 83, 71–75. [CrossRef] 31. Olsen, K.M. Pollination effectiveness and pollinator importance in a population of Heterotheca subaxillaris (Asteraceae). Oecologia 1996, 109, 114–121. [CrossRef] 32. Nielsen, L.R.; Philipp, M.; Siegismund, H.R. Selective advantage of ray florets in Scalesia affinis and S. pedunculata (Asteraceae), two endemic species from the Galápagos. Evol. Ecol. 2002, 16, 39–153. [CrossRef] 33. Goulson, D.; Cruise, J.L.; Sparrow, K.R.; Harris, A.J.; Park, K.J.; Tinsley, M.C.; Gilburn, A.S. Choosing rewarding flowers; perceptual limitations and innate preferences influence decision making in bumblebees and honeybees. Behav. Ecol. Sociobiol. 2007, 61, 1523–1529. [CrossRef] 34. Blarer, A.; Keasar, T.; Shmida, A. Possible mechanisms for the formation of flower size preferences by foraging bumblebees. Ethology 2002, 108, 341–351. [CrossRef] 35. Grüter, C.; Moore, H.; Firmin, N.; Helanterä, H.; Ratnieks, F.L.W. Flower constancy in honey bee workers (Apis mellifera) depends on ecologically realistic rewards. J. Exp. Biol. 2011, 214, 1397–1402. [CrossRef] [PubMed] 36. Balfour, N.J.; Gandy, S.; Ratnieks, F.L. Exploitative competition alters bee foraging and flower choice. Behav. Ecol. Sociobiol. 2015, 69, 1731–1738. [CrossRef] 37. Wasserthal, L.T. The pollinators of the Malagasy star orchids Angraecum sesquipedale, A. sororium and A. compactum and the evolution of extremely long spurs by pollinator shift. Bot. Acta 1997, 110, 343–359. [CrossRef] 38. Martínez-Harms, J.; Hadar, R.; Márquez, N.; Menzel, R.; Shmida, A.; Stavenga, D.G.; Vorobyev, M. Enhanced UV-reflection facilitated a shift in the pollination system of the red poppy, Papaver rhoeas (Papaveraceae). Plants 2020, 9, 927. [CrossRef] [PubMed] 39. Weiss, M.R. Floral color change: A widespread functional convergence. Am. J. Bot. 1995, 82, 167–185. [CrossRef] Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
10.1016_j.neo.2023.100910
Neoplasia 42 (2023) 100910 Contents lists available at ScienceDirect Neoplasia journal homepage: www.elsevier.com/locate/neo Original Research Genomics driven precision oncology in advanced biliary tract cancer improves survival ✩ , ✩✩ Chandan Kumar-Sinha a , b , 3 , Pankaj Vats a , b , 1 , 3 , Nguyen Tran c , 2 , 3 , Dan R. Robinson a , b , Valerie Gunchick c , Yi-Mi Wu a , b , Xuhong Cao a , b , Yu Ning a , Rui Wang a , Erica Rabban a , Janice Bell a , Sunita Shankar a , Rahul Mannan a , Yuping Zhang a , Mark M. Zalupski c , d , Arul M. Chinnaiyan a , b , d , e , 4 , ∗ , Vaibhav Sahai c , d , 4 , ∗ a Michigan Center for Translational Pathology, University of Michigan, Ann Arbor, MI 48109, USA b Department of Pathology, University of Michigan, Ann Arbor, MI 48109, USA c Department of Internal Medicine, University of Michigan, Ann Arbor, MI 48109, USA d Rogel Cancer Center, University of Michigan, Ann Arbor, MI 48109, USA e Howard Hughes Medical Institute, Chevy Chase, MD 20815, USA a r t i c l e i n f o a b s t r a c t Keywords: Cholangiocarcinoma Next-generation sequencing Precision oncology Background: Biliary tract cancers (BTCs) including intrahepatic, perihilar, and distal cholangiocarcinoma as well as gallbladder cancer, are rare but aggressive malignancies with few effective standard of care therapies. Methods: We implemented integrative clinical sequencing of advanced BTC tumors from 124 consecutive pa- tients who progressed on standard therapies (N = 92 with MI-ONCOSEQ and N = 32 with commercial gene panels) enrolled between 2011-2020. Results: Genomic profiling of paired tumor and normal DNA and tumor transcriptome (RNA) sequencing identified actionable somatic and germline genomic alterations in 54 patients (43.5%), and potentially actionable alterations in 79 (63.7%) of the cohort. Of these, patients who received matched targeted therapy (22; 40.7%) had a median overall survival of 28.1 months compared to 13.3 months in those who did not receive matched targeted therapy (32; P < 0.01), or 13.9 months in those without actionable mutations (70; P < 0.01). Additionally, we discovered recurrent activating mutations in FGFR2 , and a novel association between KRAS and BRAF mutant tumors with high expression of immune modulatory protein NT5E (CD73) that may represent novel therapeutic avenues. Conclusions: Overall, the identification of actionable/ potentially actionable aberrations in a large proportion of cases, and improvement in survival with precision oncology supports molecular analysis and clinical sequencing for all patients with advanced BTC. Introduction Biliary tract cancers (BTCs) arise from the epithelial lining of the biliary ducts and comprise of intrahepatic and extrahepatic (perihilar and distal) cholangiocarcinoma (CCA), and gallbladder cancer. The in- cidence of these cancers is rising, driven predominantly by intrahepatic CCA [1–3] . A majority of these tumors are advanced and unresectable at diagnosis [1] . Prognosis of patients with advanced BTC remains poor ✩ Grant Support: This work was supported by grants from the NCI Early Detection Research Network (U01CA214170) and NIH/NCI Outstanding Investigator Award (R35CA231996) ✩✩ Disclosures: MMZ – Institutional grant funding from AstraZeneca, MedImmune and Seattle Genetics. VS – Institutional grant funding from Agios, Bristol-Myers Squibb, Celgene, Clovis, Cornerstone, Exelixis, Fibrogen, Incyte, Ipsen, Medimmune, Merck, NCI, Rogel Cancer Center, Repare, Relay, Servier, Syros and Transthera; and consultant fees from AstraZeneca, Autem, Cornerstone, Delcath Systems, GlaxoSmithKline, Helsinn, Histosonics, Incyte, Ipsen, Kinnate, Lynx Group, QED, Servier and Taiho. Corresponding authors at: Division of Hematology and Oncology, Rogel Cancer Center, University of Michigan Medical School, 1500 E. Medical Center Dr., C412 ∗ MIB, Ann Arbor, MI 48109-5948, USA. E-mail addresses: arul@umich.edu (A.M. Chinnaiyan), vsahai@umich.edu (V. Sahai) . Work completed at University of Michigan; currently employee at Nvidia Work completed at University of Michigan; currently at Department of Oncology, Mayo Clinic, Rochester, MN equal contribution, co-first authorship Co-senior author 1 2 3 4 https://doi.org/10.1016/j.neo.2023.100910 Received 11 May 2023; Accepted 12 May 2023 1476-5586/© 2023 Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license ( http://creativecommons.org/licenses/by-nc-nd/4.0/ ) C. Kumar-Sinha, P. Vats, N. Tran et al. Neoplasia 42 (2023) 100910 with median overall survival (OS) from diagnosis of less than 12 months [4 , 5] , and five-year survival rate of about 5% despite therapy [6] . Cur- rent systemic chemotherapy options for patients with advanced BTCs remain nonspecific and suboptimal, thus it is imperative to further our understanding of the molecular biology of this disease and to define more targeted and effective therapeutic options. Molecular profiling of BTCs has identified many common drivers [7–11] , as well as molecular aberrations associated with specific anatomical subgroups, for example, FGFR2 fusions, and mutations in IDH1/2, BAP1, ARID1A , and KRAS predominantly seen in intrahepatic CCA (iCCA) [12–14] ; KRAS, TP53 , and ARID1A mutations or amplifi- cation of ERBB2 or ERBB3 in extrahepatic CCA; and TP53, ERBB2/3, CDKN2A/B , and ARID1A in gallbladder cancer [13 , 15] . Among these, targeted therapy options have recently become available for BTC pa- tients with FGFR2 and IDH1 aberrations. Pan-FGFR inhibitors pemiga- tinib and infigratinib received accelerated FDA approval for use in pa- tients with FGFR2 fusion or translocation who had progressed on first- line therapy, following multiple clinical trials that demonstrated clinical benefit in refractory BTC patients with objective response rates (ORR) ranging from 25% to 36%, and disease control rates as high as 70% to 80% [16–18] . Ivosidenib, an IDH1 inhibitor, demonstrated a statistically significant improvement in median progression-free survival (PFS) from 1.4 months to 2.7 months and disease control rate from 28% to 53% in patients as compared to placebo which led to its FDA approval in 2021 [19] . Other promising targets in BTC under clinical investigation include BRAF V600E [20] and ERBB2/HER-2 amplification [21] . Addi- tionally, application of immune checkpoint inhibitors (ICI) in BTCs has yielded variable benefit across several trials evaluating role of ICI as sin- gle agent or dual therapy, and combination with chemotherapy. In the frontline setting, the combination of chemotherapy with ICI resulted in response rates ranging from 27-73% and median PFS of 4.3–11.0 months and median OS of 10.6–20.7 months [22–25] . In the refractory popula- tion, single agent or combination immunotherapy demonstrated an ORR ranging from 5.8 to 23% with median PFS of 1.5–3.6 months and me- dian OS of 4.3–14.23 months [24 , 26–30] . In many of these trials, the median duration of response had not been reached suggesting a subset of patients had durable response. Fewer than 5% of patients with BTC have underlying microsatellite instability /deficient mismatch repair or high tumor mutational burden for which ICI has received FDA approval in a tissue-type agnostic manner. Herein, we summarize findings from clinical sequencing of 124 BTC patients with a focus on defining the spectrum of molecularly matched therapeutic options for this rare cancer and assess their impact on the clinical management of patients. Materials and methods Sequencing was performed via the MI-ONCOSEQ program us- ing standard protocols under Institutional Review Board (IRB HUM00046018, HUM00067928, HUM00056496) approved studies at Michigan Center for Translational Pathology a Clinical Laboratory Im- provement Amendments (CLIA) compliant sequencing lab at University of Michigan [31–34] . Patients enrolled in the MI-ONCOSEQ study pro- vided written informed consent to perform comprehensive molecular profiling of tumor/germline exomes and tumor transcriptome on either fresh tumor biopsies or (FFPE) tissue blocks. In addition, patient data was collected from the electronic medical records under IRB applica- tion HUM00165244. Next generation sequencing library preparation Sample details, including age, gender, and disease stage are sum- marized in Table 1 and Supplementary Table S1. Tissue acquisition, pathology review and sequencing of matched pair (tumor/normal DNA) exome, and tumor only transcriptome libraries were prepared using previously described protocols [32] . Samples with low tumor content were macro-dissected to enrich for tumor tissue based on pathologist assessment. “Human All Exon v4 ” Agilent exon probes and a selected target capture panel probes were used to capture tumor DNA and enriched following manufacturer’s protocol (Agilent/Roche). DNA/RNA paired-end sequencing libraries were sequenced using the Il- lumina HiSeq 2000 or HiSeq 2500 (2 × 100 nucleotide read length) (Illumina Inc. San Diego, CA). Exome sequencing analysis Whole Exome paired end Fastq sequence files were aligned to GRCh37 genome build using Novoalign multithreaded (version 2.08.02, Novocraft). Novosort and Picard (version 1.93) were used to sort, index and remove duplicates from the aligned bam files. Mutation analysis was carried out on matched normal–tumor pairs using freebayes (version 1.0.1) and pindel (version 0.2.5b9) as previously described [31 , 32 , 35] . Somatic SNV and Indel files from freebayes and pindel were postfiltered using at least 5% variant allelic fraction, minimum six variant reads, < 2% variant allelic fraction in normal with at least 20X coverage. The indel thresholds were optimized using a pool of hundreds of matched normal samples sequenced using the same protocol and platform as de- scribed [35] . Germline mutation analysis was performed using at least 10 variant reads in normal sample, with > = 20% allelic fraction and, < 1% population frequency in 1000 Genomes and ExAC. Variant anno- tation was performed using snpEff and snpSift (version 4.1g) based on refseq (from UCSC genome browser, retrieved on 8/22/2016), COSMIC v79, dbSNP v146, ExAC v0.3, and 1000 Genomes phase 3 databases. Copy number aberration analysis was performed on exome data using DNAcopy (version 1.48.0) to get CBS segments, regions were normalized for GC content, and log2-transformed exon coverage ratio between tumor/normal samples across the targeted regions were calculated as previously described [32 , 35] . Cohort wise copy number analysis was performed by merging all the segment files used as input to gistic version 2, and maftools was used to generate cumulative copy number plot. RNA sequencing data analysis Strand-specific RNA sequencing (RNA-seq) libraries were used for gene expression and fusion analysis. Gene expression quan- tification was performed using kallisto version 0.43.1, transcript per million (TPM) values were used as input for qlucore omics ( https://www.qlucore.com/ ) software for downstream expression anal- ysis. Genes with transcripts with < 1 TPM in at least 95% of the co- hort were removed and the data was transformed to log2. The expres- sion data was normalized for preservation method (FFPE/Fresh Frozen), biopsy sites and tumor content. Unsupervised hierarchical clustering was performed for the 69 immune marker genes, including 66 genes recently evaluated (Cancer Genome Atlas Research Network) plus IFN- 𝛾 responsive chemokines (CXCL9-11). Fusion calling was performed us- ing a combination of CRISP, CODAC MI-ONCOSEQ pipeline [32 , 35] , fusioncatcher_v1.10 [36] and arriba_v1.1.0. [37] The fusions calls were compiled and reported in Supplementary Table S8. Mutation burden estimation Freebaye’s mutation calls were used for the mutation burden estima- tion. Mutations were filtered for coverage ( > = 10x) and variant allelic fraction ( > = 6%). Mutation burden was expressed as (number of muta- tions/ total covered bases) × 10 6 . Varscan2 processed VCF files from TCGA CCA cohort (N = 51) were downloaded from the GDC data portal and lifted-over from the GRCh38 to GRCh37 reference genome using CrossMap for comparison with the MI-ONCOSEQ cohort. Pathogenic germline variant analysis Pathogenicity of germline variants were determined through review of the published literature, public databases including but not limited 2 C. Kumar-Sinha, P. Vats, N. Tran et al. Table 1 Patient Characteristics. Total, N (%) Age, years Median Range Sex, N (%) Female Male Race, N (%) White or Caucasian Asian or Asian American Black or African American American Indian Other or missing Primary tumor site, N (%) Intrahepatic Extrahepatic perihilar Extrahepatic distal Gallbladder Mixed hepatocellular/ cholangiocarcinoma First-line systemic therapy, N (%) Gemcitabine/platinum + /- agent Gemcitabine/taxane 5-fluorouracil based regimen Immunotherapy only Other Unknown or none NGS Platform, N (%) MI-ONCOSEQ Other Stage at biopsy, N (%) Resectable Locally advanced unresectable Metastatic Biopsy specimen, N (%) Primary Metastatic Biopsy in relation to chemotherapy, N (%) Pre-chemotherapy Post-chemotherapy Neoplasia 42 (2023) 100910 All Actionable Matched Treated Actionable Matched Untreated Non-actionable 124 (100) 22 (17.7) 59 17-80 65 (52) 59 (48) 111 (89.5) 5 (4.0) 4 (3.2) 2 (1.6) 2 (6) 88 (71) 10 (8) 8 (6) 13 (10) 5 (4) 90 (72.6) 4 (3.2) 15 (12.1) 5 (4.0) 4 (3.2) 6 (4.8) 92 (74.2) 32 (25.8) 16 (12.9) 11 (8.9) 97 (78.2) 61 (49) 63 (51) 41 (33) 83 (67) 56 27-75 13 (59.1) 9 (40.1) 21 (95.5) 1 (4.5) 21 (95.4) 1 (4.5) 16 (72.7) 2 (9.1) 1 (4.5) 3 (13.6) 14 (63.6) 8 (36.4) 2 (9.1) 1 (4.5) 19 (86.4) 13 (59.1) 9 (40.9) 5 (22.7) 17 (77.3) 32 (25.8) 58 17-72 19 (59.4) 13 (40.6) 30 (93.8) 1 (3.1) 1 (3.1) 27 (84.4) 1 (3.1) 2 (6.3) 1 (3.1) 1 (3.1) 24 (75.0) 3 (9.4) 1 (3.1) 1 (3.1) 3 (9.4) 27 (84.4) 5 (15.6) 3 (9.4) 4 (12.5) 25 (78.1) 20 (62.5) 12 (37.5) 13 (40.6) 19 (59.4) 70 (56.5) 61 20-80 33 (47.1) 37 (52.9) 59 (84.3) 3 (4.3) 5 (7.1) 1 (1.4) 2 (2.9) 40 (57.1) 9 (12.9) 6 (8.6) 12 (17.1) 3 (4.3) 50 (71.4) 1 (1.4) 12 (17.1) 3 (4.3) 1 (1.4) 3 (4.3) 51 (81.4) 19 (27.1) 11 (15.7) 6 (8.6) 53 (75.7) 29 (41.4) 41 (58.6) 23 (32.9) 47 (67.1) NGS, next-generation sequencing; MI-ONCOSEQ, Michigan Oncology Sequencing to ClinVar, the Human Genome Mutation Database, Leiden Open Vari- ation Databases, and variant-specific databases. Only cancer-relevant germline variants that had been previously categorized as pathogenic or likely pathogenic in ClinVar, or adjudicated at the precision molecu- lar tumor board as pathogenic were included in the study. Survival analysis Subject efficacy data was manually extracted from review of elec- tronic medical records. OS was defined as the duration of time from the date of advanced unresectable or metastatic disease until death from any cause. Follow-up time was censored at the date of last disease evalua- tion. The survival analysis was estimated using the product-limit method of Kaplan and Meier (GraphPad Prism 8, San Diego, CA). The analyzes should be considered post hoc, and the results herein exploratory with the intention to guide further definitive studies. A significance threshold for P value was arbitrarily set to 0.05 for all statistical tests. In order to associate patient outcome with reported molecular alter- ations, we included all consecutive subjects with BTC with targeted gene panel analysis completed using alternative CLIA platforms at our institu- tion. Gene alterations predictive of response to an FDA approved drug(s) were classified as ‘actionable (Tier 1)’, and aberrations associated with potential responsiveness to experimental drugs based on emerging data from ongoing clinical trials or compelling pre-clinical evidence were designated as ‘potentially actionable (Tier 2)’, and frequent aberrations noted in this cohort for which currently no therapeutic approach is avail- able were deemed non-actionable (Supplementary Table S9). 3 Results Somatic aberration landscape of advanced BTCs presents diverse therapeutic avenues Clinical sequencing data was obtained from a total of 124 consecu- tive patients with advanced BTC (from a total of 239 patients enrolled between September 2011 and February 2020 (Fig. S1). The sequencing cohort was comprised of 52% women, median age of 59 (range, 17– 80) years), including intrahepatic (N = 88), perihilar (N = 10), and distal (N = 8) CCA, mixed hepatocellular/CCA (N = 5), and gallbladder cancer (N = 13), with 83 (67%) cases being post-chemotherapy and 63 (51%) metastatic ( Table 1 and Supplementary Table S1). We obtained high quality exome sequencing data from 92 tumor/normal samples at MI- ONCOSEQ as indicated by the 94% median alignment rate (range, 53- 97%), mean coverage of 203X for whole exome (WXS) and 506X for target capture panel, and overall low PCR duplication rate averaging 8% (range 0.6–79%) (Supplementary Table S2). In parallel, high qual- ity capture transcriptome sequencing data from 85 tumor tissues was analyzed for gene fusions (Supplementary Table S7), and gene expres- sion (Supplementary Table S8). Additionally, tumors from 32 cases were analyzed through other CLIA-approved gene panels from commercial vendors including 30 from Foundation Medicine, and 2 from Guardant Health) (Supplementary Table S4). Somatic mutations (Supplementary Tables S3 and S4), copy num- ber aberrations (Supplementary Table S6), gene fusions (Supplemen- tary Table S7), and tumor mutation burden (Supplementary Table S5) C. Kumar-Sinha, P. Vats, N. Tran et al. Neoplasia 42 (2023) 100910 Fig. 1. Integrative landscape of molecular alterations in biliary cancer. Landscape of molecular alterations in biliary tract cancer cohort (n = 124). Each column represents a sample, identified at the bottom, each row represents a gene shown on the left, arranged according to functional groups as indicated. The bar on top displays cases with tumor mutation burden (TMB > 10 mutations/Mb). Mutation types are color coded: missense (green), indels (red), nonsense (black), splice site (orange), pathogenic germline (cyan), fusion (purple), copy number deletion (blue), copy number amplification (dark red), multiple hits (grey), and promotor mutation (dark pink). Respective percentages of cases with gene aberrations are shown on the right. from all the cases were assessed for potential clinical relevance and summarized in executive reports returned to the treating oncologists. The most frequently altered genes in the cohort included TP53 (N = 43, 34.7%) and CDKN2A (N = 36, 29%) followed by ARID1A (N = 25, 20.2%), IDH1/2 (N = 23, 18%), BAP1 (N = 19, 15.3%), FGFR2/3 (N = 25 20.2%), KRAS (N = 19, 15%), and PBRM1 (N = 13, 10%) ( Fig. 1 ). Among driver aberrations, FGFR2, KRAS and IDH1 were largely mutually exclusive except for one patient with concurrent mutations in IDH1 and KRAS . Actionable genomic alterations (Tier 1), defined as predictive of re- sponse to FDA approved drug(s) in any cancer, were noted in a to- tal of 54 cases (43.5%), and included hotspot activating/hotspot muta- tions in IDH1/2 (N = 23) , gene fusions in FGFR2/3 (N = 15), BRAF V600E mutation (N = 4), ERBB2 amplification (N = 4), deleterious mutations in BRCA1/BRCA2 (N = 4), and KRAS G12C (N = 3). Apart from mutations, high mutation burden in tumors also defines an actionable aberration, that can be potentially matched with checkpoint blockade immunother- apy. Enumeration of mutation burden in the MI_Oncoseq cohort iden- tified three cases with high tumor mutation burden, defined as > 10 mutations/Mb (Supplementary Table S5). These included MO_1347, a 46-year-old male with metastatic CCA (and a history of ampullary car- cinoma) previously treated with gemcitabine and cisplatin; a lymph node biopsy from this case, histologically seen as poorly differenti- ated high-grade adenocarcinoma admixed with prominent inflamma- tion, was found to harbor 225 mutations/MB and high microsatellite instability (MSI-high) score, consistent with a biallelic loss of function of the mismatch repair deficiency gene MSH2 (with truncating germline mutation MSH2 c.2494G > T; p.Glu832Ter; dbSNP: rs863225396), cou- pled with the somatic loss of heterozygosity through the splice accep- tor mutation, MSH2 c.1662-1G > A. No specific mutation or extrinsic etiology could be associated with the high mutation burden of 109 mutations/MB in the tumor from TP_2475, a 62-year-old female with stage IV metastatic CCA, “mixed ” subtype (CMS-HCC) previously treated with CDDP/gemcitabine. The third case with high mutation burden, TP_2703 with 25.3 mutations/MB displayed mutation signature 4 (as- sociated with tobacco smoking [18 , 19] ), consistent with the patient’s 30 pack-year history of smoking. The average mutation burden in the MI_Oncoseq cohort, after excluding 2 cases with low tumor content and one, MO_1347 with MSI high associated outlier mutation burden (Supplementary Table S5), was calculated as 4.3 mutations/Mb (range 0.45 to 108.9 mutations/Mb). This mutation burden in the cohort of ad- vanced, metastatic tumors was found to be significantly higher than that of the TCGA-BTC cohort comprised of primary tumors (Wilcoxon rank test p-value 0.05 ∗ , Fig. 2 A), consistent with similar observations across tumor types (for example, Robinson et al [32] ). No significant difference was noted in the mutation burden of tumor samples post-chemotherapy (N = 52), compared to advanced tumors prior to chemotherapy (N = 39). Gene fusion analysis using RNA-seq data identified known [12] and novel translocation events in 12 (9.7%) patients, including FGFR2, FGFR3 and YAP1 fused in frame with known and novel partners (Fig. S2). FGFR translocations were enriched in iCCA subtype (N = 9) with 4 C. Kumar-Sinha, P. Vats, N. Tran et al. Neoplasia 42 (2023) 100910 Fig. 2. Frequent genomic aberrations seen in biliary cancer. ( A ) Violin plots showing mutation burden across cholangiocarcinoma cohorts: TCGA-CHOL, MI- ONCOSEQ pre- and post-chemotherapy BTC cohort. The overall mutation burden is significantly higher in MI-ONCOSEQ advanced biliary tract cohort relative to the primary tumors in the TCGA cohort. ( B ) A summary of mutations in chromatin modifier genes in the BTC cohort. Individual samples represented in columns, genes in rows. The bar graphs on the right summarize the mutation frequency and number of samples mutated for each gene. Molecular alterations are color coded as described in Fig. 1 . ( C ) Copy number landscape of the BTC cohort (GISTIC2, Q < 0.05), significant losses or gains with interesting genes are represented in blue and red color, respectively. three cases of FGFR2-BICCI , and one each with FGFR2-KIAA1967, FGFR2-AFF4, FGFR2-AHCYL1 and FGFR2-CCDC6 fusion, and one each with novel partners including , FGFR2-TAX1BP1 and MATN4-FGFR2 . One gallbladder patient was identified to have FGFR3-TACC3 fusion, and one patient with mixed hepatocellular and CCA subtype had FGFR2- BICCI fusion (Fig. S1). All the FGFR rearrangements were found to retain the kinase domain and all the FGFR fusion partners potentially exhibited oligomerization capability, suggesting a shared mode of kinase activa- tion as noted previously [12] . In addition to FGFR2 gene fusions, two samples had hotspot ac- tivating mutations p.Y375C (also reported as p.Y276C) and p.C382R (also reported as p.C383R), and two cases had a novel in-frame indel p.H167_N173del located in the extracellular domain ( Figs. 2 B and S3). Importantly, a significant upregulation of FGFR2 gene expression (P = 0.029) was noted in the FGFR2 mutants (N = 4) as compared to the wild type cases ( Fig. 2 C), suggesting that the two patients with the novel in- dels represent a potentially activating aberration. Moreover, the median OS of patients in the fusion cohort (N = 12; 21.3 months) and mutation cohort was similar (N = 4; 21.5 months). Patient MO_1778 with perihilar CCA exhibited two known, recur- rent driver oncogenic fusions: FGFR2- BICC1 and YAP1-MAML2. While the FGFR2 fusion is a known driver in iCCA, recurrent YAP1-MAML2 fusion associated with aberrant Hippo pathway signaling has not been reported in CCA, but has been previously identified in other cancers [38–41] . The YAP1-MAML2 fusion encodes TEAD1, WW1 and WW2 domains from YAP1 and loss of Notch interaction domain in MAML2 , associated with transactivation of TEAD target genes leading to dedif- ferentiation or proliferation [39 , 42 , 43] . Potentially actionable aberrations and novel avenues for targeted therapies in advanced BTCs A total of 79 (63.7%) cases harbored one or more potentially ac- tionable (Tier 2) aberrations for which preliminary clinical/ preclinical rationale is available to match with experimental targeted therapeutics in ongoing clinical trials (Supplementary Table S9). Among the most common aberrations in this category were 35 patients with homozy- gous deletion or biallelic loss of Cyclin Dependent Kinase Inhibitor 2A, CDKN2A (p16INK4), associated with potential sensitivity to CDK4/6 in- hibitors.; 21 patients with activating mutations in KRAS/NRAS (and one case with deleterious mutation in NF1 ), that may be considered for treat- ment with novel KRAS and/or MEK inhibitors; and 23 patients with truncating mutations in ARID1A , a SWI-SNF pathway regulator, poten- tially associated with synthetic lethality to PARPi, ATRi or EZH2i. Ad- ditional cases with potentially actionable aberrations included tumors with amplification of MDM2 (with wild type TP53 ); CDK4 and CCND1 (with wild type RB1 ), NTRK1, MYC and CCNE1 ( Fig. 2 D), supported by outlier expression of these genes (data not shown). We identified 5 C. Kumar-Sinha, P. Vats, N. Tran et al. Neoplasia 42 (2023) 100910 recurrent in frame indels in FGFR2 that may represent gain of function mutations responsive to FGFR inhibitors as recently shown, and non- BRAF V600 mutations (class II and class III) [44] that may respond to MEK inhibitors. Overall, 105 out of 124 (84.7%) cases analyzed, were determined to harbor one or more actionable or potentially actionable aberrations that could be matched with FDA approved or experimental therapies in ongoing clinical trials. Germline alterations Pathogenic germline mutations were noted in 6 patients in the MI_Oncoseq cohort (6.5%) with majority in DNA damage repair path- way genes (2 cases of MUTYH , one each of BRCA1, BRCA2, ATM and MSH2) , and one case with germline mutation in FH, an essential gene in the tricarboxylic acid cycle ( Fig. 1 ; Supplementary Tables S3 and S4). Three patients with pathogenic germline mutations in MSH2, MUTYH and BRCA2 were found to harbor a second somatic aberration in the tumor resulting in biallelic loss. Targeted therapies and survival The median OS for the expanded cohort (MI-ONCOSEQ and other CLIA platforms) from date of diagnosis of advanced unresectable or metastatic disease was 15.2 months (range, 1.5–96.9), and from date of diagnosis was 19.2 months (range, 1.3–166). The median follow-up from date of release of genomic analysis report was 8.6 months (range, -1.4–61.3). We observed no significant imbalances in the baseline char- acteristics and treatment variables between the actionable matched and unmatched cohorts, including gender, age, ECOG performance status, FGFR status, exposure to platinum therapy, or number of lines of ther- apy, or distance from cancer center ( P > 0.05 by Fisher’s exact test or Wilcoxon test; data not shown). In the actionable cohort (N = 54; 43.5%), defined by the presence of a molecular aberration that can be matched with an FDA approved therapeutic, 22 (40.7%) subjects received a molecularly matched ther- apy (matched treated cohort) off-label or on clinical trials ( Table 2 ), while 32 (59.3%) patients did not receive molecularly matched ther- apy (matched untreated cohort). The remaining patients were defined as non-actionable for this analysis (70; 56.5%). In the matched treated group, patients received matched therapy after failure of systemic chemotherapy with the exception of one subject who received cobime- tinib and vemurafenib off-label as first-line therapy for BRAF V600E mutation ( Table 2 , Fig. 3 A). Patients with actionable mutations who re- ceived a matched therapy (N = 22) had significantly longer OS than the 70 patients in the non-actionable group, or 32 patients in the matched untreated group (28·1 months, 13.3 months and 13.9 months, respec- tively; P < 0.01). The median OS between the matched treated and un- treated arms of the actionable cohort had a hazard ratio of 0·33 (95% CI, 0·18-0·60, P < 0·01). However, median OS did not differ between the untreated actionable group and non-actionable group (HR 1.13, 95% CI, 0·80-2.00, P = 0·31) ( Fig. 3 B). A novel association between BRAF/ KRAS mutations and immune-modulator NT5E Apart from somatic mutations or copy number aberrations, we used RNAseq data (in MI_Oncoseq cohort) to help inform precision oncol- ogy avenues. This included sensitive detection of FGFR2 gene fusions in a partner-agnostic manner, and corroboration of outlier expression in cases with amplification of targetable genes such as ERBB2, CCND1, CCNE1 , and MDM2 . Additionally, querying individual driver aberrations for therapeutically informative gene expression correlates, we discov- ered a remarkable association between tumors with RAS/RAF muta- tions and expression of 5 ′ -Nucleotidase Ecto, NT5E (CD73), a membrane protein that converts extracellular nucleotides to membrane-permeable nucleosides, associated with promotion of tumor immunosuppression. Fig. 4 A shows a significantly higher level of NT5E in BTC cases in the MI_Oncoseq cohort with activating mutations in KRAS and BRAF , the lat- ter being significantly higher than KRAS . To assess this correlation in an external dataset, we accessed TCGA pan-cancer dataset from cBioportal, and compared NT5E expression in tumors with (1) BRAF V600E muta- tion, (2) KRAS G12/13 or Q61 substitutions, and (3) wild-type BRAF and KRAS . Tumors with other mutations in BRAF or KRAS , and cases with mutations in NRAS or HRAS , as well as cases with amplification or deletion of NT5E were excluded from this analysis to ensure rela- tively discreet comparisons. As seen in Fig. 4 B, tumors with activating KRAS/BRAF mutations showed significantly higher levels of NT5E ex- pression, with BRAF mutated tumors showing relatively higher expres- sion than KRAS mutated. In the context of BTCs, we corroborated the association between BRAF/KRAS mutations and NT5E expression level by IHC staining of select tumor tissue sections, as indicated ( Fig. 4 C-D). This association suggests follow up investigations for combination ther- apy with MEK and NT5E inhibitors in KRAS/BRAF mutant cases. Discussion Recent large-scale sequencing efforts like TCGA, ICGC and TARGET have provided insights into underlying molecular mechanisms in vari- ety of cancer types. In this study, we analyzed a cohort of 124 patients with advanced BTC and subjected data to integrative clinical sequenc- ing. Overall, a sizeable 43% of BTC patients harbored actionable muta- tions, of which, the 40.7% that received matched therapy had signifi- cantly longer OS by approximately 15 months compared to the cohort with actionable mutations that did not receive matched therapy. This suggests that patients with well-defined actionable molecular alterations derive considerable benefit in survival from receiving matched targeted therapy. Admittedly, the definition of actionability varies in literature but we used a common and perhaps stringent interpretation to include only FDA approved therapies for specific molecular alterations in any can- cer unless BTC-specific data suggested lack of benefit, such as palboci- clib monotherapy in cases with CDKN2A deletion [45] . Unfortunately, only 40.7% of the actionable cohort received matched targeted ther- apy. The most common reason was lack of available early phase clinical trial (n = 15), but other reasons included, the molecular analysis report preceded clinical trial investigation/availability (n = 9), inability to ob- tain off-label targeted therapy for those who did not meet trial eligi- bility (n = 5), decline in functional status or demise of the patient prior to release of the molecular analysis report (n = 2), or patient refusal to participate in clinical trial (n = 1). These outcomes suggest that precision oncology has a substantial clinical impact in patients with biliary cancer and warrants consideration of genomic analysis in all patients particu- larly earlier in their treatment course, and continued investigation of novel biomarkers and therapeutics in this rare cancer. We found the mutational burden in our cohort to be significantly higher compared to the TCGA cohort perhaps since majority of the pa- tients in our cohort had sequencing on tissue obtained at advanced dis- ease (89%), biopsies mostly included metastatic sites (55%), and pa- tients had had prior exposure to chemotherapy (56%). In comparison the TCGA cohort includes tissue obtained at primary resection. This find- ing supports the hypothesis that tumor mutational burden may worsen with prior exposure to chemotherapy and perhaps during the natural progression of the cancer. The overall tumor mutational burden is still low, however, compared to other cancers [46] , and only a small per- centage of tumors have a high enough mutational burden (3% with ≥ 10 mutations/Mb in our cohort) to leverage potential therapeutic benefit from immune checkpoint blockade [47 , 48] . Tumors with DDR gene mutations have been associated with sen- sitivity to DNA damaging chemotherapy, including platinum agents, as well as PARP inhibition [49–51] . Germline mutants of BRCA1 and 2 without defined locus specific loss of heterozygosity (LOH) in tu- mors have been associated with functional homologous recombination 6 Table 2 Patients in the study who underwent treatment with experimental agents. Case ID Gender Age (years) Anatomic subtype MO_1022 M 49 iCCA Stage at diagnosis Localized resectable Pre-biopsy treatment(s) for advanced disease Capecitabine/ oxaliplatin Actionable Alterations Potentially Actionable Alterrations Other Notable Alterations Post-biopsy Treatment(s) MO_1175 M MO_1203 F MO_1338 M 66 46 42 CUP (iCCA) Metastatic iCCA iCCA Localized resectable Metastatic Gemcitabine/carboplatin; FOLFOX/ bevacizumab Gemcitabine/ cisplatin; FOLFOX; IAP antagonist (CT) Gemcitabine/ cisplatin g BRCA2 E13fs FGFR2-CCAR2 fusion STK11 deletion Gemcitabine/capecitabine/bevacizumab (OL); Tivantinib/gemcitabine; Erlotinib/bevacizumab; MEK inhibitor (CT); cabozantinib (OL) MEK inhibitor (CT) SMAC mimetic (CT); EGFR antibody (CT) 5FU (CT); mTOR inhibitor (OL) MEK inhibitor (CT) KRAS G12D , TP53 R273C , GNAS R201H BAP1 Y173C , PBRM1 E991fs SMAD4 deletion, KEAP1 deletion, GNA11 deletion, TP53 P98fs, g FH K477dup- likely pathogenic/ VUS TP53 R174G and R158H, BAP1 V27fs, TET2 K534 ∗ TP53 R248Q Ivosidenib (CT) FOLFIRI; FOLFOX; Trastuzumab (OL) Gemcitabine/cisplatin; P emigatinib (CT) ; Anti-LAG3 antibody (CT) FOLFIRI; Pemigatinib (CT) FOLFIRI; CDK4/6 inhibitor (CT) BAP1 R252fs + LOH, CDKN2A deletion Hypermutation (TMB > 10), g MSH2 E832 ∗ + s MSH2 c.1662-1G > A splA IDH1 R132C EGFR Q787R , BRCA2 T3033fs , CDKN2A P105fs , ARID1A P1326fs ERBB2 amplification & outlier expression FGFR2-CCDC6 fusion FGFR2-MATN4 fusion MDM2 amplification + outlier expression PIK3CA E545G g ATM R2832C ARID1A p.Q177 ∗ + LOH MO_1347 M 45 iCCA Metastatic Gemcitabine/cisplatin MO_1369 F 61 iCCA Localized resectable MO_1388 M 44 iCCA Metastatic 7 Gemcitabine/ cisplatin; RFA; Capecitabine/RT; SBRT; Capecitabine/ oxaliplatin; Capecitabine/ gemcitabine Gemcitbaine/ cisplatin CUP (iCCA) Metastatic None MO_1518 F MO_1595 M MO_1613 M MO_1642 M MO_1723 F MO_1778 M MO_1780 M 63 50 70 53 50 27 71 iCCA iCCA iCCA Mixed CUP (iCCA) Metastatic Locally advanced Metastatic Locally advanced Metastatic Mixed Metastatic Gemcitabine/ cisplatin; FOLFOX Gemcitabine/ nab-paclitaxel (CT); Gemcitabine/ cisplatin; FOLFOX Gemcitabine/ cisplatin; Gemcitabine/ carboplatin Gemcitabine/ carboplatin; SIRT; FOLFOX; TACE Cisplatin/ etoposide; FOLFIRINOX FOLFOX/ bevacizumab; Capecitabine/ bevacizumab; Sorafenib Gemcitabine/ cisplatin MO_1794 F 51 iCCA Metastatic FGFR2-BICC1 fusion MO_1883 F MO_2057 F MO_2127 M 57 63 20 GB Metastatic Gemcitabine/ cisplatin CUP (iCCA) iCCA Locally advanced Metastatic Carboplatin/ etoposide; Capecitabine/ temozolomide Gemcitabine/ cisplatin/anti-PD-1 antibody (CT) IDH2 R172W FGFR2-AHCYL1 fusion FGFR2-BICC1 fusion Pemigatinib (CT) CDKN2A p.H83Y + LOH Pan FGFR inhibitor (CT) NRAS G12D Anti-PD-L1 antibody/TAK-659 (CT) BAP1 K421fs + c.375 + 1G > T splice donor BAP1 T203K, PBRM1 R690fs CDKN2A deletion , ARID1A Q479 ∗ ARID1A D2178fs + LOH TP53 R175H + LOH, SMAD4 G386D + LOH RASA1 p.L870fs + LOH CCND1 amplification + outlier expression, ARID1A Q335 ∗ + LOH TGFBR1 I109fs + LOH, APC P1594fs + LOH, CDKN2A deletion Anti-PD-1 antibody (OL) FOLFOX; Pemigatinib (CT) ; FOLFIRI; SIRT; Gemcitabine/paclitaxel; Sunitinib (CT) Gemcitabine/capecitabine; CTLA-4 antibody/PD-1 antibody (CT) Gemcitabine/oxaliplatin; Enasidenib (OL) FOLFOX ( continued on next page ) C . K u m a r - S i n h a , P . V a t s , N . T r a n e t a l . N e o p l a s i a 4 2 ( 2 0 2 3 ) 1 0 0 9 1 0 Table 2 ( continued ) Case ID Gender Age (years) Anatomic subtype Stage at diagnosis Pre-biopsy treatment(s) for advanced disease Actionable Alterations Potentially Actionable Alterrations Other Notable Alterations Post-biopsy Treatment(s) MO_2549 F 52 iCCA Metastatic Gemcitabine/ cisplatin TP_2475 F 62 Mixed Locally advanced None TP_2495 F 67 iCCA Metastatic None ERBB2 amplification + outlier expression Hypermutation (TMB > 10) MDM2 amplification + outlier expression ARID1A P153A , PTEN S129G BRAF D594G , ARID1A p.E1836fs , NRAS p.V14_Gdel in frame NRAS G13R TP_2541 M TP_2670 F TP_2682 M 59 65 43 iCCA iCCA iCCA Localized resectable Metastatic Localized resectable None None Gemcitabine/ cisplatin TSC1 R786 ∗ + LOH IDH1 R132G ARID1A P120fs RASA1 I859fs TP_2694 F 55 iCCA Metastatic None BT_7001 M BT_7019 M BT_7036 M 57 51 59 iCCA iCCA iCCA 8 Metastatic None IDH1 R132G Metastatic Metastatic Anti-CTLA-4 antibody/anti-PD-1 antibody (CT) Anti-CTLA-4 antibody/anti-PD-1 antibody (CT) BT_7047 M 61 iCCA Locally Advanced Gemcitabine/ cisplatin/ anti-PD-1 antibody (CT) BT_7201 F 54 iCCA Metastatic None BT_7304 F 66 GB Metastatic Capecitabine/ gemcitabine ARID1A Q67 ∗ + LOH CCNE1 amplification + moderate expression ARID1A p.R1528 ∗ + LOH ARID1A E2250fs + LOH , AKT2 amplification G360.01 M M FDX001 FDX002 F G360.02 FDX005 FDX006 FDX026 F F F M 75 54 60 57 55 52 64 iCCA iCCA iCCA iCCA iCCA iCCA iCCA Metastatic Metastatic None FOLFIRI plus nab-paclitaxel (CT) BRAF V600E FGFR2-SLMAP fusion PIK3CA H1047L Metastatic Gemcitabine/ cisplatin Metastatic Metastatic Metastatic Metastatic None Capecitabine None Anti-CTLA-4 antibody/anti-PD-1 antibody (CT) FGFR2-SHROOM3 fusion IDH1 R132C BRAF V600E VCL-FGFR2 fusion IDH1 R132C PTEN D268fs ∗ 30 , PIK3CA K593fs ∗ 8 CDKN2A/B deletion PARP inhibitor/PD-1 antibody (CT) BAP1 p.E31 ∗ p.S567 ∗ + LOH + LOH , RB1 gMUTYH G393D, TP53 N29fs + Splice Donor Gemcitabine/cisplatin; Anti-CTLA-4 antibody/anti-PD-1 antibody (CT); Capecitabine/oxaliplatin Gemcitabine/cisplatin; FOLFOX; Regorafenib (CT) ; FOLFIRI CDKN2A deletion + LOH, NF2 + LOH, CDKN2A MAX R33 ∗ K44 ∗ deletion FOXQ1 E147 ∗ KRAS G12V + E107K, BAP1 L100R + LOH NRAS G12D, CDKN2A deletion TP53 p.R175H + LOH, TP53 H178fs + LOH, SMAD4 p.E49 ∗ CDKN2A deletion, KRAS G12D, KDM6A deletion TP53 p.R306 ∗ CDKN2A p.H83Y + LOH + LOH, BAP1 D672G + LOH, SMAD4 splice acceptor + LOH TP53 G279E BAP1 splice site 255 + 2T > G, PBRM1 L848fs ∗ 16 CDKN2A A17_G23 > GR, TERT promoter -146C > T MYC A299V TP53 K132N Gemcitabine/cisplatin/anti-PD-1 antibody (CT) PARP inhibitor/anti-PD-1 antibody (CT) Capecitabine/gemcitabine; 5-FU/nal-irinotecan/anti-PD-1 antibody (CT) Gemcitabine/cisplatin; 5-FU/nal-irinotecan/anti-PD-1 antibody (CT); FOLFOX; Gemcitabine/nab-paclitaxel Gemcitabine/cisplatin/anti-PD-1 antibody (CT); Ivosidenib (CT) Gemcitabine/cisplatin/nab-paclitaxel; Capecitabine/gemcitabine Gemcitabine/cisplatin; FOLFOX SIRT; resection followed by Gemcitabine/cisplatin, SBRT Gemcitabine/cisplatin; PARP inhibitor/anti-PD-1 antibody (CT) ; Capecitabine/oxaliplatin; Gemcitabine/nab-paclitaxel 5-FU/nal-irinotecan/anti-PD-1 antibody (CT); Capecitabine/oxaliplatin Cobimetinib + vemurafenib (OL) Pemigatinib (CT) Pemigatinib (CT) Ivosidenib (CT) Cobimetinib + vemurafenib (OL) Pemigatinib (CT) Ivosidenib (CT) C . K u m a r - S i n h a , P . V a t s , N . T r a n e t a l . N e o p l a s i a 4 2 ( 2 0 2 3 ) 1 0 0 9 1 0 Bold denotes matched treatment to actionable or potentially actionable mutation; iCCA, intrahepatic cholangiocarcinoma; GB, gall bladder carcinoma; mixed, mixed hepatocellular/ cholangiocarcinoma; CUP, carcinoma of unknown primary; OL, off-label; CT, clinical trial C. Kumar-Sinha, P. Vats, N. Tran et al. Neoplasia 42 (2023) 100910 Fig. 3. Targeted therapies and survival. (A) Swimmers plot highlighting actionable alterations and progression-free survival for molecularly matched therapies (n = 22). Arrow denotes ongoing therapy. (B) Kaplan-Meier plot showing overall survival (OS) from diagnosis of advanced cancer in patients with actionable alterations treated with matched targeted therapy (blue; n = 22 in Fig. 3C), actionable alterations without matched therapy (red; n = 32), and patients with no actionable alterations (green; n = 70) with an overall P < 0.01. Fig. 4. Biliary cancers with KRAS / BRAF mutations show increased expression of immune-modulatory membrane protein CD73. ( A ) Box-plot representation of NT5E expression in MI-ONCOSEQ BTC cases with indicated status of KRAS and BRAF mutation. The y-axis shows NT5E expression levels (RPKM, reads per kb per million reads). ( B ) Violin-plot representation of NT5E expression in TCGA pan-cancer cohort with indicated status of KRAS and BRAF mutation. The y-axis shows NT5E expression levels (RPKM, reads per kb per million reads). ( C ) NT5E overexpression in BRAF/ KRAS mutated cholangiocarcinoma by immunohistochemistry (IHC), as indicated. Scale bar represents 50 uM (3B) and 10 uM (3C) respectively. 9 C. Kumar-Sinha, P. Vats, N. Tran et al. Neoplasia 42 (2023) 100910 deficiency as the wild-type allele may be inactivated via alternative mechanisms, such as promoter methylation. However, in absence of LOH inactivation they may lack sensitivity to DNA damaging agents [52] . Results from the phase 3 POLO trial showed significant im- provement in median PFS when patients with germline BRCA1 or 2 mutated metastatic pancreatic adenocarcinoma were treated with olaparib as maintenance therapy after platinum-based chemotherapy in patients compared to placebo [53] . It is worthwhile to hypothe- size a similar benefit in BRCA-mutated BTC treated with PARP in- hibitors, and indeed multiple clinical trials with PARP inhibitors alone (NCT04298021, NCT04042831), or in combination with anti-PD1 an- tibody (NCT03639935) are accruing patients with BTC. In addition to BRCA1 or 2 (incidence of 3–5%) [54] in BTC, other DDR mutated genes have also been identified including ATM and PALB2 (5 patients in our cohort; 4%) that may also benefit with PARP inhibitors [55 , 56] . Furthermore, patients with IDH1 or IDH2 hotspot mutations (10–15%) may also be susceptible to PARP inhibitors due to production of (R)-2- hydroxyglutarate (2HG), an oncometabolite that may impair homolo- gous recombination by inhibiting the function of histone demethylases [57] . Thus, there is a potential for significant benefit in up to 30% of patients with BTC with PARP inhibitors. A significantly high frequency (34.6%) of deleterious mutations in epigenetic modifiers, ARID1A, BAP1 and PBRM1 in the SWI/SNF acces- sory subunit highlights the role of dysregulated chromatin remodeling in BTC. ARID1A, BAP1 and PBRM1 encode subunits of the SWI/SNF chromatin-remodeling genes and were mutated in 20%, 15% and 10% of the samples, respectively; these have previously been shown to be drivers of progression in iCCA [15] . ARID1A and BAP1 have been shown to impair homologous repair in vitro [58 , 59] and therefore increase sus- ceptibility to PARP inhibitors; clinical trials to test this hypothesis are ongoing (e.g. NCT03207347). In addition, epigenetic inhibitors such as HDAC and EZH2 inhibitors [60] , proteolysis targeting chimera (PRO- TAC) degraders [61] , anti PD-1 antibodies [62] , and Aurora kinase A in- hibitors [63] may also hold promise in targeting these mutations in BTC. In addition to the mutations in the SWI/SNF complex, other epige- netic regulators such as IDH1 (and less commonly IDH2, FH ) hotspot mutations have been described in iCCA [64 , 65] . As noted above, the 2- HG oncometabolite is a byproduct of the IDH1 mutation and is known to dysregulate the function of the histone methylases [57] . Recently an IDH1 inhibitor, ivosidenib showed significant improvement in median PFS and OS compared to placebo in a phase 3 clinical trial [66] . Interest- ingly, 6 out of 23 (26%) patients with IDH1/2 mutations had concurrent mutations in either ARID1A, BAP1 or PBRM1 thus suggesting potential benefit from a combination of an IDH inhibitor and histone modifying agents such as HDAC or demethylating inhibitors in this subset, similar to AML [67 , 68 , 69] . ERBB2 amplification was identified in 4% of our cohort consis- tent with other studies with fluke-negative BTCs [70] . Data from in vitro experiments [71] , retrospective case series [72] , and prospective phase 1/2 trial [73] support the ongoing investigation of ERBB2 tar- geted therapies in clinical trials in BTC (NCT02693535, NCT03613168, NCT01953926, NCT04466891). We also identified amplifications in CCND1 [74] , MDM2 [75] and NTRK [76 , 77] and when targeted have shown modest preliminary clinical data noted in other cancers, and are under further investigation in integral biomarker trials. Other molecular alterations with less than 5% incidence in BTC that have shown promising activity include BRAF V600E mutation [20] . We identified 12 (9.7%) patients with BRAF mutation in our cohort of which 7 patients had non-V600E activating mutations, including class III (D594N, D594E, D594G), and undefined kinase domain mutations, K483E, M693V, G466E as well as N661K. Cells with BRAF class III mu- tations have been shown to be responsive to MEK inhibitors [44] . BRAF K483E is a recurrent mutation, shown to be transforming in culture [78] , and thus may represent a therapeutic target. Additionally, one case had a gene TRIM24-BRAF fusion, previously reported in a case of melanoma, sensitive to MEK inhibitor [79] . The MI-ONCOSEQ study first described the FGFR2 fusions across diverse cancers in 2013. Herein, we describe that FGFR2 activating mutations also lead to upregulation of gene expression similar to the fu- sions. Moreover, the median OS in the FGFR fusion cohort was similar to the FGFR activating mutation cohort (21.3 versus 21.5 months, respec- tively; data not shown) although the cohort sizes are small (N = 15 and 4, respectively). The median OS of patients in the FGFR cohort (fusions or activating mutations) was higher compared to the FGFR wild type (21.3 versus 14.0 months, respectively; p value 0.07; data not shown). Of the 19 patients in the FGFR fusion/activating mutation cohort, 8 patients were treated with pan FGFR inhibitors and had a median OS of 22.8 months compared to 17.3 months in the untreated arm (p value of 0.31; data not shown). These data suggest that FGFR fusions (and potentially activating mutations) are both prognostic and predictive biomarkers in this rare cancer. We also identified a FGFR3-TACC3 fusion in a patient with gallbladder cancer, and to our knowledge this is the first report of a FGFR3 fusion in gallbladder cancer. Multiple FGFR fusion partners have been previously identified of which BICC1 is the most commonly noted [16] . Herein, we describe additional novel fusion partners, specifically the FGFR2-TAX1BP1, and MATN4-FGFR2 . Clinical sequencing efforts like MI-ONCOSEQ which incorporate transcriptome analysis for gene fusions are important to identify targetable FGFR fusions due to the combinatorial possibilities of FGFR family fusion to a variety of oligomerization partners, as well as other rare fusions [80 , 81] . The discovery of novel association between BRAF/KRAS mutations and the expression of immunomodulatory target NT5E may define dual- precision therapeutic targets in a subset of cancers including the rela- tively intractable KRAS driven cancers. Notably, CD73 inhibitors are under intense clinical investigation for therapy across various can- cers, wherein some exciting results were noted in pancreatic cancer, a predominantly KRAS driven malignancy. In a Phase I ARC-8 trial (NCT04104672), treatment with small-molecule CD73 inhibitor AB680 in combination with gemcitabine and nab-paclitaxel and PD-1 inhibitor zimberelimab, in previously untreated patients with metastatic pancre- atic adenocarcinoma demonstrated effectiveness, with ORR 41%. Tu- mors reportedly shrank or stabilized in 11 of 13 patients who received the treatment for at least 16 weeks [82] , spurring dose-expansion and placebo-controlled phase II trials. We acknowledge the limitations of sample resources including neo- plastic cellularity which reduced the sample size in RNA-seq and im- mune cluster analysis. Our study also merged data from different se- quencing platforms (whole exome and targeted sequencing), thus lim- iting our analysis across the cohort to genomic regions common across the platforms. However, a uniform MI-ONCOSEQ analysis pipeline was used to ensure consistency and concordance across samples. These re- sults may not be applicable in the community for multiple reasons, in- cluding the use of a more inclusive genomic analysis platform such as MI-ONCOSEQ, lack of clinical trials at many non-academic sites, patient willingness to travel to an academic institution which may represent a more motivated sub-group (preserved performance status, younger age), and use of non-MI-ONCOSEQ genomic analysis reports in our expanded cohort includes a biased group of patients referred specifically for open clinical trials. Conclusion This study highlights the importance of integrative clinical sequenc- ing in defining molecularly matched targeted therapy options for bil- iary tract cancer, a rare yet anatomically and molecularly diverse malig- nancy with an aggressive clinical course, poor long-term prognosis due to limited therapeutic options. We observed significant improvement in survival when patients with actionable targets can receive matched therapies, and also enumerate several potentially actionable targets that provide a basis for matching with investigational drugs in ongoing clin- ical trials. Furthermore, we describe novel FGFR activating mutations 10 C. Kumar-Sinha, P. Vats, N. Tran et al. Neoplasia 42 (2023) 100910 and novel FGFR2 fusion partners which are likely to have direct impact on patient care, and diagnostic and therapeutic investigation. The novel association between KRAS/BRAF mutant tumors and the immunomod- ulatory target NT5E merits further investigation as a potential dual tar- geting modality in subsets of BTCs (as well as other cancers). These data provide evidence to strongly consider molecular analysis of tumors in patients with this rare cancer and the role of investigational therapies. Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. CRediT authorship contribution statement Chandan Kumar-Sinha: Conceptualization, Data curation, Formal analysis, Writing – review & editing, Writing – original draft. Pankaj Vats: Conceptualization, Data curation, Formal analysis, Writing – re- view & editing. Nguyen Tran: Conceptualization, Data curation, Formal analysis, Writing – review & editing, Writing – original draft. Dan R. Robinson: Data curation, Formal analysis, Writing – original draft. Valerie Gunchick: Data curation, Writing – original draft. Yi-Mi Wu: Data curation, Writing – original draft, Formal analysis, Writing – orig- inal draft. Xuhong Cao: Data curation, Writing – original draft. Yu Ning: Data curation, Writing – original draft. Rui Wang: Data curation, Writing – original draft. Erica Rabban: Data curation, Writing – orig- inal draft. Janice Bell: Data curation, Writing – original draft. Sunita Shankar: Data curation, Writing – original draft. Rahul Mannan: Data curation, Writing – original draft. Yuping Zhang: Data curation, Writ- ing – original draft. Mark M. Zalupski: Data curation, Writing – original draft. Arul M. Chinnaiyan: Data curation, Formal analysis, Writing – original draft. Vaibhav Sahai: Conceptualization, Data curation, For- mal analysis, Writing – review & editing, Writing – original draft. Supplementary materials Supplementary material associated with this article can be found, in the online version, at doi: 10.1016/j.neo.2023.100910 . References [1] N. Patel, B. Benipal, Incidence of cholangiocarcinoma in the USA from 2001 to 2015: a US cancer statistics analysis of 50 states, Cureus 11 (2019) e3962 . [2] K.J. Yao, S. Jabbour, N. Parekh, et al., Increasing mortality in the United States from cholangiocarcinoma: an analysis of the national center for health statistics database, BMC Gastroenterol. 16 (2016) 117 . [3] J.W. Valle, R.K. Kelley, B. Nervi, et al., Biliary tract cancer, Lancet 397 (2021) 428–444 . [4] V. Sahai, P.J. Catalano, M.M. Zalupski, et al., Nab-paclitaxel and gemcitabine as first-line treatment of advanced or metastatic cholangiocarcinoma: a phase 2 clinical trial, JAMA Oncol. (2018) . [5] J. Valle, H. Wasan, D. Palmer, et al., Cisplatin plus gemcitabine versus gemcitabine for biliary tract cancer, N. Engl. J. Med. 4 (2010) 395–397 . [6] H. Nathan, T.M. Pawlik, C.L. Wolfgang, et al., Trends in survival after surgery for cholangiocarcinoma: a 30-year population-based SEER database analysis, J. Gas- trointest. Surg. 11 (2007) 1488–1497 . [7] D. Sia, Y. Hoshida, A. Villanueva, et al., Integrative molecular analysis of intrahep- atic cholangiocarcinoma reveals 2 classes that have different outcomes, Gastroen- terology 144 (2013) 829–840 . [8] C.R. Churi, R. Shroff, Y. Wang, et al., Mutation profiling in cholangiocarcinoma: prognostic and therapeutic implications, PLoS One 9 (2014) e115383 . [9] M. Javle, T. Bekaii-Saab, A. Jain, et al., Biliary cancer: Utility of next-generation sequencing for clinical management, Cancer 122 (2016) 3838–3847 . [10] W. Chan-On, M.L. Nairismagi, C.K. Ong, et al., Exome sequencing identifies distinct mutational patterns in liver fluke-related and non-infection-related bile duct cancers, Nat. Genet. 45 (2013) 1474–1478 . [11] M. Simbolo, M. Fassan, A. Ruzzenente, et al., Multigene mutational profiling of cholangiocarcinomas identifies actionable molecular subgroups, Oncotarget 5 (2014) 2839–2852 . [12] Y.M. Wu, F. Su, S. Kalyana-Sundaram, et al., Identification of targetable FGFR gene fusions in diverse cancers, Cancer Discov. 3 (2013) 636–647 . [13] F. Farshidfar, S. Zheng, M.C. Gingras, et al., Integrative genomic analysis of cholan- giocarcinoma identifies distinct IDH-mutant molecular profiles, Cell Rep. 19 (2017) 2878–2880 . [14] R. Montal, D. Sia, C. Montironi, et al., Molecular classification and therapeutic tar- gets in extrahepatic cholangiocarcinoma, J. Hepatol. 73 (2020) 315–327 . [15] Y. Jiao, T.M. Pawlik, R.A. Anders, et al., Exome sequencing identifies frequent inacti- vating mutations in BAP1, ARID1A and PBRM1 in intrahepatic cholangiocarcinomas, Nat. Genet. 45 (2013) 1470–1473 . [16] G.K. Abou-Alfa, V. Sahai, A. Hollebecque, et al., Pemigatinib for previously treated, locally advanced or metastatic cholangiocarcinoma: a multicentre, open-label, phase 2 study, Lancet Oncol. 21 (2020) 671–684 . [17] M. Javle, M. Lowery, R.T. Shroff, et al., Phase II study of BGJ398 in patients with FGFR-altered advanced cholangiocarcinoma, J. Clin. Oncol. 36 (2018) 276–282 . [18] N.C. Turner, I.E. Krop, A. Bardia, et al., Abstract OT2-07-01: a phase 2 study of futibatinib (TAS-120) in metastatic breast cancers harboring fibroblast growth fac- tor receptor ( < em>FGFR < /em>) amplifications (FOENIX-MBC2), Cancer Res. 80 (2020) OT2-07-01-OT2-07-01 . [19] G.K. Abou-Alfa, T. Macarulla, M.M. Javle, et al., Ivosidenib in IDH1-mutant, chemotherapy-refractory cholangiocarcinoma (ClarIDHy): a multicentre, ran- domised, double-blind, placebo-controlled, phase 3 study, Lancet Oncol. 21 (2020) 796–807 . [20] Z.A. Wainberg, U.N. Lassen, E. Elez, et al., Efficacy and safety of dabrafenib (D) and trametinib (T) in patients (pts) with BRAF V600E–mutated biliary tract cancer (BTC): a cohort of the ROAR basket trial, J. Clin. Oncol. 37 (2019) 187-187 . [21] M. Javle, C. Churi, H.C. Kang, et al., HER2/neu-directed therapy for biliary tract cancer, J. Hematol. Oncol. 8 (2015) 58 . [22] X. Chen, X. Wu, H. Wu, et al., SHR-1210 plus GEMOX as first line treatment in biliary tract cancer: results from a single-arm exploratory study, J. Clin. Oncol. 37 (2019) 4092-4092 . [23] V. Sahai, K.A. Griffith, M.S. Beg, et al., A multicenter randomized phase II study of nivolumab in combination with gemcitabine/cisplatin or ipilimumab as first-line therapy for patients with advanced unresectable biliary tract cancer (BilT-01), J. Clin. Oncol. 38 (2020) 4582-4582 . [24] M. Ueno, M. Ikeda, C. Morizane, et al., Nivolumab alone or in combination with cisplatin plus gemcitabine in Japanese patients with unresectable or recurrent bil- iary tract cancer: a non-randomised, multicentre, open-label, phase 1 study, Lancet Gastroenterol. Hepatol. 4 (2019) 611–621 . [25] D.Y. Oh, K.H. Lee, D.W. Lee, et al., Phase II study assessing tolerability, efficacy, and biomarkers for durvalumab (D) ± tremelimumab (T) and gemcitabine/cisplatin (GemCis) in chemo-naïve advanced biliary tract cancer (aBTC), J. Clin. Oncol. 38 (2020) 4520-4520 . [26] Y.J. Bang, M. Ueno, D. Malka, et al., Pembrolizumab (pembro) for advanced bil- iary adenocarcinoma: results from the KEYNOTE-028 (KN028) and KEYNOTE-158 (KN158) basket studies, J. Clin. Oncol. 37 (2019) 4079-4079 . [27] C.S. Floudas, C. Xie, G. Brar, et al., Combined immune checkpoint inhibition (ICI) with tremelimumab and durvalumab in patients with advanced hepatocellular car- cinoma (HCC) or biliary tract carcinomas (BTC), J. Clin. Oncol. 37 (2019) 336-336 . [28] R.D. Kim, V. Chung, O.B. Alese, et al., A phase 2 multi-institutional study of nivolumab for patients with advanced refractory biliary tract cancer, JAMA Oncol. 6 (2020) 1–8 . [29] J. Kang, C. Yoo, J.H. Jeong, et al., Efficacy and safety of pembrolizumab in patients with PD-L1 positive advanced biliary tract cancer (BTC): a prospective cohort study, J. Clin. Oncol. 37 (2019) 4082-4082 . [30] O. Klein, D. Kee, A. Nagrial, et al., Combination immunotherapy with ipilimumab and nivolumab in patients with advanced biliary tract cancers, J. Clin. Oncol. 38 (2020) 4588-4588 . [31] D. Robinson, E.M. Van Allen, Y.M. Wu, et al., Integrative clinical genomics of ad- vanced prostate cancer, Cell 162 (2015) 454 . [32] D.R. Robinson, Y.M. Wu, R.J. Lonigro, et al., Integrative clinical genomics of metastatic cancer, Nature 548 (2017) 297–303 . [33] R.J. Mody, Y.M. Wu, R.J. Lonigro, et al., Integrative clinical sequencing in the man- agement of refractory or relapsed cancer in youth, JAMA 314 (2015) 913–925 . [34] S. Roychowdhury, M.K. Iyer, D.R. Robinson, et al., Personalized oncology through integrative high-throughput sequencing: a pilot study, Sci. Transl. Med. 3 (2011) 111ra121 . [35] Y.M. Wu, M. Cieslik, R.J. Lonigro, et al., Inactivation of CDK12 delineates a distinct immunogenic class of advanced prostate cancer, Cell 173 (2018) 1770–1782 e14 . [36] D. Nicorici, M. Ş atalan, H. Edgren, et al., < strong>FusionCatcher < /strong>– a tool for finding somatic fusion genes in paired-end RNA-sequencing data, bioRxiv (2014) 011650 . [37] Arriba. We used Arriba https://github.com/suhrig/arriba/ to detect gene fusions from RNA-Seq data. [38] A. Agaimy, R. Stoehr, L. Tögel, et al., YAP1-MAML2-rearranged poroid squamous cell carcinoma (squamoid porocarcinoma) presenting as a primary parotid gland tumor, Head Neck Pathol. (2020) . [39] M. Vivero, P. Davineni, V. Nardi, et al., Metaplastic thymoma: a distinctive thymic neoplasm characterized by YAP1-MAML2 gene fusions, Mod. Pathol. 33 (2020) 560–565 . [40] G. Picco, E.D. Chen, L.G. Alonso, et al., Functional linkage of gene fusions to cancer cell fitness assessed by pharmacological and CRISPR-Cas9 screening, Nat. Commun. 10 (2019) 2198 . [41] A. Valouev, Z. Weng, R.T. Sweeney, et al., Discovery of recurrent structural variants in nasopharyngeal carcinoma, Genome Res. 24 (2014) 300–309 . [42] G. Tonon, S. Modi, L. Wu, et al., t(11;19)(q21;p13) translocation in mucoepidermoid carcinoma creates a novel fusion product that disrupts a Notch signaling pathway, Nat. Genet. 33 (2003) 208–213 . 11 C. Kumar-Sinha, P. Vats, N. Tran et al. Neoplasia 42 (2023) 100910 [43] L. Wu, J. Liu, P. Gao, et al., Transforming activity of MECT1-MAML2 fusion onco- protein is mediated by constitutive CREB activation, EMBO J. 24 (2005) 2391–2402 . [44] Z. Yao, R. Yaeger, V.S. Rodrik-Outmezguine, et al., Tumours with class 3 BRAF mu- tants are sensitive to the inhibition of activated RAS, Nature (2017) advance online publication . [45] T.A. Baghdadi, S. Halabi, E. Garrett-Mayer, et al., Palbociclib in patients with pan- creatic and biliary cancer with CDKN2A alterations: results from the targeted agent and profiling utilization registry study, JCO Precis. Oncol. (2019) 1–8 . [46] Z.R. Chalmers, C.F. Connelly, D. Fabrizio, et al., Analysis of 100,000 human cancer genomes reveals the landscape of tumor mutational burden, Genome Med. 9 (2017) 34 . [47] M.A. Lowery, R. Ptashkin, E. Jordan, et al., Comprehensive molecular profiling of in- trahepatic and extrahepatic cholangiocarcinomas: potential targets for intervention, Clin. Cancer Res. (2018) . [48] G. Spizzo, A. Puccini, J. Xiu, et al., Molecular profile of BRCA-mutated biliary tract cancers, ESMO Open 5 (2020) e000682 . [49] M.W. Audeh, J. Carmichael, R.T. Penson, et al., Oral poly(ADP-ribose) polymerase inhibitor olaparib in patients with BRCA1 or BRCA2 mutations and recurrent ovarian cancer: a proof-of-concept trial, Lancet 376 (2010) 245–251 . [50] J. Ledermann, P. Harter, C. Gourley, et al., Olaparib maintenance therapy in plat- inum-sensitive relapsed ovarian cancer, N. Engl. J. Med. 366 (2012) 1382–1392 . [51] K.A. Gelmon, M. Tischkowitz, H. Mackay, et al., Olaparib in patients with recur- rent high-grade serous or poorly differentiated ovarian carcinoma or triple-negative breast cancer: a phase 2, multicentre, open-label, non-randomised study, Lancet On- col. 12 (2011) 852–861 . [52] K.N. Maxwell, B. Wubbenhorst, B.M. Wenz, et al., BRCA locus-specific loss of het- erozygosity in germline BRCA1 and BRCA2 carriers, Nat. Commun. 8 (2017) 319 . [53] T. Golan, P. Hammel, M. Reni, et al., Maintenance olaparib for germline BRCA-mu- tated metastatic pancreatic cancer, N. Engl. J. Med. 381 (2019) 317–327 . [54] G. Spizzo, A. Puccini, J. Xiu, et al., Frequency of BRCA mutation in biliary tract cancer and its correlation with tumor mutational burden (TMB) and microsatellite instability (MSI), J. Clin. Oncol. 37 (2019) 4085-4085 . [55] E.M. O’Reilly, J.W. Lee, M. Zalupski, et al., Randomized, multicenter, phase II trial of gemcitabine and cisplatin with or without veliparib in patients with pancreas adenocarcinoma and a germline BRCA/PALB2 mutation, J. Clin. Oncol. 38 (2020) 1378–1388 . [56] S.K. Radhakrishnan, D.G. Bebb, SP. Lees-Miller, Targeting ataxia-telangiectasia mu- tated deficient malignancies with poly ADP ribose polymerase inhibitors, Transl. Cancer Res. 2 (2013) 155–162 . [57] P.L. Sulkowski, C.D. Corso, N.D. Robinson, et al., 2-Hydroxyglutarate produced by neomorphic IDH mutations suppresses homologous recombination and induces PARP inhibitor sensitivity, Sci. Transl. Med. 9 (2017) . [58] H. Yu, H. Pak, I. Hammond-Martel, et al., Tumor suppressor and deubiquitinase BAP1 promotes DNA double-strand break repair, Proc. Natl. Acad. Sci. 111 (2014) 285–290 . [59] J.H. Park, E.J. Park, H.S. Lee, et al., Mammalian SWI/SNF complexes facilitate DNA double-strand break repair by promoting gamma-H2AX induction, EMBO J. 25 (2006) 3986–3997 . [60] B.G. Bitler, N. Fatkhutdinov, R. Zhang, Potential therapeutic targets in ARID1A-mutated cancers, Expert Opin. Ther. Targets 19 (2015) 1419–1422 . [61] W. Farnaby, M. Koegl, M.J. Roy, et al., BAF complex vulnerabilities in cancer demon- strated via structure-based PROTAC design, Nat. Chem. Biol. 15 (2019) 672–680 . [62] J. Shen, Z. Ju, W. Zhao, et al., ARID1A deficiency promotes mutability and poten- tiates therapeutic antitumor immunity unleashed by immune checkpoint blockade, Nat. Med. 24 (2018) 556–562 . [63] C. Wu, J. Lyu, E.J. Yang, et al., Targeting AURKA-CDC25C axis to induce synthetic lethality in ARID1A-deficient colorectal cancer cells, Nat. Commun. 9 (2018) 3212 . [64] W.G. Kaelin, S.L. McKnight, Influence of metabolism on epigenetics and disease, Cell 153 (2013) 56–69 . [65] P. Wang, Q. Dong, C. Zhang, et al., Mutations in isocitrate dehydrogenase 1 and 2 occur frequently in intrahepatic cholangiocarcinomas and share hypermethylation targets with glioblastomas, Oncogene 32 (2013) 3091–3100 . [66] G.K. Abou-Alfa, T. Macarulla Mercade, M. Javle, et al., ClarIDHy: a global, phase III, randomized, double-blind study of ivosidenib (IVO) vs placebo in patients with advanced cholangiocarcinoma (CC) with an isocitrate dehydrogenase 1 (IDH1) mu- tation, Ann. Oncol. 30 (2019) v872-v873 . [67] Dinardo C.D., Stein A.S., Stein E.M., et al. Mutant IDH1 inhibitor ivosidenib (IVO; AG-120) in combination with azacitidine (AZA) for newly diagnosed acute myeloid leukemia (ND AML). 2019;37:7011–7011. [68] G.H. Kim, S.Y. Choi, T.I. Oh, et al., IDH1(R132H) Causes Resistance to HDAC In- hibitors by Increasing NANOG in Glioblastoma Cells, Int. J. Mol. Sci. 20 (2019) 2679 . [69] C.D. DiNardo, A.S. Stein, A.T. Fathi, et al., Mutant isocitrate dehydrogenase (mIDH) inhibitors, enasidenib or ivosidenib, in combination with azacitidine (AZA): prelim- inary results of a phase 1b/2 study in patients with newly diagnosed acute myeloid leukemia (AML), Blood 130 (2017) 639-639 . [70] A. Jusakul, I. Cutcutache, C.H. Yong, et al., Whole-genome and epigenomic land- scapes of etiologically distinct subtypes of cholangiocarcinoma, Cancer Discovery 7 (2017) 1116–1135 . [71] A.R. Nam, J.W. Kim, Y. Cha, et al., Therapeutic implication of HER2 in advanced biliary tract cancer, Oncotarget 7 (2016) 58007–58021 . [72] M. Javle, C. Churi, H.C. Kang, et al., HER2/neu-directed therapy for biliary tract cancer, J. Hematol. Oncol. 8 (2015) 58-58 . [73] B.T. Li, T. Li, M.L. Johnson, et al., Safety and efficacy of pyrotinib in patients with NSCLC and other advanced solid tumors with activating HER2 alterations: A phase I basket trial, J. Clin. Oncol. 38 (2020) 3510-3510 . [74] M.J. Edelman, M.W. Redman, K.S. Albain, et al., SWOG S1400C (NCT02154490)-a phase II study of palbociclib for previously treated cell cycle gene alteration-positive patients with stage IV squamous cell lung cancer (lung-MAP substudy), J. Thorac. Oncol. 14 (2019) 1853–1859 . [75] F. Meric-Bernstam, N. Somaiah, S.G. DuBois, et al., 475P - A phase IIa clinical trial combining ALRN-6924 and palbociclib for the treatment of patients with tumours harboring wild-type p53 and MDM2 amplification or MDM2/CDK4 co-amplification, Ann. Oncol. 30 (2019) v179-v180 . [76] D.S. Hong, T.M. Bauer, J.J. Lee, et al., Larotrectinib in adult patients with solid tumours: a multi-centre, open-label, phase I dose-escalation study, Ann. Oncol. 30 (2019) 325–331 . [77] D. Hempel, T. Wieland, B. Solfrank, et al., Antitumor activity of larotrectinib in esophageal carcinoma with NTRK gene amplification, Oncologist 25 (2020) e881–e886 . [78] P.K. Ng, J. Li, K.J. Jeong, et al., Systematic functional annotation of somatic muta- tions in cancer, Cancer Cell 33 (2018) 450–462 e10 . [79] K.E. Hutchinson, D. Lipson, P.J. Stephens, et al., BRAF fusions define a distinct molecular subset of melanomas with potential sensitivity to MEK inhibition, Clin. Cancer Res. 19 (2013) 6696–6702 . [80] M.J. Borad, G.J. Gores, L.R. Roberts, Fibroblast growth factor receptor 2 fusions as a target for treating cholangiocarcinoma, Curr. Opin. Gastroenterol. 31 (2015) 264–268 . [81] Y. Arai, Y. Totoki, F. Hosoda, et al., Fibroblast growth factor receptor 2 tyrosine ki- nase fusions define a unique molecular subtype of cholangiocarcinoma, Hepatology 59 (2014) 1427–1434 . [82] Blocking CD73 can shrink pancreatic tumors, Cancer Discov. 11 (2021) OF4-OF4 . 12
10.1073_pnas.2304611120
RESEARCH ARTICLE | BIOCHEMISTRY OPEN ACCESS A biophysical framework for double- drugging kinases Chansik Kima,b,1 , Adelajda Hadzipasica,b,2, Steffen Kuttera,b,3, Vy Nguyena,b,4, and Dorothee Kerna,b,5 , Hannes Ludewiga,b Edited by Melanie Cobb, The University of Texas Southwestern Medical Center, Dallas, TX; received March 26, 2023; accepted July 6, 2023 Selective orthosteric inhibition of kinases has been challenging due to the conserved active site architecture of kinases and emergence of resistance mutants. Simultaneous inhibition of distant orthosteric and allosteric sites, which we refer to as “double- drugging”, has recently been shown to be effective in overcoming drug resistance. However, detailed biophysical characterization of the cooperative nature between orthosteric and allosteric modulators has not been undertaken. Here, we provide a quantitative framework for double- drugging of kinases employing isothermal titration calorimetry, Förster resonance energy transfer, coupled- enzyme assays, and X- ray crystallography. We discern positive and negative cooperativity for Aurora A kinase (AurA) and Abelson kinase (Abl) with different combinations of orthosteric and allosteric modulators. We find that a confor- mational equilibrium shift is the main principle governing cooperativity. Notably, for both kinases, we find a synergistic decrease of the required orthosteric and allosteric drug dosages when used in combination to inhibit kinase activities to clinically relevant inhibition levels. X- ray crystal structures of the double- drugged kinase complexes reveal the molecular principles underlying the cooperative nature of double- drugging AurA and Abl with orthosteric and allosteric inhibitors. Finally, we observe a fully closed confor- mation of Abl when bound to a pair of positively cooperative orthosteric and allosteric modulators, shedding light on the puzzling abnormality of previously solved closed Abl structures. Collectively, our data provide mechanistic and structural insights into rational design and evaluation of double- drugging strategies. kinase | conformational equilibrium | cooperativity | double- drugging Protein allostery is one of the fundamental regulatory mechanisms involved in various biological processes (1). Specifically, the allosteric regulation of protein kinases has been found essential for signaling cascades. Thus, dysregulation and overexpression of protein kinases are often related to many human diseases, including various cancers. However, due to the highly conserved catalytic site architecture of kinases, specific orthosteric inhi- bition is often unsuccessful, causing off- target effects (2). In addition, cancers often develop resistant mutations circumventing treatments with orthosteric drugs (3, 4). To overcome these problems, the field has been exploring allosteric sites of kinases for specific and efficacious inhibition (5, 6). A recently approved allosteric inhibitor of Abelson kinase (Abl), asciminib, has been highly effective in inhibiting Abl in vitro and in vivo (7–12). Remarkably, dual inhibition of Abl with this allosteric inhibitor combined with the orthosteric inhibitors (including imatinib, nilotinib, and ponatinib), which we refer to as “double- drugging”, has been impressively successful in abolishing the emergence of resistant mutants for Abl (12–14). Considering this clinical benefit, this approach has been applied to inhibit other targets such as EGFR kinase and SHP2 phosphatase (15, 16). However, the biophysical mechanisms underlying double- drugging of distant orthosteric and allosteric sites have not been well studied. Herein, we provide the quantitative framework for double- drugging using two targets: Aurora A kinase (AurA) and Abl. Both kinases participate in various cellular pathways, and their dysregulation results in a multitude of cancers, such as breast cancer and leukemia (17–19). Common obstacles faced by orthosteric inhibitors for AurA and Abl include cytotoxicity, off- target effects, and emergence of resistance mutants (3, 4, 20, 21). For both systems, we exploit a rational selection of ligands to probe positive and negative coopera- tivity between remote orthosteric and allosteric sites using isothermal titration calorimetry (ITC), Förster resonance energy transfer (FRET), and coupled- enzyme assays. We find that both orthosteric and allosteric ligands exhibit preferred binding to the active or inactive states and that cooperativity occurs by shifting this active–inactive conformational equi- librium through long- range allosteric networks that are encoded for natural regulation of those kinases. X- ray crystal structures of the double- drugged complexes shed light on the atomistic mechanisms of cooperativity. After we determine negative cooperativity for the double- drug combination used by Novartis in their clinical trials, we rationally chose a different orthosteric inhibitor, Src inhibitor 1 (SKI), for positive cooperativity with Significance While immensely successful, drugging kinases by active site inhibitors has faced major challenges. Selectivity issues leading to side effects and emergence of resistance mutations rendered treatments targeting active sites ineffective. Double- drugging via active and allosteric sites is a recently developed approach to overcome these obstacles. Using Aurora A and Abelson kinase, we provide a quantitative biophysical evaluation of double- drugging by rationally selecting inhibitor combinations with positive cooperativity. The results shed light on the interplay of kinase conformational equilibria and inhibitor- dose requirements for effective inhibition. Due to our rational selection of a positively cooperative drug combination for Abl, we deliver a fully closed, inactive Abl structure, including regulatory SH3 and SH2 domains. Collectively, this biophysical framework aids future rational double- drug designs. Preprint: This manuscript has been submitted to bioRxiv under a CC- BY 4.0 International license. This article is a PNAS Direct Submission. Copyright © 2023 the Author(s). Published by PNAS. This open access article is distributed under Creative Commons Attribution License 4.0 (CC BY). 1Present address: NoveltyNobility, Gyeonggi- do, Seongnam- si 13477, Republic of Korea. 2Present address: Novartis Institutes for Biomedical Research, Inc., Oncology Drug Discovery, Cambridge, MA 02139. 3Present address: Schrödinger, Inc., Natick, MA 01760. 4Present address: Relay Therapeutics, Cambridge, MA 02139. 5To whom correspondence may be addressed. Email: dkern@brandeis.edu. This article contains supporting information online at https://www.pnas.org/lookup/suppl/doi:10.1073/pnas. 2304611120/- /DCSupplemental. Published August 17, 2023. PNAS  2023  Vol. 120  No. 34  e2304611120 https://doi.org/10.1073/pnas.2304611120   1 of 11 asciminib. This double- drug combination forms a unique ternary complex, revealing a fully closed Abl structure. Results Cooperative Binding between Orthosteric and Allosteric Modulators of AurA. In solution, AurA exists in a conformational equilibrium between active and inactive states (22–24). We previously designed monobodies (Mbs) that are fully selective allosteric modulators, which bind to the natural allosteric regulatory site of AurA on the N- terminal lobe (N- lobe), the binding site for the natural coactivator protein TPX2 (25, 26). Different monobodies either act as activators or inhibitors depending on how they shift the active/inactive conformational equilibrium of AurA (25). To achieve double- drugging on AurA, we combined these Mbs with the orthosteric inhibitor danusertib (PHA739358) that tightly binds to AurA [IC50 = 13 nM, Ki = 0.87 ± 1.44 nM] (27). Since it has been shown that danusertib preferentially binds to the inactive conformation of AurA (22, 28), we hypothesized that inhibiting Mbs would bind tighter to AurA when in complex with the orthosteric inhibitor danusertib. Conversely, binding of activating Mbs to AurA should be weakened in the presence of danusertib (Fig. 1A). Aligning with our hypothesis, we find that the binding affinity of activating monobody (Mb1) to AurA weakens 16- fold when AurA is presaturated with danusertib (Fig. 1B). To test whether the binding of Mb1 and danusertib is mutually exclusive, we repeated this experiment by preincubating AurA with a higher concentration of danusertib (SI Appendix, Fig. S1E). Identical Mb1 binding affin- ities, independent of saturating danusertib concentrations, reveal that the simultaneous binding of Mb1 and danusertib to AurA is possible. Thus, we reason that this 16- fold negative cooperativity for Mb1 binding arises from a conformational equilibrium shift of AurA to the inactive state induced by danusertib. To achieve desired positive cooperativity between allosteric and orthosteric binders to AurA, we chose the inhibiting monobodies Mb2 and Mb3 because i) Mb2 is an inhibiting monobody for which we had obtained an X- ray crystal structure in complex with AurA, ii) Mb3 exhibits larger inhibition than Mb2, and iii) AurA- Mb3 complex exists in a monomeric form unlike the dimeric AurA- Mb2 complex (25). We indeed measure a twofold tighter binding of Mb2 to the AurA- danusertib complex compared to apo AurA (Fig. 1B). Using the equilibrium constant for active/inactive states of AurA previously determined [Keq= 0.67 (22)] and assuming identical affin- ities of Mb2 to the inactive states of apo- or danusertib- bound AurA, we fit our apparent affinities to a reversible two- state allosteric model. We find that the twofold positive cooperativity can be explained solely by the shift in the conformational equilibrium (SI Appendix, Fig. S2). Thus, a further increase in positive cooperativity would only be possible if the Mb affinity was tighter to the inactive state of the AurA- danusertib complex than to the inactive state of apo AurA. We indeed observed a threefold positive cooperativity for Mb3 with danusertib (Fig. 1B). We speculate that this increased affinity of Mb3 to the AurA- danusertib complex compared to apo AurA could result from favorable interactions with a closed activation loop, since danu- sertib binding shifts the equilibrium of the activation loop toward such conformation (23, 28). To confirm whether the mechanism of cooperativity between Mbs and danusertib follows a classic allosteric model, we tested binding of Mb6 to the AurA- danusertib complex. Despite high affinity, Mb6 binding does not change AurA’s activity, implying that Mb6 binding does not shift the active/inactive conformational equilibrium of AurA (25). Indeed, the binding affinity of Mb6 to AurA is not changed in the presence of danusertib (Fig. 1B and SI Appendix, Fig. S3). Fig.  1. Double- drugging of AurA kinase with orthosteric drug danusertib and different allosteric modulators. (A) Schematic representation of active/ inactive equilibrium of AurA [green, PDB- ID: 5G15, and orange, PDB- ID: 6C83 (25)]. Arrows indicate binding of danusertib and Mbs to their preferred AurA conformations. The table represents the rationale of positive and negative cooperativity for double- drugging of AurA. (B) Effect of preincubation with danusertib on observed dissociation constants (apparent Kd) of different monobodies measured by ITC. Activating monobody Mb1 shows 16- fold negative cooperativity, while inhibiting monobodies, Mb2 and Mb3, show twofold and threefold positive cooperativity, respectively. Mb6 binding is not affected by the presence of danusertib (SI Appendix, Fig. S3). (C) Reversal of preincubation order during affinity measurements shows identical cooperativities for orthosteric/allosteric ligand combinations (SI  Appendix, Fig. S1). Errors in (B and C) ITC data bar graph represent 68.3% CI (±1 SD) of the fit of the data. (D) Kinase inhibition curves of AurA and AurA in the presence of saturating concentrations of Mb1, Mb2 and Mb3 as a function of danusertib concentration. Enzyme assays were conducted (n = 2, mean ± SDM) under kcat/Km condition with 3 mM Lats2 peptide, measuring observed activity (kobs). With inhibiting Mb2 and Mb3, fourfold and 20- fold lower concentrations of danusertib, respectively, are required to inhibit to 10% residual AurA activity [(Danusertib)10% act., and dashed line]. Errors in this bar graph were determined by jackknifing the inhibition curve data. For a reversible two- state allosteric model, the same fold- change of cooperativity must be observed when reversing the order of binding. To measure changes in the affinity of danusertib upon Mb binding, we had to employ competitive replacement ITC with adenosine 5′- (α, β- methylene) diphosphate (AMPCP), since 2 of 11   https://doi.org/10.1073/pnas.2304611120 pnas.org danusertib binds too tightly to AurA for direct measurement. Indeed, the measured cooperativities are matching quantitatively regardless of the binding order (Fig. 1C). Double- Drugging of AurA Lowers Inhibitor Concentration Needed for Efficacious Inhibition. Next, we probed biological relevance of these observed cooperativities by measuring the inhibition of AurA kinase activity using Lats2 peptide as a substrate with cellular ATP concentrations. Preincubation of inhibiting Mbs resulted in a vastly decreased amount of danusertib required to cause 90% inhibition of AurA activity (Fig. 1D). This combined inhibition effect is the direct consequence of positive cooperativity. For instance, Mb3, which displays a larger degree of positive cooperativity than Mb2, causes a larger reduction in required amount of danusertib for effective inhibition. X- Ray Crystal Structures of Ternary Complexes: AurA- danusertib- Mb1 and AurA- danusertib- Mb2. We solved X- ray crystal structures of double- drugged AurA complexes to further understand the structural features responsible for the positive and negative cooperativity between Mbs and danusertib (SI Appendix, Table S1). The complex of AurA- danusertib- Mb1 [active, DFGin, BLAminus (29, 30)] displays hallmarks of an active kinase, such as an intact regulatory spine, the α- C helix in the “in” position, and the “DFG- in” state (Fig. 2A). However, we note that in contrast to the AurA- AMPPCP- Mb1 structure (PDB- ID: 5G15), D274 is rotated away from danusertib to avoid a steric clash with the terminal phenyl ring of danusertib (SI Appendix, Fig. S4A). This crystal structure corroborates the capability of danusertib to bind to the active conformation of AurA as we had tested biochemically (SI Appendix, Fig. S1B). The most interesting structure for “double- drugging” with maxi- mal inhibition is of course the ternary complex of AurA- danusertib- Mb2 [inactive, DFGinter (29, 30)]. Like AurA- AMPPCP- Mb2 (PDB- ID: 6C83), this ternary complex displays features of an inactive kinase: α- C helix “out”, “DFG- out,” as well as both a broken regu- latory spine and a broken canonical salt bridge (K162- E181) (Fig. 2B). This is expected due to the conformational equilibrium shift caused by Mb2 binding and the preferential binding of danus- ertib to inactive AurA. Furthermore, the activation loop is fully shifted Fig.  2. Proposed molecular mechanism for negative and positive cooperativity for double- drugging AurA with danusertib in combination with Mb1 and Mb2, respectively. (A and B) Zoom- in of X- ray crystal structures of AurA (gray) complexed with danusertib and either Mb1 (A, green) or Mb2 (B, gold). An intact regulatory spine, DFG- in conformation and extended activation loop in A is contrasted to a broken regulatory spine, DFG- out, and closed activation loop in (B). This closed activation loop provides additional hydrophobic interaction to the terminal ring of danusertib. (C–H) Orthosteric binding sites for six different AurA states reveal why danusertib has higher affinity for inactive AurA (22, 27, 31). K162 and E181, which form the canonical salt bridge in active AurA, and D274 and F275 (DFG- motif), are shown in stick representation. (D–E) While K162- E181 salt bridge is established in an active AurA conformations, (C) this salt bridge is broken in AurA- danusertib- Mb1 structure as K162 interacts with danusertib. (F–H) In the inactive AurA conformations, DFG- out F275 is positioned between K162 and E181, physically blocking the salt- bridge interaction, thereby prepositioning K162 for danusertib binding. (A–H) Oxygen, nitrogen, and phosphorous atoms are colored in red, blue, and orange, respectively. Carbon atoms are colored according to their respective protein cartoon. PNAS  2023  Vol. 120  No. 34  e2304611120 https://doi.org/10.1073/pnas.2304611120   3 of 11 toward the active site, providing additional hydrophobic interactions to the terminal ring of danusertib (Fig. 2B). This shifted activation loop is a major structural feature of an inactive AurA (23, 28), as observed in AurA bound to the orthosteric inhibitor MLN8054 (PDB- ID: 2WTV) (32). The structure of AurA- AMPPCP- Mb2 (PDB- ID: 6C83) displays a similar activation loop, however, with an extended portion being disordered (residues 276 to 290) to circum- vent clashing with the β- and γ- phosphate groups of AMPPCP (SI Appendix, Fig. S4B). The binary complex between AurA and danusertib (PDB- ID: 2J50) did not exhibit such a shift in the acti- vation loop. However, it is unclear whether the activation loop conformation in the AurA- danusertib structure reflects the solution state since the activation loop is directly involved in crystal contacts (SI Appendix, Fig. S5). Our crystal structures and ITC experiments showed that danu- sertib can bind to both the AurA- Mb2 and AurA- Mb1 complexes. To reveal why danusertib, however, binds with much higher affin- ity to AurA- Mb2 than AurA- Mb1 (there is no steric hindrance), we scrutinized the thermodynamic parameters of our ITC studies on danusertib binding to different AurA- Mb complexes (SI Appendix, Fig. S1 A–D). We find that the enthalpy for danu- sertib binding to AurA- Mb1 is reduced by 22.8 kJ/mol compared to AurA- Mb2, which approximates the equivalence of one salt bridge [12.6 to 20.9 kJ/mol (33)]. The canonical salt bridge between K162 and E181 is a feature of an active AurA, in both its apo form and bound to AMPPNP (PDB: 6CPE and 2DWB, respectively) (22) (Fig. 2 D and E). In contrast, the ternary com- plex of AurA- danusertib- Mb1 displays a broken salt bridge, as K162 interacts now with danusertib, while maintaining the α- C helix in the “in” position (Fig. 2C). Thus, we propose that the K162- E181 salt bridge in AurA- Mb1 must be broken for danus- ertib binding, as reflected by the lowered binding enthalpy. To confirm that apo AurA- Mb1 complex establishes the K162- E181 salt bridge, we deleted danusertib from the structure of AurA- danusertib- Mb1 and carried out molecular dynamics simulations in triplicate. We observed that K162- E181 indeed forms this salt bridge on average 80.8% in a 10 ns simulation (SI Appendix, Fig. S6A). However, in the presence of danusertib, we observe K162 to rather form a hydrogen bond with O- 27 of danusertib’s methoxy moiety than with E181 in the MD simulation, which is the state sampled in our crystal structure as well (Fig. 2C and SI Appendix, Fig. S6 B and C). In the inactive conformations of AurA, the broken K162- E181 salt bridge stems from the α- C helix and DFG- motif being posi- tioned in the “out” conformation such that F275 positions between K162 and E181 (PDB: 4C3R and 2J50) (27, 31) (Fig. 2 G and H). Thus, we propose that K162 in inactive AurA conforma- tions, such as AurA- Mb2 complex, is prepositioned for danusertib binding (Fig. 2F), which results in the tighter binding of danus- ertib to the inactive state of AurA. Cooperative Effect of Imatinib and Asciminib Binding on Abl. Intrigued by our mechanistic insights into double- drugging of AurA, we turned to Abl, the only target currently in clinical trials for double- drugging. It has been shown that the combination of the orthosteric inhibitor imatinib and the allosteric inhibitor asciminib abolishes the emergence of resistance mutations (7–12), an impressive breakthrough. Therefore, Abl embodies a powerful target to delineate the biophysical constraints, or “framework”, for successful double- drugging. Since the quantitative biophysical parameters for this drug combination are not known, we set out to biophysically investigate the cooperativity and modulation of Abl’s open/closed conformational equilibrium first using this exact combination of orthosteric and allosteric inhibitors. Note that we use the well- established relevant construct of SH3- SH2- KD Abl (Abl64– 510) (SI Appendix, Fig. S7). Abl exists in a conformational equilibrium between open, active, and closed, inactive conformations (34–36) (Fig.  3A). In the open conformation, the regulatory domains are elongated so that the SH2 domain moves onto the N- lobe of the kinase domain, forming a “top- hat” conformation (35). In the closed conformation, the regulatory domains tightly interact with the kinase domain, SH3:N- lobe and SH2:C- lobe, the latter facilitated by the bent C- terminal α- I helix (12, 35, 37) (Fig. 3A). This conformational equilibrium is susceptible to modulation by single agents such as imatinib and asciminib (12, 34, 38). In ITC experiments, we find that imatinib binds fivefold tighter to AblKD, which exists exclusively in the open conformation, than to Abl64– 510 (Fig. 3B). This confirms imatinib’s preferential binding to the open state of Abl (34). In full agreement with this model, imatinib binds to Abl64– 510 with a fourfold decreased affin- ity in the presence of asciminib, since asciminib shifts the equi- librium to the closed state (12) (Fig. 3B). We conclude that this fourfold negative cooperativity between imatinib and asciminib stems from a shift in the conformational equilibrium of Abl, where both drugs preferentially bind to the open and closed conforma- tion, respectively. Akin to AurA, preincubation of Abl64– 510 with increased concentration of asciminib did not result in a weakened imatinib affinity, confirming the simultaneous binding of the two inhibitors (SI Appendix, Fig. S8B). Surprisingly, we found that imatinib and asciminib display a twofold negative cooperativity for AblKD (Fig. 3B). This implies the presence of an additional conformational equilibrium within the kinase domain itself (39) and that asciminib and imatinib shift this equilibrium in opposite directions. We refer herein to the asciminib- favoring conformation as the “closing- competent” conformation of AblKD. Importantly, we measure identical negative cooperativities between imatinib and asciminib on both AblKD and Abl64– 510, regardless of binding order, within the range of errors (Fig. 3C and SI Appendix, Fig. S8E). Due to the tight binding of asciminib, its affinity was measured via competitive replacement ITC using N- Myr peptide as a weak- binding ligand (12, 40). Collectively, we conclude that the binding of imatinib to the orthosteric site and asciminib to the allosteric site in Abl64– 510 follow a two- state allosteric model, in which the two drugs favor the closed and open conformation, respectively. Positive Cooperativity between SKI and Asciminib on Abl. Considering the negative cooperativity between imatinib and asciminib described by our ITC experiments, we wanted to rationally select an orthosteric inhibitor that exhibits positive cooperativity with asciminib. We chose Src inhibitor 1 (SKI), an orthosteric inhibitor that tightly binds to Src kinase (IC50 = 44 nM) (41, 42), because Bannister et al. recently measured that SKI preferentially binds to the α- C helix out, and thus closed- inactive conformation of Src kinase, despite the DFG- motif being in the “in” position (SKI was therefore traditionally classified as type I inhibitor) (Unpublished data, Bannister et  al.). Due to Abl and Src kinases’ close structural homology, we hypothesized that SKI would bind to Abl in a similar fashion, thus exhibiting positive cooperativity with asciminib by preferentially binding to the closed state of Abl. Since SKI binding to Abl did not result in a detectable heat change in ITC, we turned to FRET experiments to quantify this interaction (Fig.  3D and SI  Appendix, Fig.  S9 and Fig.  S10). SKI indeed binds preferentially to the closed conformation of Abl64– 510, as seen by the fivefold tighter binding of SKI to Abl64– 510 than to AblKD. Furthermore, we observe a modest positive cooperativity between SKI and asciminib binding in AblKD, indicating that SKI binds to the “closing- competent” 4 of 11   https://doi.org/10.1073/pnas.2304611120 pnas.org Fig. 3. Double- drugging of Abl kinase. (A) Schematic representation of conformational equilibrium in Abl kinase. Arrows indicate binding of orthosteric inhibitors imatinib and SKI, and allosteric inhibitor asciminib to preferred Abl conformations. X- ray crystal structures from PDB- ID: 1OPL (green) and PDB- ID: 5MO4 (red) were used for open and closed Abl structures, respectively (12, 35). Note that the SH3 domain in the open structure is missing due to lacking electron density. (B) With ITC experiments, we observe twofold and fourfold negative cooperativity for open- conformation binder imatinib when AblKD and Abl64– 510, respectively, are preincubated with closed- conformation binder asciminib (SI Appendix, Fig. S8). (C) Matching fold- change of negative cooperativity is observed reversing the order of modulators used in preincubation and titration, using ITC (SI Appendix, Fig. S8). (B and C) Errors in the ITC data bar graphs represent 68.3% CI (±1 SD) of the fit of the data. (D) FRET experiments to detect SKI binding (10 nM of enzyme in all experiments. Data (n = 2 to 5, mean ± SDM) have been fitted to quadratic binding equation. Unlike imatinib, SKI binds tighter to AblKD- asciminib and Abl64– 510 than to AblKD. In AblKD, asciminib exhibits small positive cooperativity with the binding of SKI. Kd errors are SE of the fit. conformation induced by asciminib. Unexpectedly, we did not find a difference between the binding affinities of SKI to Abl64– 510 and Abl64– 510- asciminib. We interpret this result as evidence that the conformational equilibrium of apo Abl is already far shifted to the closed conformation. Hence, the binding of asciminib had no effect on this equilibrium. This is, in fact, in agreement with a NMR study by Grzesiek and colleagues reporting overlapping chemical shifts between apo and GNF- 5 (a predecessor of asciminib) bound Abl for open/closed equilibrium markers (34). Effect of Orthosteric and Allosteric Modulators on Abl Activity. Interestingly, it had been reported that allosteric inhibitors of Abl other than asciminib (such as GNF- 2, GNF- 5, myristate, and myristoyl- peptide) actually do not inhibit the catalytic activity despite binding to AblKD (40, 43, 44). However, with ITC, we observed that asciminib shifts the conformation of AblKD to the “closing- competent” conformation (Fig.  3 B and C). Is this “closing- competent” conformation of the kinase domain a catalytically inactive state of Abl? Inhibition curves of AblKD generated using a coupled- enzyme assay with Srctide as substrate and asciminib as an inhibitor reveal 30% inhibition at saturating asciminib concentration (Fig.  4A). We conclude that the closing- competent conformation of the kinase domain is indeed catalytically inactive and that asciminib shifts the conformational equilibrium of AblKD to be 30% in this conformation by binding to the C- lobe and allosteric propagation to the orthosteric site. This model also reconciles the moderate synergistic effect of SKI and asciminib binding to AblKD (Fig. 3D). We note that this unique allosteric propagation by asciminib could contribute to its increased potency relative to other myristate pocket binders. Most importantly, and stressing the importance of studying full- length kinases in drug development, asciminib causes a 93% inhibition of Abl64– 510 at saturating concentration (Fig.  4A). This vastly increased inhibition is caused by the closing of the regulatory domains leading to an inactive kinase. Next, we quantified the inhibition of AblKD and Abl64– 510 by the two orthosteric inhibitors imatinib and SKI. In agreement with our affinity measurements (Fig. 3 B–D), imatinib exhibited a lower IC50 for AblKD than for Abl64– 510, while SKI exhibited a higher IC50 for AblKD than for Abl64– 510. Second, preincubation of AblKD with asciminib increased the IC50 for imatinib. This negative cooperativity arises from binding preferences of imatinib and asciminib to opposite conformations. In contrast, double- drugging of AblKD with SKI and asciminib resulted in a reduced IC50 since both have a binding preference to the same, “closing competent” conformation, highlighting their positive cooperativity (Fig. 4A). We note that SKI’s IC50 is higher than imatinib’s IC50 with respect to AblKD, Abl64– 510, and AblKD- asciminib (Fig. 4A), whereas this trend is reversed in our binding experiments (Fig. 3 B–D). We ascribe this discrepancy to the presence of ATP in the coupled- enzyme assay: AMPPCP binds twofold tighter to AblKD than Abl64– 510 (SI Appendix, Fig. S11). Thus, under our assay condition, we reason that the ATP shifts the conformational equilibrium of PNAS  2023  Vol. 120  No. 34  e2304611120 https://doi.org/10.1073/pnas.2304611120   5 of 11 Fig. 4. Catalytic activities of Abl kinase under double- drugging conditions. (A) Inhibition curves of AblKD and Abl64– 510 with asciminib, imatinib, and SKI. IC50 values shift according to the favored binding conformations of the corresponding inhibitor. (B and C) In vitro synergy studies using (B) imatinib and asciminib (C) or SKI to examine cooperativity between both inhibitors for Abl64– 510 activity. (C, Left) Numbers in the grid represent kobs with respective concentrations of inhibitor combinations. (C, Middle and Right) graphic representation of the data to illustrate inhibitor concentration needed to achieve 10% residual kinase activity [(Imatinib)10% act. and (SKI)10% act, dashed line]. Note that SKI required for 10% residual kinase activity decreases with increasing asciminib, especially with higher fold- change than that observed for imatinib. All assays were measured (n = 4 for 0 nM orthosteric inhibitor, n = 2 for all other assays, mean ± SDM) under kcat/Km condition with 2 mM Srctide. Errors in IC50 are SE of the fit. Errors in the bar graphs were determined by jackknifing the inhibition curve data. Abl64– 510 to the open state, which is favored by imatinib over SKI binding. Inhibition of Abl Kinase Activity under Double- Drugging Condition. The key question for clinical application is: What is the effect of different dosing concentration combinations of the two inhibitors on Abl’s kinase activity? Therefore, we performed synergy studies on Abl64– 510 kinase activity varying the concentration of both orthosteric and allosteric inhibitors (Fig. 4 B and C). These experiments underscore the negative cooperativity between imatinib and asciminib and corroborate the positive cooperativity between SKI and asciminib. First, we find a more pronounced inhibition of Abl64– 510 by SKI than by imatinib in the presence of asciminib. On the other hand, when used as a single agent, imatinib inhibits Abl64– 510 stronger than SKI, highlighting the difference in cooperativity. Second, we observe that in the presence of asciminib, less SKI is required for 90% inhibition of Abl activity compared to imatinib due to the positive cooperativity between SKI and asciminib (Fig.  4 B and C). X- Ray Crystal Structure of the Ternary Complex of Abl64– 510- SKI- Asciminib. Intrigued by the synergistic effect of SKI and asciminib on Abl activity, we structurally characterized this ternary complex by cocrystallization, resulting in a 2.86 Å crystal structure of Abl64– 510- SKI- asciminib [inactive, DFGin, BLBplus (29, 30)] (Fig. 5A and SI Appendix, Fig. S12 and Table S1). Surprisingly, this Abl structure adopts a closed conformation with striking differences to previously reported closed structures; Abl in complex with nilotinib and asciminib (PDB- ID: 5MO4), as well as in complex with PD166326 and myristic acid, a groundbreaking structure of full- length Abl in the inhibited state (PDB- ID: 1OPK) (Fig. 5B and SI Appendix, Fig. S13) (12, 35). First, we note that the entire N- terminal lobe is ~30° twisted only for Abl64– 510- SKI- asciminib, when aligned by the 6 of 11   https://doi.org/10.1073/pnas.2304611120 pnas.org regulatory domains (Fig. 5B and SI Appendix, Fig. S13). Second, the α- C helix is adopting the “out” position resulting from this N- lobe twist, since an α- C helix “in position” would clash with strands β4 and β5 (Fig. 5 B and C). In consequence, the canonical salt bridge between K290 and E305, a hallmark of an active kinase, is broken in our structure, whereas D400 (DFG- motif) is positioned in the “in” position. Paradoxically, the two other closed ternary complexes of Abl possess an α- C helix located in the “in” position and an established canonical salt bridge (K290- E305), both reminiscent of an active kinase conformation (SI Appendix, Fig. S15), while their DFG- motif is in the “out” position. This highlights i) the importance of the α- C helix conformation and the canonical salt bridge, and not only the DFG- motif, in determining open/closed conformation which is directly correlated to active/inactive states in full- length kinases, and ii) that through binding of SKI and asciminib, we were able to capture the strictly closed and inactive conformation of Abl with regulatory domains. We note, that the orthosteric site is fully occupied by SKI and the twisted N- lobe aids in forming this tightly packed binding pocket (Fig. 5C and SI Appendix, Fig. S13). In fact, K290 located on β3 strand is wrapping over SKI burying the inhibitor in Abl’s orthosteric site. Besides extensive van der Waals interactions between SKI and Abl, the quinazoline ring of SKI shares two hydrogen bonds with Abl, one between the side chain hydroxyl of T334 on β- strand 5 and N- 2 of SKI as well as between the amide of M337 and N- 0 of SKI (Fig. 5C). When compared to other closed Abl structures, we find an extended domain interface between SH3, linker, and N- lobe of the kinase domain, which explains the positive cooperativity Fig. 5. X- ray structure of ternary Abl64– 510- SKI- asciminib complex reveals a fully closed conformation compared to previous “energetically frustrated” ternary closed Abl structures. (A) Abl64– 510 bound to SKI and asciminib. (B) Superposition of Abl64– 510- SKI- asciminib (blue) and Abl- nilotinib- asciminib (pink, PDB- ID: 5MO4) (12). When superimposed by the regulatory SH2 and SH3 domains, the N- lobe of Abl64– 510- SKI- asciminib twists and exhibits α- C helix “out” position. (C) Zoom into the SKI and nilotinib binding sites. Van der Waals radii for the interacting Abl residues (spheres) with SKI (orange) and nilotinib (green) show more confined binding pocket for SKI than nilotinib. (D) Comparison of interface residues between N- SH3 domain (S94, D96, T98), linker (V247, S248), and N- lobe of kinase domain (W280, K282, Y283, S284, and L285) for Abl64– 510- SKI- asciminib (blue), Abl- nilotinib- asciminib (pink, PDB- ID: 5MO4), and Abl- PD166326- myristate (yellow, PDB- ID: 1OPK) (12, 35). Due to the twist in the N- lobe for Abl64– 510- SKI- asciminib, residues in the domain/domain interface exhibit better packing. For Abl64– 510- SKI- asciminib, an additional hydrogen bond is established between S248 and D96 which contributes to this extended interface. Oxygen and nitrogen atoms are colored in red and blue, respectively. Carbon atoms are colored according to their respective protein cartoon. PNAS  2023  Vol. 120  No. 34  e2304611120 https://doi.org/10.1073/pnas.2304611120   7 of 11 between SKI and asciminib. This improved interface is a direct result of the twisted N- lobe. The repositioned β2 and β3 strands cause Y283 to be completely buried within this interface. In other ternary complexes of Abl, this interface is only partially formed (Fig. 5D). Moreover, S248 (located on the linker) forms a hydro- gen bond with D96 in the SH3 domain, which is only present in the ternary complex of Abl64– 510- SKI- asciminib. Strikingly, S248P was identified as a resistant mutation for GNF- 2 and asciminib in cell culture- based screening (7, 45) with no mechanistic under- standing, given that this mutation is far away from the allosteric inhibitor binding site. Our structure now reveals the important role of S248 in allosteric closing of Abl, and hence asciminib inhibition! We conclude that our complex of Abl64– 510- SKI- asciminib represents the only example of a fully closed and inactive Abl structure. Discussion To combat on- target cancer drug resistance, double- drugging holds promise to be a powerful strategy. The rationale behind is multiplication of individual resistance mutational probabilities for each drug. Impressively, combinations of asciminib and various orthosteric inhibitors, including imatinib, indeed abolish the emergence of Abl resistance mutants and this double- drugging of Abl is currently in clinical trials (12). Given these groundbreaking clinical results, we used Abl kinase to interrogate the biophysical mechanism underlying this drug combination to learn a quanti- tative biophysical framework for successful double- drugging. In a second step, we used our knowledge of conformational equilibria in kinases to rationally select alternative orthosteric drugs exhib- iting improved synergy with the allosteric drug. Our results have major implications: i) Knowledge of conformational equilibria in drug targets indeed enables rational selection of inhibitor combi- nations with positive cooperativity and therefore better synergy. ii) Our Abl structure solves the apparent mystery of all previous closed Abl structures with the α- C helix in the active “in position” that contradicted the common features of inactive- closed kinases with the α- C helix in the canonical “out position”. Structural investigation of our double- drugged ternary complex with SKI and asciminib reveals a true α- C helix “out” state observed in Abl structures (SI Appendix, Fig. S14). This originates from an SKI- induced twist in the N- lobe causing a fully closed conformation, thus, releasing an energetically frustrated conformation observed in other double- drugged Abl complexes, since SKI and asciminib both preferentially bind to the closed state of Abl to cause positive cooperativity. In contrast, double- drugging with an open conforma- tion binder (nilotinib) and a closed conformation binder ( asciminib) results in an energetically frustrated Abl structure (PDB- ID: 5MO4). This structural study highlights how understanding of conforma- tional equilibria crucially aids the discovery of further inhibited states. Our finding of the energetically frustrated conformation agrees with previous NMR experiments reporting an opposing bind- ing preference of imatinib and GNF- 5 for Abl (34). In contrast, Johnson et al. claimed that such an antagonism arises from mutually exclusive binding of orthosteric and allosteric inhibitors (38). This conclusion contradicts previous studies characterizing Abl- imatinib- GNF- 5 by NMR as well as crystallographic studies on the ternary complex of Abl bound to both nilotinib and asciminib (12, 34). Our ITC studies resolve this controversy by ruling out mutual exclusivity for binding of imatinib and asciminib. iii) Our Abl data solve a heated debate: Recently, Kalodimos and colleagues argued that imatinib opens Abl via binding to its allosteric site (reported Kd >10 µM), and not via binding to its active site (46). This is in disagreement with NMR and cellular studies by Grzesiek et al. (34, 47, 48). Tighter binding of imatinib to open AblKD (Kd = 15 nM) compared to closed Abl64– 510 (Kd = 72.4 nM) and negative cooperativity between imatinib and asciminib buttress Grzesiek’s model where imatinib’s preferential binding to the open conformation of Abl arises from its orthosteric site binding with nanomolar affinity. Double- drugging has been applied to two additional targets, SHP2 phosphatase (16) and EGFR kinase (15, 49). Fodor et al. used a combination of two allosteric binders, SHP099 and SHP504, to inhibit the phosphatase SHP2 (16).The authors demonstrate that the combination reduces the dosage require- ments of these allosteric inhibitors to achieve effective inhibition of SHP2; however, SHP504 is a very weak binder with an IC50 of 21 μM (16). For EGFR kinase, a combination of the inhibitor JBJ- 04- 125- 02 binding right next to the irreversible orthosteric inhibitor osimertinib has been found to be more efficacious, than single agents, for inhibiting tumor growth in a mouse model. Furthermore, Jänne and colleagues demonstrated that this double- drugging resulted in the reduced emergence of resistance mutants in cellular assays (49). Here, the allosteric inhibitor bind- ing site is in immediate proximity to the orthosteric site, resulting in direct interactions between the two inhibitors potentially driv- ing positive cooperativity (15, 49). In contrast, we investigated the mechanism of dual inhibition in AurA and Abl kinase targeting a distant allosteric site that is involved in natural regulation, in combination with active site drugs. Rationally targeting those natural allosteric sites has the advantage that it assures allosteric coupling to activity. We demon- strate with our amateur attempts on both kinases that rational selection of double- drug combinations with positive cooperativity, and hence increased synergy, is possible based on knowledge of involved conformational equilibria. Furthermore, we note that such kinase activity- based synergy studies could easily be per- formed in a high- throughput manner to test orthosteric and allosteric inhibitor combinations. In summary, this work proposes a biophysical framework for designing and evaluating double- drugging synergy utilizing ortho- steric and allosteric modulators. As highlighted here, positive coop- erativity is desirable for double- drugging approaches improving selectivity and dosage requirements. However, while extreme negative cooperativity is undesirable, the clinical success of Novartis’ drug combination for Abl (12) with fourfold negative cooperativity as measured here suggests a clinical efficacy window ranging from small negative to strong positive cooperativity, given single- drug efficacy. Single drug efficacy is crucial, as otherwise a single resistance muta- tion abolishing binding of one drug would render the dual treatment to combat drug resistance essentially ineffective. Methods Cloning and Purification of Aurora A and Monobodies. AurA (residues 122 to 403, TEV- cleavable, N- terminal His6- tagged, kanamycin- resistance) in pET28a and LPP (#79748) from Addgene were cotransformed in BL21(DE3) cells and plated on Kan/Spec LB plate. Expression cultures were grown in TB to OD = 0.6–0.8 and induced with 0.6 mM IPTG for 16 h at 21 °C. Harvested cells were resuspended in 50 mM Tris–HCl, 300 mM NaCl, 20 mM MgCl2, and 10% glycerol, pH 8.0, and sonicated in the presence of EDTA- free protease inhibitor cocktail, lysozyme and DNAse. Clarified lysate was purified via Ni- NTA columns. AurA was eluted in 100% of 50 mM Tris–HCl, 300 mM NaCl, 500 mM imidazole, 20 mM MgCl2, and 10% glycerol, pH 8.0, which was combined with TEV and GST- LPP, and then dialyzed overnight against 50 mM Tris–HCl, 300 mM NaCl, 1 mM MnCl2, 5 mM TCEP, and 10% glycerol, pH 7.5 at 4 °C. Cleaved Aurora A was purified with Ni- NTA and GST columns and subsequently polished with a 26/600 S200 pg gel filtration column equilibrated in 20 mM Tris–HCl, 200 mM NaCl, 20 mM MgCl2, 5 mM TCEP, and 8 of 11   https://doi.org/10.1073/pnas.2304611120 pnas.org 10% glycerol, pH 7.5. Pure fractions were pooled and concentrated to around 40  μM, and stored in −80 °C. Monobodies (TEV- cleavable, N- terminal His6- tagged) were purified with on- column refolding as described in Zorba et al. (25). Cloning and Purification of AblKD and Abl64– 510. AblKD (residues 229 to 510, TEV- cleavable, N- terminal MBP- His6- tagged) and Abl64– 510 (residues 64 to 510, TEV- cleavable, N- terminal MBP- His6- tagged) were cloned into pETm41 (GenScript) (SI Appendix, Fig.  S7). Residue numbering follows Abl1b isoform that naturally consists of N- myristoylation. All Abl constructs were cotransformed with phosphatase YOPH (streptomycin- resistance) in BL21(DE3) cells and plated on Kan/Strep LB plate. Expression was performed in TB media and induced at OD = 0.6–0.8 with 0.1 mM (for AblKD) or 0.2 mM IPTG (for Abl64– 510) for 16 to 20 h at 18 °C. Harvested cells were resuspended in 50 mM Tris–HCl, 500 mM NaCl, and 1 mM TCEP, pH 8.0 (buffer A). Cells were sonicated in the presence of EDTA- free protease inhibitor cocktail, lysozyme, and DNAse. Clarified lysate was with Ni- NTA columns. The protein was eluted with 100% 50 mM Tris–HCl, 500 mM NaCl, 500 mM imidazole, and 1 mM TCEP, pH 8.0 and combined with TEV and CIP (#M0525, NEB) and dialyzed overnight against buffer A at 4 °C. Cleaved Abl was further purified with Ni- NTA and a Q column (gradient elution with 50 mM Tris–HCl, 1 M NaCl, and 1 mM TCEP, pH 8.0). Prior to anion exchange chroma- tography Abl was dialyzed into 50 mM Tris–HCl, 1 mM TCEP, and 10% glycerol, pH 8.0. Dephosphorylated Abl fractions were polished with 26/600 S75 pg (for AblKD) or 26/600 S200 pg (for Abl64– 510) gel filtration column with buffer A. Pure fractions were aliquoted to around 40 μM and stored in −80 °C. ITC. All titrations were carried out using Nano ITC (TA Instruments) and analyzed via the NanoAnalyze software either using the independent fit model or compet- itive replacement model. The first injection of each experiment was discarded according to the software manual. For AurA, danusertib (Selleckchem #S1107) was reconstituted to 100 mM in 100% DMSO and was diluted to appropriate concentration to match final 5% DMSO (vol/vol) for each experiment. An ADP- analogue, AMPCP, was used for competitive replacement experiments to measure and fitting of the binding of danusertib to AurA. All proteins were dialyzed in 20 mM Tris–HCl, 200 mM NaCl, 10% (vol/vol) glycerol, and 5 mM TCEP, pH 7.5. AMPCP was resuspended with the same buffer and was matched to pH 7.5. DMSO was added prior to each experiment to match 5% between titrant and titrand. Each injection was added in 2 µL increments with 180 s interval at a constant stirring speed of 300 rpm and at 25 °C. Concentrations used for the experiments are noted in SI Appendix. For Abl, N- Myr peptide (Myr- GQQPGKVLGDQR), ordered from GenScript, was used for competitive replacement experiments to measure and fitting of the binding of asciminib to Abl. All proteins were dialyzed in 50 mM Tris–HCl, 500 mM NaCl, and 1 mM TCEP, pH 8.0. Imatinib- mesylate (Sigma #SML- 1027) and asci- minib (MedKoo #206490) were reconstituted to 10 mM in 100% DMSO and were diluted to appropriate concentration for each experiment. N- Myr peptide was resuspended with the same buffer and was matched to pH 8.0. DMSO was added prior to each experiment to match 5% between titrant and titrand. Each injection was added in 1 to 1.5 µL increments with 180 s interval at a constant stirring speed of 300 rpm and at 25 °C. Concentrations used for the experiments are noted in SI Appendix. In Vitro Kinase Assay. To measure the IC50 of danusertib to AurA, ADP- GloTM Max assay (Promega #V7001) was used. 20 nM AurA in the absence or presence of either saturating concentration of Mb1 or Mb2 or Mb3 was incubated with 3 mM Lats2 (ATLARRDSLQKPGLE), 0.6 mg/mL BSA, and varying concentrations of danusertib with final 5% (vol/vol) of DMSO at 25 °C in 20 mM Tris–HCl, 200 mM NaCl, 10% (vol/vol) glycerol, and 5 mM TCEP, pH 7.50. The bolded and underlined residue indicates site of phosphorylation. The reaction was initiated by adding 5 mM ATP, and the final samples were collected after 2 h for AurA- Mb1 complex, 10 h for apo AurA, and 20 h for AurA- Mb2 and AurA- Mb3 complexes. The amount of ADP in the samples was measured by following the manufacturer’s protocol and used to calculate the observed rate. Assays for Abl were performed at 25 °C with half- well 96- well plate (Corning #3994) in 50 mM Tris–HCl, 500 mM NaCl, and 1 mM TCEP, pH 8.0, supplemented with 20 nM Abl kinase (AblKD or Abl64– 510), 2 mM Srctide (EIYGEFKK), 0.6 mg/mL BSA, 20 mM MgCl2, 750 µM NADH, 6 mM PEP, and 2.5 units of PK/LDH (Sigma #P0294). The bolded and underlined residue indicates site of phosphorylation. Oxidation of NADH at A340 was monitored using SpectraMAX by starting the assay with 1 mM ATP. The final volume of the assay was 100 µL. The observed rate (kobs) was calculated following Zorba et al. (25). All data were processed using GraphPad Prism and fitted to a four- parameter dose- response model. Molecular Dynamics Simulation. All- atom molecular dynamics simulations were conducted using OpenMM 7.6 (50) and “Making it rain” cloud- based notebook environment (51). The structure of AurA- danusertib- Mb1 was used as an initial model. To mimic danusertib binding to AurA- Mb1 under ITC con- ditions, we created such structure via removal of danusertib from our ternary complex AurA- danusertib- Mb1 [since the published AurA- Mb1 structure (PDB- ID: 5G15) has AMPPCP bound to active site (25). Parameterization for all MD runs was conducted using LEaP (52) with Amber ff14SB force field (53), GAFF2 (54) for ligand, and TIP3P (55, 56) water model. The systems were neutralized with NaCl at 0.2 mM, following the ITC conditions, and box size was set at 20 Å. AurA- Mb1 and AurA- danusertib- Mb1 structures were equilibrated to 298 K via Langevin dynamics (57) and 1 bar via Monte Carlo barostat (58) with 2 fs integration time. We set 10,000 steps of energy minimization with 1,000 kJ/mol of harmonic position restraints. The systems were equilibrated for 0.2 ns and 1 ns for AurA- Mb1 and AurA- danusertib- Mb1, respectively, in the NVT ensemble. Then, with accordingly equilibrated systems, triplicates of 10 ns production runs were done in the NPT ensemble. Trajectories were analyzed using VMD 1.9.4a53 (59). FRET Measurements. FluoroMax- 4 (Horiba Scientific) with temperature con- troller (water bath) was used to measure FRET between intrinsic tryptophan fluorescence and SKI. Either 10 nM Abl or 10 nM Abl + 200 nM asciminib was preincubated with varying concentrations of SKI for 40 min at 25 °C before meas- urements. An increase in the fluorescence was measured when the complex, specifically tryptophan, was excited at 295 nm to emit at 340 nm, which then excites SKI to emit at 460 nm (SI Appendix, Fig. S9). Both 5 nm of excitation and emission slit width were used. Control experiments (buffer- only, protein- only, and inhibitor- only) were confirmed that the increase of fluorescence is caused by the fluorescence energy transfer. The fluorescence intensity at 460 nm versus SKI concentration was fitted to the quadratic equation below in GraphPad Prism to obtain apparent Kd. + ( I [ ] Et] [ + K d) − F = F0 + A √( I + + K d)2 − 4 [Et] [I] Et] [ [ ] 2[Et] We simulated curves with tighter Kd for comparison to ensure that the fitted curves are not step functions due to the high enzyme concentration (SI Appendix, Fig. S10). Crystallographic Methods. Crystals of AurA in complex with Mb1 and danus- ertib were obtained by combining 2 µL of 300 µM (10 mg/mL) AurA + 315 µM (4 mg/mL) Mb1 + 2 mM AMPPCP + 4 mM MgCl2 with 2 µL reservoir of 0.1 M MES pH 6.5 + 0.2 M ammonium sulfate + 4% (v/v) 1,3- propanediol + 15 to 18% PEG 8,000. Streak seeding was used to obtain bigger crystals. Crystals were grown at 18 °C by hanging drop. The crystals were transferred to a drop of fresh reservoir for 30 s to remove excess nucleotides from the crystal sur- face. Then, the crystals were transferred to a drop with reservoir with 1 mM danusertib for 16 h of soaking. For cryoprotection, the crystals were transferred into 17.5% PEG 400, 17.5% ethylene glycol, 15% reservoir, and 50% water for a few seconds. Crystals of AurA in complex with Mb2 and danusertib were obtained by com- bining 0.5 µL of 300 µM (10 mg/mL) AurA + 315 µM (4 mg/mL) Mb2 + 1 mM danusertib with 0.5 µL of 0.1 M BIS–TRIS pH 5.5 + 0.2 M Ammonium acetate + 25% PEG3350. Crystals were grown at 18 °C by sitting drop. Crystals were harvested and subsequently flash frozen. Diffraction data for AurA- danusertib- Mb1 and AurA- danusertib- Mb2 were collected at 100 K Advanced Light Source (Lawrence Berkeley National Laboratory) at beamlines BL821 and BL501, respectively, and were integrated with XIA2 (60) or XDS (61). Data were scaled and merged with AIMLESS (62). Initial phases were obtained with molec- ular replacement programs MOLREP (63) and PHASER (64) by using AurA + Mb1 PNAS  2023  Vol. 120  No. 34  e2304611120 https://doi.org/10.1073/pnas.2304611120   9 of 11 + AMPPCP (PDB- ID: 5G15) for AurA- danusertib- Mb1 structure and AurA + AMPPCP (PDB- ID: 4C3R) and HA4Mb (PDB- ID: 3K2M) for AurA- danusertib- Mb2 structure using two molecules each in the asymmetric unit. The structures were iteratively refined using refmac and phenix.refine (Version1.19.1) (65) followed by manual model building in COOT (66). Models were validated with MolProbity (67). Molecular structures were represented and rendered with ChimeraX (68, 69). Crystals of Abl64– 510 in complex with SKI and asciminib were obtained by combin- ing 0.3 µL of 600 µM Abl64– 510 + 700 µM SKI + 700 µM asciminib (~32 mg/mL) in 5% DMSO with 0.4 µL reservoir of 0.1 M Tris–HCl pH 8 + 1.75 M Ammonium sulfate + 2% (v/v) polypropylene glycol 400 (PPG 400). The final stock of complex was con- centrated from 1 µM Abl64– 510 with ~1.2 µM SKI/asciminib after incubation at 4 °C for 6 h. Screening around this condition yielded crystals in a transparent diamond- shaped or plate- shaped crystals. Crystals were grown at 18 °C by sitting drop for a few days. The crystals were transferred to a drop of fresh reservoir containing 20% xylitol with matching concentration of inhibitors in 5% DMSO for few seconds for cryoprotection. Single crystal X- ray diffraction data were collected at 100 K at Advanced Light Source Berkeley (BL201). Data were integrated with XDS (61) as well as scaled and merged with AIMLESS (62). Analysis of processed data with phenix.xtriage (70) found outliers in the dataset and further revealed substantial translational noncrystallographic symmetry with a Patterson peak of 56.63% height relative to origin, complicating refinement. Initial phases were obtained by molecular replacement (PHASER) (64) using Abl- nilotinib- asciminib (PDB- ID: 5MO4) as a search model with two molecules in the asymmetric unit. The kinase domain, SH3, and SH2 (regulatory domains) were individually placed during molecular replacement. Refinement and manual model building were performed by phenix.refine (version 1.19.1) and Coot, respectively (65, 66). Models were validated with MolProbity (67). Molecular structures were represented and rendered with ChimeraX (68, 69) and PyMol (71). Data, Materials, and Software Availability. Structure factors and refined coordinates obtained from X- ray crystallography have been deposited into the Protein Data Bank (www.wwpdb.org) under PDB accession codes: 8SSP (72) (AurA- danusertib- Mb1), 8SSO (73)  (AurA- danusertib- Mb2), and 8SSN (74) (Abl64– 510-SKI- asciminib). ACKNOWLEDGMENTS. D.K. is supported by the Howard Hughes Medical Institute (HHMI). The Berkeley Center for Structural Biology is supported by the HHMI, Participating Research Team members, and the NIH, National Institute of General Medical Sciences, ALS- ENABLE grant P30 GM124169. The Advanced Light Source is a Department of Energy Office of Science User Facility under Contract No. DE- AC02- 05CH11231. The Pilatus detector on beamline 2.0.1 was funded under NIH grant S10OD021832. The Pilatus detector on beamline 5.0.1 was funded under NIH grant S10OD026941. Author affiliations: aDepartment of Biochemistry, Brandeis University, Waltham, MA 02454; and bHHMI, Brandeis University, Waltham, MA 02454 Author contributions: C.K., A.H., V.N., and D.K. designed research; C.K., H.L., A.H., S.K., and V.N. performed research; C.K., H.L., A.H., S.K., V.N., and D.K. analyzed data; and C.K., H.L., and D.K. wrote the paper. Competing interest statement: D.K. is co- founder of Relay Therapeutics and MOMA Therapeutics. The remaining authors declare no competing interests. 1. 2. 3. 4. 5. J.- P. Changeux, Allostery and the monod- wyman- changeux model after 50 years. Annu. Rev. Biophys. 41, 103–133 (2012). S. K. Hanks, A. M. Quinn, T. Hunter, The protein kinase family: Conserved features and deduced phylogeny of the catalytic domains. Science 241, 42–52 (1988). S. A. Rosenzweig, Acquired resistance to drugs targeting tyrosine kinases. Adv. Cancer Res. 138, 71–98 (2018). S. Lu et al., Emergence of allosteric drug- resistance mutations: New challenges for allosteric drug discovery. Drug Discov. Today 25, 177–184 (2020). X. Lu, J. B. Smaill, K. Ding, New promise and opportunities for allosteric kinase inhibitors. Angew. Chem. Int. Ed. Engl. 59, 13764–13776 (2020). 25. A. Zorba et al., Allosteric modulation of a human protein kinase with monobodies. Proc. Natl. Acad. Sci. U.S.A. 116, 13937–13942 (2019). 26. R. Bayliss, T. Sardon, I. Vernos, E. Conti, Structural basis of Aurora- A activation by TPX2 at the mitotic spindle. Mol. Cell 12, 851–862 (2003). 27. D. Fancelli et al., 1, 4, 5, 6- tetrahydropyrrolo [3, 4- c] pyrazoles: Identification of a potent Aurora kinase inhibitor with a favorable antitumor kinase inhibition profile. J. Med. Chem. 49, 7247–7251 (2006). 28. E. W. Lake et al., Quantitative conformational profiling of kinase inhibitors reveals origins of selectivity for Aurora kinase activation states. Proc. Natl. Acad. Sci. U.S.A. 115, E11894–E11903 (2018). 6. H. Mobitz, W. Jahnke, S. W. Cowan- Jacob, Expanding the opportunities for modulating kinase 29. V. Modi, R. L. Dunbrack Jr., Defining a new nomenclature for the structures of active and inactive 7. targets with allosteric approaches. Curr. Topics Med. Chem. 17, 59–70 (2017). J. Zhang et al., Targeting Bcr- Abl by combining allosteric with ATP- binding- site inhibitors. Nature 463, 501–506 (2010). kinases. Proc. Natl. Acad. Sci. U.S.A. 116, 6818–6827 (2019). 30. V. Modi, R. L. Dunbrack Jr., Kincore: A web resource for structural classification of protein kinases and their inhibitors. Nucleic Acids Res. 50, D654–D664 (2022). 8. D. Rea et al., A phase 3, open- label, randomized study of asciminib, a STAMP inhibitor, vs bosutinib 31. A. Zorba et al., Molecular mechanism of Aurora A kinase autophosphorylation and its allosteric 9. in CML after 2 or more prior TKIs. Blood 138, 2031–2041 (2021). T. P. Hughes et al., Asciminib in chronic myeloid leukemia after ABL kinase inhibitor failure. New Engl. J. Med. 381, 2315–2323 (2019). 10. J. Schoepfer et al., Discovery of Asciminib (ABL001), an Allosteric Inhibitor of the Tyrosine Kinase Activity of BCR- ABL1 (ACS Publications, 2018). 11. P. W. Manley, L. Barys, S. W. Cowan- Jacob, The specificity of asciminib, a potential treatment for chronic myeloid leukemia, as a myristate- pocket binding ABL inhibitor and analysis of its interactions with mutant forms of BCR- ABL1 kinase. Leukemia Res. 98, 106458 (2020). 12. A. A. Wylie et al., The allosteric inhibitor ABL001 enables dual targeting of BCR–ABL1. Nature 543, 733–737 (2017). 13. C. A. Eide et al., Combining the allosteric inhibitor asciminib with ponatinib suppresses emergence of and restores efficacy against highly resistant BCR- ABL1 mutants. Cancer Cell 36, 431–443.e435 (2019). activation by TPX2. Elife 3, e02667 (2014). 32. C. A. Dodson et al., Crystal structure of an Aurora- A mutant that mimics Aurora- B bound to MLN8054: Insights into selectivity and drug design. Biochem. J. 427, 19–28 (2010). 33. D. E. Anderson, W. J. Becktel, F. W. Dahlquist, pH- induced denaturation of proteins: A single salt bridge contributes 3–5 kcal/mol to the free energy of folding of T4 lysozyme. Biochemistry 29, 2403–2408 (1990). 34. L. Skora, J. Mestan, D. Fabbro, W. Jahnke, S. Grzesiek, NMR reveals the allosteric opening and closing of Abelson tyrosine kinase by ATP- site and myristoyl pocket inhibitors. Proc. Natl. Acad. Sci. U.S.A. 110, E4437–E4445 (2013). 35. B. Nagar et al., Structural basis for the autoinhibition of c- Abl tyrosine kinase. Cell 112, 859–871 (2003). 36. T. Saleh, P. Rossi, C. G. Kalodimos, Atomic view of the energy landscape in the allosteric regulation of Abl kinase. Nat. Struct. Mol. Biol. 24, 893–901 (2017). 14. K. V. Gleixner et al., Asciminib and ponatinib exert synergistic anti- neoplastic effects on CML cells 37. W. Jahnke et al., Binding or bending: Distinction of allosteric Abl kinase agonists from antagonists expressing BCR- ABL1T315I- compound mutations. Am. J. Cancer Res. 11, 4470 (2021). by an NMR- based conformational assay. J. Am. Chem. Soc. 132, 7043–7048 (2010). 15. T. S. Beyett et al., Molecular basis for cooperative binding and synergy of ATP- site and allosteric EGFR 38. T. K. Johnson et al., Synergy and antagonism between allosteric and active- site inhibitors of abl inhibitors. Nat. Commun. 13, 1–11 (2022). tyrosine kinase. Angew. Chem. 133, 20358–20361 (2021). 16. M. Fodor et al., Dual allosteric inhibition of SHP2 phosphatase. ACS Chem. Biol. 13, 647–656 39. T. Xie, T. Saleh, P. Rossi, C. G. Kalodimos, Conformational states dynamically populated by a kinase (2018). 17. J. Colicelli, ABL tyrosine kinases: Evolution of function, regulation, and specificity. Sci. Signaling 3, re6 (2010). 18. N. Ertych et al., Increased microtubule assembly rates influence chromosomal instability in colorectal cancer cells. Nat. Cell Biol. 16, 779–791 (2014). determine its function. Science 370, eabc2754 (2020). 40. O. Hantschel et al., A myristoyl/phosphotyrosine switch regulates c- Abl. Cell 112, 845–857 (2003). 41. G. Tian, M. Cory, A. A. Smith, W. B. Knight, Structural determinants for potent, selective dual site inhibition of human pp60 c- src by 4- anilinoquinazolines. Biochemistry 40, 7084–7091 (2001). 42. J. Bain et al., The selectivity of protein kinase inhibitors: A further update. Biochem. J. 408, 297–315 19. H. Zhou et al., Tumour amplified kinase STK15/BTAK induces centrosome amplification, aneuploidy (2007). and transformation. Nat. Genet. 20, 189–193 (1998). 43. Y. Choi et al., N- myristoylated c- Abl tyrosine kinase localizes to the endoplasmic reticulum upon 20. R. Du, C. Huang, K. Liu, X. Li, Z. Dong, Targeting AURKA in cancer: Molecular mechanisms and binding to an allosteric inhibitor. J. Biol. Chem. 284, 29005–29014 (2009). opportunities for cancer therapy. Mol. Cancer 20, 1–27 (2021). 44. D. Fabbro et al., Inhibitors of the Abl kinase directed at either the ATP- or myristate- binding site. 21. C. O. De Groot et al., A cell biologist’s field guide to aurora kinase inhibitors. Front. Oncol. 5, 285 Biochim. Biophys. Acta 1804, 454–462 (2010). (2015). 45. W. Qiang et al., Mechanisms of resistance to the BCR- ABL1 allosteric inhibitor asciminib. Leukemia 22. W. Pitsawong et al., Dynamics of human protein kinase Aurora A linked to drug selectivity. Elife 7, 31, 2844–2847 (2017). e36656 (2018). 46. T. Xie, T. Saleh, P. Rossi, D. Miller, C. G. Kalodimos, Imatinib can act as an Allosteric Activator of Abl 23. J. A. Gilburt et al., Dynamic equilibrium of the aurora A kinase activation loop revealed by single- Kinase. J. Mol. Biol. 434, 167349 (2022). molecule spectroscopy. Angew. Chem. 129, 11567–11572 (2017). 24. S. Cyphers, E. F. Ruff, J. M. Behr, J. D. Chodera, N. M. Levinson, A water- mediated allosteric network governs activation of Aurora kinase A. Nat. Chem. Biol. 13, 402 (2017). 47. R. Sonti, I. Hertel- Hering, A. J. Lamontanara, O. Hantschel, S. Grzesiek, ATP site ligands determine the assembly state of the Abelson kinase regulatory core via the activation loop conformation. J. Am. Chem. Soc. 140, 1863–1869 (2018). 10 of 11   https://doi.org/10.1073/pnas.2304611120 pnas.org 48. S. Grzesiek, J. Paladini, J. Habazettl, R. Sonti, Imatinib disassembles the regulatory core of Abelson kinase by binding to its ATP site and not by binding to its myristoyl pocket. Magn. Res. 3, 91–99 (2022). 49. C. To et al., Single and dual targeting of mutant EGFR with an allosteric inhibitor. Cancer Discov. 9, 926–943 (2019). 61. W. Kabsch, XDS. Acta Crystallogr. D Biol. Crystallogr. 66, 125–132 (2010). 62. P. R. Evans, G. N. Murshudov, How good are my data and what is the resolution? Acta Crystallogr. D Biol. Crystallogr. 69, 1204–1214 (2013). 63. A. Vagin, A. Teplyakov, MOLREP: An automated program for molecular replacement. J. Appl. 50. P. Eastman et al., OpenMM 7: Rapid development of high performance algorithms for molecular Crystallogr. 30, 1022–1025 (1997). dynamics. PLoS Comput. Biol. 13, e1005659 (2017). 51. P. R. Arantes, M. D. Polêto, C. Pedebos, R. Ligabue- Braun, Making it rain: Cloud- based molecular simulations for everyone. J. Chem. Inf. Model. 61, 4852–4856 (2021). 52. D. A. Case et al., Amber 2021 (University of California, San Francisco, 2021). 53. J. A. Maier et al., ff14SB: Improving the accuracy of protein side chain and backbone parameters 64. A. J. McCoy et al., Phaser crystallographic software. J. Appl. Crystallogr. 40, 658–674 (2007). 65. P. V. Afonine et al., Towards automated crystallographic structure refinement with phenix. refine. Acta Crystallogr. D Biol. Crystallogr. 68, 352–367 (2012). 66. P. Emsley, K. Cowtan, Coot: Model- building tools for molecular graphics. Acta Crystallogr. D Biol. Crystallogr. 60, 2126–2132 (2004). from ff99SB. J. Chem. Theory Comput. 11, 3696–3713 (2015). 67. C. J. Williams et al., MolProbity: More and better reference data for improved all- atom structure 54. X. He, V. H. Man, W. Yang, T.- S. Lee, J. Wang, A fast and high- quality charge model for the next validation. Protein Sci. 27, 293–315 (2018). generation general AMBER force field. J. Chem. Phys. 153, 114502 (2020). 68. E. F. Pettersen et al., UCSF ChimeraX: Structure visualization for researchers, educators, and 55. H. J. Berendsen, J. P. Postma, W. F. van Gunsteren, J. Hermans, “Interaction models for water in developers. Protein Sci. 30, 70–82 (2021). relation to protein hydration” in Intermolecular Forces, B. Pullman, Eds. (Springer, Dordrecht, 1981), https://doi.org/10.1007/978- 94- 015- 7658- 1_21. pp. 331–342. 69. T. D. Goddard et al., UCSF ChimeraX: Meeting modern challenges in visualization and analysis. Protein Sci. 27, 14–25 (2018). 56. W. L. Jorgensen, J. Chandrasekhar, J. D. Madura, R. W. Impey, M. L. Klein, Comparison of simple 70. P. D. Adams et al., PHENIX: A comprehensive Python- based system for macromolecular structure potential functions for simulating liquid water. J. Chem. Phys. 79, 926–935 (1983). solution. Acta Crystallogr. D Biol. Crystallogr. 66, 213–221 (2010). 57. P. Turq, F. Lantelme, H. L. Friedman, Brownian dynamics: Its application to ionic solutions. J. Chem. 71. W. L. DeLano, Pymol: An open- source molecular graphics tool. CCP4 Newsl. Protein Crystallogr. 40, Phys. 66, 3039–3044 (1977). 82–92 (2002). 58. J. Åqvist, P. Wennerström, M. Nervall, S. Bjelic, B. O. Brandsdal, Molecular dynamics simulations of 72. C. Kim et al., AurA bound to danusertib and activating monobody Mb1. Protein Data Bank. https:// water and biomolecules with a Monte Carlo constant pressure algorithm. Chem. Phys. Lett. 384, 288–294 (2004). 59. W. Humphrey, A. Dalke, K. Schulten, VMD: Visual molecular dynamics. J. Mol. Graphics 14, 33–38 (1996). 60. G. Winter, xia2: An expert system for macromolecular crystallography data reduction. J. Appl. www.rcsb.org/structure/8SSP. Accessed 8 May 2023. 73. C. Kim et al., AurA bound to danusertib and inhibiting monobody Mb2. Protein Data Bank. https:// www.rcsb.org/structure/8SSO. Accessed 8 May 2023. 74. C. Kim et al., Abl kinase in complex with SKI and asciminib. Protein Data Bank. https://www.rcsb.org/ Crystallogr. 43, 186–190 (2010). structure/8SSN. Accessed 8 May 2023. PNAS  2023  Vol. 120  No. 34  e2304611120 https://doi.org/10.1073/pnas.2304611120   11 of 11
10.7554_elife.86784
RESEARCH ARTICLE Structural insight into guanylyl cyclase receptor hijacking of the kinase–Hsp90 regulatory mechanism Nathanael A Caveney1*, Naotaka Tsutsumi1,2†, K Christopher Garcia1,2* 1Departments of Molecular and Cellular Physiology, and Structural Biology, Stanford University School of Medicine, Stanford, United States; 2Howard Hughes Medical Institute, Stanford University School of Medicine, Stanford, United States Abstract Membrane receptor guanylyl cyclases play a role in many important facets of human physiology, from regulating blood pressure to intestinal fluid secretion. The structural mechanisms which influence these important physiological processes have yet to be explored. We present the 3.9 Å resolution cryo- EM structure of the human membrane receptor guanylyl cyclase GC- C in complex with Hsp90 and its co- chaperone Cdc37, providing insight into the mechanism of Cdc37 mediated binding of GC- C to the Hsp90 regulatory complex. As a membrane protein and non- kinase client of Hsp90–Cdc37, this work shows the remarkable plasticity of Cdc37 to interact with a broad array of clients with significant sequence variation. Furthermore, this work shows how membrane receptor guanylyl cyclases hijack the regulatory mechanisms used for active kinases to facilitate their regulation. Given the known druggability of Hsp90, these insights can guide the further development of membrane receptor guanylyl cyclase- targeted therapeutics and lead to new avenues to treat hypertension, inflammatory bowel disease, and other membrane receptor guanylyl cyclase- related conditions. eLife assessment In this important study, the human membrane receptor guanyl cyclase GC- C was expressed in hamster cells, co- purified in complex with endogenous HSP90 and CDC37 proteins, and the struc- ture of the complex was determined by cryo- EM. The study shows that the pseudo- kinase domain of GC- C associates with CDC37 and HSP90, similarly to how the bona fide protein kinases CDK4, CRAF and BRAF have been shown to interact. The methodology used is state of the art and the evidence presented is compelling. Introduction Cyclic guanosine monophosphate (cGMP) is an important second messenger for signaling in mamma- lian physiology, with roles in platelet aggregation, neurotransmission, sexual arousal, gut peristalsis, bone growth, intestinal fluid secretion, lipolysis, phototransduction, cardiac hypertrophy, oocyte maturation, and blood pressure regulation (Potter, 2011). Largely, cGMP is produced in response to the activation of guanylyl cyclases (GC), a class of receptors that contains both heteromeric soluble receptors (α1, α2, β1, and β2 in humans) and five homomeric membrane receptors (GC- A, GC- B, GC- C, GC- E, and GC- F in humans). Of note are the membrane receptor guanylyl cyclases (mGC) GC- A and GC- B, also known as natriuretic peptide receptors A and B (NPR- A and NPR- B), respectively, and GC- C, all of which have been a focus of therapeutic development. In the case of NPR- A and B, their role in regulating blood pressure in response to natriuretic peptide hormones (ANP, BNP, and *For correspondence: ncaveney@stanford.edu (NAC); kcgarcia@stanford.edu (KCG) Present address: †Graduate School of Medicine, Dentistry and Pharmaceutical Sciences, Okayama University, Okayama, Japan Competing interest: The authors declare that no competing interests exist. Funding: See page 9 Sent for Review 25 February 2023 Preprint posted 16 March 2023 Reviewed preprint posted 02 May 2023 Reviewed preprint revised 12 July 2023 Version of Record published 03 August 2023 Reviewing Editor: Mohamed Trebak, University of Pittsburgh, United States Copyright Caveney et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 1 of 12 Research article CNP) has led to the exploration of agonists for use in the treatment of cardiac failure (Kobayashi et al., 2012). Meanwhile, GC- C is the target of clinically approved laxative agonists, linaclotide, and plecanatide (Miner, 2020; Yu and Rao, 2014), which increase intestinal fluid secretion. These membrane receptor GCs consist of an extracellular ligand binding domain (ECD), which acts as a conformational switch to drive intracellular rearrangements to activate the receptor (He et al., 2001) a transmembrane region (TM); a kinase homology domain or pseudokinase domain (PK); a dimerization domain; and a GC domain, which acts to produce cGMP. The PK domain is largely thought to be involved in scaffolding and physical transduction of the extracellular rearrangements to the GC domain, in some respects similar to the role of the PK domain in the Janus kinases of the cytokine signaling system (Glassman et al., 2022). In addition, the PK domains of mGCs are regulated through phosphorylation (Potter and Garbers, 1992; Potter and Hunter, 1998; Vaandrager et al., 1993) and via association with heat shock proteins (Hsp) (Kumar et al., 2001). While the role of the phosphorylation state on mGC activity has been explored in relative detail, how the heat shock protein 90 (Hsp90) is able to regulate mGC activity is largely unknown. It has been shown that GC- A activity can be regulated through the association of Hsp90 and the co- chap- erone Cdc37 (Kumar et al., 2001). The chaperone Cdc37 is known to assist in the Hsp90 regulation of around 60% of active kinases, both in soluble and membrane receptor form (Taipale et al., 2012). Given the sequence and structural similarities between the PK domains of mGCs and the active kinase domains which Hsp90–Cdc37 regulates, it is possible that mGCs have evolved to hijack the regulatory mechanisms that are more broadly deployed for active kinases. Here, we report the 3.9 Å resolution structure of the GC- C–Hsp90–Cdc37 regulatory complex. In this structure, the core dimer of Hsp90 forms its canonical closed conformation, while Cdc37 and the C- lobe of the GC- C PK domain asymmetrically decorate the complex. The client (GC- C) is unfolded into the channel formed at the interface between the Hsp90 dimers. To our knowledge, this is the first structure of a membrane protein client of Hsp90 and the first structure of a non- kinase client of the Hsp90–Cdc37 regulatory system. This work provides a pivotal understanding of the mechanism and structural basis of kinase fold recruitment to the Hsp90–Cdc37 regulatory complex. This increased understanding can guide the further development of mGC- targeted therapeutics and lead to new avenues to treat hypertension, inflammatory bowel disease (IBD), and other mGC- related conditions. In addition, the general insights into the recruitment of Hsp90–Cdc37 clients can guide the further development of Hsp90 targeting therapeutics in cancer treatment. Results Structure of the GC-C–Hsp90–Cdc37 regulatory complex Membrane receptor guanylyl cyclases have been largely recalcitrant to structural analysis by x- ray crystallography and electron microscopy, apart from various crystal structures of both liganded and unliganded ECDs (He et al., 2001; He et al., 2006; Ogawa et al., 2004; Ogawa et al., 2010; van den Akker et al., 2000). Given the relative disparity of our structural understanding, we sought to develop a stable construct to image and gain a crucial understanding of the regulatory and functional aspects of mGCs which occur intracellularly. By replacing the ligand- responsive ECD with a homod- imeric leucine zipper, we mimic the ligand- activated geometry of the ECD (He et al., 2001), while reducing complexity of the imaged complex and increasing stability (Figure 1A). This complex was recombinantly expressed in mammalian cells, purified with anti- FLAG affinity chromatography, and vitrified on grids for cryo- EM analysis. The purified sample had a substantial portion of imaged particles for which the native regulatory heat shock protein, Hsp90, and its co- chaperone, Cdc37, are bound. The Cricetulus griseus HSP90β and Cdc37 show remarkable sequence conservation in comparison to the human equivalents, at 99.7 and 94.2% identity, respectively. This native pulldown strategy contrasts with the structures of Hsp90– Cdc37 in complex with soluble kinases (García- Alonso et  al., 2022; Oberoi et  al., 2022; Verba et al., 2016), for which Hsp90 and Cdc37 had to be overexpressed to obtain complex suitable for imaging. Three- dimensional reconstruction of our GC- C–Hsp90–Cdc37 particles generated a 3.9 Å resolution map of the regulatory complex (Figure  1, Figure  1—figure supplements 1 and 2). A second, unsharpened map from subsequent heterogeneous refinement resolves additional density Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 2 of 12 Biochemistry and Chemical Biology Research article A ECD TM TM PK DD GC B GC-C Gn, Uro zipper zipper Cdc37 Cdc37 TMTM TM TM PK PK DD DD GC GC PK PK DD DD GC GC Hsp90 Hsp90 Hsp90 Hsp90 GTP cGMP signal C GC-C 90Å D NTD NTD Cdc37 coiled- coil MD Hsp90 Hsp90 Cdc37 coiled- coiled- coil coil MD CTD DDDD PK PK C-lobe C-lobe E NTD NTD Cdc37 coiled- coil MD MD GC-C PK C-lobe CTD 90° NTD MD MD MD GC-C PK PK C-lobe C-lobe MD 90° GC-C PK C-lobe CTD CTD CTD MD CTD Figure 1. Composition and cryo- EM structure of the GC- C–Hsp90–Cdc37 regulatory complex. (A) Cartoon representation of the components of guanylyl cyclase C (GC- C) signaling and Hsp90–Cdc37 regulation and the zippered and activated GC- C. GC- C is colored in red, guanylin/uroguanylin (Gn/Uro) in yellow, Hsp90 in blue and teal, and Cdc37 in purple. Extracellular domains (ECD), transmembrane domain (TM), pseudokinase domain (PK), dimerization domain (DD), and guanylyl cyclase domain (GC) are labeled. In the rightmost cartoon, the regions unobserved in the cryo- EM density are in a lighter shade with a dashed outline. (B) The refined and sharpened cryo- EM density map of GC- C–Hsp90–Cdc37, colored as in A, with a transparent overlay of an unsharpened map with additional DD density resolved. Cdc37 coil- coiled and middle domain (MD) are labeled. (C) Reference- free 2D averages for the GC- C–Hsp90–Cdc37 complex. (D) The refined and sharpened cryo- EM density map of GC- C– Hsp90–Cdc37, colored as in A and B, labeled with all domains as in A and B, with the addition of Hsp90 N- terminal domain (NTD), middle domain (MD), and C- terminal domain (CTD). (E) Ribbon representation of a model of GC- C– Hsp90–Cdc37 complex, colored and labeled as in A, B, and C. The online version of this article includes the following figure supplement(s) for figure 1: Figure 1 continued on next page Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 3 of 12 Biochemistry and Chemical Biology Research article Figure 1 continued Figure supplement 1. GC- C–Hsp90–Cdc37 complex cryo- EM data processing. Figure supplement 2. Representative density of GC- C–Hsp90–Cdc37. for the dimerization domain, extending outward from the PK domain (Figure 1B, Figure 1—figure supplement 1). The resultant GC- C–Hsp90–Cdc37 complex is a hetero- tetramer formed by one resolved monomer of the GC- C receptor bound to a dimer of Hsp90 and one Cdc37 co- chaperone (Figure  1D). As observed with most Hsp90–client structures, the bulk of the complex is composed of the C2 pseudo- symmetric, ATP bound, closed state Hsp90 dimer. Building on this dimeric core, the Cdc37 protrudes outward from one side with its characteristic long, coiled- coil, α-hairpin. On one face of the Hsp90 dimer core, Cdc37 interacts with the PK domain of GC- C, while an extended β-sheet wraps around to the other face, lying across and extending a β-sheet in the middle domain (MDHsp90) of one Hsp90 monomer. At the opposite face, the globular and α-helical Cdc37 middle domain (MDCdc37) is formed. The C- lobe of the GC- C PK domain packs against the N- terminal region of Cdc37 on one face of the dimeric Hsp90 core, with the N- lobe unfolding through the dimer core to interface with the MDCdc37 on the opposite face. N- terminal to the PK N- lobe is the TM region, the density for which was unob- served in our reconstructions. C- terminal to the PK C- lobe, we observe some poorly resolved density for the likely mobile dimerization domain in our unsharpened map. This would precede the GC domain, which is not observed in the density of our reconstructions (Figure 1B). Together, we can use our understanding of mGC topology and our reconstruction to orient the complex as it would sit on a membrane (Figure 1B), providing insight into how Hsp90 is able to access and regulate membrane protein clients. No density is observed for the second GC- C of the dimer, though it is sterically unlikely that an additional regulatory complex is forming on the second GC- C in a concurrent fashion, given the large size of the first Hsp90–Cdc37 and the requisite proximity of the second GC- C. In addi- tion, this disruption of the native state of GC- C, as observed in our structure, would likely leave GC domains out of each other’s proximity, precluding their catalytic activity while Hsp90 is bound. Cdc37 mediated GC-C recruitment and Hsp90 loading Despite the recognized plasticity of Cdc37 co- chaperone binding to approximately 60% of kinases (Taipale et  al., 2012), the importance of the Hsp90–Cdc37 complex for pseudokinase domain- containing proteins in the human proteome is not well studied. Thus, the structural basis for how Cdc37 can recruit GC- C to the Hsp90 regulatory complex is of particular interest. In our structures, we see that Cdc37 is displacing the N- lobe of the pseudokinase domain of GC- C, binding to the C- lobe at the N–C interface, and guiding the unfolded N- lobe into the Hsp90 dimer (Figure 2). The Cdc37–GC- C interface is relatively modest in size, with a calculated mean surface area of 689 Å2 (as calculated by PISA Krissinel and Henrick, 2007). This interface is partly driven to form via charge complementarity, with positive contributions from a cluster of arginine residues on Cdc37 (R30, R32, R39) at the periphery of the interaction interface interacting with D609 and the polar residues Y580 and T586 (Figure 2B). Beyond this, the interface is likely largely driven via shape- complementarity, due to a minimal contribution from hydrogen bonding, salt- bridge formation, and aromatic packing contributions – in line with the ability of Cdc37 to chaperone such a diverse array of clients and client sequences. As the unfolded PK N- lobe extends away from Cdc37, it enters the channel formed at the interface between the dimer of Hsp90 (Figure 2C). Here, GC- C residues 528–544 (VKLDTMIFGVIEYCERG) lie across the upper region of the Hsp90 CTDs, which form the floor of the channel. These CTDs form the bulk of the interaction interface as the unfolded N- lobe passes through this channel, yet there are minor contributions from the loop regions of the β-sheet from the MDHsp90 which extend downward into this channel region. The unfolded region is relatively poorly resolved in the density, with some reconstructions from earlier refinement having no resolvable density in this channel region – indicative of the low stability and high mobility of the unfolded N- lobe as it passes through this region. Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 4 of 12 Biochemistry and Chemical Biology Research article A Hsp90 Hsp90 NTD Hsp90 NTD Cdc37 coiled- coil B Cdc37 R39 B GC-C PK C-lobe MD C MD T586T586 L581 T608 Y580 K585 D609D609 W31W31 R30 R30 R32R32 L28L28 I23I23 E18E18 N22N22 H20H20 K606 V604V604 K576K576 GC-C E Y C EE II R G544G544 GC-CGC-C Hsp90 Hsp90 CTD CTD C Hsp90 Hsp90 V528V528 D TT KK LL I M G V F Figure 2. Cdc37 mediated guanylyl cyclase C (GC- C) recruitment and heat shock protein 90 (Hsp90) loading interfaces. (A) Ribbon representation of a model of GC- C–Hsp90–Cdc37 complex. GC- C is colored in red, Hsp90 in blue and teal, and Cdc37 in purple. Pseudokinase (PK), coil- coiled, middle (MD), C- terminal (CTD), and N- terminal (NTD) domains are labeled. (B) The Cdc37–GC- C interface in ribbon representation, with interacting residues drawn in sticks, colored as in A. (C) The unfolded N- lobe of GC- C PK domain as it passes between the Hsp90 dimer, in ribbon representation, with interacting residues drawn in sticks, colored as in A and B. This region’s sequence is: VKLDTMIFGVIEYCERG. The online version of this article includes the following figure supplement(s) for figure 2: Figure supplement 1. Conservation of Cdc37 mediated heat shock protein 90 (Hsp90) regulation. Figure supplement 2. Regulatory mechanisms for membrane receptor guanylyl cyclase (mGC) activity. Conservation of Cdc37 mediated Hsp90 regulation The core structural principles of Cdc37 mediated client recruitment to Hsp90 appear to remain constant across its large range of client diversity. Across other clients–Hsp90–Cdc37 complexes with canonical soluble kinase clients (Cdk4, RAF1, B- raf) (García- Alonso et al., 2022; Oberoi et al., 2022; Verba et al., 2016), we see a conserved role for Cdc37 in client recruitment by associating with the C- lobe at the N-, C- lobe interface (Figure 2—figure supplement 1A, B). In these complexes, we see high levels of structural conservation for the Hsp90–Cdc37 (Cα RMSDs of 1.4–3.3 Å for Hsp90 and 1.5–2.5 Å for Cdc37), while the client is structurally most homogenous at the interface with Cdc37, though less structurally conserved overall (Cα RMSDs of 3.5–11.6 Å). Perhaps unsurprisingly, GC- C is one of the most divergent of these clients from a sequence perspective (Figure 2—figure supplement 1C), with sequence homology between the GC- C PK domain and the other client kinase domains ranging from 19 to 25% identity and 31 to 41% homology. This highlights the plasticity required of this system which can service such a vast array of clients across a broad range of sequence variations, yet more restricted fold architecture. Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 5 of 12 Biochemistry and Chemical Biology Research article Discussion The present cryo- EM structure of GC- C–Hsp90–Cdc37 resolves the loading of GC- C, via its PK domain and interaction with Cdc37, to the Hsp90 core dimer (Figures 1 and 2). This complex shows signifi- cant structural similarity to the mechanism that regulates soluble active kinases (García- Alonso et al., 2022; Oberoi et al., 2022; Verba et al., 2016) and presumably membrane receptor kinases in the human proteome. This structural and mechanistic conservation is largely driven by the co- chaperone Cdc37, which serves as the central binding platform for these clients by associating to the fold of the kinase (or pseudokinase in the case of mGC) domain, relatively independent of sequence identity. A model whereby recruitment is largely driven by both the fold complementarity and the specific stability properties of the kinase fold has been proposed previously (Taipale et  al., 2012). In this model, instability of a fully folded kinase domain results in partial unfolding of the C- lobe, leading Cdc37 to bind the partially unfolded state. Given the lack of functional and sequence conservation for GC- C as a client of Cdc37, our data largely fits with this model for client recruitment. It is likely that the pseudokinase domains of mGC have largely evolved to facilitate regulatory mechanisms for these receptors, both via their phosphorylation and by hijacking the regulatory mechanisms used by active soluble and membrane receptor kinases. In the case of GC- A, previous work has shown that it associates with the Hsp90–Cdc37 complex to regulate GC activity (Kumar et  al., 2001). The authors showed that adding geldanamycin, an Hsp90 inhibitor, reduces the overall cGMP output of cells in response to ANP stimulation while also reducing the association of the Hsp90 to GC- A. While this initially may seem counterintuitive, this data fits with a model of ligand- induced activity potentiating the instability of the PK domain, which then facilitates binding of the regulatory complex to ‘re- fold’ GC- A for further catalysis and cGMP produc- tion – in a core regulatory complex structurally similar to that which we observe for GC- C in this work (Figure 2—figure supplement 2). In the case of the Hsp90 inhibitor, this would release the Hsp90 and only allow full catalytic activity for the receptor until the receptor falls into the partially unfolded state, as the Hsp90 would no longer be able to re- engage at the C- lobe when inhibited (Figure 2—figure supplement 2). Interestingly there may be an additional layer of regulation involved, with crosstalk between the phosphorylation and Hsp90 regulatory mechanisms of mGC. The phosphatase PP5 is known to interact with the Hsp90–Cdc37 system and dephosphorylate Hsp90, Cdc37, and the system’s kinase clients (Oberoi et  al., 2022). PP5 has been implicated in this role for mGC (Chinkers, 1994), though this interaction was unable to be detected by a pull- down in a second study (Kumar et al., 2001). In this way, mGC association with the Hsp90–Cdc37 complex could result in multiple fates and resultant activity profiles for the receptor. When the PK of an activated mGC falls into a destabilized state, this would result in the recruitment of the Hsp90–Cdc37. First, the regula- tory complex could refold the receptor to maintain the activity of the receptor (Figure 2—figure supplement 2i). In another scenario, the Hsp90–Cdc37 complex could additionally recruit PP5 to dephosphorylate the mGC (Figure  2—figure supplement 2ii). Particularly in the case of GC- A and GC- B, and to some extent GC- C (Potter and Garbers, 1992; Potter and Hunter, 1998; Vaandrager et al., 1993), this would impair the signaling activity of the mGC, though this could be rescued through the kinase re- association and phosphorylation. In a final scenario, the binding of the Hsp90–Cdc37 complex could result in the association of ubiquitin E3 ligases (Schopf et al., 2017; Figure 2—figure supplement 2iii), which would ubiquitinate the mGC client, leading to the removal of the receptor. The regulation of mGC is influenced by a network of factors working in harmony to ensure proper signaling and physiological response for these important receptors. The structure of the core regu- latory complex shown in this work is key to many facets of mGC regulation. We hope that the struc- tural basis for the Hsp90 regulatory platform for mGC will drive renewed investigation into these diverse mechanisms and lead to the therapeutic manipulation of these mechanisms to improve mGC targeting therapies. Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 6 of 12 Biochemistry and Chemical Biology Research article Key resources table Reagent type (species) or resource Methods Designation Source or reference Identifiers Additional information Cell line (Cricetulus griseus) Chinese hamster ovary kidney cells GIBCO ExpiCHO Recombinant DNA reagent pD649- GCN4- TM- GC- C_ICD (plasmid) This paper See: Methods - Cloning and protein expression Software, algorithm Data collection software SerialEM SerialEM Software, algorithm Data processing software Software, algorithm Data sharpening software Structura Biotechnology Inc. cryoSPARC Sanchez- Garcia et al., 2021 DeepEMhancer Software, algorithm Initial modeling software Jumper et al., 2021 AlphaFold Software, algorithm Graphics software Pettersen et al., 2021 UCSF ChimeraX Software, algorithm Software, algorithm Modeling and refinement software Adams et al., 2010 Phenix Modeling and refinement software Emsley and Cowtan, 2004 Coot Software, algorithm Model validation software Chen et al., 2010 MolProbity Cloning and protein expression For cryo- EM studies, a construct containing an HA secretion signal (MKTIIALSYIFCLVFA), a FLAG peptide (DYKDDDD), linker and 3  C cleavage site (KGSLEVLFQGPG), GCN4 homodimeric zipper ( RMKQ LEDK VEEL LSKN YHLE NEVA RLKK LVGER), human GC- C regions corresponding to the small extracellular linker region, TM, and intracellular domains (residues 399–1,053), a second linker and 3 C cleavage site (AAALEVLFQGPGAA), a Protein C epitope tag (EDQVDPRLIDGK), and an 8 x His tag were cloned into a pD649 mammalian expression vector. This construct contains all domains of the native GC- C, with the exception of the ECD (Supplementary file 1). Protein was expressed using ExpiCHO Expression System Kit (Thermo Fisher). Briefly, ExpiCHO cells were maintained in ExpiCHO Expression Media at 37  °C with 5% CO2 and gentle agitation, and transiently transfected by the expression construct and cultured according to the manufacturer’s protocol. Cells were pelleted and stored at –80 °C. Protein purification Cells were resuspended in 20 mM HEPES- Na pH 8.0, 300 mM NaCl, 1 mM TCEP, protease inhibitor cocktail (Sigma), and benzonase (Sigma). Cells were lysed by Dounce homogenizer and cellular debris was pelleted by low- speed centrifugation at 500 × g. Membranes were collected by centrifugation at 46,000 × g and stored at –80 °C until use. Membranes were thawed and solubilized with the addition of 1% n- dodecyl β-D- maltoside (DDM) and 0.1% cholesteryl hemisuccinate (CHS) (10:1) (Anatrace). Debris and unsolubilized membranes were pelleted by centrifugation at 46,000 × g. The superna- tant was subsequently used in FLAG affinity chromatography. The supernatant was applied to M1 anti- FLAG resin. The resin was washed with 20 bed volumes of 20 mM HEPES- Na pH 8.0, 300 mM NaCl, 1 mM TCEP, 0.005% lauryl maltose neopentyl glycol (LMNG), 0.0005% CHS (10:1) (Anatrace), and 5  mM ATP. The protein complex was eluted with the addition of 200  μg/mL of FLAG peptide (DYKDDDD) (GenScript). Protein was subsequently concentrated to >2 mg/mL and used for cryo- EM imaging. Cryo-electron microscopy Aliquots of 3 μL of complex were applied to glow- discharged 300 mesh UltrAuFoil (1.2/1.3) grids. The grids were blotted for 3 s at 100% humidity with an offset of 3 and plunged frozen into liquid ethane using a Vitrobot Mark IV (Thermo Fisher). Grid screening and dataset collection occurred at Stanford cEMc on a 200 kV Glacios microscope (Thermo Fisher) equipped with a K3 camera (Gatan). Movies Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 7 of 12 Biochemistry and Chemical Biology Research article Table 1. Cryo- EM data collection, refinement, and validation statistics. GC- C–Hsp90–Cdc37 complex PDB 8FX4 EMD- 29523 GC- C–Hsp90–Cdc37 complex with DD density Data collection and processing Nominal magnification Acceleration voltage (kV) Electron exposure (e-/Å2) Defocus range (µm) Pixel size (Å) Symmetry imposed Final particle images 165,635 Map resolution FSC threshold Map resolution (Å) 3.9 Refinement Initial model used (PDB) 5FWK, 7ZR5, AlphaFold 45,000 200 58.8 0.8–2.0 0.9273 C1 0.143 48,283 6.3 Model resolution FSC threshold (Å) Model resolution (Å) Model Composition Non- hydrogen atoms Protein residues Ligands B- factors (Å2) Protein Ligand R.m.s. deviations Bond lengths (Å) Bond angles (°) Validation MolProbity score Clashscore Rotamer outliers (%) Ramachandran plot Favored (%) Allowed (%) Outliers (%) 0.5 4.2 13,478 1,654 2 119.49 102.85 0.004 0.914 2.14 13.88 0.67 92.0 7.6 0.4 Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 8 of 12 Biochemistry and Chemical Biology Research article were collected at a magnification corresponding to a 0.9273 Å per physical pixel. The dose was set to a total of 58.8 electrons per Å2. Automated data collection was carried out using SerialEM with a nominal defocus range set from –0.8 to –2.0 μM. Image processing All processing was performed in cryoSPARC (Punjani et al., 2017) unless otherwise noted (Figure 1— figure supplement 1). 8788 movies were motion- corrected using patch motion correction. The contrast transfer functions (CTFs) of the flattened micrographs were determined using patch CTF and an initial stack of particles was picked using Topaz picker (Bepler et al., 2019). Successive rounds of reference- free 2D classification were performed to generate a particle stack of 165,635 particles. These particles were then used in ab- initio reconstruction, followed by non- uniform refinement (Punjani et al., 2020) and finally local refinement with a loose mask around the entire complex. This resulted in a 3.9 Å reconstruction of the GC- C–Hsp90–Cdc37 complex which was sharpened with deepEMhancer (Sanchez- Garcia et al., 2021). These particles were also used in a 4- class heterogeneous refinement to pull out a volume containing some resolved density for the dimerization domain of GC- C. Model building and refinement The Cdk4–Hsp90β–Cdc37 (PDB 5FWK), PP5–B- Raf–Hsp90β–Cdc37 (PDB 7ZR5), and AlphaFold models for GC- C (Jumper et al., 2021; Mirdita et al., 2022) were docked into the map using UCSF Chimera X (Pettersen et al., 2021). A resultant hybrid model was then manually curated to contain the correct Cricetulus griseus sequences for Hsp90β–Cdc37 and run through Namdinator (Kidmose et  al., 2019). This was followed by automated refinement using Phenix real space refine (Adams et al., 2010) and manual building in Coot (Emsley and Cowtan, 2004). The final model produced a favorable MolProbity score of 2.14 (Chen et al., 2010) with 0.4% Ramachandran outliers (Table 1). Model building and refinement software was installed and configured by SBGrid (Morin et al., 2013). Acknowledgements We thank Liz Montabana and Stanford cEMc for microscope access for data collection. We thank Paul LaPointe and Kevin Jude for their insightful discussion of the Hsp90 structure and regulatory mech- anisms. NAC is a CIHR postdoctoral fellow. KCG is an investigator with the Howard Hughes Medical Institute. KCG is supported by National Institutes of Health grant R01- AI51321, the Mathers Founda- tion, and the Ludwig Foundation. Additional information Funding Funder Canadian Institutes of Health Research National Institutes of Health Mathers Foundation Ludwig Foundation Grant reference number Author Postdoctoral Fellowship Nathanael A Caveney R01-AI51321 K Christopher Garcia K Christopher Garcia K Christopher Garcia The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Author contributions Nathanael A Caveney, Conceptualization, Formal analysis, Investigation, Methodology, Writing - orig- inal draft, Writing – review and editing; Naotaka Tsutsumi, Formal analysis, Investigation, Writing – review and editing; K Christopher Garcia, Supervision, Funding acquisition, Project administration, Writing – review and editing Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 9 of 12 Biochemistry and Chemical Biology Research article Author ORCIDs Nathanael A Caveney Naotaka Tsutsumi K Christopher Garcia http://orcid.org/0000-0003-4828-3479 https://orcid.org/0000-0002-3617-7145 https://orcid.org/0000-0001-9273-0278 Peer review material Reviewer #1 (Public Review): https://doi.org/10.7554/eLife.86784.3.sa1 Reviewer #2 (Public Review): https://doi.org/10.7554/eLife.86784.3.sa2 Reviewer #3 (Public Review): https://doi.org/10.7554/eLife.86784.3.sa3 Author Response https://doi.org/10.7554/eLife.86784.3.sa4 Additional files Supplementary files • MDAR checklist • Supplementary file 1. Plasmids used in this study. Data availability Cryo- EM maps and atomic coordinates for the GC- C- Hsp90- Cdc37 complex have been deposited in the EMDB (EMD- 29523) and PDB (8FX4). Material availability: The plasmids used in this study are uploaded in (Supplementary file 1). The following datasets were generated: Author(s) Caveney NA, Garcia CK Year 2023 Caveney NA, Garcia KC 2023 Dataset title Dataset URL Database and Identifier Cryo- EM maps and atomic coordinates for the GC- C- Hsp90- Cdc37 complex have been deposited Cryo- EM maps and atomic coordinates for the GC- C- Hsp90- Cdc37 complex have been deposited https://www. ebi. ac. uk/ emdb/ EMD- 29523 EMDataResource, EMD- 29523 https://www. rcsb. org/ structure/ 8FX4 RCSB Protein Data Bank, 8FX4 References Adams PD, Afonine PV, Bunkóczi G, Chen VB, Davis IW, Echols N, Headd JJ, Hung LW, Kapral GJ, Grosse- Kunstleve RW, McCoy AJ, Moriarty NW, Oeffner R, Read RJ, Richardson DC, Richardson JS, Terwilliger TC, Zwart PH. 2010. PHENIX: A comprehensive python- based system for macromolecular structure solution. Acta Crystallographica. Section D, Biological Crystallography 66:213–221. DOI: https://doi.org/10. 1107/S0907444909052925, PMID: 20124702 Aiyer S, Zhang C, Baldwin PR, Lyumkis D. 2021. Evaluating local and directional resolution of cryo- EM density maps. Methods in Molecular Biology 2215:161–187. DOI: https://doi.org/10.1007/978-1-0716-0966-8_8, PMID: 33368004 Bepler T, Morin A, Rapp M, Brasch J, Shapiro L, Noble AJ, Berger B. 2019. Positive- unlabeled convolutional neural networks for particle picking in cryo- electron micrographs. Nature Methods 16:1153–1160. DOI: https:// doi.org/10.1038/s41592-019-0575-8, PMID: 31591578 Chen VB, Arendall WB, Headd JJ, Keedy DA, Immormino RM, Kapral GJ, Murray LW, Richardson JS, Richardson DC. 2010. MolProbity: all- atom structure validation for macromolecular crystallography. Acta Crystallographica. Section D, Biological Crystallography 66:12–21. DOI: https://doi.org/10.1107/ S0907444909042073, PMID: 20057044 Chinkers M. 1994. Targeting of a distinctive protein- serine phosphatase to the protein kinase- like domain of the atrial natriuretic peptide receptor. PNAS 91:11075–11079. DOI: https://doi.org/10.1073/pnas.91.23.11075, PMID: 7972012 Emsley P, Cowtan K. 2004. Coot: model- building tools for molecular graphics. Acta Crystallographica. Section D, Biological Crystallography 60:2126–2132. DOI: https://doi.org/10.1107/S0907444904019158, PMID: 15572765 García- Alonso S, Mesa P, de la Puente Ovejero L, Aizpurua G, Lechuga CG, Zarzuela E, Santiveri CM, Sanclemente M, Muñoz J, Musteanu M, Campos- Olivas R, Martínez- Torrecuadrada J, Barbacid M, Montoya G. 2022. Structure of the RAF1- HSP90- CDC37 complex reveals the basis of RAF1 regulation. Molecular Cell 82:3438–3452. DOI: https://doi.org/10.1016/j.molcel.2022.08.012, PMID: 36055235 Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 10 of 12 Biochemistry and Chemical Biology Research article Glassman CR, Tsutsumi N, Saxton RA, Lupardus PJ, Jude KM, Garcia KC. 2022. Structure of a Janus kinase cytokine receptor complex reveals the basis for dimeric activation. Science 376:163–169. DOI: https://doi.org/ 10.1126/science.abn8933, PMID: 35271300 He XI, Chow DC, Martick MM, Garcia KC. 2001. Allosteric activation of a spring- loaded natriuretic peptide receptor dimer by hormone. Science 293:1657–1662. DOI: https://doi.org/10.1126/science.1062246, PMID: 11533490 He X, Dukkipati A, Garcia KC. 2006. Structural determinants of natriuretic peptide receptor specificity and degeneracy. Journal of Molecular Biology 361:698–714. DOI: https://doi.org/10.1016/j.jmb.2006.06.060 Jumper J, Evans R, Pritzel A, Green T, Figurnov M, Ronneberger O, Tunyasuvunakool K, Bates R, Žídek A, Potapenko A, Bridgland A, Meyer C, Kohl SAA, Ballard AJ, Cowie A, Romera- Paredes B, Nikolov S, Jain R, Adler J, Back T, et al. 2021. Highly accurate protein structure prediction with AlphaFold. Nature 596:583–589. DOI: https://doi.org/10.1038/s41586-021-03819-2, PMID: 34265844 Kidmose RT, Juhl J, Nissen P, Boesen T, Karlsen JL, Pedersen BP. 2019. Namdinator - automatic molecular dynamics flexible fitting of structural models into cryo- EM and crystallography experimental maps. IUCrJ 6:526–531. DOI: https://doi.org/10.1107/S2052252519007619, PMID: 31316797 Kobayashi D, Yamaguchi N, Takahashi O, Deshpande GA, Fukui T. 2012. Human atrial natriuretic peptide treatment for acute heart failure: A systematic review of efficacy and mortality. The Canadian Journal of Cardiology 28:102–109. DOI: https://doi.org/10.1016/j.cjca.2011.04.011, PMID: 21908161 Krissinel E, Henrick K. 2007. Inference of macromolecular assemblies from crystalline state. Journal of Molecular Biology 372:774–797. DOI: https://doi.org/10.1016/j.jmb.2007.05.022, PMID: 17681537 Kumar R, Grammatikakis N, Chinkers M. 2001. Regulation of the atrial natriuretic peptide receptor by heat shock protein 90 complexes. The Journal of Biological Chemistry 276:11371–11375. DOI: https://doi.org/10.1074/ jbc.M010480200, PMID: 11152473 Miner PB. 2020. Plecanatide for the treatment of constipation- predominant irritable bowel syndrome. Expert Review of Gastroenterology & Hepatology 14:71–84. DOI: https://doi.org/10.1080/17474124.2020.1722101, PMID: 31985305 Mirdita M, Schütze K, Moriwaki Y, Heo L, Ovchinnikov S, Steinegger M. 2022. ColabFold: making protein folding accessible to all. Nature Methods 19:679–682. DOI: https://doi.org/10.1038/s41592-022-01488-1, PMID: 35637307 Morin A, Eisenbraun B, Key J, Sanschagrin PC, Timony MA, Ottaviano M, Sliz P. 2013. Collaboration gets the most out of software. eLife 2:e01456. DOI: https://doi.org/10.7554/eLife.01456, PMID: 24040512 Oberoi J, Guiu XA, Outwin EA, Schellenberger P, Roumeliotis TI, Choudhary JS, Pearl LH. 2022. HSP90- CDC37- PP5 forms a structural platform for kinase dephosphorylation. Nature Communications 13:7343. DOI: https:// doi.org/10.1038/s41467-022-35143-2, PMID: 36446791 Ogawa H, Qiu Y, Ogata CM, Misono KS. 2004. Crystal structure of hormone- bound atrial natriuretic peptide receptor extracellular domain: rotation mechanism for transmembrane signal transduction. The Journal of Biological Chemistry 279:28625–28631. DOI: https://doi.org/10.1074/jbc.M313222200, PMID: 15117952 Ogawa H, Qiu Y, Philo JS, Arakawa T, Ogata CM, Misono KS. 2010. Reversibly bound chloride in the atrial natriuretic peptide receptor hormone- binding domain: possible allosteric regulation and a conserved structural motif for the chloride- binding site. Protein Science 19:544–557. DOI: https://doi.org/10.1002/pro.332, PMID: 20066666 Pettersen EF, Goddard TD, Huang CC, Meng EC, Couch GS, Croll TI, Morris JH, Ferrin TE. 2021. UCSF Chimerax: structure visualization for researchers, educators, and developers. Protein Science 30:70–82. DOI: https://doi.org/10.1002/pro.3943, PMID: 32881101 Potter LR, Garbers DL. 1992. Dephosphorylation of the guanylyl cyclase- A receptor causes desensitization. The Journal of Biological Chemistry 267:14531–14534 PMID: 1353076. Potter LR, Hunter T. 1998. Identification and characterization of the major phosphorylation sites of the B- type natriuretic peptide receptor. The Journal of Biological Chemistry 273:15533–15539. DOI: https://doi.org/10. 1074/jbc.273.25.15533, PMID: 9624142 Potter LR. 2011. Guanylyl cyclase structure, function and regulation. Cellular Signalling 23:1921–1926. DOI: https://doi.org/10.1016/j.cellsig.2011.09.001, PMID: 21914472 Punjani A, Rubinstein JL, Fleet DJ, Brubaker MA. 2017. cryoSPARC: algorithms for rapid unsupervised cryo- EM structure determination. Nature Methods 14:290–296. DOI: https://doi.org/10.1038/nmeth.4169, PMID: 28165473 Punjani A, Zhang H, Fleet DJ. 2020. Non- uniform refinement: adaptive regularization improves single- particle cryo- EM reconstruction. Nature Methods 17:1214–1221. DOI: https://doi.org/10.1038/s41592-020-00990-8, PMID: 33257830 Sanchez- Garcia R, Gomez- Blanco J, Cuervo A, Carazo JM, Sorzano COS, Vargas J. 2021. DeepEMhancer: a deep learning solution for cryo- EM volume post- processing. Communications Biology 4:874. DOI: https://doi. org/10.1038/s42003-021-02399-1, PMID: 34267316 Schopf FH, Biebl MM, Buchner J. 2017. The Hsp90 chaperone machinery. Nature Reviews. Molecular Cell Biology 18:345–360. DOI: https://doi.org/10.1038/nrm.2017.20, PMID: 28429788 Taipale M, Krykbaeva I, Koeva M, Kayatekin C, Westover KD, Karras GI, Lindquist S. 2012. Quantitative analysis of Hsp90- client interactions reveals principles of substrate recognition. Cell 150:987–1001. DOI: https://doi. org/10.1016/j.cell.2012.06.047, PMID: 22939624 Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 11 of 12 Biochemistry and Chemical Biology Research article Vaandrager AB, van der Wiel E, de Jonge HR. 1993. Heat- stable enterotoxin activation of immunopurified guanylyl cyclase C. Modulation by adenine nucleotides. The Journal of Biological Chemistry 268:19598–19603 PMID: 8103520. van den Akker F, Zhang X, Miyagi M, Huo X, Misono KS, Yee VC. 2000. Structure of the dimerized hormone- binding domain of a guanylyl- cyclase- coupled receptor. Nature 406:101–104. DOI: https://doi.org/10.1038/ 35017602, PMID: 10894551 Verba KA, Wang RYR, Arakawa A, Liu Y, Shirouzu M, Yokoyama S, Agard DA. 2016. Atomic structure of Hsp90- Cdc37- Cdk4 reveals that Hsp90 traps and stabilizes an unfolded kinase. Science 352:1542–1547. DOI: https:// doi.org/10.1126/science.aaf5023, PMID: 27339980 Yu SWB, Rao SSC. 2014. Advances in the management of constipation- predominant irritable bowel syndrome: the role of linaclotide. Therapeutic Advances in Gastroenterology 7:193–205. DOI: https://doi.org/10.1177/ 1756283X14537882, PMID: 25177366 Caveney et al. eLife 2023;12:RP86784. DOI: https://doi.org/10.7554/eLife.86784 12 of 12 Biochemistry and Chemical Biology
10.2196_39479
JMIR HUMAN FACTORS Original Paper White et al Understanding the Subjective Experience of Long-term Remote Measurement Technology Use for Symptom Tracking in People With Depression: Multisite Longitudinal Qualitative Analysis Katie M White1, BSc; Erin Dawe-Lane2, MSc; Sara Siddi3, PhD; Femke Lamers4, PhD; Sara Simblett2, PhD; Gemma Riquelme Alacid3, MSc; Alina Ivan1, MSc; Inez Myin-Germeys5, PhD; Josep Maria Haro3, PhD; Carolin Oetzmann1, MSc; Priya Popat1, BSc; Aki Rintala5, PhD; Elena Rubio-Abadal3, PhD; Til Wykes2, PhD; Claire Henderson6, PhD; Matthew Hotopf1, PhD; Faith Matcham1,7, PhD 1Department of Psychological Medicine, King's College London, London, United Kingdom 2Department of Psychology, King's College London, London, United Kingdom 3Parc Sanitari Sant Joan de Déu, Fundació Sant Joan de Déu, CIBERSAM, Universitat de Barcelona, Barcelona, Spain 4Department of Psychiatry, Amsterdam University Medical Center, Vrije Universiteit, Amsterdam, Netherlands 5Center for Contextual Psychiatry, Department of Neurosciences, UK Leuven, Leuven, Belgium 6Health Service & Population Research Department, King's College London, London, United Kingdom 7School of Psychology, University of Sussex, Falmer, Sussex, United Kingdom Corresponding Author: Katie M White, BSc Department of Psychological Medicine King's College London Institute of Psychiatry, Psychology and Neuroscience 16 de Crespigny Park London, SE5 8AB United Kingdom Phone: 44 7850684847 Email: katie.white@kcl.ac.uk Abstract Background: Remote measurement technologies (RMTs) have the potential to revolutionize major depressive disorder (MDD) disease management by offering the ability to assess, monitor, and predict symptom changes. However, the promise of RMT data depends heavily on sustained user engagement over extended periods. In this paper, we report a longitudinal qualitative study of the subjective experience of people with MDD engaging with RMTs to provide insight into system usability and user experience and to provide the basis for future promotion of RMT use in research and clinical practice. Objective: We aimed to understand the subjective experience of long-term engagement with RMTs using qualitative data collected in a longitudinal study of RMTs for monitoring MDD. The objectives were to explore the key themes associated with long-term RMT use and to identify recommendations for future system engagement. In this multisite, longitudinal qualitative research study, 124 semistructured interviews were conducted with 99 Methods: participants across the United Kingdom, Spain, and the Netherlands at 3-month, 12-month, and 24-month time points during a study exploring RMT use (the Remote Assessment of Disease and Relapse-Major Depressive Disorder study). Data were analyzed using thematic analysis, and interviews were audio recorded, transcribed, and coded in the native language, with the resulting quotes translated into English. Results: There were 5 main themes regarding the subjective experience of long-term RMT use: research-related factors, the utility of RMTs for self-management, technology-related factors, clinical factors, and system amendments and additions. Conclusions: The subjective experience of long-term RMT use can be considered from 2 main perspectives: experiential factors (how participants construct their experience of engaging with RMTs) and system-related factors (direct engagement with the technologies). A set of recommendations based on these strands are proposed for both future research and the real-world https://humanfactors.jmir.org/2023/1/e39479 XSL•FO RenderX JMIR Hum Factors 2023 | vol. 10 | e39479 | p. 1 (page number not for citation purposes) JMIR HUMAN FACTORS White et al implementation of RMTs into clinical practice. Future exploration of experiential engagement with RMTs will be key to the successful use of RMTs in clinical care. (JMIR Hum Factors 2023;10:e39479) doi: 10.2196/39479 KEYWORDS remote measurement; technology; qualitative; engagement; telehealth; depression; mental health; mobile phone Introduction (MDD) Background Depressive disorders, characterized by periods of persistent low mood and anhedonia, are the third leading cause of disability worldwide is [1]. Major depressive disorder characterized by a longitudinal trajectory of relapse and remission [2]. The economic burden of MDD is currently estimated at US $326 billion [3], with high recurrence associated with increased comorbidity burden and health care resource use [4]. Traditional assessment of MDDs is limited in its ability to detect moment-by-moment symptom changes because it relies on retrospective questionnaires completed at sporadic time points, is prone to recall bias, and is often only undertaken at the point of relapse [5]. Working toward the timely diagnosis and treatment of MDD remains an urgent priority [5]. Novel remote measurement technologies (RMTs) have the potential to become an asset for chronic disease management. Multiparametric RMT systems can provide real-time, longitudinal symptom tracking by combining active symptom reporting via smartphone apps (active RMT) with physiological and behavioral wearable sensor data (passive RMT) [6]. Continuous data can be collected on mood variability [7], sociability [8], physical activity [9], cognition [10], speech acoustics [11], and sleep [12]. Integration of RMT data into MDD care may help to more accurately assess, monitor, and predict depressive symptom trajectories, ultimately enabling personalized interventions [13]. and The promise of remote tracking in MDD depends almost entirely on user engagement. Engagement with mobile health (mHealth) technologies comprises the initial and sustained active use of a device [14]. High engagement with RMTs is imperative given the high-frequency data needed to identify symptom patterns and changes over time. Several systematic reviews have highlighted the heterogeneity of engagement metrics reported in remote tracking studies [15-17]. The Remote Assessment of Disease Relapse-Major Depressive Disorder (RADAR-MDD) study is currently the largest multisite longitudinal study of a multiparametric RMT system for tracking depression [6]. The RADAR-MDD study has recently reported promising engagement, both in terms of initial recruitment rates [18] and sustained retention and data availability [19] over a 2-year follow-up of 623 participants across 3 European sites (United Kingdom, Spain, and the Netherlands). A large proportion of participants (79.8%) completed follow-up, and approximately 50% of the participants had >76% data completion for passive data streams [19]. When evaluating engagement, an understanding of the subjective experience of using RMTs should complement objective data https://humanfactors.jmir.org/2023/1/e39479 XSL•FO RenderX completion statistics [17]. Subjective engagement with mHealth technologies can be understood as an experiential construct of what it feels like [20]. Exploring subjective engagement with RMTs provides a richer insight into system usability and perceived utility of, and satisfaction with, the technology [17]. The drivers for sustained user engagement with RMT systems, which, in contrast to typical mHealth technologies, require long periods of use for little direct rewards or intervention [21], are currently unknown. Several studies have qualitatively explored subjective engagement with RMTs for depression. A multisite exploration of the perceived barriers and facilitators to RMT use by Simblett et al [22] informed the design of the RADAR-MDD study. Functional (technological convenience, accessibility, and intrusiveness) and nonfunctional (user cognition, perceived rewards) factors influenced patients when considering remote symptom tracking [22]. These findings have been replicated across patient and physician perspectives [23-25]. Two systematic reviews [26,27] on broader mHealth technologies for depression explored the experiences of participants’ actual use for up to 1 year. Factors such as lower symptom severity, perceived usefulness of the technology, lower privacy concerns, lack of technical issues, and access to responsive personal support were associated with enhanced motivation to engage with technologies [26,27]. A handful of studies have also suggested the beneficial effects of symptom monitoring, including [28], adaptation of self-management strategies [29], and access to a “safety net” of support [30]. However, typically use hypothetical scenarios or evaluate short-term system use. As a result, little is known about the subjective experience of long-term, real-world use of RMTs. increased self-awareness these studies Objective This study aims to understand the subjective experience of long-term engagement with RMTs for monitoring depression symptoms. It uses qualitative data from the RADAR-MDD study as an example of sustained RMT use across a 2-year follow-up period. This study builds on previous qualitative work by Simblett et al [22] on perceived barriers to and facilitators of intended RMT use in depression, providing a comparison with user experiences over 2 years of sustained engagement. Our objectives were (1) to explore key themes associated with long-term RMT use and (2) to identify recommendations for future system engagement. The findings will complement the objective engagement data and provide a basis for further promotion of engagement with RMTs for symptom tracking in research and clinical practice. JMIR Hum Factors 2023 | vol. 10 | e39479 | p. 2 (page number not for citation purposes) JMIR HUMAN FACTORS Methods Design This study used a multisite longitudinal qualitative research [31] approach with thematic analysis. Semistructured interviews were conducted with participants at 3-, 12-, and 24-month time points at 3 RADAR-MDD sites: King’s College London (London, United Kingdom), Centro de Investigación Biomédica en Red (Barcelona, Spain), and Amsterdam University Medical Centre (Amsterdam, the Netherlands). The design of the interview topic guide was informed by recent work on the barriers to and facilitators of RMT use in those living with depression [16,22]. Procedure The RADAR-MDD study used the RADAR-base system [32] for data collection. The study active RMT smartphone app fortnightly validated mood and self-esteem delivered questionnaires and 6-weekly, high-frequency experience sampling methodology (ESM) questionnaires on current state, cognitive games, and a speech task. The study passive RMT smartphone app collected passive data on ambient noise and light, Bluetooth connection, and GPS location. Participants were provided with a wearable device, the Fitbit Charge (Fitbit Inc), measuring their step count, sleep, and physical activity. Further information on the RADAR-MDD procedure is available in the protocol paper by Matcham et al [6]. Eligibility criteria for inclusion in this study were (1) current participation in RADAR-MDD (full eligibility criteria provided in the study by Matcham et al [6]) and (2) willingness to participate in a 1:1 interview with a researcher discussing their experiences of the study. Participants provided written informed consent for the interviews as part of their RADAR-MDD study participation. The interviews were managed by the research team lead at each site. Participants were recruited using convenience sampling at each time point to maximize data collection. Interviews were face-to-face (at the respective research site) or via telephone or video call (United Kingdom and the Netherlands only). All interviewers were female and part of the participant-facing research team. Face-to-face interviews were not conducted during the COVID-19 pandemic lockdown. Participants were reimbursed for relevant travel costs and paid per interview (£10 or €10 [US $1.2]). The interviews were semistructured using open-ended questions, designed to elicit discussions around using the study technology in daily life (Multimedia Appendix 1). The content of each topic guide reflected the expected differences between time points. For example, the 3-month guide focused on immediate problem-solving and troubleshooting, where later interviews included data sharing. White et al The topic guides were translated from English into Spanish and Dutch, and interviews were conducted by native speakers at each site. The interviews lasted between 30 and 60 minutes and were conducted between February 2018 and April 2021. Ethics Approval The semistructured interviews were approved by the ethics committee of RADAR-MDD [6]. Ethical approvals for conducting the study were obtained from Camberwell St Giles Research Ethics Committee (reference: 17/LO/1154) in London, from Clinical Research Ethics Committee Fundacio Sant Joan de Déu (CI: PIC-128-17) in Barcelona, and from Medische Ethische Toetsingscommissie VUms (2018.012–NL63557. 029.17) in the Netherlands. Data Analysis Strategy The interviews were audio recorded and transcribed verbatim. A preliminary coding framework was developed in English based on previous findings of barriers to and facilitators of RMT use in hypothetical scenarios [22]. All sites first coded example interviews for a cross-site consistency check and a discussion on revisions to the coding framework, accounting for novel codes. Each site then proceeded to recode all interviews in the native language using NVivo software (version 12; QSR International [33]) according to the final coding framework (Multimedia Appendix 2 provides a comparison of the preliminary and final coding framework). The coding was performed by independent researchers at each site. Each site sent coded NVivo data sets to the London site, with all quotes translated into English by a third-party translator briefed on the study topic [34]. The data were stored on a secure server at the London site. Multisite data were merged into one data set and thematic maps for 3-month, 12-month, and 24-month time points were developed by 3 researchers (KW, EDL, and PP), identifying key themes and subthemes. To align with previous longitudinal qualitative research work [31], data are presented not as a longitudinal narrative but as contributing to each theme. Results Participant Characteristics A total of 124 interviews with 99 participants were conducted across 3 sites. Of these 124 interviews, 40 (32.2%) interviews were conducted at the 3-month time point (15/40, 38% in United Kingdom; 15/40, 38% in Spain; and 10/40, 25% in the Netherlands), 42 (33.9%) at the 12-month time point (16/42, 38% at United Kingdom; 16/42, 38% at Spain; 10/42, 24% at the Netherlands), and 42 (33.9%) at the 24-month time point (15/42, 36% at United Kingdom; 16/42, 38% at Spain; 11/42, 26% at the Netherlands). A total of 17 participants took part in an interview at 2 time points; 4 participants were interviewed across all 3 time points. Participant characteristics according to time points are shown in Table 1. https://humanfactors.jmir.org/2023/1/e39479 XSL•FO RenderX JMIR Hum Factors 2023 | vol. 10 | e39479 | p. 3 (page number not for citation purposes) JMIR HUMAN FACTORS White et al Table 1. Participant characteristics by interview time point. Characteristics Site, n United Kingdom Spain the Netherlands Age (years), mean (SD) Female, n (%) Depression severity categorya, n (%) None Mild Moderate Severe Very severe Not reported Anxiety severity categoryb, n (%) None Mild Moderate Severe Not reported Time point 3-month (n=40) 15 15 10 44.6 (12.1) 30 (75) 4 (10) 7 (18) 10 (25) 7 (18) 11 (28) 1 (3) 7 (18) 7 (18) 12 (30) 13 (33) 1 (3) 12-month (n=42) 24-month (n=42) 16 16 10 49.4 (13.5) 32 (76) 3 (7) 5 (12) 13 (31) 10 (24) 9 (21) 2 (5) 5 (12) 10 (24) 13 (31) 12 (329) 2 (4) 15 16 11 51.9 (15.0) 29 (69) 5 (12) 5 (12) 7 (17) 6 (14) 5 (12) 14 (33) 7 (17) 8 (19) 7 (17) 5 (12) 15 (36) aMeasured as the Inventory of Depressive Symptomatology-Self Report total score nearest to the interview time for each participant. None=0-13, mild=14-25, moderate=26-38, severe=39-48, and very severe=49-84. bMeasured as the Generalized Anxiety Disorder-7 item total score nearest to the interview time for each participant. None=0-5, mild=6-10, moderate=11-15, and severe=16-21. Themes This study aimed to explore the subjective experience of long-term engagement with RMTs over a 2-year follow-up period. We present our results under five themes: (1) research-related factors, (2) the utility of RMTs for self-management, (3) technology-related factors, (4) clinical factors, and (5) system amendments and additions. Research-Related Factors When considering initial motivations for engaging with an RMT study, contributing toward novel research findings was the most prevalent reason for taking part. Across all time points, research team support was also a key facilitator of sustained engagement in the study. Altruism and Academia Taking part in the study was an opportunity to use personal experiences of depression to help others, to advance scientific understanding, and to “give back” to the system: I’ve suffered with depression the whole of my adult life, I’ve obviously had a lot out of the system. If I can do anything to put back, do you see what I mean—I will. [P30, 24 months, United Kingdom] https://humanfactors.jmir.org/2023/1/e39479 XSL•FO RenderX Taking part for “the future, for the people who come after me” (P8, 3 months, Spain) was a strong theme that arose in all sites when discussing reasons for enrolling in the research study. Altruistic motivations continued across later time points regardless of whether participants felt they had experienced any direct benefits: I am actually quite proud to say that I am doing this as part of research. Some people will ask me what it is [the wearable], and I say well it is good if more people get to know about it. And for the long-term benefits, might not be for me but for other people, because it might show. [P18, 12 months, United Kingdom] With regard to the RMT aspect of the study, some mentioned that it “piqued my interest” (P37, 24 months, United Kingdom) and “I was very intrigued by a study that kind of has consistent monitoring” (P39, 24 months, United Kingdom). However, many participants signed up with limited knowledge of the study procedure, or of the use of RMTs for health care monitoring. Thus, a lack of prior understanding of RMTs is not a barrier to initial engagement. Privacy was not a barrier to participants upon entering the study or throughout their participation. A key reason for this was that JMIR Hum Factors 2023 | vol. 10 | e39479 | p. 4 (page number not for citation purposes) JMIR HUMAN FACTORS White et al the research was conducted in a clinical and academic setting. In the Spanish cohort, one participant viewed the study as parallel to their clinical care: It’s not data about, about privacy, things about you, no, it’s related to a medical condition, isn’t it? A case of depression, that’s what it’s about. So if they ask you for medical data, well, it’s normal. [P25, 24 months, Spain] Any initial privacy or data security concerns were largely alleviated by the 3-month point through conversations with the research team. At later time points, privacy was not discussed frequently. Research Team Support Support from the research team was a facilitator to continued engagement with the RMTs. This was primarily practical; at 3 months, the research team provided support on how to use the devices and study apps, which was often imperative to successful enrollment into the study: I tried it once [the wearable] and wasn’t able to...to...put it on the phone. If it hadn’t been for [researcher name]’s help I wouldn’t have made it. [P1, 3 months, Spain] The need for practical support remained a key theme at 12 months, this time concerning technological malfunctions. Ability to contact the research team through various methods and receiving a timely reply was important. Some felt comfortable with initiating support themselves: “I didn’t need that much contact personally, I could get in contact easily, if it were necessary” (P21, 24 months, the Netherlands). Others wanted more contact, for example, more points of researcher-initiated contact, or specific contact from specialists. At-hand support was essential for continued participation: I think it is really important to have the practical support ‘cause you don’t want to be offline or not working for long than is necessary. Otherwise it goes against the purpose of the study really. [P18, 12 months, United Kingdom] There was a consensus at all time points that the research team was approachable, patient, and reassuring, helping to alleviate technological concerns. The research team also provided emotional support to the participants. Some participants sought comfort in the knowledge that they were being monitored as part of a study: “I liked it a lot because, jeez knowing, I felt safe, you know? Because knowing that you were there...” (P25, 24 months, Spain). Others had specific examples of receiving mental health support from the research team. One participant in the British cohort received direct signposting, which was noted in both their 12-month and 24-month interview as a crucial part of their study experience: because of the letter from [researcher] to the GP clinic I was able to get an immediate referral, and the problem is if you’re the system it’s great, if you’re not in the system it’s difficult to get in. I couldn’t have done it on my own. [P27, 12 months, United Kingdom] https://humanfactors.jmir.org/2023/1/e39479 XSL•FO RenderX Benefits of RMTs for Self-management Despite primarily engaging with the study for altruistic reasons, many participants experienced unexpected benefits of using RMTs for symptom monitoring during their time in the study. These comprised symptom awareness and communication, both of which were integrated into self-management of depression. Symptom Monitoring and Awareness Across all 3 time points, the most frequently reported benefit was an increase in symptom awareness. Monitoring various factors related to depression, for example, mood, sleep, and exercise, increased self-reflection, and the ability to identify patterns. For example, having access to objective sleep data provided clarification and reassurance: I loved that [the wearable data], I found that so reassuring to just relax, of course you’ve slept and then you go ok, the next time you’re lying in bed you go I’m not ever gonna sleep again but actually you have, you’ve seen that you do I think that’s brilliant, really reassuring. [P14, 3 months, United Kingdom] Although the app did not provide feedback on symptom scores, many felt that the act of answering the questionnaires prompted them to analyze how they had been feeling: I’m more aware of it, the questions on the questionnaire, especially those that ask how I’m feeling right now raise my awareness, I feel quite average or, I’m feeling not great, sometimes you ignore these things. And if you can take more time to think about these things...maybe I need to meditate more, I really feel self-conscious... [P10, 3 months, the Netherlands] For some, answering the questionnaires and viewing the Fitbit data simply provided an understanding of their experience of depression: “I have noticed that my answers have gotten more positive throughout the year” (P22, 24 months, the Netherlands). For others, these data directly motivated behavior changes. At 3 months, the discussion focused on the motivational effects of the Fitbit data; participants felt encouraged to complete their daily step count or achieve target physical activity “badges.” Toward the later time points, these data came to act as prompts for self-care, for example, increased exercise or relaxation: Wearing a watch and knowing that my activity matters, you know? I mean, like the steps I take have a direct effect on my health, both physical and mental, all my activity makes me more aware of it, more conscious of it and it has also been like a driving force for me to put my batteries in sport or stress management...a habit forever, so I do not want to do without it. [P26, 24 months, Spain] This became especially apparent during the 24-month interviews, when the Fitbit data were used to monitor sleep and mood symptom changes during the COVID-19 pandemic. Disruption to usual routines during this time allowed some to reflect more than ever on the benefit of monitoring exercise: I knew in theory, exercising and getting out and so on was good for your mental health, but over Covid, JMIR Hum Factors 2023 | vol. 10 | e39479 | p. 5 (page number not for citation purposes) JMIR HUMAN FACTORS White et al the monitor helped, and the benefit would have been even better. I think I might have been worse during Covid without it. [P36, 24 months, United Kingdom] Communication At each time point, the RMT data were also used for communicating personal experiences to others. Participants used their increased understanding of their depression to inform others: “For the first time it kind of occurred to me to let me partner know when I could feel it was starting...so if you see my behaviour change or I’m unresponsive this is why” (P39, 24 months, United Kingdom). Access to the Fitbit data also facilitated joint decision-making, both for immediate symptom management and long-term strategies: There are also days that I don’t reach 5000 steps, which will make me think oh I haven’t done that many today...my spouse will say that too, go for another walk. [P2, 3 months, the Netherlands] Overall Value and Utility There was a consensus throughout that the benefits of participating in the study outweighed the costs, of which there were relatively few. Many had not envisioned any personal benefits when enrolling as they were aware that they would not receive personalized outcomes; however, had been pleasantly surprised by the integration of RMT data into their depression self-management, as early as the 3-month time point: I think its empowering to know more about myself to understand more so I think once I can see more what the data is from collecting from data when the other apps are working and being able to see what the data is and notice any correlations then I think that will be really valuable. [P12, 3 months, United Kingdom] Technology-Related Factors Experience of the technology used in the study (smartphone apps and Fitbit) was the most widely cited theme across all sites. This covered the convenience of integrating the RMTs into daily life, the usability of the technology, technological malfunctions that occurred, and the extent to which participants found the technologies intrusive. Convenience Using a mobile phone and wearing a watch were already an integral part of many participants’ daily routine. The Fitbit device, “it’s basically wearing a watch” (P7, 3 months, United Kingdom), collected data passively without the need to input information, and continual wear, syncing, and charging were integrated into the routine as early as the 3-month time point. Reminder messages across the system were useful in the process of long-term integration. One aspect that participants found more difficult to integrate into their routine was the app questionnaires. Timing of the questionnaires was often inconvenient, for example when at work, driving, or in social situations: “Obviously I’m less likely to stop my conversation to be like oh this questionnaire, because that’s a bit rude” (P4, 3 months, United Kingdom). Frequency https://humanfactors.jmir.org/2023/1/e39479 XSL•FO RenderX of the ESM questionnaires was also too high from some: “it’s impossible to have a routine with that. If you have a full-time job, it’s always a bother” (P17, 24 months, the Netherlands). The participants were rarely able to change their routine to accommodate answering the questionnaires, which sometimes caused guilt. One participant in the Spanish cohort reflected on how work affected their ability to respond to app notifications during their 2-year participation: At the beginning it was a bit difficult because I was working, then as I was on sick leave for two years, the truth is that I’ve been able to adapt quite well. And in the end, when I went back to work again, it was a bit difficult... [P1, 24 months, Spain] Usability For those who received a smartphone upon enrollment, a large technological barrier was the process of “relearning” a new operating system. This was described by some as “more difficult than anticipated” (P3, 3 months, United Kingdom), particularly during the 3-month interviews, owing to adapting to a new user interface and decreased connectivity with other devices. At 24 months, some participants had adjusted to using the new device, whereas others planned to swap back upon study completion: No, my only peeve was that I’m an Apple user and having this bloody awful Android phone, the first thing I shall do on April 1st is take my SIM card out of the Motorola thingy. [P35, 24 months, United Kingdom] Technological Malfunctions The participants reported a range of technological malfunctions that affected their participation in the study. Issues with the study apps were particularly prevalent during the 3-month interviews owing to ongoing technological challenges during the early phases of the study. These included not receiving notifications, apps crashing, apps logging out, and difficulties with rescanning QR codes. Participants sometimes had limited time or motivation to report issues to the team: I tried opening a questionnaire I wouldn’t be able to see it, I wouldn’t be able to do it and there was no way of saying this is happening or why this is happening so maybe I should have contacted you about it but I just kind of ignored it. [P4, 3 months, United Kingdom] Issues with missing data persisted throughout the 3 time points. Participants were aware of the times when the active app had been unable to submit the completed data, or the passive app had ceased monitoring. Such malfunctions often led to anxiety or guilt that they were not “correctly” participating: “Well, yes, when it didn’t work, I became a bit nervous...” (P15, 3 months, Spain). Participants also reported frequent missing data with the Fitbit, caused either by a syncing error or inaccurate recording. These issues caused some to question the integrity of the study: “It just didn’t work and that’s not what you expect from a research study” (P18, 24 months, the Netherlands). JMIR Hum Factors 2023 | vol. 10 | e39479 | p. 6 (page number not for citation purposes) JMIR HUMAN FACTORS White et al A participant in the Spanish cohort reflected on how these technological malfunctions affected not only their ability to participate in the study but also their experience of being able to use the resulting data: There is data that I have missed here, and of course I was analyzing it with me in important situations of how I was, and that I have missed them, for more than a month. [P32, 24 months, Spain] Intrusiveness Generally, the concept of remote monitoring, or the use of the technologies, was not regarded as intrusive. Rather, passive data collection was noted as a preferable method because “at some point you don’t notice it. You don’t notice that you’re wearing it anymore” (P18, 24 months, the Netherlands). However, one area that caused disruption was the wearability of the Fitbit device. Several issues associated with the Fitbit strap were reported, including skin irritation, increased sweating, and allergic reactions. Some had briefly chosen to remove the device while experiencing discomfort, whereas others had purchased straps with alternative materials. At 12 months, many reported that their strap had broken, and by 24 months, some had to apply for a full device replacement. One participant felt guilty when asking the research team for their device to be repaired: I know that the money allocated to research programs or projects is minimal, and of course, when the strap broke or the Fitbit wouldn’t charge me and then I felt really bad because I thought “oh my God, now they have to change my Fitbit.” [P26, 24 months, Spain] Waiting for a replacement strap or device meant that participants were unable to continue to use the Fitbit for self-management: if I was going to continue and for the others who will be continuing, it will probably begin to happen more and more depending on how much people are actually exercising with them on. It only grows, that’s the problem, in my experience with the other Fitbit, that definitely happens. [P3, 12 months, United Kingdom] Clinical Factors The participants were asked to reflect on whether and how they could see the RMT data being used in a clinical setting. Discussions included the extent to which participants felt comfortable sharing the data, how they envisioned clinicians using the data, and how feasible this was in the current climate. Views on Data Sharing At the 12- and 24-month time points, the participants were specifically asked to comment on data sharing with medical professionals. In general, allowing trusted clinicians to view RMT data alongside medical records was acceptable, or even essential: “let’s say my whole history, my doctor already has it, if she has it more extensive, then all the better for me.” (P30, 24 months, Spain). There was some discrepancy over whether these data should automatically be available to clinicians or mediated by the patient. Some thought that medical professionals “would be in a better position to evaluate what they needed https://humanfactors.jmir.org/2023/1/e39479 XSL•FO RenderX from it than me to decide that” (P32, 24 months, United Kingdom). Others worried about interpretation of the data without context: I suppose, [I would like to] understand what it is that is proposed to be shared, and if there’s something there that would not be appropriate at that time, because I don’t know what it is until I see it, then yes, I would like to have a choice...I would want to make sure that my health record reflects actuality rather than something that can be interpreted by people incorrectly. [P31, 24 months, United Kingdom] Clinical Uses of RMT Data The participants suggested several ways in which they might expect RMT data to be beneficial in clinical care. These included (1) allowing the clinician to view the “whole picture” of individual experience, (2) allowing the clinician insight into new symptoms, (3) as a way for patients to report specific areas of concern, and finally (4) as a basis for making decisions about suitable treatment or care. Importantly, treatment decisions should be reached as a joint decision involving the clinician, the patient, and the data: I think they could actually look at the data that’s being produced, and that could assist them in helping me to come to another decision. Like, if I was deciding that I would like to move my medication down, but they’ve got the data that says, no you’re not...but if it backs it up as well, so it can work both ways, so I think it does have those benefits. [P33, 24 months, United Kingdom] Sleep data were repeatedly cited as a data stream that would cause change in treatment. Participants from all sites provided examples of conversations with their mental health clinicians. One participant in the British cohort also discussed their experience of integrating the sleep data into their sleep clinic appointments: It’s too expensive for the NHS to keep on doing [sleep tests]...I said, well, actually, I can show you any time in the last six months or so...an indication of when I’m sleeping...It helped them choose what exercises I needed to do and what therapy was required, so, yes, it was extremely helpful. [P22, 12 months, United Kingdom] Presentation of objective sleep data was seen as helpful “proof” of the participant’s recent experiences: You can tell your GP that you sleep terribly, but of course your GP can also think that you’re just worried, but with the data it’s a fact that you can prove, so that’s nice, that you have concrete info...whether you worry or complain about it or not doesn’t matter, the facts are there. [P10, 12 months, the Netherlands] Current Clinical Utility of RMTs Although the potential for RMTs in clinical care was recognized, 2 key barriers to their implementation were envisioned. First, the level of technological acceptance of medical professionals JMIR Hum Factors 2023 | vol. 10 | e39479 | p. 7 (page number not for citation purposes) JMIR HUMAN FACTORS White et al influenced participant views on the long-term utility of the data. Participants in the Spanish cohort, who were recruited through their clinical care, generally reported acceptance of the study from their clinicians: “even my psychiatrist here and in Barcelona had the same way of thinking and saw that this was very useful for me and encouraged me” (P9, 24 months, Spain). Others described more negative experiences, often causing them to question the use of the data: I thought it would be more relevant for my neurologist, but my neurologist wasn’t particularly interested when I told him about what I was doing in the study. [P17, 12 months, United Kingdom] Second, lack of funding, resources, and time was perceived as a major roadblock to using RMT data in appointments. This was particularly apparent in the British cohort with regard to the National Health Service. For the data to be monitored and reflected on, new procedures would need to be put in place: I would be amazed if there was sufficient funding for that...I don’t believe that the NHS have got the resources to have people monitoring this sort of stuff. [P22, 12 months, United Kingdom] Given the perceived lack of resources to effectively use RMT data in the National Health Service, some have considered how best to come to a compromise: I think realistically, if they had that [data] and I went to them with a problem, then I would like them to be able to use it at that point. But I don’t see it as something that they would be—so, for example, if I went to them with something and if somehow, it was a part of my NHS records, if they could access that, that might be helpful to them. But I don’t see them using it other than that really. [P32, 24 months, United Kingdom] System Amendments and Additions Participants discussed various changes or additions to the RMT system used in this study to further encourage long-term engagement. These included suggestions for questionnaire data collection and feedback. Data Collection Across all sites and time points, the most prevalent suggestions for changes to the study design were the content of active RMT questionnaires. Participants felt that they were frequently being asked to complete the same questions, particularly within the ESM schedule, which often prompted them to provide the same answers, for example, with regard to mood changes. This affected motivation: At first, I was more excited about it, but as time has passed, sometimes I don’t feel much like answering since the same questions get repeated. [P19, 12 months, Spain] Some also suggested the ability to postpone questionnaires if feeling too low to complete them and the ability to provide contextual information. As early as the 3-month time point, some noted that external factors affecting their mood were not https://humanfactors.jmir.org/2023/1/e39479 XSL•FO RenderX being monitored within the validated mood and self-esteem questionnaires: “I notice that when my home situation isn’t great, I also fill in the questionnaires less positively” (P5, 3 months, the Netherlands). On reflection, some would have liked to have given more information at certain points: The answers are very closed, so you can’t really answer what you feel. You know? It’s very...it’s very up in the air. [P1, 24 months, Spain] Data Feedback When asked how they might wish to view their symptom data in future use, the majority felt that this was best displayed visually through in-app graphs. Many also expressed that this would need to be accompanied by a “human explanation for what those things mean” (P3, 12 months, United Kingdom). There was a discrepancy between when these data would be best received; some only expected to receive it at the end of the study, some felt that it would be more useful in real time, whereas others were cautious that receiving data during periods of low mood would be detrimental: If I’m well I want to see it, if I’m unwell, no. If I was reporting that I was feeling suicidal I don’t think I’d want to revisit it. [P27, 24 months, United Kingdom] Furthermore, some participants considered the potential for RMT data to provide feedback on symptom patterns and changes over time, correlations with other factors, and depressive relapse prediction. Specific examples included relationships between exercise and mood, sleep and mood, and mood and concentration: “At some point I had a burn out. I’m very curious as to how my ability to concentrate changed, and if that maybe shows on the THINC-it app” (P3, 24 months, the Netherlands). It was generally accepted that having access to data of this nature would be useful for both self-management and integration into clinical care. Looking forward at the 24-month time point, one participant at the British site explained their hopes for the future of this field: I think trends are really quite important for me in managing what is going on...I think one of the things I am thinking would be good to come out of this is an ability to see patterns over time and then maybe being able to use that as a predictor or, I need to do some intervention here so that I don’t end up there again if that makes sense. [P30, 24 months, United Kingdom] Discussion Principal Findings An exploration of the subjective experience of long-term engagement with RMTs for depression symptom management could prove a necessary complement to objective engagement statistics, providing insights into technology usability, user experience, and facilitators of sustained use. This study aimed to (1) explore the key themes associated with long-term RMT use and (2) identify recommendations for future engagement through longitudinal qualitative analysis at 3-month, 12-month, and 24-month time points of the RADAR-MDD study. JMIR Hum Factors 2023 | vol. 10 | e39479 | p. 8 (page number not for citation purposes) JMIR HUMAN FACTORS White et al The themes uncovered suggest that long-term engagement with RMTs can be understood from two main perspectives: (1) experiential factors and (2) system-related factors (Figure 1). Experiential factors relate to the ways in which participants construct their experiences of engaging with RMTs for symptom monitoring. Experiential factors comprise research altruism, support from a professional team, and the benefits of using RMTs for depression management. System-related factors refer to direct engagement with the RMT systems. The factors include the usability, convenience, and intrusiveness of the technologies and the recommended system improvements for successful clinical implementation. On the basis of these perspectives, we present a set of considerations for the promotion of engagement with RMTs for depression. Given the breadth of use cases proposed for RMTs in MDD, we focused on two areas: (1) engagement with research and (2) engagement with real-world implementation. Recommendations for engagement with future RMT research are outlined in Multimedia Appendix 3. Although our data were derived from research participants, we believe that our findings can also be useful when considering implementation into clinical practice. Participants identified the following opportunities for RMTs in clinical care: (1) provision of feedback-informed care, (2) strengthening the therapeutic relationship, and (3) the specific clinical value of sleep monitoring. However, this potential was acknowledged with the caveat of a perceived lack of time and resources in clinical care across all 3 countries. Our findings indicate that a large difference between engagement with RMTs for research and long-term clinical engagement could be research altruism. In this study, an important facilitator of both initial and sustained engagement was the experiential factor of taking part in a novel, academic study to advance understanding and help others. To this end, participants forewent privacy concerns and initial receipt of personal benefit. They were also willing to engage despite the implementation concerns. In the absence of research altruism, Figure 1 can be used to identify further experiential facilitators that could instead be harnessed to promote engagement when RMTs become integrated into evidence-based practice. For example, clinical onboarding sessions could include a clear summary of the proposed uses and benefits of RMT data and symptom monitoring for an individual’s care. Multimedia Appendix 4 provides a set of considerations for the implementation of RMTs into clinical care based on the experiential and system-related factors identified. Figure 1. Experiential and system-related factors in the subjective experience of longitudinal remote measurement technology (RMT) use. https://humanfactors.jmir.org/2023/1/e39479 XSL•FO RenderX JMIR Hum Factors 2023 | vol. 10 | e39479 | p. 9 (page number not for citation purposes) JMIR HUMAN FACTORS White et al Figure 2. Recommendations for future remote measurement technology (RMT) use in observational research. Figure 3. Considerations for remote measurement technology (RMT) implementation in real-world clinical settings. Comparison With Previous Work This study builds on previous qualitative analyses of the barriers to and facilitators of intended RMT use for depression management. The functional and nonfunctional requirements set out by Simblett et al [22] roughly align with the system and experiential factors found here. However, a comparison of coding frameworks (Multimedia Appendix 2) revealed several differences in this study. First, nonfunctional, user-related factors such as cognition, symptom severity, and emotional resources were not acknowledged as barriers to long-term RMT engagement. Second, the overall utility of RMTs was discussed mainly in terms of benefits and rewards, and less so in terms of costs such as privacy and security. Third, studying long-term RMT use has revealed an additional layer of understanding https://humanfactors.jmir.org/2023/1/e39479 XSL•FO RenderX surrounding nonfunctional requirements; experiential factors include the impact of professional support and the effects of symptom monitoring on self-awareness and communication. When comparing our findings with those from the wider mHealth literature, technological and system-related factors remained a common theme. Borghouts et al [26] and Patel et al [27] found that lack of technical issues, flexible usability of the platform, personalization, and access to training were associated with increased long-term engagement with digital health intervention platforms. One clear difference with digital health intervention work is the focus on “a desire to actively improve one’s health” [27] as a main facilitator of initial and sustained engagement. Our work has shown that in the absence of a direct or tangible benefit, users remain willing to interact with RMTs for long periods within a research context. JMIR Hum Factors 2023 | vol. 10 | e39479 | p. 10 (page number not for citation purposes) JMIR HUMAN FACTORS White et al Experiential factors such as advancing scientific understanding and, at later periods, experiencing indirect benefits of mood tracking, seem to operate as a supplement to the user-related factors currently reported in the field. Strengths and Limitations To the best of our knowledge, this is the largest study to qualitatively explore long-term RMT use for depression across multiple countries. Data collection and analyses were conducted in the native language of each country and only quotes were translated into English, aiding the transfer of meaning process [34]. However, this study has some limitations. First, where we did not anticipate any major intercountry differences in terms of attitudes toward remote mental health tracking, participants in the Spanish cohort were invited to participate by the clinicians involved in their care. This might have overinflated some themes in our analyses; for example, perceived benefits of the interviews were conducted via technologies. Second, convenience sampling of the participants who remained enrolled at each time point. This increased the risk of selection bias; those who enjoyed using the RMTs were more likely to continue to engage and as a result more likely to agree to an interview. This could explain the absence of themes relating to symptom severity or cognitive barriers present in the current work, although recent analyses have suggested that these factors did not contribute to sustained engagement in the study [35]. in 21 participants Convenience sampling also resulted completing the interviews at ≥2 time points. Preliminary sensitivity checks on a subset of this sample showed no clear signs of changes in themes over time. The data were not deemed rich enough to undertake a full, longitudinal analysis on this sample. Third, because of resource constraints, no sites undertook double coding. Fourth, data-driven themes were not explored in relation to demographic or clinical factors, as this was deemed beyond the scope of this study. Although previous work suggests that perceived usability, and actual use, of the RADAR-base system remains robust across severity of clinical characteristics [35], understanding demographic differences in subjective engagement is an important avenue for future research. Finally, the COVID-19 pandemic occurred during the study follow-up period. Given the transition to remote working and health care across all 3 countries during this time, the subjective experience of using RMTs might have been positively skewed; for example, with regard to the positive impact of the research team during social isolation. It should also be noted that the topic guide primarily asked participants to review their experience of using RMTs for this specific research project, and specific use cases for clinical implementation were not outlined by interviewers. Thus, the themes that arose from this work relate primarily to long-term engagement with RMT research, and the transferability of the findings to engagement in clinical care should be taken with caution. for clear Applications for Future Research Future work should continue to explore subjective engagement with RMTs, conceptualized in terms of both experiential and system-related factors. Where system-related factors often represent technological recommendations improvements, understanding the experiential effects of engaging with RMTs is a novel finding that could prove fundamental in promoting future engagement. A recent systematic review [17] found that 5 studies have begun to explore the correlational relationship between objective and subjective engagement with RMTs. Higher daily assessment counts from an active RMT app were correlated with increased app satisfaction ratings at 3-month and 6-month time points [36,37]. Understanding the link between experiential factors, such as increased self-awareness, and objective engagement could bolster this field further. Our findings explore the initial and sustained engagement with RMTs for depression symptom monitoring in a research setting. The next step would be to replicate this work in a clinical setting. Recent qualitative analyses have reported positive views from patients and clinicians on the potential for implementation of RMT into psychological services [38]. This paper provides considerations for adapting RMT systems for use in clinical settings and a framework for continuing to analyze the subjective experience of long-term clinical engagement to allow for further iterations. Conclusions This study aimed to understand the subjective experience of long-term engagement with RMTs for depression symptom monitoring as a complement to the high rates of objective engagement observed in the RADAR-MDD study. Key experiential and system-related themes associated with long-term RMT use were identified along with a set of recommendations and considerations for promoting future system use in both research and clinical settings. Further understanding of the construction of the “experience” of using RMTs will be key to promoting long-term engagement in clinical care and depression management in comparison with general mHealth interventions that offer immediate or tangible rewards. In the wake of the rapid expansion of this field, we urge professionals to continue monitoring the subjective experience of RMT engagement to maximize the potential of remote monitoring as both a method for data collection and a tool for symptom management. Acknowledgments This paper represents an independent research part funded by the National Institute for Health Research (NIHR) Maudsley Biomedical Research Centre at South London and Maudsley National Health Service (NHS) Foundation Trust and King’s College London. The views expressed are those of the authors and not necessarily those of the NHS, the NIHR, or the Department of Health and Social Care. The authors would like to thank all the members of the Remote Assessment of Disease and Relapse-Central Nervous System (RADAR-CNS) patient advisory board for their contribution to the device selection procedures and their invaluable advice https://humanfactors.jmir.org/2023/1/e39479 XSL•FO RenderX JMIR Hum Factors 2023 | vol. 10 | e39479 | p. 11 (page number not for citation purposes) JMIR HUMAN FACTORS White et al throughout the study protocol design. The authors would also like to acknowledge the work of Felice Fernhout in conducting coding on the data. Participant recruitment in Amsterdam was partially accomplished through Hersenonderzoek.nl, a Dutch web-based registry that facilitates participant recruitment for neuroscience studies [39]. Hersenonderzoek.nl is funded by ZonMw-Memorabel (project number 73305095003), a project in the context of the Dutch Deltaplan Dementie, Gieskes-Strijbis Foundation, the Alzheimer’s Society in the Netherlands, and Brain Foundation Netherlands. Participants in Spain were recruited through the following institutions: Parc Sanitari Sant Joan de Déu network of mental health services (Barcelona), Institut Català de la Salut primary care services (Barcelona), Institut Pere Mata-Mental Health Care (Terrassa), and Hospital Clínico San Carlos (Madrid). The authors would like to thank all Genetic Links to Anxiety and Depression study volunteers for their participation and gratefully acknowledge the NIHR BioResource, NIHR BioResource centers, NHS Trusts, and staff for their contributions. The authors would also like to acknowledge NIHR Biomedical Research Centre (BRC), King’s College London, South London and Maudsley NHS Trust and King’s Health Partners. The authors would like to thank the NIHR, NHS Blood and Transplant, and Health Data Research United Kingdom, as part of the Digital Innovation Hub Program. This research was reviewed by a team with experience of mental health problems and their caregivers, who were specially trained to advise on research proposals and documentation through the Feasibility and Acceptability Support Team for Researchers (FAST-R): a free, confidential service in England provided by the NIHR Maudsley BRC via King’s College London and South London and Maudsley NHS Foundation Trust. The RADAR-CNS project received funding from the Innovative Medicines Initiative (IMI) 2 Joint Undertaking under grant 115902. This joint undertaking received support from the European Union’s Horizon 2020 research and innovation program and European Federation of Pharmaceutical Industries and Associations. This communication reflects the views of the RADAR-CNS consortium and neither IMI nor the European Union and European Federation of Pharmaceutical Industries and Associations are liable for any use that may be made of the information contained herein. The funding body was not involved in study design, data collection or analysis, or data interpretation. Data Availability The data supporting the findings of this study are available from the corresponding author upon reasonable request. Authors' Contributions KMW contributed to the design and coordination of the study in London as well as to the data processing, coding, analysis, and writing of this manuscript. EDL contributed to the data coding and analysis. S Siddi contributed to the design and coordination of the study in Barcelona as well as the data coding. FL contributed to the design and coordination of the study in Amsterdam, as well as the data coding. S Simblett contributed to the development and design of the study and advised on the analyses. GRA has contributed to data coding. AI contributed to the study conducted in London. IM-G contributed to the development and design of the study. JMH contributed to the development and design of the study. CO contributed to the study conducted in London. PP contributed to the data coding and analysis. AR contributed to the development and design of the study. ER contributed to participant recruitment for the study. TW contributed to the development and design of the study. CH contributed to data interpretation and supervision of the first author. MH secured funding and is the principal investigator of the study, and contributed to the overall study design and conduct. FM contributed to the design and coordination of the study. Patient advisory board members contributed to the design and development of the study. Conflicts of Interest MH is the principal investigator of the RADAR-CNS program, a precompetitive public-private partnership funded by the Innovative Medicines Initiative and the European Federation of Pharmaceutical Industries and Associations. The program received support from Janssen, Biogen, Merck & Co, Union Chimique Belge, and Lundbeck. JMH has received economic compensation for participating in advisory boards or giving educational lectures from Eli Lilly & Co, Sanofi, Lundbeck, and Otsuka. CO is supported by the UK Medical Research Council (MR/N013700/1) and King’s College London member of the MRC Doctoral Training Partnership in Biomedical Sciences. Multimedia Appendix 1 Main interview questions at 3-month, 12-month, and 24-month follow-up time points. [DOCX File , 21 KB-Multimedia Appendix 1] Multimedia Appendix 2 Preliminary and final codes in the coding framework. https://humanfactors.jmir.org/2023/1/e39479 XSL•FO RenderX JMIR Hum Factors 2023 | vol. 10 | e39479 | p. 12 (page number not for citation purposes) JMIR HUMAN FACTORS White et al [DOCX File , 20 KB-Multimedia Appendix 2] Multimedia Appendix 3 Recommendations for future remote measurement technology (RMT) use in observational research. [DOCX File , 319 KB-Multimedia Appendix 3] Multimedia Appendix 4 Considerations for remote measurement technology (RMT) implementation in real-world clinical settings. [DOCX File , 366 KB-Multimedia Appendix 4] References 1. 2. 3. 4. GBD 2017 DiseaseInjury IncidencePrevalence Collaborators. Global, regional, and national incidence, prevalence, and years lived with disability for 354 diseases and injuries for 195 countries and territories, 1990-2017: a systematic analysis for the Global Burden of Disease Study 2017. Lancet 2018 Nov 10;392(10159):1789-1858 [FREE Full text] [doi: 10.1016/S0140-6736(18)32279-7] [Medline: 30496104] Verduijn J, Verhoeven JE, Milaneschi Y, Schoevers RA, van Hemert AM, Beekman AT, et al. Reconsidering the prognosis of major depressive disorder across diagnostic boundaries: full recovery is the exception rather than the rule. BMC Med 2017 Dec 12;15(1):215 [FREE Full text] [doi: 10.1186/s12916-017-0972-8] [Medline: 29228943] Greenberg PE, Fournier AA, Sisitsky T, Simes M, Berman R, Koenigsberg SH, et al. The economic burden of adults with major depressive disorder in the United States (2010 and 2018). Pharmacoeconomics 2021 Jun;39(6):653-665 [FREE Full text] [doi: 10.1007/s40273-021-01019-4] [Medline: 33950419] Gauthier G, Mucha L, Shi S, Guerin A. Economic burden of relapse/recurrence in patients with major depressive disorder. J Drug Assess 2019;8(1):97-103 [FREE Full text] [doi: 10.1080/21556660.2019.1612410] [Medline: 31192030] 5. Moshe I, Terhorst Y, Opoku Asare K, Sander LB, Ferreira D, Baumeister H, et al. Predicting symptoms of depression and anxiety using smartphone and wearable data. Front Psychiatry 2021 Jan 28;12:625247 [FREE Full text] [doi: 10.3389/fpsyt.2021.625247] [Medline: 33584388] 6. Matcham F, Barattieri di San Pietro C, Bulgari V, de Girolamo G, Dobson R, Eriksson H, RADAR-CNS consortium. Remote assessment of disease and relapse in major depressive disorder (RADAR-MDD): a multi-centre prospective cohort study protocol. BMC Psychiatry 2019 Feb 18;19(1):72 [FREE Full text] [doi: 10.1186/s12888-019-2049-z] [Medline: 30777041] Torous J, Staples P, Shanahan M, Lin C, Peck P, Keshavan M, et al. Utilizing a personal smartphone custom app to assess the patient health questionnaire-9 (PHQ-9) depressive symptoms in patients with major depressive disorder. JMIR Ment Health 2015 Mar 24;2(1):e8 [FREE Full text] [doi: 10.2196/mental.3889] [Medline: 26543914] Zhang Y, Folarin AA, Sun S, Cummins N, Ranjan Y, Rashid Z, et al. Predicting depressive symptom severity through individuals' nearby bluetooth device count data collected by mobile phones: preliminary longitudinal study. JMIR Mhealth Uhealth 2021 Jul 30;9(7):e29840 [FREE Full text] [doi: 10.2196/29840] [Medline: 34328441] Kheirkhahan M, Nair S, Davoudi A, Rashidi P, Wanigatunga AA, Corbett DB, et al. A smartwatch-based framework for real-time and online assessment and mobility monitoring. J Biomed Inform 2019 Jan;89:29-40 [FREE Full text] [doi: 10.1016/j.jbi.2018.11.003] [Medline: 30414474] 7. 8. 9. 10. McIntyre RS, Best MW, Bowie CR, Carmona NE, Cha DS, Lee Y, et al. The THINC-integrated tool (THINC-it) screening assessment for cognitive dysfunction. J Clin Psychiatry 2017 Aug 23;78(7):873-881. [doi: 10.4088/jcp.16m11329] 11. Dineley J, Lavelle G, Leightley D, Matcham F, Siddi S, Peñarrubia-María MT, The RADAR-CNS Consortium. Remote smartphone-based speech collection: acceptance and barriers in individuals with major depressive disorder. Proc Interspeech 2021:631-635 [FREE Full text] [doi: 10.21437/Interspeech.2021-1240] 12. Zhang Y, Folarin AA, Sun S, Cummins N, Bendayan R, Ranjan Y, RADAR-CNS Consortium. Relationship between major depression symptom severity and sleep collected using a wristband wearable device: multicenter longitudinal observational study. JMIR Mhealth Uhealth 2021 Apr 12;9(4):e24604 [FREE Full text] [doi: 10.2196/24604] [Medline: 33843591] 13. Lee S, Kim H, Park MJ, Jeon HJ. Current advances in wearable devices and their sensors in patients with depression. Front Psychiatry 2021 Jun 17;12:672347 [FREE Full text] [doi: 10.3389/fpsyt.2021.672347] [Medline: 34220580] 14. O'Brien HL, Toms EG. What is user engagement? A conceptual framework for defining user engagement with technology. J Am Soc Inf Sci 2008 Apr;59(6):938-955 [FREE Full text] [doi: 10.1002/asi.20801] 15. De Angel V, Lewis S, White K, Oetzmann C, Leightley D, Oprea E, et al. Digital health tools for the passive monitoring 16. of depression: a systematic review of methods. NPJ Digit Med 2022 Jan 11;5(1):3 [FREE Full text] [doi: 10.1038/s41746-021-00548-8] [Medline: 35017634] Simblett S, Greer B, Matcham F, Curtis H, Polhemus A, Ferrão J, et al. Barriers to and facilitators of engagement with remote measurement technology for managing health: systematic review and content analysis of findings. J Med Internet Res 2018 Jul 12;20(7):e10480 [FREE Full text] [doi: 10.2196/10480] [Medline: 30001997] https://humanfactors.jmir.org/2023/1/e39479 XSL•FO RenderX JMIR Hum Factors 2023 | vol. 10 | e39479 | p. 13 (page number not for citation purposes) JMIR HUMAN FACTORS White et al 17. White KM, Williamson C, Bergou N, Oetzmann C, de Angel V, Matcham F, et al. A systematic review of engagement reporting in remote measurement studies for health symptom tracking. NPJ Digit Med 2022 Jun 29;5(1):82 [FREE Full text] [doi: 10.1038/s41746-022-00624-7] [Medline: 35768544] 18. Oetzmann C, White KM, Ivan A, Julie J, Leightley D, Lavelle G, RADAR-CNS consortium. Lessons learned from recruiting into a longitudinal remote measurement study in major depressive disorder. NPJ Digit Med 2022 Sep 03;5(1):133 [FREE Full text] [doi: 10.1038/s41746-022-00680-z] [Medline: 36057688] 19. Matcham F, Leightley D, Siddi S, Lamers F, White KM, Annas P, RADAR-CNS consortium. Remote Assessment of Disease and Relapse in Major Depressive Disorder (RADAR-MDD): recruitment, retention, and data availability in a longitudinal remote measurement study. BMC Psychiatry 2022 Feb 21;22(1):136 [FREE Full text] [doi: 10.1186/s12888-022-03753-1] [Medline: 35189842] Perski O, Blandford A, West R, Michie S. Conceptualising engagement with digital behaviour change interventions: a systematic review using principles from critical interpretive synthesis. Transl Behav Med 2017 Jun 13;7(2):254-267 [FREE Full text] [doi: 10.1007/s13142-016-0453-1] [Medline: 27966189] 20. 21. McGrady E, Conger S, Blanke S, Landry BJ. Emerging technologies in healthcare: navigating risks, evaluating rewards. J 22. Healthc Manag 2010;55(5):353-365. [doi: 10.1097/00115514-201009000-00011] Simblett S, Matcham F, Siddi S, Bulgari V, Barattieri di San Pietro C, Hortas López J, RADAR-CNS Consortium. Barriers to and facilitators of engagement with mHealth technology for remote measurement and management of depression: qualitative analysis. JMIR Mhealth Uhealth 2019 Jan 30;7(1):e11325 [FREE Full text] [doi: 10.2196/11325] [Medline: 30698535] 23. Alpert JM, Manini T, Roberts M, Kota NS, Mendoza TV, Solberg LM, et al. Secondary care provider attitudes towards patient generated health data from smartwatches. NPJ Digit Med 2020 Mar 03;3(1):27 [FREE Full text] [doi: 10.1038/s41746-020-0236-4] [Medline: 32140569] 24. Andrews JA, Craven MP, Jamnadas-Khoda J, Lang AR, Morriss R, Hollis C, RADAR-CNS Consortium. Health care 25. professionals' views on using remote measurement technology in managing central nervous system disorders: qualitative interview study. J Med Internet Res 2020 Jul 24;22(7):e17414 [FREE Full text] [doi: 10.2196/17414] [Medline: 32706664] Patoz M, Hidalgo-Mazzei D, Blanc O, Verdolini N, Pacchiarotti I, Murru A, et al. Patient and physician perspectives of a smartphone application for depression: a qualitative study. BMC Psychiatry 2021 Jan 29;21(1):65 [FREE Full text] [doi: 10.1186/s12888-021-03064-x] [Medline: 33514333] 26. Borghouts J, Eikey E, Mark G, De Leon C, Schueller SM, Schneider M, et al. Barriers to and facilitators of user engagement 27. 28. with digital mental health interventions: systematic review. J Med Internet Res 2021 Mar 24;23(3):e24387 [FREE Full text] [doi: 10.2196/24387] [Medline: 33759801] Patel S, Akhtar A, Malins S, Wright N, Rowley E, Young E, et al. The acceptability and usability of digital health interventions for adults with depression, anxiety, and somatoform disorders: qualitative systematic review and meta-synthesis. J Med Internet Res 2020 Jul 06;22(7):e16228 [FREE Full text] [doi: 10.2196/16228] [Medline: 32628116] Schwartz S, Schultz S, Reider A, Saunders EF. Daily mood monitoring of symptoms using smartphones in bipolar disorder: a pilot study assessing the feasibility of ecological momentary assessment. J Affect Disord 2016 Feb;191:88-93 [FREE Full text] [doi: 10.1016/j.jad.2015.11.013] [Medline: 26655117] 29. Bauer AM, Iles-Shih M, Ghomi RH, Rue T, Grover T, Kincler N, et al. Acceptability of mHealth augmentation of collaborative care: a mixed methods pilot study. Gen Hosp Psychiatry 2018 Mar;51:22-29 [FREE Full text] [doi: 10.1016/j.genhosppsych.2017.11.010] [Medline: 29272712] 30. Gustavell T, Sundberg K, Langius-Eklöf A. Using an interactive app for symptom reporting and management following pancreatic cancer surgery to facilitate person-centered care: descriptive study. JMIR Mhealth Uhealth 2020 Jun 17;8(6):e17855 [FREE Full text] [doi: 10.2196/17855] [Medline: 32554375] 31. Calman L, Brunton L, Molassiotis A. Developing longitudinal qualitative designs: lessons learned and recommendations for health services research. BMC Med Res Methodol 2013 Feb 06;13(1):14 [FREE Full text] [doi: 10.1186/1471-2288-13-14] [Medline: 23388075] 32. Ranjan Y, Rashid Z, Stewart C, Conde P, Begale M, Verbeeck D, Hyve, RADAR-CNS Consortium. RADAR-Base: open source mobile health platform for collecting, monitoring, and analyzing data using sensors, wearables, and mobile devices. JMIR Mhealth Uhealth 2019 Aug 01;7(8):e11734 [FREE Full text] [doi: 10.2196/11734] [Medline: 31373275] 33. NVIVO homepage. NVIVO. URL: https://www.qsrinternational.com/nvivo-qualitative-data-analysis-software/home 34. [accessed 2022-05-01] van Nes F, Abma T, Jonsson H, Deeg D. Language differences in qualitative research: is meaning lost in translation? Eur J Ageing 2010 Dec 19;7(4):313-316 [FREE Full text] [doi: 10.1007/s10433-010-0168-y] [Medline: 21212820] 35. Matcham F, Carr E, White KM, Leightley D, Lamers F, Siddi S, et al. Predictors of engagement with remote sensing 36. technologies for symptom measurement in Major Depressive Disorder. Journal of Affective Disorders 2022 Aug;310:106-115. [doi: 10.1016/j.jad.2022.05.005] Jamison RN, Jurcik DC, Edwards RR, Huang CC, Ross EL. A pilot comparison of a smartphone app with or without 2-way messaging among chronic pain patients: who benefits from a pain app? Clin J Pain 2017 Aug;33(8):676-686 [FREE Full text] [doi: 10.1097/AJP.0000000000000455] [Medline: 27898460] https://humanfactors.jmir.org/2023/1/e39479 XSL•FO RenderX JMIR Hum Factors 2023 | vol. 10 | e39479 | p. 14 (page number not for citation purposes) JMIR HUMAN FACTORS White et al 37. 38. Jamison RN, Mei A, Ross EL. Longitudinal trial of a smartphone pain application for chronic pain patients: predictors of compliance and satisfaction. J Telemed Telecare 2016 Nov 10;24(2):93-100. [doi: 10.1177/1357633x16679049] de Angel V, Lewis S, White KM, Matcham F, Hotopf M. Clinical targets and attitudes toward implementing digital health tools for remote measurement in treatment for depression: focus groups with patients and clinicians. JMIR Ment Health 2022 Aug 15;9(8):e38934 [FREE Full text] [doi: 10.2196/38934] [Medline: 35969448] 39. Hersenziekten de wereld uit helpen kan alleen met onderzoek. hersenonderzoek nl. URL: https://hersenonderzoek.nl/ [accessed 2023-01-13] Abbreviations ESM: experience sampling methodology MDD: major depressive disorder mHealth: mobile health RADAR-MDD: Remote Assessment of Disease and Relapse-Major Depressive Disorder RMT: remote measurement technology Edited by A Kushniruk; submitted 13.05.22; peer-reviewed by A AL-Asadi, H Hsin; comments to author 06.10.22; revised version received 07.10.22; accepted 07.11.22; published 26.01.23 Please cite as: White KM, Dawe-Lane E, Siddi S, Lamers F, Simblett S, Riquelme Alacid G, Ivan A, Myin-Germeys I, Haro JM, Oetzmann C, Popat P, Rintala A, Rubio-Abadal E, Wykes T, Henderson C, Hotopf M, Matcham F Understanding the Subjective Experience of Long-term Remote Measurement Technology Use for Symptom Tracking in People With Depression: Multisite Longitudinal Qualitative Analysis JMIR Hum Factors 2023;10:e39479 URL: https://humanfactors.jmir.org/2023/1/e39479 doi: 10.2196/39479 PMID: an open-access ©Katie M White, Erin Dawe-Lane, Sara Siddi, Femke Lamers, Sara Simblett, Gemma Riquelme Alacid, Alina Ivan, Inez Myin-Germeys, Josep Maria Haro, Carolin Oetzmann, Priya Popat, Aki Rintala, Elena Rubio-Abadal, Til Wykes, Claire Henderson, Matthew Hotopf, Faith Matcham. Originally published in JMIR Human Factors (https://humanfactors.jmir.org), 26.01.2023. This the Creative Commons Attribution License is (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Human Factors, is properly cited. The complete bibliographic information, a link to the original publication on https://humanfactors.jmir.org, as well as this copyright and license information must be included. article distributed under terms of the https://humanfactors.jmir.org/2023/1/e39479 XSL•FO RenderX JMIR Hum Factors 2023 | vol. 10 | e39479 | p. 15 (page number not for citation purposes)
10.1016_j.isci.2023.106902
iScience ll OPEN ACCESS Article Development of an in vitro method for activation of X-succinate synthases for fumarate hydroalkylation Mary C. Andorfer, Devin T. King- Roberts, Christa N. Imrich, Balyn G. Brotheridge, Catherine L. Drennan cdrennan@mit.edu Highlights X-succinate synthases (XSSs) use a glycyl radical to functionalize hydrocarbons Soluble XSS activating enzymes (XSS-AEs) were found through genome mining A soluble XSS-AE allowed in vitro glycyl radical installation within XSSs An auxiliary subunit is critical for fumarate hydroalkylation using XSSs Andorfer et al., iScience 26, 106902 June 16, 2023 ª 2023 The Author(s). https://doi.org/10.1016/ j.isci.2023.106902 iScience ll OPEN ACCESS Article Development of an in vitro method for activation of X-succinate synthases for fumarate hydroalkylation Mary C. Andorfer,1,3 Devin T. King-Roberts,2 Christa N. Imrich,4 Balyn G. Brotheridge,1,4 and Catherine L. Drennan1,3,4,5,6,7,* SUMMARY Anaerobic microbial degradation of hydrocarbons is often initiated through addi- tion of the hydrocarbon to fumarate by enzymes known as X-succinate synthases (XSSs). XSSs use a glycyl radical cofactor, which is installed by an activating enzyme (XSS-AE), to catalyze this carbon-carbon coupling reaction. The activa- tion step, although crucial for catalysis, has not previously been possible in vitro because of insolubility of XSS-AEs. Here, we take a genome mining approach to find an XSS-AE, a 4-isopropylbenzylsuccinate synthase (IBSS)-AE (IbsAE) that can be solubly expressed in Escherichia coli. This soluble XSS-AE can activate both IBSS and the well-studied benzylsuccinate synthase (BSS) in vitro, allowing us to explore XSSs biochemically. To start, we examine the role of BSS subunits and find that the beta subunit accelerates the rate of hydro- carbon addition. Looking forward, the methodology and insight gathered here can be used more broadly to understand and engineer XSSs as synthetically use- ful biocatalysts. INTRODUCTION Hydrocarbons are abundant within both natural (e.g., marine hydrocarbon seeps) and artificial (e.g., oil pipelines) environments. Aerobic microbial degradation of hydrocarbons has been well characterized and used in bioremediation of crude-oil-polluted environments; however, hydrocarbons inevitably end up in marine and terrestrial anoxic environments as well.1,2 Even within oxic zones, intensive respiration of facultative microbes creates anoxic microenvironments. Further understanding of how hydrocarbons are anaerobically degraded by microbes is necessary to create better tools for bioremediation and to inhibit microbial corrosion3 within crude-oil-containing facilities. Characterization of microbial commu- nities within these anaerobic environments remains an important, active area of research,2 and as more anaerobic degraders are discovered, it is also critical that we understand the underlying molecular mech- anisms that allow these microbes to accomplish hydrocarbon degradation in the absence of molecular oxygen.1 One of the key reactions catalyzed by these organisms involves metabolism of hydrocarbon sub- strates via addition to fumarate. This hydroalkylation reaction proceeds via homolytic cleavage of a C–H bond on the hydrocarbon substrate, addition of the hydrocarbyl radical to the C=C bond of fumarate, and addition of hydrogen to the resulting succinyl radical. Beyond environmental significance, this type of reaction is a synthetically attractive method for forming C–C bonds, as it has the potential to rapidly build structural complexity within small molecules without byproducts or pre-functionalized substrates. The growing class of enzymes that catalyze this impressive reaction is known as the X-succinate syn- thases (XSSs). In oxic environments, hydrocarbons are activated by enzymes that use iron, copper, and flavin cofactors to oxygenate using O2.4 When molecular oxygen is not present, other cofactors must be used to initiate radical chemistry for hydrocarbon activation. The XSS enzymes use a simple glycyl radical cofactor to initiate radical-based catalysis, which makes them members of the large glycyl radical enzyme (GRE) super- family.5 Members of the GRE superfamily are structurally comprised of a 10-stranded b/a barrel with Gly and Cys loops located within the barrel (Figure 1A). The only structurally characterized XSS to date is ben- zylsuccinate synthase (BSS), which catalyzes the formation of benzylsuccinate from fumarate and toluene.6,7 Once a radical is formed on the Gly residue within the Gly loop of BSS, it can form a transient thiyl radical on 1Department of Biology, Massachusetts Institute of Technology, Cambridge, MA 02139, USA 2Department of Biological Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, USA 3Howard Hughes Medical Institute, Massachusetts Institute of Technology, Cambridge, MA 02139, USA 4Department of Chemistry, Massachusetts Institute of Technology, Cambridge, MA 02139, USA 5Center for Environmental Health, Massachusetts Institute of Technology, Cambridge, MA 02139, USA 6Bio-inspired Solar Energy Program, Canadian Institute for Advanced Research (CIFAR), Toronto, ON M5G 1M1, Canada 7Lead contact *Correspondence: cdrennan@mit.edu https://doi.org/10.1016/j.isci. 2023.106902 iScience 26, 106902, June 16, 2023 ª 2023 The Author(s). This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). 1 ll OPEN ACCESS iScience Article Figure 1. Overview of BSS mechanism X-ray crystallography of benzylsuccinate synthase (BSS) has provided insight into the molecular mechanisms involved in fumarate addition to toluene.6,7 (A) Snapshots of the active site of BSS (PDB: 5BWE)7 show how substrates are positioned for radical catalysis. The glycyl radical cofactor is harbored within the Gly loop (yellow), which is proximal to the Cys loop (purple). (B) The glycyl radical is proposed to be in equilibrium with a thiyl radical formed on the Cys residue within the Cys loop. (C) The thiyl radical is proposed to abstract a hydrogen atom from the methyl group of toluene. The resulting benzylic group can add to the alkene of fumarate, resulting in a succinate radical which can abstract a hydrogen atom from Cys to ultimately form R-benzylsuccinate and regenerate the thiyl radical.8–10 (D) Working model for glycyl radical installation based on X-ray crystal structures. Note that PDB: 5BWE is a structure with glycine, not with the glycyl radical. 1 depicts the crystal structure of BSSag (PDB: 5BWD), which contains a partially open, though still buried, glycyl radical domain (GRD, dark gray) as compared to the crystal structure of BSSabg (4). It is proposed that this slightly open conformation of BSSag (1) could be in equilibrium with a fully open conformation of the GRD (2). The fully open conformation of BSSag allows binding of the activating enzyme (AE, green) to the GRD. Once bound, the AE can install the glycyl radical cofactor, thus activating BSS for catalysis (2-> 3). The AE can dissociate and BSSb can bind to the BSSag complex to stabilize the closed conformation (4). It is proposed that a closed conformation of the enzyme is required for toluene binding and fumarate addition. the neighboring Cys residue within the Cys loop (Figure 1B). The thiyl radical selectively abstracts a hydrogen atom from the methyl group of toluene to form a benzylic radical, which can add to the alkene of fumarate in a Giese-like reaction. The newly-formed succinate radical can abstract a hydrogen atom from Cys to form a single enantiomer, R-benzylsuccinate (Figure 1C).8–10 The radicals involved in this mecha- nism, including the glycyl and thiyl radicals, are thought to be protected from off-cycle reactivity by the bar- rel structure of the GRE. In the currently available structural data for GREs, if the Gly loop is able to be modeled into electron den- sity, the Gly and Cys loops are close in proximity to one another deep within the GRE barrel (Figure 1A); however, this ‘‘closed’’ conformation of GREs cannot be the sole GRE conformation, as it does not allow for glycyl radical installation to be accomplished. Based on structural and biochemical data, it is thought that a small domain on the C-terminus of the GRE called the glycyl radical domain (GRD) must flip out of the barrel and interact with a partner protein known as an S-adenosyl methionine (AdoMet) dependent activating enzyme (AE) (Figures 1D and 1D1–1D3).6,11–13 Only one GRE-AE has been structurally character- ized to date, known as pyruvate formate lyase (PFL)-AE.14 PFL-AE contains an active site [4Fe–4S] cluster coordinated by 3 Cys residues.12 The unique Fe site can coordinate an equivalent of AdoMet.15,16 On one-electron reduction of the [4Fe–4S]2+ cluster to a [4Fe–4S]1+ cluster, the C–S bond of AdoMet can be homolytically cleaved, thus forming a 50-deoxyadenosyl radical. The exact mechanism is an active topic of research,17 but canonically it is thought that this radical can abstract an H-atom from the substrate, in this case the Gly residue of the GRE. Based on the peptide-bound crystal structure of PFL-AE, the Gly res- idue is positioned well for this H-atom transfer to occur; however, there is no way to dock this peptide onto the full ‘‘closed’’ GRE structure without severe steric clashes.14,18,19 Thus, it is proposed that ‘‘open’’ con- formations must exist of GREs in which the Gly loop moves out of the GRE active site and into the GRE-AE 2 iScience 26, 106902, June 16, 2023 iScience Article ll OPEN ACCESS active site (Figures 1D, 1D2, and 1D3).11,14 Once glycyl radical is installed, it must move back into the GRE barrel, where it can catalyze multiple rounds of turnover. Though most characterized GREs consist of a single protein subunit, there are a few exceptions that contain additional smaller subunits, including 4-hydroxyphenylacetate decarboxylase (HPAD)20 and BSS.21 HPAD contains one additional 9.5 kDa subunit, which harbors two [4Fe–4S] clusters.22 BSS contains two extra sub- units – BSSg, which is 7 kDa, and BSSb, which is 9 kDa, in addition to the 98 kDa large catalytic subunit BSSa.23 Like the extra subunit of HPAD, both BSSg and BSSb contain [4Fe–4S] clusters.24 BSSa cannot be solubly expressed in Escherichia coli without BSSg.23 The structure of BSSag shows that part of BSSg fits into a hydrophobic surface pocket of BSSa, which is most likely the reason for increased solubility when the two subunits are coexpressed.6In vivo, it is known that BSSg is necessary for organism survival on toluene,25 but the native function of this subunit and its [4Fe–4S] cluster is unknown. When comparing the structures of BSSag and BSSabg, it was observed that the Gly loop of BSSag moves out of the active site by 2 A˚ and the protein begins to partially open in a clam-shell like motion (Figure S2).6 It is proposed that the BSSag structure has captured a ‘‘partially open’’ snapshot of BSS, where the Gly loop is starting to move out of the GRE active site (Figure 1D1). Based on these movements, BSSb is proposed to play a role in the conformational changes needed to move between the ‘‘closed’’ state where the Gly loop is near the Cys loop (Figures 1D–1D4) and the ‘‘open’’ state where the Gly loop moves out of the active site completely so it can bind to the GRE-AE for cofactor installation (Figures 1D2 and 1D3).6,7 Although this working model for BSSb0s role in activation fits the structural data (Figure 1D), it has not been explored biochemically because of the inability to install the glycyl radical within BSS in vitro. Multiple at- tempts have been made to express BSS-AE from Thauera aromatica (BSS-AETa) as soluble protein in E. coli, but only very small amounts of folded protein could be obtained, even when expressed as fusion pro- teins.7,26 Activation experiments have shown that these small amounts of BSS-AE are not able to install glycyl radical on BSS.26 This problem with in vitro activation has not only hampered our understanding of XSS mechanism, which is unique among the GREs in that it includes multiple subunits, but it has also severely limited the amount of mutagenesis experiments and engineering efforts using these enzymes. A robust system for glycyl radical cofactor installation would open the door to exploring members of this environmentally and synthetically important enzyme class to gain fundamental insight into molecular mechanism and to engineering green, selective biocatalysts for organic synthesis. Ever since the discovery of BSS in the 1990’s,27,28 the radical chemistry used to carry out this challenging olefin hydroalkylation reaction has been the topic of multiple reports; however, the inability to install the radical cofactor in vitro limited the types of experiments that could be accomplished with this system. Here, we have found an XSS-AE that can be recombinantly produced in E. coli and used to form glycyl radical on its native XSS in vitro. Moreover, we have shown that it has cross-reactivity with BSS. This cross-reactivity is atypical for the GRE superfamily, as GRE-AEs are typically observed to be highly specific for their native GRE partner.12,29 We have also gleaned important insights about the roles of XSS subunits in glycyl radical formation subsequent fumarate hydroalkylation. These studies not only complement previ- ous structural data in determining the function for BSSb, but also serve as a starting point for future biochemical investigations of XSS enzymes and directed evolution campaigns. This is enabling technology that will allow the larger community to more rapidly mutate XSS enzymes to both understand how hydro- carbons are anaerobically degraded and to engineer this class for asymmetric C–C bond formation. RESULTS Finding an XSS-AE that can be produced in E. coli Many gene clusters containing putative XSSs have been discovered and sequenced from anaerobic organ- isms that are capable of degrading aromatic hydrocarbons, such as toluene. Given the reported insolubility of BSS-AETa,7,26 we wondered whether recalcitrance to heterologous expression in E. coli was a hallmark of XSS-AEs, or whether unexplored homologs could be more easily obtained as pure enzyme. We performed a BLAST search using BSS-AETa as a query sequence and chose 6 putative XSS-AEs with different se- quences (see Percent Identity Matrix in Table S1) from different organisms (Figure 2, Table S2 and Fig- ure S1). These 6 XSS-AEs, along with BSS-AETa, were cloned into a pET28a vector that included a C-terminal His6-tag. Like all AdoMet-dependent enzymes, the XSS-AEs were predicted to have 3 Cys residues that co- ordinate an active site [4Fe–4S] cluster (Figure S3). In addition, 8 other Cys residues are conserved within two CX2CX2CX3C motifs, leading us to hypothesize the existence of an additional 2 auxiliary [4Fe–4S] iScience 26, 106902, June 16, 2023 3 ll OPEN ACCESS iScience Article Figure 2. Initial screen of XSS-AEs for solubility SDS-PAGE gel of BSS-AETa and 6 XSS-AE homologs after anaerobic immobilized metal affinity chromatography (IMAC) purification. Lane numbers correspond to entry numbers in the table below. Varying amounts of soluble XSS-AEs were observed for the homologs (asterisks denote bands corresponding to XSS-AEs); however, no soluble protein was observed for BSS-AETa. Impurities in samples were observed because of the low yields of protein obtained in this initial solubility screen; however, upon optimization, pure XSS-AE can be obtained. For each XSS-AE, the native organism, % sequence identity to BSS-AETa, yield of semi-purified protein, and Fe/protein are reported. clusters (Figure S3). The domain harboring these auxiliary clusters has been reported for many GRE-AEs; however, its function is still largely unknown. We anaerobically expressed and purified the 7 XSS-AEs in par- allel. In accordance with previous reports,7,26BSS-AETa was found exclusively in inclusion bodies, with no observable soluble protein (Figure 2, Lane 7). However, the BSS-AE homologs all produced some amount of soluble enzyme, although the range varied considerably, from 0.1 to 1.4 mg of protein per L of culture (Figure 2). The amount of iron in elution fractions was analyzed to estimate how many clusters each of the homologs contained. Five XSS-AEs contained between 5 and 8.3 Fe/protein (Figure 2), which is consistent with these AEs purifying with 1–2 [4Fe–4S] clusters. One homolog, 5, contained more than 12 Fe/protein, which would mean more than 3 [4Fe–4S] clusters could be present. Based on the Cys content of 5, this seems unlikely. It seems more plausible that experimental error in this preliminary solubility screen could be inflating this number. For these studies, instead of following up on 5, which was low yielding, we wanted to continue investigation of the highest yielding homolog, 3, using electron paramagnetic resonance (EPR) spectroscopy and LC-MS (Figures 3A and 3B). This particular XSS-AE, known as 4-isopropylbenzylsuccinate synthase AE (IbsAE), is from a strain of Thauera that is known to degrade p-cymene.30,31 On optimization of expression conditions for IbsAE, the protein yield was increased to 6 mg/L. Iron content was consistent with one [4Fe–4S] cluster (4.3 Fe/protein, Figure 3E). To determine whether full reconstitution of all 3 [4Fe–4S] clusters could be accomplished, we tried reconstituting the remaining 2 clusters in vitro. Similar to previous reports on similar AEs,32 full reconstitution was not observed (7.6 Fe/protein, Figure 3E). In addition, 85– 90% of the protein precipitated or aggregated as a result of reconstitution (Figure S4). IbsAE characterization via EPR spectroscopy and AdoMet cleavage assays Our overarching goal was to produce high enough yields of an XSS-AE to determine conditions suitable for in vitro glycyl radical installation and hydroalkylation assays (Figures 3C and 3D). With this goal in mind, we wanted to know if we could use IbsAE as purified instead of reconstituting, as initial attempts to optimize the reconstitution still led to dramatic losses in yield (Figure S4). Based on previous work with other GRE- AEs, we hypothesized that the 4.3 Fe/prot we see in IbsAE corresponds to the active site [4Fe–4S] cluster. To investigate the differences between the ‘IbsAE as purified’ and ‘IbsAE reconstituted’, we turned to EPR spectroscopy. We used 5-deazariboflavin as a photoreductant, which is commonly used for GRE activation assays, to reduce the [4Fe–4S]2+ cluster(s). We observed a signal consistent with a [4Fe–4S]1+ cluster for the as purified IbsAE (Figure S5). On reduction of the reconstituted IbsAE, we observe a mixture of signals, most likely corresponding to a [4Fe–4S]1+ cluster and [3Fe–4S]1+ cluster (Figure S5). We reasoned that a 4 iScience 26, 106902, June 16, 2023 iScience Article ll OPEN ACCESS Figure 3. Overview of transformations monitored (A–D) and constructs used (E) in these studies (A) The active site [4Fe–4S] cluster of IbsAE is reduced before catalyzing coupled AdoMet cleavage/glycyl radical installation. (B) The active site [4Fe–4S]+ cluster can coordinate an equivalent of AdoMet and reductively cleave it to form methionine and 50-deoxyadenosine (dAdo). (C) If IbsAE is bound to its partner XSS when AdoMet is reductively cleaved, the essential Gly residue in the XSS can be converted to a glycyl radical through H-atom abstraction by the intermediate 50-deoxyadenosyl radical. (D) The glycyl radical of the XSS can form a transient thiyl radical on a neighboring Cys residue. This thiyl radical initiates hydrocarbon (e.g. toluene) addition to fumarate. (E) SDS-PAGE gel of purified proteins used in these studies. Lanes 1 and 2 denote the two IbsAE enzymes, ‘‘as purified IbsAE’’ which has not been reconstituted with iron and sulfide and ‘‘reconstituted IbsAE.’’ MW of IbsAE is 40.3 kDa. Lanes 3 and 5 correspond to the IBSSag and BSSag complexes, respectively. MWs of IBSSa and BSSa are 98.6 and 99.0 kDa, respectively. MWs of IBSSg and BSSg are 6.7 and 6.9 kDa, respectively. IBSSb and BSSb were purified as separate constructs, not in complex with the IBSSa and BSSa subunits. Lanes 4 and 6 correspond to IBSSb and BSSb after affinity tag removal, and their MWs are 8.3 and 9.2 kDa, respectively. stronger reductant may be necessary to reduce the [3Fe–4S]1+ cluster to the EPR silent state. When we reduce the as purified and reconstituted IbsAE with dithionite, we see signal for [4Fe–4S]1+ cluster without interfering [3Fe–4S]1+ cluster signal (Figure 4A, g-values: 1.94 and 2.01). Comparing the double integrals for the two spectra shows that the reconstituted IbsAE contains approximately double the amount of [4Fe–4S]1+ cluster, consistent with our iron quantification. Temperature studies corroborate assignment of the signal as [4Fe–4S]1+ clusters, where the signal decreases as temperature is increased from 10K to 40K (Figure S6). After verifying that we did indeed have [4Fe–4S] clusters in both IbsAE as purified and IbsAE reconstituted and that we could reduce these clusters, we assessed the enzymes’ ability to cleave AdoMet in the pres- ence and absence of the corresponding XSS, 4-isopropylbenzylsuccinate synthase (IBSS) (Figure 3B). When we incubate IbsAE with AdoMet following reduction of the [4Fe–4S] cluster, we do observe AdoMet cleavage (Figure 4B, 1.1 mM dAdo), whereas none is observed in the control without IbsAE enzyme (Table S3). Reconstitution of the auxiliary clusters in IbsAE does not affect AdoMet cleavage under these conditions (Figure 4B, 1.1 mM dAdo with as purified IbsAE and 1.4 mM dAdo with reconstituted IbsAE). Oftentimes, low levels of AdoMet cleavage are observed without the substrate bound to IbsAE, as in this particular case; however, typically AdoMet cleavage is accelerated by addition of substrate. In iScience 26, 106902, June 16, 2023 5 ll OPEN ACCESS iScience Article Figure 4. Reduction of the FeS clusters in IbsAE and AdoMet cleavage assays (A) EPR spectra of IbsAE before (as purified) and after (reconstituted) reconstitution of FeS clusters. IbsAE was incubated with dithionite (1 mM) for 1 h prior to freezing. The primary signal observed in both samples is consistent with a [4Fe–4S]1+. Conditions of measurement: T = 10 K; microwave power = 50 mW; microwave frequency = 9.37 GHz; modulation amplitude = 10 G; [IbsAE] = 60 mM. (B) AdoMet cleavage by IbsAE was quantified by measuring formation of dAdo by LCMS (n = 3). Trace amounts of dAdo were observed in controls without IbsAE (shown in Table S3). (C) Time course monitoring AdoMet cleavage by IbsAE as measured by LCMS in the presence of either IBSSag or IBSSabg (n = 3). this case, the substrate for IbsAE is the protein complex IBSS. Like BSS, IBSS contains two additional [4Fe– 4S]-containing subunits (IBSSb corresponds to BSSb and IBSSg corresponds to BSSg) in addition to the cat- alytic IBSSa subunit that harbors the glycyl radical. We obtained the genes for IBSSabg with the chaperone protein IbsE. Previous structural and proteolytic data have led to the hypothesis that BSSb could play an important role in regulating the large conformational changes that must occur to make the catalytic glycine residue physically available to BSS-AE (Figure 1D).6 For this reason, we wanted to be able to control the amount of the b subunit that we add to assays. We purified IBSSag as a single complex, as previous studies had demonstrated the a subunit does not solubly express without the g subunit.23 We separately ex- pressed and purified IBSSb and subsequently removed the N-terminal His tag with a TEV cleavage site. Repeating these assays with addition of IBSSag does lead to a large increase in AdoMet cleavage 6 iScience 26, 106902, June 16, 2023 iScience Article ll OPEN ACCESS Figure 5. Glycyl radical can be installed in IBSS by IbsAE, thus activating IBSS for catalysis (A–C) (A) Representative EPR spectrum for the glycyl radical in IBSSag. Conditions of measurement: T = 80 K; microwave power = 1.26 mW; microwave frequency = 9.37 GHz; modulation amplitude = 3 G; [IBSSag] = 50 mM; [IbsAE] = 50 mM. Activation reactions were conducted for different lengths of time and frozen for EPR analysis. Double integrals of EPR spectra were calculated using Xenon software and compared to double integrals of known concentrations of Fremy’s salt standards to calculate concentration of radical in mM. These concentrations of radical were plotted versus time to produce plots (B) (comparison of radical installation in IBSSag versus IBSSabg) and (C) (6 h time course of radical installation in IBSSag). (Figure 4B, 47.6 mM dAdo with as purified IbsAE and 46.7 mM dAdo with reconstituted IbsAE). As expected based on our working model (Figure 1D), the addition of IBSSb with IBSSag yields less dAdo product in endpoint assays (Figure 4B, 7.2 mM dAdo with as purified IbsAE and 5.4 mM dAdo with reconstituted IbsAE). The rate of AdoMet cleavage by IbsAE is also slower when IBSSb is present (Figure 4C and Table S4). Glycyl radical formation on IBSS Following validation that IbsAE is able to cleave AdoMet, we next wanted to determine whether glycyl radical within IBSS could be observed by EPR spectroscopy (Figure 3C). We tried activating IBSSag with and without IBSSb using our as purified IbsAE stock, which is missing its auxiliary [4Fe–4S] clusters. Consis- tent with AdoMet cleavage assays as well as our working model, we only observe significant quantities of glycyl radical without IBSSb (Figures 5A and 5B and Table S5). The observation that IbsAE is able to form a glycyl radical on IBSSag without full reconstitution of its auxiliary [4Fe–4S] clusters is consistent with work showing that 4-Hpad-AE can also activate its corresponding GRE without the auxiliary clusters.33 However, the persistence of the glycyl radical was significantly affected in previous studies of 4-Hpad-AE, and within 16 min, most of the radical was gone.33 Time courses of activation reactions with our as purified IbsAE demonstrate that radical persistence is not an issue with this system for at least up to 6 h (Figure 5C and Table S6). It was also found that the as purified IbsAE installed glycyl radical within IBSSag faster than the reconstituted IbsAE (Figure S7), which is convenient given our low yields of reconstituted IbsAE. Cross-reactivity is observed for glycyl radical installation on BSS To date, BSS remains the only structurally characterized XSS, and significantly more is known about the scope and mechanism for this enzyme than XSSs that function on substrates beyond toluene (e.g. IBSS). We wondered whether IbsAE could activate BSS as well. Although GRE-AEs are typically highly specific for their partner GRE,12,23,29 we do observe an EPR signal consistent with the glycyl radical when BSSag is incubated with reduced IbsAE and AdoMet (Figure 6A). We wanted to test the effects of BSSb on glycyl radical installation as well as the persistence of radical on BSSag. Activation time courses were performed for BSSag and BSSabg over the course of 4 h. Similar to IBSS, less radical is formed when BSSb is added to the reactions. Similar amounts of radical are formed on BSSag as IBSSag, and this radical persists for the timescale of the experiment (Figure 6B and Table S7). BSSb is necessary for high benzylsuccinate production After demonstrating that BSSag can be activated, we wanted to determine whether we could observe hydro- alkylation activity in vitro. We activated BSSag for 3 h and subsequently added fumarate (2 mM final conc.) and toluene (6 mM added as a solution of toluene and MeOH). BSSb was added to some reactions to test the effects of this subunit on hydroalkylation yields. We detected and quantified product formation using iScience 26, 106902, June 16, 2023 7 ll OPEN ACCESS iScience Article Figure 6. Cross-reactivity is observed between IbsAE and BSS (A and B) (A) Representative EPR spectrum for the glycyl radical in BSSag (black) and BSSabg (gray). Conditions of measurement: T = 80 K; microwave power = 1.26 mW; microwave frequency = 9.37 GHz; modulation amplitude = 3 G; [BSSag] = 50 mM; [IbsAE] = 50 mM. Activation reactions were conducted for different lengths of time and frozen for EPR analysis. Double integrals of EPR spectra were calculated using Xenon software and compared to double integrals of known concentrations of Fremy’s salt standards to calculate concentration of radical in mM. These concentrations of radical were plotted versus time to produce plot (B), comparing radical installation in BSSag versus BSSabg. high-resolution QToF-LCMS. Nine replicates of each reaction condition were conducted in parallel. In reac- tions with BSSag and BSSabg, a peak with the same retention time and the same exact mass as a benzylsuc- cinate (BS) authentic standard was observed (Figure 7). Yields increased dramatically in the presence of BSSb (Figure 7 and Table S8, from 0.7% to 92.3% assay yield). Control reactions without BSSag produced no detect- able BS in 7 of 9 samples and trace levels of BS in 2 of 9 samples (Figure 7 and Table S8). DISCUSSION In this work, we set out to solve the long-standing problem of in vitro glycyl radical cofactor installation in BSS. Having an in vitro system is useful for both probing the molecular mechanism of XSSs as well as for developing XSSs as biocatalysts. XSSs are crucial to anaerobic hydrocarbon degradation within microbes, and thus understanding how these enzymes work within their native cellular environments remains an important question. By developing an in vitro activation method, we were able to explore the molecular mechanism of BSS activation and catalysis in ways that were not previously possible. Prior biochemical in- vestigations and crystal structures showed that BSS contains two accessory subunits, each with a [4Fe–4S] cluster bound.6,23 Here, we explore hypotheses regarding the function of one of these subunits – BSSb. Based on crystallographic data, we previously proposed that BSSb is needed to control conformational dy- namics of the glycyl radical domain and plug the hydrocarbon substrate channel.6,7 XSSs have to position hydrocarbon substrates for radical catalysis and, unlike substrates of single-component GREs, XSS sub- strates have no functional handles to help control positioning. Moreover, GREs are in equilibrium between an ‘‘open’’ state, where GRE-AE can install glycyl radical, and a ‘‘closed’’ state, where catalysis can occur. It is proposed that changing conditions, such as GRE-AE concentration, can shift this equilibrium.6,11 Our hy- pothesis was that BSSb binding shifts the equilibrium of BSS to the closed state, which allows for tighter control over hydrocarbon binding. Based on this model, we would expect two clear observations: (1) when BSSb is present, AE should not be able to activate BSS as well, and (2) BSSb should be necessary for catalysis. We wanted to test this hypothesis biochemically, but we could not obtain soluble BSS-AE. Here, we use genome mining to find a soluble XSS-AE with cross-reactivity for BSS. When we tested the role of BSSb on activation and catalysis, we do indeed observe that BSSb inhibits activation and is necessary for catalysis. Many more questions remain about the molecular details of XSS activation and hydroalkylation activity, which can now be more readily probed. For example, the roles of the metalloclusters within the accessory subunits are still poorly understood. In BSSg, when the [4Fe–4S] cluster is removed (either through metal chelation or mutagenesis), it appears to dissociate from BSSa.23 BSSg binding is known to enhance BSSa solubility when heterologously expressed in E. coli. The current data point to a model where the [4Fe–4S] cluster is necessary for BSSg to adopt a conformation to bind BSSa, and this binding is necessary to plug a hydrophobic patch on BSSa for solubility. We are still unsure if this is BSSg0s native role, or if these are artifacts of heterologous expression. Moreover, there is no proposed role for BSSb0s [4Fe–4S] cluster. Marsh et al. showed that they can create a mutant of BSSb that does not bind iron but were not able to 8 iScience 26, 106902, June 16, 2023 iScience Article ll OPEN ACCESS Figure 7. Activated BSSabg can catalyze the addition of toluene to fumarate High resolution LCMS was used to monitor formation of benzylsuccinate. A standard curve was prepared for benzylsuccinate using commercially available benzylsuccinate and L-tryptophan as an internal standard. Assay yields were determined by integrating the EIC spectrum for benzylsuccinate and calculating yield using the standard curve, with fumarate as the limiting reagent (n = 9). assess its effects on activation or catalysis.23 Using our in vitro system, these experiments could be revisited to determine the role of BSSb0s [4Fe–4S] cluster. Beyond arylalkyl-succinate synthases like BSS, there are numerous alkyl-succinate synthases that functionalize saturated hydrocarbons. Even less is known about these enzymes that can directly and selectively functionalize saturated alkanes; for example, the sub- unit/cofactor architecture of these enzymes is still unknown and is thought to include an additional sub- unit.34 Could we use a similar genome mining approach to find soluble alkyl-SS-AEs as well? XSSs also hold the potential to be useful synthetic tools for building small molecules. They use abundant feedstocks (hydrocarbons and olefins) to form new Csp3–Csp3 bonds using radical hydroalkylation. Hydroal- kylation chemistry is an attractive method for forming C–C bonds as it has the potential to set multiple stereo- centers at once; however, controlling the stereoselectivity remains challenging. Inside radical enzymes, substrate positioning can provide control over stereoselectivity, for example, toluene addition to fumarate to form exclusively R-benzylsuccinate in BSS. Beyond BSS, other XSSs exist that are able to perform this chem- istry using a wide range of hydrocarbons, including saturated hydrocarbons. Given that the transformations they catalyze could prove so useful, why have XSSs not been widely explored as biocatalysts? One major hur- dle has been activating the glycyl radical cofactor. Until now, activation has only been accomplished in whole cells. Purification of activated BSS from whole cells results in rapid loss of hydroalkylation activity.23 The inability to generate pure, activated BSS has limited the types of studies that can be done, including charac- terization of variants made through mutagenesis. Recently, other groups have also developed tools to circumvent these issues. In 2021, Heider et al. developed a heterologous expression and activation system in Aromatoleum evansii, which importantly cannot degrade toluene, to assess BSS variants.26 In addition to providing insight into the mechanism of substrate recognition, Heider et al. showed that the olefin sub- strate is not restricted to dicarboxylic acids.26 Even more recently, Cirino et al. developed a heterologous system for producing alkylsuccinates in E. coli using a BSS homolog, allowing even more rapid access to sub- strate scope studies.34 Although these tools will accelerate the development of XSSs as biocatalysts, a key limitation still existed – in vitro activation and subsequent hydroalkylation using purified enzymes. Whole cell activation for screening of XSS variants is attractive from a high-throughput standpoint. However, with the ability to activate in vitro, we now can conduct reactions using purified enzymes to verify findings from whole cell screening and rationalize the effects of key mutations. By combining approaches, mutagenesis studies can much more rapidly be accomplished for this enzyme class. XSSs have fascinated and challenged scientists for decades, and numerous studies5–7,21,23–28,34 have helped to shed light on their mechanism. Methods developments reported here and described above are likely to rapidly accelerate our understanding of XSS enzyme mechanisms, XSS substrate scope, and enable XSS protein engineering efforts. We are excited to see how these future efforts will reshape the way we view XSSs and what we understand about them. Limitations of the study This study focuses on the development of methodology for studying XSSs but does not explore substrate scope for these enzymes. Also, it was shown that reconstitution of auxiliary clusters within the ferredoxin iScience 26, 106902, June 16, 2023 9 ll OPEN ACCESS iScience Article domain of IbsAE is not necessary for enzyme activity, but the function of this domain and its clusters was not thoroughly investigated. In addition, in Figure 2, initial yields and Fe content of XSS-AE homologs were calculated using the protein fractions shown in the gel in Figure 2, which do contain impurities that affect the accuracy of the calculated yield and Fe content. Because the results in Figure 2 were meant to serve as an initial screen to determine which homolog was the highest yielding, expression and purification of only IbsAE was optimized for improved yield and purity. STAR+METHODS Detailed methods are provided in the online version of this paper and include the following: d KEY RESOURCES TABLE d RESOURCE AVAILABILITY B Lead contact B Materials availability B Data code and availability d EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS d METHOD DETAILS B Construction of expression plasmids B Expression and purification of constructs B EPR spectroscopy of [4Fe–4S] clusters B Activations to install glycyl radical B EPR spectroscopy to quantitate glycyl radical B LCMS/MS assays d QUANTIFICATION AND STATISTICAL ANALYSIS SUPPLEMENTAL INFORMATION Supplemental information can be found online at https://doi.org/10.1016/j.isci.2023.106902. ACKNOWLEDGMENTS We would like to thank the Marsh lab for sending us BSS and BSS-AETa plasmids. This work was completed in part with resources at the MIT Department of Chemistry Instrumentation Facility with the help of John Grimes, Walt Massefski, and Mohan Kumar. This work was supported in part by National Institutes of Health (NIH) grants R35 GM126982 (C.L.D.), F32 GM129882 (M.C.A.), and K99 GM145910 (M.C.A.). This work was also completed in part with resources at the MIT Center for Environmental Health Sciences core facility, which is funded by a core center grant P30- ES002109 from the National Institute of Environmental Health Sciences, NIEHS. C.L.D. is a Howard Hughes Medical Institute (HHMI) Investigator and a fellow of the Bio- inspired Solar Energy Program, Canadian Institute for Advanced Research. D.T.K.R. and B.G.B. were funded by the MIT UROP office. The content is solely the responsibility of the authors and does not neces- sarily represent the official views of the National Institutes of Health. AUTHOR CONTRIBUTIONS Experimental work was carried out by M.C.A., D.T.K.R., C.N.I., and B.G.B. under the direction of C.L.D. The manuscript was written by M.C.A. and C.L.D. with input from all other authors. DECLARATION OF INTERESTS The authors declare no competing interests. INCLUSION AND DIVERSITY One or more of the authors of this paper self-identifies as an underrepresented ethnic minority in their field of research or within their geographical location. One or more of the authors of this paper self-identifies as a gender minority in their field of research. One or more of the authors of this paper self-identifies as a member of the LGBTQIA+ community. One or more of the authors of this paper self-identifies as living with a disability. One or more of the authors of this paper received support from a program designed to increase minority representation in their field of research. 10 iScience 26, 106902, June 16, 2023 iScience Article Received: April 19, 2023 Revised: May 8, 2023 Accepted: May 12, 2023 Published: May 19, 2023 REFERENCES ll OPEN ACCESS 1. Wartell, B., Boufadel, M., and Rodriguez- Freire, L. (2021). An effort to understand and improve the anaerobic biodegradation of petroleum hydrocarbons: a literature review. Int. Biodeterior. Biodegrad. 157, 105156. https://doi.org/10.1016/j.ibiod.2020.105156. 11. Peng, Y., Veneziano, S.E., Gillispie, G.D., and Broderick, J.B. (2010). Pyruvate formate-lyase, evidence for an open conformation favored in the presence of its activating enzyme. J. Biol. Chem. 285, 27224–27231. https://doi.org/10. 1074/jbc.M109.096875. 20. Selmer, T., and Andrei, P.I. (2001). p- Hydroxyphenylacetate decarboxylase from Clostridium difficile. A novel glycyl radical enzyme catalysing the formation of p-cresol. Eur. J. Biochem. 268, 1363–1372. https://doi. org/10.1046/j.1432-1327.2001.02001.x. 2. Laczi, K., Erdeine´ Kis, A´ ., Szila´ gyi, A´ ., Bounedjoum, N., Bodor, A., Vincze, G.E., Kova´ cs, T., Ra´ khely, G., and Perei, K. (2020). New frontiers of anaerobic hydrocarbon biodegradation in the multi-omics era. Front. Microbiol. 11, 590049. 3. Zhang, Y., Zhai, X., Guan, F., Dong, X., Sun, J., Zhang, R., Duan, J., Zhang, B., and Hou, B. (2022). Microbiologically influenced corrosion of steel in coastal surface seawater contaminated by crude oil. Npj Mater. Degrad. 6, 35. https://doi.org/10.1038/ s41529-022-00242-4. 4. Widdel, F., and Musat, F. (2019). Diversity and common principles in enzymatic activation of hydrocarbons: an introduction. In Aerobic Utilization of Hydrocarbons, Oils, and Lipids, F. Rojo, ed. (Springer International Publishing), pp. 3–32. https://doi.org/10. 1007/978-3-319-50418-6_50. 5. Backman, L.R.F., Funk, M.A., Dawson, C.D., and Drennan, C.L. (2017). New tricks for the glycyl radical enzyme family. Crit. Rev. Biochem. Mol. Biol. 52, 674–695. https://doi. org/10.1080/10409238.2017.1373741. 6. Funk, M.A., Judd, E.T., Marsh, E.N.G., Elliott, S.J., and Drennan, C.L. (2014). Structures of benzylsuccinate synthase elucidate roles of accessory subunits in glycyl radical enzyme activation and activity. Proc. Natl. Acad. Sci. USA 111, 10161–10166. https://doi.org/10. 1073/pnas.1405983111. 7. Funk, M.A., Marsh, E.N.G., and Drennan, C.L. (2015). Substrate-bound structures of benzylsuccinate synthase reveal how toluene is activated in anaerobic hydrocarbon degradation. J. Biol. Chem. 290, 22398– 22408. https://doi.org/10.1074/jbc.M115. 670737. 8. Himo, F. (2002). Catalytic mechanism of benzylsuccinate synthase, a theoretical study. J. Phys. Chem. B 106, 7688–7692. https://doi. org/10.1021/jp014627v. 9. Qiao, C., and Marsh, E.N.G. (2005). Mechanism of benzylsuccinate synthase: stereochemistry of toluene addition to fumarate and maleate. J. Am. Chem. Soc. 127, 8608–8609. https://doi.org/10.1021/ ja051972f. 10. Li, L., and Marsh, E.N.G. (2006). Mechanism of benzylsuccinate synthase probed by substrate and isotope exchange. J. Am. Chem. Soc. 128, 16056–16057. https://doi. org/10.1021/ja067329q. 12. Shisler, K.A., and Broderick, J.B. (2014). Glycyl radical activating enzymes: structure, mechanism, and substrate interactions. Arch. Biochem. Biophys. 546, 64–71. https://doi. org/10.1016/j.abb.2014.01.020. 13. Knappe, J., Neugebauer, F.A., Blaschkowski, H.P., and Ga¨ nzler, M. (1984). Post- translational activation introduces a free radical into pyruvate formate-lyase. Proc. Natl. Acad. Sci. USA 81, 1332–1335. https:// doi.org/10.1073/pnas.81.5.1332. 14. Vey, J.L., Yang, J., Li, M., Broderick, W.E., Broderick, J.B., and Drennan, C.L. (2008). Structural basis for glycyl radical formation by pyruvate formate-lyase activating enzyme. Proc. Natl. Acad. Sci. USA 105, 16137–16141. https://doi.org/10.1073/pnas.0806640105. 15. Krebs, C., Broderick, W.E., Henshaw, T.F., Broderick, J.B., and Huynh, B.H. (2002). Coordination of adenosylmethionine to a unique iron site of the [4Fe-4S] of pyruvate formate-lyase activating enzyme: a Mo¨ ssbauer spectroscopic study. J. Am. Chem. Soc. 124, 912–913. https://doi.org/10. 1021/ja017562i. 16. Walsby, C.J., Hong, W., Broderick, W.E., Cheek, J., Ortillo, D., Broderick, J.B., and Hoffman, B.M. (2002). Electron-nuclear double resonance spectroscopic evidence that S -adenosylmethionine binds in contact with the catalytically active [4Fe(cid:2)4S] + cluster of pyruvate formate-lyase activating enzyme. J. Am. Chem. Soc. 124, 3143–3151. https:// doi.org/10.1021/ja012034s. 17. Byer, A.S., Yang, H., McDaniel, E.C., Kathiresan, V., Impano, S., Pagnier, A., Watts, H., Denler, C., Vagstad, A.L., Piel, J., et al. (2018). Paradigm shift for radical S-Adenosyl- L-methionine reactions: the organometallic intermediate U is central to catalysis. J. Am. Chem. Soc. 140, 8634–8638. https://doi.org/ 10.1021/jacs.8b04061. 21. Leuthner, B., Leutwein, C., Schulz, H., Ho¨ rth, P., Haehnel, W., Schiltz, E., Scha¨ gger, H., and Heider, J. (1998). Biochemical and genetic characterization of benzylsuccinate synthase from Thauera aromatica: a new glycyl radical enzyme catalysing the first step in anaerobic toluene metabolism. Mol. Microbiol. 28, 615–628. https://doi.org/10.1046/j.1365- 2958.1998.00826.x. 22. Martins, B.M., Blaser, M., Feliks, M., Ullmann, G.M., Buckel, W., and Selmer, T. (2011). Structural basis for a Kolbe-type decarboxylation catalyzed by a glycyl radical enzyme. J. Am. Chem. Soc. 133, 14666– 14674. https://doi.org/10.1021/ja203344x. 23. Li, L., Patterson, D.P., Fox, C.C., Lin, B., Coschigano, P.W., and Marsh, E.N.G. (2009). Subunit structure of benzylsuccinate synthase. Biochemistry 48, 1284–1292. https://doi.org/10.1021/bi801766g. 24. Hilberg, M., Pierik, A.J., Bill, E., Friedrich, T., Lippert, M.-L., and Heider, J. (2012). Identification of FeS clusters in the glycyl- radical enzyme benzylsuccinate synthase via EPR and Mo¨ ssbauer spectroscopy. J. Biol. Inorg. Chem. 17, 49–56. https://doi.org/10. 1007/s00775-011-0828-1. 25. Coschigano, P.W. (2002). Construction and characterization of insertion/deletion mutations of the tutF, tutD, and tutG genes of Thauera aromatica strain T1. FEMS Microbiol. Lett. 217, 37–42. https://doi.org/10.1111/j. 1574-6968.2002.tb11453.x. 26. Salii, I., Szaleniec, M., Zein, A.A., Seyhan, D., Sekuła, A., Schu¨ hle, K., Kaplieva-Dudek, I., Linne, U., Meckenstock, R.U., and Heider, J. (2021). Determinants for substrate recognition in the glycyl radical enzyme benzylsuccinate synthase revealed by targeted mutagenesis. ACS Catal. 11, 3361– 3370. https://doi.org/10.1021/acscatal. 0c04954. 18. Becker, A., Fritz-Wolf, K., Kabsch, W., 27. Biegert, T., Fuchs, G., and Heider, J. (1996). Knappe, J., Schultz, S., and Volker Wagner, A.F. (1999). Structure and mechanism of the glycyl radical enzyme pyruvate formate-lyase. Nat. Struct. Biol. 6, 969–975. https://doi.org/ 10.1038/13341. 19. Leppa¨ nen, V.M., Merckel, M.C., Ollis, D.L., Wong, K.K., Kozarich, J.W., and Goldman, A. (1999). Pyruvate formate lyase is structurally homologous to type I ribonucleotide reductase. Structure 7, 733–744. https://doi. org/10.1016/S0969-2126(99)80098-7. Evidence that anaerobic oxidation of toluene in the denitrifying bacterium Thauera aromatica is initiated by formation of benzylsuccinate from toluene and fumarate. Eur. J. Biochem. 238, 661–668. https://doi. org/10.1111/j.1432-1033.1996.0661w.x. 28. Beller, H.R., and Spormann, A.M. (1998). Analysis of the novel benzylsuccinate synthase reaction for anaerobic toluene activation based on structural studies of the product. J. Bacteriol. 180, 5454–5457. iScience 26, 106902, June 16, 2023 11 iScience Article (2022). Engineering Escherichia coli for anaerobic alkane activation: biosynthesis of (1-methylalkyl)succinates. Biotechnol. Bioeng. 119, 315–320. https://doi.org/10. 1002/bit.27956. (2018). 35. McCarthy, E.L., and Booker, S.J. Biochemical approaches for understanding iron–sulfur cluster regeneration in Escherichia coli synthase during catalysis. Methods Enzymol. 606, 217–239. https://doi.org/10. 1016/bs.mie.2018.06.006. lipoyl 36. Backman, L.R., Huang, Y.Y., Andorfer, M.C., Gold, B., Raines, R.T., Balskus, E.P., and Drennan, C.L. (2020). Molecular basis for catabolism of the abundant metabolite trans-4-hydroxy-L-proline by a microbial glycyl radical enzyme. Elife 9, e51420. https://doi.org/10.7554/eLife. 51420. ll OPEN ACCESS 29. Frey, M., Rothe, M., Wagner, A.F., and Knappe, J. (1994). Adenosylmethionine- dependent synthesis of the glycyl radical in pyruvate formate-lyase by abstraction of the glycine C-2 pro-S hydrogen atom. Studies of [2H]glycine-substituted enzyme and peptides homologous to the glycine 734 site. J. Biol. Chem. 269, 12432–12437. https://doi.org/10.1016/S0021-9258(18) 99892-3. 30. Harms, G., Rabus, R., and Widdel, F. (1999). Anaerobic oxidation of the aromatic plant hydrocarbon p-cymene by newly isolated denitrifying bacteria. Arch. Microbiol. 172, 303–312. https://doi.org/10.1007/ s002030050784. 31. Strijkstra, A., Trautwein, K., Jarling, R., Wo¨ hlbrand, L., Do¨ rries, M., Reinhardt, R., Drozdowska, M., Golding, B.T., Wilkes, H., and Rabus, R. (2014). Anaerobic activation of p-cymene in denitrifying betaproteobacteria: methyl group hydroxylation versus addition to fumarate. Appl. Environ. Microbiol. 80, 7592–7603. https://doi.org/10.1128/AEM. 02385-14. 32. Selvaraj, B., Pierik, A.J., Bill, E., and Martins, B.M. (2013). 4-Hydroxyphenylacetate decarboxylase activating enzyme catalyses a classical S-adenosylmethionine reductive cleavage reaction. J. Biol. Inorg. Chem. 18, 633–643. https://doi.org/10.1007/s00775- 013-1008-2. 33. Selvaraj, B., Pierik, A.J., Bill, E., and Martins, B.M. (2014). The ferredoxin-like domain of the activating enzyme is required for generating a lasting glycyl radical in 4-hydroxyphenylacetate decarboxylase. J. Biol. Inorg. Chem. 19, 1317–1326. https:// doi.org/10.1007/s00775-014-1189-3. 34. Wang, Y., Nguyen, N., Lee, S.H., Wang, Q., May, J.A., Gonzalez, R., and Cirino, P.C. 12 iScience 26, 106902, June 16, 2023 iScience Article STAR+METHODS KEY RESOURCES TABLE REAGENT or RESOURCE Bacterial and virus strains E. coli DH5a cells E. coli T7 Express cells Chemicals, peptides, and recombinant proteins Benzylsuccinate potassium nitrosodisulfonate S-Adenosyl methionine iron standard Ferene Oligonucleotides ll OPEN ACCESS SOURCE IDENTIFIER New England BioLabs New England BioLabs C2987H C2566H Sigma Aldrich Sigma Aldrich Sigma Aldrich Alfa Aesar Sigma Aldrich CAS 884-33-3 CAS 14293-70-0 CAS 86867-01-8 EINECS 231-714-2 CAS 79551-14-7 N/A N/A N/A N/A N/A N/A N/A All primers used are reported in ‘‘method details’’ and Table S2 Sigma Aldrich Recombinant DNA All purchased XSS-AE genes are reported in Table S2 XSS genes are reported in ‘‘method details’’ BSS-AETa gene Software and algorithms Xenon Software Clustal2.1 MassHunter Software RESOURCE AVAILABILITY Lead contact Twist Biosciences Twist Biosciences Ref #23, Marsh lab Bruker Ref #1 in supplemental information Agilent Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Catherine L. Drennan (cdrennan@mit.edu). Materials availability All unique plasmids generated in this study are available from the lead contact without restriction. No unique reagents were generated. Data code and availability All information required to reanalyze the data in this report is presented in the Supporting Information or from the lead contact upon request. EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS Cell lines used in this study, E. coli DH5a cells and T7 Express cells, were purchased from New England BioLabs. Growth conditions are reported in the method details section below. METHOD DETAILS Construction of expression plasmids BSS-AETa homologs The genes for the six BSS-AETa homologs were identified through a BLAST search using the amino acid sequence of BSS-AETa. Results were narrowed down based on literature precedent (i.e. the native iScience 26, 106902, June 16, 2023 13 ll OPEN ACCESS iScience Article organism had been characterized as an anaerobic aromatic hydrocarbon degrader) and sequence similarity (i.e. genes with very high similarity to one another were excluded, see Table S1). The amino acid sequences for the 6 genes (see Table S1 for gene identifiers) were used for codon optimization for expression in E. coli K12. The six genes were purchased from Twist Bioscience, where they were cloned into pET28a at restriction sites BamHI and HindIII. The resulting plasmids contained both N-ter- minal and C-terminal hexa-His-tags. The N-terminal His-tag was removed from all AE constructs using Q5(cid:2) Site-Directed Mutagenesis Kit (New England Biolabs) using the primers in Table S2. All primers were designed using NEBaseChanger(cid:3). All genes were confirmed through Sanger sequencing by Quin- tara Biosciences. BSS-AETa We received the gene for BSS-AETa from the Marsh lab.23 The BSS-AETa gene was amplified and overhangs forward primer, 50- were added with complementarity to pET28a using the following primers: CTTTAAGAAGGAGATATACCATGAAAATTCCATTAGTCAC-30 and reverse primer, 50-TCGAGTGCGG CCGCAAGCTTCCTTTTCGGGTGGGTCTCTT-30. The pET28a vector was amplified and overhangs were added with complementarity to BSS-AETa using the following primers: forward primer, 50- AAGAGACCC ACCCGAAAAGGAAGCTTGCGGCCGCACTCGA-30 and reverse primer, 50- GTGACTAATGGAATTTTCA TGGTATATCTCCTTCTTAAAG-30. PCRs were conducted using the Q5(cid:2) Site-Directed Mutagenesis Kit (New England Biolabs). PCR products were purified over 1% agarose gels using a Qiagen gel extraction kit. Gibson assembly reactions using NEBuilder(cid:2) HiFi DNA assembly were set up at 50(cid:3)C for 1 hour with an insert:vector ratio of 3:1. A small aliquot (5 mL) of the reaction was transformed into E. coli DH5a cells (New England BioLabs). The resulting construct was verified through Sanger sequencing by Quintara Biosciences. IBSS and BSS The plasmids used to express BSSabg were published previously. Briefly, BSSa and BSSg were cloned into a pET-DUET vector into sites NdeI/KpnI and NcoI/HindIII, respectively. BSSa was C-terminally His6 tagged. BSSb and TutH were cloned into a pRSF-DUET plasmid into sites NdeI/XhoI and NcoI/HindIII. For expres- sion of BSSag with the TutH chaperone, the b-subunit was removed from the pRSF-DUET plasmid using the Q5(cid:2) Site-Directed Mutagenesis Kit (New England Biolabs) and the following primers: forward primer, 50- CTCGAGTCTGGTAAAGAAAC-30 and reverse primer, 50- ATTTCGATTATGCGGCCG-30. His6-BSSb, which included a TEV cleavage site after the N-terminal His6 tag, was constructed using the Q5(cid:2) Site-Directed Mutagenesis Kit (New England Biolabs) and the following primers: forward primer, 50- AACGACC GAGAATCTTTATTTTCAGGGATCCGAGGGCAGCAACATGGAA-30 and reverse primer, 50- GGATCG TGATGGTGATGGTGATGGCTGCTAGCCATATGTATATCTCCTTCTTATACTTAACTAATATAC-30. For the IBSS complex, the amino acid sequences (IBSSa - UniProt ID: A0A096ZNX3, IBSSb - UniProt ID: A0A096ZP03, IBSSg - UniProt ID: A0A096ZNX6) and putative chaperone protein (IbsE – UniProt ID: A0A096ZNY2) were used for codon optimization for expression in E. coli K12. The genes were purchased from Twist Bioscience as linear g-blocks with overhangs for Gibson Assembly. A pET-DUET plasmid was amplified using the Q5(cid:2) Site-Directed Mutagenesis Kit (New England Biolabs) and the following primers: forward primer, 50-AGCGCAGCTTAATTAACCT-30 and reverse primer, 50-GGTATATCTCCTTCTTAAAGT TAAACAA-30. PCR product was purified over 1% agarose gels using a Qiagen gel extraction kit. C-termi- nally His6 tagged IBSSa and IBSSg were assembled into the linearized pET-DUET vector using NEBuilder(cid:2) HiFi DNA assembly kit. A pRSF-DUET plasmid was amplified using the Q5(cid:2) Site-Directed Mutagenesis Kit (New England Biolabs) and the following primers: forward primer, 50-CTCGAGTCTGGTAAAGAAAC-30 and reverse primer, 50-CCATGGTATATCTCCTTATTAAAG-30. PCR product was purified over 1% agarose gels using a Qiagen gel extraction kit. IbsE and IBSSb were assembled into the linearized pET-RSF vector using NEBuilder(cid:2) HiFi DNA assembly kit. For expression of IBSSag, the IBSSb was removed from the pRSF-DUET plasmid using the Q5(cid:2) Site-Directed Mutagenesis Kit (New England Biolabs) and the following primers: forward primer, 50- CTCGAGTCTGGTAAAGAAAC-30 and reverse primer, 50- TTAGACACGCGCTTT TGC-30. His6-IBSSb, which included a TEV cleavage site after the N-terminal His6 tag, was constructed using the Q5(cid:2) Site-Directed Mutagenesis Kit (New England Biolabs) and the following primers: forward primer, 50- AACGACCGAGAATCTTTATTTTCAGGGATCCGCTAATGTGCAGACCCAG-30 and reverse primer, 50- CAATTCATATTCTTCCTCTATATGTATACCGATCGTCGGTAGTGGTAGTGGTAGTGCTAGG-30. All BSS and IBSS constructs were verified through Sanger sequencing by Quintara Biosciences. 14 iScience 26, 106902, June 16, 2023 ll OPEN ACCESS iScience Article Expression and purification of constructs XSS-AE solubility screen XSS-AE constructs were transformed into T7 Express cells (New England BioLabs) and a single colony was used to make a glycerol stock of each. Starter cultures were inoculated from glycerol stocks and grown overnight in LB containing 50 mg/mL kanamycin for each XSS-AE at 37(cid:3)C at 220 rpm. LB media (1L) contain- ing 50 mg/mL kanamycin, 150 mg iron(II) ammonium sulfate hexahydrate (CAS: 783-85-9), and 47 mg L-cysteine was inoculated with 10 mL of starter culture. Expression cultures were grown at 37(cid:3)C at 220 rpm to an OD600 = 0.8, at which point they were induced with 1 mM IPTG (GoldBio). Induced cultures were expressed for 4 h at 22(cid:3)C at 100 rpm. Cells were pelleted by centrifugation, flash frozen in liquid ni- trogen, and stored at (cid:2)80(cid:3)C until lysis. Cell lysis and protein purification were performed anaerobically in an MBraun chamber. All buffers were sparged with argon before use. For lysis of cells, cell paste was resus- pended in 15 mL lysis buffer (lysis buffer: 50 mM HEPES pH 8.0, 300 mM NaCl, 2 EDTA-free protease inhib- itor pellet (cOmplete, Roche Diagnostics), 100 mg lysozyme (Sigma Aldrich), and 8 mL benzonase (EMD Millipore). Cells were resuspended by mashing cell paste with a spatula. Resuspended cells were incubated for 30 min at 4(cid:3)C, after which cells were sonicated for a 1 min cycle of 2 s on and 15 s off at an amplitude of 10 (Qsonica). Lysate was clarified by centrifugation for 45 min at 28,000 g and subsequently filtered (0.22 mm) before purification. XSS-AEs were purified in parallel on 0.5 mL of TALON resin, which was gravity-packed into 2 mL plastic spin columns (Thermo Scientific(cid:3) Pierce(cid:3) Centrifuge Columns). Columns were equili- brated with 10 mL equilibrations buffer (50 mM HEPES pH 8.0, 300 mM NaCl) before passing cell lysate through by gravity. Columns were washed with 10 mL of wash buffer (50 mM HEPES pH 8.0, 300 mM NaCl, 5 mM imidazole) and eluted into new 15 mL falcon tubes with 4 mL of elution buffer (50 mM HEPES pH 8.0, 300 mM NaCl, 100 mM imidazole). Concentration of protein in each eluent was determined using a Bradford assay. Iron quantification was conducted using a ferene assay35 (ferene purchased from Sigma Aldrich, CAS 79551-14-7) and iron standards (EINECS 231-714-2). IbsAE large scale expression and purification After optimization of IbsAE expression conditions, the following protocol was found to yield the highest amounts of IbsAE. Starter cultures were inoculated from glycerol stocks and grown overnight in LB contain- ing 50 mg/mL kanamycin at 37(cid:3)C at 220 rpm. Expression cultures were inoculated with 10 mL of starter culture per 1 L of TB containing 50 mg/mL kanamycin, 150 mg iron(II) ammonium sulfate hexahydrate (CAS: 783-85-9), and 47 mg L-cysteine. Eight liters total of culture were grown per round of expression and purification, split into 1 L cultures in 2.5 L flasks. Expression cultures were grown at 37(cid:3)C at 220 rpm to an OD600 = 0.8, at which point they were induced with 1 mM IPTG (GoldBio). Induced cultures were ex- pressed overnight (16–20 h) at 22(cid:3)C at 100 rpm. Cells were pelleted by centrifugation, flash frozen in liquid nitrogen, and stored at (cid:2)80(cid:3)C until lysis. Cell lysis and protein purification were performed anaerobically in an MBraun chamber. All buffers were sparged with argon before use. For lysis of cells, cell paste from 2 L of culture was resuspended in 25 mL lysis buffer (lysis buffer: 50 mM HEPES pH 8.0, 300 mM NaCl), with an EDTA-free protease inhibitor pellet (cOmplete, Roche Diagnostics), lysozyme (1 mg lysozyme/ml buffer, Sigma Aldrich), and 2 mL benzonase (EMD Millipore). Cells were resuspended by mashing cell paste with a spatula. Resuspended cells were incubated for 30 min at 4(cid:3)C, after which cells were sonicated for 2 3 1 min cycles of 2 s on and 15 s off at an amplitude of 10 (Qsonica). Lysate was clarified by centrifugation for 45 min at 28,000 g and subsequently filtered (0.22 mm) before purification. IbsAE was purified on 6 mL of TALON resin, which was gravity-packed into two 10 mL plastic spin columns (Thermo Scientific(cid:3) Pierce(cid:3) Centrifuge Columns). Columns were equilibrated with 30 mL equilibration buffer (50 mM HEPES pH 8.0, 300 mM NaCl) before passing cell lysate through by gravity. Columns were washed with 30 mL of wash buffer (50 mM HEPES pH 8.0, 300 mM NaCl, 5 mM imidazole) and eluted into new 50 mL falcon tubes with (cid:4)20 mL of elution buffer (50 mM HEPES pH 8.0, 300 mM NaCl, 100 mM imidazole). IbsAE was buffer exchanged into 50 mM HEPES pH 8.0, 300 mM NaCl, concentrated to (cid:4)300–500 mM, aliquoted and flash frozen. Concentration of protein was determined using a Bradford assay. Iron quantification was conducted using a ferene assay35 and iron standards (EINECS 231-714-2). Reconstitution of IbsAE IbsAE purified with an intact active site cluster following the purification protocol described above. For reconstitution of the auxiliary clusters, (cid:4)2 mL of purified IbsAE ((cid:4)100 mM) was reconstituted at a time. IbsAE was thawed in an MBraun chamber at 4(cid:3)C at which point DTT was added to a final concentration of 10 mM and was incubated for an hour. Five molar equivalents of Fe(III)Cl3 were added to the protein, iScience 26, 106902, June 16, 2023 15 ll OPEN ACCESS iScience Article which was immediately mixed. Five molar equivalents of Na2S were added to the protein, which was imme- diately mixed. The reconstitution was allowed to incubate for 30 minutes; then 5 more equivalents of both Fe(III)Cl3 and Na2S were added as described above. The reconstitution was incubated for 2 more hours, was spun at 14,000 g for 10 minutes, and was filtered (0.22 mm). After filtration, reconstituted IbsAE was purified by size exclusion chromatography on an S200 16/60 column (50 mM HEPES pH 8.0, 300 mM NaCl, 1 mM DTT). The monomer peak was collected, concentrated, and flash frozen. BSS and IBSS large scale expression and purification All BSS and IBSS constructs were transformed into T7 Express cells (New England BioLabs) and a single colony was used to make a glycerol stock of each. Starter cultures were inoculated from glycerol stocks and grown overnight at 37(cid:3)C at 220 rpm in LB containing either 50 mg/mL kanamycin and 100 mg/mL ampi- cillin for IBSSag and BSSag or 50 mg/mL kanamycin for IBSSb and BSSb. Expression cultures were inocu- lated with 10 mL of starter culture per 1 L of LB containing the corresponding antibiotics, 150 mg iron(II) ammonium sulfate hexahydrate (CAS: 783-85-9), and 47 mg L-cysteine. Expression, TALON purification, and concentration were carried out anaerobically as described above in ‘‘IbsAE large scale expression and purification.’’ For IBSSb and BSSb, the N-terminal His-tag was cleaved with His-tagged TEV protease at a ratio of 10:1 (b subunit:TEV protease, w/w). The reaction was gently mixed and left at 4(cid:3)C for (cid:4)24 hours (or until >80% completion as determined by SDS-PAGE) without agitation. The reaction mixture was puri- fied on TALON resin as detailed above. Fractions containing pure IBSSb or BSSb, with the Histag removed, were pooled and buffer exchanged into 50 mM HEPES pH 8.0, 300 mM NaCl. EPR spectroscopy of [4Fe–4S] clusters Reduction of the [4Fe–4S] clusters In many cases, flavin derivatives are used to reduce the active site cluster of GRE-AEs to initiate glycyl radical installation. Two flavin derivatives, acriflavine36 and 5-deazariboflavin,11 were tested for their ability to reduce the active site cluster of as purified IbsAE. In a Coy anaerobic chamber, IbsAE (60 mM) was incu- bated with either acriflavine and bicine (100 mM and 50 mM, respectively) or deazariboflavin (100 mM) in acti- vation buffer (20 mM Tris pH 7.5, 100 mM KCl) for 30 minutes. Most IbsAE precipitated out of solution when acriflavine was added, so it was no longer pursued as a photoreductant. IbsAE remained in solution with deazariboflavin and an EPR signal consistent with a [4Fe–4S]1+ cluster was observed. When the protocol using deazariboflavin was used to reduce the reconstituted IbsAE, a mix of signals was observed, corresponding to [4Fe–4S]1+ cluster and [3Fe–4S]1+ cluster. Reductions for both as purified IbsAE and re- constituted IbsAE were repeated with dithionite (1 mM final concentration) and incubated for an hour. Di- thionite-reduced samples produced primarily signals consistent with [4Fe–4S]1+ clusters. EPR parameters EPR spectra were collected in a Bruker EMX-Plus spectrometer at temperatures between 10–40 K with a Bruker/ColdEdge 4K waveguide cryogen-free cryostat. Xenon 1.1b.155 software was used to collect and process spectra. Spectra were recorded at 9.37 GHz with a modulation amplitude of 10 G, microwave po- wer of 50 mW, and a 100 kHz modulation frequency. A center field of 3500 G, a sweep time of 60 s, and a sweep width of 2000 G were used. Each spectrum shown is an average of 10 scans. The double integrals of the two spectra in Figure 4A were calculated using Xenon software and compared to one another to determine the relative amount of [4Fe–4S]1+ cluster in each. Activations to install glycyl radical In a Coy anaerobic chamber, reduction reactions were conducted by combining activation buffer (20 mM Tris pH 7.5, 100 mM KCl), 5-deazariboflavin (200 mM final conc.), DTT (2 mM final conc.), and IbsAE (100 mM final conc.). The reduction was gently mixed and illuminated using an LED light for 30 minutes. The reduc- tion was diluted with activation buffer such that the final concentration of IbsAE was 50 mM. IBSSag or BSSag (50 mM final conc.), IBSSb or BSSb (0 or 50 mM final conc.), and AdoMet (1.5 mM final conc., Sigma Aldrich CAS 86867-01-8) were added and the reaction was gently mixed. Reactions were conducted at room temperature without agitation by the LED lamp for 0.3–6 hours, at which point they were either used in AdoMet cleavage assays or hydroalkylation reactions, or anaerobically frozen in liquid nitrogen for EPR spectroscopy. 16 iScience 26, 106902, June 16, 2023 iScience Article ll OPEN ACCESS EPR spectroscopy to quantitate glycyl radical EPR spectra of the glycyl radical was collected at 80 K. Spectra were recorded at 9.37 GHz with a modulation amplitude of 3 G, microwave power of 1.26 mW, and a 100 kHz modulation frequency. A center field of 3350 G, a sweep time of 21 s, and a sweep width of 200 G were used. Each spectrum shown is an average of 10 scans. Potassium nitrosodisulfonate (Fremy’s salt, Sigma Aldrich) was used as a standard. The double integrals of each spectrum were calculated using Xenon software and compared to the double integrals obtained from Fremy’s standard to obtain concentrations of glycyl radical. LCMS/MS assays Product formation in AdoMet cleavage and hydroalkylation assays was quantified using a Q-TOF LC/MS (Agilent 6545 mass spectrometer coupled to an Agilent Infinity 1260 liquid chromatography system) and a Zorbax reversed-phase column (300SB-C18, 3.5 mm, 2.1 3 50 mm, Agilent). Solvent A was H2O with 0.1% acetic acid, and solvent B was acetonitrile with 0.1% acetic acid. Flow rate was 0.4 mL/min. The LC method for all assays was as follows: 0–2 min, 1% B; 2–4 min, gradient from 1 to 50% B; 4–6 min, gradient from 50 to 100% B; 6–7 min, 100% B; 7–8 min, gradient from 100 to 1% B. AdoMet cleavage Cluster reductions and AdoMet cleavage reactions were conducted as described above. Endpoint assays were conducted for 2 hours and time courses were conducted for 1.5–120 minutes. Reactions were quenched with one volume of methanol and 100 mM of L-tryptophan was added as an internal standard. Quenched reactions were removed from the Coy and protein was pelleted by centrifugation. The resulting supernatant was filtered through a 0.22 mm filter and used for LC/MS analysis. The LC method described above was used with the MS in positive ion mode. Extracted ion counts for 50-deoxyadenosine (dAdo) and L-tryptophan were obtained, and the concentration of product was determined using a standard curve made from known amounts of dAdo and L-Trp. Hydroalkylation Cluster reductions and glycyl radical installation were conducted as described above. BSSb was not added to glycyl radical installation reactions. Three hours after glycyl radical installation reactions were initiated, fumarate (2 mM final conc.) and toluene (6 mM final conc. added as a stock solution in MeOH, 3% v/v) were added. Reactions were diluted such that the final conc. of BSSag was 40 mM, and BSSb (40 mM final conc.) was also added to some reactions. Control reactions contained all components except BSSag. Reactions were conducted in a Coy anaerobic chamber in a 96-well microtiter plate with a final volume of 25 mL for each reaction. Hydroalkylation reactions were quenched with two volumes of methanol and 100 mM of L-tryptophan was added as an internal standard. Quenched reactions were removed from the Coy and pro- tein was pelleted by centrifugation. The resulting supernatant was diluted 20-fold and filtered through a 0.22 mm filter and used for LC/MS analysis. The LC method described above was used with the MS in nega- tive ion mode. The retention times for fumarate, L-Trp, and benzylsuccinate were 1.218, 4.393, and 5.240 min, respectively. Product concentration was determined by the extracted ion count ratio of benzyl- succinate and internal standard L-Trp, multiplied by response factor 0.21, which was established via a cali- bration curve with known amounts of benzylsuccinate (Sigma Aldrich, CAS 884-33-3) and L-Trp. The assay yield was defined as 100x[BS]/2 mM, where 2 mM represents the initial concentration of the limiting re- agent, fumarate. QUANTIFICATION AND STATISTICAL ANALYSIS In Figure 4B/Table S3 and Figure 4C/Table S4, number of assays conducted for each condition was equal to 3 (n = 3), and the mean and the standard deviations were calculated in Excel. In Figure 7/Table S8, number of assays conducted for each condition was equal to 9 (n = 9), and the mean and the standard deviations were calculated in Excel. iScience 26, 106902, June 16, 2023 17
10.7554_elife.85413
ReSeaRCH aRtICLe X- chromosome target specificity diverged between dosage compensation mechanisms of two closely related Caenorhabditis species Qiming Yang1,2†, Te- Wen Lo1,2†‡, Katjuša Brejc1,2, Caitlin Schartner1,2§, Edward J Ralston1,2, Denise M Lapidus1,2, Barbara J Meyer1,2* 1Howard Hughes Medical Institute, Berkeley, United States; 2Department of Molecular and Cell Biology, University of California, Berkeley, Berkeley, United States Abstract An evolutionary perspective enhances our understanding of biological mechanisms. Comparison of sex determination and X- chromosome dosage compensation mechanisms between the closely related nematode species Caenorhabditis briggsae (Cbr) and Caenorhabditis elegans (Cel) revealed that the genetic regulatory hierarchy controlling both processes is conserved, but the X- chromosome target specificity and mode of binding for the specialized condensin dosage compensation complex (DCC) controlling X expression have diverged. We identified two motifs within Cbr DCC recruitment sites that are highly enriched on X: 13 bp MEX and 30 bp MEX II. Mutating either MEX or MEX II in an endogenous recruitment site with multiple copies of one or both motifs reduced binding, but only removing all motifs eliminated binding in vivo. Hence, DCC binding to Cbr recruitment sites appears additive. In contrast, DCC binding to Cel recruitment sites is synergistic: mutating even one motif in vivo eliminated binding. Although all X- chromosome motifs share the sequence CAGGG, they have otherwise diverged so that a motif from one species cannot function in the other. Functional divergence was demonstrated in vivo and in vitro. A single nucleotide position in Cbr MEX can determine whether Cel DCC binds. This rapid divergence of DCC target specificity could have been an important factor in establishing reproductive isolation between nematode species and contrasts dramatically with the conservation of target specificity for X- chromosome dosage compensation across Drosophila species and for transcription factors controlling developmental processes such as body- plan specification from fruit flies to mice. Editor's evaluation This important study uses state- of- the- art methods to explore the evolution of dosage compensa- tion between two closely related nematode species. The evidence supporting the rapid evolution of the recruitment motifs on the X chromosome, despite a general conservation of the dosage compensation machinery, is compelling. This work will be of broad interest to cell biologists and evolutionary biologists. Introduction Comparative studies have shown that different facets of metazoan development exhibit remark- ably different degrees of conservation across species (Carroll, 2008). At one extreme, homeobox- containing Hox genes and Wnt- pathway signaling genes play conserved roles in body plan formation (Hox) and cell- fate determination, neural patterning, or organogenesis (Wnt) across clades diverged *For correspondence: bjmeyer@berkeley.edu †These authors contributed equally to this work Present address: ‡Department of Biology, Ithaca College, Ithaca, United States; §Roche Diagnostics, Santa Clara, Canada Competing interest: See page 35 Funding: See page 35 Preprinted: 05 December 2022 Received: 07 December 2022 Accepted: 21 March 2023 Published: 23 March 2023 Reviewing Editor: Luisa Cochella, Johns Hopkins University School of Medicine, United States Copyright Yang, Lo et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 1 of 39 Research article by more than 600 million years (MYR) (Malicki et al., 1990; De Kumar and Darland, 2021; Rim et al., 2022). Distant orthologous genes within these ancestral pathways can substitute for each other. For example, both the mouse Small eye (Pax- 6) gene (Hill et al., 1991) and the fruit fly eyeless (ey) gene (Quiring et al., 1994; Halder et al., 1995) control eye morphogenesis and encode a transcription factor that includes a paired domain and a homeodomain. Ectopic expression of mouse Pax- 6 in different fruit fly imaginal disc primordia can induce morphologically normal ectopic compound eye structures on fruit fly wings, legs, and antennae (Halder et  al., 1995). Hence, at a deep level, eye morphogenesis is under related genetic and molecular control in vertebrates and insects, despite profound differences in eye morphology and mode of development. At the other extreme are aspects of development related to sex. For example, chromosomal strat- egies to determine sexual fate in mice, fruit flies, and nematodes (XY or XO males and XX females or hermaphrodites) and the mechanism needed to compensate for the consequent difference in X- chro- mosome dose between sexes have diverged greatly. To balance X gene expression between sexes, female mice randomly inactivate one X chromosome (Yin et al., 2021; Loda et al., 2022), while male fruit flies double expression of their single X chromosome (Samata and Akhtar, 2018; Rieder et al., 2019), and hermaphrodite worms halve expression of both X chromosomes (Meyer, 2022a; Meyer, 2022b). The divergence in these pathways is so great that comparisons among animals of the same genus can provide useful evolutionary context for understanding the developmental mechanisms that distin- guish the sexes. Therefore, we determined the genetic and molecular specification of sexual fate and X- chromosome dosage compensation in the nematode C. briggsae and compared it to the wealth of knowledge amassed about these processes in C. elegans. These two species have diverged by 15–30 MYR (Cutter, 2008). In C. elegans, the sex determination and dosage compensation pathways are linked by genes that coordinately control both processes. For example, in XX embryos, the switch gene sdc- 2 sets the sex determination pathway to the hermaphrodite mode and triggers the binding of a DCC onto both X chromosomes to reduce X gene expression by half and thereby match X expression with that from XO males (Meyer, 2022a). The DCC shares subunits with condensin, a protein complex that controls the structure, resolution, and segregation of mitotic and meiotic chromosomes from yeast to humans (Yatskevich et al., 2019; Meyer, 2022b). We determined the extent to which the sex- specific gene regulatory hierarchy is conserved between C. elegans and C. briggsae and the extent to which subunits of the C. briggsae DCC corre- spond to those of the C. elegans DCC. We also defined the cis- acting regulatory sites that confer X- chromosome specificity and recruit the C. briggsae DCC. We found that the DCC itself and the regulatory hierarchy that determines sex and directs the DCC to X have been conserved, but remark- ably, both the X- chromosome target specificity of the C. briggsae DCC and its mode of binding to X have diverged. Results Conservation between C. briggsae and C. elegans of the core dosage compensation machinery and genetic hierarchy that regulates dosage compensation The pivotal hermaphrodite- specific regulatory protein that coordinately controls both sex determi- nation and dosage compensation in C. elegans is a 350  kDa protein called SDC- 2. It directs the DCC to both X chromosomes of XX embryos to achieve dosage compensation and also activates the hermaphrodite program of sexual differentiation (Chuang et al., 1996; Dawes et al., 1999; Chu et al., 2002; Pferdehirt et al., 2011). Loss of Cel sdc- 2 causes XX- specific lethality due to excessive X- chromosome gene expression and masculinization of escaper animals (Nusbaum and Meyer, 1989; Kruesi et al., 2013). SDC- 2 has no known homologs outside of nematodes and only a coiled- coil domain as a predicted structural feature (Meyer, 2022a). Among five Caenorhabditis species compared, the entire SDC- 2 protein has 23–29% identity and 38–45% similarity (Figure  1—figure supplement 2A). Between Cbr and Cel, the entire SDC- 2 protein shows 26% identity and 43% similarity (Figure  1—figure Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 2 of 39 Chromosomes and Gene Expression Research article supplements 1 and 2A). To assess the conservation of gene function, we deployed genome- editing technology in C. briggsae to knockout sdc- 2. Using a PCR- based molecular strategy to identify insertions and deletions induced by DNA repair following directed mutagenesis with zinc finger nucleases, we recovered several independent Cbr sdc- 2 mutant lines (Figure 1—figure supplement 3). Homozygous Cbr sdc- 2 mutations caused exten- sive XX- specific lethality, consistent with a defect in dosage compensation and the conservation of gene function (Figure 1A). Nearly all Cbr sdc- 2 hermaphrodites died as embryos or young larvae; rare XX survivors exhibited slow growth and masculinization. Cbr sdc- 2 males were viable (Figure 1A) and had wild- type body morphology. To determine whether the hermaphrodite- specific lethality of Cbr sdc- 2 mutants was caused by defects in dosage compensation, we first identified components of the C. briggsae DCC and then asked whether DCC binding to X is disrupted by mutation of Cbr sdc- 2, as it is by mutation of Cel sdc- 2. In C. elegans, five of the ten known DCC proteins are homologous to subunits of condensin, an evolutionarily conserved protein complex required to restructure and resolve chromosomes in prepa- ration for cell divisions in mitosis and meiosis (Figure 1B; Chuang et al., 1994; Lieb et al., 1996; Lieb et al., 1998; Chan et al., 2004; Tsai et al., 2008; Csankovszki et al., 2009; Mets and Meyer, 2009; Yatskevich et al., 2019; Meyer, 2022a). The evolutionary time scale over which condensin subunits were co- opted for dosage compensation in nematodes had not been explored. Several lines of evidence indicate that a condensin complex mediates dosage compensation in C. briggsae as well. First, BLASTP searches revealed C. briggsae orthologs of all known C. elegans DCC condensin subunits (Figure 1B). Alignment of DPY- 27 protein revealed 38% identity and 56% simi- larity between C. elegans and C. briggsae (Figure 1—figure supplement 2B). Immunofluorescence experiments using antibodies against Cbr DPY- 27, the SMC4 ortholog of the only Cel DCC condensin subunit (Cel DPY- 27) not associated with mitotic or meiotic condensins (Chuang et al., 1994), revealed X chromosome- specific localization in hermaphrodites, but not males, indicating conservation of func- tion (Figure 1C and Figure 2A and B). Specificity of DPY- 27 antibodies was demonstrated by Western blot analysis (Figure 1—figure supplement 4A). Second, disruption of Cbr dpy- 27 conferred hermaphrodite- specific lethality, with rare XX escaper animals exhibiting a dumpy (Dpy) phenotype, like the disruption of Cel dpy- 27 (Figure 1G). Immuno- fluorescence experiments with Cbr DPY- 27 antibodies revealed diffuse nuclear distribution of DPY- 27 in Dpy escapers of dpy- 27(y436) mutants instead of X localization, consistent with lethality (Figure 1D). Third, co- immunoprecipitation of proteins with rabbit Cbr DPY- 27 antibodies followed by SDS- PAGE and mass spectrometry of excised trypsinized protein bands identified Cbr MIX- 1 (Table  1; Materials and methods), the SMC2 condensin subunit ortholog found in the Cel DCC complex (Lieb et al., 1998; Figure 1B). Both DPY- 27 and MIX- 1 belong to the SMC family of chromosomal ATPases that dimerize and participate in condensin complexes (Figure 1B). Fourth, immunofluorescence experiments using Cbr MIX- 1 antibodies (Figure 1—figure supple- ment 4B) revealed co- localization of Cbr MIX- 1 with Cbr DPY- 27 on hermaphrodite X chromosomes (Figure 1E). Cbr MIX- 1 protein did not bind to X chromosomes in Cbr dpy- 27(y436) mutant animals (Figure  1F). Instead, MIX- 1 exhibited diffuse nuclear distribution, like DPY- 27, consistent with the two proteins participating in a complex and the dependence of MIX- 1 on DPY- 27 for its binding to X (Figure 1F). These data demonstrate that condensin subunits play conserved roles in the dosage compensation machinery of both C. briggsae and C. elegans. In contrast to DPY- 27, MIX- 1 shows 55% identity and 72% similarity between C. elegans and C. briggsae. Not only does MIX- 1 participate in the DCC, it also participates in two other distinct Caenor- habditis condensin complexes that are essential for the proper resolution and segregation of mitotic and meiotic chromosomes (Mets and Meyer, 2009; Csankovszki et al., 2009). Conserved roles in chromosome segregation complexes would constrain MIX- 1 sequence divergence, thereby explaining its greater conservation between species. Evidence that DCC binding defects underlie the XX- specific lethality caused by Cbr sdc- 2 muta- tions is our finding that neither Cbr DPY- 27 (Figure 2C) nor Cbr MIX- 1 (not shown) binds to X chromo- somes in Cbr sdc- 2 mutant hermaphrodites. Instead, we found a low level of diffuse nuclear staining. Thus, the role of sdc- 2 in the genetic hierarchies that activate dosage compensation is also conserved. We next explored why maternally supplied DCC subunits fail to bind to the single X chromosome of C. briggsae males. In C. elegans XO embryos, the master switch gene xol- 1 (XO lethal) represses Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 3 of 39 Chromosomes and Gene Expression Research article A sdc-2 mutations in C. briggsae cause XX-specific lethality B C. elegans DCC C. briggsae DCC wild-type XX sdc-2(y467) XX (814) sdc-2(y469) XX (848) wild-type XO sdc-2(y467) XO sdc-2(y469) XO (1840) (208) (217) (262) 0 10 20 30 40 50 100 % viability (expected no. of progeny) 60 70 80 90 DAPI DPY-27 FISH MERGE X III DPY-27 MIX-1 DPY-27 MIX-1 SDC-3 DPY-21 SDC-1 DPY-30 DPY-28 DPY-26 SDC-2 CAPG-1 SDC-2 identified by BLASTP DAPI DPY-27 MERGE D ) + ( 7 2 - y p d ) 6 3 4 y ( 7 2 - y p d DAPI DPY-27 MIX-1 MERGE F DAPI MIX-1 MERGE ) + ( 7 2 - y p d ) 6 3 4 y ( 7 2 - y p d C e a s g g i r b . C e p y t d l i w e p y t d l i w E e a s g g i r b . C e p y t d l i w G Scheme to characterize dpy-27 mutants Viability of dpy-27 mutants grandmother non-Dpy mother progeny + / - genotyped + / + genotyped + / - genotyped - / - genotyped + / + + / + + / - - / - - / - 100% 80% 60% 40% 20% 0% WT very Dpy dead +/+ 15 4452 297 +/- 13 3792 292 -/- 5 28 6 +/+ 15 4477 299 +/- 10 2953 295 -/- 5 22 4 maternal genotype broods embryo progeny avg. brood size dpy-27(y436) dpy-27(y705) 632 bp deletion: 5’ UTR, exon 1, intron 1, exon 2 52 bp deletion: exon 4 Figure 1. Conservation of X- chromosome dosage compensation machinery between C. briggsae and C. elegans. (A) sdc- 2 mutations cause XX- specific lethality in C. briggsae. Graph shows percent viability of wild- type and Cbr sdc- 2 mutant XX and XO adults. Viability of homozygous XX and hemizygous XO Cbr sdc- 2 mutants is expressed as the percentage of live adults for each karyotype relative to the number expected (shown in parentheses) in the progeny of a cross if all mutant animals were viable. Crosses and calculations are described in Materials and methods. Sequence changes of Figure 1 continued on next page Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 4 of 39 Chromosomes and Gene Expression Research article Figure 1 continued sdc- 2 mutations derived from genome editing using zinc- finger nucleases are shown in Figure 1—figure supplement 3A. (B) Schematic of the C. elegans dosage compensation complex (left) and C. briggsae orthologs identified by BLASTP (right). The C. elegans dosage compensation complex (DCC) includes homologs of all core condensin subunits (MIX- 1, DPY- 27, DPY- 26, DPY- 28, and CAPG- 1). C. briggsae DCC components identified and characterized in this study are shown in color; other orthologs are in gray. DPY- 27 and MIX- 1 belong to the SMC (Structural Maintenance of Chromosomes) family of chromosomal ATPases. Each has nucleotide- binding domains (NBDs) at its N- and C- termini that are linked by two long coiled- coil domains separated by a hinge domain. Each SMC protein folds back on itself to form a central region of two anti- parallel coiled coils flanked by the NBDs and the hinge. DPY- 27 and MIX- 1 dimerize through interactions between their hinge domains and their NBD domains. The globular NBDs bind to the three non- SMC condensin DCC subunits (DPY- 26, DPY- 28, and CAPG- 1) (See Meyer, 2022a). (C) Condensin subunit DPY- 27 binds X chromosomes and mediates dosage compensation in C. briggsae. Confocal images of C. briggsae hermaphrodite gut nuclei co- stained with the DNA dye DAPI (gray), antibodies to Cbr DPY- 27 (green), and FISH probes to either 5% of X (red, top), or 1% of chromosome III (red, bottom) show that Cbr DPY- 27 co- localizes with X but not III, consistent with a role in dosage compensation. Scale bars, 1 μm. (D) Confocal images of C. briggsae gut nuclei from dpy- 27(+) or dpy- 27(y436) mutant XX adult hermaphrodites co- stained with DAPI (blue) and the Cbr DPY- 27 rabbit antibody (red). DPY- 27 shows subnuclear localization in a dpy- 27(+) gut nucleus (top), as expected for X localization. The mutant gut nucleus (bottom) shows diffuse nuclear distribution of DPY- 27, as anticipated for a mutant SMC- 4 condensin ortholog that lacks most of the N- terminal part of the ATPase domain and, therefore, has no ATP binding or hydrolysis. Scale bars, 1 μm. (E) Confocal images of a C. briggsae gut nucleus from wild- type adult hermaphrodites co- stained with DAPI (gray) and antibodies to Cbr DPY- 27 (green) and Cbr MIX- 1 (red) show that Cbr MIX- 1 co- localizes with Cbr DPY- 27 on X in wild- type hermaphrodites. Scale bars, 1 μm. (F) Association of Cbr MIX- 1 (red) with X found in a dpy- 27(+) nucleus (top) is disrupted in a Cbr dpy- 27(y436) nucleus (bottom), in accord with participation of Cbr MIX- 1 in a protein complex with Cbr DPY- 27. Scale bars, 1 μm. (G) Viability of dpy- 27 mutant XX C. briggsae animals. The left panel shows the genetic scheme to characterize the effect of maternal genotype on viability of dpy- 27 null XX mutants. Comparison is made between homozygous null dpy- 27 progeny from heterozygous or homozygous non- Dpy mutant mothers. The genotype of non- DPY mothers was established through PCR analysis. The right panel shows the percent viability of progeny from wild- type hermaphrodites and heterozygous or homozygous dpy- 27 mutant hermaphrodites. The maternal genotype, number of broods, total number of embryo progeny from all broods, and average brood size are provided for two null alleles of dpy- 27. Molecular characterization of mutations is shown below the graph and in Figure 1—figure supplement 3B. Almost all progeny of dpy- 27 null mutant mothers are dead; a homozygous dpy- 27 null strain cannot be propagated. More than 20% of progeny of dpy- 27/+heterozygous mutant mothers are very Dpy or dead, indicating that a wild- type DPY- 27 maternal contribution has minimal effect on suppressing the deleterious effect of the homozygous null zygotic genotype. The complete XX lethality is consistent with a major role for condensin subunit DPY- 27 in dosage compensation. The online version of this article includes the following source data and figure supplement(s) for figure 1: Figure supplement 1. Protein sequence alignment comparing SDC- 2 proteins in C. elegans and C. briggsae. Figure supplement 2. Conservation of SDC- 2 and DPY- 27 proteins in the Caenorhabditis genus. Figure supplement 3. DNA sequence changes mediated by genome editing. Figure supplement 4. Specificity of Cbr DPY- 27 and MIX- 1 antibodies. Figure supplement 4—source data 1. Source data for DPY- 27 and MIX- 1 antibody specificity. the hermaphrodite- specific sdc- 2 gene required for DCC binding to X and thereby prevents other DCC subunits from functioning in males (Miller et al., 1988; Rhind et al., 1995; Dawes et al., 1999; Meyer, 2022a). Loss of Cel xol- 1 activates Cel sdc- 2 in XO embryos, causing DCC binding to X, reduction in X- chromosome gene expression, and consequent death. We isolated the null mutant allele Cbr xol- 1(y430) by PCR screening of a C. briggsae deletion library (Supplementary file 1). We found that the Cbr xol- 1 mutation caused inappropriate binding of the DCC to the single X of XO embryos (Figure 2D) and fully penetrant male lethality (Figure 3B), as expected from the disruption of a gene that prevents the DCC machinery from functioning in C. briggsae males. Cbr xol- 1 mutant XX hermaphrodites appeared wild- type. To investigate the hierarchical relationship between Cbr xol- 1 and Cbr sdc- 2, we asked whether a Cbr sdc- 2 mutation could suppress the male lethality caused by a Cbr xol- 1 mutation. Both genes are closely linked in C. briggsae, prompting us to use genome editing technology to introduce de novo mutations in cis to pre- existing lesions without relying on genetic recombination between closely linked genes. If Cbr xol- 1 controls Cbr sdc- 2, then mutation of Cbr sdc- 2 should rescue the male lethality of Cbr xol- 1 mutants (Figure 2E). This prediction proved to be correct. XO males were observed among F1 progeny from mated Cbr xol- 1 hermaphrodites injected with ZFNs targeting Cbr sdc- 2 (Figure 3A, B and D). Insertion and deletion mutations were found at the Cbr sdc- 2 target site in more than twenty tested F1 males (examples are in Figure 1—figure supplement 3C and D). Quantification of male viability in four different xol- 1 sdc- 2 mutant lines revealed nearly full rescue (Figure 3B), with a concomitant absence of DCC binding on the single X chromosome (Figure  2E). Therefore, Cbr xol- 1 functions upstream of Cbr sdc- 2 to repress it and thereby prevents DCC binding to the male X Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 5 of 39 Chromosomes and Gene Expression Research article Figure 2. Conserved genetic hierarchy targets the C. briggsae dosage compensation complex (DCC) to the X chromosomes of hermaphrodites. (A–E) Schematic depiction of the genetic hierarchy controlling sex- specific DCC recruitment to C. briggsae X chromosomes (left) paired with representative immunofluorescence experiments exemplifying DCC localization (right). Scale bars, 5 μm. Gut nuclei (A, B, C, E) or embryos (D) were co- stained with DAPI (red) and antibodies to Cbr DPY- 27 (green). In wild- type XX, but not XO gut nuclei (A, B), DPY- 27 co- localizes with X chromosomes, consistent Figure 2 continued on next page Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 6 of 39 Chromosomes and Gene Expression Research article Figure 2 continued with a role for condensin subunit DPY- 27 in dosage compensation (see also Figure 1C). (C) SDC- 2 is required for recruitment of DPY- 27 to the X chromosomes of hermaphrodites. Failure of the DCC to bind X chromosomes of sdc- 2 XX mutants underlies the XX- specific lethality. Shown is the gut nucleus of a rare XX sdc- 2 mutant escaper near death. sdc- 2 mutant XX escaper animals are masculinized. (D) Lethality of Cbr xol- 1(y430) XO animals corresponds to inappropriate binding of the DCC to the single X in embryos. (E) Mutation of the DCC recruitment factor Cbr sdc- 2 in a Cbr xol- 1 XO mutant prevents DCC recruitment to X and suppresses the XO lethality. See Figure 3B for quantification. chromosome. In summary, not only is the core condensin dosage compensation machinery conserved between Caenorhabditis species, but so also are the key features of the genetic hierarchy that confers sex- specificity to the dosage compensation process. Conservation between C. briggsae and C. elegans of the genetic hierarchy that regulates early stages of sex determination Mechanisms controlling sex determination and differentiation are dynamic over evolutionary time; major differences can exist even within an individual species. For example, males within the house fly species Musca domestica can utilize one of many different male- determining factors on autosomes and sex chromosomes to determine sex depending on a factor’s linkage to other beneficial traits (Meisel et al., 2016). Within the Caenorhabditis genus, similarities and differences occur in the genetic pathways governing the later stages of sex determination and differentiation (Haag, 2005). For example, three sex- determination genes required for C. elegans hermaphrodite sexual differentiation but not dosage compensation, the transformer genes tra- 1, tra- 2, and tra- 3, are conserved between C. elegans and C. briggsae and play very similar roles. Mutation of any one gene causes virtually identical masculinizing somatic and germline phenotypes in both species (Kelleher et al., 2008). Moreover, the DNA binding motif for both Cel and Cbr TRA- 1 (Berkseth et al., 2013), a Ci/GL1 zinc- finger transcription factor that acts as the terminal regulator of somatic sexual differentiation (Zarkower and Hodgkin, 1992), is conserved between the two species. At the opposite extreme, the mode of sexual reproduction, hermaphroditic versus male/female, dictated the genome size and reproductive fertility of Caenorhabditis species diverged by only Table 1. MALDI- TOF identification of Cbr MIX- 1 peptides. m/z Submitted MH+ Matched Delta ppm Peptide Missed Cleavage Database Sequence 916.47 916.46 1163.59 1214.65 1224.63 1263.74 1285.69 1350.69 1881.97 1886.89 2064.01 2377.18 1163.58 1214.66 1224.62 1263.74 1285.69 1350.70 1881.98 1886.91 2064.00 2377.16 9.5 3.3 –3.6 8.8 –0.87 –2.8 –8.9 –2.3 –6.8 3.4 5.6 674–680 375–384 631–641 713–723 524–534 631–641 656–666 134–150 86–101 460–477 385–415 0 1 0 1 0 0 0 0 0 0 1 (K)YHENVVR(L) (K)LRGELEGMSR(G) (R)VLIESQCLPGR(R) (R)EVAYTDGVKSR(T) (R)DVEGLVLHLIR(L) (R)VLIESQCLPGR(R) (R)YTIINDQSLQR(A) (R)GVGLNVNNPHFLIMQGR(I) (K)QSPFGMDHLDELVVQR(H) (K)ITQQVQSLGYNADEDVQR(R) (R)GTVTNDKGEHVSLETYIQETR(A) This table lists the mass- to- charge ratio (m/z) of measured peptides, the predicted masses (MH+ Matched), and the deviation from predicted masses (Delta ppm). The ID of each measured peptide is described by the residue range within full- length MIX- 1 (Peptide) and its corresponding amino acid sequence (Database Sequence). The number of uncut tryptic peptide bonds is listed for each peptide (Missed Cleavage). In addition to MIX- 1, MALDI- TOF analysis of excised protein bands in the molecular weight range of condensin subunits excised from an SDS- PAGE gel revealed peptides corresponding to four common high- molecular weight contaminants: the three vitellogenin yolk proteins VIT- 2, VIT- 4, VIT- 5, and CBG14234, an ortholog of VIT- 4. No protein bands corresponding to the molecular weights of SDC- 2 or SDC- 3 were visible on the SDS- PAGE gel. Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 7 of 39 Chromosomes and Gene Expression Research article Figure 3. sdc- 2 controls dosage compensation and sex determination in C. briggsae. (A) Diagram of the screening strategy to recover Cbr sdc- 2 mutations as suppressors of the XO- specific lethality caused by a xol- 1 mutation. Cbr xol- 1 XX hermaphrodites were mated with males carrying a gfp- marked X chromosome to allow F1 XO males to be monitored for the parental origin of the X chromosome. Animals with mating plugs (indicating Figure 3 continued on next page Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 8 of 39 Chromosomes and Gene Expression Research article Figure 3 continued successful mating) were injected with mRNAs to sdc- 2 zinc- finger nucleases, and all F1 males were examined for GFP fluorescence. Non- green males necessarily inherited an X chromosome carrying a Cbr- xol- 1 mutation and, assuming conservation of the dosage compensation complex (DCC) regulatory hierarchy, would be inviable without a concomitant Cbr sdc- 2 mutation. GFP- positive males arose at low frequency from fertilization of nullo- X oocytes (caused by non- disjunction of the maternal X chromosome) with gfp- X- bearing sperm. These false positives were discarded from further study. (B) Cbr sdc- 2 mutations rescue Cbr xol- 1(y430) XO lethality. Graph shows percent viability of wild- type XO animals and mutant XO animals carrying combinations of Cbr xol- 1 and Cbr sdc- 2 mutations. The % XO viability is expressed as the percentage of live XO adults relative to the number expected (shown in parentheses) in the progeny of the cross. Formulae for viability calculations are given in the Materials and methods. Sequence changes of sdc- 2 mutations are shown in Figure 1—figure supplement 3C and D. (C) sdc- 2 activates the program for Cbr hermaphrodite sexual development. DIC images show the comparison of tail morphologies for Cbr L4 animals of different genotypes. sdc- 2 mutations, but not dpy- 27 mutations, cause masculinization of XX animals. Scale bar, 20 μm. (D) DIC images show tail morphologies of wild- type or doubly mutant Cbr adults. An sdc- 2 mutation suppresses both the XO lethality and feminization caused by a xol- 1 mutation, consistent with a role for sdc- 2 in controlling both dosage compensation and sex determination. xol- 1 sdc- 2 XO animals are viable, fertile males, indicating that the sdc- 2 mutation suppressed the lethality and feminization caused by xol- 1 mutations in XO animals. A dpy- 27 mutation suppresses the XO lethality but not feminization caused by a xol- 1 mutation, consistent with a role for dpy- 27 in dosage compensation but not sex determination. dpy- 27; xol- 1 XO animals are fertile hermaphrodites. Scale bar, 20 μm. 3.5 million years (Yin et  al., 2018; Cutter et  al., 2019). Species that evolved self- fertilization (e.g. C. briggsae or C. elegans) lost 30% of their DNA content compared to male/female species (e.g. C. nigoni or C. remanei), with a disproportionate loss of male- biased genes, particularly the male secreted short (mss) gene family of sperm surface glycoproteins (Yin et  al., 2018). The mss genes are necessary for sperm competitiveness in male/female species and are sufficient to enhance it in hermaphroditic species. Thus, sex has a pervasive influence on genome content. In contrast to these later stages of sex determination and differentiation, the earlier stages of sex determination and differentiation had not been analyzed in C. briggsae. Therefore, we asked whether xol- 1 and sdc- 2 control sexual fate as well as dosage compensation in C. briggsae, as they do in C. elegans, over the 15–30 MYR that separates them. Our analysis of Cbr sdc- 2 XX mutant phenotypes revealed intersexual tail morphology in the rare animals that survived to the L3/L4 stage (Figure 3C), indicating a role for Cbr sdc- 2 in sex determination. Sexual transformation to the male fate was unlikely to have resulted from a disruption in dosage compensation since such transforma- tion was never observed in Cbr dpy- 27 XX mutants (Figure  3C). Analysis of sexual phenotypes in double mutant strains confirmed that Cbr sdc- 2 controls sex determination. Specifically, Cbr xol- 1 Cbr sdc- 2 double mutant XO animals develop as males, whereas Cbr dpy- 27; Cbr xol- 1 double mutant XO animals develop as hermaphrodites (Figure 3C and D). That is, both Cbr sdc- 2 and Cbr dpy- 27 mutations suppress the XO lethality caused by a xol- 1 mutation, but only Cbr sdc-2 mutations also suppress the sexual transformation of XO animals into hermaphrodites. These results show that both sdc- 2 and dpy- 27 function in C. briggsae dosage compensation, but only sdc- 2 also functions in sex determination. Thus, the two master regulatory genes that control the earliest stages of both sex determination and X- chromosome dosage compensation, xol- 1 and sdc- 2, are conserved between C. briggsae and C. elegans. DCC recruitment sites isolated from C. briggsae X chromosomes fail to bind the C. elegans DCC Discovery that the dosage compensation machinery and the gene regulatory hierarchy that controls sex determination and dosage compensation are functionally conserved between C. briggsae and C. elegans raised the question of whether the cis- acting regulatory sequences that recruit dosage compensation proteins to X chromosomes are also conserved. In C. elegans, the DCC binds to recruit- ment elements on X (rex) sites and then spreads across X to sequences lacking autonomous recruit- ment ability (Csankovszki et al., 2004; Jans et al., 2009; Pferdehirt et al., 2011; Albritton et al., 2017; Anderson et  al., 2019). Within rex sites, combinatorial clustering of three DNA sequence motifs directs synergistic binding of the DCC (Fuda et al., 2022). To compare X- recruitment mech- anisms between species, DNA binding sites for the Cbr DCC recruitment protein SDC- 2 and the Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 9 of 39 Chromosomes and Gene Expression Research article Cbr DCC condensin subunit DPY- 27 were defined by chromatin immuno- precipitation experiments followed by sequencing of captured DNA (ChIP- seq experiments) (Figure  4A). SDC- 2 sites were obtained with anti- FLAG antibodies from a genome- engineered Cbr strain encoding a FLAG- tagged version of endogenous SDC- 2. DPY- 27 sites were obtained from either a wild- type Cbr strain with DPY- 27 antibodies or from a genome- engineered strain encoding endogenous FLAG- tagged DPY- 27 with anti- FLAG antibodies. A consistent set of twelve large, overlapping SDC- 2 ChIP- seq peaks and DPY- 27 ChIP- seq peaks emerged from the studies (Figure 4A), representing less than one- fourth the number of DCC peaks than on the C. elegans X chromosome, which is smaller (17.7 Mb for Cel vs. 21.5 Mb for Cbr). SDC- 2 and DPY- 27 binding to autosomes was indistinguishable from that of the IgG control (Figure  4— figure supplement 1A and B). To determine whether DNA from these peaks acts as autonomous recruitment sites that confer X- chromosome target specificity to the dosage compensation process, we conducted DCC recruitment assays in vivo (Figure 4B). Assays were modeled on rex assays devel- oped for C. elegans (Materials and methods and Fuda et al., 2022). Embryos carrying extrachromo- somal arrays composed of multiple copies of DNA from a single ChIP- seq peak were stained with DPY- 27 antibodies and a FISH probe to the array. DPY- 27 localized to 80–90% of extrachromosomal arrays carrying DNA from each of the individual peaks (Figure 4C and E and Table 2A). In contrast, extrachromosomal arrays made from three regions of X lacking DCC binding in ChIP- seq experiments showed minimal recruitment (0–6% of nuclei with arrays) (Figure  4E and Table  2A). In strains with arrays comprised of Cbr DCC binding sites, the X chromosomes rarely exhibited fluorescent signal, because the arrays titrated the DCC from X (Figure 4C). The titration was so effective that brood sizes of array- bearing hermaphrodites were very low, and hermaphrodite strains carrying arrays could not be maintained. Thus, the twelve high- occupancy Cbr DCC binding sites identified by ChIP- seq were named recruitment elements on X (rex sites) (Table 3), like the C. elegans DCC binding sites, due to their ability to recruit the DCC when detached from X. To determine whether rex sites from C. briggsae and C. elegans had functional overlap in DCC binding specificity, we asked whether a rex site from one species could recruit the DCC from the other. We made extrachromosomal arrays in C. elegans with DNA from C. briggsae rex sites and extrachromosomal arrays in C. briggsae with DNA from C. elegans rex sites. Arrays in C. elegans with C. briggsae rex sites failed to recruit the Cel DCC or to titrate the Cel DCC from Cel X chromosomes (Figure 4C, Cbr rex- 8), indicating evolutionary divergence in rex sites between the two Caenorhab- ditis species. Reciprocally, extrachromosomal arrays made in C. briggsae with Cel rex sites failed to bind the Cbr DCC or titrate it from the Cbr X, confirming divergence in rex sites (Cel rex- 33 in Figure 4D; Cel rex- 33 and Cel rex- 4 in Table 2B). In contrast, controls showed that 100% of extrach- romosomal arrays made in C. elegans with DNA from either Cel rex- 33 or Cel rex- 4 recruited the Cel DCC (Table 2B). Because X chromosomes and extrachromosomal arrays have different topologies, histone modifi- cations, DNA binding proteins, and positions within nuclei, we devised a separate assay to assess the divergence of rex sites between species in a more natural chromosomal environment. We inserted six Cbr rex sites with a range of ChIP- seq scores into a location on the endogenous Cel X chromo- some that lacked DCC binding (15, 574, 674 bp) (Figure 5 and Table 3). Proof of principle for the experiment came from finding that insertion of Cel rex- 32, a high- affinity Cel DCC binding site, into the new location on X resulted in DCC binding that was not significantly different from binding at its endogenous location on X (p=0.2, Figure 5). All Cbr rex sites except rex- 1, which will be discussed later, failed to recruit the Cel DCC when inserted into Cel X chromosomes, confirming the divergence of rex sites between species. Identification of motifs on Cbr X chromosomes that recruit the Cbr DCC To understand the mechanisms underlying the selective recruitment of the Cbr DCC to X chromo- somes, but not autosomes, and the basis for the divergence in X- chromosome targeting between Caenorhabditis species, we searched for DNA sequence motifs that are enriched in the twelve Cbr rex sites (Figure 6—figure supplement 1A) using the website- based program called Multiple Em for Motif Elicitation (MEME) (Version 5.4.1) (Bailey and Elkan, 1994; Bailey et al., 2015) and compared them to motifs in C. elegans rex sites important for recruiting the Cel DCC to X (Figure 6A and B). Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 10 of 39 Chromosomes and Gene Expression Research article Figure 4. Identification of C. briggsae dosage compensation complex (DCC) recruitment elements on X. (A) ChIP- seq profiles of Cbr SDC- 2 and Cbr DPY- 27 binding to X chromosomes. ChIP- seq experiments were performed using an anti- FLAG antibody to immunoprecipitate SDC- 2 from a strain encoding FLAG- tagged SDC- 2, and the same anti- FLAG antibody was used in ChIP- seq experiments to immunoprecipitate DPY- 27 from a strain encoding FLAG- tagged DPY- 27. The control IgG ChIP- seq profile on X is also shown. Peaks that correspond to recruitment elements on X (rex sites), as Figure 4 continued on next page Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 11 of 39 Chromosomes and Gene Expression Research article Figure 4 continued determined by the assay in (B), are indicated in orange above the ChIP- seq profiles. RPKM is the abbreviation for reads per kilobase per million reads mapped. (B) Assay performed in vivo to determine whether DNAs from ChIP- seq peaks recruit the DCC when detached from X. XX embryos carrying extrachromosomal arrays with multiple copies of DNA from a ChIP- seq peak in (A) were stained with a DNA FISH probe to the array (red) and DPY- 27 antibody (green). If the DNA from a peak failed to recruit the DCC, DPY- 27 staining would identify X chromosomes but not the array. If DNA from a peak encoded a recruitment site (rex site), DPY- 27 staining would co- localize with the array and the X chromosome. In the merged image, the array would appear yellow and the X chromosome would appear green. Often, an array carries enough copies of a rex site that it titrates most of the DCC from X, and only the array itself shows evidence of DCC binding, appearing yellow in the merged image. In that case, the X chromosome is not detectable by DPY- 27 antibody staining. XX strains carrying rex arrays that titrate the DCC from X cannot be propagated due to the defect in dosage compensation caused by DCC titration. (C) C. briggsae rex sites recruit the C. briggsae DCC but not the C. elegans DCC. Shown is a C. briggsae or C. elegans XX gut nucleus carrying an extrachromosomal array containing multiple copies of the C. briggsae DCC recruitment site rex- 8. Nuclei were stained with appropriate species- specific C. briggsae or C. elegans antibodies to the DCC subunit DPY- 27 (green), DAPI (gray), and an array FISH probe (red). In C. briggsae, DPY- 27 bound to arrays in about 40% of the 52 scored nuclei carrying a Cbr rex- 8 array, and the DCC was titrated from X. In C. elegans, DPY- 27 bound to arrays in 0% of the 27 scored nuclei carrying a Cbr rex- 8 array, and DPY- 27 binding to the C. elegans X was evident. Scale bar, 5 μm. (D) C. elegans rex sites do not recruit the C. briggsae DCC. Shown is a C. elegans or C. briggsae XX gut nucleus carrying an extrachromosomal array containing multiple copies of the C. elegans recruitment site rex- 33 with three MEX motifs (ln[P] scores of −13.13,–15.33, –15.35). Nuclei were stained with C. elegans or C. briggsae antibodies to DCC subunit DPY- 27 (green), DAPI (gray), and an array FISH probe (red). In C. elegans, DPY- 27 bound to arrays in 100% of the 63 scored nuclei carrying a Cel rex- 33 array, and the DCC was titrated from X. In C. briggsae, DPY- 27 bound to arrays in 0% of the 53 scored nuclei carrying a Cel rex- 33 array, but did bind to Cbr X chromosomes in the same nuclei (Table 2). Scale bar, 5 μm. (E) Quantification of exemplary Cbr recruitment assays in vivo using extrachromosomal arrays containing multiple copies of DNA from Cbr DCC ChIP- seq peaks that define rex sites. Data are shown for DPY- 27 recruitment to DNA from four strong Cbr ChIP- seq peaks and a control region of DNA lacking a DCC peak (flat one containing the gene mom- 1). Shown are the locations of the sites on X, the total number of embryonic nuclei scored for DPY- 27 recruitment to the array, and the percent of nuclei recruiting the DCC. Arrays carrying rex sites recruit the DCC but arrays carrying the control flat region fail to recruit the DCC. Results of DCC recruitment assays in vivo for all rex sites are presented in Table 2. The online version of this article includes the following figure supplement(s) for figure 4: Figure supplement 1. ChIP- seq profiles of Cbr SDC- 2 and Cbr DPY- 27 binding to chromosomes X and V. We found two motifs enriched within Cbr rex sites that are highly enriched on Cbr X chromosomes compared to autosomes (Figure 6A; Figure 7A and B; Table 3). A 13 bp motif named MEX (Motif Enriched on X) is enriched up to 12- fold on X chromosomes versus autosomes, and a 30 bp motif named MEX II is enriched up to 30- fold on X versus autosomes (Figure 7A and B). All but rex- 11 and rex- 12 had either MEX, MEX II, or both. The similarity of a motif to the consensus motif is indicated by the ln(P) score, which is the natural log of the probability that the 13- mer for MEX or the 30- mer for MEX II matches the respective consensus motif matrix as calculated by the Patser program (Hertz and Stormo, 1999). The lower the score, the better the match. For both MEX and MEX II, the lower the ln(P) score, and hence the better the match to the consensus sequence, the more highly enriched is the motif on X chromosomes compared to autosomes (Figure 7A and B). Our analysis revealed that only the Cbr MEX (Figure 7C) or MEX II (Figure 7D) motifs on X that are located within rex sites are bound by SDC- 2. Negligible SDC- 2 binding was found at single, isolated MEX (Figure 7C) or MEX II (Figure 7D) motifs on X that are not in rex sites. These results implicate MEX and MEX II as important elements for Cbr DCC recruitment to rex sites. Neither of the Cbr motifs is enriched on the X chromosomes of C. elegans, indicating motif diver- gence between species (Figure 7A and B). No additional enriched C. briggsae motif candidates were found when the sequences of the two motifs in the twelve rex sites were eliminated from the search by converting them to N’s and searches for potential motifs were conducted again. In addition, motif analysis of DNA from SDC- 2 and DPY- 27 ChIP- seq peaks with intermediate or low levels of DCC binding (i.e. lower than for rex- 2) (Figure 6—figure supplement 1B) revealed no motif candidates that correlate with DCC binding. In C. elegans, two motifs are highly enriched on X chromosomes relative to autosomes: a 12 bp motif also called MEX and a 26 bp motif called MEX II (Figure 6B; Fuda et al., 2022). These C. elegans X- enriched motifs are not enriched on C. briggsae X chromosomes (Figure 6B and Figure 7—figure supplement 1A, B). Cbr MEX as well as Cel MEX and Cel MEX II share a common core sequence of CAGGG (Figure 6), which is necessary but not sufficient for DCC binding in C. elegans (Fuda et al., 2022). The core is likely indicative of a common evolutionary history between species. However, the Cbr and Cel motifs diverged sufficiently that the motifs from one species are not enriched on the Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 12 of 39 Chromosomes and Gene Expression Research article Table 2. Results of DCC recruitment assays in vivo. (A) Cbr rex DNA fragments assayed in C. briggsae and (B) Identical Cel rex DNA fragments assayed in C. elegans and in C. briggsae. (A) C. briggsae DCC binds C. briggsae DCC recruitment sites. Cbr rex Site Cbr Chr X Peak Position Cbr SDC- 2 RPKM Cbr Array Assay in vivo % Recruitment (No. of Nuclei) rex- 1 rex- 2 rex- 3 rex- 4 rex- 5 rex- 6 rex- 7 rex- 8 rex- 9 rex- 10 rex- 11 rex- 12 flat 2 flat 3 10,780,533 12,642,866 19,468,721 6,358,591 3,153,011 18,811,390 8,026,460 16,578,214 3,135,562 895,450 4,563,250 19,564,937 11,762,995 20,918,257 2890 999 3219 3915 3562 2203 2964 3217 1029 3605 830 1786 2890 999 92% 90% 88% 85% 98% 74% 97% 37% 85% 80% 89% 79% 6% 0% (59) (101) (74) (68) (45) (68) (65) (52) (62) (55) (54) (77) (48) (144) (B) C. briggsae DCC does not bind C. elegans DCC recruitment sites. Cel rex Site Cel Chr X Peak Position rex- 4 rex- 33 11,522,205 6,296,501 Cel Array Assay in vivo % Recruitment (No. of Nuclei) Cbr Array Assay in vivo % Recruitment (No. of Nuclei) 100% 100% (16) (63) 1% 0% (116) (53) (a) Extrachromosomal arrays composed of DNA fragments (2 kb) that were PCR- amplified from C. briggsae X chromosome regions corresponding to Cbr SDC- 2 ChIP- seq peaks were tested for their ability to recruit the Cbr DCC. Gut nuclei from C. briggsae transgenic lines were scored for the presence of the array using a FISH probe against the myo- 2::gfp vector and the presence or absence of DCC binding to the array by immunofluorescence signal using Cbr DPY- 27 antibodies. The % recruitment is the percentage of total scored array- bearing nuclei that showed DPY- 27 bound to the array. (B) Identical DNA fragments encoding individual C. elegans DCC recruitment sites (rex) were injected into C. elegans and C. briggsae to create extrachromosomal arrays containing multiple copies of the rex site. Gut nuclei from C. elegans or C. briggsae transgenic lines were scored for the presence of the array using a FISH probe against the myo- 2::gfp vector and for the presence or absence of DCC binding to the array by immunofluorescence signal from the species- matched DPY- 27 antibody. The % recruitment is the percentage of total scored array- bearing nuclei that showed DCC binding to the array. X chromosomes of the other species. Moreover, the Cbr MEX motif has a nucleotide substitution that would render the Cel MEX motif incapable of binding to the Cel DCC. Predominantly, the C. elegans consensus MEX motif has a cytosine nucleotide located two nucleotides 5' to the core CAGG G sequence: 5'- TCGCGCAG GGAG -3' (Figure  6B). Mutational analysis in C. elegans demonstrated that replacing that nucleotide with a guanine greatly reduced DCC binding both in vivo and in vitro (Fuda et al., 2022). The consensus Cbr MEX motif has a guanine at that critical location, and in prin- ciple, the Cbr MEX motif would not function as a Cel DCC binding motif (Figure 6), thereby offering insight into the divergence of X- chromosome binding sites between species. Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 13 of 39 Chromosomes and Gene Expression Research article Table 3. Motifs within rex sites. The ln(P) values for MEX II motifs are underlined, and the values for MEX motifs are not underlined. Cbr rex Site Chr X Peak Position SDC- 2 RPKM rex- 1 rex- 2 rex- 3 rex- 4 rex- 5 rex- 6 rex- 7 rex- 8 rex- 9 rex- 10 rex- 11 rex- 12 10,780,533 12,642,866 19,468,721 6,358,591 3,153,011 18,811,390 8,026,460 16,578,214 3,135,562 895,450 4,563,250 19,564,937 2890 999 3219 3915 3562 2203 2964 3217 1029 3605 830 1786 Cbr MEX motif ln(P) < –12 Cbr MEX II ln(P) < –12 –15.57 (13 bp) –15.57 (106 bp) –14.63 (14 bp) –14.47 (93 bp) –27.58 –14.25 (73 bp) –22.69 –12.36 (178 bp) –20.04 –19.09 (33 bp) –13.80 –18.98 –15.43 (289 bp) –13.35 –18.72 (85 bp) –12.26 (22 bp) –12.58 –13.00 (60 bp) –14.31 (69 bp) –13.22 (23 bp) –13.52 –12.8 –12.60 (63 bp) –14.68 Listed are the rex sites analyzed in this study and their motifs. Motif cutoffs used include MEX with ln(P) < –12 and MEX II with ln(P) < –12. The distances between adjacent motifs (in bp) is listed in parenthesis between motifs. Also listed are the coordinates (in bp) with the maximum SDC- 2 ChIP- seq signal in each rex site and the maximum SDC- 2 ChIP signal in reads per kilobase per million reads mapped (RPKM) within a 50 bp window. MEX and MEX II are not likely to be the only DNA sequence features within rex sites that contribute to DCC binding, since rex- 11 and rex- 12 lack these motifs with ln(P) values < –12. In C. elegans, a 9 bp motif called Motif C also participates in Cel DCC recruitment to X but lacks enrichment on X (Figure 6B; Fuda et al., 2022). Sequences between the clustered Motif C variants within a Cel rex site are also critical for DCC binding (Fuda et al., 2022). Evidence that C. elegans Motif C fails to participate in Cbr DCC recruitment to Cbr X chromosomes is our finding that Cbr SDC- 2 binding is negligible at Cel Motif C variants on Cbr X, except in the case of rare variants (0.26% of all Cel Motif C variants on X) that are within bona fide MEX or MEX II motifs in Cbr rex sites (Figure 7—figure supplement 1C). The likely reason that Cbr rex- 1 recruits the Cel DCC when inserted into Cel X chromosomes (Figure 5) is that each of the four Cbr MEX motifs includes a strong match to the consensus Cel Motif C (Figure 5 legend), and DNA sequences surrounding the Cel Motif C variants in Cbr rex- 1 are highly conserved with the syntenic region of C. elegans, which includes Cel rex- 34. Both Cel rex- 34 and Cbr rex- 1 are within coding regions of orthologous pks- 1 genes. In contrast, Cbr rex- 7 also contains Motif C variants but lacks the necessary surrounding sequences to permit Cel DCC binding when inserted on the Cel X (Figure 5). Mutational analysis of motifs on endogenous C. briggsae X chromosomes showed that combinatorial clustering of motifs in rex sites facilitates Cbr DCC binding but some binding can still occur with only a single motif in a rex site To assess further the importance of the Cbr motifs and the divergence of motifs between species, we performed mutational analyses of the two Cbr X- enriched motifs. Initial demonstration that both Cbr MEX and Cbr MEX II motifs participate in DCC binding at Cbr rex sites in C. briggsae came from analysis using extrachromosomal arrays carrying wild- type and mutant forms of Cbr rex- 1 (Figure 8— figure supplement 1). Eighty- nine percent of C. briggsae nuclei carrying extrachromosomal arrays composed of wild- type rex- 1 sequences recruited the DCC and titrated it away from X. In contrast, only 24% of nuclei carrying arrays with mutant copies of rex- 1 lacking MEX II recruited the DCC, demonstrating the importance of MEX II. Only 38% of nuclei carrying arrays with mutant copies of rex- 1 lacking all four MEX motifs recruited the DCC, demonstrating the importance of MEX. DCC Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 14 of 39 Chromosomes and Gene Expression Research article Figure 5. C. briggsae rex sites integrated into the C. elegans X chromosome by genome editing failed to recruit the C. elegans dosage compensation complex (DCC). Binding of C. elegans DCC protein Cel SDC- 3 and an IgG control were examined by ChIP- qPCR for Cel rex- 32 at its endogenous location on X, and for six C. briggsae rex sites (Cbr rex- 1, Cbr rex- 2, Cbr rex- 4, Cbr rex- 5, Cbr rex- 7, and Cbr rex- 9) plus the control Cel rex- 32 that were inserted by Cas9 genome editing into position 15,574,674 bp of the C. elegans X chromosome. (A) Schematic shows the location of Cbr rex insertions in the Cel X chromosomes and shows the different combinations of Cbr MEX and MEX II motifs in the inserted Cbr rex sites. (B) The graph of Cel SDC- 3 ChIP- qPCR data shows that all Cbr rex sites except rex- 1 exhibited SDC- 3 binding that was not significantly different from that of the autosomal negative control. Cbr rex- 1 contains a Cel Motif C variant within each Cbr MEX motif, thereby accounting for the exceptional SDC- 3 binding. The Motif Figure 5 continued on next page Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 15 of 39 Chromosomes and Gene Expression Research article Figure 5 continued C variants within Cbr rex- 1 MEX include GGGC AGGG T (–11.68), GGGC AGGG G (–14.16), GCGC AGGG C (–12.06), and CGGC AGGG G (–10.72). A fifth Motif C variant lies between the –14.16 and –12.06 variants: TCCA AGGG G (–9.84). Cel SDC- 3 levels for each replicate were normalized to the average levels at the five Cel rex sites: Cel rex- 8, Cel rex- 16, Cel rex- 32, Cel rex- 48, and Cel rex- 35. Error bars represent the SD for three replicates of Cel rex- 32 and Cbr rex- 1 and two replicates for each of Cbr rex- 2, Cbr rex- 4, Cbr rex- 5, Cbr rex- 7, and Cbr rex- 9. Cel SDC- 3 binding to the endogenous Cel rex- 32 site and the inserted rex- 32 site were not significantly different (p=0.2). Cel SDC- 3 binding to all Cbr rex sites except Cbr rex- 1 was significantly lower than binding to the Cel rex- 32 insertion (p=0.01, Student’s t- test). Cel SDC- 3 binding at Cel rex- 32 versus Cbr rex- 1 is not significantly different (p=0.3). binding was reduced to 6% of arrays carrying mutant copies of rex- 1 lacking both MEX II and the four MEX motifs. Hence, both motifs contribute to DCC binding. This conclusion was reinforced by using genome editing to mutate the MEX II sequence or all MEX II and MEX sequences in the endogenous rex- 1 site on C. briggsae X chromosomes and then assaying DCC binding (Figure 8A–C). ChIP- seq analysis revealed a significant reduction in DPY- 27 binding at rex- 1 lacking MEX II sequences and negligible DPY- 27 binding at rex- 1 lacking both MEX and MEX II sequences. Hence, the clustering of motifs in the endogenous rex- 1 on X is important for DCC binding (Figure 8). To evaluate more precisely the participation of different Cbr motifs in DCC binding, we used genome editing at three endogenous rex sites to evaluate the interplay between MEX and MEX II motifs, only MEX II motifs, or only MEX motifs. Eliminating either MEX or MEX II in rex- 4 reduced binding significantly, but the binding was evident at the remaining motif (Figure 9A–C and Figure 9— figure supplement 1A–C). Binding was dramatically reduced when both motifs were mutated. This result demonstrates that an individual MEX or MEX II motif can confer significant DCC binding at a rex site, but both motifs are needed for full DCC binding. Equivalent results were found by mutating either of the two MEX II motifs in rex- 3 or combinations of the three MEX motifs in rex- 7. For rex- 3, DCC binding was reduced significantly when one of the two MEX II motifs was mutated, but significant binding occurred at either of the remaining MEX II motifs (Figure 10A–C and Figure 10—figure supplement 1A–C). Binding was greatly reduced when both motifs were mutated. For rex- 7, DCC binding at the endogenous site lacking the MEX motif with the best match to the consensus sequence (–18.22) was not significantly different from binding at the wild- type site. In contrast, mutating different combinations of two motifs (–18.72 and –12.26 or –18.7 and –12.58) reduced binding significantly (Figure 11A–C and Figure 11—figure supplement 1A–C). Mutating all three motifs reduced binding severely. Results with the four Cbr rex sites, rex- 1, rex- 3, rex- 4, and rex- 7 demonstrate that combinatorial clustering of motifs achieves maximal DCC binding at Cbr rex sites, but significant binding can occur at a single motif. These results contrast with results in C. elegans. Mutating individual motifs, either MEX, MEX II, or Motif C, at an endogenous C. elegans rex site with multiple different motifs dramatically reduced DCC binding in vivo to nearly the same extent as mutating all motifs, demonstrating synergy in DCC binding (Fuda et al., 2022). Hence, not only have the motifs diverged between species, the mode of binding to motifs has also changed. Functional divergence of motifs demonstrated by Cel DCC binding studies in vivo and in vitro to a Cel rex site with Cbr MEX and MEX II motifs replacing Cel motifs To explore the divergence in motifs between species in greater detail, we replaced each of the two MEX II motifs of the endogenous Cel rex- 39 site on X with a copy of MEX II from Cbr rex- 3 and assayed the level of Cel SDC- 3 binding in vivo by ChIP- qPCR (Figure 12A and B). SDC- 3 binding in vivo was negligible at the Cel rex- 39 site with the Cbr MEX II motifs and indistinguishable from binding at the Cel rex- 39 site with two scrambled MEX II motifs, thus demonstrating the high degree of func- tional divergence between MEX II motifs of different species (Figure 12B). We performed a similar analysis for MEX motifs. We replaced the three MEX motifs in endogenous Cel rex- 33 with the three Cbr MEX motifs from endogenous Cbr rex- 7 (Figure 12D). SDC- 3 binding in vivo was negligible at the Cel rex- 33 site with the Cbr MEX motifs and indistinguishable from binding at the Cel rex- 33 site with three scrambled MEX motifs, demonstrating the functional divergence between MEX motifs of different species (Figure 12E). Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 16 of 39 Chromosomes and Gene Expression Research article Figure 6. Comparison of C. briggsae and C. elegans DNA motifs on X that occur within respective rex sites and recruit respective dosage compensation complex (DCC) complexes. (A) Shown are the C. briggsae consensus motifs for the 13 bp MEX and 30 bp MEX II variants that recruit the DCC. Also shown are the C. elegans consensus motifs for the 12 bp MEX, 26 bp MEX II, and 9 bp Motif C variants that recruit the Cel DCC (B). The sequences were aligned relative to the conserved adenine in the 5'-CAGGG- 3' common core of the motifs. Predominantly, the Cel MEX motif has a cytosine in the fourth Figure 6 continued on next page Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 17 of 39 Chromosomes and Gene Expression Research article Figure 6 continued position of the motif. Mutating it to a guanine (C4G) severely reduced DCC binding in assays conducted in vivo and in vitro. The consensus Cbr MEX motif has a guanine at the equivalent position relative to the CAGGG core. Hence, the Cbr MEX motif is predicted not to function as a DCC recruitment motif in C. elegans. The online version of this article includes the following figure supplement(s) for figure 6: Figure supplement 1. C. briggsae SDC- 2 ChIP- seq peak profiles for rex sites and non- rex sites on X. As a second approach, we conducted Cel DCC binding studies in vitro (Materials and methods). In brief, this assay (Fuda et al., 2022) utilized embryo extracts made from a Cel nematode strain encoding a 3xFLAG- tagged Cel SDC- 2 protein expressed from an extrachromosomal array. Wild- type or mutant 651 bp DNA fragments with biotinylated 5' ends were coupled to streptavidin- coated magnetic beads and incubated with embryo extracts. The bound proteins were eluted, spotted onto a nitrocellulose membrane, and probed with a monoclonal mouse anti- FLAG antibody. Antigen- antibody complexes were visualized and quantified by chemiluminescence using an imager. The advantage of this assay is that Cel SDC- 2 is capable of binding to a single motif on an in vitro template, perhaps because that DNA lacks the competing binding of nucleosomes and general transcription factors that occurs in vivo (Fuda et al., 2022). We assayed Cel DCC binding to a Cel rex- 39 site with two Cbr MEX II motifs (Figure 12C) and to the Cel rex- 33 site with the three Cbr MEX motifs (Figure 12F). If either of the Cbr MEX II motifs inserted into the Cel rex- 39 site were functional or if any of the three Cbr MEX motifs inserted into the Cel rex- 33 site were functional, we would detect Cel SDC- 2 binding to the template in vitro. The in vitro assay demonstrated robust binding of Cel SDC- 2 to the wild- type Cel rex- 39 template (Figure 12C) and to the wild- type Cel rex- 33 template (Figure 12F), as shown previously (Fuda et al., 2022). However, Cel SDC- 2 binding at the Cel rex- 39 site with substituted Cbr MEX II motifs was indistinguishable from binding to the mutant Cel rex- 39 template with two scrambled Cel MEX II motifs or to the negative control template made from Cel X DNA at a site lacking Cel DCC binding in vivo (Figure 12C). Similarly, Cel SDC- 2 binding at the Cel rex- 33 site with substituted Cbr MEX motifs was indistinguishable from binding to the mutant Cel rex- 33 template with three scrambled Cel MEX motifs or to the negative control template (Figure 12F). Thus, the in vitro assay demonstrates that substituting Cbr MEX II or MEX motifs for Cel MEX II or MEX motifs in a Cel rex site eliminates Cel DCC binding. A single nucleotide position in the consensus Cbr MEX motif acts as a critical determinant for whether the Cel DCC can bind to Cbr MEX In contrast to the many nucleotide changes that mark the difference between MEX II motifs in C. briggsae versus C. elegans, the MEX motifs are strikingly similar in nucleotide composition and core CAGGG sequence between species (Figure  6). A significant change between the consensus MEX motifs is the substitution in Cbr MEX of a guanine for the cytosine in Cel MEX located two nucleotides 5' from the CAGGG core of both motifs (Figure 13A). That C4G transversion was never found in a functional Cel MEX motif in vivo or in vitro. Moreover, a C4G change in either the MEX motif of endog- enous Cel rex- 1 or in an in vitro Cel DNA template reduced binding to the level of a rex- 1 deletion or negative control lacking a MEX motif (Fuda et al., 2022). Hence, the Cel DCC would be unable to bind to any Cbr MEX motif with C4G. In principle, that single cytosine- to- guanine transversion could be a critical evolutionary change in MEX motifs that render the motifs incapable of binding the DCC from the other species. To test this hypothesis, we made the C4G transversion in each of the three MEX motifs within the endogenous Cel rex- 33 site (Figure  13B). Cel SDC- 3 binding in vivo to the C4G- substituted Cel rex- 33 site was reduced to the same level of binding as that at the Cel rex- 33 site with all three Cel MEX motifs scrambled, confirming the functional significance of the nucleotide substitution between species (Figure 13B). Our in vitro assay comparing Cel SDC- 2 binding to the C4G- substituted and the MEX- scrambled Cel rex- 33 DNA templates produced the same negative result (Figure 13C). If the evolutionary transversion of that C to G between Cel and Cbr MEX motifs represents an important step in the divergence of motif function, then making a G- to- C change within the Cbr MEX motifs (G7C) inserted into Cel rex- 33 should enhance Cel DCC binding. The substitution would not be Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 18 of 39 Chromosomes and Gene Expression Research article Figure 7. Enrichment of Cbr MEX and Cbr MEX II motifs on X chromosomes between C. briggsae and C. elegans. (A, B) Graphs show the enrichment (y- axis) of Cbr MEX (A) or Cbr MEX II (B) variants (x- axis) on X chromosomes compared to autosomes in the C. briggsae (green circles) and C. elegans (orange circles) genomes. For MEX, the ln(P) is the natural log of the probability that a 13- mer matches the MEX consensus motif matrix (shown above the graphs) as calculated by the Patser program. For MEX II, the ln(P) is the natural log of the probability that a 30- mer matches the MEX II consensus Figure 7 continued on next page Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 19 of 39 Chromosomes and Gene Expression Research article Figure 7 continued motif matrix (shown above the graphs) as calculated by Patser. The lower the score, the better the match. The maximum theoretical ln(P) value for MEX is –18.7 and for MEX II is –29.3. The best MEX score found on Cbr X is –18.7 and for MEX II is - 27.58. The graphs reflect cumulative scores. For example, the 12- fold X:A enrichment of MEX for C. briggsae at –17.58 reflects all motifs with ln(P) ≤ –17.58. The C. elegans X chromosome is not enriched for the Cbr MEX or MEX II consensus motifs that are enriched on Cbr X chromosomes and that are pivotal for Cbr DCC recruitment to Cbr X, as we show subsequently. (C) The graph plots the mean (dark blue) and standard error (light blue) of Cbr SDC- 2 ChIP- seq signal (RPKM) at various distances from MEX motifs (< –12) in rex sites versus the mean (dashed red) and standard error (light red) of SDC- 2 signal at varying distances from MEX motifs (< –12) on X but not in rex sites. Abundant SDC- 2 binding was found at MEX motifs in rex sites, but negligible SDC- 2 binding was found at individual MEX motifs on X that were not in rex sites or at MEX motifs on autosomes. n, total number of MEX motifs in each category. (D) The graph plots the mean (dark blue) and standard error (light blue) of Cbr SDC- 2 ChIP- seq signal (RPKM) at various distances from MEX II motifs (< –12) in rex sites versus the mean (dashed red line) and standard error (light red) of SDC- 2 signal at varying distances from MEX II motifs (< –12) on X but not in rex sites. Abundant SDC- 2 binding was found at MEX II motifs in rex sites, but negligible SDC- 2 binding was found at individual MEX II motifs on X that were not in rex sites or at MEX II motifs on autosomes. n, total number of MEX II motifs in each category. The online version of this article includes the following figure supplement(s) for figure 7: Figure supplement 1. The C. briggsae X chromosome is not enriched for the C. elegans MEX (A) or MEX II (B) motifs that are highly enriched on Cel X chromosomes and pivotal for DCC binding to Cel X chromosomes in vivo. expected to restore Cel DCC binding fully, because other sequences within the Cbr motif contribute to a lower match to the Cel consensus sequence and hence lower Cel binding affinity. However, no other identified single nucleotide substitution within a known Cbr MEX motif besides C4G is expected to eliminate Cel DCC binding (Fuda et al., 2022). Indeed, the G7C change to Cbr MEX within Cel rex- 33 increased the Cel SDC- 3 binding in vivo by 4.2- fold and increased the specific Cel SDC- 2 binding in vitro by 4.3- fold. The G7C change increased Cel SDC- 3 binding in vivo to 18% of its binding at wild- type Cel rex- 33 (Figure 13B) and increased Cel SDC- 2 binding in vitro to 44% of its the specific binding at the wild- type Cel rex- 33 template (Figure 13C). Hence, the cytosine- to- guanine transver- sion between MEX motifs of C. elegans versus C. briggsae is important for the functional divergence in motifs. Discussion Comparison of X- chromosome dosage compensation mechanisms between the closely related Caenorhabditis species C. briggsae and C. elegans revealed that both the dosage compensation machinery and the regulatory hierarchy that directs it to hermaphrodite X chromosomes have been conserved, but remarkably, the X- chromosome target specificity of the C. briggsae machinery and its mode of binding to X have diverged, as well as the density of DCC recruitment sites. The extent of evolutionary changes in dosage compensation mechanisms between species diverged by only 15–30 MYR is in striking contrast to mechanisms that control somatic sex determination and differentiation in the same species. The master regulator of hermaphrodite sexual fate, TRA- 1, is conserved between both species, as is its DNA target specificity (Berkseth et al., 2013; Zarkower and Hodgkin, 1992). Moreover, the divergence of Caenorhabditis dosage compensation mechanisms contrasts with the conservation of Drosophila dosage compensation mechanisms (Alekseyenko et  al., 2013; Kuzu et al., 2016) and the conservation of mechanisms controlling developmental processes such as body- plan specification and eye morphogenesis from fruit flies to mice (Malicki et al., 1990; Halder et al., 1995), which utilize highly conserved transcription factors and cis- acting DNA regulatory sequences. Central to the dosage compensation machinery of both species is a specialized condensin complex. Here we identified two C. briggsae dosage compensation proteins (DPY- 27 and MIX- 1) that are ortho- logs of the SMC (structural maintenance of chromosome) subunits of condensin and bind to hermaph- rodite X chromosomes. As in C. elegans (Chuang et al., 1994; Lieb et al., 1998), mutation of dpy- 27 causes hermaphrodite- specific lethality in C. briggsae, and MIX- 1 fails to bind X in the absence of DPY- 27, consistent with both proteins acting in a complex. We also found that the hermaphrodite- specific Cbr sdc- 2 gene triggers binding of the condensin subunits to X and activates the hermaphrodite mode of sexual differentiation, as in C. elegans. Mutation of Cbr sdc- 2 causes XX- specific lethality, and rare XX animals that escape lethality develop as masculinized larvae. SDC- 2 and condensin subunits are prevented from binding to the single X of males by the action of xol- 1, the master sex- determination gene that controls both sex determination and dosage compensation and triggers the male fate by repressing sdc- 2 function. Mutation of xol- 1 kills XO animals because the DCC assembles Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 20 of 39 Chromosomes and Gene Expression Research article Figure 8. Combinatorial clustering of MEX and MEX II motifs in Cbr rex- 1 facilitates dosage compensation complex (DCC) binding to the endogenous rex- 1 site on X. (A) Shown is an enlargement of the SDC- 2 ChIP- seq peak profile for Cbr rex- 1 with its associated MEX (purple) and MEX II (green) motifs and their ln(P) scores. (B) DPY- 27 ChIP- seq analysis was performed using anti- FLAG antibody on an otherwise genetically wild- type C. briggsae strain encoding FLAG- tagged DPY- 27 and on FLAG- tagged DPY- 27 C. briggsae mutant variants carrying either a scrambled (scr) version of MEX II or a Figure 8 continued on next page Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 21 of 39 Chromosomes and Gene Expression Research article Figure 8 continued scrambled version of MEX II and all four MEX motifs. The control IgG ChIP- seq analysis was performed on the C. briggsae strain encoding FLAG- tagged DPY- 27 carrying wild- type copies of all rex sites. DPY- 27 and control IgG ChIP- seq profiles are also shown for Cbr sites rex- 7 and rex- 4 as an internal standard since DPY- 27 binding was not disrupted at these sites. (C) Sequences of the wild- type Cbr rex- 1 MEX motifs and their scrambled versions. Underlined is the Cel Motif C variant within each Cbr MEX motif. For analyzing MEX II, two different MEX II mutant variants were used, as indicated by asterisks. Numbers between motifs indicate the base pairs separating the motifs. ChIP- seq profiles reveal that mutating only MEX II reduces some DCC binding at rex- 1, and mutating MEX II and all MEX motifs eliminates DCC binding. The motifs act cumulatively to recruit the DCC. The online version of this article includes the following figure supplement(s) for figure 8: Figure supplement 1. Clustering of MEX and MEX II motifs in Cbr rex- 1 confers dosage compensation complex (DCC) binding in vivo. on the single male X, thereby reducing gene expression inappropriately. Mutations in sdc- 2 or dpy- 27 suppress the XO- specific lethality caused by xol- 1 mutations, but only mutations in sdc- 2 permit the rescued animals to develop as males. Just as in C. elegans, XO animals rescued by dpy- 27 mutations develop as hermaphrodites, consistent with dpy- 27 controlling only dosage compensation and sdc- 2 controlling both sex determination and dosage compensation. Hence, the two master regulators that control sexual fate and dosage compensation are functionally conserved between the two Caenor- habditis species, as is the condensin dosage compensation machinery. In both species, SDC- 2 recruits the condensin DCC subunits to X and is the likely protein to interact directly with X DNA. Despite their central roles in dosage compensation, these 350 kDa proteins lack homology to proteins outside of Caenorhabditis, and their only predicted structural feature is a coiled- coil region. Alignment of SDC- 2 proteins in five Caenorhabditis species revealed only 23–29% identity and 38–45% similarity across the entire protein, with two regions that show greater conservation (Figure 1—figure supplements 1 and 2A). One region is N- terminal to the coiled- coil domain and shares 36–45% identity and 57–63% similarity. A second region resides in the C- terminal part of the protein and shows 24–32% identify and 39–51%  similarity. Neither region, nor any segment of the protein, has a predicted DNA binding domain. The discovery of any such domain requires ongoing biochemical and structural analysis. DCC condensin subunits have variable conservation across species, depending on whether they function only in the DCC or participate in other condensin complexes as well. DPY- 27, the only condensin subunit specific to the DCC, has only limited conservation: 34% identity and 56% similarity across the Caenorhabditis genus (Figure  1—figure supplement 2B). In contrast, DCC condensin subunit MIX- 1, which also participates in the two condensin complexes required for mitotic and meiotic chromosome segregation, shows greater identity and similarity between both species: 55% and 72%, respectively. In comparison, SMC- 4, an ortholog of DPY- 27 and a conserved SMC chromo- somal ATPase that interacts with MIX- 1 in the mitotic and meiotic condensin complexes, but not in the DCC condensin complex (Hagstrom et al., 2002; Csankovszki et al., 2009 and Mets and Meyer, 2009), shares even greater conservation between C. elegans and C. briggsae, commensurate with its universal role in chromosome segregation: 62% identity and 76% similarity. Participation of MIX- 1 and SMC- 4 in condensin complexes dedicated to chromosome segregation constrains their divergence, thereby accounting for their higher conservation than DPY- 27. Although C. elegans and C. briggsae have conserved DCC machinery, the DCC binding sites have diverged, as has their density on X. ChIP- seq analysis of C. briggsae SDC- 2 and DPY- 27 revealed twelve sites of binding on X that were validated by functional analysis in vivo as being strong autono- mous recruitment (rex) sites. Even though the X chromosome of C. briggsae (21.5 Mb) is larger than the X of C. elegans (17.7 Mb), it has only one- fourth the number of recruitment sites. The C. briggsae sites are sufficiently strong that extrachromosomal arrays carrying multiple copies of a single site can titrate the DCC from X and cause dosage- compensation- defective phenotypes in XX animals, including death, as in C. elegans. In contrast, extrachromosomal arrays of C. briggsae rex sites made in C. elegans fail to recruit the C. elegans DCC, and vice versa, indicating that rex sites have diverged between the two species. As a more rigorous test of divergence, individual C. briggsae rex sites were inserted in a single copy into C. elegans X chromosomes and assayed for Cel DCC binding. The C. elegans DCC failed to bind to the five C. briggsae rex sites inserted into C. elegans X chromosomes. Not only have the rex sites diverged, the mechanism by which the Cbr DCC binds to X motifs differs from that of the Cel DCC. We identified two motifs within C. briggsae rex sites that are highly enriched on X, the 13 bp MEX motif and the 30 bp MEX II motif. Mutating one copy of either motif Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 22 of 39 Chromosomes and Gene Expression Research article Figure 9. MEX and MEX II motifs are critical for dosage compensation complex (DCC) binding to Cbr rex- 4 in vivo. (A) Shown is an enlargement of the SDC- 2 ChIP- seq profile for rex- 4, a schematic of the MEX (purple) and MEX II (green) motifs in rex- 4, and the location of primers (E and F, dashed lines) to evaluate DCC binding in vivo using ChIP- qPCR. Motifs are separated by 33 bp. (B) The graph shows ChIP qPCR levels for SDC- 2 (dark blue) and control IgG (light blue) at endogenous wild- type rex- 4, at endogenous rex- 4 with different combinations of motif mutations created by genome editing, and at a negative control site on X of 107 bp that lacks DCC binding centered at (7,000,213 bp). Strains carrying wild- type and mutant motifs encoded FLAG- tagged SDC- 2. SDC- 2 levels for each replicate were normalized to the average level of five endogenous non- edited rex sites (Cbr rex- 1, Cbr rex- 2, Cbr rex- 5, and Cbr rex- 9). Error bars represent the standard deviation (SD) of three replicates. Asterisks of the same color specify data compared using the Student’s t- test. If more than two motif combinations are compared, the schematic to the right of the p- value indicates the motif combination to which the other combinations were compared. (C) DNA sequences of wild- type and mutant motifs (scr) are shown below the graph. Both MEX and MEX II motifs are critical for DCC binding at rex- 4. Mutating each motif independently causes an equivalent reduction in DCC binding, and mutating both motifs is necessary to eliminate DCC binding. ChIP- qPCR analysis of SDC- 2 binding at intervals across the entire peak is presented in Figure 9—figure supplement 1. The online version of this article includes the following figure supplement(s) for figure 9: Figure supplement 1. MEX and MEX II motifs are critical for SDC- 2 binding to Cbr rex- 4 in vivo. Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 23 of 39 Chromosomes and Gene Expression Research article Figure 10. Both MEX II motifs are critical for dosage compensation complex (DCC) binding to Cbr rex- 3 in vivo. (A) Shown is an enlargement of SDC- 2 ChIP- seq profile for Cbr rex- 3 with its associated MEX II motifs (green) and their ln(P) scores. Motifs are separated by 178 bp. Locations of primers (F and G, dashed lines) to evaluate DCC binding in vivo using ChIP- qPCR are shown. (B) The graph shows ChIP qPCR levels for SDC- 2 (dark blue) and control IgG (light blue) at endogenous wild- type rex- 3, at endogenous rex- 3 with different combinations of motif mutations created by genome editing, and at a negative control site on X that lacks DCC binding. Strains carrying wild- type and mutant motifs encoded FLAG- tagged SDC- 2. SDC- 2 levels for each replicate were normalized to the average level of five endogenous non- edited rex sites (Cbr rex- 1, Cbr rex- 2, Cbr rex- 5, and Cbr rex- 9). Error bars represent the standard deviation (SD) of three replicates. Symbols of the same color specify data compared using the Student’s t-test. If more than two motif combinations are compared, the schematic to the right of the p- value indicates the motif combination to which the other combinations were compared. (C) DNA sequences of wild- type and mutant motifs (scr). Both MEX II motifs are critical for DCC binding at rex- 3. Mutating each motif independently causes an equivalent reduction in DCC binding, and mutating both motifs is necessary to eliminate DCC binding. ChIP- qPCR analysis of SDC- 2 binding at intervals across the entire peak is presented in Figure 10—figure supplement 1. The online version of this article includes the following figure supplement(s) for figure 10: Figure supplement 1. Both MEX II motifs are critical for dosage compensation complex (DCC) binding to Cbr rex- 3 in vivo. Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 24 of 39 Chromosomes and Gene Expression Research article Figure 11. Multiple MEX motifs in Cbr rex- 7 contribute to dosage compensation complex (DCC) binding in vivo. (A) Shown is an enlargement of SDC- 2 ChIP- seq profile for Cbr rex- 7 with its associated MEX motifs (purple) and their ln(P) scores. Motifs are separated by 85 bp and 22 bp. Locations of primers (D and E, dashed lines) to evaluate DCC binding in vivo using ChIP- qPCR are shown. (B) The graph shows ChIP qPCR levels for SDC- 2 (dark blue) and control IgG (light blue) at endogenous wild- type rex- 7, at endogenous rex- 7 with different combinations of motif mutations created by genome editing, and at a negative control site on X that lacks DCC binding. Strains carrying wild- type and mutant motifs encoded FLAG- tagged SDC- 2. SDC- 2 levels for each replicate were normalized to the average level of five endogenous non- edited rex sites (Cbr rex- 1, Cbr rex- 2, Cbr rex- 5, and Cbr rex- 9). Error bars represent the standard deviation (SD) of three replicates. Symbols of the same color specify data compared using the Student’s t- test. If more than two motif combinations are compared, the schematic to the right of the p- value indicates the motif combination to which the other combinations were compared. (C) Sequences of wild- type and mutant motifs (scr). Multiple MEX motifs contribute to DCC binding at rex- 7. Mutating the first MEX motif has an insignificant effect on DCC binding, but mutating the first MEX motif and either of the other two motifs reduces Figure 11 continued on next page Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 25 of 39 Chromosomes and Gene Expression Research article Figure 11 continued binding equivalently. Mutating all three MEX motifs eliminates DCC binding. ChIP- qPCR analysis of SDC- 2 binding at intervals across the entire peak is presented in Figure 11—figure supplement 1. The online version of this article includes the following figure supplement(s) for figure 11: Figure supplement 1. Multiple MEX motifs in Cbr rex- 7 contribute to dosage compensation complex (DCC) binding in vivo. in endogenous rex sites with multiple motifs reduced binding, but significant binding still occurred at the sites. Binding was eliminated only when all motifs were removed. Hence, DCC binding to motifs in C. briggsae rex sites appears additive. In contrast, mutating one motif in C. elegans rex sites that have multiple different combinations of motifs reduced binding to nearly the same extent as mutating all motifs, indicating synergy in C. elegans DCC binding (Fuda et al., 2022). Additional factors, such as yet- unidentified DNA binding proteins might alter the specificity of DCC binding between species as well as aid DCC binding at Cbr rex sites. Precedent exists in the home- odomain family of Hox DNA binding proteins that have remarkably similar DNA specificities for DNA binding in vitro but a wide range of specificities in vivo due to interactions with heterologous cofactors required for functional specificities, such as Pbx- Hox complexes (Chang et al., 1996). The need for synergy in DCC binding to Cel rex sites is likely caused by competition between DCC binding and nucleosome formation, since nucleosomes preferentially bind to rex sites when DCC binding is precluded by mutations (Fuda et al., 2022). The status of nucleosomes on C. briggsae X chromosomes remains to be determined. Although a single MEX or MEX II motif enables some DCC binding to a Cbr rex site, equivalent motifs on X that are not in rex sites appear to lack DCC binding. Nucleosome formation may preclude DCC binding at those motifs. The X may have a paucity of DNA- binding proteins that interact with core histones and open compacted chromatin to enable DCC binding. Although the X- chromosome motifs of both species share the core consensus sequence CAGGG, the motifs have diverged such that they function in only one species. This functional divergence was demonstrated through DCC binding studies in vivo and in vitro to C. elegans rex sites engineered with C. briggsae motifs substituted for C. elegans motifs. We replaced the two MEX II motifs in the endogenous C. elegans rex- 39 site with C. briggsae MEX II motifs and the three MEX motifs in Cel rex- 33 with Cbr MEX motifs while maintaining motif spacing appropriate for C. elegans. We found negligible C. elegans DCC binding in vivo and in vitro. A feature of the in vitro assay is that Cel SDC- 2 is capable of binding to a single motif on a DNA template, likely because the DNA lacks competing binding of nucleosomes that occurs in vivo. If either Cbr MEX II or MEX motif were functional in C. elegans we would have detected binding. While the MEX II motif has diverged sufficiently that evolutionary tracing is difficult, the diver- gence of MEX motifs provides important insight into their evolution. A major difference in MEX motifs between the two species is the preference for a guanine instead of a cytosine two nucleotides 5' of the conserved CAGGG sequence. We demonstrated that converting that C to G in the three Cel MEX motifs of Cel rex- 33 eliminated DCC binding in vitro. Conversely, replacing the G nucleotide in each Cbr MEX motif inserted into Cel rex- 33 with a C nucleotide partially restored Cel DCC binding in vivo and in vitro, indicating that the single nucleotide change can be important in the evolutionary diver- gence of this motif. The evolutionary C- to- G substitution in the Cbr MEX motif is sufficient to prevent it from functioning in the closely related C. elegans species. In contrast to the divergence of X- chromosome target specificity between Caenorhabditis species, X- chromosome target specificity has been conserved among Drosophila species. A 21  bp GA- rich sequence motif on X is utilized across Drosophila species to recruit the dosage compensation machinery, although it may not be the sole source of X target specificity (Alekseyenko et al., 2008; Kuzu et al., 2016; Ellison and Bachtrog, 2019; Alekseyenko et al., 2013). Conservation of DNA target specificity among species is also a common theme among devel- opmental regulatory proteins that participate in multiple, unrelated developmental processes, such as Drosophila Dorsal in the body- plan specification (Schloop et al., 2020) or Caenorhabditis TRA- 1 in hermaphrodite sexual differentiation and male neuronal differentiation (Berkseth et al., 2013; Bayer et  al., 2020). Typically, for such multi- purpose proteins, target- site specificity is evolution- arily constrained: protein function is changed far more by changes in the number and location of conserved cis- acting target sequences than by changes in the target sequences themselves (Carroll, Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 26 of 39 Chromosomes and Gene Expression Research article Figure 12. Functional divergence of X motifs demonstrated by C. elegans dosage compensation complex (DCC) binding studies in vivo and in vitro to Cel rex sites engineered to replace Cel motifs with Cbr MEX and MEX II motifs. (A) Comparison of DNA sequences for the two MEX II motifs in wild- type Cel rex- 39 (Cel ln[P] of –21.23 and –20.74) with the Cbr MEX II motifs (Cbr ln[P] of –20.04 and Cel ln[P] > –9 for both) that replaced them. DNA sequences of the spacer region between wild- type Cel MEX II motifs and inserted Cbr MEX II motifs are shown, as are sequences of the scrambled Cel MEX II Figure 12 continued on next page Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 27 of 39 Chromosomes and Gene Expression Research article Figure 12 continued motifs used as negative controls. Schematics show keys for rex sites analyzed for Cel SDC- 3 binding in vivo and Cel SDC- 2 binding in vitro: wild- type Cel rex- 39 (orange, MEX II motifs), Cel rex- 39 with Cbr MEX II motifs (green), Cel rex- 39 with scrambled Cel MEX II motifs (orange outline). (B) Graph shows ChIP qPCR levels for Cel SDC- 3 (dark blue) and control IgG (light blue) at wild- type Cel rex- 39 and mutant rex- 39 with Cbr MEX II motifs in vivo. Cel SDC- 3 binds in vivo to endogenous Cel rex- 39 sites with wild- type MEX II motifs but not to mutant Cel rex- 39 sites with either scrambled Cel MEX II motifs or Cbr MEX II motif replacements. SDC- 3 levels for each replicate were normalized to the average SDC- 3 level at 7 control rex sites (Cel rex- 8, Cel rex- 14, Cel rex- 16, Cel rex- 32, Cel rex- 35, Cel rex- 36, and Cel rex- 48). Error bars represent the standard deviation (SD) of three replicates. Statistical comparisons were calculated using the Student’s t-test. (C) Graph of in vitro assay assessing Cel SDC- 2 binding to a wild- type Cel rex- 39 DNA template and a mutant rex- 39 template with Cbr MEX II motifs. Cel SDC- 2 binds to the Cel rex- 39 template with wild- type MEX II motifs but not to mutant rex- 39 templates with either scrambled Cel MEX II motifs or Cbr MEX II motif replacements. Cel SDC- 2 does not bind to the control template (beige) made of DNA from a site on the Cel X that lacks SDC- 2 binding in vivo. SDC- 2 levels detected for the mutant variants of rex- 39 templates are shown as the percentage (%) of SDC- 2 binding to the wild- type rex- 39 template. The plot represents the average of three independent experiments, with error bars indicating SD. Statistical comparisons were calculated using the Student’s t-test. (D) Comparison of DNA sequences for the three MEX motifs in wild- type Cel rex- 33 and the Cbr MEX motifs that replaced them. Also shown are sequences for the scrambled Cel MEX motifs used as negative controls. Schematics show keys for rex sites analyzed for Cel SDC- 3 binding in vivo and Cel SDC- 2 binding in vitro: wild- type Cel rex- 33 (black, MEX motifs), Cel rex- 33 with Cbr MEX motifs (purple), Cel rex- 39 with scrambled Cel MEX motifs (black outline). (E) Graph shows ChIP qPCR levels for Cel SDC- 3 (dark blue) and control IgG (light blue) at wild- type Cel rex- 33 and mutant rex- 33 with Cbr MEX motifs in vivo. Cel SDC- 3 binds to endogenous Cel rex- 33 sites with wild- type MEX motifs but not to mutant Cel rex- 33 sites with either scrambled Cel MEX motifs or Cbr MEX motif replacements. Details of the experiment and graph are the same as in (B). (F) Graph of in vitro assay assessing Cel SDC- 2 binding to a wild- type Cel rex- 33 DNA template and a mutant rex- 33 template with Cbr MEX motifs. Cel SDC- 2 binds to the Cel rex- 33 template with wild- type MEX motifs but not to mutant Cel rex- 33 templates with either scrambled Cel MEX motifs or Cbr MEX motif replacements. Cel SDC- 2 does not bind to the control template (beige). SDC- 2 levels detected for the mutant variant rex- 33 templates are shown as the percentage (%) of SDC- 2 binding to the wild- type rex- 33 template. The plot represents the average of three independent experiments, with error bars indicating SD. Statistical comparisons were calculated using the Student’s t- test. 2008; Nitta et  al., 2015). Hence, the divergence in X- chromosome target specificity across the Caenorhabditis genus is atypical among developmental regulatory complexes with highly diverse target genes and could have been an important factor for establishing reproductive isolation between species. Our finding is reminiscent of the discovery that centromeric sequences and their corresponding centromere- binding proteins have co- evolved rapidly (Malik and Henikoff, 2001; Henikoff et  al., 2001; Talbert and Henikoff, 2022). The occurrence of rapidly changing DNA targets and their corresponding DNA- binding proteins (see also Liénard et al., 2016; Ting et al., 1998; Ting et al., 2004; Sun et al., 2004) is an increasingly dominant theme contributing to repro- ductive isolation. Materials and methods All key resources have been provided in Supplementary files 1–6. Procedures for mutant isolation Procedures for sdc- 2 mutant isolation were described previously by Wood et al., 2011. xol- 1(y430), dpy- 27(y436), and mix- 1(y435) were isolated from a C. briggsae deletion library provided by E. Haag using primers listed in Supplementary file 2. The resulting strains are listed in Supplementary file 1. Protein sequence analysis of SDC-2 and DPY-27 Sequence alignments of SDC- 2 proteins from C. elegans (UniProtKB G5EBL3), C. brenneri (UniProtKB G0M6S8), C. japonica (WormBase JA61524), C. tropicalis (this study), and C. briggsae (Uniprot A8XQT3) were generated using Clustal Omega (Madeira et al., 2022) and ESPript 3.0 server (https:// espript.ibcp.fr) (Robert and Gouet, 2014). The coiled- coil annotations were predicted using the web server version of DeepCoil (Ludwiczak et al., 2019), part of the MPI Bioinformatics Toolkit (Zimmer- mann et  al., 2018; Gabler et  al., 2020). Pairwise sequence comparisons of SDC- 2 proteins were performed with EMBOSS Needle (Madeira et al., 2022). Pairwise sequence comparisons of DPY- 27 proteins from C. elegans (Uniprot P48996), C. briggsae (Uniprot A8XX62), C. brenneri (WormBase CN00825), and C. tropicalis (this study) were generated using EMBOSS Needle. Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 28 of 39 Chromosomes and Gene Expression Research article Figure 13. A nucleotide position in the consensus Cbr MEX motif can act as a critical determinant for whether Cel dosage compensation complex (DCC) binds in vivo and in vitro. (A) Shown are DNA sequences of three wild- type or mutant Cel or Cbr MEX motifs within Cel rex- 33 assayed for Cel SDC- 3 binding in vivo (B) and Cel SDC- 2 binding in vitro (C). The ln(P) scores for the wild- type Cel MEX motifs in rex- 33 are −13.13, –15.33, and –15.35. The Cel ln(P) scores for the 3 substituted Cbr MEX motifs are all greater than –9. The three Cbr ln[P] scores for those substituted Cbr MEX motifs are Figure 13 continued on next page Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 29 of 39 Chromosomes and Gene Expression Research article Figure 13 continued −18.72, –12.26, and −12.58. The Cel ln(P) scores for the 3 Cel MEX motifs with the C4G change are −9.58, –11.20, and –11.26. The Cel ln(P) scores for the three Cbr MEX motifs with the G7C change are −12.20, –11.16, and –10.84. The Cel ln(P) scores for the Cel rex- 33 scrambled MEX motifs are all greater than –9. (B) Graph shows normalized ChIP qPCR levels for Cel SDC- 3 (dark blue) and control IgG (light blue) in vivo at endogenous Cel rex- 33 with wild- type or mutant Cel MEX motifs and wild- type or mutant Cbr MEX motifs. Replacing the critical cytosine (red asterisk) in each of the three MEX motifs of endogenous Cel rex- 33 with a guanine (C4G) eliminates Cel SDC- 3 binding, as does scrambling the three Cel MEX motifs. Substituting three Cbr MEX motifs for Cel MEX motifs also severely reduces Cel DCC binding. Each Cbr MEX motif has a guanine instead of a cytosine in the critical location. Replacing the guanine with a cytosine (G7C) in each of the Cbr MEX motifs increased Cel SDC- 3 binding 4.2- fold, resulting in a Cel SDC- 3 binding level representing 18% of that at wild- type rex- 33. SDC- 3 levels for each replicate were normalized to the average SDC- 3 level at seven control rex sites (Cel rex- 8, Cel rex- 14, Cel rex- 16, Cel rex- 32, Cel rex- 35, Cel rex- 36, and Cel rex- 48). Error bars represent the standard deviation (SD) of three replicates. Statistical comparisons were calculated using the Student’s t-test. (C) Graph of the in vitro Cel SDC- 2 binding assay shows that replacing the critical cytosine (red asterisk) in each of the three MEX motifs of Cel rex- 33 with a guanine (C4G) eliminates Cel SDC- 2 binding, as does scrambling the three MEX motifs. Substituting three Cbr MEX motifs for Cel MEX motifs severely reduces Cel DCC binding. Each Cbr MEX motif has a guanine instead of a cytosine in the critical location. Replacing the guanine with a cytosine (G7C) in each of the Cbr MEX motifs increases specific Cel SDC- 2 binding 4.3- fold and restores it to 44% of that at the wild- type rex- 33 DNA template. SDC- 2 levels detected for the mutant variants of rex- 33 templates are shown as the percentage (%) of SDC- 2 binding to the wild- type rex- 33 template. The plot represents the average of three independent experiments, with error bars indicating SD. Statistical comparisons were calculated using the Student’s t-test. Preparation of FISH probes Chromosome FISH probes were prepared from 1 mg of total DNA, which included multiple C. briggsae BACs listed in Supplementary file 3 (BACPAC Resources Center, CHORI, Oakland, CA). BACs were purified using the QIAGEN midiprep kit (catalog number 12243). Chromosomal FISH probes were made with the Invitrogen DNA FISH- tag kit. X- chromosome probes (10 BACS covering approximately 5% of the chromosome) were labeled with AlexaFluor 594 (Molecular Probes, F32949), and chro- mosome III probes (three BACS covering approximately 1% of the chromosome) were labeled with AlexaFluor 488 (Molecular Probes, F32947). Preparation of gut nuclei for FISH and immunofluorescence Adult worms were dissected in 4 µl egg buffer (25 mM HEPES, pH 7.4, 118 mM NaCl, 48 mM KCl, 0.2  mM CaCl2, 0.2  mM MgCl2) on a 18 mm X 18  mm coverslip. 4 µl of 4% formaldehyde (in egg buffer) were added, and the solution was mixed by tapping the coverslip before it was placed onto a Superfrost/Plus glass slide (Fisherbrand, 12- 550- 15). Fixed samples were incubated for 5  min at room temperature in a humid chamber, then frozen in liquid nitrogen for at least 1 min. Coverslips were removed quickly with a razor blade, and slides were placed immediately into PBS- T (PBS with 1 mM EDTA and 0.5% Triton X- 100). Slides were subjected to three 10 min washes in PBS- T at room temperature. Slides were dehydrated in 95% ethanol for 10  min at room temperature followed by either the FISH or immunofluorescence protocol below. FISH Following dehydration of the slides, excess ethanol was removed, 15 µl of hybridization solution (50% formamide, 3  X SSC, 10% dextran sulfate, 10  ng labeled DNA probe in water) was added, and a coverslip was placed on each slide. Slides were placed into a slide chamber, and the FISH incubation protocol was conducted in a PCR machine overnight (80 °C for 10 min, 0.5 °C/sec to 50 °C, 50 °C for 1 min, 0.5 °C/sec to 45 °C, 45 °C for 1 min, 0.5 °C/sec to 40 °C, 40 °C for 1 min, 0.5 °C/sec to 38 °C, 38 °C for 1 min, 0.5 °C/sec to 37 °C, 37 °C overnight). After overnight incubation at 37 °C, slides were washed at 39 °C using the following regime: three times (15 min each) in 2 X SSC (0.3 M NaCl and 30 mM Na3C6H5O7) in 50% formamide, three times (10 min each) in 2 X SSC in 25% formamide, three times (10 min each) in 2 X SSC, and three times (1 min each) in 1 X SSC. Samples were incubated in PBS- T for 10 min at room temperature, and immunofluorescence staining was performed as described below. Immunofluorescence of gut nuclei Following dehydration of slides subjected to immunofluorescence only or to PBS- T treatment (after FISH protocol), the excess liquid was removed (either ethanol from the dehydration step, or PBS- T from FISH protocol) and 20  µl of affinity- purified primary antibodies (Cbr DPY- 27 and Cbr MIX- 1 peptide antibodies [Covance, Inc.]) in PBS- T were added at 1:200 dilution. Samples were incubated Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 30 of 39 Chromosomes and Gene Expression Research article in a humid chamber for between 4 hr and overnight. Slides were washed three times (10 min each) in PBS- T at room temperature and then incubated in secondary antibodies for 3–6  hr. Slides were washed three times (10 min each) in PBS- T at room temperature before Prolong (Molecular Probes, P36934) with DAPI (1 µg/ml) was added, and the samples were imaged using a Leica TCS SP2 AOBS. Antibodies used: anti- DPY- 27 rabbit antibody raised to Cbr DPY- 27 C- terminal peptide DVQS EAPS AGRP VETD REGS YTNFD, anti- DPY- 27 guinea pig antibody raised to the same Cbr DPY- 27 peptide, anti- MIX- 1 rabbit antibody raised to Cbr MIX- 1 C- terminal peptide EATK KPSK KSAK KAVQ NTDDEME, Alexa Flour 488 goat anti- rabbit antibody (Molecular Probes, A11034), Alexa Flour 488 goat anti- guinea pig antibody (Molecular Probes, A11073), and Alexa Flour 594 goat anti- rabbit antibody (Molecular Probes, A11037). Immunofluorescence of embryos Embryos were picked into 4 µl of water on poly- lysine- treated slides. After adding a coverslip, slides were frozen in liquid nitrogen for at least 1 min. Coverslips were removed rapidly with a razor blade and samples were dehydrated in 95% ethanol for 10 min. Next, 40 µl of fix solution (2% paraformalde- hyde in egg buffer) was added and slides were incubated in a humid chamber for 10 min. Slides were washed three times (10 min each) in PBS- T at room temperature. Antibody staining was performed as described above for gut nuclei. Immunoprecipitation analysis using DPY-27 antibodies To determine whether DPY- 27, an SMC4 condensin subunit ortholog, interacts with MIX- 1, the SMC2 condensin subunit ortholog, we immunoprecipitated proteins with rabbit DPY- 27 antibodies and performed mass spectrometry of trypsinized protein bands excised from an SDS- PAGE gel, using protocols from Mets and Meyer, 2009. Analysis was performed on proteins within the molecular weight range expected for condensin subunits. In addition to MIX- 1 peptides, MALDI- TOF analysis revealed peptides from four common high- molecular weight contaminants in immunoprecipitation experiments (Table  1): the three vitellogenin yolk proteins VIT- 2, VIT- 4, VIT- 5, and CBG14234, an ortholog of VIT- 4. No protein bands corresponding to the molecular weights of SDC- 2 or SDC- 3 were visible by SDS- PAGE. Western blot analysis of anti-DPY-27 and anti-MIX-1 antibodies Fifty adult hermaphrodites from strain AF16 [wild- type C. briggsae], strain TY5774 Cbr dpy- 27(y706), 3xFLAG- tagged Cbr dpy- 27, or strain TY5005 [Cbr dpy- 27(y436)] were picked into 25  µL of water, diluted with 25 µL of 2 x SDS Sample Buffer, and heat denatured at 98 °C for 4 min. Samples (20 µL) were fractionated with 3–8% Tris Acetate SDS- PAGE electrophoresis and transferred onto nitrocellu- lose membranes using standard conditions (60 min at 100 V). Membranes were immunoblotted with either rabbit polyclonal anti- DPY- 27 (this study) or rabbit polyclonal anti- MIX antibody (this study). Following incubation with a primary antibody, membranes were incubated with a secondary donkey anti- rabbit HRP antibody (Jackson ImmunoResearch, #711- 035- 152, RRID: AB_10015282). Nitrocel- lulose membranes were then incubated in WesternBright Sirius ECL solution (Advansta Corporation, #K- 12043- D20) for 2 min, and the chemiluminescence signal was acquired using Image Lab software (Bio- Rad). Calculation of viability for C. briggsae sdc-2 mutants XX animals: sdc- 2 / + hermaphrodites were crossed to JU935 males, which carry a gfp transgene inte- grated on the X chromosome, and the hermaphrodite cross progeny (sdc- 2 + / + gfp) were moved to individual plates. Three classes of genotype were expected among the self- progeny of sdc- 2 + / + gfp hermaphrodites. Two classes, (+ gfp / + gfp and sdc- 2 + / + gfp) express GFP, whereas the third class, (sdc- 2 + / sdc- 2 +) does not. If sdc- 2 + / sdc- 2 + animals are 100% viable, the expected proportion of non- green animals among the self- progeny of sdc- 2 + / + gfp hermaphrodites is 25%. In each case, the expected number of viable non- green adult progeny is shown in parentheses, and the observed proportion is depicted in the chart as a percentage of the expected number. Wild- type XX viability was calculated among the self- progeny of + + / + gfp animals. XO animals: sdc- 2 + / + gfp hermaphrodites were crossed with + + / O (wild- type) males. Success- fully mated hermaphrodites were identified by the presence of a copulatory plug and then moved to Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 31 of 39 Chromosomes and Gene Expression Research article individual plates. Two classes of genotypes were expected among the progeny of this cross. One class (+ gfp / O) expresses GFP, whereas the other (sdc- 2 + / O) does not. If sdc- 2 + / O animals are 100% viable, the expected proportion of non- green animals among the male progeny is 50%. In each case, the expected number of non- green animals is shown in parentheses, and the observed proportion is depicted in the chart as a percentage of the expected number. Wild- type XO viability was calculated among the male cross- progeny of + + / + gfp hermaphrodites and + + / O males. Calculation for rescue of xol-1 XO-specific lethality in C. briggsae by an sdc-2 mutation The percent viability of wild- type XO animals and mutant XO animals carrying combinations of xol- 1 and sdc- 2 mutations was calculated by formulae that follow. For wild- type XO or xol- 1(y430) XO progeny from crosses of wild- type or xol- 1(y430) hermaphrodites mated with wild- type males, the formula is [(number of F1 males)/(total F1 progeny/2)] × 100, a calculation that assumes successful mating and the potential for 50% male cross progeny among the F1. For xol- 1 sdc- 2 XO double mutants, xol- 1 sdc- 2 / xol- 1 hermaphrodites were mated with wild- type males. Given that xol- 1 XO progeny are inviable, xol- 1 sdc- 2 F1 males should make up 1/3 of viable F1s. Thus, % XO rescue is calculated as [(number of males)/(total progeny/3)] × 100. Genome editing using CRISPR-Cas9 The Cbr dpy- 27(y705) (Figure 1F), Cbr rex- 1 (Figure 8), Cbr rex- 3 (Figure 10), Cbr rex- 4 (Figure 9), and Cbr rex- 7 (Figure 11) mutations, as well as Cel site 2 insertions (Figure 5) and substitutions of Cbr MEX motifs into Cel rex- 33 and substitution of Cbr MEX II motifs into Cel rex- 39 (Figure  12) were made with the CRISPR- Cas9 co- conversion technique using Cas9 RNP injections and species- appropriate co- injection markers (Farboud et al., 2019). C. elegans editing utilized the dpy- 10 roller marker, and C. briggsae editing utilized the ben- 1 marker. The tracrRNA and crRNA guides (Dhar- macon) were resuspended in 600 μM of nuclease- free water (Ambion AM9937). The Cas9 RNP mixture for injections included 5 μl Cas9 protein (UC Berkeley QB3 MacroLab, 10 mg/ml), 1.15 μl 2 M HEPES, pH 7.5, 0.35 μl 0.5 M KCl, 0.5 μl 600 μM dpy- 10 crRNA, 1 μl target crRNA (Supplementary file 4), 5 μl tracrRNA, and 7 μl nuclease- free water. The Cas9 RNP mix was incubated at 37 °C for 15 min, and 1 μl of the resulting Cas9 RNP mix was combined with 0.5 μl 10 μM dpy- 10 repair oligo (IDT), 0.5 μl 10 μM rex repair oligo (IDT), and 8 μl nuclease- free water. After centrifuging at 16,100 × g for 10 min, the Cas9 RNP mix was injected into gonads of adult hermaphrodites. The target- specific sequences for Cas9 guide RNAs are listed in Supplementary file 4. The DNA sequences for the repair templates are listed in Supplementary file 5. For C. elegans, injected adults were placed on NGM plates. After three days of growth at 25 °C, progeny with the roller phenotype were picked to individual plates and allowed to lay embryos. The roller parents were picked into lysis buffer, and the edited site was amplified and sequenced to identify the worms that were edited. The homozygous progeny from properly edited worms were backcrossed twice to wild- type (N2) worms before usage in experiments. For C. briggsae, mutants were isolated as published (Farboud et al., 2019). The homozygous progeny from those were backcrossed twice to AF16 worms before usage in experiments. Primers used for genotyping are in Supplementary file 2. C. briggsae ChIP extract preparation Mixed- stage animals were grown on MYOB agar plates with concentrated HB101 bacteria at 20 °C. Animals were cross- linked with 2% formaldehyde for 10 min and quenched with 100 mM Tris- HCl, pH 7.5. Cross- linked animals were resuspended in 1 ml of FA Buffer (150 mM NaCl, 50 mM HEPES- KOH, pH 7.6, 1 mM EDTA, 1% Triton X- 100, 0.1% sodium deoxycholate, 1 mM DTT, and protease inhibitor cocktail [Calbiochem, #539134]) for every 1 gram of animals. This mixture was frozen in liquid nitrogen and then ground under liquid nitrogen by mortar and pestle for 3 min. Once thawed, the mixture was then homogenized with 50 strokes in a Dounce homogenizer. The chromatin was sheared using the Covaris S2 (20% duty factor, power level 8, 200 cycles per burst) for a total of 30 min processing time (60 sec ON, 45 sec OFF, 30 cycles). The concentration of protein in each extract was quantified using the BCA assay (Thermo Fisher, #23228). C. elegans ChIP-seq extract preparation Mixed- stage embryos were harvested from hermaphrodites grown on MYOB agar plates with concen- trated HB101 bacteria at 20  °C. Embryos were cross- linked with 2% formaldehyde for 10  min and Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 32 of 39 Chromosomes and Gene Expression Research article quenched with 100 mM Tris- HCl, pH 7.5. Cross- linked embryos were resuspended in 1 ml of FA Buffer (150 mM NaCl, 50 mM HEPES- KOH, pH 7.6, 1 mM EDTA, 1% Triton X- 100, 0.1% sodium deoxycho- late, 1 mM DTT, and protease inhibitor cocktail [Calbiochem, #539134]) for every 1 gram of embryos and homogenized with 50 strokes in a Dounce homogenizer. The chromatin was sheared using the Covaris S2 (20% duty factor, power level 8, 200 cycles per burst) for a total of 30 min processing time (60 sec ON, 45 sec OFF, 30 cycles). The concentration of protein in each extract was quantified using the BCA assay (Thermo Fisher, #23228). C. briggsae ChIP reactions To perform the ChIP reactions, a 50 µl bed volume of protein A Dynabeads (Thermo Fisher, #10001D) was re- suspended in 1 ml of FA Buffer (150 mM NaCl, 50 mM HEPES- KOH, pH 7.6, 1 mM EDTA, 1% Triton X- 100, 0.1% sodium deoxycholate, 1  mM DTT, and protease inhibitor cocktail [Calbiochem, #539134]). The beads were incubated in a microcentrifuge tube with 5 µg of anti- FLAG antibodies (Sigma- Aldrich, #F1804) and 5  µg of rabbit anti- mouse IgG antibodies (Jackson ImmunoResearch, #315- 005- 003), or 5  µg of mouse IgG (Sigma- Aldrich, #I5381), and 5  µg of rabbit anti- mouse IgG antibodies (Jackson ImmunoResearch, #315- 005- 003), for 90  min at room temperature. Tubes with incubated beads were placed on a magnetic rack, and the liquid was discarded. Extracts containing 2 mg of protein ChIPs were increased in volume to 1 ml with FA buffer and then added to each tube of Dynabeads for a 90 min incubation. The Dynabead- extract mixture was washed at room temperature twice with FA Buffer (150 mM NaCl), once with FA Buffer (1 M NaCl), once with FA Buffer (500 mM NaCl), once with TEL buffer (10 mM Tris- HCl, pH 8.0, 250 mM LiCl, 1% IGEPAL CA- 630 [Sigma- Aldrich, #I3021], 1% sodium deoxycholate, 1 mM EDTA), and twice with TE Buffer (10 mM Tris, pH 8.0, 1 mM EDTA). Protein and DNA were eluted with 250 µl of buffer (1% SDS, 250 mM NaCl, 1 mM EDTA) at 65 °C for 20 min. C. elegans ChIP reactions To perform the ChIP reactions, a 25 µl bed volume of protein A Dynabeads (Thermo Fisher, #10001D) was re- suspended in 1 ml of FA Buffer (150 mM NaCl, 50 mM HEPES- KOH, pH 7.6, 1 mM EDTA, 1% Triton X- 100, 0.1% sodium deoxycholate, 1  mM DTT, and protease inhibitor cocktail [Calbiochem, #539134]). The beads were incubated in a microcentrifuge tube with 3  µg rabbit anti- SDC- 3 (lab stock), or 3 µg rabbit IgG (Jackson Immunoresearch, #301- 005- 003) for 90 min at room temperature. Tubes with incubated beads were placed on a magnetic rack and the liquid was discarded. Protocols for the incubation of extract with beads and elution of protein and DNA from beads were the same as those described for C. briggsae ChIP reactions. ChIP-seq, illumina sequencing, and data processing Sequencing libraries were prepared with the eluted materials from ChIP reactions as published (Zhong et al., 2010) with minor changes: sequencing adapters were obtained from Bioo (NEXTflex), and adapters were ligated using the NEB Quick Ligation Kit (M2200). Libraries were sequenced on the Illumina HiSeq 4000 platforms. After barcode removal, reads were aligned uniquely to the C. briggsae CB4 genome using the default settings in Bowtie version 2.3.4.3. To account for read depth, ChIP signal was normalized to the total number of reads that uniquely aligned to the genome. C. elegans qPCR To perform qPCR reactions, protein, and DNA from a C. elegans ChIP reaction or from 50% of a control extract (1 mg protein) were de- crosslinked at 65 °C for at least 4 hr with 150 μg/ml Proteinase K (Sigma, #3115887001). DNA from each ChIP reaction or from the control extract was isolated using the Qiagen PCR purification kit and diluted to a final volume of 200 μl with (10 mM Tris- HCl, pH 8.5). For quantitative PCR, the immunoprecipitated DNAs were quantified by comparing their threshold cycle to the standard curve from control DNA (10% and three serial 10- fold dilutions). For the site two insertions, the DCC levels at each inserted rex site were calculated for each biological replicate as a ratio of the average DCC level at five control rex sites (rex- 8, rex- 16, rex- 32, rex- 35, and rex- 48). For all experiments involving endogenous Cel rex- 39 in Figure 12B or involving endogenous Cel rex- 33 in Figure 12E, the DCC levels at each inserted rex site were calculated for each biological replicate as a Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 33 of 39 Chromosomes and Gene Expression Research article ratio of the average DCC level at seven control rex sites (rex- 8, rex- 14, rex- 16, rex- 32, rex- 35, rex- 36, and rex- 48). Primers used for qPCR are listed in Supplementary file 2. C. briggsae qPCR To perform the qPCR reactions, protein, and DNA from a C. briggsae ChIP reaction or from 50% of a control extract (1 mg protein) were de- crosslinked at 65 °C for at least 4 hr with 150 μg/ml Proteinase K (Sigma, #3115887001). DNA from each ChIP reaction or from the control extract was isolated using the Qiagen PCR purification kit and diluted to a final volume of 400 μl with (10 mM Tris- HCl, pH 8.5). For quantitative PCR, the immunoprecipitated DNAs were quantified by comparing their threshold cycle to the standard curve from control DNA (10% and three serial 10- fold dilutions). For the endog- enous rex site mutations, the DCC levels at each inserted rex site were calculated for each biological replicate as a ratio of the average DCC level at four control rex sites (rex- 1, rex- 2, rex- 5, and rex- 9). Primers used for qPCR are listed in Supplementary file 2. Identification of C. briggsae DCC binding motifs The 500  bp DNA sequence centered on each C. briggsae SDC- 2 ChIP- seq peak location for the 12 Cbr rex sites was isolated from the CB4 reference genome. Motif candidates were obtained by inputting twelve 500 bp sequences onto MEME on the MEME- suite website (Bailey and Elkan, 1994; Bailey et al., 2015). The settings used to identify motif candidates were the classic mode and any number of repetitions (anr). The X:A enrichment was calculated for motif candidates. The two motif candidates enriched on the Cbr X chromosomes were named Cbr MEX for the 13 bp motif and Cbr MEX II for the 30 bp motif (Figure 6). X:A motif enrichment calculation The Patser program (version 3e) (Hertz and Stormo, 1999) was used to calculate the natural log of the probability (ln[P]) of finding a match to the Cbr MEX motif, Cbr MEX II, Cel MEX motif, and Cel MEX II motif at all positions along each chromosome, as explained in Fuda et  al., 2022. For each threshold value, the number of motifs with ln[P] values less than the value (better match) was summed for X and for autosomes. The number of autosomal motifs was divided by the total number of auto- somal base pairs to find the number of motifs per base pair. The number of motifs per base pair of X was calculated similarly. The final X:A ratio was calculated by dividing the motifs per base pair for X by the motifs per base pair for the autosomes. C. elegans DCC binding assay performed in vitro The in vitro Cel DCC binding assays (Figure 12 and Figure 13) were performed as described previ- ously in Fuda et al., 2022. Briefly, protein extracts for the assays were made from C. elegans strain TY4573 [sdc- 2(y74) X; yEx992], in which the extrachromosomal array yEx992 carried multiple copies of a transgene that encoded Cel SDC- 2 tagged with 3xFLAG at its 5’ end. Synchronized gravid TY4573 animals were bleached to yield embryos that were resuspended in homogenization buffer (50 mM HEPES, pH 7.5, 140 mM KCl, 1 mM EDTA, 10% v/v glycerol, 0.5% v/v IGEPAL CA- 630, 5 mM DTT, 1 mM PMSF, and protease inhibitor cocktail), flash frozen in liquid nitrogen, and stored at –80 °C. After thawing on ice, the embryo suspension was sonicated (Covaris S2) in 1 mL batches for 6 min (duty cycle 10%, intensity five, cycles/burst 200) and centrifuged at 16,100 × g for 30 min at 4 °C to pellet embryo debris. The supernatant was removed, flash frozen in liquid nitrogen, and stored at –80 °C. Total protein concentration was determined using the BCA assay (Thermo Fisher, #PI23227). Both the 601 bp wild- type rex DNA and negative control DNA (np1) were obtained by amplifying DNA from worm lysates with oligonucleotides listed in Supplementary file 6. Amplified worm DNA was cloned into the TOPOBlunt vector (Thermo Fisher, #450245). The TOPOBlunt- specific oligonucle- otides kb157 and kb184r were used to amplify cloned DNA fragments, and the final DNA products had 22  bp (5’- CAGT GTGC TGGA ATTC GCCC TT) and 28  bp (5’- GTGA TGGA TATC TGCA GAAT TCGC CCTT -3’) sequences added to the 5’ and 3’ ends, respectively. The TOPOBlunt- specific oligonucle- otide kb157 contained a 5’ Biotin- TEG moiety (IDTDNA). Mutant and de novo- designed DNA probes were obtained by amplifying gblock fragments using the kb157/kb184r primer pair. The final products assayed in vitro were 651 bp DNA fragments. Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 34 of 39 Chromosomes and Gene Expression Research article For the DNA pull- down assays, Dynabeads M- 280 Streptavidin (Thermo Fisher, #11205D) were washed and coupled with the biotinylated DNA (110 ng of DNA per µg of beads) according to manu- facturer instructions. After incubation, the beads were washed with buffer A (50 mM Hepes pH 7.5, 70  mM KCl, 10  mM MgCl2, 10% glycerol, 2  mM DTT, 8.5  mg/mL BSA, 1  mM PMSF, and protease inhibitor cocktail), and re- suspended at 15 ng/µL. Embryo extract was thawed on ice and centrifuged at 16,100 × g for 30 min at 4 °C to remove any aggregates. Embryo extract (800 µg) was incubated with 10 µL of beads (150 ng coupled DNA) in final buffer B (50 mM Hepes pH 7.6, 105 mM KCl, 5 mM MgCl2, 10% v/v glycerol, 0.5 mM EDTA pH 8.0, 0.25% v/v Igepal CA- 630, 1 mM DTT, 4.25 mg/mL BSA, 1 mM PMSF, protease inhibitors cocktail, and 1 µg poly(dI- dC)). After a 3–4 hr incubation at 4 °C, samples were centrifuged briefly, incubated on a magnetic rack for 2 min, and the supernatant was removed. Beads were washed with 300 µL buffer B with a short vortex step and placed on ice for 2 min. After incubation, tubes were centrifuged briefly, incubated on a magnetic rack for 2 min, and the supernatant was removed. The wash step was repeated two additional times. For the elution step, magnetic beads were re- suspended in 50 µL buffer C (50 mM Hepes pH 7.5, 2 M NaCl, 5 mM MgCl2, 10% v/v glycerol, 0.5 mM EDTA, 1 mM DTT, and protease inhibitor cocktail) and incubated on ice for 30–45 min. The eluate was transferred to a clean tube, flash frozen in liquid nitrogen, and stored at –80 °C. For Western dot blot assays, eluted samples (3.5  μL) were spotted in triplicates onto dry nitro- cellulose membranes and left to dry for 1 hr at room temperature. Nitrocellulose membranes were incubated in blocking buffer containing milk (5%  w/v milk in 1  x TBS supplemented with 0.1%  v/v Tween- 20) for 1 hr, followed by incubation with primary anti- FLAG antibody (Sigma- Aldrich, #F1804, RRID: AB_262044) for 1  hr, and with secondary donkey anti- mouse HRP antibody (Jackson Immu- noResearch, #715- 035- 151, RRID: AB_2340771) for 1 hr. Nitrocellulose membranes were incubated in WesternBright Sirius ECL solution (Advansta Corporation, #K- 12043- D20) for 3 min, and the chemi- luminescence signal was acquired using Image Lab software (Bio- Rad). The dot blot intensities were quantified using the Volume option in Image Lab software. Acknowledgements We are grateful to E Haag and his laboratory for generously providing expertise and reagents to make and screen C. briggsae deletion pools, D King for the MALDI- TOF analysis, A Wood for initiating ZFN mutagenesis of C. briggsae sdc- 2, D Stalford for figure preparation, T Cline and laboratory members for valuable discussions, and the QB3 Genomics Facility (RRID:SCR_022170) for DNA sequencing. This work was supported in part by NIH Grant R35 GM131845 (to BJM). BJM is an investigator of the Howard Hughes Medical Institute. Additional information Competing interests Caitlin Schartner: Caitlin Schartner is affiliated with Roche Diagnostics. The author has no financial interests to declare. The other authors declare that no competing interests exist. Funding Funder Howard Hughes Medical Institute National Institutes of Health Grant reference number Author Barbara J Meyer R35 GM131845 Barbara J Meyer The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 35 of 39 Chromosomes and Gene Expression Research article Author contributions Qiming Yang, Katjuša Brejc, Conceptualization, Resources, Data curation, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing – review and editing; Te- Wen Lo, Conceptualiza- tion, Resources, Data curation, Formal analysis, Supervision, Validation, Investigation, Visualization, Methodology, Writing – review and editing; Caitlin Schartner, Resources, Formal analysis, Validation, Investigation, Visualization, Methodology; Edward J Ralston, Conceptualization, Resources, Data curation, Formal analysis, Supervision, Validation, Investigation, Visualization, Methodology; Denise M Lapidus, Resources, Data curation, Investigation, Methodology; Barbara J Meyer, Conceptualiza- tion, Resources, Data curation, Formal analysis, Supervision, Funding acquisition, Validation, Investi- gation, Visualization, Methodology, Writing – original draft, Project administration, Writing – review and editing Author ORCIDs Qiming Yang Te- Wen Lo Katjuša Brejc Barbara J Meyer http://orcid.org/0000-0003-1419-868X http://orcid.org/0000-0002-1231-5531 http://orcid.org/0000-0002-4562-6109 http://orcid.org/0000-0002-6530-4588 Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.85413.sa1 Author response https://doi.org/10.7554/eLife.85413.sa2 Additional files Supplementary files •  MDAR checklist •  Supplementary file 1. List of alleles and strains used in this study. •  Supplementary file 2. List of primers. •  Supplementary file 3. Chromosome- specific BACs used to generate FISH probes. •  Supplementary file 4. List of target- specific sequences for guide RNAs used in CRISPR / Cas9 genome editing experiments. •  Supplementary file 5. DNA sequences of repair templates used in CRISPR / Cas9 genome editing experiments. •  Supplementary file 6. DNA templates used for in vitro DCC binding assays. Data availability GEO GSE214714 is the accession number for the ChIP- seq data reported in this manuscript. The following dataset was generated: Author(s) Yang Q, Lo T, Brejc K, Schartner CM, Ralston EJ, Lapidus DM, Meyer BJ Year 2022 Dataset title Dataset URL Database and Identifier Caenorhabditis briggsae SDC- 2 and DPY- 27 ChIP- seq https://www. ncbi. nlm. nih. gov/ geo/ query/ acc. cgi? acc= GSE214714 NCBI Gene Expression Omnibus, GSE214714 References Albritton SE, Kranz AL, Winterkorn LH, Street LA, Ercan S. 2017. Cooperation between a hierarchical set of recruitment sites targets the X chromosome for dosage compensation. eLife 6:e23645. DOI: https://doi.org/10. 7554/eLife.23645, PMID: 28562241 Alekseyenko AA, Peng S, Larschan E, Gorchakov AA, Lee OK, Kharchenko P, McGrath SD, Wang CI, Mardis ER, Park PJ, Kuroda MI. 2008. A sequence motif within chromatin entry sites directs MSL establishment on the Drosophila X chromosome. Cell 134:599–609. DOI: https://doi.org/10.1016/j.cell.2008.06.033, PMID: 18724933 Alekseyenko AA, Ellison CE, Gorchakov AA, Zhou Q, Kaiser VB, Toda N, Walton Z, Peng S, Park PJ, Bachtrog D, Kuroda MI. 2013. Conservation and de novo acquisition of dosage compensation on newly evolved sex Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 36 of 39 Chromosomes and Gene Expression Research article chromosomes in Drosophila. Genes & Development 27:853–858. DOI: https://doi.org/10.1101/gad.215426. 113, PMID: 23630075 Anderson EC, Frankino PA, Higuchi- Sanabria R, Yang Q, Bian Q, Podshivalova K, Shin A, Kenyon C, Dillin A, Meyer BJ. 2019. X chromosome domain architecture regulates Caenorhabditis elegans lifespan but not dosage compensation. Developmental Cell 51:192–207.. DOI: https://doi.org/10.1016/j.devcel.2019.08.004, PMID: 31495695 Bailey TL, Elkan C. 1994. Fitting a mixture model by expectation maximization to discover motifs in biopolymers. Proceedings. International Conference on Intelligent Systems for Molecular Biology :28–36 PMID: 7584402. Bailey TL, Johnson J, Grant CE, Noble WS. 2015. The MEME suite. Nucleic Acids Research 43:W39–W49. DOI: https://doi.org/10.1093/nar/gkv416, PMID: 25953851 Bayer EA, Sun H, Rafi I, Hobert O. 2020. Temporal, spatial, sexual and environmental regulation of the master regulator of sexual differentiation in C. elegans. Current Biology 30:3604–3616. DOI: https://doi.org/10.1016/j. cub.2020.06.060, PMID: 32707064 Berkseth M, Ikegami K, Arur S, Lieb JD, Zarkower D. 2013. TRA- 1 chip- seq reveals regulators of sexual differentiation and multilevel feedback in nematode sex determination. PNAS 110:16033–16038. DOI: https:// doi.org/10.1073/pnas.1312087110, PMID: 24046365 Carroll SB. 2008. Evo- Devo and an expanding evolutionary synthesis: a genetic theory of morphological evolution. Cell 134:25–36. DOI: https://doi.org/10.1016/j.cell.2008.06.030, PMID: 18614008 Chan RC, Severson AF, Meyer BJ. 2004. Condensin restructures chromosomes in preparation for meiotic divisions. The Journal of Cell Biology 167:613–625. DOI: https://doi.org/10.1083/jcb.200408061, PMID: 15557118 Chang CP, Brocchieri L, Shen WF, Largman C, Cleary ML. 1996. Pbx modulation of Hox homeodomain amino- terminal arms establishes different DNA- binding specificities across the Hox locus. Molecular and Cellular Biology 16:1734–1745. DOI: https://doi.org/10.1128/MCB.16.4.1734, PMID: 8657149 Chu DS, Dawes HE, Lieb JD, Chan RC, Kuo AF, Meyer BJ. 2002. A molecular link between gene- specific and chromosome- wide transcriptional repression. Genes & Development 16:796–805. DOI: https://doi.org/10. 1101/gad.972702, PMID: 11937488 Chuang PT, Albertson DG, Meyer BJ. 1994. DPY- 27: a chromosome condensation protein homolog that regulates C. elegans dosage compensation through association with the X chromosome. Cell 79:459–474. DOI: https://doi.org/10.1016/0092-8674(94)90255-0, PMID: 7954812 Chuang PT, Lieb JD, Meyer BJ. 1996. Sex- specific assembly of a dosage compensation complex on the nematode X chromosome. Science 274:1736–1739. DOI: https://doi.org/10.1126/science.274.5293.1736, PMID: 8939870 Csankovszki G, McDonel P, Meyer BJ. 2004. Recruitment and spreading of the C. elegans dosage compensation complex along X chromosomes. Science 303:1182–1185. DOI: https://doi.org/10.1126/science.1092938, PMID: 14976312 Csankovszki G, Collette K, Spahl K, Carey J, Snyder M, Petty E, Patel U, Tabuchi T, Liu H, McLeod I, Thompson J, Sarkeshik A, Yates J, Meyer BJ, Hagstrom K. 2009. Three distinct condensin complexes control C. elegans chromosome dynamics. Current Biology 19:9–19. DOI: https://doi.org/10.1016/j.cub.2008.12.006, PMID: 19119011 Cutter AD. 2008. Divergence times in Caenorhabditis and Drosophila inferred from direct estimates of the neutral mutation rate. Molecular Biology and Evolution 25:778–786. DOI: https://doi.org/10.1093/molbev/ msn024, PMID: 18234705 Cutter AD, Morran LT, Phillips PC. 2019. Males, outcrossing, and sexual selection in Caenorhabditis nematodes. Genetics 213:27–57. DOI: https://doi.org/10.1534/genetics.119.300244, PMID: 31488593 Dawes HE, Berlin DS, Lapidus DM, Nusbaum C, Davis TL, Meyer BJ. 1999. Dosage compensation proteins targeted to X chromosomes by a determinant of hermaphrodite fate. Science 284:1800–1804. DOI: https://doi. org/10.1126/science.284.5421.1800, PMID: 10364546 De Kumar B, Darland DC. 2021. The Hox protein conundrum: the specifics of DNA binding for Hox proteins and their partners. Developmental Biology 477:284–292. DOI: https://doi.org/10.1016/j.ydbio.2021.06.002, PMID: 34102167 Ellison C, Bachtrog D. 2019. Contingency in the convergent evolution of a regulatory network: dosage compensation in Drosophila. PLOS Biology 17:e3000094. DOI: https://doi.org/10.1371/journal.pbio.3000094, PMID: 30742611 Farboud B, Severson AF, Meyer BJ. 2019. Strategies for efficient genome editing using CRISPR- Cas9. Genetics 211:431–457. DOI: https://doi.org/10.1534/genetics.118.301775, PMID: 30504364 Fuda NJ, Brejc K, Kruesi WS, Ralston EJ, Bigley R, Shin A, Okada M, Meyer BJ. 2022. Combinatorial clustering of distinct DNA motifs directs synergistic binding of Caenorhabditis elegans dosage compensation complex to X chromosomes. PNAS 119:e2211642119. DOI: https://doi.org/10.1073/pnas.2211642119, PMID: 36067293 Gabler F, Nam S- Z, Till S, Mirdita M, Steinegger M, Söding J, Lupas AN, Alva V. 2020. Protein sequence analysis using the MPI bioinformatics toolkit. Current Protocols in Bioinformatics 72:e108. DOI: https://doi.org/10. 1002/cpbi.108, PMID: 33315308 Haag ES. 2005. The evolution of nematode sex determination: C. elegans as a reference point for comparative biology. WormBook 1:1–14. DOI: https://doi.org/10.1895/wormbook.1.120.1, PMID: 18050417 Hagstrom KA, Holmes VF, Cozzarelli NR, Meyer BJ. 2002. C. elegans condensin promotes mitotic chromosome architecture, centromere organization, and sister chromatid segregation during mitosis and meiosis. Genes & Development 16:729–742. DOI: https://doi.org/10.1101/gad.968302, PMID: 11914278 Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 37 of 39 Chromosomes and Gene Expression Research article Halder G, Callaerts P, Gehring WJ. 1995. Induction of ectopic eyes by targeted expression of the eyeless gene in Drosophila. Science 267:1788–1792. DOI: https://doi.org/10.1126/science.7892602, PMID: 7892602 Henikoff S, Ahmad K, Malik HS. 2001. The centromere paradox: stable inheritance with rapidly evolving DNA. Science 293:1098–1102. DOI: https://doi.org/10.1126/science.1062939, PMID: 11498581 Hertz GZ, Stormo GD. 1999. Identifying DNA and protein patterns with statistically significant alignments of multiple sequences. Bioinformatics 15:563–577. DOI: https://doi.org/10.1093/bioinformatics/15.7.563, PMID: 10487864 Hill RE, Favor J, Hogan BL, Ton CC, Saunders GF, Hanson IM, Prosser J, Jordan T, Hastie ND, van Heyningen V. 1991. Mouse small eye results from mutations in a paired- like homeobox- containing gene. Nature 354:522– 525. DOI: https://doi.org/10.1038/354522a0, PMID: 1684639 Jans J, Gladden JM, Ralston EJ, Pickle CS, Michel AH, Pferdehirt RR, Eisen MB, Meyer BJ. 2009. A condensin- like dosage compensation complex acts at a distance to control expression throughout the genome. Genes & Development 23:602–618. DOI: https://doi.org/10.1101/gad.1751109, PMID: 19270160 Kelleher DF, de Carvalho CE, Doty AV, Layton M, Cheng AT, Mathies LD, Pilgrim D, Haag ES. 2008. Comparative genetics of sex determination: masculinizing mutations in Caenorhabditis briggsae. Genetics 178:1415–1429. DOI: https://doi.org/10.1534/genetics.107.073668, PMID: 18245372 Kruesi WS, Core LJ, Waters CT, Lis JT, Meyer BJ. 2013. Condensin controls recruitment of RNA polymerase II to achieve nematode X- chromosome dosage compensation. eLife 2:e00808. DOI: https://doi.org/10.7554/eLife. 00808, PMID: 23795297 Kuzu G, Kaye EG, Chery J, Siggers T, Yang L, Dobson JR, Boor S, Bliss J, Liu W, Jogl G, Rohs R, Singh ND, Bulyk ML, Tolstorukov MY, Larschan E. 2016. Expansion of GA dinucleotide repeats increases the density of clamp binding sites on the X- chromosome to promote Drosophila dosage compensation. PLOS Genetics 12:e1006120. DOI: https://doi.org/10.1371/journal.pgen.1006120, PMID: 27414415 Lieb JD, Capowski EE, Meneely P, Meyer BJ. 1996. DPY- 26, a link between dosage compensation and meiotic chromosome segregation in the nematode. Science 274:1732–1736. DOI: https://doi.org/10.1126/science.274. 5293.1732, PMID: 8939869 Lieb JD, Albrecht MR, Chuang PT, Meyer BJ. 1998. MIX- 1: an essential component of the C. elegans mitotic machinery executes X chromosome dosage compensation. Cell 92:265–277. DOI: https://doi.org/10.1016/ s0092-8674(00)80920-4, PMID: 9458050 Liénard MA, Araripe LO, Hartl DL. 2016. Neighboring genes for DNA- binding proteins rescue male sterility in Drosophila hybrids. PNAS 113:E4200–E4207. DOI: https://doi.org/10.1073/pnas.1608337113, PMID: 27357670 Loda A, Collombet S, Heard E. 2022. Gene regulation in time and space during X- chromosome inactivation. Nature Reviews. Molecular Cell Biology 23:231–249. DOI: https://doi.org/10.1038/s41580-021-00438-7, PMID: 35013589 Ludwiczak J, Winski A, Szczepaniak K, Alva V, Dunin- Horkawicz S. 2019. DeepCoil- a fast and accurate prediction of coiled- coil domains in protein sequences. Bioinformatics 35:2790–2795. DOI: https://doi.org/10.1093/ bioinformatics/bty1062, PMID: 30601942 Madeira F, Pearce M, Tivey ARN, Basutkar P, Lee J, Edbali O, Madhusoodanan N, Kolesnikov A, Lopez R. 2022. Search and sequence analysis tools services from EMBL- EBI in 2022. Nucleic Acids Research 50:W276–W279. DOI: https://doi.org/10.1093/nar/gkac240, PMID: 35412617 Malicki J, Schughart K, McGinnis W. 1990. Mouse hox- 2.2 specifies thoracic segmental identity in Drosophila embryos and larvae. Cell 63:961–967. DOI: https://doi.org/10.1016/0092-8674(90)90499-5, PMID: 1979525 Malik HS, Henikoff S. 2001. Adaptive evolution of Cid, a centromere- specific histone in Drosophila. Genetics 157:1293–1298. DOI: https://doi.org/10.1093/genetics/157.3.1293, PMID: 11238413 Meisel RP, Davey T, Son JH, Gerry AC, Shono T, Scott JG. 2016. Is multifactorial sex determination in the house fly, Musca domestica (L.), stable over time? The Journal of Heredity 107:615–625. DOI: https://doi.org/10. 1093/jhered/esw051, PMID: 27540102 Mets DG, Meyer BJ. 2009. Condensins regulate meiotic DNA break distribution, thus crossover frequency, by controlling chromosome structure. Cell 139:73–86. DOI: https://doi.org/10.1016/j.cell.2009.07.035, PMID: 19781752 Meyer BJ. 2022a. Mechanisms of sex determination and X- chromosome dosage compensation. Genetics 220:iyab197. DOI: https://doi.org/10.1093/genetics/iyab197, PMID: 35100381 Meyer BJ. 2022b. The X chromosome in C. elegans sex determination and dosage compensation. Current Opinion in Genetics & Development 74:101912. DOI: https://doi.org/10.1016/j.gde.2022.101912, PMID: 35490475 Miller LM, Plenefisch JD, Casson LP, Meyer BJ. 1988. Xol- 1: a gene that controls the male modes of both sex determination and X chromosome dosage compensation in C. elegans. Cell 55:167–183. DOI: https://doi.org/ 10.1016/0092-8674(88)90019-0, PMID: 3167975 Nitta KR, Jolma A, Yin Y, Morgunova E, Kivioja T, Akhtar J, Hens K, Toivonen J, Deplancke B, Furlong EEM, Taipale J. 2015. Conservation of transcription factor binding specificities across 600 million years of bilateria evolution. eLife 4:e04837. DOI: https://doi.org/10.7554/eLife.04837, PMID: 25779349 Nusbaum C, Meyer BJ. 1989. The Caenorhabditis elegans gene sdc- 2 controls sex determination and dosage compensation in XX animals. Genetics 122:579–593. DOI: https://doi.org/10.1093/genetics/122.3.579, PMID: 2759421 Pferdehirt RR, Kruesi WS, Meyer BJ. 2011. An MLL/COMPASS subunit functions in the C. elegans dosage compensation complex to target X chromosomes for transcriptional regulation of gene expression. Genes & Development 25:499–515. DOI: https://doi.org/10.1101/gad.2016011, PMID: 21363964 Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 38 of 39 Chromosomes and Gene Expression Research article Quiring R, Walldorf U, Kloter U, Gehring WJ. 1994. Homology of the eyeless gene of Drosophila to the small eye gene in mice and aniridia in humans. Science 265:785–789. DOI: https://doi.org/10.1126/science.7914031, PMID: 7914031 Rhind NR, Miller LM, Kopczynski JB, Meyer BJ. 1995. Xol- 1 acts as an early switch in the C. elegans male/ hermaphrodite decision. Cell 80:71–82. DOI: https://doi.org/10.1016/0092-8674(95)90452-2, PMID: 7813020 Rieder LE, Jordan WT, Larschan EN. 2019. Targeting of the dosage- compensated male X- chromosome during early Drosophila development. Cell Reports 29:4268–4275. DOI: https://doi.org/10.1016/j.celrep.2019.11.095, PMID: 31875538 Rim EY, Clevers H, Nusse R. 2022. The Wnt pathway: from signaling mechanisms to synthetic modulators. Annual Review of Biochemistry 91:571–598. DOI: https://doi.org/10.1146/annurev-biochem-040320-103615, PMID: 35303793 Robert X, Gouet P. 2014. Deciphering key features in protein structures with the new endscript server. Nucleic Acids Research 42:W320–W324. DOI: https://doi.org/10.1093/nar/gku316, PMID: 24753421 Samata M, Akhtar A. 2018. Dosage compensation of the X chromosome: A complex epigenetic assignment involving chromatin regulators and long noncoding rnas. Annual Review of Biochemistry 87:323–350. DOI: https://doi.org/10.1146/annurev-biochem-062917-011816, PMID: 29668306 Schloop AE, Bandodkar PU, Reeves GT. 2020. Formation, interpretation, and regulation of the Drosophila dorsal/NF- bs.ctdb.2019.11.007, PMID: 32143742 κ b gradient. Current Topics in Developmental Biology 137:143–191. DOI: https://doi.org/10.1016/ Sun S, Ting CT, Wu CI. 2004. The normal function of a speciation gene, odysseus, and its hybrid sterility effect. Science 305:81–83. DOI: https://doi.org/10.1126/science.1093904, PMID: 15232104 Talbert P, Henikoff S. 2022. Centromere drive: chromatin conflict in meiosis. Current Opinion in Genetics & Development 77:102005. DOI: https://doi.org/10.1016/j.gde.2022.102005, PMID: 36372007 Ting CT, Tsaur SC, Wu ML, Wu CI. 1998. A rapidly evolving homeobox at the site of A hybrid sterility gene. Science 282:1501–1504. DOI: https://doi.org/10.1126/science.282.5393.1501, PMID: 9822383 Ting CT, Tsaur SC, Sun S, Browne WE, Chen YC, Patel NH, Wu CI. 2004. Gene duplication and speciation in Drosophila: evidence from the odysseus locus. PNAS 101:12232–12235. DOI: https://doi.org/10.1073/pnas. 0401975101, PMID: 15304653 Tsai CJ, Mets DG, Albrecht MR, Nix P, Chan A, Meyer BJ. 2008. Meiotic crossover number and distribution are regulated by a dosage compensation protein that resembles a condensin subunit. Genes & Development 22:194–211. DOI: https://doi.org/10.1101/gad.1618508, PMID: 18198337 Wood AJ, Lo TW, Zeitler B, Pickle CS, Ralston EJ, Lee AH, Amora R, Miller JC, Leung E, Meng X, Zhang L, Rebar EJ, Gregory PD, Urnov FD, Meyer BJ. 2011. Targeted genome editing across species using zfns and TALENs. Science 333:307. DOI: https://doi.org/10.1126/science.1207773, PMID: 21700836 Yatskevich S, Rhodes J, Nasmyth K. 2019. Organization of chromosomal DNA by SMC complexes. Annual Review of Genetics 53:445–482. DOI: https://doi.org/10.1146/annurev-genet-112618-043633, PMID: 31577909 Yin D, Schwarz EM, Thomas CG, Felde RL, Korf IF, Cutter AD, Schartner CM, Ralston EJ, Meyer BJ, Haag ES. 2018. Rapid genome shrinkage in a self- fertile nematode reveals sperm competition proteins. Science 359:55–61. DOI: https://doi.org/10.1126/science.aao0827, PMID: 29302007 Yin H, Wei C, Lee JT. 2021. Revisiting the consequences of deleting the X inactivation center. PNAS 118:e2102683118. DOI: https://doi.org/10.1073/pnas.2102683118, PMID: 34161282 Zarkower D, Hodgkin J. 1992. Molecular analysis of the C. elegans sex- determining gene tra- 1: A gene encoding two zinc finger proteins. Cell 70:237–249. DOI: https://doi.org/10.1016/0092-8674(92)90099-x, PMID: 1339311 Zhong M, Niu W, Lu ZJ, Sarov M, Murray JI, Janette J, Raha D, Sheaffer KL, Lam HYK, Preston E, Slightham C, Hillier LW, Brock T, Agarwal A, Auerbach R, Hyman AA, Gerstein M, Mango SE, Kim SK, Waterston RH, et al. 2010. Genome- wide identification of binding sites defines distinct functions for Caenorhabditis elegans PHA- 4/ FOXA in development and environmental response. PLOS Genetics 6:e1000848. DOI: https://doi.org/10.1371/ journal.pgen.1000848, PMID: 20174564 Zimmermann L, Stephens A, Nam S- Z, Rau D, Kübler J, Lozajic M, Gabler F, Söding J, Lupas AN, Alva V. 2018. A completely reimplemented MPI bioinformatics toolkit with A new hhpred server at its core. Journal of Molecular Biology 430:2237–2243. DOI: https://doi.org/10.1016/j.jmb.2017.12.007, PMID: 29258817 Yang, Lo et al. eLife 2023;12:e85413. DOI: https:// doi. org/ 10. 7554/ eLife. 85413 39 of 39 Chromosomes and Gene Expression
10.7554_elife.85521
RESEARCH ARTICLE Origin of wiring specificity in an olfactory map revealed by neuron type–specific, time- lapse imaging of dendrite targeting Kenneth Kin Lam Wong1, Tongchao Li1*†, Tian- Ming Fu2‡, Gaoxiang Liu3, Cheng Lyu1, Sayeh Kohani1, Qijing Xie1, David J Luginbuhl1, Srigokul Upadhyayula3,4,5, Eric Betzig2,3,6, Liqun Luo1* 1Department of Biology, Howard Hughes Medical Institute, Stanford University, Stanford, United States; 2Howard Hughes Medical Institute, Janelia Research Campus, Ashburn, United States; 3Advanced Bioimaging Center, Department of Molecular and Cell Biology, University of California, Berkeley, Berkeley, United States; 4Molecular Biophysics and Integrated Bioimaging Division, Lawrence Berkeley National Laboratory, Berkeley, United States; 5Chan Zuckerberg Biohub, San Francisco, United States; 6Departments of Molecular and Cell Biology and Physics, Howard Hughes Medical Institute, Helen Wills Neuroscience Institute, University of California, Berkeley, United States Abstract How does wiring specificity of neural maps emerge during development? Formation of the adult Drosophila olfactory glomerular map begins with the patterning of projection neuron (PN) dendrites at the early pupal stage. To better understand the origin of wiring specificity of this map, we created genetic tools to systematically characterize dendrite patterning across develop- ment at PN type–specific resolution. We find that PNs use lineage and birth order combinatorially to build the initial dendritic map. Specifically, birth order directs dendrite targeting in rotating and binary manners for PNs of the anterodorsal and lateral lineages, respectively. Two- photon– and adaptive optical lattice light- sheet microscope–based time- lapse imaging reveals that PN dendrites initiate active targeting with direction- dependent branch stabilization on the timescale of seconds. Moreover, PNs that are used in both the larval and adult olfactory circuits prune their larval- specific dendrites and re- extend new dendrites simultaneously to facilitate timely olfactory map organization. Our work highlights the power and necessity of type- specific neuronal access and time- lapse imaging in identifying wiring mechanisms that underlie complex patterns of func- tional neural maps. Editor's evaluation When a neuron is born it correlates with where it targets in the neuropil and this has been best demonstrated in the olfactory lobe of Drosophila. This important study uses sophisticated genetics and advanced live imaging to provide a compelling description of how neuronal dendrites explore the target field, eliminate excessive branches, and assort into the correct region during develop- ment. In the process, it develops valuable tools. The study brings us closer to a comprehensive understanding of how the birth order of a neuron translates to dendrite patterning within the Drosophila antennal lobe circuit. *For correspondence: ltongchao@outlook.com (TL); lluo@stanford.edu (LL) Present address: †Liangzhu Laboratory, MOE Frontier Science Center for Brain Science and Brain- machine Integration, State Key Laboratory of Brain- machine Intelligence, Zhejiang University, Hangzhou, China; ‡Department of Electrical and Computer Engineering, Princeton University, Princeton, United States Competing interest: The authors declare that no competing interests exist. Funding: See page 25 Received: 11 December 2022 Preprinted: 29 December 2022 Accepted: 27 March 2023 Published: 28 March 2023 Reviewing Editor: Sonia Sen, Tata Institute for Genetics and Society, India Copyright Wong et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 1 of 33 Research article eLife digest The brain’s ability to sense, act and remember relies on the intricate network of connections between neurons. Organization of these connections into neural maps is critical for processing sensory information. For instance, different odors are represented by specific neurons in a part of the brain known as the olfactory bulb, allowing animals to distinguish between smells. Projection neurons in the olfactory bulb have extensions known as dendrites that receive signals from sensory neurons. Scientists have extensively used the olfactory map in adult fruit flies to study brain wiring because of the specific connections between their sensory and projection neurons. This has led to the discovery of similar wiring strategies in mammals. But how the olfactory map is formed during development is not fully understood. To investigate, Wong et al. built genetic tools to label specific types of olfactory projection neurons during the pupal stage of fruit fly development. This showed that a group of projection neurons directed their dendrites in a clockwise rotation pattern depending on the order in which they were born: the first- born neuron sent dendrites towards the top right of the antennal lobe (the fruit fly equivalent of the olfactory bulb), while the last- born sent dendrites towards the top left. Wong et al. also carried out high- resolution time- lapse imaging of live brains grown in the labora- tory to determine how dendrites make wiring decisions. This revealed that projection neurons send dendrites in all directions, but preferentially stabilize those that extend in the direction which the neurons eventually target. Also, live imaging showed neurons could remove old dendrites (used in the larvae) and build new ones (to be used in the adult) simultaneously, allowing them to quickly create new circuits. These experiments demonstrate the value of imaging specific types of neurons to understand the mechanisms that assemble neural maps in the developing brain. Further work could use the genetic tools created by Wong et al. to study how wiring decisions are determined in this and other neural maps by specific genes, potentially yielding insights into neurological disorders associated with wiring defects. Introduction Organization of neuronal connectivity into spatial maps occurs widely in the nervous systems across species (Luo and Flanagan, 2007; Cang and Feldheim, 2013; Luo, 2021). For example, in the retino- topic map of the visual system, nearby neurons in the input field project axons to nearby neurons in the target field (Cang and Feldheim, 2013). Such a continuous organization preserves spatial relation- ships in the visual world. Contrary to retinotopy, the olfactory glomerular map consists of discrete units called glomeruli in which input neurons connect with the cognate output neurons based on neuronal type rather than soma position (Mombaerts et al., 1996; Gao et al., 2000; Vosshall et al., 2000). This discrete map represents a given odor by the combinatorial activation of specific glomeruli. Whereas continuous maps are readily built using gradients of guidance cues (Cang and Feldheim, 2013), how glomeruli are placed at specific locations in discrete maps is less clear (Murthy, 2011). Understanding the developmental origins of these neural maps is fundamental for deciphering the logic of their func- tional organization through which information is properly represented and processed. The adult Drosophila olfactory map in the antennal lobe (the equivalent of the vertebrate olfactory bulb) has proven to be a powerful model for studying mechanisms of wiring specificity, thanks to the type- specific connections between the presynaptic olfactory receptor neurons (ORNs) and the cognate postsynaptic projection neurons (PNs). Molecules and mechanisms first identified in this circuit have been found to play similar roles in the wiring of the mammalian brain (e.g. Hong et al., 2012; Berns et al., 2018; Pederick et al., 2021). Assembly of the fly olfactory map begins with dendritic growth and patterning of PNs derived primarily from the anterodorsal (adPNs) and lateral (lPNs) lineages and born with an invariant birth order within each lineage (Jefferis et al., 2001; Jefferis et al., 2004; Marin et al., 2005; Yu et al., 2010; Lin et al., 2012; Figure 1A and B). This patterning creates a prototypic olfactory map, prior to ORN axon innervation, indicative of the PN- autonomous ability to target dendrites into specific regions. However, earlier studies could only unambiguously follow the development of one single PN type – DL1 PNs (Jefferis et al., 2004). It remains unclear to date how the prototypic olfactory map is organized and what cellular mechanisms PN dendrites use to achieve Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 2 of 33 Developmental Biology | Neuroscience Research article Figure 1. Organization and development of the adult olfactory circuit in Drosophila. (A, B) Timeline (A) and schematic illustration (B) of Drosophila olfactory circuit development. Green, red, and blue circles denote the birth of embryonic- born anterodorsal projection neuron (adPN), larval- born adPN, and larval- born lPN, respectively. At the onset of metamorphosis, the larval- specific olfactory circuit degenerates; larval olfactory receptor neurons (ORNs) die while embryonic- born adPNs prune their larval- specific processes and re- extend new processes into the adult- specific olfactory circuit. In the adult- specific olfactory circuit, projection neuron (PN) dendrites extend first and form a prototypic map. This is followed by an extension of ORN axons and synaptic partner matching between cognate PN dendrites and ORN axons to form a mature map. Solid and open arrowheads in A indicate onset of innervation for PN dendrites and ORN axons, respectively. (C) Overview of this study investigating the logic of dendritic patterning (C1; see Figures 3 and 4) as well as cellular mechanisms of dendrite targeting specificity (C2; see Figures 6 and 7) and re- wiring (C3; see Figure 8) that contribute to the developmental origin of the adult Drosophila olfactory map. (D) Staining of fixed brains at indicated stages showing dendrite development of adPNs (VT033006+ run+ ; labeled in yellow) and lPNs (VT033006+ run–; labeled in cyan). As run- FLP is expressed before 0 h APF in adPN but not lPN neuroblasts, we can use it to label adPNs and lPNs with two distinct colors using an intersectional reporter (see Materials and methods for the genotype). Yellow arrowheads in (D1) mark larval- and adult- specific dendrites of adPNs in larval- and adult- specific antennal lobes, respectively. Cyan arrowheads in (D3) denote specific targeting of lPN dendrites at the opposite ends of the dorsomedial- ventrolateral axis. (D1): N=12; (D2): N=7; (D3): N=17; (D4): N=10; (D5): N=12. Common notations in this study: Unless otherwise indicated, all images in this and subsequent figures are partial z projections of confocal stacks of representative images. N indicates the number of antennal lobes imaged. Antennal lobe neuropils are revealed by N- Cadherin (Ncad; in blue) staining. Adult- specific (developing) antennal lobe is outlined with a white solid line. Larval- specific antennal lobe is outlined with an orange line (dashed line used to denote the degeneration stage) and is distinguished from the developing antennal lobe by the more intense nc82 staining as shown in Figure 1—figure supplement 1 (nc82 channel not shown here). Asterisks (*) indicate PN cell bodies, which are outside the antennal lobe neuropil (and sometimes appear on top because of the z- projections). Arrowheads mark PN dendrites. Arrows mark PN axons projecting towards higher olfactory centers (see Figure 1—figure supplement 2 for PN axons at their targets in the mushroom body and lateral horn). h APF: hours after puparium formation; h ALH: hours after larval hatching. DL: dorsolateral; DM: dorsomedial; VM: ventromedial; VL: ventrolateral. Scale bar = 10 µm. The online version of this article includes the following figure supplement(s) for figure 1: Figure supplement 1. Visualization of larval- and adult- specific antennal lobes by co- staining of Ncad and nc82. Figure supplement 2. Projection neuron (PN) axon development across pupal stages. Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 3 of 33 Developmental Biology | Neuroscience Research article targeting specificity (Figure 1C1- 2). The initial map formation is further complicated by circuit remod- eling during which embryonic- born PNs used in both the larval and adult circuits reorganize their neurites (Marin et  al., 2005). How embryonic- born PNs coordinate remodeling with re- integration into the adult circuit is not known (Figure 1C3). Here, we set out to explore the origin of the olfactory map by performing a systematic and compar- ative study of PN dendrite development at type- specific resolution in vivo, and two- photon– and adaptive optical lattice light- sheet microscope–based time- lapse imaging of PN dendrites in early pupal brain explants. As our overarching goal is to understand how the wiring specificity between ORNs and PNs arises, we focus on PNs that project to single glomeruli. Neurons from the lateral lineage that innervate multiple glomeruli or project to other regions of the adult brain (Lin et  al., 2012) are not studied here. Our study uncovers wiring logic that directs PN dendrites to create an organized olfactory map, dendritic branch dynamics that lead to directional selectivity, and a novel re- wiring mechanism that facilitates timely olfactory map formation. These wiring strategies used in the initial map organization lay the foundation of precise synaptic connectivity between PNs and ORNs in the final glomerular map. Results Overview of Drosophila olfactory circuit development at a lineage- specific resolution We first described the development of the Drosophila olfactory circuit using pupal brains double- labeled for adPNs and lPNs (Figure 1D; see the genetic design in Figure 2). At the onset of metamor- phosis (0 hr after puparium formation; 0 hr APF), the adult- specific antennal lobe (also referred to as ‘developing antennal lobe’) remained relatively small, located dorsolateral and posterior to the larval- specific antennal lobe (also referred to as ‘degenerating antennal lobe’) (Figure 1D1). As PN dendrites continued to grow and innervate the developing antennal lobe, its size increased considerably (Figure 1D1–3). By 12 hr APF, PNs already appeared to be sorting their dendrites into specific regions to form a prototypic map, as revealed by the heterogeneous patterning of lPN dendrites (arrowheads in Figure  1D3). From 21 hr to 50 hr APF, dendrites of adPNs and lPNs gradually segregated and eventually formed intercalated but non- overlapping glomeruli (Figure 1D4–5). The development of the adult- specific antennal lobe partially overlapped with the degeneration of the larval- specific antennal lobe, as indicated by fragmentation of the larval- specific dendrites of embryonic- born PNs at 3  hr APF (Figure 1D2). This gross characterization at the resolution of two PN lineages was consistent with earlier studies (Jefferis et al., 2004; Marin et al., 2005). However, the resolution was not sufficiently high to answer the questions we raised in the Introduction (Figure 1C). Expanded genetic toolkit for type-specific labeling of PNs during early pupal development To reveal how PN dendrites initiate olfactory map formation at the high spatiotemporal resolution, we needed genetic access to specific PN types during early pupal development. From our recently deciphered single- cell PN transcriptomes (Xie et  al., 2021), we searched for genetic markers that are expressed strongly and persistently in single or a few PN types across pupal development. This transcriptome- instructed search led to the identification of CR45223 (in place of this non- coding gene, we used the adjacent CG14322 that exhibits nearly identical expression pattern), lov, and tsh (Figure 2A and B; Figure 2—figure supplement 1). Next, using CRISPR/Cas9, we generated knock- in transgenic QF2 expression driver lines in which T2A- QF2 (or T2A- FLP for intersection) was inserted immediately before the stop codon of the endog- enous gene (Figure  2—figure supplement 2). The self- cleaving peptide T2A allows QF2 to be expressed in the same pattern as the endogenous gene (Diao and White, 2012). With these new QF2 lines together with existing GAL4 lines that label additional PN types (Xie et al., 2019), we now have an expanded toolkit accessing PNs ranging from early- to late- born PNs, from adPN to lPN lineages, and from PNs with neighboring glomerular projections to those with distant projections in the adult antennal lobe (Figure 2C and D). As QF2/QUAS and GAL4/UAS expression systems operate orthogo- nally to each other (Potter et al., 2010; Riabinina et al., 2015), we crossed our QF2 lines with existing GAL4 lines for simultaneous labeling of distinct PN types in the same brain (see inset in Figure 2C). Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 4 of 33 Developmental Biology | Neuroscience Research article Figure 2. Expanded genetic toolkit for dual- color, type- specific labeling of projection neurons (PNs). (A) tSNE plot of PN single- cell transcriptomes, color- coded according to CR45223 expression level in [log2(CPM +1)], where CPM stands for transcript counts per million reads. Zoom- in of boxes in the tSNE plot (left) is shown on the right, and color- coded according to PN types and developmental stages. (B) Dot plot showing the expression of acj6, vvl, CR45223, CG14322, lov, and tsh in 0 hr APF PNs arranged according to their birth order and lineage (green: embryonic- born anterodorsal projection neuron (adPNs); red: larval- born adPNs; blue: larval- born lPNs). Unit of expression is [log2(CPM +1)] as in A. Data from panels A are B are from Xie et al., 2021. (C) Birth orders of adPNs and lPNs summarized by Lin et al., 2012; Yu et al., 2010 and genetic tools used to access them. Left: Accessible PN types are colored. Circles beneath the PN types denote QF2/GAL4 drivers used to access them. Asterisks beneath the PN types denote access by MARCM. Gray arrowhead marks neuroblast (NB) rest. Right: Genetic tools. Inset shows the combinatorial use of QF2/FLP and GAL4 (linked by dashed lines) for comparative analyses of dendrite development of two groups of PNs in the same animal. (D) Schematic of glomerular projections of QF2/ GAL4- accessible PNs in the adult antennal lobe. Indicated glomeruli are color- coded based on the genetic tools used to access them. See the color code in C. (E, F) Schematic of intersectional logic gates for dual- color labeling of PNs. See Figure 2—figure supplement 2 for newly generated FLP- out reporters. The online version of this article includes the following figure supplement(s) for figure 2: Figure supplement 1. Expression of projection neuron (PN) marker genes across development. Figure supplement 2. Generation of T2A- QF2/FLP transgenic flies by CRISPR/Cas9. Figure supplement 3. Design of single- and dual- color FLP- out reporters. This combinatorial use of driver lines permitted comparative analyses of the development of distinct PN types with minimal biological and technical variations (Supplementary file 1). To limit driver expression only in PNs, we applied intersectional logic gates (AND and NOT gates) using our newly generated conditional reporters genetically encoding either mGreenLantern, Halo tags, and/or SNAP tags (Kohl et al., 2014; Sutcliffe et al., 2017; Campbell et al., 2020; Figure 2E and F; Figure 2—figure supplement 3). These reporters can be broadly used in other systems. Finally, Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 5 of 33 Developmental Biology | Neuroscience Research article we used MARCM (Lee and Luo, 1999) to label PNs that remain inaccessible due to a lack of drivers (Figure 2C; discussed in Figure 3). Early larval-born adPN dendrites initially share similar targeting regions Using the new genetic tools, we first re- visited the dendrite development of DL1 PNs—the first larval- born adPN type—using pupal brains double- labeled for DL1 PNs (labeled by 71B05- GAL4) and adPNs (Figure  3A). Consistent with our previous study (Jefferis et  al., 2004), DL1 PNs already showed robust dendritic growth at the wandering third instar larval stage (Figure 3—figure supplement 1A). At 0 hr APF, DL1 PN dendrites extended radially outwards from the main process, reaching nearly the entire developing antennal lobe and often overshooting it (white arrowheads in Figure 3A1), likely surveying the surroundings. By 6 hr APF, most of the dendrites already occupied the dorsolateral (DL) corner of the antennal lobe (Figure 3A2). As the antennal lobe continued to grow, this dorsolateral positioning of the DL1 PN dendrites remained largely unchanged (Figure  3A3–6). From 21  hr APF onwards, the dendrites underwent progressive refinement: they were restricted into a smaller area by 30 hr APF (Figure 3A4–5), and eventually formed a compact, posterior glomerulus by 50 hr APF (Figure 3A6 showing a single z section). To assess whether other PN types follow the same developmental trajectory, we next examined CG14322+ PNs, which include DL1 PNs and DA3 PNs—the first and second larval- born adPN types, respectively. In the same brain, we also labeled with a different fluorophore DC2 PNs—the third larval- born adPN type (Figure  3B). The dendritic pattern of DL1/DA3 PNs appeared indistinguish- able from that of DL1 PNs from 0 hr to 12 hr APF (compare the yellow channel of Figure 3B1–3 with Figure 3A1–3), suggesting that DL1 and DA3 PN sent dendrites to the same region in the antennal lobe. We began to see differences in 21 hr APF pupal brains in which DL1/DA3 PN dendrites not only occupied the dorsolateral region but also spread ventrally (white arrowhead in Figure 3B4; compare with Figure  3A4). The more ventrally targeted dendrites likely belong to DA3 PNs. This suggests that ~21 hr APF marks the beginning of dendritic segregation of DL1 and DA3 PNs. By 30 h APF, DL1 and DA3 dendrites were clearly separable (Figure 3B5), which respectively formed more posteriorly and anteriorly targeted glomeruli at 50 hr APF (Figure 3B6; see single z sections in Figure 3—figure supplement 1C). Next, we focused on the third- born—DC2 PNs labeled by 91G04- GAL4 (Figure 3B). This GAL4 labeled additional embryonic- born adPNs from 0 hr to 6 hr APF, but the expression in these PNs diminished afterward. As embryonic- born adPNs do not have any dendrites in the developing antennal lobe at 0 hr APF (discussed in Figure 8), dendrites found in the antennal lobe should belong to the larval- born DC2 PNs. Like DL1/DA3 PNs, DC2 PNs initiated radial dendritic extension across the antennal lobe at 0 hr APF (Figure 3B1; Figure 3—figure supplement 1B). Notably, DL1/DA3 and DC2 PN dendrites exhibited substantial overlap from 0 hr to 12 hr APF and shared a similar targeting region at the dorsolateral corner from 6 hr to 12 hr APF (Figure 3B1–3). It was not until 21 hr APF that DL1, DA3, and DC2 dendrites began to segregate from each other along both medial- lateral and anterior- posterior axes (Figure 3B4–5). By 50 hr APF, the DC2 glomerulus was separated from DL1/DA3 glomeruli by intermediate glomeruli (Figure 3B6). In summary, dendrites of consecutively larval- born DL1, DA3, and DC2 adPNs (here collectively named ‘early larval- born adPNs’; see its definition in next section) develop in a similar fashion and share a similar targeting region at early pupal stages (0–12  hr APF). This is then followed by their segregation into distinct regions close to their adult glomerular positions during mid- pupal stages (21–50 hr APF). Larval-born adPNs with distant birth order send dendrites to distinct regions The analysis of early larval- born adPNs (Figure  3A and B) led us to hypothesize that larval- born adPNs might use their birth order to coordinate dendrite targeting during early pupal stages. If this were true, we would expect dendrites of larval- born adPNs with distant birth order to occupy distinct regions. To test this hypothesis, we compared dendrite- targeting regions of early larval- born adPNs with those of later- born adPNs. We first examined DC3/VA1d adPNs (referred to as ‘mid- early larval- born adPNs’) using Mz19- GAL4 (Figure 3C). This GAL4 is expressed in three PN types from 24 hr APF to adulthood: DC3 adPNs, VA1d Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 6 of 33 Developmental Biology | Neuroscience Research article Figure 3. Birth order–dependent spatial patterning of anterodorsal projection neuron (adPN) dendrites in the developing antennal lobe. (A) Confocal images of fixed brains at indicated stages showing dendrite development of adPNs (acj6+; labeled in green) and DL1 adPNs (71B05+; labeled in yellow). Right column of A1 shows a zoom- in of the dashed box. The labeling of acj6+ adPNs outlines the developing antennal lobe and is used in dual- color Figure 3 continued on next page Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 7 of 33 Developmental Biology | Neuroscience Research article Figure 3 continued AO- LLSM imaging later (see Figure 7A–C). White arrowheads in (A1) mark dendrites overshooting the antennal lobe. (A1): N=14; (A2): N=12; (A3): N=14; (A4): N=6; (A5): N=4; (A6): N=4. (B) Confocal images of fixed brains at indicated stages showing dendrite development of DL1/DA3 adPNs (CG14322+; labeled in yellow) and DC2 adPNs (91G04+; labeled in magenta). As 91G04- GAL4 labels some embryonic- born projection neurons (PNs) from 0 to 6 hr APF, their neurites are found in the larval- specific antennal lobe (B1, 2). Right column of (B1) shows a zoom- in of the dashed box. White arrowhead in (B4) denotes the more ventrally targeted DL1/DA3 dendrites. (B1): N=6; (B2): N=5; (B3): N=12; (B4): N=4; (B5): N=7; (B6): N=2. (C) Confocal images of fixed brains at indicated stages showing dendrite development of DC3/VA1d adPNs (Mz19+ acj6+; labeled in red) and DA1 lPNs (Mz19+ acj6–; labeled in cyan). (C1): N=14; (C2): N=6; (C3): N=4; (C4): N=10; (C5): N=10; (C6): N=6; (C7): N=4. (D) Confocal images of single- cell MARCM clones (in yellow) of DL1 PNs (D1–3), mid- late larval- born adPNs (D4–6), and late larval- born adPNs (D7–9) in 12 hr APF pupal brains, generated by heat shocks (hs) at indicated times. Three biological samples are shown for each of the indicated adPN cohorts. D1–3: N=5; D4–6: N=4; D7–9: N=8. (E) Summary of wiring logic of larval- born adPN dendrites to form an olfactory map in the 12 hr APF developing antennal lobe. See Figure 1 legend for common notations. The online version of this article includes the following video, source data, and figure supplement(s) for figure 3: Figure supplement 1. Dendrite development of early larval- born projection neurons (PNs). Figure supplement 2. MARCM- labeled single- cell projection neurons (PNs) of indicated lineages in adult brains. Figure supplement 3. Dendrite development of DL1, middle larval- born, and late larval- born projection neurons (PNs) at early stages. Figure supplement 3—source data 1. Source data for Figure 3—figure supplement 3F and G. Figure 3—video 1. 3D rendering of z stacks of indicated projection neurons (PNs) in 12 hr APF antennal lobe. https://elifesciences.org/articles/85521/figures#fig3video1 adPNs, and DA1 lPNs (Jefferis et al., 2004). To distinguish adPNs from lPNs, we previously adopted an FLP- out strategy labeling Mz19+ PNs with either GFP or RFP based on their lineages and studied dendrite segregation and refinement during mid- pupal stages (Li et al., 2021; Figure 3C4–7). However, the weak GAL4 expression before 24 hr APF prevented us from visualizing any dendrites at earlier stages. To overcome this, we incorporated Halo and SNAP chemical labeling (Kohl et al., 2014) in place of the immunofluorescence approach. This modification substantially extended the detection to developmental stages as early as 12 hr APF (Figure 3C1). We found that, from 12 hr to 21 hr APF, DC3/VA1d PN dendrites targeted the ventrolateral (VL) corner of the antennal lobe (Figure 3C1–4). Thus, early (DL1/DA3/DC2) and mid- early (DC3/VA1d) larval- born adPN dendrites occupy distinct regions at 12 hr APF. As we did not have reliable drivers to access other later- born PNs at early pupal stages, we turned to MARCM (Lee and Luo, 1999) to generate heat shock- induced single- cell clones of PNs born at different times (Figure 3—figure supplement 2). We used GH146- GAL4(IV), a PN driver that labels the majority of PN types, including later- born adPNs (Figure  3—figure supplement 2D–E), with a tight temporal control of heat shock and analyzed heat shock- induced animals that were among the first to form puparium to minimize the effects of unsynchronized development among individual animals (see Materials and methods for details). These optimizations permitted a systematic clonal analysis at higher PN type- specific resolution that correlates with birth time. Based on birth timing that corresponds to the heat shock time we applied to induce single- cell MARCM clones, we assigned larval- born adPNs to approximate temporal cohorts: (1) heat shock at 0–24 hr ALH (after larval hatching): first- born (DL1), (2) heat shock at 42–48 hr ALH: early- born (DL1, DA3, DC2, and D), (3) heat shock at 66–72 hr ALH: mid- late born (VM7v, VM7d, VM2, DM6, and VA1v), and (4) heat shock at 96–100 hr ALH: late- born (DM6, VA1v, DL2v, DL2d) (Figure 3E1). We assigned DC3/VA1d PNs labeled by Mz19- GAL4 to the mid- early cohort because they are born between the early and mid- late adPNs. We note that DM6 and VA1v PNs were assigned to both cohorts of mid- late and late- born adPNs, reflecting the nature of short birth timing differences and overlaps between adjacent cohorts. Using this strategy, we could also label lPNs born at different times and assigned them into approximate temporal cohorts (Figure 3—figure supplement 2F). Clonal analysis revealed that, at 12 hr APF, the first- born DL1 adPNs sent dendrites to the dorso- lateral corner of the antennal lobe as expected (Figure  3D1–3). By contrast, dendrites of mid- late larval- born adPNs occupied a large region on the medial/dorsomedial (M/DM) side (Figure  3D4–6). Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 8 of 33 Developmental Biology | Neuroscience Research article The dendritic arborization patterns of these PNs varied widely, most likely because they belonged to different PN types. Intriguingly, late larval- born adPN dendrites targeted the peripheral, dorsomedial (abbreviated as pDM) corner where the staining of the pan- neuropil marker N- Cadherin was relatively weak (Figure 3D7–9). The weak staining implies that this area is less populated by PN dendrites (the major constituent of the antennal lobe neuropil at this stage), possibly because (1) this area is not innervated by many PNs and/or (2) the dendrites of late- born PNs innervate later and remain less elaborate than earlier- born PNs (we will explore this later). Together, our data (Figure 3A–D) suggest that larval- born adPNs with adjacent birth order send dendrites to similar regions of the developing antennal lobe whereas those with distant birth order send dendrites to distinct regions (Figure 3E2,3). Notably, the birth order of the examined PNs does not specify dendrite targeting randomly (Figure 3E4). Rather, the stereotyped dendritic pattern in the prototypic map correlates with the birth order in an organized manner (rotating clockwise in the right hemisphere when viewed from the front; anti- clockwise in the left: early↔DL; mid- early↔VL; mid- late↔M/DM; late↔pDM). One can, therefore, infer at least the approximate birth order of a larval- born adPN based on its initial dendrite targeting, and vice versa. As the antennal lobe is a 3D structure, we also visualized PN dendrite targeting in the 12 hr APF map with 3D rendering generated from z stacks with rotation along the y- axis (Figure 3—video 1). We found that, along the short anterior- posterior axis (spanning about 20  µm), PN dendrites were located primarily on the periphery of the antennal lobe, whereas the center housed the axon bundle projecting out of the antennal lobe. Some dendrites could reach almost the entire depth, suggesting active exploration of the surroundings in many directions. While 3D projections provide rich details in depth and different viewing angles, we did not find an apparent relationship between birth order and dendrite targeting along the anterior- posterior axis, at least for the examined PN types at 12 hr APF. Thus, the approximate 2D projection (Figure  3E2–4) conveys the logic of dendrite patterning effectively. Dendrite targeting timing of larval-born adPN depends on birth order Having provided evidence for birth order–dependent spatial patterning of larval- born adPN dendrites, we next asked whether the timing of dendritic extension and targeting is also influenced by birth order. We noticed that the extent of dendritic innervation of 0 hr APF first- born DL1 adPNs resembled that of 6 hr APF mid- late born adPNs (compare Figure 3—figure supplement 3A1–4 with Figure 3— figure supplement 3B5–8). Such a resemblance was also seen between 0 hr APF mid- late and 6 hr APF late- born adPNs (compare Figure  3—figure supplement 3B1–4 with Figure  3—figure supplement 3C). Quantitative analyses of the exploring volume of dendrites and the number of terminal branches showed that, at 0 hr APF, DL1 PN dendrites were more elaborate than mid- late born PN dendrites (Figure 3—figure supplement 3F). By 6 hr APF, the mid- late born appeared to catch up, showing an extent of innervation comparable to DL1 PNs. We next examined when the dendrites reach their targeting regions. We found that whereas early larval- born adPNs (DL1, DA3, DC2) concentrated their dendrites to the dorsolateral corner by 6 hr APF (Figure 3B2; Figure 3—figure supplement 3A5–8), later- born PNs concentrated their dendrites to the medial/dorsomedial or peripheral dorsomedial side at 12 hr APF (Figure 3D4- 9; Figure 3—figure supplement 3B5- 8, C). Thus, our results suggest larval- born adPN dendrites innervate and pattern the antennal lobe using a ‘first born, first developed’ strategy. Contribution of lineage to early PN dendritic patterning Both lineage and birth order of PNs contributes to the eventual glomerular choice of their dendrites (Jefferis et al., 2001). What is the involvement of lineage in the prototypic map formation? Do lPN dendrites pattern the developing antennal lobe following similar rules as adPNs? To characterize lPN dendrite development at type–specific resolution, we used tsh- GAL4 to genetically access DA1/ DL3 lPNs, and MARCM clones of lPNs as a complementary approach (Figure 4). We focused on the dendritic patterns of tsh+ DA1/DL3 lPNs from 0 hr to 12 hr APF as tsh- GAL4 labeled additional PNs from 21 hr APF onwards (Figure 4A4–6; Figure 4—figure supplement 1B4–6; Figure 4—figure supple- ment 2; Figure 2—figure supplement 1). Examination of pupal brains double- labeled with DA1/DL3 lPNs (referred to as ‘middle larval- born lPNs’) and DL1/DA3 adPNs revealed that, like the early larval- born adPNs, dendritic growth of DA1/ Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 9 of 33 Developmental Biology | Neuroscience Research article Figure 4. Birth order–dependent spatial patterning of lPN dendrites in the developing antennal lobe. (A) Confocal images of fixed brains at indicated stages showing dendrite development of DL1/DA3 adPNs (CG14322+; labeled in yellow) and DA1/DL3 lPNs (tsh+; labeled in cyan). Right column of A1 shows a zoom- in of the dashed box. (A1): N=8; (A2): N=4; (A3): N=6; (A4): N=10; (A5): N=4; (A6): N=5. (B) MARCM clones (in cyan) of early (B1–3) and late (B4–6) larval- born lPNs in 12 hr APF pupal brains, generated by heat shocks (hs) at indicated times. In (B3), (B5), and (B6), single- cell clones of anterodorsal projection neuron (adPN) (yellow arrowheads) and lPN (cyan arrowheads) lineages were simultaneously labeled. Three biological samples are shown for each of the indicated lPN cohorts. B1–3: N=4; B4–6: N=6. (C) Summary of wiring logic of larval- born lPN dendrites to form an olfactory map in the 12 hr APF developing antennal lobe. (D) Summary of determination of dendrite targeting of larval- born PNs by lineage and birth order. See Figure 1 legend for common notations. The online version of this article includes the following video and figure supplement(s) for figure 4: Figure supplement 1. Dendrite development of DL1/DA3 and DA1/DL3 projection neurons (PNs). Figure supplement 2. Expression patterns of tsh in the developing antennal lobe during mid- pupal stages. Figure 4—video 1. 3D rendering of z stacks of indicated projection neurons (PNs) in 12 hr APF antennal lobe. https://elifesciences.org/articles/85521/figures#fig4video1 DL3 lPNs was evident by the wandering third instar larval stage (Figure 4—figure supplement 1A). At this stage, most DA1/DL3 lPN dendrites innervated the antennal lobe and intermingled with those of DL1/DA3 adPNs. From 0 hr to 12 hr APF, despite a high degree of overlap among those dendrites that explored the surroundings, DA1/DL3 lPN dendrites primarily targeted an area ventrolateral to those of DL1/DA3 adPNs (Figure 4A1–3; see 3D rendering in Figure 4—video 1). Such a spatial distinction was also observed between middle larval- born adPNs and lPNs in 0  hr and 6  hr APF pupal brains where occasionally single- cell clones from both lineages were simultaneously generated by MARCM Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 10 of 33 Developmental Biology | Neuroscience Research article (Figure 3—figure supplement 3D1–4, 7–10). Thus, at least some adPNs and lPNs sort their dendrites into distinct regions very early on regardless of birth timing. Next, we used MARCM to ask if lPNs born earlier and later than DA1/DL3 lPNs would send dendrites to regions different from that of DA1/DL3 lPNs. We found that dendrites of early- born lPNs primarily occupied the medial/dorsomedial side of the antennal lobe (Figure  4B1–3); we note that adPNs born at the same time sent dendrites to the dorsolateral side (see yellow arrowhead in Figure 4B3). Also, in contrast to the ventrolateral targeting of middle- born lPN dendrites, late- born lPNs sent dendrites to the dorsomedial corner (Figures 4B4–6). Like larval- born adPNs, late- born lPNs innervated the antennal lobe later than earlier- born lPNs (Figure 3—figure supplement 3D7–12–E, G). These data suggest that, at early pupal stages, lPN dendrites pattern the developing antennal lobe following similar rules as larval- born adPNs: adjacent birth order → similar dendrite targeting; distant birth order → distinct dendrite targeting; ‘first born, first developed.’ However, unlike the correlation of birth order and target positions in a rotational manner for adPNs (Figure 3E), the lPN dendritic map formation appears binary: early↔M/DM; middle↔VL; late↔DM (Figure  4C). Our type- specific characterization corroborated with the gross examination of the lPN dendrites as previously reported (Jefferis et al., 2004): at 12 hr APF, lPN dendrites mostly occupied the opposite corners along the dorsomedial- ventrolateral axis, leaving the middle of the axis largely devoid of lPN dendrites (arrow- heads in Figure 1D3). In summary, we propose that lineage and birth order of larval- born PNs contribute to their dendrite targeting in a combinatorial fashion (Figure 4D). The wiring logic of PN dendrites in the developing antennal lobe can, therefore, be represented by [lineage, birth order]=dendrite targeting; one can deduce the unknown if the other two are known. An explant system for time-lapse imaging of PN development at early pupal stages So far, we have identified wiring logic governing the initial dendritic map formation (Figures 3 and 4) by examining specifically labeled neuron types in the fixed brain at different developmental stages. To examine dendrite targeting at the higher spatiotemporal resolution, we established an early- pupal brain explant culture system based on previous protocols (Özel et  al., 2015; Rabinovich et  al., 2015; Li and Luo, 2021; Li et  al., 2021), and performed single- or dual- color time- lapse imaging with two- photon microscopy as well as adaptive optical lattice light- sheet microscopy (AO- LLSM) (Figure  5A–C). The following lines of evidence support that our explant system recapitulates key features of in vivo olfactory circuit development. First, during normal development, the morphology of the brain lobes changes from spherical at 0 hr APF to more elongated rectangular shapes at 15 hr APF (Rabinovich et al., 2015). After 22 hr ex vivo culture, the spherical hemispheres of brains dissected at 3 hr APF became more elongated, mimicking  ~15  hr APF in vivo brains characterized by the separation of the optic lobes from the central brain (Figure 5D). Second, dual- color, two- photon imaging of PNs every 20 min for 22 hr revealed that lPNs in 3 hr APF brains initially produced dynamic but transient dendritic protrusions in many directions, followed by extensive innervation into the antennal lobe (arrowheads in Figure 5E1–3; Figure 5—video 1). In higher brain centers, lPN axons clearly showed direction- specific outgrowth of collateral branches into the mushroom body calyx as well as forward extension into the lateral horn (arrows in Figure 5E3), thus resembling in vivo development (Figure 1—figure supplement 2). Third, larval- specific dendrites observed in 0 hr APF brains cultured for 12 hr ex vivo (orange arrow- head in Figure 5F4) were no longer seen in those cultured for 24 hr ex vivo (Figure 5F5), indicative of successful pruning and clearance of larval- specific dendrites. Also, the size of the developing antennal lobe in the brains cultured for 24  hr ex vivo increased considerably (Figure  5F5). These imply that olfactory circuit remodeling (degeneration of larval- specific processes and growth of adult- specific processes) proceeds normally, albeit at a slower rate (compare with Figure 5F1–3). Fourth, dendrites from genetically identified DL1 and DA1/DL3 PNs targeted to their stereotyped locations in the antennal lobe in 0 hr APF brains cultured for 24 hr ex vivo (Figure 5G), mimicking in vivo development (Figure 4A). Finally, the segregation of dendrites of PNs targeting to neighboring proto- glomeruli could be recapitulated in brains dissected at 24 hr APF and cultured for 8 hr (Figure 5—figure supplement 1; Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 11 of 33 Developmental Biology | Neuroscience Research article Figure 5. Establishment of an explant system for time- lapse imaging of olfactory map formation. (A) Schematic of the anatomical organization of the olfactory circuit in early pupal brain (0–3 hr APF). Green, red, and blue denote embryonic- born adPN, larval- born anterodorsal projection neuron (adPN), and larval- born lPN, respectively. MB: mushroom body; LH: lateral horn. (B) Schematic of explant culture system for early pupal brains. Wells created in the Sylgard plate from which brains were imbedded are shown in blue. (C) Schematic of explant culture and imaging system for early pupal brains. (D)  Top: Schematic of morphological changes of brain lobes from 0 hr to ~15 hr APF during normal development. Bottom: Morphologies of a brain explant dissected at 3 hr APF and cultured for 0 hr ex vivo and cultured for 22 hr ex vivo. (E) Two- photon time- lapse imaging of adPNs (VT033006+ run+ ; labeled in magenta) and lPNs (VT033006+ run–; labeled in green) in pupal brain dissected at 3 hr APF and cultured for 0–22 hr ex vivo. Arrowheads mark dynamic but transient dendritic protrusions of lPNs in E1, 2, and extensive dendritic innervation of lPNs in (E3). Arrows in (E3) mark axonal innervation of lPNs in the mushroom body calyx and lateral horn. N=3. (F) Confocal images of antennal lobes labeled by VT033006+ projection neurons (PNs) (in green) at 0 hr (F1), 6 hr (F2), and 12 hr (F3) APF in vivo. Confocal images of antennal lobes labeled by VT033006+ PNs in pupal brains were dissected at 0 hr APF and cultured for 12 hr (F4) and 24 hr (F5) ex vivo. (F1): N=6; (F2): N=5; (F3): N=6; (F4): N=8; (F5): N=8. (G) Dendrite targeting regions of DL1 PNs (71B05+; in yellow; G1) and DA1/DL3 PNs (tsh+; in cyan; G2) in the antennal lobes in pupal brains dissected at 0 hr APF and cultured for 24 hr ex vivo. Antennal lobes are revealed by N- Cadherin (Ncad; in blue) staining. (G1): N=5; (G2): N=6. See Figure 1 legend for common notations. The online version of this article includes the following video, source data, and figure supplement(s) for figure 5: Figure supplement 1. Dendritic segregation of DC3/VA1d adPNs and DA1 lPNs targeting neighboring proto- glomeruli. Figure supplement 1—source data 1. Source data for Figure 5—figure supplement 1C and D. Figure 5—video 1. Two- photon time- lapse imaging of projection neuron (PN) development. https://elifesciences.org/articles/85521/figures#fig5video1 Figure 5—video 2. Two- photon time- lapse imaging of projection neuron (PN) dendritic segregation. https://elifesciences.org/articles/85521/figures#fig5video2 Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 12 of 33 Developmental Biology | Neuroscience Research article Figure 5—video 2). Specifically, despite constant dynamic interactions among dendrites that explore the surroundings (arrowheads in Figure 5—figure supplement 1A2–4), DC3/VA1d and DA1 PNs exhib- ited a 1–2 µm increase in the distance between centers of the two dendritic masses and a substantial decrease in the overlap of their core targeting regions (Figure 5—figure supplement 1B–D). Taken together, these data support that the explant culture and imaging system established here reliably captures key neurodevelopmental events starting from early pupal stages. Single-cell, two-photon imaging reveals active dendrite targeting Our observation in fixed brains revealed that dendrites of DL1 adPNs transition from a uniform exten- sion in the antennal lobe at 0 hr APF to concentration at the dorsolateral corner of the antennal lobe at 6 hr APF (Figure 3A). To identify mechanisms of dendrite targeting specificity that could be missed in static developmental snapshots, we performed two- photon time- lapse imaging of single- cell MARCM clones of DL1 PNs in 3 hr APF brains (Figure 6; Figure 6—figure supplement 1; Figure 6—video 1). Although we did not have a counterstain outlining the antennal lobe, we could use the background signals to discern the orientation of DL1 PNs in the brain (Figure  6—figure supplement 1A). The final targeting regions relative to the antennal lobe revealed by post hoc fixation and immunostaining confirmed proper dendrite targeting (yellow arrowhead in Figure 6A10; Figure 6—figure supplement 1B–C). Using DL1 PN in Figure  6A (pseudo- colored in yellow; Figure  6—video 1) as an example, we observed that the PN initially extended dendrites in every direction (Figure  6A1–3), like what we observed in fixed tissues (Figure 3A1). The first sign of active targeting emerged at 2 hr 20 min ex vivo when DL1 PN began to generate long, albeit transient, dendritic protrusions in the dorsolateral direc- tion; these selective protrusions were more prominent at 3 hr ex vivo (arrowheads in Figure 6A4–6). The dorsolateral targeting continued to intensify, leading to the formation of a highly focal dendritic mass seen at 13 hr ex vivo (arrowhead in Figure 6A8). As the dendrites reached the dorsolateral corner and explored locally, the change in shape appeared less pronounced (Figure 6A9). To quantitatively characterize the active targeting process, we categorized the bulk dendritic masses emanating from the main process according to their targeting directions: DL, DM, VM, and VL (Figure 6B). During the initial phase, the percentage of dendritic volume in each direction varied from 10% to 40% (Figure 6C and D), indicative of active exploration with little targeting specificity. Despite these variations, the total amount of dendritic mass seen in the VM direction over the entire imaging time (area under the graph of Figure  6C) was the smallest across all samples examined (Figure 6E). The initial phase of exploration in every direction was followed by a ~4 hr transitional phase during which DL1 PNs predominantly extended dendrites in 2 of the 4 directions (Figure 6C; Figure 6—figure supplement 1D–E). One of the 2 directions was always DL whereas the other was either DM or VL but never VM. In the final phase, DL1 PN dendrites always preferred DL out of the two available directions. Lastly, we analyzed the bulk dendritic movements. We defined bulk extension and retraction events when dendrites respectively extended and retracted more than 2 μm between two consecutive time frames. The analyses showed a striking shift from frequent extension and retraction towards stabilization, reflecting the pre- and post- targeting dynamics, respectively (Figure 6F and G). Hence, long- term two- photon imaging of single- cell DL1 PNs revealed that dendrite targeting specificity increases over time via active targeting in a specific direction and stepwise elimination of unfavorable trajectory choices (see summary in Figure 7F1–3). AO-LLSM imaging suggests a cellular mechanism underlying dendrite targeting specificity To capture fast dynamics of single dendritic branches, we performed dual- color adaptive optical lattice sheet microscopy (AO- LLSM) imaging (Chen et al., 2014; Wang et al., 2014; Liu et al., 2018) of PNs every 30 s for 15 min, following a protocol we recently established (Li et  al., 2021; Li and Luo, 2021). We selected 3 hr, 6 hr, and 12 hr APF pupal brains double- labeled with DL1 PNs and bulk adPNs (Figure 7A–C; Figure 7—videos 1–3). The labeling of adPNs with GFP outlined PN cell bodies and the developing antennal lobe but not the degenerating one, presumably because the GFP in larval- specific dendrites was quickly quenched upon glial phagocytosis (Marin et al., 2005). In the 15 min imaging window, we observed four types of terminal branches regardless of neuronal types or developmental stages: (1) stable branch that existed throughout the entire imaging time, Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 13 of 33 Developmental Biology | Neuroscience Research article Figure 6. Two- photon time- lapse imaging reveals active dendrite targeting. (A) Two- photon time- lapse imaging of MARCM- labeled DL1 projection neuron (PN) (pseudo- colored in yellow) in a brain dissected at 3 hr APF and cultured for 21 hr ex vivo (A1–9). Arrowheads in A4–6 denote protrusions of dendritic branches towards the dorsolateral direction. After 21 hr culture, the explant was fixed and immuno- stained for N- Cadherin (Ncad; in blue) to outline the developing antennal lobe (A10). Yellow and cyan arrowheads indicate DL1 PN dendrites and processes of other GH146+ cells, respectively. (B) Neurite tracing of DL1 PN at the beginning of live imaging (3 hr APF + 0 hr ex vivo). Dendrites are categorized based on the directions to which they extend and color- coded accordingly. (C) Left: Quantification of the percentage of dendritic volume in indicated direction during the time- lapse imaging period reveals a transitional phase during which dendrites were found in only two out of the four directions. Right: Schematic of the initial, transitional, and final phases during the course of targeting. ‘½’ denotes the reduction of available trajectory directions by half. Timestamp 00:00 refers to HH:mm; H, hour; m, minute. See Figure 6—source data 1. (D) Quantification of the percentage of DL1 PN dendritic volume in an indicated direction in 3 hr APF cultured brains at the beginning (0 hr ex vivo) and at/near the end of imaging (18 hr ex vivo). DL1 PN sample size = 3. t- test; *p<0.05. Timestamp 00:00 refers to HH:mm; H, hour; m, minute. (E) Quantification of the percentage of the sum of DL1 PN dendritic volume in indicated directions throughout the entire imaging time. DL1 PN sample size = 3. (F) Bulk dendrite dynamics of DL1 PN in Figure 6A. Each row represents bulk dendritic dynamics in the indicated direction (color- coded as in Figure 6B) across the 21 hr imaging period. Each block represents a 20 min window. Bulk extension (in green) and retraction (in magenta) events are defined as dendrites extending and retracting more than 2 μm between two consecutive time windows. The first and last six consecutive windows refer to the initial and final phases of imaging. (G) Quantification of the number of bulk extension and retraction events in the dorsolateral direction during the initial and final phases of imaging. DL1 PN sample size = 3. t- test; *p<0.05. The online version of this article includes the following video, source data, and figure supplement(s) for figure 6: Source data 1. Source data for Figure 6C–G and Figure 6—figure supplement 1D and E. Figure supplement 1. Two- photon time- lapse imaging of DL1 projection neuron (PNs). Figure 6—video 1. Two- photon time- lapse imaging of DL1 projection neuron (PN) dendrites. https://elifesciences.org/articles/85521/figures#fig6video1 Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 14 of 33 Developmental Biology | Neuroscience Research article (2) transient branch that was produced and eliminated within the imaging window, (3) emerging branch that was produced after imaging began, and (4) retracting branch that was eliminated within the imaging period (Figure  7—figure supplement 1A). To examine if terminal branch dynamics exhibit any directional preference, we assigned the branches according to their targeting directions (Figure  7D). Extension and retraction events were defined when the speed exceeded 0.5 μm/min. Terminal branches were selected for analyses as branches closer to the main process were too dense to resolve. Figure 7D1- 3 showed the dynamics of ~15 randomly selected terminal branches in each direction from the representative 3 hr, 6 hr, and 12 hr APF DL1 PNs (Figure 7A–C). Quantitative analyses revealed that at 3  hr APF, DL1 PNs constantly produced, eliminated, extended, and retracted dendritic branches (Figure 7A, Figure 7D1, Figure 7—video 1). Even stable branches were not immobile. Rather, they spent comparable amounts of time extending and retracting at ~1.5 μm/min (Figure 7—figure supplement 1A1, 1B). Transient, emerging, and retracting branches had similar, but more variable speeds, ranging from 1 to 2.5 μm/min. Although there was no correla- tion between targeting direction and frequency/speed of extension/retraction, the number of stable branches in the VM direction was significantly lower than in other directions across all 3 hr DL1 PN samples examined (Figure 7E1). This suggests that even though dendritic branches were developed in every direction at the early stages, those branches in the VM direction were short- lived and might be eliminated by retraction. The direction- dependent stability/lifespan of dendritic branches on the timescale of seconds uncovered from AO- LLSM imaging explains why bulk dendrites in unfavorable trajectories failed to persist in long- term two- photon imaging. From 6 hr to 12 hr APF, DL1 PNs no longer manifested direction- specific branch de/stabiliza- tion (Figure 7B–C, Figure 7D2–3, Figure 7—videos 2–3). At the same developmental stage, stable branches in one direction appeared indistinguishable from those in other directions in terms of abun- dance, frequency, and speed (Figure 7D2–3, Figure 7—figure supplement 1C–D). This suggests that the entire dendritic mass tends to stay in equilibrium upon arrival at target regions. At 12 hr APF, the abundance of stable branches of DL1 PNs was the highest (Figure 7D–E1). Also, the stable branches of 12 hr APF DL1 PNs moved at a significantly lower speed (~1 μm/min) (Figure 7E2) and spent more time being stationary than those at 3 hr and 6 hr (Figure 7—figure supplement 1B–D). The reduced branch dynamics at 12 hr APF is consistent with observations from two- photon imaging showing fewer bulk extension/retraction events in the final phase of targeting (Figure 6F–G). Despite the slowdown, dendritic arborization was evident in terminal branches of 12  hr APF DL1 PNs (Figure  7—figure supplement 1E), suggesting that PN dendrites are transitioning from simple to complex branch architectures. Although it remains unclear if there is a causal relationship between reduced branch dynamics and increased structural complexity, we propose that both contribute to the sustentation of dendrite targeting specificity. In summary, AO- LLSM imaging reveals that PNs selectively stabilize branches in the direction towards the target and destabilize those in the opposite direction, providing a cellular basis of dendrite targeting specificity. Upon arrival at the target, the specificity is sustained through branch stabilization in a direction- independent manner (summarized in Figure 7F4–7). Embryonic-born PNs timely integrate into an adult olfactory circuit by simultaneous dendritic pruning and re-extension In earlier sections, we uncovered wiring logic of larval- born PN dendritic patterning and cellular mech- anisms of dendrite targeting specificity used to initiate olfactory map formation (Figures 3–7). In this final section, we focused on embryonic- born PNs, which participate in both larval and adult olfactory circuits by reorganizing their processes (Marin et al., 2005). Our previous study demonstrates that embryonic- born PNs prune their larval- specific dendrites during early metamorphosis (Marin et al., 2005; Figure  1D1–3). Here, we examined when and how embryonic- born PNs re- extend dendrites used in the adult olfactory circuit. It is known that γ neurons of Drosophila mushroom body (γ Kenyon cells) and sensory Class IV dendritic arborization (C4da) neurons prune their processes between 4 hr and 18 hr APF and show no signs of re- extension at 18 hr APF (Lee et al., 2000; Watts et al., 2003; Lee et al., 2009). Do embryonic- born adPNs follow a similar timeframe? We first examined developing brains double- labeled for embryonic- born DA4l/VA6/VA2 adPNs (collectively referred to as ‘lov+ PNs’) and early larval- born DC2 adPNs (Figure 8A; Figure 8—figure supplement 1). We found that, by 12 hr APF, Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 15 of 33 Developmental Biology | Neuroscience Research article Figure 7. AO- LLSM time- lapse imaging reveals cellular mechanisms of dendrite targeting specificity. (A–C) AO- LLSM imaging of DL1 projection neurons (PNs) (71B05+; labeled in yellow) and anterodorsal projection neurons (adPNs) (acj6+; labeled in blue) in cultured brains dissected at 3 hr (A), 6 hr (B), and 12 hr (C) APF. Zoom- in, single z- section images of (A1), (B1), and (C1) (outlined in dashed boxes) are shown in A2, B2 and C2, respectively. (D) Single dendritic branch dynamics of 3 hr (D1), 6 hr (D2), and 12 hr (D3) DL1 PNs shown in A–C. Terminal branches are analyzed and categorized based on the directions in which they extend. Their speeds are color- coded using purple- gray- green gradients (negative speeds, retraction; positive speeds, extension). Individual branches are also assigned into four categories: stable, transient, emerging, and retracting (color- coded on the right; see Figure 7— figure supplement 1A). Each block represents a 30s window. Each row represents individual branch dynamics across the 15 min imaging period. (E) Quantification of the abundance (in percentage) of DL1 PN stable branches in indicated direction at 3 hr, 6 hr, and 12 hr (E1). Average speed of DL1 PN stable branches in indicated direction at 3 hr, 6 hr, and 12 hr (E2). DL1 PN sample size: 3 hr=4; 6 hr=3; 12 hr=3. Error bars, SEM; t-test; One- way ANOVA; *p<0.05; n.s., p≥0.05. SEM, standard error of the mean; n.s., not significant. See Figure 7—source data 1. (F) Summary of mechanisms underlying the emergence of dendrite targeting specificity revealed by two- photon and AO- LLSM imaging of DL1 PN dendrites. The online version of this article includes the following video, source data, and figure supplement(s) for figure 7: Source data 1. Source data for Figure 7E. Figure supplement 1. Analyses of DL1 projection neuron (PN) dendritic branches captured by AO- LLSM imaging. Figure 7—video 1. AO- LLSM time- lapse imaging of 3 hr DL1 projection neuron (PN) dendrites. https://elifesciences.org/articles/85521/figures#fig7video1 Figure 7—video 2. AO- LLSM time- lapse imaging of 6 hr DL1 projection neuron (PN) dendrites. https://elifesciences.org/articles/85521/figures#fig7video2 Figure 7—video 3. AO- LLSM time- lapse imaging of 12 hr DL1 projection neuron (PN) dendrites. https://elifesciences.org/articles/85521/figures#fig7video3 Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 16 of 33 Developmental Biology | Neuroscience Research article lov+ PNs already sent adult- specific dendrites to a region ventromedial to DC2 PN dendrites (green arrowhead in Figure 8A3; see 3D rendering in Figure 8—video 1). This implies that lov+ PNs have already caught up with DC2 PNs on dendrite development at this stage, and the re- extension of lov+ PN dendrites must have happened even earlier. Indeed, we observed lov+ PN dendrites innervated the developing antennal lobe extensively at 6 hr APF (Figure 8A2). Such innervation was not observed at 0  hr APF (Figure  8A1). After 12  hr APF, the time course of lov+ PN dendrite development was comparable to that of DC2 PNs (Figure 8A4–6). To characterize dendritic re- extension at single- cell resolution, we developed a sparse, stochastic labeling strategy to label single lov+ PNs (Figure  8B). We found that lov+ PNs produced nascent branches from the main process dorsal to larval- specific dendrites as early as 3 hr APF (Figure 8C2–3; arrowheads in Figure 8C6–7). At 6 hr APF, when larval- specific dendrites were completely segregated from lov+ PNs, the robust extension of adult- specific dendrites was seen across the developing antennal lobe (Figure 8C4). These data indicate that lov+ PNs re- extend their adult- specific dendrites at a more dorsal location before the larval- specific dendrites are completely pruned. Do other embryonic- born PNs prune and re- extend their dendrites simultaneously? Like lov drivers, Mz612- GAL4 labels embryonic- born PNs, one of which is VA6 PN (Marin et al., 2005). In 3 hr APF brains co- labeled for Mz612+ and lov+ PNs, we could unambiguously access three single embryonic- born PN types: (1) lov+ Mz612– PN, (2) lov– Mz612+ PN, and (3) lov+ Mz612+PN (Figure 8—figure supplement 2A–B). Tracing of individual dendritic branches showed that all these PNs already re- ex- tended dendrites to varying extents prior to the separation of larval- specific dendrites from the rest of the processes (Figure 8—figure supplement 2C). Thus, concurrent pruning and re- extension apply to multiple embryonic- born PN types. To capture the remodeling at the higher temporal resolution, we performed two- photon time- lapse imaging of single embryonic- born PNs labeled by Split7- GAL4 (Figure 8D, Figure 8—video 2, Figure 8—figure supplement 3). This GAL4 labels one embryonic- born PN (either VA6 or VA2 PN) at early pupal stages but eight PN types at 24 hr APF (Xie et al., 2021). Initially (3 hr APF + 0 hr ex vivo), no adult- specific dendrites were detected in live Split7+ PNs (Figure 8D1). The following ~3 hr ex vivo saw thickening of the main process (arrowhead in Figure 8D3). From 4 hr ex vivo onwards, re- extension occurred in the presumed developing antennal lobe located dorsal to larval- specific dendrites (arrowheads in Figure 8D4–8; see traces in Figure 8D9). Live imaging of Split7+ PNs also revealed that fragmentation of larval- specific dendrites occurred at the distal ends (Figure 8—figure supplement 3B1–5), and the process leading to larval- specific dendrites gradually disappeared as pruning approached completion (Figure 8—figure supplement 3B6–10). These observations suggest that pruning of embryonic- born PN dendrites is not initiated by severing at the proximal end. Distal- to- proximal pruning, rather than in the reversed direction, further supports concurrent but spatially segregated pruning and re- extension processes. It has been shown that dendritic pruning of embryonic- born PNs requires ecdysone signaling in a cell- autonomous manner (Marin et al., 2005). We asked if the re- extension process also depends on ecdysone signaling. We expressed a dominant negative form of ecdysone receptor (EcR- DN) in most PNs (including lov+ PNs) and monitored the development of lov+ PN dendrites (Figure  8— figure supplement 4). We found that inhibition of ecdysone signaling by EcR- DN expression not only suppressed pruning, but also blocked re- extension. This is consistent with a previous study reporting the dual requirement of ecdysone signaling in the pruning and re- extension of Drosophila anterior paired lateral (APL) neurons, although, unlike embryonic- born PNs, APL neurons prune and re- extend processes sequentially (at 6 hr and 18  hr APF, respectively) (Mayseless et  al., 2018). We currently could not distinguish if the lack of re- extension is due to defective pruning, or if ecdysone signaling controls pruning and re- extension independently. Taken together, our data demonstrate that embryonic- born PNs prune and re- extend dendrites simultaneously at spatially distinct regions, and that both processes require ecdysone signaling (Figure 8E). Such a ‘multi- tasking’ ability explains how embryonic- born PNs can re- integrate into the adult olfactory circuit and engage in its prototypic map formation in a timely manner. Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 17 of 33 Developmental Biology | Neuroscience Research article Figure 8. Embryonic- born projection neurons (PNs) timely participate in olfactory map formation via simultaneous pruning and re- extension. (A) Confocal images of fixed brains at indicated stages showing dendrite development of lov+ PNs (embryonic- born; labeled in green) and 91G04+DC2 PNs (larval- born; labeled in magenta). As 91G04- GAL4 also labels some embryonic- born PNs from 0 to 6 hr APF, their processes are found in the larval- specific antennal lobe (A1, 2). Right columns of A1, 2 show a zoom- in of the dashed boxes. Green arrowhead in (A2) indicates robust dendrite re- extension of embryonic- born PNs across the developing antennal lobe at 6 hr APF. (A1): N=6; (A2): N=12; (A3): N=9; (A4): N=12; (A5): N=9; (A6): N=5. (B) Schematic of the sparse, stochastic, and dual- color labeling strategy. In this strategy, the same cell has one copy of UAS- responsive conditional reporter 1 and one copy of QUAS- responsive reporter 2, both of which are integrated into the same 86Fb genomic locus (i.e. UAS- FRT- stop- FRT- reporter1/QUAS- FRT- stop- FRT- reporter2). FLP expression yields cis and trans recombination of FRT sites in a stochastic manner. Upon GAL4 expression, reporter 1 is expressed in cells with cis recombination, whereas reporter 2 is expressed only when cis and trans recombination events co- occur. (C) Sparse labeling of lov+ PNs (labeled in green; single- cell lov+ PNs in gray) at indicated developmental stages. (C6) and (C7) are zoom- in images of the rectangular boxes in (C2) and (C3), respectively. Arrowheads indicate nascent, adult- specific dendrites. Larval- specific dendrites are outlined by dashed orange lines. Arrows indicate axons projecting towards high brain centers. (C1): N=6; (C2–3): N=6; (C4): N=4; (C5): N=4. (D) Two- photon time- lapse imaging of a single Figure 8 continued on next page Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 18 of 33 Developmental Biology | Neuroscience Research article Figure 8 continued embryonic- born PN (Split7+; pseudo- colored in yellow) in a brain dissected at 3 hr APF and cultured for 23 hr ex vivo. Arrowhead in (D3) denote the thickening of the main process. Arrowheads in D4, 5 denote dendritic protrusions dorsal to larval- specific dendrites. (D9) shows neurite tracing of the embryonic- born PN. Triangles in (D9) indicate the degenerating larval- specific dendrites. N=3. (E) Schematic summary of remodeling of embryonic- born PN dendrites. Following simultaneous pruning and re- extension, embryonic- born PNs timely integrate into an adult olfactory circuit and, together with larval- born PNs, participate in the prototypic map formation. The online version of this article includes the following video and figure supplement(s) for figure 8: Figure supplement 1. Dendrite development of lov+ embryonic- born projection neurons (PNs). Figure supplement 2. Dendrite re- extension of lov+ and Mz612+ embryonic- born projection neurons (PNs). Figure supplement 3. Two- photon time- lapse imaging of Split7+ projection neuron (PN) dendrites. Figure supplement 4. Dual requirement of ecdysone signaling in pruning and re- extension of embryonic- born projection neuron (PN) dendrites. Figure 8—video 1. 3D rendering of z stacks of indicated projection neurons (PNs) in 12 hr APF antennal lobe. https://elifesciences.org/articles/85521/figures#fig8video1 Figure 8—video 2. Two- photon time- lapse imaging of Split7+ projection neuron (PN) dendrites. https://elifesciences.org/articles/85521/figures#fig8video2 Discussion Wiring logic for the prototypic olfactory map Prior to this study, no apparent logic linking PN lineage, birth order, and adult glomerular position has been found. Our systematic analyses of dendritic patterning at the resolution of specific PN types across development identified wiring logic underlying the spatial organization of the prototypic olfac- tory map (Figures 3 and 4). We found that PNs of a given lineage and temporal cohort share similar dendrite targeting spec- ificity and timing. Notably, dendrites of adPNs and lPNs respectively pattern the antennal lobe in rotating and binary manners following birth order. Based on our new observations and previous find- ings, we discuss possible mechanisms that execute the wiring logic to form the initial map: (1) speci- fication of the initial dendrite targeting through combinatorial inputs from lineage and birth order, (2) PN dendrite- dendrite interactions, and (3) contribution of the degenerating larval- specific antennal lobe. The spatial distinctions of cell bodies (e.g. Figure 1D1), axons (e.g. Figure 1—figure supplement 2A), and dendrites (e.g. Figure 4A1) of adPNs and lPNs observed in 0 hr APF pupal brain suggest that lineage endows projection specificity very early on. Lineage- specific transcription factors have been identified to instruct PN neurite targeting (Komiyama et al., 2003; Komiyama and Luo, 2007; Li et al., 2017; Xie et al., 2022), which might explain the differences between the adPN and lPN dendritic maps. Nonetheless, lineage alone does not account for the characteristic dendritic patterns. Rather, dendrite targeting can be predicted using combinatorial inputs from lineage and birth order. This combinatorial strategy is also seen in neuronal fate diversification and wiring of the Drosophila optic lobe and ventral nerve cord (Erclik et  al., 2017; Mark et  al., 2021), suggesting that it is a general principle in wiring the fly brain and likely also used in vertebrates (Holguera and Desplan, 2018; Sen, 2023). Substantial advances have been made in understanding how temporal patterning arises for intra- lineage specification (Doe, 2017; Miyares and Lee, 2019). For instance, the embry- onic ventral nerve cord neuroblasts sequentially express a cascade of temporal transcription factors (TTFs) to specify temporal identity (Isshiki et al., 2001). Larval optic lobe neuroblasts also deploy the same strategy but use a completely different TTF cascade (Li et al., 2013). Earlier studies show Chinmo, a TTF, and RNA- binding proteins that regulate Chinmo translation, control neuronal cell fate of the adPN lineage (Zhu et al., 2006; Liu et al., 2015). Specifically, DL1 PNs mutant for Chinmo project dendrites to D glomerulus that is targeted by the fourth larval- born adPNs (Zhu et al., 2006), demonstrating temporal order specifies final glomerular targeting. However, whether approximate temporal cohorts of a given PN lineage we described arise from sequential expression of temporal factors, and how such factors translate into initial dendrite patterning remains a fertile ground for future studies. Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 19 of 33 Developmental Biology | Neuroscience Research article Our time- lapse imaging data reveals robust PN dendritic dynamics during the initial targeting process (Figures  5–8), suggesting that cellular interactions among PN dendrites contribute to the initial map formation. This appears to contrast with the PN- ORN map in the mature antennal lobe, which is highly stable; connection specificity remains largely unchanged upon genetic ablation of their synaptic partners (Berdnik et al., 2006). Future works using early- onset genetic drivers for specific PN types for ablation can be used to investigate interactions between different PN groups, such as adPNs and lPNs, in the construction of the initial PN dendrite map. Does the degenerating larval- specific antennal lobe contribute to the initial dendrite patterning of the developing adult- specific antennal lobe? Earlier studies found that the larval- specific ORN axons secrete semaphorins, Sema- 2a and Sema- 2b, which act as repulsive ligands for dendrites of Sema- 1a- expressing PNs (including DL1 PNs) (Komiyama et  al., 2007; Sweeney et  al., 2011). As the larval- specific lobe is located ventromedial to the adult- specific lobe, Sema- 2a/b and Sema- 1a form opposing gradients along the dorsolateral- ventromedial axis. When DL1 PNs (the first- born/ developed) begin to target their dendrites, this repulsive action could destabilize branches in the ventromedial direction and thus favor dorsolateral targeting. This provides a plausible explanation as to why the adPN rotation pattern begins at the dorsolateral position. It would be interesting to see if the pattern is perturbed upon ablation of larval- specific ORNs. Our new tools for labeling and genetic manipulation of distinct PN types (Figure 2) will now enable in- depth investigations into the potential cellular interactions and molecular mechanisms leading to the initial map organization. Wiring logic evolves as development proceeds After the initial map formation at 12  hr APF, dendrite positions in the antennal lobe could change substantially in the next 36 hr (for example, see DC2 PNs in Figure 3B4–6 and DA1 and VA1d/DC3 PNs in Figure 3C4–7). These changes occur when dendrites of PNs with neighboring birth order begin to segregate and when ORN axons begin to invade the antennal lobe. Accordingly, the ovoid- shaped antennal lobe turns into a globular shape (30–50 hr APF; Figure 3C6- 7). These PN- autonomous and non- autonomous changes likely mask the initial wiring logic, explaining why previous studies, which mostly focused on examining the final glomerular targets in adults (Jefferis et al., 2001), have missed the earlier organization. Interestingly, the process of PN dendritic segregation coincides with the peak of PN transcriptomic diversity at 24 hr APF (Li et al., 2017; Xie et al., 2021). Recent proteomics and genetic analyses have indicated that PN dendrite targeting is mediated by cell- surface proteins cooperating as a combinatorial code (Xie et al., 2022). The evolving wiring logic, which is consistent with the stepwise assembly of an olfactory circuit (Hong and Luo, 2014), suggests the combinatorial codes are not static. We propose that PNs use a numerically simpler code for initial dendrite targeting. Following the expansion of transcriptomic diversity, PNs acquire a more complex code mediating dendritic segregation of neighboring PNs and matching of PN dendrites and ORN axons. Functional characterization of differentially expressed genes between 12 hr and 24 hr APF PNs may provide molecular insights into how the degree of discreteness in the olfactory map arises. Although the initial wiring logic is not apparent in the final map, several lines of evidence suggest the final map depends on the initial map. First, as mentioned above, the change of the temporal identity of DL1 PNs affects glomerular targeting (Zhu et al., 2006). Second, loss of Sema- 1a in DL1 PNs occasionally causes mistargeting in areas outside of the antennal lobe, and dendrite mistargeting phenotype along the dorsolateral- ventromedial axis is persistent across development as well as in adulthood (Komiyama et  al., 2007). Our work thus demonstrates that identification of the wiring logic in the early stages should help us better resolve the architectures in complex neural circuits. Selective branch stabilization as a cellular mechanism for dendrite targeting Utilizing an early pupal brain explant culture system coupled with two- photon and AO- LLSM imaging (Figure 5), we presented the first time- lapse videos following dendrite development of a specific PN type – DL1 PNs (Figures 6 and 7). We found that DL1 PN dendrites initiate active targeting towards their dorsolateral target with direction- dependent branch stabilization. This directional selectivity provides a cellular basis for the emerging targeting specificity of PN dendrites at the beginning of olfactory map formation. Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 20 of 33 Developmental Biology | Neuroscience Research article Although selective branch stabilization as a mechanism to achieve axon targeting specificity has been described in neurons in the vertebrate and invertebrate systems (e.g. Yates et  al., 2001; Li et  al., 2021), our time- lapse imaging showed, for the first time to our knowledge, that selective branch stabilization is also used to achieve dendrite targeting specificity. Furthermore, AO- LLSM imaging revealed that selective stabilization and destabilization of dendritic branches occur on the timescale of seconds. As the rate of olfactory circuit development in the brain explants was slower than normal development (Figure 5F), we might have captured PN dendritic dynamics in slow motion. Using AO- LLSM for high spatiotemporal resolution imaging, we just begin to appreciate how fast PN dendrites are coordinating trajectory choices with branch stabilization to make the appropriate deci- sion. Having characterized the dendritic branch dynamics of the wild- type DL1 PNs, we have set the stage for future studies addressing how positional cues and the downstream signaling instruct wiring, and whether other PN types follow similar rules as DL1 PNs. Simultaneous pruning and re-extension as novel remodeling mechanism for neuronal remodeling Our data on embryonic- born adPN dendrite development reveals a novel mode of neuronal remod- eling during metamorphosis (Figure  8). In mushroom body γ neurons and body wall somatosen- sory neurons, two well- characterized systems, larval- specific neurites are first pruned, followed by re- extension of adult- specific processes (Watts et al., 2003; Williams and Truman, 2005; Yaniv and Schuldiner, 2016). However, embryonic- born adPNs prune larval- specific dendrites and re- extend adult- specific dendrites simultaneously but at spatially separated subcellular compartments. Such spatial segregation suggests that regional external cues could elicit compartmentalized downstream signals leading to opposite effects on the dendrites. Subcellular compartmentalization of signaling and cytoskeletal organization has been observed in diverse neuron types across species (Rolls et al., 2007; Kanamori et al., 2013; O’Hare et al., 2022). Why do embryonic- born adPNs ‘rush’ to re- extend dendrites? During normal development, it takes at least 18 hr for embryonic- born adPNs to produce and properly target dendrites (growth at 3–6 hr APF, initial targeting at 6–12  hr APF, and segregation at 21–30  hr APF). Given that the dendritic re- extension of embryonic- born PNs is ecdysone dependent (Figure 8—figure supplement 4), if the PNs did not re- extend dendrites at 3 hr APF, they would have to wait for the next ecdysone surge at ~20 hr APF (Thummel, 2001), which might be too late for their dendrites to engage in the proto- typic map formation. Thus, embryonic- born PNs develop a remodeling strategy that coordinates with the timing of systemic ecdysone release. By simultaneous pruning and re- extension, embryonic- born adPNs timely re- integrate into the adult prototypic map that readily serves as a target for subsequent ORN axon innervation. In conclusion, our study highlights the power and necessity of type- specific neuronal access and time- lapse imaging to identify wiring logic and mechanisms underlying the origin of an olfactory map. Applying similar approaches to other developing neural maps across species should broaden our understanding of the generic and specialized designs that give rise to functional maps with diverse architectures. Materials and methods Drosophila stocks and husbandry Flies were maintained on a standard cornmeal medium at 25 °C. Fly lines used in this study included GH146- FLP (Hong et  al., 2009), QUAS- FRT- stop- FRT- mCD8- GFP (Potter et  al., 2010), UAS- mCD8- GFP (Lee and Luo, 1999), UAS- mCD8- FRT- GFP- FRT- RFP (Stork et al., 2014), VT033006- GAL4 (Tirian and Dickson, 2017), Mz19- GAL4 (Jefferis et al., 2004), 91G04- GAL4 (Jenett et al., 2012), Mz612- GAL4 (Marin et al., 2005), 71B05- GAL4 (Jenett et al., 2012), Split7- GAL4 (Xie et al., 2021), QUAS- FLP (Potter et al., 2010), and UAS- EcR.B1-ΔC655.F645A (Cherbas et al., 2003). The following GAL4 lines were obtained from Bloomington Drosophila Stock Center (BDSC): tsh- GAL4 (BDSC #3040) and lov- GAL4 (BDSC #3737). The following two stocks were used for MARCM analyses: (1) UAS- mCD8- GFP, hs- FLP; FRTG13, tub- GAL80;; GH146- GAL4, and (2) FRTG13, UAS- mCD8- GFP (Lee and Luo, 1999). Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 21 of 33 Developmental Biology | Neuroscience Research article The following lines were generated in this study: UAS- FRT10- stop- FRT10- 3xHalo7- CAAX (on either II or III chromosome), UAS- FRT- myr- 4xSNAPf- FRT- 3xHalo7- CAAX (III), UAS- FRT- myr- mGreenLantern- FRT- 3xHalo7- CAAX (II), QUAS- FRT- stop- FRT- myr- 4xSNAPf (III), run- T2A- FLP (X), acj6- T2A- FLP (X), acj6- T2A- QF2 (X), CG14322- T2A- QF2 (III), and lov- T2A- QF2 (II). Drosophila genotypes tub- GAL80/FRTG13, UAS- mCD8- GFP;; supplement 1B: GH146- FLP/UAS- FRT10- stop- FRT10- 3xHalo7- CAAX; supplement 1A: GH146- FLP/UAS- FRT10- stop- FRT10- 3xHalo7- CAAX; Figure 1D, Figure 1—figure supplement 1, Figure 1—figure supplement 2: run- T2A- FLP/+; UAS- mCD8- FRT- GFP- FRT- RFP/+; VT033006- GAL4/+ Figure  3A: acj6- T2A- QF2/+; GH146- FLP, QUAS- FRT- stop- FRT- mCD8- GFP/UAS- FRT10- stop- FRT10- 3xHalo7- CAAX; 71B05- GAL4/+ Figure  3B, Figure  3—figure supplement 1C: GH146- FLP/UAS- FRT10- stop- FRT10- 3xHalo7- CAAX; 91G04- GAL4/CG14322- T2A- QF2, QUAS- FRT- stop- FRT- myr- 4xSNAPf Figure 3C: acj6- T2A- FLP/+; Mz19- GAL4; UAS- FRT- myr- 4xSNAPf- FRT- 3xHalo7- CAAX/+ Figure 3D, Figure 3—figure supplement 2, Figure 3—figure supplement 3: UAS- mCD8- GFP, hs- FLP/+; FRTG13, tub- GAL80/FRTG13, UAS- mCD8- GFP;; GH146- GAL4 (IV)/+ Figure  3—figure 71B05- GAL4/+ Figure  3—figure 91G04- GAL4/+ Figure 3—video 1: Please refer to Figure 3 for genotypes. Figure  4A, Figure  4—figure supplement 1: GH146- FLP, UAS- FRT10- stop- FRT10- 3xHalo7- CAAX/tsh- GAL4; CG14322- T2A- QF2, QUAS- FRT- stop- FRT- myr- 4xSNAPf/+ Figure  4B: UAS- mCD8- GFP, hs- FLP/+; FRTG13, GH146- GAL4 (IV)/+ Figure 8—figure supplement 2: acj6- T2A- FLP/+; tsh- GAL4, UAS- mCD8- FRT- GFP- FRT- RFP Figure 4—video 1: Please refer to Figure 4 for genotypes. Figure  5E, Figure  5—video 1: run- T2A- FLP/+; UAS- FRT- myr- mGreenLantern- FRT- 3xHalo7- CAAX/+; VT033006- GAL4/+ Figure 5F: UAS- mCD8- GFP/+; VT033006- GAL4/+ Figure 5G1: GH146- FLP/UAS- FRT10- stop- FRT10- 3xHalo7- CAAX; 71B05- GAL4/+ Figure 5G2: GH146- FLP/tsh- GAL4; UAS- FRT10- stop- FRT10- 3xHalo7- CAAX/+ Figure  5—figure supplement 1, Figure  5—video 2: acj6- T2A- FLP/+; Mz19- GAL4/ UAS- FRT- myr- mGreenLantern- FRT- 3xHalo7- CAAX Figure  6A, Figure  6—figure supplement 1, Figure  6—video 1: UAS- mCD8- GFP, hs- FLP/+; FRTG13, tub- GAL80/FRTG13, UAS- mCD8- GFP;; GH146- GAL4 (IV)/+ Figure  7A–C, Figure  7—figure supplement 1, Figure  7—videos 1–3: acj6- T2A- QF2/+; QUAS- FRT- stop- FRT- mCD8- GFP/UAS- FRT10- stop- FRT10- 3xHalo7- CAAX; GH146- FLP, 71B05- GAL4/+ Figure 8A, Figure 8—figure supplement 1: GH146- FLP, QUAS- FRT- stop- FRT- mCD8- GFP/lov- T2A- QF2; UAS- FRT10- stop- FRT10- 3xHalo7- CAAX/91G04- GAL4 8C: Figure QUAS- FRT- stop- FRT- myr- 4xSNAPf Figure  8D, Figure  8—figure supplement 3, Figure  8—video 2: UAS- mCD8- GFP/+; Split7- GAL4 (i.e. FlyLight SS01867: 72C11- p65ADZp; VT033006- ZpGDBD)/+ Figure  8—figure supplement 2: GH146- FLP, QUAS- FRT- stop- FRT- mCD8- GFP/lov- T2A- QF2, Mz612- GAL4; UAS- FRT10- stop- FRT10- 3xHalo7- CAAX/+ Figure  8—figure UAS- mCD8- FRT- GFP- FRT- RFP Figure  8—figure supplement 4B: lov- T2A- QF2, QUAS- FLP/UAS- EcR- DN; VT033006- GAL4/ UAS- mCD8- FRT- GFP- FRT- RFP Figure 8—video 1: Please refer to Figure 8 for genotypes. lov- T2A- QF2, QUAS- FLP/+; VT033006- GAL4/ UAS- FRT10- stop- FRT10- 3xHalo7- CAAX/ GH146- FLP/lov- GAL4; supplement 4A: Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 22 of 33 Developmental Biology | Neuroscience Research article MARCM clonal analyses MARCM clonal analyses have been previously described (Lee and Luo, 1999). Larvae of the genotype UAS- mCD8- GFP, hs- FLP/+; FRTG13, tub- GAL80/FRTG13, UAS- mCD8- GFP;; GH146- GAL4/+ were  heat shocked at 37 °C for 1 hr. To label the first- born DL1 PNs, heat shock was applied at 0–24 hr after larval hatching (ALH). MARCM clones of early, middle (mid- late for adPNs), and late larval- born PNs were generated by applying heat shocks at 42–48 hr, 66–72 hr, and 96–100 hr ALH, respectively. As larvae developed at different rates (Tennessen and Thummel, 2011), we reasoned that even if we could collect 0 hr–2 hr ALH larvae, their development might have varied by the time of heat shock. To minimize the effects of unsynchronized development, we selected those heat- shocked larvae that were among the first to form puparia and collected these white pupae in a ~3 hr window for the clonal analyses. Transcriptomic analyses Transcriptomic analyses have been described previously (Xie et al., 2021). tSNE plots and dot plots were generated in Python using PN single- cell RNA sequencing data and code available at https:// github.com/Qijing-Xie/FlyPN_development (Xie, 2021). Generation of T2A-QF2/FLP lines To generate a T2A- QF2/FLP donor vector for acj6 (we used the same strategy for run, CG14322 and lov), a ~2000  bp genomic sequence flanking the stop codon of acj6 was PCR amplified and introduced into pCR- Blunt II- TOPO (ThermoFisher Scientific #450245), forming pTOPO- acj6. To build pTopo- acj6- T2A- QF2, T2A- QF2 including loxP- flanked 3xP3- RFP was PCR amplified from pBPGUw- HACK- QF2 (Addgene #80276), followed by insertion into pTOPO- acj6 right before the stop codon of acj6 by DNA assembly (New England BioLabs #E2621S). To generate T2A- FLP, we PCR- amplified FLP from the genomic DNA of GH146- FLP strain. QF2 in pTopo- acj6- T2A- QF2 was then replaced by FLP through DNA assembly. Using CRISPR Optimal Target Finder (Gratz et al., 2014), we selected a 20 bp gRNA target sequence that flanked the stop codon and cloned it into pU6- BbsI- chiRNA (Addgene #45946). If the gRNA sequence did not flank the stop codon, silent mutations were introduced at the PAM site of the donor vector by site- directed mutagenesis. Donor and gRNA vectors were co- injected into Cas9 embryos in- house or through BestGene. Generation of FLP-out reporters To generate pUAS- FRT10- stop- FRT10- 3xHalo7- CAAX, FRT10- stop- FRT10 was PCR amplified from pUAS- FRT10- stop- FRT10- mCD8- GFP (Li et  al., 2021) and inserted into pUAS- 3xHalo7- CAAX (Addgene #87646) through NotI and DNA assembly. To generate pUAS- FRT- myr- 4xSNAPf- FRT- 3xHalo7- CAAX, we first PCR amplified myr- 4xSNAPf from pUAS- myr- 4xSNAPf (Addgene #87637) using FRT- containing primers. FRT- myr- 4xSNAPf- FRT was then introduced into pCR- Blunt II- TOPO, forming pTOPO- FRT- myr- 4xSNAPf- FRT. Using NotI- containing primers, FRT- myr- 4xSNAPf- FRT was PCR amplified and subcloned into pUAS- 3xHalo7- CAAX through NotI. To generate pUAS- FRT- myr- mGreenLantern- FRT- 3xHalo7- CAAX, we first PCR amplified mGreen- Lantern from pcDNA3.1- mGreenLantern (Addgene #161912). Using MluI and XbaI, we replaced 4xSNAPf in pUAS- myr- 4xSNAPf with mGreenLantern to build pUAS- myr- mGreenLantern. myr- mGreenLantern was PCR amplified with the introduction of FRT sequence, followed by insertion into pCR- Blunt II- TOPO. Using the NotI- containing primers, FRT- myr- mGreenLantern- FRT was PCR ampli- fied and subcloned into pUAS- 3xHalo7- CAAX through NotI. To generate pQUAS- FRT- stop- FRT- myr- 4xSNAPf, we first PCR amplified FRT- stop from pJFRC7- 20XUAS- FRT- stop- FRT- mCD8- GFP (Li et al., 2021) and inserted it into pTOPO- FRT- myr- 4xSNAPf- FRT through DNA assembly to form pTOPO- FRT- stop- FRT- myr- 4xSNAPf- FRT. Using NotI- containing forward and KpnI- containing reverse primers, FRT- stop- FRT- myr- 4xSNAPf was PCR amplified and subcloned into p10XQUAST. p10XQUAST was generated using p5XQUAS (Addgene #24349) and p10xQUAS- CsChrimson (Addgene #163629). attP24 and 86Fb landing sites were used for site- directed integration. Immunofluorescence staining and confocal imaging Fly brain dissection for immunostaining and live imaging has been described (Wu and Luo, 2006). Briefly, brains were dissected in phosphate- buffered saline (PBS) and fixed with 4% Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 23 of 33 Developmental Biology | Neuroscience Research article paraformaldehyde in PBS for 20 min on a nutator at room temperature. Fixed brains were washed with 0.1% Triton X- 100 in PBS (PBST) for 10  min twice. After blocking with 5% normal donkey serum in PBST for 1 hr at room temperature, the brains were incubated with primary antibodies overnight at 4  °C. After PBST wash, brains were incubated with secondary antibodies (1:1000; Jackson ImmunoResearch) in dark for 2  hr at room temperature. Washed and mounted brains were imaged with confocal laser scanning microscopy (ZEISS LSM 780; LSM 900 with Airyscan 2). Images were processed with ImageJ. Neurite tracing images were generated using Simple Neurite Tracer (SNT) (Arshadi et al., 2021). Primary antibodies used included chicken anti- GFP (1:1000; Aves Lab #GFP- 1020), rabbit anti- DsRed (1:500; TaKaRa #632496), rat anti- Cadherin DN (1:30; Developmental Studies Hybridoma Bank DSHB DN- Ex#8 supernatant), and mouse anti- Bruchpilot (1:30; DSHB nc82 supernatant). Chemical labeling Chemical labeling of Drosophila brains has been described (Kohl et al., 2014). Janelia Fluor (JF) Halo and SNAP ligands (stocks at 1 mM) were gifts from Dr. Luke Lavis (Grimm et al., 2017; Grimm et al., 2021). Fixed brains were washed with PBST for 5  min, followed by incubation with Halo and/or SNAP ligands (diluted in PBS) for 45  min at room temperature. Brains were then washed with PBST for 5  min, followed by blocking and immunostaining if necessary. For the co- incubation of Halo and SNAP ligands, JF503- cpSNAP (1:1000) and JF646- Halo (1:1000) were used. Alternatively, JFX650- SNAP (1:1000) and JFX554- Halo (1:10,000) were used. When only Halo ligands were needed, either JF646- Halo or JF635- Halo (1:1000) was used. For live brain imaging, dissected brains were incubated with Halo ligands diluted in culture media (described below) for 30 min at room temperature. For two- photon imaging, JF570- Halo was used at 1:5000. For AO- LLSM imaging, following JF646- Halo incubation at 1:1000, the brains were incubated with 1 µM Sulforhodamine 101 (Sigma) for 5 min at room temperature. The brains were then briefly washed with culture media before imaging. Brain explant culture setup and medium preparation Brain explant culture setup was modified based on Li et al., 2021; Li and Luo, 2021. A Sylgard plate with a thickness of ~2 millimeters was prepared by mixing base and curing agent at 10:1 ratio (DOW SYLGARD 184 Silicone Elastomer Kit). The mixture was poured into a 60 mm × 15 mm dish in which it was cured for two days at room temperature. Once cured, the plate was cut into small squares (~15 mm × ~15 mm). Indentations were created based on the size of an early pupal brain using a No.11 scalpel. Additional slits were made around the indentations for attaching imaginal discs which served as anchors to hold the brain position. A square Sylgard piece was then placed in a 60 mm × 15 mm dish or on a 25 mm round coverslip in preparation for two- photon/AO- LLSM imaging. Culture medium was prepared based on published methods (Rabinovich et al., 2015; Li and Luo, 2021; Li et al., 2021). The medium contained Schneider’s Drosophila Medium (ThermoFisher Scientific #21720001), 10% heat- inactivated Fetal Bovine Serum (ThermoFisher Scientific #16140071), 10 µg/mL human recombinant insulin (ThermoFisher Scientific #12585014; stock = 4 mg/mL), 1:100 Penicillin- Streptomycin (ThermoFisher Scientific #15140122). For 0 hr–6 hr APF brain culture, 0.5 mM ascorbic acid (Sigma #A4544; stock concentration = 50 mg/mL in water) was included. 20- hydroxyecdysone (Sigma #H5142; stock concentration = 1 mg/mL in ethanol) was used for 0 hr–6 hr and 12 hr brain explants at 20 µM and 2 µM, respectively. Culture medium was oxygenated for 20 min before use. Single- and dual-color imaging with two-photon microscopy Single- and dual- color imaging of PNs were performed at room temperature using a custom- built two- photon microscope (Prairie Technologies) with a Chameleon Ti:Sapphire laser (Coherent) and a 16 X water- immersion objective (0.8 NA; Nikon). Excitation wavelength was set at 920 nm for GFP imaging, and at 935 nm for co- imaging of mGreenLantern and JF570- Halo. z- stacks were obtained at 4 µm increments (10 µm increments for Figure 5—video 1). Images were acquired at a resolution of 1024 × 1024 pixel2 (512 × 512 for Figure 5—video 1), with a pixel dwell time of 6.8 µs and an optical zoom of 2.1, and at a frequency every 20 min for 8–23 hr. Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 24 of 33 Developmental Biology | Neuroscience Research article Dual-color imaging with AO-LLSM For AO- LLSM- based imaging, the excitation and detection objectives along with the 25 mm coverslip were immersed in ~40 mL of culture medium at room temperature. Explant brains held on Sylgard plate were excited simultaneously using 488 nm (for GFP) and 642 nm (for JF- 646) lasers operating with ~2–10 mW input power to the microscope (corresponding to ~10–50 µW at the back aperture of the excitation objective). An exposure time of 20–50 msec was used to balance imaging speed and signal- to- noise ratio (SNR). Dithered lattice light- sheet patterns with an inner/outer numerical aperture of 0.35/0.4 or 0.38/0.4 were used. The optical sections were collected by an axial step size of 250 nm in the detection objective coordinate, with a total of 81–201 steps (corresponding to a total axial scan range of 20–50 µm). Emission light from GFP and JF- 646 was separated by a dichromatic mirror (Di03- R561, Semrock, IDEX Health & Science, LLC, Rochester, NY) and captured by two Hamamatsu ORCA- Fusion sCMOS cameras simultaneously (Hamamatsu Photonics, Hamamatsu City, Japan). Prior to the acquisition of the time series data, the imaged volume was corrected for optical aberrations using a two- photon guide star- based adaptive optics method (Chen et al., 2014; Wang et al., 2014; Liu et al., 2018). Each imaged volume was deconvolved using Richardson- Lucy algorithm on HHMI Janelia Research Campus’ or Advanced Bioimaging Center’s computing cluster (https://github.com/ scopetools/cudadecon, Lambert et al., 2023; https://github.com/abcucberkeley/LLSM3DTools, Ruan and Upadhyayula, 2020) with experimentally measured point spread functions obtained from 100 or 200 nm fluorescent beads (Invitrogen FluoSpheres Carboxylate- Modified Microspheres, 505/515 nm, F8803, FF8811). The AO- LLSM was operated using a custom LabVIEW software (National Instruments, Woburn, MA). Statistics For data analyses, t- test and one- way ANOVA were used to determine p values as indicated in the figure legend for each graph, and graphs were generated using Excel. Exact p values were provided in source data files. Material and data availability All reagents generated in this study are available from the lead corresponding author upon request. Figure 3—figure supplement 3—source data 1, Figure 6—source data 1, and Figure 7—source data 1 contain the numerical and statistical data used to generate the figures. The confocal imaging dataset is available at Brain Image Library under DOI https://doi.org/10.35077/g.933. Acknowledgements We thank the Luo lab members for constructive feedback on the manuscript; Tzumin Lee for sharing equipment at Janelia Research Campus; Luke Lavis for sharing JF dyes. This work was supported by a grant from NIH (R01 DC005982 to LL). TL was supported by NIH 1K99DC01883001. GL and SU are funded by Philomathia Foundation. SU is funded by the Chan Zuckerberg Initiative Imaging Scientist program. SU is a Chan Zuckerberg Biohub Investigator. EB and LL are HHMI investigators. Additional information Funding Funder National Institutes of Health Philomathia Foundation Chan Zuckerberg Initiative National Institutes of Health Grant reference number Author R01 DC005982 Liqun Luo Gaoxiang Liu Srigokul Upadhyayula Srigokul Upadhyayula 1K99DC01883001 Tongchao Li Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 25 of 33 Developmental Biology | Neuroscience Research article Funder Grant reference number Author Howard Hughes Medical Institute Eric Betzig Liqun Luo The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Author contributions Kenneth Kin Lam Wong, Conceptualization, Data curation, Formal analysis, Investigation, Visualization, Methodology, Writing - original draft; Tongchao Li, Resources, Investigation, Methodology, Writing – review and editing; Tian- Ming Fu, Gaoxiang Liu, Resources, Data curation, Investigation, Meth- odology, Writing – review and editing; Cheng Lyu, Resources, Methodology, Writing – review and editing; Sayeh Kohani, Data curation, Investigation; Qijing Xie, Data curation, Investigation, Writing – review and editing; David J Luginbuhl, Resources, Data curation, Writing – review and editing; Srigokul Upadhyayula, Eric Betzig, Resources, Supervision, Writing – review and editing; Liqun Luo, Conceptualization, Supervision, Funding acquisition, Investigation, Methodology, Project administra- tion, Writing – review and editing Author ORCIDs Kenneth Kin Lam Wong Tian- Ming Fu Liqun Luo http://orcid.org/0000-0001-6265-0859 http://orcid.org/0000-0001-5467-9264 http://orcid.org/0000-0002-5597-4051 Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.85521.sa1 Author response https://doi.org/10.7554/eLife.85521.sa2 Additional files Supplementary files • Supplementary file 1. Sample variability among individual brains. A supplemental table describing the biological and technical variations we observed among individual brain samples, and measures we took to minimize them, if possible. • MDAR checklist Data availability Figure 3—source data 1, Figure 5—source data 1, Figure 6—source data 1, and Figure 7—source data 1 contain the numerical and statistical data used to generate the figures. The confocal imaging dataset is available at Brain Image Library under DOI https://doi.org/10.35077/g.933. The following dataset was generated: Author(s) Wong KLK Year 2023 Dataset title Dataset URL Database and Identifier https:// doi. org/ 10. 35077/ g. 933 Brain Image Library, 10.35077/g.933 Origin of wiring specificity in an olfactory map revealed by neuron type- specific, time- lapse imaging of dendrite targeting: Confocal imaging of developing fly brain Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 26 of 33 Developmental Biology | Neuroscience Research article The following previously published dataset was used: Dataset title Dataset URL Database and Identifier Temporal evolution of single- cell transcriptomes of Drosophila olfactory projection neurons https://www. ncbi. nlm. nih. gov/ geo/ query/ acc. cgi? acc= GSE161228 NCBI Gene Expression Omnibus, GSE161228 Year 2021 Author(s) Xie Q, Brbic M, Horns F, Kolluru SS, Jones RC, Li J, Reddy AR, Xie A, Kohani S, Li Z, McLaughlin CN, Li T, Xu C, Vacek D, Luginbuhl DJ, Leskovec J, Quake SR, Luo L Li H References Arshadi C, Günther U, Eddison M, Harrington KIS, Ferreira TA. 2021. Snt: a unifying toolbox for quantification of neuronal anatomy. Nature Methods 18:374–377. DOI: https://doi.org/10.1038/s41592-021-01105-7, PMID: 33795878 Berdnik D, Chihara T, Couto A, Luo L. 2006. Wiring stability of the adult Drosophila olfactory circuit after lesion. The Journal of Neuroscience 26:3367–3376. DOI: https://doi.org/10.1523/JNEUROSCI.4941-05.2006, PMID: 16571743 Berns DS, DeNardo LA, Pederick DT, Luo L. 2018. Teneurin- 3 controls topographic circuit assembly in the hippocampus. Nature 554:328–333. DOI: https://doi.org/10.1038/nature25463, PMID: 29414938 Campbell BC, Nabel EM, Murdock MH, Lao- Peregrin C, Tsoulfas P, Blackmore MG, Lee FS, Liston C, Morishita H, Petsko GA. 2020. MGreenLantern: a bright monomeric fluorescent protein with rapid expression and cell filling properties for neuronal imaging. PNAS 117:30710–30721. DOI: https://doi.org/10.1073/pnas.2000942117, PMID: 33208539 Cang J, Feldheim DA. 2013. Developmental mechanisms of topographic MAP formation and alignment. Annual Review of Neuroscience 36:51–77. DOI: https://doi.org/10.1146/annurev-neuro-062012-170341, PMID: 23642132 Chen BC, Legant WR, Wang K, Shao L, Milkie DE, Davidson MW, Janetopoulos C, Wu XS, Hammer JA, Liu Z, English BP, Mimori- Kiyosue Y, Romero DP, Ritter AT, Lippincott- Schwartz J, Fritz- Laylin L, Mullins RD, Mitchell DM, Bembenek JN, Reymann AC, et al. 2014. Lattice light- sheet microscopy: imaging molecules to embryos at high spatiotemporal resolution. Science 346:1257998. DOI: https://doi.org/10.1126/science. 1257998, PMID: 25342811 Cherbas L, Hu X, Zhimulev I, Belyaeva E, Cherbas P. 2003. EcR isoforms in Drosophila: testing tissue- specific requirements by targeted blockade and rescue. Development 130:271–284. DOI: https://doi.org/10.1242/dev. 00205, PMID: 12466195 Diao F, White BH. 2012. A novel approach for directing transgene expression in Drosophila: T2A- gal4 in- frame fusion. Genetics 190:1139–1144. DOI: https://doi.org/10.1534/genetics.111.136291, PMID: 22209908 Doe CQ. 2017. Temporal patterning in the Drosophila CNS. Annual Review of Cell and Developmental Biology 33:219–240. DOI: https://doi.org/10.1146/annurev-cellbio-111315-125210, PMID: 28992439 Erclik T, Li X, Courgeon M, Bertet C, Chen Z, Baumert R, Ng J, Koo C, Arain U, Behnia R, del Valle Rodriguez A, Senderowicz L, Negre N, White KP, Desplan C. 2017. Integration of temporal and spatial patterning generates neural diversity. Nature 541:365–370. DOI: https://doi.org/10.1038/nature20794, PMID: 28077877 Gao Q, Yuan B, Chess A. 2000. Convergent projections of Drosophila olfactory neurons to specific glomeruli in the antennal lobe. Nature Neuroscience 3:780–785. DOI: https://doi.org/10.1038/77680, PMID: 10903570 Gratz SJ, Ukken FP, Rubinstein CD, Thiede G, Donohue LK, Cummings AM, O’Connor- Giles KM. 2014. Highly specific and efficient CRISPR/Cas9- catalyzed homology- directed repair in Drosophila. Genetics 196:961–971. DOI: https://doi.org/10.1534/genetics.113.160713, PMID: 24478335 Grimm JB, Muthusamy AK, Liang Y, Brown TA, Lemon WC, Patel R, Lu R, Macklin JJ, Keller PJ, Ji N, Lavis LD. 2017. A general method to fine- tune fluorophores for live- cell and in vivo imaging. Nature Methods 14:987– 994. DOI: https://doi.org/10.1038/nmeth.4403, PMID: 28869757 Grimm JB, Xie L, Casler JC, Patel R, Tkachuk AN, Falco N, Choi H, Lippincott- Schwartz J, Brown TA, Glick BS, Liu Z, Lavis LD. 2021. A general method to improve fluorophores using deuterated auxochromes. JACS Au 1:690–696. DOI: https://doi.org/10.1021/jacsau.1c00006, PMID: 34056637 Holguera I, Desplan C. 2018. Neuronal specification in space and time. Science 362:176–180. DOI: https://doi. org/10.1126/science.aas9435, PMID: 30309944 Hong W, Zhu H, Potter CJ, Barsh G, Kurusu M, Zinn K, Luo L. 2009. Leucine- Rich repeat transmembrane proteins instruct discrete dendrite targeting in an olfactory MAP. Nature Neuroscience 12:1542–1550. DOI: https://doi. org/10.1038/nn.2442, PMID: 19915565 Hong W, Mosca TJ, Luo L. 2012. Teneurins instruct synaptic partner matching in an olfactory MAP. Nature 484:201–207. DOI: https://doi.org/10.1038/nature10926, PMID: 22425994 Hong W, Luo L. 2014. Genetic control of wiring specificity in the fly olfactory system. Genetics 196:17–29. DOI: https://doi.org/10.1534/genetics.113.154336, PMID: 24395823 Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 27 of 33 Developmental Biology | Neuroscience Research article Isshiki T, Pearson B, Holbrook S, Doe CQ. 2001. Drosophila neuroblasts sequentially express transcription factors which specify the temporal identity of their neuronal progeny. Cell 106:511–521. DOI: https://doi.org/10.1016/ s0092-8674(01)00465-2, PMID: 11525736 Jefferis GSXE, Marin EC, Stocker RF, Luo L. 2001. Target neuron prespecification in the olfactory map of Drosophila. Nature 414:204–208. DOI: https://doi.org/10.1038/35102574, PMID: 11719930 Jefferis G, Vyas RM, Berdnik D, Ramaekers A, Stocker RF, Tanaka NK, Ito K, Luo L. 2004. Developmental origin of wiring specificity in the olfactory system of Drosophila. Development 131:117–130. DOI: https://doi.org/10. 1242/dev.00896, PMID: 14645123 Jenett A, Rubin GM, Ngo T- TB, Shepherd D, Murphy C, Dionne H, Pfeiffer BD, Cavallaro A, Hall D, Jeter J, Iyer N, Fetter D, Hausenfluck JH, Peng H, Trautman ET, Svirskas RR, Myers EW, Iwinski ZR, Aso Y, DePasquale GM, et al. 2012. A GAL4- driver line resource for Drosophila neurobiology. Cell Reports 2:991– 1001. DOI: https://doi.org/10.1016/j.celrep.2012.09.011, PMID: 23063364 Kanamori T, Kanai MI, Dairyo Y, Yasunaga K, Morikawa RK, Emoto K. 2013. Compartmentalized calcium transients trigger dendrite pruning in Drosophila sensory neurons. Science 340:1475–1478. DOI: https://doi. org/10.1126/science.1234879, PMID: 23722427 Kohl J, Ng J, Cachero S, Ciabatti E, Dolan MJ, Sutcliffe B, Tozer A, Ruehle S, Krueger D, Frechter S, Branco T, Tripodi M, Jefferis G. 2014. Ultrafast tissue staining with chemical tags. PNAS 111:E3805–E3814. DOI: https:// doi.org/10.1073/pnas.1411087111, PMID: 25157152 Komiyama T, Johnson WA, Luo L, Jefferis GSXE. 2003. From lineage to wiring specificity: POU domain transcription factors control precise connections of Drosophila olfactory projection neurons. Cell 112:157–167. DOI: https://doi.org/10.1016/s0092-8674(03)00030-8, PMID: 12553905 Komiyama T, Luo L. 2007. Intrinsic control of precise dendritic targeting by an ensemble of transcription factors. Current Biology 17:278–285. DOI: https://doi.org/10.1016/j.cub.2006.11.067, PMID: 17276922 Komiyama T, Sweeney LB, Schuldiner O, Garcia KC, Luo L. 2007. Graded expression of semaphorin- 1a cell- autonomously directs dendritic targeting of olfactory projection neurons. Cell 128:399–410. DOI: https://doi. org/10.1016/j.cell.2006.12.028, PMID: 17254975 Lambert L, Shao L, dmilkie. 2023. cudaDecon. GitHub. https://github.com/scopetools/cudadecon Lee T, Luo L. 1999. Mosaic analysis with a repressible cell marker for studies of gene function in neuronal morphogenesis. Neuron 22:451–461. DOI: https://doi.org/10.1016/s0896-6273(00)80701-1, PMID: 10197526 Lee T, Marticke S, Sung C, Robinow S, Luo L. 2000. Cell- Autonomous requirement of the USP/ecr- B ecdysone receptor for mushroom body neuronal remodeling in Drosophila. Neuron 28:807–818. DOI: https://doi.org/10. 1016/s0896-6273(00)00155-0, PMID: 11163268 Lee HH, Jan LY, Jan YN. 2009. Drosophila IKK- related kinase ik2 and katanin p60- like 1 regulate dendrite pruning of sensory neuron during metamorphosis. PNAS 106:6363–6368. DOI: https://doi.org/10.1073/pnas. 0902051106, PMID: 19329489 Li X, Erclik T, Bertet C, Chen Z, Voutev R, Venkatesh S, Morante J, Celik A, Desplan C. 2013. Temporal patterning of Drosophila medulla neuroblasts controls neural fates. Nature 498:456–462. DOI: https://doi.org/10.1038/ nature12319, PMID: 23783517 Li H, Horns F, Wu B, Xie Q, Li J, Li T, Luginbuhl DJ, Quake SR, Luo L. 2017. Classifying Drosophila olfactory projection neuron subtypes by single- cell RNA sequencing. Cell 171:1206–1220. DOI: https://doi.org/10.1016/ j.cell.2017.10.019, PMID: 29149607 Li T, Fu TM, Wong KKL, Li H, Xie Q, Luginbuhl DJ, Wagner MJ, Betzig E, Luo L. 2021. Cellular bases of olfactory circuit assembly revealed by systematic time- lapse imaging. Cell 184:5107–5121. DOI: https://doi.org/10.1016/ j.cell.2021.08.030, PMID: 34551316 Li T, Luo L. 2021. An explant system for time- lapse imaging studies of olfactory circuit assembly in Drosophila. Journal of Visualized Experiments 2021:176. DOI: https://doi.org/10.3791/62983, PMID: 34723938 Lin S, Kao CF, Yu HH, Huang Y, Lee T. 2012. Lineage analysis of Drosophila lateral antennal lobe neurons reveals notch- dependent binary temporal fate decisions. PLOS Biology 10:e1001425. DOI: https://doi.org/10.1371/ journal.pbio.1001425, PMID: 23185131 Liu Z, Yang CP, Sugino K, Fu CC, Liu LY, Yao X, Lee LP, Lee T. 2015. Opposing intrinsic temporal gradients guide neural stem cell production of varied neuronal fates. Science 350:317–320. DOI: https://doi.org/10.1126/ science.aad1886, PMID: 26472907 Liu TL, Upadhyayula S, Milkie DE, Singh V, Wang K, Swinburne IA, Mosaliganti KR, Collins ZM, Hiscock TW, Shea J, Kohrman AQ, Medwig TN, Dambournet D, Forster R, Cunniff B, Ruan Y, Yashiro H, Scholpp S, Meyerowitz EM, Hockemeyer D, et al. 2018. Observing the cell in its native state: imaging subcellular dynamics in multicellular organisms. Science 360:eaaq1392. DOI: https://doi.org/10.1126/science.aaq1392, PMID: 29674564 Luo L, Flanagan JG. 2007. Development of continuous and discrete neural maps. Neuron 56:284–300. DOI: https://doi.org/10.1016/j.neuron.2007.10.014, PMID: 17964246 Luo L. 2021. Architectures of neuronal circuits. Science 373:eabg7285. DOI: https://doi.org/10.1126/science. abg7285 Marin EC, Watts RJ, Tanaka NK, Ito K, Luo L. 2005. Developmentally programmed remodeling of the Drosophila olfactory circuit. Development 132:725–737. DOI: https://doi.org/10.1242/dev.01614, PMID: 15659487 Mark B, Lai SL, Zarin AA, Manning L, Pollington HQ, Litwin- Kumar A, Cardona A, Truman JW, Doe CQ. 2021. A developmental framework linking neurogenesis and circuit formation in the Drosophila CNS. eLife 10:e67510. DOI: https://doi.org/10.7554/eLife.67510, PMID: 33973523 Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 28 of 33 Developmental Biology | Neuroscience Research article Mayseless O, Berns DS, Yu XM, Riemensperger T, Fiala A, Schuldiner O. 2018. Developmental coordination during olfactory circuit remodeling in Drosophila. Neuron 99:1204–1215.. DOI: https://doi.org/10.1016/j. neuron.2018.07.050, PMID: 30146303 Miyares RL, Lee T. 2019. Temporal control of Drosophila central nervous system development. Current Opinion in Neurobiology 56:24–32. DOI: https://doi.org/10.1016/j.conb.2018.10.016, PMID: 30500514 Mombaerts P, Wang F, Dulac C, Chao SK, Nemes A, Mendelsohn M, Edmondson J, Axel R. 1996. Visualizing an olfactory sensory map. Cell 87:675–686. DOI: https://doi.org/10.1016/s0092-8674(00)81387-2, PMID: 8929536 Murthy VN. 2011. Olfactory maps in the brain. Annual Review of Neuroscience 34:233–258. DOI: https://doi. org/10.1146/annurev-neuro-061010-113738, PMID: 21692659 O’Hare JK, Gonzalez KC, Herrlinger SA, Hirabayashi Y, Hewitt VL, Blockus H, Szoboszlay M, Rolotti SV, Geiller TC, Negrean A, Chelur V, Polleux F, Losonczy A. 2022. Compartment- Specific tuning of dendritic feature selectivity by intracellular Ca2+ release. Science 375:eabm1670. DOI: https://doi.org/10.1126/science. abm1670, PMID: 35298275 Özel MN, Langen M, Hassan BA, Hiesinger PR. 2015. Filopodial dynamics and growth cone stabilization in Drosophila visual circuit development. eLife 4:e10721. DOI: https://doi.org/10.7554/eLife.10721, PMID: 26512889 Pederick DT, Lui JH, Gingrich EC, Xu C, Wagner MJ, Liu Y, He Z, Quake SR, Luo L. 2021. Reciprocal repulsions instruct the precise assembly of parallel hippocampal networks. Science 372:1068–1073. DOI: https://doi.org/ 10.1126/science.abg1774, PMID: 34083484 Potter CJ, Tasic B, Russler EV, Liang L, Luo L. 2010. The Q system: a repressible binary system for transgene expression, lineage tracing, and mosaic analysis. Cell 141:536–548. DOI: https://doi.org/10.1016/j.cell.2010.02. 025, PMID: 20434990 Rabinovich D, Mayseless O, Schuldiner O. 2015. Long term ex vivo culturing of Drosophila brain as a method to live image pupal brains: insights into the cellular mechanisms of neuronal remodeling. Frontiers in Cellular Neuroscience 9:327. DOI: https://doi.org/10.3389/fncel.2015.00327, PMID: 26379498 Riabinina O, Luginbuhl D, Marr E, Liu S, Wu MN, Luo L, Potter CJ. 2015. Improved and expanded Q- system reagents for genetic manipulations. Nature Methods 12:219–222, . DOI: https://doi.org/10.1038/nmeth.3250, PMID: 25581800 Rolls MM, Satoh D, Clyne PJ, Henner AL, Uemura T, Doe CQ. 2007. Polarity and intracellular compartmentalization of Drosophila neurons. Neural Development 2:7. DOI: https://doi.org/10.1186/1749- 8104-2-7, PMID: 17470283 Ruan X, Upadhyayula S. 2020. Llsm3Dtools: Tools for the analysis of 3D live images from lattice light- sheet microscopy (LLSM). d482a69. GitHub. https://github.com/abcucberkeley/LLSM3DTools Sen SQ. 2023. Generating neural diversity through spatial and temporal patterning. Seminars in Cell & Developmental Biology 142:54–66. DOI: https://doi.org/10.1016/j.semcdb.2022.06.002, PMID: 35738966 Stork T, Sheehan A, Tasdemir- Yilmaz OE, Freeman MR. 2014. Neuron- Glia interactions through the heartless FGF receptor signaling pathway mediate morphogenesis of Drosophila astrocytes. Neuron 83:388–403. DOI: https://doi.org/10.1016/j.neuron.2014.06.026, PMID: 25033182 Sutcliffe B, Ng J, Auer TO, Pasche M, Benton R, Jefferis GSXE, Cachero S. 2017. Second- Generation Drosophila chemical tags: sensitivity, versatility, and speed. Genetics 205:1399–1408. DOI: https://doi.org/10.1534/ genetics.116.199281, PMID: 28209589 Sweeney LB, Chou YH, Wu Z, Joo W, Komiyama T, Potter CJ, Kolodkin AL, Garcia KC, Luo L. 2011. Secreted semaphorins from degenerating larval Orn axons direct adult projection neuron dendrite targeting. Neuron 72:734–747. DOI: https://doi.org/10.1016/j.neuron.2011.09.026, PMID: 22153371 Tennessen JM, Thummel CS. 2011. Coordinating growth and maturation- insights from Drosophila. Current Biology 21:R750–R757. DOI: https://doi.org/10.1016/j.cub.2011.06.033, PMID: 21959165 Thummel CS. 2001. Molecular mechanisms of developmental timing in C. elegans and Drosophila. Developmental Cell 1:453–465. DOI: https://doi.org/10.1016/s1534-5807(01)00060-0, PMID: 11703937 Tirian L, Dickson BJ. 2017. The VT GAL4, LexA, and Split- GAL4 Driver Line Collections for Targeted Expression in the Drosophila Nervous System. bioRxiv. DOI: https://doi.org/10.1101/198648 Vosshall LB, Wong AM, Axel R. 2000. An olfactory sensory map in the fly brain. Cell 102:147–159. DOI: https:// doi.org/10.1016/s0092-8674(00)00021-0, PMID: 10943836 Wang K, Milkie DE, Saxena A, Engerer P, Misgeld T, Bronner ME, Mumm J, Betzig E. 2014. Rapid adaptive optical recovery of optimal resolution over large volumes. Nature Methods 11:625–628. DOI: https://doi.org/ 10.1038/nmeth.2925, PMID: 24727653 Watts RJ, Hoopfer ED, Luo L. 2003. Axon pruning during Drosophila metamorphosis: evidence for local degeneration and requirement of the ubiquitin- proteasome system. Neuron 38:871–885. DOI: https://doi.org/ 10.1016/s0896-6273(03)00295-2, PMID: 12818174 Williams DW, Truman JW. 2005. Cellular mechanisms of dendrite pruning in Drosophila: insights from in vivo time- lapse of remodeling dendritic arborizing sensory neurons. Development 132:3631–3642. DOI: https://doi. org/10.1242/dev.01928, PMID: 16033801 Wu JS, Luo L. 2006. A protocol for dissecting Drosophila melanogaster brains for live imaging or immunostaining. Nature Protocols 1:2110–2115. DOI: https://doi.org/10.1038/nprot.2006.336, PMID: 17487202 Xie Q, Wu B, Li J, Xu C, Li H, Luginbuhl DJ, Wang X, Ward A, Luo L. 2019. Transsynaptic fish- lips signaling prevents misconnections between nonsynaptic partner olfactory neurons. PNAS 116:16068–16073. DOI: https://doi.org/10.1073/pnas.1905832116, PMID: 31341080 Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 29 of 33 Developmental Biology | Neuroscience Research article Xie Q. 2021. Flypn_Development. bf76a37. GitHub. https://github.com/Qijing-Xie/FlyPN_development Xie Q, Brbic M, Horns F, Kolluru SS, Jones RC, Li J, Reddy AR, Xie A, Kohani S, Li Z, McLaughlin CN, Li T, Xu C, Vacek D, Luginbuhl DJ, Leskovec J, Quake SR, Luo L, Li H. 2021. Temporal evolution of single- cell transcriptomes of Drosophila olfactory projection neurons. eLife 10:e63450. DOI: https://doi.org/10.7554/ eLife.63450, PMID: 33427646 Xie Q, Li J, Li H, Udeshi ND, Svinkina T, Orlin D, Kohani S, Guajardo R, Mani DR, Xu C, Li T, Han S, Wei W, Shuster SA, Luginbuhl DJ, Quake SR, Murthy SE, Ting AY, Carr SA, Luo L. 2022. Transcription factor acj6 controls dendrite targeting via a combinatorial cell- surface code. Neuron 110:2299–2314.. DOI: https://doi. org/10.1016/j.neuron.2022.04.026, PMID: 35613619 Yaniv SP, Schuldiner O. 2016. A fly’s view of neuronal remodeling. Wiley Interdisciplinary Reviews. Developmental Biology 5:618–635. DOI: https://doi.org/10.1002/wdev.241, PMID: 27351747 Yates PA, Roskies AL, McLaughlin T, O’Leary DD. 2001. Topographic- specific axon branching controlled by ephrin- as is the critical event in retinotectal MAP development. The Journal of Neuroscience 21:8548–8563. DOI: https://doi.org/10.1523/JNEUROSCI.21-21-08548.2001, PMID: 11606643 Yu HH, Kao CF, He Y, Ding P, Kao JC, Lee T. 2010. A complete developmental sequence of A Drosophila neuronal lineage as revealed by twin- spot MARCM. PLOS Biology 8:e1000461. DOI: https://doi.org/10.1371/ journal.pbio.1000461, PMID: 20808769 Zhu S, Lin S, Kao CF, Awasaki T, Chiang AS, Lee T. 2006. Gradients of the Drosophila chinmo BTB- zinc finger protein govern neuronal temporal identity. Cell 127:409–422. DOI: https://doi.org/10.1016/j.cell.2006.08.045, PMID: 17055440 Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 30 of 33 Developmental Biology | Neuroscience Research article Appendix 1 Appendix 1—key resources table Reagent type (species) or resource Designation Source or reference Identifiers Additional information Genetic reagent (D. melanogaster) GH146- FLP DOI: 10.1038/nn.2442 Genetic reagent (D. melanogaster) QUAS- FRT- stop- FRT- mCD8- GFP DOI: 10.1016 /j. cell.2010.02.025 Genetic reagent (D. melanogaster) UAS- mCD8- GFP DOI: 10.1016 / s0896- 6273(00)80701–1 Genetic reagent (D. melanogaster) UAS- mCD8- FRT- GFP- FRT- RFP DOI: 10.1016 /j. neuron.2014.06.026 Genetic reagent (D. melanogaster) VT033006- GAL4 Genetic reagent (D. melanogaster) Mz19- GAL4 Genetic reagent (D. melanogaster) 91 G04- GAL4 Genetic reagent (D. melanogaster) Mz612- GAL4 Genetic reagent (D. melanogaster) 71B05- GAL4 Genetic reagent (D. melanogaster) Split7- GAL4 Genetic reagent (D. melanogaster) QUAS- FLP DOI: 10.1101/198648 DOI: 10.1242/dev.00896 DOI: 10.1016 /j. celrep.2012.09.011 DOI: 10.1242/dev.01614 DOI: 10.1016 /j. celrep.2012.09.011 DOI: 10.7554/eLife.63450 FlyLight:SS01867 DOI: 10.1016 /j. cell.2010.02.025 Genetic reagent (D. melanogaster) UAS- EcR.B1-ΔC655.F645A DOI: 10.1242/dev.00205 Genetic reagent (D. melanogaster) tsh- GAL4 Genetic reagent (D. melanogaster) lov- GAL4 Bloomington Drosophila Stock Center BDSC:3040 Bloomington Drosophila Stock Center BDSC:3737 Genetic reagent (D. melanogaster) UAS- mCD8- GFP, hs- FLP; FRTG13, tub- GAL80;; GH146- GAL4 DOI: 10.1016 / s0896- 6273(00)80701–1 Genetic reagent (D. melanogaster) FRTG13, UAS- mCD8- GFP DOI: 10.1016 / s0896- 6273(00)80701–1 Genetic reagent (D. melanogaster) UAS- FRT10- stop- FRT10- 3xHalo7- CAAX this paper Genetic reagent (D. melanogaster) UAS- FRT- myr- 4xSNAPf- FRT- 3xHalo7- CAAX this paper Genetic reagent (D. melanogaster) UAS- FRT- myr- mGreenLantern- FRT- 3xHalo7- CAAX this paper Genetic reagent (D. melanogaster) QUAS- FRT- stop- FRT- myr- 4xSNAPf this paper Genetic reagent (D. melanogaster) run- T2A- FLP Genetic reagent (D. melanogaster) acj6- T2A- FLP Genetic reagent (D. melanogaster) acj6- T2A- QF2 this paper this paper this paper Genetic reagent (D. melanogaster) CG14322- T2A- QF2 this paper Genetic reagent (D. melanogaster) lov- T2A- QF2 this paper Antibody chicken polyclonal anti- GFP Aves Lab Appendix 1 Continued on next page on either II or III chromosome; see Materials and methods on III chromosome; see Materials and methods on II chromosome; see Materials and methods on III chromosome; see Materials and methods on X chromosome; see Materials and methods on X chromosome; see Materials and methods on X chromosome; see Materials and methods on III chromosome; see Materials and methods on II chromosome; see Materials and methods RRID:AB_10000240; Aves Lab:GFP- 1020 (1:1000) Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 31 of 33 Developmental Biology | Neuroscience Research article Appendix 1 Continued Reagent type (species) or resource Designation Source or reference Identifiers Additional information Antibody rabbit polyclonal anti- DsRed TaKaRa RRID:AB_10013483; TaKaRa:632496 (1:500) Antibody rat monoclonal anti- Cadherin DN Developmental Studies Hybridoma Bank RRID:AB_528121; DSHB:DN- Ex#8 (1:30) Antibody mouse monoclonal anti- Bruchpilot Developmental Studies Hybridoma Bank RRID:AB_2314866; DSHB:nc82 supernatant (1:30) Recombinant DNA reagent Recombinant DNA reagent Recombinant DNA reagent Recombinant DNA reagent Recombinant DNA reagent pBPGUw- HACK- QF2 Addgene RRID:Addgene_80276 pU6- BbsI- chiRNA Addgene RRID:Addgene_45946 pUAS- 3xHalo7- CAAX Addgene RRID:Addgene_87646 pUAS- myr- 4xSNAPf Addgene RRID:Addgene_87637 pcDNA3.1- mGreenLantern Addgene RRID:Addgene_161912 Recombinant DNA reagent p5XQUAS Addgene RRID:Addgene_24349 Recombinant DNA reagent p10xQUAS- CsChrimson Addgene RRID:Addgene_163629 Recombinant DNA reagent pUAS- FRT10- stop- FRT10- 3xHalo7- CAAX this paper Recombinant DNA reagent pUAS- FRT- myr- 4xSNAPf- FRT- 3xHalo7- CAAX this paper Recombinant DNA reagent pUAS- FRT- myr- mGreenLantern- FRT- 3xHalo7- CAAX this paper Recombinant DNA reagent Recombinant DNA reagent pUAS- myr- mGreenLantern this paper pQUAS- FRT- stop- FRT- myr- 4xSNAPf this paper Chemical compound, drug SYLGARD 184 Silicone Elastomer Kit DOW Schneider’s Drosophila Medium ThermoFisher Scientific Fetal Bovine Serum ThermoFisher Scientific Human recombinant insulin ThermoFisher Scientific Penicillin- Streptomycin ThermoFisher Scientific backbone from pUAS- 3xHalo7- CAAX; see Materials and methods backbone from pUAS- 3xHalo7- CAAX; see Materials and methods backbone from pUAS- 3xHalo7- CAAX; see Materials and methods backbone from pUAS- myr- 4xSNAPf; see Materials and methods backbone from p5XQUAS; see Materials and methods DOW:2646340 ThermoFisher Scientific:21720001 ThermoFisher Scientific:16140071 ThermoFisher Scientific:12585014 ThermoFisher Scientific:15140122 used at 10% used at 10 µg/mL (1:100) Ascorbic acid Sigma Sigma:A4544 used at 50 mg/mL in water 20- hydroxyecdysone Sigma Sigma:H5142 used at 20 µM and 2 µM JF503- cpSNAP Chemical compound, drug JF646- Halo Chemical compound, drug Chemical compound, drug JFX650- SNAP JFX554- Halo Appendix 1 Continued on next page DOI: 10.1038/nmeth.4403; DOI: 10.1021/jacsau.1c00006 DOI: 10.1038/nmeth.4403; DOI: 10.1021/jacsau.1c00006 DOI: 10.1038/nmeth.4403; DOI: 10.1021/jacsau.1c00006 DOI: 10.1038/nmeth.4403; DOI: 10.1021/jacsau.1c00006 (1:1000); gift from Dr. Luke Lavis (1:1000); gift from Dr. Luke Lavis (1:1000); gift from Dr. Luke Lavis (1:10000); gift from Dr. Luke Lavis Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 32 of 33 Developmental Biology | Neuroscience Research article Appendix 1 Continued Reagent type (species) or resource Designation Source or reference Identifiers Additional information Chemical compound, drug JF635- Halo Chemical compound, drug JF570- Halo DOI: 10.1038/nmeth.4403; DOI: 10.1021/jacsau.1c00006 DOI: 10.1038/nmeth.4403; DOI: 10.1021/jacsau.1c00006 (1:1000); gift from Dr. Luke Lavis (1:5000); gift from Dr. Luke Lavis Chemical compound, drug Sulforhodamine 101 Sigma Sigma:S7635 used at 1 µM Software, algorithm ZEN Carl Zeiss RRID:SCR_013672 Software, algorithm ImageJ National Institutes of Health RRID:SCR_003070 Software, algorithm Python Programming Language Python RRID:SCR_008394 http://www.python.org/ Wong et al. eLife 2023;12:e85521. DOI: https://doi.org/10.7554/eLife.85521 33 of 33 Developmental Biology | Neuroscience
10.7554_elife.85970
RESEARCH ARTICLE Neural circuit- wide analysis of changes to gene expression during deafening- induced birdsong destabilization Bradley M Colquitt1,2*†, Kelly Li1,2, Foad Green1,2‡, Robert Veline1,2§, Michael S Brainard1,2* 1Howard Hughes Medical Institute, Chevy Chase, United States; 2Department of Physiology, University of California, San Francisco, San Francisco, United States *For correspondence: colquitt@ucsc.edu (BMC); michael.brainard@ucsf.edu (MSB) Present address: †Department of Molecular, Cell, and Developmental Biology, University of California- Santa Cruz, Santa Cruz, United States; ‡Syapse, Inc, San Francisco, United States; §The Advanced Science Research Center, The City University of New York, The Graduate Center at the City University of New York, New York, United States Competing interest: See page 30 Funding: See page 30 Preprinted: 14 December 2022 Received: 05 January 2023 Accepted: 17 April 2023 Published: 07 June 2023 Reviewing Editor: Anne E West, Duke University School of Medicine, United States Copyright Colquitt et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Abstract Sensory feedback is required for the stable execution of learned motor skills, and its loss can severely disrupt motor performance. The neural mechanisms that mediate sensorimotor stability have been extensively studied at systems and physiological levels, yet relatively little is known about how disruptions to sensory input alter the molecular properties of associated motor systems. Songbird courtship song, a model for skilled behavior, is a learned and highly structured vocalization that is destabilized following deafening. Here, we sought to determine how the loss of auditory feedback modifies gene expression and its coordination across the birdsong senso- rimotor circuit. To facilitate this system- wide analysis of transcriptional responses, we developed a gene expression profiling approach that enables the construction of hundreds of spatially- defined RNA- sequencing libraries. Using this method, we found that deafening preferentially alters gene expression across birdsong neural circuitry relative to surrounding areas, particularly in premotor and striatal regions. Genes with altered expression are associated with synaptic transmission, neuronal spines, and neuromodulation and show a bias toward expression in glutamatergic neurons and Pvalb/Sst- class GABAergic interneurons. We also found that connected song regions exhibit correla- tions in gene expression that were reduced in deafened birds relative to hearing birds, suggesting that song destabilization alters the inter- region coordination of transcriptional states. Finally, lesioning LMAN, a forebrain afferent of RA required for deafening- induced song plasticity, had the largest effect on groups of genes that were also most affected by deafening. Combined, this integrated transcriptomics analysis demonstrates that the loss of peripheral sensory input drives a distributed gene expression response throughout associated sensorimotor neural circuitry and iden- tifies specific candidate molecular and cellular mechanisms that support the stability and plasticity of learned motor skills. Editor's evaluation This is an important study that uses the song system in a bird model to understand the transcrip- tional mechanisms underlying neuronal adaptations to sensory deprivation. The manuscript offers compelling data in support of the authors' hypothesis that these transcriptional changes are related to song plasticity. The work will be of interest to biologists who study neuronal plasticity mechanisms. Introduction The accurate and stable performance of motor skills relies on sensory feedback (Todorov, 2004). The loss of this feedback, for example through hearing or vision loss from injury or neurodegeneration, can Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 1 of 37 Research article lead to increased errors in the execution of even well- learned motor behaviors, such as speech and walking (Lane and Webster, 1991; Waldstein, 1990; Wood et al., 2011). Yet it is poorly understood how such peripheral sensory loss influences the properties of central motor circuits to drive neural plasticity and how these effects in turn influence motor output. The courtship song of songbirds, a learned motor skill subserved by a dedicated and discrete neural architecture, offers a tractable system in which to characterize the neural mechanisms that underlie sensorimotor integration and motor skill stability. Juvenile birds produce unstructured and variable vocalizations that, over the course of several months of learning, become more structured, less vari- able, and more similar to adult song (Brainard and Doupe, 2013). In finches, after this developmental learning period, birdsong performance remains extraordinarily consistent from rendition- to- rendition over the course of a bird’s life and is said to be ‘crystallized.’ However, auditory feedback plays an essential role in maintaining this stability; modifying auditory feedback or completely removing audi- tory input through deafening can drive changes to birdsong (Brainard and Doupe, 2001; Brainard and Doupe, 2000; Fukushima and Margoliash, 2015; Leonardo and Konishi, 1999; Lombardino and Nottebohm, 2000; Nordeen and Nordeen, 1992; Okanoya and Yamaguchi, 1997; Tschida and Mooney, 2012; Woolley and Rubel, 1997). The neural mechanisms that underlie these changes have been studied in terms of physiology and morphology, yet we lack a transcriptome- and circuit- wide understanding of how altered auditory feedback influences gene expression in song sensorimotor circuitry — a critical biological vantage point to understand how sensory information intersects with the nervous system to influence motor plasticity. To gain insight into the molecular and cellular factors that regulate the stability of adult bird- song, we analyzed gene expression alterations in birdsong sensorimotor circuitry and surrounding non- song regions in response to deafening, a strong driver of song destabilization. We developed a gene expression profiling approach that enabled large- scale analysis of gene expression in spatially defined brain regions. Using this technique, we identified a suite of expression changes across the song system, including region- specific and common transcriptional responses as well as altered gene expression correlations across regions. Using a previously generated single- cell atlas of the premotor portion of song neural circuitry, we identified the cellular types that experience the greatest transcrip- tional change following deafening. Finally, we examined how input from a song region required for deafening- induced song plasticity influences gene expression in its song premotor target and found a diverse set of expression changes, with substantial overlap with those elicited by deafening. Results Deafening destabilizes birdsong and increases song variability Songbirds rely on auditory feedback to maintain the quality of their songs (Figure  1A). Past work has shown that experimentally removing this feedback by deafening results in the gradual deteri- oration of both song spectral structure and temporal ordering of the individual elements (syllables) that comprise song (Nordeen and Nordeen, 1992; Okanoya and Yamaguchi, 1997; Woolley and Rubel, 1997). Deafening also drives a range of physiological, cellular, and molecular changes in the song system, including alterations to neuronal turnover (Scott et al., 2000; Wang et al., 1999) (but see Pytte et al., 2012), dendritic spine morphology (Peng et al., 2012a; Peng et al., 2013; Tschida and Mooney, 2012; Zhou et al., 2017), neuronal excitability (Tschida and Mooney, 2012), and gene expression (Watanabe et al., 2002). We reasoned that comparisons of gene expression in birds undergoing song destabilization following deafening would uncover molecular pathways involved in either promoting or limiting song plasticity. We, therefore, generated a cohort of eighteen adult male Bengalese finches (Lonchura striata domestica) that were either deafened through bilateral cochlear removal or underwent a sham surgery (nine birds for each condition, Figure 1B). This cohort was further divided into sets of birds that were euthanized 4, 9, or 14 days post- procedure (three birds per procedure type and time point). This range of time points was used to generate a diversity of song destabilization values for subse- quent gene expression analysis. Birds were euthanized two hours after lights- on. As in previous studies (Brainard and Doupe, 2000; Okanoya and Yamaguchi, 1997; Tschida and Mooney, 2012; Woolley and Rubel, 1997), deafening caused a gradual decay of song quality over the course of several days, while sham surgery induced relatively little song change (Figure 1B–H, Figure 1—figure supplement Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 2 of 37 Chromosomes and Gene Expression | Neuroscience Research article Figure 1. Rapid and global destabilization of the song following deafening. (A) Song destabilization through the removal of auditory feedback. Adult songbirds use auditory feedback to evaluate their own song production and maintain song quality and consistency. Loss of auditory feedback results in the gradual destabilization of song. (B) Experimental overview. After a baseline period of song recording, Bengalese finches (Lonchura striata domestica) were either deafened through bilateral cochlear removal or underwent a sham surgery. After 4, 9, or 14 days post- surgery, birds were euthanized for gene expression analysis. Bengalese finch graphic obtained from Sainburg, 2020a. (C) Example spectrograms from one hearing (sham) and one deaf (bilateral cochlear removal) bird. Songs are shown from before the procedure and 14 days following the procedure. Labels below each spectrogram correspond to discrete categories of song units (‘syllables’). kHz, kiloHertz. (D) Uniform Manifold Approximation and Projection (UMAP) representation of syllable spectrograms (see Methods) across the entire recording period for each bird (4 days before to 14 days after the procedure). Data are split into ‘pre’-procedure (4–1 day before surgery) and ‘post’-procedure (1–14 days after surgery) subsets. For reference, gray points in each plot correspond to data from the other subset. Example syllable spectrograms are placed adjacent to their position in UMAP space. (E) Density plots of UMAP projections for the syllables from one deafened bird (shown in panel (A)) at two timepoints, one day before and 13 days after deafening. (F) Subtraction of UMAP densities in (E) from the average pre- procedure density. (G) Mean sum of UMAP density differences for syllables from birds that were either deafened (deaf) or underwent a sham surgery (hearing). For each bird and each day, positive UMAP density differences were summed and then averaged across birds. Error bands are standard errors of the mean. Color bars indicate days in which values were significantly different between deaf and hearing birds (Student’s t- test, two- sided, p<0.05). (H) UMAP plot of one syllable from one deafened bird colored by the day following deafening. (I) Comparison of fundamental frequency (FF) variability between hearing and deafened birds. Rolling coefficient of variation (CV, window size 11 syllables) was calculated for the fundamental frequencies of each harmonic stack for each bird. Shown are two example syllables from one hearing and one deafened bird, plotted across the number of days relative to the procedure date (sham or cochlear removal). (J) Mean FF CV in the 7–9 days following sham or cochlear removal normalized to FF CV in the 2 days before the procedure. Linear mixed- effects regression (see Methods) was used to estimate the group post vs. Figure 1 continued on next page Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 3 of 37 Chromosomes and Gene Expression | Neuroscienceiiiiibcddefiiiiibcddef891011pre14 days postiibcddefiiiiiibcddefiiiiiiiibcdddefiiibcdd91011121314150.2 s80kHzhearing (sham)deaf (cochlea removed)pre14 days postiiiiiibbcddefbcdddffbcdd0123408Frequency, kHzoutput_1446918112CDdaysrelative todeafeningsyllable cor37pu44HE-1 days13 daysUMAP densitydifference from pre-proceduremean UMAP density−0.0100.02F0.020-1 days13 daysfiiififiiffddfcidiiidiifiiiiididiiddiiiffiddici234567Time, s1020304050−4−202468101214deafdeafn=9hearingn=9mean summed difference from baselinedays relative to procedureGUMAP 1UMAP 2UMAP 1UMAP 2UMAP 1UMAP 2soundisolation18 Bengalese finches4-14 dayseuthanizecontinuous song recordingday of euthanasia2 hourscochlear removalor sham3 days−4−202468101214−1000100200days post−procedure−4−202468101214−1000100200300days post−procedureor38pu45 - hearingor37pu44 - deafRolling FF CV% of baselineIJ0204060FF CV, % of baselinehearingdeaf0.002NSA−50510AuditoryfeedbackErrorevaluationSong productionStablesongErrorevaluationSong productionDestabilizedsongauditory feedback removalBAuditoryfeedback Research article Figure 1 continued pre- procedure FF CV difference for hearing and deaf birds. P- values are obtained from the regression model using Satterthwaite’s degrees of freedom method. The online version of this article includes the following figure supplement(s) for figure 1: Figure supplement 1. Additional quantification of deafening- induced changes to the song. 1A–C). To visualize deafening- induced changes to a song, we calculated spectrograms for each syllable and used uniform manifold approximate projection (UMAP) to project these spectrograms onto a latent space, using an approach described in Sainburg et al., 2020c (Figure 1C, Figure 1— figure supplement 1A–C). Following deafening, the projections of syllable spectrograms gradually shifted to occupy different locations, indicating a change from a pre- procedure song (Figure 1D–H, Figure  1—figure supplement 1A and B). Syllable spectral changes after deafening were complex but generally trended toward an increase in syllable ‘noisiness’ (Wiener entropy) (Figure 1—figure supplement 1D and E). Although adult birdsong is a highly precise motor skill, its features vary slightly from rendition- to- rendition, similar to other motor skills. Past work has demonstrated that this variability is in part gener- ated by central neural mechanisms in the forebrain and is modulated by social context, indicating that song variability is an actively regulated component of birdsong (Kao et al., 2005; Kao and Brainard, 2006; Kojima et al., 2013; Moorman et al., 2021; Sakata et al., 2008). To assess how song variability changes following deafening, we focused on a single spectral feature, fundamental frequency (FF), and calculated its coefficient of variation (CV) across renditions (Figure 1J and K). Deafening resulted in a gradual increase in the CV of FF (day 7–9 post- procedure mean ± SEM, 57 ± 18%) while sham surgery elicited no change (0.11 ± 2.3%). This increase in rendition- to- rendition FF variability is consis- tent with reports describing an increase in within- syllable frequency modulation following deafening (Brainard and Doupe, 2001). These results indicate that deafening elicits both shifts in the structure of song as well as decreases in stereotypy across renditions. Neural circuit-wide analysis of gene expression Birdsong is generated by a dedicated and anatomically discrete neural circuit called the song system (Figure 2A and B). This defined architecture allows interrogation of how the molecular and cellular properties of each region (termed ‘song nuclei’) influence and is influenced by birdsong performance and learning. Four primary song nuclei reside in the telencephalon: HVC (proper name), RA (robust nucleus of the arcopallium), LMAN (lateral magnocellular nucleus of the anterior nidopallium), and Area X. HVC and RA comprise the song motor pathway (SMP) and are necessary for song performance (Nottebohm et  al., 1976; Simpson and Vicario, 1990). HVC influences the timing and temporal structure of song and projects to RA, which provides descending motor control of song via projections to syringeal and respiratory brainstem regions, which send recurrent connections back into the SMP to influence spectral and temporal features of the song (Goldberg and Fee, 2012; Vicario, 1991; Wild, 1993). HVC also projects to the striatal nucleus Area X that, together with the pallial region LMAN and thalamic region DLM, form the ’anterior forebrain pathway (AFP),’ which contributes to song plasticity both during song acquisition in juveniles and song adaptation in adults (Andalman and Fee, 2009; Bottjer et al., 1984; Brainard and Doupe, 2000; Charlesworth et al., 2012; Nordeen and Nordeen, 1993; Scharff and Nottebohm, 1991; Sohrabji et al., 1990; Warren et al., 2011; Williams and Mehta, 1999). Each region is embedded in a larger anatomical domain that lies outside of song control circuitry but shares similar molecular, connectivity, and functional properties (Figure  2B; Feenders et al., 2008; Helduser et al., 2013; Kröner and Güntürkün, 1999). HVC is located in the dorsal part of the caudal nidopallium (NC); RA is located in the arcopallium (Arco.); Area X is located in the striatum (Stri.); and LMAN is located in the rostral nidopallium (NR). In effect, these regions serve as ‘non- song’ comparators for each song region that enable the identification of molecular and cellular features that are specific to song- related perturbations. Disruptions to song that follow the loss of auditory feedback are associated with both local alter- ations to song nuclei as well as to the interactions among connected components of the song system neural circuit (Brainard and Doupe, 2000; Hamaguchi et al., 2014; Kojima et al., 2013; Watanabe et  al., 2006). To examine how song destabilization influences gene expression in the song system Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 4 of 37 Chromosomes and Gene Expression | Neuroscience Research article Figure 2. Neural circuit transcriptomics using Serial Laser Capture RNA- sequencing (SLCR- seq). (A) Schematic overview of the song system. HVC, proper name; RA, robust nucleus of the arcopallium; LMAN, lateral magnocellular nucleus of the nidopallium; Av, avalanche; DLM, medial portion of the dorsolateral thalamic nucleus; D, dorsal; P, posterior. (B) Circuit diagram of the song system. Arrowheads and closed circles indicate excitatory and inhibitory connections, respectively. NC, caudal nidopallium; Arco., arcopallium; NR, rostral nidopallium; Stri., striatum. (C) Schematic of SLCR- seq. Fresh- frozen brains were cryosectioned for Laser Capture Microdissection (LCM). Individual sections of regions of interest were collected into wells of 96- well plates, and then total RNA was purified using an optimized solid phase reversible immobilization (SPRI) protocol. After RNA purification, 3’-end sequencing libraries were prepared containing unique molecular identifiers (UMI) using a custom protocol. (D) Left: Experimental overview of SLCR- seq on hearing and deaf birds. After a baseline period of song recording, birds were either deafened through bilateral cochlear removal or underwent a sham surgery. After 4, 9, or 14 days post- surgery, birds were euthanized and SLCR- seq libraries were prepared from HVC, NC, RA, Arco., LMAN, NR, Area X, and Stri. Right: Uniform Manifold Approximation and Projection (UMAP) plot of SLCR- seq data colored by section position. Each point reflects the gene expression profile of a single SLCR- seq sample. Samples show segregation by broad anatomical area — striatal (Area X), nidopallial (HVC, NC, LMAN, NR), arcopallial (RA, Arco.) — and song system nuclei from surrounding areas. (E) Normalized log gene expression data of three example genes — SLC17A6, PVALB, and AR. Each point is gene expression in a single SLCR- seq sample. SLC17A6 is a marker for glutamatergic cells and is distinctly depleted in the striatal samples; PVALB and AR are two genes known to be enriched in song system nuclei. (F) Coronal anatomical atlas representation of the expression of the three genes shown in panel (F). Each region is colored according to the log gene expression value. D, dorsal; L, lateral. The online version of this article includes the following figure supplement(s) for figure 2: Figure supplement 1. Additional validation of Serial Laser Capture RNA- seq (SLCR- seq) data. at both local and circuit levels, we developed a protocol for sample collection and RNA- seq library construction that addresses three goals: (1) precise collection of histologically defined samples, (2) ease of collecting multiple replicates per animal and brain region, and (3) reduced per- sample cost for library preparation and sequencing. This approach, termed Serial Laser Capture RNA- seq (SLCR- seq), combines the anatomical precision of laser capture microdissection (LCM) with the capacity to work with large numbers of low- input RNA samples provided by single- cell RNA- sequencing protocols (Figure  2C, see Methods). Brains were flash- frozen without fixation and then cryosectioned onto slides suitable for LCM. We visualized song nuclei using an optimized rapid Nissl staining protocol, collected single cryosections from regions of interest in 96- well plates using LCM, then purified total RNA using a custom solid phase reversible immobilization protocol. The preparation produces high- quality RNA (RIN = 9.1 ± 0.5) and yields that are sufficient for library preparation (one 20 μm- thick section with an area of 100,000 μm2 yields 1–2 ng; RA area is ~125,000 μm2). From this total RNA, we then prepared 3’-localized RNA- seq libraries containing unique molecular identifiers adapted from protocols previously developed for single- cell RNA- sequencing (Islam et  al., 2014; Kivioja et  al., 2012; Picelli et al., 2014). Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 5 of 37 Chromosomes and Gene Expression | Neuroscience Research article We used SLCR- seq to generate RNA- seq libraries for each bird from eight brain regions — HVC, RA, LMAN, and Area X, and four paired non- song regions — (Figure 2D) with multiple LCM sections (2–6) collected per region per bird, yielding 598 samples after quality control filtering for the number of detected genes in each sample (mean ± s.d. number of sections per region per bird = 4.5 ± 1.5). Gene expression variation across this dataset segregated samples into three broad clusters corre- sponding to the region of origin — arcopallium (RA and Arco.), nidopallium (HVC, NC, LMAN, and NR), and striatum (Area X and Stri.) — consistent with the different functional properties and devel- opmental origins of these regions (Figure 2D). Furthermore, these broad clusters were subdivided into adjacent but distinct song/non- song pairs, reflecting known similarities between song nuclei and adjacent neural regions (Nevue et  al., 2020). To further validate this approach, we inspected the SLCR- seq expression values of genes with known variation across the songbird brain or enrichment in the song system (Figure 2E and F and Figure 2—figure supplement 1). The glutamatergic neuron marker SLC17A6 was strongly depleted in the striatal regions Area X and Stri, consistent with the rela- tive scarcity of excitatory neurons in these regions. Genes with known variation in expression across the system, e.g., parvalbumin (PVALB) and androgen receptor (AR), corroborated a strong correspon- dence between SLCR- seq expression values and in situ hybridization signal intensities (Lovell et al., 2020; Figure 2E and F and Figure 2—figure supplement 1). Song system-wide transcriptional signatures of song destabilization To provide a single statistic that reflects the extent of song change for each bird, we used a previously developed method (Mets and Brainard, 2018) that builds statistical models of song in two conditions (e.g. pre and post- procedure) and calculates the distance between probability distributions gener- ated from these models using Kullback- Leibler divergence (‘Song DKL,’ see Figure 3—figure supple- ment 1A and Methods). Here, higher values indicate a larger divergence of post- procedure songs compared to pre- procedure songs, therefore providing a measure of song change from baseline. The UMAP quantification used in Figure 1 was used to summarize this condensed visual representation into a single statistic. However, we choose Song DKL for subsequent analysis over the UMAP- based quantification because Song DKL is a previously validated statistical modeling approach that robustly captures alterations to a wide range of song types (Mets and Brainard, 2018). For each bird, we calculated Song DKL between songs recorded during the two days before the procedure (deafening or sham) and those recorded on the day of euthanasia and the preceding day. Deafening resulted in a significant increase in Song DKL relative to sham (Figure 3A, hearing 0.14 ± 0.041 log Song DKL mean ± SEM; deaf 0.52 ± 0.085 mean ± SEM, two- sided Wilcoxon rank- sum test p=5e−4). We did not include the number of days from procedure (sham or deafening) as an explicit variable in subse- quent analyses since the Song DKL measure more directly captured the amount of alteration to song. Singing influences gene expression in the song system (Feenders et al., 2008; Horita et al., 2012; Jarvis et al., 1998; Sasaki et al., 2006; Wada et al., 2006; Warren et al., 2010; Whitney et al., 2014; Whitney and Johnson, 2005), and previous work has indicated that recent singing influences song plasticity and variability (Chen et al., 2013; Hayase et al., 2018; Hilliard et al., 2012; Miller et al., 2010; Ohgushi et al., 2015). Therefore, we also included terms for the number of songs sung on the day of euthanasia and the average number of songs sung per day in the pre- procedure period to control for constitutive differences in singing propensity. These values varied widely across birds (Figure 3B) but did not differ significantly between hearing and deafened birds (number of songs on date of euthanasia: hearing 52 ± 18 mean ± SEM, deaf 55 ± 17, two- sided Wilcoxon rank- sum test p=0.7; pre- procedure songs/day: hearing 332 ± 39 mean ± SEM, deaf 339 ± 43, two- sided Wilcoxon rank- sum test p=1). We used multiple regression to identify genes whose expression varied with song destabilization (Song DKL) (Figure  3C). A subset of hearing and deaf birds showed overlapping Song DKL values (three birds in each condition). To detect genes with expression differences associated with song destabilization, we compared birds in each group that had Song DKL values outside of this overlap- ping range (Figure 3A, six hearing birds with ‘low’ Song DKL values, and six deaf birds with ‘high’ Song DKL values). In general, birds that had been deafened for longer (9 and 14 days) had Song DKL values outside of the hearing Song DKL range while those deafened for less time (4 days) showed less song destabilization. Exceptions to this pattern include one 4- day- deafened bird that showed partic- ularly strong song destabilization and one 9- day- deafened bird that showed modest song change. Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 6 of 37 Chromosomes and Gene Expression | Neuroscience Research article Figure 3. Song destabilization is associated with song system- wide alterations to gene expression. (A) Relative spectral distance between syllables pre- and post- procedure, represented as the mean Kullback- Leibler (KL) distance between Gaussian mixture models or ‘Song DKL’ (see Methods for calculation). Song DKL trends higher with increasing days from deafening. Significance was calculated using a two- sided Wilcoxon rank- sum test. (B) (left) Number of songs sung on euthanasia date (i.e. within the two hour period between lights- on and euthanasia) and (right) log mean number of songs sung per day pre- procedure for each bird grouped by hearing or deaf. Significance was calculated using a two- sided Wilcoxon rank- sum test. (C) Differential expression analysis of song destabilization. Multiple regression using voom/limma provided estimates of gene expression fold change with variation in song deviation (Song DKL), number of songs on the day of euthanasia, and baseline differences in singing rate. (D) Differential expression gene (DEG) scores from gene expression regressions. Positive values reflect genes with increased expression, while negative values indicate genes with reduced expression. Scores are the sum of the –1 * log10(adjusted p- values) of high vs. low Song DKL regression coefficients. Each score is multiplied by the sign of the coefficient to obtain a signed value. Separate coefficients were estimated for each neural region. (E) Volcano plots of –1 * log10(adjusted p- values) versus the log fold- change of gene expression in RA and Area X in high vs. low Song DKL birds. Signed adjusted p- values above five were assigned values of five to aid visualization. Labeled are the 10 genes with the highest signed adjusted p- value. ‘ncRNA- 1’ accession is LOC116184561, ‘ncRNA- 2’ accession is LOC116183441. (F) Similarity of song destabilization differential gene expression across the song system and surrounding regions. Heatmaps show the -log10(p- value) from hypergeometric tests comparing the expected versus observed overlap of the top 250 differentially expressed genes for each compared Figure 3 continued on next page Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 7 of 37 Chromosomes and Gene Expression | Neuroscience Research article Figure 3 continued region, divided into genes with increased expression with song destabilization (lower left triangle) and those with decreased expression (upper right triangle). (G) Distribution of values from (F) comparing song versus song regions, song versus non- song regions, and non- song versus non- song regions. Box middle is the median, box upper and lower bounds are the 25th and 75th percentile, and whisker ends lie at 1.5 times the inter- quartile range. (H) Gene set enrichment analysis (GSEA) of song destabilization- associated genes. Shown are the Gene Ontology (GO) terms that are significant in at least one song or non- song region (adjusted p- value <0.1, see Methods). Heatmap represents the signed log10(adjusted p- value) for each GO term and region, with the sign indicating that a given term is associated with increased or decreased expression in Song DKL high versus low birds. Terms are ordered by hierarchical clustering (euclidean distance, Ward squared method). Representative terms are listed for each cluster. (I) Song destabilization gene expression responses of genes in two gene sets — ‘DNA binding transcription activator’ (GO:0001216) and ‘Hormone activity’ (GO:0005179) — that have differential expression across song regions. Shown are the top 20 leading edge genes from GSEA (gray lines) and the top six of these are labeled at right. The mean expression change for these 20 genes is shown as a colored line in each panel. The online version of this article includes the following figure supplement(s) for figure 3: Figure supplement 1. Extended analysis of song destabilization- associated gene expression. To identify regional patterns of differential expression, we used a differential expression gene (DEG) score that incorporates the number of differentially expressed genes and their adjusted p- values (Figure  3D and Supplementary file 2, see Methods). Song regions generally showed greater expression differences than non- song regions for both Song DKL and singing- rate (Figure  3D and Figure 3—figure supplement 1B, C). Across the song regions, the largest changes to gene expres- sion between high and low Song DKL occurred for the forebrain motor output nucleus RA and the striatal component of the song system Area X (Figure 3D, red and green bars). For these regions, the most significantly modulated genes (adjusted p- value <0.1) were equally likely to be upregulated versus downregulated in deafened versus hearing birds (Figure 3E; for RA, 14 genes upregulated and 11 genes downregulated; for Area X, 15 genes up, 16 genes down). Among the most highly upregulated genes in RA were the plasticity- associated gene protein kinase C β (PRKCB) (Chu et al., 2014; Fioravante et al., 2014), whose protein levels were previously shown to be upregulated in RA following deafening (Watanabe et al., 2002); the microtubule- destabilizing protein stathmin 1 (STMN1), which has roles in long term potentiation and fear memory formation (Shumyatsky et al., 2005); CD24, a surface protein that influences neurite extension (Gilliam et al., 2017); and the lipid processing enzyme lipoprotein lipase (LPL), which is implicated in memory formation and Alzhei- mer’s disease pathology (Wang and Eckel, 2012; Yu et al., 2015). Likewise, among the most down- regulated genes in RA were secreted neuromodulatory proteins including corticotropin- releasing hormone binding protein (CRHBP), somatostatin (SST), and insulin growth factor 2 (IGF2), which each have described roles in regulating neuronal physiology and neural plasticity (Chen et al., 2011; Li et al., 2016; Song et al., 2021). To further examine the neural- circuit- wide structure of gene expression across the song system and surrounding regions, we pairwise intersected the lists of the top Song DKL differentially expressed genes (250 genes with the lowest p- values) for each region and calculated the degree of overlap using a hypergeometric test (Figure 3F and G). Song destabilization- associated differential gene expression was more similar between song regions than between both song and non- song pairs and non- song regions with each other (Figure  3G), indicating that the song system exhibits, in part, a common transcriptional response to song destabilization that is not shared in adjacent regions. We performed gene set enrichment analysis (GSEA) of differentially expressed genes to identify pathways that show coherent gene expression responses to song destabilization (Figure 3H and Supplementary file 3). Several pathways exhibited similar expression responses across all four song regions, including those related to transcription regulation, glia differentiation, and hormone activity (Figure  3H1). Genes related to synaptic transmission were differentially expressed across multiple pallial regions, including song regions RA, HVC, and LMAN as well as the non- song region NCL. Neuron spine- associated genes were upregulated across RA, Arco., and HVC, consistent with previous reports of altered spine dynamics in the song motor pathway following deafening (Peng et al., 2012a; Peng et al., 2013; Tschida and Mooney, 2012; Zhou et al., 2017). Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 8 of 37 Chromosomes and Gene Expression | Neuroscience Research article Correlated gene modules associated with song destabilization The regression analysis described in Figure 3 identified differential expression at the level of individual genes but may have missed subtler expression responses that are correlated across multiple genes. To better identify groups of differentially expressed genes with similar responses to song destabi- lization, we leveraged the large sample numbers of the SLCR- seq dataset to perform gene- gene correlation network analysis for each region separately using MEGENA (Multiscale Embedded Gene Co- expression Network Analysis), an approach that generates sparse networks of covarying genes by applying a topological constraint to co- expression networks (Song and Zhang, 2015; Figure 4A and Figure 4—figure supplement 1). Using this method, we constructed gene- gene correlation networks for each song and non- song region separately, combining SCLR- seq samples across all birds, both hearing and deafened. Mapping song destabilization fold- change of each gene onto the RA network showed a segregation between genes with increased and decreased expression, indicating that expression differences associated with song destabilization state are prominent drivers of network structure (Figure  4B). This segregation was also seen for the HVC, LMAN, and Area X networks (Figure  4—figure supplement 1A). MEGENA employs a hierarchical module detection algorithm to identify correlated sets of genes at different levels of resolution (Supplementary file 4). For each module in each region’s network, we averaged high- vs- low Song DKL regression coefficients for its member genes and compared these observed mean values to a shuffled distribution of mean values, generated by sampling the same number of module genes across the network at random. Several modules in each region’s network showed significantly higher or lower mean fold- change relative to a distribution of mean fold- changes from sets of randomly selected genes (100 random samplings of genes from the network, shuffled p- value <0.01, Figure 4C and Figure 4—figure supplement 1B). To assess the similarity between modules in the correlation networks of one brain region and those in the networks of other brain regions, we calculated a module preservation score (see Methods) and found that RA destabilization- associated modules were preserved to different degrees in networks for the other song regions. In addition, several RA modules showed similar response patterns in other song regions, such that modules upregulated in RA were upregulated in HVC, LMAN, and Area X and likewise for downregulated modules (Figure 4D and E and Figure 4—figure supplement 1B). This pattern is consistent with the overall similarity in differential expression seen among song regions using the regression analysis described in Figure 3 (Figure 3F). Gene set enrichment analysis indi- cated that differential RA modules are enriched for a range of biological pathways (Figure 4F and Figure 4—figure supplement 1C). Notably, the top downregulated module (M74) was enriched for secreted proteins, such as CRHBP, SST, and CHGB (Figure 4G). Upregulated modules were enriched for genes involved in development, morphogenesis, and gene regulation, including PBX1, NR2F2, ZHX3, and ANKRD11. Cell type expression of song destabilization-associated genes After establishing a circuit- wide view of gene expression responses to song destabilization, we inves- tigated the cellular specificity of these responses to understand what cell classes exhibit the most substantial transcriptional changes and may play a role in deafening- induced song plasticity. To do so, we integrated the SLCR- seq data with a previously generated single- nucleus and single- cell RNA- sequencing dataset from HVC and RA of hearing adult male finches (Colquitt et al., 2021; Figure 5A). In that work, we compared songbird neuronal classes in HVC and RA to those in mammals and identi- fied a high degree of transcriptional similarity across several neuronal classes (Figure 5B). For each gene, we computed a cell type destabilization score — the product of a gene’s cell type specificity with its fold- change between high and low Song DKL — to assay cellular biases of destabilization- associated expression (Figure 5C and D, and Figure 5—figure supplement 1A, B). In RA, which showed the strongest transcriptional changes as described above (Figure 3D), differentially expressed genes were most strongly localized to neurons. In particular, genes with reduced expres- sion during song destabilization, such as CRHBP, SST, NPY, and CHGB, showed a bias toward Sst- and Pvalb- class interneurons (GABA- 2/3/4). In addition, several upregulated genes, such as PRKCB, EPHB1, and DNM1, were biased toward RA glutamatergic neurons. HVC showed a similar pattern of cell- type expression, with genes that had reduced expression biased toward Sst- class interneurons as well as LGE- class GABA- 1 and MGE- class GABA- 7 interneurons (Figure 5—figure supplement 1A and B). These cellular expression biases could arise from increases or decreases in the abundances of Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 9 of 37 Chromosomes and Gene Expression | Neuroscience Research article Figure 4. Correlated modules of gene expression associated with song destabilization. (A) To identify correlated patterns of gene expression, gene- gene correlation networks were constructed for each region using MEGENA (Multiscale Embedded Gene Co- expression Network Analysis). These networks were then used to identify correlated sets of gene modules. Estimated regression coefficients were mapped onto correlation networks to identify covariate- associated expression modules. (B) Gene- gene correlation network for RA. Each node is colored by the log fold- change expression between deaf and hearing birds. (C) Average song destabilization gene expression changes for each RA module. Error bars are null distributions generated by repeatedly sampling the network (100 times) for the number of nodes in a given module and then averaging their high vs. low Song DKL fold- changes. Dots that are colored have mean coefficient values that are lower or higher than 1% or 99% of the sampled distribution, respectively. (D) Average change in expression of RA modules across each song system and non- song system region. (E) Preservation scores for RA modules in the correlation networks of other song and non- song system regions. Only significant values are shown (Bonferroni p- values <0.01), and values are scaled to the maximum and minimum for each module to show relative levels of preservation across regions. (F) Gene set enrichment analysis (GSEA) of RA modules with significant gene expression alteration with song destabilization. Mean fold- change values for each module, as represented in (B), are shown at the top of the GSEA plot. Shown are at most the top five significant Gene Ontology (GO) terms (GSEA adjusted p- values <0.2). (G) Network diagrams for three modules (M61, M60, and M74) that show large deviations with song destabilization. Labeled Figure 4 continued on next page Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 10 of 37 Chromosomes and Gene Expression | Neuroscience Research article Figure 4 continued and highlighted are selected hub genes for each module (see Methods for classification). Node colors indicate log fold- change expression between deaf and hearing birds (scale given in (A)). The online version of this article includes the following figure supplement(s) for figure 4: Figure supplement 1. Network analysis of song destabilization- associated gene expression. defined cell populations. To determine if these cellular biases do reflect a bulk loss of particular cell types, we analyzed the differential expression of marker genes for each cell type (Figure 5—figure supplement 1C). In both song nuclei, markers for each neuronal cell type (see Methods for definition) showed no significant difference between high and low Song DKL conditions (median of fold- changes Figure 5. Cell- type specificity of destabilization- modulated genes. (A) Schematic of the approach to determine the cell type expression biases of genes that are differentially regulated with song destabilization. A previously generated cell- resolved gene expression dataset for RA (robust nucleus of the arcopallium) and HVC (proper name) (Colquitt et al., 2021) was combined with RA and HVC song destabilization regression coefficients from this study to compute a cell- type bias score (see Methods). Shown also is a uniform manifold approximate projection (UMAP) plot of the full dataset with major cell type groups indicated. OPC, oligodendrocyte precursor cell. (B) Schematic of neuronal cell types in the song motor pathway, as previously defined in Colquitt et al., 2021. HVC glutamatergic neurons are broadly similar to intratelencephalic (IT) mammalian neocortical neurons from multiple layers, and RA neurons are similar to extratelencephalic (ET) neurons from layer 5. Eight primary GABAergic clusters are found equally in both HVC and RA and are organized into clusters corresponding to subpallial regions of origin — lateral, medial, and caudal ganglionic eminences (LGE/MGE/CGE). The LGE- class GABA- 1 has no known correspondence with mammalian neocortical neurons; GABA- 2 is transcriptionally similar to Sst- class neurons, GABA- 4 is similar to Pvalb- class neurons, and GABA- 3 is transcriptionally intermediate between GABA- 2/4. (C) Integration of cell type specificity scores and song destabilization differential expression to identify cell type- associated transcriptional effects of song destabilization. For the top 50 differentially expressed genes, cell type specificity was multiplied by log fold- change between high and low Song DKL birds. These values were then split by sign then summed within each cell type to yield a cell type Song DKL score. Gray bars indicate the distribution (1–99%) of Song DKL- cell type specificity scores for 100 random sets of 50 genes. (D) RA cell type specificity scores for top Song DKL differentially expressed genes, divided into upregulated and downregulated genes. Values are scaled for each gene such that the cell type with the highest specificity score equals 1 and that with the lowest equals 0. At the top of each specificity score heatmap is the log fold- change expression for high vs. low Song DKL. The online version of this article includes the following figure supplement(s) for figure 5: Figure supplement 1. Cell type- associated differential expression. Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 11 of 37 Chromosomes and Gene Expression | Neuroscience Research article greater than 99% or less than 1% of medians from randomly selected genes), indicating that the cell type biases of destabilization associated- genes are not due to changes in cell type abundance, but rather to the expression levels of specific genes within defined cell classes. Inter-region correlation of gene expression is reduced in deafened birds The foregoing analyses focused on comparing gene expression responses to deafening that are local to each region. However, the song system is an interconnected neural circuit, and gene expression in one region could be correlated with that in others due to shared patterns of neural activity, common responses to hormonal signaling, or to baseline expression differences across regions that vary in a concerted fashion across individuals. By similar logic, manipulations such as deafening that could disrupt global patterns of neural or hormonal signaling might result in alterations in the patterns of inter- region correlations in gene- expression. To determine whether and how deafening alters inter- region correlations in gene expression, we first identified genes that have correlated expression between brain regions across birds. Briefly, for each gene, we calculated correlation values for each pairwise combination of brain regions, yielding region- by- region correlation matrices (Figure 6A). We identified significant correlations as those that were less than the 2.5% quantile or greater than the 97.5% quantile of a shuffled distribution (see Methods). We calculated the across- bird gene expression similarity between regions as the number of thresholded correlations. This analysis revealed several notable relationships among brain regions. First, each song nucleus had the highest gene correlation strength with its paired non- song region (with the exception of HVC which had generally weaker correlation strengths with other regions), consistent with the shared molecular profiles of each song nucleus with the surrounding tissue. Second, the nuclei of the vocal motor pathway (HVC and RA) and anterior forebrain pathway (LMAN and Area X) were more correlated with each other than with nuclei in the other pathway. Third, normalizing correlation strength for each song nucleus recovered known connections between nuclei (Figure 6B): HVC displayed strong correlations with its target RA, and LMAN was strongly correlated with both of its direct targets, RA and Area X. Interestingly, we found relatively weaker gene correlation strength between HVC and its target Area X. To identify genes that have shared patterns of inter- region correlation across multiple song nuclei, we next clustered genes by the similarity of the correlations between song nuclei known to be directly connected, HVC- RA, LMAN- RA, HVC- X, and LMAN- X (Figure  6C). This analysis generated a diver- sity of patterns with most genes showing correlated expression among the three pallial song nuclei, HVC- RA and LMAN- RA (cluster 3). Gene set enrichment analysis indicated that this cluster is enriched for genes that are associated with signaling receptor binding and that are responsive to neural activity (Figure  6D). Indeed, the genes most strongly associated with HVC- RA and LMAN- RA correlations included the activity- dependent genes CRHBP, NR4A3, and NRN1 (Figure 6E). We then assessed how deafening alters gene expression correlations across the song system. To do so, we computed pairwise correlations for each gene between each region for hearing and deaf birds separately, then computed a differential matrix comparing absolute correlations in deaf birds to those in hearing birds (Figure 6F–H). Differentially correlated genes were defined as those with a deaf versus hearing value less than (decorrelation) or greater than (correlation gain) the extreme values of a shuffled distribution calculated for each pairwise comparison (2.5% or 97.5%, respectively). Overall, each directly connected pair of song regions had a greater number of genes with reduced correlation in deaf versus hearing birds than increased correlation (Figure  6F and G and Supplementary file 5). Two of the most strongly decorrelated genes highlight this effect. Expression of the neurotrophic factor BDNF was positively correlated between LMAN and RA in hearing birds but was uncorrelated in deaf birds; similarly, expression of the nuclear receptor PPARG was negatively correlated between LMAN and Area X in hearing birds but was uncorrelated in deaf birds (Figure 6H). Loss of afferent input to the motor pathway affects the expression of song destabilization-associated genes The output nucleus of the anterior forebrain pathway, LMAN, is required for adaptive plasticity to song and moment- by- moment song variability and is one of the two major afferents to the motor nucleus RA (Andalman and Fee, 2009; Kao et al., 2005; Nottebohm et al., 1982; Olveczky et al., 2005; Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 12 of 37 Chromosomes and Gene Expression | Neuroscience Research article Figure 6. Inter- region gene expression correlation and decorrelation with song destabilization. (A) Inter- region gene expression correlations across song and non- song regions. For each region, the 500 genes with the highest variability across birds were selected (see Methods), then the expression of each gene was correlated across regions. Significant genes were called as those with an observed correlation less than 1% or greater than 99% of a shuffled correlation distribution (100 shuffles, calculated for each pairwise comparison between regions). The number of significantly correlated genes was Z- scored across the set of pairwise comparisons to highlight the relative strength of inter- region expression correlations. Within- region correlations were excluded from the Z- scoring and are colored gray. (B) Representation of the data in (A) showing the strength of gene expression correlation between each song system nucleus and other assayed regions. (C) Patterns of inter- region gene expression correlations. Genes were clustered into eight clusters by their pairwise correlation values between HVC- RA, LMAN- RA, HVC- Area X, and LMAN- Area X. Heatmap shows mean correlations within each cluster (rows) and region comparison (columns). Barplots represent the number of genes in each row or column. Highlighted is cluster 3, the HVC- RA and LMAN- RA correlation cluster, which has the greatest number of genes. (D) Gene set enrichment analysis (GSEA) indicates that cluster 3 is enriched for genes that are activity- dependent and have a signaling- related function. (E) Expression of three cluster 3 genes across pairs of song system regions with direct projections — HVC to RA, LMAN to RA, HVC to Area X, and LMAN to Area X. Each point is the z- scored expression estimate for each nucleus in one bird. (F) Schematic of inter- region gene expression differential correlation analysis between hearing and deaf birds. For each gene, region pair, and condition, a Pearson correlation was calculated, then a differential correlation was calculated as the difference between unsigned correlations for hearing versus Figure 6 continued on next page Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 13 of 37 Chromosomes and Gene Expression | Neuroscience Research article Figure 6 continued deaf conditions. To determine significance, 100 random permutations of the expression data were made for each gene, region pair, and condition and differential correlation was computed in the same manner as for the observed values. Genes with observed differential correlations in the top or bottom 2.5% of the shuffled distribution were considered significant. (G) Analysis of inter- region gene expression correlations compared between hearing and deaf birds. Heatmap shows the number of genes that became decorrelated (bottom- left) or gained correlation (top- right) in deaf versus hearing birds. (H) Total number of genes that show significant correlation gain or decorrelation in the four pairs of assayed song system regions with direct projections. (I) Examples of genes that show decorrelation in deafened birds relative to hearing birds. BDNF (brain- derived neurotrophic factor) expression is correlated between RA (y- axis) and LMAN (x- axis) in hearing birds but shows no inter- region correlation in deafened birds. Similarly, PPARG expression is correlated between LMAN (y- axis) and Area X (x- axis) in hearing birds but is uncorrelated in deaf birds. Each point is the z- scored expression estimate for one bird. Warren et al., 2011; Williams and Mehta, 1999; Figure 2A and B). Lesions of this nucleus result in reduced song variability (Kao and Brainard, 2006) and, when performed before cochlear removal, prevent song destabilization (Brainard and Doupe, 2000), indicating that deafening generates plas- ticity signals that require inputs from LMAN. We hypothesized that lesions of LMAN would establish a molecular state in RA similar to that found in other low variability and low plasticity conditions, such as that in normal hearing adult birds (versus deaf birds). To assess LMAN’s influence on gene expres- sion in the song motor pathway, we unilaterally lesioned LMAN in five adult male birds (Figure 7A and Figure 7—figure supplement 1). Unilateral LMAN lesions did not grossly alter song, and song stability as measured by Song DKL was similar to that for unlesioned hearing birds (Figure 7—figure supplement 1). Sixteen days after lesioning, we collected HVC, NCL, RA, Arco., and the primary audi- tory area Field L for SLCR- seq (Figure 7B, 91 libraries total). Field L, a region that is easily identifiable using the rapid Nissl stain protocol used in SLCR- seq, was added here to provide a control region that was outside of the song motor pathway. Song regions from each hemisphere were collected independently to allow within- bird comparisons between regions ipsi- and contralateral to the lesion. LMAN was not substantially lesioned in one bird (lesion extent ~0% of LMAN volume, see Methods), and samples from each hemisphere for this bird were treated as unlesioned. Unlike mammals, birds do not have an interhemispheric connection at the level of the forebrain, such that there is no direct connectivity between song system nuclei across hemispheres (Nottebohm et al., 1982; Nottebohm et al., 1976). We reasoned that gene expression modulated directly by LMAN activity would show specific effects in its direct target RA relative to regions that do not receive direct afferents from LMAN such as HVC and surrounding regions that are not part of the song system (Arco., NCL, and Field L). For each brain region, we performed comparisons between the region ipsilateral to the LMAN lesion to that in the contralateral hemisphere (Figure 7C and Supplementary file 6). As expected, RA exhibited the greatest expression changes between sides ipsilateral and contralateral to the lesion (35 genes with reduced and 40 genes with increased expression in ipsilateral, adjusted p- value <0.1) compared to ipsilateral to contralateral comparisons of non- direct targets of LMAN (Figure 7C and D). Genes that were more highly expressed ipsilateral to the lesion were enriched for immune- responsive genes likely reflecting an injury response in RA to the afferent lesion (Figure 7D and E). In contrast, genes with reduced expression ipsilateral to the lesion were enriched for a range of biolog- ical processes, including activity- dependent delayed primary response genes (Tyssowski et al., 2018), neuron cellular homeostasis, metalloendopeptidase activity, and potassium channels (Figure 7E). To examine more broadly how these expression alterations compared to those associated with deafening- induced song destabilization, we calculated the average ipsilateral versus contralateral fold change for the destabilization- associated gene modules described in Figure 4 (Figure 7F). If LMAN lesions impose a molecular state associated with low variability and low plasticity, we would expect to see an inverse pattern of expression between Song DKL and lesion differential expression. Indeed, on the whole, modules that had increased expression in RA with higher Song DKL had lower expression ipsilateral to the lesion, and vice versa (Figure 7G). However, one module, M74, showed an opposite pattern — it was the most strongly reduced module both with increased song destabilization and with LMAN lesions. M74 hub genes CRHBP and SST were reduced specifically in RA ipsilateral to the lesion and showed no change in other assayed regions (Figure 7H and Figure 7—figure supplement 2A–C). This module is enriched for secreted neuropeptides, and the similarity of its expression change Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 14 of 37 Chromosomes and Gene Expression | Neuroscience Research article Figure 7. Loss of afferent input to the motor pathway nucleus RA (robust nucleus of the arcopallium) alters destabilization- associated gene expression. (A) Schematic of unilateral LMAN (lateral magnocellular nucleus of the anterior nidopallium) lesions and sample collection for Serial Laser Capture RNA- seq (SLCR- seq). Five birds received unilateral LMAN lesions (three left and two right hemisphere). After 16 days, birds were euthanized, and HVC (proper name), NC, RA, arcopallium (Arco.), and the primary auditory region Field L were collected for SLCR- seq. Each hemisphere was processed separately to examine the ipsilateral versus the contralateral influence of LMAN lesioning on gene expression in hearing birds. (B) UMAP plot of SLCR- seq samples colored by region. (C) Within- bird differential expression analysis of the influence of LMAN lesions on different regions of the songbird brain. Shown are differential expression gene (DEG) scores for each region assayed. Positive values reflect genes with increased expression, while negative values indicate genes with reduced expression. DEG scores are calculated as the sum of the –1 * log10(adjusted p- values) of regression coefficients for gene expression in brain regions ipsilateral versus contralateral to the LMAN lesion. Each score is multiplied by the sign of the coefficient to obtain a signed value. Separate coefficients were estimated for each neural region. (D) Volcano plot showing the genes that had the most significant difference in expression (red points) between RA on the ipsilateral versus the contralateral side of LMAN lesion, quantified as –1 * log10(adjusted p- values) versus the log fold- change of gene expression for RA ipsilateral versus contralateral to the LMAN lesion. Signed adjusted p- values above five were assigned values of five to aid visualization. (E) Gene set enrichment analysis of differential expression in RA ipsilateral versus contralateral to the LMAN lesion side. Top leading edge genes are listed at right. NES, normalized enrichment score. (F) Average expression change between ipsilateral and contralateral RA for each Song DKL module that was identified in Figure 4. Error bars are null distributions generated by repeatedly sampling the network (100 times) for the number of nodes in a given module and then averaging their differential ipsilateral versus contralateral coefficients. Dots are colored that have mean coefficient values that are lower or higher than 1% or 99% of the sampled distribution, respectively. (G) Comparison of mean fold- change expression differences in RA between high- vs- low Song DKL and contralateral versus ipsilateral to LMAN lesions. Blue line indicates linear regression through the data after excluding outlier modules M74 and M61. (H) Expression of two M74 hub genes, CRHBP and SST, between RA and Arco contralateral or ipsilateral to the LMAN lesion. Each dot is the estimated gene expression within a given bird and region, and error bars are standard errors of this estimate. Adjusted p- values were obtained from the ipsilateral versus contralateral regression analysis. The online version of this article includes the following figure supplement(s) for figure 7: Figure supplement 1. Validation and quantification of unilateral lateral magnocellular nucleus of the anterior nidopallium (LMAN) lesions. Figure supplement 2. Validation of unilateral lateral magnocellular nucleus of the anterior nidopallium (LMAN) lesion effects on robust nucleus of the arcopallium (RA) gene expression. Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 15 of 37 Chromosomes and Gene Expression | Neuroscience Research article across both deafening and LMAN lesions could reflect its sensitivity to altered neural activity in RA, either through the loss of auditory input or through the direct loss of a major afferent to RA. We integrated ipsi- vs- contralateral lesion differential expression with cell type specificity, as described above in the deafening analysis, to examine the cellular expression biases of genes that are influenced by the loss of LMAN. Upregulated genes were primarily expressed in non- neuronal cells and in particular in microglia and astrocytes, consistent with an injury response (Figure 7—figure supplement 2D and E). Indeed, marker genes for microglia show a strong increase in expression in RA ipsilateral to LMAN lesions, suggesting an increase in microglia abundance in RA (Figure 7—figure supplement 2F). This bias is consistent with a glial injury response to the lesion or alternatively may reflect glia- mediated synaptic plasticity. In contrast, downregulated genes were largely expressed in neurons (Figure 7—figure supplement 2D and E), namely glutamatergic projection neurons (Glut- 1) and MGE- derived GABAergic interneurons such as Sst- class (GABA- 2), Pvalb- class (GABA- 4), and cholinergic neurons (GABA- 8). Discussion Sensory feedback is necessary for the reliable and successful execution of learned motor skills, and its loss can lead to increased motor errors and aberrant motor plasticity. The deprivation of sensory expe- rience has been used effectively to characterize plasticity within sensory systems and its underlying cellular and molecular mechanisms. In contrast, how sensory deprivation drives plasticity in associated sensorimotor circuitry at cellular and molecular levels is comparatively poorly understood. Here, we used the experimental advantages of birdsong, a highly precise learned motor skill that has a dedi- cated neural circuitry, to identify molecular pathways in sensorimotor circuits that are influenced by the loss of auditory input and associated vocal motor destabilization. This model has particular relevance for understanding the neural basis of speech alterations caused by deafening that occurs after speech acquisition (post- lingual deafening). Similar to the effects of auditory feedback loss to birdsong, post- lingual deafening in humans reduces the rendition- to- rendition precision of speech production (Lane and Webster, 1991; Waldstein, 1990) and alters spectrotemporal features of speech (Lane and Webster, 1991; Schenk et  al., 2003). Finally, the deterioration of both speech and birdsong is more extreme when deafening occurs at earlier ages, suggesting that there are similar age- dependent mechanisms of vocal stabilization in both systems (Brainard and Doupe, 2001; Cowie et al., 1982; Lombardino and Nottebohm, 2000; Waldstein, 1990). It is an open question how the different components of speech production neural circuitry respond to hearing loss at various biological levels, from molecular to physiological. For the songbird, prior studies have identified a variety of circuit, cellular, and molecular mecha- nisms that may contribute to deafening- induced song- destabilization (Brainard and Doupe, 2000; Kojima et al., 2013; Mandelblat- Cerf et al., 2014; Mori and Wada, 2015; Peng et al., 2012a; Peng et al., 2013; Peng et al., 2012b; Scott et al., 2000; Tschida and Mooney, 2012; Wang et al., 1999; Watanabe et  al., 2002; Zhou et  al., 2017). These prior demonstrations, which have focused on a disparate set of song control structures and specific candidate mechanisms, motivated our interest in applying a circuit- wide and unbiased approach in this system to identify molecular responses to auditory deprivation- induced motor destabilization. Understanding these responses in the songbird vocal control system could provide insight into the neural mechanisms underlying the plasticity and resilience of both learned vocalizations and other well- learned motor skills. Molecular localization of song destabilization Past work on the neural mechanisms underlying song plasticity has largely focused on changes occur- ring in one or two brain regions at a time. Song destabilization in adult songbirds is associated with a variety of changes to the morphology and physiology of neurons in the song system including changes to dendritic spine stability and synapse densities in HVC and RA (Tschida and Mooney, 2012; Zhou et al., 2017); alterations to song tuning responses in LMAN (Roy and Mooney, 2007) and decreased synaptic inputs onto and increased intrinsic excitability of HVC projection neurons (Hamaguchi et al., 2014; Tschida and Mooney, 2012). Our neural circuit- wide analysis of gene expression responses to deafening allowed us to investigate which regions of the song system show the strongest transcrip- tional changes during song destabilization, providing a readout of the molecular correlates of neural Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 16 of 37 Chromosomes and Gene Expression | Neuroscience Research article plasticity. We found that the motor output nucleus RA showed the highest differential expression upon song destabilization, with substantial changes also found in Area X and to a lesser extent in HVC and LMAN. RA lies at the nexus of the motor pathway and the anterior forebrain pathway, a circuit required for song plasticity, and is a major locus of neural plasticity during juvenile song learning and adult song adaptation (Garst- Orozco et al., 2014; Miller et al., 2017; Ölveczky et al., 2011). This position in the song neural circuit makes it well situated to integrate neural activity associated with the stable motor program with AFP- generated contributions to deafening- induced plasticity. Some of the most salient pathways upregulated in RA were associated with synaptic transmis- sion and neuron spines, consistent with previous reports that found that deafening increases synapse densities, spine densities, and spine lengths in RA (Peng et  al., 2012a; Zhou et  al., 2017). These terms were also enriched to some extent for differentially expressed genes in HVC, in which neurons projecting to Area X exhibit decreased spine stability following deafening (Tschida and Mooney, 2012). Past work on several of the top differentially expressed genes in RA supports a general role for altered synapse and spine dynamics during deafening- induced song plasticity. In particular, stathmin 1 (STMN1) is located at synapses and binds to tubulin to inhibit microtubule formation (Curmi et al., 1997; Shumyatsky et al., 2005). Knockout of STMN1 in mice results in impaired long- term poten- tiation in the amygdala and reduced memory in fear- conditioning tasks (Shumyatsky et al., 2005). Furthermore, STMN1 is differentially phosphorylated during fear conditioning, altering its activity and AMPA receptor localization to the synapse (Uchida et al., 2014). Similarly, the surface glycoprotein CD24, which is upregulated in RA, influences neurite outgrowth (Gilliam et  al., 2017) as well as synapse formation and transmission (Jevsek et al., 2006). Lastly, the lipid processing enzyme lipo- protein lipase (LPL) is upregulated in RA during song destabilization. Knockouts of LPL in mice result in impaired learning and memory, decreased presynaptic vesicles in the hippocampus (Xian et  al., 2009), and reduced AMPA receptor expression (Yu et al., 2015). The expression of neuropeptides was also broadly reduced following deafening across multiple song nuclei. This result suggests that song plasticity is a product of not only alterations to synapse structure and neurotransmitter- mediated signaling but also changes in neuromodulation. It has not been well- examined how secreted neuropeptides influence birdsong plasticity and neural activity in birdsong neural circuitry. However, extensive evidence garnered in other systems indicates that neuro- peptide signaling has a powerful effect on neural circuit activity, plasticity, and behavioral output (Bargmann, 2012; Marder, 2011). The specific signaling systems altered following deafening in this study provide a set of candidate mechanisms that may influence song. For example, corticotropin- releasing hormone binding protein (CRHBP) , one of the most strongly downregulated genes in RA following deafening, modulates activity in the CRH signaling pathway (Kemp et al., 1998), which has diverse effects on long- term potentiation, neuronal excitability, and spine dynamics in central circuits (Aldenhoff et al., 1983; Blank et al., 2003; Chen et al., 2008; Fox and Gruol, 1993; Kratzer et al., 2013; Li et al., 2016). Such evidence suggests that the dynamic modulation of neuropeptides could play a prominent role in regulating birdsong stability and plasticity and may similarly influence the control of other stable sensorimotor skills such as human speech. Although we included the amount of singing in the two hours prior to euthanasia (a proxy for neural activity in the song system) as a variable in our regression analysis, we cannot fully dissociate the influences of motor destabilization per se and alterations to neural activity driven by the loss of audi- tory activity. Future work could combine similar circuit- wide gene expression analysis with disruptions to auditory input that do not alter hearing generally (such a delayed auditory feedback [Leonardo and Konishi, 1999]) to further characterize song plasticity- specific expression responses. Similarly, manipulations that induce song plasticity without altering hearing, such as tracheosyringeal nerve cuts (Roy and Mooney, 2007), may help disambiguate motor- vs- auditory expression responses. Ulti- mately, direct manipulations of gene expression in song regions (through knockdown or overexpres- sion) combined with analyses of song destabilization would help clarify the causal roles of candidate genes in promoting or limiting song plasticity. A previous study examined how the loss of auditory input before song learning in juveniles influ- ences gene expression in HVC and RA (Mori and Wada, 2015). In that work, the authors iden- tify a strong gene expression signature that varies with developmental age but is independent of whether the birds are hearing or deaf. Follow- up experiments examined the expression of a subset of developmentally- regulated genes between hearing and adult- deafened birds (similar to the approach Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 17 of 37 Chromosomes and Gene Expression | Neuroscience Research article used here) and found no significant change. That work identified an important separation between developmentally- driven and experience- dependent molecular responses in the song system, but its aims were distinct from the present study, which sought to identify gene expression responses to deafening- induced song plasticity. Neuronal contributors to song plasticity By integrating song system- wide and cell- resolved expression profiles, we can make initial predic- tions about which cell classes exhibit transcriptional changes during song destabilization. Glutama- tergic projection neurons in RA are similar to layer 5 extratelencephalic neurons in the mammalian neocortex, both in terms of their projections to subcerebral structures and their expression profiles (Colquitt et al., 2021; Nevue et al., 2020; Pfenning et al., 2014; Vicario, 1991). A number of differ- entially expressed genes showed biased expression toward glutamatergic neurons in RA, including protein kinase C β (PRKCB), a calcium sensor associated with short- term plasticity (Chu et al., 2014; Fioravante et al., 2014) and previously shown to be upregulated in RA following deafening (Wata- nabe et al., 2002), as well as the lipid processing enzyme LPL, discussed in the previous section of the Discussion. Our results also point to a prominent role for GABAergic interneurons in deafening- induced song plasticity. In particular, genes that had reduced expression with song destabilization showed an expression bias toward Sst- and Pvalb- class interneurons in RA. The interneuron subclasses present in the song system are strongly similar to well- characterized interneuron types in the mammalian neocortex, suggesting deep conservation of inhibitory networks (Colquitt et al., 2021). How these specific subclasses influence network activity in the song system is an open question; however, previous work has established a general role for local inhibition in the regulation of song learning and stability (Kosche et al., 2015; Vallentin et al., 2016). In particular, song learning in juveniles — during which song becomes more structured and less variable — refines the synaptic connectivity between glutamatergic projection neurons in RA and an inhibitory neuron type that has electrophysiological properties similar to fast- spiking Pvalb- class interneurons (Miller et  al., 2017). Similarly, inhibitory input to HVC projection neurons increases and becomes more precise as song performance improves during juvenile song learning (Vallentin et al., 2016). Many of the Sst/Pvalb- biased genes affected by deafening- induced song plasticity are secreted neuropeptides that are sensitive to levels of neural activity (Hou and Yu, 2013; Tyssowski et al., 2018), suggesting that their reduced expression reflects reduced activity in these populations during birdsong destabilization. Moreover, several of these neuropeptides, including SST and CRHBP, act to inhibit neural activity, either directly through receptor binding or indirectly through interactions with other neuromodulators (Hou and Yu, 2013; Li et al., 2016; Pittman and Siggins, 1981). This role of inhibition in maintaining birdsong structure has parallels to the role of inhibition during neural plasticity in mammals. For instance, the density of synapses from Sst- and Pvalb- class interneu- rons onto pyramidal neurons in the mammalian motor cortex is modulated during motor learning in mice, with an overall reduction of Sst- class input during motor plasticity (Chen et al., 2015). Likewise, low Pvalb- class network activity in the hippocampus is associated with increased synaptic plasticity, and low Pvalb expression, itself sensitive to neural activity, is found in the motor cortex during early motor learning (Donato et al., 2013). Similarly, increased neuronal excitability and decreased inhi- bition have also been found in the mammalian auditory cortex following deafening or noise trauma (Kotak et al., 2008; Kotak et al., 2005; Seki and Eggermont, 2003). Together, these results suggest that reduced inhibition, either through altered synaptic transmission or neuromodulation, is a key component of neural plasticity in both the mammalian and avian central nervous systems. Circuit contributions to transcriptional state By sampling gene expression across the different connected components of birdsong neural circuitry in individual birds, our study allowed us to examine the correlation of gene expression in one brain area with that in another. Regions with direct projections to each other, for instance, HVC to RA and LMAN to RA, tended to have a higher number of genes with correlated expression across individuals than song or non- song regions that are not directly connected. Moreover, genes with correlated expression across the three pallial (cortical- like) song regions HVC, RA, and LMAN were enriched for activity- dependent genes and secreted neuropeptides. These results could reflect the presence Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 18 of 37 Chromosomes and Gene Expression | Neuroscience Research article of a shared molecular state across the song system, perhaps established by singing- related neural activity or a common response to a general hormonal factor reflecting some aspect of a given bird’s state (e.g. testosterone levels which may in turn be affected by sensory deprivation [Livingston et al., 2000]). Two results further support that these correlations reflect some aspect of shared activity across regions. First, deafened birds exhibited, on the whole, reduced gene expression correlation between song regions, suggesting that either the loss of auditory information or associated song destabiliza- tion disrupts inter- region coordination of gene expression. Second, lesions of LMAN altered gene expression specifically in RA, one of its primary efferents, but not other brain regions to which it does not directly project. This analysis highlights the importance of considering inter- regional influences in understanding the mechanistic basis of transcriptional responses across an integrated neural circuit. The structure of the song system offers an opportunity to better understand this general issue of how connected neural components mutually influence local properties. In particular, HVC and LMAN converge onto RA and have distinct roles in song production and learning. A number of studies have examined how disrupting these inputs influences RA neural activity, synaptic transmission, neuronal morphology, and cell survival (Akutagawa and Konishi, 1994; Johnson et al., 1997; Johnson and Bottjer, 1994; Kittelberger and Mooney, 1999; Ölveczky et  al., 2011), yet little is known about how each afferent differentially influences the molecular features of RA. A joint analysis of how each afferent alters gene expression across diverse RA neuronal types, using circuit- wide and cell- resolved gene expression approaches such as those described here, could yield insight into how converging neural inputs are integrated in target structures at the molecular level. Our current analysis focused on the transcriptional effects of unilateral LMAN lesions on target structures in hearing birds. We found that expression differences in RA following LMAN lesions were broadly the inverse of those following deafening, suggesting that the loss of LMAN establishes a transcriptional state charac- teristic of reduced plasticity. An informative followup experiment would be to perform bilateral lesions of LMAN before cochlear removal — a manipulation known to prevent deafening- induced song destabilization (Brainard and Doupe, 2000; Kojima et  al., 2013; Scott et  al., 2000) — and compare expression profiles in RA in these birds to those in birds with only bilateral LMAN lesions, only cochlear removal, or unmanipulated controls. We predict that this approach would uncover genes whose expression tracks with song destabilization across manipulations, further pinpointing relevant plasticity- associated molecular factors. These cross- regional patterns of gene expression relate to a central question of this study: how does the loss of sensory input influence gene expression in sensorimotor circuits? Our results suggest a model in which altered activity propagates through existing circuits, such that the state of one circuit component progressively modifies gene expression in its targets. Local mechanisms engaged within each region, such as synapse/spine remodeling and neuropeptidergic signaling implicated here, could then alter circuit connectivity and function, leading to behavioral plasticity. Identifying how neural activity influences gene expression across neural circuits and what specific molecular and cellular factors in turn shape circuit function will be instrumental to better understand the neural mechanisms that underlie sensorimotor stability and its impairment following sensory loss. Key resources table Materials and methods Reagent type (species) or resource Biological sample (Lonchura striata domestica) Designation Source or reference Identifiers Additional information Brain tissue Lab animal colony Sequence- based reagent RT_primer_v1 Sequence- based reagent RT_primer_v2 Continued on next page IDT IDT AAGC AGTG GTAT CAAC GCAGAGTA CNNN NNNN NNNN NNNN NNNNNNNN NNXX XXXX TTTT TTTT TTTT TTTTTTT TTTTTTTTTVN SLCR- seq primer SLCR- seq primer AAGC AGTG GTAT CAAC GCAGAGTA CNNN NNNN NNNN NNNA TCTA GCCGG CCTT TTTT TTTT TTTT TTTT TTTT TTTTTTVN Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 19 of 37 Chromosomes and Gene Expression | Neuroscience Research article Continued Reagent type (species) or resource Sequence- based reagent Sequence- based reagent Sequence- based reagent Sequence- based reagent Designation Source or reference Identifiers Additional information TSO_LNA Exiqon SLCR- seq primer AAGCAGTGGTATCAACGCAG AGTGAATrGrG +G TSO_PCR IDT SLCR- seq primer AAGC AGTG GTAT CAAC GCAGAGT P5- TSO_Hybrid IDT SLCR- seq primer Read1CustomSeqB IDT SLCR- seq primer AATG ATAC GGCG ACCA CCGAGAT CTAC ACGC CTGT CCGC GGAAGCA GTGGTATCAACGCAGAGT*A*C GCCTGTCCGCGGAAGCAGTG GTATCAACGCAGAGTAC Sequence- based reagent PCR2 IDT SLCR- seq primer CAAG CAGA AGAC GGCA TACGAGA TYYY YYYY YGTC TCGT GGGCTCGG Commercial assay or kit Commercial assay or kit Commercial assay or kit KAPA HiFi Hotstart Roche 7958927001 Qubit dsDNA HS ThermoFisher Q32851 Nextera XT Illumina FC- 131–1024 Commercial assay or kit KAPA Library Quantification Kit Roche 07960140001 Commercial assay or kit 2% BluePippin Gels Sage BEF2010 Commercial assay or kit ISH- HCR Molecular instruments ISH probes and reagents Commercial assay or kit AMPure XP Beckman Coulter A63881 Other Other molecule sieve beads Sigma cresyl violet powder Sigma 208582 255246 Used to make anhydrous ethanol solution used in 'SLCR- seq — rapid Nissl stain' Stain used in 'SLCR- seq — rapid Nissl stain' Other Guanidine thiocyanate Sigma G9277 Other Sera- Mag SpeedBeads Carboxyl Magnetic Beads, hydrophobic FisherScientific 09- 981- 123 Other EvaGreen Biotium 13000 Component of lysis solution used for RNA purification in 'SLCR- seq — SPRI RNA purification' Used to create homemade SPRI RNA purification solution, as in 'SLCR- seq — SPRI RNA purification' Dye added to library amplification to determine needed number of amplification cycles, as in 'SLCR- seq — library preparation' Animal care and use All Bengalese finches were from our breeding colonies at UCSF or were purchased from approved vendors. Experiments were conducted in accordance with NIH and UCSF policies governing animal use and welfare. Song recording and preprocessing Birds were individually housed in wire cages in sound isolation chambers. Song was recorded using Countryman Isomax microphones taped to the top of the wire cage. Microphones were connected to USB preamplifiers that were connected to a Linux workstation. Audio was recorded at a frame rate of 44,100 samples/second using a custom python script, and, to select for periods of singing, Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 20 of 37 Chromosomes and Gene Expression | Neuroscience Research article blocks of continuous sound with amplitudes above a manually set threshold were saved as 24- bit WAV files. Song autolabeling The analysis of specific spectral features (e.g. fundamental frequency) was performed on syllables that were labeled using a supervised machine learning approach, called hybrid- vocal- classifier or hvc (Nicholson, 2021). For each bird, 20–50 songs were manually labeled using the Matlab software evsonganaly. Using hvc, a set of spectrotemporal features was computed for each syllable (e.g. dura- tion, mean frequency, pitch goodness, and mean spectral flatness as defined in Tachibana et  al., 2014). These features and the manually defined labels were provided to hvc to train a support vector machine (SVM) with radial basis function, with a grid search across parameters C and gamma to iden- tify parameters with the highest classification accuracy. A set of models were then trained using these selected parameters and a random sample of training syllables, and model accuracy was tested on a held- out set of syllables. For each bird, the number of input syllables and parameters were adjusted until label accuracy reached 95–100%. The model with the highest accuracy was then used to predict labels on unlabeled songs. To select confidently labeled syllables, a prediction confidence score was calculated for each syllable as the entropy ( sklearn. stats. entropy) of the classification probabilities resulting from SVM model prediction. Syllables with a prediction confidence score greater than 0.5 were retained. Song dimensionality reduction To project syllable spectrotemporal structure into a reduced dimension space, we used an approach developed by Sainburg et al., 2020c with code and example scripts obtained from the AVGN Github repository (Sainburg, 2020b). Songs were first isolated from audio recordings and then segmented into syllables based on amplitude threshold crossings. Spectrograms were computed for each syllable using short- time Fourier transforms (512 window size, 0.5 ms step size, 6 ms window size, 44,100 frames per second) and frequencies between 500  Hz and 15,000  Hz were retained. Spectrograms were converted to mel scale using a mel filter with 128 channels. Syllables were compressed in the time dimension to a framerate of 640 frames per second then zero- padded to yield a standardized dimension of 128. Before dimensionality reduction, these 128 × 128 spectrograms were further reduced to 16 × 16 matrices and then flattened yielding a 256- length feature vector for each syllable. Syllable x feature vector matrices were then processed using the single- cell analysis framework Seurat v3 (Stuart et al., 2019). Principal component analysis was performed, then Uniform Manifold Approx- imation and Projection (UMAP) was performed on the first 10 principal components to produce a two- dimensional reduction. UMAP density differences To calculate global differences in syllable spectral structure before and after a manipulation (e.g. deafening), we split each bird’s song UMAP by day relative to the manipulation and computed two- dimensional kernel density estimates (R package MASS v7.3 function kde2d, 200 × 200 grid) for each of these per- day plots. A baseline UMAP structure was calculated as the mean density across the 2–4 days of singing before the manipulation, then density differences were calculated by subtracting this baseline density from each per- day density plot. Positive values from each difference plot were summed to give a single statistic for each day. Significance between hearing and deaf conditions for each day was determined using a two- sided t- test. Fundamental frequency statistics To calculate fundamental frequency (FF) for a given harmonic stack, we first computed the average spectrogram for 20 randomly selected syllables. We then identified a time within the syllable (relative to syllable onset) with stable FF and defined minimum and maximum frequency bounds to define a frequency band containing the FF. A short- time Fourier transform (STFT) was then calculated at this time point using function spec from R package seewave v2.1.8 (Sueur et al., 2008) (1024 window size, 44,100 frames per second). FF was estimated by interpolating the frequency spectrum on an output vector spanning the minimum and maximum frequency bounds with a resolution of 1  Hz (function aspline from R package akima v0.6–2.2). The maximum value of this interpolated frequency spectrum Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 21 of 37 Chromosomes and Gene Expression | Neuroscience Research article was taken as the FF. Rolling variability of FF was calculated as the coefficient of variation (CV, standard deviation/mean) over a set of FF values for a given syllable and the 10 prior syllables (of the same type). To compare variability relative to a baseline period, FF CV values were transformed to a percentage relative to the average variability before manipulation. Group estimates and significance values were obtained from mixed effects linear models using R package lme4 v1.1–27.1 (Bates et al., 2015) and function lmer (maximum likelihood criterion). The time period (before or after manipulation) was treated as a fixed effect and bird ID and syllable were treated as random effects with syllable nested under bird ID [model in lme4 notation: period + (1 | bird/syllable)]. p- values for fixed effect were obtained using ANOVA (Type II, Wald chi- square test statistic, R package car v3.0–11 function Anova, Fox and Weisberg, 2019) followed by adjustments for multiple testing using Benjamini- Hochberg correction. Song DKL To provide a single statistic that represents the amount of difference between songs in two condi- tions, we used a measure that we previously developed called Song DKL (Mets and Brainard, 2018). Songs for a given bird were divided into ‘pre’ and ‘post’-procedure groups. The ‘pre’ group consisted of songs from at most four days before the procedure up to the day preceding the procedure. The ‘post’ group contained song from two days before the day of euthanasia to the day of euthanasia. A maximum of 50 songs were sampled from each day. Syllables were identified in song WAV files by amplitude thresholding using a manually defined threshold for each bird. Mean power spectral densities (PSD) were computed for each segmented syllable using short- time Fourier transforms via R package seewave v2.1.8 (Sueur et  al., 2008) and function meanspec (window length ‘wl’=512, overlap ‘ovlp’=0%, normalized ‘norm’=T). Syllables in each dataset were split into a training dataset of 500 syllables and a held- out dataset of the remaining syllables. 50 PSDs were randomly selected from the ‘pre’ training dataset to serve as reference syllables for distance calculations. Inter- syllable spectral distances were calculated as Euclidean distances between this reference syllable set and each PSD, generating distance matrices for the ‘pre’ training and held- out datasets and the ‘post’ training and held- out datasets. Gaussian mixture models (GMMs) were fit to the ‘pre’ training distance matrix using function Mclust (5–12 mixture components, diagonal multivariate mixture model with varying volume, varying shape ‘VVI’) from R package mclust v5.4.7 (Scrucca et al., 2016). Bayesian Information Crite- rion (BIC) was computed for each model and second- order differences (difference of the difference) were calculated between the BICs for models with increasing numbers of mixture components. The model with minimum second- order BIC difference was selected for further use. A GMM was likewise fit to the ‘post’ training distance matrix using the same number of mixture components as in the selected ‘pre’ training model. The likelihood of generating each syllable in the ‘pre’ held- out dataset under the ‘pre’ and ‘post’ GMMs was calculated. This procedure was repeated ten times with different randomly selected reference syllables. The Kullback- Leibler divergence was then calculated as L1 − ( where L1 is the mean likelihood of observing a ‘pre’ held- out syllable across the ten replicated ‘pre’ models and L2 is the corresponding mean likelihood value for the ‘post’ models. These syllable- level DKL values were then averaged to give a single SongDKL for a given bird. DKL = log2 L2 ) Deafening by bilateral cochlear removal Nine Adult male Bengalese finches (103–458 days post- hatch, median ± SD of 133 ± 123) were deaf- ened by bilateral cochlear removal. Birds were anesthetized using isoflurane and an incision was made in the skin covering the ear canal to expose the canal. The tympanic membrane was ruptured, and the columella was removed using forceps. Cochlea were removed using a fine tungsten wire shaped into a hook. The incision was then resealed using VetBond (3M). For each deafened bird, a control (‘hearing’) bird underwent a sham surgery on the same day in which the bird was anesthetized, and the skin incision was made and then resealed. Birds survived for 4, 9, or 14 days (three hearing and three deaf birds for each timepoint) then were euthanized as described in Euthanasia and brain preparation. Unilateral LMAN lesions Five birds received unilateral LMAN lesions, three with left- hemisphere lesions, and two with right- hemisphere lesions. LMAN was electrolytically lesioned using a 100 kOhm platinum/iridium electrode. Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 22 of 37 Chromosomes and Gene Expression | Neuroscience Research article LMAN was stereotactically located at 4.7 mm AP, 1.7 mm ML, and 2.1 mm DV using a beak angle of 50 degrees. In one hemisphere, five penetrations were made, one at the given coordinates and four more +/-300 microns from this center position. At each penetration, 100 μA of current at the anode was passed for 60 s. At the experiment end, birds were euthanized as described in Euthanasia and brain preparation. To assess lesion completeness, 20 um coronal cryosections were collected at 100  um intervals across the anterior- posterior extent of LMAN onto SuperFrost Plus slides (Fisher- Brand), then Nissl stained as described in Standard Nissl stain, fresh- frozen cryosections. The volume of LMAN was estimated in ImageJ/Fiji (Schindelin et al., 2012) by calculating the area of LMAN on the unlesioned side in each section that it was visible, interpolating a smooth curve in R across these measured areas and the known distance between cryosections, then calculating the area under the curve. This procedure was repeated for any residual LMAN visible on the lesioned side, and the lesion percentage was calculated as 100 * (1 - [volume LMAN lesioned] / [volume LMAN unlesioned]). LMAN was considered lesioned if more than 25% of the volume was spanned by the lesion. Standard Nissl stain, fresh-frozen cryosections Frozen sections were allowed to come to room temperature for at least 20 min and then placed in a glass staining rack. Then slides were sequentially transferred to two rounds of xylenes for 5 min, two rounds of 100% ethanol for 5 min, one round of 95% ethanol for 5 min, 1 round of 70% ethanol for 5 min, water for 1 min, stained in 0.5% cresyl violet solution for 30 min, then rinsed for 1 min in water. Slides were then transferred to one round of 70% ethanol for 15–20 s (depending on desired staining intensity), one round of 95% ethanol for 30 s, two rounds of 100% ethanol for 30 s each, then two rounds of xylenes for 3 min each. DPX Mountant (Sigma) was applied, then slides were coverslipped. 0.5% cresyl violet was prepared as 300  mL water, 1  mL glacial acetic acid, and 1.5  g cresyl violet acetate. Solution was stirred for two days with no heat and then filtered. Euthanasia and brain preparation Birds were euthanized using isoflurane, decapitated, and debrained. All birds used for Serial Laser Capture RNA- seq were euthanized 2 hr after lights on at 9 AM. Brains were flash- frozen in –70 C dry ice- chilled isopentane for 12 s within 4 min from decapitation. Serial laser capture microdissection RNA-sequencing (SLCR-seq) — overview We were motivated by improvements to low- input RNA- sequencing stemming from optimized single- cell approaches to develop a method that would allow the construction of tens to hundreds of gene expression libraries from anatomically- defined regions. To achieve this we combined an optimized rapid Nissl staining protocol, laser capture microdissection, scalable RNA purification, and low- cost and low- input RNA- sequencing library construction into a single pipeline called Serial Laser Capture Microdissection RNA- sequencing (SLCR- seq). SLCR-seq — cryosectioning Surfaces in the cryostat chamber were first cleaned using a mixture of 50% RNaseZap (Ambion)/50% ethanol followed by a rinse of 70% ethanol in nuclease- free water. Flash- frozen brains were removed from –80  °C storage and allowed to equilibrate in a cryostat chamber set to –18  °C for  ~30  min. PEN membrane slides for LCM (Leica) and Superfrost Plus glass slides for histology (Fisherbrand) were placed in the cryochamber to chill. Once equilibrated, the brain was mounted onto a cryostat chuck using a small amount of OCT (TissueTek) with the posterior surface down and the anterior surface available for coronal sectioning. The brain was trimmed approximately 1.8 mm until reaching the anterior- posterior position of LMAN and Area X, which were visible as slightly darker regions. Sections were cut at 20 μm, transferred to pre- chilled membrane or glass slides, then melted onto the slides using a metal dowel that was pre- warmed on a slide warmer. Once a section was fully melted, the slide was transferred to a metal block in the cryostat chamber to refreeze the section. After sectioning through LMAN and Area X, the brain was detached from the chuck and remounted along the cut anterior surface for sectioning from the posterior surface. The brain was trimmed until reaching the anterior- position for RA (~0.8 mm from the posterior surface of the forebrain), which was also evident as a slightly darker region. Sections were collected onto membrane and glass slides as Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 23 of 37 Chromosomes and Gene Expression | Neuroscience Research article described through the level of HVC (~2.3 mm from the posterior surface) or Field L (visible as a dark curve extending from the medial surface). Once the collection was finished, slides were transferred to plastic slide mailers and stored in freezer boxes at –80 °C. Remaining brain tissue was re- wrapped in aluminum foil, placed back into a 15 mL conical tube, and stored at –80 °C. Test assays indicated that brains could be resectioned once more (for a total of two sectioning sessions) without negatively impacting RNA quality. SLCR-seq — rapid Nissl stain A fast Nissl staining procedure was developed to quickly stain cryosections before laser capture microdissection. Anhydrous 100% ethanol solution was prepared by adding 15 g of molecular sieve beads (Sigma 208582, 3  Å, 8–12 mesh) to 500  mL 100% molecular grade ethanol (Sigma E7023). Cresyl violet staining solution was prepared as by dissolving cresyl violet powder (Sigma) to 4% wt/ vol in 75% ethanol (75% molecular grade ethanol, 25% nuclease- free water), stirring for two days, then filtering through a 0.22 μm filter. Before staining, a series of 95%, 75%, and 50% ethanol solu- tions were prepared. To stain, each slide was thawed at room temperature on a bench for 20 s then transferred to 95% ethanol for 30 s, 75% ethanol for 30 s, and 50% ethanol for 30 s. 400 μL of cresyl violet staining solution was then applied to the slide for 30 s. Slides were destained and dehydrated by transferring to 50% ethanol for 30 s, 75% ethanol for 30 s, 95% ethanol for 30 s, then two rounds of 100% ethanol for 30 s each. Slides were then allowed to air dry. Time series experiments indicated that RNA quality was maintained for up to 45 min following staining. SLCR-seq — Laser capture microdissection After staining, slides were loaded onto a Leica LMD7000. Song nuclei were identified by anatomical landmarks (such as lamina and position relative to brain surfaces) and their higher intensity Nissl staining relative to surrounding regions. Sections cut from the surrounding tissue (power 45, aperture 50, speed 10, specimen balance 0, head 90%, pulse 92, offset 15) into eight- well strip caps containing 31.5 μL of RNA Lysis Buffer/PK (see SPRI RNA purification). After filling each cap with a section, the strip was placed onto a 96- well plate pre- chilled on ice and covered with an ice pack. Once a plate was filled, it was vortexed, spun down at 3250 × g for 5 min at 4 °C, then transferred to dry ice. For long- term storage, plates were stored at –80 °C. SLCR-seq — SPRI RNA purification The following solutions were prepared before LCM section collection: 50% guanidine thiocyanate (Sigma) in nuclease- free water, 5 X CN buffer (250  mM sodium citrate pH 7.0 (Sigma), 5%  NP- 40 (Sigma)), and RNA Lysis Buffer (20% guanidine thiocyanate, 1 X CN buffer). The following solutions were prepared before RNA purification: RNA Wash Buffer (25 mM sodium citrate pH 7.0, 15% guan- idine thiocyanate, 40% isopropanol) and solid phase reversible immobilization (SPRI) bead solution. SPRI bead solution was prepared by first vortexing Sera- Mag SpeedBeads Carboxyl Magnetic Beads, hydrophobic (Fisher) until fully suspended transferring 1  mL beads to a 1.5  mL tube. Beads were washed by placing the tube on a tube magnet, waiting until the solution cleared, removing the solu- tion, adding 1 mL of TE Buffer (10 mM UltraPure Tris HCl, pH 8.0 (ThermoFisher), 1 mM EDTA pH 8 (ThermoFisher)), and pipetting to mix. This wash was repeated once more, then the beads were resuspended in 1 mL TE Buffer. Separately, 9 g polyethylene glycol 8000 (Amresco), 10 mL 5 M NaCl, 500 μL 1 M UltraPure Tris HCl pH 8.0 (ThermoFisher), 100 μL 0.5 M EDTA pH 8.0 (ThermoFisher), and 500 μL 2% sodium azide (Sigma) were combined and brought to ~49 mL using nuclease- free water. Solution was mixed by inversion until PEG 8000 went into solution. Then, 137.5 μL of 20% Tween- 20 and 1 mL of beads/TE were added and mixed by inversion. This SPRI bead solution was then stored at 4 °C. Just before LCM collection, 31.5 μL of RNA Lysis Buffer/PK (1.5 μL of Proteinase K (Ambion), 30 μL of RNA Lysis Buffer) was prepared for each well. To purify RNA following LCM section collection, samples were first allowed to thaw on ice if stored at –80 °C. SPRI bead solution was allowed to come to room temperature, then 40 uL SPRI bead solution was mixed with 47.5 uL isopropanol for each sample. Samples were then lysed by incubating at 42 °C for 30 min in a thermocycler and then placed at room temperature. 87.5 uL of SPRI/isopropanol solution was added to each sample and then mixed 10 x by pipetting. Samples were incubated for 5 min at room temperature and then transferred to Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 24 of 37 Chromosomes and Gene Expression | Neuroscience Research article a magnetic plate stand. After 3  min, the solution was removed, the plate was removed from the magnetic stand, 100 uL of RNA Wash Buffer was added, and beads were resuspended by pipetting. The plate was immediately transferred back to the magnetic plate stand and held there for 2 min until the solution cleared. The solution was removed, and the plate was removed from the stand. 100 uL of 70% ethanol was added, beads were resuspended by pipetting 10 times, the plate was returned to the magnetic stand, the solution was allowed to clear for 2 min, and the solution was removed. This step was repeated for two total ethanol washes. Following the final wash, the beads were allowed to dry for 10 min while the plate remained on the stand. Residual ethanol was removed by pipetting. To elute RNA, the plate was removed from the magnet, 15 uL of nuclease- free water was added to each sample, and beads were resuspended by pipetting 10 times. Samples were incubated at room temperature for 5 min, then the plate was transferred to a low- elution volume magnetic stand. After 2 min or until the solution cleared, 10–12 μL eluted RNA was transferred to new 96- well plates on ice. Plates were sealed using foil adhesive, frozen on dry ice, then transferred to –80 °C for long- term storage. SLCR-seq — library preparation The SLCR- seq library preparation was adapted from several low- input and single- cell RNA- sequencing library protocols (Islam et al., 2014; Islam et al., 2012; Kivioja et al., 2012; Macosko et al., 2015; Picelli et  al., 2014). Barcoded unique molecular identifier (UMI) reverse transcription (RT) primers were prepared in advance in a 96- well plate (RT/TSO/dNTP mix). Each well contained 10 μM barcoded reverse transcription primer (RT_primer, IDT), 10  μM template- switching oligonucleotide with lock nucleic acids (TSO_LNA, Exiqon), and 10 mM dNTPs. Plates were sealed with foil adhesive and stored at –80  °C. Two RT primers were used in this study: one for the initial 18 bird deafening dataset (RT_primer_v1, 25 base UMI, six base barcode), and another for the 10 bird unilateral LMAN dataset (RT_primer_v2, 14 base UMI, 12 base barcode). RT_primer_v1 and RT_primer_v2 sets consisted of 24 and 48 barcodes, respectively (Supplementary file 1). Barcodes were at least one edit distance away from all other barcodes in the set. For library preparation, total RNA prepared from SPRI RNA purification was thawed on ice, then 4 μL total RNA was placed into a well of a 96- well plate chilled on ice. 1 μL RT/TSO/dNTP mix was added and mixed 10 times by pipetting. Plates were sealed with foil adhesive, incubated at 72 °C for 3 min, then snap- cooled in ice for at least 2 min. An RT Master Mix was prepared containing 1 x Enzscript RT buffer (Enzymatics), 5 mM dithiothreitol, 1 mM betaine, 12 mM MgCl2, 0.25 μL Recombi- nant Ribonuclease Inhibitor (Takara), and 10 U/μL Enzscript Moloney- Murine Leukemia Virus Reverse Transcriptase (Enzymatics). 5 μL of RT Master Mix was added to each sample and mixed by pipetting 10 times. Plates were sealed with foil adhesive and incubated in a thermocycler: 42 °C for 90 min, 70 °C for 15 min, 4 °C hold. Reactions were then pooled within a barcode set (e.g. barcodes 1–48 from RT_primer_v2 were combined into one tube). To purify cDNA, 0.6 x volume of Ampure XP bead solution was added to each pooled sample and mixed by pipetting 10 times. Samples were incubated for 5 min and transferred to a tube magnet. After the beads cleared from the solution, the solution was removed, and the beads were washed in 400 μL freshly prepared 80% ethanol for 30 s. This step was repeated for a total of two washes. After the second wash, the ethanol solution was removed, and the beads were allowed to dry for 5–10 min. Beads were then resuspended in 22 μL of nuclease- free water and incubated for 2 min. 20 μL eluted cDNA was transferred to new 1.5 mL LoBind tubes or 96- well plates and either stored at –20 °C or amplified immediately. During the purification a 40  μL cDNA Amplification Master Mix was prepared containing 10  μL KAPA HiFi 5 x Buffer, 1 μL 10 mM dNTPs, 4 μL 10 mM TSO_PCR primer, 0.5 μL 1 U/μL KAPA HiFi Hotstart DNA polymerase, and 24.5 μL nuclease- free water. 10 μL of purified cDNA was added to this master mix, pipetted 10 x to mix, then amplified under the following cycling parameters: 95 °C for min, then four cycles of 98 °C for 30 s, 65 °C for 45 s, and 72 °C for 3 min. Reactions were then placed on ice. During this initial amplification, a second master mix was prepared to determine the target number of amplification cycles by quantitative PCR. This mix contained 3 μL KAPA HiFi 5 x Buffer, 0.3 μL 10 mM dNTPs, 1.2 μL 10 mM TSO_PCR primer, 0.15 μL 1 U/μL KAPA HiFi Hotstart DNA poly- merase, 0.75 μL 20 x EvaGreen (Biotium), and 4.6 μL nuclease- free water. 5 μL of preamplified cDNA was added to this mix and amplified in a real- time PCR machine: 98 °C for 3 min, followed by 24 cycles of 98 °C for 20 s, 67 °C for 20 s, and 72 °C for 3 min, followed by 72 °C for 5 min. The target number Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 25 of 37 Chromosomes and Gene Expression | Neuroscience Research article of additional cycles was determined by identifying the Ct at 20% of the max fluorescence and then subtracting five cycles from this number. This number was generally between 5–7 additional cycles. The remaining 45 μL was placed back into the thermocycler and cycled at 98 °C for 30 s, the number of additional cycles at 98 °C for 20 s, 67 °C for 20 s, and 72 °C for 3 min, followed by 72 °C for 5 min. To purify the amplified cDNA, 0.6 x volume of Ampure XP bead solution was added to each reac- tion and mixed by pipetting 10 times. Samples were incubated for 5 min and transferred to a tube magnet. After the beads cleared from the solution, the solution was removed, and the beads were washed in 200 μL freshly prepared 80% ethanol for 30 s. This step was repeated for a total of two washes. After the second wash, the ethanol solution was removed, and the beads were allowed to dry for 5 min. Beads were then resuspended in 22 μL of nuclease- free water and incubated for 2 min. 20 μL eluted cDNA was transferred to new 1.5 mL LoBind tubes or 96- well plates and stored at –20 °C. Sample concentration was quantified using Qubit dsDNA High Sensitivity kit (ThermoFisher), then sample concentrations were standardized to 100 pg/μL. To prepare tagmented DNA, 4 μL (400 pg) of amplified cDNA was added to 10 μL Tagmentation Buffer (Buffer TD from the Nextera XT DNA Sample Prep Kit, Illumina), 1 μL nuclease- free water, and 5 μL ATM (Nextera XT). Reactions were mixed by pipetting 10 times the incubated at 55 °C for 5 min. 5 μL of Buffer NT was then added, then the reactions were incubated for 5 min at room temperature. Final libraries were constructed by first preparing a PCR master mix containing 20 μL KAPA HiFi 5 x Buffer, 2 μL 10 mM dNTPs, 5 μL 10 mM P5- TSO_Hybrid primer, 5 μL 10 mM PCR2 primer, 1 μL 1 U/ μL KAPA HiFi Hotstart DNA polymerase, and 42 μL nuclease- free water. PCR2 contains an i7 index (Supplementary file 1). The 25 μL tagmentation reaction was then added directly to the mix, and mixed by pipetting 10 times. Samples were amplified using 72 °C for 3 min; 95 °C for 3 min; followed by 16 cycles of 98 °C for 10 s, 55 °C for 30 s, and 72 °C for 30 s; followed by 72 °C for 5 min. Samples were then purified by adding 1.2 x volumes of Ampure XP, incubating for 5 min, then transferring to a tube magnet. After the beads cleared from the solution, the solution was removed, and the beads were washed in 200 μL freshly prepared 80% ethanol for 30 s. This step was repeated for a total of two washes. After the second wash, the ethanol solution was removed, and the beads were allowed to dry for 5 min. Beads were then resuspended in 22 μL of Low Elution Buffer (10 mM Tris HCl pH 8.0, 0.1 mM EDTA, 0.05% Tween- 20) and incubated for 2 min. 20 μL eluted cDNA was transferred to new 1.5 mL LoBind tubes and stored at –20 °C. Library size distributions were assessed using a Bioanalyzer High Sensitivity DNA Chip (Agilent), and library concentrations were determined using the KAPA Library Quantification Kit (Illumina Complete Kit, Roche). Samples were pooled at equal concentra- tions and then size- selected using a BluePippin and 2% BluePippin gels. DNA from 180 to 500 bp was selected and then purified using the MinElute kit (Qiagen) with two rounds of 10 μL elution in Low Elution Buffer. Samples were stored at –20 °C. RT_primer_v1 RT_primer_v2 TSO_LNA TSO_PCR AAGC AGTG GTAT CAAC GCAG AGTA CNNN NNNN NNNN NNNN NNNN NNNN NN XXXX XXTT TTTT TTTT TTTT TTTT TTTT TTTT TTVN AAGC AGTG GTAT CAAC GCAG AGTA CNNN NNNN NNNN NNNA TCTA GCCG G CCTT TTTT TTTT TTTT TTTT TTTT TTTT TTVN AAGC AGTG GTAT CAAC GCAG AGTG AATr GrG +G AAGC AGTG GTAT CAAC GCAG AGT P5- TSO_Hybrid AATG ATAC GGCG ACCA CCGA GATC TACA CGCC TGTC CGCG GAAG CAGT GGTA TCAA CGCA GAGT *A*C Read1CustomSeqB GCCT GTCC GCGG AAGC AGTG GTAT CAAC GCAG AGTA C PCR2 CAAG CAGA AGAC GGCA TACG AGAT YYYY YYYY GTCT CGTG GGCT CGG ‘N’, random nucleotide; ‘X’, barcode sequence; ‘Y’, i7 index sequence ‘V’, A or C or G; ‘r’, ribonucleic acid; ‘+’, locked nucleic acid; *, phosphorothioate RNA-sequencing preprocessing Sequencing reads were first trimmed for adaptor sequences using trim_galore (Krueger, 2020, --quality 20, --paired, --overlap 10, adaptors AAAA AAAA AA and GTAC TCTG CGTT GATA CCAC TGCT TCCG CGGA CAGG CGTG TAGA TCT). We first generated an initial alignment to the Bengalese finch genome (lonStrDom2, GCF_005870125.1) using STAR v2.7.8a (STARsolo mode, Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 26 of 37 Chromosomes and Gene Expression | Neuroscience Research article default parameters, --outFilterIntronMotifs RemoveNoncanonical) (Dobin et  al., 2013). To better annotate the 3’ UTRs of Bengalese finch gene models, we identified transcript 3’ ends by assembling transcripts using these initial alignments and the RNA- seq assembler Stringtie (Kovaka et al., 2019) (--fr -m 100). These Stringtie models were then intersected with the NCBI Bengalese finch transcrip- tome (lonStrDom2, GCF_005870125.1). New Stringtie exons were filtered by same- strandedness to the intersected reference genes, a minimal expression level (at least 10% of expression max for a given gene), and at least within 10 kilobases from the 3’ end of the gene. 3’ UTRs of the reference transcriptome were extended out to these new exons. Reads were then re- aligned to this extended transcriptome using the ‘bus’ subcommand from kallisto v0.46.1 Bray et al., 2016; Melsted et al., 2021 followed by barcode error correction using bustools v0.39.3 ‘correct,’ sorting using ‘sort,’ and read counting using ‘count.’. Differential expression analysis Gene- sample count matrices were filtered to remove lowly expressed genes, defined as having a total number of reads across samples less than the number of samples divided by eight (the number of brain regions assayed). For each sample we also calculated the ‘cellular detection rate (CDR)’ or the number of genes detected in a given sample, previously shown to substantially influence differ- ential expression analysis on single- cell RNA- sequencing samples (Finak et  al., 2015). Low- quality samples were defined as having a CDR of less than 30% of the total number of genes in the reference annotation (18,674 genes). Normalization factors were calculated using the function calcNormFactors from the R package edgeR v3.31.4 and the ‘TMMwsp’ method. The count matrix, these normaliza- tion factors, and a design matrix were then provided to the function voom from the limma package v.3.48.3 (Law et al., 2014; Ritchie et al., 2015). The design matrix was specified as: ~0 + position + position:num_songs_on_euth_date_log_scale + position:kl_mean_log_scale_ cut2_proc2 + position:kl_mean_log_scale_cut2_proc2:num_songs_on_euth_date_log_scale + position:nsongs_per_day_pre_log_scale + cdr_scale + frac_mito_scale + sv1 + sv2 where ‘position’ is an indicator for brain region, ‘num_songs_on_euth_date_log_scale’ is log- transformed total number of songs sung on the day of euthanasia, ‘kl_mean_log_scale_cut2_proc2’ is log- transformed Song DKL discretized into three equally sized bins, ‘nsongs_per_day_pre_log_scale’ is log- transformed average number of songs sung per day during the pre- procedure period, ‘cdr_scale’ is CDR, and ‘frac_mito’ is the fraction of reads mapping to mitochondrial genes in a given sample. Variables with ‘scale’ in their names were mean- subtracted and standard deviation- normalized. ‘sv1’ and ‘sv2’ correspond to the top two surrogate variables calculated using the function svaseq from the R package sva v3.40.0 (Leek, 2014), with full model specified as above and a null model given as ‘~0 + position +cdr_scaled +frac_mito_scale.’ Because SLCR- seq samples taken from the same bird and brain region are not fully independent samples, we considered these samples as technical replicates. We used the consensus correlation approach implemented in limma/voom to estimate the within- block (bird/region) expression similarity. To calculate the within- block correlation between samples, the resulting voom object was passed to duplicateCorrelation with block specified as the bird ID and brain region (for the deafening samples) or bird ID and brain hemisphere (for the unilateral LMAN lesion samples). To fit the model, the voom object, design matrix, and the consensus correlation were input to function lmFit from limma. Coefficient estimates and standard errors for each coefficient were calculated using function contrasts. fit, the function eBayes was used to compute moderated t- statistics and p- values, and the function topTable was used to adjust p- values using the Benjamini- Hochberg method. Genes were considered differentially expressed if their adjusted p- values were less than 0.1. Differentially expressed genes in the unilateral LMAN lesion SLCR- seq dataset were calculated similarily but with design specified as: ~0 + group + tags + cdr_scale where group indicates whether the region is ipsilateral or contralateral to the LMAN lesion. The variable ‘tags’ refers to bird ID tags and, therefore, controls for bird- level differences allowing pairwise comparisons of the effect of lesioning within birds. Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 27 of 37 Chromosomes and Gene Expression | Neuroscience Research article Expression estimates and standard errors for a given bird and brain region were computed using a regression approach with a design matrix specified as: ~0 + position:tags + cdr_scale +frac_mito_scale where ‘position’ is an indicator for brain region, ‘tags’ is the bird ID, ‘index2’ is a categorical vari- able indicating the sequencing run, and ‘cdr_scale’ and ‘frac_mito_scale’ are as described above. Standard errors were extracted from the linear fit model ‘fit’ as: sqrt( fit$ s2. post) * fit$stdev.unscaled. Network analysis We used the R package MEGENA (Multiscale Clustering of Geometrical Network, v1.3.7) to identify modules of genes with correlated expression across SLCR- seq data (Song and Zhang, 2015). Low- quality samples were removed by retaining samples with cellular detection rates above 0.42. Samples expression values were normalized using normalization factors calculated as described in ‘Differential expression analysis’(calcNormFactors and the ‘TMMwsp’ method) and then log- transformed with a pseudocount of 1. Samples were then split by brain region. To remove batch effects contributed by which pool a given sample was in, we used the function ComBat (Johnson et al., 2007) from the R package sva v3.40.0. For each brain region, we then selected the top 2000 variable genes as defined using the Seurat function FindVariableFeatures v4.0.4 and the ‘vst’ method. Signed Pearson correla- tions between every pair of genes were then calculated using the function calculate.correlation from MEGENA, which calculates false discovery rates by permutation (50 permutations). Correlations with an FDR less than 0.05 were retained. We passed these pairwise correlations to function calculate.PFN to generate a more sparse network that retains information edges using the MEGENA Planar Filtered Network algorithm. Module detection was then performed on this filtered network using the function do.MEGENA. To identify modules associated with behavioral features, we calculated the average of log- transformed estimates for a given coefficient across genes in a given module. To identify modules with greater (or lesser) than expected fold- changes, for each module we randomly selected the same number of genes and averaged their log- transformed coefficient estimates 100 times. Modules that had averages less than 1% or greater than 99% of this null distribution were considered signif- icant. Hub genes were designated using the approach defined in MEGENA. For each module, the link weights of the planar filtered network were permuted 100 times to generate a set of random networks. Within- module connectivities, defined as the sum of link weights with each other gene in a gene’s module, were calculated for each gene in each random network. The p- value was calculated as the probability of finding within- module connectivity values from this null distribution equal to or greater than the observed within- module connectivity. These p- values were then adjusted using the Benjamini- Hochberg method and genes with adjusted p- values less than 0.05 were designated hub genes. To compute gene module memberships, eigengenes were first determined for each module using the R package WGCNA v1.70–3 (Langfelder and Horvath, 2008), and function moduleEigengenes then the Pearson correlation was computed between each module eigengene and each gene. Module preservation statistics were calculated using the WGCNA function modulePreservation (Langfelder et al., 2011). Gene set enrichment analysis Gene Ontology lists were obtained from the Molecular Signatures Database (set C5, version 7). Gene set enrichment analysis was performed using the R package fgsea v1.18.0 (Korotkevich et al., 2021). T- statistics from voom regression or gene module membership scores from MEGENA were input into the function fgseaMultilevel (minSize = 20, maxSize = 200). Resulting pathways were filtered for those with an adjusted p- value less than 0.2 and similar pathways were pruned using collapsedPathways (pval.threshold=0.01 or 0.05). Inter-region correlation analysis To analyze inter- region gene correlations, we first selected in each region the 500 genes with the highest variability, computed as the variance- mean ratio of non- log expression across samples. For each gene, we calculated Pearson correlation values for each pairwise combination of brain regions, Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 28 of 37 Chromosomes and Gene Expression | Neuroscience Research article yielding region- by- region correlation matrices. To generate null distributions for each gene, we shuf- fled bird identities for each pairwise region- region comparison 100 times and computed Pearson correlations. We thresholded observed correlations using statistics from these shuffled distributions (correlation lesser or greater than the 2.5% or 97.5% shuffled quantiles). We calculated the across- bird expression similarity between regions as the number of thresholded correlations. To determine if deafening alters inter- region gene expression coupling in the song system, we computed pairwise Pearson correlations for each gene between each region for hearing and deaf birds separately, then took the absolute value for each matrix. The hearing absolute correlation matrix was then subtracted from the deaf absolute correlation matrix. This procedure was repeated on 100 shuffled distributions to generate a null distribution of differential absolute correlations. Differentially correlated genes were called as those with a deaf versus hearing value less than (decorrelation) or greater than (correlation gain) extreme values of a shuffled distribution calculated for each pairwise comparison (2.5% or 97.5%, respectively). Cell type specificity and differential expression scores For each gene, a specificity score was calculated as , where xn is expression divided by the sum of expression across all clusters and ¯xn is the mean of this value. Regression coefficient- cell ) type specificity scores were calculated by selecting differentially expressed genes (adjusted p- value <0.1) and then splitting genes by the sign of the coefficient. Scores were then computed as the dot- product between the cell type × gene specificity matrix and the gene × coefficient matrix. xn/ ¯xn xnlog ∑ ( Fluorescent in situ hybridization (FISH) FISH was performed using the hairpin chain reaction system from Molecular Instruments. Birds were euthanized using isoflurane, decapitated, and debrained. Brains were flash- frozen in –70 °C dry ice- chilled isopentane for 12 s within 4 min from decapitation then stored at –80 °C. Fresh- frozen brains were cryosectioned at 16 μm onto SuperFrost slides (Fisherbrand) chilled in the cryochamber then melted onto the slide using a warmed metal dowel. Slides were then transferred to –80 °C for storage. For the FISH, slides were transferred from –80 °C to slide mailers containing cold 4% PFA and incu- bated for 15 min on ice. Slides were washed three times for 5 min using DEPC- treated PBS + 0.1% Tween- 20, dehydrated in 50%, 70%, and two rounds of 100% ethanol for 3–5 min each round, then air dried. Slides were then transferred to a SlideMoat (Boekel Scientific) at 37 °C. 100 μL of v3 Hybridiza- tion Buffer (Molecular Instruments) was added to each slide, which were then coverslipped and incu- bated for 10 min at 37 °C. Meanwhile, 2 nM of each probe was added to 100 μL Hybridization Buffer and denatured at 37 °C. Pre- hybridization buffer was removed, 100 μL of probe/buffer was added, and slides were coverslipped and incubated overnight at 37 °C. The next day, coverslips were floated off in Probe Wash Buffer (PWB, 50% formamide, 5 x SSC, 9 mM citric acid pH 6.0, 0.1% Tween- 20, 50 μg/ml heparin), then washed in 75% PWB/25% SSCT (5 x SSC, 0.1% Tween- 20), 50% PWB/50% SSCT, 25% PWB/75% SSCT, 100% SSCT for 15  min each at 37  °C. This was followed by 5  min at room temperature in SSCT. Slides were incubated in 200 μL of Amplification Buffer (provided by the company) for 30 min at room temperature. Alexa fluor- conjugated DNA hairpins were denatured for 90 s at 95 °C then allowed to cool for at least 30 min in the dark at room temperature. Hairpins were added to 100 μL amplification buffer, applied to slides, and incubated overnight at room temperature. The following day, slides were washed in SSCT containing 1 ng/mL DAPI for 30 min at room tempera- ture, then SSCT for 30 min at room temperature, followed by a final 5 min in SSCT at room tempera- ture. Prolong Glass Antifade Medium (Thermofisher) was added to each slide and then coverslipped. Sections were imaged on a confocal microscope (Zeiss 710) using a 20 X objective. FISH quantification Image quantification was performed using CellProfiler v4.0.4 (Stirling et  al., 2021). DAPI- stained nuclei were first identified using the ClassifyPixels- Unet module. Areas corresponding to cells were estimated by extending nuclei boundaries by five pixels. Then signal puncta for each channel were identified and their intensities were measured. For each cell and each channel, we calculated the summed signal intensity of overlapping puncta divided by the cell area. To test for significant differ- ences in gene expression between hearing and deaf birds, a linear mixed effects model was fit using function lmer from R package lmerTest v3.1–3 (Kuznetsova et al., 2017) for each target gene and Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 29 of 37 Chromosomes and Gene Expression | Neuroscience Research article brain region as ‘intensity ~ condition + (1|bird)’ where ‘condition’ is contra or ipsi and ‘(1|bird)’ is the per- bird grouping factor. p- values were obtained by comparing this model with a reduced model ‘intensity ~ (1|bird)’ using ANOVA. Code availability Code underlying the analysis of birdsong and SLCR- seq gene expression can be found in the GitHub repository https://github.com/bradleycolquitt/deaf_gex (copy archived at Colquitt, 2023). Acknowledgements We would like to thank Andrea Hausenstaub and Christoph Schreiner for providing critical commen- tary on this manuscript, Adria Arteseros for providing technical expertise, and Mimi Kao for surgical expertise. Additional information Competing interests Foad Green: Foad Green is affiliated with Syapse, Inc. The author has no financial interests to declare. The other authors declare that no competing interests exist. Funding Funder National Institute of Neurological Disorders and Stroke Howard Hughes Medical Institute Grant reference number Author F32NS098809 Bradley M Colquitt Investigator Michael S Brainard The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Author contributions Bradley M Colquitt, Conceptualization, Resources, Data curation, Software, Formal analysis, Valida- tion, Investigation, Visualization, Methodology, Writing – original draft, Project administration, Writing – review and editing; Kelly Li, Resources, Investigation; Foad Green, Robert Veline, Investigation; Michael S Brainard, Conceptualization, Resources, Supervision, Funding acquisition, Project adminis- tration, Writing – review and editing Author ORCIDs Bradley M Colquitt http://orcid.org/0000-0001-5819-7924 Ethics All Bengalese finches (Lonchura striata domestica) were from our breeding colonies at UCSF or were purchased from approved vendors. All birds experienced a 14 hr:10 hr day:night cycle and were housed in communal cages separated by sex. Experiments were conducted in accordance with NIH and UCSF policies governing animal use and welfare (IACUC protocol number AN107972). All surgery was performed under isoflurane anesthesia, and every effort was made to minimize suffering. Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.85970.sa1 Author response https://doi.org/10.7554/eLife.85970.sa2 Additional files Supplementary files • Supplementary file 1. SLCR- seq barcode and index sequences. Full sequences for RT_primer_v1, Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 30 of 37 Chromosomes and Gene Expression | Neuroscience Research article RT_primer_v2, and PCR2. • Supplementary file 2. Deafening voom statistics Fold- change estimates and p- values for voom regression analysis of the deafening SLCR- seq dataset. • Supplementary file 3. GSEA statistics Gene set enrichment analysis of song destabilization- associated genes. • Supplementary file 4. Network module memberships Gene memberships in MEGENA modules for the RA, HVC, LMAN, and Area X networks. • Supplementary file 5. Differential correlation analysis Inter- region gene correlations results for the combined dataset and split between hearing and deaf birds. • Supplementary file 6. Unilateral LMAN lesion voom statistics Fold- change estimates and p- values for voom regression analysis of the unilateral LMAN lesion SLCR- seq dataset. • MDAR checklist Data availability SLCR- seq mapped sequencing reads, gene- by- sample count matrices, and metadata can be found at NCBI GEO for deafening (accession number GSE200663) and unilateral LMAN lesion datasets (GSE200664). The following datasets were generated: Author(s) Colquitt BM, Brainard MS Year 2022 Colquitt BM, Brainard MS 2022 Dataset title Dataset URL Database and Identifier Analysis of the effects of deafening on gene expression in birdsong neural circuitry https://www. ncbi. nlm. nih. gov/ geo/ query/ acc. cgi? acc= GSE200663 Analysis of the effects of unilateral LMAN lesioning on gene expression in birdsong neural circuitry https://www. ncbi. nlm. nih. gov/ geo/ query/ acc. cgi? acc= GSE200664 NCBI Gene Expression Omnibus, GSE200663 NCBI Gene Expression Omnibus, GSE200664 References Akutagawa E, Konishi M. 1994. Two separate areas of the brain differentially guide the development of a song control nucleus in the zebra finch. PNAS 91:12413–12417. DOI: https://doi.org/10.1073/pnas.91.26.12413, PMID: 7809051 Aldenhoff JB, Gruol DL, Rivier J, Vale W, Siggins GR. 1983. Corticotropin releasing factor decreases postburst hyperpolarizations and excites hippocampal neurons. Science 221:875–877. DOI: https://doi.org/10.1126/ science.6603658, PMID: 6603658 Andalman AS, Fee MS. 2009. A basal ganglia- forebrain circuit in the songbird biases motor output to avoid vocal errors. PNAS 106:12518–12523. DOI: https://doi.org/10.1073/pnas.0903214106, PMID: 19597157 Bargmann CI. 2012. Beyond the connectome: how neuromodulators shape neural circuits. BioEssays 34:458– 465. DOI: https://doi.org/10.1002/bies.201100185, PMID: 22396302 Bates D, Mächler M, Bolker B, Walker S. 2015. Fitting linear mixed- effects models using lme4. Journal of Statistical Software 67:i01. DOI: https://doi.org/10.18637/jss.v067.i01 Blank T, Nijholt I, Grammatopoulos DK, Randeva HS, Hillhouse EW, Spiess J. 2003. Corticotropin- Releasing factor receptors couple to multiple G- proteins to activate diverse intracellular signaling pathways in mouse hippocampus: role in neuronal excitability and associative learning. The Journal of Neuroscience 23:700–707. DOI: https://doi.org/10.1523/JNEUROSCI.23-02-00700.2003, PMID: 12533630 Bottjer SW, Miesner EA, Arnold AP. 1984. Forebrain lesions disrupt development but not maintenance of song in passerine birds. Science 224:901–903. DOI: https://doi.org/10.1126/science.6719123, PMID: 6719123 Brainard MS, Doupe AJ. 2000. Interruption of a basal ganglia- forebrain circuit prevents plasticity of learned vocalizations. Nature 404:762–766. DOI: https://doi.org/10.1038/35008083, PMID: 10783889 Brainard MS, Doupe AJ. 2001. Postlearning consolidation of birdsong: stabilizing effects of age and anterior forebrain lesions. The Journal of Neuroscience 21:2501–2517. DOI: https://doi.org/10.1523/JNEUROSCI.21- 07-02501.2001, PMID: 11264324 Brainard MS, Doupe AJ. 2013. Translating birdsong: songbirds as a model for basic and applied medical research. Annual Review of Neuroscience 36:489–517. DOI: https://doi.org/10.1146/annurev-neuro-060909- 152826, PMID: 23750515 Bray NL, Pimentel H, Melsted P, Pachter L. 2016. Near- optimal probabilistic RNA- seq quantification. Nature Biotechnology 34:525–527. DOI: https://doi.org/10.1038/nbt.3519, PMID: 27043002 Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 31 of 37 Chromosomes and Gene Expression | Neuroscience Research article Charlesworth JD, Warren TL, Brainard MS. 2012. Covert skill learning in a cortical- basal ganglia circuit. Nature 486:251–255. DOI: https://doi.org/10.1038/nature11078, PMID: 22699618 Chen Y, Dubé CM, Rice CJ, Baram TZ. 2008. Rapid loss of dendritic spines after stress involves derangement of spine dynamics by corticotropin- releasing hormone. The Journal of Neuroscience 28:2903–2911. DOI: https:// doi.org/10.1523/JNEUROSCI.0225-08.2008 Chen DY, Stern SA, Garcia- Osta A, Saunier- Rebori B, Pollonini G, Bambah- Mukku D, Blitzer RD, Alberini CM. 2011. A critical role for IGF- II in memory consolidation and enhancement. Nature 469:491–497. DOI: https:// doi.org/10.1038/nature09667, PMID: 21270887 Chen Q, Heston JB, Burkett ZD, White SA. 2013. Expression analysis of the speech- related genes FOXP1 and FOXP2 and their relation to singing behavior in two songbird species. The Journal of Experimental Biology 216:3682–3692. DOI: https://doi.org/10.1242/jeb.085886, PMID: 24006346 Chen SX, Kim AN, Peters AJ, Komiyama T. 2015. Subtype- Specific plasticity of inhibitory circuits in motor cortex during motor learning. Nature Neuroscience 18:1109–1115. DOI: https://doi.org/10.1038/nn.4049, PMID: 26098758 Chu Y, Fioravante D, Leitges M, Regehr WG. 2014. Calcium- Dependent PKC isoforms have specialized roles in short- term synaptic plasticity. Neuron 82:859–871. DOI: https://doi.org/10.1016/j.neuron.2014.04.003 Colquitt BM, Merullo DP, Konopka G, Roberts TF, Brainard MS. 2021. Cellular transcriptomics reveals evolutionary identities of songbird vocal circuits. Science 371:eabd9704. DOI: https://doi.org/10.1126/science. abd9704, PMID: 33574185 Colquitt BM. 2023. Influence of Deafening on gene expression in birdsong vocal motor circuitry. swh:1:rev:740f983ecea0c6b425dce6efa6c13b374262bc06. Software Heritage. https://archive.softwareheritage. org/swh:1:dir:f3501f06b9687a0abca13a79f550ec989bae5f1c;origin=https://github.com/bradleycolquitt/deaf_ gex;visit=swh:1:snp:6fa0c813b2d3b9df7418e955057abe2a4752afb5;anchor=swh:1:rev:740f983ecea0c6b425dc e6efa6c13b374262bc06 Cowie R, Douglas- Cowie E, Kerr AG. 1982. A study of speech deterioration in post- lingually deafened adults. The Journal of Laryngology and Otology 96:101–112. DOI: https://doi.org/10.1017/s002221510009229x, PMID: 7057081 Curmi PA, Andersen SS, Lachkar S, Gavet O, Karsenti E, Knossow M, Sobel A. 1997. The stathmin/tubulin interaction in vitro. The Journal of Biological Chemistry 272:25029–25036. DOI: https://doi.org/10.1074/jbc. 272.40.25029, PMID: 9312110 Dobin A, Davis CA, Schlesinger F, Drenkow J, Zaleski C, Jha S, Batut P, Chaisson M, Gingeras TR. 2013. Star: ultrafast universal RNA- seq aligner. Bioinformatics 29:15–21. DOI: https://doi.org/10.1093/bioinformatics/ bts635, PMID: 23104886 Donato F, Rompani SB, Caroni P. 2013. Parvalbumin- Expressing basket- cell network plasticity induced by experience regulates adult learning. Nature 504:272–276. DOI: https://doi.org/10.1038/nature12866, PMID: 24336286 Feenders G, Liedvogel M, Rivas M, Zapka M, Horita H, Hara E, Wada K, Mouritsen H, Jarvis ED. 2008. Molecular mapping of movement- associated areas in the avian brain: a motor theory for vocal learning origin. PLOS ONE 3:e1768. DOI: https://doi.org/10.1371/journal.pone.0001768, PMID: 18335043 Finak G, McDavid A, Yajima M, Deng J, Gersuk V, Shalek AK, Slichter CK, Miller HW, McElrath MJ, Prlic M, Linsley PS, Gottardo R. 2015. Mast: a flexible statistical framework for assessing transcriptional changes and characterizing heterogeneity in single- cell RNA sequencing data. Genome Biology 16:278. DOI: https://doi. org/10.1186/s13059-015-0844-5, PMID: 26653891 Fioravante D, Chu Y, de Jong AP, Leitges M, Kaeser PS, Regehr WG. 2014. Protein kinase C is a calcium sensor for presynaptic short- term plasticity. eLife 3:e03011. DOI: https://doi.org/10.7554/eLife.03011, PMID: 25097249 Fox EA, Gruol DL. 1993. Corticotropin- Releasing factor suppresses the afterhyperpolarization in cerebellar Purkinje neurons. Neuroscience Letters 149:103–107. DOI: https://doi.org/10.1016/0304-3940(93)90358-r, PMID: 8469370 Fox J, Weisberg S. 2019. An R companion to applied regression. 0.1. R Scripts. https://www.john-fox.ca/ Companion/index.html Fukushima M, Margoliash D. 2015. The effects of delayed auditory feedback revealed by bone conduction microphone in adult zebra finches. Scientific Reports 5:8800. DOI: https://doi.org/10.1038/srep08800, PMID: 25739659 Garst- Orozco J, Babadi B, Ölveczky BP. 2014. A neural circuit mechanism for regulating vocal variability during song learning in zebra finches. eLife 3:e03697. DOI: https://doi.org/10.7554/eLife.03697, PMID: 25497835 Gilliam DT, Menon V, Bretz NP, Pruszak J. 2017. The CD24 surface antigen in neural development and disease. Neurobiology of Disease 99:133–144. DOI: https://doi.org/10.1016/j.nbd.2016.12.011, PMID: 27993646 Goldberg JH, Fee MS. 2012. A cortical motor nucleus drives the basal ganglia- recipient thalamus in singing birds. Nature Neuroscience 15:620–627. DOI: https://doi.org/10.1038/nn.3047, PMID: 22327474 Hamaguchi K, Tschida KA, Yoon I, Donald BR, Mooney R. 2014. Auditory synapses to song premotor neurons are gated off during vocalization in zebra finches. eLife 3:e01833. DOI: https://doi.org/10.7554/eLife.01833, PMID: 24550254 Hayase S, Wang H, Ohgushi E, Kobayashi M, Mori C, Horita H, Mineta K, Liu WC, Wada K. 2018. Vocal practice regulates singing activity- dependent genes underlying age- independent vocal learning in songbirds. PLOS Biology 16:e2006537. DOI: https://doi.org/10.1371/journal.pbio.2006537, PMID: 30208028 Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 32 of 37 Chromosomes and Gene Expression | Neuroscience Research article Helduser S, Cheng S, Güntürkün O. 2013. Identification of two forebrain structures that mediate execution of memorized sequences in the pigeon. Journal of Neurophysiology 109:958–968. DOI: https://doi.org/10.1152/ jn.00763.2012, PMID: 23236000 Hilliard AT, Miller JE, Fraley ER, Horvath S, White SA. 2012. Molecular microcircuitry underlies functional specification in a basal ganglia circuit dedicated to vocal learning. Neuron 73:537–552. DOI: https://doi.org/10. 1016/j.neuron.2012.01.005, PMID: 22325205 Horita H, Kobayashi M, Liu WC, Oka K, Jarvis ED, Wada K. 2012. Specialized motor- driven DUSP1 expression in the song systems of multiple lineages of vocal learning birds. PLOS ONE 7:e42173. DOI: https://doi.org/10. 1371/journal.pone.0042173, PMID: 22876306 Hou ZH, Yu X. 2013. Activity- Regulated somatostatin expression reduces dendritic spine density and lowers excitatory synaptic transmission via postsynaptic somatostatin receptor 4. The Journal of Biological Chemistry 288:2501–2509. DOI: https://doi.org/10.1074/jbc.M112.419051, PMID: 23233668 Islam S, Kjällquist U, Moliner A, Zajac P, Fan JB, Lönnerberg P, Linnarsson S. 2012. Highly multiplexed and strand- specific single- cell RNA 5’ end sequencing. Nature Protocols 7:813–828. DOI: https://doi.org/10.1038/ nprot.2012.022, PMID: 22481528 Islam S, Zeisel A, Joost S, La Manno G, Zajac P, Kasper M, Lönnerberg P, Linnarsson S. 2014. Quantitative single- cell RNA- seq with unique molecular identifiers. Nature Methods 11:163–166. DOI: https://doi.org/10. 1038/nmeth.2772, PMID: 24363023 Jarvis ED, Scharff C, Grossman MR, Ramos JA, Nottebohm F. 1998. For whom the bird sings: context- dependent gene expression. Neuron 21:775–788. DOI: https://doi.org/10.1016/s0896-6273(00)80594-2, PMID: 9808464 Jevsek M, Jaworski A, Polo- Parada L, Kim N, Fan J, Landmesser LT, Burden SJ. 2006. Cd24 is expressed by myofiber synaptic nuclei and regulates synaptic transmission. PNAS 103:6374–6379. DOI: https://doi.org/10. 1073/pnas.0601468103, PMID: 16606832 Johnson F, Bottjer SW. 1994. Afferent influences on cell death and birth during development of a cortical nucleus necessary for learned vocal behavior in zebra finches. Development 120:13–24. DOI: https://doi.org/ 10.1242/dev.120.1.13, PMID: 21375056 Johnson F, Hohmann SE, DiStefano PS, Bottjer SW. 1997. Neurotrophins suppress apoptosis induced by deafferentation of an avian motor- cortical region. The Journal of Neuroscience 17:2101–2111. DOI: https://doi. org/10.1523/JNEUROSCI.17-06-02101.1997, PMID: 9045737 Johnson WE, Li C, Rabinovic A. 2007. Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics 8:118–127. DOI: https://doi.org/10.1093/biostatistics/kxj037, PMID: 16632515 Kao MH, Doupe AJ, Brainard MS. 2005. Contributions of an avian basal ganglia- forebrain circuit to real- time modulation of song. Nature 433:638–643. DOI: https://doi.org/10.1038/nature03127, PMID: 15703748 Kao MH, Brainard MS. 2006. Lesions of an avian basal ganglia circuit prevent context- dependent changes to song variability. Journal of Neurophysiology 96:1441–1455. DOI: https://doi.org/10.1152/jn.01138.2005, PMID: 16723412 Kemp CF, Woods RJ, Lowry PJ. 1998. The corticotrophin- releasing factor- binding protein: an act of several parts. Peptides 19:1119–1128. DOI: https://doi.org/10.1016/s0196-9781(98)00057-6, PMID: 9700765 Kittelberger JM, Mooney R. 1999. Lesions of an avian forebrain nucleus that disrupt song development alter synaptic connectivity and transmission in the vocal premotor pathway. The Journal of Neuroscience 19:9385– 9398. DOI: https://doi.org/10.1523/JNEUROSCI.19-21-09385.1999, PMID: 10531443 Kivioja T, Vähärautio A, Karlsson K, Bonke M, Enge M, Linnarsson S, Taipale J. 2012. Counting absolute numbers of molecules using unique molecular identifiers. Nature Methods 9:72–74. DOI: https://doi.org/10.1038/nmeth. 1778 Kojima S, Kao MH, Doupe AJ. 2013. Task- related `` cortical'' bursting depends critically on basal ganglia input and is linked to vocal plasticity. PNAS 110:4756–4761. DOI: https://doi.org/10.1073/pnas.1216308110 Korotkevich G, Sukhov V, Budin N, Shpak B, Artyomov MN, Sergushichev A. 2021. Fast Gene Set Enrichment Analysis. bioRxiv. DOI: https://doi.org/10.1101/060012 Kosche G, Vallentin D, Long MA. 2015. Interplay of inhibition and excitation shapes a premotor neural sequence. The Journal of Neuroscience 35:1217–1227. DOI: https://doi.org/10.1523/JNEUROSCI.4346-14.2015, PMID: 25609636 Kotak VC, Fujisawa S, Lee FA, Karthikeyan O, Aoki C, Sanes DH. 2005. Hearing loss raises excitability in the auditory cortex. The Journal of Neuroscience 25:3908–3918. DOI: https://doi.org/10.1523/JNEUROSCI.5169- 04.2005, PMID: 15829643 Kotak VC, Takesian AE, Sanes DH. 2008. Hearing loss prevents the maturation of GABAergic transmission in the auditory cortex. Cerebral Cortex 18:2098–2108. DOI: https://doi.org/10.1093/cercor/bhm233, PMID: 18222937 Kovaka S, Zimin AV, Pertea GM, Razaghi R, Salzberg SL, Pertea M. 2019. Transcriptome assembly from long- read RNA- seq alignments with stringtie2. Genome Biology 20:278. DOI: https://doi.org/10.1186/s13059-019-1910- 1, PMID: 31842956 Kratzer S, Mattusch C, Metzger MW, Dedic N, Noll- Hussong M, Kafitz KW, Eder M, Deussing JM, Holsboer F, Kochs E, Rammes G. 2013. Activation of CRH receptor type 1 expressed on glutamatergic neurons increases excitability of CA1 pyramidal neurons by the modulation of voltage- gated ion channels. Frontiers in Cellular Neuroscience 7:91. DOI: https://doi.org/10.3389/fncel.2013.00091, PMID: 23882180 Kröner S, Güntürkün O. 1999. Afferent and efferent connections of the caudolateral neostriatum in the pigeon (Columba livia): a retro- and anterograde pathway tracing study. The Journal of Comparative Neurology Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 33 of 37 Chromosomes and Gene Expression | Neuroscience Research article 407:228–260. DOI: https://doi.org/10.1002/(sici)1096-9861(19990503)407:2<228::aid-cne6>3.0.co;2-2, PMID: 10213093 Krueger F. 2020. Trim galore. 0.6.6. GitHub. https://github.com/FelixKrueger/TrimGalore Kuznetsova A, Brockhoff PB, Christensen RHB. 2017. Lmertest package: tests in linear mixed effects models. Journal of Statistical Software 82:i13. DOI: https://doi.org/10.18637/jss.v082.i13 Lane H, Webster JW. 1991. Speech deterioration in postlingually deafened adults. The Journal of the Acoustical Society of America 89:859–866. DOI: https://doi.org/10.1121/1.1894647, PMID: 2016436 Langfelder P, Horvath S. 2008. WGCNA: an R package for weighted correlation network analysis. BMC Bioinformatics 9:559. DOI: https://doi.org/10.1186/1471-2105-9-559, PMID: 19114008 Langfelder P, Luo R, Oldham MC, Horvath S. 2011. Is my network module preserved and reproducible? PLOS Computational Biology 7:e1001057. DOI: https://doi.org/10.1371/journal.pcbi.1001057, PMID: 21283776 Law CW, Chen Y, Shi W, Smyth GK. 2014. Voom: precision weights unlock linear model analysis tools for RNA- seq read counts. Genome Biology 15:R29. DOI: https://doi.org/10.1186/gb-2014-15-2-r29, PMID: 24485249 Leek JT. 2014. Svaseq: removing batch effects and other unwanted noise from sequencing data. Nucleic Acids Research 42:e161. DOI: https://doi.org/10.1093/nar/gku864, PMID: 25294822 Leonardo A, Konishi M. 1999. Decrystallization of adult birdsong by perturbation of auditory feedback. Nature 399:466–470. DOI: https://doi.org/10.1038/20933, PMID: 10365958 Li K, Nakajima M, Ibañez- Tallon I, Heintz N. 2016. A cortical circuit for sexually dimorphic oxytocin- dependent anxiety behaviors. Cell 167:60–72. DOI: https://doi.org/10.1016/j.cell.2016.08.067 Livingston FS, White SA, Mooney R. 2000. Slow NMDA- epscs at synapses critical for song development are not required for song learning in zebra finches. Nature Neuroscience 3:482–488. DOI: https://doi.org/10.1038/ 74857, PMID: 10769389 Lombardino AJ, Nottebohm F. 2000. Age at Deafening affects the stability of learned song in adult male zebra finches. The Journal of Neuroscience 20:5054–5064. DOI: https://doi.org/10.1523/JNEUROSCI.20-13-05054. 2000, PMID: 10864963 Lovell PV, Wirthlin M, Kaser T, Buckner AA, Carleton JB, Snider BR, McHugh AK, Tolpygo A, Mitra PP, Mello CV. 2020. Zebra: zebra finch expression brain atlas- A resource for comparative molecular neuroanatomy and brain evolution studies. The Journal of Comparative Neurology 528:2099–2131. DOI: https://doi.org/10.1002/cne. 24879, PMID: 32037563 Macosko EZ, Basu A, Satija R, Nemesh J, Shekhar K, Goldman M, Tirosh I, Bialas AR, Kamitaki N, Martersteck EM, Trombetta JJ, Weitz DA, Sanes JR, Shalek AK, Regev A, McCarroll SA. 2015. Highly parallel genome- wide expression profiling of individual cells using nanoliter droplets. Cell 161:1202–1214. DOI: https://doi.org/10.1016/j.cell.2015.05.002 Mandelblat- Cerf Y, Las L, Denisenko N, Fee MS. 2014. A role for descending auditory cortical projections in songbird vocal learning. eLife 3:e02152. DOI: https://doi.org/10.7554/eLife.02152, PMID: 24935934 Marder E. 2011. Variability, compensation, and modulation in neurons and circuits. PNAS 108 Suppl 3:15542– 15548. DOI: https://doi.org/10.1073/pnas.1010674108, PMID: 21383190 Melsted P, Booeshaghi AS, Liu L, Gao F, Lu L, Min KH, da Veiga Beltrame E, Hjörleifsson KE, Gehring J, Pachter L. 2021. Modular, efficient and constant- memory single- cell RNA- seq preprocessing. Nature Biotechnology 39:813–818. DOI: https://doi.org/10.1038/s41587-021-00870-2 Mets DG, Brainard MS. 2018. An automated approach to the quantitation of vocalizations and vocal learning in the songbird. PLOS Computational Biology 14:e1006437. DOI: https://doi.org/10.1371/journal.pcbi.1006437, PMID: 30169523 Miller JE, Hilliard AT, White SA. 2010. Song practice promotes acute vocal variability at a key stage of sensorimotor learning. PLOS ONE 5:e8592. DOI: https://doi.org/10.1371/journal.pone.0008592, PMID: 20066039 Miller MN, Cheung CYJ, Brainard MS. 2017. Vocal learning promotes patterned inhibitory connectivity. Nature Communications 8:2105. DOI: https://doi.org/10.1038/s41467-017-01914-5, PMID: 29235480 Moorman S, Ahn JR, Kao MH. 2021. Plasticity of stereotyped birdsong driven by chronic manipulation of cortical- basal ganglia activity. Current Biology 31:2619–2632.. DOI: https://doi.org/10.1016/j.cub.2021.04.030, PMID: 33974850 Mori C, Wada K. 2015. Audition- independent vocal crystallization associated with intrinsic developmental gene expression dynamics. The Journal of Neuroscience 35:878–889. DOI: https://doi.org/10.1523/JNEUROSCI. 1804-14.2015, PMID: 25609608 Nevue AA, Lovell PV, Wirthlin M, Mello CV. 2020. Molecular specializations of deep cortical layer analogs in songbirds. Scientific Reports 10:18767. DOI: https://doi.org/10.1038/s41598-020-75773-4, PMID: 33127988 Nicholson D. 2021. NickleDave/hybrid- vocal- classifier. Zenodo. https://doi.org/10.5281/zenodo.4678768 DOI: https://doi.org/10.5281/zenodo.4678768 Nordeen KW, Nordeen EJ. 1992. Auditory feedback is necessary for the maintenance of stereotyped song in adult zebra finches. Behavioral and Neural Biology 57:58–66. DOI: https://doi.org/10.1016/0163-1047(92) 90757-u, PMID: 1567334 Nordeen KW, Nordeen EJ. 1993. Long- Term maintenance of song in adult zebra finches is not affected by lesions of a forebrain region involved in song learning. Behavioral and Neural Biology 59:79–82. DOI: https:// doi.org/10.1016/0163-1047(93)91215-9, PMID: 8442737 Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 34 of 37 Chromosomes and Gene Expression | Neuroscience Research article Nottebohm F, Stokes TM, Leonard CM. 1976. Central control of song in the Canary, serinus canarius. The Journal of Comparative Neurology 165:457–486. DOI: https://doi.org/10.1002/cne.901650405, PMID: 1262540 Nottebohm F, Kelley DB, Paton JA. 1982. Connections of vocal control nuclei in the Canary telencephalon. The Journal of Comparative Neurology 207:344–357. DOI: https://doi.org/10.1002/cne.902070406, PMID: 7119147 Ohgushi E, Mori C, Wada K. 2015. Diurnal oscillation of vocal development associated with clustered singing by juvenile songbirds. The Journal of Experimental Biology 218:2260–2268. DOI: https://doi.org/10.1242/jeb. 115105, PMID: 26034125 Okanoya K, Yamaguchi A. 1997. Adult bengalese finches (Lonchura striata var. domestica) require real- time auditory feedback to produce normal song SYNTAX. Journal of Neurobiology 33:343–356 PMID: 9322153. Olveczky BP, Andalman AS, Fee MS. 2005. Vocal experimentation in the juvenile songbird requires a basal ganglia circuit. PLOS Biology 3:e153. DOI: https://doi.org/10.1371/journal.pbio.0030153, PMID: 15826219 Ölveczky BP, Otchy TM, Goldberg JH, Aronov D, Fee MS. 2011. Changes in the neural control of a complex motor sequence during learning. Journal of Neurophysiology 106:386–397. DOI: https://doi.org/10.1152/jn. 00018.2011, PMID: 21543758 Peng Z, Zeng S, Liu Y, Dong Y, Zhang H, Zhang X, Zuo M. 2012a. Comparative study on song behavior, and ultra- structural, electrophysiological and immunoreactive properties in RA among deafened, untutored and normal- hearing bengalese finches. Brain Research 1458:40–55. DOI: https://doi.org/10.1016/j.brainres.2012. 04.009, PMID: 22552113 Peng Z, Zhang X, Xi C, Zeng S, Liu N, Zuo M, Zhang X. 2012b. Changes in ultra- structures and electrophysiological properties in HVC of untutored and deafened bengalese finches relation to normally reared birds: implications for song learning. Brain Research Bulletin 89:211–222. DOI: https://doi.org/10.1016/ j.brainresbull.2012.09.004, PMID: 22982255 Peng Z, Zhang X, Liu Y, Xi C, Zeng S, Zhang X, Zuo M, xu J, Ji Y, Han Z. 2013. Ultrastructural and electrophysiological analysis of area X in the untutored and deafened bengalese finch in relation to normally reared birds. Brain Research 1527:87–98. DOI: https://doi.org/10.1016/j.brainres.2013.06.031, PMID: 23820426 Pfenning AR, Hara E, Whitney O, Rivas MV, Wang R, Roulhac PL, Howard JT, Wirthlin M, Lovell PV, Ganapathy G, Mouncastle J, Moseley MA, Thompson JW, Soderblom EJ, Iriki A, Kato M, Gilbert MTP, Zhang G, Bakken T, Bongaarts A, et al. 2014. Convergent transcriptional specializations in the brains of humans and song- learning birds. Science 346:1256846. DOI: https://doi.org/10.1126/science.1256846, PMID: 25504733 Picelli S, Björklund AK, Reinius B, Sagasser S, Winberg G, Sandberg R. 2014. Tn5 transposase and tagmentation procedures for massively scaled sequencing projects. Genome Research 24:2033–2040. DOI: https://doi.org/ 10.1101/gr.177881.114, PMID: 25079858 Pittman QJ, Siggins GR. 1981. Somatostatin hyperpolarizes hippocampal pyramidal cells in vitro. Brain Research 221:402–408. DOI: https://doi.org/10.1016/0006-8993(81)90791-5 Pytte CL, George S, Korman S, David E, Bogdan D, Kirn JR. 2012. Adult neurogenesis is associated with the maintenance of a stereotyped, learned motor behavior. The Journal of Neuroscience 32:7052–7057. DOI: https://doi.org/10.1523/JNEUROSCI.5385-11.2012, PMID: 22593073 Ritchie ME, Phipson B, Wu D, Hu Y, Law CW, Shi W, Smyth GK. 2015. Limma powers differential expression analyses for RNA- sequencing and microarray studies. Nucleic Acids Research 43:e47. DOI: https://doi.org/10. 1093/nar/gkv007, PMID: 25605792 Roy A, Mooney R. 2007. Auditory plasticity in a basal ganglia- forebrain pathway during decrystallization of adult birdsong. Journal of Neuroscience 27:6374–6387. DOI: https://doi.org/10.1523/JNEUROSCI.0894-07.2007 Sainburg T. 2020a. Bengalese finch. Zenodo. https://doi.org/10.5281/zenodo.3926440 DOI: https://doi. org/10.5281/zenodo.3926440 Sainburg T. 2020b. Animal vocalization generative network (AVGN). 0.1.0. GitHub. https://github.com/timsainb/ avgn_paper Sainburg T, Thielk M, Gentner TQ. 2020c. Finding, visualizing, and quantifying latent structure across diverse animal vocal repertoires. PLOS Computational Biology 16:e1008228. DOI: https://doi.org/10.1371/journal. pcbi.1008228, PMID: 33057332 Sakata JT, Hampton CM, Brainard MS. 2008. Social modulation of sequence and syllable variability in adult birdsong. Journal of Neurophysiology 99:1700–1711. DOI: https://doi.org/10.1152/jn.01296.2007, PMID: 18216221 Sasaki A, Sotnikova TD, Gainetdinov RR, Jarvis ED. 2006. Social context- dependent singing- regulated dopamine. The Journal of Neuroscience 26:9010–9014. DOI: https://doi.org/10.1523/JNEUROSCI.1335-06. 2006, PMID: 16943558 Scharff C, Nottebohm F. 1991. A comparative study of the behavioral deficits following lesions of various parts of the zebra finch song system: implications for vocal learning. The Journal of Neuroscience 11:2896–2913. DOI: https://doi.org/10.1523/JNEUROSCI.11-09-02896.1991, PMID: 1880555 Schenk BS, Baumgartner WD, Hamzavi JS. 2003. Effect of the loss of auditory feedback on segmental parameters of vowels of postlingually deafened speakers. Auris, Nasus, Larynx 30:333–339. DOI: https://doi. org/10.1016/s0385-8146(03)00093-2, PMID: 14656557 Schindelin J, Arganda- Carreras I, Frise E, Kaynig V, Longair M, Pietzsch T, Preibisch S, Rueden C, Saalfeld S, Schmid B, Tinevez JY, White DJ, Hartenstein V, Eliceiri K, Tomancak P, Cardona A. 2012. Fiji: an open- source Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 35 of 37 Chromosomes and Gene Expression | Neuroscience Research article platform for biological- image analysis. Nature Methods 9:676–682. DOI: https://doi.org/10.1038/nmeth.2019, PMID: 22743772 Scott LL, Nordeen EJ, Nordeen KW. 2000. The relationship between rates of hvc neuron addition and vocal plasticity in adult songbirds. Journal of Neurobiology 43:79–88. DOI: https://doi.org/10.1002/(sici)1097-4695( 200004)43:1<79::aid-neu7>3.0.co;2-p, PMID: 10756068 Scrucca L, Fop M, Murphy TB, Raftery AE. 2016. Mclust 5: clustering, classification and density estimation using Gaussian finite mixture models. The R Journal 8:289–317. DOI: https://doi.org/10.32614/RJ-2016-021, PMID: 27818791 Seki S, Eggermont JJ. 2003. Changes in spontaneous firing rate and neural synchrony in cat primary auditory cortex after localized tone- induced hearing loss. Hearing Research 180:28–38. DOI: https://doi.org/10.1016/ s0378-5955(03)00074-1, PMID: 12782350 Shumyatsky GP, Malleret G, Shin RM, Takizawa S, Tully K, Tsvetkov E, Zakharenko SS, Joseph J, Vronskaya S, Yin D, Schubart UK, Kandel ER, Bolshakov VY. 2005. Stathmin, a gene enriched in the amygdala, controls both learned and innate fear. Cell 123:697–709. DOI: https://doi.org/10.1016/j.cell.2005.08.038, PMID: 16286011 Simpson HB, Vicario DS. 1990. Brain pathways for learned and unlearned vocalizations differ in zebra finches. The Journal of Neuroscience 10:1541–1556. DOI: https://doi.org/10.1523/JNEUROSCI.10-05-01541.1990, PMID: 2332796 Sohrabji F, Nordeen EJ, Nordeen KW. 1990. Selective impairment of song learning following lesions of a forebrain nucleus in the juvenile zebra finch. Behavioral and Neural Biology 53:51–63. DOI: https://doi.org/10. 1016/0163-1047(90)90797-a, PMID: 2302141 Song WM, Zhang B. 2015. Multiscale embedded gene co- expression network analysis. PLOS Computational Biology 11:e1004574. DOI: https://doi.org/10.1371/journal.pcbi.1004574, PMID: 26618778 Song YH, Yoon J, Lee SH. 2021. The role of neuropeptide somatostatin in the brain and its application in treating neurological disorders. Experimental & Molecular Medicine 53:328–338. DOI: https://doi.org/10.1038/ s12276-021-00580-4, PMID: 33742131 Stirling DR, Swain- Bowden MJ, Lucas AM, Carpenter AE, Cimini BA, Goodman A. 2021. CellProfiler 4: improvements in speed, utility and usability. BMC Bioinformatics 22:433. DOI: https://doi.org/10.1186/ s12859-021-04344-9, PMID: 34507520 Stuart T, Butler A, Hoffman P, Hafemeister C, Papalexi E, Mauck WM, Hao Y, Stoeckius M, Smibert P, Satija R. 2019. Comprehensive integration of single- cell data. Cell 177:1888–1902.. DOI: https://doi.org/10.1016/j.cell. 2019.05.031, PMID: 31178118 Sueur J, Aubin T, Simonis C. 2008. SEEWAVE, a free modular tool for sound analysis and synthesis. Bioacoustics 18:213–226. DOI: https://doi.org/10.1080/09524622.2008.9753600 Tachibana RO, Oosugi N, Okanoya K. 2014. Semi- Automatic classification of birdsong elements using a linear support vector machine. PLOS ONE 9:e92584. DOI: https://doi.org/10.1371/journal.pone.0092584, PMID: 24658578 Todorov E. 2004. Optimality principles in sensorimotor control. Nature Neuroscience 7:907–915. DOI: https:// doi.org/10.1038/nn1309, PMID: 15332089 Tschida KA, Mooney R. 2012. Deafening drives cell- type- specific changes to dendritic spines in a sensorimotor nucleus important to learned vocalizations. Neuron 73:1028–1039. DOI: https://doi.org/10.1016/j.neuron.2011. 12.038, PMID: 22405211 Tyssowski KM, DeStefino NR, Cho JH, Dunn CJ, Poston RG, Carty CE, Jones RD, Chang SM, Romeo P, Wurzelmann MK, Ward JM, Andermann ML, Saha RN, Dudek SM, Gray JM. 2018. Different neuronal activity patterns induce different gene expression programs. Neuron 98:530–546. DOI: https://doi.org/10.1016/j. neuron.2018.04.001, PMID: 29681534 Uchida S, Martel G, Pavlowsky A, Takizawa S, Hevi C, Watanabe Y, Kandel ER, Alarcon JM, Shumyatsky GP. 2014. Learning- Induced and stathmin- dependent changes in microtubule stability are critical for memory and disrupted in ageing. Nature Communications 5:4389. DOI: https://doi.org/10.1038/ncomms5389, PMID: 25007915 Vallentin D, Kosche G, Lipkind D, Long MA. 2016. Neural circuits. Inhibition protects acquired song segments during vocal learning in zebra finches. Science 351:267–271. DOI: https://doi.org/10.1126/science.aad3023, PMID: 26816377 Vicario DS. 1991. Organization of the zebra finch song control system: II. functional organization of outputs from nucleus robustus archistriatalis. The Journal of Comparative Neurology 309:486–494. DOI: https://doi.org/10. 1002/cne.903090405, PMID: 1655832 Wada K, Howard JT, McConnell P, Whitney O, Lints T, Rivas MV, Horita H, Patterson MA, White SA, Scharff C, Haesler S, Zhao S, Sakaguchi H, Hagiwara M, Shiraki T, Hirozane- Kishikawa T, Skene P, Hayashizaki Y, Carninci P, Jarvis ED. 2006. A molecular neuroethological approach for identifying and characterizing a cascade of behaviorally regulated genes. PNAS 103:15212–15217. DOI: https://doi.org/10.1073/pnas.0607098103, PMID: 17018643 Waldstein RS. 1990. Effects of postlingual deafness on speech production: implications for the role of auditory feedback. The Journal of the Acoustical Society of America 88:2099–2114. DOI: https://doi.org/10.1121/1. 400107, PMID: 2269726 Wang N, Aviram R, Kirn JR. 1999. Deafening alters neuron turnover within the telencephalic motor pathway for song control in adult zebra finches. The Journal of Neuroscience 19:10554–10561. DOI: https://doi.org/10. 1523/JNEUROSCI.19-23-10554.1999, PMID: 10575051 Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 36 of 37 Chromosomes and Gene Expression | Neuroscience Research article Wang H, Eckel RH. 2012. Lipoprotein lipase in the brain and nervous system. Annual Review of Nutrition 32:147–160. DOI: https://doi.org/10.1146/annurev-nutr-071811-150703, PMID: 22540257 Warren WC, Clayton DF, Ellegren H, Arnold AP, Hillier LW, Künstner A, Searle S, White S, Vilella AJ, Fairley S, Heger A, Kong L, Ponting CP, Jarvis ED, Mello CV, Minx P, Lovell P, Velho TAF, Ferris M, Balakrishnan CN, et al. 2010. The genome of a songbird. Nature 464:757–762. DOI: https://doi.org/10.1038/nature08819 Warren TL, Tumer EC, Charlesworth JD, Brainard MS. 2011. Mechanisms and time course of vocal learning and consolidation in the adult songbird. Journal of Neurophysiology 106:1806–1821. DOI: https://doi.org/10.1152/ jn.00311.2011, PMID: 21734110 Watanabe A, Kimura T, Sakaguchi H. 2002. Expression of protein kinase C in song control nuclei of deafened adult male bengalese finches. Neuroreport 13:127–132. DOI: https://doi.org/10.1097/00001756-200201210- 00030, PMID: 11924874 Watanabe A, Li R, Kimura T, Sakaguchi H. 2006. Lesions of an avian forebrain nucleus prevent changes in protein kinase C levels associated with deafening- induced vocal plasticity in adult songbirds. The European Journal of Neuroscience 23:2447–2457. DOI: https://doi.org/10.1111/j.1460-9568.2006.04763.x, PMID: 16706851 Whitney O, Johnson F. 2005. Motor- induced transcription but sensory- regulated translation of ZENK in socially interactive songbirds. Journal of Neurobiology 65:251–259. DOI: https://doi.org/10.1002/neu.20187, PMID: 16155900 Whitney O, Pfenning AR, Howard JT, Blatti CA, Liu F, Ward JM, Wang R, Audet JN, Kellis M, Mukherjee S, Sinha S, Hartemink AJ, West AE, Jarvis ED. 2014. Core and region- enriched networks of behaviorally regulated genes and the singing genome. Science 346:1256780. DOI: https://doi.org/10.1126/science.1256780, PMID: 25504732 Wild JM. 1993. Descending projections of the songbird nucleus robustus archistriatalis. The Journal of Comparative Neurology 338:225–241. DOI: https://doi.org/10.1002/cne.903380207, PMID: 8308169 Williams H, Mehta N. 1999. Changes in adult zebra finch song require a forebrain nucleus that is not necessary for song production. Journal of Neurobiology 39:14–28. DOI: https://doi.org/10.1002/(SICI)1097-4695(199904) 39:13.0.CO;2-X, PMID: 10213450 Wood JM, Lacherez P, Black AA, Cole MH, Boon MY, Kerr GK. 2011. Risk of falls, injurious falls, and other injuries resulting from visual impairment among older adults with age- related macular degeneration. Investigative Opthalmology & Visual Science 52:5088. DOI: https://doi.org/10.1167/iovs.10-6644 Woolley SM, Rubel EW. 1997. Bengalese finches lonchura striata domestica depend upon auditory feedback for the maintenance of adult song. The Journal of Neuroscience 17:6380–6390. DOI: https://doi.org/10.1523/ JNEUROSCI.17-16-06380.1997, PMID: 9236246 Xian X, Liu T, Yu J, Wang Y, Miao Y, Zhang J, Yu Y, Ross C, Karasinska JM, Hayden MR, Liu G, Chui D. 2009. Presynaptic defects underlying impaired learning and memory function in lipoprotein lipase- deficient mice. The Journal of Neuroscience 29:4681–4685. DOI: https://doi.org/10.1523/JNEUROSCI.0297-09.2009, PMID: 19357293 Yu T, Taussig MD, DiPatrizio NV, Astarita G, Piomelli D, Bergman BC, Dell’Acqua ML, Eckel RH, Wang H, Andrews Z. 2015. Deficiency of lipoprotein lipase in neurons decreases AMPA receptor phosphorylation and leads to neurobehavioral abnormalities in mice. PLOS ONE 10:e0135113. DOI: https://doi.org/10.1371/journal. pone.0135113 Zhou X, Fu X, Lin C, Zhou X, Liu J, Wang L, Zhang X, Zuo M, Fan X, Li D, Sun Y. 2017. Remodeling of dendritic spines in the avian vocal motor cortex following Deafening depends on the basal ganglia circuit. Cerebral Cortex 27:2820–2830. DOI: https://doi.org/10.1093/cercor/bhw130, PMID: 27166173 Colquitt et al. eLife 2023;12:e85970. DOI: https://doi.org/10.7554/eLife.85970 37 of 37 Chromosomes and Gene Expression | Neuroscience
10.1093_plcell_koad157
ERROR: type should be string, got "https://doi.org/10.1093/plcell/koad157\n\nTHE PLANT CELL 2023: 35: 3236–3259\n\nw\ne\ni\nv\ne\nR\n\nThe pyrenoid: the eukaryotic CO2-concentrating\norganelle\n\nShan He\n\n,1,2 Victoria L. Crans\n\n1 and Martin C. Jonikas\n\n1,2,*\n\n1 Department of Molecular Biology, Princeton University, Princeton, NJ 08540, USA\n2 Howard Hughes Medical Institute, Princeton University, Princeton, NJ 08540, USA\n\n*Author for correspondence: mjonikas@princeton.edu\n\nAbstract\nThe pyrenoid is a phase-separated organelle that enhances photosynthetic carbon assimilation in most eukaryotic algae and\nthe land plant hornwort lineage. Pyrenoids mediate approximately one-third of global CO2 fixation, and engineering a pyrenoid\ninto C3 crops is predicted to boost CO2 uptake and increase yields. Pyrenoids enhance the activity of the CO2-fixing enzyme\nRubisco by supplying it with concentrated CO2. All pyrenoids have a dense matrix of Rubisco associated with photosynthetic\nthylakoid membranes that are thought to supply concentrated CO2. Many pyrenoids are also surrounded by polysaccharide\nstructures that may slow CO2 leakage. Phylogenetic analysis and pyrenoid morphological diversity support a convergent evo-\nlutionary origin for pyrenoids. Most of the molecular understanding of pyrenoids comes from the model green alga\nChlamydomonas (Chlamydomonas reinhardtii). The Chlamydomonas pyrenoid exhibits multiple liquid-like behaviors, includ-\ning internal mixing, division by fission, and dissolution and condensation in response to environmental cues and during the cell\ncycle. Pyrenoid assembly and function are induced by CO2 availability and light, and although transcriptional regulators have\nbeen identified, posttranslational regulation remains to be characterized. Here, we summarize the current knowledge of pyr-\nenoid function, structure, components, and dynamic regulation in Chlamydomonas and extrapolate to pyrenoids in other\nspecies.\n\nIntroduction\nPhotosynthesis forms the base of the food chain in most eco-\nsystems by converting CO2 from the environment into or-\nganic carbon. At the heart of these reactions is the enzyme\nRibulose-1,5-bisphosphate carboxylase/oxygenase (Rubisco),\nwhich assimilates CO2 into sugar precursors used to generate\nbiomass (Bracher et al. 2017). Despite its crucial role in\nphotosynthesis, Rubisco has two limitations: (1) it has a\nslow catalytic rate for an enzyme in central carbon metabol-\nism; and (2) it can also catalyze oxygenation, a wasteful reac-\ntion that uses O2 instead of CO2 (Flamholz et al. 2019). A\ntradeoff between Rubisco’s catalytic rate and specificity for\nCO2 over O2 appears to prevent the evolution or engineering\nof Rubisco to be both fast and specific (Tcherkez et al. 2006;\nSavir et al. 2010; Flamholz et al. 2019). To keep oxygenation at\n\ntolerably low levels, many plants use a specific but slow form\nof Rubisco and compensate for its slow catalytic rate by pro-\nducing a large amount of the enzyme (Raven 2013). This\nstrategy requires significant cellular resources, including up\nto 25% of total leaf nitrogen (Raven 2013).\n\nSome photosynthetic organisms overcome the limitations\nof Rubisco by using a CO2-concentrating mechanism (CCM)\nto deliver concentrated CO2 to the enzyme. This concen-\ntrated CO2 increases the turnover rate of Rubisco, and the\nhigher ratio of CO2 to O2 favors carboxylation and suppresses\noxygenation (Badger et al. 1980; Kupriyanova et al. 2023).\nThere is currently great interest in understanding how\nCCMs work, both because of their significant ecological\nrole (Ehleringer et al. 1991; Badger et al. 2006; Meyer et al.\n2017) and because engineering a CCM into crops has the\n\nReceived February 23, 2023. Accepted May 17, 2023. Advance access publication June 4, 2023\n© The Author(s) 2023. Published by Oxford University Press on behalf of American Society of Plant Biologists.\nThis is an Open Access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and\nreproduction in any medium, provided the original work is properly cited.\n\nOpen Access\n\n\fThe eukaryotic CO2-concentrating organelle\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\n|\n\n3237\n\npotential to increase yields (Matsuoka et al. 2001; Kajala et al.\n2011; Hanson et al. 2016; Hennacy and Jonikas 2020; Adler\net al. 2022).\n\nCCMs are categorized into two broad classes: biochemical\nand biophysical, depending on the nature of the intermedi-\nate molecules used to concentrate CO2. Biochemical\nCCMs, which\ninclude C4, C2, and crassulacean acid\nmetabolism (CAM), transiently fix CO2 into intermediate or-\nganic molecules such as oxaloacetate and malate, from which\nconcentrated CO2\nis released in proximity to Rubisco\n(Caemmerer and Furbank 2003; Sage et al. 2012; Heyduk\net al. 2019). By contrast, in biophysical CCMs, the only inter-\n−) (Hennacy and\nmediate molecule is bicarbonate (HCO3\nJonikas 2020). Biochemical CCMs are predominantly found\nin plants and typically involve multicellular structures,\nwhereas biophysical CCMs are predominantly found in mi-\ncrobes and operate at a single-cell level (Maberly and\nGontero 2017).\n\nBiophysical CCMs differ between prokaryotes and eukar-\nyotes. Both rely on a subcellular structure whose matrix\ncontains a high concentration of Rubisco, into which concen-\n− (Wang et al. 2015; Kaplan\ntrated CO2 is released from HCO3\n2017; Hennacy and Jonikas 2020; Adler et al. 2022; Ang et al.\n2022). However, the eukaryotic compartment known as\nthe pyrenoid (1-2 µm in diameter) is much bigger than the\nbacterial Rubisco-containing compartment, the carboxysome\nin diameter). Additionally, CO2 delivery to\n(∼200 nm\nRubisco in the two structures is thought to be achieved based\non different principles: in pyrenoids, CO2 delivery is mediated\nby thylakoid membranes, as discussed below (Fei et al. 2022),\n− diffus-\nwhereas in carboxysomes, CO2 is produced from HCO3\ning directly into the carboxysome matrix (Mangan et al. 2016).\nThis review will focus on the pyrenoid.\n\nPyrenoids are found inside the chloroplasts of most eu-\nkaryotic algae (including most microalgae and many macro-\nalgae) and some species of nonvascular land plants called\nhornworts (Villarreal and Renner 2012; Meyer et al. 2017; Li\net al. 2020). The pyrenoid is typically visible under light mi-\ncroscopy as a 1–2 µm punctum within the chloroplast\n(Fig. 1A).\n\nOne of the earliest records of a pyrenoid dates from 1782\nby the Danish naturalist and scientific illustrator Otto\nFrederik Müller, who drew unnamed puncta in sketches of\nthe green alga Spirogyra (formerly Conferva jugalis) (Müller\n1782), making it one of the first scientifically documented or-\nganelles. The pyrenoid was first described in a publication in\n1803 (Vaucher 1803). The term pyrenoid was conceived in\n1882 from the Greek πυρην (pyren, kernel) (Schmitz 1882).\nFrom this point on, the pyrenoid became the focus of\nmany classic morphological studies using light and electron\nmicroscopy. Further reading on the history of pyrenoid re-\nsearch can be found in a recent review by Barrett et al. and\na book chapter by Meyer et al. (Meyer et al. 2020a; Barrett\net al. 2021).\n\nResearch on pyrenoids has recently gained momentum\nand currently has three major motivations: (1) a growing ap-\npreciation for the major role of pyrenoids in the global car-\nbon cycle; (2) prospects to engineer pyrenoids into crops\nto increase yields; and (3) the unique value of the pyrenoid\nas a model for biological phase-separated condensates, a re-\ncently discovered ubiquitous class of organelles. We discuss\neach of these motivations below.\n\nPyrenoids play a major role in the global carbon cycle, me-\ndiating approximately 30% to 40% of global CO2 assimilation\neach year (Mackinder et al. 2016). Approximately one-half of\nglobal CO2 assimilation occurs in the oceans (Field et al. 1998;\n\nFigure 1. Structure of the Chlamydomonas pyrenoid. A) The Chlamydomonas pyrenoid is visible by light microscopy. Scale bar, 1 μm. B) The\nChlamydomonas pyrenoid is composed of three major compartments: the Rubisco matrix, thylakoid tubules, and starch sheath. Scale bar,\n200 nm. C) Two-dimensional and D) Three-dimensional models of the pyrenoid showing the major compartments and protein peripheral struc-\ntures. See Table 1 for a list of the known protein components of each structure. (The circled numbers indicating the sub-pyrenoid localizations in\npanel C are coordinated with the circled numbers in Table 1).\n\nACDB\f3238\n\n|\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\nHe et al.\n\nBehrenfeld et al. 2001), and most of this assimilation is attrib-\nuted to eukaryotic algae (Flombaum et al. 2013; Rousseaux\nand Gregg 2013), nearly all of which have pyrenoids (Mann\n1996; Not et al. 2004; Thierstein and Young 2004; Meyer\nand Griffiths 2013).\n\nThere is a growing interest in enhancing yields of major\nglobal crops that do not have CCMs by engineering a CCM\ninto them (Rae et al. 2017; Hennacy and Jonikas 2020).\nAmong the various CCMs that could be engineered, the\ngreen algal pyrenoid-based CCM is a particularly promising\ncandidate for engineering into non-CCM plants as a result\nof two attractive qualities: (1) it operates at the single-cell le-\nvel, which means that leaf anatomy does not need to be en-\ngineered as would be necessary for engineering of the C4\nCCM (Kajala et al. 2011); and (2) unlike the prokaryotic\ncarboxysome-based CCM, the green algal pyrenoid-based\nCCM is natively encoded in the eukaryotic nuclear genome,\nwhich could facilitate its engineering into monocot crops\nsuch as wheat (Triticum aestivum) and rice (Oryza sativa),\nwhose prokaryotic chloroplast genomes remain challenging\nto engineer.\n\nThe pyrenoid is also of interest from a fundamental science\nperspective, as it is a phase-separated organelle (Freeman\nRosenzweig et al. 2017). Biological phase separation underlies\nthe formation of many cellular structures (Shin and\nBrangwynne 2017). The pyrenoid is one of the few phase-\nseparated organelles where the functional value of conden-\nsate formation is understood, as there is a clear functional\nand fitness cost to preventing Rubisco condensation into a\nmatrix (Meyer et al. 2012; Mackinder et al. 2016; He et al.\n2020; Fei et al. 2022). Moreover, the pyrenoid of the model\nalga Chlamydomonas (Chlamydomonas reinhardtii) is one of\nthe structurally best understood phase-separated conden-\nsates, as its phase separation was reconstituted in vitro\n(Wunder et al. 2018) and the structural basis behind this phase\nseparation was determined (He et al. 2020), making it a power-\nful system for deriving the basic fundamental principles that\nunderlie the assembly of phase-separated organelles.\n\nThe vast majority of our molecular understanding of pyr-\nenoids comes from recent studies in Chlamydomonas. As a\nwell-established model organism widely used for photosyn-\nthesis studies, Chlamydomonas benefits from a thriving com-\nmunity of researchers who have produced genome\nsequences and annotations (Merchant et al. 2007; Craig\net al. 2023), genome-wide omics data (Brueggeman et al.\n2012; Fang et al. 2012; Zones et al. 2015; Strenkert et al.\n2019), mutant libraries (Li et al. 2016; Li et al. 2019), fluores-\ncently tagged lines for gene functional analysis (Mackinder\net al. 2017; Wang et al. 2022), as well as pyrenoid proteomes\nand a pyrenoid proxiome (Mackinder et al. 2017; Zhan et al.\n2018; Lau et al. 2023). Such resources are currently lacking for\nother algal species, although recent progress has been made\ntoward developing similar tools in model diatoms such as\nhigh-efficiency transformation protocols in Phaeodactylum\ntricornutum (Miyagawa et al. 2009), stably propagated epi-\nsomes in P. tricornutum and Thalassiosira pseudonana\n\n(Karas et al. 2015), proteome analyses of mitochondria and\nplastids in T. pseudonana (Schober et al. 2019), and a fluores-\ncent protein-tagging pipeline in T. pseudonana (Nam et al.\n2022).\n\nIn this review, we discuss the basic concepts of pyrenoid\nfunction as well as the current understanding of pyrenoids\nin Chlamydomonas, other algae, and hornworts. Some as-\npects of these topics have been covered in recent reviews\n(Barrett et al. 2021; Adler et al. 2022). Our review seeks to\nprovide an update on the most recent discoveries in the field\nand discuss the evolution, biogenesis, regulation, and func-\ntion of the pyrenoid-based CCM.\n\nOperating principles and evolution\nOperating principles of the pyrenoid-based CCM\nPhotosynthetic cells can obtain their carbon from two\n−. The availability\nsources in the environment: CO2 and HCO3\nof each source can vary depending on the environment (e.g.\naquatic growth or growth on surfaces exposed to air) and\nconditions (e.g. external pH). In the aquatic environment,\n− is normally more abundant than CO2, whereas CO2\nHCO3\nmay be more available to cells growing on the surface of par-\nticles in the soil.\n\nCO2 is difficult to concentrate directly within cells because\nit is a small, uncharged molecule that rapidly leaks across\nmembranes. Current models of the pyrenoid-based CCM\nsuggest that, to overcome this issue, cells use carbonic anhy-\n−, which\ndrases to convert CO2 to the charged molecule HCO3\ncannot easily diffuse across membranes and can be directed\nto subcellular compartments via transmembrane transpor-\nters. Intercompartmental pH differences are thought to\nplay a crucial role in this process by driving the interconver-\n− in the appropriate cellular compart-\nsion of CO2 to HCO3\nments (Fig. 2) (Hennacy and Jonikas 2020; Wang and\nJonikas 2020; Fei et al. 2022).\n\nA recent study (Fei et al. 2022) provides a detailed compu-\ntational model of the Chlamydomonas CCM that is consist-\nent with all available experimental evidence. We expect that\nthe model will also be generally relevant to other algae, as\nmost algae face similar biophysical challenges and the model\nis robust over broad parameter ranges.\n\nInterestingly, this model indicates that two distinct CCM\noperating modes are feasible, which share a common core\n− is accumulated in the chloroplast\nbut differ in how HCO3\nstroma (Fei et al. 2022) (Fig. 2). At the common core of the\n− is transported into\ntwo operating modes, stromal HCO3\nthe thylakoid lumen, likely by the bestrophin-like channels\nBST1 (encoded by Cre16.g662600), BST2 (Cre16.g663400),\nand/or BST3 (Cre16.g663450), although the role of each\nBST remains unclear (Mukherjee et al. 2019). Inside specia-\nlized pyrenoid-traversing regions of the thylakoid mem-\nbranes, carbonic anhydrase 3 (CAH3, Cre09.g415700)\n− to CO2, which is driven by the low pH of\nconverts HCO3\nthe thylakoid lumen (Karlsson et al. 1998; Hanson et al.\n\n\fThe eukaryotic CO2-concentrating organelle\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\n|\n\n3239\n\nstroma, as follows. The first mode uses a passive chloroplast\nCO2 uptake strategy, where CO2 passively diffuses from the\nperiplasm across the plasma membrane via the channel\nlow CO2-inducible 1 (LCI1, Cre03.g162800) (Ohnishi et al.\n2010; Kono and Spalding 2020) and across the chloroplast\nenvelope into the chloroplast stroma (Fig. 2A). CO2 diffusing\ninto the chloroplast or leaking out of the pyrenoid is con-\n− by the low-CO2-inducible B/C (LCIB/LCIC,\nverted to HCO3\nCre10.g452800/Cre06.g307500) carbonic anhydrase complex\nin a reaction driven by the high pH in the chloroplast stroma\n(Wang and Spalding 2006; Yamano et al. 2010; Jin et al. 2016;\nKasili et al. 2023).\n\nBy contrast, the second CCM operating mode uses an ac-\n− uptake strategy. CO2 is converted to\ntive chloroplast HCO3\n− at the periplasm by the carbonic anhydrases CAH1\nHCO3\n(Cre04.g223100) and CAH2 (Cre04.g223050) (Fujiwara et al.\n− crosses the plasma\n1990; Van and Spalding 1999). HCO3\nmembrane via the transporter high light activated 3 (HLA3,\nCre02.g097800) and is then concentrated across the chloro-\nplast envelope by LCIA (Cre06.g309000), which in this model\n− pump (Fig. 2B) (Miura et al. 2004; Yamano\nis an active HCO3\net al. 2015). We note that LCIA has not been experimentally\n− across a membrane, but ac-\nshown to actively pump HCO3\ntive pumping seems likely as the model indicates that passive\n− channels across the chloroplast envelope fail to\nHCO3\nachieve an effective CCM (Fei et al. 2022). CO2 leaking out\nof the pyrenoid is recaptured by the LCIB/LCIC carbonic an-\nhydrase complex, which relocalizes to the periphery of the\npyrenoid (Wang and Spalding 2014a) to enhance the effi-\n− to\nciency of CO2 recapture and avoid conversion of HCO3\nCO2 near the chloroplast envelope, which would lead to\n− (Fig. 2B) (Fei et al.\nloss of accumulated chloroplast HCO3\n2022).\n\nThe passive CO2 uptake strategy and active HCO3\n\n− uptake\nstrategy have different performance depending on external\nCO2 concentrations and pH (Fei et al. 2022). The passive\nCO2 uptake strategy is effective and energetically efficient\nunder ambient air levels of external CO2 (0.04%, 400 ppm;\nequivalent to 10 μM cytosolic in the model) but is unable\nto deliver enough CO2 to saturate Rubisco under lower levels\nof CO2 (0.004%, also known as “very low CO2”; corresponding\nto 1 μM cytosolic CO2\nin the model). Accordingly,\nChlamydomonas appears to use the passive CO2 uptake\nstrategy under air levels of external CO2 but not under\nvery low CO2, as evidenced by the severe growth defects of\nthe lcib mutant under air levels of CO2 but not very low\nCO2 (Wang and Spalding 2006, 2014a, 2014b; Duanmu\net al. 2009; Kono and Spalding 2020). In contrast to the pas-\nsive CO2 uptake strategy, modeling suggests that the active\n− uptake strategy can be effective and energetically effi-\nHCO3\ncient under both growth conditions (Fei et al. 2022).\nIntriguingly, despite the predicted good performance of the\n− uptake strategy under air levels of CO2 in silico,\nactive HCO3\nin vivo Chlamydomonas appears to reserve this strategy only\nfor very low CO2 conditions, as evidenced by O2 evolution ex-\nperiments (Wang and Spalding 2014a; Yamano et al. 2015;\n\nFigure 2. Operating principles of the pyrenoid-based CCM. The\nChlamydomonas CO2-concentrating mechanism is shown; the basic\nprinciples are likely to apply in other species, although due to conver-\ngent evolution, the specific proteins that mediate some of the reactions\nmay be phylogenetically unrelated to those in Chlamydomonas.\nMutant phenotypes and biophysical modeling (Fei et al. 2022) support\nthe existence of two operating modes of a pyrenoid-based\n− is ac-\nCO2-concentrating mechanism, which differ based on how HCO3\ncumulated in the chloroplast stroma. A) The first mode uses a passive\nchloroplast CO2 uptake strategy, where CO2 passively diffuses across\nthe chloroplast envelope into the stroma and is converted into\n− by the LCIB/LCIC carbonic anhydrase complex. This strategy is\nHCO3\nused under low CO2 (ambient air levels of external CO2). B) The second\n− uptake strategy, which relies on\nmode uses an active chloroplast HCO3\n− into the chloroplast. This strategy is used un-\nactive pumping of HCO3\nder very low CO2.\n\n2003; Blanco-Rivero et al. 2012; Burlacot et al. 2022). This CO2\ndiffuses out of the thylakoid membranes and into the pyre-\nnoid matrix, where it is captured by Rubisco. A CO2 leakage\nbarrier is thought to slow the escape of CO2 from the pyre-\nnoid, increasing CO2 concentration and decreasing energetic\ncosts (Fei et al. 2022). In the case of Chlamydomonas, mod-\neling and experimental evidence suggest that the starch\nsheath (Toyokawa et al. 2020; Fei et al. 2022) and thylakoid\nmembrane sheets (Fridlyand 1997; Fei et al. 2022) can serve\nas CO2 leakage barriers.\n\nThe two pyrenoid-based CCM operating modes use differ-\n− in the chloroplast\n\nent strategies to accumulate HCO3\n\nAB\f3240\n\n|\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\nHe et al.\n\nTable 1. Summary of Chlamydomonas pyrenoid–specific proteins whose localization has been confirmed.\n\nProtein\nname\n\nABCF6\nCPLD2\n\nEPYC1/\nLCI5\nHDA5\nrbcL\nRBCS1\nRBCS2\nRCA1\n\nGene ID\n\nSubpyrenoid localization\n\nLocalization reference\n\nReported or predicted functions\n\nCre06.g271850\nCre03.g206550\n\nCre10.g436550\n\nCre06.g290400\nCreCp.g802313\nCre02.g120100\nCre02.g120150\nCre04.g229300\n\nMatrix\nMatrix\n\nMatrix\n\nMatrix\nMatrix\n\nLau et al. (2023)\n\n①\n① Wang et al. (2022)\n\nPredicted ATP-binding cassette family F like protein\nHomolog of the Arabidopsis XuBP phosphatase\n\nCbbY (At3g48420)\n\n①\n\nMackinder et al. (2016)\n\nRubisco linker, phase-separates with Rubisco to form\n\nthe pyrenoid matrix\n\n① Wang et al. (2022)\n①\n\nHoldsworth (1971); Mackinder\n\net al. (2017)\n\nPredicted histone deacetylase\nLarge and small subunits of the Rubisco holoenzyme,\nwhich fixes CO2 and produces 3-PG and 2-PG\n\nMatrix\n\n①\n\nVladimirova et al. (1982);\n\nRubisco activase\n\nLacoste-Royal (1987); McKay\net al. (1991); Mackinder et al.\n(2017)\n\nSTR16\n\nCre13.g573250\n\nSTR18\n\nCre16.g663150\n\n–\n–\nCAH3\nCAS1\n\nCre16.g648400\nCre02.g143635\nCre09.g415700\nCre12.g497300\n\nCYN7\nCYN20-6\nDEG8\n\nCre12.g544150\nCre12.g544114\nCre01.g028350\n\nHCF136\nPSAH\nPSBP1\n\nCre06. g273700\nCre07. g330250\nCre12.g550850\n\nRBMP1\n\nCre06. g261750\n\nTEF14\n\nCre06.g256250\n\n-\n\nCre03.g172700\n\nCGLD14\n\nCre10.g446350\n\nPNU1\n\nCre03.g183550\n\nRBMP2\n\nCre09.g416850\n\nPSBP4\n\nCre08.g362900\n\nSAGA1\n\nCre11.g467712\n\nSAGA2\n\nCre09.g394621\n\nSMC7\n\nCre17.g720450\n\nSBE3\nSTA2\n\nCre10.g444700\nCre17.g721500\n\nMatrix\n\nMatrix\n\nMatrix\nMatrix\nTubules\nTubules\n\nTubules\nTubules\nTubules\n\nTubules\nTubules\nTubules\n\nTubules\n\nTubules\n\nTubules\n\nPyrenoid center\n(reticulated region)\nPyrenoid center\n(reticulated region)\nTubules (reticulated\nregion)\nThylakoid lumen\npuncta\nPuncta surrounding the\nmatrix\nInterface between\nmatrix and starch\nsheath\nPuncta surrounding the\nmatrix\nStarch sheath\nStarch sheath\n\n⑤\n\n⑨\n\n④\n\n④\n\n④\n\n③\n③\n\n① Wang et al. (2022); Lau et al.\n\nPredicted thiosulfate sulfurtransferase containing a\n\n(2023)\n\nrhodanese domain\n\n① Wang et al. (2022); Lau et al.\n\nPredicted thiosulfate sulfurtransferase containing a\n\n(2023)\n① Wang et al. (2022)\n① Wang et al. (2022)\n②\n② Wang et al. (2016)\n\nSinetova et al. (2012)\n\n② Wang et al. (2022)\n② Wang et al. (2022)\n② Wang et al. (2022)\n\n② Wang et al. (2022)\n②\n② Wang et al. (2022)\n\nMackinder et al. (2017)\n\nrhodanese domain\n\n–\n–\nAlpha-type carbonic anhydrase\nCalcium-mediated regulator of the expression of\n\nsome CCM-related genes\n\nPredicted peptidyl-prolyl cis-trans isomerase\nPredicted peptidyl-prolyl cis-trans isomerase\nPredicted DegP-type protease, Arabidopsis homolog\nis involved in the degradation of photodamaged\nPSII reaction center protein D1\nPS II stability/assembly factor HCF136\nPSI subunit\nPredicted oxygen-evolving enhancer protein 2 of PS\n\nII\n\n②\n\nMeyer et al. (2020b)\n\nProposed to mediate pyrenoid matrix connection to\n\ntubules\n\n② Wang et al. (2022)\n\nPredicted thylakoid-luminal protein, no labeled\n\ndomains\n\n②\n\nLau et al. (2023)\n\nProtein with multiple predicted Rubisco-binding\n\nmotifs\n\n⑤ Wang et al. (2022)\n\nPSBP domain-containing protein 3, conserved in the\n\n⑤ Wang et al. (2022)\n\nPROTEIN F23H11.5, has a bifunctional nuclease\n\ngreen lineage and diatoms\n\ndomain\n\nMeyer et al. (2020b)\n\nProposed to mediate pyrenoid matrix connection to\n\ntubules\n\nMackinder et al. (2017)\n\nLuminal PsbP-like protein, Arabidopsis homolog is\n\nessential for PS I assembly and function\n\nItakura et al. (2019); Meyer et al.\n\nProposed to mediate adherence of the starch sheath\n\n(2020b)\n\nto the matrix\n\nMeyer et al. (2020b)\n\nProposed to mediate adherence of the starch sheath\n\nto the matrix\n\nLau et al. (2023)\n\nProposed to have a similar function as SAGA1\n\nMackinder et al. (2017)\nMackinder et al. (2017)\n\n–\n\n–\n\nCre09.g394547\n\nStarch sheath\n\n③ Wang et al. (2022)\n\nCre09.g415600\n\nStarch sheath\n\n③ Wang et al. (2022)\n\nConserved starch-branching enzyme\nGranule-bound starch synthase involved in amylose\nbiosynthesis and the biosynthesis of long chains in\namylopectin\n\nPredicted cyclomaltodextrin glucanotransferase/\n\ncyclodextrin glycosyltransferase\n\nPredicted glucan 1,4-alpha-glucosidase/Lysosomal\nalpha-glucosidase; has a starch-binding domain\n\n(continued)\n\n\fThe eukaryotic CO2-concentrating organelle\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\n|\n\n3241\n\nTable 1. (continued)\n\nGene ID\n\nSubpyrenoid localization\n\nLocalization reference\n\nReported or predicted functions\n\nProtein\nname\n\nLCI9\n\nLCIB\n\nCre09.g394473\n\nCre10.g452800\n\nLCIC\n\nCre06.g307500\n\n–\n\nCre09.g394510\n\nGaps between starch\nplates\nGaps between starch\nplates and tubules (very\nlow CO2)\nGaps between starch\nplates and tubules (very\nlow CO2)\nStarch-matrix interface\nand gaps between\nstarch plates\n\n⑦\n\n⑧\n\n⑧\n\nMackinder et al. (2017)\n\nContains 2 starch-binding domains; may help ensure\n\nYamano et al. (2010); Wang and\nSpalding (2014b); Mackinder\net al. (2017)\n\na close fit for adjacent starch plates\n\nBeta-type carbonic anhydrase\n\nYamano et al. (2010); Mackinder\n\nBeta-type carbonic anhydrase\n\net al. (2017)\n\n④⑦⑧ Lau et al. (2023)\n\nContains a CBM20 starch-binding domain and a\nt-SNARE domain; proposed to be involved in\nmembrane remodeling of the pyrenoid tubules;\ncould be involved in membrane remodeling of the\npyrenoid tubules\nMalate dehydrogenase\nHomolog of the Arabidopsis chloroplast division site\n\nregulator MinD1\n\nMDH1\nMIND1\n\nCre03.g194850\nCre12.g522950\n\nPyrenoid periphery\nPyrenoid periphery\n\n–\n–\n\nWang et al. (2022)\nWang et al. (2022)\n\nThe circled numbers indicating the subpyrenoid localizations are coordinated with the circled numbers in Fig. 1C.\n\nKono and Spalding 2020). A possible explanation for this ob-\nservation is that when Chlamydomonas is grown under air\n− uptake strategy may\nlevel CO2 conditions, the active HCO3\nincur additional energetic costs beyond those accounted\nfor in the model. Indeed, the model only considered energet-\nic costs in the chloroplast, and it is possible that under cer-\n− uptake strategy requires\ntain conditions the active HCO3\nenergetic input outside the chloroplast, for example to\npump HCO3\n\n− across the plasma membrane.\n\nThe operation of a pyrenoid-based CCM under either air or\nvery low CO2 is estimated to be feasible for as little as the en-\nergetic equivalent of approximately 1 ATP per CO2 fixed (Fei\net al. 2022), making it energetically inexpensive relative to the\noverall cost of CO2 fixation by the Calvin-Benson-Bassham\ncycle, which is approximately energetically equivalent to 9\nATPs per CO2 fixed (Mangan et al. 2016).\n\nThe energy for operating the CCM must ultimately come\nfrom the light reactions, which directly drive the pH differ-\nence between the thylakoid lumen and stroma and indirectly\nmaintain the pH of other compartments and drive the activ-\nities of transporters. A recent study suggested that the pro-\n− to CO2 in the\ntons needed to drive conversion of HCO3\nthylakoid lumen are produced by photosynthetic cyclic elec-\ntron flow mediated by proton gradient regulation-like 1\n(PGRL1, Cre07.g340200) and pseudocyclic electron flow re-\nsulting from O2 photoreduction mediated by flavodiiron\nand Cre16.g691800)\nproteins\n(Burlacot et al. 2022). The same study also found that\nchloroplast-to-mitochondria electron flow contributes to\nenergizing the CCM, potentially by supplying ATP to drive\ntransporters.\n\n(FLVs, Cre12.g531900\n\nRubisco fixes CO2 through carboxylation of ribulose-1,5-bi-\nsphosphate (RuBP) to produce 3-phosphoglycerate (3-PG or\n3-PGA). While some 3-PG goes on to other parts of metabol-\nism, most of it is metabolized in the Calvin-Benson-Bassham\ncycle to regenerate RuBP, allowing the cycle to continue\n\n(Calvin 1962). Interestingly, in Chlamydomonas, Rubisco is\nthe only enzyme of the Calvin-Benson-Bassham cycle found\nin the pyrenoid; all other Calvin-Benson-Bassham cycle en-\nzymes and associated regulatory proteins that have been lo-\ncalized are enriched in a region of the stroma immediately\nsurrounding the pyrenoid (Fig. 1) (Küken et al. 2018; Wang\net al. 2022). Thus, the Rubisco substrate RuBP and its product\n3-PG need to exchange efficiently between the stroma and\nthe pyrenoid. The pathway and mechanism of this exchange\nare currently unknown in any organism with a pyrenoid.\n\nPyrenoids are likely the product of convergent\nevolution\nThe predominant theory regarding the origins of pyrenoids is\nbased on historical changes in atmospheric CO2 and O2 con-\ncentrations and the evolution of polyphyletic algal lineages.\nOxygenic photosynthesis first evolved in cyanobacteria ap-\nproximately 3 billion years ago (bya) (Schirrmeister et al.\n2015) at a time when atmospheric CO2 concentrations\nwere high and O2 concentrations were low (Fig. 3). The ad-\nvent of oxygenic photosynthesis led to the Great Oxidation\nEvent approximately 2.4 bya, when atmospheric O2 concen-\ntrations first began to rise (Anbar et al. 2007). Eukaryotic al-\ngae are thought to have first evolved approximately 1.5 to 2.0\nbya (Yoon et al. 2004; Sánchez-Baracaldo et al. 2017; Strassert\net al. 2021), a time when the atmospheric CO2:O2 ratio was\nstill high and CCMs were likely not necessary for efficient\ngrowth (Raven et al. 2017). Over the course of approximately\nthe next billion years, a diverse set of algal lineages arose\nseries of endosymbiotic events\nthrough a complex\n(Falkowski et al. 2004; Reyes-Prieto et al. 2007; Keeling\n2010; Dorrell et al. 2017; Jackson et al. 2018; Strassert et al.\n2021). These different lineages can be split into two broad\ngroups, green lineage algae and red lineage algae, which are\n\n\f3242\n\n|\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\nHe et al.\n\nFigure 3. Pyrenoids appear to have convergently evolved in response to declining atmospheric CO2 levels. Approximate CO2 and O2 concentrations\nover time (Berner 2006; Whitney et al. 2011) are correlated with the phylogenetic tree of photosynthetic eukaryotes below (Strassert et al. 2021).\nBranch points correlate with the approximate timing of the divergence of different groups (see Strassert et al. 2021 and Bowles et al. 2022 for dis-\ncussions on uncertainties regarding branch points). Asterisks denote the approximate timing of the acquisition of plastids through primary (*) or\nsecondary (**) endosymbiosis (Jackson et al. 2018; Strassert et al. 2021). The blue shade highlights the proposed range for the timing of CCM evo-\nlution in different photosynthetic species (Villarreal and Renner 2012; Meyer et al. 2020a). Green and red lineages are denoted as green or purple\ntext, respectively (most dinoflagellates have red plastids with the exception of Lepidodinium sp., which have green plastids) (Kamikawa et al. 2015).\nRepresentative electron micrographs of pyrenoids are shown below cartoons of four general pyrenoid types, displaying the wide variety of morph-\nologies observed in each algal lineage and the hornworts. Roman numerals on the electron micrographs denote references to their original pub-\nlications as follows: (I) Kusel-Fetzmann and Weidinger 2008; (II) Nudelman et al. 2006; (III) Zhang et al. 2008; (IV) van Baren et al. 2016; (V)\nBorowitzka 2018; (VI) Goudet et al. 2020; (VII) Mikhailyuk et al. 2014; (VIII) Duff et al. 2007; (IX) Hall and Claus 1963; (X) Nelson and Ryan 1988;\n(XI) Ford 1984; (XII) Laza-Martínez et al. 2012; (XIII) Clay and Kugrens 1999; (XIV) Shiratori et al. 2017; (XV) Ota et al. 2007; (XVI) Schnepf and\nElbräChter 1999; (XVII) Kowallik 1969; (XVIII) Hansen and Daugbjerg 2009; (XIX) Decelle et al. 2021; (XX) Bedoshvili et al. 2009; (XXI) Bedoshvili\nand Likhoshway 2012; (XXII) Buma et al. 2000; (XXIII) Fresnel and Probert 2005. This figure was created with BioRender.\n\n\fThe eukaryotic CO2-concentrating organelle\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\n|\n\n3243\n\ndistinguished by their use of different chlorophyll accessory\npigments (Falkowski et al. 2004; Keeling 2010).\n\nDuring the time that these algal lineages were evolving,\nCO2 concentrations were trending downward (Berner and\nKothavala 2001; Berner 2006), a phenomenon accelerated\nby the evolution of land plants between 500 and 360 million\nyears ago (mya), after each algal lineage had already been es-\ntablished (Fig. 3) (Berner 1997; Morris et al. 2018). This de-\ncrease in atmospheric CO2 and the simultaneous increase\nin atmospheric O2 are thought to be the main driving forces\nfor the evolution of CCMs in aquatic microorganisms, leading\nto the theory that pyrenoids and other CCMs evolved inde-\npendently via convergent evolution (Villarreal and Renner\n2012; Rae et al. 2013; Raven et al. 2017; Meyer et al. 2020a).\nThis convergent evolution theory potentially explains why\npyrenoids first evolved, but it remains difficult to pinpoint\nthe exact timing of their origin. In the absence of concrete\nevidence from the fossil record (Knoll 1992), previous reviews\nhave estimated the most likely timeline for CCM evolution\nby considering how various factors—including fluctuating\nCO2 and O2 concentrations, temperature changes, nutrient\nlevels, and the kinetic properties of different forms of\nRubisco—would have influenced the growth advantage con-\nferred by a CCM (Griffiths et al. 2017; Raven et al. 2017; Meyer\net al. 2020a). These reviews estimate that CCMs may have\nevolved in cyanobacteria and algae approximately 300 to\n450 mya (Badger and Price 2003; Griffiths et al. 2017), likely\nwhen the atmospheric CO2 concentration was 2 to 16 times\nthe present level (Raven et al. 2017). In hornworts (discussed\nfurther below), pyrenoids are estimated to have evolved ap-\nproximately 100 mya (Villarreal and Renner 2012). These es-\ntimates all correspond\ntime when different\nphotosynthetic lineages were already established and sup-\nport the convergent evolution theory.\n\nto a\n\nThe convergent evolution theory can be tested by compar-\ning the sequences of proteins thought to perform the same\nfunctions in the pyrenoids of phylogenetically distant algal\nspecies, but this is currently difficult to do because the mo-\nlecular composition of most pyrenoids is unknown. There\nare, however, three lines of molecular evidence that support\nthe convergent evolution theory. The first\nis that\nthylakoid-luminal carbonic anhydrases of different types\nare necessary for CCM function in different lineages: the\nChlamydomonas CCM requires the alpha-type carbonic an-\nhydrase CAH3 (Karlsson et al. 1998; Hanson et al. 2003;\nSinetova et al. 2012), whereas the diatom P. tricornutum re-\nquires a theta-type carbonic anhydrase (Kikutani et al.\n2016; Matsuda et al. 2017).\n\nThe second piece of molecular evidence that supports con-\nvergent evolution is based on the different forms of Rubisco\nacross lineages. There are at least four distinct types of\nRubisco enzymes within algae and cyanobacteria, which dif-\nfer greatly in their kinetic properties and holoenzyme struc-\nture (Badger et al. 1998). The fact that pyrenoids in different\nlineages package vastly different Rubisco holoenzymes sup-\nports the theory that they evolved convergently.\n\nis\n\nto\n\nThe\n\nrelated\n\nthird piece of evidence\n\nthe\nChlamydomonas protein Essential Pyrenoid Component 1\n(EPYC1, Cre10.g436550, also known as LCI5) (Turkina et al.\n2006; Mackinder et al. 2016), which is a linker protein that\nclusters Rubisco together to form the pyrenoid matrix\n(Mackinder et al. 2016; He et al. 2020). EPYC1 is necessary\nfor the Chlamydomonas CCM, but no homologs of this pro-\ntein could be identified in algae beyond the closely related\nVolvocales (Mackinder et al. 2016), suggesting that its func-\ntion is performed by other proteins that may have conver-\ngently evolved in different algal species. Repeat proteins\nwith similar predicted properties as EPYC1 have been identi-\nfied in other algae (Mackinder et al. 2016), and work is on-\nlinker\ngoing to characterize these and other putative\nproteins.\n\nIn addition to research on the algal CCM, several studies\nhave been conducted on the evolution of CCMs in horn-\nworts, which are the only land plants known to have pyre-\nnoids. Hornworts are nonvascular plants thought to have\nbeen important in the water-to-land transition during em-\nbryophyte evolution (Qiu et al. 2006). The first hornwort pyr-\nenoids evolved approximately 100 mya (Villarreal and Renner\n2012), coinciding with a drastic decline in atmospheric CO2\nlevels (Fig. 3). The presence of a pyrenoid in hornworts is cor-\nrelated with CCM activity detected by organic isotope dis-\ncrimination and mass spectrometry analyses (Smith and\nGriffiths 1996a, 1996b, 2000; Hanson et al. 2002; Meyer\net al. 2008), suggesting that pyrenoids play a similar role in\nhornwort CCMs as they do in algal CCMs. Phylogenetic evi-\ndence and ultrastructural data suggest that hornwort pyre-\nnoids were gained and lost 5 to 6 times independently\nsince they first appeared (Villarreal and Renner 2012).\nInterestingly, the distribution of pyrenoids across the green\nlineage of algae also suggests that multiple independent\nlosses and gains have occurred (Meyer and Griffiths 2013).\nThe environmental factors favoring pyrenoid loss remain\nunclear.\n\nStructure and components\nPyrenoid morphology can vary greatly depending on the spe-\ncies (Fig. 3), but the one unifying feature of all pyrenoids is\nthe Rubisco matrix, which contains densely packed Rubisco\n(Holdsworth 1971; Borkhsenious et al. 1998). Most algae\nhave one matrix per cell, although some species have mul-\ntiple Rubisco matrices that can vary in size and shape\n(Meyer et al. 2020a). The Rubisco matrix in all observed spe-\ncies is associated with thylakoid membranes (Meyer et al.\n2017), which are thought to deliver CO2 to Rubisco (Fig. 2)\n(Hennacy and Jonikas 2020). The simplest pyrenoids consist\nof a Rubisco matrix either embedded between thylakoid\nmembranes, such as that of the coccolithophore Emiliania\nhuxleyi (Buma et al. 2000), or projecting out of the chloro-\nplast into the cytoplasm, such as that of the diatom\nAttheya ussurensis (Bedoshvili et al. 2009) (Fig. 3). In most\nspecies, the thylakoid membranes traverse the pyrenoid\n\n\f3244\n\n|\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\nHe et al.\n\nmatrix either as sheets, as in the euglenophyte Euglena carter-\nae (Kusel-Fetzmann and Weidinger 2008), or tube-like struc-\ntures, as in the chlorophyte Heveochlorella hainangensis\n(Zhang et al. 2008). In some species, Rubisco matrices lack\ntraversing thylakoids but are surrounded by polysaccharide\ndeposits that potentially act as CO2 leakage barriers, as in\nthe chlorophyte Micromonas pusilla (van Baren et al. 2016).\nThe most elaborate pyrenoid morphologies consist of all\nthree sub-structures: thylakoids traversing a Rubisco matrix\nin\nthat\nChlamydomonas (Fig. 1B). In this section, we describe what\nis known about each of the pyrenoid sub-compartments in\nChlamydomonas and give a brief introduction to what is\nknown about these structures in other species.\n\nin a starch sheath, as\n\nis encased\n\nfound\n\nThe CO2-fixing pyrenoid matrix is a phase-separated\ncondensate of Rubisco and a linker protein\nRubisco makes up approximately 90% of the protein content\nof the pyrenoid matrix (Holdsworth 1971). Based on trans-\nmission electron microscopy (TEM) images, the matrix ap-\nin several species (Holdsworth 1968;\npears crystalline\nKowallik 1969; Bertagnolli and Nadakavukaren 1970) and\namorphous in others (Griffiths 1970; Meyer et al. 2012).\n\nThe Chlamydomonas pyrenoid matrix was recently shown\nto be a liquid-like phase-separated condensate (Freeman\nRosenzweig et al. 2017). Fluorescence recovery after photo-\nbleaching experiments indicated that the matrix mixes intern-\nally on a timescale of approximately 20 seconds (Freeman\nRosenzweig et al. 2017), similar to that observed for other\nliquid-like compartments such as P granules and nucleoli\n(Bracha et al. 2019). Furthermore, the Rubisco matrix exhibits\nother liquid-like behaviors, including division by fission, and\ndissolution into the chloroplast during cell division and\nunder high CO2 conditions (>0.40% CO2). Rubisco in the\nChlamydomonas pyrenoid matrix was found by cryo-electron\ntomography (cryo-ET) to lack the long-range order character-\nistic of a crystal; instead, the distribution of Rubisco fits well\nwith a simple model for the distribution of particles in a liquid\n(Freeman Rosenzweig et al. 2017). This description of the pyr-\nenoid matrix as a phase-separated condensate likely applies to\npyrenoids in other species, as it explains observations such as\nthe spheroidal shape of most pyrenoids (Fig. 3), the rapid ap-\npearance and disappearance of pyrenoids during cell division\n(Brown et al. 1967; Retallack and Butler 1970), and their div-\nision by fission (Brown and Bold 1964; Brown et al. 1967).\n\nFor decades, only two matrix proteins were known:\nRubisco and Rubisco activase (RCA1, Cre04.g229300)\n(Holdsworth 1971; Vladimirova et al. 1982; McKay and\nGibbs 1991), and the mechanism by which Rubisco is densely\nclustered in the pyrenoid matrix was a mystery. In 2016, the\nrepeat protein EPYC1 was proposed to link individual\nRubiscos\nin Chlamydomonas\nthe matrix\n(Mackinder et al. 2016). EPYC1 localizes to the pyrenoid ma-\ntrix and is one of the most abundant proteins in the pyrenoid\n\nform\n\nto\n\nafter Rubisco (Mackinder et al. 2016; Hammel et al. 2018). In\nepyc1 mutant cells, the majority of Rubisco is dispersed in the\nchloroplast outside of the pyrenoid, indicating that EPYC1\nplays a major role in Rubisco localization to the matrix\n(Mackinder et al. 2016). Purified Rubisco and EPYC1 can\nphase-separate with each other to form liquid-like droplets\nin vitro (Wunder et al. 2018), suggesting that these two pro-\nteins are sufficient for driving the formation of the liquid-like\npyrenoid matrix.\n\nRubisco is an oligomeric holoenzyme with eight identical\nlarge subunits and eight identical small subunits. A structural\nstudy using cryo-electron microscopy (cryo-EM) found that\nEPYC1 directly binds to Rubisco on the two alpha-helices\nof each Rubisco small subunit through salt bridges and\nhydrophobic interactions (Fig. 4) (He et al. 2020). This struc-\nture is consistent with previous genetic studies showing that\nthese alpha helices are important for the formation of the\npyrenoid and for the Rubisco–EPYC1 interaction (Meyer\net al. 2012; Atkinson et al. 2019). Each Rubisco holoenzyme\nhas eight EPYC1-binding sites and each EPYC1 has five\nRubisco-binding sites (Fig. 4, C to E), allowing the two pro-\nteins to form an interdependent network that clusters\nRubisco together in the pyrenoid matrix. The low binding\naffinity (approximately 3 mM) between individual EPYC1–\nRubisco binding site is consistent with the principle that bio-\npolymer phase separation is mediated by weak multivalent\ninteractions (Li et al. 2012).\n\nin\n\nin\n\nlocalization or enrichment\n\nthe Plantae and diatoms 2\n\nBeyond Rubisco, RCA1, and EPYC1, ten additional proteins\nshow exclusive\nthe\nChlamydomonas pyrenoid matrix when examined as fluores-\ncently tagged proteins (Fig. 1; Table 1). These proteins are the\nputative S-adenosyl-L-methionine-dependent methyltrans-\nferase SMM7 (Cre03.g151650) (Mackinder et al. 2017), the\npredicted xylulose-1,5-bisphosphate (XuBP) phosphatase\n(CPLD2,\nconserved\nCre03.g206550), the putative histone deacetylase HDA5\n(Cre06.g290400), uncharacterized proteins encoded by\nCre16.g648400, Cre13.g573250, Cre16.g663150, Cre02.g143635\n(Wang et al. 2022), thiosulfate sulfurtransferase16 (STR16,\nCre13.g573250), STR18 (Cre16.g663150), and ATP-binding\ncassette F like-protein 6 (ABCF6, Cre06.g271850) (Lau et al.\n2023) (Table 1). CPLD2 is the homolog of the highly selective\nArabidopsis XuBP phosphatase AtCbbY (At3g48420), which\nconverts the Rubisco inhibitor XuBP to a non-inhibitory\ncompound that can be recycled back to the Rubisco sub-\nstrate RuBP (Bracher et al. 2015). The pyrenoid localization\nof CPLD2 suggests that it may also convert XuBP to RuBP\nin the pyrenoid. Both HDA5 and the protein encoded by\nCre16.g648400 have predicted “Rubisco-binding motifs”\n(Fig. 4) (discussed in a later section) (Meyer et al. 2020b), sug-\ngesting that they bind Rubisco in the pyrenoid matrix (Wang\net al. 2022). STR16, STR18, and the proteins encoded by\nCre13.g573250 and Cre16.g663150 are all predicted to be\nthiosulfate sulfurtransferases. In addition, STR16 and STR18\ncontain a rhodanese domain, which is predicted to function\n\n\fThe eukaryotic CO2-concentrating organelle\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\n|\n\n3245\n\nFigure 4. A common Rubisco-binding motif mediates the assembly of the major compartments of the pyrenoid. A) In Chlamydomonas, many\npyrenoid-localized proteins contain at least one Rubisco-binding motif. B) The Rubisco-binding motif mediates the assembly of the three pyrenoid\nsub-compartments. The motifs on EPYC1 link Rubisco to form the pyrenoid matrix (He et al. 2020). The motifs on the tubule-localized transmem-\nbrane proteins RBMP1 and RBMP2 are proposed to connect the Rubisco to the tubules, and the motifs on the putative starch-binding proteins\nSAGA1 and SAGA2 are proposed to mediate interactions between the matrix and the surrounding starch sheath. A Rubisco-binding motif was\nalso shown to be necessary and sufficient to target a nascent protein to the pyrenoid (Meyer et al. 2020b). C) A model illustrating how EPYC1\n(red) clusters Rubisco (blue) in the pyrenoid matrix. D) The Rubisco-binding motif of EPYC1 (red) binds to the Rubisco small subunit (dark\nblue) (He et al. 2020); other Rubisco-binding motifs in Chlamydomonas are expected to bind to the same site. E) The motif binds between two\nalpha-helices of the Rubisco small subunit.\n\nin disulfide bond formation and iron-sulfur cluster biosyn-\nthesis (Lau et al. 2023). ABCF6 is a predicted member of\nthe ATP-binding cassette F\nfamily\nthat regulates translation via binding to ribosomes (Lau\net al. 2023).\n\n(ABCF) protein\n\nPyrenoid-associated membranes likely supply\nRubisco with concentrated CO2\nThe Rubisco matrix in all known pyrenoids is in contact with a\nportion of the thylakoid membranes of the chloroplast (Meyer\net al. 2017), consistent with the idea that these membranes\n\nperform the essential function of supplying CO2 to Rubisco\n(Pronina and Semenenko 1990; Raven 2008). A broad range\nof morphologies has been observed in different species for\nthese pyrenoid-associated thylakoids (Fig. 3). Some species\nhave a single traversing membrane, some have multiple parallel\nor interconnected membranes, and some have more complex\nmorphologies, such as the undulating membranes found in\nspecies of the red algal genus Porphyridium (Nelson and\nRyan 1988) (Fig. 3). Some pyrenoids, such as in the dinoflagel-\nlate Podolampas bipes (Schnepf and ElbräChter 1999) (Fig. 3),\nhave no observed traversing membranes and instead are\n\nACDEB\f3246\n\n|\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\nHe et al.\n\nembedded between thylakoid membranes in the chloroplast.\nGiven that CO2 diffuses rapidly relative to the rate of its fixation\nby Rubisco, the exact location of CO2 release within the pyre-\nnoid likely has little effect on the distribution of CO2 within the\npyrenoid; thus, a broad range of membrane morphologies can\neffectively supply CO2 to Rubisco (Fei et al. 2022).\n\nIn Chlamydomonas, the thylakoid membranes extend into\nthe Rubisco matrix to form pyrenoid tubules whose lumina\nare continuous with the thylakoid lumen (Fig. 1, C and D)\n(Sager and Palade 1954, 1957; Ohad et al. 1967a).\nTraditional 2D TEM images have shown that thylakoid sheets\nnear the pyrenoid are directed toward the gaps of the starch\nsheath (Sager and Palade 1954, 1957; Ohad et al. 1967a). This\nobservation was corroborated by cryo-ET, which showed in\n3D that thylakoid sheets merge with each other as they\npass through gaps in the starch sheath to form cylindrical\npyrenoid tubules that traverse the Rubisco matrix (Fig. 1, C\nand D) (Engel et al. 2015). In the center of the pyrenoid,\nthe tubules converge to form a complex interconnected net-\nwork known as the reticulated region (Fig. 1, C and D) (Engel\net al. 2015; Meyer et al. 2020b).\n\nInside each Chlamydomonas tubule, there are two to eight\nsmaller tubes called minitubules (Ohad et al. 1967b; Engel et al.\n2015) (Fig. 1, C and D). Minitubules appear to provide con-\nduits from the inter-thylakoid stromal space to the pyrenoid\nmatrix (Engel et al. 2015). These minitubules have been pro-\nposed to facilitate the diffusion of small molecules such as\nRuBP and 3-PG between these two compartments (Engel\net al. 2015; Küken et al. 2018); however, their internal dia-\nmeters of approximately 3.5 ± 0.5 nm are likely too small to\nmediate a substantial flux of metabolites. Curiously, minitu-\nbules have not been observed in any species other than\nChlamydomonas (Meyer et al. 2017), although this could be\ndue to limitations in TEM imaging. The function of minitu-\nbules remains unknown, and the mechanisms by which the in-\ntricate morphology of\nthe\npyrenoid-traversing membranes is achieved is unknown in\nany organism.\n\nthe different\n\nregions of\n\nIt is notable that Chlamydomonas mutants that lack a\nRubisco matrix still have tubule networks in the canonical lo-\ncation within the chloroplast (Caspari et al. 2017; He et al.\n2020), indicating that the tubules can form in the absence\nof the matrix and suggesting that the location of the pyre-\nnoid could be determined by the tubules. However, the mo-\nlecular basis for the localization of the tubules remains\nunknown.\n\nThe delivery of concentrated CO2 to Rubisco in the matrix\nis thought to be mediated by carbonic anhydrases that con-\n− into CO2 in the lumen of the pyrenoid-traversing\nvert HCO3\nmembranes (Pronina and Semenenko 1990; Raven 2008). In\nChlamydomonas, the carbonic anhydrase that mediates\nthis key step is CAH3 (Karlsson et al. 1998; Hanson et al.\n2003). Consistent with this role, Chlamydomonas mutants\nlacking functional CAH3 have a severe growth defect when\ngrown in limiting CO2 conditions (Spalding et al. 1983;\nFunke et al. 1997; Karlsson et al. 1998) and over-accumulate\n\n− within the mutant cells (Spalding et al. 1983). CAH3\nHCO3\nlocalizes to the thylakoid lumen (Karlsson et al. 1998) and be-\ncomes enriched in the pyrenoid tubules during activation of\nthe CCM in transitions from high CO2 to limiting CO2\n(Blanco-Rivero et al. 2012; Sinetova et al. 2012) and from\ndark to light (Mitchell et al. 2014). How CAH3 relocalizes\nto the tubules remains unknown.\n\nRecent studies in Chlamydomonas have identified add-\nitional pyrenoid tubule–localized proteins that perform vari-\nous functions. Like CAH3, the Ca2+-binding protein calcium\nsensing receptor (CAS, Cre12.g497300) relocalizes to the pyr-\nenoid tubules upon CCM induction (Wang et al. 2016;\nYamano et al. 2018). CAS is a putative Rhodanese-like Ca2+-\nsensing receptor that regulates the expression of several\nCCM-related genes, including HLA3 and LCIA (Wang et al.\n2016). Upon activation of the CCM, CAS switches from being\ndispersed across the chloroplast to being associated with the\npyrenoid tubules (Wang et al. 2016). This change is accom-\npanied by an increase in Ca2+ in the pyrenoid (Wang et al.\n2016). The role of Ca2+ in the pyrenoid, the mechanism of\nCAS relocalization, and the purpose of CAS signaling in the\nCCM remain unclear.\n\nTwo other tubule-localized proteins are Rubisco-binding\nmembrane protein 1 (RBMP1, encoded by Cre06.g261750)\nand RBMP2 (Cre09.g416850) (Fig. 4). Both proteins bind to\nRubisco\nin vivo\nin vitro (Meyer et al. 2020b) and\n(Mackinder et al. 2017), suggesting that they may promote\ninteractions between the Rubisco matrix and the pyrenoid\ntubules. RBMP1 is associated with peripheral tubular regions,\nwhereas RBMP2 localizes to the central reticulated region of\nthe tubules (Meyer et al. 2020b). Intriguingly, RBMP2 con-\ntains a rhodanese domain, as do STR16, STR18, and CAS,\nbut the function of these rhodanese domains in these pro-\nteins remains to be determined. The putative roles of\nRBMP1 and RBMP2 in linking matrix to tubules also remain\nto be tested.\n\nIn addition to RBMP2, the proteins pyrenoid nuclease 1\n(PNU1, Cre03.g183550) and conserved in the green lineage\nand diatoms 14 (CGLD14, Cre10.g446350) appear to localize\nto the reticulated region of the tubules (Wang et al. 2022).\nPNU1 is a bifunctional nuclease domain-containing protein.\nAs oxidized RNA was also localized to the pyrenoid in\nChlamydomonas (Zhan et al. 2015), the pyrenoid localization\nof PNU1 suggests that the pyrenoid might be a site of oxi-\ndized RNA degradation (Wang et al. 2022). CGLD14 is con-\nserved in the green lineage and diatoms and is also named\nPSBP-domain-containing protein 3 (PPD3).\n\nMultiple components of the electron transport chain are\npresent in the Chlamydomonas pyrenoid tubules, including\nsubunits of photosystem I (PSI) (Photosystem I reaction cen-\nter subunit V [PSAG, Cre12.g560950], Photosystem I reaction\ncenter subunit H [PSAH, Cre07.g330250], Photosystem I reac-\ntion center\n[PSAK, Cre17.g724300], and\nferredoxin [FDX1, Cre14.g626700]),\nchloroplast-localized\nprotein\n(PSII)\nphotosystem\n3\n[Cre08.g362900],\nCre12.g509050],\n[PSBP3,\n\n(PsbP-like\nPSBP4\n\nsubunit K\n\nII\n\n\fThe eukaryotic CO2-concentrating organelle\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\n|\n\n3247\n\nand\n\nCre16.g651050),\n\nCre06.g261000]),\n\ncytochrome\nATP\n\nPhotosystem\nII oxygen evolution enhancer protein 3\n[PSBQ, Cre08.g372450], and Photosystem II subunit R\nb6f\n[PSBR,\n(CYC6,\nsynthase\n(ATPC, Cre06.g259900) (Mackinder et al. 2017). However,\ndespite some components of the O2-evolving PSII being pre-\nsent in Chlamydomonas pyrenoid-traversing membranes,\nother PSII components (such as subunit 1 of the\nPSII oxygen-evolving enhancer protein 1 [OEE1 or PSBO1,\nCre09.g396213] and PSII intrinsic core polypeptides D2\n[psbD, CreCp.g802329] and P5 [PSBP5, Cre09.g389578]) ap-\npear to be absent based on immunogold labeling (de Vitry\net al. 1989; McKay and Gibbs 1991). Relatedly, PSII was found\nto be inactive in the pyrenoid of the red alga Porphyridium\ncruentum based on cytochemical assays in which PSII activity\nwas detected through the production of osmiophilic difor-\nmazan upon the photoreduction of tetrazolium salts\n(McKay and Gibbs 1990, 1991). Minimizing PSII activity in\npyrenoid-traversing thylakoids may be a strategy for minim-\nizing O2 levels within the pyrenoid, which may help maintain\na high CO2 to O2 ratio around Rubisco (McKay and Gibbs\n1991). The electron transport chain components found in\nthe pyrenoid may therefore be in assembly intermediates,\nin inactive complexes undergoing repair, or may have differ-\nent functions from those found in stromal thylakoids. This\nhypothesis is supported by the pyrenoid localization of\nPSBP4, which is a homolog of the Arabidopsis PSII repair pro-\ntein PSBP-LIKE PROTEIN1 (PPL1, encoded by At3g55330)\nand\nfactors\n(Mackinder et al. 2017).\n\nfour known PSI assembly\n\ninteracts with\n\nluminal\n\nfactor 14\n\nchlorophyll\n\nfluorescence\n\nEight other proteins have recently been localized to the\npyrenoid tubules\nin Chlamydomonas: Deg protease 8\n(DEG8, Cre01.g028350), the cyclophilins CYN7 (Cre12.g544150)\nand CYN20-6 (Cre12.g544114), Photosystem II subunit P1\n(PSBP1, Cre12.g550850), the PSII stability/assembly factor\nhigh\n136\n(HCF136, Cre06.g273700), the thylakoid luminal protein\nthylakoid\n(TEF14, Cre06.g256250),\nuncharacterized proteins encoded by Cre03.g198850 (Wang\net al. 2022), and Cre03.g172700 (Lau et al. 2023) (Table 1).\nDEG8 is a predicted DegP-type protease, while CYN7 and\nCYN20-6 are two predicted peptidyl-prolyl cis-trans iso-\nmerases. The pyrenoid tubules may thus be involved in pro-\ntein folding, degradation, and/or import of new proteins into\nthe pyrenoid (Wang et al. 2022). The protein encoded\nby Cre03.g172700 is predicted to contain a long central\nalpha-helix and four “Rubisco-binding motifs” (discussed in\na later section), which might allow it to act as a potential pyr-\nenoid tether between the pyrenoid matrix and tubules along-\nside RBMP1 and RBMP2 (Lau et al. 2023).\n\nA polysaccharide sheath likely serves as a CO2\ndiffusion barrier\nPolysaccharide deposits are associated with the pyrenoids of\nsome species in every major algal lineage except the diatoms\n\nand coccolithophores (Fig. 3). In red algae and green algae,\nthe polysaccharide that makes up these deposits is starch,\nwhereas different polymers are used in other lineages, such\nas paramylon in the case of euglenoid algae (Nudelman\net al. 2006; Suzuki and Suzuki 2013; Ball et al. 2015). Green\nalgae produce starch in the chloroplast, whereas all other\nlineages produce their polysaccharide deposits in the cytosol\n(with the exception of the cryptophytes, which produce\nstarch in the periplastid) (Suzuki and Suzuki 2013; Ball\net al. 2015). Presumably because of these differences, species\nfrom all lineages except the green algae only have polysac-\ncharides associated with their pyrenoids if they have a stalked\nor bulging pyrenoid that projects into the cytoplasm (Meyer\net al. 2017) (Fig. 3). In these cases, the polysaccharide struc-\ntures are separated from the Rubisco matrix by the chloro-\nplast envelope (Meyer et al. 2017). The association of\npolysaccharide deposits with pyrenoids even when separated\nby membranes further implicates these structures in pyre-\nnoid function. In Chlamydomonas, the starch sheath is a\nshell-like structure made by curved starch granules around\nthe pyrenoid matrix (Fig. 1). Small gaps between starch plates\nallow the tubules to penetrate through into the matrix.\n\nAvailable evidence suggests that the pyrenoid polysacchar-\nide sheath, when present, serves as a barrier to slow the es-\nfrom the pyrenoid, allowing a higher\ncape of CO2\nconcentration of CO2 to be maintained in the pyrenoid\nand decreasing the energetic costs of CO2 concentration\n(Toyokawa et al. 2020; Fei et al. 2022). The most convincing\nevidence to date supporting this function comes from the\ndecreased CCM efficacy observed under very low CO2 in\nthe Chlamydomonas sta2-1 mutant (defective in starch syn-\nthase 2 [STA2, encoded by Cre17.g721500]), which has a\nthinner starch sheath but otherwise apparently normal local-\nization of key proteins (Toyokawa et al. 2020).\n\nThe pyrenoid starch sheath granules in Chlamydomonas\nare different from the stromal starch granules in their shape,\ncomposition, and the conditions under which they accumu-\nlate. The granules that make up the starch sheath are more\ncurved than stromal granules, which are globular in shape.\nThe molecular composition of starch consists of alternating\namorphous layers of amylose and crystalline layers of amylo-\npectin (Zeeman et al. 2010). Compared with stromal starch,\npyrenoidal starch has less amylose but more amylopectin\ncontent (Libessart et al. 1995; Findinier et al. 2019). Both\namylose and amylopectin have been shown to decrease O2\ngas permeability in vitro (Forssell et al. 2002), which further\nsupports the possible function of the starch sheath in slowing\ndown the escape of leaking CO2 from the matrix. Relatedly,\nthe molecular structure of starch varies depending on the al-\ngal lineage (Suzuki and Suzuki 2013; Ball et al. 2015), which\ncould have implications for the ability of starch to prevent\nCO2 diffusion in different species. In addition to differences\nin their shape and composition, pyrenoid starch sheath gran-\nules and stromal starch granules also accumulate under dif-\nferent conditions. When Chlamydomonas cells are grown in\nunfavorable conditions such as during nitrogen starvation,\n\n\f3248\n\n|\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\nHe et al.\n\nstromal starch content increases while that of pyrenoid\nstarch decreases, as starch metabolism rapidly switches\nfrom pyrenoidal to storage biosynthesis (Kuchitsu et al.\n1988; Findinier et al. 2019). However, when cells are moved\nfrom high CO2 (4%) to low CO2 (air-level), pyrenoid starch\naccumulates rapidly within hours and stromal starch is de-\ngraded (Kuchitsu et al. 1988).\n\nSeveral proteins have been implicated in the formation\nand degradation of the starch sheath in Chlamydomonas.\nThe protein StArch Granules Abnormal 1 (SAGA1, encoded\nby Cre11.g467712), which contains a putative starch-binding\ndomain, localizes to distinct puncta at the pyrenoid matrix/\nstarch interface (Figs. 1, C and D and 4B) (Itakura et al. 2019;\nMeyer et al. 2020b). Abnormally elongated and thinner\nstarch granules were observed in saga1 mutant cells, indicat-\ning that SAGA1 is required for normal starch sheath forma-\ntion (Itakura et al. 2019). A recent study suggested that\nSAGA1 is also necessary for relocalizing CAS and LCIB to\nthe pyrenoid under\nlimiting CO2 conditions and for\nCAS-dependent retrograde signaling regulation of nuclear\n− transporters (Shimamura\ngenes encoding CO2 and HCO3\net al. 2023). Another protein, bimodal starch granule 1\n(BSG1, Cre02.g091750), may be involved in the degradation\nof the starch sheath during the transition from low CO2 to\nhigh CO2 (Findinier et al. 2019).\n\nHigh-throughput studies have identified other proteins\nthat could potentially be involved in the formation of the\npyrenoid starch sheath (Table 1) (Mackinder et al. 2017;\nMeyer et al. 2020b). SAGA2 (Cre09.g394621) is a protein\nthat shares 30% sequence identity with SAGA1 and also\nhas a predicted starch-binding domain (Meyer et al.\n2020b). Like SAGA1, SAGA2 also localizes to the pyrenoid\nmatrix/starch interface (Fig. 4B), although its function is cur-\nrently unknown. Two other proteins, granule-bound STA2\n(Delrue et al. 1992; Maddelein et al. 1994) and starch-\nbranching enzyme 3 (SBE3, Cre10.g444700), localize around\nthe pyrenoid periphery,\nforming a plate-like pattern\n(Mackinder et al. 2017), which suggests that they may con-\ntribute to the biosynthesis of the starch sheath (Table 1).\nLCI9 (Cre09.g394473), which contains two starch-binding do-\nmains and\nfunction as a glucan\nto\n1,4-α-glucosidase, localizes in a mesh structure between the\ngaps of the starch sheath, suggesting that it may degrade\nstarch at the gaps between starch plates to ensure a close\nfit between adjacent plates (Mackinder et al. 2017). Further\nstudies of these starch-associated proteins are needed to\nin the biogenesis of the\nunderstand their\nChlamydomonas starch sheath.\n\nis predicted\n\nfunctions\n\nSix new pyrenoid-periphery proteins were recently identi-\nfied in Chlamydomonas: MIND1 (Cre12.g522950), malate de-\nhydrogenase 1 (MDH1, Cre03.g194850), uncharacterized\nproteins encoded by Cre09.g394547, Cre09.g415600 (Wang\net al. 2022), structural maintenance of chromosomes 7\n(SMC7, Cre17.g720450), and uncharacterized protein en-\ncoded by Cre09.g394510 (Lau et al. 2023) (Table 1). The loca-\ntion of most of these newly identified proteins relative to the\n\nstarch sheath remains unclear. Interestingly, MIND1 is a\nhomolog of the Arabidopsis chloroplast division site regula-\ntor MinD1 (At5g24020), suggesting that MIND1 could po-\ntentially play a role in coordinating pyrenoid fission or\ndissolution with chloroplast division in Chlamydomonas\n(Colletti et al. 2000; Freeman Rosenzweig et al. 2017; Wang\net al. 2022). SMC7 shows a punctate localization similar to\nthat of SAGA1 and is annotated as a member of the SMC\nfamily (Lau et al. 2023). SAGA1 and SAGA2 are also anno-\ntated as members of this family, which suggests that SMC7\nmight\nSAGA2.\nThe protein encoded by Cre09.g394510 contains a starch-\nbinding domain and localizes to the starch-matrix interface\nand the gaps between starch plates. It contains a predicted\nt-SNARE domain, which mediates vesicle fusion, suggesting\nthat it may be involved in membrane remodeling of the pyr-\nenoid tubules (Lau et al. 2023).\n\nfunction\n\nsimilarly\n\nSAGA1\n\nand\n\nto\n\nA Rubisco-binding motif mediates pyrenoid assembly\nAs previously discussed, the repeat protein EPYC1 has five\nRubisco-binding regions critical for pyrenoid matrix assembly\nin Chlamydomonas (He et al. 2020). Intriguingly, similar se-\nquences to the EPYC1 Rubisco-binding region have been iden-\ntified on many other pyrenoid-localized proteins (Fig. 4, A and\nB) (Meyer et al. 2020b). These sequences, including the\nRubisco-binding region on EPYC1, have been named\n“Rubisco-binding motifs.” The motifs on other pyrenoid pro-\nteins show similar binding affinity to Rubisco as the motifs\non EPYC1 (whose KD is approximately 3 mM) (He et al. 2020;\nMeyer et al. 2020b). Due to sequence similarity, the motifs\non other proteins are believed to bind to the same alpha-\nhelices of Rubisco small subunits as EPYC1 (Fig. 4, C–E).\n\nThe Rubisco-binding motif was shown to be necessary and\nsufficient for targeting a protein to the pyrenoid matrix\n(Meyer et al. 2020b). These observations suggest that nascent\npyrenoid proteins with copies of the motif diffuse around the\nchloroplast stroma until they encounter the matrix, where\nthey are captured by binding to Rubisco (Fig. 4B). One\nopen question is how proteins are targeted to the matrix\nwhen a full starch sheath has been assembled because stro-\nmal proteins would then not have direct access to Rubisco.\nIn addition to targeting proteins to the matrix, the\nRubisco-binding motif has been proposed to anchor the\nRubisco matrix to the pyrenoid tubules and connect\nthe starch sheath to the matrix (Meyer et al. 2020b). The\nRubisco-binding motif-containing proteins RBMP1 and\nRBMP2 localize to the tubules, suggesting that they target\na layer of Rubisco to the tubules. From there, EPYC1 may\nbe able to connect additional Rubiscos, causing the matrix\nto condense around the entire tubule network. The proteins\nSAGA1 and SAGA2 also contain Rubisco-binding motifs in\naddition to their starch-binding domains, which suggests\nthat they might link the matrix to the starch sheath\n(Fig. 4B) (Meyer et al. 2020b).\n\nThe identification of the Rubisco-binding motif and the\nhypothesis that proteins with this motif link the three\n\n\fThe eukaryotic CO2-concentrating organelle\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\n|\n\n3249\n\npyrenoid sub-compartments together in Chlamydomonas\nexplains the initially puzzling difference between the pheno-\ntypes of a mutant lacking functional EPYC1 and mutants\nwith disrupted EPYC1-binding sites on Rubisco small subu-\nnits (He et al. 2020; Meyer et al. 2020b). In a mutant lacking\nEPYC1, a minimal pyrenoid can still be observed (Mackinder\net al. 2016); however, mutants with disrupted EPYC1-binding\nsites on Rubisco small subunits lack a pyrenoid altogether\n(He et al. 2020; Meyer et al. 2012, 2020b). These findings\ncan be reconciled when considering that, in the mutant lack-\ning EPYC1, proteins other than EPYC1 that have the\nRubisco-binding motif (potentially including RBMP1 and\nRBMP2) can still bind to Rubisco and form the observed\nmuch smaller pyrenoid-like structure that still contains tu-\nbules and a starch sheath but lacks a canonical matrix.\n\nThe sequences of the Rubisco-binding motifs and their\nbinding sites on Rubisco are conserved\nin the order\nVolvocales to which Chlamydomonas belongs, but the motif\nhas not been found in any other algal lineages (Meyer et al.\n2020b). Assuming that pyrenoids convergently evolved, it is\npossible that the assembly of pyrenoids via common\nRubisco-binding motifs may broadly apply to pyrenoids\nacross the algal tree of life, although the specific sequences\nmay differ in different algal lineages (Meyer et al. 2020b).\n\nOther candidate pyrenoid components in\nChlamydomonas\nPhysical interactors of proteins that localize to the pyrenoid\nwere identified using large-scale affinity-purification mass\nspectrometry (Mackinder et al. 2017). Using this method,\n513 interactions involving 398 proteins were identified\n(Mackinder et al. 2017).\n\nIn a parallel study, Chlamydomonas pyrenoids were puri-\nfied and their proteome was analyzed, identifying 190 pro-\nteins in total (Zhan et al. 2018). Of the 190 candidate\npyrenoid proteins identified, many have confirmed or pre-\ndicted functions that are known or proposed to occur in pyr-\nenoids, such as the CCM, starch metabolism, or RNA\nmetabolism and translation. Additional proteins suggestive\nof new pyrenoid functions in tetrapyrrole and chlorophyll\nsynthesis, carotenoid metabolism, or amino acid metabolism\nwere also identified. Future work on these uncharacterized\ncandidate pyrenoid proteins will yield a better understanding\nof the biogenesis, function, and regulation of the pyrenoid.\n\nPyrenoid dynamics are likely highly regulated, but the under-\nlying regulatory mechanisms remain to be discovered.\n\nThe pyrenoid forms in response to limiting CO2 levels\nunder constant light\nWhen Chlamydomonas cells are transferred from high CO2\nto limiting CO2 under constant light conditions, the pyrenoid\nmatrix grows within one hour (Kuchitsu et al. 1991;\nRamazanov et al. 1994), presumably by relocalization of\nRubisco from the chloroplast stroma to the pyrenoid matrix.\nThe starch sheath starts to form within one hour after trans-\nfer from high to low CO2 as well and is fully formed after\nabout five hours (Kuchitsu et al. 1988; Ramazanov et al.\n1994).\n\nThe expression of many CCM-related genes, including\nthose encoding confirmed pyrenoid proteins such as\nEPYC1, STA2, and CAH3, is upregulated during the transition\nfrom high to low CO2 (Brueggeman et al. 2012; Fang et al.\n2012), which is consistent with the expansion of the pyrenoid\nmatrix and the formation of the starch sheath. The\nChlamydomonas CCM “master regulator” inorganic carbon\n(Ci) acquisition 5 (CIA5, Cre02.g096300, also known as\nCCM1) is required for the transcriptional upregulation of\nCAH3, EPYC1, STA2, LCIB, LCIC, and SMM7 in response to\nthe transition from high CO2 to low CO2 (Fang et al. 2012;\nSanthanagopalan et al. 2021), although CIA5 may also have\nother non-CCM-related functions (Moroney et al. 1989;\nMarek and Spalding 1991; Miura et al. 2004; Wang et al.\n2005; Fang et al. 2012; Redekop et al. 2022). CIA5 has zinc-\nbinding activity and was proposed to be a transcription fac-\ntor (Fukuzawa et al. 2001; Xiang et al. 2001; Kohinata et al.\n2008), but this has not been confirmed because no\nDNA-CIA5 complexes have been identified. Additionally,\nthe CIA5 regulatory mechanism is not well understood.\nCIA5 transcript and CIA5 protein levels are similar in high–\nCO2-grown and low–CO2-grown unsynchronized wild-type\ncells (Wang et al. 2005; Fang et al. 2012); thus, CIA5 activity\nregulated by posttranslational modifications\nmay be\n(Fukuzawa et al. 2001; Xiang et al. 2001; Wang et al. 2005;\nBrueggeman et al. 2012; Chen 2016). CIA5 regulates the tran-\nscription\n1\n(LCR1, Cre09.g399552), which is known to directly regulate\nthe\nLCI1\nCAH1\n(Cre03.g162800), and LCI6 (Cre12.g553350) (Yoshioka et al.\n2004).\n\n(Cre04.g223100),\n\nexpression\n\nlow-CO2\n\nresponse\n\nfactor\n\nstress\n\nof\n\nDynamics and regulation\nPyrenoids in various species are dynamic, showing noticeable\nmorphological changes under different growth conditions\nand during cell division (Brown and Bold 1964; Brown et al.\n1967; Goodenough 1970; Retallack and Butler 1970). The\nnewly reported liquid-like nature of the Chlamydomonas\npyrenoid (Freeman Rosenzweig et al. 2017) provides a new\nframework for thinking about the biophysics underlying pyr-\nenoid dissolution, condensation, and division by fission.\n\nPosttranslational modifications are thought to regulate the\nfunctions of several essential pyrenoid proteins. The Rubisco\nlinker EPYC1/LCI5 was reported to be phosphorylated under\nlow CO2 conditions but not under high CO2 (Turkina et al.\n2006), although the functional implications of this phosphor-\nylation are unknown. The relocalization of CAH3 from the\nstromal thylakoids to the pyrenoid tubules under low CO2\nconditions as well as the functions of HLA3 and LCIC under\nlow CO2 and very low CO2 have also been proposed to be\nregulated by phosphorylation (Blanco-Rivero et al. 2012;\n\n\f3250\n\n|\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\nHe et al.\n\nWang et al. 2014). LCIB is glutathionylated during acclima-\ntion to limiting CO2 (Zaffagnini et al. 2012). The effect of\nthese posttranslational modifications on the functions of\nthese CCM proteins is unclear, as is the identity of the regu-\nlatory proteins upstream of these modifications. Methylation\nmay also be involved in the regulation of pyrenoid biogenesis,\nas suggested by altered pyrenoid morphologies under low\nCO2 in mutants lacking function for the putative methyl-\ntransferase CIA6 (Cre10.g437829) (Ma et al. 2011).\n\nWhen cells are transferred from limiting CO2 to high CO2,\nthe disassembly of the pyrenoid is much slower than its assem-\nbly when cells are transferred from high CO2 to limiting CO2\n(Kuchitsu et al. 1988; Ramazanov et al. 1994). The degradation\nof the starch sheath and the dissolution of the matrix can take\ntwo to three days (Ramazanov et al. 1994). This slow degrad-\nation is consistent with the slow deactivation of the CCM,\nwhich also requires about three days when cells are moved\nfrom limiting CO2 to high CO2 (Ramazanov et al. 1994). It\nhas been suggested that CCM proteins are not rapidly de-\ngraded after the transition from low CO2 to high CO2\n(Toguri et al. 1989), whereas the synthesis of new CCM pro-\nteins stops shortly after this transition (Manuel and\nMoroney 1988). These observations have not been confirmed\nfor crucial pyrenoid proteins, and the regulatory mechanisms\nremain unclear.\n\nThe pyrenoid-based CCM is induced and\ndeactivated during the diurnal cycle\nMost studies of the Chlamydomonas pyrenoid and CCM\nhave been performed with asynchronous cultures of cells\ngrown under constant light, where the induction of the\nCCM and formation of the pyrenoid are solely determined\nby the level of CO2. However, synchronous growth under di-\nurnal cycles has revealed that the CCM is regulated during\nthe course of the day/night cycle (Mitchell et al. 2014;\nTirumani et al. 2014; Zones et al. 2015; Strenkert et al. 2019).\nChlamydomonas cells can be synchronized when grown\nunder 12-h-light/12-h-dark cycles in minimal medium, with\ncells going through one cell cycle each day (Harris 2009;\nMitchell et al. 2014; Zones et al. 2015; Strenkert et al.\n2019). The CCM is downregulated at night and fully induced\none hour before dawn (Mitchell et al. 2014). During this in-\nduction, both Rubisco and CAH3 were found to relocalize\nfrom the chloroplast to the pyrenoid based on statistical ana-\nlyses of immunogold labeling (Mitchell et al. 2014), although\nit should be noted that the original electron micrographs\nwere not provided in this study. Transcriptomics studies\nhave shown that, in cells grown under diurnal cycles, genes\nencoding the master regulator CIA5 and crucial pyrenoid\nproteins reach their highest transcript levels in the first few\nhours around the transition from dark to light (Strenkert\net al. 2019; Adler et al. 2022). However, whether CIA5 is\nalso involved in CCM activation and pyrenoid formation dur-\ning diurnal cycles is not known. The pyrenoid may grow and\nexpand during the day as cells grow (Zones et al. 2015;\n\nStrenkert et al. 2019), but this has not yet been specifically\nmeasured.\n\nPyrenoid formation can be induced by\nhyperoxia and H2O2\nHyperoxia was recently reported to induce pyrenoid forma-\n− levels (Neofotis et al. 2021).\ntion, even at high CO2 or HCO3\nThe authors reasoned that because pyrenoid formation is in-\nduced by both low CO2 and hyperoxia, a metabolite that ac-\ncumulates under both conditions may serve as a signal that\ninduces pyrenoid formation. Consistent with this idea, the\nauthors observed that hydrogen peroxide (H2O2), which is\nexpected to accumulate under both conditions, induces pyr-\nenoid formation. H2O2 is a byproduct of photorespiration,\nwhich recycles 2-phosphoglycolate, the toxic product of\nthe oxygenase activity of Rubisco (Moroney et al. 2013),\nwhich is increased under low CO2 and hyperoxia. Whether\nor\nH2O2\nindirectly (e.g. via other metabolites) has not been\ndetermined.\n\nformation\n\nregulates\n\npyrenoid\n\ndirectly\n\nstudies observed\n\nelectron microscopy\n\nThe Chlamydomonas pyrenoid matrix divides by\nfission and dissolves into the surrounding chloroplast\nduring cell division\nEarly\nthe\nChlamydomonas pyrenoid dividing by fission (Goodenough\n1970). Recent live-cell microscopy observation of pyrenoid\nmatrix dynamics using fluorescently tagged Rubisco or\nEPYC1 revealed that the matrix is inherited by fission in\nmost chloroplasts and assembled de novo in others (Fig. 5)\n(Freeman Rosenzweig et al. 2017). Approximately two-thirds\nof daughter chloroplasts inherited their matrix through\nelongation and then fission of the pyrenoid matrix from\nthe mother chloroplast (Fig. 5, A to C), whereas one of the\ndaughter chloroplasts inherited the entire matrix punctum\nin the remaining cases (Fig. 5D). When the pyrenoid divided\nby fission, matrix elongation and fission occurred toward the\nend of overall chloroplast division and took approximately\nseven minutes. A “bridge” of matrix connecting the two lobes\nwas briefly visible towards the end of fission (Fig. 5B). After\nthe bridge ruptured, the daughter pyrenoids quickly reverted\nto spherical shapes, similar to the behavior of liquid droplets\n(Fig. 5B) (Stone 1994; Yanashima et al. 2012; Freeman\nRosenzweig et al. 2017). The mechanism mediating pyrenoid\nfission during cell division is unknown.\n\nA portion of the pyrenoid matrix rapidly disperses into the\nstroma approximately 20 minutes before pyrenoid division\n(Fig. 5C, at approximately 19 minutes, the light blue through-\nout the chloroplast indicates the dispersed pyrenoid matrix).\nDuring this dispersal, small puncta of matrix often transiently\nappear throughout the stroma (Fig. 5C). The dispersal of ma-\ntrix materials may facilitate equal distribution of the pyre-\nnoid matrix to daughter chloroplasts and may also help\ndecrease the surface tension or viscosity of the matrix droplet\nto facilitate fission (Freeman Rosenzweig et al. 2017). Indeed,\n\n\fThe eukaryotic CO2-concentrating organelle\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\n|\n\n3251\n\nFigure 5. The Chlamydomonas pyrenoid exhibits liquid-like behavior during cell divisions. A) Diagram depicting the timeline and morphology of a\ntypical cell division with pyrenoid fission in Chlamydomonas (adapted from Freeman Rosenzweig et al. 2017). The time point t = 0 is the moment\nthe chloroplast division furrow passes between the daughter pyrenoids. A portion of the pyrenoid matrix disperses into the chloroplast stroma\nduring the division of the pyrenoid. The approximate timing and duration of key events are shown below the timeline. B) Diagram depicting\nthe “bridge” of matrix during pyrenoid fission. C) Diagram depicting the transient appearance of small puncta of pyrenoid matrix throughout\nthe stroma during dispersal of the matrix in some dividing cells. D) Diagram depicting the de novo formation of a daughter pyrenoid when pyrenoid\nfission fails. The lower daughter cell inherits the entire pyrenoid of the mother cell. The upper cell shows de novo pyrenoid formation with the\nappearance of one or more fluorescent puncta growing or coalescing into one pyrenoid (observed in wild-type cells expressing either\nEPYC1-Venus or Rubisco-Venus).\n\nABCD\f3252\n\n|\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\nHe et al.\n\nmany of the daughter chloroplasts that did not inherit a pyr-\nenoid through fission inherited dissolved matrix building\nblocks, from which they appeared to form a pyrenoid de\nnovo (Freeman Rosenzweig et al. 2017) (Fig. 5D). In such\ncases, multiple small Rubisco or EPYC1 fluorescent puncta\nappeared, and smaller puncta shrank whereas larger ones\ngrew until the cell contained a single pyrenoid (Freeman\nRosenzweig et al. 2017) (Fig. 5D). This behavior resembles\nOstwald ripening, a physical mechanism by which larger dro-\nplets in a phase-separated system grow by acquiring building\nblocks from smaller droplets (Hyman et al. 2014; Freeman\nRosenzweig et al. 2017; Rosowski et al. 2020). The mechan-\nisms regulating the formation of multiple puncta of matrix\nmaterial and the dispersal of the matrix remain unknown.\n\nPyrenoids in other algae are likely to leverage liquid-like\nproperties during cell division in ways similar to those ob-\nserved in Chlamydomonas (Freeman Rosenzweig et al.\n2017; Barrett et al. 2021). Indeed, both division by fission\nand the rapid disappearance and reappearance of the pyre-\nnoid matrix (which would be consistent with matrix dissol-\nution and condensation) have been observed during cell\ndivision using TEM on fixed cells in some species of the green\nalgae Tetracystis, Chlorococcum, and Bulbochaete and the dia-\ntom Donkinia (Brown et al. 1967; Retallack and Butler 1970;\nCox 1981). Additional discussions on liquid-like pyrenoid be-\nhavior during cell division can be found in the recent review\nby Barrett et al. (Barrett et al. 2021). Further studies on other\nspecies will be necessary to test the generality of this\nprinciple.\n\nThe number of pyrenoids per cell appears to be\nregulated in Chlamydomonas\nWild-type Chlamydomonas cells have only one pyrenoid.\nHowever, in mutant cells lacking SAGA1, EPYC1, or CIA6,\nmultiple pyrenoids are often observed, suggesting that wild-\ntype cells may actively work to ensure that they have a single\npyrenoid, and aspects of this regulation may be disrupted in\nthese mutants.\n\nThe most striking example of multiple pyrenoids can be\nseen in the saga1 mutant, with an average of approximately\nten matrix droplets per cell (Itakura et al. 2019). The multiple\npyrenoids in saga1 are stable without obvious changes in size\nor position in living mutant cells over the course of one\nhour. Although the molecular function of SAGA1 remains\nunclear, these observations suggest that this protein is in-\nvolved in maintaining a single pyrenoid.\n\nThe two other known cases of mutants with multiple pyr-\nenoids are epyc1 and cia6. Multiple pyrenoids were observed\nby TEM in 13% of epyc1 cells compared with 3% in wild-type\ncells (Mackinder et al. 2016). A mutant of the putative\nmethyltransferase CIA6 also exhibits multiple small pyre-\nnoids, as observed by TEM or via light microscopy (Ma\net al. 2011). TEM images showed that the morphology of\nthe pyrenoids in epyc1 and cia6 are similar, with a small ma-\ntrix and thicker starch sheath than in wild-type cells (Ma\n\net al. 2011; Mackinder et al. 2016). EPYC1 has a well-\ncharacterized function in pyrenoid matrix biogenesis; thus,\nit is possible that defects in matrix biogenesis lead to multiple\npyrenoids. The mechanism by which the number of pyre-\nnoids in a cell is regulated remains to be identified.\n\nPerspective\nThe pyrenoid provides a unique opportunity to expand our\nfundamental knowledge of liquid-liquid phase separation,\ngenetically engineer photosynthetic organisms for higher\ncrop yields, and obtain insights into critical photosynthetic\ncarbon assimilation in the oceans and fresh water, all through\nthe study of one organelle.\n\nStudies of the Chlamydomonas pyrenoid have laid the\nfoundation for exploring the molecular composition of pyr-\nenoids across the photosynthetic tree of life. Of particular\ninterest is investigating whether pyrenoids in other species\nalso exhibit\nliquid-like behavior. Additionally, exploring\nwhether Rubisco-binding motifs are a common principle\nacross all algal pyrenoids could lead to a better understand-\ning of how pyrenoids first evolved. Molecular studies of CCM\nfunction in different algae could also help reveal the minimal\ncomponents that will be necessary to create a functional pyr-\nenoid in plants. More broadly, studying the mechanisms by\nwhich Rubisco condensates associate with membrane struc-\ntures and peripheral polysaccharides could provide insights\ninto general principles by which liquid-like organelles interact\nwith other cellular structures.\n\nFuture studies in Chlamydomonas as well as in other spe-\ncies of algae will bring us closer to generating the first func-\ntional pyrenoid in vascular land plants, which will help us\nmeet increasing agricultural demands as the global popula-\ntion rises. Progress is already being made toward engineering\na Chlamydomonas-based pyrenoid into land plants with the\nsuccessful reconstitution of Rubisco-EPYC1 condensates in\nArabidopsis (Atkinson et al. 2020). The generation of thyla-\nkoid tubules that traverse these condensates, delivery of\nCO2 via these tubules, and the assembly of starch around\nthese condensates will be important next steps.\n\nIn addition, by studying the regulation of pyrenoid forma-\ntion and dynamics in Chlamydomonas as well as in dominant\nmarine species such as diatoms, we can advance our under-\nstanding of the aquatic photosynthesis that is crucial for our\necosystem and the global carbon cycle.\n\n(Princeton\n\nAcknowledgments\nWe thank Alistair McCormick (University of Edinburgh),\nMoritz Meyer\nalumni), Ned Wingreen\n(Princeton University), Debashish Bhattacharya (Rutgers\nUniversity), Ben Engel (University of Basel), Lianyong Wang\n(Princeton University), Micah\n(Princeton\nUniversity), and the two anonymous reviewers for helpful\ncomments, suggestions, and clarifications. We thank Marie\nBao, as part of Life Science Editors, for help with manuscript\n\nBurton\n\n\fThe eukaryotic CO2-concentrating organelle\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\n|\n\n3253\n\nediting. This work was supported by Howard Hughes Medical\nInstitute, National Science Foundation (MCB-1935444), and\nNational Institutes of Health (R01GM140032) grants to\nM.C.J. M.C.J.\nInstitute\nInvestigator. We apologize to all colleagues whose work we\ncould not incorporate due to space constraints.\n\nis a Howard Hughes Medical\n\nAuthor contributions\nAll authors contributed to the writing and the figures of the\narticle.\n\nConflict of interest statement. None declared.\n\nBehrenfeld MJ, Randerson JT, McClain CR, Feldman GC, Los SO,\nTucker CJ, Falkowski PG, Field CB, Frouin R, Esaias WE, et al.\nBiospheric primary production during an ENSO transition. Science.\n2001:291(5513):2594–2597. https://doi.org/10.1126/science.1055071\nBerner RA. The rise of plants and their effect on weathering and atmos-\npheric CO2. Science. 1997:276(5312):544–546. https://doi.org/10.\n1126/science.276.5312.544\n\nBerner RA. GEOCARBSULF: a combined model for Phanerozoic atmos-\npheric O2 and CO2. Geochim Cosmochim Acta. 2006:70(23):\n5653–5664. https://doi.org/10.1016/j.gca.2005.11.032\n\nBerner RA, Kothavala Z. Geocarb III: a revised model of atmospheric\nCO2 over Phanerozoic time. Am J Sci. 2001:301(2):182–204. https://\ndoi.org/10.2475/ajs.301.2.182\n\nBertagnolli BL, Nadakavukaren MJ. An ultrastructural study of pyre-\nnoids from Chlorella pyrenoidosa. J Cell Sci. 1970:7(3):623–630.\nhttps://doi.org/10.1242/jcs.7.3.623\n\nReferences\nAdler L, Diaz-Ramos A, Mao Y, Pukacz KR, Fei C, McCormick AJ.\nNew horizons for building pyrenoid-based CO2-concentrating me-\nchanisms in plants to improve yields. Plant Physiol. 2022:190(3):\n1609–1627. https://doi.org/10.1093/plphys/kiac373\n\nAnbar AD, Duan Y, Lyons TW, Arnoled GL, Kendall B. A whiff of oxy-\ngen before the great oxidation event? Science. 2007:317(5846):\n1903–1906. https://doi.org/10.1126/science.1140325\n\nAng WSL, How JA, How JB, Mueller-Cajar O. The stickers and spacers\nof Rubisco condensation: assembling the heartpiece of biophysical\nCO2 concentrating mechanisms. J Exp Bot. 2022:74(2):612–626.\nhttps://doi.org/10.1093/jxb/erac321\n\nAtkinson N, Mao Y, Chan KX, McCormick AJ. Condensation of Rubisco\ninto a proto-pyrenoid in higher plant chloroplasts. Nat Commun.\n2020:11(1):6303. https://doi.org/10.1038/s41467-020-20132-0\n\nAtkinson N, Velanis CN, Wunder T, Clarke DJ, Mueller-Cajar O,\nMcCormick AJ. The pyrenoidal linker protein EPYC1 phase separates\nwith hybrid Arabidopsis-Chlamydomonas Rubisco through interac-\ntions with the algal Rubisco small subunit. J Exp Bot. 2019:70(19):\n5271–5285. https://doi.org/10.1093/jxb/erz275\n\nBadger MR, Andrews TJ, Whitney SM, Ludwig M, Yellowlees DC,\nLeggat W, Price GD. The diversity and coevolution of Rubisco, plas-\ntids, pyrenoids, and chloroplast-based CO2-concentrating mechan-\nisms in algae. Can J Bot. 1998:76(6):1052–1071. https://doi.org/10.\n1139/b98-074\n\nBadger MR, Kaplan A, Berry JA. Internal inorganic carbon pool of\nChlamydomonas reinhardtii: evidence\nfor a carbon dioxide-\nconcentrating mechanism. Plant Physiol. 1980:66(3):407–413.\nhttps://doi.org/10.1104/pp.66.3.407\n\nBadger MR, Price GD. CO2 Concentrating mechanisms in cyanobac-\nteria: molecular components, their diversity and evolution. J Exp\nBot. 2003:54(383):609–622. https://doi.org/10.1093/jxb/erg076\n\nBadger MR, Price GD, Long BM, Woodger FJ. The environmental plas-\nticity and ecological genomics of the cyanobacterial CO2 concentrat-\ning mechanism. J Exp Bot. 2006:57(2):249–265. https://doi.org/10.\n1093/jxb/eri286\n\nBall S, Colleoni C, Arias M. The transition from glycogen to starch me-\ntabolism in cyanobacteria and eukaryotes. In: Nakamura Y, editor.\nStarch metabolism and structure. Tokyo: Springer Japan KK; 2015.\np. 93–158.\n\nBarrett J, Girr P, Mackinder LCM. Pyrenoids: CO2-fixing phase sepa-\nrated\nliquid organelles. Biochim Biophys Acta Mol Cell Res.\n2021:1868(5):118949. https://doi.org/10.1016/j.bbamcr.2021.118949\nBedoshvili YD, Likhoshway YV. The cell ultrastructure of diatoms—\nimplications for phylogeny? In: Maaz K, editor. The transmission elec-\ntron microscope. InTech; 2012. p. 147–160.\n\nBedoshvili YD, Popkova TP, Likhoshway YV. Chloroplast structure of\ndiatoms of different classes. Cell Tissue Biol. 2009:3(3):297–310.\nhttps://doi.org/10.1134/S1990519X09030122\n\nBlanco-Rivero A, Shutova T, Roman MJ, Villarejo A, Martinez F.\nPhosphorylation controls the localization and activation of the lume-\nnal carbonic anhydrase in Chlamydomonas reinhardtii. PLoS One.\n2012:7(11):e49063. https://doi.org/10.1371/journal.pone.0049063\nBorkhsenious ON, Mason CB, Moroney JV. The intracellular localiza-\nin\ntion of\nChlamydomonas reinhardtii. Plant Physiol. 1998:116(4):1585–1591.\nhttps://doi.org/10.1104/pp.116.4.1585\n\nribulose-1,5-bisphosphate\n\ncarboxylase/oxygenase\n\nBorowitzka MA. Biology of microalgae. In: Levine IA, Fleurence J, edi-\ntors. Microalgae in health and disease prevention. London, England:\nAcademic Press; 2018. p. 23–72.\n\nBowles AMC, Williamson CJ, Williams TA, Lenton TM, Donoghue\nPCJ. The origin and early evolution of plants. Trends Plant Sci.\n2022:28(3):312–329. https://doi.org/10.1016/j.tplants.2022.09.009\nBracha D, Walls MT, Brangwynne CP. Probing and engineering liquid-\nphase organelles. Nat Biotechnol. 2019:37(12):1435–1445. https://\ndoi.org/10.1038/s41587-019-0341-6\n\nBracher A, Sharma A, Starling-Windhof A, Hartl FU, Hayer-Hartl M.\nDegradation of potent Rubisco inhibitor by selective sugar phosphat-\nase. Nat Plants. 2015:1:14002. https://doi.org/10.1038/nplants.2014.2\nBracher A, Whitney SM, Hartl FU, Hayer-Hartl M. Biogenesis and\nmetabolic maintenance of Rubisco. Annu Rev Plant Biol.\n2017:68(1):29–60. https://doi.org/10.1146/annurev-arplant-043015-\n111633\n\nBrown RM, Arnott HJ, Bisalputra T, Hoffman LR. The pyrenoid: its\nstructure, distribution, and function. J Phycol. 1967:3(Supp):5–6\nBrown RMJ, Bold H. Phycological studies: V. Comparitive studies of the\nalgal genera Tetracystis and Cholorococcum. University of Texas\nPublication; 1964.\n\nBrueggeman AJ, Gangadharaiah DS, Cserhati MF, Casero D, Weeks\nDP, Ladunga I. Activation of the carbon concentrating mechanism\nby CO2 deprivation coincides with massive transcriptional restruc-\nturing\nin Chlamydomonas reinhardtii. Plant Cell. 2012:24(5):\n1860–1875. https://doi.org/10.1105/tpc.111.093435\n\nBuma AGJ, van Oijen T, van de Poll W, Veldhuis MJW, Gieskes\nWWC. The sensitivity of Emiliania huxleyi (Prymnesiophyceae) to\nultraviolet-b radiation. J Phycol. 2000:36(2):296–303. https://doi.\norg/10.1046/j.1529-8817.2000.99155.x\n\nBurlacot A, Dao O, Auroy P, Cuine S, Li-Beisson Y, Peltier G.\nAlternative\nalgal\nCO2-concentrating mechanism. Nature. 2022:605(7909):366–371.\nhttps://doi.org/10.1038/s41586-022-04662-9\n\nphotosynthesis\n\npathways\n\ndrive\n\nthe\n\nCaemmerer SV, Furbank RT. The C4 pathway: an efficient CO2 pump.\nPhotosynth Res. 2003:77(2/3):191–207. https://doi.org/10.1023/A:\n1025830019591\n\nCalvin M. The path of carbon in photosynthesis: the carbon cycle is a\ntool for exploring chemical biodynamics and the mechanism of\nquantum conversion. Science. 1962:135(3507):879–889. https://doi.\norg/10.1126/science.135.3507.879\n\nCaspari OD, Meyer MT, Tolleter D, Wittkopp TM, Cunniffe NJ,\nin\n\nLawson T, Grossman AR, Griffiths H. Pyrenoid\n\nloss\n\n\f3254\n\n|\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\nHe et al.\n\nChlamydomonas reinhardtii causes limitations in CO2 supply, but\nJ Exp Bot. 2017:68(14):\nnot thylakoid operating efficiency.\n3903–3913. https://doi.org/10.1093/jxb/erx197\nfunction and\n\nin\nChlamydomonas reinhardtii [doctoral dissertation]. [Ames (IA)]:\nIowa State University; 2016.\n\nregulation of CIA5_CCM1\n\nChen B. The\n\nClay BL, Kugrens P. Characterization of Hemiselmis amylosa sp. nov.\nand phylogenetic placement of the blue-green cryptomonads H.\namylosa and Falcomonas daucoides. Protist. 1999:150(3):297–310.\nhttps://doi.org/10.1016/S1434-4610(99)70031-3\n\nColletti KS, Tattersall EA, Pyke KA, Froelich JE, Stokes KD,\nOsteryoung KW. A homologue of the bacterial cell division site-\ndetermining factor MinD mediates placement of the chloroplast div-\nision apparatus. Curr Biol. 2000:10(9):507–516. https://doi.org/10.\n1016/S0960-9822(00)00466-8\n\nCox EJ. Observations on the morphology and vegetative cell division of\nthe diatom donkinia recta. Helgoländer Meeresunters. 1981:34(4):\n497–506. https://doi.org/10.1007/BF01995921\n\nCraig RJ, Gallaher SD, Shu S, Salome PA, Jenkins JW, Blaby-Haas CE,\nPurvine SO, O’Donnell S, Barry K, Grimwood J, et al. The\nChlamydomonas Genome Project, version 6: reference assemblies\nfor mating-type plus and minus strains reveal extensive structural\nmutation in the laboratory. Plant Cell. 2023:35(2):644–672. https://\ndoi.org/10.1093/plcell/koac347\n\nDecelle J, Veronesi G, LeKieffre C, Gallet B, Chevalier F, Stryhanyuk\nH, Marro S, Ravanel S, Tucoulou R, Schiever N, et al. Subcellular\narchitecture and metabolic connection in the planktonic photosym-\nbiosis between Collodaria (radiolarians) and their microalgae.\nEnviron Microbiol. 2021:23(11):6569–6586. https://doi.org/10.1111/\n1462-2920.15766\n\nDelrue B, Fontaine T, Routier F, Decq A, Wieruszeski JM, Van Den\nKoornhuyse N, Maddelein ML, Fournet B, Ball S. Waxy\nChlamydomonas reinhardtii: monocellular algal mutants defective\nin amylose biosynthesis and granule-bound starch synthase activity\naccumulate a structurally modified amylopectin.\nJ Bacteriol.\n1992:174(11):3612–3620.\nhttps://doi.org/10.1128/jb.174.11.3612-\n3620.1992\n\nde Vitry C, Olive J, Drapier D, Recouvreur M, Wollman FA.\nPosttranslational events leading to the assembly of photosystem II\nprotein complex: a study using photosynthesis mutants from\nChlamydomonas reinhardtii. J Cell Biol. 1989:109(3):991–1006.\nhttps://doi.org/10.1083/jcb.109.3.991\n\nDorrell RG, Gile G, McCallum G, Méheust R, Bapteste EP, Klinger\nCM, Brillet-Guéguen L, Freeman KD, Richter DJ, Bowler C.\nChimeric origins of ochrophytes and haptophytes revealed through\nan ancient plastid proteome. Elife. 2017:6:e23717. https://doi.org/\n10.7554/eLife.23717\n\nDuanmu D, Miller AR, Horken KM, Weeks DP, Spalding MH.\nKnockdown of limiting-CO2-induced gene HLA3 decreases HCO3-\ntransport and photosynthetic ci affinity in Chlamydomonas rein-\nhardtii. Proc Natl Acad Sci U S A. 2009:106(14):5990–5995. https://\ndoi.org/10.1073/pnas.0812885106\n\nDuff RJ, Villarreal JC, Cargill DC, Renzaglia KS. Progress and chal-\nlenges toward developing a phylogeny and classification of the horn-\nworts. Bryologist. 2007:110(2):214–243. https://doi.org/10.1639/\n0007-2745(2007)110[214:PACTDA]2.0.CO;2\n\nEhleringer JR, Sage RF, Flanagan LB, Pearcy RW. Climate change and\nthe evolution of C4 photosynthesis. Trends Ecol Evol. 1991:6(3):\n95–99. https://doi.org/10.1016/0169-5347(91)90183-X\n\nEngel BD, Schaffer M, Kuhn Cuellar L, Villa E, Plitzko JM, Baumeister\nW. Native architecture of the Chlamydomonas chloroplast revealed\nby in situ cryo-electron tomography. Elife. 2015:4:e04889. https://\ndoi.org/10.7554/eLife.04889\n\nFalkowski PG, Katz ME, Knoll AH, Quigg A, Raven JA, Schofield O,\nTaylor FJ. The evolution of modern eukaryotic phytoplankton.\nScience. 2004:305(5682):354–360. https://doi.org/10.1126/science.\n1095964\n\nFang W, Si Y, Douglass S, Casero D, Merchant SS, Pellegrini M,\nLadunga I, Liu P, Spalding MH. Transcriptome-wide changes in\nChlamydomonas reinhardtii gene expression regulated by carbon di-\noxide and the CO2-concentrating mechanism regulator CIA5/CCM1.\nPlant Cell. 2012:24(5):1876–1893. https://doi.org/10.1105/tpc.112.\n097949\n\nFei C, Wilson AT, Mangan NM, Wingreen NS, Jonikas MC. Modelling\nthe pyrenoid-based CO2-concentrating mechanism provides insights\ninto its operating principles and a roadmap for its engineering into\ncrops. Nat Plants. 2022:8(5):583–595. https://doi.org/10.1038/\ns41477-022-01153-7\n\nField CB, Behrenfeld MJ, Randerson JT, Falkowski P. Primary produc-\ntion of the biosphere: integrating terrestrial and oceanic compo-\nnents. Science. 1998:281(5374):237–240. https://doi.org/10.1126/\nscience.281.5374.237\n\nFindinier J, Laurent S, Duchene T, Roussel X, Lancelon-Pin C, Cuine\nS, Putaux JL, Li-Beisson Y, D’Hulst C, Wattebled F, et al. Deletion of\nBSG1 in Chlamydomonas reinhardtii leads to abnormal starch gran-\nule size and morphology. Sci Rep. 2019:9(1):1990. https://doi.org/10.\n1038/s41598-019-39506-6\n\nFlamholz AI, Prywes N, Moran U, Davidi D, Bar-On YM, Oltrogge\nLM, Alves R, Savage D, Milo R. Revisiting trade-offs between\nRubisco kinetic parameters. Biochemistry. 2019:58(31):3365–3376.\nhttps://doi.org/10.1021/acs.biochem.9b00237\n\nFlombaum P, Gallegos JL, Gordillo RA, Rincon J, Zabala LL, Jiao N,\nKarl DM, Li WK, Lomas MW, Veneziano D, et al. Present and future\nglobal distributions of the marine Cyanobacteria Prochlorococcus\nand Synechococcus. Proc Natl Acad Sci U S A. 2013:110(24):\n9824–9829. https://doi.org/10.1073/pnas.1307701110\n\nFord TW. A comparative ultrastructural study of Cyanidium caldarium\nand the unicellular red alga Rhodosorus marinus. Ann Bot.\nhttps://doi.org/10.1093/oxfordjournals.aob.\n1984:53(2):285–294.\na086690\n\nForssell P, Lahtinen R, Lahelin M, Myllärinen P. Oxygen permeability\nof amylose and amylopectin films. Carbohydr Polym. 2002:47(2):\n125–129. https://doi.org/10.1016/S0144-8617(01)00175-8\n\nFreeman Rosenzweig ES, Xu B, Kuhn Cuellar L, Martinez-Sanchez A,\nSchaffer M, Strauss M, Cartwright HN, Ronceray P, Plitzko JM,\nForster F, et al. The eukaryotic CO2-concentrating organelle is\nliquid-like and exhibits dynamic reorganization. Cell. 2017:171(1):\n148–162.e19. https://doi.org/10.1016/j.cell.2017.08.008\n\nFresnel J, Probert I. The ultrastructure and life cycle of the coastal\ncoccolithophorid Ochrosphaera neapolitana (Prymnesiophyceae).\nEur\nhttps://doi.org/10.1080/\n09670260400024659\n\n2005:40(1):105–122.\n\nPhycol.\n\nJ\n\nFridlyand LE. Models of CO2 concentrating mechanisms in microalgae\ntaking into account cell and chloroplast structure. Biosystems.\n1997:44(1):41–57. https://doi.org/10.1016/S0303-2647(97)00042-7\nFujiwara S, Fukuzawa H, Tachiki A, Miyachi S. Structure and differen-\ntial expression of two genes encoding carbonic anhydrase in\nChlamydomonas\nreinhardtii. Proc Natl Acad Sci U S A.\n1990:87(24):9779–9783. https://doi.org/10.1073/pnas.87.24.9779\nFukuzawa H, Miura K, Ishizaki K, Kucho KI, Saito T, Kohinata T,\nOhyama K. Ccm1, a regulatory gene controlling the induction of a\ncarbon-concentrating mechanism in Chlamydomonas reinhardtii\nby sensing CO2 availability. Proc Natl Acad Sci U S A. 2001:98(9):\n5347–5352. https://doi.org/10.1073/pnas.081593498\n\nFunke RP, Kovar JL, Weeks DP. Intracellular carbonic anhydrase is es-\nsential to photosynthesis in Chlamydomonas reinhardtii at atmos-\npheric levels of CO2. Demonstration via genomic complementation\nof the high-CO2-requiring mutant ca-1. Plant Physiol. 1997:114(1):\n237–244. https://doi.org/10.1104/pp.114.1.237\n\nGoodenough UW. Chloroplast division and pyrenoid formation in\nhttps:\nJ\n\nChlamydomonas\nPhycol.\n//doi.org/10.1111/j.1529-8817.1970.tb02348.x\n\n1970:6:1–6.\n\nreihardi.\n\nGoudet MMM, Orr DJ, Melkonian M, Müller KH, Meyer MT,\nCarmo-Silva E, Griffiths H. Rubisco and carbon-concentrating\n\n\fThe eukaryotic CO2-concentrating organelle\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\n|\n\n3255\n\nmechanism co-evolution across chlorophyte and streptophyte green\nalgae. New Phytol. 2020:227(3):810–823. https://doi.org/10.1111/\nnph.16577\n\nGriffiths DJ. The pyrenoid. Bot Rev. 1970:36(1):29–58. https://doi.org/\n\n10.1007/BF02859154\n\nGriffiths H, Meyer MT, Rickaby REM. Overcoming adversity through\ndiversity: aquatic carbon concentrating mechanisms. J Exp Bot.\n2017:68(14):3689–3695. https://doi.org/10.1093/jxb/erx278\n\nHall W, Claus G. Ultrastructural studies on the blue-green algal sym-\nbiont in Cyanophora Paradoxa korschikoff. J Cell Biol. 1963:19(3):\n551–563. https://doi.org/10.1083/jcb.19.3.551\n\nHammel A, Zimmer D, Sommer F, Muhlhaus T, Schroda M. Absolute\nquantification of major photosynthetic protein complexes\nin\nChlamydomonas reinhardtii using quantification concatamers\n(QconCATs). Front Plant Sci. 2018:9:1265. https://doi.org/10.3389/\nfpls.2018.01265\n\nHansen G, Daugbjerg N. Symbiodinium natans sp. nov.: a “free-living”\ndinoflagellate from Tenerife (Northeast-Atlantic Ocean). J Phycol.\n2009:45(1):251–263. https://doi.org/10.1111/j.1529-8817.2008.00621.x\nHanson D, Andrews TJ, Badger MR. Variability of the pyrenoid-based\nCO2 concentrating mechanism in hornworts (Anthocerotophyta).\nFunct Plant Biol. 2002:29(3):407–416. https://doi.org/10.1071/PP01210\nHanson DT, Franklin LA, Samuelsson G, Badger MR. The\nChlamydomonas reinhardtii cia3 mutant lacking a thylakoid lumen-\nlocalized carbonic anhydrase is limited by CO2 supply to Rubisco and\nnot photosystem II function in vivo. Plant Physiol. 2003:132(4):\n2267–2275. https://doi.org/10.1104/pp.103.023481\n\nHanson MR, Lin MT, Carmo-Silva AE, Parry MA. Towards engineering\ncarboxysomes into C3 plants. Plant J. 2016:87(1):38–50. https://doi.\norg/10.1111/tpj.13139\n\nHarris EH. Chlamydomonas In the laboratory. III. Culture conditions\n5. Synchronous cultures. In: The Chlamydomonas sourcebook. 2nd\ned. Canada: Elsevier; 2009, p. 252–254.\n\nHe S, Chou HT, Matthies D, Wunder T, Meyer MT, Atkinson N,\nMartinez-Sanchez A, Jeffrey PD, Port SA, Patena W, et al. The\nstructural basis of Rubisco phase separation in the pyrenoid. Nat\nPlants. 2020:6(12):1480–1490. https://doi.org/10.1038/s41477-020-\n00811-y\n\nHennacy JH, Jonikas MC. Prospects for engineering biophysical CO2\nconcentrating mechanisms into land plants to enhance yields.\nAnnu Rev Plant Biol. 2020:71(1):461–485. https://doi.org/10.1146/\nannurev-arplant-081519-040100\n\nHeyduk K, Moreno-Villena JJ, Gilman IS, Christin PA, Edwards EJ.\nThe genetics of convergent evolution: insights from plant photosyn-\nthesis. Nat Rev Genet. 2019:20(8):485–493. https://doi.org/10.1038/\ns41576-019-0107-5\n\nHoldsworth RH. The presence of a crystalline matrix in pyrenoids of\nthe diatom, Achnanthes brevipes. J Cell Biol. 1968:37(3):831–837.\nhttps://doi.org/10.1083/jcb.37.3.831\n\nHoldsworth RH. The isolation and partial characterization of the pyr-\nenoid protein of Eremosphaera viridis. J Cell Biol. 1971:51(2):499–513.\nhttps://doi.org/10.1083/jcb.51.2.499\n\nHyman AA, Weber CA, Julicher F. Liquid-liquid phase separation in\nbiology. Annu Rev Cell Dev Biol. 2014:30(1):39–58. https://doi.org/\n10.1146/annurev-cellbio-100913-013325\n\nItakura AK, Chan KX, Atkinson N, Pallesen L, Wang L, Reeves G,\nPatena W, Caspari O, Roth R, Goodenough U, et al. A\nRubisco-binding protein is required for normal pyrenoid number\nand starch sheath morphology in Chlamydomonas reinhardtii.\nProc Natl Acad Sci U S A. 2019:116(37):18445–18454. https://doi.\norg/10.1073/pnas.1904587116\n\nJackson C, Knoll AH, Chan CX, Verbruggen H. Plastid phylogenomics\nwith broad taxon sampling further elucidates the distinct evolution-\nary origins and timing of secondary green plastids. Sci Rep. 2018:8(1):\n1523. https://doi.org/10.1038/s41598-017-18805-w\n\nJin S, Sun J, Wunder T, Tang D, Cousins AB, Sze SK, Mueller-Cajar O,\nGao YG. Structural insights into the LCIB protein family reveals a new\n\ngroup of beta-carbonic anhydrases. Proc Natl Acad Sci U S A.\n2016:113(51):14716–14721. https://doi.org/10.1073/pnas.1616294113\nKajala K, Covshoff S, Karki S, Woodfield H, Tolley BJ, Dionora MJ,\nMogul RT, Mabilangan AE, Danila FR, Hibberd JM, et al.\nStrategies for engineering a two-celled C(4) photosynthetic pathway\ninto rice. J Exp Bot. 2011:62(9):3001–3010. https://doi.org/10.1093/\njxb/err022\n\nKamikawa R, Tanifuji G, Kawachi M, Miyashita H, Hashimoto T,\nInagaki Y. Plastid genome-based phylogeny pinpointed the origin\nof the green-colored plastid in the dinoflagellate Lepidodinium chlor-\nophorum. Genome Biol Evol. 2015:7(4):1133–1140. https://doi.org/\n10.1093/gbe/evv060\n\nKaplan A. On the cradle of CCM research: discovery, development, and\nchallenges ahead. J Exp Bot. 2017:68(14):3785–3796. https://doi.org/\n10.1093/jxb/erx122\n\nKaras BJ, Diner RE, Lefebvre SC, McQuaid J, Phillips AP, Noddings\nCM, Brunson JK, Valas RE, Deerinck TJ, Jablanovic J, et al.\nDesigner diatom episomes delivered by bacterial conjugation. Nat\nCommun. 2015:6(1):6925. https://doi.org/10.1038/ncomms7925\nKarlsson J, Clarke AK, Chen ZY, Hugghins SY, Park YI, Husic HD,\nMoroney JV, Samuelsson G. A novel alpha-type carbonic anhydrase\nassociated with the thylakoid membrane in Chlamydomonas rein-\nhardtii is required for growth at ambient CO2. EMBO J. 1998:17(5):\n1208–1216. https://doi.org/10.1093/emboj/17.5.1208\n\nKasili RW, Rai AK, Moroney JV. LCIB functions as a carbonic anhy-\ndrase: evidence from yeast and Arabidopsis carbonic anhydrase\nknockout mutants. Photosynth Res. 2023:156(2):193-204. https://\ndoi.org/10.1007/s11120-023-01005-1\n\nKeeling PJ. The endosymbiotic origin, diversification and fate of plas-\ntids. Philos Trans R Soc B Biol Sci. 2010:365(1541):729–748. https://\ndoi.org/10.1098/rstb.2009.0103\n\nKikutani S, Nakajima K, Nagasato C, Tsuji Y, Miyatake A, Matsuda Y.\nThylakoid luminal theta-carbonic anhydrase critical for growth and\nphotosynthesis in the marine diatom Phaeodactylum tricornutum.\nProc Natl Acad Sci U S A. 2016:113(35):9828–9833. https://doi.org/\n10.1073/pnas.1603112113\n\nKnoll AH. The early evolution of eukaryotes: a geological perspective.\nScience. 1992:256(5057):622–627. https://doi.org/10.1126/science.\n1585174\n\nKohinata T, Nishino H, Fukuzawa H. Significance of zinc in a regula-\ntory protein, CCM1, which regulates the carbon-concentrating\nin Chlamydomonas reinhardtii. Plant Cell Physiol.\nmechanism\n2008:49(2):273–283. https://doi.org/10.1093/pcp/pcn003\n\nKono A, Spalding MH. LCI1, A Chlamydomonas reinhardtii plasma\nmembrane protein, functions in active CO2 uptake under low CO2.\nPlant J. 2020:102(6):1127–1141. https://doi.org/10.1111/tpj.14761\nKowallik K. The crystal lattice of the pyrenoid matrix of Prorocentrum\nmicans. J Cell Sci. 1969:5(1):251–269. https://doi.org/10.1242/jcs.5.1.\n251\n\nKuchitsu K, Tsuzuki M, Miyachi S. Changes of starch localization within\nthe chloroplast induced by changes in CO2 concentration during\ngrowth of Chlamydomonas reinhardtii: independent regulation of pyr-\nenoid starch and stroma starch. Plant Cell Physiol. 1988:29(8):1269–\n1278. https://doi.org/10.1093/oxfordjournals.pcp.a077635\n\nKuchitsu K, Tsuzuki M, Miyachi S. Polypeptide composition and en-\nzyme activities of the pyrenoid and its regulation by CO, concentra-\ntion in unicellular green algae. Can J Bot. 1991:69(5):1062–1069.\nhttps://doi.org/10.1139/b91-136\n\nKüken A, Sommer F, Yaneva-Roder L, Mackinder LC, Hohne M,\nGeimer S, Jonikas MC, Schroda M, Stitt M, Nikoloski Z, et al.\nEffects of microcompartmentation on flux distribution and metabol-\nic pools in Chlamydomonas reinhardtii chloroplasts. Elife. 2018:7:\ne37960. https://doi.org/10.7554/eLife.37960\n\nKupriyanova EV, Pronina NA, Los DA. Adapting from low to high: an\nupdate to CO(2)-concentrating mechanisms of cyanobacteria and\nmicroalgae. Plants (Basel) 2023:12(7):1569. https://doi.org/10.3390/\nplants12071569\n\n\f3256\n\n|\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\nHe et al.\n\nin\n\nthe\n\nKusel-Fetzmann E, Weidinger M. Ultrastructure of five Euglena spe-\nsubdivision Serpentes. Protoplasma.\ncies positioned\n2008:233(3–4):209–222. https://doi.org/10.1007/s00709-008-0005-8\nLacoste-Royal G, Gibbs SP. Immunocytochemical localization of\nRibulose-1,5-Bisphosphate Carboxylase in the pyrenoid and thyla-\nkoid region of the chloroplast of Chlamydomonas reinhardtii.\nPlant Physiol. 1987:83:602–606. https://doi.org/10.1104/pp.83.3.\n602\n\nLau CS, Dowle A, Thomas GH, Girr1 P, Mackinder LC. A phase-separated\nCO2-fixing pyrenoid proteome determined by TurboID in Chlamydomonas\nreinhardtii. Plant Cell 2023:35(9):3260–3279. https://doi.org/10.1093/\nplcell/koad131\n\nLaza-Martínez A, Arluzea J, Miguel I, Orive E. Morphological and mo-\nlecular characterization of Teleaulax gracilis sp. nov. and T. minuta sp.\nnov. (Cryptophyceae). Phycologia. 2012:51(6):649–661. https://doi.\norg/10.2216/11-044.1\n\nLi P, Banjade S, Cheng HC, Kim S, Chen B, Guo L, Llaguno M,\nHollingsworth JV, King DS, Banani SF, et al. Phase transitions in\nthe\nsignalling proteins. Nature.\n2012:483(7389):336–340. https://doi.org/10.1038/nature10879\n\nassembly of multivalent\n\nLi FW, Nishiyama T, Waller M, Frangedakis E, Keller J, Li Z,\nFernandez-Pozo N, Barker MS, Bennett T, Blazquez MA, et al.\nAnthoceros genomes illuminate the origin of land plants and the un-\nique biology of hornworts. Nat Plants. 2020:6(3):259–272. https://doi.\norg/10.1038/s41477-020-0618-2\n\nLi X, Patena W, Fauser F, Jinkerson RE, Saroussi S, Meyer MT,\nIvanova N, Robertson JM, Yue R, Zhang R, et al. A genome-wide\nalgal mutant library and functional screen identifies genes required\nfor eukaryotic photosynthesis. Nat Genet. 2019:51(4):627–635.\nhttps://doi.org/10.1038/s41588-019-0370-6\n\nLi X, Zhang R, Patena W, Gang SS, Blum SR, Ivanova N, Yue R,\nRobertson JM, Lefebvre PA, Fitz-Gibbon ST, et al. An indexed,\nmapped mutant library enables reverse genetics studies of biological\nprocesses in Chlamydomonas reinhardtii. Plant Cell. 2016:28(2):\n367–387. https://doi.org/10.1105/tpc.15.00465\n\nLibessart N, Maddelein ML, Koornhuyse N, Decq A, Delrue B,\nMouille G, D’Hulst C, Ball S. Storage, photosynthesis, and growth:\nthe conditional nature of mutations affecting starch synthesis and\nstructure\nin Chlamydomonas. Plant Cell. 1995:7(8):1117–1127.\nhttps://doi.org/10.2307/3870089\n\nMa Y, Pollock SV, Xiao Y, Cunnusamy K, Moroney JV. Identification\nof a novel gene, CIA6, required for normal pyrenoid formation in\nChlamydomonas reinhardtii. Plant Physiol. 2011:156(2):884–896.\nhttps://doi.org/10.1104/pp.111.173922\nMaberly SC, Gontero B. Ecological\n\nfor aquatic\nCO2-concentrating mechanisms. J Exp Bot. 2017:68(14):3797–3814.\nhttps://doi.org/10.1093/jxb/erx201\n\nimperatives\n\nMackinder LCM, Chen C, Leib RD, Patena W, Blum SR, Rodman M,\nRamundo S, Adams CM, Jonikas MC. A spatial interactome reveals\nthe protein organization of the algal CO2-concentrating mechanism.\nCell. 2017:171(1):133–147.e14. https://doi.org/10.1016/j.cell.2017.08.\n044\n\nMackinder LC, Meyer MT, Mettler-Altmann T, Chen VK, Mitchell\nMC, Caspari O, Freeman Rosenzweig ES, Pallesen L, Reeves G,\nItakura A, et al. A repeat protein links Rubisco to form the eukary-\notic carbon-concentrating organelle. Proc Natl Acad Sci U S A.\n2016:113(21):5958–5963. https://doi.org/10.1073/pnas.1522866113\nMaddelein ML, Libessart N, Bellanger F, Delrue B, D’Hulst C, Van\nden Koornhuyse N, Fontaine T, Wieruszeski JM, Decq A, Ball S.\nToward an understanding of the biogenesis of the starch granule.\nDetermination of granule-bound and soluble starch synthase func-\ntions in amylopectin synthesis. J Biol Chem. 1994:269(40):25150–-\n25157. https://doi.org/10.1016/S0021-9258(17)31510-7\n\nMangan NM, Flamholz A, Hood RD, Milo R, Savage DF. Ph deter-\nmines the energetic efficiency of the cyanobacterial CO2 concentrat-\ning mechanism. Proc Natl Acad Sci U S A. 2016:113(36):E5354-5362.\nhttps://doi.org/10.1073/pnas.1525145113\n\nMann DG. Chloroplast morphology, movements and inheritance in\ndiatoms. In: Cytology, genetics and molecular biology of algae.\nAmsterdam, the Netherlands: SPB Academic Publishing; 1996.\np. 249–274.\n\nManuel LJ, Moroney\n\nInorganic carbon accumulation by\nChlamydomonas reinhardtii: new proteins are made during adapta-\ntion to low CO2. Plant Physiol. 1988:88(2):491–496. https://doi.org/\n10.1104/pp.88.2.491\n\nJV.\n\nMarek LF, Spalding MH. Changes in photorespiratory enzyme activity\nin response to limiting CO2 in Chlamydomonas reinhardtii. Plant\nPhysiol. 1991:97(1):420–425. https://doi.org/10.1104/pp.97.1.420\nMatsuda Y, Hopkinson BM, Nakajima K, Dupont CL, Tsuji Y.\nMechanisms of carbon dioxide acquisition and CO2 sensing in mar-\nine diatoms: a gateway to carbon metabolism. Philos Trans R Soc\nLond B Biol Sci. 2017:372(1728):20160403. https://doi.org/10.1098/\nrstb.2016.0403\n\nMatsuoka M, Furbank RT, Fukayama H, Miyao M. Molecular engin-\neering of C4 photosynthesis. Annu Rev Plant Physiol Plant Mol\nBiol. 2001:52(1):297–314. https://doi.org/10.1146/annurev.arplant.\n52.1.297\n\nMcKay RM, Gibbs SP. Phycoerythrin is absent from the pyrenoid of\nimplications. Planta.\n\nPorphyridium cruentum: photosynthetic\n1990:180(2):249–256. https://doi.org/10.1007/BF00194004\n\nMcKay R, Gibbs S. Composition and function of pyrenoids: cytochem-\nical and immunocytochemical approaches. Can J Bot. 1991:69(5):\n1040–1052. https://doi.org/10.1139/b91-134\n\nMcKay RML, Gibbst SP, Vaughn KC. RuBisCo activase is present in the\n\npyrenoid of green algae. Protoplasma. 1991:162:38–45.\n\nFritz-Laylin\n\nSalamov A,\n\nMerchant SS, Prochnik SE, Vallon O, Harris EH, Karpowicz SJ,\nWitman GB, Terry A,\nLK,\nMaréchal-Drouard L, et al. The Chlamydomonas genome reveals\nfunctions. Science.\nthe evolution of key animal and plant\n2007:318(5848):245–250. https://doi.org/10.1126/science.1143609\nMeyer MT, Genkov T, Skepper JN, Jouhet J, Mitchell MC,\nSpreitzer RJ, Griffiths H. Rubisco small-subunit α-helices\ncontrol pyrenoid formation in Chlamydomonas. Proc Natl Acad\nSci U S A. 2012:109(47):19474–19479. https://doi.org/10.1073/\npnas.1210993109\n\nMeyer MT, Goudet MMM, Griffiths H. The algal pyrenoid. In: Larkum\nAWD, Grossman AR, Raven JA, editors. Photosynthesis in algae: bio-\nchemical and physiological mechanisms. Springer Nature Switzerland\nAG; 2020a. p. 179–203.\n\nMeyer M, Griffiths H. Origins and diversity of eukaryotic\nCO2-concentrating mechanisms: lessons for the future. J Exp Bot.\n2013:64(3):769–786. https://doi.org/10.1093/jxb/ers390\n\nMeyer MT, Itakura AK, Patena W, Wang L, He S, Emrich-Mills T, Lau\nCS, Yates G, Mackinder LCM, Jonikas MC. Assembly of the algal\nCO2-fixing organelle, the pyrenoid, is guided by a Rubisco-binding\nmotif. Sci Adv. 2020b:6(46):eabd2408. https://doi.org/10.1126/\nsciadv.abd2408\n\nMeyer M, Seibt U, Griffiths H. To concentrate or ventilate? Carbon ac-\nquisition, isotope discrimination and physiological ecology of early\nland plant\nlife forms. Philos Trans R Soc Lond B Biol Sci.\n2008:363(1504):2767–2778. https://doi.org/10.1098/rstb.2008.0039\nMeyer MT, Whittaker C, Griffiths H. The algal pyrenoid: key un-\nanswered questions. J Exp Bot. 2017:68(14):3739–3749. https://doi.\norg/10.1093/jxb/erx178\n\nMikhailyuk T, Holzinger A, Massalski A, Karsten U. Morphology and\nultrastructure of Interfilum and Klebsormidium (Klebsormidiales.\nStreptophyta) with special reference to cell division and thallus for-\nmation. Eur J Phycol. 2014:49(4):395–412. https://doi.org/10.1080/\n09670262.2014.949308\n\nMitchell MC, Meyer MT, Griffiths H. Dynamics of carbon-\nconcentrating mechanism induction and protein relocalization dur-\ning the dark-to-light transition in synchronized Chlamydomonas re-\ninhardtii. Plant Physiol. 2014:166(2):1073–1082. https://doi.org/10.\n1104/pp.114.246918\n\n\fThe eukaryotic CO2-concentrating organelle\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\n|\n\n3257\n\nMiura K, Yamano T, Yoshioka S, Kohinata T, Inoue Y, Taniguchi F,\nAsamizu E, Nakamura Y, Tabata S, Yamato KT, et al. Expression\nprofiling-based identification of CO2-responsive genes regulated by\nCCM1\nin\nChlamydomonas reinhardtii. Plant Physiol. 2004:135(3):1595–1607.\nhttps://doi.org/10.1104/pp.104.041400\n\ncarbon-concentrating mechanism\n\ncontrolling\n\na\n\nMiyagawa A, Okami T, Kira N, Yamaguchi H, Ohnishi K, Adachi M.\nResearch note: high efficiency transformation of the diatom\nPhaeodactylum tricornutum with a promoter from the diatom\nCylindrotheca fusiformis. Phycol Res. 2009:57(2):142–146. https://\ndoi.org/10.1111/j.1440-1835.2009.00531.x\n\nMoroney JV, Husic HD, Tolbert NE, Kitayama M, Manuel LJ,\nTogasaki RK.\nIsolation and characterization of a mutant of\nChlamydomonas reinhardtii deficient in the CO2 concentrating\nmechanism. Plant Physiol. 1989:89(3):897–903. https://doi.org/10.\n1104/pp.89.3.897\n\nMoroney JV, Jungnick N, Dimario RJ, Longstreth DJ. Photorespiration\nand carbon concentrating mechanisms: two adaptations to high O2,\nlow CO2 conditions. Photosynth Res. 2013:117(1–3):121–131.\nhttps://doi.org/10.1007/s11120-013-9865-7\n\nMorris JL, Puttick MN, Clark JW, Edwards D, Kenrick P, Pressel S,\nWellman CH, Yang Z, Schneider H, Donoghue PCJ. The timescale\nof early\nland plant evolution. Proc Natl Acad Sci U S A.\n2018:115(10):E2274–E2283. https://doi.org/10.1073/pnas.1719588115\nMukherjee A, Lau CS, Walker CE, Rai AK, Prejean CI, Yates G,\nEmrich-Mills T, Lemoine SG, Vinyard DJ, Mackinder LCM, et al.\nThylakoid localized bestrophin-like proteins are essential for the\nCO2 concentrating mechanism of Chlamydomonas reinhardtii.\nProc Natl Acad Sci U S A. 2019:116(34):16915–16920. https://doi.\norg/10.1073/pnas.1909706116\n\nMüller OF. Flora Danica. 1782:14:XIV. Denmark: Otto Friedrich Müller.\nNam O, Grouneva I, Mackinder LCM. Endogenous GFP tagging in the\ndiatom Thalassiosira pseudonana. bioRxiv. 2022. doi: https://doi.org/\n10.1101/2022.09.30.510313\n\nNelson WA, Ryan KG. Porphyridium purpureum (Bory) Drew et Ross\n(Porphyridiales, Rhodophyceae)—first record of a marine unicellular\nred alga in New Zealand. J R Soc N Z. 1988:18(1):127–128. https://doi.\norg/10.1080/03036758.1988.10421701\n\nNeofotis P, Temple J, Tessmer OL, Bibik J, Norris N, Pollner E, Lucker\nB, Weraduwage SM, Withrow A, Sears B, et al. The induction of\npyrenoid synthesis by hyperoxia and its implications for the natural\ndiversity of photosynthetic responses in Chlamydomonas. Elife.\n2021:10:e67565. https://doi.org/10.7554/eLife.67565\n\nNot F, Latasa M, Marie D, Cariou T, Vaulot D, Simon N. A single spe-\ncies, Micromonas pusilla (prasinophyceae), dominates the eukaryotic\npicoplankton\nin the Western English Channel. Appl Environ\nMicrobiol. 2004:70(7):4064–4072. https://doi.org/10.1128/AEM.70.7.\n4064-4072.2004\n\nNudelman MA, Leonardi PI, Conforti V, Farmer MA, Triemer RE.\nFine structure and taxonomy of Monomorphina Aenigmatica\ncomb. nov. (Euglenophyta). J Phycol. 2006:42(1):194–202. https://\ndoi.org/10.1111/j.1529-8817.2006.00170.x\n\nOhad I, Siekevitz P, Palade GE. Biogenesis of chloroplast membranes.\nII. Plastid differentiation during greening of a dark-grown algal mu-\ntant (Chlamydomonas reinhardi). J Cell Biol. 1967a:35(3):553–584.\nhttps://doi.org/10.1083/jcb.35.3.553\n\nOhad\n\nI, Siekevitz P, Palade GE. Biogenesis of chloroplast\nmembranes. I. Plastid dedifferentiation in a dark-grown algal mutant\n(Chlamydomonas reinhardi).\nJ Cell Biol. 1967b:35(3):521–552.\nhttps://doi.org/10.1083/jcb.35.3.521\n\nOhnishi N, Mukherjee B, Tsujikawa T, Yanase M, Nakano H,\nMoroney JV, Fukuzawa H. Expression of a low CO(2)-inducible pro-\ntein, LCI1, increases inorganic carbon uptake in the green alga\nChlamydomonas reinhardtii. Plant Cell. 2010:22(9):3105–3117.\nhttps://doi.org/10.1105/tpc.109.071811\n\nOta S, Ueda K, Ishida K. Taxonomic study of Bigelowiella longifila sp.\nnov. (Chlorarachniophyta) and a time-lapse video observation of the\n\nunique migration of amoeboid cells. J Phycol. 2007:43(2):333–343.\nhttps://doi.org/10.1111/j.1529-8817.2007.00316.x\n\nPronina NA, Semenenko VE. Membrane-bound carbonic anhydrase\nin algae cells. Curr Res\nhttps:\n\ntakes part\nPhotosynth.\n//doi.org/10.1007/978-94-009-0511-5_739\n\nin CO2 concentration\n\n1990:4(18):489–492.\n\nQiu Y-L, Li L, Wang B, Chen Z, Knoop V, Groth-Malonek M,\nDombrovska O, Lee J, Kent L, Rest J, et al. The deepest divergences\nin land plants inferred from phylogenomic evidence. Proc Natl Acad\nSci U S A. 2006:103(42):15511–15516. https://doi.org/10.1073/pnas.\n0603335103\n\nRae BD, Long BM, Badger MR, Price GD. Functions, compositions, and\nevolution of the two types of carboxysomes: polyhedral microcom-\npartments that facilitate CO2 fixation in cyanobacteria and some\nproteobacteria. Microbiol Mol Biol Rev. 2013:77(3):357–379.\nhttps://doi.org/10.1128/MMBR.00061-12\n\nRae BD, Long BM, Forster B, Nguyen ND, Velanis CN, Atkinson N,\nHee WY, Mukherjee B, Price GD, McCormick AJ. Progress and chal-\nlenges of engineering a biophysical CO2-concentrating mechanism\ninto higher plants. J Exp Bot. 2017:68(14):3717–3737. https://doi.\norg/10.1093/jxb/erx133\n\nRamazanov Z, Rawat M, Henk MC, Mason CB, Matthews SW,\nMoroney JV. The induction of the CO2-concentrating mechanism\nis correlated with the formation of the starch sheath around the pyr-\nenoid of Chlamydomonas reinhardtii. Planta. 1994:195(2):210–216.\nhttps://doi.org/10.1007/BF00199681\n\nRaven JA. CO2-concentrating mechanisms: a direct role for thylakoid\nlumen acidification? Plant Cell Environ. 2008:20(2):147–154.\nhttps://doi.org/10.1046/j.1365-3040.1997.d01-67.x\n\nRaven JA. Rubisco: still the most abundant protein of Earth? New\n\nPhytol. 2013:198(1):1–3. https://doi.org/10.1111/nph.12197\n\nRaven JA, Beardall J, Sanchez-Baracaldo P. The possible evolution and\nfuture of CO2-concentrating mechanisms. J Exp Bot. 2017:68(14):\n3701–3716. https://doi.org/10.1093/jxb/erx110\n\nRedekop P, Sanz-Luque E, Yuan Y, Villain G, Petroutsos D,\nGrossman AR. Transcriptional regulation of photoprotection in\ndark-to-light transition—more than just a matter of excess light en-\nergy. Sci Adv. 2022:8(22):eabn1832. https://doi.org/10.1126/sciadv.\nabn1832\n\nRetallack B, Butler RD. The development and structure of pyrenoids in\nBulbochaete hiloensis. J Cell Sci. 1970:6(1):229–241. https://doi.org/\n10.1242/jcs.6.1.229\n\nReyes-Prieto A, Weber APM, Bhattacharya D. The origin and establish-\nment of the plastid in algae and plants. Annu Rev Genet. 2007:41(1):\n147–168. https://doi.org/10.1146/annurev.genet.41.110306.130134\nRosowski KA, Sai T, Vidal-Henriquez E, Zwicker D, Style RW,\nDufresne ER. Elastic ripening and inhibition of liquid-liquid phase\nseparation. Nat Phys. 2020:16(4):422–425. https://doi.org/10.1038/\ns41567-019-0767-2\n\nRousseaux C, Gregg W. Interannual variation in phytoplankton pri-\nmary production at a global scale. Remote Sens. 2013:6(1):1–19.\nhttps://doi.org/10.3390/rs6010001\n\nSage RF, Sage TL, Kocacinar F. Photorespiration and the evolution of\nC4 photosynthesis. Annu Rev Plant Biol. 2012:63(1):19–47. https://\ndoi.org/10.1146/annurev-arplant-042811-105511\n\nSager R, Palade GE. Chloroplast structure in green and yellow strains of\nChlamydomonas. Exp Cell Res. 1954:7(2):584–588. https://doi.org/10.\n1016/S0014-4827(54)80107-8\n\nSager R, Palade GE. Structure and development of the chloroplast in\nChlamydomonas. I. The normal green cell. J Biophys Biochem\nCytol. 1957:3(3):463–488. https://doi.org/10.1083/jcb.3.3.463\n\nSánchez-Baracaldo P, Raven JA, Pisani D, Knoll AH. Early photosyn-\nthetic eukaryotes inhabited low-salinity habitats. Proc Natl Acad Sci\nU S A. 2017:114(37):E7737–E7745. https://doi.org/10.1073/pnas.\n1620089114\n\nSanthanagopalan I, Wong R, Mathur T, Griffiths H. Orchestral man-\nlight: crosstalk needed for regulation of the\n\noeuvres\n\nin the\n\n\f3258\n\n|\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\nHe et al.\n\nChlamydomonas carbon concentration mechanism. J Exp Bot.\n2021:72(13):4604–4624. https://doi.org/10.1093/jxb/erab169\n\nSavir Y, Noor E, Milo R, Tlusty T. Cross-species analysis traces adapta-\ntion of Rubisco toward optimality in a low-dimensional landscape.\nProc Natl Acad Sci U S A. 2010:107(8):3475–3480. https://doi.org/\n10.1073/pnas.0911663107\n\nSchirrmeister BE, Gugger M, Donoghue PCJ. Cyanobacteria and the\nfossils—\ngreat oxidation event: evidence\nSchirrmeister—2015—Palaeontology—Wiley\nLibrary.\nPalaeontology. 2015:58(5):769–785. https://doi.org/10.1111/pala.\n12178\n\nfrom genes and\n\nOnline\n\nSchmitz F. Die chromatophoren der algen: vergleichende untersuchun-\ngen über bau und entwicklung der chlorophyllkörper und der analo-\ngen farbstoffkörper der algen. Germany: M. Cohen & Sohn (F.\nCohen); 1882.\n\nSchnepf E, ElbräChter M. Dinophyte chloroplasts and phylogeny—a\nhttps://doi.org/10.1080/\n\n1999:38(2–3):81–97.\n\nreview. Grana.\n00173139908559217\n\nSchober AF, Ri OBRC, Bischoff A, Lepetit B, Gruber A, Kroth PG.\nOrganelle studies and proteome analyses of mitochondria and\nplastids fractions from the diatom Thalassiosira pseudonana.\nPlant Cell Physiol. 2019:60(8):1811–1828. https://doi.org/10.\n1093/pcp/pcz097\n\nShimamura D, Yamano T, Niikawa Y, Hu D, Fukuzawa H. A\npyrenoid-localized protein SAGA1 is necessary for Ca(2+)-binding pro-\ntein CAS-dependent expression of nuclear genes encoding inorganic\ncarbon transporters in Chlamydomonas reinhardtii. Photosynth Res.\n2023:156(2):181–192. https://doi.org/10.1007/s11120-022-00996-7\nShin Y, Brangwynne CP. Liquid phase condensation in cell physiology\nand disease. Science. 2017:357(6357):eaaf4382. https://doi.org/10.\n1126/science.aaf4382\n\nShiratori T, Fujita S, Shimizu T, Nakayama T, Ishida KI. Viridiuvalis\nadhaerens gen. et sp. nov., a novel colony-forming chlorarachnio-\nphyte. J Plant Res. 2017:130(6):999–1012. https://doi.org/10.1007/\ns10265-017-0961-1\n\nSinetova MA, Kupriyanova EV, Markelova AG, Allakhverdiev SI,\nPronina NA. Identification and functional role of the carbonic anhy-\ndrase Cah3 in thylakoid membranes of pyrenoid of Chlamydomonas\nreinhardtii. Biochim Biophys Acta. 2012:1817(8):1248–1255. https://\ndoi.org/10.1016/j.bbabio.2012.02.014\n\nSmith EC, Griffiths H. The occurrence of the chloroplast pyrenoid is\ncorrelated with the activity of a COe-concentrating mechanism\nand carbon isotope discrimination in lichens and bryophytes.\nPlanta. 1996a:198(1):6–16. https://doi.org/10.1007/BF00197580\n\nSmith EC, Griffiths H. A pyrenoid-based carbon-concentrating mech-\nanism is present in terrestrial bryophytes of the class Anthocerotae.\nPlanta. 1996b:200(2):203–212. https://doi.org/10.1007/BF00208310\nSmith EC, Griffiths H. The role of carbonic anhydrase in photosyn-\nthesis and the activity of the carbon-concentrating-mechanism in\nbryophytes of the class Anthocerotae. New Phytol. 2000:145(1):\n29–37. https://doi.org/10.1046/j.1469-8137.2000.00559.x\n\nSpalding MH, Spreitzer RJ, Ogren WL. Carbonic anhydrase-deficient\nmutant of Chlamydomonas reinhardii requires elevated carbon diox-\nide concentration for photoautotrophic growth. Plant Physiol.\n1983:73(2):268–272. https://doi.org/10.1104/pp.73.2.268\n\nStone HA. Dynamics of drop deformation and breakup in viscous\nfluids. Annu Rev Fluid Mech. 1994:26(1):65–102. https://doi.org/10.\n1146/annurev.fl.26.010194.000433\n\nStrassert JFH, Irisarri I, Williams TA, Burki F. A molecular timescale\nfor eukaryote evolution with implications for the origin of red algal-\nderived plastids. Nat Commun. 2021:12:1–13. https://doi.org/10.\n1038/s41467-021-22044-z\n\nStrenkert D, Schmollinger S, Gallaher SD, Salome PA, Purvine SO,\nNicora CD, Mettler-Altmann T, Soubeyrand E, Weber APM,\nLipton MS, et al. Multiomics resolution of molecular events during\na day in the life of Chlamydomonas. Proc Natl Acad Sci U S A.\n2019:116(6):2374–2383. https://doi.org/10.1073/pnas.1815238116\n\nSuzuki E, Suzuki R. Variation of storage polysaccharides in photo-\ntrophic microorganism. J Appl Glycosci. 2013:60(1):21–27. https://\ndoi.org/10.5458/jag.jag.JAG-2012_016\n\nTcherkez GG, Farquhar GD, Andrews TJ. Despite slow catalysis\nand confused substrate specificity, all ribulose bisphosphate car-\nboxylases may be nearly perfectly optimized. Proc Natl Acad Sci\nU S A. 2006:103(19):7246–7251. https://doi.org/10.1073/pnas.\n0600605103\n\nThierstein HR, Young JR. Coccolithophores: from molecular processes\n\nto global impact. Heidelberg, Germany: Springer; 2004.\n\nTirumani S, Kokkanti M, Chaudhari V, Shukla M, Rao BJ. Regulation\nof CCM genes in Chlamydomonas reinhardtii during conditions of\nlight-dark cycles\nin synchronous cultures. Plant Mol Biol.\n2014:85(3):277–286. https://doi.org/10.1007/s11103-014-0183-z\n\nToguri T, Muto S, Mihara S, Miyachi S. Synthesis and degradation of\ncarbonic anhydrase in a synchronized culture of Chlamydomonas re-\ninhardtii. Plant Cell Physiol. 1989:30(4):533–539. https://doi.org/10.\n1093/oxfordjournals.pcp.a077772\n\nToyokawa C, Yamano T, Fukuzawa H. Pyrenoid starch sheath is re-\nquired for LCIB localization and the CO2-concentrating mechanism\nin green algae. Plant Physiol. 2020:182(4):1883–1893. https://doi.\norg/10.1104/pp.19.01587\n\nTurkina MV, Blanco-Rivero A, Vainonen JP, Vener AV, Villarejo A.\nCO2 Limitation induces specific redox-dependent protein phosphor-\nylation\nin Chlamydomonas reinhardtii. Proteomics. 2006:6(9):\n2693–2704. https://doi.org/10.1002/pmic.200500461\n\nVan K, Spalding MH. Periplasmic carbonic anhydrase structural gene\nin Chlamydomonas reinhardtii. Plant Physiol.\n\n(Cah1) mutant\n1999:120(3):757–764. https://doi.org/10.1104/pp.120.3.757\n\nvan Baren MJ, Bachy C, Reistetter EN, Purvine SO, Grimwood J,\nSudek S, Yu H, Poirier C, Deerinck TJ, Kuo A, et al.\nEvidence-based green algal genomics reveals marine diversity and an-\ncestral characteristics of land plants. BMC Genomics. 2016:17(1):\n1–22. https://doi.org/10.1186/s12864-016-2585-6\n\nVaucher J-P. Histoire des conferves d’eau douce: contenant leurs\ndifférents modes de reproduction, et la description de leurs princi-\npales espèces, suivie de l’histoire des trémelles et des ulves d’eau\ndouce. Switzerland: J.J. Paschoud, Genéve; 1803.\n\nVillarreal JC, Renner SS. Hornwort pyrenoids, carbon-concentrating\nstructures, evolved and were lost at least five times during the\nlast 100 million years. Proc Natl Acad Sci U S A. 2012:109(46):\n18873–18878. https://doi.org/10.1073/pnas.1213498109\n\nVladimirova MG, Markelova AG, Semenenko VE. Identification of\nribulose bisphosphate carboxylase location in the pyrenoids of uni-\ncellular algae by the cytoimmunofluorescent method. Fiziol Rast.\n(Moscow). 1982:29:941–950. (English translation: Soviet Plant\nPhysiol 930: 725–734).\n\nWang H, Gau B, Slade WO, Juergens M, Li P, Hicks LM. The global\nphosphoproteome of Chlamydomonas reinhardtii reveals complex\norganellar phosphorylation in the flagella and thylakoid membrane.\nMol Cell Proteomics. 2014:13(9):2337–2353. https://doi.org/10.1074/\nmcp.M114.038281\n\nWang L, Jonikas MC. The pyrenoid. Curr Biol. 2020:30(10):R456–R458.\n\nhttps://doi.org/10.1016/j.cub.2020.02.051\n\nWang L, Patena W, Van Baalen KA, Xie Y, Singer ER, Gavrilenko S,\nWarren-Williams M, Han L, Harrigan HR, Chen V, et al. A chloro-\nplast protein atlas reveals novel structures an spatial organization of\nbiosynthetic pathways. bioRxiv. 2022. doi: https://doi.org/10.1101/\n2022.05.31.493820\n\nWang Y, Spalding MH. An inorganic carbon transport system respon-\nsible for acclimation specific to air levels of CO2 in Chlamydomonas\nreinhardtii. Proc Natl Acad Sci U S A. 2006:103(26):10110–10115.\nhttps://doi.org/10.1073/pnas.0603402103\n\nWang Y, Spalding MH. Acclimation to very low CO2: contribution of\nlimiting CO2 inducible proteins, LCIB and LCIA, to inorganic carbon\nuptake in Chlamydomonas reinhardtii. Plant Physiol. 2014a:166(4):\n2040–2050. https://doi.org/10.1104/pp.114.248294\n\n\fThe eukaryotic CO2-concentrating organelle\n\nTHE PLANT CELL 2023: 35; 3236–3259\n\n|\n\n3259\n\nWang Y,\n\nSpalding MH.\n\nthe Chlamydomonas\nLCIB\nCO2-concentrating mechanism. Photosynth Res. 2014b:121(2–3):\n185–192. https://doi.org/10.1007/s11120-013-9956-5\n\nin\n\nWang Y, Stessman DJ, Spalding MH. The CO2 concentrating mechan-\nism and photosynthetic carbon assimilation in limiting CO2: how\nChlamydomonas works against the gradient. Plant J. 2015:82(3):\n429–448. https://doi.org/10.1111/tpj.12829\n\nWang Y, Sun Z, Horken KM, Im C-S, Xiang Y, Grossman AR, Weeks\nDP. Analyses of CIA5, the master regulator of the carbon-\nconcentrating mechanism in Chlamydomonas reinhardtii, and its\ncontrol of gene expression. Can J Bot. 2005:83(7):765–779. https://\ndoi.org/10.1139/b05-062\n\nWang L, Yamano T, Takane S, Niikawa Y, Toyokawa C, Ozawa SI,\nTokutsu R, Takahashi Y, Minagawa J, Kanesaki Y, et al.\nChloroplast-mediated regulation of CO2-concentrating mechanism\nby Ca2+-binding protein CAS in the green alga Chlamydomonas re-\ninhardtii. Proc Natl Acad Sci U S A. 2016:113(44):12586–12591.\nhttps://doi.org/10.1073/pnas.1606519113\n\nWhitney SM, Houtz RL, Alonso H. Advancing our understanding and cap-\nacity to engineer nature’s CO2-sequestering enzyme, Rubisco. Plant\nPhysiol. 2011:155(1):27–35. https://doi.org/10.1104/pp.110.164814\nWunder T, Cheng SLH, Lai SK, Li HY, Mueller-Cajar O. The phase sep-\naration underlying the pyrenoid-based microalgal Rubisco super-\ncharger. Nat Commun. 2018:9(1):5076. https://doi.org/10.1038/\ns41467-018-07624-w\n\nXiang Y, Zhang J, Weeks DP. The cia5 gene controls formation of the\ncarbon concentrating mechanism in Chlamydomonas reinhardtii.\nProc Natl Acad Sci U S A. 2001:98(9):5341–5346. https://doi.org/10.\n1073/pnas.101534498\n\nYamano T, Sato E, Iguchi H, Fukuda Y, Fukuzawa H. Characterization\nof cooperative bicarbonate uptake into chloroplast stroma in the\ngreen alga Chlamydomonas reinhardtii. Proc Natl Acad Sci U S A.\n2015:112(23):7315–7320. https://doi.org/10.1073/pnas.1501659112\nYamano T, Toyokawa C, Fukuzawa H. High-resolution suborganellar\nlocalization of Ca(2+)-binding protein CAS, a novel regulator of\nCO2-concentrating mechanism.\n2018:255(4):\n1015–1022. https://doi.org/10.1007/s00709-018-1208-2\n\nProtoplasma.\n\nYamano T, Tsujikawa T, Hatano K, Ozawa S, Takahashi Y, Fukuzawa\nH. Light and low-CO2-dependent LCIB-LCIC complex localization in\nthe chloroplast supports the carbon-concentrating mechanism in\n\nChlamydomonas\n1453–1468. https://doi.org/10.1093/pcp/pcq105\n\nreinhardtii. Plant Cell Physiol. 2010:51(9):\n\nYanashima R, Garcia AA, Aldridge J, Weiss N, Hayes MA, Andrews\nJH. Cutting a drop of water pinned by wire loops using a superhydro-\nphobic surface and knife. PLoS One. 2012:7(9):e45893. https://doi.\norg/10.1371/journal.pone.0045893\n\nYoon HS, Hackett JD, Ciniglia C, Pinto G, Bhattacharya D. A molecu-\nlar timeline for the origin of photosynthetic eukaryotes. Mol Biol\nEvol. 2004:21(5):809–818. https://doi.org/10.1093/molbev/msh075\nYoshioka S, Taniguchi F, Miura K, Inoue T, Yamano T, Fukuzawa H.\nThe novel Myb\nthe\nCO2-responsive gene Cah1, encoding a periplasmic carbonic anhy-\ndrase\nin Chlamydomonas reinhardtii. Plant Cell. 2004:16(6):\n1466–1477. https://doi.org/10.1105/tpc.021162\n\nfactor LCR1\n\ntranscription\n\nregulates\n\nin\n\nthe photosynthetic model\n\nZaffagnini M, Bedhomme M, Groni H, Marchand CH, Puppo C,\nGontero B, Cassier-Chauvat C, Decottignies P, Lemaire SD.\norganism\nGlutathionylation\nChlamydomonas reinhardtii: a proteomic survey. Mol Cell Proteomics.\n2012:11(2):M111.014142. https://doi.org/10.1074/mcp.M111.014142\nZeeman SC, Kossmann J, Smith AM. Starch: its metabolism, evolution,\nand biotechnological modification in plants. Annu Rev Plant Biol.\nhttps://doi.org/10.1146/annurev-arplant-\n2010:61(1):209–234.\n042809-112301\n\nZhan Y, Dhaliwal JS, Adjibade P, Uniacke J, Mazroui R, Zerges W.\nLocalized control of oxidized RNA. J Cell Sci. 2015:128:4210–4219.\nhttps://doi.org/10.1242/jcs.175232\n\nZhan Y, Marchand CH, Maes A, Mauries A, Sun Y, Dhaliwal JS,\nUniacke J, Arragain S, Jiang H, Gold ND, et al. Pyrenoid functions\nrevealed by proteomics in Chlamydomonas reinhardtii. PLoS\nOne. 2018:13(2):e0185039. https://doi.org/10.1371/journal.pone.\n0185039\n\nZhang J, Huss VAR, Sun X, Chang K, Pang D. Morphology and phylo-\ngenetic position of a trebouxiophycean green alga (Chlorophyta)\ngrowing on the rubber tree, Hevea brasiliensis, with the description\nof a new genus and species. Eur J Phycol. 2008:43(2):185–193.\nhttps://doi.org/10.1080/09670260701718462\n\nZones JM, Blaby IK, Merchant SS, Umen JG. High-resolution profiling of\na synchronized diurnal transcriptome from Chlamydomonas reinhard-\ntii reveals continuous cell and metabolic differentiation. Plant Cell.\n2015:27:2743–2769. https://doi.org/10.1105/tpc.15.00498\n\n"
10.1021_acssynbio.2c00587
pubs.acs.org/synthbio Research Article Receptor Elimination by E3 Ubiquitin Ligase Recruitment (REULR): A Targeted Protein Degradation Toolbox Dirk H. Siepe, Lora K. Picton, and K. Christopher Garcia* Cite This: ACS Synth. Biol. 2023, 12, 1081−1093 Read Online ACCESS Metrics & More Article Recommendations *sı Supporting Information targeted protein degradation (TPD) of plasma ABSTRACT: In recent years, membrane proteins by hijacking the ubiquitin proteasome system (UPS) or the lysosomal pathway has emerged as a novel therapeutic avenue in drug development to address and inhibit canonically difficult targets. While TPD strategies have been in targeting cell surface receptors, these approaches are limited by the successful availability of suitable binders to generate heterobifunctional molecules. Here, we present the development of a nanobody (VHH)-based degradation toolbox termed REULR (Receptor Elimination by E3 Ubiquitin Ligase Recruitment). We generated human and mouse cross-reactive nanobodies against five transmembrane PA-TM- RING-type E3 ubiquitin ligases (RNF128, RNF130, RNF167, RNF43, and ZNRF3), covering a broad range and selectivity of tissue expression, with which we characterized the expression in human and mouse cell lines and immune cells (PBMCs). We demonstrate that heterobifunctional REULR molecules can enforce transmembrane E3 ligase interactions with a variety of disease-relevant target receptors (EGFR, EPOR, and PD-1) by induced proximity, resulting in effective membrane clearance of the target receptor at varying levels. In addition, we designed E3 ligase self-degrading molecules, “fratricide” REULRs (RNF128, RNF130, RENF167, RNF43, and ZNRF3), that allow downregulation of one or several E3 ligases from the cell surface and consequently modulate receptor signaling strength. REULR molecules represent a VHH-based modular and versatile “mix and match” targeting strategy for the facile modulation of cell surface proteins by induced proximity to transmembrane PA-TM-RING E3 ligases. KEYWORDS: targeted protein degradation, E3 ligase, receptor, induced proximity, REULR, fratricide, nanobody ■ INTRODUCTION Classical drug discovery approaches against membrane protein targets such as cell surface receptors generally rely on small molecule inhibitors and monoclonal antibodies, but the vast majority of disease-relevant cell surface receptors still remain extremely challenging to target and have been largely deemed “undruggable” by established screening strategies.1 Finding alternative strategies to target challenging plasma membrane proteins has therefore become a prime focus in recent years. Targeted protein degradation has emerged as a novel therapeutic strategy in drug development by directing proteins to the cells’ own degradation machinery (UPS).2−4 The majority of degraders such as PROTACs,2 molecular glues,5 dTags,6 or TRIM-Away7 are based on a heterobifunctional design that leads to the formation of a ternary complex between a cytosolic E3 ubiquitin ligase and a protein of to facilitate ubiquitination and subsequent 26S proteasome-dependent degradation.8 While classical degraders have been successful,1 this approach is ultimately limited to cytosolic targets, and therefore 1/3 of the protein-coding genes representing the membrane proteome are not accessible by this approach.9,10 interest More recently, targeted protein degradation approaches utilizing lysosomal degradation strategies (LYTAC and KineTac)11,12 and proteolysis-targeting antibodies (AbTac and ProTab) using WNT-related transmembrane E3 ligases (RNF43 and ZNRF3) have emerged.13,14 These approaches tether target proteins on the cell surface to either lysosome shuttling receptors or cell−surface E3 ubiquitin ligases to induce membrane clearance. Both technologies are mainly limited by the availability and specificity of shuttling receptors or transmembrane E3-binding moieties, selectivity (tissue ex- pression), design (antibody formatting), and complexity of production. In an effort to accelerate the development of targeted protein degradation tools, we present a modular and versatile nanobody (VHH)-based protein degradation toolbox termed REULR (Receptor Elimination by E3 Ubiquitin Ligase Recruitment). We generated human and mouse cross-reactive nanobodies against ECDs (extracellular domain) of five transmembrane PA- Received: November 4, 2022 Published: April 3, 2023 © 2023 The Authors. Published by American Chemical Society 1081 https://doi.org/10.1021/acssynbio.2c00587 ACS Synth. Biol. 2023, 12, 1081−1093 ACS Synthetic Biology pubs.acs.org/synthbio Research Article Figure 1. Transmembrane PA-TM-RING E3 ligase nanobodies for receptor elimination. (A) Pie chart representation of the transmembrane E3 ligase family classified by subcellular localization (upper chart). Plasma membrane-localized transmembrane E3 ligase subfamily grouped into subcellular and structurally related sub classes (lower chart). (B) Hierarchical two-way clustering heatmap of normal tissue mRNA expression data for the PA-TM- RING E3 ligase subfamily. (C) Schematic representation of the REULR concept. Enforced transmembrane E3 ubiquitin ligase recruitment to a target 1082 https://doi.org/10.1021/acssynbio.2c00587 ACS Synth. Biol. 2023, 12, 1081−1093 ACS Synthetic Biology Figure 1. continued pubs.acs.org/synthbio Research Article receptor reduces target receptor cell surface levels by E3 ligase-dependent intracellular ubiquitination and subsequent membrane clearance. (D) SPR sensograms and binding affinities of PA-TM-RING ligase-selected nanobodies (analytes) for human RNF128, RNF130, RNF167, RNF43, and ZNRF3 ECDs (ligands). (E) Cell surface staining of representative human (HEK293T, CaCo-2, YT1, and UT/7) and mouse (BaF3, 3T3, and B16) cell lines using a panel of five PA-TM-RING E3 ligase-binding nanobodies (nanobody:SA647 tetramers) and analysis by flow cytometry, full titration (1:1 dilutions; 200 nM tetramer), and Biotin:SA647 (Biotin) served as a negative control. (F) Staining data visualized in a normalized heatmap for human and mouse cell lines. (G) PBMC (Peripheral Blood Mononuclear Cells) immunophenotyping panel to identify the binding of five PA-TM- RING E3 ligase-binding nanobodies (nanobody:SA647 tetramers; 200 nM) to T cells (CD4+; CD8+), monocytes, B cells, and NK cells, analysis by flow cytometry. Biotin:SA647 served as a negative control (Biotin). Anti-PD1 and anti-CD69 were used as phenotyping control antibodies in comparison to an isotype control. (H) PBMC sub cell-type data summarized in a normalized heatmap. Data are represented as mean ± SD (n = 3). TM-RING-type E3 ubiquitin ligases (RNF128, RNF130, RNF167, RNF43, and ZNRF3), covering a broad range and selectivity of tissue expression. Next, we utilized our VHHs to characterize the expression of these five PA-TM-RING E3 ligases in commonly used human and mouse cell lines and immune cells (T cells, monocytes, B cells, and NK cells). We demonstrate that heterobifunctional REULR molecules can enforce transmembrane E3 ligase interactions with a variety of disease-relevant target receptors (EGFR, EPOR, and PD-1) by induced proximity, resulting in robust membrane attenuation of the target receptor. Furthermore, we present a strategic approach to tune transmembrane E3 ligases itself by generating homo-, heterobifunctional, and arrayed multimeric fratricide REULRs and consequently modulate signaling events of natural target receptors. ■ RESULTS PA-TM-RING E3 Ligase Nanobodies for Receptor Elimination. The human transmembrane (TM) E3 ligase family represents a class of diverse RING-type E3 ubiquitin ligases15,16 with approximately 50 members (Figure 1A upper chart). These proteins exert widespread involvement in several diseases and cancer.16,17 The family can be further grouped into subcellular and structurally related sub classes; the plasma membrane localized E3 TM ligases include RING domain- containing proteins (7), PA-TM-RING (10), RING between RING (RBR; 5), and the membrane-associated RING-CH (MARCH; 4) families (Figure 1A, lower chart). In general, E3 ligases are notoriously challenging to study, and their substrates still remain highly elusive, mostly due to the nature of the ubiquitylation cascade, which is characterized by very weak target affinities and fast kinetics.18−20 Here, we focused on the PA-TM-RING-type E3 ligases,21 a family of approximately 10 members with a broad tissue expression pattern (Figure 1B) and a unique domain is minimally defined by three conserved architecture that domains: an extracellular protease-associated (PA) domain that acts as a substrate recruitment domain, a transmembrane domain (TM), and a cytosolic catalytic RING-type E3 ligase domain (RING-H2 finger; RNF) (Figure 1C). Mechanistically, the cytosolic RING E3 ligase domain functions as an allosteric activator and scaffold that recruits the ubiquitin machinery in close proximity to a substrate, while the extracellular PA domain functions as a substrate recruitment domain. We therefore hypothesized that PA-TM-RING E3 ligases could be retasked to selectively eliminate non-natural cell surface targets by an induced proximity approach that we termed REULR: Receptor Elimination by E3 Ubiquitin Ligase Recruitment (Figure 1C). To develop a modular and versatile toolbox, we first identified five PA-TM-RING E3 ligases covering a wide range of tissue type-specific REULR approaches: expression, allowing cell RNF128 (GRAIL), RNF130 (GOLIATH), RNF167 (GOD- ZILLA), RNF43, and ZNRF3 (Figure 1B; marked in red). Therapeutic monoclonal antibodies (mABs) and antibody engineering have revolutionized cancer therapies in the last decade,22 but they are not without limitations, mainly size, complexity of formatting, expression, and modularity. In order to overcome these limitations, we took advantage of the superior pharmacokinetic properties of nanobodies (VHH) such as their small size (1/10 the size of conventional antibodies), high stability, strong antigen-binding affinity, modularity, and ease of expression.23−25 We screened a synthetic nanobody library, allowing rapid high-throughput selection by yeast display26 using the ECDs (extracellular domains) of human RNF128 (GRAIL), RNF130 (GOLIATH), RNF167 (GODZILLA), RNF43, and ZNRF3 that led to 8 nanobodies against 5 ligases with nanomolar to picomolar affinities (Figures 1D and S1A− C). A pairwise protein sequence alignment of the human and mouse ECDs of the five PA-TM-RING-type E3 ligases revealed that the ECDs are highly conserved between both species, with an average amino acid sequence identity of 97.75% (Figure S1D). We therefore tested our PA-TM-RING E3 ligase nanobodies against a panel of commonly used human (HEK293T, CaCo-2, YT1, and UT/7) and mouse (BaF3, 3T3, and B16) cell lines by cell surface staining as indicated (Figures 1E and S2A, B) and summarized in a normalized heatmap (Figure 1F). Indeed, all nanobodies tested were cross- reactive against human and mouse cell lines, which poses a significant advantage for the design and application of the REULR molecule for in vitro and in vivo studies. We next evaluated the nanobodies on primary cells using PBMCs (Primary Peripheral Blood Mononuclear Cells) to identify cell surface binding to immune cells: T cells (CD4+; CD8+), monocytes, B cells, and NK cells (Figures 1G and S2C), summarized in a normalized heatmap (Figure 1H). Similar to human and mouse cell lines, we observed some ligases like RNF167 being highly expressed in many cell types, while most other ligases tested show a more nuanced, cell type-specific expression pattern (Figure 1F,H). Receptor Elimination by E3 Ubiquitin Ligase Recruit- ment (REULR). To evaluate potential targets for our REULR approach, we performed a membrane proteome wide analysis of reported ubiquitin sites.27 On average, 45% of cell surface receptors were reported to have at least one or more ubiquitin site (Figure 2A), which represents an untapped potential for cell surface receptors to be targeted using a REULR strategy. Members of the receptor tyrosine kinase (RTK) family including cytokine receptors EpoR (via JAK2 V617F) and members of the epidermal growth factor receptor (ErbB; HER) the most common oncogenic drivers of family represent malignant carcinomas.28−30 However, despite their immense 1083 https://doi.org/10.1021/acssynbio.2c00587 ACS Synth. Biol. 2023, 12, 1081−1093 ACS Synthetic Biology pubs.acs.org/synthbio Research Article Figure 2. EGFR REULR. (A) Analysis of MS (Mass Spectrometry)-validated proteome wide ubiquitin sites matched to the human membrane proteome, subclassified by the number of transmembrane domains. (B) Schematic representation of EGFR degradation using a EGFR−REULR molecule. (C−F) HEK293T cells were transiently transfected with FLAG-tagged full-length EGFR cDNA (human) under the control of a constitutively active CMV (cytomegalovirus) promoter. 24 h post-transfection, cells were incubated with EGFR−REULR molecules (50 nM) as indicated using RNF128 (E1; E2)-, RNF130 (D1)-, or RNF167 (A5)-targeting nanobodies in combination with EGFR-binding moieties (7D12; 9G8) in varying orientations as indicated in comparison to monomeric nanobodies or PBS. After 24 h, cells were subjected to FACS analysis using a FLAG antibody (Alexa Fluor 647 conjugate) to monitor EGFR levels on the cell surface. Representative FACS histograms are visualized below the quantified data. Data are mean ± s.d. (n = 3 replicates). (G) Cell surface staining of A431 human squamous carcinoma cell using a panel of five PA-TM-RING E3 ligase-binding nanobodies (nanobody:SA647 tetramers) and analysis by flow cytometry, full titration (1:1 dilutions; 100 nM tetramer), and Biotin:SA647 (Biotin) served as a negative control. (H,I) Cell proliferation assay (CellTiter-Glo 2.0; Promega). A431 cells were seeded at 2.5k cells/ well. After 24 h, cells were treated with PBS, cetuximab, or EGFR−REULR molecules as indicated using RNF167 (A5)-targeting nanobodies in combination with EGFR-binding moieties (7D12; 9G8) (50 nM). Cells were incubated for 72 h, washed, and subjected to CellTiter-Glo (2.0) assays to measure cell proliferation, according to the manufacturer’s specifications (Promega). Data are presented as a percentage of untreated cells (n = 4). clinical relevance, conventional drug discovery approaches have shown limited efficacy and problems of resistance.31,32 These limitations are mainly due to the nature of primary and emerging secondary escape mutations in the receptor (EGFR T790M) and acquired resistance, as well as pathway mutations, e.g., JAK2 V617F (EPOR/TPOR) that lead to constitutive over activation and dysregulation with detrimental outcomes for patients.33−36 We first designed different combinations of heterobifunc- tional REULRs to EGFR using two VHH (7D12; 9G8) that were previously described to inhibit ligand binding to EGFR: nanobody 7D12 sterically blocks ligand binding to EGFR, 1084 https://doi.org/10.1021/acssynbio.2c00587 ACS Synth. Biol. 2023, 12, 1081−1093 ACS Synthetic Biology pubs.acs.org/synthbio Research Article Figure 3. EPOR REULR. (A) Schematic representation of EPOR−REULR-mediated EpoR degradation. (B−D) HEK293T cells were transiently transfected with FLAG-tagged full-length EpoR cDNA (human) under the control of a constitutively active CMV (cytomegalovirus) promoter. 24 h post-transfection, cells were incubated with EPOR−REULR molecules (50 nM) as indicated using RNF128 (E1)-, RNF43 (A7)-, or ZNRF3 (A10)- targeting nanobodies fused to a scFv (single chain fragment variable) reformatted EpoR diabody (DA10). Monomeric binding moieties or PBS were used as a negative control. After 24 h, cells were subjected to FACS analysis using a FLAG antibody (Alexa Fluor 647 conjugate) to monitor EPOR levels on the cell surface. Representative FACS histograms are visualized below the quantified data. Data are mean ± s.d. (n = three replicates). similar to cetuximab, and 9G8 acts by inhibiting high-affinity ligand binding and dimerization.37−39 To assess whether EGFR can be degraded by this set of REULR molecules, we overexpressed FLAG-tagged full-length EGFR in HEK293T cells that endogenously express PA-TM-RING E3 ligases at varying levels (Figure 1E) and treated cells with intact REULR molecules, monomeric versions, or PBS as negative controls (Figures 2C−F and S3A−D). We observed EGFR degradation after treatment with EGFR−REULR molecules at varying efficiencies, depending on the choice of the E3-targeting ligase, EGFR VHH, and orientation (Figure 2C−F). Collectively, EGFR−REULR designs using the N-terminal 9G8 nanobody in combination with C-terminal RNF128, RNF130, or RNF167- targeting nanobodies performed better and resulted in more effective EGFR degradation compared to other designs. Targeting the EGFR pathway with tyrosine kinase inhibitors (TKI; e.g., afatinib, erlotinib, gefitinib, and osimertinib) or monoclonal antibodies (e.g., cetuximab, panitumumab, nimo- tuzumab, and necitumumab) is a well-characterized strategy for treating cancers including lung adenocarcinomas (NSCLC) and squamous cell carcinoma (SCC), which are one of the most types of skin cancer.33,40 A431 cells, a human prevalent the squamous carcinoma cell EGFR gene and, as a consequence, express a high level of EGFR. In addition, A431 cells have been widely used for studying skin cancer as well as for pharmaceutical and biomedical purposes in vitro and in xenograft models.41−43 We therefore explored how EGFR REULR molecules can exert line, have amplifications of antiproliferative activity in A431 cells in comparison to cetuximab, a first-generation anti-EGFR chimeric antibody used for the treatment of metastatic colorectal cancer and head and neck cancers.44,45 We first evaluated the positive binding of our PA-TM-RING E3 ligase nanobodies in A431 cells by cell surface tetramer staining (Figure 2G) and selected RNF167 (A5)-based REULR molecules that were tested in HEK293 cells (Figure 2F) for the proliferation assays in A431 cells. Indeed, in agreement with our degradation assays, treatment of A431 cells with intact REULR molecules using nanobodies against RNF167 (A5) in combination with two EGFR nanobodies (7D12 and 9G8) resulted in a significant reduction of cell proliferation but with different efficacies compared to cetuximab (Figure 2H,I), while monomeric nanobodies or PBS served as negative controls and showed no significant change. To show modularity with other binding moieties, we reformatted an EpoR-targeting diabody (DA10)46 into an scFv (single-chain variable fragment) and fused it to RNF128-, RNF43-, and ZNRF3-targeting nanobodies (Figure 3A). intact EPOR−REULR molecules could efficiently Indeed, degrade EpoR while showing no activity when cells were treated with the monomeric version of the individual targeting arms or PBS (Figures 3B−D and S4A). Of note, the degradation efficiency did not directly correlate with the expression levels of PA-TM-RING E3 ligases observed in HEK293T cells. While RNF128 and RNF43 appear to be expressed at much lower levels than ZNRF3 (∼25×; Figure 1E), degradation using EPOR-RNF128 or EPOR-RNF43 REULR (Figure 3B,C) still 1085 https://doi.org/10.1021/acssynbio.2c00587 ACS Synth. Biol. 2023, 12, 1081−1093 ACS Synthetic Biology pubs.acs.org/synthbio Research Article Figure 4. Immune checkpoint REULR. (A) Schematic representation of PD1-REULR-mediated PD-1 degradation. (B−D) HEK293T cells were transiently transfected with FLAG-tagged full-length PD-1 cDNA (human) under the control of a constitutively active CMV (cytomegalovirus) promoter. 24 h post-transfection, cells were incubated with PD1-REULR molecules (50 nM), as indicated using RNF128 (E1; E2)-, RNF130 (A1; D1)-, or RNF167 (A5)-targeting nanobodies fused to a PD-1 binding nanobody (PD1). Monomeric binding moieties or PBS were used as negative controls. After 24 h, cells were subjected to FACS analysis using a FLAG antibody (Alexa Fluor 647 conjugate) to monitor PD-1 levels on the cell surface. Representative FACS histograms are visualized below the quantified data. Data are mean ± s.d. (n = three replicates). resulted in comparable levels of EpoR loss in comparison to a ZNRF3 based EPOR−REULR molecule (Figure 3D). Immunotherapies based on checkpoint biology have emerged as a major pillar in fighting cancer. Immune-checkpoint inhibitors (ICIs) such as antibodies targeting CTLA-4 (ipilimumab), PDL1 (atezolizumab and durvalumab), or PD1 (pembrolizumab and nivolumab) have become some of the most widely used anticancer therapies.47,48 However, immune- related adverse events (irAEs), such as autoimmune symptoms and tumor hyperprogression, present a significant challenge in the clinic49 and a need for the continuous development of immune-oncology pipeline drugs. Targeted protein degradation could provide a major expansion in the repertoire of modulating immune checkpoint receptors by directly regulating their respective cell surface levels. We therefore next generated REULR molecules targeting the immune checkpoint receptor PD-1 (programmed cell death protein 1) by fusing an anti- human PD-1 nanobody50 to several nanobodies targeting RNF128, RNF130, and RNF167 (Figure 4A−D). Similar to EGFR− and EPOR−REULRs, treatment of HEK293T cells overexpressing FLAG-tagged full-length PD-1 with a variety of PD1-REULR molecules using RNF128-, RNF130-, or RNF167- targeting nanobodies resulted in a robust and near-complete loss the PD-1 receptor from the cell surface compared to of treatment with the respective monomeric VHHs or PBS (Figures 4B−D and S5A). While RNF130-based REULR molecules worked most effectively in degrading EGFR, PD1- REULR molecules using RNF128 and ENF167 targeting nanobodies collectively resulted in the substantial elimination of PD1. Expansion of the REULR Platform to Modulate E3 Ligases Itself: Fratricide REULRs. Emerging evidence highlights the pivotal role of RING-type E3 ligases and their substrates in a wide range of human diseases, and mutation of RING-type E3s or modulation of their activity is frequently associated with pathogenesis including viral infections, neuro- degenerative disorders, autoimmune diseases, and cancer.16−18 Indeed, RNF43 mutations have been associated with aggressive tumor biology such as colorectal and endometrial cancer.51−53 To evaluate the impact of other PA-TM-RING E3 ligases in cancer, we analyzed TCGA (The Cancer Genome Atlas) tissue mRNA expression data obtained for RNF128, RNF130, RNF167, RNF43, and ZNRF3 from 17 cancer types, representing 21 cancer subtypes. The data show elevated expression of E3 ligases in various cancer types. Notably, while RNF167 is highly expressed in almost all cancers, other ligases like RNF128 (thyroid, liver, urothelial, and colorectal), RNF130 (gliomas), or RNF43 (colorectal cancers) show a more selective tissue-associated expression pattern (Figure 5A) in cancer cells. 1086 https://doi.org/10.1021/acssynbio.2c00587 ACS Synth. Biol. 2023, 12, 1081−1093 ACS Synthetic Biology pubs.acs.org/synthbio Research Article Figure 5. Homo- and heterobifunctional fratricide REULR. (A) TCGA cancer tissue RNA-seq data for RNF128, RNF130, RNF167, RNF43, and ZNRF3 was obtained from 17 cancer types, representing 21 cancer subtypes and were processed as median FPKM (number fragments per kilobase of exon per million reads) and visualized as a hierarchical clustering heatmap. (B) Schematic representation of homo- or heterobispecific fratricide REULR. (C) HEK293T cells were transiently transfected with HA-tagged full-length PA-TM-RING E3 ligase cDNA (human) under the control of a constitutively active CMV (cytomegalovirus) promoter, as indicated. 24 h post-transfection, cells were incubated with fratricide REULR molecules (50 nM), targeting RNF128, RNF130, RNF167, RNF43, or ZNRF3. (D) HEK293T cells were co-transfected with HA-tagged full-length RNF43 and MYC-tagged full-length ZNRF3 cDNA (human) and treated with a heterobispecific RNF43-ZNRF3 fratricide REULR 24 h post-transfection (monomeric binding moieties were used as negative controls). After 24 h, cells were subjected to FACS analysis using a HA antibody (Alexa Fluor 647 conjugate) or an MYC antibody (Alexa Fluor 488 conjugate) to monitor PA-TM-RING E3 ligase levels on the cell surface. Data are mean ± s.d. (n = three replicates). (E) Schematic representation of heterobispecific arrayed multimeric fratricide REULR. (F) HEK293T cells were co-transfected with HA-tagged full-length RNF128, MYC-tagged full-length ZNRF3, and FLAG-tagged full-length RNF43 cDNA (human). 24 h post-transfection, cells were treated with homo- or heterobispecific Fratricide REULR molecules as indicated or a RNF43-RNF128-ZNRF3 multimeric fratricide REULR (PBS and monomeric binding moieties were used as negative controls). After 36 h, cells were subjected to FACS analysis using an HA antibody (Alexa Fluor 647 conjugate), a MYC antibody (Alexa Fluor 488 conjugate), and a FLAG antibody (Brilliant Violet 421) to monitor RNF128 (left panel), RNF43 (middle), and ZNRF3 (right panel) E3 ligase levels on the cell surface. Data are mean ± s.d. (n = three replicates). Despite their critical role in regulating protein homeostasis and pathological signaling, our understanding of transmembrane E3 ligase-mediated signaling still remains largely fragmented and can mainly be attributed to the limited availability of tools to study TM E3 ligases. Interestingly, the activity of E3 ligases is tightly regulated by post-translational modifications, and a typical feature of most ligases is the ability to catalyze their own ubiquitination.54,55 Based on this paradigm, we postulated that self-regulation by auto-ubiquitination could be used to regulate E3 ligase-dependent signaling. We therefore proceeded in developing REULR molecules that target the PA-TM-RING E3 ligase itself, either by homodime- 1087 https://doi.org/10.1021/acssynbio.2c00587 ACS Synth. Biol. 2023, 12, 1081−1093 ACS Synthetic Biology pubs.acs.org/synthbio Research Article Figure 6. Fratricide REULR and WNT signaling potentiation. (A) Schematic representation of a RNF43- or ZNRF3-based fratricide REULR in the context of a canonical WNT signaling pathway. (B) HEK293STF cells were seeded at 10k/well and subsequently treated with RNF43 or ZNRF3 fratricide REULR for 24 h. FZD cell surface levels were measured by incubating cells with a biotinylated pan-FZD Darpin (DRPB_Fz7/8), recovered by SA647 and analyzed by flow cytometry. Data are mean ± s.d. (n = three replicates). (C) HEK293STF cells were seeded at 5k/well and treated with RNF43, ZNRF3, or RNF43-ZNRF3 fratricide REULR molecules after 24 h in the presence of 10% conditioned WNT3a media (monomeric binding moieties or PBS were used as negative controls). After 36 h, the activation of the β-catenin-dependent STF reporter by fratricide REULRs was measured. Data are mean ± s.d. (n = three replicates). rization or heterodimerization between two transmembrane E3 ligases. Using this approach would allow strategic modulation of transmembrane E3 ligases and consequently protein homeo- stasis of their natural targets, a process we termed fratricide REULRs (Figure 5B). Indeed, treatment of cells with RNF128, RNF130, RNF167, RNF43, and ZNRF3 fratricide REULR molecules resulted in an effective loss of cell surface ligase levels in HEK293T cells (Figures 5C−D and S6A). Furthermore, to demonstrate the modular nature and flexibility of the nanobody-based REULR design, we engineered a RNF43-ZNRF3 heterobifunctional REULR (Figures 5B and S6B) that would allow the elimination of two ligases using one fratricide REULR molecule. The treatment of HEK293T cells overexpressing MYC-tagged RNF43 and HA-tagged ZNRF3 with a RNF43-ZNRF3 heterobifunctional fratricide REULR (Figure 5D; right bar graph) resulted in a significant reduction of RNF43 and a near-compete loss of ZNRF3 levels comparable to RNF43 and ZNRF3 fratricide REULRs (Figure 5D; left and middle bar graph). To demonstrate the ease of formatting using PA-TM-RING E3-binding VHHs, we extended the previous design into a linear, hetero-trimeric array of VHHs targeting RNF128, RNF43, and ZNRF3 with one fratricide REULR molecule (Figures 5E and S6C). We co-expressed HA-tagged RNF128, MYC-tagged ZNRF3, and FLAG-tagged RNF43 in HEK293T cells and treated cells with RNF128 (only targets RNF128) or ZNRF3-REULR (only targets ZNRF3), heterobifunctional RNF43-ZNRF3 REULR (targets RNF43 and ZNRF3), or a hetero-trimeric RNF43-RNF128-ZNRF3 fratricide REULR that targets all three PA-TM-RING E3 ligases for degradation (Figure 5F). A RNF43-RNF128-ZNRF3-targeting VHH array was able to efficiently eliminate all three E3 ligases from the cell surface and further shows the robustness and advantage of a “mix and match” VHH-based targeting approach. The WNT signaling pathway is instrumental for embryonic development, stem cell differentiation, and regeneration of injured tissues, and modulation of WNT signaling presents an untapped potential in regenerative medicine.56−59 RNF43 and ZNRF3 are two pivotal PA-TM-RING E3 ligases known to negatively regulate the WNT signaling pathway by targeting Wnt receptors FZD and promoting receptor degradation via the UPS (Figure 6A).60,61 With well-established fratricide REULRs in hand, we explored whether RNF43 and ZNRF3-based fratricide REULR molecules have the potential to modulate FZD receptor cell surface levels and potentiate downstream WNT signaling events. We first treated HEK293T cells with RNF43 or ZNRF3 fratricide REULR molecules and monitored FZD cell surface levels using a previously developed pan-FZD (DRPB_Fz7/8) as a staining reagent due to its high affinity and broad binding spectrum for FZD receptors: FZD1, 2, 5, 7, and 8.62 We indeed observed a significant increase in the accumulation of FZD levels after RNF43 or ZNRF3 fratricide REULR treatment compared to PBS or monomeric RNF43 or ZNRF nanobodies (Figure 6B). To examine whether these results can be translated into a functional assay and elicit fratricide REULR-specific activation of canonical WNT signal- ing, we performed a series of reporter assays using HEK 293STF (SuperTopFlash) cells. In agreement with the increased FZD levels upon treatment with RNF43 or ZNRF3 fratricide REULRs, we similarly observed a robust induction of WNT signaling and increased signaling activity using a heterospecific RNF43-ZNRF3 fratricide REULR, compared to treatment with WNT, PBS or monomeric PA-TM-RING nanobodies alone (Figure 6C). ■ DISCUSSION In summary, we implemented a modular, “mix and match” human and mouse cross-reactive nanobody-based targeted protein degradation platform termed REULR by retasking five PA-TM-RING E3 ligases (RNF128, RNF130, RNF167, RNF43, and ZNRF3) to modulate cell surface receptors by induced proximity, allowing selective, tissue-specific application. REULR-based bispecific molecules can be broadly applied to modulate cell surface levels of a variety of therapeutically 1088 https://doi.org/10.1021/acssynbio.2c00587 ACS Synth. Biol. 2023, 12, 1081−1093 ACS Synthetic Biology pubs.acs.org/synthbio Research Article relevant transmembrane receptors using different binding moieties (Figures 2−4). Furthermore, we present a strategic approach to tune transmembrane E3 ligases itself by using homo-,heterobispecific and arrayed fratricide REULRs and consequently modulate the signaling events of natural target receptors (Figures 5 and 6). formatting, and modularity that might While similar approaches (AbTACs and PROTABs) have been reported to degrade PD-L1 or IGF1R, they are mainly limited by using WNT-responsive E3 ligase RNF43 and ZNRF313,14 and rely on human IgG antibody scaffolds to generate heterobifunctional-targeting molecules. Antibody- derived biologics are generally more constrained by their inherent structural properties including their large size (150 kDa), the applicability for tumor therapy. By contrast, REULR molecules take advantage of the superior pharmacokinetic properties of nanobodies (VHH), allowing for a versatile and modular design with ease of formatting into homo- or heterobifunctional dimers or arrayed multimers to target one or multiple cell surface proteins (Figures 2−5). While nanobodies have been proven to possess a low immunogenicity risk profile, it is important to consider that they can show limitations in their therapeutic lifetime due to rapid renal clearance without further engineering, e.g., half-life extension through the use of serum albumin nanobody fusions (NbHSA) which could be achieved due to the modular nature of the REULR molecules.63 limit Interestingly, we observed that the affinity and expression levels of the PA-TM-RING E3 ligase-targeting nanobody did not directly correlate with the levels of cell surface clearance. This suggests that the PA-TM-RING E3 ligases operate with a wide spectrum of cytosolic catalytic RING E3 activity rather than by abundance alone. Catalytic activity and processivity of E3 ligases are regulated by many contributing factors to safeguard substrate selection including cell-type expression levels and tightly regulated post-translational modifications including phosphorylation and sumoylation among others, as well as binding ubiquitin proteasome machinery adapters.15 Furthermore, E3 ligase protein homeostasis is regulated by ubiquitylation itself and subsequent internalization; thus, there is likely a pool of E3 ligase whose activity levels are unknown.55,64 Moreover, REULR processivity may be further influenced by the orientation and geometry of the ternary receptor−REULR−ligase complex. In addition to the therapeutic potential of REULR molecules, the monomeric-binding modules (nanobodies) themselves present invaluable tools to validate natural targets and to gain a deeper understanding into the fundamental biological function of transmembrane E3 ligases and their cellular pathways in drug discovery and the context of cancer biology. Collectively, we believe that our “mix and match” nanobody- based REULR protein degradation strategy holds tremendous promise for a large variety of targets and serves as a powerful research tool with the potential to develop novel therapeutic applications that can be easily customized by virtue of its modularity, human and mouse cross-reactivity, and tissue specificity. ■ MATERIALS AND METHODS Curation of the Human Ubiquitin Cell Surface Receptor Proteome. A raw list of reported ubiquitin sites was obtained from PhosphoSitePlus (PSP; https://www. phosphosite.org) and matched to a curated list of the human membrane proteome65 to generate a master list of cell surface receptors with reported ubiquitination sites. Database Integration. Pairwise protein sequence align- ments were performed using the Smith−Waterman algorithm to calculate alignments between human and mouse amino acid sequences obtained from UniProt (https://www.uniprot.org/). Phylogenetic homology analysis was performed to generate phylogenetic trees from multiple sequence alignments (MSA) of amino acid sequences of ECD sequences of transmembrane cell surface receptors (https://www.uniprot.org/). Briefly, MSA was performed using ClustalOmega (https://www.ebi.ac.uk/Tools/ msa/clustalo/), and alignment results were submitted to calculate phylogenetic tree parameters (https://www.ebi.ac. uk/Tools/phylogeny/simple phylogeny/), which were visual- ized by Interactive Tree of Life (iTOL; https://itol.embl.de/).66 Tissue expression datasets and normal tissue and TCGA datasets were downloaded from The Human Protein Atlas (https://www.proteinatlas.org; v21.1). TCGA cancer tissue RNA-seq data were obtained from 17 cancer types, representing 21 cancer subtypes, and were processed as median FPKM (number of fragments per kilobase of exon per million reads) and visualized as a hierarchical clustering heatmap using JMP Pro (v16). Unsupervised hierarchical clustering of normalized mRNA gene expression by tissue was performed with the Ward linkage, and correlation distances were plotted as heatmaps using JMP Pro (v16). Cell Lines. Suspension cells were grown in plain-bottom, vented flasks (Thermo); adherent cells were grown in T25 or T75 flasks (ThermoFisher). Cells were maintained at 37 °C and 5% CO2. HEK293T (CRL-3216; ATCC), and LentiX cells were maintained in DMEM supplemented with 10% FBS, 1% GlutaMax, and 1% penicillin/streptomycin. Caco-2, YT1, A431, UT/7, BaF3, 3T3, and B16 cells were obtained from ATCC and grown and maintained according to ATCC specifications. HEK293F (R79007; ThermoFisher) were grown in FreeStyle media (12338018; ThermoFisher). Expi293F (A14528; ThermoFisher) cells were grown in Expi293 Expression Medium (ThermoFisher). Cell lines tested negative for mycoplasma (MycoAlert Mycoplasma Detection kit, Lonza). Facs Staining. Cells were stained with the indicated antibodies at a 1:100 dilution or tetramer at the indicated concentration for 30 min on ice in MACS staining buffer (Miltenyi). After incubation with fluorescent antibodies or tetramers, cells were washed with MACS buffer and analyzed via flow cytometry on a Cytoflex (Beckman Coulter) instrument. Surface expression was quantified by FACS using the CytoFLEX, equipped with a high-throughput sampler. Live cells were identified after gating on the basis of forward scatter (FSC) and side scatter (SSC) and propidium iodide (PI)- negative staining. Data were analyzed using FlowJo 10.8.1 (BD). All assays were performed using independent biological replicates. The number of replicates (n) is indicated in the figure legends. The mean fluorescence intensity (MFI) was determined in FlowJo 10.8.1. ■ ANTIBODIES Primary antibodies used in this study include the anti- DYKDDDDK tag (CST, D6W5B, no. 15009), anti-HA Tag (CST, 6 × 102, no. 3444), and anti-MYC (CST, 9B11, no. 2279). Antibodies were used at 1:100 dilution in MACS staining buffer (Miltenyi). 1089 https://doi.org/10.1021/acssynbio.2c00587 ACS Synth. Biol. 2023, 12, 1081−1093 ACS Synthetic Biology pubs.acs.org/synthbio Research Article Production of Purified Proteins. Proteins were produced in Expi293F cells using transfection conditions following the manufacturer’s protocol. After harvesting of cell media, 1 M Tris, pH 8.0 was added to a final concentration of 20 mM. Ni-NTA agarose (Qiagen) was added to ∼5% media volume. 1× sterile PBS, pH 7.2 (Gibco) was added to ∼3× medium volume. The mixture was stirred overnight at 4 °C. Ni-NTA agarose beads were collected in a Buchner funnel and washed with ∼300 mL protein wash buffer (20 mM HEPES, pH 7.2, 150 mM NaCl, 20 mM imidazole). Beads were transferred to an Econo-Pak chromatography column (Bio-Rad), and the protein was eluted in 15 mL of elution buffer (20 mM HEPES, pH 7.2, 150 mM NaCl, 200 mM imidazole). The DNA encoding for pan-FZD (DRPB_Fz7/8) was cloned into pET-28 with a C-terminal AVI- 6xHIS tag and transformed into Rosetta DE3-competent cells. The cells were grown at 37 °C in 2YT media supplemented with kanamycin (40 μg/mL) until the culture reached log-phase growth. IPTG was added to the culture to induce protein expression at a final concentration of 1 mM. The culture was shaken at 37 °C for 3 h, and protein was harvested from the cells by sonication. Pan-Fzd protein was purified using Ni-NTA agarose (Qiagen), followed by biotinylation and size-exclusion chromatography with a Superdex S75 column (GE Healthcare). In general, proteins were concentrated using Amicon Ultracel filters (Millipore), and absorbance at 280 nm was measured using a Nanodrop 2000 spectrophotometer (Thermo Fisher Scientific) to determine protein concentrations. REULR Design and Expression. All proteins were cloned in-frame in a modified pD649 plasmid with a N-terminal hemagglutinin signal peptide (HAsp) and a C-terminal AVI- 6xHIS tag for protein expression and purification from Expi293F cells. REULR molecules were connected either by a LEVLFQGP (3C) or a GSLEVLFQGPGS (GS flanked 3C) linker. All VHH and scFv sequences were cloned using gBlocks (IDT), and final sequence integrity was confirmed by DNA sequencing. All amino acid sequences can be found in Table S2A−C. Biotinylation and FPLC Purification. Where indicated, proteins were biotinylated as described previously.67 Briefly, up to 10 mg of protein was incubated at 4°C overnight in 2× Biomix A (0.5 M bicine buffer), 2× Biomix B (100 mM ATP, 100 mM MgOAc, 500 μM D-biotin), and Bio200 (500 μM D-biotin) to a final concentration of 20 μM, and 60−80 units BirA ligase in a final volume of 1 mL. All proteins were further purified by size- exclusion chromatography using an S75 or a S200 Increase column (GE Healthcare), depending on protein size, on an Ä KTA Pure FPLC (GE Healthcare). Nanobody Selection. Nanobody selection was performed as previously described with minor alterations. Briefly, the synthetic yeast library was expanded overnight in -Trp media with glucose at 30 °C and induced at 10× the theoretical diversity by suspension in -Trp media with galactose, grown at 20 °C for 24 h. Surface display was assessed by flow cytometry after staining with an anti-HA antibody. Rounds 1 and 2 were first negatively selected on magnetic streptavidin beads and then positively selected on magnetic streptavidin beads loaded with biotinylated target protein. Subsequent rounds were carried out with target proteins tetramerized by streptavidin and bound to anti-fluorophore magnetic beads, followed by decreasing monomer protein concentrations and by flow cytometry. Single clones from the final round were sorted into 96 well plates, induced for 24 h at 20 °C, and grown in deep well blocks. The top 20 clones were sequenced, and unique clones were expressed in Expi293F cells and assayed for binding to the corresponding target protein by SPR. SPR Experiments. SPR experiments were performed using a Biacore T100 instrument (GE Healthcare). FPLC-purified biotinylated proteins (ligands) in HBS-P + buffer (GE Healthcare) were captured on a streptavidin (SA) series S sensor chip (GE Healthcare). Chip capture was performed in HBS-P + buffer (GE Healthcare) to aim for ∼100−200 ligand response units (RU). Flow cell 1 was left empty as a reference flow-cell for on-line subtraction of bulk solution refractive index and for evaluation of non-specific binding of the analyte to the chip surface using Biacore T100 Control Software (version 3.2) (GE Healthcare). FPLC-purified non-biotinylated protein was used as the analyte. Analytes were run in HBS-P + buffer using twofold increasing protein concentrations to generate a series of sensograms. Binding parameters were either determined based on a 1:1 Langmuir model or at equilibrium using the accompanying Biacore T100 evaluation software. A table of all SPR conditions for each ligand−analyte pair tested including the concentration range of twofold analyte dilutions, injection rate, injection and dissociation times, and regeneration conditions can be found in Table S1. FPLC traces for purified proteins used for SPR can be found in Figure S1. Cell−Surface Binding Assay with Streptavidin-Tetra- merized Proteins. To examine PPIs at the cell surface, we performed cell−surface protein binding assays using human or mouse cell lines, or primary cells (PBMCs) with streptavidin- tetramerized, biotinylated proteins. To generate streptavidin- tetramerized proteins to test for binding to cells, FPLC-purified biotinylated proteins (see above) were incubated with streptavidin tetramers conjugated to Alexa647 Fluor (SA-647) (Thermo Fisher) at a 4:1 molar ratio on ice for at least 15 min. Approximately 150,000 cells were incubated with protein:SA- 647 complexes in a final volume of 100 μL in 96-well round- bottom plates (Corning) for 30-60 min at 4 °C protected from light. Following incubation, cells were washed two times with 200 μL cold MACS buffer and resuspended in 200 μL cold MACS buffer with 1:3000 PI (Thermo Fisher Scientific). Immunofluorescence staining was analyzed using a Cytoflex (Beckman Coulter), and data were collected for 20,000 cells. Data were analyzed using FlowJo v10.4.2 software. All data report MFI. Concentration-dependent binding of protein:SA- 647 to full-length receptor-expressing, but not mock control cells, was deemed indicative of cell−surface binding. STF Luciferase Reporter Assays. HEK293STF cells were seeded for each condition in 96-well plates and stimulated with fratricide REULRs, WNT (WNT3a conditioned media; ATCC), control proteins, or PBS for 36 h. After washing cells with 1× PBS, cells in each well were lysed in 30 μL 1× passive lysis buffer (Promega). 10 μL per well of lysate was assayed using the Dual Luciferase Assay kit (Promega). Cell Proliferation Assay. A431 cells were seeded at 2.5k cell/well. After 24 h, cells were treated with PBS, Cetuximab, or different EGFR−REULR molecules (50 nM). Cells were incubated for 72 h, washed, and subjected to CellTiter-Glo (2.0) assays to measure cell proliferation, according to the manufacturer’s specifications (Promega). Data are presented as a percentage of untreated cells (n = 4). Statistics. All figures are representative of at least n = 3 (in vitro) experiments, unless otherwise noted. Statistical signifi- cance was assayed by grouped, one-way ANOVA using GraphPad Prism 9.4.1. In all figures, *P < 0.05; **P < 0.01; 1090 https://doi.org/10.1021/acssynbio.2c00587 ACS Synth. Biol. 2023, 12, 1081−1093 ACS Synthetic Biology pubs.acs.org/synthbio Research Article ***P < 0.001; ****P < 0.0001; NS: not significant. Data are represented as mean ± s.d., unless otherwise stated. ■ ASSOCIATED CONTENT Data Availability Statement All data generated or analyzed during this study are included in the manuscript and supporting files. *sı Supporting Information The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acssynbio.2c00587. Size-exclusion chromatography, SPR sensograms, SPR conditions, pairwise sequence alignments, cell surface staining gating strategies, REULR design, REULR architecture, and REULR-related amino acid sequences (PDF) ■ AUTHOR INFORMATION Corresponding Author K. Christopher Garcia − Department of Molecular and Cellular Physiology, Stanford University School of Medicine, Stanford, California 94305, United States; Department of Structural Biology and Howard Hughes Medical Institute, Stanford University School of Medicine, Stanford, California 94305, United States; Email: kcgarcia@stanford.edu orcid.org/0000-0001-9273-0278; Authors Dirk H. Siepe − Department of Molecular and Cellular Physiology, Stanford University School of Medicine, Stanford, California 94305, United States; 0009-8023 orcid.org/0000-0002- Lora K. Picton − Department of Molecular and Cellular Physiology, Stanford University School of Medicine, Stanford, California 94305, United States Complete contact information is available at: https://pubs.acs.org/10.1021/acssynbio.2c00587 Author Contributions Conceptualization, D.H.S. and K.C.G.; Methodology, D.H.S. and K.C.G.; Nanobody library screening and protein expression for in vitro studies, D.H.S. and L.K.P.; Analysis, D.H.S.; Investigation, D.H.S. and K.C.G.; Writing�Original Draft, D.H.S; Writing�Review and Editing, D.H.S. and K.C.G.; Visualization, D.H.S.; Supervision, K.C.G.; Funding Acquis- ition, K.C.G. Notes The authors declare the following competing financial interest(s): K.C.G. and D.H.S. are co-inventors on a patent (PCT/US2022/030132) based upon the technology described in this manuscript. K.C.G. is the founder of InduPro Labs, Inc. ■ ACKNOWLEDGMENTS The authors are funded by the Howard Hughes Medical Institute, NCI 2R01CA177684 and Emerson Collective (K.C.G.), and NIGMS 1RO1GM150125 (K.C.G.). ■ REFERENCES (1) Mullard, A. First Targeted Protein Degrader Hits the Clinic. Nat. Rev. Drug Discovery 2019, DOI: 10.1038/D41573-019-00043-6. (2) Sakamoto, K. M.; Kim, K. B.; Kumagai, A.; Mercurio, F.; Crews, C. M.; Deshaies, R. J. Protacs: Chimeric Molecules That Target Proteins to the Skp1-Cullin-F Box Complex for Ubiquitination and Degradation. Proc. Natl. Acad. Sci. U.S.A. 2001, 98, 8554−8559. (3) Deshaies, R. J. Prime Time for PROTACs. Nat. Chem. Biol. 2015, 11, 634−635. (4) Békés, M.; Langley, D. R.; Crews, C. M. PROTAC Targeted Protein Degraders: The Past Is Prologue. Nat. Rev. Drug Discovery 2022, 21, 181−200. (5) Dong, G.; Ding, Y.; He, S.; Sheng, C. Molecular Glues for Targeted Protein Degradation: From Serendipity to Rational Discovery. J. Med. Chem. 2021, 64, 10606−10620. (6) Nabet, B.; Roberts, J. M.; Buckley, D. L.; Paulk, J.; Dastjerdi, S.; Yang, A.; Leggett, A. L.; Erb, M. A.; Lawlor, M. A.; Souza, A.; Scott, T. G.; Vittori, S.; Perry, J. A.; Qi, J.; Winter, G. E.; Wong, K. K.; Gray, N. S.; Bradner, J. E. The DTAG System for Immediate and Target-Specific Protein Degradation. Nat. Chem. Biol. 2018, 14, 431−441. (7) Clift, D.; McEwan, W. A.; Labzin, L. I.; Konieczny, V.; Mogessie, B.; James, L. C.; Schuh, M. A Method for the Acute and Rapid Degradation of Endogenous Proteins. Cell 2017, 171, 1692−1706. (8) Li, K.; Crews, C. M. PROTACs: Past, Present and Future. Chem. Soc. Rev. 2022, 51, 5214−5236. (9) Uhlén, M.; Fagerberg, L.; Hallström, B. M.; Lindskog, C.; Oksvold, P.; Mardinoglu, A.; Sivertsson, A.; Kampf, C.; Sjöstedt, E.; Asplund, A.; Olsson, I.; Edlund, K.; Lundberg, E.; Navani, S.; Szigyarto, C. A. K.; Odeberg, J.; Djureinovic, D.; Takanen, J. O.; Hober, S.; Alm, T.; Edqvist, P. H.; Berling, H.; Tegel, H.; Mulder, J.; Rockberg, J.; Nilsson, P.; Schwenk, J. M.; Hamsten, M.; von Feilitzen, K.; Forsberg, M.; Persson, L.; Johansson, F.; Zwahlen, M.; von Heijne, G.; Nielsen, J.; Pontén, F. Tissue-Based Map of the Human Proteome. Science 2015, 347, 1260419. (10) Thul, P. J.; Åkesson, L.; Wiking, M.; Mahdessian, D.; Geladaki, A.; Ait Blal, H.; Alm, T.; Asplund, A.; Björk, L.; Breckels, L. M.; Bäckström, A.; Danielsson, F.; Fagerberg, L.; Fall, J.; Gatto, L.; Gnann, C.; Hober, S.; Hjelmare, M.; Johansson, F.; Lee, S.; Lindskog, C.; Mulder, J.; Mulvey, C. M.; Nilsson, P.; Oksvold, P.; Rockberg, J.; Schutten, R.; Schwenk, J. M.; Sivertsson, A.; Sjöstedt, E.; Skogs, M.; Stadler, C.; Sullivan, D. P.; Tegel, H.; Winsnes, C.; Zhang, C.; Zwahlen, M.; Mardinoglu, A.; Pontén, F.; von Feilitzen, K.; Lilley, K. S.; Uhlén, M.; Lundberg, E. A Subcellular Map of the Human Proteome. Science 2017, 356, 6340. (11) Banik, S. M.; Pedram, K.; Wisnovsky, S.; Ahn, G.; Riley, N. M.; Bertozzi, C. R. Lysosome-Targeting Chimaeras for Degradation of Extracellular Proteins. Nature 2020, 584, 291−297. (12) Pance, K.; Gramespacher, J. A.; Byrnes, J. R.; Salangsang, F.; Serrano, J.-A. C.; Cotton, A. D.; Steri, V.; Wells, J. A. Modular Cytokine Receptor-Targeting Chimeras for Targeted Degradation of Cell Surface and Extracellular Proteins. Nat. Biotechnol. 2022, 41, 273−281. (13) Marei, H.; Tsai, W.-T. K.; Kee, Y.-S.; Ruiz, K.; He, J.; Cox, C.; Sun, T.; Penikalapati, S.; Dwivedi, P.; Choi, M.; Kan, D.; Saenz-Lopez, P.; Dorighi, K.; Zhang, P.; Kschonsak, Y. T.; Kljavin, N.; Amin, D.; Kim, I.; Mancini, A. G.; Nguyen, T.; Wang, C.; Janezic, E.; Doan, A.; Mai, E.; Xi, H.; Gu, C.; Heinlein, M.; Biehs, B.; Wu, J.; Lehoux, I.; Harris, S.; Comps-Agrar, L.; Seshasayee, D.; de Sauvage, F. J.; Grimmer, M.; Li, J.; Agard, N. J.; de Sousa e Melo, F. Antibody Targeting of E3 Ubiquitin Ligases for Receptor Degradation. Nature 2022, 610, 182−189. (14) Cotton, A. D.; Nguyen, D. P.; Gramespacher, J. A.; Seiple, I. B.; Wells, J. A. Development of Antibody-Based PROTACs for the Degradation of the Cell-Surface Immune Checkpoint Protein PD-L1. J. Am. Chem. Soc. 2021, 143, 593−598. (15) Deshaies, R. J.; Joazeiro, C. A. P. RING Domain E3 Ubiquitin Ligases. Annu. Rev. Biochem. 2009, 78, 399−434. (16) Cai, C.; Tang, Y.-D.; Zhai, J.; Zheng, C. The RING Finger Protein Family in Health and Disease. Signal Transduction Targeted Ther. 2022, 7, 300. (17) Chen, X.; Jiang, L.; Zhou, Z.; Yang, B.; He, Q.; Zhu, C.; Cao, J. The Role of Membrane-Associated E3 Ubiquitin Ligases in Cancer. Front. Pharmacol. 2022, 13, 928794. (18) Duan, S.; Pagano, M. Ubiquitin Ligases in Cancer: Functions and Clinical Potentials. Cell Chem. Biol. 2021, 28, 918−933. 1091 https://doi.org/10.1021/acssynbio.2c00587 ACS Synth. Biol. 2023, 12, 1081−1093 ACS Synthetic Biology pubs.acs.org/synthbio Research Article (19) Komander, D.; Rape, M. The Ubiquitin Code. Annu. Rev. Biochem. 2012, 81, 203−229. (20) Metzger, M. B.; Pruneda, J. N.; Klevit, R. E.; Weissman, A. M. RING-Type E3 Ligases: Master Manipulators of E2 Ubiquitin- Conjugating enzymes and Ubiquitination. Biochim. Biophys. Acta 2014, 1843, 47. (21) Nakamura, N. The Role of the Transmembrane RING Finger Proteins in Cellular and Organelle Function. Membranes 2011, 1, 354− 393. (22) Robert, C. A Decade of Immune-Checkpoint Inhibitors in Cancer Therapy. Nat. Commun. 2020, 11, 3801. (23) Yang, E. Y.; Shah, K. Nanobodies: Next Generation of Cancer Diagnostics and Therapeutics. Front. Oncol. 2020, 10, 1182. (24) Bannas, P.; Hambach, J.; Koch-Nolte, F. Nanobodies and Nanobody-Based Human Heavy Chain Antibodies as Antitumor Therapeutics. Front. Immunol. 2017, 8, 1603. (25) Jin, S.; Sun, Y.; Liang, X.; Gu, X.; Ning, J.; Xu, Y.; Chen, S.; Pan, L. Emerging New Therapeutic Antibody Derivatives for Cancer Treat- ment. Signal Transduction Targeted Ther. 2022, 7, 39. (26) McMahon, C.; Baier, A. S.; Pascolutti, R.; Wegrecki, M.; Zheng, S.; Ong, J. X.; Erlandson, S. C.; Hilger, D.; Rasmussen, S. G. F.; Ring, A. M.; Manglik, A.; Kruse, A. C. Yeast Surface Display Platform for Rapid Discovery of Conformationally Selective Nanobodies. Nat. Struct. Mol. Biol. 2018, 25, 289−296. (27) Hornbeck, P. v.; Zhang, B.; Murray, B.; Kornhauser, J. M.; Latham, V.; Skrzypek, E. PhosphoSitePlus, 2014: Mutations, PTMs and Recalibrations. Nucleic Acids Res. 2015, 43, D512−D520. (28) Lemmon, M. A.; Schlessinger, J. Cell Signaling by Receptor Tyrosine Kinases. Cell 2010, 141, 1117−1134. (29) Schmidt-Arras, D.; Böhmer, F. D. Mislocalisation of Activated Receptor Tyrosine Kinases − Challenges for Cancer Therapy. Trends Mol. Med. 2020, 26, 833−847. (30) Gschwind, A.; Fischer, O. M.; Ullrich, A. The Discovery of Receptor Tyrosine Kinases: Targets for Cancer Therapy. Nat. Rev. Cancer 2004, 4, 361−370. (31) Chong, C. R.; Jänne, P. A. The Quest to Overcome Resistance to EGFR-Targeted Therapies in Cancer. Nat. Med. 2013, 19, 1389−1400. (32) Yang, Y.; Li, S.; Wang, Y.; Zhao, Y.; Li, Q. Protein Tyrosine Kinase Inhibitor Resistance in Malignant Tumors: Molecular Mechanisms and Future Perspective. Signal Transduction Targeted Ther. 2022, 7, 329. (33) Shi, K.; Wang, G.; Pei, J.; Zhang, J.; Wang, J.; Ouyang, L.; Wang, Y.; Li, W. Emerging Strategies to Overcome Resistance to Third- Generation EGFR Inhibitors. J. Hematol. Oncol. 2022, 15, 94. (34) Kumar, R.; George, B.; Campbell, M. R.; Verma, N.; Paul, A. M.; Melo-Alvim, C.; Ribeiro, L.; Pillai, M. R.; da Costa, L. M.; Moasser, M. M. HER Family in Cancer Progression: From Discovery to 2020 and Beyond. Adv. Cancer Res. 2020, 147, 109−160. (35) Chan, D. L. H.; Segelov, E.; Wong, R. S. H.; Smith, A.; Herbertson, R. A.; Li, B. T.; Tebbutt, N.; Price, T.; Pavlakis, N. Epidermal Growth Factor Receptor (EGFR) Inhibitors for Metastatic Colorectal Cancer. Cochrane Database Syst. Rev. 2017, 2017, CD007047. (36) Glassman, C. R.; Tsutsumi, N.; Saxton, R. A.; Lupardus, P. J.; Jude, K. M.; Garcia, K. Structure of a Janus Kinase Cytokine Receptor Complex Reveals the Basis for Dimeric Activation. Science 2022, 376, 163. (37) Roovers, R. C.; Vosjan, M. J. W. D.; Laeremans, T.; el Khoulati, R.; de Bruin, R. C. G.; Ferguson, K. M.; Verkleij, A. J.; van Dongen, G. A. M. S.; van Bergen en Henegouwen, P. M. P. A Biparatopic Anti- EGFR Nanobody Efficiently Inhibits Solid Tumour Growth. Int. J. Cancer 2011, 129, 2013−2024. (38) Schmitz, K. R.; Bagchi, A.; Roovers, R. C.; van Bergen En Henegouwen, P. M. P.; Ferguson, K. M. Structural Evaluation of EGFR Inhibition Mechanisms for Nanobodies/VHH Domains. Structure 2013, 21, 1214. (39) Wolf, A.; Schmitz, C.; Böttger, A. Changing Story of the Receptor for Phosphatidylserine-Dependent Clearance of Apoptotic Cells. EMBO Rep. 2007, 8, 465−469. (40) Cai, W. Q.; Zeng, L. S.; Wang, L. F.; Wang, Y. Y.; Cheng, J. T.; Zhang, Y.; Han, Z. W.; Zhou, Y.; Huang, S. L.; Wang, X. W.; Peng, X. C.; Xiang, Y.; Ma, Z.; Cui, S. Z.; Xin, H. W. The Latest Battles Between EGFR Monoclonal Antibodies and Resistant Tumor Cells. Front. Oncol. 2020, 10, 1249. (41) Merlino, G. T.; Xu, Y. H.; Ishii, S.; Clark, A. J. L.; Semba, K.; Toyoshima, K.; Yamamoto, T.; Pastan, I. Amplification and Enhanced Expression of the Epidermal Growth Factor Receptor Gene in A431 Human Carcinoma Cells. Science 1984, 224, 417−419. (42) Masui, H.; Castro, L.; Mendelsohn, J. Consumption of EGF by A431 Cells: Evidence for Receptor Recycling. J. Cell Biol. 1993, 120, 85. (43) Lim, Y. J.; Jeon, S. R.; Koh, J. M.; Wu, H. G. Tumor Growth Suppression and Enhanced Radioresponse by an Exogenous Epidermal Growth Factor in Mouse Xenograft Models with A431 Cells. Cancer: Res. Treat. 2015, 47, 921−930. (44) Galizia, G.; Lieto, E.; De Vita, F.; Orditura, M.; Castellano, P.; Troiani, T.; Imperatore, V.; Ciardiello, F. Cetuximab, a Chimeric Human Mouse Anti-Epidermal Growth Factor Receptor Monoclonal Antibody, in the Treatment of Human Colorectal Cancer. Oncogene 2007, 26, 3654−3660. (45) Yewale, C.; Baradia, D.; Vhora, I.; Patil, S.; Misra, A. Epidermal Growth Factor Receptor Targeting in Cancer: A Review of Trends and Strategies. Biomaterials 2013, 34, 8690−8707. (46) Moraga, I.; Wernig, G.; Wilmes, S.; Gryshkova, V.; Richter, C. P.; Hong, W. J.; Sinha, R.; Guo, F.; Fabionar, H.; Wehrman, T. S.; Krutzik, P.; Demharter, S.; Plo, I.; Weissman, I. L.; Minary, P.; Majeti, R.; Constantinescu, S. N.; Piehler, J.; Garcia, K. C. Tuning Cytokine Receptor Signaling by Re-Orienting Dimer Geometry with Surrogate Ligands. Cell 2015, 160, 1196−1208. (47) Robert, C. A Decade of Immune-Checkpoint Inhibitors in Cancer Therapy. Nat. Commun. 2020, 11, 3801. (48) Johnson, D. B.; Nebhan, C. A.; Moslehi, J. J.; Balko, J. M. Immune-Checkpoint Inhibitors: Long-Term Implications of Toxicity. Nat. Rev. Clin. Oncol. 2022, 19, 254−267. (49) Conroy, M.; Naidoo, J. Immune-Related Adverse Events and the Balancing Act of Immunotherapy. Nat. Commun. 2022, 13, 392. (50) Fernandes, R. A.; Su, L.; Nishiga, Y.; Ren, J.; Bhuiyan, A. M.; Cheng, N.; Kuo, C. J.; Picton, L. K.; Ohtsuki, S.; Majzner, R. G.; Rietberg, S. P.; Mackall, C. L.; Yin, Q.; Ali, L. R.; Yang, X.; Savvides, C. S.; Sage, J.; Dougan, M.; Garcia, K. C. Immune Receptor Inhibition through Enforced Phosphatase Recruitment. Nature 2020, 586, 779. (51) Herwaarden, Y. J.; Koggel, L. M.; Simmer, F.; Vink-Börger, E. M.; Dura, P.; Meijer, G. A.; Nagengast, F. M.; Hoogerbrugge, N.; Bisseling, T. M.; Nagtegaal, I. D. RNF43 Mutation Analysis in Serrated Polyposis, Sporadic Serrated Polyps and Lynch Syndrome Polyps. Histopathology 2021, 78, 749. (52) Matsumoto, A.; Shimada, Y.; Nakano, M.; Oyanagi, H.; Tajima, Y.; Nakano, M.; Kameyama, H.; Hirose, Y.; Ichikawa, H.; Nagahashi, M.; Nogami, H.; Maruyama, S.; Takii, Y.; Ling, Y.; Okuda, S.; Wakai, T. RNF43 Mutation Is Associated with Aggressive Tumor Biology along with BRAF V600E Mutation in Right-Sided Colorectal Cancer. Oncol. Rep. 2020, 43, 1853. (53) Fang, L.; Ford-Roshon, D.; Russo, M.; O’Brien, C.; Xiong, X.; Gurjao, C.; Grandclaudon, M.; Raghavan, S.; Corsello, S. M.; Carr, S. A.; Udeshi, N. D.; Berstler, J.; Sicinska, E.; Ng, K.; Giannakis, M. RNF43 G659fs Is an Oncogenic Colorectal Cancer Mutation and Sensitizes Tumor Cells to PI3K/MTOR Inhibition. Nat. Commun. 2022, 13, 3181. (54) Lorick, K. L.; Jensen, J. P.; Fang, S.; Ong, A. M.; Hatakeyama, S.; Weissman, A. M. RING Fingers Mediate Ubiquitin-Conjugating Enzyme (E2)-Dependent Ubiquitination. Proc. Natl. Acad. Sci. U.S.A. 1999, 96, 11364−11369. (55) de Bie, P.; Ciechanover, A. Ubiquitination of E3 Ligases: Self- Regulation of the Ubiquitin System via Proteolytic and Non-Proteolytic Mechanisms. Cell Death Differ. 2011, 18, 1393. (56) Barker, N.; Clevers, H. Mining the Wnt Pathway for Cancer Therapeutics. Nat. Rev. Drug Discovery 2006, 5, 997−1014. 1092 https://doi.org/10.1021/acssynbio.2c00587 ACS Synth. Biol. 2023, 12, 1081−1093 ACS Synthetic Biology pubs.acs.org/synthbio Research Article (57) Logan, C. Y.; Nusse, R. The Wnt Signaling Pathway in Development and Disease. Annu. Rev. Cell Dev. Biol. 2004, 20, 781− 810. (58) MacDonald, B. T.; Tamai, K.; He, X. Wnt/Beta-Catenin Signaling: Components, Mechanisms, and Diseases. Dev. Cell 2009, 17, 9−26. (59) Niehrs, C. The Complex World of WNT Receptor Signalling. Nat. Rev. Mol. Cell Biol. 2012, 13, 767−779. (60) Clevers, H.; Nusse, R. Wnt/β-Catenin Signaling and Disease. Cell 2012, 149, 1192−1205. (61) Zebisch, M.; Jones, E. Y. ZNRF3/RNF43–A Direct Linkage of Extracellular Recognition and E3 Ligase Activity to Modulate Cell Surface Signalling. Prog. Biophys. Mol. Biol. 2015, 118, 112−118. (62) Dang, L. T.; Miao, Y.; Ha, A.; Yuki, K.; Park, K.; Janda, C. Y.; Jude, K. M.; Mohan, K.; Ha, N.; Vallon, M.; Yuan, J.; Vilches-Moure, J. G.; Kuo, C. J.; Garcia, K. C.; Baker, D. Receptor Subtype Discrimination Using Extensive Shape Complementary Designed Interfaces. Nat. Struct. Mol. Biol. 2019, 26, 407−414. (63) Shen, Z.; Xiang, Y.; Vergara, S.; Chen, A.; Xiao, Z.; Santiago, U.; Jin, C.; Sang, Z.; Luo, J.; Chen, K.; Schneidman-Duhovny, D.; Camacho, C.; Calero, G.; Hu, B.; Shi, Y. A Resource of High-Quality and Versatile Nanobodies for Drug Delivery. iScience 2021, 24, 103014. (64) Weissman, A. M.; Shabek, N.; Ciechanover, A. The Predator Becomes the Prey: Regulating the Ubiquitin System by Ubiquitylation and Degradation. Nat. Rev. Mol. Cell Biol. 2011, 12, 605−620. (65) Siepe, D. H.; Henneberg, L. T.; Wilson, S. C.; Hess, G. T.; Bassik, M. C.; Zinn, K.; Garcia, K. C. Identification of Orphan Ligand-Receptor Relationships Using a Cell-Based CRISPRa Enrichment Screening Platform. Elife 2022, 11, No. e81398. (66) Letunic, I.; Bork, P. Interactive Tree Of Life (ITOL) v5: An Online Tool for Phylogenetic Tree Display and Annotation. Nucleic Acids Res. 2021, 49, W293−W296. (67) Ö zkan, E.; Carrillo, R. A.; Eastman, C. L.; Weiszmann, R.; Waghray, D.; Johnson, K. G.; Zinn, K.; Celniker, S. E.; Garcia, K. C. An Extracellular Interactome of Immunoglobulin and LRR Proteins Reveals Receptor-Ligand Networks. Cell 2013, 154, 228. 1093 https://doi.org/10.1021/acssynbio.2c00587 ACS Synth. Biol. 2023, 12, 1081−1093
10.1021_acscentsci.2c01325
http://pubs.acs.org/journal/acscii Research Article A Turn-On Fluorescent Amino Acid Sensor Reveals Chloroquine’s Effect on Cellular Amino Acids via Inhibiting Cathepsin L Michael R. Smith,# Le Zhang,# Yizhen Jin, Min Yang, Anusha Bade, Kevin D. Gillis, Sadhan Jana, Ramesh Naidu Bypaneni, Timothy E. Glass,* and Hening Lin* Cite This: ACS Cent. Sci. 2023, 9, 980−991 Read Online ACCESS Metrics & More Article Recommendations *sı Supporting Information ABSTRACT: Maintaining homeostasis of metabolites such as amino acids is critical for cell survival. Dysfunction of nutrient balance can result in human diseases such as diabetes. Much remains to be discovered about how cells transport, store, and utilize amino acids due to limited research tools. Here we developed a novel, pan- amino acid fluorescent turn-on sensor, NS560. It detects 18 of the 20 proteogenic amino acids and can be visualized in mammalian cells. Using NS560, we identified amino acids pools in lysosomes, late endosomes, and surrounding the rough endoplasmic reticulum. Interestingly, we observed amino acid accumulation in large cellular foci after treatment with chloroquine, but not with other autophagy inhibitors. Using a biotinylated photo-cross-linking chloroquine analog and chemical proteomics, we identified Cathepsin L (CTSL) as the chloroquine target leading to the amino acid accumulation phenotype. This study establishes NS560 as a useful tool to study amino acid regulation, identifies new mechanisms of action of chloroquine, and demonstrates the importance of CTSL regulation of lysosomes. ■ INTRODUCTION Maintenance of amino acid homeostasis is important for cellular function.1 In order to accomplish this task, cells have evolved different signaling and regulatory mechanisms, such as to sense and the mTOR and GCN2 signaling pathways, regulate the uptake and utilization of amino acids.1−6 It is established that under stress conditions such as amino acid deprivation, mTORC1 activity is downregulated, translation is inhibited, and cells adapt via a host of stress-response mechanisms.2,3,7,8 Dysregulation of amino acid homeostasis can lead to human diseases. For example, branched chain amino acids are critical switches in Maple syrup syndrome, mental retardation, and premature death if catabolism is dysregulated.9 As an amino acid sensing hub, mTORC1 has long been implicated in cancer and neurodegenerative diseases.4,10 These diseases highlight the importance of understanding the uptake, storage, utilization, and regulation of amino acids. Lysosomes are critical amino acid sensing and storage depots and are responsible for degradation of autophagosomes, mitochondria, and other damaged organelles.7 Autophagy inhibitors, such as Bafilomycin A1, chloroquine, and ammonium chloride, are known to disrupt these functions.11 Bafilomycin A1 (BafA1) inhibits the vacuolar (H+)-ATPase critical for maintaining low pH in lysosomes and late endosomes.12 Chloroquine (CQ), an antimalarial drug, and ammonium chloride (NH4Cl) are proposed to cross lysosomal membrane where they become protonated and accumu- late.12,13 This temporarily neutralizes lysosomes (H+ “sponge” effect) and renders them dysfunctional. Despite widespread use as a tool to inhibit autophagy, as well as in clinical trials for cancer and SARS-CoV-2 therapies,14−16 the mechanism(s) of action of chloroquine is still not well understood. To study the uptake, storage, and utilization of amino acids, it would be useful to be able to visualize and monitor amino acid pools in live cells at high resolution. There are a number of genetically encoded biosensors available that utilize fluorescence resonance energy transfer (FRET) or fluorescent protein permutations to monitor individual metabolites.17 One system, named OLIVe (optical biosensor leucine, isoleucine, and valine), uses YFP/CFP FRET technology to sense branched chain amino acids. This method is effective for branched chain amino acids but does not detect amino acids in individual organelles due to low fluorescent turn-on.18 Many studies have focused on glutamate sensors to study synaptic transmission in neurons but also lack resolution to determine organelle localization.19 Histidine or cysteine sensing follows a similar pattern.20,21 These methods are limited by localization of an overexpressed sensing domain and low relative turn-on for Received: November 5, 2022 Published: April 24, 2023 © 2023 The Authors. Published by American Chemical Society 980 https://doi.org/10.1021/acscentsci.2c01325 ACS Cent. Sci. 2023, 9, 980−991 ACS Central Science http://pubs.acs.org/journal/acscii Research Article Figure 1. Design and synthesis of NS560 as a pan-specific amino acid probe. (A) Incubation of NS560 with free amino acids results in reversible covalent attachment of the N and C termini to the aldehyde and boronic acid, respectively, producing a fluorescent adduct. (B) Synthesis of NS560. (C) In vitro incubation of NS560 with proteinogenic amino acids leads to metabolite binding and fluorescence increase. Fluorescence enhancement was determined by taking the ratio of saturated over control fluorescence signal (10 μM NS560, 25 mM HEPES, 50 mM Na2S2O3, 1% DMSO, pH 7.4, λex = 488 nm, λem = 560 nm). The fluorescence of Tyr was estimated due to low solubility. signal due to the ratiometric analysis. Enriching and purifying lysosomes for eventual metabolomics (Lyso-IP) is a quantitative method used to reveal the interplay between SLC38A9 transporter activity and mTOR signaling at lysosomes, but it is not a microscopy tool for live cells.7 We envision that a membrane-permeable small molecule that exhibits robust turn-on fluorescence upon binding amino acids would be simple to use and enable investigations on the regulation of amino acids in cells. A number of small-molecule fluorescent sensors for amino acids have been developed over the years, though few function well under physiological conditions except for cysteine.22,23 Here we report a novel fluorescent turn-on sensor for amino acids, NS560, based on our neurosensor class of fluorescent sensors.24 We demonstrate that NS560 can detect amino acids in live cells and confirm that lysosomes and late endosomes house free amino acid pools in cells. With this amino acid sensor, we were able to quickly test a number of small molecules for their ability to affect cellular amino acids utilization. Interestingly, chloroquine causes dramatic changes in cellular amino acids. Using a functionalized chloroquine analog for chemical proteomics, we discovered that chlor- oquine has affinity toward lysosomal proteins Cathepsin L (CTSL), NPC2, and PSAP. We found that chloroquine can bind and inhibit CTSL. The robust change in amino acid labeling happens mainly due to the inhibition of CTSL by chloroquine as knockdown or inhibition of CTSL led to similar NS560 amino acids labeling. Our work establishes NS560 as a 981 https://doi.org/10.1021/acscentsci.2c01325 ACS Cent. Sci. 2023, 9, 980−991 ACS Central Science http://pubs.acs.org/journal/acscii Research Article Figure 2. NS560 labels amino acids in HeLa cells. (A) Incubation of NS560 in HeLa cells for 45 min resulted in strong fluorescence at 5 μM. Scale bars, 40 μm. (B) Amino acid deprived HeLa cells incubated with NS560 showed increased labeling after replenishment with 5× essential amino acid solution. Scale bars, 30 μm. Representative images of three replicates. (C−D) Confocal microscopy reveals that NS560 signal can be surrounded by lysosomal marker Lamp1-RFP or dsRed-Rab7 under basal conditions. Scale bar, 5 μm. Representative images of three biological replicates. useful tool for rapid imaging of free amino acids in live cells and identifies CTSL as a new target for chloroquine, which provide important insights into the various reported biological activities of chloroquine. ■ RESULTS Design, Synthesis, and Characterization of a Pan- Amino Acid Biosensor, NS560. NS560 is a quinolone fluorophore with an aldehyde and a boronic acid functional group that can react with amino acids. Phenylboronic pinacol esters are susceptible to hydrolysis at physiological pH.25 After reacting with amino acids, the aldehyde is converted to an iminium ion, which has a long wavelength absorption allowing selective excitation and fluorescence upon binding (Figure 1A). While isolated carboxylate groups typically do not form boronate esters favorably, proximity promotes ester formation for NS560. Formation of the macrocycle restricts the rotation of the fluorescence the molecule (Figure 1).26 NS560 was properties of synthesized in three steps in overall 16% yield (Figure 1B, see SI Synthetic Procedures). the arylboronic acid which affects NS560 was tested as a pan-amino acid probe in vitro by titration with amino acids. Measurements of both absorbance and emission were collected for all 20 proteogenic amino acids. For example, upon titration with glutamate, NS560 gave a 50 nm red shift of the maximum absorbance and a ∼800-fold fluorescence enhancement at 560 nm using excitation at 488 nm (Figure S2). Although the red shift in absorbance is similar to other sensors in this class, the very high fluorescence enhancement was unprecedented. The apparent association constant of NS560 for glutamate was 44 M−1, which is only moderate for these types of sensors. GABA, a gamma-amino acid, also binds with moderate affinity but with much lower fluorescence enhancement and altered maximum emission wavelength (Figure S4). Thus, binding of both functional for the extremely high groups to the sensor is essential fluorescence enhancements seen with α-amino acids. Mass spectroscopic analysis of a reaction sample supports the iminium ion adduct (Figure S1). The fluorescence enhance- ment, indicated by the ratio of fluorescence emission of NS560 with or without amino acid ligands, was greater than 35-fold for all but proline and tryptophan (Figures 1C and S2−42, Table S1). The lack of enhancement with proline is due to the terminal secondary amine, reducing the nucleophilic attack on the aldehyde recognition element. Tryptophan’s indole ring quenches the fluorescence after reacting with NS560 similar to catecholamine with previous probes.24 Thus, NS560 is a pan- specific amino acid probe in vitro capable of recognizing 18 of the 20 proteogenic amino acids. NS560 Labels Amino Acids in Mammalian Cells. We next tested whether NS560 can detect amino acids in live cells. HeLa cells were incubated with NS560 for 45 min, and after washing to remove probe from the cell media, the cells were examined using fluorescence microscopy. Fluorescence signal was detected at different NS560 concentrations, and 5 μM was chosen for future experiments (Figure 2A). To ensure that NS560 fluorescence was indeed from detecting amino acids, cells were briefly starved of amino acids with EBSS buffer for 60 min, incubated with NS560 for 30 min, and then replenished with a 5-fold excess of essential amino acids (compared to standard DMEM) for 30 min. NS560 signal was greatly enhanced in cells replenished with amino acids compared with control cells (Figure 2B), suggesting that NS560 could detect amino acids in cells. Lysosomes and Late Endosomes Are Primary Storage Compartments for Free Amino Acids. Cells in basal conditions displayed mostly diffusive amino acids with occasional small punctate structures. We hypothesized these 982 https://doi.org/10.1021/acscentsci.2c01325 ACS Cent. Sci. 2023, 9, 980−991 ACS Central Science http://pubs.acs.org/journal/acscii Research Article Figure 3. Chloroquine treatment alters amino acid pools in A549 cells. (A) Five-hour chloroquine treatment at 100 μM causes drastic buildup in lysosomal amino acid labeling. Scale bars, 20 or 4 μm (for zoom-in images). (B) A549 cells treated with BafA1 (25 nM), CQ (25 μM), and NH4Cl (25 mM) for 7 h. Representative images of three biological replicates. Scale bar, 5 μm. (C) Representative quantification of NS560 foci in cells treated with different small molecules. (D) Western blot analysis of cells treated in B to detect LC3 levels (autophagy). GAPDH was used as loading control. Representative images are from three independent biological replicates. represented organelle-specific free amino acid pools. Because lysosomes are known to store amino acids, we overexpressed fluorescent markers of the lysosome and late endosomes, Lamp1-RFP and mCherry-Rab7, respectively, and analyzed the localization of these markers with NS560 signal using confocal microscopy. Both Lamp1-RFP and mCherry-Rab7 partially colocalize or surround NS560 signal in larger puncta (Figure 2C−D). We also observed NS560 signal near rough-ER (marked with mCherry-Sec61) (Figure S45A). Rough endoplasmic reticulum (ER) is the site where protein translation occurs and may need to have amino acids readily available. Other organelle markers such as dsRed2-Rab5A (early endosome) or mCherry-Rab11A (recycling endosome) did not overlap with NS560 signal (Figure S45B−C). Given that lysosomes, late endosomes, and rough ER are predictable locations for free amino acids in cells, the data further support that NS560 can detect free amino acids in cells. Chloroquine Treatment Accumulates Amino Acids in Late Endosomes and Lysosomes. As NS560 allows facile detection of free amino acids in live cells, we wanted to use it to screen for small molecules that could affect the distribution Interestingly, we found that of cellular amino acids. chloroquine can dramatically alter the distribution of cellular that all amino acids. Incubation for either 7 or 24 h with 25 μM of chloroquine in A549 cells led to a buildup of puncta of amino acids signals that colocalized with both lysosomes and late endosomes (Figures 3A and S46). The effect was unique to lysosome inhibitors bafilomycin or chloroquine, as other ammonium chloride did not result in such a phenotype despite the fact three inhibitors led to LC3-II accumulation, a hallmark of autophagy blockade (Figure 3B−D). Chloroquine, originally used as an antimalarial drug, is now commonly used as an autophagy inhibitor in cells. It is generally believed that chloroquine inhibits autophagy by raising the pH of acidic organelles.27 However, our observation of amino acid buildup after treatment is not consistent with augmented lysosomal pH, especially given that bafilomycin and ammonium chloride do not cause amino acid buildup. In our hands and in published literature, LysoTracker Red signal increases after prolonged (24 h) chloroquine treatment, indicating more acidic lysosomal organelles.12 Our observation suggests chloroquine treatment accumulates amino acids in lysosomes and late endosomes via mechanisms other than affecting lysosomal pH. Chemical Proteomics Identifies Chloroquine Binding Proteins in Lysosomes. We hypothesized that chloroquine 983 https://doi.org/10.1021/acscentsci.2c01325 ACS Cent. Sci. 2023, 9, 980−991 ACS Central Science http://pubs.acs.org/journal/acscii Research Article Figure 4. Chemical proteomics identifies lysosomal CQ targets. (A) Schematic of the proteomics strategy. Samples were prepared in triplicate for analysis. (B) Label-free proteomics results. Red bars indicate cut-offs (enrichment >2, chase rate >40%). (C) Relevant hits from proteomics data. (D) Validation of three chloroquine binding proteins from the proteomics data, CTSL, PSAP, and NPC2. CQ-X cross-linked targets were enriched via antibody and confirmed by blotting for cross-linked CQ-X via streptavidin blot. Representative Western blot for three biologically independent experiments. binds to and inhibits certain proteins, leading to free amino acids accumulation in lysosomes. In order to test this and identify these target proteins, we synthesized a chloroquine derivative (CQ-X) with a diazirine group for UV-cross-linking and an alkyne handle for copper(I)-catalyzed azide−alkyne cycloaddition to “click” on biotin (Figure S47). Importantly, CQ-X maintains the same NS560 labeling phenotypes as the parent chloroquine compound (Figure S48A). Furthermore, treatment of cells with chloroquine or CQ-X results in LC3-II accumulation (Figure S48B). A549 cells were treated for 1 h with chloroquine (control, 50 μM), CQ-X (50 μM), or CQ-X in combination with a 5-fold excess of chloroquine (chase). Cross-linking was carried out with 365 nm light. Cells were lysed, and biotin was attached via click chemistry. Proteins labeled by CQ-X were then affinity enriched with streptavidin beads and identified by MS after on-bead trypsin digestion (Figure 4A). We used label-free quantification (LFQ) to find proteins that are more abundant (≥2-fold) in the CQ-X treated sample than in the chloroquine or CQ-X/chloroquine treated samples. 243 proteins were enriched >2-fold by CQ-X, but only 29 were both enriched and chased by excess chloroquine below a ratio of 0.6 (Figure 4B−C, Table S2). The proteomics results are reliable for several reasons. First, there is a published crystal structure of chloroquine bound to saposin B (PSAP is a precursor for saposin A through saposin D and is identified as a chloroquine target in our proteomics study).28 Second, palmitoyl-protein thioesterase 1 (PPT1) is reported to bind dimeric chloroquine derivatives29 and is identified as a chloroquine target in our proteomics study. Lastly, many of the identified proteins exist in the lysosomes, which is consistent with chloroquine being a lysosomotropic agent. We confirmed the proteomics results by cross-linking with CQ-X and pulling down several protein hits identified, NPC2, CTSL, and PSAP, and then immunoblotted for streptavidin signal. We confirmed that all three targets were bound by CQ- X, but only NPC2 and CTSL were chased by excess chloroquine (Figure 4D). The results confirmed that chloroquine could bind to NPC2 and CTSL. Knockdown or Inhibition of CQ-X Binding Proteins Alters Free Amino Acids. In order to connect the CQ-X proteomics results with chloroquine induced cellular amino acids distribution, we used shRNA or siRNA to knockdown the identified chloroquine target proteins, NPC2 and CTSL. Knockdown of NPC2 resulted in larger fluorescent amino acid sites (Figure 5A−B), but the NS560 signal was not drastically increased. In contrast, siRNA knockdown of CTSL resulted in a quantifiable increase in NS560 foci compared to control (Figure 5C−E). The siRNA knockdown of CTSL efficiently reduced CTSL as detected via Western blot. Because multiple cathepsins (A, L, and Z) were identified in the proteomics study, the chloroquine effect on cathepsin may not be specific 984 https://doi.org/10.1021/acscentsci.2c01325 ACS Cent. Sci. 2023, 9, 980−991 ACS Central Science http://pubs.acs.org/journal/acscii Research Article Figure 5. Chloroquine targets CTSL and NPC2 to regulate free amino acids. A549 cells were treated with shRNA for NPC2 for 48 h (A) or transfected with siRNA for CTSL for 48−72 h (C) then labeled with NS560 for 45 min. Representative Western blots from three biological replicates show efficient knockdown of all genes (B and D). (E) Representative quantification of NS560 foci per cell for images shown in C. (F) A549 cells treated with E64d (25 μM) or Pepstatin A (25 μM) for 24 h then labeled with NS560. Scale bar, 5 μm for all images. (G) Representative quantification of NS560 foci per cell for images shown in (F). to CTSL. We treated A549 cells with an irreversible pan- cysteine protease inhibitor, E64d, or an aspartyl protease inhibitor Pepstatin A. After NS560 labeling, E64d treated cells showed an obvious free amino acid buildup in NS560 foci, while Pepstatin A had no effect (Figure 5F−G). This suggests cysteine proteases (Cathepsins L, Z, or others) but not aspartyl proteases (Cathepsins D and E) regulate free amino acid pools. Overall, genetic and chemical manipulation of CTSL inhibition could lead to changes in amino acids distribution. Chloroquine Inhibits Cathepsin L In Vitro. Next, we focused on the specific effect of chloroquine on CTSL. First, we tested if chloroquine can inhibit CTSL in vitro using commercially available purified enzyme and well-established dipeptide substrate Z-Phe-Arg-AMC. Active CTSL will cleave AMC and increase the fluorescence. After 25 min chloroquine preincubation with CTSL enzyme, AMC cleavage was inhibited by chloroquine with an IC50 of 181 μM (Figure 6A). While the IC50 seems to be on the high range, we believe this is physiologically relevant as chloroquine is a lysosomo- tropic drug and is estimated to reach concentrations of >25 mM in the lysosomes.13 Seeking further validation of chloroquine binding CTSL, we overexpressed a CTSL-Flag-Myc plasmid in HEK-293T cells and purified the enzyme using Flag beads (Figure S55). We measured chloroquine binding to CTSL by measuring the changes in the intrinsic fluorescence of CTSL by different concentrations of chloroquine. Exciting CTSL at 265 nm results in an emission peak around 308 nm. The presence of chloroquine caused an increase in the fluorescence, indicating the binding of chloroquine to CTSL (Figure 6B). To further validate the direct binding of chloroquine to CTSL, we designed a chloroquine-TAMRA derivate (CQ- TAMRA) and performed a fluorescence polarization assay by incubating this new probe with varying concentrations of CTSL. A mild milipolarization (mP) shift of 20−25 was obtained at maximum CTSL concentrations indicating mild binding (Figure S56). The data again support that chloroquine binds to CTSL. However, due to the limited CTSL we could obtain, we could not saturate the binding to get a binding constant. Using computational modeling, chloroquine can be docked to CTSL near its catalytic triad (Figure S51). This suggests that CQ is a competitive inhibitor. Based on this assumption and the IC50 value we obtained as well as the reported Km value the Z-Phe-Arg-AMC substrate,30 we can calculate a of dissociation constant of ∼35 μM using the equation of Kd = IC50/(1 + [S]/Km). Overall, our study shows that CTSL is an important protein for regulating free amino acids in cells. Prolonged chloroquine the or E64d treatment lysosomal CTSL, but inhibits 985 https://doi.org/10.1021/acscentsci.2c01325 ACS Cent. Sci. 2023, 9, 980−991 ACS Central Science http://pubs.acs.org/journal/acscii Research Article Figure 6. Chloroquine inhibits CTSL activity and increases intrinsic fluorescence in vitro. (A) CTSL hydrolysis of dipeptide substrate Z-FR-AMC was monitored by measuring the fluorescence of the released AMC. Reaction was performed at pH 5.5 in 100 mM MES-NaOH, 150 mM NaCl, and 7.5 mM DTT for 30 min after 25 min preincubation with chloroquine (CQ). Representative plot and calculation of three biological replicates. (B) Intrinsic fluorescence of purified CTSL-Flag, 200 nM, incubated with and without chloroquine. Representative data from three independent experiments. Fluorescence values were corrected for chloroquine background at these wavelengths. (C) Schematic representation of chloroquine- induced lysosome amino acid accumulation as compared with Bafilomycin treatment. degradative capacity of the lysosome remains largely intact likely due to the presence of other proteases. However, the resulting degradation products (free amino acids) are not efficiently exported due to the inhibition of CTSL by chloroquine or E64d (Figure 6C). In contrast, bafilomycin, which increases the lysosomal pH and thus inhibits all the lysosomal proteases, leads to diminished protein degradation and thus did not lead to amino acids accumulation in the lysosome. ■ DISCUSSION Our study introduces a novel chemical probe, NS560, that can be utilized to uncover important cell biology related to amino acid storage, utilization, and regulation. The probe rapidly labels free amino acid pools in cells and thus can provide important information about the levels and localizations of total amino acids. Using NS560, we screened several small molecules to see whether any of them could alter cellular amino acid levels and/ or localization. This led to the finding that chloroquine treatment causes previously unknown accumulation of amino in cells. Further chemical acids in specific organelles proteomics and biochemical studies established that this effect is due to chloroquine binding to and inhibiting lysosomal proteins CTSL (and possibly other cathepsins cysteine proteases). The study of chloroquine’s effect on cellular amino acid distribution is a nice example demonstrating the utility of NS560. Amino acids are essential for cellular life and thus have to be carefully regulated. The ability to visualize amino acids would enable us to track changes in amino acid levels and localization in a variety of conditions, such as metabolic stress, signaling activation, or disease conditions, which could provide important to understand fundamental cellular processes. insights Since its discovery as a potent antimalarial agent, chloroquine’s mechanism of action has been intensely studied. Here we observe extended (7 h or longer) chloroquine treatment results in large amino acid foci (Figure 4A). Whether these pools of amino acids are a result of increased uptake, increased protein degradation, or decreased export requires further research. There is a dichotomy in the literature pertaining to pH and hydrolytic capacity of lysosomes after laureate Christian de Duve chloroquine treatment. Nobel wrote a commentary discussing lysosomotropism that high- 986 https://doi.org/10.1021/acscentsci.2c01325 ACS Cent. Sci. 2023, 9, 980−991 ACS Central Science http://pubs.acs.org/journal/acscii Research Article lights amine-containing, weak base compounds hyper-accumu- late in lysosomes due to pH partitioning.13,27,31,32 Upon protonation, agents such as chloroquine become trapped, resulting in commonly cited increased organelle pH that inactivates hydrolytic enzymes.32−34 In stark contrast, there are reports showing that lysosomes adapt to chloroquine treatment and increase lysosomal acidity but develop other dysfunc- tions.32 Another report emphasizes that chloroquine inhibits the autophagosome-lysosome fusion step of autophagy and severely alters Golgi and endolysosomal systems but does not increase pH.12 In our hands, prolonged CQ treatment results in enlarged lysosomes, recovered acidic pH, and accumulation of amino acids. is another We employed commonly used autophagy inhibitors Bafilomycin A1 and NH4Cl to compare to chloroquine. Bafilomycin A1 inhibits the lysosomal proton pump to raise lysosomotropic lysosomal pH, while NH4Cl compound. In our hands, chloroquine or NH4Cl treatment for 1 h completely eliminates Lysotracker signal, consistent with predictions from the initial lysosomotropism analysis.13 lysosomotropic However, 7 h after addition of either autophagy inhibitor, Lysotracker returns due to signal previously reported cellular adaptation.12,32 Amino acid accumulation begins during this adaptation time period, but the drastic phenotype is only present in chloroquine-treated cells but not in bafilomycin A1 or NH4Cl-treated cells. Thus, we believe the amino acid buildup after chloroquine treatment is not a result of raised pH or inactive lysosomes. Interestingly, several reports are consistent with our findings that chloroquine (but not BafA1 or NH4Cl) leads to amino acids accumulation in the lysosomes. One study shows that after 24 h of chloroquine (but not BafA1) treatment, mTORC1 relocalizes away from lysosomes.6 Another shows that chloroquine (but not BafA1 or NH4Cl) treatment for 24 h reduces mTORC1 activity (p-S6K).11 It is likely that the accumulation of amino acids in the lysosomes by chloroquine leads to the decrease of amino acids in the cytosol, which are required for mTORC1 activation. Our data also suggest CTSL plays an important role in amino acid regulation beyond degradation of proteins. Cathepsins have become increasingly studied in relation to lysosomal dysfunction in recent years. For example, Cathepsins B and L regulate NPC2 secretion in macrophages activated with LPS, which impacts cholesterol metabolism.35 Another study reports that inhibition or genetic deletion of Cathepsins B and L (but not D) results in lysosomes with accumulated cholesterol, LC3-II, and Lysotracker.36 How CTSL inhibition by chloroquine or E64d leads to amino acids accumulation requires further studies. One hypothesis is that CTSL regulates an amino acid transporter. Regardless of the exact mechanism, our findings highlight the importance of CTSL in lysosomal regulation. CTSL has become a drug candidate for the SARS-CoV-2 pandemic because it cleaves the spike protein critical for infection.37 The inhibition of CTSL by chloroquine may also explain the reported potential beneficial effects of chloroquine in treating SARS-CoV-2 infection. Given that CTSL is not the major protease that cleaves the spike protein, our data are also consistent with the fact that chloroquine is not highly effective in treating SARS-CoV-2 infection in humans.38,39 Overall, this work highlights the usefulness of NS560 as a novel tool to visualize amino acid in cells. By employing NS560, a snapshot of the amino acid state of cells can be visualized. The use of NS560 is amenable for high-throughput screening, which will help uncover new biology related to the essential cellular building blocks, amino acids, as we show here for chloroquine. ■ METHODS General Synthetic Procedures of NS560. Chemicals were obtained from Sigma-Aldrich, Acros, Fisher, TCI America, Alfa Aesar, or Combi-Blocks and were used without further purification. Flash chromatography was performed with 32−63 μm silica gel. NMR spectra were recorded on a Bruker DRX 500 and 600. IR spectra were recorded on a Nexus 670 FT-IR E.S.P. spectrometer. Detailed synthesis and structural characterization can be found in the Supporting Information. Spectroscopic Studies of NS560. One mM stock solution of NS560 in DMSO for UV/vis spectra and fluorescence spectra was prepared and diluted to 1 mL with buffer (1.0 × 10−5 M, 25 mM HEPES, 50 mM Na2S2O3, pH 7.4, 5.0, 1% DMSO). The analytes (20 proteogenic amino acids and GABA) were prepared by dissolving the analytes in buffer sensor solution (the concentration of NS560 was the same as described above) to make sure the concentration of NS560 keeps constant. Sodium thiosulfate was used to protect aromatic analytes from oxidation in solution. UV/vis spectra were recorded on an Agilent Cary 100 UV/vis spectropho- tometer at ambient temperature. Fluorescence spectra were recorded on a Shimadzu RF-6000 PC Spectro Fluoropho- tometer at ambient temperature. Common Reagents and Antibodies. The following reagents and antibodies were purchased from commercial sources: Antibodies against β-actin HRP (sc-4777), CTSL (sc- 32320), GAPDH (sc-47724 HRP), normal mouse IgG (sc- 2025), normal rabbit IgG (sc-2027) along with Protein A/G PLUS-Agarose beads (sc-2003) were purchased from Santa Cruz. Antibodies against LC3 (Cat. #12741) and Streptavidin- HRP (Cat. #3999) were purchased from Cell Signaling Technology. Antibodies against PSAP (A1819) and NPC2 (A5413) were purchased from Abclonal. Protease inhibitor cocktail was purchased from Sigma-Aldrich (Cat. P8340). Streptavidin beads, ECL Western blotting detection reagent, and Pierce Universal nuclease were purchased from Thermo- Scientific. ClarityMax Western blotting detection reagent was purchased from BioRad (Cat. 1705062). Polyethylenimine (PEI) was purchased from Polysciences (Cat. 4765). Inhibitors used were all purchased as follows: Bafilomycin from CST (Cat. 54645), chloroquine diphosphate from TCI (C2301), ammonium chloride from Fisher Scientific (A661), E64d from SeleckChem (S7393), and Pepstatin A from Sigma (P5318). BCA assay was used for protein concentrations. Expression plasmids for organelle markers were all purchased from Addgene: mCherry-Sec61 (#49155), dsRed-Rab5 (#13050), mCherry-Rab11 (#55124), dsRed-Rab7 (#12661), and Lamp1-RFP (#1817).40−43 Cell Culture. A549, HEK-293T, and Hela cells were purchased from ATCC. A549 cells were cultured in RPMI media from Thermo (1875135) with 10% fetal bovine serum from Thermo. HeLa cells were cultured in DMEM (Gibco Cat. 11965−092) with 10% fetal bovine serum. HEK-293T cells were cultured in DMEM with 10% calf serum. For transient knockdown experiments, shRNA and siRNA were purchased from Sigma. NPC2 sh.1 (TRC0000293234), NPC2 sh.2 (TRC0000293323), CTSL si.1 (SASI_Hs01_00079400), and CTSL si.2 (SASI_Hs02_00332791). shRNA lentiviral particles 987 https://doi.org/10.1021/acscentsci.2c01325 ACS Cent. Sci. 2023, 9, 980−991 ACS Central Science http://pubs.acs.org/journal/acscii Research Article were generated by cotransfecting shRNA with psPAX pack- aging plasmid and pMD2.G enveloping plasmid. Particles were collected, filtered with 0.25 μm sterile filter, and used for future knockdown in A549 cells. In-Cell Microscopy of NS560. A549 or HeLa cells were seeded on a 35 mm glass bottom poly-D-lysine coated MatTek imaging dishes (Cat. P35GC-1.5C) 1 day prior to experiment. NS560 was added to cell media at 5 μM for 45 min prior to imaging or fixing unless indicated otherwise either live cell (Figure 2A−B). At the same time as NS560 addition, Lysotracker Deep Red (ThermoFisher Cat. L12492), Magic Red L (Immunochemistry Cat. #941), and/or Hoechst 33342 nuclear stain (ThermoFisher Cat. H3570) were added to cell media as indicated. After 45 min, cells were washed twice in PBS, and images were captured by BioTek Cytation 5 microscope (Figure 2A−B). For all detailed puncta or localization experiments, cells were washed twice with PBS, fixed with 4% PFA in PBS, and mounted using Fluoromount-G (Southern Biotech Cat. # 0100-01). Cells were then imaged the same day as the probe addition using Zeiss LSM 710 confocal microscope. NS560 and LysoTracker fluorescence will decrease if imaged the following day. Detailed analysis for NS560 foci count was performed using a macro designed with ImageJ software. Amino Acid Deprivation and Addition Studies. A549 or HeLa cells were seeded the same as described above in standard media 1 day prior to experiment. The following day, cells were washed 3 times with EBSS salt solution (Thermo- Fisher Cat. #24010043) and then grown in EBSS media without amino acids but supplemented with 2 g/L glucose (Thermo Cat. #A24940-01), vitamins (100× stock from Thermo Cat. 11120052), and 10% dialyzed FBS for 30 min following protocols known to impact mTORC1.8 Next, NS560 (5 μM) and Hoechst stain were added and incubated for 30 more minutes. Cells were again washed with EBSS to remove NS560 from the media and then incubated with the same EBSS starvation media supplemented with or with 5X MEM essential amino acids (50× stock from ThermoFisher Cat. # integral NS560 11130-051). Cells were imaged for total fluorescence using Cytation 5 microscope and data analysis. Sample Preparation for CQ-X In-Cell Cross-Linking and Click Chemistry. A549 cells were treated with chloroquine or CQ-X as indicated for 45 min. Cells were subjected to 10 min of cross-linking at 365 nm using a Boekel Scientific UV cross-linker. Next, cells were collected with cold PBS and lysed in NP-40 lysis buffer with protease inhibitor cocktail. Biotin-azide (Apex Bio Cat. A8013) was covalently attached via copper(I) catalyzed click chemistry. The reaction was run for 1 h at room temperature. Proteins were extracted with chloroform−methanol extraction methods. The protein pellet was resolubilized in buffer containing 8 M urea, 2.5% SDS, and 0.3 M NaCl. After BCA quantification, equal amounts of lysate were diluted into 0.1% NP-40 IP wash buffer and incubated with either streptavidin beads for proteomics or antibodies to validate for chloroquine protein targets proteomic results. Streptavidin beads were washed and submitted for digestion and proteomics analysis. On-Bead Trypsin Digestion for Proteomics Samples. The PBS storage buffer was removed from the beads. To denature and reduce the proteins bound to the beads, 30 μL of 50 mM TEAB (pH 8.5), 6 M urea, 2 M thiourea, 10 mM DTT were added and then incubated for 1 h at 35 °C. This was then followed by alkylation with 50 mM iodoacetamide for 45 min in the dark and quenched with a final concentration of 50 mM dithiothreitol (DTT). Samples were diluted with 50 mM TEAB pH 8.5 to a final concentration of 1 M urea. Trypsin was then added to a final concentration of 10 ng/μL and incubated overnight (16 h) at 35 °C. The digested peptides were desalted with Oasis MCX cartridge (Waters) and then dried down to ∼100 μL using speed vacuum SC110 (Thermo Savant, Milford, MA). All samples were filtered with 0.22 μm cellulose acetate spin filters (Costar). Filtered peptides were then dried down to dryness in the speed vacuum. Protein Identification by Nano LC/MS/MS Analysis. The tryptic digests were reconstituted in 2% acetonitrile (ACN) containing 0.5% formic acid (FA), and enolase (yeast) tryptic digest was added to the final concentration of 100 fmol/μL as internal standard for nanoLC-ESI-MS/MS analysis. The analysis was carried out using an Orbitrap Eclipse Tribrid (Thermo-Fisher Scientific, San Jose, CA) mass spectrometer equipped with a nanospray Flex Ion Source and coupled with a Dionex UltiMate 3000 RSLCnano system (Thermo, Sunny- vale, CA). The peptide samples (10 μL) were injected onto a PepMap C-18 RP nano trapping column (5 μm, 100 μm i.d × 20 mm) at 20 μL/min flow rate for rapid sample loading and then separated on a PepMap C-18 RP nano column (2 μm, 75 μm × 25 cm) at 35 °C. The tryptic peptides were eluted in a 90 min gradient of 5% to 35% ACN in 0.1% formic acid at 300 nL/min, followed by 8 min ramping to 90% ACN-0.1% FA and a 7 min hold at 90% ACN-0.1% FA. The column was re- equilibrated with 0.1% FA for 25 min prior to the next run. The Orbitrap Eclipse was operated in positive ion mode with spray voltage set at 1.6 kV and source temperature at 300 °C. External calibration for FT, IT, and quadrupole mass analyzers was performed. In data-dependent acquisition (DDA) analysis, the instrument was operated using FT mass analyzer in MS scan to select precursor ions followed by 3 s “Top-Speed” data- dependent CID ion trap MS/MS scans at 1.6 m/z quadrupole isolation for precursor peptides with multiple charged ions above a threshold ion count of 10,000 and normalized collision energy of 30%. MS survey scans were at a resolving power of 120,000 (fwhm at m/z 200) for the mass range of m/z 375− 1575. Dynamic exclusion parameters were set at 50 s of exclusion duration with ±10 ppm exclusion mass width. All data were acquired under Xcalibur 4.4 operation software (Thermo-Fisher Scientific). Data Analysis. The DDA raw files for CID MS/MS were subjected to database searches using Proteome Discoverer (PD) 2.5 software (Thermo Fisher Scientific, Bremen, Germany) with the Sequest HT algorithm. Processing workflow for precursor-based quantification. The PD 2.5 processing workflow containing an additional node of Minora Feature Detector for precursor ion-based quantification was used for protein identification and protein relatively quantitation analysis between samples. The database search was conducted against a Homo sapiens NCBI database that has 81,786 sequences. Two-missed trypsin cleavage sites were allowed. The peptide precursor tolerance was set to 10 ppm, and fragment to 0.6 Da. Variable modification of methionine oxidation, deamidation of aspara- gines/glutamine, acetylation, M-loss, and M-loss+acetylation on protein N-terminus and fixed modification of cysteine carbamidomethylation were set for the database search. Only high confidence peptides defined by Sequest HT with a 1% the peptide FDR by Percolator were considered for identification. ion tolerance was set 988 https://doi.org/10.1021/acscentsci.2c01325 ACS Cent. Sci. 2023, 9, 980−991 ACS Central Science http://pubs.acs.org/journal/acscii Research Article Relative quantitation of identified proteins between the control and treated samples was determined by the Label Free Quantitation (LFQ) workflow in PD 2.5. The precursor abundance intensities for each peptide identified by MS/MS in each sample were automatically determined, and their unique peptides for each protein in each sample were summed and used for calculating the protein abundance by PD 2.5 software with normalization against the spike yeast enolase protein. Protein ratios were calculated based on pairwise ratio for treatment over control samples. CTSL Activity Assay Using Z-FR-AMC. A Cytation 5 (BioTek) plate-reader was used to analyze CTSL enzymatic activity against a dipeptide substrate Z-FR-AMC (R&D Systems ES009). CTSL for this experiment was purchased from BPS Biosciences as part of the Cathepsin L Inhibitor screening assay kit (Cat. # 79591). The kit buffer was substituted with a similar buffer: 100 mM MES-NaOH pH 5.5, 7.5 mM DTT, and 150 mM NaCl. 2-Fold serial dilutions of CQ were preincubated with CTSL for 25 min on ice. The final enzyme concentration was 0.016 ng/μL, the final substrate concentration was 5 μM, and the reaction was run at 25 °C for 30 min. Three technical replicates were run for each reaction, and the experiment was performed three times. AMC fluorescence (substrate cleavage) was monitored by excitation at 340 nm and emission at 445 nm (F445). Proper controls were run for all components, as CQ fluorescence has potential to impact fluorescence spectra. Percent activity was calculated by comparing F445 for DMSO treated CTSL to the varying CQ concentrations. IC50 was calculated using Graphpad Prism analysis using log(inhibitor) vs response nonlinear regression analysis. Purification of CTSL. CTSL-Flag-Myc mammalian ex- pression vector was purchased from Origene (Cat. # RC203143). The plasmid was expressed in six 15 cm plates of HEK-293T cells (ATCC). Cells were collected and lysed in 1%-NP-40 lysis buffer. Importantly, protease inhibitors were not used in the lysis process. Instead, all steps were performed as efficiently as possible on ice. Lysates were enriched with Flag-beads for 2 h (Sigma Cat. # A2220), washed 3 times in 0.1% NP-40 wash buffer and then 2 more times in 50 mM Tris, 150 mM NaCl buffer. Flag-beads were eluted with 180 μM Flag-peptide (Biomatik) and concentrated, and total protein was quantified via Bradford assay. An aliquot of purified protein was confirmed to be CTSL via Coomassie stained SDS-page gel and Western blot for Flag signal. CTSL Intrinsic Fluorescence. Purified CTSL-Flag in- trinsic fluorescence was monitored using a Cytation 5 plate reading instrument. 200 nM of protein was added to the same, chilled, CTSL activity assay buffer (see above) and kept in a clear-bottom 96-well UV-star plate (Greiner Cat. # 655809). Exciting the protein at 265 nm and monitoring emission from 300 nm to beyond 400 nm showed a peak above the buffer background around 308 nm. 10 min preincubation of increasing CQ concentrations caused an increase in fluorescence at this peak. At concentrations of 25 and 50 μM (Figure 6B), CQ alone has lower fluorescence than buffer control at 308 nm. Thus, the data are presented with CQ background subtraction to best represent fluorescence increase of CTSL with CQ treatment. Chloroquine Computational Modeling. A reference CTSL structure was obtained from PDB database (2XU3) and loaded into MOE (2020) software as Biomolecule Assembly with default settings.44 The protein structure was then prepared using the QuickPrep function with the default parameters and thoroughly checked using the Structure Preparation function. The potential binding sites of CTSL were calculated and identified using the Site Finder tool. CQ was then docked into the identified site, and the docking poses were scored using London dG and refined based on GBVI/ WSA dG. The most confident binding pose was selected and visualized in MOE software. Fluorescence Polarization Assay. The stock solution of purified CTSL was diluted with mixture of 10× assay buffer (final concentrations of 25 mM Tris pH 8.0, 150 mM NaCl and 0.01% Tween-20), CQ-TAMRA (150 nM) and water to a total volume of 50 uL in Corning 96-well, half-area black plates. The plate was covered and left on ice for 10 min. Two technical replicates per sample type were measured. The plate was scanned 3 times on Cytation5 using a FP filter cube (Agilent, part number: 8040562, Ex: 530/25, Em: 590/35). The parallel and perpendicular fluorescence intensities of each well were recorded, and the mP values were then calculated based on the blank-subtracted data using established formula.45 ■ ASSOCIATED CONTENT *sı Supporting Information The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acscentsci.2c01325. Figures S1−S56, Table S1, and synthetic procedures for NS560 and CQ-X (PDF) Additional data (XLSX) Transparent Peer Review report available (PDF) ■ AUTHOR INFORMATION Corresponding Authors Timothy E. Glass − Department of Chemistry, University of Missouri, Columbia, Missouri 65211, United States; orcid.org/0000-0002-5064-3341; Email: glasst@ missouri.edu Hening Lin − Department of Chemistry and Chemical Biology, Cornell University, Ithaca, New York 14853, United States; Howard Hughes Medical Institute, Department of Chemistry and Chemical Biology, Cornell University, Ithaca, New York 14853, United States; orcid.org/0000-0002- 0255-2701; Email: hl379@cornell.edu Authors Michael R. Smith − Department of Chemistry and Chemical Biology, Cornell University, Ithaca, New York 14853, United States Le Zhang − Department of Chemistry, University of Missouri, Columbia, Missouri 65211, United States Yizhen Jin − Graduate Program of Biochemistry, Molecular and Cell Biology, Department of Chemistry and Chemical Biology, Cornell University, Ithaca, New York 14853, United States Min Yang − Department of Chemistry and Chemical Biology, Cornell University, Ithaca, New York 14853, United States Anusha Bade − Department of Chemistry, University of Missouri, Columbia, Missouri 65211, United States Kevin D. Gillis − Dalton Cardiovascular Research Center, Department of Bioengineering and Department of Medical Pharmacology and Physiology, University of Missouri, Columbia, Missouri 65211, United States 989 https://doi.org/10.1021/acscentsci.2c01325 ACS Cent. Sci. 2023, 9, 980−991 ACS Central Science http://pubs.acs.org/journal/acscii Research Article Sadhan Jana − Department of Chemistry and Chemical Biology, Cornell University, Ithaca, New York 14853, United States Ramesh Naidu Bypaneni − Department of Chemistry, University of Missouri, Columbia, Missouri 65211, United States Complete contact information is available at: https://pubs.acs.org/10.1021/acscentsci.2c01325 Author Contributions #These authors contributed equally. Notes The authors declare the following competing financial interest(s): Hening Lin is a founder and consultant for Sedec Therapeutics. ■ ACKNOWLEDGMENTS We thank Dr. Sheng Zhang and Dr. Qin Fu at Cornell University’s Proteomic Facility for help with the proteomics studies and HHMI for the purchase of the Orbitrap Eclipse Tribrid mass spectrometer. Imaging data were acquired through the Cornell Institute of Biotechnology’s Imaging Facility, with NIH 1S10RR025502 funding for the shared Zeiss LSM 710 Confocal Microscope. This work is supported in part by NIH/NIGMS grant R35GM131808 and by NSF grant (CHE 2203359). We also thank Dr. Gia Voeltz (mCh-Sec61), Dr. Richard Pagano (dsRed-Rab5 and dsRed-Rab7), Dr. Michael Davison (mCh-Rab11a), and Walther Mothes (Lamp1-RFP) for their gift of plasmids. ■ REFERENCES (1) Efeyan, A.; Comb, W. C.; Sabatini, D. M. Nutrient-Sensing Mechanisms and Pathways. Nature 2015, 517 (7534), 302−310. (2) Bröer, S.; Bröer, A. Amino Acid Homeostasis and Signalling in Mammalian Cells and Organisms. Biochem. J. 2017, 474 (12), 1935− 1963. (3) Goberdhan, D. C. I.; Wilson, C.; Harris, A. L. Amino Acid Sensing by MTORC1: Intracellular Transporters Mark the Spot. Cell Metab. 2016, 23 (4), 580−589. (4) Sabatini, D. M. Twenty-Five Years of MTOR: Uncovering the Link from Nutrients to Growth. Proc. Natl. Acad. Sci. U. S. A. 2017, 114 (45), 11818−11825. (5) Onorati, A. V.; Dyczynski, M.; Ojha, R.; Amaravadi, R. K. Targeting Autophagy in Cancer. Cancer 2018, 124 (16), 3307−3318. (6) Fedele, A. O.; Proud, C. G. Chloroquine and Bafilomycin A Mimic Lysosomal Storage Disorders and Impair MTORC1 Signaling. Biosci. Rep. 2020, 40 (4), No. BSR20200905, DOI: 10.1042/ BSR20200905. (7) Abu-Remaileh, M.; Wyant, G. A.; Kim, C.; Laqtom, N. N.; Abbasi, M.; Chan, S. H.; Freinkman, E.; Sabatini, D. M. Lysosomal Metabolomics Reveals V-ATPase- and MTOR-Dependent Regulation of Amino Acid Efflux from Lysosomes. Science (80-.). 2017, 358 (6364), 807−813. (8) Wyant, G. A.; Abu-Remaileh, M.; Wolfson, R. L.; Chen, W. W.; Freinkman, E.; Danai, L. V.; Vander Heiden, M. G.; Sabatini, D. M. MTORC1 Activator SLC38A9 Is Required to Efflux Essential Amino Acids from Lysosomes and Use Protein as a Nutrient. Cell 2017, 171 (3), 642−654.e12. (9) Zhang, Z. Y.; Monleon, D.; Verhamme, P.; Staessen, J. A. Branched-Chain Amino Acids as Critical Switches in Health and Disease. Hypertension 2018, 72 (5), 1012−1022. (10) Saxton, R. A.; Sabatini, D. M. MTOR Signaling in Growth, Metabolism, and Disease. Cell 2017, 168 (6), 960−976. (11) Stamenkovic, M.; Janjetovic, K.; Paunovic, V.; Ciric, D.; Kravic- Stevovic, T.; Trajkovic, V. Comparative Analysis of Cell Death Mechanisms Induced by Lysosomal Autophagy Inhibitors. Eur. J. Pharmacol. 2019, 859, 172540. (12) Mauthe, M.; Orhon, I.; Rocchi, C.; Zhou, X.; Luhr, M.; Hijlkema, K. J.; Coppes, R. P.; Engedal, N.; Mari, M.; Reggiori, F. Chloroquine Inhibits Autophagic Flux by Decreasing Autophago- some-Lysosome Fusion. Autophagy 2018, 14 (8), 1435−1455. (13) De Duve, C.; De Barsy, T.; Poole, B.; Trouet, A.; Tulkens, P.; Van Hoof, F. Lysosomotropic Agents. Biochem. Pharmacol. 1974, 23 (18), 2495−2531. (14) Ganzleben, I.; Neurath, M. F.; Becker, C. Autophagy in Cancer Therapy-Molecular Mechanisms and Current Clinical Advances. Cancers (Basel). 2021, 13 (21), 5575. (15) Marin, S.; Val, A. M.; Peligero, M. B.; Rodríguez-Bernuz, C.; Pérez-Ricart, A.; Vilaró Jaques, L.; Paredes, R.; Roca, J.; Quiñones, C. Safety of Short-Term Treatments with Oral Chloroquine and Hydroxychloroquine in Patients with and without COVID-19: A Systematic Review. Pharmaceuticals 2022, 15 (5), 634. (16) Drożdżal, S.; Rosik, J.; Lechowicz, K.; Machaj, F.; Szostak, B.; Przybyciński, J.; Lorzadeh, S.; Kotfis, K.; Ghavami, S.; Łos, M. J. An Update on Drugs with Therapeutic Potential for SARS-CoV-2 (COVID-19) Treatment. Drug Resist. Updat. 2021, 59, 100794. (17) Chandris, P.; Giannouli, C. C.; Panayotou, G. Imaging Approaches for the Study of Metabolism in Real Time Using Genetically Encoded Reporters. Front. Cell Dev. Biol. 2022, 9, 725114 DOI: 10.3389/fcell.2021.725114. (18) Yoshida, T.; Nakajima, H.; Takahashi, S.; Kakizuka, A.; Imamura, H. OLIVe: A Genetically Encoded Fluorescent Biosensor for Quantitative Imaging of Branched-Chain Amino Acid Levels inside Single Living Cells. ACS Sensors 2019, 4 (12), 3333−3342. (19) Wu, J.; Abdelfattah, A. S.; Zhou, H.; Ruangkittisakul, A.; Qian, Y.; Ballanyi, K.; Campbell, R. E. Genetically Encoded Glutamate Indicators with Altered Color and Topology. ACS Chem. Biol. 2018, 13 (7), 1832−1837. (20) Hu, H.; Gu, Y.; Xu, L.; Zou, Y.; Wang, A.; Tao, R.; Chen, X.; Zhao, Y.; Yang, Y. A Genetically Encoded Toolkit for Tracking Live- Cell Histidine Dynamics in Space and Time. Sci. Rep. 2017, 7 (1), 1− 9. (21) Singh, S.; Sharma, M. P.; Ahmad, A. Construction and Characterization of Protein-Based Cysteine Nanosensor for the Real Time Measurement of Cysteine Level in Living Cells. Int. J. Biol. Macromol. 2020, 143, 273−284. (22) Kiselyov, K.; Chen, J.; Rbaibi, Y.; Oberdick, D.; Tjon-Kon- Sang, S.; Shcheynikov, N.; Muallem, S.; Soyombo, A. TRP-ML1 Is a Lysosomal Monovalent Cation Channel That Undergoes Proteolytic Cleavage. J. Biol. Chem. 2005, 280 (52), 43218−43223. (23) Yue, Y.; Huo, F.; Yin, C. The Chronological Evolution of Small Organic Molecular Fluorescent Probes for Thiols. Chem. Sci. 2021, 12 (4), 1220−1226. (24) Zhang, L.; Liu, X. A.; Gillis, K. D.; Glass, T. E. A High-Affinity Fluorescent Sensor for Catecholamine: Application to Monitoring Norepinephrine Exocytosis. Angew. Chemie - Int. Ed. 2019, 58 (23), 7611−7614. (25) Achilli, C.; Ciana, A.; Fagnoni, M.; Balduini, C.; Minetti, G. Susceptibility to Hydrolysis of Phenylboronic Pinacol Esters at Physiological PH. Cent. Eur. J. Chem. 2013, 11, 137−139. (26) Hettie, K. S.; Glass, T. E. Coumarin-3-Aldehyde as a Scaffold for the Design of Tunable PET-Modulated Fluorescent Sensors for Neurotransmitters. Chem. - A Eur. J. 2014, 20 (52), 17488−17499. (27) Marceau, F.; Bawolak, M. T.; Lodge, R.; Bouthillier, J.; Gagné- Henley, A.; C.-Gaudreault, R.; Morissette, G. Cation Trapping by Cellular Acidic Compartments: Beyond the Concept of Lysosomo- tropic Drugs. Toxicol. Appl. Pharmacol. 2012, 259 (1), 1−12. (28) Huta, B. P.; Mehlenbacher, M. R.; Nie, Y.; Lai, X.; Zubieta, C.; Bou-Abdallah, F.; Doyle, R. P. The Lysosomal Protein Saposin B Binds Chloroquine. ChemMedChem. 2016, 11 (3), 277−282. (29) Rebecca, V. W.; Nicastri, M. C.; Fennelly, C.; Chude, C. I.; Barber-Rotenberg, J. S.; Ronghe, A.; McAfee, Q.; McLaughlin, N. P.; Zhang, G.; Goldman, A. R.; Ojha, R.; Piao, S.; Noguera-Ortega, E.; Martorella, A.; Alicea, G. M.; Lee, J. J.; Schuchter, L. M.; Xu, X.; 990 https://doi.org/10.1021/acscentsci.2c01325 ACS Cent. Sci. 2023, 9, 980−991 ACS Central Science http://pubs.acs.org/journal/acscii Research Article Herlyn, M.; Marmorstein, R.; Gimotty, P. A.; Speicher, D. W.; Winkler, J. D.; Amaravadi, R. K. PPT1 Promotes Tumor Growth and Is the Molecular Target of Chloroquine Derivatives in Cancer. Cancer Discovery 2019, 9 (2), 220−229. (30) Lecaille, F.; Chowdhury, S.; Purisima, E.; Brömme, D.; Lalmanach, G. The S2 subsites of cathepsins K and L and their contribution to collagen degradation. Protein Sci. 2007, 16 (4), 662− 70. (31) Zhitomirsky, B.; Yunaev, A.; Kreiserman, R.; Kaplan, A.; Stark, M.; Assaraf, Y. G. Lysosomotropic Drugs Activate TFEB via Lysosomal Membrane Fluidization and Consequent Inhibition of MTORC1 Activity. Cell Death Dis. 2018, 9 (12), 1−15. (32) Lu, S.; Sung, T.; Lin, N.; Abraham, R. T.; Jessen, B. A. Lysosomal Adaptation: How Cells Respond to Lysosomotropic Compounds. PLoS One 2017, 12 (3), e0173771. (33) Al-Bari, A. A. Chloroquine Analogues in Drug Discovery: New Directions of Uses, Mechanisms of Actions and Toxic Manifestations from Malaria to Multifarious Diseases. J. Antimicrob. Chemother. 2015, 70 (6), 1608−1621. (34) Poole, B.; Ohkuma, S. Effect of Weak Bases on the Intralysosomal PH in Mouse Peritoneal Macrophages. J. Cell Biol. 1981, 90 (3), 665−669. (35) Hannaford, J.; Guo, H.; Chen, X. Involvement of Cathepsins B and L in Inflammation and Cholesterol Trafficking Protein NPC2 Secretion in Macrophages. Obesity 2013, 21 (8), 1586−1595. (36) Cermak, S.; Kosicek, M.; Mladenovic-Djordjevic, A.; Smiljanic, K.; Kanazir, S.; Hecimovic, S. Loss of Cathepsin B and L Leads to Lysosomal Dysfunction, NPC-like Cholesterol Sequestration and Accumulation of the Key Alzheimer’s Proteins. PLoS One 2016, 11 (11), e0167428. (37) Zhao, M. M.; Yang, W. L.; Yang, F. Y.; Zhang, L.; Huang, W. J.; Hou, W.; Fan, C. F.; Jin, R. H.; Feng, Y. M.; Wang, Y. C.; Yang, J. K. Cathepsin L Plays a Key Role in SARS-CoV-2 Infection in Humans and Humanized Mice and Is a Promising Target for New Drug Development. Signal Transduct. Target. Ther. 2021, 6 (1), 134 DOI: 10.1038/s41392-021-00558-8. (38) Ou, T.; Mou, H.; Zhang, L.; Ojha, A.; Choe, H.; Farzan, M. Hydroxychloroquine-Mediated Inhibition of SARS-CoV-2 Entry Is Attenuated by TMPRSS2. PLoS Pathog. 2021, 17 (1), e1009212. (39) Hoffmann, M.; Mösbauer, K.; Hofmann-Winkler, H.; Kaul, A.; Kleine-Weber, H.; Krüger, N.; Gassen, N. C.; Müller, M. A.; Drosten, C.; Pöhlmann, S. Chloroquine Does Not Inhibit Infection of Human Lung Cells with SARS-CoV-2. Nature 2020, 585 (7826), 588−590. (40) Zurek, N.; Sparks, L.; Voeltz, G. Reticulon Short Hairpin Transmembrane Domains Are Used to Shape ER Tubules. Traffic 2011, 12 (1), 28−41. (41) Choudhury, A.; Dominguez, M.; Puri, V.; Sharma, D. K.; Narita, K.; Wheatley, C. L.; Marks, D. L.; Pagano, R. E. Rab Proteins Mediate Golgi Transport of Caveola-Internalized Glycosphingolipids and Correct Lipid Trafficking in Niemann-Pick C Cells. J. Clin. Invest. 2002, 109 (12), 1541−1550. (42) Sharma, D. K.; Choudhury, A.; Singh, R. D.; Wheatley, C. L.; Marks, D. L.; Pagano, R. E. Glycosphingolipids Internalized via Caveolar-Related Endocytosis Rapidly Merge with the Clathrin Pathway in Early Endosomes and Form Microdomains for Recycling. J. Biol. Chem. 2003, 278 (9), 7564−7572. (43) Sherer, N. M.; Lehmann, M. Jimenez-Soto, L. F.; Ingmundson, A.; Horner, S. M.; Cicchetti, G.; Allen, P. G.; Pypaert, M.; Cunningham, J. M.; Mothes, W. Visualization of Retroviral Replication in Living Cells Reveals Budding into Multivesicular Bodies. Traffic 2003, 4 (11), 785−801. (44) Molecular Operating Environment (MOE) 2022.02; Chemical Computing Group ULC: Montreal, Canada, 2023. (45) Lea, W. A.; Simeonov, A. Fluorescence Polarization Assays in Small Molecule Screening. Expert Opin. Drug Discovery 2011, 6 (1), 17−32. J.; 991 https://doi.org/10.1021/acscentsci.2c01325 ACS Cent. Sci. 2023, 9, 980−991
10.1017_s0260210522000614
Review of International Studies (2023), 49: 4, 763–779 doi:10.1017/S0260210522000614 R E S E A R C H A R T I C L E From savages to snowflakes: Race and the enemies of free speech Darcy Leigh* Sussex Law School, School of Law, Politics and Sociology, Freeman Centre, University of Sussex, Brighton, United Kingdom *Corresponding author. Email: d.leigh@sussex.ac.uk (Received 5 July 2021; revised 28 July 2022; accepted 1 September 2022) Abstract Right-wing free speech advocacy is increasingly shaping global politics. In IR, free speech has generally been viewed within human rights and international legal frameworks. However, this article shows that contemporary free speech advocates often ignore or oppose human rights and international law, focusing instead on (what they describe as) a defence of the nation state against the enemies of free speech. This article examines this articulation of free speech’s enemies: first historically as the ‘savage’ in John Stuart Mill’s influential formulation of free speech; and then contemporarily as the ‘snowflake’, ‘mob’, and ‘cul- tural Marxist’ by elected officials and lobbyists in the UK and US. The article argues that John Stuart Mill’s savage is figured within a racialised civilisational hierarchy of degrees of humanity. Today, right-wing free speech advocates extend and reconfigure this hierarchy, imagining the ‘snowflake’, ‘mob’, and ‘cultural Marxist’ as lesser human, subhuman, and extra-human, respectively. Thus, in contrast to rights-based analyses of free speech advocacy – which assume or assess the promotion of rights as a ‘public good’ – the article argues that narratives of free speech’s enemies are deployed by right-wing free speech advocates to underwrite racialised policy responses and global hierarchies. Keywords: Free Speech; Far Right; Race; The Human; White Supremacy Introduction On 3 July 2020, on the eve of US Independence Day, former US President Donald Trump spoke at Mount Rushmore in defence of free speech.1 According to Trump, the censorious enemies of free speech were engaged in a ‘merciless campaign to wipe out our history … erase our values, and indoctrinate our children.’2 These enemies had, in Trump’s narrative, taken over state and societal institutions, instituting ‘extreme indoctrination and bias’ in which left-wing domination was enforced through the threat of being ‘censored, banished, blacklisted, persecuted, and pun- ished’.3 Trump described Black Lives Matter (BLM) protests, then taking place globally, as a par- ticular threat: these ‘angry mobs’ were attacking the free expression of American nationalism and global civilisation. The ‘mob’ was variously criminalised (‘unleash[ing] a wave of violent crime in our cities’), lacking rationality (having ‘no idea why they are doing this’), and/or highly inten- tional (‘some know why they are doing this’).4 Throughout, Trump used the language of war. US citizens had ‘fought’, ‘struggled’, and ‘bled’ to secure freedom of speech, which was now 1Donald Trump, ‘Speech at Mount Rushmore’, South Dakota, 3 July 2020, available at: {rev.com/blog/transcripts/donald- trump-speech-transcript-at-mount-rushmore-4th-of-july-event} accessed 20 February 2021. 2Ibid 3Ibid 4Ibid © The Author(s), 2023. Published by Cambridge University Press on behalf of the British International Studies Association. This is an Open Access article, distributed under the terms of the Creative Commons Attribution licence (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted re-use, distribution and reproduction, provided the original article is properly cited. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l b u P 4 1 6 0 0 0 2 2 5 0 1 2 0 6 2 0 S / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h 764 Darcy Leigh under ‘attack’ and ‘radical assault’ from the dangerous ‘weapon’ of ‘cancel culture’.5 The American people were not ‘weak’ but ‘strong’, and ready to fight in defence of ‘the nation’s chil- dren’.6 Trump closed by announcing the creation of the ‘National Guard of American Heroes’: a ‘vast outdoor park’ in which statues of ‘the greatest Americans who have ever lived’ would defend America and civilisation against the enemies of free speech.7 Trump’s speech embodies the concerns of a right-wing free speech movement that has become increasingly voluble and influential in the Global North during the last decade.8 The speech also illustrates the failure of IR to address this development or its significance in global politics. Free speech in IR is usually viewed, as by some Constructivist IR scholars, as a human right located within international legal frameworks.9 These scholars join a rich literature beyond IR, in Philosophy, Law and Media Studies, which explores the legal or practical scope of a right to free speech.10 Yet con- temporary right-wing free speech advocates tend not to reference or act on – or are actively opposed to – international law and/or human rights.11 Further, while Constructivist IR research tends to focus on less powerful actors using rights frameworks to challenge power inequities,12 right-wing free speech advocates often have disproportionally large public platforms, which they use to consolidate existing hierarchies.13 In this light, a focus on human rights and international law is ill equipped to grasp the nature of contemporary right-wing free speech advocacy, which, as illustrated by Trump’s speech, is more often concerned with securing the nation against its enemies. If contemporary right-wing free speech advocacy does not uphold (or even address) human rights, international law and/or a defence of the voiceless, what is its function? To answer this question, this article examines the articulation of free speech’s enemies as a central feature of con- temporary free speech advocacy. The article argues that free speech advocates locate their enemies on a hierarchy of development, via an account of their proximity to whiteness, statehood, and humanity. Historically, this civilisational rationality was made integral to free speech in ‘the most famous liberal defence of free speech’,14 John Stuart Mill’s On Liberty, which also figured ‘the savage’ as a proto-enemy of free speech.15 Today, the enemies of free speech are figured 5Ibid 6Ibid 7Ibid 8See overviews of this movement in Gavin Titley, Is Free Speech Racist? (Cambridge, UK: Polity, 2020); P. Moskowitz, The Case against Free Speech: The First Amendment, Fascism, and the Future of Dissent (New York, NY: Bold Type Books, 2019). 9D. C. Thomas, ‘The Helsinki effect’, in Thomas Risse, Stephen Ropp, and Kathryn Sikkink (eds), The Power of Human Rights: International Norms and Domestic Change (Cambridge, UK: Cambridge University Press, 1999); D. C. Thomas, The Helsinki Effect (Princeton, NJ: Princeton University Press, 2001); A. Callamard and L. Bollinger (eds), Regardless of Frontiers (New York, NY: Columbia University Press, 2021). For broader Constructivist analyses of human rights norms, see Risse, Ropp, and Sikkink (eds), The Power of Human Rights; Kathryn Sikkink, ‘Transnational politics, International Relations the- ory, and human rights’, Political Science and Politics, 31:3 (1998), pp. 516–23; Martha Finnemore and Kathryn Sikkink, ‘Taking stock: The constructivist research program in International Relations and comparative politics’, Annual Review of Political Science, 4 (2001), pp. 391–416. 10Eric Barendt, Freedom of Speech (Oxford, UK: Oxford University Press, 2007); Ivan Hare and James Weinstein (eds), Extreme Speech and Democracy (Oxford, UK: Oxford University Press, 2011). 11For example, some free speech advocates who supported the UK exit from the EU oppose European human rights legis- lation and promote ‘British liberties’ as a replacement for human rights. C. R. G., ‘Murray, Magna Carta’s tainted legacy: Historic justifications for a British Bill of Rights and the case against the Human Rights Act’, in F. Cowell (ed.), The Case Against the 1998 Human Rights Act: A Critical Assessment (London, UK: Routledge, 2017). 12Finnemore and Sikkink, ‘Taking stock’. 13This is illustrated, as Will Davies argues, by professors and journalists writing about their own censorship in major news outlets. William Davies, ‘The free speech panic: How the right concocted a crisis’, The Guardian (26 July 2018), available at: {https://www.theguardian.com/news/2018/jul/26/the-free-speech-panic-censorship-how-the-right-concocted-a-crisis} accessed 6 December 2022. 14David van Mill, ‘Freedom of Speech’, Stanford Encyclopaedia of Philosophy (2017), available at: {plato.stanford.edu/ entries/freedom-speech} accessed 20 February 2021. 15John Stuart Mill and Elizabeth Rapaport, On Liberty (Cambridge, MA: Hackett Publishing, 1869). Hereafter ‘Mill, On Liberty’. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l b u P 4 1 6 0 0 0 2 2 5 0 1 2 0 6 2 0 S / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Review of International Studies 765 as ‘generation snowflake’, ‘the mob’, and the ‘cultural Marxist’. These figures – which variously repeat, extend, and refigure a Millian civilisational racial hierarchy – are deployed, the article shows, to enact and/or underwrite (especially racialised and/or colonial) statecraft and global hierarchies. The article proceeds in four sections. The first section locates right-wing free speech advocacy in IR and, empirically, in global politics. The second section develops an analytic framework based in Critical and Queer IR (on Cynthia Weber’s ‘figuration’ specifically), as well as Black Studies scholarship.16 The third section reads John Stuart Mill’s account of free speech through this framework, showing how statehood, whiteness, and free speech are connected, in the figure of ‘the savage’, through Mill’s civilisational rationality. The fourth section situates imagined contem- porary enemies of free speech – ‘generation snowflake’, ‘the mob’, and ‘cultural Marxism’ – as differently located within, informed by and/or revising Mill’s framework. The article concludes with a discussion of the implications of its analysis for the populations who these figurations are claimed to represent and for future IR research on free speech. Locating right-wing free speech advocacy: In global politics and in IR Research on free speech is relatively absent from IR. This section discusses three exceptions – regarding human rights,17 right-wing populism18 and the securitising regulation of speech19 – where IR scholarship directly or indirectly addresses an aspect of contemporary free speech advocacy. In my reading, scholarship in these fields situates free speech as a human rights discourse poten- tially open to co-optation or distortion, relating to a rising global populist movement, and entangled with narratives of defence, sovereignty, and exceptionalism. Ultimately, however, the section argues that these approaches fail to capture the significance of free speech advocacy as: part of the white supremacist histories of the US and UK, as well as global imperialism more broadly; undermining divisions between ‘moderate’ and ‘fringe’ right-wing politics; and deploy- ing ‘freedom’ in racially stratifying ways (making a turn to ‘freedom’ a problematic response to the racialised securitisation of regulation). From the 1960s to the 1980s, free speech was a central demand of left wing, Black, women’s and LGBT rights movements.20 Today, in Western liberal democracies, free speech advocacy is 16Cynthia Weber, Queer International Relations: Sovereignty, Sexuality and the Will to Knowledge (New York, NY: Oxford University Press, 2016), pp. 28–33; see also Donna Haraway, Modest Witness@Second Millennium.FemaleMan Meets OncoMouse: Feminism and technoscience (New York, NY: Routledge, 1997). 17Thomas, ‘The Helsinki effect’; Callamard and Bollinger (eds), Regardless of Frontiers; Risse, Ropp, and Sikkink (eds), The Power of Human Rights; Sikkink, ‘Transnational politics, International Relations theory, and human rights’; Finnemore and Sikkink, ‘Taking stock’. 18Sandra Destradi and Johannes Plagemann, ‘Populism and International Relations: (Un)predictability, personalisation, and the reinforcement of existing trends in world politics’, Review of International Studies, 45:5 (2019,) pp. 711–30; Bice Maiguashca, ‘Resisting the “populist hype”: A feminist critique of a globalising concept’, Review of International Studies, 45:5 (2019), pp. 768–85; Vedi Hadiz and Angelos Chryssogelos, ‘Populism in world politics: A comparative cross-regional perspective’, International Political Science Review, 38:4 (2017), pp. 399–411; Pablo de Orellana and Nicholas Michelsen, ‘Reactionary internationalism: The philosophy of the New Right’, Review of International Studies, 45:5 (2019), pp. 748–67; Jean-Francois Drolet and Michael C. Williams, ‘The radical Right, realism, and the politics of conservatism in postwar inter- national thought’, Review of International Studies, 47:3 (2021), pp. 273–93. 19Nadya Ali, ‘Seeing and unseeing prevent’s racialised borders’, Security Dialogue, 51:6 (2020), pp. 579–96; Andrew Neal, ‘University free speech as a space of exception in Prevent?’, in Ian Cram (ed.), Extremism, Free Speech and Counter-Terrorism Law and Policy (London, UK: Routledge, 2019); Randy Borum, ‘Rethinking radicalization’, Journal of Strategic Security, 4:4 (2011), pp. 1–6; P. R. Neumann, ‘The trouble with radicalization’, International Affairs, 89:4 (2013), pp. 873–93; Mark Sedgwick, ‘The concept of radicalization as a source of confusion’, Terrorism and Political Violence, 22:4 (2010), pp. 479– 94; see also Rita Floyd, ‘Parallels with the hate speech debate: The pros and cons of criminalising harmful securitising requests’, Review of International Studies, 44:1 (2017), pp. 43–6. 20Cynthia Enloe and Review of International Studies, ‘Interview with Professor Cynthia Enloe’, Review of International Studies, 27:4 (2001), pp. 649–66. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l b u P 4 1 6 0 0 0 2 2 5 0 1 2 0 6 2 0 S / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h 766 Darcy Leigh more often associated with a range of right-wing movements, including those identified as centre- right, right-wing populist, libertarian, and/or conservative. Constructivist scholarship is one of the few fields in IR where free speech has been addressed, either as an explicit and central object of analysis or, more often, within a broader package of international human rights norms or legal frameworks.21 For example, Daniel Thomas examines how norms surrounding the right to free speech circulate internationally, as well as how shared ideas, identities, or information contribute to (or inhibit) the implementation of international law.22 Often focusing on authoritarian or post- authoritarian states, such analyses tend to view free speech and its advocacy, along with rights more broadly, as a public good and challenge to the powerful by the powerless.23 This leaves con- structivist approaches ill-equipped to account for contemporary right-wing free speech advocacy in Western liberal democracies, which often opposes human rights and international law, or consolidates rather than challenging existing hierarchies. Nonetheless, a rights-based approach illuminates some aspects of the landscape of contempor- ary free speech politics. Assessed against the ‘successful’ diffusion or implementation of the right to free speech, contemporary right-wing free speech advocates can be viewed as claiming but failing to protect free speech as a right.24 Or, contemporary free speech advocates might be viewed as deploying free speech rhetoric to legitimise right-wing political activities and/or to have ‘co-opted’ free speech from ‘the left’ and/or from international human rights advocates. This argument is made in recent longform journalism by William Davies and Nesrine Malik.25 Yet this narrative alone misapprehends the history of free speech activism, which, as I show elsewhere26 and illustrate in the discussion of Mill below, has been co-constituted with racialised state formation and empire since the 1800s.27 That is, the racial stratification of modern state formation was expressed and extended through free speech advocacy long before its recent uptake by right-wing advocates. The implications of this history are obscured if we assume that right-wing free speech advocacy can be fully explained as a ‘recent’ ‘co-optation’ of human rights discourse. In this way, the article situates free speech within the co-constitution of liberalism, modern statehood, and empire, observed by Critical IR scholars.28 For Mill, however, free speech is not simply one of many rights constituting state citizenship but the principle upon which both statehood and international order are based.29 This article argues that this state-forming role is taken up and rearticulated in contemporary right-wing free speech advocates’ accounts of the enemies of free speech: in their accounts of their enemies free speech advocates are not simply failing or dishonest in their claims to promote rights, but are engaged in a long-running project of colonial and racialised statecraft enacted in the name of free speech.30 This chronology undermines any straightforward narrative that the ‘public good’ of free speech has been appropriated for harmful ends. In fact, this chronology suggests that even 1960s 21For example, Risse, Ropp, and Sikkink (eds), The Power of Human Rights; Sikkink, ‘Transnational politics, International Relations theory, and human rights’; Finnemore and Sikkink, ‘Taking stock’. 22Thomas, ‘The Helsinki effect’. 23Finnemore and Sikkink describe this as a trend in Constructivist research in general. Finnemore and Sikkink, ‘Taking stock’. 24Moskowitz shows that right-wing free speech advocates are often more invested in controlling or constraining speech than ‘freeing’ it. Moskowitz, The Case against Free Speech. 25Davies, ‘The free speech panic’; Nesrine Malik, ‘The myth of the free speech crisis’, The Guardian (3 September 2019), available at: {https://www.theguardian.Com/world/2019/sep/03/the-myth-of-the-free-speech-crisis} accessed 6 December 2022. 26Darcy Leigh, ‘The settler coloniality of free speech’, International Political Sociology, 16:3 (2022), pp. 1–16. 27I argue elsewhere that this is true from the emergence of modern free speech as a concept in the 1700s, but say 1800s here because this is the time period addressed in this article. Leigh, ‘The settler coloniality of free speech’. 28Jens Bartelson, A Genealogy of Sovereignty (Cambridge, UK: Cambridge University Press, 1995); Jens Bartelson, The Critique of the State (Cambridge, UK: Cambridge University Press, 2001). 29Mill, On Liberty. 30Leigh, ‘The settler coloniality of free speech’. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l b u P 4 1 6 0 0 0 2 2 5 0 1 2 0 6 2 0 S / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Review of International Studies 767 left-wing free speech activism might – in a similar vein to Critical and Queer IR analyses of other human rights movements31– be revisited and resituated in light of the racialised history of advo- cacy for the right to free speech. This is not to say that all free speech advocacy is determined by or reducible to the civilisational rationality embodied in Mill’s ‘savage’, or to foreclose how a range of movements might be situated within the Mill’s legacy (resistance or alternative to that legacy may be possible). Rather, this suggestion underscores the potential implications of interrupting the chronology implied by a narrative of free speech as recently co-opted by the right, and refuses to assume that left-wing expressions of free speech are unshaped by a racialising heritage. A second field in IR that addresses an aspect of right-wing free speech advocacy is the growing body of scholarship on the rise of the neofascist populist far-right and right-wing extremism.32 Although this scholarship does not address free speech itself, free speech is a central component of the emergent far-right populist ‘reactionary internationalism’,33 which IR scholars show is reshaping international politics. Free speech advocacy should be viewed, like far-right populist, neofascist, and extremist movements, as international: even when free speech advocacy is expressed as a concern with the decline of the nation,34 or an intrusion into the expression of nationalism,35 these concerns are taken up and deployed internationally on both practical and ideological levels.36 As such, despite this article’s focus on the UK and US, it addresses a move- ment that spans Western Europe, North America, Australia, and Aotearoa/New Zealand. However, not only are ‘fringe’, ‘extremist’, neofascist, far right, or populist politics not the pri- mary object of this article, but the article calls into question an exceptionalist delineation of those politics. The article shows that the figuration of free speech’s enemies is one way in which the neofascist, extremist, and/or populist far right and more ‘moderate’ free speech advocates are connected and collaborate: the enemies of free speech are figured similarly or jointly across a wide spectrum of right-wing politics. In this way, right-wing free speech advocacy is entangled with populist far-right politics via the figuration of the enemy of free speech. As such, rather than addressing the populist far right directly, by centring the imagined enemies of free speech, this article undermines any clear lines or exceptionalism surrounding far right populism. A final field of IR scholarship relating to free speech addresses the regulation or constraint of speech in the name of ‘counter-terror’37 and ‘deradicalisation’.38 In these cases, some speech is designated as threatening to the security of the nation-state and in need of (often exceptional or violent) constraint. This securitisation of the regulation of speech, some Critical IR scholars 31These arguments are often focused on the roles of women’s and LGBT rights in military intervention, border policies, and neocolonialism, see, for example, Weber, Queer International Relations and Jasbir Puar, Terrorist Assemblages: Homonationalism in Queer Times (Durham, NC: Duke University Press, 2007). 32Destradi and Plagemann, ‘Populism and International Relations’; Maiguashca, ‘Resisting the “populist hype”: A feminist critique of a globalising concept’; de Orellana and Michelsen, ‘Reactionary internationalism’; Drolet and Williams, ‘The rad- ical right’. 33This term is borrowed from de Orellana and Michelsen, ‘Reactionary internationalism’. 34As in Greg Lukianoff and Jonathan Haidt, The Coddling of the American Mind: How Good Intentions and Bad Ideas are Setting up a Generation for Failure (London, UK: Penguin, 2018); Hara Estroff Marano, A Nation of Wimps: The High Cost of Invasive Parenting (New York, NY: Broadway Books, 2008). 35As in Trump, ‘Speech at Mount Rushmore’. 36This was evidenced in March 2018, when Martin Sellner, the Austrian leader of far-right European group Generation Identity, was denied entry to the United Kingdom. UK-based far-right leader Tommy Robinson then delivered Sellner’s speech in his stead, citing the refused entry as censorship. Later it was revealed that both activists collaborate to circulate funds internationally. James Poulter, ‘The far right are uniting around their right to free speech’, Vice (20 March 2018), avail- able at: {https://www.vice.com/en/article/j5ax9d/the-far-right-are-uniting-around-their-right-to-free-speech} accessed 20 February 2021; Ben Quinn, ‘Far-right fundraising not taken seriously by UK, report finds’, The Guardian (31 May 2019), available at: {https://www.theguardian.com/world/2019/may/31/far-right-fundraising-not-taken-seriously-uk-government- extremists} accessed 20 February 2021. 37Ali, ‘Seeing and unseeing Prevent’s racialised borders’; Neal, ‘University free speech as a space of exception in Prevent?’. 38Borum, ‘Rethinking radicalization’; Neumann, ‘The trouble with radicalization’; Sedgwick, ‘The concept of radicalization as a source of confusion’; Floyd, ‘Parallels with the hate speech debate’. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l b u P 4 1 6 0 0 0 2 2 5 0 1 2 0 6 2 0 S / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h 768 Darcy Leigh argue, underwrites white supremacy and other racial hierarchies. For example, analyses by Nadia Ali39 and Andrew Neal40 show how defence of the state against terrorism via regimes of speech is racialised, whether by assigning whiteness to narratives of the state41 or targeting com- munities of colour in practice.42 While this scholarship addresses specific policy contexts (for example, Prevent in the UK), and does not consider free speech or its imagined enemies expli- citly, it does reflect the concerns of contemporary free speech advocates when it comes to figuring the enemies of free speech, as well as the securitising and racially stratifying effects of this figuration. Yet focusing solely on the regulation of speech implies that the ‘unfreedom’ of regulation is in some way tied to the ‘unfreedom’ of racialised state suppression43 – or, to put it another way, that the racialised constraint of speech is an affront to free speech and/or could be corrected with freer speech. Without disputing the observation that speech is restricted along racial lines, the current article com- plicates any simple turn to ‘free speech’ or its advocacy as a response to the racialised constraint of speech: the article shows that, through the racialised figuration of free speech’s enemies, calls for free speech can restrict freedoms and enact white supremacy as much as calls for restriction do. Overall, when free speech has been considered in IR, it has been primarily addressed within a framework of rights as a ‘social good’ or international legal norm. This not only fails to account for the contemporary right-wing expression of free speech, but risks obscuring a history in which free speech is articulated through state-formation and racialised state violence. While free speech is a concern of right-wing populist, extremist, or neofascist movements, centring the figuration of the enemies of free speech shows that these movements are not exceptional nor fully distinct from more ‘moderate’ politics. Finally, while calls for the regulation of speech highlight speech as a site of racialised securitisation, they fail to address the ways in which, through references to an imagined enemy, calls for free speech do not necessarily oppose, but rather extend, racially hierarchical state formation. The following section further situates the current article within IR scholarship, developing a methodology grounded in Critical, Queer, and Decolonial IR. Analytic framework: Figuration, developmental temporality, and racialised degrees of humanity Since Richard Ashley’s 1989 account of ‘statecraft as mancraft’,44 which shows how sovereign state formation is underwritten by the articulation of ‘sovereign man’, Critical, Feminist, and Queer IR scholars have identified a range of figures through which modern statehood is constituted. Echoing Ashley’s identification of both ‘man’ and ‘his others’ as constitutive of sovereign state formation,45 IR scholarship on figures has focused both on those that stand in for the modern state, and on the others, outsiders and threats, against which statehood is articulated. Such figures include, for example, soldiers and statesmen,46 ‘mothers, monsters and whores’,47 diplomats,48 39Ali, ‘Seeing and unseeing’. 40Neal, ‘University free speech as a space of exception in Prevent?’. 41Ali, ‘Seeing and unseeing’. 42Neal, ‘University free speech as a space of exception in Prevent?’. 43This is illustrated by Neal’s discussion of whether or not Prevent unfairly targets or constrains people of colour in uni- versities. Neal, ‘University free speech as a space of exception in Prevent?’. 44Richard Ashley ‘Living on border lines: Man, poststructuralism, and war’, in James Der Derian and Michael Shapiro (eds), International/Intertextual Relations (New York, NY: Lexington Books, 1989), pp. 260–313. 45Ibid. 46Christine Sylvester, Feminist Theory and International Relations in a Postmodern Era (Cambridge, UK: Cambridge University Press, 1994). 47Laura Sjoberg and Caron E. Gentry, Mothers, Monsters, Whores: Women’s Violence in Global Politics (London, UK: Zed Books, 2007). (2020), pp. 573–93. 48Ann Towns, ‘“Diplomacy is a feminine art”: Feminised figurations of the diplomat’, Review of International Studies, 46:5 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l b u P 4 1 6 0 0 0 2 2 5 0 1 2 0 6 2 0 S / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Review of International Studies 769 and, beyond the discipline of IR, the ‘monster, terrorist [and/or] fag’,49 and ‘the soldier and the terrorist’.50 More recently, in a study of figures of ‘the homosexual’, Cynthia Weber labels the pro- cess through which figures are articulated in global politics ‘figuration’, setting out a framework for analysing figuration in IR.51 This section draws on and adapts Weber’s framework, centring Weber’s focus on developmental temporality. It draws on Black Studies scholarship to add an emphasis on the racialisation of ‘the human’ (or humanisation and dehumanisation). The article subsequently locates the enemies of free speech among the many figures identified by IR scholars as sites of global politics. Weber describes how figures come to be seen as extant and stable through the process of figuration, which occurs in practices, policies, ideas, and rhetoric.52 Figures do not correspond to the groups they are claimed to represent, but are instead mobilised as statecraft to underwrite policies and/or global hierarchies. For example, Weber shows how the figure of the ‘normal LGBT rights holder’53 marks Western states as developed nations, legitimises their dominance in the international sphere, and obscures inaction on issues affecting queer populations not represented as normal (for example, on queer migration or homelessness). In contrast, the figure of the ‘perverse’ homosexual immigrant or terrorist justifies border and deportation policies aimed at securing Western states against a ‘racially darkened’ dangerous threat, as well as international intervention in the name of ‘development’.54 Weber’s analysis provides a framework for analysing free speech advocates’ focus on the developmental status of their enemies. Weber argues that figuration relies on and reproduces a developmental temporality, which subsequently underpins the policies and hierarchies enacted by figuration.55 In doing so, Weber echoes Critical IR scholarship on temporality, which shows that a developmental temporality is constitutive of liberal statehood and modern colonial global order.56 Weber’s analysis shows that the relationship of figures to this temporality is com- plex, eschewing binaries of ‘developed’ vs ‘underdeveloped’ or ‘past’ vs ‘present’. For example, the ‘normal’ LGBT rights holder is located as both advanced in comparison with the underdeveloped ‘perverse’ homosexual, and temporally universal in contrast to the provincial ‘perverse’ homosex- ual.57 At the same time, some ‘perverse’ homosexuals are located as less developed within linear- progressive time (as ‘underdeveloped’), or as stuck in the past or prior-to developmental time (as ‘undevelopable’).58 In the case of Weber’s homosexual, it is this developmental temporality that informs, for example, the interventionist or anti-immigration policies and other statecraft justified by these figures. Given free speech advocates’ emphasis on the humanity (or lack thereof) of the enemies of free speech, it is worth noting how ‘the human’ is situated within Weber’s developmental temporality. Weber argues that ‘the human’ of human rights is situated within the universal, which is equated 49Jasbir Puar and Amit Rai, ‘Monster, terrorist, fag: The war on terrorism and the production of docile patriots’, Social Text, 20:3 (2002) pp. 117–48. 50Adi Kuntsman, ‘The soldier and the terrorist: Sexy nationalism, queer violence’, Sexualities, 11:1–2 (2008), pp. 142–70. 51Weber, Queer International Relations; Weber borrows this term and concept from Haraway, Modest Witness@Second Millennium.FemaleMan Meets OncoMouse 52Weber’s use of the term ‘figuration’ as both a verb and a noun emphasises the ongoing-ness of any figure that appears as stable. Here, however, I use both ‘figure’ and ‘figuration’ for ease of reading: the term ‘figure’ should be read as expressing the same unfolding process as ‘figuration’. Weber, Queer International Relations. 53Weber, Queer International Relations, p. 29. 54Ibid., pp. 31–5. 55Ibid., pp. 29–31; drawing on Donna Haraway, Modest Witness@Second Millennium.FemaleMan Meets OncoMouse. 56See, for example, Kimberly Hutchings, ‘Happy Anniversary! Time and critique in International Relations theory’, Review of International Studies, 33:S1 (2007), pp. 71–89; Anna Agathangelou and Kyle Killian (eds), Time, Temporality and Violence in International Relations: (De)Fatalizing the Present, Forging Radical Alternatives (New York, NY: Routledge, 2016). 57Weber, Queer International Relations, quotations from p. 32, argument made throughout book. 58Ibid. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l b u P 4 1 6 0 0 0 2 2 5 0 1 2 0 6 2 0 S / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h 770 Darcy Leigh with progress and development.59 This is underscored by poststructuralist,60 posthuman,61 and decolonial IR62 scholars, who show that ‘the human’ more broadly is often articulated as a white, non-disabled, heterosexual, Christian and male citizen-subject. This scholarship shows that the figuration of this human – standing in for progress, citizenship, security, and sovereign statehood – is integral to developmental and colonising global politics. The racialisation of ‘the human’ – and the implications of this figuration for global politics – is underscored in Black Studies scholarship on the dehumanisation of blackened figures.63 This scholarship shows that blackness is often figured as animal, object, and/or otherwise sub- human.64 As Zakiyyah Iman Jackson describes, blackness has been repeatedly dehumanised, bestialised, or objectified, with a lack of (perceived) development or civilisation cited as evi- dence of a lack of full humanity.65 This blackened subhumanity has legitimised and informed anti-Black state formation, not least the transatlantic slave trade and imperialism. Especially relevant to figurations of the enemy of free speech – who is often viewed as lacking the capacity for rationality – Jackson draws attention to the ways that lack of development or humanity is articulated through an assessment of Black minds and rationality as lacking self-conscious rationality, or ‘the clarity of self-knowledge’.66 Both blackness and irrationality have also, Jackson argues, been feminised and/or articulated in relation to deviant or ‘uncivilised’ femin- inity. As I describe below, this blackened dehumanisation is especially, but not exclusively, res- onant with right-wing free speech advocates’ narratives surrounding the ‘uncivilised’ ‘threat’ posed by anti-racist or Black activism. Methodologically, then, the current article follows an adapted version of Weber’s approach to figuration. It analyses books, articles, and speeches by right-wing free speech advocates – specifically elected politicians and lobbyists – as sites of the figuration of free speech’s enemies. The selection of these texts is not comprehensive, but each captures or circulates a particularly central or influential narrative among free speech advocates (e.g., they coined a term, informed a political response and/or are by high ranking politicians). The article does not treat ‘snowflakes’, ‘the mob’, or ‘cultural Marxists’ as existent subjects, but rather inves- tigates how their figuration in free speech advocacy informs policy and hierarchies. Like Weber, the article emphasises temporality, situating free speech advocates’ own emphasis on temporality within the developmental temporality of state formation and international relations. Finally, following Jackson, the article considers the degrees of humanity attributed to the enemies of free speech, especially when these are racialised and/or signalled by a perceived lack of rationality. 59Weber, Queer International Relations. 60See, for example, Ashley, ‘Living on border lines’. 61Audra Mitchell, ‘Only human? A worldly approach to security’, Security Dialogue, 45:1 (2014), pp. 5–22; Erika Cudworth, Stephen Hobden, and Emilian Kavalski (eds), Posthuman Dialogues in International Relations (London, UK: Routledge, 2018); Erika Cudworth, and Stephen Hobden, Posthuman International Relations: Complexity, Ecologism and Global Politics (London, UK: Zed, 2011). 62Vicki Squire, ‘Migration and the politics of “the human”: Confronting the privileged subjects of IR’, International Relations, 34:3 (2020), pp. 290–308; Louisa Odysseos, ‘Prolegomena to any future decolonial ethics: Coloniality, poetics and “being human as praxis”’, Millennium, 45:3 (2017), pp. 447–72; Audra Mitchell, International Intervention in a Secular Age: Re-Enchanting Humanity? (London, UK: Routledge, 2014). 63Sylvia Wynter, ‘Unsettling the coloniality of being/power/truth/freedom: Towards the human, after man, its over- representation – an argument’, The New Centennial Review, 3:3 (2003), pp. 257–337; Bénédicte Boisseron, Afro-Dog: Blackness and the Animal Question (New York, NY: Columbia University Press, 2018); Zakiyyah Iman Jackson, Becoming Human: Matter and Meaning in an Antiblack World (New York, NY: New York University Press, 2020). 64Ibid. 65Jackson’s discussion dehumanisation takes place in the introduction to Becoming Human, which subsequently seeks to displace this analysis as the sole register in which blackness and humanity are analysed together. Jackson, Becoming Human, p. 7. 66Ibid., p. 5. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l b u P 4 1 6 0 0 0 2 2 5 0 1 2 0 6 2 0 S / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Review of International Studies 771 John Stuart Mill’s civilisational free speech and its ‘savage’ other Working in the East India Company for thirty years, Mill was a colonial official in the mid-1800s whose work shaped European empire and state-formation.67 Today, Mill is widely recognised as ‘the most influential liberal thinker’68 on free speech. His well-known defence of free speech in On Liberty posits free speech as the most important principle in liberal states, with free expression driving societal progress.69 This section shows how Mill’s theory of free speech operates through the developmental temporality described by Weber, as well as the (connected) whitened version of the human and rationality described by Jackson. I argue that Mill’s ‘savage’ other to free speech, while not always viewed as a ‘threat’ as such, is nonetheless a proto-enemy of contempor- ary figurations of free speech’s enemies. While Mill is not the only nor even the original free speech theorist (John Locke before him advocated for greater ‘toleration’),70 he is exceptionally influential. The analysis of his work offered here is deployed later in the article to illuminate the civilisational logics that continue to underpin – or are otherwise taken up and rearticulated by – contemporary right-wing free speech advocacy. That a colonial framework underpins Mill’s work in general is well established.71 Yet the rela- tionship between this civilisational framework and Mill’s account of free speech – not least as expressed in Mill’s figure of ‘the savage’ – remains largely unexamined. One exception is my own work on the settler colonial dimension of the genealogy of free speech, where I detail how Mill articulates free speech through his colonising civilisational framework and vice versa, making the two inseparable.72 In my reading of On Liberty, Mill makes the following set of (somewhat circular) arguments: because statehood is the most rational and civilised form of gov- ernance, state formation indicates that a society is civilised and rational, while the absence of state formation indicates an absence of civilisation or rationality; because sovereign statehood is the most civilised and rational form of governance, and free speech drives towards rationalism and progressive civilisation, free speech should lead organically to state formation; only those societies that are civilised and rational already (again, signalled by the occurrence of state formation), should be granted free speech, and with it other citizenship rights and sovereign statehood.73 These are not abstract arguments, nor accounts of why colonial subjects did not speak (freely or otherwise). Rather, these arguments legitimised ‘despotism’74 over colonial subjects, including exclusion from participation in colonial states, and repression of Indigenous and Black cultures, languages, and political systems. They also authorised colonial expansion and governance in the 1800s more broadly.75 Departing from this analysis, I deploy Weber’s framework of figuration here to situate Mill’s ‘savage’ as central to his account of civilisational free speech. Mill’s ‘savage’ or ‘barbarian’ is figured as living in ‘… those backward states of society in which the race itself may be considered as in its [infancy].’76 Mill describes the ‘savage’ as ‘wandering or thinly scattered over a vast tract of country’), lacking ‘commerce’, ‘manufactures’, ‘agriculture’, ‘law’, ‘administration of justice’, ‘property’, or ‘intelligence’.77 For Mill, these forms of life define savagery as well as constituting 67Lynn Zastoupil, John Stuart Mill and India (Stanford, CA: Stanford University Press, 1994). 68van Mill, ‘Freedom of speech’. 69Mill, On Liberty; Barendt, Freedom of Speech. 70John Locke, An Essay Concerning Toleration (Indianapolis: Liberty Fund, 1685); for a reading of Locke’s work on free speech in relation to Mill’s, see Leigh, The Settler Coloniality of Free Speech. 71Jahn, ‘Barbarian thoughts’; Zastoupil, John Stuart Mill and India; Mehta, Liberalism and Empire. 72Leigh, ‘The settler coloniality of free speech’, pp. 8–11. 73This reading of the first chapter of Mill, On Liberty, is given in Leigh, ‘The settler coloniality of free speech’, pp. 8–11. 74Mill, On Liberty, pp. 9–10. 75Uday Singh Mehta, Liberalism and Empire: A Study in Nineteenth-Century British Liberal Thought (Chicago, IL: University of Chicago Press, 1999); Zastoupil, John Stuart Mill and India. 76Mill, On Liberty, pp. 9–10. 77John Stuart Mill, On Civilization (1836), p. 120. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l b u P 4 1 6 0 0 0 2 2 5 0 1 2 0 6 2 0 S / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h 772 Darcy Leigh a failure to form states or capitalist agricultural arrangements. Mill also figures the ‘savage’ with direct reference to their unreadiness for free expression, as living in a ‘… state of things anterior to the time when mankind have become capable of being improved by free and equal discus- sion.’78 To reiterate, ‘being improved by free and equal discussion’ would, for Mill, mean state- formation. Here we see how the figure of the ‘savage’ embodies Mill’s civilisational colonial framework of free speech described above. We also see both Weber’s developmental temporality and Jackson’s dehumanisation. The terms ‘infancy’ and ‘anterior to’ signal the developmental temporal relations between the ‘savage’ or ‘barbarian’ and what Mill describes as ‘human beings in the maturity of their faculties’.79 The emphasis on ‘the maturity of their faculties’ ties (what Mill sees as) the development of the human mind to both the practice of and right to sovereign state formation.80 Significantly for today’s free speech advocates, this infantilisation places ‘the savage’ outside the realm of legitimate political participation. The circularity of the argument means that colonised peoples are only entitled to ‘freedom’ of speech so long as that freedom is not expressed outside or against European state formation or colonial governance. Otherwise, in the name of rationality and civilisation, they are figured as unready for such freedom. However, in the same way that today’s free speech advocates imagine a varied set of enemies of free speech, so too Mill differentiated free speech’s others within a civilisational hierarchy. Different colonial subjects were, for Mill, located at different points within the temporality of development, with correlate rationales for varied regimes of British colonial governance in the name of development and civilisation.81 In some cases, Mill deemed figures as more capable of or susceptible to assimilation into rationality, civilisation, and statehood (this made Mill’s work ‘progressive’ – and Mill a ‘radical’ – in contrast to his predecessors in colonial governance). For example, Mill argued that Indian religious elites should be recruited by colonial officials to assist in governing or civilising other Indians.82 In contrast, Indigenous peoples in Europe’s settler colonies were figured as more lacking in modern human individuality, rationality, and civilised political organisation, justifying violent tactics of colonial occupation. In these ways, Mill establishes the tradition of free speech advocacy within a developmental temporality and in relation to racialised degrees of humanity. He figures the ‘savage’ as the ‘other’ to free speech and is concerned with the savage’s lack of rationality and/or inability to self- govern (and thus exclusion from the realm of the political). The following section turns to the contemporary figuration of free speech’s enemies and shows how each is figured within, extends or departs from a Millian hierarchy of civilisation. Contemporary right-wing free speech advo- cates, it argues, follow Mill in promoting or enacting (often racialised) state policies based on the civilisational status assigned to its enemies. The civilisational status accorded speech’s enemies today not only echoes and repeats, but also refigures and reworks Mill’s framework, not least by extending it through the hyper- or extra- human ‘cultural Marxist’. Contemporary figurations of free speech’s enemies: The lesser-human infantile ‘snowflake’, subhuman animalistic ‘mob’, and extra-human puppeteer ‘cultural Marxist’ This section argues that today the enemies of free speech are figured as infantile (‘the snowflake’), subhuman and animalistic (‘the mob’), and extra-human (‘the cultural Marxist’) in relation to Mill’s civilisational hierarchy. Overall, the section argues that the enemies of free speech function to inform policies and politics that ‘defend’ a whitened state against a racially darkened ‘enemy’ – 78Mill, On Liberty, pp. 9–10. 79Ibid. 80Ibid. 81Mehta, Liberalism and Empire. 82Zastoupil, John Stuart Mill and India, pp. 28–50. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l b u P 4 1 6 0 0 0 2 2 5 0 1 2 0 6 2 0 S / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Review of International Studies 773 not least by placing anti-racist and other activism outside the realm of legitimate participation in state politics. Mill’s ‘savage’ or civilisational framework are not uniformly reproduced in later iterations of free speech advocacy – these latter iterations not only reproduce and extend, but also – especially through the figure of the ‘cultural Marxist’ – rearticulate the racialised rationality of free speech in new ways. Infantile generation snowflake The trope of ‘generation snowflake’ – now in wide public circulation – centres on the figure of the young as weak, infantile, overly emotional, irrational, feminised, racialised, and/or deindividua- lised.83 Generation snowflake is figured as a censorious threat to free speech but also a victim of infantilisation by policymakers, educators, and parents (and, in turn, as a threat to and/or marker of threatened national character).84 In this way, the snowflake is a lesser and undeveloped human, but not always inhuman, and sometimes recoverable or developable. In 2016, Claire Fox, a peer in the UK House of Lords and former Member of European Parliament, as well as director of the think tank Academy of Ideas, offered an early public articu- lation of ‘generation snowflake’. Fox says younger generations are weaker than previous genera- tions (here she introduces the temporality of decline) and lacking in the robustness required for free debate. Fox describes ‘generation snowflake’ as ‘thin-skinned’,85 ‘febrile’,86 ‘fragile’,87 and ‘too mollycoddled and infantilised for the rough and tumble of real life’.88 According to Fox, weakness is joined with emotionality to cloud the judgement of generation snowflake and makes it unable to confront ideas or arguments as such (or as ‘just words’). Instead, as Fox argues elsewhere, when faced with ideas and arguments they disagree with, generation snowflake becomes ‘hysterical’ and ‘can’t cope’.89 Describing the reaction of some school students who objected to her views on sex- ual violence, Fox says, ‘Some of the girls were sobbing and hugging each other … while others shrieked.’90 Similarly, describing a group of Muslim girls approaching her after another speech to express their disagreement with her views on Islam, Fox says that their emotional reactions prevented them from receiving her rational argument rationally.91 Here, Weber’s developmental temporality is visible in the figuration of generation snowflake. Fox argues that members of ‘generation snowflake’ are underdeveloped, or wrongly developed, at the level of their individual life experiences. At the same time, by articulating this as generational and a departure from the trajectory of previous generations, Fox suggests this is a societal or national developmental problem. Concerns with ‘the human’ embodied in an individual rational mind are also present. Figuring the threat to free speech as generational deindividualises members of generation snowflake. When a younger person objects to Fox’s speech, this objection is framed as part of a generational ‘trend’, rather than political expression by an individual with the capacity for thought or political agency. Fox also racialises and genders the irrational ‘snowflake’ enemy of free speech by repeatedly associating it with Islam. Even when talking about non-Muslims, Fox uses the term ‘offense 83As in Fox, I Find That Offensive!. 84As in Lukianoff and Haidt, The Coddling of the American Mind; Marano, A Nation of Wimps. 85Fox, I Find That Offensive!, p. 7. 86Ibid., p. 17. 87Ibid., p. 37. 88Ibid., p. 9. 89Claire Fox, ‘Why today’s young women are just so FEEBLE’, Mail Online (9 June 2016), available at: {https://www.daily- mail.co.uk/femail/article-3632119/Why-today-s-young-women-just-FEEBLE-t-cope-ideas-challenge-right-view-world-says- academic.html} accessed 20 February 2021. 90Ibid. 91Fox, I Find That Offensive!, pp. 6–7. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l b u P 4 1 6 0 0 0 2 2 5 0 1 2 0 6 2 0 S / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h 774 Darcy Leigh fatwas’ to associate what she sees as over-emotional irrationality with Islam more broadly.92 In the story above, Fox also draws on misogynist tropes of shrieking and hysteria. She combines these with racialisation and deindividualisation into the ultimate ’snowflakes’: a group of emo- tional and irrational Muslim girls. The figuration of ‘generation snowflake’ informs a particular political response, as illustrated by Greg Lukianoff and Jonathan Haidt’s influential The Coddling of the American Mind. Drawing on a Cognitive Behavioural Therapy based psychological approach, Lukianoff and Haidt not only analyse generation snowflake, but set out a programme to address the threat posed by ‘snowflakes’ to free speech. The programme draws on Cognitive Behavioural Therapy techniques, along with metaphors of free debate as a ‘mental gymnasium’ or boxing ring.93 They argue that young people need to participate in debate as they would a gym or sparring session, in order to develop their strength for debate and disagreement, and to stop seeing themselves as weak. In the spirit of this argument, Haidt founded and now codirects the impactful US free speech organisation Foundation for Individual Rights in Education, which supports legal action against US univer- sities for perceived free speech violations (among other activities). In contrast to Fox, Lukianoff and Haidt reindividualise generation snowflake. Yet the effects are equally depoliticising. By suggesting the maldevelopment of generation snowflake can be cor- rected through individual psychological redevelopment, Lukianoff and Haidt further deny the rational thought and political agency of generation snowflake: they do not see collective youth organising as political expression, instead figuring it as an individualised psychological problem. They thus legitimise an interventionist, individualised, and pathologised response to opposition to right-wing politics.94 In all these ways, the figuration of ‘generation snowflake’ echoes Mill’s account of the ‘savage’ and those who ‘lack the maturity of their faculties’.95 Unlike Mill’s savage, however, ‘generation snowflake’ is also sometimes a victim of indoctrination. Yet like Mill’s ‘savage’, ‘snowflakes’ are often seen as developable. This may be because ‘the snowflake’ is associated with universities, which are, in turn, associated with whiteness, proximity to the state and access to institutions. Overall, however, in the absence of such development or assimilation, ‘generation snowflake’ is infantilised and depoliticised. The criminal, animalistic, and subhuman ‘mob’ The trope of ‘the mob’ figures the enemies of free speech as animalistic, criminal, and often black- ened. Here, I discuss the blackened animality, criminality, and threat to security of ‘the mob’, before showing how, as with the snowflake, opponents of right-wing free speech advocates are articulated as irrational, deindividualised, depoliticised. Unlike the snowflake, however, I suggest that the mob appears as entirely subhuman, threatening and unassimilable within the terms of free speech. I begin by discussing the blackened BLM ‘mob’, then consider the more generic ‘social justice mob’. As illustrated by the Trump speech with which this article opened, ‘the mob’ is often asso- ciated with anti-racist protesters, especially BLM and the removal or destruction of statues. When BLM protests and statue removal took place in mid-2020, UK and US governments framed their responses not as related to the politics of racism or antiracism, but with the rhetoric of free speech. BLM protestors were figured as a censorious ‘mob’. The ‘mob’ figured by UK and US gov- ernments in response to BLM was dehumanised and depoliticised through two key figurative moves. 92Ibid., p. 18. 93As in Lukianoff and Haidt, The Coddling of the American Mind, p. 18. 94Ibid. 95Mill, On Liberty, pp. 9–10. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l b u P 4 1 6 0 0 0 2 2 5 0 1 2 0 6 2 0 S / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Review of International Studies 775 First, ‘the mob’ was repeatedly articulated as animalistic and irrational. For example, then UK Secretary of State for Housing, Communities and Local Government, Robert Jenrick, called pro- testors ‘a baying mob’,96 equating BLM protestors with animals (‘baying’ is a noise made by a pack of dogs). This directly echoes the white supremacist articulation of blackness as animalistic described by Jackson. Jenrick’s bestialisation of BLM also figures the political expression of opposition to racism – including the toppling of statues – as a noise unintelligible to humans. Dehumanisation and animalisation were further expressed through claims that BLM protestors were unable or unwilling to express their dissent through rational and civilised state channels. For example, Jenrick argued that ‘what has stood for generations should be considered thought- fully, not removed on a whim’,97 as if BLM protestors had not ‘thought’ or ‘considered’ their actions but instead acted on some animalistic urge. Second, the mob was repeatedly figured as criminal. UK Secretary of State Priti Patel and Trump both reduced the protests to criminal acts, rarely mentioning BLM by name or even using the words ‘race’ or ‘protest’. Trump (2020) variously called BLM protestors a ‘mob’, ‘van- dals’, ‘violent extremists’, and arsonists, advocating ‘the full force of the law’ in response.98 Patel similarly called the BLM protests ‘hooliganism and thuggery’.99 Criminalising the protests in this way not only evoked stereotypes of working class and Black criminality, but also places interac- tions between the state and BLM within the realms of criminal justice or exceptional security, rather than politics. The enemy of free speech is not figured as ‘the mob’ solely in response to BLM protests. The term is also applied to left-wing activists or ‘social justice warriors’ more broadly.100 For example, students protesting right-wing free speech advocates visiting campuses across the UK and US are often figured as ‘mobs’ threatening free speech.101 Here, the racialisation of the enemy of free speech by free speech activists functions in complex ways. While these mobs may not be black- ened or otherwise racialised in the same way as BLM protestors, they may be implicitly racialised via their articulation as animalistic, irrational, and uncivilised. At the same time, the naming of these ‘social justice mobs’ as such avoids naming the politics of the groups the figure of ‘the mob’ is claimed to represent, which are often anti-racist or Black politics. In this way, race is evoked to further criminalise the mob, or goes unnamed in order to depoliticise opposition to racism. However, this does not mean the joining of blackness and animality in the trope of ‘the mob’ affects all those targeted by free speech activists equally. For example, while a majority white stu- dent anti-racist group may be described as an animalistic mob by free speech activists, they may also be figured as ‘snowflakes’, and it is unlikely that they will be responded to with the same state violence as, for example, the majority black participants in a BLM protest. Images of the white ‘mob’ – from KKK lynching to the ‘storming’ of the US Capitol building in 2021 – further com- plicate and extend this picture. Perhaps the ‘mob’ must be blackened to be fully criminalised and securitised. It is also possible that applying the language of the ‘mob’ to white supremacist violent risks naming animality or incivility rather than white supremacy as ‘the problem’. 96Cited in ‘Statues to get protection from "baying mobs"’, BBC News (17 January 2021), available at: {https://www.bbc.co. uk/news/uk-55693020} accessed 20 March 2021. 97Ibid. 98Trump, ‘Speech at Mount Rushmore’. 99Speech to UK Conservative Party Conference 2020, cited in Patrick Daly, ‘Priti Patel slams XR and BLM activists for “hooliganism and thuggery” during protests’, The Scotsman (4 October 2020), available at: {https://www.scotsman.com/ news/politics/priti-patel-slams-xr-and-blm-activists-hooliganism-and-thuggery-during-protests-2992424} accessed 20 April 2021. 100See, for example, by Stella Morabito, ‘What to learn from the social justice warrior who was eaten by his own mob’, The Federalist (18 July 2018), available at: {https://thefederalist.com/2018/07/18/learn-social-justice-warrior-eaten-mob/} accessed 20 April 2021. 101See, for example, by Mathew Goodwin, ‘Mob rule is crushing free speech on campus’, The Times (30 June 2019), avail- able at: {https://www.thetimes.co.uk/article/mob-rule-is-crushing-free-speech-on-campus-30269p6q9} accessed 20 March 2021. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l b u P 4 1 6 0 0 0 2 2 5 0 1 2 0 6 2 0 S / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h 776 Darcy Leigh Finally, the figuration of the ‘social justice mob’ as emerging in universities illustrates the over- lapping of different figurations of free speech’s enemies – in this case ‘the mob’ and ‘the snow- flake’. Often both tropes are mobilised simultaneously and in interconnected ways. Both deindividualise and depoliticise the political opponents of right-wing free speech activists. Both deny some degree of humanity, civilisation, and development among those opponents, with a focus on their lack of capacity for rational thought, rational discussion, or political subject- hood. However, while generation snowflake is brought into the realm of psychology (articulated as over-emotional), the mob is situated in the realm of criminality and security (articulated as violent and threatening). While the snowflake is articulated as vulnerable, the mob is articulated as threatening. In these ways, while both the snowflake and the mob can be understood in relation to Mill’s civilisational hierarchy, they are located differently within this hierarchy. Generation snowflake is articulated as a lesser human threat to national character or progress and in need of rescue or development (in need of CBT); the mob is articulated as subhuman and undevelop- able threats to the rule of law (in need of incarceration or a military response). The extra-human ‘cultural Marxist’ The trope of ‘cultural Marxism’ articulates a behind-the-scenes international conspiracy of Jewish intellectuals who are taking over liberal institutions and replacing free speech with indoctrin- ation.102 This section shows how figurations of the enemy of free speech as a ‘cultural Marxist’ rely on pre-existing antisemitic tropes of Jews as scheming, rich, and power-hungry. I argue that the ‘cultural Marxist’ is figured as extra-human and hyper-modern in its organisation and power, and as such as a threat to national sovereignty and state institutions. To understand the figuration of ‘cultural Marxism’ it is necessary to understand how this figure is deployed across ‘fringe’ neo-Nazi and alt-right groups (e.g., formed part of Norwegian mass shooter Anders Breivik’s manifesto),103 as well as ‘mainstream’ party politics (described below). The term originates with an explicit naming of cultural Marxists as Jews, Jews as dangerous intellectuals or and builds on an antisemitic tradition that paints Bolsheviks, wandering and thus disloyal to states, and/or controlling or taking over world polit- ics.104 Elected officials and lobbyists, however, tend to omit mentioning this heritage of the term or explicitly naming Jews, even while all other elements of the far right conspiracy theory remain intact. In this way, ‘cultural Marxism’ functions as a ‘dog whistle’ through which antisemitism is expressed in state politics in a plausibly deniable way.105 it A 2019 speech by Member of the UK Parliament and free speech advocate Suella Braverman captures the way that ‘cultural Marxists’ are figured as enemies of free speech.106 Braverman argues that, as a result of the overwhelming aims and power of ‘cultural Marxists’, ‘banning things is becoming de rigueur’, ‘freedom of speech is becoming a taboo’ and ‘our universities … are being shrouded in censorship and a culture of no-platforming’.107 This cultural Marxist takeover 102Tanner Mirrlees, ‘The Alt-Right’s discourse on “cultural Marxism”, Atlantis, 39:1 (2018), pp. 49–69. 103Andrew Berwick, A European Declaration of Independence (2011). This is searchable online but, following Sarah Ahmed’s politics of citation, I decline to link to it here. See Sara Ahmed, Living a Feminist Life (Durham, NC: Duke University Press, 2017). A survey of white supremacist texts deploying the trope of including Berwick’s manifesto, can be found in Mirrlees, ‘The Alt-Right’s discourse on “cultural Marxism”’. ‘cultural Marxism’, 104Bill Berkowitz, ‘Cultural Marixsm Catching On’, Southern Poverty Law Centre (15 August 2003), available at: {https:// www.splcenter.org/fighting-hate/intelligence-report/2003/cultural-marxism-catching} accessed 20 April 2021. 105For an analysis of this process, illustrated by a case study of the Australian far right, see Rachel Busbridge, Benjamin ‘Cultural Marxism: Far-right conspiracy theory in Australia’s culture wars’, Social 106Cited in Peter Walker, ‘Tory MP criticised for using antisemitic term “cultural Marxism”’, The Guardian (26 March {https://www.theguardian.com/news/2019/mar/26/tory-mp-criticised-for-using-antisemitic-term-cul- Moffitt, and Joshua Thorburn, Identities, 26:6 (2020), pp. 722–38. 2019), available at: tural-marxism} accessed 20 March 2021. 107Ibid. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l b u P 4 1 6 0 0 0 2 2 5 0 1 2 0 6 2 0 S / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Review of International Studies 777 was, for Braverman, ‘absolutely damaging for our spirit as British people, and our genius, whether it’s for innovation and science, or culture and civilisation … for statecraft’.108 As such, Braverman argues, ‘Conservatives are engaged in a battle’ against these enemies.109 A similar enemy of free speech was also figured by Trump at Mount Rushmore, as taking over ‘our schools, our news- rooms, even our corporate boardrooms’. Here, ‘cultural Marxists’ are viewed not simply as the political opponents of right-wing free speech advocates, but rather – via their imagined threat to free speech – as the enemies of the British nation and civilisation. In addition to being seen as disloyal threats to nationhood, and as power-hungry or scheming, they are attributed the power and coordination necessary to take over state institutions (rather than, for example, being seen as relatively limited and disem- powered student, left wing, or Jewish groups).110 Once again, the relationships between different figurations of free speech’s enemies are blurry. Is the cultural Marxist preying on vulnerable ‘snowflake’ youth, or creating them through a cen- sorious orthodoxy? Are the same ‘coddled’ university students also predatory ‘cultural Marxists’? For example, Braverman accused cultural Marxists of ‘putting everyone in cotton wool’, arguing that ‘a risk-averse mentality is now taking over’.111 ‘Cotton wool’ is often, as it is for Fox, a sig- nifier of ‘generation snowflake’.112 There is no one specific manifestation of the relationship of ‘cultural Marxism’ to other enemies of free speech: a range of narratives attendant to each circu- late between and are combined multiply by right-wing free speech advocates. This echoes Weber’s account of the complex interrelated developmental temporalities of figuration. In all these ways, like the ‘snowflake’ and ‘mob’, the ‘Cultural Marxist’ is deindividualised, fig- ured not as a human individual but a mass conspiracy. However, unlike the ‘snowflake’ and ‘mob’, the ‘Cultural Marxist’ is represented as hyper-rational and over-intelligent, rather than irrational or incapable of thought. The cultural Marxist is not a ‘normal’ rational human citizen- subject, but nor is this enemy a vulnerable infant or subhuman (despite sometimes overlapping or connecting with vulnerable youth and ‘snowflakes’). Instead, this enemy of free speech is figured as extra-human, hyper-strategic, and hyper-influential. The location of the ‘cultural Marxist’ does not appear within Weber’s analysis of developmental temporality or Jackson’s analysis of the human. Nor is it discussed by Mill in relation to civilisation. Instead, contemporary figurations of ‘cultural Marxism’ extend the developmental temporality with which racialised degrees of humanity are articulated into a distorted and threatening futurity. Conclusion This article has shown that Mill’s civilisational framework for free speech – embodied in his fig- uration of ‘the savage’ – is reproduced and rearticulated in contemporary free speech advocates’ articulation of their enemies. The ‘snowflake’, ‘mob’, and ‘cultural Marxist’ are all figured through and/or extend this framework. The article has further argued that the figuration of the enemies of free speech as ‘generation snowflake’, ‘the mob’, and ‘cultural Marxism’ authorise right-wing free speech advocates’ policymaking, depoliticise their opponents, and/or underwrite racialised hier- archies. Before closing, I now consider some possible implications of this analysis. First, for the populations which figured enemies are claimed to represent. Second, for researching free speech advocacy beyond right-wing electoral expressions in the UK and US. As Weber (2016) describes, figures do not correspond to the lived experience of subjects. In fact, this article has observed how right-wing free speech advocates often apply ‘generation 108Ibid. 109Ibid. 110Berkowitz, ‘Cultural Marixsm Catching On’; Mirrlees, ‘The Alt-right’s discourse on “cultural Marxism”’; Moffitt and Thorburn, ‘Cultural Marxism’. 111Moffitt and Thorburn, ‘Cultural Marxism’. 112Fox, I Find That Offensive!, p. 31. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l b u P 4 1 6 0 0 0 2 2 5 0 1 2 0 6 2 0 S / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h 778 Darcy Leigh snowflake’, ‘the mob’, and ‘the cultural Marxist’ (or aspects of these figures) to the very same populations. This is clear in free speech advocates’ opposition to BLM protestors, who are ima- gined both as ‘the mob’ and as a ‘cultural Marxist’ takeover. Similarly, university students are framed as both sensitive ‘snowflake’ victims, and a ‘censorious Marxist mob’ stifling free expres- sion. Given that each figure comes with its own political logic and implications – for example, rescue, development or incarceration/securitisation – it is possible that how and when popula- tions are figured as a particular ‘enemy’ reflects the broader (often racialised) politics of free speech advocates in relation to those populations. This would account for the shifting and multi- ply applied figurations of free speech’s enemies by free speech advocates depending on the context. While figurations do not correspond to the lived lives of subjects, the populations that figures are claimed to represent may engage – or be forced to engage – the process of figuration. According to Weber, particular figurations may be inhabited performatively and intentionally or forcibly. For example, Weber suggests that some ‘“homosexuals” welcome the opportunity to inhabit the image of the “LGBT rights holder’”, while others may find this figure constraining and/or inaccessible. In a very different context, some Black Studies scholars argue that wilfully embracing uncivility, the non-human and animality may be an opportunity for political solidar- ity, agency, and organising.113 They note, however, that this comes with risks in a context where the figuration of black people as subhuman is enforced, and might be co-opted, as a core function of white supremacist violence. With regards to the enemies of free speech, it is likely that the loca- tion of a figure within a civilisational framework determines, to some degree, the costs and oppor- tunities embracing that figure represents: a Black activist embracing the criminality of ‘the mob’ may find themselves at greater risk than, for example, a white activist embracing that same figure, or of either embracing the (potentially whitened) category of ‘generation snowflake’. At the same time, perhaps the same outsider status of ‘the mob’, which legitimises violence may also make it a politically potent and disruptive category. The question of whether or how the figures of ‘gener- ation snowflake’, ‘the mob’, and/or ‘cultural Marxism’ might be embraced or inhabited remains open. Finally, what does this article’s analysis of free speech’s enemies mean for how we understand free speech advocacy more broadly? The article has focused on right-wing conservative, libertar- ian, and populist elected politicians and lobbyists in the UK and US. This focus reflects the increasing dominance and influence of right-wing free speech politics in the Global North today, which has not been accounted for by research in IR that tends to view free speech as solely a public good, human right, and/or matter of international law. This leaves a wide range of con- temporary free speech advocacy unexamined. In the US and UK, this includes both those who identify as neo-Nazis or overt white supremacists and as left wing (notable examples of the latter in the US are academics facing university censure for criticism of the state of Israel or use of ‘Critical Race Theory’). In other countries, it includes movements countering state censorship, such as journalists and academics in Turkey, or religious minorities in China. In contrast to the right-wing free speech advocates examined here, who often have disproportionately large public platforms despite their claims to being victims of free speech’s enemies, some of these other free speech advocates face severe, even carceral or lethal, penalties for advocating free speech. While the specifics of these varied cases put them beyond the scope of this article, and it is absolutely not my intention to homogenise or dismiss all free speech advocacy, the article none- theless raises questions free speech advocacy beyond its right-wing electoral expression in the US and UK. At the very least, the article calls into question the framework of human rights, inter- national law, and norm diffusion as the de facto sole lens through which all free speech advocacy must be viewed. As I describe above, though such a lens might usefully assess free speech 113Jackson, Becoming Human; Bénédicte Boisseron, Afro-Dog. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l b u P 4 1 6 0 0 0 2 2 5 0 1 2 0 6 2 0 S / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h Review of International Studies 779 advocacy as more or less successful or disingenuous, it fails to capture the potentially productive function of such advocacy within global racial hierarchies. More specifically, without foreclosing the answer, the article raises the question of whether and how free speech advocates beyond UK and US right-wing advocacy figure, racialise and/or (de)humanise their enemies. For those work- ing within Mill’s legacy – which includes not only right-wing advocacy but also liberal multicul- turalism and ‘equality and diversity’ agendas114 – the question is raised as to whether and how Mill’s ‘savage’ and civilisational rationality persist or, perhaps, can be resisted. In these ways, this article expands and updates the small IR literature on free speech that has focused primarily on human rights diffusion, international law, and/or ‘progressive’ advocacy for free speech. It does so empirically, by examining recent right-wing free speech advocacy in the US and UK that often explicitly opposes human rights and international law. It does so methodo- logically, by addressing how free speech advocates figure the enemies of free speech, including how those enemies are racialised as human, subhuman, or extra-human. This shifts the analysis of free speech away from instrumental questions about rights implementation towards discursive and political ones. Free speech becomes visible as integral to a range of core IR concerns, not least (in Mill’s account) sovereignty and (in Trump’s account) national security. Free speech’s enemies become located among the constitutive figures of international politics. Acknowledgements. This article has been improved by comments and/or support from Daniel Bulley, Harry Josephine Giles, Laura Jung, Louiza Odysseos, Maddie Breeze, Matthew Evans, attendees of the Pan-European Conference on International Relations 2019, members of the Centre for Rights and Anti-Colonial Justice at the University of Sussex, the Editors of Review of International Studies, and anonymous reviewers. Dr Darcy Leigh is a Lecturer in Law at the University of Sussex, where she researches the history and ongoing present of the British Empire, with a focus on its settler colonial dimension and/or expression in gender and sexuality. Dr Leigh also teaches about colonialism, gender, and sexuality in university, activist, and community contexts, using democratic and creative pedagogies. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l b u P 4 1 6 0 0 0 2 2 5 0 1 2 0 6 2 0 S / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h 114I explore the question of multicultural policies and note its relevance to safer spaces activism elsewhere. Leigh, ‘The settler coloniality of free speech’. Cite this article: Leigh, D. 2023. From savages to snowflakes: Race and the enemies of free speech. Review of International Studies 49, 763–779. https://doi.org/10.1017/S0260210522000614
10.1016_j.enpol.2022.113313
What causes energy and transport poverty in Ireland? Analysing demographic, economic, and social dynamics, and policy implications Lowans, C., Foley, A., Furszyfer Del Rio, D., Caulfield, B., Sovacool, B. K., Griffiths, S., & Rooney, D. (2023). What causes energy and transport poverty in Ireland? Analysing demographic, economic, and social dynamics, and policy implications. Energy Policy , 172, Article 113313. https://doi.org/10.1016/j.enpol.2022.113313 Published in: Energy Policy Document Version: Publisher's PDF, also known as Version of record Queen's University Belfast - Research Portal: Link to publication record in Queen's University Belfast Research Portal Publisher rights Copyright 2022 The Authors. This is an open access article published under a Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution and reproduction in any medium, provided the author and source are cited. General rights Copyright for the publications made accessible via the Queen's University Belfast Research Portal is retained by the author(s) and / or other copyright owners and it is a condition of accessing these publications that users recognise and abide by the legal requirements associated with these rights. Take down policy The Research Portal is Queen's institutional repository that provides access to Queen's research output. Every effort has been made to ensure that content in the Research Portal does not infringe any person's rights, or applicable UK laws. If you discover content in the Research Portal that you believe breaches copyright or violates any law, please contact openaccess@qub.ac.uk. Open Access This research has been made openly available by Queen's academics and its Open Research team. We would love to hear how access to this research benefits you. – Share your feedback with us: http://go.qub.ac.uk/oa-feedback Download date:03. Aug. 2024 Contents lists available at ScienceDirect Energy Policy journal homepage: www.elsevier.com/locate/enpol What causes energy and transport poverty in Ireland? Analysing demographic, economic, and social dynamics, and policy implications Christopher Lowans a, *, Aoife Foley a, b, Dylan Furszyfer Del Rio a, c, Brian Caulfield b, Benjamin K. Sovacool c, g,i, Steven Griffiths e, David Rooney h a School of Mechanical and Aerospace Engineering, Queen’s University Belfast, Belfast, United Kingdom b Department of Civil, Structural, and Environmental Engineering, Trinity College Dublin, The University of Dublin, Dublin, 2, Ireland c Science Policy Research Unit, University of Sussex, Brighton, United Kingdom e Khalifa University of Science and Technology, Abu Dhabi, United Arab Emirates g Department of Business Technology and Development, Aarhus University, Denmark h School of Chemistry and Chemical Engineering, Queen’s University of Belfast, Belfast, United Kingdom i Earth and Environment, Boston University, United States A R T I C L E I N F O A B S T R A C T Keywords: Energy poverty Transport poverty Covid-19 Ireland Nationally representative survey Energy and transport poverty have been postulated as conditions linked by overlapping causal factors such as structural economic inequality or housing stock and affecting overlapping demographics such as family size or income. The strength of the overlap of these conditions and their causal mechanisms has not been assessed across Ireland prior to this study. We apply and analyse existing and novel energy and transport poverty metrics in a survey of 1564 participants across Ireland and consider results from expenditure and consensual data examining causal mechanisms and correlations. We find that energy and transport poverty rates are broadly similar across Ireland at approximately 14% for energy poverty and 18% for transport poverty using the half-median metric, while participant knowledge of causal factors, such as lack of domestic energy efficiency and perceived desir- ability of potential poverty solutions, such as increased public transport provision, are low. Furthermore, we find that self-reported data concerning energy and transport expenditures and preferences do not correspond to ex- pected outcomes. We thus conclude that ever refined targeting of individuals and households for support measures is not optimal for either decarbonisation or alleviation of energy and transport poverty conditions and suggest some salient policy implications. 1. Introduction Energy and transport poverty can co-occur and reinforce each other leading to a “double energy vulnerability”. Historically, energy and transport poverty were treated as different problems with their own causes and consequences (Simcock et al., 2021). Recently, however, it has been postulated that these conditions are not distinct and have overlapping causes and links (Mattioli et al., 2017). One of the key characteristics of this double energy vulnerability is that it could force individuals or groups of individuals to choose which service to prioritise, for example, choosing between heating the home or paying for school transport (Sovacool and Furszyfer Del Rio, 2022). This dichotomic issue ought to take more policy relevance. It has been shown that as many as 6% of neighbourhoods, or 3 million people in England, are at risk of “double energy vulnerability” clustered in isolated rural areas, due to a lack of both energy and transport infra- structure (Robinson and Mattioli, 2020). The current energy crisis is expected to place enormous pressure on households and public services during the winter of 2022 (Bolton and Stewart, 2022) (Torjesen, 2022). We position our research focused on the extent of energy and transport poverty and their causal mechanisms with a view to uncovering routes to their alleviation across the Island of Ireland. The literature has defined fuel (or energy) poverty as the inability to secure materially and socially-necessitated energy services, such as heating a home or using appliances (Bouzarovski and Petrova, 2015). This lack of energy provision results in a range of physical health, mental * Corresponding author. School of Mechanical and Aerospace Engineering, Queen’s University Belfast, Ashby Building, Stranmillis Road, Belfast, BT9 5AH, United Kingdom. E-mail address: clowans01@qub.ac.uk (C. Lowans). https://doi.org/10.1016/j.enpol.2022.113313 Received 21 April 2022; Received in revised form 3 October 2022; Accepted 20 October 2022 EnergyPolicy172(2023)113313Availableonline4November20220301-4215/©2022TheAuthors.PublishedbyElsevierLtd.ThisisanopenaccessarticleundertheCCBYlicense(http://creativecommons.org/licenses/by/4.0/). C. Lowans et al. health and social impacts, including increased risk of circulatory and respiratory disease, increased social isolation and thousands of excess winter deaths annually (National Audit Office, 2003) (Rudge and Gil- christ, 2005) (Marmot Review Team and Friends of the Earth, 2011). Those most vulnerable to energy poverty are those least able to adapt to it, i.e., low-income households with children, the elderly, and the disabled, or those whose pre-existing health vulnerabilities are most acutely exacerbated (Bednar and Reames, 2020). Assessing energy poverty typically takes the form of either expenditure measures, where energy expenses are measured against a certain threshold, or via consensual measures, which assess the subjective lived experiences of households to determine poverty. Concerning expenditure measures, either modelled expenditure or need to spend (to maintain a certain heating regime) is used, such as with the widespread 10% metric (where household expenditure on energy exceeds 10% of their income after deductions), or actual spend, such as in the half-median metric (where household expenditure on energy is less than half the sample median) (Thomson et al., 2017), can be used. Transport poverty, meanwhile, deals with the lack of mobility ser- vices necessary for participation in society, resulting from the inacces- sibility, unaffordability or unavailability of transport (Lucas et al., 2016; Mattioli et al., 2017; Mullen and Marsden, 2016). Depending upon the definition, up to 90% of households may be affected by transport poverty (Lucas et al., 2016). The consequences of transport poverty are no less severe than those of energy poverty, given the increased likeli- hood of low-income and marginalised groups being exposed to transportation-related air pollution, violence, sexual harassment and crime (Furszyfer Del Rio and Sovacool, 2023). Other effects include restricted access to employment and the increased difficulties posed to the disabled (Lucas et al., 2016). Transport poverty metrics focus on one of the aspects of inaccessibility, unaffordability or unavailability of transport that were outlined by Lucas et al. (2016). However, as there is no standard definition of transport poverty, the metrics applied are not (yet) as sophisticated as those in the energy poverty domain. Assessments of causal mechanisms and options for the alleviation of energy poverty are scarce but increasing in number. For example, one recent study has assessed the causal instruments of energy poverty in eleven countries, while another has examined the global potential for alleviating energy poverty with renewable energy (Rao et al., 2022) (Zhao et al., 2022). Such assessments are also increasing in the transport poverty literature, for example, with recent studies examining the relationship between income and commute satisfaction in China and the accessibility of public transport in Oslo (Shi et al., 2022) (Lunke, 2022). Furthermore, studies are also increasingly paying attention to these is- sues as a joint subject, for instance, recent work in Iceland concerning the lived experience of energy and transport poverty (Upham et al., 2022). Assessing transport and energy poverty together, however, is not a simple task. Research in this area has concluded that measuring these problems poses a key challenge for researchers and impacts real-world outcomes (Mattioli et al., 2017). Challenges for uniting their measure- ment begin with the unit of measurement; households are energy poor while individuals are transport poor. Furthermore, while there are standards for household energy use, there are no standards for transport use. To compound this issue, a problem arises as to which comes first, data collection or metric definition, which creates the common chicken and egg problem. Additionally, no single metric captures all aspects of either condition, so using multiple metrics simultaneously is required for a more complete picture. The introduction of vulnerability lenses, i.e., assessing who is more likely to be vulnerable to each condition, is less technically challenging than measurement. We have argued in previous work that this ought to be used in conjunction with energy or transport poverty metrics (Lowans et al., 2021). Of contextual relevance is the Covid-19 pandemic. The Covid-19 pandemic had a profound impact on energy consumption in 2020, causing demand to contract by 5% (International Energy Agency, 2020). Beyond consumption, the Covid-19 pandemic has been noted to have implications for energy justice, energy poverty, and transport poverty (Sovacool et al., 2020). However, the pandemic also has been noted to create opportunities for sustainable responses (Griffiths et al., 2021). Consequently, a more thorough understanding of the impacts of the pandemic at the household level is required to assess the impacts of the pandemic and the opportunities arising from it. This research examines the Island of Ireland, which is comprised of 2 distinct political and legal jurisdictions, yet shares a common market for electricity, in addition to many areas of economic interdependence allowing for the comparison of causal factors. The Island of Ireland has a large proportion of rural dwellers (a demographic known to be vulner- able to energy and transport poverty), yet no assessment of energy and transport poverty as a joint issue exists for either jurisdiction. Further- more, assessments of energy and transport poverty are not up to date in either jurisdiction; we, therefore, aim to be more comprehensive with recent empirical and original data. To fill data gaps and examine the intersections between energy and transport poverty and the decarbonisation of the energy and transport systems, we conducted a nationally representative survey (n = 1564) with participants from the Island of Ireland. This work is the first cross border study across the Island of Ireland, which examines the conditions of energy and transport poverty simultaneously to inform future research on decarbonising each area in a just manner which is of particular importance given the ongoing energy price crisis. This paper has three aims, which are presented here and examined in Sections 4 and 5. 1. Assess and record self-reported expenditures on energy and transport services and use these and other collected data to assess contempo- rary energy and transport poverty on the Island of Ireland. 2. Assess the strength of the causal mechanisms of these conditions and assess their overlap. 3. Analyse how the Covid-19 pandemic affected energy and transport usage. The outcomes and conclusions of this research will be useful to re- searchers and practitioners seeking to alleviate energy and transport poverty, individually or as a joint issue. The article proceeds as follows. First, we begin by outlining the context of energy and transport trends across the Island of Ireland. Second, we discuss our research design and subsequently present our results and discuss them. Last, we derive conclusions from our findings and suggest some policy implications of these findings. 2. Contextualising energy and mobility trends and energy and transport poverty in Ireland As remarked in the introduction, insufficient provision of modern energy and transport services contributes significantly to deprivation across the developed world. Considering energy in Northern Ireland (NI) first, the latest House Condition Survey showed that in 2016, 22% of households in NI were in fuel poverty1, decreasing to 18% in 2018 due to a reduction in fuel prices (Northern Ireland Housing Executive, 2016). According to this official data, energy poverty data in NI is based upon modelled expenditure, which ignores actual spending patterns, exposing a data gap. In the Republic of Ireland, energy poverty rates are calcu- lated using data from the EU SILC database and have also historically been assessed using the 10% metric. During the development of in- dicators for EU wide comparison, the Energy Poverty Observatory found energy poverty rates in Ireland to range from 5% to 18%, depending on the chosen indicator (Energy Poverty Advisory Hub, 2020). 1 Note that our previous research has already criticised current metrics for being insufficient, and thus this number may be an underestimate. EnergyPolicy172(2023)1133132 C. Lowans et al. Work is emerging in interrogating the causal mechanisms of energy poverty and its effects on household incomes in the Republic of Ireland. Researchers have found that (using EU SILC data) fuel poverty is indistinct from general deprivation as defined by the National Measure of Deprivation for Ireland finding that when aspects of fuel poverty are included in the National Measure of Deprivation, fuel poverty and deprivation are subsequently indistinct (Watson and Maitre, 2015). Transport poverty remains somewhat indirectly quantified and has not been directly examined for quite some time, but NI’s problem in this area is considerable (General Consumer Council Northern Ireland, 2001). In NI, transport data forms the Travel Survey for Northern Ireland (TSNI), while the equivalent in Ireland is the National Travel Survey (NTS) (Department for Infrastructure, 2020) (Central Statistics Office, 2021). Common survey questions and outcomes include items such as average journey length and the main mode of transport. However, none of the data collected are used to explicitly measure transport poverty. Indicators of energy poverty are used for high level monitoring of energy poverty rates. However, access to support is often devolved to sub-national governments and subject to stringent targeting criteria such as having a very low household income, and often the incentives for landlords to access such schemes are greatly diminished. In Northern Ireland for example, access to the main retrofit funding scheme is a “postcode lottery” where access is only available in areas where fuel poverty is highest (Northern Ireland Housing Executive, 2022). Transport poverty indicators and vulnerability lenses are typically not used explicitly in any context but are implicitly acknowledged in accessing transport related supports. For example, in Ireland, people with disabilities are eligible for the Motorised Transport Grant, provided they require a vehicle to access employment, cannot use public trans- port, and are subjected to a means test (Citizensinformation.ie, 2022). However, these support measures can ignore the fact that the causal mechanism of transport poverty can often be related to the built envi- ronment. For example the Irish National Travel Survey notes that the greatest contributor to encouraging more cycling would be safer cycling routes (Central Statistics Office, 2021). In the literature concerning the alleviation of energy poverty, access to support measures for those in energy poverty to undertake building fabric upgrades is seen as nearly essential. Middlemiss and Gillard find that social housing providers are the most common source of lasting built fabric improvements, and that most respondents would not consider debt mechanisms to improve their dwelling fabric (Middlemiss and Gillard, 2015). It is noted that in similar research for the Scottish Government, most participants’ awareness about the availability of support is low, and few believe that they require help or advice and thus would not actively seek either (Ipsos MORI Scotland and Alembic Research Ltd., 2020). Regarding alleviating transport poverty, many barriers relate to the insufficient provision of or high cost of public transport or are related to infrastructural issues. Indeed, research concludes that street connectiv- ity, bus provision and neighbourhood safety are more significant con- tributors to spatial variation in transport use than demographic factors (Lucas et al., 2018). However, some barriers are more related to perception. In Northern Ireland, for instance, fear of travelling into unknown areas arises. Thus, not only must more transport options be available, but they must also be considered safe by users (Crisp et al., 2017). Overcoming transport poverty, therefore, requires changes to the provision of public transport and should avoid exacerbating existing inequalities. Unfortunately, subsidies for more sustainable mobility options such as EVs, which are noted as inequitable, have been found in Ireland (Caulfield et al., 2022). Furthermore, means of alleviating energy poverty and transport are often linked to climate goals in that they also present effective mecha- nisms for emissions reduction and are frequently key components of sectoral targets in climate change laws or plans. Northern Ireland and the Republic of Ireland have passed net-zero emissions laws with a deadline of 2050 for net-zero emissions and with various sub-sector targets (Minister for Agriculture, Environment and Rural Affairs, 2022) (Oireachtas, 2021). Sufficient support for vulnerable groups will be essential to meet climate goals. Our results also discuss their effec- tiveness for energy and transport poverty alleviation and climate change mitigation. 3. Research design Our survey instrument combines existing energy and transport poverty measurements as adapted from the EU Energy Poverty Advisory Hub with new assessments (Energy Poverty Advisory Hub, 2020). Moreover, it aims to determine how households think about financial trade-offs between energy and transport and thus points the way for prioritising current and future solutions. The survey asked questions according to the research aims of the project as well as to glean infor- mation beyond the data represented in current statistics (e.g., how households trade-off between energy/transport services and other es- sentials). The key objectives of this household survey are the following: 1. Record self-reported expenditure on energy and transport services and consequently assess energy and transport poverty in the same data set. 2. Assess the strength of the causal mechanisms and measure the rela- tionship between energy and transport poverty. 3. Provide an analysis on the effects of the Covid-19 pandemic on re- spondents’ energy and transport use. Once the survey data was processed, it was applied to previously unused energy and transport poverty metrics. These metrics, subjective experiences and the overlap of these conditions are the key knowledge gaps we seek to fill. Furthermore, although some data being collected already exists for Ireland, collecting it again in tandem with data from Northern Ireland allows for an accurate cross-jurisdiction comparison across the Island. 3.1. Expenditure metrics of energy and transport poverty Expenditure metrics can be subcategorised according to the expen- diture used: either modelled or actual spending. The collection of data regarding actual energy and transport expenditure is advantageous as fuel poverty figures in NI are based upon modelled expenditure (i.e., what a household needs to spend, according to a household energy model to maintain a certain heating regime) ignoring actual spending patterns, which are the means of measurement in Ireland. The main drawback of actual expenditure is that it makes it difficult to assess whether a certain level of energy expenditure indicates financial cir- cumstances or deliberate choice of the household (Lowans et al., 2021). Hence, we have also collected data for consensual measures. The mea- sures applied, drawing from the EU Energy Poverty Observatory and other sources, are as follows. • 2Mexp: a household (energy) or individual (transport) can be considered energy/transport poor if expenditure on energy/trans- port exceeds twice the sample median. For energy, this metric may capture households that are energy inefficient and spend an exces- sive amount. However, it may also or instead capture the richest individuals who have the most to spend and may not therefore be limited to its ability to measure energy poverty (Energy Poverty Advisory Hub, 2020). We use this metric for two additional reasons. First, it comprises half of Mattioli’s “Car related economic stress” metric in transport (Mattioli et al., 2016). Second, due to data limi- tations in our survey, we were not able to collect household income data, only individual respondent income. • M/2: a household (energy) or individual (transport) is energy/ transport poor if its absolute energy/transport expenditure (in financial terms) is below half the national median or abnormally low. EnergyPolicy172(2023)1133133 C. Lowans et al. This could be due to high energy efficiency standards but may also be indicative of households that are dangerously under-consuming energy.2 3.2. Consensual measures of energy and transport poverty In addition to expenditure metrics, we used consensual metrics of energy and transport poverty as follows. • Arrears on bills: households that report falling into arrears on their energy (or transport) bills once or more during the past 12 months. This metric is useful for uncovering households that self-report financial difficulties in paying for energy or transport, which may not be revealed by the M/2 indicator (Energy Poverty Advisory Hub, 2020). • Inability to keep warm: households that self-report the inability to keep their home adequately warm when needed are considered en- ergy poor under this metric. This can uncover either financial hardship caused by energy bills or the effects of buildings in poor condition (Energy Poverty Advisory Hub, 2020). • Essentiality of car ownership: an individual is transport poor if they consider a car essential to meet their needs. This borrows from Mattioli’s “Forced Car Ownership” metric, but makes this a consen- sual measure rather than a financial one (Mattioli, 2017). • Adequacy of public transport: as a compliment to the essentiality of car ownership, individuals can be considered transport poor if they do not believe public transport in their area is sufficient to meet their needs. This borrows from research by the Social Exclusion Unit identifying the availability and accessibility of private and public transport to be a barrier to social inclusion (Social Exclusion Unit, 2003). 3.3. The survey instrument The main aim of designing the survey was to achieve empirical novelty rather than aiming for conceptual or methodological novelty, which is becoming an established practice in the social sciences (Sova- cool et al., 2021). As with this prior referenced work, the survey designed and conducted here had no theoretical framework. The aims of the overall project require quantitative data as a deliverable. We did not wish to retrofit our hypotheses to fit the collected data (Sovacool et al., 2021). The questionnaire was designed to take 15–20 min to complete and consisted of 39 questions. The first section assessed the demographics of the respondents. The second section assessed respondents’ attitudes and behaviours regarding domestic energy use. The third section asked re- spondents questions regarding their attitudes and behaviours regarding transport energy use. A mix of answer types were used, ranging from allowing respondents to input numerical values to ranking categorical values. Lastly, some questions were open ended (e.g., allowing re- spondents to describe how they cope with and manage their energy expenditures.) The survey was implemented online by the market research company Dynata, which used a representative respondent panel. Dynata scripted the survey using their software, which the research team checked before being sent to respondents. These re- spondents agreed to participate in Dynata’s respondent panels in return for incentives from Dynata: the researchers had no contact with the respondents and were not involved in providing incentives. All re- spondents were at least 18 and resident in one of the study jurisdictions. A standard data assessment procedure of inspection for incorrect or inconsistent data, cleaning for removal of anomalies, visual inspection 2 Note that we have not used metrics which include a relative threshold for income as we have been unable to collect household incomes, as mentioned above. and verification, and a recording of the changes made to the stored data was followed. A total of 328 respondents were removed based on quality checks. These quality checks included “flat-liners,” i.e., where re- spondents gave straight-line responses on blocks of questions; those who gave incomplete, contradictory, or unrealistic responses; and re- spondents who had unrealistically fast survey completion times. The final sample comprised 1564 respondents, with 431 in Northern Ireland and 1133 in Ireland. These provided a representative sample of each respective jurisdiction and the Island as a whole and are illustrated in Table 1. 3.4. Statistical testing and analysis Our analysis has used multiple methods of testing to determine the strength of relationships between variables. The need for multiple methods arises from the multiple formats of data collected. The methods we use are: linear and logistic regression, Pearson correlation coeffi- cient, chi-square tests, point biserial correlations, Spearman’s rank correlation, and Cramer’s V tests. Regression analysis was carried out on the collected survey data to determine the strength of the drivers of energy and transport poverty. The Pearson correlation coefficient is used to determine the linear cor- relation between data sets. The Chi-square test is used to determine whether there is a statistically significant difference between observed and expected outcomes in categorical variables. Point biserial correla- tion coefficients are used when one variable is binary and the other is continuous. It is equivalent to the Pearson correlation coefficient, which applies to two continuous variables. Spearman’s rank correlation coef- ficient is used to test the strength of the association between two ranked variables, or one ranked variable and one continuous variable. Here we have used it to measure the correlation between 2 binary variables. Cramer’s V test is another test of association based upon the chi-squared test, used to measure the association between nominal variables, and may be used on variables with multiple categories. All significance tests are conducted at the 0.05 level. Depending upon the statistical test used, we calculate significance either as a P- value or with a two-tailed test. 3.5. Demographics of respondents The full demographic profile of our respondents is outlined in Table 1 below. Table 1 shows our respondents’ demographic and socioeconomic profiles, which were ensured to be representative for the Island of Ireland in terms of dwelling type, dwelling tenure, personal income, and location. However, we cannot guarantee representativeness beyond these categories (e.g., educational attainment). The survey was completed by respondents in November 2021, making our results very up to date at the time of publication, albeit preceding 2022 international energy crisis and the consequences of Russia’s invasion of Ukraine (European Commission, 2022) (International Energy Agency, 2022). Note that energy and transport poverty are calculated for each jurisdiction using the median for each jurisdiction, and when displayed together as a rate for the whole Island this is the sum of the number of energy or transport poor for each jurisdiction as a percentage of the sample size. 3.6. Study limitations Overall, we identify three key potential limitations related to surveying as a methodology, namely, the acquiescence bias in responses (Messick, 1966) (Furr, 2011), perceived social desirability of responses (Fisher, 1993) (Huang et al., 1998) and respondent knowledge (Mel- chert, 2011) (van de Mortel, 2008) (Kruger and Dunning, 1999). These will be discussed in turn. The acquiescence bias in responses is a phenomenon exhibited by EnergyPolicy172(2023)1133134 C. Lowans et al. Table 1 Demographic profile of respondents. Demographics Jurisdiction of residence Northern Ireland Republic of Ireland Number of Household inhabitants 1 2 3 4 5 6 7 8 9 Age of respondent 18–24 25–39 40–49 50–59 60–74 75+ Gender of respondent Frequency Percent 431 1133 27.6% 72.4% Frequency Percent 234 444 336 329 131 60 21 7 2 15% 28.4% 21.5% 21% 8.4% 3.8% 1.3% 0.4% 0.1% Frequency Percent 127 548 356 242 253 38 8.1% 35% 22.8% 15.5% 16.2% 2.4% Frequency Percent Male Female Other Is the respondent a member of the Black, Asian or other ethnic minority community 637 922 5 40.7% 59% 0.3% Yes No Area of residence Frequency Percent 112 1452 7.2% 92.8% Frequency Percent Armagh Belfast Derry/Londonderry Lisburn Newry Dublin Cork Limerick Waterford Galway Large Town (18,000 inhabitants to 75,000 inhabitants) Small/Medium Town (4500 inhabitants to 10000 inhabitants) 6 65 8 9 1 136 29 17 13 19 453 342 0.4% 4.2% 0.5% 0.6% 0.1% 8.7% 1.9% 1.1% 0.8% 1.2% 29% 21.9% Intermediate Settlement/Village (1000 inhabitants to 4500 150 9.6% inhabitants) Small Village/Hamlet/Open Country (less than 1000 316 20.2% inhabitants) Respondent employment status Frequency Percent Working full-time Not Working Retired Permanently Sick/Disabled or Looking After Family/Home Working part-time Respondent home ownership status 845 134 195 135 255 54% 8.6% 12.5% 8.6% 16.3% Owned by respondent Rented Social Housing Owned by respondent’s family Respondent dwelling type Bungalow Terraced House Semi-Detached House Detached House Frequency Percent 926 391 79 168 59.2% 25% 5.1% 10.7% Frequency Percent 231 263 466 372 14.8% 16.8% 29.8% 23.8% Table 1 (continued ) Demographics Flat/Apartment Caravan Other Respondent monthly income Northern Ireland [GBP] 0–1000 1001–2000 2001–3000 3001–4000 4001+ Republic of Ireland [EUR] 0–1000 1001–2000 2001–3000 3001–4000 4001+ 215 2 15 13.7% 0.1% 1% Frequency Percent 93 163 107 42 26 Frequency 175 291 354 204 109 21.6% 37.8% 24.8% 9.7% 6% Percent 15.4% 25.7% 31.2% 18% 9.6% respondents whereby they have a tendency to give positive answers to questions, regardless of the content of the question, and do not consider their “true” response (Messick, 1966). Furthermore, a related known contributor to this trend occurs when there is an unbalance of positively or negatively described items in a survey, i.e., a string of positively described items exacerbates tendencies to respond affirmatively (Furr, 2011). The perceived social desirability of responses presents an issue such that respondents may edit their responses in order to be perceived in a more favourable light (Fisher, 1993). This presents an issue in that the reported responses do not reflect respondents’ true behaviour. Indeed the respondents may over or under-report so that their answers could be seen as more moderate than their true response (Huang et al., 1998). Respondent knowledge presents issues in several ways. Firstly, re- spondents unaware of their emotions may not fully understand or be aware of their behaviours or tendencies and so may give misleading answers (Melchert, 2011). Secondly, some respondents may prefer to present a façade rather than be truthful in their responses (note this is related to but not the same as responding in a way the respondent sus- pects would be perceived as more socially desirable) (van de Mortel, 2008). Thirdly, a widely recognised issue is that people hold overly favourable views of their capabilities in many intellectual and social domains, that do not reflect their true capabilities (Kruger and Dunning, 1999). Additionally, the length of the survey instrument may have contributed to some fatigue in responses. This risk was deemed acceptable when considering the survey aim of understanding energy and transport poverty among the same respondent panel. Our survey relies entirely on self-reported data, which can present limitations e.g., some respondents may not accurately recall informa- tion. Moreover, we expected collecting household income data to be highly inaccurate in rented properties, and to have posed ethical ques- tions for respondents who may be unwilling or unable to disclose in- formation regarding others in their household who in turn might not consent. Relatedly, some weakness exists in our methodology in that we have had to minimise the questions asked, and thus the data collected, to measure both energy and transport poverty while not exhausting re- spondents. This has come at the expense of slightly imperfect methods for each of energy and transport poverty measurement. Ideally, we would find household incomes in addition to individual incomes so that we could also run analysis using relative as well as absolute thresholds for our expenditure metrics of energy and transport poverty. In these circumstances, we and others argue that the main drawback of actual expenditure is that it makes it difficult to assess whether a certain level of energy or transport expenditure indicates financial circumstances or deliberate choice of the household. However, we believe the merits of collecting the data to examine the overlap of these conditions outweigh EnergyPolicy172(2023)1133135 C. Lowans et al. the drawbacks of limiting our data collection in each sub-area. When calculating expenditure rates of energy and transport poverty, it was our intention during data collection to account for the support measures that individuals receive. However, when examining these collected data, many erroneous entries were noted e.g., where re- spondents claimed to receive more in supports than is possible. Thus, we omitted all supports data from energy (and transport) poverty expen- diture metric calculations. This has unavoidably affected the calculation of energy and transport poverty rates and possibly also the correlation analysis, but we are unable to say by how much in either case. Finally, due to constraints on the length of this paper, we cannot deeply analyse all 39 questions in this paper or present the entirety of the results. Thus, the results which are most relevant to the research aims are presented alongside the most pertinent analysis. 4. Results & analysis This section presents our results and analysis, beginning with the energy and transport questions, and lastly, their overlap as outlined by our paper aims in Section 3. Note that the abbreviations NI and ROI should be taken to mean Northern Ireland and Republic of Ireland respectively. 4.1. Patterns of energy use and expenditures In this section, we discuss the results and analysis pertinent to re- spondents’ domestic energy use. This pertains to all three of our paper aims: to uncover the extent of energy poverty, assess the causal mech- anisms, and examine the effects of the Covid-19 pandemic on energy usage Ireland wide. Firstly, Fig. 1 and Table 2 below show respondents’ monthly energy bills. The distribution of self-reported energy expenses is displayed in Fig. 1, whilst the median and mean of self-reported energy expenses are listed in Table 2. As can be seen, the median and mean energy bill rose across this time by 14% in NI, 10% in ROI and 17% in NI and 16% in ROI, respectively. In Tables 3 and 4 we show responses to questions concerning thermal comfort and dwelling issues. The first questions deal with consensual Table 2 Median energy expenditures in each jurisdiction. Median Monthly energy Bill before pandemic Median monthly energy bill during the pandemic Mean Monthly energy Bill before pandemic Mean monthly energy bill during the pandemic 150 171 181 212 138 151 177 206 Jurisdiction Northern Ireland [GBP] Republic of Ireland [EUR] Table 3 Respondents’ ability to keep their household comfortably warm when needed. Q13. Can your household keep your home comfortably warm when needed? Frequency Percentage Island wide Northern Ireland Republic of Ireland Yes No Yes No Yes No 1340 224 381 50 959 174 85% 14% 88% 12% 85% 15% Table 4 Showing the percentage of respondents per jurisdiction reporting issues with their dwelling. Do you have any of the following problems with your dwelling/accommodation? Percentage of respondents Island wide Northern Ireland Republic of Ireland A leaking roof Damp walls/floors/foundation Rot in window frames or floor Other structural problem(s) I have none of these problems with my dwelling 7% 17% 8% 8% 71% 5% 13% 5% 4% 80% 7% 18% 8% 10% 68% Fig. 1. Respondent’s energy bills in each jurisdiction, prior to and during the Covid-19 pandemic. Panel A) Northern Ireland, Panel B) Republic of Ireland. EnergyPolicy172(2023)1133136 C. Lowans et al. measures of energy poverty that concern respondents’ self-reported ability to keep their home comfortably warm and the presence of the problems with their dwelling that are known to relate to energy poverty. Under the consensual measure of ability to keep their home comfortably warm, as shown in Table 3, some 14% of respondents can be considered energy poor, whilst 17% of respondents report at least one problem with their dwelling, as shown in Table 4. In Table 5 we correlate answers displayed in Table 3 with expendi- ture metrics of energy poverty. Chi-squared tests and Spearman’s ρ tests show weak, statistically insignificant correlations between these metrics. As stated in section 3.2, a key consensual measure of energy poverty is the presence of arrears on household energy bills (Energy Poverty Advisory Hub, 2020). As shown in Table 6 below, when asked “In the past 12 months, have you been unable to pay for such a heating or electricity bill on time due to financial difficulties?” 10% of respondents reported falling into arrears at least once, and 11% reported falling into arrears twice or more; thus, the rate of falling into arrears is 21%. Table 7 below shows the correlations between the results from Table 6 and financial metrics of energy poverty. Chi-squared tests and Cramer’s V tests show weak, statistically insignificant correlations be- tween the falling into arrears on an energy bill and financial metrics of energy poverty, despite the design aim of this metric; to capture households experiencing difficulties paying for energy due to financial circumstance. The results displayed in Table 8 below pertain to respondents’ heating fuels. This question provides important context as to which fuels the energy poor are consuming, and the ease with which the switch to low carbon options might occur. Regarding heating fuels, when asked what the primary heating fuel of their dwelling is, 38% of respondents reported oil, 33% reported gas, 20% reported electricity, with the remainder made up of coal, peat, other, and “don’t know". Table 9 shows that chi-squared tests show strong associations be- tween primary heating fuel and financial metrics of energy poverty; however, these are statistically insignificant except for the 2Mexp metric for energy expenditure before Covid-19, which is significant; that is that over-expenditure on energy is correlated significantly to heating fuel. Cramer’s V test on the same 2Mexp metric shows a weak but significant correlation. All other associations found by Cramer’s V tests are weak and not statistically significant. Table 10 shows a list of energy technologies and whether re- spondents have these installed. When asked which energy related technologies they have installed at home, or plan to install within the next 12 months, except for low energy lightbulbs, most respondents did not have, nor planned to install solar PV, smart meters, smart appliances, EV chargepoints, or “other”. These results suggest that either re- spondents have limited knowledge that alternative technologies could lower their energy bills, or that they are aware of these opportunities yet have no plans to install such technologies regardless of the potential savings. As shown in Table 11 below, when asked to identify which option would make the greatest difference in meeting their domestic energy needs, 46% of respondents identified lower costs, 26% identified more income, and 25% identified more efficient home and appliances i.e., even though more efficient homes and appliances would decrease Table 5 Correlations between self-reported ability to keep homes warm and financial metrics of energy poverty. Can you keep your household comfortably warm when needed? Chi Squared Spearman’s Rho P value M/2 (pre-Covid-19) M/2 (during Covid-19) 2Mexp (pre-Covid-19) 2Mexp (during Covid-19) 0.066 0.116 0.275 0.036 0.007 0.09 0.013 (cid:0) 0.05 0.797 0.734 0.6 0.849 Table 6 Respondents’ reporting difficulty paying energy bills. In the past 12 months, have you been unable to pay for such a heating or electricity bill on time due to financial difficulties? Percentage of respondents Yes, once Yes, twice or more No Island wide Northern Ireland 10% 11% 79% 8% 6% 85% Republic of Ireland 10% 13% 77% Table 7 Correlations between the inability to pay for an energy bill and financial metrics of energy poverty. Correlations between the inability to pay for an energy bill in the past 12 months and financial metrics of energy poverty Chi Squared Cramer’s V P value M/2 (pre-Covid-19) M/2 (during Covid-19) 2Mexp (pre-Covid-19) 2Mexp (during Covid-19) 1.324 0.175 1.71 2.195 0.029 0.011 0.033 0.037 0.516 0.916 0.425 0.334 Table 8 Respondents’ primary heating fuel. Primary heating fuel Percentage of respondents Oil Gas Electricity Coal Peat Other Don’t know Island wide Northern Ireland Republic of Ireland 38% 33% 20% 3% 4% 2% 1% 52% 35% 10% 2% 1% 2% 33% 33% 32% 24% 3% 5% 2% 1% Table 9 Correlations between primary heating fuel and financial metrics of energy poverty. Correlations between primary heating fuel and financial metrics of energy poverty Chi Squared Cramer’s V P value M/2 (pre-Covid-19) M/2 (during Covid-19) 2Mexp (pre-Covid-19) 2Mexp (during Covid-19) 9.217 6.399 12.369 12.082 0.077 0.064 0.09 0.088 0.162 0.38 0.049 0.06 energy bills, lower unit costs are the more popular means of improving the ability of respondents to meet their needs. In Fig. 2 below, we depict the rates of energy poverty calculated from our data, as per our first paper’s aim. Calculated rates of energy poverty show that under the M/2 metric, both jurisdictions have approximately the same rates of energy poverty. This suggests that the proportion of the population that under-consumes energy is roughly the same in each jurisdiction and has remained approximately constant from before the Covid-19 pandemic to during the pandemic. However, under the 2Mexp metric, over-expenditure on energy is roughly 4% higher in ROI than in NI. Table 12 below shows the correlations between income and financial metrics of energy poverty and illustrates that point biserial correlations uncover no associations between self-reported monthly post-tax income and financial metrics of energy poverty. Furthermore, a binomial regression model using all demographic factors together as predictors could not predict instances of energy poverty. The coefficient results of this model are not presented here to conserve space. EnergyPolicy172(2023)1133137 C. Lowans et al. Table 10 Technologies installed, or planned to be installed at respondents’ dwellings. Technologies installed, or planned to be installed at respondents’ dwellings Percentage of responses Solar panels This is installed at my home This will be installed at my home within the next 12 9% 6% months I have no plans to install this 85% Smart meter 20% 18% 63% Smart appliances (networked) Electric vehicle chargepoint Low energy lightbulbs 16% 15% 69% 4% 8% 88% 71% 11% 18% Other 8% 6% 73% Table 11 Respondents’ perception of item which would make greatest difference to meeting household energy needs. Respondents’ perception of item which would make greatest difference to meeting household energy needs More income A more efficient home and appliances Island wide 26% 25% Lower heating and electricity 46% costs Other 3% Northern Ireland Republic of Ireland 25% 23% 48% 4% 27% 26% 45% 3% 4.2. Patterns of transport use and expenditures In this section we will discuss the results and analysis pertinent to respondents’ transport energy use. This pertains to all three of our paper aims: to uncover the extent of transport poverty, assess the causal mechanisms, and examine the effects of the Covid-19 pandemic on transport usage Ireland wide. Firstly, Fig. 3 and Table 13 below show respondents’ monthly transport bills. The distribution of self-reported transport expenses (comprising expenses on cars, public transport, and taxis) is displayed in Fig. 3, whilst the median and mean of self-reported transport expenses are listed in Table 13. As can be seen, the median and mean transport bill fell across this time by 22% in NI, 27% in ROI and 17% in both NI and ROI respectively. In Tables 14 and 15 we show the results of our questions pertaining to perceptions of different transport modes. We see in Table 14 that across the Island, over 90% of responses say that owning a motor vehicle is essential to fully participate in society. However this contrasts with the results in Table 15, showing that a combined 48% of respondents said that public transport in their area is sufficient to meet most or all of their needs. Table 16 shows the correlations between the perception based re- sponses in Table 14 with financial metrics of transport poverty. Chi- squared tests and Spearman’s ρ tests show no statistically significant correlations between the belief in the necessity of vehicle ownership and financial metrics of transport poverty. Table 17 shows the correlations between the perception based re- sponses in Table 15 with financial metrics of transport poverty. Chi- squared tests and Cramer’s V tests show no statistically significant cor- relations between the belief in the sufficiency of public transport and financial metrics of transport poverty, except for a very weak correlation Table 12 Statistical associations between monthly post-tax income and financial metrics of energy poverty. Statistical associations between monthly post-tax income and financial metrics of energy poverty Point biserial correlation Significance test M/2 (pre-Covid-19) M/2 (during Covid-19) 2Mexp (pre-Covid-19) 2Mexp (during Covid-19) 0.008 0.031 (cid:0) 0.004 (cid:0) 0.005 0.752 0.225 0.877 0.848 Fig. 2. Calculated rates of energy poverty across the Island of Ireland using the M/2 and 2Mexp metrics. Note NI = Northern Ireland. ROI = Republic of Ireland. Island = both. EnergyPolicy172(2023)1133138 C. Lowans et al. Fig. 3. Respondent’s transport bills in each jurisdiction, prior to and during the Covid-19 pandemic. Panel A) Northern Ireland, Panel B) Republic of Ireland. Table 13 Median and mean transport bills before and during the Covid-19 pandemic in NI and ROI. Table 16 Statistical associations between belief in the necessity of owning a vehicle and financial metrics of transport poverty. Jurisdiction Median Monthly transport Bill before pandemic Median monthly transport bill during the pandemic Mean Monthly transport Bill before pandemic Mean monthly transport bill during the pandemic NI [GBP] ROI [EUR] 231 237 174 173 298 318 248 265 Statistical associations between belief in the necessity of owning a vehicle and financial metrics of transport poverty Chi Squared Spearman’s Rho P Value M/2 (pre-Covid-19) M/2 (during Covid-19) 2Mexp (pre-Covid-19) 2Mexp (during Covid-19) 0.587 0.803 0.85 1.669 (cid:0) 0.018 (cid:0) 0.017 (cid:0) 0.020 (cid:0) 0.029 0.498 0.520 0.435 0.268 Table 14 Respondents’ perception of the essentiality of owning their motor vehicle. Respondents’ perception of the essentiality of owning their motor vehicle Island wide Northern Ireland Republic of Ireland Essential Not-essential N/A 85% 7% 3% 85% 5% 4% 85% 7% 3% Table 15 Respondents’ perception of public transport sufficiency. Respondents’ perception of the sufficiency of public transport in their area Sufficient to meet most transport needs Island wide 37% Sufficient to meet all transport 11% needs Not sufficient 52% Northern Ireland Republic of Ireland 43% 11% 46% 35% 11% 55% with the M/2 metric during Covid-19 as shown in Table 17. The results of asking respondents what would aid in meeting their transport needs are shown in Table 18. Namely, 28% say they require more income, 30% say they require lower fuel costs whilst only 16% would like greater public transport provision. The remaining 25% is Table 17 Statistical associations between belief in public transport sufficiency and financial metrics of transport poverty. Statistical associations between belief in public transport sufficiency and financial metrics of transport poverty Chi Squared Cramer’s V P Value M/2 (pre-Covid-19) M/2 (during Covid-19) 2Mexp (pre-Covid-19) 2Mexp (during Covid-19) 4.315 6.782 9.019 1.966 0.053 0.066 0.076 0.035 0.116 0.034 0.011 0.374 divided across increased vehicle efficiency, greater access to vehicles, an increased ability to work from home, increased provision of EV public charging and “other". Regarding the inability to pay for personal car use and public transport, the responses for car and public transport have moved in the same direction following the Covid-19 pandemic, as shown in Table 19. Before the pandemic, 13% of respondents report an inability to pay for their car expenses at least once, but this falls to 10% during the pandemic, perhaps due to decreasing volumes of travel. As for public transport, 8% report an inability to pay for public transport at least once prior to the pandemic, falling to 7% during the pandemic. Chi-squared tests and Cramer’s V tests, as shown in Tables 20 and 21, show no statistically significant correlations between respondents’ inability to pay for any of their means of transport before and during the EnergyPolicy172(2023)1133139 C. Lowans et al. Table 18 Respondents’ perception of item which would make greatest difference to meeting transport needs. Table 21 Statistical associations between inability to pay for transport and financial metrics of transport poverty, during the Covid-19 19 pandemic. Respondents’ perception of item which would make greatest difference to meeting transport needs Statistical associations between inability to pay for transport and financial metrics of transport poverty, during the Covid-19 pandemic Northern Ireland Republic of Ireland Car Chi Squared Cramer’s V P Value More income A more efficient vehicle Lower petrol/diesel/electricity costs Island wide 28% 9% 30% More public transport Owning or having access to more 16% 2% vehicles Ability to work from home more More public electric vehicle 10% 2% charging stations Other 3% 24% 10% 35% 16% 1% 8% 2% 4% 30% 9% 29% 16% 2% 10% 2% 2% Table 19 Respondents’ self reported inability to pay for modes of transport, before and during the pandemic. Respondents’ self reported inability to pay for modes of transport, before and during the pandemic Island wide Northern Ireland Republic of Ireland Personal vehicle – before the pandemic Yes, once Yes, twice or more No I have adapted my behaviour 5% 5% 84% 6% instead of not paying Personal vehicle – during the pandemic Yes, once Yes, twice or more No I have adapted my behaviour 7% 6% 79% 8% instead of not paying Public transport – before the pandemic Yes, once Yes, twice or more No I have adapted my behaviour 3% 4% 88% 5% instead of not paying Public transport – during the pandemic Yes, once Yes, twice or more No I have adapted my behaviour 5% 3% 86% 6% instead of not paying 4% 4% 88% 4% 7% 3% 84% 6% 4% 2% 90% 4% 6% 2% 87% 6% 5% 6% 83% 6% 8% 7% 76% 9% 3% 4% 88% 6% 4% 4% 86% 6% pandemic and financial metrics of transport poverty. In Fig. 4 we show the rates of transport poverty calculated from our data, as per our first aim. Expenditure rates of transport poverty show Table 20 Statistical associations between inability to pay for transport and financial metrics of transport poverty, prior to the Covid-19 19 pandemic. Statistical associations between inability to pay for transport and financial metrics of transport poverty, prior to the Covid-19 pandemic Car Chi Squared Cramer’s V P Value M/2 2Mexp 6.305 2.492 Public transport 0.063 0.04 0.098 0.477 Chi Squared Cramer’s V P Value M/2 2Mexp 6.347 4.803 0.064 0.055 0.096 0.187 M/2 2Mexp 1.019 2.48 Public transport 0.026 0.04 0.797 0.479 Chi Squared Cramer’s V P Value M/2 2Mexp 1.776 4.496 0.034 0.054 0.620 0.213 that under the M/2 metric, both jurisdictions have approximately the same rate of transport poverty during the pandemic, and additionally the rate was lower prior to the pandemic. This suggests that the pro- portion of the population that under-consumes transport is roughly the same in each jurisdiction. As with energy, under the 2Mexp metric, over- expenditure on transport is higher in ROI than in NI, but the difference is small. Table 22 illustrates that, there are no statistically significant corre- lations between monthly post-tax income and financial metrics of transport poverty except for the M/2 metric during the pandemic which is significant, but the association is very weak and negative. Further- more, as with energy poverty, a binomial regression model using all demographic factors together as predictors could not predict instances of transport poverty. The coefficient results are not presented here to conserve space. 4.3. Intersection of energy and mobility poverty In this section, we will discuss the results and analysis concerning the overlap in respondents’ domestic energy and transport use. This pertains to our second aim concerning the overlap of energy and transport poverty, and to our third aim of examining the effects of the Covid-19 pandemic on these issues. Table 23 shows there are statistically significant correlations be- tween all measures of energy and transport poverty and that these are all significant. This is likely explained by a strong association between being not energy poor and not transport poor, i.e., between respondents measuring as a 0 on each of the binary measures. Table 24 shows statistically significant but very weak correlations between monthly energy and monthly transport bills. That is to say that very little of the change in energy bills is explained by a change in transport bills, despite the significance of the finding, with the magni- tude of this influence decreasing by a factor of ten during the Covid-19 pandemic, as would be expected with greatly reduced travel behaviour. 5. Discussion This paper had three aims. First, we sought to record self-reported energy and transport services expenditure and assess consensual and financial metrics of energy and transport poverty. Second, we sought to determine the strength of the causal mechanisms for energy and trans- port poverty and measure the relationship between these conditions. Third and finally, we sought to evaluate the impact of the Covid-19 pandemic on respondents’ energy and transport expenditures. For aims 1 and 3, our results indicate that mean and median energy expenses rose from the period preceding the pandemic to the period during the pandemic, while mean and median transport expenses fell over this period. We found no statistically significant associations be- tween self-reported incomes and energy or transport bills. This is perhaps surprising given that one might expect higher earners to spend more on these services. EnergyPolicy172(2023)11331310 C. Lowans et al. Fig. 4. Calculated rates of transport poverty across the Island of Ireland using the M/2 and 2Mexp metrics. Table 22 Statistical associations between monthly post-tax income and financial metrics of transport poverty. Statistical associations between monthly post-tax income and financial metrics of transport poverty Point biserial correlation Significance test M/2 (pre-Covid-19) M/2 (during Covid-19) 2Mexp (pre-Covid-19) 2Mexp (during Covid-19) (cid:0) 0.024 (cid:0) 0.056 (cid:0) 0.009 (cid:0) 0.032 0.349 0.027 0.731 0.213 Regarding financial metrics of energy poverty and aim 1 of this paper, the rates of energy poverty under the M/2 metric are similar across the Island of Ireland, while overconsumption is higher in Ireland than in Northern Ireland. This pattern holds for financial metrics of transport poverty although the differences are smaller. Possible expla- nations for these patterns include higher wages or higher fuel prices in ROI, but we have not been able to determine the causal mechanism from the data collected. As for consensual energy poverty measures, up to 21% of re- spondents can be considered energy poor. However, no statistical as- sociation was uncovered between financial metrics of energy poverty and the consensual measures. This is surprising, as we would expect to find an association between the M/2 metric, designed to uncover under- consumption of energy, and those who report an inability to keep their home warm or arrears on bills. As for transport poverty, the lack of as- sociation between arrears on bills and the M/2 metric persists. Regarding motor vehicles, 90% of respondents considered owning a Table 23 Statistical associations between metrics of energy and transport poverty. Statistical associations between metrics of energy and transport poverty motor vehicle a necessity, yet 48% of respondents stated that public transport in their area is sufficient to meet most or all of their needs. This result suggests that reasons for owning cars or other motor vehicles extend beyond meeting “needs” and includes wants based on cultural norms and perhaps negative preconceptions concerning public transport (Mattioli et al., 2020). As with energy, there are no meaningful corre- lations between financial and consensual measures of transport poverty. This would suggest that under-expenditure and over-expenditure on transport in our results are distinct from the perceptions of the factors that indicate transport poverty. A consequence of collecting self-reported individual incomes, which we cannot verify, and lacking full household income responses is that we do not have the data needed to determine if energy and transport poverty are distinct from income poverty to corroborate or refute research outlined in section 2 regarding energy poverty in the Republic of Ireland (Watson and Maitre, 2015). However, determining this distinction may not be very useful to the research or policy communities given that our results show that with each change of indicator, there is a change in who is identified as energy or transport poor. This concurs with our earlier research suggesting that there is no single perfect Table 24 Regression analysis results showing the relationship between monthly energy bills and monthly transport bills before and during the Covid-19 19 pandemic. Monthly transport bills Before the pandemic During the pandemic Monthly energy bills R2 0.032 0.0037 P value 0.000 0.000 Transport Energy M/2 Chi Sq 89.035 93.376 4.433 6.681 M/2 M/2 Covid-19 2Mexp 2Mexp Covid-19 P value 0.000 0.000 0.035 0.01 M/2 Covid-19 Chi Sq 92.862 100.216 3.895 6.488 P value 0.000 0.000 0.048 0.011 2Mexp Chi Sq 15.036 18.209 20.128 21.007 P value 0.000 0.000 0.000 0.000 2Mexp Covid-19 Chi Sq 20.106 22.163 19.373 24.996 P value 0.000 0.000 0.000 0.000 EnergyPolicy172(2023)11331311 C. Lowans et al. indicator, nor should one be sought. Rather an appropriate set of in- dicators should be used (Lowans et al., 2021). Given this indicator imperfection, we suggest that more broad approaches should be adopted for identification and alleviation of energy and transport poverty. In policy terms this means identification of the energy poor should continue to be devolved to local governments who know local situations best. The Affordable Warmth Scheme in Northern Ireland is one example of this. The same should apply for transport poverty schemes and criteria for identifying these people should be widened beyond the current stringent conditions. Regarding aim 2 of this paper concerning causal mechanisms, our analysis has not uncovered any statistically significant correlations be- tween the demographic data and the energy and transport poverty re- sults (under expenditure metrics), nor between demographic data and self-reported energy and transport bills. That is to say that vulnerabil- ities known to contribute to energy and transport poverty (such as age, income etc.) do not, according to our results, have a statistical associa- tion with being energy or transport poor. As for the relationship between energy and transport poverty, we have observed statistically significant associations between all financial energy and transport poverty metrics as outlined in Table 23. However, we suspect that what is being iden- tified is the link between being not energy poor and not transport poor as most respondents gave these answers. Concerning energy poverty alleviation, despite a desire for lower domestic energy costs, respondents are broadly unwilling to install new technologies to reduce these costs. Hence, perceived barriers opposing the uptake of such technologies must be considered. A household might, for instance, object to taking out debt to finance a solar PV installation even though the installation would reduce the carbon footprint of the dwelling and possibly lower per-unit energy costs. Although the net financial impact on the household might be positive, the perceived benefits of undertaking the installation may not outweigh perceived costs (Middlemiss and Gillard, 2015). Therefore, for successful out- comes, not only must energy cost burdens be lifted, but also any related barriers must be addressed as well. Lastly, as indicated by low uptake rates, the reliance on market signals to trigger mass retrofits is insuffi- cient, which could be overcome by expanding the groups subject to targeted interventions. The Republic of Ireland has a 2030 climate goal to roll out 2.7 TWh of district heating in cities and to install 400,000 heat pumps in existing homes (Department of the Environment Climate and Communications, 2021). However, at the time of writing, Northern Ireland lacks housing-specific decarbonisation targets (in the form of estimated numbers of installations and retrofits per year) (Northern Ireland Department for the Economy, 2021). Hence, greater efforts are needed to match policy goals with implementation in the Republic of Ireland. As for transport poverty, the most prevalent responses indicated that more income or lower fuel costs would make the most difference. Despite widespread policy recognition that technology and modal shifts in energy and transport, if managed correctly, would benefit consumers, there is very little recognition of this by consumers themselves. This finding, in conjunction with the finding in the energy poverty literature from Middlemiss and Gillard that support schemes should actively seek participants, suggests that there is much more work to be done yet by governments to provide and promote sustainable mobility (Middlemiss and Gillard, 2015). Our finding regarding the perceived need for car ownership suggests once again that much more work is required to reach the Irish Government’s 2030 goal of reducing the amount of “fossil fuelled distance” by 10% (Department of the Environment Climate and Communications, 2021). If most respondents do not believe public transport to be sufficient to meet their needs, they are very unlikely to forego personal vehicles for another transport mode. As with domestic energy, Northern Ireland lacks quantified targets for transport decar- bonisation and modal shift. Lastly, the lack of statistical correlations between our expenditure metrics, causal factors, and consensual metrics highlights the challenge associated with defining energy and transport poverty and categorising those impacted. Our results contradict findings that the drivers of energy and transport consumption are those that are accounted for in housing energy models and vulnerability lenses (e.g., house age or dwelling location). That is, we have found self-reported spending on energy and transport is distinct from expected behaviour, but we cannot determine why this is the case. Furthermore, we have found no discernible single or multiple root causes when examining self-reported energy and transport poverty, nor can we explain why we cannot find these causes. 6. Conclusion and policy implications The first policy implication of our work is that in the absence of revamped national surveying in Northern Ireland to collect actual expenditure alongside modelled data, the focus on modelled expendi- ture data that is currently collected will remain necessary going forward for monitoring energy poverty rates at the national level and should remain in place for consistent monitoring of “need to spend” and for assessing energy performance gap purposes in the future. The second policy implication of our work is that we believe it necessary in both jurisdictions for official transport poverty indicators to be adopted and collected alongside energy poverty indicators to monitor overlaps and characteristics at a national level. A third policy implication is that the energy or transport poor should be anyone identifiable by any of a series of energy or transport poverty indicators or vulnerability lenses, as opposed to stringent targeting criteria. As we have not been able to correlate our findings with the outcomes we expected, we believe further refining of targeted support is a poor policy approach. Indeed, we have noted that support schemes are most effective when they are comprehensive and when local govern- ments proactively reach out to the vulnerable, rather than the other way around. It is widely recognised that national domestic retrofit programs, active travel schemes and improvement of public transport services are among the measures required for meeting decarbonisation targets and at deployment rates exceeding what is currently observed. With continued locally devolved selection of support recipients, the more widely defined and identified energy or transport poor can be the first to access the support schemes or necessary infrastructure. As we have noted, many transport poverty barriers are infrastructural which require solutions in the built environment. Improved and sustainable public transport and active travel schemes should thus be the focus of transport policy efforts. Furthermore, and as noted in the discussion, debt mechanisms are the least attractive means of support for the energy poor (and by analogy, the same could be said of the transport poor). In many cases, re- spondents were not inclined to acquire technologies that may ameliorate their energy and/or transport poverty situation. Therefore, as a final policy implication of this work, and building on other literature, support measures should not pose a debt burden to vulnerable households and should be large enough to enact lasting change rather than merely lessening the financial burden of consumption of contemporary energy and transport services. Regarding furthering the literature, we have two recommendations. First, we recommend similar studies are carried out in other jurisdictions (within and beyond Europe) to explore the reported outcomes further and to see if the reasons for the difference between actual and expected outcomes can be determined, which is a key weakness of our study. Second, we recommend that detailed surveys of vulnerable groups, as identified by vulnerability lenses, are conducted to determine the rea- sons for the difference between actual and expected outcomes, or to see if more focused surveys contradict our findings. We are excited to see the outcomes of such studies regardless of whether they agree or contradict our findings. EnergyPolicy172(2023)11331312 C. Lowans et al. CRediT authorship contribution statement Christopher Lowans: Conceptualization, Methodology, Investiga- tion, Data curation, Software, Formal analysis, Writing – original draft, Writing – review & editing, Project administration. Aoife Foley: Conceptualization, Methodology, Writing – original draft, Writing – review & editing, Funding acquisition. Dylan Furszyfer Del Rio: Conceptualization, Methodology, Writing – original draft, Writing – review & editing. Brian Caulfield: Conceptualization, Methodology, Writing – original draft, Writing – review & editing. Benjamin K. Sovacool: Conceptualization, Methodology, Writing – original draft, Writing – review & editing, Funding acquisition. Steven Griffiths: Conceptualization, Methodology, Writing – original draft, Writing – review & editing. David Rooney: Writing – review & editing. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Data availability The data that has been used is confidential. Acknowledgements Mr Christopher Lowans and Dr Aoife Foley’s research is supported by the Department for the Economy (DfE), Northern Ireland. The views and opinions expressed in this document do not necessarily reflect those of DfE. Dr. Dylan Furszyfer del Rio, Dr Steve Griffiths and Professor Benjamin Sovacool gratefully acknowledge financial support from UK Research and Innovation through the Centre for Research into Energy Demand Solutions, grant reference number EP/R035288/1, as well as Khalifa University of Science and Technology “High Impact Grant." Nomenclature & Abbreviations 2Mexp A household (energy) or individual (transport) is energy/ transport poor if its absolute energy/transport expenditure (in financial terms) is below half the national median European Union EU EU SILC European Union Statistics on Income and Living Conditions EUR EV GBP M/2 Euro Electric Vehicle Pound Sterling A household (energy) or individual (transport) can be considered energy/transport poor if expenditure on energy/ transport exceeds twice the sample median Northern Ireland Republic of Ireland Terawatt Hours NI ROI TWh References Bednar, D.J., Reames, T.G., 2020. Recognition of and response to energy poverty in the United States. Nat. Energy 5, 432–439. https://doi.org/10.1038/s41560-020-0582- 0. Bolton, P., Stewart, I., 2022. Domestic Energy Prices Research Briefing. House of Commons Library (House of Commons Library Research Briefing No. 9491). Bouzarovski, S., Petrova, S., 2015. A global perspective on domestic energy deprivation: overcoming the energy poverty–fuel poverty binary. Energy Res. Social Sci. 10, 31–40. https://doi.org/10.1016/j.erss.2015.06.007. Caulfield, B., Furszyfer, D., Stefaniec, A., Foley, A., 2022. Measuring the equity impacts of government subsidies for electric vehicles. Energy 248, 123588. https://doi.org/ 10.1016/j.energy.2022.123588. Central Statistics Office, 2021. National Travel Survey [WWW Document]. URL, 3.21.2022. https://www.cso.ie/en/methods/transport/nationaltravelsurvey/. Citizensinformation.ie, 2022. Motorised Transport Grant [WWW Document]. URL, 3.26.2022. https://www.citizensinformation.ie/en/travel_and_recreation/transport _and_disability/motorised_transport_grant.html. Crisp, R., Gore, T., McCarthy, L., 2017. Addressing Transport Barriers to Work in Low Income Neighbourhoods: A Review of Evidence and Practice. Sheffield Hallam University. https://doi.org/10.7190/cresr.2017.3465773384. Department for Infrastructure, 2020. Travel Survey for Northern Ireland [WWW Document]. Infrastructure. URL, 3.21.2022. https://www.infrastructure-ni.gov. uk/articles/travel-survey-northern-ireland. Department of the Environment, Climate and Communications, 2021. Climate Action Plan 2021. Government of Ireland. Energy Poverty Advisory Hub, E.C., 2020. Energy Poverty Observatory: Indicators [WWW Document]. URL, 3.10.2022. https://energy-poverty.ec.europa.eu/ene rgy-poverty-observatory/indicators_en. European Commission, 2022. Joint European action for more affordable, secure energy [WWW Document]. Eur. Comm. - Eur. Comm. URL, 4.19.2022. https://ec.europa. eu/commission/presscorner/detail/en/ip_22_1511. Fisher, R.J., 1993. Social desirability bias and the validity of indirect questioning. J. Consum. Res. 20, 303–315. https://doi.org/10.1086/209351. Furr, R.M., 2011. Scale Construction and Psychometrics for Social and Personality Psychology. SAGE Publications Ltd. https://doi.org/10.4135/9781446287866, 1 Oliver’s Yard, 55 City Road, London EC1Y 1SP United Kingdom. Furszyfer Del Rio, D.D., Sovacool, B.K., 2023. Of cooks, crooks and slum-dwellers: exploring the lived experience of energy and mobility poverty in Mexico’s informal settlements. World Dev.. 161, 106093. https://doi.org/10.1016/j.worlddev.2022.10 6093. General Consumer Council Northern Ireland, 2001. THE TRANSPORT TRAP - How Transport Disadvantages Poorer People (No. 2), the Price of Being Poor. Griffiths, S., Furszyfer Del Rio, D., Sovacool, B., 2021. Policy mixes to achieve sustainable mobility after the COVID-19 crisis. Renew. Sustain. Energy Rev. 143, 110919 https://doi.org/10.1016/j.rser.2021.110919. Huang, C.Y., Liao, H.Y., Chang, S.H., 1998. Social desirability and the clinical self-report inventory: methodological reconsideration. J. Clin. Psychol. 54, 517–528 https:// doi.org/10.1002/(sici)1097-4679(199806)54:4<517::aid-jclp13>3.0.co;2-i. International Energy Agency, 2020. World Energy Outlook 2020 464. International Energy Agency, 2022. Russian supplies to global energy markets [WWW Document]. IEA. URL, 4.19.2022. https://www.iea.org/reports/russian-suppl ies-to-global-energy-markets. Kruger, J., Dunning, D., 1999. Unskilled and unaware of it: how difficulties in recognizing one’s own incompetence lead to inflated self-assessments. J. Pers. Soc. Psychol. 77, 1121–1134. https://doi.org/10.1037//0022-3514.77.6.1121. Lowans, C., Furszyfer Del Rio, D., Sovacool, B.K., Rooney, D., Foley, A.M., 2021. What is the state of the art in energy and transport poverty metrics? A critical and comprehensive review. Energy Econ. 101, 105360 https://doi.org/10.1016/j. eneco.2021.105360. Lucas, K., Mattioli, G., Verlinghieri, E., Guzman, A., 2016. Transport poverty and its adverse social consequences. Proc. Inst. Civ. Eng. - Transp. 169, 353–365. https:// doi.org/10.1680/jtran.15.00073. Lucas, K., Philips, I., Mulley, C., Ma, L., 2018. Is transport poverty socially or environmentally driven? Comparing the travel behaviours of two low-income populations living in central and peripheral locations in the same city. Transplant. Res. Part Policy Pract. 116, 622–634. https://doi.org/10.1016/j.tra.2018.07.007. Lunke, E.B., 2022. Modal accessibility disparities and transport poverty in the Oslo region. Transplant. Res. Part Transp. Environ. 103, 103171 https://doi.org/ 10.1016/j.trd.2022.103171. Marmot Review Team, Friends of the Earth, 2011. The Health Impacts of Cold Homes and Fuel Poverty. Friends of the Earth & the Marmot Review Team, London. Mattioli, G., 2017. ‘Forced car ownership’ in the UK and Germany: socio-spatial patterns and potential economic stress impacts. Soc. Incl. 5, 147–160. https://doi.org/ 10.17645/si.v5i4.1081. Mattioli, G., Wadud, Z., Lucas, K., 2016. Developing a novel approach for assessing the transport vulnerability to fuel price rises at the household level. In: Transportation Research Procedia. Presented at the World Conference on Transport Research - WCTR 2016 Shanghai. Elsevier B. V., Shanghai. Mattioli, G., Lucas, K., Marsden, G., 2017. Transport poverty and fuel poverty in the UK: from analogy to comparison. Transport Pol. 59, 93–105. https://doi.org/10.1016/j. tranpol.2017.07.007. Mattioli, G., Roberts, C., Steinberger, J.K., Brown, A., 2020. The political economy of car dependence: a systems of provision approach. Energy Res. Social Sci. 66, 101486 https://doi.org/10.1016/j.erss.2020.101486. Melchert, T., 2011. Foundations of Professional Psychology, first ed. Elsevier B. V. first ed. Messick, S., 1966. The psychology of acquiescence: an interpretation of research evidence. ETS Res. Bull. Ser. 1966 https://doi.org/10.1002/j.2333-8504.1966. tb00357.x i–44. Middlemiss, L., Gillard, R., 2015. Fuel poverty from the bottom-up: characterising household energy vulnerability through the lived experience of the fuel poor. Energy Res. Social Sci. 6, 146–154. https://doi.org/10.1016/j.erss.2015.02.001. Minister for Agriculture, Environment and Rural Affairs, 2022. Climate Change (No. 2) Bill. Northern Ireland Assembly. MORI Scotland, Ipsos, Alembic Research Ltd, 2020. Research into the Lived Experience of Fuel Poverty in Scotland, People, Communities and Places. Scottish Government. Mullen, C., Marsden, G., 2016. Mobility justice in low carbon energy transitions. Energy Res. Soc. Sci., Energy demand for mobility and domestic life: new insights from energy justice 18, 109–117. https://doi.org/10.1016/j.erss.2016.03.026. EnergyPolicy172(2023)11331313 C. Lowans et al. National Audit Office, 2003. Warm Front: Helping to Combat Fuel Poverty - National Audit Office (NAO) Report [WWW Document]. Natl. Audit Off. URL, 3.18.2022. htt ps://www.nao.org.uk/report/warm-front-helping-to-combat-fuel-poverty/. Northern Ireland Department for the Economy, 2021. Energy Strategy - Path to Net Zero Energy. Department for the Economy. Northern Ireland Housing Executive, 2016. House Condition Survey [WWW Document]. URL, 3.18.2022. https://www.nihe.gov.uk/Working-With-Us/Research/House- Condition-Survey. Northern Ireland Housing Executive, 2022. Affordable Warmth Scheme [WWW Document]. URL, 3.26.2022. https://www.nihe.gov.uk/Housing-Help/Affordable -Warmth-Boiler-Replacement/Affordable-Warmth-Scheme. Oireachtas, H. of the, 2021. Climate Action and Low Carbon Development (Amendment) Act 2021 – No. 32 of 2021 – Houses of the Oireachtas [WWW Document]. URL, 4.19.2022. https://www.oireachtas.ie/en/bills/bill/2021/39. Rao, F., Tang, Y.M., Chau, K.Y., Iqbal, W., Abbas, M., 2022. Assessment of energy poverty and key influencing factors in N11 countries. Sustain. Prod. Consum. 30, 1–15. https://doi.org/10.1016/j.spc.2021.11.002. Robinson, C., Mattioli, G., 2020. Double energy vulnerability: spatial intersections of domestic and transport energy poverty in England. Energy Res. Social Sci. 70, 101699 https://doi.org/10.1016/j.erss.2020.101699. Rudge, J., Gilchrist, R., 2005. Excess winter morbidity among older people at risk of cold homes: a population-based study in a London borough. J. Public Health 27, 353–358. https://doi.org/10.1093/pubmed/fdi051. Social Exclusion Unit, 2003. Making the Connections: Final Report on Transport and Social Exclusion [WWW Document]. URL, 3.10.2022. http://www.ilo.org/ emppolicy/pubs/WCMS_ASIST_8210/lang–en/index.htm. Sovacool, B.K., Furszyfer Del Rio, D.D., 2022. We’re not dead yet!“: extreme energy and transport poverty, perpetual peripheralization, and spatial justice among Gypsies and Travellers in Northern Ireland. Renew. Sustain. Energy Rev. 160, 112262 https://doi.org/10.1016/j.rser.2022.112262. Sovacool, B.K., Furszyfer Del Rio, D., Griffiths, S., 2020. Contextualizing the Covid-19 pandemic for a carbon-constrained world: insights for sustainability transitions, energy justice, and research methodology. Energy Res. Social Sci. 68, 101701 https://doi.org/10.1016/j.erss.2020.101701. Sovacool, B.K., Martiskainen, M., Furszyfer Del Rio, D.D., 2021. Knowledge, energy sustainability, and vulnerability in the demographics of smart home technology diffusion. Energy Pol. 153, 112196 https://doi.org/10.1016/j.enpol.2021.112196. Thomson, H., Bouzarovski, S., Snell, C., 2017. Rethinking the measurement of energy poverty in Europe: a critical analysis of indicators and data. Indoor Built Environ. 26, 879–901. https://doi.org/10.1177/1420326X17699260. Torjesen, I., 2022. NHS leaders warn of “intolerable” winter pressure without action on rising energy prices. BMJ 378, o2060. https://doi.org/10.1136/bmj.o2060. Upham, P., Sovacool, B.K., Monyei, C.G., 2022. Energy and transport poverty amidst plenty: exploring just transition, lived experiences and policy implications in Iceland. Renew. Sustain. Energy Rev. 163, 112533 https://doi.org/10.1016/j. rser.2022.112533. Shi, K., Yang, Y., De Vos, J., Zhang, X., Witlox, F., 2022. Income and commute van de Mortel, T.F., 2008. Faking it: social desirability response bias in self-report satisfaction: on the mediating roles of transport poverty and health conditions. Travel Behav. Soc. 29, 297–307. https://doi.org/10.1016/j.tbs.2022.07.004. Simcock, N., Jenkins, K.E.H., Lacey-Barnacle, M., Martiskainen, M., Mattioli, G., Hopkins, D., 2021. Identifying double energy vulnerability: a systematic and narrative review of groups at-risk of energy and transport poverty in the global north. Energy Res. Social Sci. 82, 102351 https://doi.org/10.1016/j. erss.2021.102351. research - southern Cross University. Aust. J. Adv. Nurs. 25, 40–48. Watson, D., Maitre, B., 2015. Is fuel poverty in Ireland a distinct type of deprivation? Econ. Soc. Rev. 46, 267–291. Zhao, J., Dong, K., Dong, X., Shahbaz, M., 2022. How renewable energy alleviate energy poverty? A global analysis. Renew. Energy 186, 299–311. https://doi.org/10.1016/ j.renene.2022.01.005. EnergyPolicy172(2023)11331314
10.1007_s00127-023-02428-w
Social Psychiatry and Psychiatric Epidemiology (2023) 58:569–579 https://doi.org/10.1007/s00127-023-02428-w ORIGINAL PAPER Five‑year illness trajectories across racial groups in the UK following a first episode psychosis Siân Lowri Griffiths1 Linda Everard2 · Peter B. Jones3 · David Fowler4 · Joanne Hodgekins5 · Tim Amos6 · Nick Freemantle7 · Paul McCrone8 · Swaran P. Singh9 · Max Birchwood9 · Rachel Upthegrove1  · Tumelo Bogatsu1 · Mia Longhi1 · Emily Butler1 · Beel Alexander1 · Mrunal Bandawar1 · Received: 2 May 2022 / Accepted: 12 January 2023 / Published online: 30 January 2023 © The Author(s) 2023 Abstract Purpose Psychosis disproportionally affects ethnic minority groups in high-income countries, yet evidence of disparities in outcomes following intensive early intervention service (EIS) for First Episode Psychosis (FEP) is less conclusive. We investigated 5-year clinical and social outcomes of young people with FEP from different racial groups following EIS care. Method Data were analysed from the UK-wide NIHR SUPEREDEN study. The sample at baseline (n = 978) included White (n = 750), Black (n = 71), and Asian (n = 157) individuals, assessed during the 3 years of EIS, and up to 2 years post- discharge (n = 296; Black [n = 23]; Asian [n = 52] and White [n = 221]). Outcome trajectories were modelled for psychosis symptoms (positive, negative, and general), functioning, and depression, using linear mixed effect models (with random intercept and slopes), whilst controlling for social deprivation. Discharge service was also explored across racial groups, 2 years following EIS. Results Variation in linear growth over time was accounted for by racial group status for psychosis symptoms—positive (95% CI [0.679, 1.235]), negative (95% CI [0.315, 0.783]), and general (95% CI [1.961, 3.428])—as well as for functioning (95% CI [11.212, 17.677]) and depressive symptoms (95% CI [0.261, 0.648]). Social deprivation contributed to this vari- ance. Black individuals experienced greater levels of deprivation (p < 0.001, 95% CI [0.187, 0.624]). Finally, there was a greater likelihood for Asian (OR = 3.04; 95% CI [2.050, 4.498]) and Black individuals (OR = 2.47; 95% CI [1.354, 4.520]) to remain in secondary care by follow-up. Conclusion Findings suggest variations in long-term clinical and social outcomes following EIS across racial groups; social deprivation contributed to this variance. Black and Asian individuals appear to make less improvement in long-term recov- ery and are less likely to be discharged from mental health services. Replication is needed in large, complete data, to fully understand disparities and blind spots to care. Keywords Outcomes · Early psychosis · Ethnicity · Deprivation · Inequities Max Birchwood and Rachel Upthegrove shared joint senior authorship. * Siân Lowri Griffiths s.l.griffiths@bham.ac.uk 1 Institute for Mental Health, University of Birmingham, Edgbaston, Birmingham B15 2TT, UK 2 Birmingham and Solihull Mental Health Foundation Trust, Birmingham, UK 3 Department of Psychiatry, University of Cambridge and CAMEO, Cambridge and Peterborough NHS Foundation Trust, Fulbourn, UK 4 Department of Psychology, University of Sussex, Brighton, UK 5 Norwich Medical School, University of East Anglia, Norwich, UK 6 Academic Unit of Psychiatry, University of Bristol, Bristol, UK 7 8 Institute of Clinical Trials and Methodology, University College London, London, UK Institute for Life Course Development, University of Greenwich, London, UK 9 Mental Health and Wellbeing Warwick Medical School, University of Warwick, Coventry, UK Vol.:(0123456789)1 3 570 Introduction The incidence of psychosis disproportionally affects eth- nic minority groups in high-income countries [1]. Black Caribbean individuals are five times more likely to develop psychosis in the UK, compared to the White British popu- lation, but such incidence rates are not mirrored in Carib- bean countries [2]. Further, for individuals of Pakistani, Bangladeshi, or of Mixed ethnic backgrounds in England, the incidence rates are twice as high compared to the White population [1]. It is well documented that inequalities exist in access to mental health care, for example, Black individuals are more likely to experience adverse pathways to care [3]. Differences may also exist in the type of care offered and received by ethnic minorities within mental health ser- vices. Black Caribbean and Black African individuals with psychosis are 15–30% less likely to receive Cogni- tive–Behavioural Therapy (CBT) compared to White indi- viduals with psychosis in the UK [4]. National clinical audit data has also highlighted inequalities in the offer of clozapine; Black individuals are up to 44% less likely to be offered this evidenced-based medication for treatment- resistant psychosis [5]. Despite this, less robust and consistent research has been carried out on the impact of this potential disparity on course and outcome of psychosis [6, 7]. A systematic review has provided evidence that migrant groups are more likely to achieve remission but have higher rates of involuntary admission and disengagement compared to host populations [8, 9]. In studies comparing outcomes across ethnic groups, poorer social and clinical outcomes for Black individuals are reported compared to White indi- viduals [6, 7, 10–15]. For other racial groups, outcomes have been reported to be more benign. For example, in an exploratory study by Birchwood et  al., relapse and readmission rates were the highest for Black Caribbean individuals, and lowest in those of Asian heritage, when compared to White British individuals [10]. Family struc- ture, quicker access to care, and employment status have been proposed to mitigate these effects [10]. However, inconsistencies and methodological constraints, such as small sample size, high attrition, short follow-up, and ret- rospective designs, make it difficult to draw on conclusions regarding any differences in clinical and social outcomes for ethnic minority individuals with psychosis, and ques- tions remain over why such differences exist [7, 12]. In a more recent longitudinal study, the AESOP-10 cohort study investigated ethnic disparities in illness out- comes between Black minority ethnic and White British individuals 10 years following a first episode psychosis (FEP) [6]. Compared to the White British group, the Black Social Psychiatry and Psychiatric Epidemiology (2023) 58:569–579 Caribbean group had poorer clinical, social, and service use outcomes. There was also some initial evidence sug- gesting social disadvantage and isolation contributed to the differences in symptoms and social outcomes [6]. It is important to understand malleable factors and social inequalities related to illness incidence, but also whether underlying factors continue to drive enduring impair- ment and poorer outcomes. We aim to extend the evidence on social inequalities and ethnic variation in outcomes after FEP, using large, national longitudinal dataset of patients receiving gold standard EIP care. We wish to establish whether: (1) Black and Asian racial minority individuals with FEP differ in their long-term symptoms (psychosis symptoms and depression) and func- tional outcomes, compared to White individuals; (2) social deprivation contributes to later clinical and social outcomes across racial groups; (3) discharge services 2 years after EIS, differ by racial group. Method Study design This was a secondary analysis of the National Evaluation of the Development and Impact of Early Intervention Services (NEDEN) study, a prospective longitudinal study of young people with a first episode of psychosis (FEP), across 14 early intervention services (EIS) in the UK [16]. The details of the original study methodology are reported elsewhere [16], but in brief, participants were initially recruited and assessed over the first 12 months of service as part of the NEDEN study. SUPEREDEN is the follow-on study, pro- spectively assessing the same cohort of individuals up until discharge from EIS (approximately at 3 years from baseline), and then up to 2 years post-discharge from EIS. Individuals with lived experience were involved in the study implemen- tation and delivery and were regularly consulted throughout the SUPEREDEN project. Sample The initial sample had a total of 978 participants. Participants were aggregated into 3 racial groups: a Black minority racial group (n = 71; 6.9%), Asian minority racial group (n = 157; 15%), and a White racial group (n = 750; 73%). The Black racial minority group included individuals who identified as Black Caribbean, Black African, and Black ‘other’. The White group included participants who identified as White British, White Irish or White ‘other’. The Asian group included par- ticipants who identified as Pakistani, Bangladeshi, Indian, or other Asian background. Participants met diagnostic criteria outlined in International Classification of Diseases under the 1 3 Social Psychiatry and Psychiatric Epidemiology (2023) 58:569–579 571 following codes: F20, F25, F29, F31, F32–F32.1, and F32.3 [17]. Written and verbal consent was obtained for all partici- pants. Ethical approval was given by Suffolk Local Research Ethics Committee, UK. REC reference number: 05/Q0102/44. characteristics between racial groups at baseline and final follow-up (approximately 5 years from baseline). Model building Measures Outcome variables Assessments were undertaken by research assistants who were trained and had no clinical involvement with the partic- ipants. A robust reliability protocol is detailed in the original research [16]. The following measures were used to assesses outcomes: Positive and Negative Syndrome Scale (PANSS) [18], Calgary Depression Scale for Schizophrenia (CDSS) [19], Global Assessment of Functioning Disability Scale (GAF Disability) [20], and Duration of Untreated Psycho- sis (DUP) [21]. Covariate: social deprivation A social deprivation proxy was derived at each time point by summing the presence of the following demographic factors: (1) unemployed, (2) single marital status, (3) living alone, and (4) living in temporary/supported accommodation or social housing, with each of these factors being assigned a score of 1 if present (maximum score = 4). A score of ‘1’ for living alone may also be indicative of financial stability or independence; however, a high score on our proxy measure (i.e., score of 4) is within the context of being unemployed, single and in supported or temporary accommodation, and hence more likely to signify social deprivation. To validate the summation of these items, reliability sta- tistics were inspected. Given that reliability coefficients such as Cronbach’s alpha are sensitive to the number of items in a scale and often lower with a smaller number of items, we interpreted this coefficient alongside the optimal mean inter- item correlations (r = 0.2–0.4), and explored the dimension- ality of the data using a factor analysis [22, 23]. Correlations between items were significant (p < 0.01), and the mean inter-item correlations fell within the recommended range (r = 0.366), with a Cronbach’s alpha of 0.69 (Supplementary Material 1) [24]. The exploratory factor analysis confirmed the uni-dimensionality of the data, with all items loading strongly on a single component (Supplementary Material 1). Statistical analysis Descriptive statistics Chi-square tests for categorical, and between Analy- sis of Variance (ANOVA) tests for continuous variables were performed on the demographic, clinical, and social To determine the longitudinal relationship between ethnic status and clinical and social outcomes, hierarchical lin- ear mixed effect models were constructed within Statisti- cal Package for the Social Sciences (SPSS v.25). Multi- level models were constructed in the following manner for PANSS Positive, PANSS Negative, PANSS General, GAF Disability, and Calgary Depression. At level 1, fixed and randomly varying time components were added to the model to examine the rate of change on the outcome for partici- pants across the 5-year study period. Graphs were initially inspected to provide an indication of the shape of the growth trajectory alongside model fit indices to determine which rate of growth provided the best model fit. Lower scores on the Schwartz’s Bayesian Criterion indicated that a lin- ear time component (coded as 0 for baseline and 1–4 for subsequent follow-ups) provided better model fit for each outcome and was therefore used to model the growth tra- jectories (Supplementary Material 2). At level 2, race was added to the covariance model to see if any variation in the (random) time slopes and intercepts for each of the outcomes were accounted for by racial group (Supplementary Mate- rial 2). At level 3, a social deprivation proxy was added as a covariate to determine its influence on the outcome when all variables were added (and controlled for) in the model. Models were estimated using a restricted maximum-likeli- hood (REML) method. REML was selected as it provides unbiased parameter estimates and is robust against large missing data and unbalanced designs [25–27]. Simulation studies have demonstrated that using REML to estimate the linear mixed models is preferable to multiple imputation for handling missing data when the mechanisms of missingness is assumed to be random; data imputation introduces greater noise into the models, rendering them more unstable [28, 29]. Results of the missing data analyses are reported on page 9. Finally, a diagonal covariance structure was used for the repeated and random effects which assumes heter- ogenous variances and no correlation between any of the elements [27]. Discharge services A binary logistic regression was employed to explore the discharge destinations of the racial minority groups com- pared to the White racial group, 2 years following discharge from EIS. The binary outcome was coded as ‘1’ for sec- ondary care (i.e., specialist mental health service support), or a ‘0’ for primary care (i.e., non-specialist community care from a general physician, on a needs basis). Electronic 1 3 572 Social Psychiatry and Psychiatric Epidemiology (2023) 58:569–579 medical record data were accessed for this part of the analy- sis meaning that more complete (85.2%) data were obtained (n = 833; Black = 57; Asian = 147; White = 628). to be missing at random. A restricted maximum-likelihood method (REML) was considered appropriate to fit the linear mixed models [25, 28]. Results Sample description Missing data At baseline, outcome data were available for n = 912 partici- pants (male = 632, 69.3%; mean age = 21.9 years), with an average retention rate of 33% (n = 296) by the final follow-up (5 years from baseline). This included data on 34% (n = 23) of the Black racial group, 37% (n = 52) of the Asian group, and 32% (n = 221) of the White group, by year 5. The great- est attrition was observed when participants reconsented into the SUPEREDEN study (Supplementary Material 3). To determine any bias in the patterns of missingness on the outcome variables, we conducted an exploratory analysis comparing individuals who remained in the study compared to those who did not. We did not find significant differences on any of the outcome measures at baseline (Supplementary Material 4), and there were no differences by racial group (X2 = 1.165, p = 0.559). We therefore assumed that missing- ness was not related to the outcomes of interest, and likely Demographic and clinical characteristics At baseline and at 2 years post-discharge, there was a higher frequency of individuals within the Black racial group who were living alone, single, and living in temporary or sup- ported accommodation (Table  1). They were also more likely to be unemployed at baseline, but there were no sig- nificant differences by follow-up. There were no significant differences in qualifications levels across racial groups. The clinical characteristics of the sample are provided in Table 2. There were no significant differences across the groups with age of onset; however, the White group had a significantly longer median DUP, and a significantly higher percentage of the White racial group had reported self-harm and used cannabis persistently. There were no significant dif- ferences between the racial groups on medication adherence and prescriptions of clozapine or psychological therapies (Table 2). Over the follow-up period, the Black racial group had a higher average score on our proxy measure of depri- vation (b = 0.406, p < 0.001, 95% CI [0.187, 0.624]), whilst the Asian group had a lower score (b = − 0.322, p < 0.001, 95% CI [− 0.477, − 0.168]) compared to the White group. Table 1 Demographic breakdown of racial groups at baseline and 2 years post-discharge from early intervention service Black N = 71 Asian N = 157 White N = 750 66 (93%) 23 (79.3%) 15 (21%) 12 (42.9%) 52 (73%) 21 (75%) 18 (27%) 32 (48%) 17 (25%) 4 (6%) 40 (62%) 25 (36%) 34 (49%) 11 (15%) 115 (73%) 25 (43.9%) 660 (88%) 196 (71%) 6 (3.8%) 5 (9.3%) 102 (65%) 34 (63%) 42 (28%) 59 (39%) 39 (26%) 12 (8%) 121 (79%) 117 (76%) 24 (16%) 13 (8%) 106 (14.1%) 80 (31.9%) 419 (56%) 149 (59.8%) 168 (23%) 293 (40%) 192 (26%) 79 (11%) 726 (97%) 396 (56%) 250 (35%) 66 (9%) Statistical signifi- cance p < 0.001 p < 0.001 p < 0.001 p < 0.001 p < 0.001 NS NS p < 0.001 p < 0.001 Single marital status  Baseline 2 years post-discharge Living alone  Baseline  2 years post-discharge Unemployed  Baseline  2 years post-discharge Qualifications  None  GSCE/NVQ  A-level/BTEC  Degree Place of birth: UK Housing type  Owned/parents own  Rented  Temporary or supported NS non-significant 1 3 Social Psychiatry and Psychiatric Epidemiology (2023) 58:569–579 Table 2 Clinical characteristics across racial groups Black N = 71 Asian N = 157 White N = 750 573 Statistical signifi- cance Presentation factors  Delay of untreated psychosis (weeks; median)a  Age of onset (years; mean/SD) Ongoing factors Cannabis useb (persistent) Self-harm (n; %)c Yes/no Treatment factors  Medication Non-adherenced Clozapinee Psychological therapyf 6.43 8.64 12.71 p < 0.05 21.72 (4.7) 21.05 (4.17) 21.4 (5.17) NS 3 (4.2%) 1; 53 (1.9%) 8 (5.5%) 5; 117 (4.1%) 107 (14.7%) 85; 510 (14.3%) p < 0.001 p < 0.002 10 (14.1%) 22 (14%) 89 (11.9%) 2 (2.8%) 13 (18.3%) 8 (5.1%) 28 (17.8%) 14 (1.86%) 125 (16.7%) NS NS NS NS non-significant a Independent median test b Persistent cannabis use = continued cannabis use over 12 months derived from the Drug Check [30] c Client reported self-harm; any incidence of self-harm over the initial 12 months of treatment d Medication adherence derived as an average score from the clinician-rated ‘Service Engagement Scale’ [31] e Prescribed clozapine within the first year of EIS treatment f Received an individualised form of therapy, e.g., cognitive behavioural therapy across the full study period Table 3 Linear mixed model fixed effects analysis of recovery out- comes over the 5-year study period Beta SE p value Lower-95 Upper-95 PANSS positive − 0.502 PANSS negative − 0.335 PANSS general − 1.037 CDSS GAF disability 0.077 < 0.001 − 0.652 0.071 < 0.001 − 0.474 0.126 < 0.001 − 1.284 − 0.473 − 0.065 < 0.001 − 0.600 1.582 0.243 < 0.001 1.104 − 0.352 − 0.197 − 0.789 − 0.345 2.060 PANSS Positive and Negative Syndrome Scale, CDSS Calgary Depression Syndrome for Schizophrenia Racial group differences on recovery outcomes Linear time effect (level 1) Over the follow-up period, there were significant main effects of time for PANSS positive, PANSS negative, PANSS general, and Calgary Depression, with symp- toms decreasing on average over the follow-up period. GAF disability scores on average increased over the study period, with higher scores indicating improved function- ing (Table 3). Illness trajectories and race (level 2) The random covariance analysis indicated significant vari- ation in the intercepts and linear slopes across the racial groups for PANSS positive (b = 0.140; 95% CI [0.679, 1.235]), negative (b = 0.497; 95% CI [0.315, 0.783]), and general symptoms (b = 2.593; 95% CI [1.961, 3.428]), as well as GAF disability (b = 14.078, 95% CI [11.212, 17.677]) and depression (b = 0.684; 95% CI [0.261, 0.648]). The growth trajectories are summarised in Table 4, and visu- alisation of the trajectories are provided in Figs. 1, 2, 3, 4 and 5 (see also Supplementary Material 5 for means and standard deviations). Steeper slopes were observed for the White racial group. The Black group showed no significant variation in growth for PANSS positive and Calgary Depres- sion. Lower symptom scores were observed for the Black group at baseline, whilst the White group had higher scores, except for negative symptoms, where the Asian group were observed to have higher scores at baseline. Social deprivation, race, and outcome (level 3) A ‘social deprivation proxy’ was added as a covariate in the linear mixed models for each of the outcome variables described above. Social deprivation proxy score significantly contributed to variance in outcomes across the racial groups. Higher scores on the social deprivation proxy was associated 1 3 574 Social Psychiatry and Psychiatric Epidemiology (2023) 58:569–579 Fig. 1 Graphs depicting illness trajectories across the racial groups on PANSS positive over the 5-year follow-up Fig. 4 Graphs depicting illness trajectories across the racial groups on CDSS depression symptoms over the 5-year follow-up Fig. 2 Graphs depicting illness trajectories across the racial groups on PANSS negative over the 5-year follow-up Fig. 5 Graphs depicting illness trajectories across the racial groups on GAF disability over the 5-year follow-up with higher PANNS positive scores (b = 0.710, SE = 0.103, p < 0.001, 95% CI [0.509, 0.912]), PANSS negative scores, b = 0.875, SE = 0.106, p < 0.001, 95% CI [0.667, 1.083]), PANNS general score (b = 1.390; SE = 0.174, p < 0.001, 95% CI [1.050, 1.731]), and Calgary depression scores (b = 0.455, SE = 0.099, p < 0.001, 95% CI [0.261, 0.648]). Finally, a higher social deprivation score on our proxy measure was associated with lower GAF scores (b = − 5.116, SE = 0.328, p < 0.001, 95% CI [− 5.758, − 4.473]). Discharge trajectories A binary logistic regression comparing discharge services across racial groups 2 years following EIS, showed that, compared to their White counterparts, there was a greater likelihood for the Asian (OR = 3.04; 95% CI [2.050, 4.498]; p = < 0.001) and Black racial group (OR = 2.47; 95% CI Fig. 3 Graphs depicting illness trajectories across the racial groups on PANSS general symptoms over the 5-year follow-up 1 3 Social Psychiatry and Psychiatric Epidemiology (2023) 58:569–579 575 Table 4 Linear growth parameters across racial groups over time for each of the outcome variable Beta Wald Z p value Lower-95 Upper-95 PANSS positive   Intercepta  Group × timeb   Black   Asian   White PANSS negative   Intercepta  Group × timeb   Black   Asian   White PANSS general  Group × timeb   Black   Asian   White CDSS Intercepta Group × timeb   Black   Asian   White GAF disability   Intercepta  Group × timeb   Black   Asian   White 4.824 4.894 < 0.001 3.232 7.200 0.706 1.372 1.121 NS 1.935 3.4943 < 0.001 < 0.001 6.656 0.257 0.783 0.835 1.944 2.406 1.505 7.275 6.568 < 0.001 5.398 9.805 1.541 0.929 0.609 11.554a 2.988 2.695 2.944 2.500 2.838 4.487 4.381 2.418 3.175 6.784 0.0124 0.0045 < 0.001 < 0.001 0.704 0.466 0.394 7.387 0.016 0.002 < 0.001 1.329 1.454 2.205 3.374 1.85 0.943 18.073 6.721 4.997 3.930 6.1502a 8.770 < 0.001 4.9185 7.690 0.074 0.434 0.812 0.424 2.214 6.406 NS 0.033 < 0.001 0.001 0.1791 0.598 7.561 1.053 1.103 81.516 9.351 < 0.001 66.103 100.524 8.537 12.125 14.933 0.0471 1.986 3.530 < 0.001 7.8943 < 0.001 3.182 6.959 11.650 22.909 21.127 19.142 a Random covariance parameter for the intercepts across racial groups b Random covariance slope parameter for time × racial group NS non-significant at 0.05 alpha level [1.354, 4.520]; p = < 0.001) to remain in secondary care (i.e., treatment within mental health services) by follow-up. Discussion In this large, prospective FEP cohort, recovery outcomes significantly improved across the follow-up period, which included the duration of EIS care and up to 2 years post-discharge. The rate of improvement varied by racial group, with the White group showing more growth in their recovery trajec- tories. Social deprivation further contributed to this variance in growth across racial groups. Two years following EIS care, the Asian and Black individuals were less likely to be discharged from mental health services. To our knowledge, this is the first study to report long- term outcomes across different racial minority groups fol- lowing EIS care [6, 11–15]. Our findings hint at the potential compounded impact of the intersectional challenges of racial minority status and deprivation [6, 12, 32–34]. However, our findings are nuanced; deprivation was not uniform across minority racial groups. The Black group had significantly greater levels of deprivation, whilst the Asian group experi- enced less social deprivation over the study period. Despite improving more, the White group typically had similar levels of symptoms to the minority racial groups by follow-up, possibly suggesting a ceiling effect in recovery trajectories for the minority racial groups. This was likely the case for the Black group who showed no change in growth over time on the Calgary Depression Scale, but this was in the context of low, stable symptoms across the time frame. Similarly, self-harm was less frequent in the Black group; a finding supported by previous research [6]6. Nevertheless, we showed that minority individuals were more likely to be receiving mental health treatment follow- ing discharge from EIS, suggesting that they may not have achieved the same level of recovery as their White counter- parts. This possible enduring nature of psychosis for minor- ity groups would support the work of Morgan et al., where Black individuals were more likely to have a continuous, non-remitting illness course, as opposed to an episodic tra- jectory [6]. Confounding treatment factors It is well documented that a prolonged delay of untreated psychosis (DUP) is associated with poorer recovery out- comes [35, 36]. We did not find a longer DUP for racial minority groups. Instead, the White group had a significantly longer DUP; a finding supported by other studies [37–40]. This may account for the higher symptom scores for the White group at baseline, yet the White group typically showed more growth in their trajectories over time. This raises the question as to why the trajectories of the racial minority groups may appear less responsive to the support offered within current service models. Linked to this notion, our initial inspection showed no differences in treatment factors that are likely to influence recovery outcomes, such as medication adherence, treat- ment with clozapine, and receiving a psychological therapy [41–43]. Further, we found no differences in age of onset of illness, but there were significant differences in persistent cannabis use, which was more frequent in the White group. However, as previously reported by the EDEN consortium, 1 3 576 Social Psychiatry and Psychiatric Epidemiology (2023) 58:569–579 the influence of cannabis on poor outcomes was shown to be independent of ethnic status [44]. Proposed mechanisms driving inequalities in outcomes Socio-economic status, experiences of racism, linguistic dis- tance, and social exclusion and discrimination may lead to a psychological ‘disempowerment’ [45] or ‘social defeat’ [46]. Such processes are likely to play an important role in the aetiology and pathogenesis of psychosis [34, 46, 47]. Indeed, in our study, for the Black racial group, deprivation was already apparent at baseline, likely reflecting a longstand- ing trajectory of deprivation. This not only exposes these individuals to psychotic illness, but is likely to be mutually reinforcing, where psychosis symptoms drive further depri- vation and exclusion, and vice versa, resulting in enduring impairment, marginalisation, and further feelings of disem- powerment [6, 33, 48]. On the other hand, the Asian minority group expe- rienced less deprivation compared to the other groups, which suggests that other factors are also likely to play a part. Indeed, previous studies have reported racial-ethnic differences in receiving evidence-based interventions and family psychoeducation once in treatment following a first episode of schizophrenia [49]. Compulsory treatment is also frequently reported [9]. Themes of mistrust in services, stigma, and coerciveness have also featured in the narra- tives of Black and minority individuals receiving mental health treatment [50]. Thus, treatment trajectories, including pathways out of EIS, warrant in-depth exploration, particu- larly as our racial minority groups were less likely to be discharged out of mental health service 2 years following EIS discharge. The lived experiences of these individuals will be essential to fully understand the processes behind these disparities. Strengths and limitations There are several strengths to this study. The EDEN studies comprised a large, perspective cohort of participants who had experienced FEP across distinct and varied geographi- cal areas in England, making it representative of the UK’s diverse population, but also representing socioeconomic variability. We add to past literature by further including a comparison with individuals of Asian heritage, which has not been robustly reported within the literature. Finally, we explore a range of outcome variables and model the hetero- geneity in illness trajectories across the duration of EIS care and the subsequent 2 years following discharge. However, there are important study limitations to consider. First, whilst over a thousand participants originally con- sented to the study, our target minority racial groups were substantially smaller, reducing our statistical power. Given the high prevalence of psychosis within ethnic minority groups in high-income countries, our small group size in this study may reflect lack of engagement of minority individu- als in research, thus placing limit on the representativeness of our findings and potentially biasing the sample. Second, there were high levels of attrition across each time point, potentially introducing bias in our findings. We were, how- ever, able to demonstrate that missingness did not differ by racial group, and there were no differences by racial group on the main outcomes at baseline for those who continued in the study compared to those who dropped out. In such situations where mechanisms of missingness are assumed to be random, the REML algorithm (used within the analy- sis) is shown to be robust to large missing data and unbal- anced designs [25, 28, 29]. Third, for reasons of statistical power, we were not able to explore intergroup differences. For example, there is evidence pointing to differential out- comes in Black Caribbean, as opposed to Black African individuals [6, 51]. We also did not include a mixed racial group in our analysis because of the limited sample size; this should be investigated further. Finally, as this was a second- ary analysis of existing data, this restricted our examina- tions into other potential factors influencing the observed differences. This also meant that a proxy estimate was used to quantify social deprivation. Future research may wish to build on these findings using a more robust measure of social deprivation, which also considers the premorbid levels of deprivation, compared with the deprivation synergistically linked to psychosis. Implications and future directions Methodological issues place limit on how much we can extrapolate our findings, but they nevertheless add to a growing body of research indicating differential outcomes for racial minorities recovering from a first episode psycho- sis. In addition to replication, further research is also needed to understand the key drivers of these disparities that may serve as pivotal points for intervention. Our findings may suggest wider contextual and societal factors feeding into illness trajectories. Systemic barriers and social structures inherent within our society are likely to permeate into health care and place limit on one’s outcome. Breaking this cycle should not only be a priority for EIS, but a shared priority for public health and social policy [6]. There is growing interest looking into area-level inter- ventions to mitigate the psychological consequences of belonging to a disempowered minority group. For example, increasing access to social capital is proposed to dampen the social stress associated with deprivation and discrimination, and thus foster an environment that is more conducive to 1 3 Social Psychiatry and Psychiatric Epidemiology (2023) 58:569–579 577 recovery [11, 52, 53]. Though promising, implementing such interventions is complex given their nuanced and context- dependent nature [54]. At a service level, there may be a need to develop clinicians’ cultural competencies, in addi- tion to offering culturally sensitive interventions to improve service provision for underserved groups. Co-produced work will be an important step towards achieving this goal [55]. Finally, exploring the disempowerment experienced by such individuals may also be an important target for clinical inter- vention [47]. adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/. Conclusion In a large FEP cohort, our findings suggest variations in long-term clinical and social outcomes following EIS for racial minority groups. Social deprivation contributed to this variance, with Black individuals experiencing the most dep- rivation. Black and Asian individuals were also less likely to be discharged from mental health service by follow-up. Though replication is needed, our findings hint at the need for targeted, and culturally sensitive service provision, that mitigates the impact of discrimination and deprivation and promotes long-term recovery following FEP. Supplementary Information The online version contains supplemen- tary material available at https:// doi. org/ 10. 1007/ s00127- 023- 02428-w. Acknowledgements M.B. and S.P.S are part funded by the National Institute for Health Research through the Applied Research Collabo- ration West Midlands (ARC-WM). P.B.J. is part funded by the NIHR ARC East of England. The views expressed in this publication are those of the authors and not necessarily those of the NHS, NIHR, or Department of Health. Birmingham and Solihull NHS Foundation Trust acted as study sponsor. We would like to thank the participants of the National EDEN study and the UK Clinical Research Network for study support. Data Availability The datasets generated during and/or analysed dur- ing the current study are not publicly available under current ethical approvals but are available from the corresponding author on reason- able request. Declarations Conflict of interest RU reports grants from Medical Research Council, grants from National Institute for Health Research: Health Technol- ogy Assessment, grants from European Commission—Research: The Seventh Framework Programme, and personal fees from Sunovion, outside the submitted work. Ethical standards Ethical approval was given by Suffolk Local Research Ethics Committee, UK, in accordance with the ethi- cal standards laid down in the 1964 Declaration of Helsinki and its later amendments. Research Ethics Committee reference number: 05/ Q0102/44. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, References 1. Kirkbride JB et al (2012) Incidence of schizophrenia and other psychoses in England, 1950–2009: a systematic review and meta- analyses. PLoS One 7:e31660. https:// doi. org/ 10. 1371/ journ al. pone. 00316 60 2. Bhugra D et al (1996) First-contact incidence rates of schizo- phrenia in Trinidad and one-year follow-up. Br J Psychiatry 169:587–592. https:// doi. org/ 10. 1192/ bjp. 169.5. 587 3. Halvorsrud K, Nazroo J, Otis M, Brown Hajdukova E, Bhui K (2018) Ethnic inequalities and pathways to care in psychosis in England: a systematic review and meta-analysis. BMC Med 16:223. https:// doi. org/ 10. 1186/ s12916- 018- 1201-9 4. Morris RM et al (2020) Ethnicity and impact on the receipt of cognitive-behavioural therapy in people with psychosis or bipo- lar disorder: an English cohort study. BMJ Open 10:e034913. https:// doi. org/ 10. 1136/ bmjop en- 2019- 034913 5. RCPsych, R. C. o. P. (2021) National clinical audit of psy- chosis—national report for the early intervention in psychosis audit 2020/2021. Healthcare Quality Improvement Partnership, London 6. Morgan C et al (2017) Ethnicity and long-term course and out- come of psychotic disorders in a UK sample: the ÆSOP-10 study. Br J Psychiatry 211:88–94. https:// doi. org/ 10. 1192/ bjp. bp. 116. 193342 7. Chorlton E, McKenzie K, Morgan C, Doody G (2012) Course and outcome of psychosis in black Caribbean populations and other ethnic groups living in the UK: a systematic review. Int J Soc Psychiatry 58:400–408. https:// doi. org/ 10. 1177/ 00207 64011 403070 8. Maguire J, Sizer H, Mifsud N, O’Donoghue B (2020) Outcomes for migrants with a first episode of psychosis: a systematic review. Schizophr Res 222:42–48. https:// doi. org/ 10. 1016/j. schres. 2020. 05. 048 9. Maguire J et al (2021) Symptomatic, functional and service uti- lization outcomes of migrants with a first episode of psychosis. Soc Psychiatry Psychiatr Epidemiol 56:1389–1397. https:// doi. org/ 10. 1007/ s00127- 020- 02011-7 10. Birchwood M et al (1992) The influence of ethnicity and family structure on relapse in first-episode schizophrenia: a comparison of Asian, Afro-Caribbean, and White patients. Br J Psychiatry 161:783–790. https:// doi. org/ 10. 1192/ bjp. 161.6. 783 11. Hodgekins J et al (2015) Investigating trajectories of social recovery in individuals with first-episode psychosis: a latent class growth analysis. Br J Psychiatry 207:536–543. https:// doi. org/ 10. 1192/ bjp. bp. 114. 153486 12. Hall M-H, Holton KM, Öngür D, Montrose D, Keshavan MS (2019) Longitudinal trajectory of early functional recovery in patients with first episode psychosis. Schizophr Res 209:234– 244. https:// doi. org/ 10. 1016/j. schres. 2019. 02. 003 1 3 578 Social Psychiatry and Psychiatric Epidemiology (2023) 58:569–579 13. Morgan C et al (2008) Cumulative social disadvantage, ethnicity and first-episode psychosis: a case-control study. Psychol Med 38:1701–1715 14. Li H, Eack SM, Montrose DM, Miewald JM, Keshavan M (2011) Longitudinal treatment outcome of African American and Caucasian patients with first episode psychosis. Asian J Psychiatr 4:266–271. https:// doi. org/ 10. 1016/j. ajp. 2011. 08. 004 15. Kam SM, Singh SP, Upthegrove R (2015) What needs to follow early intervention? Predictors of relapse and functional recov- ery following first-episode psychosis. Early Interv Psychiatry 9:279–283 16. Birchwood M et al (2014) The UK national evaluation of the development and impact of Early Intervention Services (the National EDEN studies): study rationale, design and baseline characteristics. Early Interv Psychiatry 8:59–67. https:// doi. org/ 10. 1111/ eip. 12007 17. WHO (2004) ICD-10: international statistical classification of diseases and related health problems: tenth revision. ICD-10 18. Kay SR, Fiszbein A, Opler LA (1987) The positive and nega- tive syndrome scale (PANSS) for schizophrenia. Schizophr Bull 13:261–276. https:// doi. org/ 10. 1093/ schbul/ 13.2. 261 19. Addington D, Addington J, Schissel B (1990) A depression rating scale for schizophrenics. Schizophr Res 3:247–251. https:// doi. org/ 10. 1016/ 0920- 9964(90) 90005-R 20. Roche E et al (2016) The prognostic value of formal thought disorder following first episode psychosis. Schizophr Res 178:29–34. https:// doi. org/ 10. 1016/j. schres. 2016. 09. 017 21. Larsen TK, McGlashan TH, Moe LC (1996) First-episode Schizo- phrenia: I. Early course parameters. Schizophr Bull 22:241–256. https:// doi. org/ 10. 1093/ schbul/ 22.2. 241 22. Briggs SR, Cheek JM (1986) The role of factor analysis in the devel- opment and evaluation of personality scales. J Pers 54:106–148. https:// doi. org/ 10. 1111/j. 1467- 6494. 1986. tb003 91.x 23. Piedmont RL (2014) In: Michalos AC (ed) Encyclopedia of quality of life and well-being research. Springer Netherlands, Heidelberg, pp 3303–3304 24. Pallant J (2001) SPSS survival manual 25. Gilmour AR, Thompson R, Cullis BR (1995) Average information REML: an efficient algorithm for variance parameter estimation in linear mixed models. Biometrics 51:1440–1450. https:// doi. org/ 10. 2307/ 25332 74 26. Snijders TA, Bosker RJ (2011) Multilevel analysis: An introduction to basic and advanced multilevel modeling, Sage 27. Heck RH, Thomas SL, Tabata LN (2013) Multilevel and longitudi- nal modeling with IBM SPSS. Routledge, New York 28. Twisk J, de Boer M, de Vente W, Heymans M (2013) Multiple impu- tation of missing values was not necessary before performing a lon- gitudinal mixed-model analysis. J Clin Epidemiol 66:1022–1028. https:// doi. org/ 10. 1016/j. jclin epi. 2013. 03. 017 29. Peters SAE et al (2012) Multiple imputation of missing repeated outcome measurements did not add to linear mixed-effects models. J Clin Epidemiol 65:686–695. https:// doi. org/ 10. 1016/j. jclin epi. 2011. 11. 012 30. Kavanagh DJ et al (2011) The drugcheck problem list: a new screen for substance use disorders in people with psychosis. Addict Behav 36:927–932 31. Tait L, Birchwood M, Trower P (2002) A new scale (SES) to meas- ure engagement with community mental health services. J Ment Health 11:191–198. https:// doi. org/ 10. 1080/ 09638 23002 00235 70-2 32. Suvisaari J et al (2018) Is it possible to predict the future in first- episode psychosis? Front Psychiatry. https:// doi. org/ 10. 3389/ fpsyt. 2018. 00580 33. Puntis S, Oke J, Lennox B (2018) Discharge pathways and relapse following treatment from early intervention in psychosis services. BJPsych Open 4:368–374. https:// doi. org/ 10. 1192/ bjo. 2018. 50 34. Jongsma HE et al (2020) Social disadvantage, linguistic distance, ethnic minority status and first-episode psychosis: results from the EU-GEI case–control study. Psychol Med. https:// doi. org/ 10. 1017/ S0033 29172 00002 9X 35. Marshall M et al (2014) Impact of early intervention services on duration of untreated psychosis: data from the National EDEN pro- spective cohort study. Schizophr Res 159:1–6. https:// doi. org/ 10. 1016/j. schres. 2014. 07. 005 36. Drake RJ et al (2020) Effect of delaying treatment of first-episode psychosis on symptoms and social outcomes: a longitudinal analysis and modelling study. Lancet Psychiatry 7:602–610. https:// doi. org/ 10. 1016/ S2215- 0366(20) 30147-4 37. Singh SP et al (2015) Ethnicity and pathways to care during first episode psychosis: the role of cultural illness attributions. BMC Psychiatry 15:287 38. Morgan C et al (2006) Duration of untreated psychosis and ethnicity in the AESOP first-onset psychosis study. Psychol Med 36:239–247. https:// doi. org/ 10. 1017/ s0033 29170 50066 04 39. Anderson KK, Flora N, Archie S, Morgan C, McKenzie K (2014) Race, ethnicity, and the duration of untreated psychosis: a systematic review. Soc Psychiatry Psychiatr Epidemiol 49:1161–1174. https:// doi. org/ 10. 1007/ s00127- 013- 0786-8 40. Bhui K, Ullrich S, Coid JW (2014) Which pathways to psychi- atric care lead to earlier treatment and a shorter duration of first- episode psychosis? BMC Psychiatry 14:72. https:// doi. org/ 10. 1186/ 1471- 244x- 14- 72 41. Stokes I et al (2020) Prevalence of treatment resistance and clozap- ine use in early intervention services. BJPsych Open 6:e107. https:// doi. org/ 10. 1192/ bjo. 2020. 89 42. Wykes T, Steel C, Everitt B, Tarrier N (2007) Cognitive behavior therapy for schizophrenia: effect sizes, clinical models, and meth- odological rigor. Schizophr Bull 34:523–537. https:// doi. org/ 10. 1093/ schbul/ sbm114 43. Demjaha A et al (2017) Antipsychotic treatment resistance in first- episode psychosis: prevalence, subtypes and predictors. Psychol Med 47:1981–1989. https:// doi. org/ 10. 1017/ s0033 29171 70004 35 44. Seddon JL et al (2016) Cannabis use is associated with increased psychotic symptoms and poorer psychosocial functioning in first- episode psychosis: a report from the UK national EDEN study. Schizophr Bull 42:619–625. https:// doi. org/ 10. 1093/ schbul/ sbv154 45. Marmot MG (2006) Status syndrome: a challenge to medicine. JAMA 295:1304–1307. https:// doi. org/ 10. 1001/ jama. 295. 11. 1304 46. Selten JP, Cantor-Graae E (2005) Social defeat: risk factor for schiz- ophrenia? Br J Psychiatry 187:101–102. https:// doi. org/ 10. 1192/ bjp. 187.2. 101 47. Baker SJ, Jackson M, Jongsma H, Saville CWN (2021) The ethnic density effect in psychosis: a systematic review and multilevel meta- analysis. Br J Psychiatry J Ment Sci 219:632–643. https:// doi. org/ 10. 1192/ bjp. 2021. 96 48. Griffiths SL, Wood SJ, Birchwood M (2018) Vulnerability to psy- chosocial disability in psychosis. Epidemiol Psychiatr Sci. https:// doi. org/ 10. 1017/ s2045 79601 80004 95 49. Oluwoye O et al (2018) Racial-ethnic disparities in first-episode psychosis treatment outcomes from the RAISE-ETP study. Psychiatr Serv 69:1138–1145. https:// doi. org/ 10. 1176/ appi. ps. 20180 0067 50. Devonport TJ et al (2022) A systematic review of inequalities in the mental health experiences of Black African, Black Caribbean and black-mixed UK populations: implications for action. J Racial Ethn Health Disparities. https:// doi. org/ 10. 1007/ s40615- 022- 01352-0 51. Ajnakina O et al (2017) Patterns of illness and care over the 5 years following onset of psychosis in different ethnic groups; the GAP-5 study. Soc Psychiatry Psychiatr Epidemiol 52:1101–1111 52. Handley CE, Oakley D, Saville CWN (2020) Residential churn moderates the relationship between economic deprivation and psy- chiatric admission: evidence from Wales. J Epidemiol Community Health 74:560–564. https:// doi. org/ 10. 1136/ jech- 2019- 213351 1 3 Social Psychiatry and Psychiatric Epidemiology (2023) 58:569–579 579 53. Das-Munshi J et al (2012) Ethnic density as a buffer for psychotic experiences: findings from a national survey (EMPIRIC). Br J Psy- chiatry 201:282–290 54. Kirkbride JBBJ, Ploubidis GB, Morgan C, Dazzan P, McKenzie K, Murray RM, Jones PB (2008) Testing the association between the incidence of schizophrenia and social capital in an urban area. Psychol Med 38:1083–1094. https:// doi. org/ 10. 1017/ S0033 29170 70020 85 55. Naz S, Gregory R, Bahu M (2019) Addressing issues of race, ethnic- ity and culture in CBT to support therapists and service managers to deliver culturally competent therapy and reduce inequalities in mental health provision for BAME service users. Cogn Behav Ther 12:e22. https:// doi. org/ 10. 1017/ S1754 470X1 90000 60 1 3
10.1016_j.jrurstud.2023.01.003
Contents lists available at ScienceDirect Journal of Rural Studies journal homepage: www.elsevier.com/locate/jrurstud Rural co-working: New network spaces and new opportunities for a smart countryside Gary Bosworth a, *, Jason Whalley a, Anita Fuzi b, Ian Merrell c, f, Polly Chapman d, Emma Russell e a Newcastle Business School, Northumbria University, Newcastle Upon Tyne, UK b Cushman & Wakefield, London, UK c Rural Policy Centre, Scotland’s Rural University College, Edinburgh, Scotland d Impact Hub Inverness, Inverness, Scotland, UK e DIGIT Research Centre, University of Sussex, UK f National Innovation Centre for Rural Enterprise, Newcastle University, UK A R T I C L E I N F O A B S T R A C T Keywords: Co-working Rural entrepreneurship Digital economy Network-immiscibility Smart countryside Coworking has been a largely urban phenomenon although new initiatives are emerging in rural areas. Rural coworking is partly a response to the growing need for ICT, which is unevenly provided across rural areas, and partly to the social needs of freelancers and home-workers. By combining technological and social functions, coworking spaces can play key roles in the progress of a Smart Countryside, supporting digital, knowledge-based and creative entrepreneurs within rural places, thus reducing the need for extensive commuting and out- migration, particularly among younger and higher-skilled workers. As working practices evolve in the aftermath of Covid-19, these new physical spaces are expected to facilitate new network connections. Castells’ Network Society provides a valuable lens through which to investigate how coworking founders and managers promote a mix of internal and external networks that might create new, and superior, entrepreneurial opportunities. The research highlights strategies to promote collaboration as well as methods of adapting to meet new demands from rural workers in a range of rural settings. As an array of different rural coworking models evolve, we also reflect on the importance of inclusivity and identity in determining their relationship with other actors in the local economy. 1. Introduction The digitalisation of information and communications in the Global Network Society has facilitated working beyond traditional offices, so long as individuals have the requisite network connectivity (Castells, 2004) and the skills required for digital and remote working (Helsper and van Deursen, 2017; OECD, 2019). Remote working offers the po- tential to create a so-called “cyber-utopia” without traffic jams or urban overcrowding (Malecki and Moriset, 2008, p150), but this vision was only unexpectantly realised as a consequence of the lockdown measures adopted during the Covid-19 global pandemic, which were anything but utopian. Despite the earlier, relatively slow development of coworking, particularly in more rural settings, many commentators suggest that elements of these new ways of working will perpetuate in varying forms in a post-Covid economy (Clark, 2020; Kitagawa et al., 2021; Marcus, 2022; Tomaz et al., 2021; Reuschke et al., 2021). In this article, we define coworking spaces as, “flexible, shared, rentable and community-oriented workspaces occupied by professionals from diverse sectors” that are “designed to encourage collaboration, creativity, idea sharing, networking, socializing, and generating new business opportunities for small firms, start-ups and freelancers” (Füzi, 2015, p462). Coworking offers the potential to reverse or slow down the relentless expansion of commuting and other business travel (Fior- entino, 2019; Ohnmacht et al., 2020), which can have major impacts for the environment as well as the economic and social geography of both cities and rural regions. Uncertainty about the future intensity of city-centre office working in the wake of Covid-19 (Glaeser, 2021; Florida et al., 2020; Marcus, 2022; Nathan and Overman, 2020) along with increased investment in rural digital connectivity to address the long-standing “digital divide” (Salemink et al., 2017) and increasing * Corresponding author. E-mail addresses: gary.bosworth@northumbria.ac.uk (G. Bosworth), jason.whalley@northumbria.ac.uk (J. Whalley), anita.fuezi@gmail.com (A. Fuzi), ian. merrell@sruc.ac.uk (I. Merrell), polly.champman@impacthub.net (P. Chapman), emma.russell@sussex.ac.uk (E. Russell). https://doi.org/10.1016/j.jrurstud.2023.01.003 Received 26 February 2022; Received in revised form 23 December 2022; Accepted 9 January 2023 JournalofRuralStudies97(2023)550–559Availableonline13January20230743-0167/©2023TheAuthors.PublishedbyElsevierLtd.ThisisanopenaccessarticleundertheCCBYlicense(http://creativecommons.org/licenses/by/4.0/). G. Bosworth et al. demand for rural living (Property Wire, 2020) make this a critical time to investigate the new entrepreneurial dynamics that might be activated and sustained by rural coworking spaces. We apply the lens of the Network Society (Castells, 2004), which emphasises both social and technological processes, to assess the role of coworking in so called “smart rural futures” that are themselves dependent upon knowledge and innovation supported by advances in communications technology (Naldi et al., 2015). Applying this lens, our analysis focuses on two objectives: Firstly, to examine the new networks that are emerging within rural coworking spaces and the strategies of coworking operators that nurture collaborative communities; and sec- ondly, to examine linkages that are developing between coworking spaces and their wider rural and regional economies. As rural develop- ment is influenced by both internal and external drivers of growth, requiring a similar mix of network connections (Ray, 2006; Bock, 2016), we are fundamentally concerned with the roles that rural coworking spaces can play in integrating local and extra-local economies. Our research examines whether coworking spaces build new con- nections within their local communities and economies (i.e., are highly embedded) to boost the local entrepreneurial ecosystem (Mason and Brown, 2014), or whether they exist more as urban exclaves serving the needs of urban-centric businesses and remote working practices among urban employees. Just as Castells observed the potential for unequal access to networks and resources in his Network Society, a study of a London coworking venue, identified that the value of openness could “constitute new geographies of exclusion, enclosure and exploitation” (Lorne, 2019, p761). The diversity that is championed as a driver of innovation reifies the entrepreneurial personality who is comfortable is that space, but potentially alienates other kinds of diversity. This dilemma helps us to frame the two objectives of this paper around the internal and external dynamics of coworking. In line with these objectives, we developed a qualitative approach to engage with a range of coworking operators located in, and/or serving, rural areas. After an initial review of the literature on the emergence of coworking and the theoretical foundations of the Network Society and Smart Rural Development, we present the full methodology and then report on findings from interviews and focus groups. We finish by of- fering conclusions and recommendations. 1.1. Rural coworking: the story pre-covid Telework centres (Oestmann and Dymond, 2001) or telecottages (Paavonen, 1999) developed through the 1990–2000s with early ver- sions recognising the need of homeworkers to create physical and mental separation between home and work, to access superior tech- nology and to replicate the “buzz” of a traditional office setting (Malecki and Moriset, 2008). Many early examples struggled to transition from public funding into sustainable business models (Mokhtarian and Bag- ley, 2000) but, moving into the 2010s, the number of coworking spaces grew globally (Clifton et al., 2019). Although the sector has evolved more slowly in rural areas, the impact of the Covid-19 pandemic has drawn attention to more peripheral and rural working environments (Akhavan et al., 2021). Coworking spaces take a number of forms and operate with different ownership and management structures (Fiorentino, 2019). Private en- terprises can be single facilities or global companies operating a network of venues. There are also a wide range of publicly-run and community-led initiatives, filling these gaps left by private enterprise or creating alternative spaces tailored to niche user-demands. Focusing on rural regions, venues vary from informal community spaces, often retro-fitted to take up otherwise redundant space, through to dedicated spaces co-located with enterprise hubs or business incubators offering users the option to rent fixed workspace as well as hot-desks (Merrell et al., 2022). The spread of coworking spaces into more rural areas has been enabled by rapid advances in digital technologies and increased coverage of Wi-Fi enabled broadband (Houghton et al., 2018; Nambisan et al., 2019). The range of jobs that can be carried out beyond the traditional workplace is also increasing, so long as the requisite con- nectivity is available (Kane and Clark, 2019). In particular, the indi- vidualisation of work, combined with low-cost software and an explosion of cloud-based and mobile app-based digital services allow co-workers to operate relatively independently (Vallas and Schor, 2020). Sole-traders can streamline a range of administration activities, customer services and accounts (Atherton, 2016; Jordan, 2021), changing the traditional professional service function for both service user and service provider and creating new spaces for innovation. Dig- ital technologies are also accelerating the inception, scaling and evolu- tion of new ventures and leading to some radical re-thinking of creative endeavours that span traditional industry/sectoral boundaries (Nambi- san et al., 2019). Coworking was traditionally most attractive to smaller start-up businesses, creative industries, freelancers and solo consultancies (Füzi, 2015), with only a few examples identifying their appeal to homeworkers employed by larger institutions, including the public sector (Houghton et al., 2018). The essential values of coworking include work-life balance, reduced commuting and new network op- portunities, whether for collaboration and knowledge-sharing or to help homeworkers to overcome isolation (Spinuzzi 2012; Füzi, 2015) and create important markers between work and home life (Russell and Grant, 2020; Merrell et al., 2022). While pre-pandemic research has shown that homeworking can enhance the well-being of many groups of workers, especially employees, isolation of self-employed workers was found to have impacts on the perceived financial situation of the household in addition to feeling of loneliness (Reuschke, 2019). The social value of coworking spaces extends to the provision of a stronger collective voice to their members in local development policy circles with the ability to lobby for better business support and infrastructure improvements (Kolehmainen et al., 2016). Whether just small-talk and companionship or more business focused benefits of knowledge exchange and collaboration, the social functions of coworking spaces have been linked to better time management, personal and psychological health benefits and serendipitous moments that trigger learning and innovations (Kov´acs and Zolt´an 2017). In rural settings, this can extend to community well-being impacts too, partic- ularly as coworking spaces have the potential to engage different com- munity groups as well as businesses (Stojmenova Duh and Kos, 2016). Where coworking spaces develop to become embedded as part of the relational assets (Storper, 1997) of a local innovative milieu (Camagni, 1995) or entrepreneurial ecosystem (MasonandBrown, 2014), their in- fluence can transcend the value to members by enhancing the image of a place, providing a hub of activity to sustain other nearby enterprise and providing support to a range of community initiatives (Hill, 2022). This embedding role of coworking spaces fits with narratives of the influence of social and community factors on rural entrepreneurship practices (Korsgaard et al., 2015; Bosworth and Turner, 2018). The benefits of interacting and collaborating with people from different professions is frequently cited (Houghton et al., 2018; ˇ Sebestov´a et al., 2017), but research suggests that co-location alone is not sufficient to generate cross-fertilization and innovation outcomes (Füzi, 2015; Johns and Hall, 2020). Successful collaboration is depen- dent on internal facilitators and the wider entrepreneurial environments in which they are located (Kov´acs and Zolt´an, 2017; Clifton et al., 2019). In particular, more facilitated models of coworking with skilled hosts/managers were found to be important to support younger entre- preneurs and start-ups, mirroring some of the more established learning from business incubators (Füzi, 2015). This highlights the need to better understand the nature of new network configurations that will form within and beyond coworking spaces and the outcomes that may follow. Predictions that rural coworking will advance through a combination of tailored policies coupled with bottom-up initiatives (Akhavan et al., 2021) lead us to examine these complex relationships through the lenses JournalofRuralStudies97(2023)550–559551 G. Bosworth et al. of the Network Society and “smart” rural development. 2. Smart rural development and the Network Society The likely impact of new connectivity and mobility technologies mean that smart rural futures need to be framed differently from smart cities (Cowie et al., 2020), and need to take account of different rural and remote working patterns and coworking spaces. From a sustain- ability perspective, new technologies within coworking hubs can reduce commuting and carbon footprints and shorten supply chains, offering the potential to revitalise rural economies (Zavratnik et al., 2019) and helping to address the smart vs sustainable growth conundrum (Naldi et al., 2015). To be effective, these technological developments depend on social factors too, which are central to understanding the Network Society. The Network Society is defined as: “The social structure that results from the interaction between social organisation, social change, and a tech- nological paradigm constituted around digital information and communica- tion technologies” (Castells, 2004, xvii). Although most references to Castells’ work focus on the global reach of digital networks and examine his “space of flows” concept (Simonsen, 2004; Zhen et al., 2020), Cas- tells himself recognises the importance of different cultures, power and localised networks being integral to understanding and shaping the Network Society. While the Network Society connects many cultures on one level, people’s local experiences can be “fragmented, customized [and] individualized” (Castells, 2004, p30). The Network Society allows people to participate in multiple net- worked spaces of communication centred around mass media and the Internet, and not necessarily embedded in the local community. This spatial-social dichotomy is not unique to the online world, as shown by research into rural migration and commuting patterns (Champion et al., 2009; Bosworth and Venhorst, 2018), but the proliferation of digital communications exacerbates fragmentation. The irony of framing coworking spaces, which are themselves dependent on digital technol- ogy, as the antidote for rural society to reconnect around “place” is not lost on us, but we see their emergence as a key component of smart rural development (Naldi et al., 2015; Slee, 2019). Just as smart growth is founded on knowledge and innovation supported by advances in com- munications technology (Naldi et al., 2015), the Network Society also views economic growth as being dependent on global flows of infor- mation structured around socio-technological networks (Castells, 2004). Castells makes no particular reference to rural areas, suggesting that rural spaces sit rather low in the hierarchy of network nodes (Murdoch, 2000) and at the periphery of knowledge-based networks (Benneworth and Charles, 2005). However, a more positive outlook is that mecha- nisms to enhance access to these global flows of information could break down old spatial divisions such as the urban-rural divide (Murdoch, 2000). Coworking is one such mechanism, which brings the added advantage that it can help to address the digital divide (Salemink et al., 2017) by providing greater access to new technologies and supporting the digital skills and social networks needed to promote local entre- preneurship and innovation (Gerli and Whalley, 2022). This reinforces the importance of places as mediators of technological change (Cowie et al., 2020) as well as the environments in which meaningful cultural and social existence occurs (Fisker et al., 2021). The global nature of the Network Society demands cultural distinc- tiveness as the cornerstone of communication and knowledge exchange. Castells argues that “cultural identities become the trenches of auton- omy” (2004, p39) offering the potential for “complementarity and reciprocal learning” (2004, p42) between cultures. This requires local actors to have sufficient agency to balance top-down and bottom-up processes and develop a strong voice in dialogues with external orga- nisations. In the language of the Network Society, actors need the means to communicate and understand different cultures with the necessary openness to allow the permeation of new ideas across diverse networks. To advance “smart” forms of place-based development, local actors need to draw upon the value and distinctiveness of local resources, knowledge and traditions when engaging in wider networks (Naldi et al., 2015; OECD, 2018). Castells refers to cultures having their relevance as “nodes of a net- worked system of cultural dialogue” (2004, p42) and Murdoch describes “a constellation of networks that can be found in the contemporary countryside” (2006, p172). While this shows that rural areas have an important place in a global Network Society, we need to understand more about the different types of networks, their resources, their inter-connections and their reach. Where rural nodes become discon- nected from dominant, resource-rich networks, their value is diminished and individuals become excluded (Hacker et al., 2009). Exclusion from networks relegates actors to the space of place alone, bypassed by the network flows that are essential facilitators of social mobility as well as entrepreneurship (Baker et al., 2017). Therefore, the spaces and pro- cesses that create and sustain networks within rural spaces are critical to explaining entrepreneurial and innovative potential. Returning to Cas- tells, “We must place at the centre of the analysis the networking capacity of institutions, organisations, and social actors, both locally and globally. Connectivity and access to networks become essential” (2004, p42). Local social and economic dynamics see rural entrepreneurs draw on a range of resources to create distinctive business opportunities that satisfy both economic and lifestyle goals (Korsgaard et al., 2015). Too much emphasis on high growth, high-tech and innovative entrepre- neurship within the entrepreneurial ecosystem literature constrains our understanding of entrepreneurial enablers and dynamics in rural con- texts (Mu˜noz and Kimmitt, 2019). Instead, capitalising on the value of multiple, heterogenous rural assets requires networks through which their distinctive values can be communicated effectively, thus strengthening the identity of network nodes themselves. As Horlings et al. observe, “The nature of a place is not just a matter of its internal (perceived) features, but a product of its connectivity with other places. Places are nodes in networks, integrating the global and the local” (2020. P.356). The value of networks depends upon the utility of their nodes and the wider access that they provide (Anttiroiko, 2016; Varnelis, 2008). The sparser networks of firms in rural areas may diminish some network advantages, such as access to information, business support or training, but they still motivate innovation and entrepreneurship (Copus and Skuras 2006) and provide conduits through which firms can develop and communicate their distinctive values and capabilities (Malecki 1997). Indeed, the greater propensity for self-employment (Phillipson et al., 2019) and greater overlap of social and economic imperatives among many rural businesses (Steiner and Atterton, 2014) may see rural net- works becoming more start-up oriented and mutually supportive, drawing on a collective identity outside of urban networks. Within this space, new combinations of local and extra-local knowledge and re- lationships can spark new entrepreneurial ideas and opportunities. In rural regions experiencing increased rates of counterurbanisation and return migration, these trends add further to the network diversity (Kalantaridis and Bika, 2011; Mitchell and Madden, 2014). Until now, the economic potential of rural areas has been limited by slower and inferior provision of communications infrastructure compared to urban areas (Grubesic and Mack, 2017). The disadvantages that this created for rural areas are, however, narrowing through the collective impact of policy initiatives, government investment and entrepreneurial activity (Gerli et al., 2020; Sadowski, 2017). As a result, new opportunities are emerging for entrepreneurs to combine distinc- tive features of rurality with the benefits of digital technologies – reaching new markets, interacting more with customers and developing new products and services as well as new working practices and business models that reflect distinctive values attributed to rural places (Hill, 2022; Bosworth and Turner, 2018). Rural coworking spaces form part of this evolution, challenging conventional institutional and organisational cultures and affording greater importance to individuals’ networks in their communities of JournalofRuralStudies97(2023)550–559552 G. Bosworth et al. place (Mazur and Duchlinski, 2020). Recognising that rural coworking is opening up to employees as well as freelancers, the idea that one shares information with one’s coworking neighbour, in another firm or another industry, before sharing it with one’s work colleague may be unsettling for managers but transformative for innovation. With Covid-19 stimu- lating a rapid increase in remote working, the “buzz” of urban locations may be compromised, and the value of rural environments and their community connections are accentuated. The weakening gravitational pull of clusters, especially in the tech- nology sector (Feldman et al., 2020), challenges conventional regional economic theories and represents a major U-turn for firms who have spent years investing in attractive, comfortable and collaborative workplace environments (Dahl and Sorensen, 2020). Echoing calls from Gruber and Soci (2010) a decade ago, such transformation calls for greater attention to be afforded to the local dynamics of peripheral re- gions, not just to dominant (traditionally urban-centric) network nodes. While cities will recover, their functions may change and the new-found acceptance of nomadic forms of working will see different features of local environments attracting workers with the flexibility to work remotely. Just as Castells observed, though, this will have implications for those who are less able to engage in this new labour market and whose jobs require a physical presence in fixed premises (Florida et al., 2020; Marcus, 2022). Reframing the Network Society to consider the uniqueness of rural economies identifies that networks are not just spaces of flows but they are fundamental to shaping and narrating rural places. However, the configuration of networks within a spatially defined node and the extent to which actors are embedded in more locally or externally-oriented networks are essential to understanding the implications for rural pla- ces. For example, more innovative services have been associated with the need for stronger external networks connecting into nodes higher up the urban hierarchy (Shearmur and Doloreux, 2015) yet other creative businesses thrive as a result of their rural locations (Townsend et al., 2017). The new spaces of rural coworking hubs and the increased va- riety of remote-working practices prompted by the Covid-19 pandemic, provide the context for rethinking the meaning and influence of rural places becoming more vibrant and active nodes within the Network Society. The co-location of employees and entrepreneurs across a range of sectors forms part of the entrepreneurial potential of rural coworking, supporting an emerging literature on sector fluidity that views industries sectors being less fixed or bounded (De Massis et al., 2018) and collaborating in a quadruple helix relationship (Kolehmainen et al., 2016). Rather than a sector-focused set of relationships, rural coworking provides a greater emphasis on the social and cultural environment, inspiration and opportunities from where entrepreneurs derive (Anderson et al., 2010; Honig and Samuelsson, 2021). At this hyper-local scale, coworking spaces foster individual relationships and knowledge exchange that erode boundaries between firms and sectors. This is not technology breaking down barriers in the traditional lan- guage of the Network Society but a hybrid space where re-localisation presents a new nexus of opportunities and enterprising actors (Shane and Venkataraman, 2000) combined with networks connecting to external enablers (Davidsson, 2015). To better understand these emerging entrepreneurial spaces, both the internal and external dynamics of rural coworking spaces are investigated. Recognising that digitization is offering the tools to sup- port collective approaches to the pursuit of entrepreneurship (Nambi- san, 2017), and combining this with analysis of the network structures that surround rural coworking spaces, the methodology reflects contemporary understanding of a smart countryside. 3. Methodology Since the research took place during the Covid-19 pandemic, all data collection was conducted online. This included a series of 17 semi- structured video interviews with coworking operators/developers, supplemented by two policy-maker focus groups, an interview with the managing director of the Flexible Workspace Association and a larger online workshop. In total, the research engaged with around 80 discrete participants between September 2020 and June 2021. Additional data was collected from analysis of website content to explore the marketing messages used to describe the advantages of coworking, their key fea- tures and the rationales behind their establishment. This captured the perspectives of operators as well as the representation of rural cow- orking that they seek to communicate externally – mirroring the twin objectives of understanding both internal and external dynamics of rural coworking. The inability to access users of coworking spaces was a limitation of the research project, something which is planned to be addressed in future research. However, the framing of this paper means that the founders and managers are best placed to explain their strategies and give an informed overview of the evolving nature of rural coworking based on their experiences. They were asked to comment on the reasons that their members and customers gave for using their venues as well as explaining their marketing strategies, business models, workspace and technology provision, and the ways that they adapted to stay in contact with their members through the various periods of Covid-19 lockdown. The video interviews were audio-recorded and participants gave their consent to transcribe the conversations. The online workshop was staged on the Collab online conferencing platform (https://collabvirtual world.com) and attracted 60 delegates, mainly coworking operators along with a small number of researchers and policy-makers. This began with a presentation of emerging findings after which participants were asked to join one of a selection of “virtual tables” where members of the research team led structured break-out discussions as one might do in a global caf´e style event. Focus group participants were recruited through an email to members of the Rural Services Network, a membership organisation for rural Local Authorities and associated rural develop- ment stakeholders. Each focus group was conducted on Microsoft Teams with three members of the research team joined by 11 participants split across two sessions. Thematic analysis of the interview transcripts, focus groups and workshop notes focused on key themes of coworking practices, intra- group networks, wider connections within and beyond the rural econ- omy, the impacts of Covid-19 and the role of technology. For this paper, we focused principally on the interview data and analyse the transcripts to draw out references to “internal collaboration and networks” and “external networks and spillover effects”. Quotations were collected that picked up both positive and negative features relating to each broad theme and then arranged according to secondary themes of social or economic factors, formal or informal networks and the degree to which place was important in shaping the activities or networks being analysed. 4. Findings The sample of coworking spaces identified a wide range of organi- sations with different business models, premises, clientele and future aspirations. These ranged from social enterprises focusing on the needs of small local communities through to wholly for-profit ventures with growth plans across multiple settlements. We also spoke to operators of coworking retreats that were more targeted towards digital nomads at the national and even international scale as well as some in larger towns and cities who served a heavily rural region and others in much smaller and more remote locations. A summary of the 16 interviewees is pro- vided in Table 1. Although it is possible to identify a number of different coworking models across the operators we interviewed (Author et al., 2022), this section focuses on common elements of coworking that nurture sup- portive networks and community identities internally, while building extensive connections that help to develop their external profiles. Before JournalofRuralStudies97(2023)550–559553 G. Bosworth et al. Table 1 Interview sample characteristics. Interviewee (pseudonym) Location type Type of Organisation Annie Ben Connie David Ernie Freddy Graham Harriet Ian Julia Kenny Louise Martin Neil Olive Peter Rachel Small city Open countryside Village 2 Market towns Market town Village Market town Village Small city Island town Market town Village Market town Market town Market town 2 village locations Village Non-profit Private limited company Private limited company Private limited company Private limited company and social enterprise Family business Local Authority Family Business Community Interest Company Part of a private limited company Private company Private company Private company Private limited company Private company Private company Opened/ registered 2020 2016 2019 2017 2012 2020 2009/10 2021 2017 2018 2016 2017 2020 2021 2017 2013 Informal group 2020 exploring these networks, it is important to contextualise the research in relation to the importance of the rural location as portrayed among coworking operators. The interviewees identified both nature-based and community-based values for co-workers, for whom connections with the environment has been shown to benefit their wider well-being too (Merrell et al., 2022). Whether moving into rural areas or already embedded in the locality, many operators were very passionate about the location as highlighted in the selected quotations below: “We set it up in the countryside because we had identified … that people actually wanted to not just go [to the countryside] for the weekend or for a holiday but actually spend a longer amount of time, and if they could they’d like to work on their projects outside of the city. So we developed it as a way to help people escape the city” (Louise) “You don’t just get a nice desk. You get an AONB landscape out your window and wetlands habitat and opportunity to plant trees or whatever it might be. I think being out in the countryside around green space can help with productivity [and] creative thinking” (Neil) “One of the advantages that we really have here is that we’re on the coast and that in your lunch hour you can walk down to the beach and have your picnic lunch there” (Harriet) And operators were well aware of the marketing potential that rural locations offered too: “We definitely play on our rustic feel, like we can’t offer sleek city centre kind of facilities. This is very much a country house with views of the [mountains] and I guess it’s the location that sells it but the house itself is rustic … so to be honest it kind of suits my style.” (Connie) Emphasising the distinctiveness of the location as a strong base from which to communicate with the wider world is a good example of how the Network Society can empower rural places to take advantage of their distinctive characteristics. While urban coworking spaces may be rela- tively homogenous, focusing on hi-spec and hi-tech office space that is familiar to mobile workers wherever they happen to be, rural spaces have the scope to position themselves differently. First impressions from our research sample indicated that creating the “buzz” of urban loca- tions requires alternative approaches to community-building as well as efforts to raise awareness about coworking. These differences give rise to a number of questions to explore, in terms of how these distinctive identities are formed and the extent to which they are inclusive and representative of their wider communities. 4.1. Internal networking The literature on networking among rural firms and co-workers in- dicates that simply being close together does not guarantee collabora- tion, but it provides a foundation for new connections to emerge. Therefore, in addition to functional responsibilities, a key role for coworking operators is to promote an entrepreneurial and supportive culture within their organisation. As David observed “We always find that people think they need a desk and Wi-Fi and when people are in what keeps them in is the community.” The consensus among interviewees was that collaboration cannot be forced upon people, only facilitated, but it was very rewarding for founders when this worked: “One of the nicest parts of running a coworking space is seeing those connections being made and facilitating it, or it happening auto- matically. It’s very enjoyable. I love that. I love when people interact and they find each other and it works out and it’s very positive”. (Ben) The value of softer networks was illustrated by interviewees referring to “socializing” more than business networking. Examples included the value of being able to share the success of winning a new contract (online workshop conversation), sharing the frustration of IT problems (Rachel) or simply the need for companionship: “[One member], he comes just for company really. But he needs complete silence to work so he has his own office then comes down for coffee and lunch to meet everyone. We have a couple of people that just like to come in and know that there’s people to speak to if they need to, but they just find their own space. And then the rest of us come in and chat and then we work and then we chat a little bit more and then we work again.” (Connie) This culture was reinforced by another interview with a founder of a high street coworking venue who described one member being “a little bit too pushy” when it came to business networking: “There’s one member … he wants us to have lunches where we talk about what we do and maybe share some presentations, but [among the wider group] it’s quite overwhelmingly an interest in socialising and not talking about your business … and that actually becomes a little bit of a thing because he’s not interested in socialising, he wants to talk business and nobody else wants to.” (Annie) Later in the same interview Annie said: “We always kind of look to who’s in our building first when we look for collaborators. And I also think that very much draws people to us”, highlighting that collaborative working for mutual gain is part of their aspiration – but there is a culturally acceptable way to facilitate it. A second example from Scot- land identified similar collaborations that support members to bid for larger contracts: “we’ve formed a consortium … together we are able to bid for contracts. A lot of these contracts come along and you need to have something like £5 million worth of public liability, or some kind of insurance that is vast sums. And none of these individuals will have it whereas we’ve got it” (Ian). Stimulating this type of collaboration was also important for Local Authority focus group participants who are looking at how cow- orking might translate into rural economic growth. Whether providing a supportive social environment or actively facilitating collaborative working, there is no prescription for what makes an entrepreneurial culture. It might be relaxed, professional, focused, sociable or collaborative, each requiring different combinations of events, branding and spaces to support their members. The selection of furniture, the layout of the venue and d´ecor of rooms all contribute to JournalofRuralStudies97(2023)550–559554 G. Bosworth et al. the identity of the coworking group, often reflecting the attitudes of the founders: “Everything is community for us. We use second-hand furniture as much as possible for environmental reasons [and] … so we don’t spend millions of pounds on fitting out space. We’d much rather spend that money on activities that happen within the space.” (David) “It was important for us to have a variety of workspace types … that’s why we had this caf´e type space. That’s where people can be more social. They can have little meetings, little coffee meetings, either with their colleagues or for a break. The library is also more of a shared space, a little bit more casual. But then we have the really dedicated workspaces” (Louise) “We’re professional but we’re not formal” (Harriet) This focus on “community”, as something over and above the fundamental provision of ICT, is a clear example of Castells’ argument that nodes within the Network Society are defined by their internal cultural identity. The functional or tangible elements of the service are largely homogenous so can be accessed anywhere, but social capital and community identity are seen by the coworking founders/managers as being unique. In the case of founders who work in the space, it is often a personal reflection of their own working culture too. Without this, the homogeneity of a single Global Network Society becomes the dominant trope of how new (digital) technologies influence working practices but the response among coworking operators appears to engender a clear desire for diversity. Following this logic, spaces designed to facilitate different types of behaviour and interaction are paramount to the success of coworking spaces and consistently it was the kitchen area that was most discussed. This is where people are “off-duty” and relaxing as themselves, so the tone of the conversation is different and people become more open and more interested in each other since the pressure of the next task, the next phone call or next email is in another room: “[the kitchen] should be the heart of a coworking space because that’s where everyone collaborates and talks, and that should be right in the middle of the building and it should be where everyone goes and you should base everything around that coffee pod. (Ernie) “In [the local region] you meet people in their kitchens so we designed the front of the office to be a kitchen. So we’ve got a new dishwasher, we’ve got the toaster, we’ve got everything else in there. People come in and have their breakfast … That’s where you learn stuff” (Ian) As well as internal network building, common spaces allow for non- members to see the coworking space and for new users or event at- tendees to interact with established members. Breakfast clubs, caf´e’s open to the public and rooms dedicated to community functions all provided opportunities for events to widen the reach of the venue. Where co-workers were able to host external guests, this also helped to build a sense of community ownership among members (David). So long as external events were not disruptive for co-workers, they become a key foundation for external network connections. 4.2. Building external networks Coworking spaces represent new network nodes that can strengthen connections between rural and urban economies. A particular example was cited in Scotland where bringing together sole-traders or very small businesses allowed them to bid for larger projects outside of their lo- cality (Ian). Not only did this help others realise that a geographically peripheral business location was not a barrier to working further afield, but it is also provides a practical demonstration of how internal net- works can be leveraged externally. While the internal dynamics of the coworking “node” are critical for generating the scale of activity and cultural distinctiveness to engage in complimentary and reciprocal learning within the Network Society (Castells, 2004), interviewees were equally aware of their wider responsibilities. These include business support programmes, networking events, boosting trade for other local businesses and engaging in wider outreach activities. A number of comments capture this mentality: “We actively try and do stuff outside of our four walls which is why we’ve recruited, two years ago we recruited an outreach manager. It was her job to go out and run courses for people, so it’s a big part of what we do.” (Ernie) “We have a lot of partnerships with local businesses … I don’t think it’s a nice thing to have a project in the community where you don’t interact with the community” (Louise) “We don’t just want our spaces being another coworking space, we’re really set on a mission to make our spaces the hub of the ecosystem … we work really hard to try to get that set in people’s minds that it becomes a functional hub for the stakeholders” (David) In some cases, building external networks to support rural economic development was part of the founding principle of establishing a cow- orking space too: “The decision to start a rural hub really came from part of our pur- pose which is to improve the connections between rural and urban entrepreneurs, to see some of their learning spread a little bit further than just within the city, [and] to see the rural entrepreneurs benefiting from what’s happening in the vibrant start-up scene, which is often city based” (Olive). The bridging role of coworking spaces encompasses both the urban- rural scale and more local connections beyond the traditional digital or creative freelancer groups of co-workers. One opportunity at the local scale is presented by the anticipated growth of homeworking among salaried employees who are seeking to reduce their commuting fre- quency following the impacts of the Covid-19 pandemic. This potential new source of demand was a foundation of Neil’s business model and a major topic of conversation in the research workshop sessions. From a Local Authority perspective, potential new demand stimulated enthu- siasm to promote coworking as part of a regeneration strategy to raise the profile and appeal of small towns and failing High Streets. Although there were mixed opinions about the role of the public sector as risk- taking founder or arms’ length facilitator, there was optimism that small town coworking could boost the footfall on the High Street and support other town centre businesses. Despite positive ambitions and rhetoric around the wider value of coworking spaces, only one attempted to quantify their contribution: “It’s bringing people here, has a pretty big impact so I estimate that for the local business every year we generate about €1.2 million for accommodation, for food, for transportation, for stuff that people buy here.” (Kenny) More typical, were comments such as: “These people come here, spend money, spend time, accommoda- tion, other services … I think we are a very good addition to the landscape of [our] area” (Martin) Beyond financial benefits, the research identified a variety of con- tributions yielding more social value. A good example is Peter, the founder of a rural coworking and co-living destination, who explained that they involve local retired people in events because “they don’t need the money … they need conversations.” Peter and his business partner have also set up an educational programme where they “teach the skills of digital nomads to people who want to become digital nomads” because “we want to teach people who don’t want to leave their villages to work, but to stay at home.” In a Network Society sense, the growth of digital JournalofRuralStudies97(2023)550–559555 G. Bosworth et al. nomadism is an illustration that the urban-rural connectivity can be a two-way dynamic where people chose to visit rural locations for certain types of work. Thus, the rural coworking venue is not solely a mecha- nism to reduce out-commuting from rural places but also a location that attracts inward commuters that strengthens its role as a node linking (rural and urban) places together. The chance to support young people was echoed by Neil who felt that they struggle to access to the same training and career development opportunities as people in the big cities and recognised coworking as part of a solution that offers “a stepping-stone to seeing new career op- portunities [and] … a real opportunity for rural areas.” The sense that coworking is a point of connection between places reflects the Network Society but it also extends to a psychological connection where rural places can be perceived as being less isolated and offering greater equality in terms of access to skills and skilled employment. Once the purpose and identity of a rural coworking space is under- stood as something distinctive and place-based, the opportunity for a range of community-focused activities emerge – both promoting the space to other potential users and helping to develop a unique identity. For example, another recent start-up explained her social values in relation to future development plans: “There’s another building that I want to refurbish … we were kind of thinking like a gallery or an exhibition space or something for artists or creatives … they could run workshops there because we’ve already got a link with a local artist and she’s keen to set up chil- dren’s activities and then also do a programme for 16-24 year-olds that aren’t engaging that well with school. So that kind of thing … as well as the desks I’d like to be doing some projects that actually help people as well” (Harriet) While Harriet and her family are firmly embedded in the local area, and approach the community function from that perspective, an in- comer in a similarly remote location gave an interesting perspective on the integrative function that coworking can play. “90 per cent, maybe even 95 per cent of the people who use the Business Hub are incomers. I don’t know whether locals just feel like they don’t need it because they’ve got enough contacts and they know enough places where they can find space to work themselves, so it’s the people who don’t have those connections in the commu- nity who are coming to me. And I’m an incomer myself.” (Julia) These examples highlight the potential for coworking spaces to provide the connectivity and access to networks that are essential to the Network Society. The combined social and technological functions also highlight how this application of Network Society thinking is commensurate with “Smart” rural development. As well as highlighting the local/extra-local connections promoted by coworking, the final quotation also opens up a new set of questions about the inclusiveness of rural coworking. In the early phases of development, and with the need to build communities of users, it ap- pears inevitable that some cliques will emerge and not all people will feel able to participate. This is where the variety of rural coworking models can broaden3 accessibility far more than the corporate struc- tures that have predominated in big cities. Introducing a range of social and community activities that welcome different people into coworking venues offers the potential to build new connections among increasingly mobile, but less cohesive, rural populations. The inclusiveness of indi- vidual coworking spaces is a question for future research with co- workers but the variety of local spaces as interconnected and heterog- enous nodes aligns with Castells’ conceptualization of cultural nodes in the Network Society. 5. Discussion: conceiving diverse impacts for rural places The two areas of findings have highlighted that network relation- ships are critical to the development of rural coworking. In each case, facilitation of soft, informal networks is a key role for coworking oper- ators that was supported by a range of strategies from the design of the space, particularly communal spaces like kitchens, the staging of events (including some that were online during the pandemic) and the creation of a collective identity that engages co-workers. As in urban coworking spaces, collaboration and innovation occur through serendipitous meetings of like-minded people, not through formal networking meet- ings or hard-sell approaches. The difference in rural coworking spaces arises when communities of users develop particular identities, often based around place and nourished by the efforts of managers to create distinctive community identities. As a result, rural coworking venues become more heterogenous, shaped by combinations of social, cultural and environmental factors, and represented through the interactions of co-workers in different settings. The local environment, the character- istics of the building itself, the range of non-business activities, the personal characteristics of the owner and their ambitions to grow or diversify the membership all contribute to a particular feel for each venue. This was evident in the marketing messages of coworking web- sites too, where quotations frequently drew on their location to communicate opportunities to interact with nature, to socialise and to enhance well-being: “Pack your swimming trunks, take your to-do list and then nothing like going out to the country” “There is nowhere else can you surf in the morning and be in central London by lunch time. This is a pure manifestation of the perfect work/ life balance we all strive for” “We want the freelancers that ultimately form the creative group at NAME to feel like family” “With its own garden, high ceilings, lots of light, natural finishes and loads of plants, NAME is an energising, enjoyable place to work” “You will gain inspiration while you work, and exchange experi- ences, tips, ideas and contacts” Through the examples here, aspects of creativity, inspiration and collaboration are evident, but all were presented as part of something more holistic in terms of the work/life experience that coworking can provide. To realise this, coworking operators have to provide the right working spaces, complete with both social and technological in- frastructures – the twin pillars of smart rural development in microcosm. Each pillar has implications for the internal and external network structures, and the communications that evolve within and beyond coworking spaces. In other words, the social and technological context of rural coworking shapes the ways in which co-workers engage in the Network Society and influences the balance of local and external factors that shape business opportunities and identities. The economic spillovers, although hard to quantify, appeared to stem from building a community of co-workers with a sense of connection to their locality. Through this, businesses are able to collaborate with another and recognise opportunities to work with other local firms. Business events and training, as well as more community- focused events in some venues, all expanded the social networks around coworking spaces, increasing their external visibility and often building a sense of identity within the group – the local culture that emerges provides a sense of autonomy and empowerment aligned with that of the Network Society. The importance of the collective, can also be explained in game-theory terms since if all members sought to exploit the group for business growth, the working environment would become a deterrent. In reality, the only way to foster collaboration over time is to prioritise and develop the collective well-being of the group. Shifting the locus of networking from corporate to community spaces raises a number of questions about the agency of individuals within social networks (Taselli and Kilduff, 2021); particularly the extent to which they actively build new connections that spark the potential for innovation and new network configurations. Where home-workers and JournalofRuralStudies97(2023)550–559556 G. Bosworth et al. entrepreneurs interact in rural coworking spaces, the locality affords a common frame of reference and shared identity out of which new ideas can emerge. If these ideas are place-dependent, bringing characteristics of a rural location to the fore, the cultural identities that evolve might become new “trenches of autonomy” (Castells, 2004) that can sustain rural social innovation as well as profit-motivated entrepreneurship. In essence, where agency shifts to the local level, yet the actor remains influentially connected into wider networks, this reflects the philosophy of neo-endogenous development too (Ray, 2006). Re-engaging with Network Society theory is especially timely because of the new connections to ‘place’ deriving from the Covid-19 pandemic (Newman, 2020). In some interpretations, the Network So- ciety emphasises networks to the detriment of places (Zhen et al., 2020) where, rather than being in the right place, being in the right network counts (Anttiroiko, 2016). Here, we argue that such a dichotomy be- tween place and networks can be bridged by new remote-working and coworking practices that build and sustain new network connections within rural places while strengthening and extending connections beyond. Furthermore, creating these new nodes offers significant po- tential innovation, opportunity-creating and professional support networks associated with agglomeration (relatively homogenous) while simultaneously strength- ening heterogeneous, place-based identities and social networks that capture distinctive qualities of their rural context. rural communities replicate the for to The growing diversity of rural businesses in the UK context has been linked with professional incomers and rural returnees (Kalantaridis and Bika, 2011; Stockdale, 2015). These mobile professionals (Keeble and Nachum, 2002) and members of the rural creative class (Herslund, 2012) are better equipped to draw on valuable experience and con- nections beyond the constraints of the local rural context (Bosworth and Bat Finke, 2020); a feature aided by advances in communications technology across rural areas. However, not all forms of employment can benefit from digitalisation and the new ways of working that this enables, with a notable divide between knowledge intensive and manual occupations for example (Dingel and Neiman, 2020). Throughout the Covid pandemic, the housing market has seen increased demand for rural living, indicating that remote working practices are likely to increase in popularity. Combined with the continuing spread of online working and education, this likely to result in further decentralisation of skilled work, with migration more aligned to lifestyle choices and natural amenity values associated with the rural creative class (McGranahan and Wojan, 2007) rather than proximity to workplaces. On one hand, this offers opportunities for coworking, as identified by several research participants, but it also reinforces the perception that coworking is exclusively for mobile professionals and skilled workers. In the Network Society, Castells framed this in terms of differences in education and a person’s ability to work in the informa- tion economy, not as class conflict (Ampuja and Koivsito, 2014). This is reinforced by findings from research into homeworking during the Covid-19 pandemic too, where personal and household factors were key factors determining changes in worker productivity (Felstead and Reuschke, 2021; Hackney et al., 2022; Kitagawa et al., 2021). Given that there are multiple factors that influence workers’ productivity and their ability to participate equally in new ways of working, there is a risk that localised professional networks lead to a two-tier rural society with increased social and economic inequalities. Rural coworking is a possible cause and a possible solution to this problem. The research has identified that many coworking spaces pro- vide opportunities for community activities, training and inclusion. This is essential to avoid the perils of “network immiscibility” (Bosworth and Venhorst, 2018) where, just like the chemical properties of oil and water, networks may co-exist in a place but they require catalysts to stimulate new interactions to bridge between different sub-groups. Where coworking spaces adopt an integrating role, they can facilitate the human, social and financial capital in their networks to contribute to local development. By contrast, if they become exclusive professional spaces more integrated into urban economies, they will exacerbate the marginalisation of other sections of rural society less equipped to participate in the Network Society, perhaps lacking (access to) digital, social or professional skills. As rural coworking evolves, the challenge for operators and policymakers will be to ensure that other parts of the rural economy can benefit, even if they are not active in coworking themselves. 6. Conclusions As creative industries and knowledge-intensive business services continue to grow in rural areas (Townsend et al., 2017; Johnston and Huggins, 2016), facilitated by improved digital connectivity (European Commission, 2020; Ofcom, 2020) and the opportunity to work outside of congested, costly city locations, they are likely to shape the next phase of rural coworking development. In a post-Covid economy, there is every likelihood that rural residential preferences and digitally-enabled homeworking will fuel further demand for coworking too (McKinsey, 2021). Such a shift could challenge certain urban-centric assumptions of the Network Society based on the greater density of flows of people, knowledge and ideas that can fuel urban economic growth. Instead, rural regions can be supported in catching up with their urban coun- terparts if these flows of resources become increasingly accessible to rural entrepreneurs. As evidenced by those participating in our research, this can be facilitated through enhanced communications technologies, personal mobility and extensive networks. Rural coworking spaces can play important roles in elevating their localities to become more significant network nodes, combining local and extra-local networks around a space that depends upon both social and digital infrastructures. Conceptually, this emphasis on social and technological processes confirms that coworking can be an integral component of smart rural development too (Naldi et al., 2015). The potential for innovative mixing between sectors and professions adds a further dimension to rural coworking as a driver of new economic op- portunities. By fulfilling a combination of functions, they can be simultaneously remote network bridges connecting urban centres and urban firms and they can integrate rural economy actors into new networks. If, as a consequence of Covid-19, increased remote working becomes the norm to the extent that we conceive of ‘remote employers’ rather than ‘remote workers’, it is likely that the co-worker with rural business connections will be strongly positioned. Conversely, if the growth of remote working wanes, the potential functions of rural coworking nodes become less clear. We argue that a critical mass of human and social capital operating in rural places is integral to the development of cow- orking spaces as hubs for enterprising businesses. Through improved connectivity, which may take the form of better physical infrastructure or digital networks, rural areas are then better able to draw on a wider array of resources, which, in turn, can be leveraged to enhance the attractiveness of rural places and generate new economic activities. If resulting forms of entrepreneurship are socially embedded and digitally enabled, they can contribute to new dynamics of smart rural develop- ment that valorise spatial diversity (Naldi et al., 2015). Our paper has sought to re-invigorate the Network Society by applying its core ideas in the context of dominant place-based and “smart” rural development paradigms. This has revealed significant opportunities to promote new networks built around the social and technological needs of contemporary ways of working. Moreover, the strategies of rural coworking operators highlight the importance of identity, or “cultural distinctiveness” (Castells, 2004), in addition to the connectivity and openness to engage in heterogenous networks that characterise the Network Society. The research has also identified a challenge for rural policymakers and coworking operators to facilitate networks that bridge spatial, social and skills divides while supporting local cohesion and integration. We suggest that the most promising avenues to achieve this require rural coworking spaces to enhance their JournalofRuralStudies97(2023)550–559557 G. Bosworth et al. place-based distinctiveness by providing services to more isolated and marginalised groups, as well as the essential facilities and network brokerage demanded by rural co-workers. Author statement Gary Bosworth: Funding acquisition, Conceptualization, Method- ology, Investigation, Original draft. Jason Whalley: Conceptualization, Reviewing and editing. Anita Fuzi: Methodology, Investigation. Ian Merrell: Investigation, Reviewing and editing. Polly Chapman: Meth- odology, investigation, Reviewing and editing. Emma Russell: Funding acquisition, Reviewing and editing. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Data availability Data will be made available on request. Acknowledgements We would like to acknowledge the Digital Futures at Work Research Centre (Digit) for their funding and support throughout the project. References Akhavan, M., Mariotti, I., Rossi, F., 2021. The rise of coworking spaces in peripheral and rural areas in Italy. Territorio - Sezione Open Access (97-Supplemento). https://doi. org/10.3280/tr2021-097-Supplementooa12925. Ampuja, M., Koivisto, J., 2014. From “post-industrial” to “network society” and beyond: the political conjunctures and current crisis of information society theory. TripleC 12 (2), 447–463. Anderson, A., Dodd, S., Jack, S., 2010. Network practices and entrepreneurial growth. Scand. J. Manag. 26, 121–133. Anttiroiko, A.-V., 2016. Castells’ network concept and its connections to social, economic and political network analyses. J. Soc. Struct. 16, 1–18. Atherton, B., 2016. The rise and rise of the tech-savvy sole trader, Xero blog. Accessed 23rd February 2022 at: https://www.xero.com/blog/2016/01/the-rise-and-rise-of-th e-tech-savvy-sole-trader/. Baker, S., Warburton, J., Hodgkin, S., Pascal, J., 2017. The supportive network: rural disadvantaged older people and ICT. Ageing Soc. 37, 1291–1309. Benneworth, P., Charles, D., 2005. University spin-off policies and economic development in less successful regions: learning from two decades of policy practice. Eur. Plann. Stud. 13 (4), 537–557. Bock, B., 2016. Rural marginalisation and the role of social innovation; a turn towards nexogenous development and rural reconnection. Sociol. Rural. 56 (4), 552–573, 2016. Bosworth, G., Bat Finke, H., 2020. Commercial counterurbanisation and the changing roles of rural businesses. Environ. Plann. 52 (3), 654–674. Bosworth, G., Turner, R., 2018. Interrogating the meaning of a rural business through a rural capitals framework. J. Rural Stud. 60 (1), 1–10. Bosworth, G., Venhorst, V., 2018. Economic linkages between urban and rural regions what’s in it for the Rural? Reg. Stud. 52 (8), 1075–1085. Camagni, R., 1995. The concept of innovative milieu and its relevance for public policies in European lagging regions. Pap. Reg. Sci. 74 (4), 317–340. Castells, M., 2004. The Network Society: A Cross-Cultural Perspective. Edward Elgar, Cheltenham. Champion, T., Coombes, M., Brown, D., 2009. Migration and longer-distance commuting in rural england. Reg. Stud. 43 (10), 1245–1259. Clark, P., 2020. Here’s what the Office of 2021 Should Look like, 6 December. Financial Times available at: www.ft.com. Clifton, N., Füzi, A., Loudon, G., 2019. Coworking in the Digital Economy: Context, Motivations, and Outcomes Futures, 102439. https://doi.org/10.1016/j. futures.2019.102439. Copus, A., Skuras, D., 2006. Business networks and innovation in selected lagging areas of the European union: a spatial perspective. Eur. Plann. Stud. 14 (1), pp79–93. Cowie, P., Townsend, L., Salemink, K., 2020. Smart rural futures: will rural areas be left behind in the 4th industrial revolution? J. Rural Stud. 79, 169–176. Dahl, M., Sorenson, O., 2020. Will Covid-19 Take the Air Out of Silicon Valley? Entrepreneur & Innovation Exchange, 2020. https://doi.org/10.32617/586- 5fdfae3ca0cbf. Published online at EIX.org on December 20th. Davidsson, P., 2015. Entrepreneurial opportunities and the entrepreneurship nexus: a re- conceptualization. J. Bus. Ventur. 30, 674–695. De Massis, A., Kotlar, J., Wright, M., Kellermanns, E.W., 2018. Sector-based entrepreneurial capabilities and the promise of sector studies in entrepreneurship. Enterpren. Theor. Pract. 42 (1), 47–69. Dingel, J., Neiman, B., 2020. How many jobs can be done at home? National Bureau of Economic Research. Accessed 14th February 2022 at: https://www.nber.org/papers /w26948. European Commission, 2020. Digital Economy and Society Index accessed 26th February 2022 at: https://digital-strategy.ec.europa.eu/en/policies/desi. Feldman, M., Guy, F., Iammarino, S., 2020. Regional income disparities, monopoly and finance. Camb. J. Reg. Econ. Soc. 14 (1), 25–49. Felstead, A., Reuschke, D., 2021. A flash in the pan or a permanent change?. In: The Growth of Homeworking during the Pandemic and its Effect on Employee Productivity in the UK. Information Technology and People, pp. 1–22. https://doi. org/10.1108/ITP-11-2020-0758. Fiorentino, S., 2019. Different typologies of ‘coworking spaces’ and the contemporary dynamics of local economic development in Rome. Eur. Plann. Stud. 27 (9), 1768–1790. Fisker, J.K., Kwiatkowski, G., Hjalager, A.M., 2021. The translocal fluidity of rural grassroots festivals in the network society. Soc. Cult. Geogr. 22 (2), 250–272. Florida, R., Rodrıguez-Pose, A., Storper, M., 2020. Cities in a post-COVID world, papers in evolutionary economic geography (PEEG). Accessed 5th December 2022 at: http://econ.geo.uu.nl/peeg/peeg2041.pdf. Füzi, A., 2015. Coworking spaces for promoting entrepreneurship in sparse regions: the case of South Wales. Regional Studies, Regional Science 2 (1), 462–469. Gerli, P., Whalley, J., 2022. Digital entrepreneurship in a rural context: the implications of the rural-urban digital divide. In: Keyhani, M., Kollmann, T., Ashjari, A., Sorgner, A., Hull, C.E. (Eds.), Handbook of Digital Entrepreneurship. Edward Elgar, Cheltenham, UK. Gerli, P., Matteuci, N., Whalley, J., 2020. Infrastructure provision on the margins: an assessment of Broadband Delivery UK. Int. J. Publ. Adm. 43 (6), 540–551. Glaeser, E., 2021. Urban Resilience. National Bureau of Economic Research, p. WP29261. Accessed 23rd February 2022 at: www.nber.org/papers/w29261. Gruber, S., Soci, A., 2010. Agglomeration, agriculture and the perspective of the periphery. Spatial Econ. Anal. 5 (1), 43–72. Grubesic, T.H., Mack, E.A., 2017. Broadband Telecommunications and Regional Development. Routledge, London, UK. Hacker, K., Mason, S., Morgan, E., 2009. Digital disempowerment in a network society. Int. J. Electron. Govern. Res. 5 (2), 57–71. Hackney, A., Yung, M., Somasundram, K.G., Nowrouzi-Kia, B., Oakman, J., Yazdani, A., 2022. Working in the digital economy: a systematic review of the impact of work from home arrangements on personal and organisational performance and productivity. PLoS One 17 (10), e0274728. Helsper, E.J., van Deursen, A.J.A.M., 2017. Do the rich get digitally richer? Quantity and quality of support for digital engagement. Inf. Commun. Soc. 20 (5), 700–714. Herslund, L., 2012. The rural creative class: counterurbanisation and entrepreneurship in the Danish countryside. Sociol. Rural. 52 (2), 235–255. Hill, I., 2022. Rural Arts Entrepreneurs’ Placemaking – How ‘entrepreneurial Placemaking’ Explains Rural Creative Hub Evolution during COVID-19 Lockdowns. In: Local Economy (in press). Honig, B., Samuelsson, M., 2021. Business planning by intrapreneurs and entrepreneurs under environmental uncertainty and institutional pressure. Technovation 2021, 102124. Horlings, L., Roep, D., Mathijs, E., Marsden, T., 2020. Exploring the transformative capacity of place-shaping practices. Sustain. Sci. 15, 353–362. Houghton, K., Foth, M., Hearn, G., 2018. Working from the other office: trialling coworking spaces for public servants. Aust. J. Publ. Adm. 77 (4), 757–778. Johns, J., Hall, S., 2020. I have so little time [. . .] I got shit I need to do’: critical perspectives on making and sharing in Manchester’s FabLab. Environ. Plann. 52 (7), 1292–1312. Johnshton, Huggins, 2016. Drivers of university-industry links: the case of knowledge- intensive business service firms in rural locations. Reg. Stud. 50 (8), 1330–1345. Jordan, A., 2021. Best Payment Processors for UK Small Businesses – 11 of the Best, Small Business, 18 October available at: https://smallbusiness.co.uk. Kalantaridis, C., Bika, Z., 2011. Entrepreneurial origin and the configuration of innovation in rural areas: the case of Cumbria. In: Environment and Planning A, vol. 43. North West England, 886-884. Kane, K., Clark, 2019. Mapping the landscape of urban work: home-based businesses and the built environment. Environ. Plann. 51 (2), 323–350. Keeble, D., Nachum, L., 2002. Why do business service firms cluster? Small consultancies, clustering and decentralization in London and southern England. Trans. Inst. Br. Geogr. 27 (1), 67–90. Kitagawa, R., Kuroad, S., Okudaira, H., Owan, H., 2021. Working from home and productivity under the COVID-19 pandemic: using survey data of four manufacturing firms. PLoS One 16 (12), e0261761. Kolehmainen, J., Irvine, J., Stewart, L., Karacsonyi, Z., Szab´o, T., Alarinta, J., Norberg, A., 2016. Quadruple Helix, innovation and the knowledge-based development: lessons from remote, rural and less-favoured regions. Journal of the Knowledge Economy 7 (1), 23–42. Korsgaard, S., Muller, S., Tanvig, H., 2015. Rural entrepreneurship or entrepreneurship in the rural – between place and space. Int. J. Entrepreneurial Behavior and Research 21 (1), 5–26. Kov´acs, J., Zolt´an, E., 2017. Rural enterprise hub supporting rural entrepreneurship and innovation – case studies from Hungary. Eur. Countrys. 9 (3), 473-385. Lorne, C., 2019. The limits to openness: coworking, design and social innovation in the neoliberal city. Environ. Plann. 52 (4), 747–765. JournalofRuralStudies97(2023)550–559558 G. Bosworth et al. Malecki, E.J., 1997. Technology and Economic Development: the Dynamics of Local, Regional and National Competitiveness. Longman, Harlow. Malecki, E., Moriset, B., 2008. The Digital Economy; Business Organization, Production Ray, C., 2006. Neo-endogenous rural development in the EU. In: Cloke, P., Marsden, T., Mooney, P. (Eds.), The Handbook of Rural Studies. SAGE Publications Ltd), London, pp. 278–291. Processes and Regional Developments. Routledge, Abingdon, UK. Reuschke, D., 2019. The subjective well-being of homeworkers across life domains. Marcus, S., 2022. COVID-19 and the Shift to Remote Working. Policy Contribution 09/ Environ. Plann. 51 (6), 1326–1349. 22, June, Accessed 15 December at. https://www.bruegel.org. Mason, C., Brown, R., 2014. Entrepreneurial ecosystems and growth oriented entrepreneurship. Final Report to OECD, Paris 30 (1), 77–102. Mazur, P., Duchlinski, P., 2020. Credibility and creativity in network society. Creativity Studies 13 (1), 53–63. https://journals.vgtu.lt/index.php/CS/article/view/6585/ 9626. McGranahan, D., Wojan, T., 2007. Recasting the creative class to examine growth processes in rural and urban counties. Reg. Stud. 41 (2), 197–216. McKinsey, 2021. The future of work after COVID-19. Accessed 2nd February 2022 at: www.mckinsey.com/featured-insights/future-of-work/the-future-of-work-after-c ovid-19. Merrell, I., Fuzi, A., Russell, E., Bosworth, G., 2022. How rural coworking hubs can Reuschke, D., Clifton, N., Long, J., 2021. Remote Working - Spatial Implications in Wales. In: A Report Commissioned by the Welsh Parliament. Accessed 5th December 2022 at: https://research.senedd.wales/research-articles/new-publication-remote- working-spatial-implications-in-wales/. Russell, E., Grant, C., 2020. Agile Working and Well-Being in the Digital Age. Palgrave MacMillan, Cham, Switzerland. Sadowski, B.M., 2017. Advanced users and the adoption of high-speed broadband: results of a living lab in The Netherlands. Technol. Forecast. Soc. Change 1155, 1–14. Salemink, K., Strijker, D., Bosworth, G., 2017. Rural Development in the Digital Age: a systematic literature review on unequal ICT availability, adoption, and use in rural areas. J. Rural Stud. 54, 360–371. ˇ ˇ Sebestov´a, J., Sperka, R., Małecka, J., Łuczka, T., 2017. Coworking centres as a potential facilitate well-being through the satisfaction of key psychological needs. Local Econ 36 (7–8), 606–626. supportive network for cross-border business co-operation. Forum Scientiae Oeconomia 5 (4), 23–34. Mitchell, C., Madden, M., 2014. Re-Thinking commercial counterurbanisation: evidence Shane, S.A., Venkataraman, S., 2000. The promise of entrepreneurship as a field of from rural nova scotia, Canada. J. Rural Stud. 36, 137–148. Mokhtarian, P., Bagley, M., 2000. Modeling employees’ perceptions and proportional preferences of work locations: the regular workplace and telecommuting alternatives. Transport. Res. Pol. Pract. 34, 643–675. Mu˜noz, P., Kimmitt, J., 2019. Rural entrepreneurship in place: an integrated framework. Enterpren. Reg. Dev. 31 (9–10), 842–873. research. Acad. Manag. Rev. 25 (1), 217–226. Shearmur, R., Doloreux, D., 2015. Central places or networks? Paradigms, metaphors, and spatial configurations of innovation-related service use. Environ. Plann. 47 (7), 1521–1539. Simonsen, K., 2004. Networks, flows, and fluids – reimagining spatial analysis? Environ. Plann. 36 (8), 1333–1337. Murdoch, J., 2000. Networks – a new paradigm of rural development? J. Rural Stud. 16, Slee, B., 2019. Delivering on the concept of smart villages – in search of an enabling 407–419. Murdoch, J., 2006. Networking rurality: emergent complexity in the countryside. In: Cloke, P., Marsden, T., Mooney, P.H. (Eds.), Handbook of Rural Studies. Sage, London, pp. 171–184. Naldi, L., Nilsson, P., Westlund, H., Wixe, S., 2015. What is smart rural development? J. Rural Stud. (40) 190-101. Nambisan, S., 2017. Digital entrepreneurship: toward a digital technology perspective of entrepreneurship. Enterpren. Theor. Pract. 41 (6), 1029–1055. Nambisan, S., Wright, M., Feldman, M., 2019. The digital transformation of innovation and entrepreneurship: progress, challenges and key themes. Res. Pol. 48 (8), 103773. Nathan, M., Overman, H., 2020. Will coronavirus cause a big city exodus? Environment and Planning B: Urban Analytics and City Science 47 (9), 1537–1542. Newman, P., 2020. Covid, cities and climate: historical precedents and potential transitions for the new economy. Urban Science 4 (3), 32. OECD, 2018. Rural 3.0: a framework for rural development. Accessed 23rd at: https:// www.oecd.org/rural/rural-development-conference/documents/. OECD, 2019. Skills outlook – thriving in a digital world. Accessed 23rd December at: https://www.oecd-ilibrary.org/education/oecd-skills-outlook-2019_df80bc12-en. Oestmann, S., Dymond, A.C., 2001. Telecentres—experiences, lessons and trends. Telecentres: Case studies and key issues 1, 1–15. theory. Eur. Countrys. 11 (4), 634–650. Spinuzzi, C., 2012. Working alone together: coworking as emergent collaborative activity. J. Bus. Tech. Commun. 26 (4), 399–441. Steiner, A., Atterton, J., 2014. The contribution of rural businesses to community resilience. Local Econ. 29 (3), 228–244. Stockdale, A., 2015. Contemporary and ‘messy’ rural in-migration processes: comparing counterurban and lateral rural migration. Popul. Space Place 22 (6), 599–616. Stojmenova Duh, E., Kos, A., 2016. Fablabs as drivers for open innovation and co- creation to foster rural development. In: International Conference on Identification, Information and Knowledge in the Internet of Things. https://doi.org/10.1109/ IIKI2016.70. Storper, M., 1997. Regional economies as relational assets. In: Lee, R., Wills, J. (Eds.), Geographies of Economics. Arnold, London, pp. 248–258. Tasselli, S., Kilduff, M., 2021. Network agency. Acad. Manag. Ann. 15 (1) https://doi. org/10.5465/annals.2019.0037. Tomaz, E., Moriset, B., Teller, J., 2021. Rural coworking spaces in the Covid-19 era. A window of opportunity? Accessed 16th December 2021 at: https://halshs.archives-ou vertes.fr. Townsend, L., Wallace, C., Fairhurst, G., Anderson, A., 2017. Broadband and the creative industries in rural Scotland. J. Rural Stud. 54, 451–458. Ofcom, 2020. Connected Nations 2020 - UK report, 17 December, available at: www.ofc Vallas, S., Schor, J.B., 2020. What do platforms do? Understanding the gig economy. om.gov.uk. Ohnmacht, T., Z’Rotz, J., Dang, L., 2020. Relationships between Coworking Spaces and CO2 emissions in work-related commuting: first empirical insights for the case of Switzerland with regard to urban-rural differences. Environmental Research Communications 2 (12), 125004. Paavonen, W., 1999. Telecottages and other work centre experiments. Telektronikk 95 (4), 64–68. Phillipson, J., Tiwasing, P., Gorton, M., Maioli, S., Newbery, R., Turner, R., 2019. Shining a spotlight on small rural businesses: how does their performance compare with urban? J. Rural Stud. 68, 230–239. Annu. Rev. Sociol. 46, 273–294. Varnelis, K., 2008. Networked Publics. MIT Press, Cambridge, MA. Wire, Property, 2020. Demand for homes in rural areas rises due to the pandemic. available at: www.propertywire.com. Zavratnik, V., Superina, A., Stojmenova Duh, E., 2019. Living labs for rural areas: contextualization of living lab frameworks, concepts and practices. Sustainability 11 (14), 3797. Zhen, F., Jia, T., Wang, X., 2020. How does castells’s the rise of the network society contribute to research in human geography? A citation content and context analysis. Prof. Geogr. 72 (1), 96–108. JournalofRuralStudies97(2023)550–559559
10.3847_1538-4357_ac721b
The Astrophysical Journal, 943:19 (14pp), 2023 January 20 © 2023. The Author(s). Published by the American Astronomical Society. https://doi.org/10.3847/1538-4357/ac721b DeepZipper. II. Searching for Lensed Supernovae in Dark Energy Survey Data with Deep Learning , , , , R. Morgan1,2,3 R. A. Gruendl10,11 , B. Nord2,4 , K. Bechtol1,5 , E. J. Buckley-Geer2,12 T. M. C. Abbott19, M. Aguena14 D. L. Burke9,23, M. Carrasco Kind10,11 , J. De Vicente33 M. Costanzi30,31,32 , J. García-Bellido37 J. Frieman2,4 , A. Möller6 , A. J. Shajib12,56 , F. Andrade-Oliveira20, J. Annis2 , W. G. Hartley7, S. Birrer8,9 , S. J. González1 , A. Carnero Rosell13,14,15 , C. Lidman16,17 , M. Martinez1 , T. Collett18 , D. Bacon18 , S. Bocquet21 , J. Carretero24 , S. Desai34 , F. J. Castander25,26 , C. Conselice27,28 , P. Doel22, S. Everett35, I. Ferrero36, B. Flaugher2 , E. Gaztanaga25,26 , D. Gruen21 , G. Gutierrez2 D. L. Hollowood35 R. Miquel24,44 , K. Honscheid39,40 , A. Palmese45 J. Prat4,12, M. Rodriguez-Monroy49, A. K. Romer50 , K. Kuehn41,42 , F. Paz-Chinchón10,46 , N. Kuropatkin2 , M. E. S. Pereira47, A. Pieres14,29 , O. Lahav22 , A. Roodman9,23 , E. Sanchez33 , V. Scarpine2, I. Sevilla-Noarbe33 , , D. Brooks22 , L. N. da Costa14,29, , D. Friedel10, , , S. R. Hinton38 , M. Lima14,43, F. Menanteau10,11 , A. A. Plazas Malagón48 , M. Smith51 , E. Suchyta52 , M. E. C. Swanson51, G. Tarle20 , D. Thomas18, and T. N. Varga53,54,55 1 Physics Department, University of Wisconsin-Madison, Madison, WI 53706, USA; robert.morgan@wisc.edu 2 Fermi National Accelerator Laboratory, P.O. Box 500, Batavia, IL 60510, USA 3 Legacy Survey of Space and Time Corporation Data Science Fellowship Program, USA 4 Kavli Institute for Cosmological Physics, University of Chicago, Chicago, IL 60637, USA 5 Legacy Survey of Space and Time, 933 North Cherry Avenue, Tucson, AZ 85721, USA 6 Centre for Astrophysics & Supercomputing, Swinburne University of Technology, Victoria 3122, Australia 7 Department of Astronomy, University of Geneva, ch. d’Ecogia 16, CH-1290 Versoix, Switzerland 8 Kavli Institute for Particle Astrophysics and Cosmology, Department of Physics, Stanford University, Stanford, CA 94305, USA 9 SLAC National Accelerator Laboratory, Menlo Park, CA 94025, USA 10 Center for Astrophysical Surveys, National Center for Supercomputing Applications, 1205 West Clark St., Urbana, IL 61801, USA 11 Department of Astronomy, University of Illinois at Urbana-Champaign, 1002 W. Green Street, Urbana, IL 61801, USA 12 Department of Astronomy and Astrophysics, University of Chicago, Chicago, IL 60637, USA 13 Instituto de Astrofisica de Canarias, E-38205 La Laguna, Tenerife, Spain 14 Laboratório Interinstitucional de e-Astronomia - LIneA, Rua Gal. José Cristino 77, Rio de Janeiro, RJ—20921-400, Brazil 15 Universidad de La Laguna, Dpto. Astrofísica, E-38206 La Laguna, Tenerife, Spain 16 Centre for Gravitational Astrophysics, College of Science, The Australian National University, ACT 2601, Australia 17 The Research School of Astronomy and Astrophysics, The Australian National University, ACT 2601, Australia 18 Institute of Cosmology and Gravitation, University of Portsmouth, Portsmouth PO1 3FX, UK 19 Cerro Tololo Inter-American Observatory, NSF’s National Optical-Infrared Astronomy Research Laboratory, Casilla 603, La Serena, Chile 20 Department of Physics, University of Michigan, Ann Arbor, MI 48109, USA 21 University Observatory, Faculty of Physics, Ludwig-Maximilians-Universität, Scheinerstr. 1, D-81679 Munich, Germany 22 Department of Physics & Astronomy, University College London, Gower Street, London WC1E 6BT, UK 23 Kavli Institute for Particle Astrophysics & Cosmology, P.O. Box 2450, Stanford University, Stanford, CA 94305, USA 24 Institut de Física d’Altes Energies (IFAE), The Barcelona Institute of Science and Technology, Campus UAB, E-08193 Bellaterra (Barcelona), Spain 25 Institut d’Estudis Espacials de Catalunya (IEEC), E-08034 Barcelona, Spain 26 Institute of Space Sciences (ICE, CSIC), Campus UAB, Carrer de Can Magrans, s/n, E-08193 Barcelona, Spain 27 Jodrell Bank Center for Astrophysics, School of Physics and Astronomy, University of Manchester, Oxford Road, Manchester M13 9PL, UK 28 University of Nottingham, School of Physics and Astronomy, Nottingham NG7 2RD, UK 29 Observatório Nacional, Rua Gal. José Cristino 77, Rio de Janeiro, RJ—20921-400, Brazil 30 Astronomy Unit, Department of Physics, University of Trieste, via Tiepolo 11, I-34131 Trieste, Italy 31 INAF-Osservatorio Astronomico di Trieste, via G.B. Tiepolo 11, I-34143 Trieste, Italy 32 Institute for Fundamental Physics of the Universe, Via Beirut 2, I-34014 Trieste, Italy 33 Centro de Investigaciones Energéticas, Medioambientales y Tecnológicas (CIEMAT), Madrid, Spain 34 Department of Physics, IIT Hyderabad, Kandi, Telangana 502285, India 35 Santa Cruz Institute for Particle Physics, Santa Cruz, CA 95064, USA 36 Institute of Theoretical Astrophysics, University of Oslo. P.O. Box 1029 Blindern, NO-0315 Oslo, Norway 37 Instituto de Fisica Teorica UAM/CSIC, Universidad Autonoma de Madrid, E-28049 Madrid, Spain 38 School of Mathematics and Physics, University of Queensland, Brisbane, QLD 4072, Australia 39 Center for Cosmology and Astro-Particle Physics, The Ohio State University, Columbus, OH 43210, USA 40 Department of Physics, The Ohio State University, Columbus, OH 43210, USA 41 Australian Astronomical Optics, Macquarie University, North Ryde, NSW 2113, Australia 42 Lowell Observatory, 1400 Mars Hill Rd, Flagstaff, AZ 86001, USA 43 Departamento de Física Matemática, Instituto de Física, Universidade de São Paulo, CP 66318, São Paulo, SP, 05314-970, Brazil 44 Institució Catalana de Recerca i Estudis Avançats, E-08010 Barcelona, Spain 45 Department of Astronomy, University of California, Berkeley, 501 Campbell Hall, Berkeley, CA 94720, USA 46 Institute of Astronomy, University of Cambridge, Madingley Road, Cambridge CB3 0HA, UK 47 Hamburger Sternwarte, Universität Hamburg, Gojenbergsweg 112, D-21029 Hamburg, Germany 48 Department of Astrophysical Sciences, Princeton University, Peyton Hall, Princeton, NJ 08544, USA 49 Centro de Investigaciones Energéticas, Medioambientales y Tecnológicas (CIEMAT), Madrid, Spain 50 Department of Physics and Astronomy, Pevensey Building, University of Sussex, Brighton BN1 9QH, UK 51 School of Physics and Astronomy, University of Southampton, Southampton, SO17 1BJ, UK 52 Computer Science and Mathematics Division, Oak Ridge National Laboratory, Oak Ridge, TN 37831, USA 53 Excellence Cluster Origins, Boltzmannstr. 2, D-85748 Garching, Germany 1 The Astrophysical Journal, 943:19 (14pp), 2023 January 20 Morgan et al. 54 Max Planck Institute for Extraterrestrial Physics, Giessenbachstrasse, D-85748 Garching, Germany 55 Universitäts-Sternwarte, Fakultät für Physik, Ludwig-Maximilians Universität München, Scheinerstr. 1, D-81679 München, Germany Received 2022 April 11; revised 2022 May 17; accepted 2022 May 20; published 2023 January 23 Abstract Gravitationally lensed supernovae (LSNe) are important probes of cosmic expansion, but they remain rare and difficult to find. Current cosmic surveys likely contain 5–10 LSNe in total while next-generation experiments are expected to contain several hundred to a few thousand of these systems. We search for these systems in observed Dark Energy Survey (DES) five year SN fields—10 3 sq. deg. regions of sky imaged in the griz bands approximately every six nights over five years. To perform the search, we utilize the DeepZipper approach: a multi-branch deep learning architecture trained on image-level simulations of LSNe that simultaneously learns spatial and temporal relationships from time series of images. We find that our method obtains an LSN recall of 61.13% and a false-positive rate of 0.02% on the DES SN field data. DeepZipper selected 2245 candidates from a magnitude-limited (mi < 22.5) catalog of 3,459,186 systems. We employ human visual inspection to review systems selected by the network and find three candidate LSNe in the DES SN fields. Unified Astronomy Thesaurus concepts: Strong gravitational lensing (1643); Supernovae (1668) 1. Introduction light Galaxy-scale gravitational lensing occurs when the gravitational potential of a foreground galaxy (positioned along an observer’s line of sight to a background galaxy) is large enough to deflect the photons of the background galaxy on their journey to an observer. This process produces arcs and/or multiple images of the background galaxy (Treu 2010). For the specific case in which the background galaxy contains a supernova (SN), the photons that contribute to each of the multiple images of the lensed supernova (LSN) travel different paths and distances to the observer and encounter different depths of gravitational potential depending on the distribution of the foreground galaxy’s mass. Because the speed of the distinct paths correspond to distinct arrival times of the photons from each SN image. Combining this time delay with a model of the foreground galaxy’s mass distribution enables the direct inference of the rate of expansion of the universe today, H0, as well as other cosmological parameters (Refsdal 1964). is constant, Historically, LSNe are rare—only a few detections have been made in total (Amanullah et al. 2011; Quimby et al. 2014; Kelly et al. 2015; Rodney et al. 2015, 2021; Goobar et al. 2017). However, modern optical time-domain survey data sets, such as those collected in the southern hemisphere by the Dark Energy Survey’s (DES; Abbott et al. 2016; Diehl 2020) SN fields, in the northern hemisphere by the Zwicky Transient Facility (Graham et al. 2019) and the Young Supernova Experiment (Jones et al. 2021), and over the next decade by the Vera C. Rubin Observatory’s Legacy Survey of Space and Time (LSST; Ivezić et al. 2019), are promising places to search for LSNe. Based on imaging depth, sky area, and duration of observations, the DES SN fields are expected to contain ∼0.5–2 LSNe, and the LSST wide field is expected to contain ∼2000 LSNe (Oguri 2019). These data sets, which contain hundreds of millions to tens of billions of objects that are not LSNe, pose a significant challenge for searches (Marshall et al. 2017; Abbott et al. 2021). In particular, it is vital to identify an LSN rapidly to enable follow-up observations before the SN fades during the weeks to months after 56 NHFP Einstein Fellow. the explosion (Mihalas 1963). To keep pace with the data streams of large surveys and identify candidate LSNe promptly, we require fast and robust algorithms. In Morgan et al. (2022)—hereafter referred to as “DZ1”—we designed a deep learning detection architecture (“ZipperNet”) for LSNe and demonstrated its performance on four simulated optical survey data sets that mimic DES and LSST. In this work, we use a ZipperNet to search the DES SN fields (Abbott et al. 2021) for LSNe. We also discuss the data collection and data reduction steps necessary to carry out a comprehensive LSN search in an optical survey data set. We have made all code for data processing and deep learning available at DZ1. We present this work as follows. In Section 2, we describe the characteristics of the DES SN field data. In Section 3, we describe the training and optimization of our deep learning approach. In Section 4, we quantify the performance of this architecture on the DES SN field data, as well as current candidate LSN systems. In Section 5, we discuss the significance of the results and the outlook for detecting LSNe in Rubin Observatory data. We conclude in Section 6. 2. Data Collection 2.1. The DES SN Fields DES SN field data were collected (a) to facilitate the Type Ia SN (SN Ia) cosmology analyses in DES that use the single-epoch images and (b) to enable galaxy population modeling (near the detection limits of the DES wide-field survey) that uses coadded images. All data were collected with DECam (Flaugher et al. 2015) on the Victor M. Blanco telescope from the Cerro-Tololo Inter- American Observatory in Chile between 2012 and 2018. There are 10 3 sq. deg. fields: eight shallow fields (X1, X2, E1, E2, C1, C2, S1, and S2) observed to a single-visit depth of ∼23.5 mag, and two deep fields (X3 and C3) observed to a single-visit depth of ∼24.5 mag. Each field was imaged in the griz bands approximately every six nights over five years, subject to Sun, Moon, and weather conditions. The median full-width-at-half-maximum point-spread functions (PSFs) (“seeing”) for the SN field images used in this analysis (after the downsampling discussed in Section 2.2) were 1 37, 1 26, 1 15, and 1 08 for the griz. Original content from this work may be used under the terms of the Creative Commons Attribution 4.0 licence. Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. 2.2. Candidate System Selection and Data Reduction We begin our search for candidate LSNe with all cataloged objects of DES Data Release 1 (also referred to as the “Year 3 2 The Astrophysical Journal, 943:19 (14pp), 2023 January 20 Gold Catalog”; Abbott et al. 2018). We construct an initial sample by requiring the object to be positioned within one of the SN fields and requiring all griz MAG_AUTO measurements to be brighter than 27.5 mag. Then, within that sample, we require the i band MAG_AUTO only to be brighter than 22.5 mag to restrict the total number of objects in this first search of the DES SN fields. Also within the initial sample, we require a catalog-level parameter size measurement (CM_T) to be greater than 0.05, which excludes non-extended objects (e.g., stars) with approximately 99% galaxy purity and 98% galaxy completeness. To evaluate the purity and completeness, we take a nearest-neighbor machine-learning classifier that com- bines DES photometry with near-infrared photometry as truth, which has shown near-perfect performance at magi < 22.5 (Hartley et al. 2021). These cuts produce a sample of 3,459,186 candidate systems for our analysis. We next introduce a selection on the images that are used in the LSNe search across all five years of DES SN field exposures (Abbott et al. 2021). If a system has two images on the same night in the same band, we choose the image for which the object was observed with the better seeing. For each image, we also require the cataloged object’s centroid to be positioned more than 23 pixels from all CCD edges: this permits constructing image cutouts (45 pixel by 45 pixel) without producing partial images. Finally, to enforce cadence uniformity and simplify data processing, we require the same the griz bands. We number of observations in each of determine the band with the fewest useful observations and exclude images from the other bands to match it. In doing so, the time series in we exclude images from regions of descending order of the sampling rate. Thus, for each candidate lens galaxy in the SN fields selected from the DES catalog, we obtain a time series image set with the same number of images in each band of griz. A typical length for a time series image set is ∼ 20–35 epochs. We process each year of DES data independently. 3. Deep Learning Methods 3.1. Training Set Construction Our approach for detecting LSNe in the DES deep fields requires samples of LSNe (positives) and non-LSNe (nega- tives) to train the ZipperNet in a binary classification scheme. To construct the training set, we used ∼2% of the total data set —76,203 time series image sets. Due to the lack of real LSN examples, we create the positive class using gravitational lensing simulation software (deeplenstronomy; Morgan et al. 2021) to add LSNe to DES images in the training set. For the negative class, we use time series image sets selected at random from the data set. Even given the erroneous case where a real LSN is randomly selected for the negative training class, LSNe are expected to be sufficiently rare in the DES SN fields such that this error would be infrequent and not affect the training. Nevertheless, likely types of false the two most positives will be non-lensed SNe and strongly lensed galaxies without SNe; and unfortunately, both these types of systems are also expected to be rare in our data set. Therefore, to prepare a training set with boosted representation of systems that we to be more challenging to classify, we also use expect deeplenstronomy to inject lensed source galaxies and non-LSNe into a fraction of the negative-class images. Morgan et al. includes all types of astronomical systems that The process of injecting simulated light sources into real time series image sets has multiple benefits. The training data the set ZipperNet will classify because it is chosen from the total data set. Also, the properties of the simulated source galaxies and SNe are drawn from real data, maintaining all inherent physical correlations. We join the DES Year 3 Gold catalog and DES Year 1 morphological catalog (Tarsitano et al. 2018) to obtain a sample of ∼100,000 galaxies from which we draw parameter values for simulations. The simulated source galaxies are modeled with Sérsic light profiles that have a color-indepen- dent ellipticity, a Sérsic profile index, a band-wise half-light radius, a band-wise magnitude, and a photometric redshift—all measured within DES pipelines. As in DZ1, the injected SNe were simulated using public rest-frame SN spectral energy distributions (Kessler et al. 2010) available in deeplenstr- onomy, which redshifts the distribution and calculates the observed magnitude in each band. The injected SNe reach peak brightness within the interval of 20 days before the first observation and 20 days after the final observation: the data set and contains complete lightcurves (∼70%). falling-only (∼15%), rising-only (∼15%), To calculate the lensing effects of the real galaxy on the simulated source light, we use the measured photometric redshift of the lens galaxy, select an Einstein radius at random from the interval [0 4, 1 8], and model the mass distribution of the lens as a singular isothermal ellipsoid following similar approaches in the literature (Rojas et al. 2022). For simplicity, the mass distribution shares the measured center position and ellipticity values with the light from the real lens galaxy. This simplification is not expected to greatly affect performance because these parameters are expected to be positively correlated. From the mass profile, we calculate the lensed positions of the source galaxy and LSN, as well as account for the time delays of the separate SN images. The output of the deeplenstronomy simulation are time series image sets with three kinds of objects added to real DES images—LSNe, lensed source galaxies, and non-LSNe. In total, 25% of the 76,203 time series image sets placed aside for training are injected with an LSN Ia and 25% are injected with a lensed core-collapse SN (LSN CC) to construct the positive class. Also, 16.5% of the training time series image sets are left untouched, 16.5% are injected with a galaxy– galaxy strong lens, 8.25% are injected with an SN Ia, and 8.25% are injected with an SN CC. The positive and negative training classes are equal to maintain a balanced data set throughout training. We describe the details of the training in Section 3.4, but it is worth noting here that, given our choice of loss function, balancing the classes is essential to prevent class representation biasing the learned feature representation. The remainder of subsection describes this simulation–injection process in detail. Examples of objects in the training data set are collected in Figure 1. in total number this 3.2. Preprocessing Before we train the ZipperNet and apply it to the observed data set, we apply a series of standardization steps. We first truncate the time series image sets to 10 “time steps” in each band. A time step refers to a single exposure in the sequence in the DES SN fields, a time step is of observations; approximately 6–7 days. If an image set contains more time steps, we separate it into multiple 10-time step sequences: 3 The Astrophysical Journal, 943:19 (14pp), 2023 January 20 Morgan et al. Figure 1. Examples of systems from our training data set. The composite image is an RGB visualization of the averaged gri images and the scaled brightnesses are the values extracted from the g (blue “×”), r (green triangles), i (orange circles), z (red squares) images at each time step in the time series image set using the aperture method presented in DZ1. time steps 1–10 are a single sequence; time steps 2–11 are a second sequence, etc. Then, for each 10-step image sequence, we extract the total brightness as a function of time using the background-subtracted aperture technique presented in DZ1 with an aperture radius of 15 pixels. Importantly, when extracting the total brightness, the zero-point of the image is not used to maintain independence from all non-image data products. This choice produces noise-dominated extracted brightness lightcurves, such as those in Figure 1, though it is shown in the remainder of the analysis that the ZipperNet can still identify the temporal signatures of LSNe within the noise. Next, we average the images within each band to obtain a single image in each band for the 10-step image sequence. Finally, we scale the pixel values of the averaged images and the extracted brightness values linearly to range 0 to 1 on a per- example basis. The resulting input to the ZipperNet is two different kinds of data: (1) a scaled image in each of the griz bands as a 4 × 45 × 45-element array and (2) a scaled 10-step lightcurve in each of the griz bands as a 4 × 10-element array. After processing the training data set into 10-step sequences and downsampling to maintain equal representation of the positive and negative classes, we have a total of 1,000,012 training examples. We split these examples into 90% training and 10% validation data sets. 3.3. ZipperNet The two-branch architecture of ZipperNet was first presented and validated in DZ1, and we summarize here. One branch receives scaled, time-averaged images in each band as inputs to a block that extracts convolutional features. The other branch receives scaled extracted brightness–time series as inputs to a block that extracts sequence features. The outputs from the feature-extraction blocks are flattened and concatenated. A series of fully connected layers then weights and condenses the concatenated feature representation to produce an output score that the input system contains an LSN. The ZipperNet used in this paper is similar to Figure 2 of DZ1, and the exact hyperparameter settings for this analysis are presented in Table 1. We performed a full hyperparameter optimization of the architecture and learning algorithm using the validation data set. Small changes to hyperparameter settings from the prototype ZipperNet in DZ1 reflect a specialization for the real DES images used in the training data. We find that the the addition of addition of another convolutional another long short-term memory (LSTM) layer, minor tweaks to convolutional layer kernel and stride settings, and the removal of dropout layers leads to boosted performance. The selected settings for the learning algorithm are presented in Section 3.4. layer, 4 The Astrophysical Journal, 943:19 (14pp), 2023 January 20 Morgan et al. Table 1 ZipperNet Layer Specifications Specifications Conv2D—(k: 10, p: 2, s: 1)—(4 → 16) MaxPool2D (k: 2) Conv2D—(k: 5, p: 2, s: 1)—(16 → 32) MaxPool2D (k: 2) Conv2D—(k: 3, p: 2, s: 1)—(32 → 64) MaxPool2D (k: 2) Reshape (12 × 12 × 64 → 9216 × 1) Fully connected (9216 → 408) Fully connected (408 → 25) LSTM (h: 128) LSTM (h: 128) LSTM (h: 128) Fully connected (128 → 50) Concatenate fc2 and fc3 Outputs Fully connected (75 → 6) Fully connected (6 → 2) Layer conv1a maxpool conv2a maxpool conv3a maxpool flatten fc1a fc2a lstm1 lstm2 lstm3 fc3a concat fc4a fc5b Notes. We adopt the following shorthand: kernel size (k), padding (p), stride (s), and hidden units (h). Arrows indicate the change in the size of the data representation as it is passed through the layer. a Indicates a Rectified Linear Unit (ReLU) activation function. b Indicates a LogSoftmax activation function. In total, our model contains 4,148,225 trainable parameters. 3.4. Training To train the ZipperNet, we implemented a distributed setup on five computers (two machines with Intel 3.2 GHz processors and 256 GB RAM, one machine with an AMD 2.2 GHz processor and 512 GB RAM, and two machines with IntelX 2.6 GHz processors and 768 GB RAM) on the DES cluster at Fermilab. The training data set was split into five equal chunks, each placed on an independent computer. On each computer, we instantiated a ZipperNet and initialized the weights at the same randomly selected values. We then begin passing the chunks of training data through the ZipperNet instances on each of the five computers. At regular intervals (every 1/15 of a chunk), we collect the parameters of each of the five ZipperNet instances and average the values of the parameters. Mathematically, the averaging operation is equivalent to the training, provided the weights being updated by normal learning rate is scaled by the number of network instances. Within this setup, we use a batch size of five examples and use stochastic gradient descent with a Nesterov momentum coefficient of 0.9, a constant learning rate of 0.001, and categorical cross-entropy loss to update the weights at each training step. We refer to the exhaustion of all data in a chunk as a “training iteration” and cycle back to the beginning of the chunk once the data has all been passed through the network instance. We allow training to continue for five training iterations and reach a final validation set accuracy of 93.0%. This raw accuracy is dependent on the representations of the different types of negative examples in the validation data set. In Section 4, we assess the performance using physically meaningful metrics. 3.5. Candidate Selection Criteria The output of the trained ZipperNet on an input (pair of an averaged image and a lightcurve) is a score with a value 5 typically between −100.0 and 50.0. Based on the minimum and maximum values of this range in our validation data set, we linearly scale the ZipperNet output scores to the range [0.0, 1.0], such that they are similar to probabilities. Next, we select a threshold ZipperNet score above which we include the candidate system in our final sample and below which we exclude the candidate system. We select this threshold by iterating through possible threshold values and analyzing the fraction of LSNe that scored higher than the threshold compared to the fraction of galaxies that scored higher than the threshold. The left panel of Figure 2 shows the attainable values of these quantities for different thresholds. We expect galaxies to be the largest background: the number of galaxies in a given area of sky is orders of magnitude higher than the number of strong lenses (SLs) or SNe. Therefore, we select the threshold by reducing the fraction of galaxies scored higher than the threshold to the lowest value before the fraction of LSNe scored higher than the threshold starts to decline rapidly. Based on this analysis, we select an operating threshold for the scaled ZipperNet scores of 0.76. This threshold value is contextualized with the ZipperNet scores for the systems in our validation data set in the right panel of Figure 2. We develop a final selection criterion to narrow the sample of candidate systems selected by ZipperNet. We leverage the aspect of our data processing from Section 3.2 in which time series image sets with more than 10 epochs are split into 10 epoch subsequences, which are then classified independently by ZipperNet. In analyzing the ZipperNet classifications made on all subsequences of a time series image set, we find that LSNe are more likely than galaxies to have multiple detections. This relationship is illustrated in Figure 3 using our validation data set, which we use as motivation to develop a criterion on the aggregate detections in a time series image set. Importantly, the total length of the time series image sets in our training and validation data was not required to match the real data as a result of our preprocessing methods, so it would be inaccurate to set a strict requirement on the number of ZipperNet detections (score above the threshold) based on the validation data set. Rather, to put the validation data set and the real data on the same footing, we set a requirement on the ratio of number of detections to number of subsequences. Therefore, we select the threshold for this ratio such that the false-positive rates (FPRs) are minimized to the point where human inspection of the final sample becomes feasible. We choose to require at least 60% of the subsequences to have a ZipperNet score above 0.76 for the candidate system to be included in our final sample of candidate LSNe. The 60% threshold and the 0.76 ZipperNet score threshold were determined simulta- neously by computing the LSN recall and galaxy FPR at all possible values. 4. Results 4.1. Performance Metrics To evaluate the performance of the fully trained ZipperNet, we define quantities and metrics of interest and compute them on the validation data set. We introduce two terms that describe classification score thresholds: “classified as an LSN” means the candidate system had a ZipperNet score greater than the threshold in at least 60% of subsequences; and “classified as background” means the candidate system had a ZipperNet than 60% of than the threshold in fewer score greater The Astrophysical Journal, 943:19 (14pp), 2023 January 20 Morgan et al. Figure 2. Left: receiver operating characteristic curve showing the lensed supernova (LSN) true-positive rate and LSN false-positive rate for all possible values of the ZipperNet operating threshold. The operating threshold of 0.760 is chosen to minimize the false positive rate to the point immediately prior to the true positive rate declining rapidly. Right: histograms of the scaled ZipperNet scores for each class in the validation data set. The selected operating threshold limits false positives from all systems in the negative class while keeping the majority of the positive class. Table 2 Metrics for Evaluating the Performance of ZipperNet and Our Final Sample Selection That are Robust against the Class Representations of the Validation Dataset Metric ZipperNet LSN Recall LSNIa Recall LSNCC Recall FPRGalaxy FPRSL FPR -SN Ia FPR -SN CC 0.8447 0.8426 0.8468 0.0157 0.2448 0.0049 0.0046 + Final 0.6113 0.5949 0.6273 0.0002 0.0046 0.0001a 0.0001a Equation Equation (1) Equation (2) Equation (2) Equation (3) Equation (3) Equation (3) Equation (3) Note. All metrics are defined in Section 4.1. a Indicates the use of an upper limit on the metric value resulting from limited statistical precision. the LSN-type-specific recall is LSN Recall type = TP type / ( TP type + FN type ) , 2 ( ) where type is “Ia” or “CC”; and the FPR for each type of negative class is FPR type = FP type / ( TN type + FN type ) , 3 ( ) where type is “Galaxy,” “SL,” “SN-Ia,” “SN-CC.” The values of these metrics are collected in Table 2 for ZipperNet alone and for the combination of ZipperNet with our final sample- selection criterion. There are a few key results from these metrics worth highlighting. The ZipperNet LSN recall indicates that approxi- mately 84% of all LSNe in the validation data set are scored above the operating threshold. The ZipperNet galaxy FPR indicates that roughly 1.5% of galaxies will be scored above our operating threshold and erroneously populate our candidate sample. By itself, the ZipperNet is a powerful classifier, but the minimized galaxy FPR is still large enough where the resulting candidate sample would be too large for visual inspection. With the selection criterion on the number of the addition of ZipperNet detections for each constituent subsequence, the performance is boosted. Critically, the final galaxy FPR is reduced, facilitating visual inspection of the full final candidate Figure 3. Number of time series image set subsequences scored above the ZipperNet threshold for each type of object in our validation data set. On average, LSNe time series image sets are scored above the ZipperNet threshold in a higher fraction of their subsequences than all types of negative examples. subsequences. We define the following terms regarding metrics based on the threshold score: 1. a true positive (TP) is an LSN, and it is classified as an LSN; 2. a false positive (FP) is a galaxy, galaxy–galaxy lens, or unlensed SN, and it is classified as an LSN; 3. a true negative (TN) is a galaxy, galaxy–galaxy lens, or unlensed SN, and it is classified as background; and 4. a false negative (FN) is an LSN, and it is classified as background. Using these quantities, common metrics like accuracy are straightforward to compute; however, those metrics are misleading due to the boosted representation of rare physical systems in our training and validation data sets. We instead focus on class-specific metrics that carry physical meaning and are robust against the class representation in the validation data set: the LSN recall is LSN Recall = TP / ( TP + FN ; ) 1 ( ) 6 The Astrophysical Journal, 943:19 (14pp), 2023 January 20 sample. This stricter selection has the consequence of reducing the final LSN recall. However, most of the removed LSNe are those that peak before or after the window of observations. 4.2. Searching the DES SN Fields Applying our trained ZipperNet and additional selection criterion to the DES SN field data produces 2245 candidate LSNe, approximately half of which had ZipperNet detections in multiple years of DES data. We expect the majority of these systems to have resolvable features based on two aspects of the analysis. First, these 2245 candidate LSNe were identified in the magnitude-limited sample of the DES galaxies, leading to a tendency for low-redshift, nearby galaxies to be more highly represented than high-redshift, distant galaxies. Second, based on the physical selection function of the ZipperNet on this data set (shown in Appendix A), LSNe in systems with large Einstein radii and better seeing are more likely to be recalled. Therefore, because the majority of the systems in this candidate sample should have resolvable features, human visual inspec- tion becomes a viable approach for identifying the most interesting candidate LSNe. A team of strong lensing experts within DES inspected six year coadded, color-composite images of the 2245 candidate LSNe systems to search for lensing, similar to how precursor strong lensing searches have been carried out. The team assigned all objects a score using the following system: 1. the detection is an image artifact, such as a diffraction spike or contamination from a bright foreground star; 2. there is a single object, such as a galaxy or star; 3. there are multiple objects with no evidence of lensing, such as SNe or clusters of galaxies; 4. there are multiple objects with evidence of lensing. Using this system and the median score for each object, the team of inspectors identified 522, 802, 871, and 50 objects with scores “1,” “2,” “3,” and “4,” respectively. For the 50 systems with evidence of lensing, we extracted aperture lightcurves for each object in each system from the DES single-epoch images. Three systems from the 50 systems with evidence of lensing were identified to have SN-like time variability, and we upgraded their overall score to a “5.” Candidate systems scored as a “4” or “5” are presented in Figures 4 and 5, respectively, and have their properties collected in Table B1. The 50 candidate systems scored at or above a “4” found in this analysis show evidence of lensing in their images. There are still non-lenses in this sample: for example, DES- 700492744 is a high-proper-motion white dwarf appearing as a red object between two blue point sources; nevertheless, we include all systems labeled as interesting by the labeling team for completeness. Some of these systems also show evidence of point sources within the lensing configurations: there are nearly circular objects positioned within the lensing configuration. Going further, we analyze the time variability of the candidate systems by extracting five year background-subtracted light- curves for each source in the images. The objects scored as a “5” show evidence of SN-like time variability: a short rise followed by a steady decay in brightness over the course of approximately one month as shown by Figure 5. The objects scored as a “4” do not show this temporal behavior; however, the possibility remains that some of the objects scored as a “4” are strongly lensed systems and potentially house a lensed Morgan et al. quasar. Section 4.3 contains a detailed presentation of the three objects scored as a “5”. Lastly, we cross-match the 2245 ZipperNet-identified systems with the systems identified during the DES five-year photometric SN Ia cosmology analysis (Möller et al. 2022). In Möller et al., difference imaging (Kessler et al. 2015) identified 31,636 transients and SALT-II SN Ia lightcurve fitting (Guy et al. 2010) identified 2381 single-season SNe from that sample of transients. The SNe selected by lightcurve fitting are more likely to be SNe Ia than SNe CC, and most SNe CC in the total sample are also excluded by the fitting. Furthermore, this selection procedure searches for normal SNe Ia and is not adapted for possible changes in the lightcurves from the lensing. In total, there is an overlap (using a 5″ radius) of 104 systems among the ZipperNet sample and the DES SN analysis transient sample. All but four overlapping systems—DES- 691702170, DES-699127397, DES-699340227, and DES- 700977591—were scored as either a “2” or a “3” by the labeling team, indicating no convincing evidence for lensing. The locations of the detected transients are marked in Figure 4. Only the transient in DES-699127397 passed the SALT-II SN Ia lightcurve fitting. The difference-imaging detections in DES- 699340227 and DES-700977591 appear to be spurious detections due to image subtraction errors. Lastly, while the transients detected in DES-691702170 and DES-699127397 are likely SNe, these systems do not appear to be lenses and likely should have received lower grades from the labeling team; DES-691702170 lacks an obvious lensing galaxy and the positions of the galaxies in DES-699127397 are more likely a cluster of galaxies than multiple images of the same back- ground galaxies due to their asymmetric alignment. Based on the SN FPRs in Table 2, this overlap is consistent with the expected ZipperNet SN background. The three systems scored as a “5” by the visual inspection team, indicating the presence of both lensing and SN-like temporal behavior, were not included in the overlapping sample. We believe the faintness of the SNe or foreground contamination for the lensing galaxy may have contributed to the non- detection from difference imaging, though a full understanding of this discrepancy is beyond the focus of our analysis. 4.3. Final LSN Candidates The three most interesting systems identified by the ZipperNet and subsequent human visual inspection are DES-691022126, DES-701263907, and DES-699919273. We present five-year color-composite coadded images of these systems and extract lightcurves for each object of interest within them in Figure 5. From the lightcurves, the five observing seasons of the DES SN program are easily distinguishable, and we refer to each observing season as “Y1” through “Y5.” We extract the lightcurves from the single-epoch images by summing the pixels in the aperture displayed in the coadded image, subtracting the sky back- ground measured by DES, and converting to a magnitude using the zero-point measured by DES. Importantly, the magnitudes are the combination of all objects within the aperture, so for example an SN lightcurve will contain contamination from its host galaxy. All estimated Einstein radii have been obtained by measuring the angular separa- tions between objects, as opposed to a full modeling of the lensing system. We choose to present only the z-band lightcurves for these visualizations for simplicity, though all 7 The Astrophysical Journal, 943:19 (14pp), 2023 January 20 Morgan et al. Figure 4. Candidate systems detected by ZipperNet that showed evidence of lensing but do not show SNe-like variability in their lightcurves. The properties of these candidates are collected in Table B1. Difference imaging detections from the DES SN group are shown with white star markers. four bands were assessed to identify SN-like temporal behavior. The bluer bands such as g and r have larger PSFs in this data set compared to the redder i and z bands, leading to noisier aperture photometry measurements. Furthermore, LSNe are likely to be at high redshifts, leading to a tendency for LSN temporal signatures to be most visible in the redder bands. DES-691022126 is a system of four objects labeled in the top panel of Figure 5 as A, B, C, and D. We interpret objects C and D to be galaxies based on their constant brightness over 8 The Astrophysical Journal, 943:19 (14pp), 2023 January 20 Morgan et al. Figure 5. Candidate LSNe identified by ZipperNet and human visual inspection. The aperture used to extract the magnitude measurement from each source is show and annotated on the coadded image. The properties of the candidates are collected in Table B1. time. Objects A and B are much redder, and display a greater degree of brightness variability when looking at the typical size of the magnitude error bars compared to the five-year median z- band magnitude for each object. Furthermore, the lightcurves for objects A and B both contain a period of linear decline in magnitude on month timescales: object A in Y5 and object B in Y3. ZipperNet detected the system in Y2 and Y3, but not in Y5. We believe it detected the linear decline of object B in Y3 and that perhaps object C contained light from a SN between Y2 and Y3 of which ZipperNet detected the beginning. The fact that the linear decline of object A’s brightness in Y5 was not detected by ZipperNet is likely due to object A being the faintest source in the system and the selection function of ZipperNet (see the bottom right panel of Figure A). The SN- like lightcurve features that are shared between objects A and B, when combined with the evidence for lensing with an Einstein radius of approximately 1 7, support the claim of the system as a LSN. By comparison, DES-701263907 is a much more compli- cated system, shown in the middle panel of Figure 5. A large foreground galaxy (SIMBAD source LEDA 135660) at redshift 0.03 dominates the image. Object B (SIMBAD source SDSS J024352.54-003708.4) is cataloged as a galaxy also at redshift 0.03, but may also be a dense, star-forming region based on its blue color. This dense area, combined with the gravitational potential of LEDA 135660 itself would have a large lensing cross-section, increasing the likelihood that background objects would be lensed. Because the lightcurve extraction method 9 The Astrophysical Journal, 943:19 (14pp), 2023 January 20 used in the lightcurves of Figure 5 does not subtract the effect of the host galaxy, the variability of these objects cannot be assessed without difference imaging techniques outside the scope of this paper. Nonetheless, we identify object A as the most variable source in the system given the foreground contamination. In the image cutout for DES-701263907, we also note the location of an SN detected in 2020 September (AT2020scq). It is possible that object B acts as a primary lensing galaxy, object A is an LSN identified by ZipperNet in 2018, and AT2020scq is a second appearance of object A delayed by approximately two years. Given the large fore- ground galaxies at redshift 0.03 and a potential Einstein radius of ≈ 3 0, this time delay would be consistent with an LSN. Lastly, DES-699919273 is another four-object system that we enumerate as A, B, C, and D in the bottom panel of Figure 5. We interpret object C as the lensing galaxy, object D as an image of the source galaxy without an SN, and objects A and B as images of the source galaxy, where an SN was present at some point during DES observations. The Einstein radius for this system is ≈ 2 1. Particularly, we note a linear decline in z- band magnitude for object A in Y3 and a nearly identical linear decline in z-band magnitude for object B in Y5. ZipperNet detected the linear decline in Y5, but had no such detection in Y3. We interpret this event as another manifestation of the less- than-perfect recall of the classifier. Nonetheless, the SN-like temporal signal appearing in two of the images within a lensing geometry is evidence for the presence of an LSN. 5. Discussion The method presented in this analysis contains a few areas where improvements could increase the LSN recall while decreasing the FPR. One such change is to add centroiding to account for sub-pixel-level shifts in position prior to stacking and averaging the images. While the offsets are small, the ≈0 25 scale can cause image-based misalignment at features to become less sharp and harder for a convolutional layer to identify. When stacking the images, it may also boost performance to only include the images with high image quality (e.g., seeing above some quality threshold and/or cloudiness below some quality threshold): this would ensure that the ability to resolve features in the resulting composite image is only limited by the instrumentation. These possibi- lities focus on improving the appearance of spatial features in the data to boost the ZipperNet’s ability to learn relationships and are motivated by the analysis of the physical selection function of our approach, which is described in Appendix A. The lightcurve extraction step of the data preprocessing also could be improved by discarding common artifacts such as diffraction spikes and saturated pixels to avoid contaminating the extracted brightness. Similarly, an analysis of the clarity of features in the lightcurves as a function of the aperture radius used in the lightcurve extraction may find that a different aperture radius leads to higher performance. It is possible that scaling the time series image sets to have a standardized mean and a standardized variance of pixel value prior to preproces- the sing would lead to smoother preprocessing steps used in this analysis down-selected images to standardize the cadence, though other approaches have demonstrated success with arbitrary numbers of images in the time series (Kodi Ramanah et al. 2022). Removing the need for a standardized cadence would greatly improve the applicability lightcurves. Finally, 10 Morgan et al. of this approach to real-time LSNe identification and remove the need for images to be discarded. It is possible that the machine-learning aspect of the analysis could be improved with subtle changes to the training set. For example, when simulating lensed systems, we made the the mass profile ellipticity simplifying approximation that was equivalent in angle and strength to the light profile. While there is likely to be a strong correlation between the mass and light profiles, the exclusion of training set examples with different relationships between mass profile ellipticity and light profile ellipticity may bias LSN selection to systems in which these quantities are highly correlated. We also employed a uniform distribution of Einstein radii, and it is possible that an approach such as that of Kodi Ramanah et al. (2022) with a physically motivated distribution could lead to improved performance. the first In consideration of a real-time LSN detection pipeline, a couple of changes to the methodology may improve perfor- mance. We envision the 10-epoch time series image sets being constructed as observations are ongoing: after a new image of a image in the time series is system is collected, discarded and a new 10-epoch sequence is created. There are two downsides to that approach: (1) there is an implicit requirement of 10 epochs before the trained ZipperNet can be utilized, and (2) our final selection criterion on the fraction of subsequences scored above the ZipperNet threshold requires additional epochs to create and track multiple subsequences. With the improvements to the preprocessing discussed above, it may be possible to sufficiently boost the ZipperNet perfor- mance to the point where the additional selection criterion can be removed. Furthermore, we did not experiment with time series image sets with fewer than 10 epochs, and it is possible that the analysis can be performed with a less strict requirement on the total number of epochs. With the current configuration, we have successfully reduced a catalog of 3,459,186 objects to 2245 with our deep learning approach, and proceeded to identify 50 systems of interest through human visual inspection, three of which show some evidence of an LSN. While we do not confirm or further characterize these three systems of interest, they all contain lensing features and the presence of point sources as found during the human visual inspection. Full characterization would entail Scene Modeling Photometry (Brout et al. 2019) to obtain lightcurves without host galaxy contamination, photometric classification of time-series photometry, redshift measurements for all objects in the system, and lens modeling, which are beyond the scope of this search. Because any detected LSNe would have faded by now, follow-up observations to confirm them are unlikely to provide any additional information apart from redshifts. However, several of the systems of interest were detected by ZipperNet in multiple years throughout DES operations. Therefore, these persistent lensed systems with point sources offer interesting candidates searches. The three most (DES-691022126, DES-701263907, interesting candidates and DES-699919273) are the most likely LSNe found by our ZipperNet in the magnitude-limited five-year DES data set utilized in this analysis. Given the approximate time delays and Einstein radii of the systems, spectroscopic redshifts and lens modeling could produce three independent measurements of H0. lensed quasar that for The Astrophysical Journal, 943:19 (14pp), 2023 January 20 light The ZipperNet architecture itself provides a new and powerful LSN identification tool going forward. The accom- panying code for this analysis (DZ1) also makes the data collection, processing, simulation, training, classification, and candidate selection routines available for future analyses. With first from the Vera C. Rubin Observatory quickly approaching, setting up a pipeline to detect LSNe is vital for time-delay cosmography measurements. The analysis presented here and suggested improvements provide a template for one such pipeline that would facilitate real-time detection of LSNe in short time series sequences of images without a dependence on traditional and computationally expensive image processing algorithms. 6. Conclusion This analysis presents the application of a deep learning LSN detection algorithm to an observed optical survey data set. The algorithm utilizes a novel neural network architecture called a ZipperNet that simultaneously learns characteristic features from image and temporal data to identify LSNe in DES data. Using a ZipperNet trained on simulated LSNe that are injected into the DES SN field data—along with a selection criterion on the number of detections for each system—our approach performs with an LSN recall of 61.13% and an FPR of 0.02%. This technique identified 2245 candidate LSN systems in the DES SN fields, and a human visual inspection found 50 systems of interest, three of which contained evidence of a time-variable lensed source. Confirmation of these candidates is left for future work, and these systems may of interest facilitate direct measurements of H0 when fully characterized. Looking to the Rubin Observatory era, the approach developed in DZ1 and implemented on the DES SN fields here has the potential to aid in the identification of several hundred LSNe. R.M. thanks the Universities Research Association Fermilab Visiting Scholars Program for funding his work on this project. R.M. also thanks the LSSTC Data Science Fellowship Program, which is funded by LSSTC, NSF Cybertraining grant #1829740, the Brinson Foundation, and the Moore Foundation; his participation in the program has benefited this work. We acknowledge the Deep Skies Lab as a community of multi-domain experts and collaborators who have facilitated an environment of open discussion, idea-generation, and colla- boration. This community was important for the development of this project. This material is based upon work supported by the National Science Foundation Graduate Research Fellowship Program under grant No. 1744555. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation. Work supported by the Fermi National Accelerator Labora- tory, managed and operated by Fermi Research Alliance, LLC under Contract No. DE-AC02-07CH11359 with the U.S. Department of Energy. The U.S. Government retains and the publisher, by accepting the article for publication, acknowl- edges that the U.S. Government retains a non-exclusive, paid- up, irrevocable, world-wide license to publish or reproduce the published form of this manuscript, or allow others to do so, for U.S. Government purposes. Morgan et al. Funding for the DES Projects has been provided by the U.S. Department of Energy, the U.S. National Science Foundation, the Ministry of Science and Education of Spain, the Science and Technology Facilities Council of the United Kingdom, the Higher Education Funding Council for England, the National Center for Supercomputing Applications at the University of Illinois at Urbana-Champaign, the Kavli Institute of Cosmo- logical Physics at the University of Chicago, the Center for Cosmology and Astro-Particle Physics at the Ohio State University, the Mitchell Institute for Fundamental Physics and Astronomy at Texas A&M University, Financiadora de Estudos e Projetos, Fundação Carlos Chagas Filho de Amparo à Pesquisa do Estado do Rio de Janeiro, Conselho Nacional de Desenvolvimento Científico e Tecnológico and the Ministério da Ciência, Tecnologia e Inovação, the Deutsche Forschungs- gemeinschaft and the Collaborating Institutions in the Dark Energy Survey. the University of Edinburgh, The Collaborating Institutions are Argonne National Labora- tory, the University of California at Santa Cruz, the University of Cambridge, Centro de Investigaciones Energéticas, Med- the University of ioambientales y Tecnológicas-Madrid, Chicago, University College London, the DES-Brazil Con- sortium, the Eidgenössische Technische Hochschule (ETH) Zürich, Fermi National Accel- Illinois at Urbana- the University of erator Laboratory, Champaign, the Institut de Ciències de l’Espai (IEEC/CSIC), the Institut de Física d’Altes Energies, Lawrence Berkeley National Laboratory, the Ludwig-Maximilians Universität München and the associated Excellence Cluster Universe, the University of Michigan, NSF’s NOIRLab, the University of Nottingham, The Ohio State University, the University of Pennsylvania, the University of Portsmouth, SLAC National Accelerator Laboratory, Stanford University, the University of Sussex, Texas A&M University, and the OzDES Membership Consortium. Based in part on observations at Cerro Tololo Inter- American Observatory at NSF?s NOIRLab (NOIRLab Prop. ID 2012B-0001; PI: J. Frieman), which is managed by the Association of Universities for Research in Astronomy (AURA) under a cooperative agreement with the National Science Foundation. The DES data management system is supported by the National Science Foundation under grant Numbers AST- 1138766 and AST-1536171. The DES participants from Spanish institutions are partially supported by MICINN under grants ESP2017-89838, PGC2018-094773, PGC2018-102021, and MDM-2015-0509, SEV-2016-0588, SEV-2016-0597, some of which include ERDF funds from the European Union. IFAE is partially funded by the CERCA program of the Generalitat de Catalunya. Research leading to these results has received funding from the European Research Council under the European Union’s Seventh Framework Program (FP7/ 2007-2013) including ERC grant agreements 240672, 291329, and 306478. We acknowledge support from the Brazilian Instituto Nacional de Ciência e Tecnologia (INCT) e-Universe (CNPq grant 465376/2014-2). This paper has gone through internal review by the DES collaboration. This manuscript has been authored by Fermi Research Alliance, LLC under Contract No. DE-AC02- 07CH11359 with the U.S. Department of Energy, Office of Science, Office of High Energy Physics. 11 The Astrophysical Journal, 943:19 (14pp), 2023 January 20 Software: astropy (Astropy Collaboration et al. 2013), deeplenstronomy (Morgan et al. 2021), lenstronomy (Birrer & Amara 2018; Birrer et al. 2021), matplotlib (Hunter 2007), numpy (Harris et al. 2020), pandas (McKinney 2010), PlotNeuralNet (Iqbal 2018), PyTorch (Paszke et al. 2019), Scikit-Learn (Pedregosa et al. 2011), scipy (Virtanen et al. 2020). Appendix A Lensed SN Physical Selection Function We present the selection function for our ZipperNet and our selection criterion for the number of detections. In this section, we analyze four central properties to understand the selection function of our approach: the Einstein radius, the seeing, the brightness of the source galaxy, and the brightness of the LSN. is calculated as a function of these four The LSN recall properties in Figure A1. We find that our approach has an easier time identifying larger Einstein radii than smaller Einstein radii. Similarly, we Morgan et al. find that our approach has an easier time identifying LSNe in good seeing conditions than in average or poor seeing these properties shed light on the conditions. Both of importance of the clarity of spatial features in the images. Poor seeing or small Einstein radii are both situations in which image resolution is compromised and consequently spatial features become difficult or impossible to realize. This trend in algorithm performance points to data quality characteristics as opposed to a selection bias introduced by our approach. Lastly, we find that source galaxy brightness has little impact on the performance for the range applicable to this magnitude-limited analysis. We observe a similar trend for LSNe brighter than 22 mag, but notice a reduction in performance for fainter LSNe. This magnitude threshold is near the single-epoch limiting magnitude for the DES SN fields and is likely due to Malmquist bias, but there could be second-order selection effects in the detectability of the LSN images. Figure A1. Physical selection function for simulated LSNe in our validation data set. We measure the LSN recall (defined in Section 4.1) as a function of Einstein radius, seeing, source galaxy unlensed magnitude, and LSN simulated unlensed magnitude. Error bars show a statistical uncertainty of one standard deviation. 12 The Astrophysical Journal, 943:19 (14pp), 2023 January 20 Morgan et al. Appendix B Candidate Metadata This appendix lists properties of systems detected by ZipperNet inspection and scored as a “4” or a “5” by human visual (Table B1). Table B1 Properties of the Systems Detected by ZipperNet That Received a Score of “4” or “5” by Human Visual Inspection No. Coadd Id. R.A. (deg.) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 DES-691022126 DES-701263907 DES-699919273 DES-690157493 DES-690456076 DES-690583502 DES-690642061 DES-690918939 DES-691032289 DES-691068769 DES-691440047 DES-691442767 DES-691524775 DES-691664180 DES-691702170 DES-691896609 DES-691902610 DES-691947063 DES-691968734 DES-692023723 DES-692243027 DES-692639734 DES-693331974 DES-693351134 DES-695852037 DES-696865317 DES-697161182 DES-697274399 DES-697446876 DES-697521552 DES-698587357 DES-698925976 DES-699088459 DES-699127397 DES-699219206 DES-699235372 DES-699340227 DES-699466457 DES-699478563 DES-699621639 DES-699723043 DES-699926736 DES-700364825 DES-700492744 DES-700541568 DES-700548040 DES-700863020 DES-700977591 DES-701328706 DES-701662201 53.898910 40.969218 10.155917 53.394495 55.010602 55.066197 54.464014 54.410941 54.420139 53.476146 55.134258 55.214489 54.476519 53.787326 54.607503 53.148074 52.888864 52.597080 52.958718 52.061383 53.581253 53.269373 52.064812 53.144038 34.616442 35.170551 35.442353 36.589451 36.990923 36.249985 7.138332 9.245344 7.447012 7.528665 9.869087 9.611497 10.241319 9.455661 8.862263 8.923322 10.869276 8.351206 42.449669 42.157070 41.136120 41.369059 41.521851 41.022356 42.256074 41.406713 Decl. (deg.) −28.912293 −0.619054 −44.437515 −26.716658 −26.549543 −27.347839 −27.418446 −28.417641 −29.048729 −29.331470 −29.302642 −29.331276 −29.498563 −29.869823 −29.829073 −27.405501 −27.157390 −27.675669 −27.916592 −28.071403 −27.886280 −28.909563 −28.509826 −28.581257 −4.670152 −6.631484 −6.948239 −3.896260 −4.185003 −4.382867 −42.415813 −43.112557 −43.647847 −43.465050 −43.142798 −43.351731 −43.413401 −43.915105 −44.227511 −44.133374 −44.022876 −43.536464 0.176652 −0.524332 −0.444152 −0.530169 −0.190829 −0.785606 −1.098086 −1.916363 Magi 21.73 17.31 18.95 21.82 20.79 21.44 21.62 20.65 21.80 20.89 20.72 21.79 21.37 21.05 19.42 22.04 22.37 22.22 21.93 22.20 21.18 21.01 22.47 20.66 21.25 20.56 22.25 20.39 22.47 22.23 21.02 20.18 18.95 21.21 20.47 20.46 20.32 22.19 20.67 19.12 21.71 21.14 20.46 19.30 21.63 21.29 20.52 18.75 20.76 21.23 Redshift L 0.030024 0.556 L L L L L L L L 0.139740 L L L 0.725 L L 0.610106 0.949 0.739 0.471666 L 0.815070 L L L 0.435 0.463 0.798 L 0.318563 L 0.657900 L L L 0.469 0.751 0.235 L L L L L L L 0.287564 L L Field Years Detected Inspection C2 S2 E2 C1 C1 C1 C1 C2 C2 C2 C2 C2 C2 C2 C2 C3 C3 C3 C3 C3 C3 C3 C3 C3 X1 X2 X2 X3 X3 X3 E1 E1 E1 E1 E2 E2 E2 E2 E2 E2 E2 E1 S1 S1 S2 S2 S2 S2 S2 S2 Y2 Y3 Y5 Y5 Y1 Y1 Y2 Y3 Y4 Y1 Y2 Y3 Y4 Y5 Y2 Y2 Y2 Y1 Y2 Y4 Y5 Y2 Y4 Y2 Y5 Y4 Y1 Y2 Y4 Y2 Y1 Y2 Y1 Y2 Y3 Y4 Y5 Y4 Y2 Y4 Y2 Y5 Y1 Y2 Y4 Y5 Y1 Y2 Y4 Y5 Y1 Y2 Y4 Y5 Y1 Y2 Y3 Y4 Y5 Y2 Y4 Y5 Y2 Y3 Y4 Y5 Y1 Y2 Y4 Y2 Y1 Y2 Y3 Y4 Y5 Y2 Y2 Y1 Y2 Y3 Y4 Y5 Y1 Y2 Y1 Y2 Y1 Y2 Y3 Y5 Y1 Y2 Y4 Y5 Y4 Y5 Y4 Y1 Y2 Y4 Y2 Y1 Y2 Y5 Y1 Y3 5 5 5 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 Note. The “Coadd Id.” is from the DES Y3 GOLD Catalog. The “Years Detected” indicate the years of DES data collection during which the candidate was selected by ZipperNet. The “Redshift” values are either photometric estimates from DES (shown to three significant digits) or spectroscopic measurements from OzDES (Yuan et al. 2015) and refer to the candidate lensing galaxy. 13 The Astrophysical Journal, 943:19 (14pp), 2023 January 20 Morgan et al. ORCID iDs https://orcid.org/0000-0002-7016-5471 https://orcid.org/0000-0001-6706-8972 https://orcid.org/0000-0001-8156-0429 https://orcid.org/0000-0001-8211-8608 https://orcid.org/0000-0003-3195-5507 R. Morgan B. Nord K. Bechtol A. Möller S. Birrer S. J. González M. Martinez R. A. Gruendl E. J. Buckley-Geer A. J. Shajib A. Carnero Rosell C. Lidman T. Collett M. Aguena J. Annis D. Bacon S. Bocquet D. Brooks M. Carrasco Kind J. Carretero F. J. Castander C. Conselice M. Costanzi J. De Vicente S. Desai B. Flaugher J. Frieman J. García-Bellido E. Gaztanaga D. Gruen G. Gutierrez S. R. Hinton D. L. Hollowood K. Honscheid K. Kuehn N. Kuropatkin O. Lahav F. Menanteau R. Miquel A. Palmese F. Paz-Chinchón A. Pieres A. A. Plazas Malagón A. K. Romer A. Roodman https://orcid.org/0000-0001-7282-3864 https://orcid.org/0000-0002-8397-8412 https://orcid.org/0000-0002-4588-6517 https://orcid.org/0000-0002-3304-0733 https://orcid.org/0000-0002-5558-888X https://orcid.org/0000-0003-3044-5150 https://orcid.org/0000-0003-1731-0497 https://orcid.org/0000-0001-5564-3140 https://orcid.org/0000-0001-5679-6747 https://orcid.org/0000-0002-0609-3987 https://orcid.org/0000-0002-2562-8537 https://orcid.org/0000-0002-4900-805X https://orcid.org/0000-0002-8458-5047 https://orcid.org/0000-0002-4802-3194 https://orcid.org/0000-0002-3130-0204 https://orcid.org/0000-0001-7316-4573 https://orcid.org/0000-0003-1949-7638 https://orcid.org/0000-0001-8158-1449 https://orcid.org/0000-0001-8318-6813 https://orcid.org/0000-0002-0466-3288 https://orcid.org/0000-0002-2367-5049 https://orcid.org/0000-0003-4079-3263 https://orcid.org/0000-0002-9370-8360 https://orcid.org/0000-0001-9632-0815 https://orcid.org/0000-0003-3270-7644 https://orcid.org/0000-0003-0825-0517 https://orcid.org/0000-0003-2071-9349 https://orcid.org/0000-0002-9369-4157 https://orcid.org/0000-0002-6550-2023 https://orcid.org/0000-0003-0120-0808 https://orcid.org/0000-0003-2511-0946 https://orcid.org/0000-0002-1134-9035 https://orcid.org/0000-0002-1372-2534 https://orcid.org/0000-0002-6610-4836 https://orcid.org/0000-0002-6011-0530 https://orcid.org/0000-0003-1339-2683 https://orcid.org/0000-0001-9186-6042 https://orcid.org/0000-0002-2598-0514 https://orcid.org/0000-0002-9328-879X https://orcid.org/0000-0001-5326-3486 E. Sanchez I. Sevilla-Noarbe M. Smith E. Suchyta G. Tarle https://orcid.org/0000-0002-9646-8198 https://orcid.org/0000-0002-1831-1953 https://orcid.org/0000-0002-3321-1432 https://orcid.org/0000-0002-7047-9358 https://orcid.org/0000-0003-1704-0781 References Abbott, T., Abdalla, F. B., Aleksić, J., et al. 2016, MNRAS, 460, 1270 Abbott, T. M. C., Abdalla, F. B., Allam, S., et al. 2018, ApJS, 239, 18 Abbott, T. M. C., Adamów, M., Aguena, M., et al. 2021, ApJS, 255, 20 Amanullah, R., Goobar, A., Clément, B., et al. 2011, ApJL, 742, L7 Astropy Collaboration, Robitaille, T. P., Tollerud, E. J., et al. 2013, A&A, 558, A33 Birrer, S., & Amara, A. 2018, PDU, 22, 189 Birrer, S., Shajib, A. J., Gilman, D., et al. 2021, JOSS, 6, 3283 Brout, D., Sako, M., Scolnic, D., et al. 2019, ApJ, 874, 106 Diehl, H. T. 2020, The Dark Energy Survey and Operations: Year 6 – The Finale. https://www.osti.gov/biblio/1596042/ Flaugher, B., Diehl, H. T., Honscheid, K., et al. 2015, AJ, 150, 150 Goobar, A., Amanullah, R., Kulkarni, S. R., et al. 2017, Sci, 356, 291 Graham, M. J., Kulkarni, S. R., Bellm, E. C., et al. 2019, PASP, 131, 078001 Guy, J., Sullivan, M., Conley, A., et al. 2010, A&A, 523, A7 Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020, Natur, 585, 357 Hartley, W. G., Choi, A., Amon, A., et al. 2021, MNRAS, 509, 3547 Hunter, J. D. 2007, CSE, 9, 90 Iqbal, H. 2018, PlotNeuralNet v1.0.0, Zenodo, doi:10.5281/zenodo.2526396 Jones, D. O., Foley, R. J., Narayan, G., et al. 2021, ApJ, 908, 143 Ivezić, Ž., Kahn, S. M., Tyson, J. A., et al. 2019, ApJ, 873, 111 Kelly, P. L., Rodney, S. A., Treu, T., et al. 2015, Sci, 347, 1123 Kessler, R., Bassett, B., Belov, P., et al. 2010, PASP, 122, 1415 Kessler, R., Marriner, J., Childress, M., et al. 2015, AJ, 150, 172 Kodi Ramanah, D., Arendse, N., & Wojtak, R. 2022, MNRAS, 512, 5404 Marshall, P., Anguita, T., Bianco, F. B., et al. 2017, arXiv:1708.04058 McKinney, W. 2010, in Proc. of the 9th Python in Science Conf., 445, ed. Stéfan van der Walt & Jarrod Millman (Austin, TX: SciPy), 56 Mihalas, D. 1963, PASP, 75, 256 Möller, A., Smith, M., Sako, M., et al. 2022, MNRAS, 514, 5159 Morgan, R. 2022, DeepZipper II: Analysis and Data Processing Code, Zenodo, doi:10.5281/zenodo.6419722 Morgan, R., Nord, B., Bechtol, K., et al. 2022, ApJ, 927, 109 Morgan, R., Nord, B., Birrer, S., Lin, J. Y.-Y., & Poh, J. 2021, JOSS, 6, 2854 Oguri, M. 2019, RPPh, 82, 126901 Paszke, A., Gross, S., Massa, F., et al. 2019, arXiv:1912.01703 Pedregosa, F., Varoquaux, G., Gramfort, A., et al. 2011, JMLR, 12, 2825 Quimby, R. M., Oguri, M., More, A., et al. 2014, Sci, 344, 396 Refsdal, S. 1964, MNRAS, 128, 307 Rodney, S. A., Brammer, G. B., Pierel, J. D. R., et al. 2021, NatAs, 5, 1118 Rodney, S. A., Patel, B., Scolnic, D., et al. 2015, ApJ, 811, 70 Rojas, K., Savary, E., Clément, B., et al. 2022, A&A, 668, A73 Tarsitano, F., Hartley, W. G., Amara, A., et al. 2018, MNRAS, 481, 2018 Treu, T. 2010, ARA&A, 48, 87 Virtanen, P., Gommers, R., Oliphant, T. E., et al. 2020, NatMe, 17, 261 Yuan, F., Lidman, C., Davis, T. M., et al. 2015, MNRAS, 452, 3047 14
10.1007_s11538-023-01121-y
Bulletin of Mathematical Biology (2023) 85:18 https://doi.org/10.1007/s11538-023-01121-y O R I G I N A L A R T I C L E On Parameter Identifiability in Network-Based Epidemic Models István Z. Kiss1 · Péter L. Simon2,3 Received: 15 August 2022 / Accepted: 3 January 2023 / Published online: 27 January 2023 © The Author(s) 2023 Abstract Modelling epidemics on networks represents an important departure from classical compartmental models which assume random mixing. However, the resulting models are high-dimensional and their analysis is often out of reach. It turns out that mean- field models, low-dimensional systems of differential equations, whose variables are carefully chosen expected quantities from the exact model provide a good approx- imation and incorporate explicitly some network properties. Despite the emergence of such mean-field models, there has been limited work on investigating whether these can be used for inference purposes. In this paper, we consider network-based mean-field models and explore the problem of parameter identifiability when obser- vations about an epidemic are available. Making use of the analytical tractability of most network-based mean-field models, e.g. explicit analytical expressions for leading eigenvalue and final epidemic size, we set up the parameter identifiability problem as finding the solution or solutions of a system of coupled equations. More precisely, subject to observing/measuring growth rate and final epidemic size, we seek to iden- tify parameter values leading to these measurements. We are particularly concerned with disentangling transmission rate from the network density. To do this, we give a condition for practical identifiability and we find that except for the simplest model, parameters cannot be uniquely determined, that is, they are practically unidentifiable. This means that there exist multiple solutions (a manifold of infinite measure) which give rise to model output that is close to the data. Identifying, formalising and analyt- ically describing this problem should lead to a better appreciation of the complexity involved in fitting models with many parameters to data. Keywords Epidemics · Inference · Identifiability B István Z. Kiss i.z.kiss@sussex.ac.uk 1 Department of Mathematics, University of Sussex, Falmer, Brighton BN1 9QH, UK 2 Institute of Mathematics, Eötvös Loránd University, Budapest, Hungary 3 Numerical Analysis and Large Networks Research Group, ELKH-ELTE, Budapest, Hungary 123 18 Page 2 of 17 1 Introduction I. Z. Kiss, P. L. Simon Differential-equation-based models are widespread in modelling population dynamics be that in problems arising in ecology, evolution or epidemiology (Anderson and May 1992; Blasius et al. 2007; Diekmann and Heesterbeek 2000; Kiss et al. 2017). Such systems are relatively straightforward to set up and the theory of dynamical systems offers tools to analyse them. Over the past two decades, differential-equation-based models have gained a lot of popularity in modelling epidemics on networks (Porter and Gleeson 2016; Kiss et al. 2017). Such models, often referred to mean-field models, aim to approximate the expected behaviour of some quantities of interest (e.g. expected number of infected individuals in time) and rely on closure assumptions which are needed to produce tractable systems. The major difference between mean-field models arising from modelling epidemics on networks and classic compartmental models is that in the former the assumption of homogeneous random mixing between individuals can be relaxed. This sometimes comes at the expense of having to keep track of multiple variables, such as the number of nodes with different number of contacts (Pastor-Satorras and Vespignani 2001) and different disease status or write down differential equations for all nodes in the network (Van Mieghem et al. 2008). The complexity of such network-based mean-field models is highly dependent on the heterogeneity in the contact network and how the modelling is performed. For example, the edge-based compartmental model (Miller et al. 2012) is able to retain all the information about the distribution of contacts in the form of the corresponding probability generating function and the resulting system consists of one single differential equation. While such models have been studied extensively and have provided the means to understand the impact of contact heterogeneity on the epidemic threshold, final epidemic size and other epidemic characteristics, there has been relatively little work on using such models for inference purposes. For example, often information about the network of contacts is not available or patchy. Hence, plac- ing such models in an inference framework where network parameters are inferred along disease dynamic parameters may reveal important information about the under- lying contact network which can be used for the design and implementation of control measures. Many of the network-based mean-field models provide explicit or implicit analytical expressions for quantities such as the basic reproduction number (or leading eigenvalue based on the linear stability analysis around the disease-free steady state), timing and/or peak prevalence, final epidemic size etc. Hence, given a synthetic or real epidemic and being able to measure a number of the aforementioned quantities, it is of interest to investigate whether parameters of the epidemic model, including that of the contact network, that generated the data can be inferred or determined. In this paper, we focus on the pairwise (Keeling 1999) and the edge-based compartmental model (Miller et al. 2012). These choices are motivated by the popularity and wide use of the pairwise model and the compact nature of the edge-based compartmental model. Fitting epidemic models to synthetic or real-world data is of great interest as it allows us to infer model parameters which in turn helps us to (i) learn more about the disease, (ii) implement and test control scenarios via simulations, and (iii) make short- or long-term predictions about the epidemic (Chowell 2017; King et al. 2015). In many 123 On Parameter Identifiability in Network-Based Epidemic Models Page 3 of 17 18 0.4 0.2 l e c n e a v e r P 0 0 0.1 0.05 s e s a c y l i a D 0 0 20 40 60 Time 20 40 60 Time Fig. 1 (Color Figure Online) Illustration of how distinct pairs of average degree and transmission rate, (n, τ ), lead to almost indistinguishable time evolution of the prevalence and daily new cases. Baseline values of the parameters are: average degree n = 6, τ = γ R0/((n − 1) − R0) = 0.1429, with R0 = 2.5, rate of recovery, γ = 1/7, number of nodes N = 10000 and epidemic started with one infected individual, with the corresponding output shown by the thick grey lines. The black and red-dashed lines correspond to (n, τ ) = (8.46, 0.09) and (n, τ ) = (2.454, 1.091), respectively cases, such models can and will be used for parameter estimation and prediction and can suffer of the well-known problem of parameter redundancy and unidentifiability (Cole 2019; Villaverde et al. 2016; Gallo et al. 2022). This problem has also been highlighted in network-based epidemic models, for example in Britton and O’Neill (2002). This problem is not model specific. For example, in Fig. 1, we show that for the pairwise model, Eqs. (3)–(6), it is possible to find distinct sets of parameters whereby the time evolution of prevalence and daily new cases are near indistinguishable. Of course this also implies that the initial growth rate and final epidemic size are also close. We note that this figure is for illustration purposes only. The pairwise model is discussed in the main body of the paper, and its full understanding in the context of the figure is not necessary. Parameter identifiability can be considered in two different ways. On the one hand, different parameter values may lead to identical observations, which is called struc- tural, or a priori unidentifiability (Anstett-Collin et al. 2020; Villaverde et al. 2016). On the other hand, the observations belonging to two sets of parameters can be very close to each other, referred to as practical unidentifiability, see e.g. (Wieland et al. 2021). Struc- tural identifiability has been studied in several epidemic models. For example, in Roosa and Chowell (2019), the authors consider the problem of parameter identifiability in a number of increasingly complex compartmental epidemic models. As the number of states in the model increases so does the number of parameters. While the parame- ters remained identifiable, in particular the basic reproduction number, the uncertainty around the estimate increased in models with more parameters. The paper (Massonis et al. 2021) relates the identifiability problem to observability in a higher dimensional augmented system, and investigates structural identifiability in more than fifty com- partmental epidemic models by using the formalism of observability-identifiability conditions. On the other hand in Gallo et al. (2022), the authors provide a framework to quantify how the uncertainty in the data affects the determination of the parame- ters and the evolution of the unmeasured variables of a given model. Their approach allows them to characterise different regimes of identifiability and argue that in some 123 18 Page 4 of 17 I. Z. Kiss, P. L. Simon cases, such as COVID-19 the lack of identifiability may prevent reliable predictions of the epidemic dynamics. Finally, in Villaverde et al. (2016), the authors argue that structural identifiability in every model should be checked before using the model for inference. But this is seldom done since it involves either complex analytical or numerical calculations. In this paper, we show that practical unidentifiability is present in a number of network-based epidemic models. This is, however, not due to hidden or unmeasured variables. Moreover, our inference is making use of available analytical formulas for leading eigenvalue, or equivalently growth rate, and final epidemic size. While many of the previous works are concerned with local changes, that is, quantifying change in observations induced by a small local change in parameter values, we show that in our models varying the parameters globally leads to small local changes in measurement. The paper is structured as follows. In Sect. 2, we describe the general mathematical approach and suggest some definition and ways to formalise the identifiability problem. In Sect. 3, we start with simple models such as the well-mixed susceptible-infected- recovered (SIR) compartmental model, followed by more complex models such as, the pairwise (Sect. 4), and the edge-based compartmental model (Sect. 5). We show that except for the simplest of models, there are clear parameter identifiability problems which we map out and explain analytically, where possible. In models with a larger number of parameters, it is often the case that many different combinations of the model parameters (with many individual parameters being far from their true values) result in output which is consistent with the true epidemic. Finally, we provide some discussion and future directions of research. 2 General Approach We are given a system of ODEs involving some parameters: ˙x(t) = f (x(t), μ), where x(t) ∈ Rn is the state vector of the system and μ ∈ Rk is the vector of parameters. We observe a derived quantity (e.g. final epidemic size, growth rate) for which data is available. This is given by an observation function h : Rn → Rm, i.e. the observation y is y(t) = h(x(t), μ). The goal is to solve the inverse problem, namely to determine the parameter μ based on the observation y(t); note that the observation does not need to be time dependent. This is line with the formulation of a general inverse problem, see (Cole 2019). Our question here is parameter identification, namely to understand whether it may happen that the observations y and y corresponding to different parameters μ and μ are identical or very close to each other. The first one is called structural (or a priori) unidentifiability, while the second one is referred to as practical unidentifiability, see 123 On Parameter Identifiability in Network-Based Epidemic Models Page 5 of 17 18 e.g. (Wieland et al. 2021). Our main focus here is practical identifiability, and hence we deal with structural identifiability only briefly. The main idea of investigating structural identifiability can be explained as follows. Since the observation y(t) is known for all time values in an interval, its derivatives are also known (measured). Differentiating the equation of the observation function and substituting t = 0 yields ˙y(0) = h(cid:4)(x(0), μ) f (x(0), μ) that is an equation for the n + k unknowns: the coordinates of x(0) and those of μ. Hence we need further equations to determine the parameters. Differentiating y(t) n + k − 1 times leads to n + k equations for the n + k unknowns. (Note that the zeroth- order derivative can also be used.) These derivatives are called the Lie derivatives of the output along the trajectories of the governing dynamical system. This system of equations is nonlinear, and hence its unique solvability is determined by the Jaco- bian matrix containing the partial derivatives with respect to the coordinates of x(0) and those of μ. This matrix is called the Observability-identifiability Matrix, see e.g. (Massonis et al. 2021). The application of the implicit function theorem yields that the parameters can be locally uniquely determined if this matrix has full rank, i.e. the Observability-identifiability Condition (OIC) holds. More detailed approaches lead to slightly different definitions, the relationships of which are studied in Anstett-Collin et al. (2020). The OIC condition has been studied in the case of many compartmental epidemic models. An exhaustive summary is presented in Massonis et al. (2021). As an illustration of the results derived there, we mention that the traditional SIR model is structurally identifiable when the observation is the number of infected individuals, I (t). However, in the more realistic case when one can observe only an unknown proportion of the infected individuals, q I (t), the parameter q is not structurally iden- tifiable. It is important to note that the definition of structural identifiability is related to Kalman’s observability condition in an augmented system (where the phase space is extended by new artificial variables representing the parameters), see e.g. (Villaverde et al. 2016). Let us turn now to the main focus of our study, practical identifiability, when significantly different parameter values yield observation which are almost identical. Generally speaking, defining practical identifiability, we use some ε accuracy of the observation. The accuracy of the observation can be measured in two different ways. The first is when the time dependence of the observation is known for all time values, or at least for an observation time-window and accuracy is defined as some norm of the difference of the two functions, see e.g. (Gallo et al. 2022; Wieland et al. 2021). The second is when we have formulas for some characteristic quantities of the observation. For example, the derivative of the quantity being observed at the initial instant, ˙y(0), or its limit for large time, y(∞). These formulas typically involve the unknown parameter values and hence define a system of equations for them. We note that structural identifiability can be defined in this case as well, namely the parameters can be identified by the model, if this system can be uniquely solved for the parameters. We speak about practical unidentifiability when clearly distinct parameter values satisfy the above system of equations but with some small error. 123 18 Page 6 of 17 I. Z. Kiss, P. L. Simon This can be formulated as follows. Let the system of equations for the parameters take the form F(μ) = 0. We call the problem structurally unidentifiable if the system of equations F(μ) = 0 has more than one (typically infinitely many) solution. The problem is called practically unidentifiable if the system |F(μ)| < ε is satisfied by a large set of μ values for any ε > 0. In fact, we will show that in our cases the set of μ values solving |F(μ)| < ε has infinite measure. We note that this does not exclude that the equation F(μ) = 0 has a unique solution. This notion of unidentifiability is related but not identical to the question of sensitive dependence on parameters, which is a notion of local nature. That is expressed in terms of the Jacobian of F at the solution of F(μ) = 0. The problem fits into the framework of error analysis and sensitivity analysis that are widely studied important fields of parameter inference. We refer the interested reader to Stigter and Molenaar (2015) and to the books (Cacuci et al. 2005; Einarsson 2005), in which both the introduction to the topic and elaborated examples are available. Comparing our definition to those in Gallo et al. (2022), Wieland et al. (2021), the main novelty in ours is that the inequality |F(μ)| < ε holds globally in the parameter space. Another difference between our definition and previous ones is that both (Gallo et al. 2022; Wieland et al. 2021) infer parameters from the time dependence of the solutions, while we use exact (not numerical) formulas for some characteristic quantities (leading eigenvalue and final epidemic size). Hence, the parameter inference is done by solving a system of equations instead of fitting to time-dependent curves. We apply this general theory to the SIR compartmental, pairwise and edge-based compartmental models when the observations are the leading eigenvalue at the disease- free steady state and the final epidemic size. We note that the leading eigenvalue of the disease-free steady state results from the linear stability analysis around it. The largest eigenvalue of the resulting Jacobian is the leading eigenvalue. For the three different models analysed in this paper, these are given in Eqs. (1), (7) and (18). The parameters to be determined, given these observations, are the infection rate τ , the recovery rate γ and the average degree of the underlying network, n. We will show that for these models (in fact for several other models as well) the leading eigenvalue can be expressed in terms of these parameters as λ = τl(n) − γ , where l(n) is a linear function depending on the model. We will derive an implicit equation for the final epidemic size in each case. It will turn out that this implicit equation contains the parameters τ and γ linearly and n in a nonlinear way. The equation can be written in the form τ = A(n)γ for all cases, where A(n) is a nonlinear function. The problem of parameter identification can be formulated as follows. Depending on the choice of the model, i.e. the choice of the functions l(n) and A(n), can the parameters be recovered by solving the two equations above? Since we have two equations for three parameter values, it is obvious that one of the parameters has to be 123 On Parameter Identifiability in Network-Based Epidemic Models Page 7 of 17 18 assumed to be given. The recovery rate is more appropriate for being a known value since it depends on epidemiological characteristics. While τ and especially n are more dependent on the behaviour of the agents and on the network, these are more difficult to determine. Our goal will be to solve the above equations for τ and n with a given value of γ and also with given initial conditions. (We note that the initial conditions could also be considered as parameters which makes the problem even more complicated in a real-life situation.) These equations define two curves in the (n, τ ) parameter plane. The parameter values leading to the desired values of the final size and leading eigenvalue can be obtained as the intersection point of the two curves. The main novelty of the paper is the observation that these curves are very close to each other, and hence relatively different parameter values may yield very similar final size and leading eigenvalue. Thus, noisy data may preclude the correct identification of the values of these parameters. The above system is linear in τ and γ when n is considered to be known. Hence its solvability is easy to check by computing the determinant. For the sake of complete- ness, this will also be carried out below in each case. 3 Identifiability in the Compartmental SIR Model The well-known SIR compartmental model takes the form ˙S = −τ n I ˙I = +τ n I S N S N , − γ I . Simple differentiation at the disease-free steady state (S = N , I = 0) yields that the leading eigenvalue is λ = τ n − γ . On the other hand, the final epidemic size is given by the solution of the following implicit equation R∞ = N − S0 exp (−τ n R∞/N γ ) as it is given in (4.12) in Kiss et al. (2017). Let us assume, for sake of simplicity, that S0 = N , that is, initially there are very few infected and recovered nodes. Then the final size equation can be rearranged to -τ nr∞ = γ ln(1 − r∞), where we introduced the fraction r∞ = R∞/N . Thus the system relating the measured characteristic quantities λ and R∞ to the parameters, τ , γ and n takes the form τ n − γ = λ, τ nr∞ + γ ln(1 − r∞) = 0. (1) (2) This system is linear in τ and γ , and hence apart from exceptional cases it has a unique solution for τ and γ if n is known and the characteristic quantities of the epidemic, λ and R∞, are measured. That is, knowing/measuring the leading eigenvalue and final epidemic size, it is possible to uniquely determine τ and γ . However, the parameters τ and n cannot be obtained from this system, since only their product is determined by the equations. That is, knowing/measuring the leading eigenvalue and final epidemic 123 18 Page 8 of 17 I. Z. Kiss, P. L. Simon size, it is not possible to determine the infection rate τ and average degree n. This is the case of structural unidentifiability when the system of equations has infinitely many solutions (if it has a solution at all). 4 Identifiability in the Pairwise SIR Model The pairwise model focuses on a hierarchical construction where expected number of nodes in state A at time t, [A](t), depends on the expected number of pairs of various types (e.g. [AB]) and then, these in turn depend on triples such as [ABC]. Here the counting is done in all possible directions meaning that [SS] pairs are counted twice and and that [SI] = [I S]. With this in mind, the pairwise model becomes (see e.g. in Kiss et al. (2017)) [ ˙S] = −τ [SI]; [ ˙I ] = τ [SI] − γ [I ]; [ ˙R] = γ [I ], [ ˙S I ] = −(τ + γ )[SI] + τ ([SSI] − −[ISI]); [ ˙SS] = −2τ [SSI]. This system is not self-consistent as pairs depend on triples and equations for these are needed. To tackle this dependency on higher-order moments, the triples in the equation above are closed using the following relation: [ASB] = κ [AS][SB] [S] , where A, B ∈ {S, I }. Common choices for κ are (n − 1)/n and 1. In Kiss et al. (2022), it was shown that both these closures are exact in the limit of large networks when the contact distribution is binomial and Poisson, respectively. We use the former as the average degree appears explicitly and is subject to inference. Applying κ = n−1 n leads to [ ˙S] = −τ [SI], [ ˙I ] = τ [SI] − γ [I ], [ ˙S I ] = −(τ + γ )[SI] + τ n − 1 n [ ˙SS] = −2τ n − 1 n [SS][SI] [S] , [SI]([SS] − [SI]) [S] , (3) (4) (5) (6) which is now a self-contained system. The leading eigenvalue, resulting from the linear stability analysis around the disease-free steady state, ([S], [I ], [SS], [SI]) = (N , 0, n N , 0)), can be easily com- puted from Eq. (3)-(6) as λ = τ (n − 2) − γ . (7) An implicit equation for the final number of recovered and susceptible nodes can be derived as it is shown in Section 4.3.4 in Kiss et al. (2017), see also below 123 On Parameter Identifiability in Network-Based Epidemic Models Page 9 of 17 18 N (τ + γ ) (S∞) 2 1 n = N γ S n 0 (S∞) 1 2 n + τ S∞ S n 0 . (8) Equation (8) there yields the final number of susceptible nodes, S∞ = N − R∞. Let us assume again, for sake of simplicity, that S0 = N , that is, initially there are very few infected and recovered nodes. Then introducing s∞ = S∞/N in Eq. (8) leads to (cid:2) ∞ − s2/n s1/n ∞ s∞ − s2/n ∞ = 0. + γ (cid:2) (cid:3) (cid:3) τ Thus the system relating the measured characteristic quantities λ and s∞ to the param- eters, τ , γ and n takes the form (cid:2) τ s∞ − s2/n ∞ (cid:3) + γ τ (n − 2) − γ = λ, (cid:2) ∞ − s2/n s1/n ∞ = 0. (cid:3) (9) (10) This system is linear in τ and γ , and hence apart from exceptional cases it has a unique solution for τ and γ if n is known and the characteristic quantities of the epidemic, λ and R∞, are measured. That is, knowing/measuring the leading eigenvalue and final epidemic size, it is possible to uniquely determine τ and γ . Let us turn now to the parameters τ and n. Now γ is considered to be given, and the characteristic quantities of the epidemic, λ and s∞ are measured. We can express τ from the equations above yielding , λ + γ τ = n − 2 ∞ − s2/n τ = γ s1/n ∞ s2/n ∞ − s∞ . (11) (12) In order to show unidentifiability visually, let us plot the curves given by the above equations in the (τ, n) plane. We can see in Fig. 2 (bottom panel) that the two curves are practically indistinguishable. In fact, they have a single intersection point, i.e. the system has a unique solution, but any value of τ yields a value of n on the hyperbola- like curve, that is an approximate solution with high accuracy. In fact the experiment that we setup here, and in some of the cases that follow, is that we start with a known set of parameters, often referred to as master set of values. These generate a particular numerical value for the lead eigenvalue, final epidemic size and time evolution of the prevalence or daily new cases. We then ask the questions: are there any other parameter combination (τ, n) that give rise to daily new cases in time that are similar to that obtained by using the master values? The top panel in Fig. 2 shows the Euclidean distance between the master daily cases vector and those resulting from (τ, n) pairs chosen between the bounds seen in the figure. There are several important features to note about the surface showing the distances. First, there is a clear hyperbola-like valley of minimum points, where any choice of (τ, n) seems to be close enough to the output based on the master values. Several minima are observed which indicate that any kind of optimiser may struggle to find 123 18 Page 10 of 17 I. Z. Kiss, P. L. Simon 10 8 6 4 2 0 1 0.9 0.8 0.75 0.65 0.6 0.5 0.4 0.2 ( ,n ) m m L. Eigv FES 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Fig. 2 (Color Figure Online) Top panel: illustration of the distance profile D between the daily new infections in time for a fixed set of values (τm , nm ) = (γ R0/((n − 1) − R0) = 0.1429, 6) (magenta diamond), with γ = 1/7 and R0 = 2.5, compared to daily new cases for different choices for the values of (τ, n) pair. Distance measured using an Euclidian norm scaled by the population size N = 10000. Bottom panel: the same as above as contour plot with two additional curves given by the equations for the lead eigenvalue (denoted as L. Eigv in the legend) and final epidemic size (denoted as FES in the legend) Eqs. (11) and (12), respectively, where the lead eigenvalue λ and s∞ are calculated with (τm , nm ) as given above the global optimum. Of course in this thought experiment, there is a unique (τ, n) pair that makes D = 0. However, given noisy observations, it is easy to see that any values along the hyperbola-like valley may return an acceptable fit, such as the one in Fig. 2. The empirical experiment and observations above can be made more substantial by considering the bottom panel in Fig. 2. The contour plot is based on the same data as in the surface plot above but with the addition of two curves: that of the leading eigenvalue and final epidemic size, which have unique numerical values determined by the master values and fixed γ . It is clear that these two curves are indeed close to each other and that they capture the hyperbola-like valley of small values in distance. Beyond this visualisation of unidentifiability, we formally prove it in terms of the definition given in Sect. 2. First, we reduce system (11)–(12) to a single equation as follows: 123 On Parameter Identifiability in Network-Based Epidemic Models Page 11 of 17 18 Fig. 3 (Color Figure Online) Plots of function f (n). For different values of s∞ 1.4 1.2 1 0 20 40 60 where λ + γ = γ f (n), ∞ − s2/n f (n) = s1/n ∞ s2/n ∞ − s∞ (n − 2). We can assume without loss of generality that the two curves have a common point, i.e. there is a value n∗ of n satisfying λ + γ = γ f (n∗). Otherwise, the measurement was so inaccurate that no values of τ and n could lead to the measured value of λ and s∞. Thus the single equation to be solved for the unknown n takes the form f (n) = f (n∗). We will prove that this equation does not identify the value of n in the practical sense. In order to do so, we determine the characteristic properties of function f . These properties can be easily visualised by plotting the graph of the function for n > 2, see Fig. 3. It turns out that the function is very close to a constant and its value changes only slightly from n = 2 to infinity. For example, in the case s∞ = 0.9, the functions grow from 1.027 (at n = 2) to 1.054 as n tends to infinity, so the function is constant with accuracy 0.027. Simple application of L’Hospital’s rule yields that the limits of f as n tends to 2 or to infinity exist and their values are f (n) = 2 √ s∞ − s∞ s∞ ln s∞ := f2 f (n) = ln s∞ s∞ − 1 := f∞ lim n→2 lim n→∞ The next proposition expresses the fact that the measure of the range of this function is small. Proposition 1 There exists a number 0 < a < 1 such that s∞ > a implies that f is increasing and f2 < f (n) < f∞ for all n > 2. That is, the range of f is the interval ( f2, f∞). 123 18 Page 12 of 17 I. Z. Kiss, P. L. Simon Proof Introducing a = s∞, x = 1/n and the function g(x) = a2x − ax a − a2x (cid:5) − 2 , (cid:4) 1 x we have f (n) = g(1/n), leading to f (cid:4)(n) = −g(cid:4)(1/n) 1 that g(cid:4)(x) < 0 for all x ∈ (0, 1/2). Simple differentiation shows that g(cid:4)(x) < 0 is equivalent to n2 . Hence it is enough to prove (cid:5) (cid:6) − 2 (2a2x − ax )(a − a2x ) + 2a2x (a2x − ax ) (cid:4) 1 x (cid:7) ln a < 1 x 2 (a2x − ax )(a − a2x ), that can be rearranged to (by multiplying by x 2) (cid:6) (1 − 2x) 2a2x+1 − ax+1 − a3x (cid:7) ln ax < (a2x − ax )(a − a2x ). Introducing the new variable b = ax and returning to n instead of x, the desired inequality takes the form (after dividing by b3) 0 < n(1 − b)(1 − bn−2) + (n − 2)(1 + bn−2 − 2bn−1) ln b := h(b). This newly defined function satisfies h(1) = 0, and elementary differentiation shows that h(cid:4)(1) = 0 = h(cid:4)(cid:4)(1). Moreover, the inequality h(cid:4)(cid:4)(cid:4)(1) < 0 holds. Based on this inequality, it is easy to check that h is positive in a left neighbourhood of 1, that is, there exists a number b < 1, such that h(b) > 0 holds when b < b < 1. Let us define the desired number a as a = b . On the other hand, n > 2 and b < 1 imply that b2 > bn, hence b2 > bn > b , yielding b > b leading to h(b) > 0. This is equivalent to g(cid:4)(x) < 0 that we wanted to prove. (cid:8)(cid:9) . Then a > a is equivalent to bn > b 2 2 2 We note that numerical evidence shows that the number a given by the proposition is relatively small, e.g. a < 0.1. That is, for reasonable values of s∞ the assertions of the proposition hold. The proposition yields practical unidentifiability as follows. The value of γ is considered to be given, and the characteristic quantities of the epidemic, λ and s∞ are measured. These determine the unique intersection point (n∗, τ ∗) of the curves given by (11)–(12). In other words, n∗ is the trivial solution of the reduced single equation f (n) = f (n∗). An approximate solution n satisfies | f (n) − f (n∗)| < ε with a given positive value of ε. The proposition implies that | f (n) − f (n∗)| < ε holds for any n > 2 if ε > f∞ − f2, which is a small number. An even smaller ε is achieved if the measured data λ and s∞ yield a value of n∗ which is larger, i.e. f (n∗) is closer to f∞. Then the value of ε can be chosen as ε = f∞ − f (n∗) and then | f (n) − f (n∗)| < ε holds for n values in a half-line, i.e. in a set of measure infinity. This was defined as practical unidentifiability. 123 On Parameter Identifiability in Network-Based Epidemic Models Page 13 of 17 18 10 8 6 4 2 L. Eigv FES Master 10 8 6 4 2 L. Eigv FES Master 0 0.5 1 0 0.2 0.4 Fig. 4 (Color Figure Online) Left panel: From left to right curves correspond to solving Eq. (11) and (12) with the leading eigenvalue and the final epidemic size being set to values obtained by using τ = 0.26, 0.33, 0.47, γ = 1 and n = 6. Right panel: Curves given by Eqs. (21) and (22) for values of the transmission rate τ = 0.03, 0.045, 0.07 (from left to right). Other parameters are γ = 1/7 and n = 6. For both plots, the black curve represent (τ, n) pairs where the leading value is that determined by the master values shown as a diamond magenta. Similarly, the red star represent (τ, n) pairs where the final epidemic size is equal to that given by the master values In Fig. 4, we explore the dependency of the practical unidentifiability in the pairwise model on the precise parameters used in the model. The left panel of this figure shows that this feature seems to hold for different parameter combinations and that we can find infinitely many (τ, n) pairs that lead to a desired eigenvalue and final epidemic size. Moreover, we emphasise again that the two curves do overlap to a great extent and over a large range of parameters. Before investigating the same problem in a different model, we note that the same calculations for the leading eigenvalue and final epidemic size can be done when the pairwise model is closed with κ = 1. These calculations lead to τ = τ = λL + γ n − 1 , [SS](0) ([S](0))2 γ S∞(ln(S∞) − ln([S](0)) (cid:2) (S∞)2 − S∞(ln(S∞) − ln([S](0)) − (13) (14) [S](0) + [SS](0) [SI](0) [S](0) (cid:3) . S∞ By using the disease-free initial condition, [S](0) = N , [SS](0) = n N , [SI](0) = 0 and using that s∞ = S∞/N , the equations above lead to τ = τ = λL + γ n − 1 , γ ln(s∞) ns∞ − ln(s∞) − n . (15) (16) It turns out that the formulas above are identical to those that we obtain later on for the edge-based compartmental model. 123 18 Page 14 of 17 I. Z. Kiss, P. L. Simon 5 Identifiability in the Edge-Based Compartmental Model The edge-based compartmental model is given by ˙θ = −τ θ + τ φS(0) ψ (cid:4)(θ ) ψ (cid:4)(1) + γ (1 − θ ) + τ φR(0) = f (θ ), (17) where θ denotes the probability that a random neighbour ν of a random, initially susceptible test node u has not yet passed infection to u. Furthermore, φS(0) and φR(0) are the probabilities that, at t = 0, the random neighbour ν of a random, initially susceptible test node u is susceptible and recovered, respectively. The disease-free steady state is given by θ (0) = 1, φR(0) = 0, θ (0) = 1 and φS(0) = 1. However, in order to generate an epidemic curve, when the system is above the epidemic threshold, one can perturb the steady state above by setting φS(0) = 1 − ε. We now consider the case of ψ(x) = exp(n(x − 1)), that is, a network with Poisson degree distribution with mean n. Linearising around θ = 1, we obtain f (cid:4)(θ )|θ=1 = τ φS(0) ψ (cid:4)(cid:4)(1) ψ (cid:4)(1) − τ − γ = τ n2 n − τ − γ = τ (n − 1) − γ = λ, (18) The final epidemic size can also be worked out by finding limt→∞ θ (t) = θ∞ and using that the final proportion of susceptible left in the population is s∞ = ψ(θ∞). Setting the right-hand side of Eq. (17) to zero, an implicit equation for θ∞ follows: (τ + γ )θ∞ − γ − τ ψS(0)en(θ∞−1) = 0. (19) Since s∞ = exp(n(θ∞ − 1)), Eq. (19) can be recast in terms of s∞ and yields τ nψS(0)s∞ − (τ + γ ) ln(s∞) − τ n = 0. (20) We are now in a position to write down a system of equations based on (18) and (20) τ = τ = λ + γ n − 1 , γ ln(s∞) nψS(0)s∞ − ln(s∞) − n (21) (22) These curves are shown in the right panel of Fig. 4. It can be seen that the two curves are close to each other. The coincidence is more emphasised when s∞ is larger, i.e. the final epidemic size is smaller. Beyond this visualisation of unidentifiability, we formally prove that in terms of the definition given in Sect. 2. First, we reduce the above system to a single equation as follows: λ + γ = γ q f (n), 123 On Parameter Identifiability in Network-Based Epidemic Models Page 15 of 17 18 where f (n) = n − 1 n − q , and q = ln s∞ s∞ − 1 > 1. We can assume without loss of generality that the two curves have a common point, i.e. there is a value n∗ of n satisfying λ + γ = γ q f (n∗). Otherwise, the measurement was so inaccurate that no values of τ and n could lead to the measured value of λ and s∞. Thus the single equation to be solved for the unknown n takes the form f (n) = f (n∗). This equation does not identify the value of n in the practical sense. By plotting the graph of f , it turns out that the function is very close to a constant, its value changes only slightly from large values of n to infinity. For example, in the case s∞ = 0.9, the functions change from 1.006 (at n = 10) to 1 as n tends to infinity, so the function is constant with accuracy 0.006 in the infinite half-line n > 10. In general, one can directly see that f is decreasing and its limit is 1 as n tends to infinity. Similarly to the case of the pairwise model, practical unidentifiability follows from the fact that the function f is to a constant. 6 Discussion In this paper, we study the identifiability of parameters in network-based epidemic models. We find that network density and the transmission rate cannot be disentan- gled. More formally this means that when considering these parameters, the model is structurally not identifiable. Preliminary analysis suggests that combinations of n and τ and other parameters are better behaved, for example when packaged into the expres- sion for R0; this is in line with how to deal with identifiability problems (Villaverde et al. 2016). In an ideal situation, the leading eigenvalue and final epidemic size can be measured to any desired accuracy. Assuming that this is the case, an exhaustive search in the parameter space, again to arbitrary precision, would be able to identify the precise parameters which generated the data. However, real-life observations are noisy and even a small measurement error can lead to a significant shift in the values of the inferred parameters. This leads to practical unidentifiability. Contact patterns and the transmission of the disease across a link are strongly related and often are difficult to disentangle. Intuitively, it is known that dense networks with low transmission rate and spare networks with high transmission rate can produce similar epidemics. In fact, our hyperbolas trace out and connect these regimes. Of course, in this case a Bayesian approach may alleviate the problem in the sense that good informative priors are likely to reduce the dimensionality of the parameter space or at least the range of parameters. With more and more mobility data becoming available as well as data from contact surveys, contact networks can be characterised sufficiently in order to produce meaningful estimates from complex models. 123 18 Page 16 of 17 I. Z. Kiss, P. L. Simon In terms of future work, we believe that there is value in carrying out a systematic search over the parameter space to identify areas, both in terms of parameter subsets and ranges in parameter values, where the unidentifiability is the most significant. Our preliminary analysis shows that this is both model and parameter dependent. We also note that unidentifiability seems to be more marked for less severe epidemics. For larger epidemics, the overlap between the two hyperbolas decreases, meaning that parameters are easier to identify. Furthermore, similar analysis can be extended to different disease dynamics such the susceptible-exposed-infected-recovered (SEIR) model or alternative network-based mean-field models. Acknowledgements István Z. Kiss acknowledges support from the Leverhulme Trust for the Research Project Grant RPG-2017-370. Péter L. Simon acknowledges support from the Hungarian Scientific Research Fund, OTKA, (grant no. 135241) and from the Ministry of Innovation and Technology NRDI Office within the framework of the Artificial Intelligence National Laboratory Program. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. References Anderson RM, May RM (1992) Infectious diseases of humans: dynamics and control. Oxford university press Anstett-Collin F, Denis-Vidal L, Millérioux G (2020) A priori identifiability: an overview on definitions and approaches. Annu Rev Control 50:139–149 Blasius B, Kurths JJ, Stone L (2007) Complex population dynamics: nonlinear modeling in ecology, Epi- demiology, and Genetics, vol 7. World Scientific Britton T, O’Neill PD (2002) Bayesian inference for stochastic epidemics in populations with random social structure. Scand J Stat 29(3):375–390 Cacuci DG, Ionescu-Bujor M, Navon IM (2005) Sensitivity and uncertainty analysis, volume II: applications to large-scale systems. CRC press Chowell G (2017) Fitting dynamic models to epidemic outbreaks with quantified uncertainty: a primer for parameter uncertainty, identifiability, and forecasts. Infect Dis Model 2(3):379–398 Cole DJ (2019) Parameter redundancy and identifiability in hidden markov models. Metron 77(2):105–118 Diekmann O, Heesterbeek JAP (2000) Mathematical epidemiology of infectious diseases In: model building, analysis and interpretation, vol 5. John Wiley & Sons Einarsson Bo (2005) Accuracy and reliability in scientific computing. SIAM Gallo L, Frasca M, Latora V, Russo G (2022) Lack of practical identifiability may hamper reliable predictions in covid-19 epidemic models. Sci Adv 8(3):eabg5234 Keeling MJ (1999) The effects of local spatial structure on epidemiological invasions. Proc R Soc Lond B 266(1421):859–867 King AA, De Cellés MD, Magpantay FMG, Rohani P (2015) Avoidable errors in the modelling of outbreaks of emerging pathogens, with special reference to Ebola. Proceed Royal Soc B Biol Sci 282(1806):0–6 Kiss Istvan Z, Kenah Eben, Rempala Grzegorz A (2022) Necessary and sufficient conditions for exact closures of epidemic equations on configuration model networks. arXiv preprint arXiv:2208.07983 Kiss István Z, Miller Joel C, Simon Péter L et al (2017) Mathematics of epidemics on networks. Cham: Springer 598:31 123 On Parameter Identifiability in Network-Based Epidemic Models Page 17 of 17 18 Massonis G, Banga JR, Villaverde AF (2021) Structural identifiability and observability of compartmental models of the covid-19 pandemic. Annu Rev Control 51:441–459 Miller JC, Slim AC, Volz EM (2012) Edge-based compartmental modelling for infectious disease spread. J R Soc Interface 9(70):890–906 Pastor-Satorras R, Vespignani A (2001) Epidemic spreading in scale-free networks. Phys Rev Lett 86(14):3200 Porter Mason A, Gleeson James P (2016) Dynamical Systems on Networks. pp. 49–51 Roosa K, Chowell G (2019) Assessing parameter identifiability in compartmental dynamic models using a computational approach: application to infectious disease transmission models. Theor Biol Med Model 16(1):1–15 Stigter JD, Molenaar J (2015) A fast algorithm to assess local structural identifiability. Automatica 58:118– 124 Van Mieghem P, Omic J, Kooij R (2008) Virus spread in networks. IEEE/ACM Trans Netw 17(1):1–14 Villaverde AF, Barreiro A, Papachristodoulou A (2016) Structural identifiability of dynamic systems biology models. PLoS Comput Biol 12(10):e1005153 Wieland F-G, Hauber AL, Rosenblatt M, Tönsing C, Timmer J (2021) On structural and practical identifi- ability. Curr Opin Syst Biol 25:60–69 Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. 123
10.3390_vision7010010
Communication Full-Thickness Compressive Corneal Sutures with Removal of Anterior Chamber Air Bubble in the Management of Acute Corneal Hydrops Zahra Ashena 1 , Ritika Mukhija 1 and Mayank A. Nanavaty 1,2,* 1 2 Sussex Eye Hospital, Brighton & Sussex University Hospitals NHS Trust, Eastern Road, Brighton BN2 5BF, UK Brighton & Sussex Medical School, University of Sussex, Falmer, Brighton BN1 9PX, UK * Correspondence: mayank.nanavaty@nhs.net Abstract: Acute hydrops is a rare complication of corneal ectatic disease, which occurs secondary to Descemet membrane break. Spontaneous resolution of this condition is associated with longstanding ocular discomfort and corneal scar. Intracameral gas/air injection with or without corneal suturing, anterior segment ocular coherence tomography (ASOCT)-guided drainage of intrastromal fluid, and penetrating keratoplasty are some of the described surgical interventions to manage this condition. The purpose of our study was to assess the effect of full-thickness corneal suturing as a solo treatment in the management of acute hydrops. A total of five patients with acute hydrops received full- thickness corneal sutures perpendicular to their Descemet break. A complete resolution of symptoms and corneal oedema was observed between 8 to 14 days post-operation with no complications. This technique is simple, safe, and effective in the management of acute hydrops and saves patients from a corneal transplant in an inflamed eye. Keywords: full thickness suturing; keratoconus; hydrops 1. Introduction Acute corneal hydrops is a visually debilitating complication of ectatic corneal diseases, which results from a tear in the Descemet’s membrane (DM) followed by the entry of aqueous humour into the corneal stroma [1]. Vernal keratoconjunctivitis (VKC), steeper keratometry, atopy, Down syndrome, and eye rubbing are important risk factors for corneal ectasia such as keratoconus where corneal hydrops can develop [2,3]. Patients present with intense photophobia, pain, and reduced visual acuity due to significant corneal oedema [4]. Formerly, the management of hydrops was often conserva- tive as it can resolve spontaneously with extensive scarring. Medical treatment, including topical steroids, antibiotics, cycloplegics, hypertonic saline, ocular antihypertensive, and lubricants, aims to reduce inflammation and provide symptomatic relief. However, spon- taneous resolution of corneal hydrops takes longer (up to 36 weeks) [5], and persistent oedema causes prolonged discomfort and complications such as infection, scarring, neo- vascularization, and permanent visual loss [6,7]. Some interventions have been described in the literature to accelerate the resolution of hydrops and hasten visual recovery. These include injection of intra-cameral air or gas with or without compressive sutures, pre-Descemet suturing [8], anterior segment ocular coherence tomography (ASOCT)-guided drainage of intra-stromal fluid [9,10], am- niotic membrane transplant [11], mini-Descemet’s membrane endothelial keratoplasty (DMEK) [12], penetrating keratoplasty, and, in the rare event of corneal perforation, applica- tion of tissue adhesives [6]. We aim to describe the outcomes of full-thickness compressive corneal suturing as a primary and sole treatment in managing acute corneal hydrops. Citation: Ashena, Z.; Mukhija, R.; Nanavaty, M.A. Full-Thickness Compressive Corneal Sutures with Removal of Anterior Chamber Air Bubble in the Management of Acute Corneal Hydrops. Vision 2023, 7, 10. https://doi.org/10.3390/ vision7010010 Received: 16 October 2022 Revised: 10 January 2023 Accepted: 20 January 2023 Published: 28 January 2023 Copyright: © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). Vision 2023, 7, 10. https://doi.org/10.3390/vision7010010 https://www.mdpi.com/journal/vision vision Vision 2023, 7, 10 2 of 9 2. Methods/Surgical Procedure All patients had ASOCT at presentation to locate the Descemet tear. However, this diagnostic tool was not conclusive in a few cases with significant corneal oedema and stromal cleft. After obtaining informed consent, the theatre procedure was performed under local or general anaesthesia, based on the patient’s preference. A small air bubble was injected into the anterior chamber to visualize the Descemet tear, which revealed the break/s and demonstrated their orientation and length within 10 min. This technique was especially beneficial in cases with extensive oedema, where ASOCT failed to locate the Descemet break. This was followed by placing 3 to 6 full-thickness interrupted 10-0 nylon corneal sutures perpendicular to the Descemet break/s (Video S1, Figure 1a–f). Care was taken not to touch the crystalline lens. The sutures were tied tighter than usual, expecting some loosening on the resolution of corneal oedema and the knots were buried in the stroma. The small air bubble was removed at the end of the procedure, and intracameral cefuroxime was injected. The eyes were covered with a clear shield. Patients were prescribed topical Tobramycin and Dexamethasone combination eye drops (Tobradex®, Alcon Laboratories, Fort Worth, Texas, USA) four times a day for two weeks, and a follow-up was arranged in one week. Figure 1. (a) Injecting the air bubble in the AC via paracentesis, (b,c) the appearance of two Descemet breaks, (d) placing the perpendicular full-thickness suture, (e) making the suture tight, and (f) four full-thickness suture perpendicular to the Descemet breaks. 3. Report of the Cases Case 1: A 23-year-old Caucasian male, known keratoconus since the age of 17, pre- sented with acute hydrops in his left eye. His uncorrected visual acuity (UCVA) was counting fingers (CF) on examination. He received four full-thickness corneal compres- sive sutures under general anaesthesia (GA). The procedure helped provide significant symptomatic relief and reduced his apex pachymetry on OCT from 1096 µ to 466 µ in 10 days (Figure 2a). A total of four weeks later, the sutures were loose and hence removed (Figure 2b). Two months after the initial presentation, his UCVA was CF with further Vision 2023, 7, x FOR PEER REVIEW 2 of 10 2. Methods/Surgical Procedure All patients had ASOCT at presentation to locate the Descemet tear. However, this diagnostic tool was not conclusive in a few cases with significant corneal oedema and stromal cleft. After obtaining informed consent, the theatre procedure was performed un-der local or general anaesthesia, based on the patient’s preference. A small air bubble was injected into the anterior chamber to visualize the Descemet tear, which revealed the break/s and demonstrated their orientation and length within 10 min. This technique was especially beneficial in cases with extensive oedema, where ASOCT failed to locate the Descemet break. This was followed by placing 3 to 6 full-thickness interrupted 10-0 nylon corneal sutures perpendicular to the Descemet break/s (Video S1, Figure 1a–f). Care was taken not to touch the crystalline lens. The sutures were tied tighter than usual, expecting some loosening on the resolution of corneal oedema and the knots were buried in the stroma. The small air bubble was removed at the end of the procedure, and intracameral cefuroxime was injected. The eyes were covered with a clear shield. Patients were pre-scribed topical Tobramycin and Dexamethasone combination eye drops (Tobradex®, Al-con Laboratories, Fort Worth, Texas, USA) four times a day for two weeks, and a follow-up was arranged in one week. Figure 1. (a) Injecting the air bubble in the AC via paracentesis, (b,c) the appearance of two Descemet breaks, (d) placing the perpendicular full-thickness suture, (e) making the suture tight, and (f) four full-thickness suture perpendicular to the Descemet breaks. 3. Report of the Cases Case 1: A 23-year-old Caucasian male, known keratoconus since the age of 17, pre-sented with acute hydrops in his left eye. His uncorrected visual acuity (UCVA) was counting fingers (CF) on examination. He received four full-thickness corneal compres-sive sutures under general anaesthesia (GA). The procedure helped provide significant symptomatic relief and reduced his apex pachymetry on OCT from 1096 µ to 466 µ in 10 days (Figure 2a). A total of four weeks later, the sutures were loose and hence removed (Figure 2b). Two months after the initial presentation, his UCVA was CF with further Vision 2023, 7, 10 3 of 9 improvement to 6/18 with pinhole, similar to his pre-hydrops visual acuity. He was not keen on rigid gas permeable contact lenses (RGPCL) since he had a good unaided vision in his fellow eye. Figure 2. (a) Anterior segment OCT: pre-op (left) with apex pachymetry of 1096 µ and 10 days post-op (right) with apex pachymetry of 466 µ. The arrow shows the level of the scan. (b) Loose corneal sutures after resolution of corneal oedema. Case 2: A 20-year-old female of Afro-Caribbean origin, diagnosed with keratoconus since the age of 18, presented with acute hydrops in her right eye. Her best spectacle- corrected visual acuity (BSCVA) was hand movement (HM). Her maximum keratometry increased from 83 dioptres (D) to 124D, and her apex pachymetry changed from 340 µ to 1380 µ. ASOCT confirmed the presence of stromal clefts (Figure 3). Conservative management was started, and within one week she received four full-thickness compressive corneal sutures, as described above. There was a complete resolution of her symptoms and corneal oedema in 8 days. The sutures were removed in four weeks. She was not keen on RGPCL. Her final BSCVA improved to 6/36 in the affected eye, one line lower than her pre-hydrops vision of 6/24. Case 3: A 41-year-old male of South Asian origin, with a strong history of atopy and known keratoconus since the age of 19, presented with acute hydrops in his left eye. His best-corrected visual acuity (BCVA) from 6/9 with scleral contact lenses pre- hydrops dropped to hand movement. His apex pachymetry increased to 1473 µ from 398 µ. ASOCT showed large stromal clefts (Figure 4a). He received six full-thickness compressive corneal sutures perpendicular to his Descemet break two days after presentation. The com- pressive sutures significantly helped with his symptoms and corneal oedema (Figure 4b) and his apex pachymetry decreased to 532 µ in two weeks. His corneal sutures were removed in 4 weeks, and his final BCVA, using a scleral contact lens, was CF three months from hydrops. Vision 2023, 7, x FOR PEER REVIEW 3 of 10 improvement to 6/18 with pinhole, similar to his pre-hydrops visual acuity. He was not keen on rigid gas permeable contact lenses (RGPCL) since he had a good unaided vision in his fellow eye. Figure 2. (a) Anterior segment OCT: pre-op (left) with apex pachymetry of 1096 µ and 10 days post-op (right) with apex pachymetry of 466 µ. The arrow shows the level of the scan. (b) Loose corneal sutures after resolution of corneal oedema. Case 2: A 20-year-old female of Afro-Caribbean origin, diagnosed with keratoconus since the age of 18, presented with acute hydrops in her right eye. Her best spectacle-corrected visual acuity (BSCVA) was hand movement (HM). Her maximum keratometry increased from 83 dioptres (D) to 124D, and her apex pachymetry changed from 340 µ to 1380 µ. ASOCT confirmed the presence of stromal clefts (Figure 3). Conservative manage-ment was started, and within one week she received four full-thickness compressive cor-neal sutures, as described above. There was a complete resolution of her symptoms and corneal oedema in 8 days. The sutures were removed in four weeks. She was not keen on RGPCL. Her final BSCVA improved to 6/36 in the affected eye, one line lower than her pre-hydrops vision of 6/24. Vision 2023, 7, 10 4 of 9 Figure 3. Anterior segment OCT: pre-op with a stromal cleft (top 2 scans) and apex pachymetry of 1380 µ, and 8 days post-op (bottom 2 scans) with apex pachymetry of 548 µ. The arrow shows the level of the scan. Figure 4. (a) Anterior segment OCT: pre-op (top) with apex pachymetry of 1473 µ, 15 days post-op (middle) with apex pachymetry of 532 µ, and 8 weeks post-op (bottom) with apex pachymetry of 402 µ; (b) Anterior segment photo: pre-op (left) and 15 days post-op (right). The arrow shows the level of the scan. Vision 2023, 7, x FOR PEER REVIEW 4 of 10 Figure 3. Anterior segment OCT: pre-op with a stromal cleft (top 2 scans) and apex pachymetry of 1380 µ, and 8 days post-op (bottom 2 scans) with apex pachymetry of 548 µ. The arrow shows the level of the scan Case 3: A 41-year-old male of South Asian origin, with a strong history of atopy and known keratoconus since the age of 19, presented with acute hydrops in his left eye. His best-corrected visual acuity (BCVA) from 6/9 with scleral contact lenses pre-hydrops dropped to hand movement. His apex pachymetry increased to 1473 µ from 398 µ. ASOCT showed large stromal clefts (Figure 4a). He received six full-thickness compressive corneal sutures perpendicular to his Descemet break two days after presentation. The compres-sive sutures significantly helped with his symptoms and corneal oedema (Figure 4b) and his apex pachymetry decreased to 532 µ in two weeks. His corneal sutures were removed in 4 weeks, and his final BCVA, using a scleral contact lens, was CF three months from hydrops. Vision 2023, 7, x FOR PEER REVIEW 5 of 10 Figure 4. (a) Anterior segment OCT: pre-op (top) with apex pachymetry of 1473 µ, 15 days post-op (middle) with apex pachymetry of 532 µ, and 8 weeks post-op (bottom) with apex pachymetry of 402µ; (b) Anterior segment photo: pre-op (left) and 15 days post-op (right). The arrow shows the level of the scan Case 4: A 36-year-old Caucasian male, diagnosed with keratoconus at 27 years of age, presented with right eye acute hydrops, where conservative treatment was initiated. He was thereafter reviewed for a follow-up in one month in the cornea clinic when he was still symptomatic with pain and photophobia. On assessment, ASOCT showed significant stromal oedema and his BCVA had dropped from 6/9 with scleral contact lenses pre-hy-drops to counting fingers (CF). His apex pachymetry measured 1194 µ. He was offered compressive corneal suturing perpendicular to his Descemet break. In nine days, his apex pachymetry decreased to 430 µ. The corneal sutures were removed five weeks later (Fig-ure 5a), and his BCVA with scleral contact lenses improved to 6/9, which was similar to his pre-hydrops vision. Case 5: A 41-year-old Caucasian male, known with keratoconus since the age of 22 years, presented with left corneal hydrops. His VA at presentation was HM with no im-provement with pinhole, and his apex pachymetry was 1024 µ. He received four compres-sive sutures to his Descemet break seven days from the presentation. Two weeks post-procedure, his discomfort had settled, and apex pachymetry was measured at 493 µ (Fig-ure 5b,c). Since then, he has returned to his local hospital, and we have no further infor-mation on his visual outcome. Vision 2023, 7, 10 5 of 9 Case 4: A 36-year-old Caucasian male, diagnosed with keratoconus at 27 years of age, presented with right eye acute hydrops, where conservative treatment was initiated. He was thereafter reviewed for a follow-up in one month in the cornea clinic when he was still symptomatic with pain and photophobia. On assessment, ASOCT showed significant stromal oedema and his BCVA had dropped from 6/9 with scleral contact lenses pre- hydrops to counting fingers (CF). His apex pachymetry measured 1194 µ. He was offered compressive corneal suturing perpendicular to his Descemet break. In nine days, his apex pachymetry decreased to 430 µ. The corneal sutures were removed five weeks later (Figure 5a), and his BCVA with scleral contact lenses improved to 6/9, which was similar to his pre-hydrops vision. Figure 5. (a) Anterior segment photo of the right eye 4 weeks post-suturing (left) and after suture removal (right); (b) Anterior segment OCT: pre-op (left) with apex pachymetry of 1024 µ and 14 days post-op (right) with apex pachymetry of 493 µ; (c). Anterior segment photo of left eye post-suturing. Case 5: A 41-year-old Caucasian male, known with keratoconus since the age of 22 years, presented with left corneal hydrops. His VA at presentation was HM with no improvement with pinhole, and his apex pachymetry was 1024 µ. He received four compressive sutures to his Descemet break seven days from the presentation. Two weeks post-procedure, his discomfort had settled, and apex pachymetry was measured at 493 µ (Figure 5b,c). Since then, he has returned to his local hospital, and we have no further information on his visual outcome. 4. Discussion Conventional management of acute hydrops mainly focused on providing symp- tomatic relief. This included topical sodium chloride eye drops to decrease corneal oedema, cycloplegic eye drops to address the photosensitivity, bandage contact lenses to help with the pain, aqueous suppressants to lower the intraocular pressure and encourage corneal dehydration, topical steroids and non-steroids to alleviate the inflammation, and topical antibiotics to prevent the secondary infection [1,4,6]. However, recovery from acute corneal Vision 2023, 7, x FOR PEER REVIEW 6 of 10 Figure 5. a. Anterior segment photo of the right eye 4 weeks post-suturing (left) and after suture removal (right); b. Anterior segment OCT: pre-op (left) with apex pachymetry of 1024 µ and 14 days post-op (right) with apex pachymetry of 493 µ; c. Anterior segment photo of left eye post-suturing. 4. Discussion Conventional management of acute hydrops mainly focused on providing sympto-matic relief. This included topical sodium chloride eye drops to decrease corneal oedema, cycloplegic eye drops to address the photosensitivity, bandage contact lenses to help with the pain, aqueous suppressants to lower the intraocular pressure and encourage corneal dehydration, topical steroids and non-steroids to alleviate the inflammation, and topical antibiotics to prevent the secondary infection [1,4,6]. However, recovery from acute cor-neal hydrops may take a few months without surgical intervention. This is associated with prolonged pain, photosensitivity, stromal vascularization, and inflammation and can lead to corneal scaring and worse visual outcomes. Emergency penetrating keratoplasty used to be an alternative surgical option to manage acute hydrops [1]. However, this is not the recommended treatment at present as no report has been published on the survival of penetrating keratoplasty in acute hydrops, and corneal transplant in an inflamed eye is associated with a significantly higher risk of rejection and failure [13]. A few surgical interventions have been described to hasten the recovery in acute cor-neal hydrops. Miyata et al. tried intracameral air injection and although the final visual outcome was similar between their intervention and control group, a significantly quicker resolution of corneal oedema was observed in the intervention group, but a few injections were required [14]. Shah et al. used perfluoropropane (C3F8) to save patients from re-peated intracameral air bubble injections. Two intracameral injections of non-expansile C3F8 gas were given five days apart, and complete resolution of corneal oedema was no-ticed shortly after the second injection without any complications [15]. Basu et al. con-ducted a comparative study on using non-expansile C3F8 gas to manage acute hydrops in 62 eyes and compared the outcome with 90 eyes in the control group. The resolution of symptoms and corneal oedema was significantly quicker in the intervention group, with Vision 2023, 7, 10 6 of 9 hydrops may take a few months without surgical intervention. This is associated with prolonged pain, photosensitivity, stromal vascularization, and inflammation and can lead to corneal scaring and worse visual outcomes. Emergency penetrating keratoplasty used to be an alternative surgical option to manage acute hydrops [1]. However, this is not the recommended treatment at present as no report has been published on the survival of penetrating keratoplasty in acute hydrops, and corneal transplant in an inflamed eye is associated with a significantly higher risk of rejection and failure [13]. A few surgical interventions have been described to hasten the recovery in acute corneal hydrops. Miyata et al. tried intracameral air injection and although the final visual outcome was similar between their intervention and control group, a significantly quicker resolution of corneal oedema was observed in the intervention group, but a few injections were required [14]. Shah et al. used perfluoropropane (C3F8) to save patients from re- peated intracameral air bubble injections. Two intracameral injections of non-expansile C3F8 gas were given five days apart, and complete resolution of corneal oedema was noticed shortly after the second injection without any complications [15]. Basu et al. con- ducted a comparative study on using non-expansile C3F8 gas to manage acute hydrops in 62 eyes and compared the outcome with 90 eyes in the control group. The resolution of symptoms and corneal oedema was significantly quicker in the intervention group, with reversible pupillary block and elevated IOP being the main complication of their intervention. No significant difference was noticed between the groups regarding the final visual outcome [16]. Rajaraman et al. reported outcomes of combined C3F8 and compressive sutures in managing acute hydrops in 17 patients with keratoconus. They noticed that intracameral C3F8 alone was not enough to resolve corneal oedema in the pres- ence of stromal clefts. To prevent pupillary block and glaucoma, they applied immediate post-op pupillary dilation, and no complication was reported in their series of patients2. Cherif et al. introduced the pre-Descemet sutures to manage Descemet detachment due to corneal hydrops [8]. This technique was combined with intra-cameral air injection and proved safe and effective in shortening the hydrops and resolving oedema. The challenge with this technique is limited visualization of the cornea and pre-Descemet area in the presence of significant stromal oedema. To overcome this challenge, further modifications were suggested, such as full-thickness compressive suture instead of pre-Descemet suture or the use of intra-operative OCT (iOCT) to visualize the pre-Descemet cornea. Mohebbi et al. tried the combination of full-thickness compressive suture and injection of SF6. They found this method very effective and safe, with complete resolution of corneal oedema in 5–24 days without any complication [3]. OCT-guided drainage of intrastromal fluid, using venting incisions combined with intracameral air injection, has also been tried [9]. This is reported to achieve Descemet attachment on the first day post-op with complete resolution of corneal oedema within 2–3 weeks. Later, iOCT was utilized in another study to remove intra-stromal fluid pockets and place pre-Descemet sutures in two patients with acute hydrops. The authors injected intracameral SF6 to tamponade the detached Descemet and achieved good results [10]. Venting incisions to compress the stromal fluid combined with intracameral injection of 20% SF6 was successfully tried in another study [17]. Bachmann et al. reported a mini-DMEK technique to treat three patients with acute hydrops [12]. The procedure was successful in two patients and needed repeating in one patient. To apply this technique, iOCT is necessary due to the poor visibility, a donor graft is required on an urgent basis, and the AC needs to be filled with gas. While iOCT is not readily available in all centers, and donor tissue shortage is another concern in many countries, and the use of anterior chamber gas can be associated with potential complications such as pupillary block and cataract. Our study reports the full-thickness compressive corneal suturing as a solo treatment in managing acute hydrops. This is a relatively simple technique that surgeons can easily adopt. Furthermore, although iOCT is an excellent tool for these cases, this technique could be performed in absence of iOCT since the sutures are full-thickness. The key is to Vision 2023, 7, 10 7 of 9 visualize the Descemet breaks using a small anterior chamber air bubble before suturing. This enables the surgeon to place the sutures perpendicular to the break. Most of the previous reports combined this technique with the injection of air or gas into the anterior chamber. Although some studies with a small number of patients did not come across the pupillary block [3], others with larger sample size noticed up to 16% rate of pupillary block in their patients [16], as it is already understood very well [18]. To prevent this complication, either prophylactic peripheral iridotomy (PI) is to be performed, which is not only difficult to perform via oedematous cornea with poor visibility but also associated with potential complications such as hyphaema [19], IOP spike [19], dysphotopsia [20], and similar, or the need for topical cycloplegic drops post-operatively, as long as the gas is present in the AC. However, with our technique, there is no requirement for prophylactic PI or post-op cycloplegic agents. In all five reported cases, we achieved good anatomical outcomes without leaving any air bubbles in the anterior chamber. Moreover, although opacification of the crystalline lens was not reported in the previ- ous reports where they used air/gas bubbles to manage acute hydrops, we do not have detailed information on the duration of gas remaining in the anterior chamber and the duration of post-procedure follow-up in those reports. The posterior segment gas bubble induces cataracts [21], and several reports have been published on IOL opacification with anterior chamber air/gas bubble following endothelial keratoplasty [22]. Considering the young age of these patients and further challenges in intraocular lens power calculation and visual rehabilitation, it is best to avoid leaving any air in the anterior chamber of these patients, which can potentially induce cataract. Regarding the visual outcome, post-procedure BCVA, compared to the BCVA at presentation, improved in four patients and was unknown in one patient who did not attend the follow-up. Two of our patients achieved their pre-hydrops visual acuity; one patient experienced a one-line drop on the Snellen chart, and another patient, with extensive involvement of the central cornea, had a significant drop on the Snellen chart. Cherif et al. also compared the BCVA at presentation with post-suturing BCVA, which showed a significant improvement [8]. Furthermore, two of their patients required PK once the hydrops settled. Although the visual outcome is not provided in all patients, the graft patients achieved BCVA of 20/20 and 20/32 and another non-grafted eye achieved BCVA of 20/20 [8]. Therefore, the 6-month post-suturing BCVA in their study reflects not only the suturing outcome but also the added PK benefit [8]. Likewise, the visual outcome in the mini-DMEK technique is not comparable to our study as they provided unaided visual acuity in their three patients, which is understandably poor in eyes with keratoconus regardless of hydrops [12]. Nevertheless, a critical factor determining the final visual outcome is the extent and location of DM break and corneal oedema. The involvement of the central cornea will have a remarkable effect on the final visual outcome. In contrast, paracentral or peripheral DM break and oedema’s impact on final vision will be less significant. Some previous studies reported no significant difference in visual outcomes between their intervention and control groups [14,16]. Instead of comparing the final visual acuity between two groups, we suggest comparing each patient’s pre-hydrops and post-treatment/observation visual acuity to provide a more accurate result. Although we did not encounter any complications when placing full-thickness corneal suture, care should be taken not to damage the adjacent structures. Also, to avoid infection, the use of intracameral antibiotics is highly recommended. Placing corneal sutures can also be associated with suture-tract linear scarring. This was noticed with both pre-Descemet [8] and full-thickness suturing [3]. We also came across this complication, however, similar to the previous studies, it did not appear to cause a significant visual problem, and good vision was achieved with RGPCL. In summary, full-thickness corneal suturing as a solo technique in managing acute hydrops is effective and safe. This procedure can be performed without iOCT and does not necessitate techniques to tackle pupillary block or raised IOP. This technique pre- vents prolonged discomfort and corneal oedema, leading to vascularization and scar- Vision 2023, 7, 10 8 of 9 ring, and depending on the extent of the Descemet break, it can save patients from penetrating keratoplasty. 4.1. What Was Known before Spontaneous resolution of corneal hydrops is a lengthy process and associated with discomfort and corneal scar. Approximation pre-Descemet or full thickness suturing with intracameral air or gas is effective in the management of corneal hydrops. 4.2. What This Paper Adds Full-thickness corneal suturing without leaving any gas in the anterior chamber is a simple and safe method to treat acute hydrops, which can be performed in the absence of intraoperative OCT. Supplementary Materials: The following supporting information can be downloaded at: https:// www.mdpi.com/article/10.3390/vision7010010/s1, Video S1: Surgical Procedure. Author Contributions: Conceptualization, M.A.N. and Z.A.; methodology, M.A.N. and Z.A; soft- ware, Z.A., R.M.; validation, Z.A., R.M., M.A.N.; formal analysis, Z.A.; investigation, Z.A. and R.M.; resources, R.M., M.A.N.; data curation, Z.A. and R.M.; writing—original draft preparation, Z.A. and R.M.; writing—review and editing, M.A.N., Z.A. and R.M.; visualization, M.A.N.; supervision, M.A.N.; project administration, R.M.; funding acquisition: not applicable. All authors have read and agreed to the published version of the manuscript. Funding: No funding was available for this case repot. Institutional Review Board Statement: Ethical review and approval were waived for this study as this study was an audit of cases treated with the similar technique. Informed Consent Statement: Informed consent was obtained from all subjects involved in the study. Data Availability Statement: Not applicable. Conflicts of Interest: The authors declare no conflict of interest. References 1. 2. Fan Gaskin, J.C.; Patel, D.V.; McGhee, C.N. Acute corneal hydrops in keratoconus—New perspectives. Am. J. Ophthalmol. 2014, 157, 921–928. [CrossRef] Rajaraman, R.; Singh, S.; Raghavan, A.; Karkhanis, A. Efficacy and safety of intracameral perfluoropropane (C3F8) tamponade and compression sutures for the management of acute corneal hydrops. Cornea 2009, 28, 317–320. [CrossRef] 4. 7. 6. 5. 3. Mohebbi, M.; Pilafkan, H.; Nabavi, A.; Mirghorbani, M.; Naderan, M. Treatment of Acute Corneal Hydrops with Combined Intracameral Gas and Approximation Sutures in Patients with Corneal Ectasia. Cornea 2020, 39, 258–262. [CrossRef] [PubMed] Tuft, S.J.; Gregory, W.M.; Buckley, R.J. Acute corneal hydrops in keratoconus. Ophthalmology 1994, 101, 1738–1744. [CrossRef] [PubMed] Basu, S.; Vaddavalli, P.K.; Vemuganti, G.K.; Ali, M.H.; Murthy, S.I. Anterior segment optical coherence tomography features of acute corneal hydrops. Cornea 2012, 31, 479–485. [CrossRef] Sharma, N.; Maharana, P.K.; Jhanji, V.; Vajpayee, R.B. Management of acute corneal hydrops in ectatic corneal disorders. Curr. Opin. Ophthalmol. 2012, 23, 317–323. [CrossRef] [PubMed] Grewal, S.; Laibson, P.R.; Cohen, E.J.; Rapuano, C.J. Acute hydrops in the corneal ectasias: Associated factors and outcomes. Trans. Am. Ophthalmol. Soc. 1999, 97, 187–198; discussion 198–203. [CrossRef] Yahia Chérif, H.; Gueudry, J.; Afriat, M.; Delcampe, A.; Attal, P.; Gross, H.; Muraine, M. Efficacy and safety of pre-Descemet’s membrane sutures for the management of acute corneal hydrops in keratoconus. Br. J. Ophthalmol. 2015, 99, 773–777. [CrossRef] Vajpayee, R.B.; Maharana, P.K.; Kaweri, L.; Sharma, N.; Jhanji, V. Intrastromal fluid drainage with air tamponade: Anterior segment optical coherence tomography guided technique for the management of acute corneal hydrops. Br. J. Ophthalmol. 2013, 97, 834–836. [CrossRef] Siebelmann, S.; Händel, A.; Matthaei, M.; Bachmann, B.; Cursiefen, C. Microscope-Integrated Optical Coherence Tomography- Guided Drainage of Acute Corneal Hydrops in Keratoconus Combined with Suturing and Gas-Aided Reattachment of Descemet Membrane. Cornea 2019, 38, 1058–1061. [CrossRef] 10. 9. 8. 11. Wylegala, E.; Tarnawska, D. Amniotic membrane transplantation with cauterization for keratoconus complicated by persistent hydrops in mentally retarded patients. Ophthalmology 2006, 113, 561–564. [CrossRef] Vision 2023, 7, 10 9 of 9 12. Bachmann, B.; Handel, A.; Siebelmann, S.; Matthaei, M.; Cursiefen, C. Mini-Descemet Membrane Endothelial Keratoplasty for the Early Treatment of Acute Corneal Hydrops in Keratoconus. Cornea 2019, 38, 1043–1048. [CrossRef] [PubMed] 13. Rahman, I.; Carley, F.; Hillarby, C.; Brahma, A.; Tullo, A.B. Penetrating keratoplasty: Indications, outcomes, and complications. Eye 2009, 23, 1288–1294. [CrossRef] [PubMed] 14. Miyata, K.; Tsuji, H.; Tanabe, T.; Mimura, Y.; Amano, S.; Oshika, T. Intracameral air injection for acute hydrops in keratoconus. 15. Am. J. Ophthalmol. 2002, 133, 750–752. [CrossRef] [PubMed] Shah, S.G.; Sridhar, M.S.; Sangwan, V.S. Acute corneal hydrops treated by intracameral injection of perfluoropropane (C3F8) gas. Am. J. Ophthalmol. 2005, 139, 368–370. [CrossRef] 16. Basu, S.; Vaddavalli, P.K.; Ramappa, M.; Shah, S.; Murthy, S.I.; Sangwan, V.S. Intracameral perfluoropropane gas in the treatment 17. of acute corneal hydrops. Ophthalmology 2011, 118, 934–939. [CrossRef] [PubMed] Sayadi, J.J.; Lam, H.; Lin, C.C.; Myung, D. Management of acute corneal hydrops with intracameral gas injection. Am. J. Ophthalmol. Case Rep. 2020, 20, 100994. [CrossRef] 18. Röck, D.; Bartz-Schmidt, K.U.; Röck, T.; Yoeruek, E. Air Bubble-Induced High Intraocular Pressure After Descemet Membrane Endothelial Keratoplasty. Cornea 2016, 35, 1035–1039. [CrossRef] 19. Kam, J.P.; Zepeda, E.M.; Ding, L.; Wen, J.C. Resident-performed laser peripheral iridotomy in primary angle closure, primary angle closure suspects, and primary angle closure glaucoma. Clin. Ophthalmol. 2017, 11, 1871–1876. [CrossRef] 20. Vera, V.; Naqi, A.; Belovay, G.W.; Varma, D.K.; Ahmed, I.I. Dysphotopsia after temporal versus superior laser peripheral iridotomy: A prospective randomized paired eye trial. Am. J. Ophthalmol. 2014, 157, 929–935. [CrossRef] 21. Kanclerz, P.; Grzybowski, A. Complications Associated with the Use of Expandable Gases in Vitrectomy. J. Ophthalmol. 2018, 2018, 8606494. [CrossRef] [PubMed] 22. Morgan-Warren, P.J.; Andreatta, W.; Patel, A.K. Opacification of hydrophilic intraocular lenses after Descemet stripping automated endothelial keratoplasty. Clin. Ophthalmol. 2015, 9, 277–283. [CrossRef] [PubMed] Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
10.1371_journal.pone.0262792
RESEARCH ARTICLE Age-related changes in tau and autophagy in human brain in the absence of neurodegeneration Shreyasi Chatterjee1,2☯, Megan Sealey1☯, Eva Ruiz1, Chrysia M. Pegasiou1,3, Keeley Brookes2, Sam Green1, Anna Crisford1, Michael Duque-Vasquez1, Emma LuckettID Paul Grundy5, Diederik Bulters5,6, Christopher Proud1,7, Mariana Vargas-CaballeroID Amritpal MudherID 1,4, Rebecca Robertson1, Philippa Richardson1, Girish Vajramani5, 1* 1*, a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Chatterjee S, Sealey M, Ruiz E, Pegasiou CM, Brookes K, Green S, et al. (2023) Age-related changes in tau and autophagy in human brain in the absence of neurodegeneration. PLoS ONE 18(1): e0262792. https://doi.org/10.1371/journal. pone.0262792 Editor: Hemant K. Paudel, McGill University, CANADA Received: September 19, 2021 Accepted: August 19, 2022 Published: January 26, 2023 Peer Review History: PLOS recognizes the benefits of transparency in the peer review process; therefore, we enable the publication of all of the content of peer review and author responses alongside final, published articles. The editorial history of this article is available here: https://doi.org/10.1371/journal.pone.0262792 Copyright: © 2023 Chatterjee et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: All relevant data are within the manuscript and its Supporting Information files. 1 School of Biological Sciences, University of Southampton, Southampton, United Kingdom, 2 Department of Biochemistry, School of Science and Technology, Nottingham Trent University, Nottingham, United Kingdom, 3 School of Life Sciences, University of Sussex, Brighton, United Kingdom, 4 Department of Neuroscience, KU Leuven, Leuven, Belgium, 5 Department of Neurosurgery, Wessex Neurological Centre, University Hospital Southampton, Southampton, United Kingdom, 6 Faculty of Medicine, Clinical Neurosciences, Clinical and Experimental Sciences, University of Southampton, Southampton, United Kingdom, 7 Lifelong Health, South Australian Health and Medical Research Institute, SAHMRI, and School of Biological Sciences, University of Adelaide, Adelaide, Australia ☯ These authors contributed equally to this work. * A.Mudher@soton.ac.uk (AM); M.Vargas-Caballero@soton.ac.uk (MV-C) Abstract Tau becomes abnormally hyper-phosphorylated and aggregated in tauopathies like Alzhei- mers disease (AD). As age is the greatest risk factor for developing AD, it is important to understand how tau protein itself, and the pathways implicated in its turnover, change during aging. We investigated age-related changes in total and phosphorylated tau in brain sam- ples from two cohorts of cognitively normal individuals spanning 19–74 years, without overt neurodegeneration. One cohort utilised resected tissue and the other used post-mortem tis- sue. Total soluble tau levels declined with age in both cohorts. Phosphorylated tau was undetectable in the post-mortem tissue but was clearly evident in the resected tissue and did not undergo significant age-related change. To ascertain if the decline in soluble tau was correlated with age-related changes in autophagy, three markers of autophagy were tested but only two appeared to increase with age and the third was unchanged. This implies that in individuals who do not develop neurodegeneration, there is an age-related reduction in soluble tau which could potentially be due to age-related changes in autophagy. Thus, to explore how an age-related increase in autophagy might influence tau-mediated dysfunc- tions in vivo, autophagy was enhanced in a Drosophila model and all age-related tau pheno- types were significantly ameliorated. These data shed light on age-related physiological changes in proteins implicated in AD and highlights the need to study pathways that may be responsible for these changes. It also demonstrates the therapeutic potential of interven- tions that upregulate turnover of aggregate-prone proteins during aging. PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 1 / 24 PLOS ONE Funding: AM Gerald Kerkut Trust https://www. kerkut-trust.org.uk/ Alzheimer’s Research UK https://www.alzheimersresearchuk.org/ No. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: The authors have declared that no competing interests exist. Age-related changes in tau and autophagy in human brain Introduction Alzheimer’s disease (AD) is the most common cause of dementia in the elderly, and with a growing population of older people, it is becoming increasingly more prevalent. Age is one of the greatest risk factors for developing AD, with incidence increasing from 1 in 14 above that age of 65 to 1 in 6 at age >80 (Alzheimer’s Association report 2019). This clearly demonstrates that age is a significant risk factor for developing this condition. Since many of the neuropathological hallmarks of AD and the clinical features that charac- terise it, such as declining working memory, short-term recall and processing speed, are also observed in normal aging [1–3], it has been proposed that these age-related changes are a con- tinuum of AD [4]. However, the relationship between aging and AD is complex and interwo- ven, and though there are many areas of overlap [5], it is clear that AD is not an inevitable part of the physiological CNS aging process [6]. It therefore becomes imperative to understand what about the aging brain makes it vulnerable to AD compared to the younger brain; to understand the physiological changes that occur in aging brains and the possible mechanisms by which deviation from the normal aging process occurs and leads to pathology and neurode- generative disease. This topic has been the subject of numerous reviews, each discussing the many theories that have been proposed to explain how aging confers risk of AD and other age- related proteinopathies [5,7]. One area of interest in this regard is the role of proteostatic path- ways such as autophagy. Both macro-autophagy and chaperone-mediated autophagy are implicated in the turnover of tau, the microtubule associated protein that constitutes the neu- rofibrillary tangles, one of the two defining pathological hallmarks of AD [8–13]. One view is that there is a functional decline in autophagic clearance capabilities in the brains of individu- als who are affected by neurodegenerative proteinopathies, and this manifests in the accumula- tion of misfolded protein aggregates [14,15]. Many mechanisms have been proposed to account for this functional decline. A large body of evidence suggests that autophagy is acti- vated as a reactive response to misfolded protein formation at early stages of disease, but that the pathogenic proteins have a toxic effect on autophagy, compromising it such that it eventu- ally becomes overwhelmed in proteinopathies [16]. Another view is that with increasing age, there is reduced autophagy and this is conducive to promoting aggregation and accumulation of misfolded proteins in the aged brain [17]. While many studies have demonstrated impaired autophagy in the brains of neurodegenerative disease patients, there are relatively few studies investigating normal physiological changes in basal autophagy with increasing age in the absence of neurodegeneration [17]. This is clearly an area with controversial findings and mer- its further investigation like the study we conducted. In a similar vein, very few studies have shown how proteins like tau, which accumulate and misfold in Tauopathies and are believed to be substrates for autophagy, actually change physio- logically in the human brain during healthy aging without the confounding influences of underlying neurodegeneration. The only study that assessed this in human brain was pub- lished over 20 years ago and reported a decrease in total soluble tau levels [18]. Though a mechanism was not demonstrated, that study nonetheless showed that the decrease in soluble tau was not due to a reciprocal age-related increase in insoluble tau. The current study addresses these questions by investigating physiological age-related changes in both tau and autophagy in human brain samples taken from young and old subjects who did not have any neurodegenerative disease. In agreement with the study by Mukaetova- Ladinska et al [18], our findings also suggest that with increasing age, total soluble tau levels decline. Strengthening that previous finding, we demonstrate that this reduction in tau is seen in resected human brain tissue as well as post-mortem brain, confirming that it is not due to a post-mortem artefact. We also investigated whether hyperphosphorylated tau levels indicative PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 2 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain of AD pathology were altered in these non-AD cohorts and found no immunoreactivity in the post-mortem brain tissues. We then go on to investigate if this change in total soluble tau could be explained by an age-related change in autophagy. Though we do demonstrate an increase in expression of some markers of autophagy in older brains, others are unchanged. This data implies that an age-related increase in autophagy could influence total soluble tau levels and protect against tau-mediated neurodegeneration. To test this, a Drosophila model was employed in which co-expression of a pro-autophagic gene was found to significantly improve age-related tau phenotypes. Materials and methods Human tissue samples Samples of frozen cortical sections from post-mortem human brains (six 20–33 year old brains and seven 70–74 year old brains) were obtained from Sudden Death Brain Bank (Edinburgh, United Kingdom, cortical area: superior temporal gyrus, case numbers included: SD023/08, SD041/05, SD027/06, SD034/08, SD030/09, SD001/11, SD008/12, SD020/06, SD053/14, SD025/17, SD024/17, SD012/17, SD008/17). No pathological diagnosis of Alzheimer’s Disease or any other pathology was reported in these tissues. Resected human Brain tissue was obtained during neurosurgical procedures at the Wessex Neurological Centre at University Hospital Southampton and the samples were processed according to the Human Tissue Act 2004 with approval from the Faculty of Medicine Ethics Committee and the Southampton Research Biorepository following written informed consent (Study reference number SRB002/14). The patients were being treated for medial temporal epilepsy, cavernous or arteriovenous malformations or glioma (Table 1). Any surplus tissue that was removed for access to these lesions that would otherwise be discarded was retained and subsequently processed as described by Pegasiou et al. [19]. Following resection cortical tissue was immediately put into artificial cerebrospinal fluid (ACSF) (110 mM choline chlo- ride, 26 mM NaHCO3, 10 mM D-glucose, 116 mM C6H7NaO6, 7 mM MgCl2, 3.1 mM C3H3NaO3, 2.5 mM KCl, 1.25 mM NaH2PO4 and 0.5 mM CaCl2.) The tissues were collected from patients in the age group 19 to 70 years and age-matched with the post-mortem brain samples. Thirteen samples were used for our experiments. These tissues were snap frozen on dry ice within 10 min of removal from the patient. Both post-mortem and resected tissue samples were then lysed in homogenization buffer (10 mM Tris Base, 150 mM NaCl, 0.05% Tween-20, 1mM Na orthovanadate, 10 mM NaF, 10 μM Staurosporine, 1X Protease Inhibitor cocktail and 1 mM Tyrophostin) on ice. The sam- ples were spun briefly at 8000 x g for 5 min and the pellets were discarded. A protein assay was conducted by the Bradford method to normalize the protein concentration. The samples were then boiled for in 2x SDS sample buffer at 95˚C for 5 min for subsequent Western Blotting analysis. Electrophysiology Following previously published methods [20], slices were transferred to a submerged-style recording chamber and superfused with recording ACSF at a rate of *1.5 mL/min. Whole- cell voltage clamp recordings were performed using glass pipettes (~5 MO) pulled from boro- silicate glass, yielding a series resistance of ~15 MO. Recordings were made at room tempera- ture (21-25˚C) using K-gluconate-based intracellular solution, containing (in mM): 120 Potassium gluconate; 10 KCl; 10 Hepes; 0.3 GTP; 4 Mg-ATP; and pH was titrated to 7.25 with KOH. The final osmolarity was 285 mOsmol−1. Neuronal excitability was tested by delivering square pulses of current in increments of 50 pA. PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 3 / 24 PLOS ONE Table 1. Case study details of the human post-mortem brain and surgically resected tissues. Post-mortem brain tissues were obtained from the Sudden Death Brain Bank and divided into young (20 to 33 years) and old (70 to 74 years) cohorts. There was no significant difference in the post-mortem intervals between the young and old cohorts (p = 0.48). Age-matched resected tissues (19 to 70 years) were collected from the Southampton General Hospital. Neuropathological abnormalities were found in cases 0011 and 0020 but no neurodegenerative changes were detected in any of the human tissues. Age-related changes in tau and autophagy in human brain Case Number Sex Age Diagnosis SD023/08 F 24 Small vessel disease SD041/05 M 24 No significant abnormality SD027/06 M 25 Cerebral oedema. Recent hypoxia. History of butane gas inhalation. Other Medical History Not known Not known Not known SD034/08 M 70 No significant abnormality SD030/09 F 71 No significant abnormality SD001/11 M 74 No significant abnormality SD008/12 M 25 No neuropathological findings SD020/06 F 20 No significant abnormality SD053/14 M 33 No significant abnormalities Not known Not known Not known Not known Not known Not known SD025/17 M 73 No significant abnormalities Not known SD024/17 M 72 No significant abnormalities Not known SD012/17 SD008/17 F F 71 No significant abnormalities Not known 71 No significant abnormalities Not known Resected Tissue Case Data 0019 0026 0014 0023 0016 0028 0011 0024 0012 0017 0025 0013 0020 M 19 Arteriovenous malformation Not known M 27 Mesial temporal DNET with signal Not known changes in the hippocampus M 32 Hippocampal Sclerosis M 35 Cavernoma F F F F 36 Hippocampal sclerosis 38 42 50 Epilepsy Glioma Cavernous malformation M 55 Glioblastoma Not known Not known Not known Asthma Not known Not known Not known F F F F 62 Hippocampal Sclerosis Depression, hypertension, gastroreflux disease 62 Metastasis—Adenocarcinoma Not known of the lung 69 High Grade Glioma Not known 70 Arteriovenous malformation Hysterectomy, urostomy, acute myocardial infarction, hypertension, polymyalgia rheumatic, hypercholesteromia, recurrent urinary tract infections Brain area Temporal lobe Temporal lobe Temporal lobe Temporal lobe Temporal lobe Temporal lobe Temporal Superior Temporal Superior Superior temporal gyrus BA41/42 Superior temporal gyrus BA41/42 Superior temporal gyrus BA41/42 Superior temporal gyrus BA41/42 Superior temporal gyrus BA41/42 Left frontal lobe Right anterior temporal lobe Left anterior temporal lobe Left frontal lobe Right anterior temporal lobe Right anterior temporal lobe Left anterior temporal lobe Right lateral temporal lobe Left parietal lobe Right anterior temporal lobe Right frontal lobe Right frontal lobe Right lateral temporal lobe (superior temporal gyrus) Note: F, female; M, male; DNET, dysembryoplastic neuroepithelial tumor. https://doi.org/10.1371/journal.pone.0262792.t001 Electron microscopy Sections of freshly resected cortex were placed into a fixative solution consisting of 3% glutaral- dehyde and 4% formaldehyde in 0.1 M PIPES and stored at 4˚C for a minimum of 1 h. Brain PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 4 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain samples, approx. 1 mm by 1 mm in size, were washed with 0.1 M PIPES and then incubated in 1% osmium tetroxide for 1 h. After further washes in 0.1 M PIPES, samples were incubated in 2% uranyl acetate for 20 min. Dehydration followed using increasing concentrations of etha- nol. After a final incubation in acetonitrile for 10 min, samples were incubated overnight in 50:50 acetonitrile:TAAB resin. Samples were then incubated for 6 h in TAAB resin and were then left to polymerise for 24 h at 60˚C. 90–100 nm sections were produced from the samples using the Reichert OM-U3 ultra-microtome and placed onto copper palladium grids. Samples were then placed onto droplets of lead citrate for 3–5 min for further contrast before being imaged at x87,000 magnification using an FEI Tecnai T12 transmission electron microscope. RT PCR RNA from the tissue samples was isolated using an in-house developed methodology [21]. Briefly, the Covaris Cryo Prep system was employed to crush the tissue to increase the surface area for allowing efficient cell lysis in 1 ml Trizol (Invitrogen). This was followed by RNA extraction with the RNAeasy Minikit from QIAGEN using standard protocols. The RNA con- centration was measured in the nanodrop at an absorbance of 260 nm and the samples were processed for RT PCR. Briefly, 1 ug of RNA was treated with RNAse-free DNase I at 37˚C for 30 mins to remove genomic DNA contamination. Following this treatment the DNAse was inactivated at 65˚C and RT-PCR was performed by using the Superscript IV One-step RT-PCR kit (Invitrogen) with Platinum Taq Polymerase. Single strand cDNA synthesis was followed by PCR using 1 ug of RNA. For amplification of the 5’ half of the tau mRNAs tau1 fwd primer (5’-ATGGCTGAGCCCCGCCAGGAG-3’) and tau 4 reverse primer (5’-CCCAGCTCT GGTGAACCTCCA-3’) were used [22]. These primers amplify tau sequences including exon 2 and 3, exon 2 alone or without these exons. The PCR was performed using the following con- ditions, pre-denaturation (94˚C for 2mins) followed by 40 cycles of amplification (denatur- ation, 94˚C for 15s, annealing 60˚C for 30s; extension 720C for 60s) and a final extension of 720C for 5 min. Another RT PCR was performed with human-β actin forward (5’-CCTCGCC TTTGCCGATCC-3’) and reverse (5’GGATCTTCATGAGGTAGTGAGTC-3’) primers for nor- malization. Products of RT-PCR were analyzed in 2% Agarose gels in TAE. Western blotting Human brain samples: 20 to 30 μg of the homogenates were separated on 10% SDS-PAGE gels and transferred to nitrocellulose membranes/PVDF membranes. After blocking the mem- branes in 5% BSA in 1X TBS, these were incubated overnight at 4˚C with the following anti- bodies. Anti-human tau (Dako, 1:10,000), the phosphorylation specific anti-tau antibody Ser396/Ser404 (PHF1, 1:500) (a gift from Dr. Peter Davies, USA), AT8 (Invitrogen, 1:1000), phospho-Ser-262 (Thermofisher, 1:1000) Acetylated α−tubulin (Sigma Aldrich, 1:1000), Tyro- sinated α−tubulin (Sigma Aldrich, 1:2000), Total α-tubulin (Sigma Aldrich, 1:1000), p62 (Millipore, 1:1000), LC3 (Abcam, 1:500), GAPDH (Abcam, 1:2500), Beclin1 (Abcam, 1:500), Cathepsin D (Abcam, 1:1000), Rabbit β-actin (Abcam, 1:3000) and Mouse β-actin (Abcam, 1:1000) were used as loading controls. Membranes were then incubated with secondary anti- bodies (Alexa Fluor 680 goat anti-mouse or Alexa Fluor 800 goat anti-rabbit) at 1:20,000 for one hour at room temperature. Antigens were visualised using an Odyssey scanner (Li-Cor Biosciences). Fly brain samples: Adult flies were snap frozen in liquid nitrogen and fly heads were dis- sected. Ten fly heads were then homogenized in homogenization buffer in a ratio of 1 head: 10 μl buffer (150 mM NaCl, 50 mM MES, 1% Triton-X, protease inhibitor cocktail, 30 mM NaF, 40 mM 2-glycerophosphate, 20 mM Na-pyrophosphate, 3.5 mM Na orthovanadate, 1% PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 5 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain SDS, 10 μM Staurosporine). Samples were centrifuged at 3000 x g and pellets were discarded. 2x SDS sample buffer was added to each tube and boiled at 95˚C for 5 min. Western Blotting was then done as described above. Fly stocks Elav-GeneSwitch, UAS-hTau0N3R, UAS-Atg1 were obtained from Bloomington Drosophila Stock Center, Indiana, US. Bigenic flies (UAS-hTau0N3R/UAS-Atg1) were created using stan- dard genetic recombination methods. The flies were raised and maintained at 25˚C on SYA diet using standard protocols. RU486 treatment A 100 mM stock solution of RU486 was prepared in 100% ethanol. Post eclosion, 0–3 day-old flies were moved to either food containing 200 μM RU486 or to control diet for biochemistry and behavioural studies; this was to ensure gene expression was switched on only in adult flies. Only male flies were used for our studies to remove confounding impact of egg laying in the climbing assay that was done in these animals. Climbing assay The climbing assay was performed according to standard procedures developed in our labora- tory [23]. Briefly, five cohorts of 50 flies were transferred to 50 ml measuring cylinders without anasthetisation and the distance climbed was recorded at 10 seconds after tapping down. The assays was repeated three times with 1 min rest between each trial and the mean was calculated. Statistics GraphPad Prism 5.0 software was used for statistical analysis. Two-tailed t tests were used to analyse the difference between two groups. For multiple comparisons one-way or two-way ANOVA with Bonferroni’s post-hoc correction for pairwise comparisons was used. Error bars depict standard error of the mean (SEM) as indicated in the figure legend. Statistical signifi- cance is depicted by the following n.s = not significant (p>0.05), �p<0.05, ��p<0.01, ���p<0.001, ����p<0.0001. Results Resected human brain tissue is physiologically normal To study impact of aging on tau in the absence of neurodegeneration in human brain, total and phosphorylated levels were assessed in non-pathological cortical human brain tissues taken from young and old individuals assessed by routine neuropathological analysis and by the lack of inflammatory markers [19]. Importantly, as previous studies [24–26], including ones from our own lab [27], have demonstrated that the post-mortem (PM) interval influences the phosphorylation state of tau; our initial studies were conducted both in resected human tis- sue following collection in the operating theatre and rapid processing (no PM interval) as well as in human tissue from national brain banks (standard PM interval). Resected tissue was col- lected at the neurosurgery unit of University Hospital Southampton from a spectrum of surgi- cal cases ranging from 19 to 70 years of age. Tissue was either quickly frozen and kept at -80˚C or processed for electrophysiology or electron microscopy (see methods). As described recently [19], the tissue utilised for these studies was physiologically and morphologically nor- mal unless otherwise indicated. The rapid collection permits the preservation of function and PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 6 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain ultrastructure. We are confident that the resected tissue that we have utilised in this study rep- resents normal physiological tissue without confounding disease. This is illustrated in S1 Fig where an electrophysiological recording taken from one such case (S1A Fig) showing robust action potential firing as a response to depolarising current injection, together with an ultra- structural image showing pre and postsynaptic structures (S1B Fig). Further electrophysiologi- cal recordings on these resected human tissue are shown in Table 1 [19]. Since resected tissue is a precious and scarce resource, only the analyses of total and phosphorylated tau, which we know from previous findings are sensitive to PM delay [28], were conducted using this tissue. Age dependent decrease in total tau in non AD brain Post-mortem cortical human brain tissues from individuals without AD or other neurodegener- ative disorders were obtained from the Sudden Death Brain Bank (Centre for Clinical Brain Sci- ences, Edinburgh, UK). These tissues were divided into young (aged 20 to 33 years) and elderly cohorts (aged 70 to 74 years) and age-related alterations in the levels of total and phospho-tau were assessed by western blotting. The post mortem intervals in the young and old cohorts were not significantly different (p = 0.48). Similarly, age-matched cortical resected tissues were freshly collected from individuals ranging in age from 19 to 70 years who were undergoing sur- gery (Wessex Neurological Centre, University Hospital Southampton) but did not have neuro- degenerative disease. As with the post-mortem tissue, the resected tissues were divided into young (aged 19 to 42 years) and old cohorts (aged older brains 50–70 years). Due to limited availability, the resected human tissue could not be segregated into very young and very old, in contrast to the PM tissue, where samples of desired ages could be requested from brain banks. Nonetheless, the segregation of the resected tissue into young and old brain groups was per- formed based on previous findings that cognitive decline begins from 45 years onwards [29]. A significant decrease (p<0.0001) in the levels of total tau was observed in the post-mortem brain samples from the older cohorts compared to the younger cohorts (Fig 1A i and iii). This significant reduction in total soluble tau levels with increasing age was also observed in age- matched resected tissues (p<0.05) (Fig 1B i and ii). Interestingly, a robust phospho-tau signal (immunoreactive to the PHF1 antibody) was evident in resected tissues (Fig 1B i), but no such signal was detected in the post-mortem tissues (Fig 1A i) possibly due to the post-mortem delay. We did not observe a significant age-dependent increase of PHF-tau signal in the resected tissues (Fig 1B iii). The observation of robust PHF-1 positive tau in non-AD brain is unusual as tau is not believed to be significantly phosphorylated in normal adult brain (i.e., in the absence of neurodegeneration). This demonstrates the advantage of comparing signals in resected and PM tissues alongside each other to identify changes that may otherwise of be lost due to tissue processing and PM delay. We also tested two other AD associated phospho-epi- topes AT8 and p-262 in the post-mortem tissues. While no signal was detected with the AT8 antibody, p-262 elicited weak antibody response in the samples (S2A and S2B Fig). Next, we analyzed if this decrease of tau was due to transcriptional repression. We isolated total RNA from the post-mortem brain tissues and performed RT-PCR with tau-specific prim- ers. We detected no significant difference in the 2N (567 bp), 1N (480 bp) or 0N (393 bp) tau transcripts between the youngers vs older cohorts (S3A and S3B Fig) suggesting that this varia- tion of tau was at the translational level. In order to confirm that the decrease in total tau levels in the older cohorts is not due to an age-related loss of neurons, the level of the neuronal marker NeuN was assessed by western blotting. No significant difference in NeuN levels was observed between the younger and older cohorts (Fig 1A ii and iii). Thus, we can conclude that the age-dependent decrease in the total soluble tau levels is not due to neuronal loss. PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 7 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain Fig 1. Total tau levels are significantly decreased in the healthy older brains compared to the younger brains. Representative western blot images for total tau, PHF1 and Actin in post-mortem brains (A i) and quantification of total tau normalized to Actin (A iii). Representative western blot image for NeuN and Actin in post-mortem brains (A ii) and quantification of NeuN normalized to Actin (A iii). Representative western blot images for total tau, PHF1 and Actin in resected tissues (B i) and quantification of total tau normalized to Actin (B ii). Regression plot showing an increased trend of PHF1/Tau signal in the older cohorts compared to the younger group in resected tissue samples (B iii) In post-mortem samples n = 6 (older brains 70–74 years), n = 6 (younger brains 20–33 years), in resected tissue samples n = 13 (older brains 50–70 years and younger brains 19–42 years). ���p<0.0001, �p<0.05 by two-tailed unpaired t test. Data represent mean ± SEM (A iii and B ii). An increased level of PHF1 immunoreactivity was observed in older brains of resected tissues although it was not significant (B iii). https://doi.org/10.1371/journal.pone.0262792.g001 We next wanted to test whether the change in tau is correlated with an age-related alter- ation in tau-mediated function. Microtubule stability measures are similar in young and old brains One of the primary functions of tau is the binding and polymerization of microtubules [30,31]. It has been shown that the hyperphosphorylated and PHF tau in AD brains can reduce the number and length of microtubules [32]. Further, it has also been shown that microtubule-stabilizing drugs are able to reverse axonal defects in animal models of tauopa- thy suggesting a direct connection between pathological tau and microtubule dysfunction [33]. In order to assess whether the loss of tau in the elderly cohorts compromises its func- tion of polymerization and stabilization of neuronal microtubules, the levels of acetylated and tyrosinated alpha tubulin were assessed by western blotting. These post-translational modifications of tubulin are common indicators of microtubular integrity [34]. While acety- lation is a marker for the stability of microtubules, tyrosination of tubulin has been shown to influence neuronal organization [34,35]. No difference in the levels of acetylated tubulin (Fig 2A i and ii) or tyrosinated tubulin (Fig 2B i and ii) was detected between the younger and elderly cohorts in the post-mortem brain samples. This implies that despite the decrease in cortical tau levels in healthy elderly cohorts, the stability of the microtubules in the same brain region is not affected. PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 8 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain Fig 2. The levels of total α-tubulin and its post-translational modifications are not altered with age in post-mortem brains. Representative western blot images for acetylated α-tubulin, total α-tubulin and GAPDH in post-mortem brains (A i) and quantification of actetylated α-tubulin relative to total α-tubulin normalized to GAPDH (A ii). Representative western blot images for tyrosinated α-tubulin, total α-tubulin and GAPDH in post-mortem brains (B i) and quantification of tyrosinated α-tubulin relative to total α-Tubulin normalized to GAPDH (B ii). In post-mortem samples n = 6 (older brains) and n = 6 (younger brains). p values are not significant by two-tailed unpaired t test. Data represent mean ± SEM (A ii and B ii). https://doi.org/10.1371/journal.pone.0262792.g002 Late autophagy markers are significantly increased in the post-mortem brains of the elderly cohorts compared to the younger cohorts To investigate the potential cause for the age-related decline in total tau levels, markers of autophagy were studied in the young and old post-mortem brain samples. p62 is a classical autophagy marker, its function being the ubiquitination of a host of cellular proteins that are then targeted for degradation [36]. As it is destroyed during autophagy, accumulation of p62 implies an impairment of autophagy while a reduction of p62 is indicative of activated autop- hagic clearance. A significant decrease of p62 (p<0.05) was evident in the post-mortem brains of the older cohorts compared to the younger cohorts (Fig 3A i and ii). As there is no signifi- cant difference in the PM delay in these samples which may be influence this readout, these results potentially point to increased markers of autophagy in the brains of the elderly cohorts. To investigate this further, additional markers of autophagy were investigated. In mamma- lian systems, another classical marker for late-stage autophagosome formation is the lysosomal marker LC3-II, and an increase in this signifies an enhancement of autophagy [37]. In addition to decreased p62, a significant increase in the LC3-II/LC3-I ratio (p<0.05) as well as an increase in total LC3-II levels was evident in the older post-mortem brains compared to the younger brain tissues (Fig 3B i-iii). Since both p62 and LC3II/I ratios are markers of late-stage autophagy, we next investigated whether early stage markers were also altered with age. For this we assessed levels of Beclin 1, a key autophagy protein known to be decreased in AD brains [38]. Interestingly, we did not see a significant difference in the levels of this early autophagy marker–neither did this marker go up or down with age in our young and old non-AD cohorts (Fig 4 i and ii). Another important pathway for tau degradation; particularly in the AD brains is the lysosomal clearance pathway. Therefore, we assessed the levels of the lysosomal protease Cathepsin D in the younger and older non-AD cohorts. However, no significant difference was observed (S4 Fig). PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 9 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain Fig 3. Autophagy is significantly increased in the healthy older brains compared to the younger brains in post-mortem tissues. Representative western blot images for p62 and Actin in post-mortem brains (A i) and quantification of p62 normalized to Actin (A ii). Representative western blot images for LC3 and Actin in post-mortem brains (B i) and quantification of LC3-II/LC3-I normalized to Actin (B ii) and LC3-II normalised to Actin (B iii). In post-mortem samples n = 7 (older brains), n = 6 (younger brains). �p<0.05, by two-tailed unpaired t test. Data represent mean ± SEM (A ii and B ii). https://doi.org/10.1371/journal.pone.0262792.g003 Taken together our results indicate that expression of some late-stage autophagy markers is increased in the older compared to the younger cohorts. However as the one early stage marker we examined appeared unaltered, it is not clear how overall autophagy changes with age, in the absence of neurodegeneration. Nonetheless it is clear that there is no age-related decline in autophagy as has been the consensus view. Age-related tau-induced behavioural changes are suppressed by activation of autophagy in a Drosophila model of Tauopathy The data presented thus far show that in non-AD brains, soluble tau levels are decreased with age, whilst expression of some specific markers of autophagy is increased. One interpretation of this is that an age-related increase in autophagy is a normal physiological response to aging which likely prevents the accumulation of total tau and protects against formation of tau aggre- gates in non-demented elderly subjects. To test this hypothesis, we assessed the impact of inducing autophagy on some age-related tau phenotypes in a Drosophila model of Tauopathy PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 10 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain Fig 4. Early autophagy gene Beclin 1 is not significantly increased in the older brains compared to the younger brains in post-mortem tissues. Representative western blot images for Beclin 1 and Actin in post-mortem brains (i) and quantification of Beclin 1 normalized to Actin (ii). In post-mortem samples n = 7 (older brains), n = 6 (younger brains). p values are not significant by two-tailed unpaired t test. Data represent mean ± SEM. https://doi.org/10.1371/journal.pone.0262792.g004 in which hTau0N3R is expressed. 0N3R is one of the human Tau isoforms that has three micro- tubule binding repeats in the C-terminal region and no N-terminal repeats. In human brains the accumulation of the 0N3R isoform of tau is mainly causative of Pick’s Disease but 3R tau is also important for AD pathology [39]. In this model, autophagy was induced by upregulation of the autophagy-specific kinase gene Atg1 [40]. To avoid developmental impact of inducing autophagy, pan-neuronal expression of both hTau0N3R and the autophagy gene manipulation was initiated in adults using the inducible GeneSwitch system in which the drug RU486 was fed to newly eclosed (emerged) adults to induce expression in adult flies which were then aged for a 4 week period. In parallel, the tau/Atg1 bigenics were similarly fed RU486 upon eclosion PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 11 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain Fig 5. Upregulation of autophagy decreases the age-related accumulation of hyperphosphorylated tau and ameliorates tau-induced locomotor defects in human Tau (0N3R) expressing transgenic flies. Pan-neuronal expression of human 0N3R tau significantly increases the age-dependent accumulation of hyperphosphorylated tau that is restricted by coexpression of autophagy marker Atg1. Representative western blot images for PHF1, total tau and Actin (A i) and quantification of PHF1 relative to total tau normalized to Actin (A ii) in htau0N3R and htau0N3R/Atg1 overexpressing Drosophila models respectively. �p<0.01 by one-way ANOVA with Bonferroni’s correction. Data represent mean ± S.E.M. Expression of 03NR human tau pan-neuronally in adult flies induces locomotor deficits that are rescued by coexpression of Atg1 (B) Comparison of the climbing ability with age over a period of 4 weeks for hTau0N3R, Atg1 and hTau0N3R/Atg1 transgenics (n = 50). (2 way ANOVA; �p<0.05, ��p<0.001, ����p<0.0001). Error bars are plotted as ± S.E.M. Genotypes: hTau0N3R = {w; Elav- Geneswitch/UAS-htau0N3R}, Atg1 = {w; Elav-Geneswitch/+; UAS-Atg1/+}, hTau0N3R/Atg1 = {w; Elav-Geneswitch /UAS-htau0N3R; UAS-Atg1/+}, on an OreR background. https://doi.org/10.1371/journal.pone.0262792.g005 and aged for the same period of time. This treatment led to induction of expression of both hTau0N3R and autophagy (S5 Fig). With this mode of adult-onset hTau0N3R expression, a sig- nificant age-related increase (p<0.01) in phosphorylated tau was evident which was abrogated by co-expression of Atg1 (Fig 5A i/ii). We have previously demonstrated that overexpression of hTau0N3R using the non-tempo- rally sensitive Elav-GAL4 pan-neural driver causes an age-dependent defect in the climbing ability of the flies [13]. Such a reduction in climbing ability was beginning to emerge in the flies expressing hTau0N3R following RU486 consumption, though it was less pronounced over the 4-week period tested, possibly because, with the GeneSwitch method employed here, tau expression only begins upon eclosion, but is likely to begin several days before eclosion with the Elav-Gal4 expression system previously used (Fig 5B). Induction of autophagy by co- expression of Atg1 significantly (p<0.0001) improved the climbing behaviour of the hTau0N3R flies. In young flies (up to 1wk old) there was no difference in the climbing ability of flies expressing hTau0N3R alone when compared to hTau0N3R flies that co-expressed Atg1. However, as the hTau0N3R/Atg1 bigenics started aging, their climbing ability was significantly rescued by coexpression of Atg1 compared to the age-matched hTau0N3R flies alone. After 2 weeks, the PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 12 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain climbing ability of the hTau0N3R/Atg1 bigenic animals became significantly better (p<0.001) than that of hTau0N3R flies so that by 4 weeks their climbing was nearly 100% better (p<0.0001). This effect was tau-specific and not a non-specific effect of over-expressing Atg1 since there was no similar age-dependent improvement in the climbing ability of flies express- ing Atg1 alone. These results show that induction of autophagy prevents the age-related accu- mulation of hyperphosphorylated tau and improves the climbing behaviour of tau-expressing flies in an age-related manner. Curiously, the climbing behaviour of the hTau0N3R;Atg1 bigenics was not just improved compared to those of either hTau0N3R or Atg1 alone flies, it was also better than the unex- pressed bigenic controls. One potential explanation for this is that, as well as mediating a toxic effect, expression of hTau may have some other beneficial effect not evident in the unexpressed controls. One such effect may relate to endogenous Drosophila (dTau) as we have previously demonstrated that the two interact [41]. To explore this, the impact of hTau0N3R expression on dTau was examined. dTau levels were found to accumulate with age in wild-type flies but not in flies expressing hTau0N3R (S6 Fig). It is likely that, in the presence of hTau, endogenous dTau expression is reduced as a compensatory measure. Consequently any detrimental effects caused by dTau accumulation would not be evident in hTau expressing flies but would still be present in unexpressed controls [42,43]. Discussion Age is the biggest risk factor for Alzheimer’s disease [44,45] and yet the mechanism(s) by which increasing age confers risk of disease are not well understood. In this study the impact of age on total and phosphorylated tau levels was studied in human brain tissue taken from young and old subjects, together with an analysis of age-related changes in autophagy, a path- way implicated in tau clearance [46]. With increasing age, a decrease in total tau levels and an alteration in two markers indicative of increased autophagy were evident in the absence of overt neurodegeneration. In order to test the impact of increased autophagy on human tau during aging, autophagy was genetically upregulated and tau-related behavioral changes were studied in young and old transgenic Drosophila. Co-expression of an autophagy gene, Atg1, with human tau ameliorated the accumulation of phosphorylated tau and significantly improved their climbing behaviour in an age-related manner in this model. Collectively these data suggest that age-related increases in autophagy may prevent the accumulation of tau and thus protect against tau-mediated pathogenic changes in vivo. It remains to be seen whether this is the mechanism by which elderly individuals who do not develop AD are protected and is worthy of further investigation. Similarly, whether it is a deviation from this trait that leads to compromised autophagy in some individuals as they age which then causes tau accumula- tion and downstream degeneration also needs further exploration (S7 Fig). Age-related changes in tau in non-AD brain Aging is defined as a time-dependent deterioration of the physiological functions of our body. Whilst the biological process of aging is complex and not well defined, it has been established that age is the primary risk factor for Alzheimer’s Disease (AD) with 90% of the AD cases occurring when the patients are 65 years or older (Alzheimer’s Association 2019). There are many theories to explain how aging confers risk of developing Alzheimer’s disease [5,6]. One view is that the for- mation of neuritic plaques and neurofibrillary tangles with increasing age [47] is a continuum of AD and that individuals with intermediate levels of these pathologies display mild cognitive symp- toms whilst those with higher levels of pathology display full blown dementia [4]. Though attrac- tive, this view is challenged by the finding that a significant number of individuals with neuritic PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 13 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain plaque and tangle pathology are not cognitively impaired [48,49]. This indicates that increased formation of such pathologies with increasing age cannot entirely explain how aging increases risk of developing dementia, and that other factors may also play a role [50]. Nonetheless, since there is a tight correlation between formation of neurofibrillary pathol- ogy and cognitive decline, it is worth investigating how tau proteins change in normal, non- demented individuals with increasing age. Whilst many studies have studied tau protein in brain and CSF from Alzheimer patients, and by default in most of these studies an analysis of age-matched controls is included, relatively few studies have assessed how tau protein itself changes with aging by directly comparing it in CNS tissues from young and old non-demented individuals. The data presented here demonstrates that total tau levels in cortical samples from older individuals are significantly reduced when compared to those found in younger individ- uals. This is likely to reflect an age-related decline in total soluble tau rather than a non-specific consequence of neuronal death during aging since the neuronal marker NeuN [51] does not change, indicating that there is no significant age-related neuronal loss in these cohorts, a phe- nomenon reported by others as well [5]. Our data cannot be attributed to brain-bank specific peculiarities in processing of tissues, as our samples were obtained from two distinct brain banks. Nor can it be attributed to age related changes in the tau mRNA levels. Additionally our data is in agreement with one other study that has similarly compared tau levels in young and old non-cognitively impaired human brain samples. They also showed that soluble total tau levels decreased with age in indi- viduals ranging from 19–88 years, but were not able to identify a potential mechanism. None- theless they did show the reduction in soluble tau could not be explained by a correlative increase in insoluble tau suggesting that there must be another explanation for the age-related reduction of soluble tau that both they and we observe [18]. As well as this, one more recent observation that also supports this result showed immunohistochemical decreases in tau in ret- inal biopsies taken from older subjects compared to those from younger subjects [52]. Building on these findings, we show here that in the brain samples where soluble tau is decreased, there is also an age-related increase in expression of late-stage markers of autophagy, but it is unclear whether this contributes to the age-related reduction in soluble tau. To address this issue we investigated whether that the reduction of soluble tau levels in the older subjects could also be due to transcriptional repression of the tau-gene. Interestingly, we observed equal amounts of tau isoforms indicating that the decrease of total tau that we observe in older cohorts is not due to decreased transcription in these samples. Even fewer studies have reported age-related changes in tau phosphorylation in non- demented individuals, let alone assess how these change with age. This is because the consen- sus view is that tau is not extensively phosphorylated in cortical regions of non-demented indi- viduals who do not have overt pathology [53] with only a limited number of studies reporting some physiologically phosphorylated epitopes in non-AD brains [54,55]. In agreement with this, very little phosphorylation of tau was evident, at the disease epitopes PHF-1, AT8 and phospho-serine 262 epitopes that were examined in this study, in any of the cortical brain tis- sues taken from non-demented individuals where there was a post-mortem delay. However, phosphorylation at the PHF1 site was clearly evident in resected cortical tissues taken from both young and old non-demented individuals, and there was a trend for this to increase with age, though this was not significant. One study published over 20 years ago made similar observations, showing that tau derived from biopsy non-AD brain tissue was more highly phosphorylated at several sites, thought to be primarily AD-specific, including PHF-1, AT8 and 12E8 and that phosphorylation at most of these sites was dramatically reduced if a post-mortem delay was artificially simulated, even if for 5–10 mins [24]. We have previously shown that this is also true for endogenously PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 14 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain phosphorylated rodent tau, where strong AT8 immunoreactivity is evident in hippocampal neu- rones (prominently localised to the somatodendritic compartment), which is reduced by nearly 75% when a post-mortem interval is simulated [27]. As well as a lack of PM delay, one cannot exclude the impact of one other difference between the resected tissue and the PM tissue which is that in the former the patients underwent anaesthesia, which has been shown to increase tau phosphorylation [56,57]. However, we have previously demonstrated that in rodent brain, even in the absence of anaesthesia, tau is significantly more phosphorylated when tissue is processed quickly without PM delay compared to when it is processed with a PM delay [27]. A post-mor- tem delay of several minutes is also created by default during trans-cardial perfusion fixation in rodent models of tauopathy, and this may also influence the phosphorylation status of any solu- ble tau species in those models. Collectively, these findings and our own strongly suggest that soluble tau is more phosphorylated in cortical tissues from non-demented individuals than is generally reported, but that the phosphorylation status is significantly reduced by the post-mor- tem delay. Phosphorylation status is adequately preserved in resected tissue, that though taken from individuals undergoing neurosurgery, is taken from brain regions that are physiologically and ultra-structurally normal as shown by us (in S1 Fig) and previously by others [12,15]. Stud- ies in such tissues may be useful in future to assess if physiological phosphorylation of soluble tau increases with age, enhancing age-related risk of aggregation. Tau aggregates composed of soluble and insoluble oligomers are one of the crucial factors driving the spread of tau pathology in the AD brain [58]. However, previous studies from Ladinska et al. has shown that the decrease in soluble tau observed in non-demented elderly cohorts does not correlate with the increase of insoluble tau in these brain samples [18]. There- fore, we did not investigate the levels of tau aggregation in these samples. Age-related changes in cytoskeletal integrity Since microtubule stabilisation is a key physiological function of tau, one may hypothesize that the age-related reduction in total tau may manifest in reduced cytoskeletal integrity. To assess the impact of age-related reduction of total tau, on the neuronal cytoskeleton, acetylation and tyrosination of alpha-tubulin, two post-translational modifications that may indicate its stabil- ity, were probed [59–61]. Despite the significant reduction in total tau levels in the elderly cohorts compared to the younger cohorts, no change was observed in the acetylation or tyrosi- nation levels of alpha-tubulin between the two groups. This implies that the loss of tau in the healthy control brains does not affect altered stability of microtubules with increasing age. It is possible that microtubule stability is only disrupted in the situation of tau hyperphosphoryla- tion and profound neuronal loss that is observed in AD but not in aging control brains, or that age-related loss of tau is compensated for by an upregulation of other microtubule-associated proteins thereby preserving the neuronal cytoskeleton. This result is in disagreement with that of Cash et al. [31], who reported an age-related reduction in cytoskeletal integrity in cortical biopsy samples from non-AD patients. However, their cortical samples were taken from more elderly individuals than ours, and their age range spanned 2 decades, ranging from 62–80 years, whereas ours spanned 5 decades ranging from 24–74 years. Age-related increase in markers of autophagy and age-related decline in total tau In order to examine the underlying biochemical pathways possibly responsible for the age-related reduction of total tau, we explored whether there were age-related changes in macroautophagy, a pathway implicated in tau turnover [62,63]. There is a lot of interest in protoeostatic pathways such as autophagy and the ubiquitin-proteasome system, as they have been shown to regulate PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 15 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain turnover of tau, and its accumulation in Tauopathies is postulated to arise due to an age-related functional decline in the efficiency of these pathways [64–66]. However, although there are numerous studies reporting autophagic induction in AD brains and comparing that to age- matched control brain [67], there are only limited reports comparing markers of autophagy in brains of young and old non-AD subjects. Two studies that have examined this suggest that macroautophagy decreases with increasing age in normal brain, as shown by an age-related down- regulation of several autophagy related genes [68] and Beclin 1, a protein essential for autophagy [69]. The data presented here suggests that the situation is more complicated and one cannot nec- essarily assume that there is an age-related reduction in autophagy as the consensus view suggests. We show that two markers of late-stage autophagy are increased with age, the ratio of LC3 II/I, a sensitive marker of macroautophagy [70], with a corresponding reduction in p62 (which is destroyed during autophagy), thus implying greater autophagy [71], whilst one marker of early stage autophagy, Beclin 1 is unchanged. Indeed increased autophagy with advancing age has been reported in other tissues [72,73] and in choroid epithelial cells in brains of non-AD subjects [74]. This is further supported by some studies of aging in experimental models, where markers of autophagy including LC3 II/I ratios and pro-autophagic regulators like BAG-3 are higher in brains of old animals when compared to young animals [75]. However, there are conflicting findings suggesting age-related decline in autophagy in other experimen- tal models [76]. Confusing the issue further, autophagy is believed to be upregulated with increasing age in long-lived animals and in healthy long-lived humans [77,78]. Next we analyzed Cathepsin D which is a lysosomal protease that has been found to be upregulated in AD [79]. In our study the levels of Cathepsin D were not significantly altered between the younger and the older cohorts in the post-mortem brain despite seeing a signifi- cant increase in some autophagy markers in the older cohorts. Although there is an intimate connection between Cathepsin D and autophagy, this has been mostly reported in the neuro- logical disease conditions [80]. Given that these samples were collected from non-AD subjects there may not be detectable alterations in the levels of Cathepsin D. The findings presented here and the conflicting reports from other studies on physiological age-related changes in autophagic capacity, very few of which were conducted using material from human brain, collectively imply that the situation is complex and requires further investi- gation [81]. Such investigations should also assess the cellular localisation of markers of autop- hagy, since some markers, like LC3-II, may appear to increase in conditions when autophagic flux is blocked, complicating interpretation of biochemical readouts [71]. Only then could a consensus be reached as to whether autophagic capacity changes with age, especially in human brain, and whether this per se confers higher risk of neurodegenerative proteinopathies in the elderly. Whilst the evidence for changes in autophagy with age is confusing, it has consistently been shown by many that autophagic clearance can regulate tau turnover [20,81]. The age-related decline in soluble tau that we report may therefore occur at least in part due to the age-related increase in autophagy suggested by the increased markers of autophagy that is also evident in these brains. However further studies will be needed to confirm this and the role of lysosomal proteases in degrading tau in post-mortem non-demented AD brains. Furthermore, since autophagy deficiency leads to age-dependent neurodegeneration [62], it is possible that the lack of neurodegeneration in the brains of the older subjects in this study is a consequence of their healthy autophagic capacity. One may speculate that an age-related increase in effective autophagic capacity is a central component of healthy aging and protects healthy individuals from developing AD as they age (S7A Fig blue line). However, deviation from this, for example if the soluble tau (or amyloid beta) load overwhelms the autophagic system, could trigger the onset of AD in some individuals (S7A Fig, red line). Indeed, many studies have shown that PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 16 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain macroautophagy is impaired in AD brains [67] as well as in animal models of tauopathy [82]. An alternative hypothesis, based on the consensus view discussed earlier, is that a reduction in autophagic capacity is evident in all individuals with increasing age, and this is the reason why some individuals might develop AD (S7B Fig). Though the latter does not explain why only some aged indiviudals go on to develop proteineopathy. Another interesting study conducted by Kang et al, demonstrated that both normal and pathological tau is spread in the neurons by autophagy inducers [83]. Hence, further studies, using additional markers of autophagy in a larger cohort are needed to distinguish between these two scenarios. Upregulation of autophagy rescues age-related tau phenotypes Our observation of increased markers of autophagy in aging human brains, evident in the face of decreased total tau levels and the absence of tau pathology, led us to hypothesize that induc- tion of autophagy could ameliorate tau abnormalities in an age-dependent manner in a Dro- sophila model. To test this hypothesis, we assessed the impact of upregulating autophagy in a Drosophila model of Tauopathy in which we have previously described age-related tau pheno- types [22]. Drosophila is an ideal organism in which to test such hypotheses because it has been used for decades for genetically dissecting the pathways that underpin aging, autophagy and tau-mediated neurodegenerative diseases [84–86]. Indeed others have previously also shown that induction of autophagy in both rodent and Drosophila models of tauopathy reduces tau pathology and phenotypes [87–92] but the impact of aging has not been investigated. In agreement with these studies, our data show that co-expression of Atg1, a key autophagy protein ameliorates the age-dependent increase in tau phosphorylation and significantly improves the climbing ability of hTau0N3R expressing flies, a behaviour we have previously shown is associated with tau hyperphosphorylation [23]. Interestingly, the beneficial effect of inducing autophagy was age-dependent, being more evident in older flies compared to youn- ger flies. That increased autophagy is benefical for aging flies, as our data indicates, is sup- ported by several other reports which demonstrate that both pharmacological and genetic upregulation of autophagy extends lifespan in invertebrate and vertebrate models of aging [81,90], which in some cases is accompanied by suppression of brain aging and a reduction in total tau [91,92]. The consistent conclusion that can be drawn from all such studies is that robust autophagic capacity may be key for preventing the age-related accumulation and thus toxicity of aggregate- prone proteins such as tau, thus enabling healthy brain aging and protecting against neurodegen- eration. Our studies also add to the growing weight of evidence suggesting that the activation of autophagy could be an important therapeutic target for the treatment of tauopathies [93–97]. Conclusions Our data highlight the importance of studying the effect of aging on the normal physiological turnover of aggregate-prone proteins like tau as well as the mechanistic pathways that might regulate this in the absence of disease. It is especially important to undertake such studies using normal human brain tissue to identify the impact of age-related changes without the confounding influence of underlying neurodegenerative disease. This is particularly relevant in understanding how aging confers risk of proteinopathies such as Alzheimer’s disease. Supporting information S1 Fig. Electrophysiological recordings and ultrastructure of cortical resected tissue. Whole-cell patch clamp recording from a layer II-III human cortical pyramidal neuron PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 17 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain showing active voltage responses to current injection (A). Electron micrograph showing a cor- tical synapse. Scale Bar 200 nm (B). (TIF) S2 Fig. Phospho-tau immunoreactivity (AT8 and phospho-262) in the post-mortem brain samples was low. Representative Western Blot images for AT8, ps-262 and β-actin in post- mortem brains (A and B). There was no significant difference in the AT8 and ps-262 immuno- reactivities between the younger vs older cohorts. (TIF) S3 Fig. Electrophoresis of RT-PCR amplification products of tau mRNA show no signifi- cant difference between the younger and the older cohorts of the post-mortem brain sam- ples. The three PCR products correspond to the 2N, 1N and 0N forms of tau in the top panel and Actin in the bottom panel (A). Quantitation of 2N, 1N and 0N tau normalized to Actin does not display a significant difference between the cohorts (B). (TIF) S4 Fig. Expression of the lysosomal protease Cathepsin D was not significantly different between the younger and the older cohorts in the post-mortem brain tissues. Representa- tive western blot probed with Cathepsin D and b-actin loading control (A). Quantitation of Cathepsin D normalised to β-Actin displays no significant difference. (TIF) S5 Fig. Unexpressed controls. Elav Geneswitch (GS) expression system allows for temporal control of pan-neural hTau0N3R expression following introduction of RU486 upon eclosion. No expression is evident when no drug is given (0μm RU486) and a dose dependent increase in tau expression is seen with increasing doses of RU486. 200 μM RU486 gives tau expression that is comparable to that seen with the standard Elav Gal4 so this dose was chosen for all the experiments in this study (A). Atg8 staining in Elav GS/UAS-Atg1, Elav GS/hTau0N3R and Elav GS/htau0N3R;Atg1 transgenics after treatment with 200 μM RU486 upon eclosion. During autophagy, Atg8-I is conjugated with a lipid moiety and the lapidated form (Atg8-II) is recruited to the autophagosomal membranes. An elevated Atg8-II level therefore serves as an indicator of autophagy activation. Atg8-II levels are evident in the Atg1 and hTau0N3R;Atg1 flies but not to that extent in the hTau0N3R alone flies (B). (TIF) S6 Fig. Impact of hTau expression on endogenous dTau. Drosophila tau (dTau) was probed using an antibody specific only to dTau (gift from Johnston lab University of Cambridge) and in wild type (wt) OreR flies, dTau levels were found to increase with age (p = 0.0193). This age- related increase was not evident upon Elav Gal4 expression of hTau0N3R. The dTau expres- sion in 6 week wt flies that are expressing hTau0N3R was significantly less than that evident in 6 week wt flies (p = 0,0309) (data is fold change compared to dTau levels in wt flies at 0 week; unpaired two tailed t-test; n = 7–14). (TIF) S7 Fig. Two hypotheses for how autophagy may change with age and influence accumula- tion of misfolded proteins. In individuals with a healthy autophagic capacity, autophagic responses increase with age to deal with age-related insults. This may contribute to a decline in soluble phosphorylated tau levels and protect from tau accumulation and sunsequent degener- ation in healthy long-lived elderly subjects who do not develop tauopathy (red line) (A). An age-related decline in autophagic capacity is evident in all elderly but in some subjects (blue line) this is not significant enough to encourage tau accumulation and they do not develop PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 18 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain degeneration or tauopathy in other subjects (red line), the age-related impairment occurs to a greater extent and this precipates accumulation of misfolded proteins and pathogenesis of neurodegenerative diseases like tauopathy. (TIF) S1 Raw images. (PDF) Acknowledgments We thank Dr. Anton Page for his electron microscopy training and advice and Ms. Hannah Warming for tissue collection and retrieval. Author Contributions Conceptualization: Amritpal Mudher. Data curation: Megan Sealey. Formal analysis: Shreyasi Chatterjee, Megan Sealey. Funding acquisition: Amritpal Mudher. Investigation: Shreyasi Chatterjee, Megan Sealey, Eva Ruiz, Anna Crisford, Emma Luckett, Rebecca Robertson, Philippa Richardson, Paul Grundy, Diederik Bulters, Mariana Vargas- Caballero. Methodology: Shreyasi Chatterjee, Megan Sealey, Eva Ruiz, Keeley Brookes, Sam Green, Anna Crisford, Michael Duque-Vasquez, Emma Luckett, Rebecca Robertson, Philippa Richard- son, Girish Vajramani, Paul Grundy, Diederik Bulters, Mariana Vargas-Caballero. Project administration: Shreyasi Chatterjee. Resources: Chrysia M. Pegasiou, Girish Vajramani, Mariana Vargas-Caballero. Supervision: Amritpal Mudher. Writing – original draft: Shreyasi Chatterjee. Writing – review & editing: Christopher Proud, Mariana Vargas-Caballero, Amritpal Mudher. References 1. Yankner B.A., Lu T. and Loerch P. (2008). "The aging brain." Annu Rev Pathol 3: 41–66. https://doi. org/10.1146/annurev.pathmechdis.2.010506.092044 PMID: 18039130 2. Bishop N.A., Lu T. and Yankner B.A. (2010). "Neural mechanisms of ageing and cognitive decline." Nature 464(7288): 529–535. https://doi.org/10.1038/nature08983 PMID: 20336135 3. Bancher C., Braak H., Fischer P. and Jellinger K.A. (1993). “Neuropathological Staging of Alzheimer Lesions and Intellectual Status in Alzheimer’s and Parkinson’s Disease Patients”. Neurosci Lett Nov 12; 162(1–2):179–82. https://doi.org/10.1016/0304-3940(93)90590-h PMID: 8121624 4. Serrano-Pozo A., Frosch M.P., Masliah. E. and Hyman. B.T. (2011). "Neuropathological alterations in Alzheimer disease." Cold Spring Harb Perspect Med 1(1): a006189. https://doi.org/10.1101/ cshperspect.a006189 PMID: 22229116 5. Xia X., Jiang Q., McDermott J. and Han J.J. (2018). "Aging and Alzheimer’s disease: Comparison and associations from molecular to system level." Aging Cell 17(5): e12802. https://doi.org/10.1111/acel. 12802 PMID: 29963744 PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 19 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain 6. Vemuri P. (2018). ""Exceptional brain aging" without Alzheimer’s disease: triggers, accelerators, and the net sum game." Alzheimers Res Ther 10(1): 53. https://doi.org/10.1186/s13195-018-0373-z PMID: 29859131 7. Peters R. (2006). "Ageing and the brain." Postgrad Med J 82(964): 84–88. https://doi.org/10.1136/ pgmj.2005.036665 PMID: 16461469 8. Piras A., Collin L., Gruninger F., Graff C. and Ronnback A. (2016). "Autophagic and lysosomal defects in human tauopathies: analysis of post-mortem brain from patients with familial Alzheimer disease, cor- ticobasal degeneration and progressive supranuclear palsy." Acta Neuropathol Commun 4: 22. https:// doi.org/10.1186/s40478-016-0292-9 PMID: 26936765 9. Ward W.F. (2002). "Protein degradation in the aging organism." Prog Mol Subcell Biol 29: 35–42. https://doi.org/10.1007/978-3-642-56373-7_3 PMID: 11908071 10. Caballero B., Wang Y., Diaz A., Tasset I., Juste Y.R., Stiller B., et al. (2018). "Interplay of pathogenic forms of human tau with different autophagic pathways." Aging Cell 17(1). https://doi.org/10.1111/acel. 12692 PMID: 29024336 11. Verhoog M. B., Goriounova N.A., Obermayer J., Stroeder J., Hjorth J.J., Testa-Silva G., et al. (2013). "Mechanisms underlying the rules for associative plasticity at adult human neocortical synapses." J Neurosci 33(43): 17197–17208. https://doi.org/10.1523/JNEUROSCI.3158-13.2013 PMID: 24155324 12. Dolan P. J. and Johnson G.V. (2010). "A caspase cleaved form of tau is preferentially degraded through the autophagy pathway." J Biol Chem 285(29): 21978–21987. https://doi.org/10.1074/jbc.M110. 110940 PMID: 20466727 13. Wang A. L., Lukas T.J., Yuan M., Du N., Tso M.O. and Neufeld A.H. (2009). "Autophagy, exosomes and drusen formation in age-related macular degeneration." Autophagy 5(4): 563–564. https://doi.org/ 10.4161/auto.5.4.8163 PMID: 19270489 14. Kaushik S. and Cuervo A.M. (2015). "Proteostasis and aging." Nat Med 21(12): 1406–1415. https://doi. org/10.1038/nm.4001 PMID: 26646497 15. Bar-Yosef T., Damri O. and Agam G. (2019). "Dual Role of Autophagy in Diseases of the Central Ner- vous System." Front Cell Neurosci 13: 196. https://doi.org/10.3389/fncel.2019.00196 PMID: 31191249 16. Scrivo A., Bourdenx M., Pampliega O. and Cuervo A.M. (2018). "Selective autophagy as a potential therapeutic target for neurodegenerative disorders." Lancet Neurol 17(9): 802–815. https://doi.org/10. 1016/S1474-4422(18)30238-2 PMID: 30129476 17. Loeffler D.A. (2019). "Influence of Normal Aging on Brain Autophagy: A Complex Scenario." Front Aging Neurosci 11: 49. https://doi.org/10.3389/fnagi.2019.00049 PMID: 30914945 18. Mukaetova-Ladinska E.B., Harrington C.R., Roth M. and Wischik C.M. (1996). "Alterations in tau protein metabolism during normal aging." Dementia 7(2): 95–103. https://doi.org/10.1159/000106861 PMID: 8866683 19. Pegasiou C.M., Zolnourian A., Gomez-Nicola D., Deinhardt K., Nicoll J.A.R., Ahmed A.I., et al. (2020). "Age-Dependent Changes in Synaptic NMDA Receptor Composition in Adult Human Cortical Neurons." Cereb Cortex. https://doi.org/10.1093/cercor/bhaa052 PMID: 32191258 20. Lopez A., Lee S.E., Wojta K., Ramos E.M., Klein E., Chen J., et al. (2017). "A152T tau allele causes neurodegeneration that can be ameliorated in a zebrafish model by autophagy induction." Brain 140(4): 1128–1146. https://doi.org/10.1093/brain/awx005 PMID: 28334843 21. Uddin M. S., Stachowiak A., Mamun A. A., Tzvetkov N. T., Takeda S., Atanasov A. G., et al. (2018). "Autophagy and Alzheimer’s Disease: From Molecular Mechanisms to Therapeutic Implications." Front Aging Neurosci 10: 04. https://doi.org/10.3389/fnagi.2018.00004 PMID: 29441009 22. Boutajangout A., Boom A., Leroy K., & Brion J. P. (2004). Expression of tau mRNA and soluble tau iso- forms in affected and non-affected brain areas in Alzheimer’s disease. FEBS Letters, 576(1–2). https:// doi.org/10.1016/j.febslet.2004.09.011 PMID: 15474035 23. Sealey M.A., Vourkou E., Cowan C.M., Bossing T., Quraishe S., Grammenoudi S., et al. (2017). "Dis- tinct phenotypes of three-repeat and four-repeat human tau in a transgenic model of tauopathy." Neuro- biol Dis 105: 74–83. https://doi.org/10.1016/j.nbd.2017.05.003 PMID: 28502805 24. Seubert P., Mawal-Dewan M., Barbour R., Jakes R., Goedert M., Johnson G.V., et al. (1995). "Detec- tion of phosphorylated Ser262 in fetal tau, adult tau, and paired helical filament tau." J Biol Chem 270 (32): 18917–18922. https://doi.org/10.1074/jbc.270.32.18917 PMID: 7642549 25. Schneider A., Biernat J., von Bergen M., Mandelkow E. and Mandelkow E.M. (1999). "Phosphorylation that detaches tau protein from microtubules (Ser262, Ser214) also protects it against aggregation into Alzheimer paired helical filaments." Biochemistry 38(12): 3549–3558. https://doi.org/10.1021/ bi981874p PMID: 10090741 26. Matsuo E.S., Shin R.W., Billingsley M.L., Van deVoorde A., O’Connor M., Trojanowski J.Q. et al. (1994). "Biopsy-derived adult human brain tau is phosphorylated at many of the same sites as PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 20 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain Alzheimer’s disease paired helical filament tau." Neuron 13(4): 989–1002. https://doi.org/10.1016/ 0896-6273(94)90264-x PMID: 7946342 27. Mudher A.K., Woolley S.T., Perry V.H. and Greene J.R. (1999). "Induction of hyperphosphorylated tau in living slices of rat hippocampal formation and subsequent detection using an ELISA." J Neurosci Methods 88(1): 15–25. https://doi.org/10.1016/s0165-0270(99)00006-0 PMID: 10379575 28. Eckenhoff R. G. and Planel E. (2012). "Postoperative cognitive decline: where art tau?" Anesthesiology 116(4): 751–752. https://doi.org/10.1097/ALN.0b013e31824be8e1 PMID: 22343470 29. Singh-Manoux A., Kivimaki M., Glymour M.M., Elbaz A., Berr C., Ebmeier K. P., et al. (2012). "Timing of onset of cognitive decline: results from Whitehall II prospective cohort study." BMJ 344: d7622. https:// doi.org/10.1136/bmj.d7622 PMID: 22223828 30. Buee L., Bussiere T., Buee-Scherrer V., Delacourte A. and Hof P.R. (2000). "Tau protein isoforms, phosphorylation and role in neurodegenerative disorders." Brain Res Brain Res Rev 33(1): 95–130. https://doi.org/10.1016/s0165-0173(00)00019-9 PMID: 10967355 31. Breuzard G., Hubert P., Nouar R., De Bessa T., Devred F., Barbier P., et al. (2013). "Molecular mecha- nisms of Tau binding to microtubules and its role in microtubule dynamics in live cells." J Cell Sci 126(Pt 13): 2810–2819. https://doi.org/10.1242/jcs.120832 PMID: 23659998 32. Cash A.D., Aliev G., Siedlak S.L., Nunomura A., Fujioka H., Zhu X., et al. (2003). "Microtubule reduction in Alzheimer’s disease and aging is independent of tau filament formation." Am J Pathol 162(5): 1623– 1627. https://doi.org/10.1016/s0002-9440(10)64296-4 PMID: 12707046 33. Brandt R. and Bakota L. (2017). "Microtubule dynamics and the neurodegenerative triad of Alzheimer’s disease: The hidden connection." J Neurochem 143(4): 409–417. https://doi.org/10.1111/jnc.14011 PMID: 28267200 34. Hammond J.W., Cai D. and Verhey K.J. (2008). "Tubulin modifications and their cellular functions." Curr Opin Cell Biol 20(1): 71–76. https://doi.org/10.1016/j.ceb.2007.11.010 PMID: 18226514 35. 36. 37. Janke C. and Montagnac G. (2017). "Causes and Consequences of Microtubule Acetylation." Curr Biol 27(23): R1287–R1292. https://doi.org/10.1016/j.cub.2017.10.044 PMID: 29207274 Liu W.J., Ye L., Huang W.F., Guo L.J., Xu Z.G., Wu H.L., et al. (2016). "p62 links the autophagy path- way and the ubiqutin-proteasome system upon ubiquitinated protein degradation." Cell Mol Biol Lett 21: 29. https://doi.org/10.1186/s11658-016-0031-z PMID: 28536631 Tanida I., Ueno T. and Kominami E. (2008). "LC3 and Autophagy." Methods Mol Biol 445: 77–88. https://doi.org/10.1007/978-1-59745-157-4_4 PMID: 18425443 38. Pickford F., Masliah E., Britschgi M., Lucin K., Narasimhan R., Jaeger P.A., et al. (2008). "The autop- hagy-related protein beclin 1 shows reduced expression in early Alzheimer disease and regulates amy- loid beta accumulation in mice." J Clin Invest 118(6): 2190–2199. https://doi.org/10.1172/JCI33585 PMID: 18497889 39. Kraus A., Saijo E., Metrick M. A., Newell K., Sigurdson C. J., Zanusso G., et al. (2019). Seeding selectiv- ity and ultrasensitive detection of tau aggregate conformers of Alzheimer disease. Acta Neuropatholo- gica, 137(4). https://doi.org/10.1007/s00401-018-1947-3 PMID: 30570675 40. Scott R.C., Juhasz G. and Neufeld T.P. (2007). "Direct induction of autophagy by Atg1 inhibits cell growth and induces apoptotic cell death." Curr Biol 17(1): 1–11. https://doi.org/10.1016/j.cub.2006.10. 053 PMID: 17208179 41. Kijak E. and Pyza E. (2017). "TOR signaling pathway and autophagy are involved in the regulation of cir- cadian rhythms in behavior and plasticity of L2 interneurons in the brain of Drosophila melanogaster." PLoS One 12(2): e0171848. https://doi.org/10.1371/journal.pone.0171848 PMID: 28196106 42. Cowan C. M., Bossing T., Page A., Shepherd D. and Mudher A. (2010). "Soluble hyper-phosphorylated tau causes microtubule breakdown and functionally compromises normal tau in vivo." Acta Neuropathol 120(5): 593–604. https://doi.org/10.1007/s00401-010-0716-8 PMID: 20617325 43. Ubhi K. K., Shaibah H., Newman T. A., Shepherd D. and Mudher A. (2007). "A comparison of the neuro- nal dysfunction caused by Drosophila tau and human tau in a Drosophila model of tauopathies." Invert Neurosci 7(3): 165–171. https://doi.org/10.1007/s10158-007-0052-4 PMID: 17636367 44. Guerreiro R. and Bras J. (2015). "The age factor in Alzheimer’s disease." Genome Med 7: 106. https:// doi.org/10.1186/s13073-015-0232-5 PMID: 26482651 45. Edwards Iii G.A., Gamez N., Escobedo G., Calderon O. and Moreno-Gonzalez I. (2019). "Modifiable Risk Factors for Alzheimer’s Disease." Front Aging Neurosci 11: 146. https://doi.org/10.3389/fnagi. 2019.00146 PMID: 31293412 46. Limanaqi F., Biagioni F., Gambardella S., Familiari P., Frati A. and Fornai F. (2020). "Promiscuous Roles of Autophagy and Proteasome in Neurodegenerative Proteinopathies." Int J Mol Sci 21(8). https://doi.org/10.3390/ijms21083028 PMID: 32344772 PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 21 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain 47. Braak H., Braak E. and Bohl J. (1993). “Staging of Alzheimer-related cortical destruction.” Eur Neurol 33(6):403–8. https://doi.org/10.1159/000116984 PMID: 8307060 48. Arriagada P.V., Growdon J.H., Hedley-Whyte E.T. and Hyman B.T. (1992). "Neurofibrillary tangles but not senile plaques parallel duration and severity of Alzheimer’s disease." Neurology 42(3 Pt 1): 631– 639. https://doi.org/10.1212/wnl.42.3.631 PMID: 1549228 49. Tanprasertsuk J., Johnson E.J., Johnson M.A., Poon L.W., Nelson P.T., Davey A., et al. (2019). "Clin- ico-Neuropathological Findings in the Oldest Old from the Georgia Centenarian Study." J Alzheimers Dis 70(1): 35–49. https://doi.org/10.3233/JAD-181110 PMID: 31177211 50. Chappell S., Patel T., Guetta-Baranes T., Sang F., Francis P. T., Morgan K., et al. (2018). Observations of extensive gene expression differences in the cerebellum and potential relevance to Alzheimer’s dis- ease. BMC Research Notes, 11(1). https://doi.org/10.1186/s13104-018-3732-8 PMID: 30180886 51. Mullen R.J., Buck C.R. and Smith A.M. (1992). "NeuN, a neuronal specific nuclear protein in verte- brates." Development 116(1): 201–211. https://doi.org/10.1242/dev.116.1.201 PMID: 1483388 52. Bouras C., Hof P.R., Giannakopoulos P., Michel J.P. and Morrison J.H. (1994). "Regional distribution of neurofibrillary tangles and senile plaques in the cerebral cortex of elderly patients: a quantitative evalua- tion of a one-year autopsy population from a geriatric hospital." Cereb Cortex 4(2): 138–150. https:// doi.org/10.1093/cercor/4.2.138 PMID: 8038565 53. Nag T.C, Kathpalia K. P. and Wadhwa S. (2020). “Microtubule alterations may destabilize photorecep- tor integrity: Age-related microtubule changes and pattern of expression of MAP-2, Tau and hyperpho- sphorylated Tau in aging human photoreceptor cells.” Exp Eye Res 198: 1–15. https://doi.org/10.1016/ j.exer.2020.108153 PMID: 32710889 54. Lowe V.J., Wiste H.J., Senjem M.L., Weigand S.D., Therneau T.M., Boeve B.F., et al. (2018). "Wide- spread brain tau and its association with ageing, Braak stage and Alzheimer’s dementia." Brain 141(1): 271–287. https://doi.org/10.1093/brain/awx320 PMID: 29228201 55. Dickey C., Kraft C., Jinwal U., Koren J., Johnson A., Anderson L., et al. (2009). "Aging analysis reveals slowed tau turnover and enhanced stress response in a mouse model of tauopathy." Am J Pathol 174 (1): 228–238. https://doi.org/10.2353/ajpath.2009.080764 PMID: 19074615 56. Metaxas A., Thygesen C., Kempf S.J., Anzalone M., Vaitheeswaran R., Petersen S., et al. (2019). "Age- ing and amyloidosis underlie the molecular and pathological alterations of tau in a mouse model of famil- ial Alzheimer’s disease." Sci Rep 9(1): 15758. https://doi.org/10.1038/s41598-019-52357-5 PMID: 31673052 57. Planel E., Richter K.E., Nolan C.E., Finley J.E., Liu L., Wen Y., et al. (2007). "Anesthesia leads to tau hyperphosphorylation through inhibition of phosphatase activity by hypothermia." J Neurosci 27(12): 3090–3097. https://doi.org/10.1523/JNEUROSCI.4854-06.2007 PMID: 17376970 58. Ercan-Herbst E., Ehrig J., Schondorf D.C., Behrendt A., Klaus B., Gomez Ramos B., et al. (2019). "A post-translational modification signature defines changes in soluble tau correlating with oligomerization in early stage Alzheimer’s disease brain." Acta Neuropathol Commun 7(1): 192. https://doi.org/10. 1186/s40478-019-0823-2 PMID: 31796124 59. Wloga D., Joachimiak E. and Fabczak H. (2017). "Tubulin Post-Translational Modifications and Microtu- bule Dynamics." Int J Mol Sci 18(10). https://doi.org/10.3390/ijms18102207 PMID: 29065455 60. Zhang F., Su B., Wang C., Siedlak S.L., Mondragon-Rodriguez S., Lee H.G., et al. (2015). "Posttransla- tional modifications of alpha-tubulin in alzheimer disease." Transl Neurodegener 4: 9. 61. Hempen B. and Brion J.P. (1996). "Reduction of acetylated alpha-tubulin immunoreactivity in neurofi- brillary tangle-bearing neurons in Alzheimer’s disease." J Neuropathol Exp Neurol 55(9): 964–972. https://doi.org/10.1097/00005072-199609000-00003 PMID: 8800092 62. 63. Inoue K., Rispoli J., Kaphzan H., Klann E., Chen E.I., Kim J., et al. (2012). "Macroautophagy deficiency mediates age-dependent neurodegeneration through a phospho-tau pathway." Mol Neurodegener 7: 48. https://doi.org/10.1186/1750-1326-7-48 PMID: 22998728 Ling D. and Salvaterra P.M. (2009). "A central role for autophagy in Alzheimer-type neurodegenera- tion." Autophagy 5(5): 738–740. https://doi.org/10.4161/auto.5.5.8626 PMID: 19395865 64. Mputhia Z., Hone E., Tripathi T., Sargeant T., Martins R. and Bharadwaj P. (2019). "Autophagy Modula- tion as a Treatment of Amyloid Diseases." Molecules 24(18). https://doi.org/10.3390/ molecules24183372 PMID: 31527516 65. Djajadikerta A., Keshri S., Pavel M., Prestil R., Ryan L. and Rubinsztein D.C. (2020). "Autophagy Induc- tion as a Therapeutic Strategy for Neurodegenerative Diseases." J Mol Biol 432(8): 2799–2821. https:// doi.org/10.1016/j.jmb.2019.12.035 PMID: 31887286 66. Vargas-Caballero M., Willaime-Morawek S., Gomez-Nicola D., Perry V.H., Bulters D. and Mudher A. (2016). "The use of human neurons for novel drug discovery in dementia research." Expert Opin Drug Discov 11(4): 355–367. https://doi.org/10.1517/17460441.2016.1154528 PMID: 26878555 PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 22 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain 67. Nixon R.A., Wegiel J., Kumar A., Yu W.H., Peterhoff C., Cataldo A. et al. (2005). "Extensive involvement of autophagy in Alzheimer disease: an immuno-electron microscopy study." J Neuropathol Exp Neurol 64(2): 113–122. https://doi.org/10.1093/jnen/64.2.113 PMID: 15751225 68. Lipinski M.M., Zheng B., Lu T., Yan Z., Py B.F., Ng A., et al. (2010). "Genome-wide analysis reveals mechanisms modulating autophagy in normal brain aging and in Alzheimer’s disease." Proc Natl Acad Sci U S A 107(32): 14164–14169. https://doi.org/10.1073/pnas.1009485107 PMID: 20660724 69. Shibata M., Lu T., Furuya T., Degterev A., Mizushima N., Yoshimori T., et al. (2006). "Regulation of intracellular accumulation of mutant Huntingtin by Beclin 1." J Biol Chem 281(20): 14474–14485. https://doi.org/10.1074/jbc.M600364200 PMID: 16522639 70. Karim M.R., Kanazawa T., Daigaku Y., Fujimura S., Miotto G. and Kadowaki M. (2007). "Cytosolic LC3 ratio as a sensitive index of macroautophagy in isolated rat hepatocytes and H4-II-E cells." Autophagy 3(6): 553–560. https://doi.org/10.4161/auto.4615 PMID: 17617739 71. Klionsky D.J., Abdalla F.C., Abeliovitch H., Abraham R.T. et al (2012). “Guidelines for the Use and Inter- pretation of Assays for Monitoring Autophagy” Autophagy 8(4):445–544. https://doi.org/10.4161/auto. 19496 PMID: 22966490 72. Grimm A. and Eckert A. (2017). "Brain aging and neurodegeneration: from a mitochondrial point of view." J Neurochem 143(4): 418–431. https://doi.org/10.1111/jnc.14037 PMID: 28397282 73. Ward W. and Richardson A. (1991). "Effect of age on liver protein synthesis and degradation." Hepatol- ogy 14(5): 935–948. https://doi.org/10.1002/hep.1840140529 PMID: 1937398 74. Ward W.F. (2002). "Protein degradation in the aging organism." Prog Mol Subcell Biol 29: 35–42. https://doi.org/10.1007/978-3-642-56373-7_3 PMID: 11908071 75. Gamerdinger M., Hajieva P., Kaya A.M., Wolfrum U., Hartl F.U. and Behl C. (2009). "Protein quality con- trol during aging involves recruitment of the macroautophagy pathway by BAG3." EMBO J 28(7): 889– 901. https://doi.org/10.1038/emboj.2009.29 PMID: 19229298 76. Bakhoum M.F., Bakhoum C.Y., Ding Z., Carlton S.M., Campbell G.A. and Jackson G.R. (2014). "Evi- dence for autophagic gridlock in aging and neurodegeneration." Transl Res 164(1): 1–12. https://doi. org/10.1016/j.trsl.2014.01.016 PMID: 24561013 77. De Biase D., Costagliola A., Pagano T.B., Piegari G., Wojcik S., Dziewiatkowski J., et al. (2017). "Amy- loid precursor protein, lipofuscin accumulation and expression of autophagy markers in aged bovine brain." BMC Vet Res 13(1): 102. https://doi.org/10.1186/s12917-017-1028-1 PMID: 28407771 78. Nakamura S. and Yoshimori T. (2018). "Autophagy and Longevity." Mol Cells 41(1): 65–72. https://doi. org/10.14348/molcells.2018.2333 PMID: 29370695 79. Chai Y. L., Chong J. R., Weng J., Howlett D., Halsey A., Lee J. H., et al. (2019). Lysosomal cathepsin D is upregulated in Alzheimer’s disease neocortex and may be a marker for neurofibrillary degeneration. Brain Pathology, 29(1). https://doi.org/10.1111/bpa.12631 PMID: 30051532 80. Mrschtik M., & Ryan K. M. (2015). Lysosomal proteins in cell death and autophagy. In FEBS Journal (Vol. 282, Issue 10). https://doi.org/10.1111/febs.13253 PMID: 25735653 81. Berger Z., Ravikumar B., Menzies F.M., Oroz L.G., Underwood B.R., Pangalos M.N., et al. (2006). "Rapamycin alleviates toxicity of different aggregate-prone proteins." Hum Mol Genet 15(3): 433–442. https://doi.org/10.1093/hmg/ddi458 PMID: 16368705 82. Schaeffer V., Lavenir I., Ozcelik S., Tolnay M., Winkler D.T. and Goedert M. (2012). "Stimulation of autophagy reduces neurodegeneration in a mouse model of human tauopathy." Brain 135(Pt 7): 2169– 2177. https://doi.org/10.1093/brain/aws143 PMID: 22689910 83. Kang S., Son S. M., Baik S. H., Yang J., Mook-Jung I., & Kim B. (2019). Autophagy-Mediated Secretory Pathway is Responsible for Both Normal and Pathological Tau in Neurons. Journal of Alzheimer’s Dis- ease, 70(3). https://doi.org/10.3233/JAD-190180 PMID: 31256134 84. Jackson G. R., Wiedau-Pazos M., Sang T.K., Wagle N., Brown C.A., Massachi S. et al. (2002). "Human wild-type tau interacts with wingless pathway components and produces neurofibrillary pathology in Drosophila." Neuron 34(4): 509–519. https://doi.org/10.1016/s0896-6273(02)00706-7 PMID: 12062036 85. Chatterjee S., Sang T.K., Lawless G.M. and Jackson G.R. (2009). "Dissociation of tau toxicity and phos- phorylation: role of GSK-3beta, MARK and Cdk5 in a Drosophila model." Hum Mol Genet 18(1): 164– 177. https://doi.org/10.1093/hmg/ddn326 PMID: 18930955 86. Shulman J. M., Imboywa S., Giagtzoglou N., Powers M.P., Hu Y., Devenport D., et al. (2014). "Func- tional screening in Drosophila identifies Alzheimer’s disease susceptibility genes and implicates Tau- mediated mechanisms." Hum Mol Genet 23(4): 870–877. https://doi.org/10.1093/hmg/ddt478 PMID: 24067533 87. Wang B. and Kundu M. (2017). "Canonical and noncanonical functions of ULK/Atg1." Curr Opin Cell Biol 45: 47–54. https://doi.org/10.1016/j.ceb.2017.02.011 PMID: 28292700 PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 23 / 24 PLOS ONE Age-related changes in tau and autophagy in human brain 88. Folwell J., Cowan C.M., Ubhi K.K., Shiabh H., Newman T.A., Shepherd D. et al. (2010). "Abeta exacer- bates the neuronal dysfunction caused by human tau expression in a Drosophila model of Alzheimer’s disease." Exp Neurol 223(2): 401–409. https://doi.org/10.1016/j.expneurol.2009.09.014 PMID: 19782075 89. McPhee C. K. and Baehrecke E.H. (2009). "Autophagy in Drosophila melanogaster." Biochim Biophys Acta 1793(9): 1452–1460. https://doi.org/10.1016/j.bbamcr.2009.02.009 PMID: 19264097 90. Caccamo A., Magri A., Medina D.X., Wisely E.V., Lopez-Aranda M.F., Silva A.J. et al. (2013). "mTOR regulates tau phosphorylation and degradation: implications for Alzheimer’s disease and other tauopa- thies." Aging Cell 12(3): 370–380. https://doi.org/10.1111/acel.12057 PMID: 23425014 91. Bjedov I., Toivonen J.M., Kerr F., Slack C., Jacobson J., Foley A. et al. (2010). "Mechanisms of life span extension by rapamycin in the fruit fly Drosophila melanogaster." Cell Metab 11(1): 35–46. https://doi. org/10.1016/j.cmet.2009.11.010 PMID: 20074526 92. Kolosova N. G., Vitovtov A.O., Muraleva N.A., Akulov A.E., Stefanova N.A. and Blagosklonny M.V. (2013). "Rapamycin suppresses brain aging in senescence-accelerated OXYS rats." Aging (Albany NY) 5(6): 474–484. https://doi.org/10.18632/aging.100573 PMID: 23817674 93. Shibuya Y., Niu Z., Bryleva E.Y., Harris B.T., Murphy S.R., Kheirollah A., et al. (2015). "Acyl-coenzyme A:cholesterol acyltransferase 1 blockage enhances autophagy in the neurons of triple transgenic Alz- heimer’s disease mouse and reduces human P301L-tau content at the presymptomatic stage." Neuro- biol Aging 36(7): 2248–2259. https://doi.org/10.1016/j.neurobiolaging.2015.04.002 PMID: 25930235 94. Wahl D., Solon-Biet S. M., Wang Q. P., Wali J. A., Pulpitel T., Clark X., et al. (2018). "Comparing the Effects of Low-Protein and High-Carbohydrate Diets and Caloric Restriction on Brain Aging in Mice." Cell Rep 25(8): 2234–2243 e2236. https://doi.org/10.1016/j.celrep.2018.10.070 PMID: 30463018 95. Wahl D., Coogan S. C., Solon-Biet S. M., de Cabo R., Haran J. B., Raubenheimer D., et al. (2017). "Cognitive and behavioral evaluation of nutritional interventions in rodent models of brain aging and dementia." Clin Interv Aging 12: 1419–1428. https://doi.org/10.2147/CIA.S145247 PMID: 28932108 96. Wahl D., Cogger V. C., Solon-Biet S. M., Waern R. V., Gokarn R., Pulpitel T., et al. (2016). "Nutritional strategies to optimise cognitive function in the aging brain." Ageing Res Rev 31: 80–92. https://doi.org/ 10.1016/j.arr.2016.06.006 PMID: 27355990 97. Wahl D., Solon-Biet S. M., Cogger V. C., Fontana L., Simpson S. J., Le Couteur, et al. (2019). "Aging, lifestyle and dementia." Neurobiol Dis 130: 104481. https://doi.org/10.1016/j.nbd.2019.104481 PMID: 31136814 PLOS ONE | https://doi.org/10.1371/journal.pone.0262792 January 26, 2023 24 / 24 PLOS ONE
10.1177_16094069231207011
Regular Article Researching Sensitive Topics With Children and Young People: Ethical Practice and Blurry Boundaries Katie Ellis1 , Kristine Hickle2, and Camille Warrington3 International Journal of Qualitative Methods Volume 22: 1–12 © The Author(s) 2023 DOI: 10.1177/16094069231207011 journals.sagepub.com/home/ijq Abstract Despite representing a vast and global concern, the narratives of children who experience child sexual exploitation (CSE) and access associated services are marginalised within research. As an outcome, relatively little is known about how children cope with the impact and consequences of their experiences. This paper draws together methodological insights from researchers reflecting upon three distinct pieces of qualitative fieldwork conducted with children and young people considered ‘vulnerable’ to, and ‘at risk of’, CSE. In doing so, we seek to recognise the challenges encountered when conducting research with vulnerable populations and explore the ‘blurry boundaries’ that researchers tread in order to balance competing power dynamics. This paper will discuss potential safeguarding concerns that arise when conducting sensitive research and will share our experiences of supporting young people to take part in research around child sexual exploitation. We will reflect upon the research process to highlight some of the strategies adopted to enable young people to engage in data collection safely. We consider the dynamic ethical practices that take place in the moment of research encounters, alongside the framework of procedural ethics, to conclude that both are fundamental to enable meaningful participation in research. Keywords ethical inquiry, focus groups, methods in qualitative inquiry, social, justice, mixed methods, ethnography Introduction Child sexual abuse is a term used to describe a range of sexual offenses perpetrated against minors, who are legally unable to consent (World Health Organization, 2003). While child sexual abuse represents a global concern, occurrences are often hidden and its prevalence is largely unknown (McClain & Amar, 2013). UK government statistics suggest that as many as 1 in 20 children experience some kind of sexual abuse between the ages of 11 and 17, equating to around 220,000 children nationally, at any one time (HM Government, 2017). Estimates in the United States range from 1 in 4 to 1 in 13 (Centres for Disease Control and Prevention, 2022; Finkelhor, et al., 2014) and the Council of Europe estimates as many as one in five children experience sexual abuse or violence (European Commission, 2020). Yet most cases of sexual abuse remain unreported (Silverio et al., 2021) and less than 10% of cases are reported to someone in authority (Radford et al., 2011). Despite representing a vast and global concern, relatively little is known about how victims of child sexual abuse cope with the impact and consequences of their ex- periences (Palmer & Foley, 2017) and young victim narratives often remain marginalised within research (McClain & Amar, 2013; Silverio et al., 2021). This paper draws together methodological insights from researchers reflecting upon three distinct pieces of qualitative fieldwork conducted with children and young people who had been victim of, or were considered to be ‘at risk of’, child sexual exploitation (CSE), a particular form of child sexual 1Faculty of Health, University of Sheffield, UK 2Department of Social Work and Social Care, University of Sussex, UK 3Institute Of Applied Social Research, University of Bedfordshire, UK Corresponding Author: Katie Ellis, University of Sheffield, The University of Sheffield, Barber House, Sheffield S10 2HQ, UK. Email: k.ellis@sheffield.ac.uk Creative Commons CC BY: This article is distributed under the terms of the Creative Commons Attribution 4.0 License (https:// creativecommons.org/licenses/by/4.0/) which permits any use, reproduction and distribution of the work without further permission provided the original work is attributed as specified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/ open-access-at-sage). 2 International Journal of Qualitative Methods abuse (Beckett & Walker, 2017). We will consider the po- tential challenges of including the perspectives of child sexual abuse victims in research and explore the institutional chal- lenges for researchers that are presented by focusing on this cohort of participants. In addition, we will contemplate the role of the researcher in minimising risk, whilst at the same time supporting potentially vulnerable participants to share their experiences and perspectives. This paper explores the realities of doing research with children and young people that upholds their best interests - while recognising that the concept of ‘best interests’ is itself not straightforward and that is sometimes used by professionals when overruling children’s own wishes (Daniel, 2010; James & James, 2004). We highlight the need to better recognise the dynamic and rela- tional practice of research and to foreground the ethical dilemmas that arise ‘in the moment’ but yet are rarely documented or accounted for in institutional ethical processes. In doing so, this paper considers the ethics involved in col- lecting informed consent, explores some of the methods that are appropriate to traverse unequal power relationships and contemplates the blurry boundaries that are walked by re- searchers who explore sensitive topics with children and young people. Research Ethics Academic research is strictly bound by ethical guidelines arising from the Nuremberg Code and the Helsinki Decla- ration (Guillemin & Gillam, 2004; Moriña, 2021). These guidelines continue to shape research practice with a con- tinuing legacy purporting that research should do ‘no harm’ (Dickson-Swift et al., 2009; Melrose, 2011). In this context, formal ethical guidelines and checklists are engineered to support researchers to plan and undertake research in line with the principles of their institution and the norms of the wider social research community. Guillemin and Gillam (2004) refer to these formal processes of seeking approval for research as ‘procedural ethics’. However, there are substantial differences between procedural ethics and the realities of behaving ethically in practice. While ‘checklists’ are helpful in en- couraging researchers to consider potential harms that might arise from the research and the minimum requirements of ethical research practice (Guillemin & Gillam, 2004), they are not exhaustive and cannot pre-empt every potential ethical dilemma that might occur (Taquette & Borges da Matta Souza, 2022). Our experience has shown, that in practice, unforeseen ethical issues emerge and present complex dilemmas, which researchers must react to in the moment and as the need arises (Graham et al., 2015). The ethics of research with vulnerable children Historically children have been excluded from research, with the views of professionals, practitioners and parents called to account for particular issues that may affect the everyday experiences of children (Christensen & James, 2017; Gilligan, 2016; Graham et al., 2015). The formulation of UNCRC (1989), and particularly Article 12, which details that children have the right to be heard about the issues that most affect them, paved the way for an increase in research around the experiences of children and young people. Subsequently a body of research has both aimed to further understanding about children’s everyday lives (Hilpp¨o et al., 2016) and been used to formulate more inclusive policy and practice (Brady et al., 2018; Ellis & Johnston, 2019; Mannay, 2015). The drive to include the perspectives of children and young people has been embraced by researchers globally, across disciplines, including health (Birch et al., 2007; Lees et al., 2017), social work (Ellis, 2018; Hickle, 2020; Moore et al., 2018; Warrington et al., 2016), youth justice (Phoenix & Kelly, 2013; Sharpe, 2012; Smithson et al., 2021), politics (Hadfield-Hill & Christensen, 2021) and education (Buchanan et al., 2022; Forde et al., 2018; Mayall, 2002). While these examples evidence the commitment of researchers to capture and share the views and experiences of children, it is also important to acknowledge the positionality of researchers within the wider structure of knowledge. As Beazley et al. (2009) explain, ‘the researcher is not the knower of truth, but rather the recorder and interpreter of multiple ‘other’ social subjectivities’ (p. 369). In this, we acknowledge that research is filtered and controlled by adults who take on the decision about whose voice to share and in which contexts (Faldet & Nes, 2021; James, 2007). While research frequently aims to be inclusive, access to vulnerable populations is often tightly controlled by gate- keepers (such as headteachers, clinicians, social work man- agers), who may decline research participation on behalf of others (Gilligan, 2016; Hasking et al., 2023). However, while purported as being a protective act, to ensure the ‘best in- terests’ of vulnerable populations, gatekeeping research ac- cess can enable the silencing of those who may go on to share something uncomfortable, from a service point of view. In- deed, the perspectives of those who are not shared are often those who could be considered to have the most to say, albeit perhaps providing difficult information to hear. The studies explored in this paper sought to highlight the perspectives of young people who have been noticeably ab- sent within public debates around CSE. Our work aimed to address this gap in knowledge by sharing the perspectives of children and young people who had experienced abuse. In doing so, we sought to engage in research processes that minimised the ‘filtering’ or controlling of the narratives of children and young people (James, 2007) with the aim of shifting the understanding of CSE so that future support might be better targeted to meet the needs of those most affected. Although the commitment to researching experiences of those classed as ‘extremely vulnerable’ is well established as valuable (Pearce, 2009), their inclusion in research still ca- talyses significant ethical tensions (Hackett, 2017). Central to Ellis et al. 3 these tensions is the fact that participants are approached and invited to participate in studies that cannot be assumed to directly (or obviously) benefit them (Gullimen, 2004), and that as research participants they are primarily serving the needs of researchers, funders and wider practice and policy audiences. This is coupled with one of the most pertinent risks when working with children affected by sexual abuse, that par- ticipating may re-surface traumatic memories that cause ad- ditional psychological harm. Though it is hoped that research will generate awareness and potentially improve services for similar and future children, it is true that the primary benefits of research will not usually affect research participants di- rectly, and that the time lag between research publication and impact is known to be protracted. Despite some exceptions (Bovarnick & Cody, 2021; Ellis, 2019; Warrington et al., 2016) these tensions go part way to explain the paucity of children and young people’s views in debates around child sexual abuse and exploitation (Gilligan, 2016). Yet, enabling the participation of young people in a research context offers an important and unique opportunity to ensure that their perspectives are recorded and heard, even when competing voices continue to dominate (Jessiman et al., 2017). While we assert the importance of listening to those who are considered vulnerable, we also acknowledge that the experiences and perspectives of children and young people impacted by trauma can make for uncomfortable listening. However, by not including their voices, and instead relying on professionals or practitioners to share their views instead, we risk further marginalising the first hand experiences of chil- dren and young people (Ellis, 2018, 2019; Jessiman et al., 2017). In addition, when received sensitively and empathet- ically, narrative retelling is noted to offer a potentially cathartic experience for victims of trauma (Kearney, 2007; McClain & Amar, 2013). Yet caution must be taken, for while retelling previous trauma may be cathartic, it is also emotionally de- manding for both parties (Silvario et al., 2021), and can at times mean that the researcher is balancing competing ethical priorities, which we will explore further in more detail. This paper will consider the tensions present in the context of each of our related fieldwork experiences to challenge the absence of methodological accounts which capture the lived realities of negotiating ethical research dilemmas with children and young people impacted by trauma. We seek to identify the approaches and decision-making we have em- ployed to address questions of inclusion, consent and the power dynamics that are especially apparent when con- ducting research with children who have experienced child sexual exploitation. We will share examples of practical methods adopted to navigate research in these contested ethical spaces, exploring both the feasibility and ethical incentive to support children to share their views on sensitive subjects safely. In considering the issues raised above, we will explore the ways in which researchers can carefully balance ethical considerations to demonstrate the importance of listening to children at the margins and to explore some of the potential ethical challenges that can be safely traversed along the way. The research The content of this paper is drawn from three separate qualitative studies conducted by the authors in distinct settings which provided support to young people who had been identified as at risk of, or the victim of, child sexual abuse and exploitation. While diverse in their focus, the studies all shared a commitment to foreground the per- spectives of young people considered to be marginalised. Children and young people engaged in each of the research settings were mostly known to social services, and because of their exploitation, were receiving interventions from multiple welfare professionals. Despite the commonalities between the young people receiving services, there were important differences between each of the research settings. For example, while the services referred to in projects two and three encouraged voluntary engagement from young people, young people’s service engagement in the setting of project one was mandatory. In each setting, the professional welfare contexts informed the tone and expected power dynamics for these young people when engaging with professionals. Specifically, the presence of ongoing ‘safe- guarding concerns’ meant participants in all three studies were required to tread careful boundaries with professionals and demonstrate heightened awareness about types of dis- closure or information that would trigger safeguarding responses. Additional project information is detailed below: Project one was conducted with girls in a local authority secure children’s home (LASCH) by Katie Ellis, who con- ducted ethnographic research over a period of 12 months. Ellis conducted three life history interviews with fifteen of the girls living in the home at that time. Participants were aged between 13-16 years old. Although young people’s perspectives were centred in the research, Ellis also gained consent to access individual case files so that she could explore the ways in which young people’s voices were described by professionals working with them. Further findings from this project can be found in Ellis, 2016, 2018; 2021. Project two is an ongoing project funded by the Arts and Humanities Research Council conducted by Kristine Hickle and Camille Warrington with girls and young women (age 13–25) accessing services provided for young people victi- mised by sexual abuse and sexual and criminal exploitation. The project involved the use of participatory photography and creative methods to understand experiences of individual and collective resistance to interpersonal violence and related forms of institutional harm. Participants were involved in a series of workshops in which they engaged in creative ac- tivities including image making and poetry. Ethnographic methods were also used, including participant observation. Hickle and Warrington gained consent to use some of the 4 International Journal of Qualitative Methods participants’ images and voice recordings alongside tran- scriptions of their discussions and individual interviews. Project three was a doctoral study conducted by Camille Warrington. The study involved in-depth qualitative inter- views with twenty children and young people (age 14–29) through seven different voluntary sector specialist sexual exploitation services across the UK, alongside ten practitioner interviews. Interviews with young people were undertaken both individually and in small groups and explored children and young people’s experiences of participation rights and involvement in decision making in welfare services after sexual exploitation. The project sought to understand the meaning of participation rights in the context of day-to-day service provision and support. Further details of the methods and findings can be found in Warrington (2013). Children’s researchers cite using a range of research methods to include children and young people in research (Greene & Hogan, 2005; Groundwater-Smith, et al., 2014; Mannay, 2015). Christensen and James (2017) assert that rather than being ‘child friendly’ researchers should be ‘participant friendly’ and mindful to create methods ‘appro- priate for the people involved in the study [….] and for the kinds of research questions that are being posed’ (p. 15). In this way, our studies sought to be ‘participant friendly’ and were designed with the intention of being sensitive and ap- propriate to all of our participants, regardless of their age. Each of the research sites presented above offered different specialist services to children and young people in need of support. Specialisms included: support for gang-associated young women; young people affected by exploitation; young people at risk of going missing; and young people who had been deprived of their liberty for their own protection. In choosing to undertake research in each of these sites, our work was borne from a shared commitment to the value and ethical necessity of including children and young people affected by sexual abuse and exploitation in research. Despite differences between the studies, this shared commitment required each of us to engage with risk in ways which we found little sufficient ethical guidance or procedures. Instead, across these studies we shared the experience of responding to unanticipated ethical challenges ‘in the moment’ and coming to understand research ethics as a dynamic and relational practice. In the section below, we explore a number of examples from this practice. These examples highlight issues around consent and privilege; the power dynamics that are omnipresent and in- formed by institutional contexts when conducting research with populations considered vulnerable; and the blurry boundaries that we, and other researchers, must carefully tread when working in contested spaces. Procedural Ethics and Informed Consent University governance procedures are fundamentally re- sponsible for protecting the reputations of research institu- tions, which may sometimes place them at odds with researchers seeking to empower participation from those whose voices may otherwise be marginalised or silenced (Whittington, 2019). Unsurprisingly, and due to the fact that the research focused on the sensitive topic of child sexual abuse, ethics procedures for each of the studies discussed here were lengthy and involved submitting examples of carefully crafted information sheets, consent forms and the details of additional support available for participants. The construction of these documents created practical tensions around meeting the requirements of an ethics committee whilst simultaneously being accessible to young people. For instance, although university systems require that strict protocols are followed, the protocols usually seen in these documents may not align with the needs of participants who are, 1) diverse in their literacy skills; 2) have prior exposure to multiple forms being handed to them by professionals and 3) are being asked to engage in participatory methods that evolve with the research. For example, some participatory methods intentionally allow space for participants to help shape the research design and make key decisions regarding modes of data production (Mannay, 2015). Participatory and other creative or group based research can therefore often necessitate a more fluid approach to research consent than traditional ethical approval processes were designed to accommodate and can instead require flexibility about the research methods being proposed, a variety of mechanisms to share information and an increased value on continuously negotiating informed consent. These tensions are explored further in an example from study two where a group of young people (age 16–25) initially refused to read and complete project paperwork. Following a verbal explanation that directly summarised the content of the information letter, all of the young people agreed to sign a consent form, apart from Jo. Despite being reluctant to sign an official form, Jo asked if she could continue to take part in the research workshop. She was welcomed into the group and engaged in all the research activities, with the understanding that her participation would not be recorded in any way. On the second day of activities, Jo attended and contributed to all research discussions with enthusiasm. Afterwards, the re- searcher thanked Jo for contributing powerful opinions and asked again whether she would consider signing the consent form so that her views could be shared more widely, outside of the project. Jo agreed that she would like others to understand her perspective and so agreed to take part in the research and formally completed the consent documents required for procedural ethics. The above example highlights what Whittington (2019) observed when she reflected on her own participatory research with young people exploring sexual consent. She noted im- portant parallels between how sexual consent was described as ‘fluid, constantly renegotiated, communicated, verbally and nonverbally, voluntary, mutual, and withdrawable’ (p. 205), and how these same principles could apply to participatory research. Whilst not all qualitative research with young people utilises a participatory design, it is worth considering how the Ellis et al. 5 concept of consent translates across methods and requires reflexivity of the researcher as an intentional ethical research practice. In our example, Jo subsequently became one of the most engaged research participants on the project, yet her participation was only obtainable by enabling consent to be renegotiated when she understood precisely what the research process would involve, how it would feel to take part, and believed the researchers were trustworthy. Furthermore, consent given by Jo in this moment could be considered to be more ‘informed’ than that consent given by others previously, since Jo gave consent when she had gained a clear under- standing of the research process and of the researcher’s in- tentions and not simply upon their first meeting. Conceptualising consent as a complex and moveable el- ement of research is a perspective that has been called for by others (e.g. Boddy & Oliver, 2010; David et al., 2001; de St Croix & Doherty, 2022; Sonne, et al., 2013; Whittington, 2019) who criticise an overreliance on traditional informed consent procedures. In these debates, it is argued that for- malised ethical procedures remain relatively static and diffi- to challenge, even when participants’ access and cult engagement with the research is possible only through em- bracing its complexities. In all three examples from research shared here, researchers felt compelled to rely on traditional methods of obtaining consent (i.e. written information sheet and a ‘tick box’ consent form) - which provided a degree of accountability, but also employed a broader conceptualisation of consent as dynamic and ongoing and a more flexible, re- lational approach to engaging with the form based procedures. David et al. (2001) previously questioned whether procedural ethics need to be revisited entirely. In their research on children in school settings, they noted that the mode of as- certaining consent aligned too closely with the kind of in- formation delivery that children and young people associated with education and school settings. DeSt de St Croix and Doherty (2022) found that young people in youth work set- tings felt similarly, as research processes that evoked a feeling of ‘school’ for young people were rejected, given the negative experiences in school settings that many participants had undergone. For young people in our research, it may have echoed both experiences within educational settings and the experience of contact with safeguarding professionals who have discussed informed consent and confidentiality with them, in the context of service provision. If these experiences have been negative, then employing a process that is con- sidered to be standard practice in relation to procedural ethics could inhibit young people from consenting to research they may actually want to engage in. Similarly while we recognise that recording consent represents an important aspect of ac- countability and governance to safeguard against abuse, it is rarely enough to ensure meaningful consent nor necessarily requires a form based method. In addition to ensuring the process of informed consent remains sufficiently fluid and flexible, our practice highlights issues related to what how researchers also encounter Guillemin and Gillam (2004) refer to as ‘ethics in practice’. This work queries whether there are resources to draw upon within the tradition of qualitative research for help in dealing with ethical aspects of the research or if these resources should be sought elsewhere; in research with young people, this may include multidisciplinary intervention that is ethically com- plex in relation to issues of both informed consent and (limited) confidentiality. For example, in a systematic review of professionals’ perspectives on informed consent and confidentiality in work with young people, Thannhauser et al. (2021) found that there is widespread inconsistency in ethical decision-making amongst practitioners, sometimes related to a fundamental difference in how the concepts of confidentiality and informed consent are understood. However, when re- searchers fail to ensure that informed consent is understood by young people, they risk compromising potentially trusting relationships when they encounter an ethical dilemma that they feel needs to be reported. Thereby, researchers must be mindful about the decisions that they make in the moment, to ensure that the wellbeing of participants is valued and pro- tected, and that participants are made aware of the types of information that cannot be held in confidence. We will discuss this issue, and the blurry boundary sometimes walked by researchers, later in this paper. Considering Power Dynamics within Research Power dynamics between an adult researcher and a child participant are always apparent, but they are especially pro- nounced in settings where young people are socially bound to follow the instructions of adults (Woodhead & Faulkner, 2008). This is true in most forms of research with children, but even more so when embarking on research with those who are considered vulnerable and may have previously experi- enced difficult or contentious relationships with professionals or practitioners (Ellis, 2016; Faldet & Nes, 2021). Support services, like the ones described in this paper, are often delivered within a socio-political context of continuous uncertainty, where funding is time-limited, and service evaluations are constrained by narrow understandings of ‘what works’ (Boddy, 2023). Yet, in research, the relationships we make with stakeholders are key, since they are our route to ensuring the successful engagement of children and young people. When services agree to host a researcher, they also engage in a transfer of trust in which they believe that the researcher will act in the best interests of participants. It is often this trust that reassures young people that it is safe to take part in research. Our partners also provide the resources and support that helps to assess risk and to provide tailored support during the research process, and after, if required. It is therefore important that we, as researchers, are able to maintain positive relationships within the research site. To do this, we balance competing challenges to maintain research access, often within narrowly defined research objectives, 6 International Journal of Qualitative Methods whilst simultaneously protecting young people’s active par- ticipation. While different actors within the research site may have competing interests, using a strengths-based approach, like appreciative inquiry (Liebling et al., 1999), can help to identify positive aspects of practice and can give confidence to those working within complex structures and strict hierar- chies. Strengths-based approaches can encourage researchers to recognise that professionals and practitioners are often working to achieve positive outcomes in particular contexts that are largely outside of their control. While working collaboratively with stakeholders is im- portant, maintaining distance and independence from them is key, and often vital to ensure the confidentiality of views shared by potentially vulnerable participants. There are often conflicting views about the potential of research and it is the role of the researcher to manage expectations about which information will be shared and with whom. For instance, we discovered that while managers may want research to vouch for the success of their service, practitioners sometimes want feedback around working styles and management practices, and service funders seek assurance that the programme is being delivered as originally intended. Yet the research may deliver none of these results, especially if they are in direct conflict with views put forward by young people who engage in services, sometimes against their will. Though promises about confidentiality are made to participants before they take part in research, there are examples where the actions of adult professionals who support children can challenge these agreements and relationships of trust. This can happen when adults ask researchers about responses given by in- dividual children (Christensen & Prout, 2002) and also where adult professionals casually volunteer personal in- formation about individual children to researchers without explicit consent to do so (Warrington, 2013). In maintaining confidentiality promised to children, it is crucial to explain very clearly when responses must be shared, for instance, if a participant discloses that abuse is being perpetrated against them or by them. In response to the power differences that are endemic when conducting research, we, like others (Moore et al., 2018) took extra care to remind participants of their rights at multiple points. As well as revisiting consent, to ensure that participants had not changed their mind about taking part, we also offered creative ways to participate in offering consent. Using more flexible research methods can empower more meaningful consent within the research process. For instance, in place of traditional interviews, question cards can be laid out for participants to see, read and potentially discard. Similarly, collaborative visual mapping with participants to outline key themes at the outset of an interview can help participants to prioritise what they would like to talk about and to point out anything they would like to avoid (Warrington et al., 2016). These methods can also facilitate discrete ways of opting out and participants can be encouraged to look at the cards or themes in turn and to disregard those that they would prefer not to discuss. As well as offering easy ways to opt out, and reminding participants that they can withdraw from research at any time (without sharing their reasons), researchers can also share phrases that participants can use if they are feeling uncom- fortable with a particular topic, such as ‘can we move on please?’. While some young people might be able to verbalise their reluctance to participate, in our research, we were es- pecially mindful of the non-verbal cues given by participants. We previously shared the example of Jo, who wanted to take part but did not want to sign a form; in study one, the re- searcher experienced the opposite when she became aware in the context of an interview, that Imogen, who signed her consent form willingly, did not want to participate in the research. Although verbally presenting as willing, Imogen began her interview looking at the floor with her arms and legs tightly crossed. Again, this highlighted the limits of procedural ethics for enabling meaningful consent. Instead the researcher had to respond in the moment, and to decide how best to ethically proceed. In this case, the researcher talked again about informed consent and offered the view that participants declining to take part was an equally valuable position to take, since non-participation demonstrated that the research was absolutely voluntary and conducted ethically. Imogen decided not to take part and was thanked for her honesty, meanwhile, the researcher reiterated the voluntary nature of the research to practitioners who had been active in recruiting young people. Conducting research with young people who have expe- rienced CSE can be difficult and potentially emotionally distressing for all parties. It is therefore important that re- searchers enter the field with relational skills which allow them to build rapport whilst being comfortable to sit with, and respond to, strong emotions. While research is not intended to be therapeutic, it can facilitate a feeling of connection and it is important that research does not contribute to deepening isolation, stigma or self-blame. Beazley et al. (2009, p. 374) caution researchers that to disregard children’s own percep- tions of self when describing their life circumstances is to ‘violate their dignity’. However, when working with young people who are considered extremely vulnerable, it might be necessary to occasionally (and very carefully) challenge their narratives. For instance, in study one, Robyn, talks about being punished ‘a lot’ at home because she ‘was a really naughty child’. Punishments cited ranged from being beaten, locked away, shouted at, and other abusive behaviours which finally triggered Robyn being taken into care. Although Robyn was nonchalant in her reporting of these incidents, admissions of being ‘naughty’ occurred frequently. Despite seeking to empower Robyn to tell her story, it became in- creasingly uncomfortable to hear a narrative of a preschool child deserving to be beaten because they were ‘naughty’. As this example shows, it is not always ethical (or possible) to listen without challenging young people’s accounts (albeit with care and sensitivity). In this scenario, the researcher engaged the participant in reflective listening, asking her to consider the ways in which very young children could be Ellis et al. 7 ‘really naughty’, exploring misdemeanours that young chil- dren cannot perform (e.g. armed robbery, fraud, vehicle theft, etc.). The exchange resulted in laughter, after which Robyn reconsidered her definition of ‘naughty’ and conceded that ‘maybe it wasn’t my fault’ that she was beaten and placed in care. This exchange illustrates an example of Guillemin and Gillam (2004) ‘ethically important moments’ and it was felt that by keeping silent, the researcher would unintentionally reinforce Robyn’s potentially harmful beliefs about her abuse. Such conversations occurred with regularity across all three studies and we all encountered instances in which we gently challenged young people’s narratives of both self and others. Here again evidence emerges of the limits of procedural ethics, which though able to help researchers to consider some of these instances, it remains the responsibility of the researcher to be prepared to act in the moment to make snap decisions about how to respond. Such challenges are rarely simple and so while silence, neutrality or passive listening can be con- sidered harmful, we equally need to heed caution from Rothman et al. (2018) who encourage researchers not to overstep the line, to resist taking on advocacy roles that were not intended for them and to ensure that they do not inad- vertently cause harm to the support already being given by targeted services. Behaving Ethically ‘in the Moment’ Across all these examples, and their recurring demonstrations of the limits of procedural ethics and the need to consider relational dynamics, a persistent theme emerges around the difficulty of holding and applying ‘absolute’ ethical rules in the moment. We recognise that the firm boundaries and static ‘rules’ of research engagement, outlined to meet thresholds of tolerance for an ethics reviewing committee, are often impossible to hold tightly to within our lived experience of research where individual needs and contexts matter. Fur- thermore our experiences suggest that ethical research practice demands flexibility and responsiveness to changing circum- stances, albeit within the safety of considered (and where possible shared) decision making. What this means however is that on entering a research interaction, there is no clear ‘blueprint’ that can sufficiently lay out ‘how to behave eth- ically’ in all scenarios. Subsequently in all of our research practices we experience a sense of ‘walking blurry bound- aries’ and sometimes having to live with a degree of uncer- tainty about the ‘right thing to do’. The realities of undertaking research in the contexts we describe means that there are often competing ethical con- siderations, like those described above – such as the moment where a researcher steps out of a boundaried researcher role to challenge young people’s narratives; or consent processes are reworked to respond to individual needs and forms of com- munication that are non-verbal. Further tensions arise in relation to individual versus group needs and promoting in- clusion versus minimising risk. An example of this was illustrated by a scenario in study three, which centres on a group interview where three young women, with previous experience of working together, who were invited (through their CSE service) to participate. Prior to undertaking the interview, the researcher was informed that one young person, Sally, was unlikely to engage due to ‘a chaotic life and some indifference to services’. At the pre-arranged time of the in- terview, Sally unexpectedly arrived and joined the other two with a friend (a non-service user) and her friend’s baby. She also announced on arrival that she had very little time and could ‘only stay for a bit’. In this initial moment, she was informed that her friend could not participate in the interview and would have to wait for her outside. Unfazed, Sally, who was sensed to hold considerable sway amongst her peers, asked the other two participants in somewhat loaded terms: ‘you don’t care if my friend stays here do you?’ to which they unsurprisingly replied ‘no’. Within this moment, it became clear to the researcher that insisting upon a formal application of the ethical framework (i.e. no non-research participants in a group interview setting) was likely to result in this young woman leaving with her instinctively this felt problematic. In this friend and that context, the researcher represented an unfamiliar visitor, on Sally’s territory, asking for her support but setting out new parameters about the terms on which she was welcome within her own project. In addition, the young woman’s very pres- ence demonstrated her commitment to participate in the research with no obvious benefits for herself. While on the one hand asking her to leave would allow adherence to ethical procedure, such a stance would also represent a rejection and an invalidation of her view, which had already been identified by the worker as hard to capture. The subsequent decision to allow Sally to stay was com- plex and was based on multiple pieces of contextual infor- mation. Key among these was the project workers’ surprise at her attendance and the fact that Sally was identified as rep- resenting an important perspective that might otherwise be difficult to access given her minimal contact with most ser- vices. Asking her to leave meant a risk of both silencing her important perspective and communicating a rejection of her offer to support the research. In addition, the researcher was aware of her announcement that she would shortly leave and her knowledge that all three invited participants were over 16 and had a long history of group work together. Subsequently Sally (with her friend and baby) were wel- comed and invited to stay. So began a dynamic process of trying to navigate consent and confidentiality in an unantic- ipated and challenging context, in which the researcher at- tempted to highlight the limits of confidentiality – hoping to encourage informed decision making from participants about what they did and didn’t share (Warrington, 2013). The re- searcher notes beginning with broad, impersonal questions which steered away from encouraging any personal revela- tion, whilst largely focusing questions on Sally, who spoke candidly and with passion about her experiences of services. 8 International Journal of Qualitative Methods Perhaps most significantly she spoke of her exclusion and marginalisation and the signifiers of power and inequality she saw around her. When asked about whether she felt listened to by professionals Sally laughed and explained her experience of attending review meetings: There’s me in me trackies and hoodie and there’s all them in their proper suits.. [laughter]… All they see is some common young child – that is it. Oh it’s another hoodie causes trouble… rarr rarr rarr. As these words highlight, formal processes of listening to ‘marginalised’ young people have the potential to exclude as well as empower. A clear parallel exists for our research where a desire to hold too tightly to prescribed processes, conceived and approved far from lived research, can work counter to intentions. Overly narrow ideas about young people’s ‘best interests’ easily overlook the wider (or less obvious) benefits of capturing marginalised perspectives and conversely the harms created through exclusion or ‘silencing’ those young people for whom our formal processes are neither meaningful or easy to engage with. Furthermore it can entrench the power dynamics between researcher and researched. Each of the studies presented in this paper shared a commitment to creating a space for research interactions that could be accessed by diverse young people identified as marginalised by other formal service and consultation pro- cesses. Within these spaces, we sought to enable young people to share their experiences, whilst being free from judgement and disapproval, with the view to enabling research en- gagement to be a route for individual advocacy or influencing policy change at a broader level. In order to achieve these aims there was a need to reframe what is meant by protection within the research space. Rather than protecting young people by avoiding the discussion of ‘sensitive issues’, we recognise the protective benefits to young people in being able to share their perspectives safely, whilst feeling valued and listened to by wider audiences. Research has shown that providing a safe and reflective space to discuss sensitive topics can be cathartic and go some way to reduce the stigma experienced by those who are victims of CSE (McClain & Amar, 2013). In support of this, young people in all three studies reflected on the benefits of engaging in a reflective research interview or the potential for young people’s workshop. Furthermore, views and experiences to influence longer-term policy and practice narratives was also recognised as potentially sup- portive of young people’s longer-term wellbeing. Discussion Research that seeks to gain the views of children who have experiences that are identified as ethically sensitive, has a tendency to be flagged as ‘risky’, and as a consequence diverse perspectives of those with lived experience are unintentionally marginalised (Silverio et al., 2021). In these scenarios, the potential for distress and re-traumatisation is often fore- grounded and yet in all of our experiences we acknowledge that alongside these risks (which are legitimate and real) we also need to recognise benefits. The act of research can provide a powerful moment of recognition and legitimation for those who have experienced harm and create opportunities for those who have been previously marginalised to contribute to generating change at both a practice and policy level. As Bovarnick & Cody (2021) note, risk in this field needs to be put into perspective. By working flexibly, creatively and in- clusively, it is possible for research to centre these margin- alised narratives and to ensure that knowledge is safely built on, and with, lived experience. In this way, research can play a role in contributing to tackling injustice; create opportunities to counter missing perspectives in public narratives; and thereby aid in democratising the production and dissemination of knowledge. In doing so, research makes inroads towards dismantling some of the patterns of power perpetuated by traditional patterns of knowledge production, which tends to elevate particular narratives while obscuring or missing out others. While we must be careful not to over claim the contribution of our research to redressing these relationships, how we undertake it and who we manage to include can challenge some of the injustices that are entrenched in knowledge generation, albeit in small ways. While we concur that the inclusion of previously excluded perspectives are fundamental in creating knowledge, it is important to acknowledge the process of collecting this in- formation can raise important ethical considerations. Along- side the importance of advance planning and ethical review processes, our experiences have taught us that researchers cannot anticipate every potential ethical scenario. Rather our collective research experience suggests a need to be prepared to engage with relational dynamics that make up the ‘ethically important moments’ discussed by Guillemin and Gillam (2004). While keen to share young people’s voices, and herald their views as representing first-hand experience of CSE and associated services, we must also consider our own positionalities within the wider structure of knowledge. As the ‘recorder’ of multiple truths (Beazley et al., 2009), it is not the role of the researcher to police information put forward by participants. This matter is made altogether more complex when conducting research with participants who have pre- vious experience of abuse, coercion and manipulation, as shared previously in relation to Robyn and her understanding of ‘naughty’. Although it is not the role of a researcher to contradict participants, researchers must be mindful that they are not complicit in reinforcing harmful narratives given to children by those who may have previously harmed them. Empathy is a much needed skill for qualitative researchers (Dickson-Swift et al., 2009), especially when the intention of the research is to engage participants in discussions that prompt a recollection of previous traumatic experiences, which potentially renders them vulnerable (Melrose, 2011). Of course the balance of power is complex in a research setting, Ellis et al. 9 and it is the role of the researcher to help participants feel comfortable and engaged in research spaces. As demonstrated in the sections above, power imbalances can present in dif- ferent ways, and it is important that research spaces are continually (re)negotiated, with the best interests of partici- pants at the centre of all research decisions. For these reasons, it is important that participatory research remains fluid and able to adapt to changing circumstances (Lenette, 2022). We draw upon research scenarios in which Imogen, Jo, Robyn and Sally each presented dilemmas that were negotiated by re- searchers in the moment. In these examples, we demonstrate that the wellbeing of participants was granted precedence over formal procedural ethics in a way that did not contravene the parameters set by the ethical approval granted for each study. Instead, we recognised that young people have different re- sponses and that consent has to be negotiated, individually and with sensitivity to the circumstances that young people are in at the time. A final ethical challenge for researchers, in relation to consent and confidentiality, is ensuring that young people’s understanding of these constructs remain ‘live’ throughout the project, so that as their experience of participating in the their understanding of these constructs research changes, move with them. In this way, participants continue to have control over their participation and the information they choose to share. We thereby suggest that researchers consider building in opportunities for young people’s choice and decision-making, even when research designs are not intended to be participatory. This may include intentionally applying for ethical approval in stages, so that feedback from young people regarding what feels most comfortable and accessible to them is considered as a project progresses. Ethical boundaries are important and the formal structures that set out guidance for ethical research practice are funda- mental in ensuring that there are recognised parameters on the limits of the topic being investigated; accountability of re- searchers; commitments to safety and best interests of par- ticipants and management of risk. However it is vital that such structures also leave space for situational idiosyncrasies and encourage the space for reflection; skills for relational prac- tice; and mechanisms in place to enable shared consideration to address complex issues during the research process. These reflective processes help to facilitate new ways of knowing, that can arise as the research progresses (de St Croix & Doherty, 2022; Lenette, 2022). Procedural ethics cannot capture all possible scenarios and the realities of ethical practice cannot always be anticipated, we therefore urge re- searchers to recognise the needs of different individuals, in different spaces and to ensure that institutional ethics do not override the needs of individual research participants. Our experiences have shown that it is the application and reflection of principles rather than fixed procedures that enable ethical research practice and consider the observation by McLeod (2007:285) that ‘a prerequisite for adults working with dis- affected youth is sensitivity towards issues of power’. In the field of sexual violence, it remains vital for researchers to take the previous experiences of trauma and the into account marginalisation of ‘vulnerable’ participants and to enable young people to shape the nature of their own engagement in research. Our research sought to platform the perspectives of those who have been previously marginalised and to thereby create a valuable and safe space for participants to share their expe- riences. As such, we recognise that while presented as a linear process, research can encompass a number of potentially ‘sticky’ situations that have to be considered ‘in the moment’. While important in securing research integrity, a commitment to working ethically can impact upon the timelines of projects, and thereby has implications for funders. Yet, researchers have a duty of care towards participants, and must strive to make research engagement safe and positive by creating the con- ditions to ensure that young people’s decisions are informed, as well as providing space and time for participants to reconsider their engagement. In this paper, we highlight the importance of building choice and control for participants into research processes, and in doing so, consider that by creating the space for participants to engage in research in meaningful ways, we must also hold and embrace the space for partici- pants to meaningfully withdraw from research. Our approach to ethics must then be responsive and reflective, to both facilitate engagement and empower young people to act self- protectively, in their own best interests. Conclusions The research included in this paper sought exclusively to collect the views of young people who had experienced sexual abuse and exploitation. As such, participants in all three studies were identified in a range of professional contexts as ‘vulnerable’, including both the organisations providing ser- vices to them and in the context of university ethical review processes. Accessing this group of participants is incredibly important and helps to ensure that their voices are not silenced in favour of those who claim to act in their ‘best interests’ or by the policy and practice that shapes their experiences of sup- port. Yet, research in this area is not straightforward and the need to balance unequal power dynamics to ensure that re- search creates a safe ethical space often requires careful consideration. It is therefore necessary to balance competing priorities in order to empower young people to set the scope of their participation. In this paper, we highlight some of the challenges researchers encounter and the blurry boundaries that must be negotiated in order to maintain positive relationships with research partners whilst ensuring friendly’. While ac- that knowledging that procedural ethics are important in ensuring that researchers carefully contemplate the potential risks associated with their research, they are not sufficient on their own, and it is thereby vital that researchers consider the wider implications and the need to behave ethically ‘in the moment’. research remains ‘participant that 10 International Journal of Qualitative Methods Safe research practice, particularly in relation to unantic- ipated or relational ethical issues, are enhanced by structures of peer review, reflective discussion and shared decision making. Relatedly the process of coming together to write this article has further highlighted to us the value of peer support and spaces for honest reflection about the realities of research practice in this field. We thereby encourage researchers working in sensitive spaces to support one another and to be reflective in their own practice so that research continues to be a safe space in which those who are perceived to be especially vulnerable can be supported to share their experiences. Declaration of Conflicting Interests The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article. Funding The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Economic and Social Research Council (ES/F021860/ 1), Arts and Humanities Research Council (AH/T003685/1). ORCID iD Katie Ellis  https://orcid.org/0000-0003-4185-1912 References Beazley, H., Bessell, S., Ennew, J., & Waterson, R. (2009). The right to be properly researched: Research with children in a messy, real world. Children’s Geographies, 7(4), 365–378. https://doi. org/10.1080/14733280903234428 Beckett, H., & Walker, J. (2017). Words matter: Reconceptualising the conceptualisation of child sexual exploitation. In H. Beckett, & J. Pearce (Eds), Understanding and responding to child sexual exploitation. Taylor and Francis. Birch, J., Curtis, P., & James, A. (2007). Sense and sensibilities: In search of the child-friendly hospital. Built Environment, 33(4), 405–416. https://doi.org/10.2148/benv.33.4.405 Boddy, J. (2023). Engaging with uncertainty: Studying child and family welfare in precarious times. Families, Relationships, Societies, 21(1), 127–141. https://doi.org/10.1332/ 204674321X16704251101652 Boddy, J., & Oliver, C. (2010). Research governance in children’s services: The scope for new advice”, research report No. DFE- RR072. Department for Education, HM Government. Bovarnick, S., & Cody, C. (2021). Putting risk into perspective: Lessons for children and youth services from a participatory advocacy project with survivors of sexual violence in Albania, Moldova and Serbia. Children and Youth Services Review, 126, 106003. https://doi.org/10.1016/j.childyouth.2021.106003 Brady, L., Hathway, F., & Roberts, E. (2018). A case study of children’s participation in health policy and practice. In Social policy first hand. Policy Press. Retrieved Dec 8, 2022, from. https://bristoluniversitypressdigital.com/view/book/ 97814473325 Buchanan, D., Hargreaves, E., & Quick, L. (2022). Schools closed during the pandemic: Revelations about the well-being of ‘lower-attaining’ primary-school children. Education, 3, 1–14. https://doi.org/10.1080/03004279.2022.2043405 Centres for Disease Control and Prevention (2022). Preventing child sexual abuse. Retrieved 19 June, 2022. https://www.cdc.gov/ violenceprevention/childsexualabuse/fastfact.html#print Christensen, P., & James, A. (2017). Research with children: Per- spectives and practices. Routledge. Christensen, P., & Prout, A. (2002). Working with ethical symmetry in social research with children. Childhood, 9(4), 477–497. https://doi.org/10.1177/0907568202009004007 Daniel, B. (2010). Concepts of adversity, risk, vulnerability and resilience: A discussion in the context of the ‘child protection system. Social Policy and Society, 9(2), 231–241. https://doi. org/10.1017/S1474746409990364 David, M., Edwards, R., & Alldred, P. (2001). Children and school- based research:‘informed consent ’or ‘educated consent. British Educational Research Journal, 27(3), 347–365. https://doi.org/ 10.1080/01411920120048340 de St Croix, T., & Doherty, L. (2022). Capturing the magic’: Grassroots perspectives on evaluating open youth work. Journal of Youth Studies, 1–17. Dickson-Swift, V., James, E. L., Kippen, S., & Liamputtong, P. (2009). Researching sensitive topics: Qualitative research as emotion work. Qualitative Research, 9(1), 61–79. https://doi. org/10.1177/1468794108098031 Ellis, K. (2016). ‘He’s got some nasty impression of me he has’: Listening to Children in the Secure Estate. British Journal of Social Work, 46(6), 1553–1567. https://doi.org/10.1093/bjsw/ bcv114 Ellis, K. (2018). Contested vulnerability: A case study of girls in secure care. Children and Youth Services Review, 88, 156–163. https://doi.org/10.1016/j.childyouth.2018.02.047 Ellis, K. (2019). Blame and culpability in children’s narratives of child sexual abuse. Child Abuse Review, 28(6), 405-417. https:// doi.org/10.1002/car.2590 Ellis, K., & Johnston, C. (2019). Pathways to university from care: Findings report one. The University of Sheffield. https://doi.org/ 10.15131/shef.data.9578930.v3 European Commission (2020). Communication from the commission to the European parliament, the Council, the European eco- nomic and social committee and the committee of the regions. https://home-affairs.ec.europa.eu/system/files/2020-07/ 20200724_com-2020-607-commission-communication_en.pdf Faldet, A., & Nes, K. (2021). Valuing vulnerable children’s voices in educational research. International Journal of Inclusive Edu- cation, 1–16. https://doi.org/10.1080/13603116.2021.1956602 Finkelhor, D., Shattuck, A., Turner, H. A., & Hamby, S. L. (2014). The lifetime prevalence of child sexual abuse and sexual as- sault assessed in late adolescence. Journal of Adolescent Health, 55(3), 329–333. https://doi.org/10.1016/j.jadohealth. 2013.12.026 Forde, C., Horgan, D., Martin, S., & Parkes, A. (2018). Learning from children’s voice in schools: Experiences from Ireland. Ellis et al. 11 Journal of Educational Change, 19(4), 489–509. https://doi.org/ 10.1007/s10833-018-9331-6 Gilligan, P. (2016). Turning it around: What do young women say helps them to move on from child sexual exploitation? Child Abuse Review, 25(2), 115–127. https://doi.org/10.1002/car.2373 Graham, A., Powell, M. A., & Taylor, N. (2015). Ethical research involving children: Encouraging reflexive engagement in re- search with children and young people. Children and Society, 29, 331–343. https://doi.org/10.1111/chso.12089 Greene, S., & Hogan, D. (Eds), (2005). Researching children0 s experience: Approaches and methods. Sage. Groundwater-Smith, S., Dockett, S., & Bottrell, D. (2014). Partic- ipatory research with children and young people. Sage. Guillemin, M., & Gillam, L. (2004). Ethics, reflexivity and “ethically important moments” in research. Qualitative Inquiry, 10(2), 261–280. https://doi.org/10.1177%2F1077800403262360 Hackett, S. (2017). Researching child sexual assault: Towards a child sensitive methodology. In M. Cowburn, L. Gelsthorpe, & A. Wahidin (Eds), Research ethics in criminology: Dilemmas, is- sues and solutions (pp. 133–149). Routledge. Hadfield-Hill, S., & Christensen, P. (2021). ‘I’m big, you’re small. I’m right, you’re wrong’: The multiple P/politics of ‘being young’ in new sustainable communities. Social and Cultural Geography, 22(6), 828–848. https://doi.org/10.1080/14649365. 2019.1645198 Hasking, P., Lewis, SP., Staniland, L., et al. (2023). Further silencing the voiceless: The role of gatekeepers in accessing information about self-injury. The Journal of Nervous and Mental Disease, 211(1), 5–10. https://doi.org/10.1097/NMD.0000000000001575 Hickle, K. (2020). Introducing a trauma-informed capability ap- proach in youth services. Children and Society, 34(6), 537–551. Hilpp¨o, J., Lipponen, L., Kumpulainen, K., & Virlander, M. (2016). Sense of agency and everyday life: Children;s perspective. Learning, Culture and Social Interaction, 10, 50–59. https://doi. org/10.1097/NMD.0000000000001575 Hm Government (2017) Tackling child sexual exploitation progress report. https://www.gov.uk/government/uploads/system/ uploads/attachment_data/file/592588/Tackling_Child_Sexual_ Exploitation_-_Progress_Report__web_.pdf James, A. (2007). Giving voice to children’s voices: Practices and problems, pitfalls and potentials. American Anthropologist, 109(2), 261–272. https://doi.org/10.1525/aa.2007.109.2.261 James, A., James (2004). Constructing childhood. Theory, policy and Lenette, C. (2022). Participatory action research: Ethics and de- colonization. Oxford University Press. Liebling, A., Price, D., Elliott, C., et al. (1999). Appreciative inquiry and relationships in prison. Punishment & Society, 1(1), 71. https://doi.org/10.1177/14624749922227711 Mannay, D. (2015). Visual, narrative and creative research methods: Application, reflection and ethics. Routledge. Mayall, B. (2002). Children in action at home and school Children’s childhoods (pp. 120–133). Routledge. McClain, N., & Amar, A. F. (2013). Female survivors of child sexual abuse: Finding voice through research participation. Issues in Mental Health Nursing, 34(7), 482–487. https://doi.org/10. 3109/01612840.2013.773110 Melrose, M. (2011). Regulating social research: Exploring the im- plications of extending ethical review procedures in social re- search. Sociological Research Online, 16(2), 49–58. https://doi. org/10.5153/sro.2377 Moore, T. P., McArthur, M., & Noble-Carr, D. (2018). More a marathon than a hurdle: Towards children’s informed consent in a study on safety. Qualitative Research, 18(1), 88–107. https:// doi.org/10.1177/1468794117700708 Moriña, A. (2021). When people matter: The ethics of qualitative research in the health and social sciences. Health and Social Care in the Community, 29(5), 1559–1565. https://doi.org/10. 1111/hsc.13221 Palmer, E., & Foley, M. (2017). ‘I Have My Life Back’: Recovering from Child Sexual Exploitation. British journal of social work, 47(4), 1094-1110. Pearce, J. (2009). Young people and sexual exploitation: ‘It’s not hidden, you just aren’t looking’ routledge. Routledge. Phoenix, J., & Kelly, L. (2013). ‘You have to do it for yourself’: Responsibilization in youth justice and young people’s situated knowledge of youth justice practice. British Journal of Crim- inology, 53(3), 419–437. https://doi.org/10.1093/bjc/azs078 Radford, L., Corral, S., Bradley, C., Fisher, H., Bassett, C., Howat, N., & Collishaw, S. (2011). Child abuse and neglect in the UK today. NSPCC. Rothman, E. F., Farrell, A., Bright, K., & Paruk, J. (2018). Ethical and practical considerations for collecting research-related data from commercially sexually exploited children. Behavioral Medicine, 44(3), 250–258. https://doi.org/10.1080/08964289.2018. 1432550 Sharpe, G. (2012). Offending girls: Young women and youth justice. social practice. Palgrave. Routledge. Jessiman, P., Hackett, S., & Carpenter, J. (2017). Children’s and carers’ perspectives of a therapeutic intervention for children affected by sexual abuse. Child and Family Social Work, 22(2), 1024–1033. https://doi.org/10.1111/cfs.12322 Kearney, R. (2007). Narrating pain: The power of catharsis. Para- graph, 30(1), 51–66. Lees, A., Payler, J., Ballinger, C., Lawrence, P., Faust, S. N., & Meads, G. (2017). Positioning children’s voice in clinical trials research: A new model for planning, collaboration, and re- flection. Qualitative Health Research, 27(14), 2162–2176. https://doi.org/10.1177/1049732317726760 Silverio, S. A., Bewley, S., Montgomery, E., Roberts, C., Richens, T., Maxted, F., Sandall, J., & Montgomery, J. (2021). Disclosure of non-recent (historic) childhood sexual abuse: What should re- searchers do? Journal of Medical Ethics, 47(12), 779–783. https://doi.org/10.1136/medethics-2020-106343 Smithson, H., Gray, P., & Jones, A. (2021). ‘They really should start listening to you’: The benefits and challenges of Co-producing a participatory framework of youth justice practice. Youth Justice, 21(3), 321–337. https://doi.org/10.1177/1473225420941598 Sonne, S. C., Andrews, J. O., Gentilin, S. M., Oppenheimer, S., Obeid, J., Brady, K., Wolf, S., Davis, R., & Magruder, K. (2013). 12 International Journal of Qualitative Methods Development and pilot testing of a video-assisted informed consent process. Contemporary Clinical Trials, 36(1), 25–31. https://doi.org/10.1016/j.cct.2013.05.011 Taquette, S. R., & Borges da Matta Souza, L. M. (2022). Ethical dilemmas in qualitative research: A critical literature review. International Journal of Qualitative Methods, 22, 1–15. https:// doi.org/10.1177/16094069221078731 Thannhauser, R. E., Morris, Z. A., & Gamble, N. (2021). Informed consent, confidentiality, and practitioner disclosure in thera- peutic work with youth: A systematic review of practitioners’ perspectives. Adolescent Research Review, 7, 1–28. United Nations Convention on the Rights of the Child (1989). United Nations Treaty (vol. 1577, pp. 3–178). https://treaties.un.org/ doc/Treaties/1990/09/19900902-03-14-AM/Ch_IV_11p.pdf Warrington, C. (2013). Helping me find my own way’: sexually in decision-making exploited young people’s involvement their care. Doctoral about fordshire. Luton. thesis. University of Bed- Warrington, C., Ackerley, E., Beckett, H., Walker, M., & Allnock, D. (2016). Making Noise: Children’s voices for positive change after sexual abuse’. University of Bedfordshire/ Children’s Commissioner. Whittington, E. (2019). Co-producing and navigating consent in participatory research with young people. Journal of Children’s Services, 14(3), 205–216. Woodhead, M., & Faulkner, D. (2008) (In press). Subjects, objects or participants? Dilemmas of psychological research with children. In P Christiansen & A James (Eds.), Research With Children: Perspectives and Practices (pp. 10–39). Falmer Press / Routledge. World Health Organization. (2003). Guidelines for medico-legal care of victims of sexual violence.
10.7554_elife.85814
TOOLS AND RESOURCES Continuous muscle, glial, epithelial, neuronal, and hemocyte cell lines for Drosophila research Nikki Coleman- Gosser1†, Yanhui Hu2†, Shiva Raghuvanshi1†, Shane Stitzinger1†, Weihang Chen2, Arthur Luhur3, Daniel Mariyappa3, Molly Josifov1, Andrew Zelhof3, Stephanie E Mohr2, Norbert Perrimon2,4*, Amanda Simcox1,5* 1Department of Molecular Genetics, Ohio State University, Columbus, United States; 2Drosophila RNAi Screening Center and Department of Genetics, Harvard Medical School, Boston, United States; 3Drosophila Genomics Resource Center and Department of Biology, Indiana University, Bloomington, United States; 4Howard Hughes Medical Institute, Chevy Chase, United States; 5National Science Foundation, Alexandria, United States Abstract Expression of activated Ras, RasV12, provides Drosophila cultured cells with a prolifer- ation and survival advantage that simplifies the generation of continuous cell lines. Here, we used lineage- restricted RasV12 expression to generate continuous cell lines of muscle, glial, and epithe- lial cell type. Additionally, cell lines with neuronal and hemocyte characteristics were isolated by cloning from cell cultures established with broad RasV12 expression. Differentiation with the hormone ecdysone caused maturation of cells from mesoderm lines into active muscle tissue and enhanced dendritic features in neuronal- like lines. Transcriptome analysis showed expression of key cell- type- specific genes and the expected alignment with single- cell sequencing and in situ data. Overall, the technique has produced in vitro cell models with characteristics of glia, epithelium, muscle, nerve, and hemocyte. The cells and associated data are available from the Drosophila Genomic Resource Center. Editor's evaluation This valuable work describes the establishment and characterization of new cell lines derived from specific tissues of the fruit fly Drosophila. The evidence supporting the claims of the authors is convincing, with rigorous characterization of the cell lines and incorporation of their transcriptomes into Drosophila Gene Expression Tool website for user- friendly access. These lines will be a valuable resource that complements in vivo Drosophila genetics, improving biochemistry and facilitating high- throughput screening. Introduction The use of cell cultures has been important for studying biological processes that are not easily acces- sible in whole organisms (Klein et  al., 2022). A number of advances in mammalian cell cultures, for instance, development of 3D/organoid cultures (Rossi et  al., 2018), improved genome editing tools to manipulate induced pluripotent stem cells (Hockemeyer and Jaenisch, 2016), and better optimized media formulations for recombinant protein expression Ritacco et al., 2018 have further enhanced the utility of mammalian cell culture systems. These advances are accompanied by the avail- ability of several distinct mammalian cell lines derived from different tissue types. Similarly, the use of *For correspondence: perrimon@genetics.med. harvard.edu (NP); simcox.1@osu.edu (AS) †These authors contributed equally to this work Competing interest: The authors declare that no competing interests exist. Funding: See page 20 Received: 28 December 2022 Preprinted: 19 January 2023 Accepted: 12 July 2023 Published: 20 July 2023 Reviewing Editor: Erika A Bach, New York University School of Medicine, United States This is an open- access article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose. The work is made available under the Creative Commons CC0 public domain dedication. Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 1 of 27 Tools and resources eLife digest Fruit flies are widely used in the life and biomedical sciences as models of animal biology. They are small in size and easy to care for in a laboratory, making them ideal for studying how the body works. There are, however, some experiments that are difficult to perform on whole flies and it would be advantageous to use populations of fruit fly cells grown in the laboratory – known as cell cultures – instead. Unlike studies in humans and other mammals, which – for ethical and practical reasons –heavily rely on cell cultures, few studies have used fruit fly cell cultures. Recent work has shown that having an always active version of a gene called Ras in fruit fly cells helps the cells to survive and grow in cultures, making it simpler to generate new fruit fly cell lines compared with traditional methods. However, the methods used to express activated Ras result in cell lines that can be a mixture of many different types of cell, which limits how useful they are for research. Here, Coleman- Gosser, Hu, Raghuvanshi, Stitzinger et al. aimed to use Ras to generate a collection of cell lines from specific types of fruit fly cells in the muscle, nervous system, blood and other parts of the body. The experiments show that selectively expressing activated Ras in an individual type of cell enables them to outcompete other cells in culture to generate a cell line consisting only of the cell type of interest. The new cell lines offer models for experiments that more closely reflect their counterparts in flies. For example, the team were able to recapitulate how fly muscles develop by treating one of the cell lines with a hormone called ecdysone, which triggered the cells to mature into active muscle cells that spontaneously contract and relax. In the future, the new cell lines could be used for various experiments including high throughput genetic screening or testing the effects of new drugs and other compounds. The method used in this work may also be used by other researchers to generate more fruit fly cell lines. insect cell lines also complements whole organismal studies and helped to illuminate many aspects of insect cell biology (Luhur et al., 2019) including development (Sato and Siomi, 2020), immunity (Goodman et al., 2021; Chen et al., 2021), host–pathogen relationships (Smagghe et al., 2009), in addition to biotechnological applications (Hong et al., 2022). Fruit fly (Drosophila melanogaster) cell cultures are among the most widely used invertebrate cell cultures (Luhur et al., 2019). Drosophila cell lines are relatively homogenous, and highly scalable for both biochemical and high- throughput functional genomic analyses (Debec et al., 2016, Baum and Cherbas, 2008; Zirin et al., 2022; Mohr, 2014; Viswanatha et al., 2019). These features underlie their status as an important workhorse for scientific discovery in organismal development and as models for human disease. There are approximately 250 distinct Drosophila cell lines housed by the Drosophila Genomics Resource Center (DGRC) (Luhur et al., 2019). The majority of these cell lines, initially established by independent laboratories worldwide, were donated to the DGRC. A subset of 25 of these lines was subjected to transcriptome analysis, with the results demonstrating that approximately half of the transcripts expressed by each of these lines were unique such that even cell lines derived from the same tissue had distinct transcriptomic profiles (Cherbas et  al., 2011). Furthermore, the transcriptional profiles of several imaginal disc lines analyzed were found to match profiles of cells from distinct spatial locations in the respective discs (Cherbas et al., 2011). All lines exhibited transcript profiles indicative of cell growth and cell division, and not cellular differentiation, as expected for proliferating cells (Cherbas et al., 2011). Thus, the transcriptional profiles of several Drosophila cell lines provided a platform for subsequent analyses. For instance, a few examples of the impact of this work include research into better understanding crosstalk between signaling pathways (Ammeux et al., 2016), exploring transcription factor networks (Rhee et al., 2014), establishing small RNA diversity (Wen et al., 2014), characterizing signaling pathways (Neal et al., 2019), nucleosomal organization (Martin et al., 2017) among multiple other utilities reviewed extensively (Cherbas and Gong, 2014; Luhur et al., 2019). Over two- thirds of the D. melanogaster cell lines listed in the DGRC catalog were derived from whole embryos and the remainder are from various larval imaginal discs, the larval central nervous system, larval hemocytes, or adult ovaries. The potential of cells from these different sources to Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 2 of 27 Cell Biology | Developmental Biology Tools and resources Table 1. Cell lines analyzed. Tissue- type alignment Genotype Glial Repo- Gal4; RasV12; bratdsRNA Lines analyzed* DGRC stock name and number RRID Rbr6 (parental) Rbr6- 2 Rbr6- 4 Rbr6- F9 repo>Ras bratdsRNA- L6, 282 repo>Ras bratdsRNA- L6- Clone2, 326 repo>Ras bratdsRNA- L6- Clone4, 327 repo>Ras bratdsRNA- L6- CloneF9, 328 Epithelial btl- Gal4; UAS- P35; UAS- RasV12 Btl3 (parental) btl>Ras attP- L3, 332 btl- Gal4; UAS- P35; attP, UAS- RasV12 Btl7 (parental) Btl8 (parental) btl>Ras attP- L7, 285 btl>Ras attP- L8, 286 Muscle 24B- Gal4; attP, UAS- RasV12 24B- Gal4; UAS- GFP; attP, UAS- RasV12 24B5 (parental) 24B5- B8 24B5- D8 24B>Ras attP- L5, 284 24B>Ras attP- L5- CloneB8, 323 24BG1 (parental) 24BG1- F3† 24BG1- G1† 24B>Ras attP GFP- L1, 283 24B>Ras attP- G1- CloneF3, 325 24B>Ras attP- G1- CloneG1, 324 RRID:CVCL_XF57 RRID:CVCL_C7G9 RRID:CVCL_C7GA RRID:CVCL_C7GB RRID:CVCL_B3N7 RRID:CVCL_XF53 RRID:CVCL_XF54 RRID:CVCL_XF52 RRID:CVCL_C7G6 RRID:CVCL_XF51 RRID:CVCL_C7G8 RRID:CVCL_C7G7 Neuronal Act5C- GeneSwitch- Gal4; UAS- GFP; attP, UAS- RasV12 ActGSB- 6‡ ActGSI- 2 Act5C- GS>Ras attP- LB- Clone6, 329 Act5C- GS>Ras attP- GFP- LI- Clone2, 330 RRID:CVCL_C7GC RRID:CVCL_C7GD Blood Act5C- GeneSwitch- Gal4; UAS- GFP; attP, UAS- RasV12 ActGSI- 3 Act5C- GS>Ras attP- GFP- LI- Clone3, 331 RRID:CVCL_C7GE *Clones unless indicated. †Do not differentiate into active muscle. ‡These cells do not express GFP, the reason for this is not known. differentiate into adult cell types is not known. However, temporal transcriptional profiling of the Ecdysone response of 41 cell lines (Stoiber et al., 2016) provided evidence that cell lines exhibited varying levels of ecdysone sensitivity and potential for cellular differentiation, suggesting the possi- bility of developing cell- type- specific cell lines with the capacity to differentiate. As well as having unknown cellular origins, most Drosophila cell lines arose spontaneously, and the time needed to develop a continuous cell line was often protracted. In contrast, expression of activated Ras, RasV12, using the Gal4- UAS system, resulted in the rapid and reproducible generation of continuous cell lines from primary embryonic cultures (Simcox et al., 2008b). The Ras method was used to develop an array of mutant cell lines by using appropriate genotypes to establish the primary cultures (Simcox et al., 2008a, Lee et al., 2015; Kahn et al., 2014; Lim et al., 2016; Nakato et al., 2019). To date all lines have been generated using ubiquitous expression of UAS- Ras with Act5C- Gal4 and therefore the cell type in a given line is unknown. Here, we describe a second- generation version of the Ras method in which RasV12 expression is restricted to a lineage by using tissue- specific Gal4 drivers. This genetic ‘dissection’ provides only the targeted cells with the survival and proliferation advantage conferred by RasV12 expression (Simcox et al., 2008b). As we show, the approach has been successful and resulted in the generation of cell lines with glial, epithelial, and muscle characteristics. Lines generated by broad RasV12 expression should also include those of specific cell types and by using single- cell cloning and cell type char- acterization (marker gene expression and RNAseq) we identified lines with neuronal and hemocyte characteristics. Collectively, these cell lines provide in vitro models for five different cell types and are expected to be a valuable resource for high- throughput and biochemical approaches, which rely on large numbers of homogeneous cells. Results Primary cultures were established from embryos in which UAS- RasV12 expression was restricted to glial, tracheal epithelial, and mesodermal cells using lineage- specific Gal4 drivers (Table 1, Supplementary file 1). A subset of continuous cell lines derived from each type of primary culture was analyzed with regard to cell morphology, the presence of proteins characteristic of specific cell types, and other attributes (Table 1, Supplementary file 1, Supplementary file 2; Figure 1). We also analyzed lines with neuronal- or hemocyte- like characteristics that were cloned from parental lines resulting from ubiquitous expression of UAS- RasV12 (Table 1, Supplementary file 1, Supplementary file 2; Figure 1). Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 3 of 27 Cell Biology | Developmental Biology Tools and resources Figure 1. Morphology of cells. (A–C) Glial- lineage clones. The cells have an elongated morphology with variable lengths from approximately 20 to >50 µm (red arrowheads). (D–F) Tracheal- lineage cells. Btl3 and Btl7 cells form squamous epithelial sheets. Btl8 are closely associated but do not abut each other to form a sheet. (G–J) Mesodermal- lineage cells. The cells have a bipolar morphology. Multinucleate cells are frequently found in 24BGI- F3 and 24BG1- GI clones (red arrowheads). (K, L) Neuronal- like clones. ActGSB- 6 cells are mainly bipolar; however, some have asymmetric processes or thin Figure 1 continued on next page Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 4 of 27 Cell Biology | Developmental Biology Tools and resources Figure 1 continued processes (red arrowheads). ActGSI- 2 are bipolar. (M) Hemocyte- like clone ActGSI- 3. The cells form floating clusters that increase in cell number as they proliferate. Individual cells have a round morphology. (N) Schneider’s S2 cells. The cells are thought to be of hemocyte type and grow as single round cells in suspension. Scale bar = 10 µm. We further analyzed the cell lines by RNAseq to determine the transcriptome and signaling path- ways (Figure 2 and Figure 2—figure supplements 1–3). The gene expression values (Fragments Per Kilobase per Million mapped fragments, FKPM) are provided in Supplementary file 3. The dataset (Ras cell lines) has been imported into the Drosophila Gene Expression Tool (DGET) database (https:// www.flyrnai.org/tools/dget/web/), which is the bulk RNAseq data portal at Drosophila RNAi Screening Center (DRSC) (Hu et al., 2017). The TM4 package was used for making the plot in Figure 2 (Wang et al., 2017). As expected, the transcriptomes of the new cell lines are distinct from those of existing cell lines (Cherbas et al., 2011; Figure 2—figure supplement 1) and new cell lines derived from the same Gal4 driver cluster with one another (Figure 2—figure supplement 2). Moreover, comparison of differentially expressed (DE) genes with RNAseq data from single- cell RNAseq data (Li et al., 2022; Table  2) or with known cell type- associated transcription factors (Figure  2—figure supplement 3) reveals that these cells express genes characteristic of specific cell types. The results of our detailed characterization are described according to cell type in the sections below. Glial-lineage cell lines Repo is expressed exclusively in glial cells (Xiong et al., 1994). A repo- Gal4 driver that recapitulates Repo expression was used to express UAS- RasV12 (Ogienko et  al., 2020; Sepp et  al., 2001). This led to robust production of primary cultures however these failed to survive beyond early passages (Supplementary file 1). To counter potential cell death or modulate growth signaling, additional genotypes were tested including co- expression of UAS- transgenes encoding the P35 baculovirus cell survival factor, dsRNAs targeting tumor suppressors, or the Gal4 inhibitor Gal80ts (Supplementary file 1). Co- expression of a UAS- bratdsRNA or expression of tub- Gal80ts each produced a single line of cells that could be propagated for extended passages however the latter line was difficult to maintain and eventually lost (Supplementary file 1). The repo- Gal4: UAS- bratdsRNA; UAS- RasV12 (Rbr6) line has been passaged more than 50 times. The parental Rbr6 line and three clonal derivatives (Rbr6- 2, Rbr6- 4, and Rbr6- F9) have an elongated morphology and stained positive for Repo (Table 1; Figures 1 and 3, and Figure 3—figure supplement 1). A few cells expressed neuronal markers (Figure 3—figure supplement 1; Supplementary file 2). To induce differentiation, we gave cells two 24 hr ecdysone treatments separated by 24 hr to approximate the pulses of ecdysone during the larval to pupal tran- sition. Cells from each of the clones survived treatment with ecdysone suggesting they are of adult type, two clones showed morphological changes and formed a network, and all continued to express Repo (Figure 3 and Figure 3—figure supplement 2). The results of RNAseq analysis revealed that the three Rbr6 clones have very similar expression patterns (Figure  2—figure supplement 2). In addition, their DE gene signatures are also a close match to gene signatures of glial cells as identified by single- cell RNAseq (Table  2) and to glial- associated genes reported in the literature. For example, zydeco (zyd), which encodes a potassium- dependent sodium/calcium exchanger, is upregulated in all three clones, consistent with the literature (Zwarts et al., 2015; Featherstone, 2011), and gcm2, a transcription factor, is upregulated in two clones (Figure 2—figure supplement 3). These data suggest the Rbr6 clones will be a useful in vitro source of glial cells. Tracheal epithelium-lineage cell lines Breathless is expressed in the tracheal epithelium and a btl- Gal4 driver was used to express UAS- RasV12 (Shiga et al., 1996). Patches of cells with epithelial morphology proliferated in primary cultures and several continuous lines were generated (Table  1, Supplementary file 1). We were unable to derive clones of these using dilution or selection methods, which were successful for other cell types. Correspondingly, three parental lines were examined: Btl3, Btl7, and Btl8 (Table  1). All showed expression of the epithelial marker Shotgun/E- Cadherin (Shg/Ecad) and two grew in a squamous epithelial sheet with Ecad expression at the cell periphery (Figures 1 and 4, and Figure 4—figure supplement 1). In comparison S2 did not show peripheral expression of Ecad (Figure 4). Treatment of Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 5 of 27 Cell Biology | Developmental Biology Tools and resources Figure 2 continued Figure supplement 2. Principal component analysis (PCA) of RNAseq data from the lineage- restricted Ras cell lines. Figure supplement 3. Relative expression of transcription factors associated in the literature with specific tissue lineages in the lineage- restricted Ras cell lines. the squamous epithelial cells (Btl3 and Btl7) with ecdysone caused aggregation and formation of large multicellular clusters (Figure  4, Figure  4— figure supplement 2). RNAseq data analysis comparing the top upregulated genes in the Btl cell lines with scRNAseq datasets revealed that the lines closely match the signatures of the adult trachea, a network of epithelial tubules (Table  2) and Btl3 expresses trachealess (trh) a master regulator of tracheal identity (Wilk et  al., 1996; Figure  2— figure supplement 3). Overall, the morpholog- ical and molecular characteristics of the lines are consistent with an epithelial cell type of tracheal origin. Mesodermal-lineage cell lines The 24B- Gal4 driver is an insertion in held out wings (how) and is expressed in mesoderm and muscle cells (Brand and Perrimon, 1993; Zaffran et al., 1997). Expression of UAS- RasV12 with 24B- Gal4 readily produced continuous lines (Table 1, Supplementary file 1). Four clones (24B5- B8, 24B5- D8, 24BG1- F3, and 24BG1- G1) derived from two parental lines (24B5 and 25BG1) were analyzed in more detail (Table 1). The cells had a bipolar shape and expressed mesoderm markers including Twist and Mef2 (Figures  1 and 5, and Figure  5—figure supplement 1). When treated with ecdysone, cells from both parental lines and clones 24B5- B8 and 24B5- D8 elongated, fused as indicated by multinucleate cells, formed a network, and expressed Myosin heavy chain (Mhc) (Figure  5 and Figure  5—figure supplement 2). There was also extensive cell lysis. Beginning 2 days after the second ecdysone treatment, the cells began to contract spontaneously. Contrac- tion of cells from the 24B5 parental line and the two derivative clones (24B5- B8 and 24B5- D8) was visible in real time (Videos 1 and 2), whereas contraction of parental line 24BG1 cells was much slower and visualized more clearly in time- lapse (Videos  3 and 4). The clones 24BG1- F3 and 24BG1- G1 underwent morphological change but did not express Mhc or contract (Figure 5— figure supplements 2 and 3). In later passages, Figure 2. Expression levels of ligands and receptors for major signaling pathways. The ligand and receptor annotation for major signaling pathways was obtained from FlyPhoneDB (https://www.flyrnai.org/tools/ fly_phone/web/). The expression levels of ligands and receptors are represented as a heatmap of FPKM values. The online version of this article includes the following figure supplement(s) for figure 2: Figure supplement 1. Comparison of lineage- restricted Ras cell lines with previously isolated Drosophila cell lines. Figure 2 continued on next page Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 6 of 27 Cell Biology | Developmental Biology Tools and resources Table 2. RNAseq data analysis. Tissue type Glial Cell line Cell cluster scRNAseq Enrichment p value scRNAseq scRNAseq dataset Cell type based on in situ data Enrichment p value in situ Rbr6- 2 Rbr6- 4 Adult reticular neuropil- associated glial cell 8.13E−05 Whole body Glia 4.84E−05 Cell body glial cell 7.56E−04 Whole body Rbr6- F9 Adult glial cell 8.13E−05 Whole body Glia 3.42E−02 Epithelial Btl3 Adult tracheal cell 2.61E−06 Whole body Tracheal 1.08E−01 Btl7 Btl8 Adult tracheal cell 8.81E−04 Oenocyte Adult tracheal cell 2.72E−02 Body Tracheal 2.05E−02 Muscle 24B5- B8 Muscle cell 2.93E−6 Male reprod glands 24BG1- F3 Muscle cell 1.66E−04 Antenna 24BG1- G1 Muscle 8.83E−02 Neuronal ActGSI- 2 leg muscle motor neuron system 5.79E−03 Whole body Neuron 6.68E−02 ActGSB- 6 adult ventral nervous 7.56E−04 Whole body Neuron 5.71E−02 Blood ActGSI- 3 hemocyte 1.00E−25 Whole body Circulatory system 1.29E−01 Analysis using the Drosophila RNAi Screening Center’s single- cell DataBase (DRscDB), all datasets used are from FCA 10x Sequencing (https://flycellatlas.org/). The in situ data were from the BDGP (https://insitu.fruitfly.org/cgi- bin/ex/insitu.pl) and the enrichment p value was calculated by a hypergeometric test. the 24BG1 parental line also lost expression of Mhc and the ability to contract (Figure  5—figure supplement 2). This highlights the importance of using early passage cells and avoiding extended passaging that could alter the phenotypic (and genotypic) characteristics of the cells. We also attempted to derive lines from Mef2- Gal4 because Mef2 regulates muscle development and is expressed in muscle progenitors and differentiated muscle suggesting Mef2- Gal4 would be a good candidate for deriving cell lines (Bour et  al., 1995; Gossett et  al., 1989; Lilly et  al., 1995; Ranganayakulu et al., 1995). However, only rare primary cultures had some proliferating cell patches, and none progressed to continuous lines (Supplementary file 1; Figure 5—figure supple- ment 4). Analysis of larvae from the cross (Mef2- Gal4/+; UAS- GFP/UAS- RasV12) and control larvae (Mef2- Gal4/+; UAS- GFP/+) showed that RasV12 expression disrupted muscle development, suggesting that the prevalent amorphous GFP- positive cells observed in primary cultures were abnormal muscle cells (Figure 5—figure supplement 4). The RNAseq analysis for 24B- Gal4- derived cell lines, identified the cells as muscle (Table  2). 24B5- B8 cells express high levels of the transcription factors nautilus (nau) and twist (twi) (Figure 2— figure supplement 3; Figure  5—figure supplement 1; Table  2), and high levels of myoblast city (mbo), which encodes an unconventional bipartite GEF with a role in myoblast fusion (Erickson et al., 1997). The capacity of these mesoderm- derived cell lines to differentiate into active muscle shows that the cells are muscle precursors and thus should be a useful reagent to analyze muscle physiology and development. Neuronal-like cell lines To target neuronal cells, we expressed UAS- RasV12 with the pan- neural drivers scratch- Gal4 and elav- Gal4, however none of the primary cultures resulted in continuous cell lines (Supplementary file 1; Figure 6—figure supplement 1). In previous work, we made primary cultures from embryos with ubiquitous expression of UAS- RasV12 using the Act5C- Gal4 driver (Simcox et  al., 2008b). The cells growing in these cultures included neuronal cells (Simcox et al., 2008b). Here, we used an Act5C- GeneSwitch- Gal4 driver to express UAS- RasV12. GeneSwitch- Gal4 is only active in the presence of the drug, RU486/mifepristone, which provides the advantage of being able to regulate RasV12 expression Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 7 of 27 Cell Biology | Developmental Biology Tools and resources Figure 3. Glial clone Rbr6- 2 cells express Repo. Cells were grown in plain medium (A, C) or treated with ecdysone (B, D). (A, B) After ecdysone treatment, cells make a lace- like network. (C, D) Cells express Repo with or without ecdysone treatment. Inset: DAPI (4′,6- diamidino- 2- phenylindole), DNA. The online version of this article includes the following figure supplement(s) for figure 3: Figure supplement 1. Marker gene expression in glial- lineage clones. Figure supplement 2. Glial cell morphology with and without ecdysone treatment. Figure supplement 3. Gross karyotypes of glial cell clones. (Nicholson et al., 2008; Osterwalder et al., 2001). Several continuous lines were generated (Supple- mentary file 1). Clones derived from two of these (ActGSB- 6 and ActGSI- 2) (Table 1) were positive for the neuronal marker, HRP (horseradish peroxidase) (Figure 6, Figure 6—figure supplements 2 and 3). After differentiation with ecdysone, expression of Futsch/MAPB1 (Hummel et al., 2000) and Fas2 (Mao and Freeman, 2009) was enhanced and revealed axonal- like outgrowths from the cells (Figure 6 and Figure 6—figure supplement 3). Differentiated cells also showed enhanced expression of Elav, which is commonly used as a marker for postmitotic neurons (Figure 6 and Figure 6—figure supplement 3; Robinow and White, 1991). Elav is also expressed transiently in glial cells and prolif- erating neuroblasts Berger et al., 2007; however, the cells were negative for the glial marker Repo (Supplementary file 2). RNAseq analysis revealed that many neuronal genes are upregulated in these cell lines, including Glutamic acid decarboxylase 1 (Gad1), slowpoke (slo), 5- hydroxytryptamine (serotonin) receptor 1A (5- HT1A), Protein C kinase 53E (Pkc53E), Diuretic hormone 31 Receptor (Dh31- R), and straight- jacket (stj). In addition, comparison of the top upregulated genes in these cells to marker genes from scRNAseq data identifies a cell type of neuronal origin as the best match (Table 2). The cells should be a useful source of neuronal cells. Hemocyte-like cell line Cells of clone ActGSI- 3 derived from the ActGSI parental line (UAS- RasV12 expression with Act5C- GeneSwitch- Gal4; Table 1, Supplementary file 1) show characteristics of hemocytes and express the hemocyte marker Hemese (Figure  7; Kurucz et  al., 2003). They are also positive for HRP, but not other neuronal markers (Figure 7—figure supplement 1). ActGSI- 3 cells divide in floating clusters, contrasting with S2 cells, which are also thought to be hemocytes, that grow as single cells (Figures 1 and 7). RNAseq analysis demonstrated that many hemocyte genes are upregulated in these cells, including serpent (srp), Hemese (He), eater, u- shaped (ush), Cecropin A2 (CecA2), and Cecropin C (CecC). Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 8 of 27 Cell Biology | Developmental Biology Tools and resources Figure 4. Tracheal- lineage cells of line Btl3 express the epithelial cadherin Ecad/Shotgun. All panels show Btl3 cells except (B) that shows S2 cells. Cells were grown in plain medium (A–C, E) or treated with ecdysone (D, F). (A) Btl3 cells form a squamous epithelial sheet and express Ecad/Shotgun at cell peripheries. (B) S2 cells grow as single cells and Ecad expression is diffuse. (C) Btl3 cells form a sheet with small cell clusters and expressed Ecad at the cell boundaries (E). (D) Ecdysone- treated cells form large multicellular clusters that expressed Ecad (F). Insets in E and F show nuclei with DAPI. The online version of this article includes the following figure supplement(s) for figure 4: Figure supplement 1. Marker gene expression in tracheal- lineage lines. Figure supplement 2. Morphology of tracheal epithelial parental lines after ecdysone treatment. Figure supplement 3. Gross karyotypes of tracheal epithelial parental cell lines. Comparison of top upregulated genes with scRNAseq data showed that the cells have a strong match to the top marker genes of hemocytes (Table 2). Growth, karyotype, and transfection efficiency of cell lines We determined the cell density at confluence for the cell lines (Table 3). The cells in each line grow to confluence attached to the tissue- culture surface, except ActGSI- 3, which grow as floating cell clusters (Figure 8). The cells are not contact inhibited and cell clusters are formed allowing cells to grow to higher density (Figure 8). We determined the doubling time of 13 cell lines and clones using growth curves (Table 3; Figure 8—figure supplement 1). Most had doubling times within a range of approx- imately 20–40 hr (Table 3). The hemocyte- like clone ActGSI- 3 was an outlier with a longer doubling time of 70 hr (Table 3). In cells from clones ActGSB- 6, ActGSI- 2, and ActGSI- 3, expression of RasV12 is dependent on GeneSwitch Gal4, which is active only in the presence of mifepristone. In the absence of the drug the cells become quiescent (Figure 8—figure supplement 1). We determined the gross karyotype of 13 cell lines and clones. In keeping with previous findings for RasV12 expressing cell lines, most (8) were diploid, or near diploid (Simcox et al., 2008b; Table 3; Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 9 of 27 Cell Biology | Developmental Biology Tools and resources Figure 5. Mesodermal- lineage cells of Clone 24B5- B8 express Myosin heavy chain after differentiation. Cells were grown in plain medium (A, C) or treated with ecdysone (B, D–F). (A) Cells have a bipolar shape. (B) Ecdysone- treated cells elongate and contract. (C) Control cells do not express Mhc. (D) Ecdysone- treated cells express the muscle marker Mhc. Inset: DAPI, DNA. (E, F) Differentiated 24B5- B8 fuse to form muscle fibers that contain multiple nuclei (white arrowheads), some differentiate without fusing with other cells and have single nucleus (blue arrowhead), and some fail to differentiate and remain spherical with a single nucleus (red arrowhead). The online version of this article includes the following figure supplement(s) for figure 5: Figure supplement 1. Marker gene expression in mesodermal- lineage clones. Figure supplement 2. Immunostaining of mesodermal- lineage cells for Myosin heavy chain. Figure supplement 3. Mesodermal cells showed altered morphology after ecdysone treatment. Figure supplement 4. Mef2- Gal4; UAS- GFP; UAS- RasV12 cultures. Figure supplement 5. Gross karyotypes of mesodermal cell clones. Figure 3—figure supplement 3; Figure 4—figure supplement 3; Figure 5—figure supplement 5; Figure 6—figure supplement 4; Figure 7—figure supplement 2). Related clones had similar karyo- types, which likely indicates that parental lines may also be clonal as a result of selective pressure for cells that grow well in culture. Some lines were polyploid and common aneuploid conditions include loss of an X chromosome and varying numbers of chromosome 4 (Table 3). Nine parental and clonal lines were transfected with an Act5C- EGFP plasmid and the fraction of GFP- positive cells was determined after 48 hr. Cells from all lines tested could be transfected. The range of efficiency was from 16% to 34% with most lines showing transfection of approximately one quarter of the cells (Table 3). Similarly treated, cells from the S2 line showed an efficiency of 53%. Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 10 of 27 Cell Biology | Developmental Biology Tools and resources Video 1. 24B- Gal4B5- B8 cells contract spontaneously after differentiation with ecdysone. https://elifesciences.org/articles/85814/figures#video1 Video 3. 24B- Gal4GI cells contract spontaneously after differentiation with ecdysone. Time- lapse video, view looping. https://elifesciences.org/articles/85814/figures#video3 Discussion Expressing activated Ras, RasV12, in primary cells provides a growth and survival advantage and leads to the rapid and reliable generation of continuous cell lines—the so- called Ras method (Simcox et al., 2008b). In a second- generation version of the Ras method, we found that restricting RasV12 expression with lineage- specific Gal4 drivers gave the targeted cells a competitive advantage and produced continuous lines with expected cell- type- specific phenotypes. With this approach we produced glial, epithelial, and muscle cell lines using the repo-, btl-, and 24B/how- Gal4 drivers, respectively. In theory, the approach could be used to produce cell lines corresponding to any cell type for which there is an appropriate Gal4 driver. We tried to derive lines with Mef2- Gal4, a muscle master regulator gene, and the pan- neuronal driver elav- Gal4; however, no continuous lines were produced (Supplementary file 1; Figure  5—figure supplement 4 and Figure  6—figure supplement 1). In both cases, RasV12 expression appeared to disrupt growth of the targeted cell type. In the case of the muscle lineage, 24B/how- Gal4 was efficient at producing cell lines. The success with one and not the other muscle driver shows that in practice, it may be necessary to test multiple Gal4 lines for a given lineage. Drivers with very specific expression patterns may prove useful, including those generated by the Split Gal4 system (Luan et al., 2006). As with any tissue- culture system, the unnatural conditions of growing in vitro may select for ‘generic’ cells that survive well in culture and lose their lineage iden- tities. This means that characterizing cell lines after generation for a battery of features (morpholog- ical, physiological, and molecular) is an essential step in assessing whether cells represent the tissue of origin expected for a given Gal4 driver. repo- Gal4 is a pan- glial driver and many primary cultures expressing RasV12 with this driver reached confluence and could be passaged several times Video 2. 24B- Gal4B5- B8 cells contract spontaneously after differentiation with ecdysone. Video 4. 24B- Gal4GI cells contract spontaneously after differentiation with ecdysone. Time- lapse video, view looping. https://elifesciences.org/articles/85814/figures#video2 https://elifesciences.org/articles/85814/figures#video4 Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 11 of 27 Cell Biology | Developmental Biology Tools and resources Figure 6. Neuronal- like clone ActGSI- 2 expresses neuronal markers. ActGSI- 2 cells were grown in three conditions: RU486 (A, D, G, J, M); RU486 and ecdysone (B, E, H, K, N), or with no additives (C, F, I, L, O). RU486/mifepristone is required for GeneSwtch- Gal4 activation, transgene expression, and cell proliferation. (A) In the growing condition, cells reach confluence and continue to grow by piling up. (B) After ecdysone treatment cells elongated and developed axonal- like outgrowths. (C) In the quiescent state (no RU), cells do not proliferate and fail to reach confluence. (D–F) Cells in all conditions are positive for HRP. (G–I) Expression of Elav, is elevated after ecdysone treatment (H). (J–L) Expression of Futsch/MAP1B- like protein (recognized by antibody 22C10) is elevated after ecdysone treatment (K). (M–O) Fas2 neural- adhesion protein. Cells show elevated expression after ecdysone treatment (N). Insets: DAPI, DNA. Figure 6 continued on next page Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 12 of 27 Cell Biology | Developmental Biology Tools and resources Figure 6 continued The online version of this article includes the following figure supplement(s) for figure 6: Figure supplement 1. elav- G4; UAS- GFP; UAS- RasV12 cultures. Figure supplement 2. Marker gene expression in neuronal- like clones. Figure supplement 3. Neuronal- like clone ActGSB- 6. Figure supplement 4. Gross karyotypes of neuronal- like cell clones. but did not produce continuous lines (Supplementary file 1). We tested different genotypes to deter- mine if the success rate could be improved by modulation of RasV12 expression (co- expression of the Gal4 inhibitor Gal80ts), co- expression of the p35 baculovirus survival factor, or growth stimulation by downregulation of tumor suppressors (dsRNA for warts or brat). One line, also harboring a Gal80ts transgene, reached passage 25; however, the line was unstable and in early passages the cells vari- ably lost Repo expression and changed morphologically. The one continuous glial line generated expresses a transgene that targets the tumor suppressor, brat (repo- Gal4; UAS- RasV12; UAS- bratdsRNA). Given a single success, it is not clear if downregulation of brat contributed to derivation of the line. Moreover, there is no evidence that these genotypic variations enhanced cell line generation with other drivers, as primary cultures expressing RasV12 without modulation or a survival factor produced lines with similar success rates for the btl- Gal4 or 24B/how- Gal4 drivers (Supplementary file 1). As with all types of tissue culture, best practices involve maintaining frozen aliquots of cell lines at relatively low passage numbers. Aliquots of cells from the lines and clones described here, on which RNAseq was performed, have been archived at similar passage numbers as those used for the RNAseq analysis. This will allow users to start experimentation with the lines in a known state. The importance of this is exemplified by line 24BG1, which lost the ability to contract and express the muscle protein Mhc after multiple passages (Figure 5—figure supplement 2). Figure 7. Hemocyte- like Clone ActGSI- 3 morphology and marker expression. Cells were grown in three conditions: RU486 (A, D); RU486 and ecdysone (B, E), or with no additives (C, F). (A) In the growing condition, cells formed floating clusters of multiple cells. (B) After ecdysone treatment cells formed large aggregates and there was cell lysis. (C) In the quiescent state (no RU), individual round cells are seen. (D–F) Cells in all conditions express the hemocyte cell marker Hemese, as recognized by the antibody H2. Inset: DAPI, DNA. The online version of this article includes the following figure supplement(s) for figure 7: Figure supplement 1. Marker gene expression in hemocyte- like clone. Figure supplement 2. Gross karyotypes of hemocyte cell clone. Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 13 of 27 Cell Biology | Developmental Biology Tools and resources Table 3. Confluent density, growth, karyotype, and transfection efficiency of cell lines. Tissue type Line Confluent density (×106)* Doubling time (hr) Karyotype Transfection efficiency (%) Glial Epithelial Rbr6- 2 Rbr6- 4 Rbr6- F9 Btl3 Btl7 Btl8 1.8 2.4 3.4 3.7 2.6 2.7 24B5- B8 1.4 24B5- D8 24BG1- G1 Mesodermal 24BG1- F3 ActGSB- 6 Neuronal ActGSI- 2 Blood ActGSI- 3 S2 5.1 2.8 2.7 2.9 8.1 1.9 6.2 20 20 19 33 37 22 29 23 21 35 23 27 70 ND 8, XY 8, XY 8, XY 7, XY, –4 Abnormal tetraploid, XX, variable 4 Abnormal tetraploid, XX, –4 Abnormal tetraploid, XXY, variable 4 Abnormal tetraploid, XX, variable 4 8, XY (some –4) 8, XY (some –4) 7, XO 8, XX Abnormal tetraploid, XX, variable 4 ND 24 28 22 26 34 16 23 27 ND ND 29 ND ND 53 *Confluent density in one well of a 12- well plate, 3.5 cm2 surface area (average of three wells). The mesodermal, neuronal, and glial cells represent in vitro counterparts of the tissues of origin that can be used for studying development and physiology in an accessible and reproducible system. The mesodermal cells that differentiate into active muscle will allow investigation of muscle fusion, as the cells are multinucleate (Figure 5), as well as muscle physiology and function. For example, the cells contract spontaneously and in apparent waves (Videos 1 and 2); however, the mechanism for stimulation (if any) and regulation have not been investigated and may cast light on in vivo processes. Given a variety of cell types, it will also be interesting to examine cell form and function in co- cultures, for example, of glia and neurons. The method and the cells will be useful for generating disease models. New lineage- specific lines could be generated in the desired mutant background by establishing primary cultures from embryos in which only the mutant genotype expresses RasV12 giving these cells a growth and survival advan- tage (Simcox et al., 2008a). Derivative lines should include those of the desired cell type and geno- type. Alternatively, the existing cell lines could be edited using CRISPR, or insertion of transgenes using the attP site that most lines and clones contain (Supplementary file 1; Bateman et al., 2006; Manivannan et al., 2015). The cells with epithelial morphology derived from the tracheal lineage (Btl3 and Btl7) will provide good models for investigating assay conditions that promote polarization and 3D cell interactions that could allow the cells to manifest a more complex tissue architecture. In keeping with this possibility, treating these cells with ecdysone to induce differentiation showed cell clumping suggestive of a multicellular structure (Figure 4—figure supplement 2). RNAseq analysis of cells from the ActGSI- 3 cell clone showed a striking similarity to hemocytes, and the cells may be a good model for studying immunity (Table  2). The cells lyse after ecdysone treatment suggesting they are of embryonic origin (Figure 7). The cells grow as floating cell clumps (Figures 1 and 7) that may recapitulate subepidermal clusters of sessile hemocytes of the larva (Leitão and Sucena, 2015; Márkus et al., 2009). The most significantly upregulated marker genes in each cell line are significantly enriched for top marker genes from expected cell types based on the single- cell RNAseq data from Fly Cell Atlas in Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 14 of 27 Cell Biology | Developmental Biology Tools and resources Figure 8. Morphology of confluent cultures. (A–C) Glial- lineage clones. The cells grow in dense sheets and ridges with swirl patterns. (D–F) Tracheal- lineage cells. Btl3 and Btl7 cells form squamous epithelial sheets with raised clusters of cells. Btl8 grow densely however individual cells remain separate. (G–J) Mesodermal- lineage cells. The cells grow densely, and form raised clusters. (K, L) Neuronal- like clones. ActGSB- 6 cells grow densely and form Figure 8 continued on next page Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 15 of 27 Cell Biology | Developmental Biology Tools and resources Figure 8 continued peaks and valleys. ActGSI- 2 cells grow densely with scattered raised clusters. (M) Hemocyte- like clone ActGSI- 3. The cells form floating clusters that coalesce into large cell rafts. (N) Schneider’s S2 cells. The cells grow to high density in suspension. Scale bar = 200 µm. The online version of this article includes the following figure supplement(s) for figure 8: Figure supplement 1. Growth curves. most cases. This indicates the potential value of these cell lines as corresponding in vitro models for studying these cell types. While the cells will prove to be valuable models, it should be noted that even those showing a clear differentiated phenotype exhibit unexpected patterns of gene expression. For example, some cells in the mesodermal clone, 24B5- B8, are positive for HRP (Figure 5—figure supplement 1; Supplementary file 2) and the two neuronal- like lines express a mesodermal marker, Twist (Figure  6—figure supplement 2; Supplementary file 2). This anomalous gene expression is likely to be an effect of Ras activation on downstream pathways and genes. Ras/MAPK has a key role in muscle cell determination (Buff et al., 1998; Carmena et al., 2002; Halfon et al., 2000) and acti- vates downstream muscle determination genes. It will also be important to consider what genes are not expressed by a given cell line, for example, glial cell missing (gcm) is not differentially expressed in the three glial- lineage cell clones and gcm2 is differentially expressed in only two of the three clones. Further, trachealess (trh) is only differentially expressed in one of the three tracheal- lineage cell lines. Similarly, the muscle- specific transcription factors twist (twi), nautilus (nau), snail (sna), and Mef2 show variable expression in the muscle- lineage cell clones. It should also be noted that the expression patterns were determined for undifferentiated cells and expression levels could change after hormone exposure. The cells will have value for both low- and high- throughput approaches, including genetic or compound screens for which screening in the relevant cell type will result in identifying targets that are more likely to be of physiological relevance. Most of the cells have an attP- flanked cassette (Table 1), which makes them amenable to insertion of transgenes such as reporters by Recombination Mediated Cassette Exchange (RMCE) (Bateman et al., 2006; Manivannan et al., 2015). Moreover, cells compe- tent for RMCE can be modified by stable expression of Cas9 and then used for genome- wide CRISPR pooled screening. With this approach, a library of single guide RNAs (sgRNAs) are integrated at RMCE sites (Viswanatha et al., 2018; Viswanatha et al., 2019). This generates a pool of cells, each with a different sgRNA, that can be subjected to a screen assay. Results are identified by PCR amplification of inserted sgRNAs followed by next- generation sequencing to detect sgRNAs that are enriched or depleted in the experimental cell pool as compared with a control. To date, pooled CRISPR screens in Drosophila have only been performed in S2 cells, which have hemocyte- like features. The availability of new cell lines with muscle, glial, and epithelial characteristics will enable screens designed to inter- rogate biological processes specific to these cell types. There are hundreds of Drosophila cell lines; however, the number corresponding to known cell types is low. This is due in part to the lack of a method for generating cell lines from specific tissues. We expect that the method described here, using restricted expression of RasV12, will be a tractable approach for investigators to generate lines of cell types of interest. Single- cell cloning followed by cell characterization (immunohistochemistry and RNAseq) also proved to be a useful method to iden- tify cell- type- specific lines and this approach could identify additional valuable lines in the existing collection at the DGRC. In summary, we show that lineage- restricted Ras expression and cell cloning has produced a set of new cell lines that will be of immediate value for analyses in the five cell types they represent. Materials and methods Fly stocks The following fly stocks were used to create primary cell lines: Gal4 drivers: 24B/how- Gal4, w[*]; P(w[+ mW. hs]=GawB)how[24B] (BL 1767); repo- Gal4, P(GAL4)repo (BL 7415); btl- Gal4, P(GAL4- btl.S)3- 2 (BL 78328); Act5C- GeneSwitch- Gal4, P(UAS- GFP.S65T)Myo31DF[T2]; P(Act5C(- FRT)GAL4.Switch.PR)3 (BL 9431). Transgenes: UAS- RasV12 (3), P(w[+mC]=UAS- Ras85D.V12)TL1 (BL 64195); UAS- RasV12 (2), P(w[+mC]=UAS- Ras85D.V12)2 (BL 64196); UAS- RasV12 with RMCE site (3), P(w[+mC]=UAS- Ras85D. Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 16 of 27 Cell Biology | Developmental Biology Tools and resources V12)TL1, P(w[+mC]=attP.w[+].attP)JB89B (BL 64197); UAS- GFP nuclear, P( UAS- GFP. nls) 14 (BL 4775); bratdsRNA, P(y[+t7.7] v[+t1.8]=TRiP.HMS01121)attP2 (BL 34646); UAS- p35 baculovirus death inhibitor, P(w[+mC]=UAS- p35.H)BH1 (BL 5072) and Gal80ts, w[*]; P(w[+mC]=tubP- GAL80[ts])20 (BL 7019). Setting up primary cultures This follows a detailed method, which has additional information (Simcox, 2013), except that no yeast paste is used on the egg collection plates. Yeast paste, even when sterilized, promotes contam- ination in the cultures. Crosses were made between the Gal- 4 driver lines and UAS- RasV12 lines. Some RasV12 stocks had additional alleles as noted in Supplementary file 1. Approximately 200 males and 200 females of a cross were transferred into a laying cage, with a fluted Whatman 3MM paper insert to increase surface area, and eggs were collected using 60- mm Petri dishes containing egg laying medium. Egg collections were made during the day for 8 hr at room temperature or 16 hr overnight at 17°C. After collection, approximately 3 ml of TXN (NaCl [0.7%], Triton X [0.02%] in water) was added to the plate. Any hatched larvae, which rise to the surface, were removed and the unhatched embryos were dislodged using a large soft paint brush to gently release them from the surface. Embryos were tipped off with the liquid into a sieve. Additional rinsing and brushing were used to ensure most embryos were dislodged and collected in the sieve. After thorough rinsing of the embryos with TXN from a squirt bottle, the sieve was upended over a 15- ml Falcon tube and a stream of TXN was used to transfer the embryos into the tube. Once the embryos settled, the TXN was removed and replaced with 3 ml of 50% bleach (Clorox) in water. The tube was capped and inverted three to five times and subsequently the embryos were treated using sterile techniques. The embryos were allowed to settle at the bottom of the tube and the bleach was removed after 3–5 min. The bleach dechorionates and surface sterilizes the embryos. The embryos were rinsed 2× with 4 ml of sterile TXN and transferred to a fresh tube of TXN to minimize bleach contamination. After two additional TXN rinses the embryos were transferred to TXN in a 5- ml glass homogenizer (with Teflon pestle). Embryos were rinsed in 3 ml of water followed by a rinse in 1 ml of Schneider’s S2 medium (supplemented with 10% heat inactivated fetal bovine serum and 1× Pen- strep solution). Embryos tend to clump in the Schneider’s S2 medium and stick to the sides of the homogenizer and pipette and care is needed to remove the medium without disturbing the embryos. 3  ml of fresh Schneider’s S2 medium was added to the homogenizer and the embryos were disrupted by three gentle strokes with the pestle. Care was taken to minimize bubbles by not withdrawing the pestle beyond the surface of the liquid. The homogenate was allowed to settle for 2 min and the super- natant was transferred to a 15- ml Falcon tube leaving the large cell clumps and any whole embryos in the bottom of the homogenizer. 3 ml of fresh Schneider’s S2 medium was added to the homoge- nizer and three more strokes, with a twist at the bottom, were used to disrupt remaining tissue and embryos. The second homogenate was added to the Falcon tube. The tube was centrifuged in a benchtop centrifuge at 1400 × g. The supernatant was discarded, and the pellet was resuspended in 3- ml Schneider’s S2 medium and centrifugation step and washing with Schneider’s S2 medium was repeated twice more. The final pellet size was estimated and plated in 1 or more 12.5 cm2 T- flasks with 2–3 ml Schneider’s S2 medium. The number of flasks needed for a given pellet size can also be estimated from the volume of packed embryos with approximately 30 µl of packed embryos being sufficient for one flask. Culture conditions for new cell lines Cells were grown in 25 cm2 T- flasks at 25°C in Schneider’s S2 medium and were passaged at between 90% and full confluence (Figure 8) using trypsin to release cells from the tissue- culture surface. Trypsin is needed as cells in all the lines are adherent except ActGSI3 cells that float freely (Figures 1 and 8). Cells were pelleted and approximately 20–25% of the cells were plated in a new flask. Cells were checked using an inverted microscope approximately every 5 days. The medium was changed on cultures showing signs of poor cell health (extended processes, little growth). This was sometimes necessary for cell types that are more metabolically active and acidify the medium, including the mesodermal lines. Cells were passaged every 5–7 days. Cell freezing (Schneider’s S2 medium with 20% heat inactivated fetal bovine serum and 10% DMSO (Dimethyl sulfoxide)) was used to keep a supply of frozen aliquots so that cells with similar passage numbers were used in experiments. Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 17 of 27 Cell Biology | Developmental Biology Tools and resources Cell cloning For puromycin selection, 2–6 × 105 cells in a 35- mM well were transfected with 0.4 µg of DNA encoding a puro resistance plasmid (pCoPURO, Addgene #17533) using Effectene Transfection Reagent (QIAGEN). After 24  hr, cells were selected with puromycin at 0.5–2.5  µg/ml for 5 days. After 2–4 weeks, colonies were isolated and expanded. For dilution cloning, cells were seeded into a 96- well plate at a concentration of 0.5–1 cell/well in 100 µl conditioned media (Housden et al., 2015). Hormone treatment To simulate the major pulse of ecdysone at the larval to pupal transition, cells were treated with two 24 hr doses of β-ecdysone (Sigma 5289- 74- 7) at 1 µg/ml separated by 24 hr in non- supplemented medium. Immunohistochemistry Cells were fixed with 4% paraformaldehyde (Electron Microscopy Sciences) for 15 min or 3.5% form- aldehyde (Sigma) for 30  min at room temperature, and then rinsed twice with 0.1% Tween- 20 in phosphate- buffered saline (PBS- T). Cells were permeabilized (0.2% Triton X- 100 in PBS) for 10 min at room temperature. Cells were blocked (5% bovine serum albumin in PBS- T) for 30  min at room temperature and incubated with diluted primary antibodies overnight at 4°C. Cells were washed three times with PBS- T and incubated with diluted secondary antibodies in blocking buffer for 1 hr at room temperature or overnight at 4°C. Cells were washed three times with PBS- T and mounted in Vecta- Shield with DAPI (Vector Laboratories). For the Dcad2 antibody, cells were fixed and processed as described in Oda et al., 1994. The following primary antibodies and dilutions were used: HRP (rabbit polyclonal, Jackson ImmunoResearch 323- 005- 021, 1:500), 22C10 (mouse monoclonal anti- Futsch, Developmental Studies Hybridoma Bank, DSHB, 1:100), ELAV (rat monoclonal, DSHB 7E8A10, 1:100), Repo (mouse monoclonal, DSHB 8D12, 1:100), FasII (mouse monoclonal, DSHB 1D4, 1:100), Twist (a gift from M. Levine, UC Berkeley, CA, guinea pig 1:500), MHC (mouse monoclonal, DSHB 3E8- 3D3, 1:100), Dcad2 (rat monoclonal, DSHB, 1:100), and DMef2 (a gift from J. R. Jacobs [Vanderploeg et al., 2012], rabbit polyclonal, 1:500), H2 (mouse monoclonal, [Kurucz et al., 2003], 1:10). Cells were incubated with the following secondary antibodies at the indicated dilutions: Cy3- conjugated goat anti- mouse (Jackson ImmunoResearch 115- 165- 003, 1:1000), Cy3- conjugated goat anti- rat (Jackson ImmunoResearch 112- 165- 003, 1:1000), Cy3- conjugated goat anti- guinea pig (Jackson ImmunoRe- search 106- 165- 003, 1:1000), Cy3- conjugated goat anti- rabbit (Jackson ImmunoResearch 111- 165- 045, 1:1000), and Alexa Fluor 488- conjugated donkey anti- rabbit (Invitrogen A- 21206, 1:1000). Growth curve analysis 1–2 × 105 cells were plated in a 12- well plate. Cells were counted from triplicate wells every 3 days over a 9- day period. Doubling time was calculated using log2 cell numbers (Roth, 2006). Karyotype analysis Cells were grown to 50–90% confluence and incubated with 0.05 µg/ml KaryoMAX (Gibco- Thermo Fisher 15212012) for 3–18 hr. Cells were processed for analysis using the method in Lee et al., 2014, which uses 0.5% sodium citrate as a hypotonic solution and a 3:1 ice cold mix of methanol and acetic acid as a fix. After dropping fixed cells, slides were air dried and mounted in VectaShield with DAPI (Vector Laboratories) and viewed with an Olympus BX41 microscope. Transfection Cells in a 6- well plate (approximately 70% confluent) were transfected with 0.4 µg of an Actin5C- EGFP plasmid (pAc5.1B- EGFP, Addgene #21181) using Effectene Transfection Reagent (QIAGEN). The frac- tion of GFP- positive cells was scored after 48 hr. RNA extraction and RNAseq Cell cultures were grown and expanded in their respective media. All cell lines were cultured in Schneiders Drosophila Medium (Gibco Cat # 21720001), supplemented with 10% fetal bovine serum (Cytiva Hyclone Cat SH30070.03). For Act5C- GS>Ras attP- GFP- LI- Clone 2, Act5C- GS>Ras attP- GFP- LI- Clone 3, and Act5C- GS>Ras attP- GFP- LB- Clone 6 cultures were grown in the same basal media Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 18 of 27 Cell Biology | Developmental Biology Tools and resources supplemented with 10 nM of Mifepristone (Thermo Fisher Cat# H11001). Cultures were allowed to grow in T- 25 flasks to become confluent before treatment with trypsin (Gibco Cat# 12604013) for 4 min to dislodge the cell monolayer from the growth surface. The cells were resuspended in 4 ml of their respective media and 1 ml of the cell suspension was collected for pelleting, followed by washing in 1× PBS, and then flash- freezing in liquid nitrogen. All cell samples were processed in triplicates. Total RNA was isolated from the pellets using the TRIzol reagent (Life Technologies [Ambion], Cat#:15596018) as per the manufacturer’s instructions. The isolated total RNA was subjected to further purification using the RNeasy Mini Kit (QIAGEN, Cat#74104) and the RNA post- cleanup was eluted in RNase- free water. The eluted total RNA was confirmed to have a A260/A280 ratio >1.8 and RIN >7. Upon passing the quality control parameters, Illumina TruSeq libraries were constructed using TruSeq stranded mRNA HT kit (Illumina, Cat# RS- 122- 2103). Paired end sequencing was performed on an Illumina NextSeq 500 with a 150- cycle high output kits (Illumina, Cat# FC- 404- 2002). RNAseq data analysis Raw data processing was performed using the STAR sequence aligner (https://github.com/alexdobin/ STAR; Dobin et al., 2013). Reads were aligned to the Drosophila genome and featureCounts were used to get gene counts from all samples into a count matrix for downstream analysis. A principal component analysis plot was produced using heatmaply. FPKM values were calculated using fpkm(- DEseq2) using gene length output by featureCounts. The reference genome used was FB2022_05, dmel_r6.48 (FlyBase) (Jenkins et al., 2022). Both raw sequencing reads and the count matrix were deposited in the NCBI Gene Expression Omnibus (GEO) database under the accession number GSE219105. The processed dataset has also been imported into DGET database for user to mine gene(s) of interest or search for genes with similar expression pattern (https://www.flyrnai.org/tools/ dget/web/). Each sample was compared against all other samples by using DESeq2 ( Love et  al., 2014) to determine differentially expressed genes (DE calling). The set of top DE genes for each cell line was compared with the top 100 markers in single- cell RNAseq datasets corresponding to cell types in the Fly Cell Atlas 10× datasets (Li et al., 2022). Enrichment analysis was conducted using the DRscDB tool to identify the Fly Cell Atlas cell type that matched closely to each cell line (Hu et al., 2021). We also compared the DE genes with the genes identify in various tissues in embryo and larval based on in situ data (PMID: 24359758, 17645804, 12537577) and majority of the best matching tissues are consistent with the analysis using scRNAseq datasets (Table 2). The RNAseq data for the cell lines described in this work were also compared with RNAseq datasets determined previously for 24 other Drosophila cell lines (Cherbas et al., 2011). The comparison was conducted by hierarchical clustering analysis using Pearson correlation coefficient scores. To survey the activities of major signaling pathways in the cell lines, we specifically selected the ligands and receptors annotated at FlyPhoneDB (PMID: 35100387) to plot their expression levels using heatmap. Materials availability All cell lines described here have been deposited to the Drosophila Genomics Resource Center (DGRC) at Indiana University. The lines are available for distribution to the research community. Acknowledgements We thank M Levine, J R Jacobs, and D Hultmark for antibodies and the Bloomington Stock Center for fly stocks. We thank Mikhail Kouzminov for help with data analysis. Funding This work is supported by the National Institutes of Health (NIH Office of the Director R24 OD019847 to NP, SEM, and AS, P40OD010949 to the DGRC, and NIH NIGMS P41 GM132087 to the DRSC- BTRR), the National Science Foundation (IOS 1419535 to AS, and support while serving at the National Science Founda- tion to AS), the Howard Hughes Medical Institute (NP), and a grant from Women & Philanthropy at the Ohio State University (to AS). Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 19 of 27 Cell Biology | Developmental Biology Tools and resources Additional information Funding Funder Grant reference number Author National Institutes of Health (NIH) Office of the Director R24 OD019847 Norbert Perrimon Stephanie E Mohr Amanda Simcox National Institutes of Health National Institutes of Health National Science Foundation P40OD010949 Andrew Zelhof P41 GM132087 Norbert Perrimon Stephanie E Mohr IOS 1419535 Amanda Simcox Howard Hughes Medical Institute Women & Philanthropy at The Ohio State University Grant National Science Foundation Support while serving at the National Science Foundation Norbert Perrimon Amanda Simcox Amanda Simcox The funders had no role in study design, data collection, and interpretation, or the decision to submit the work for publication. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation. Author contributions Nikki Coleman- Gosser, Shane Stitzinger, Formal analysis, Investigation, Writing – review and editing; Yanhui Hu, Formal analysis, Writing – review and editing; Shiva Raghuvanshi, Molly Josifov, Investiga- tion; Weihang Chen, Formal analysis; Arthur Luhur, Daniel Mariyappa, Investigation, Writing – review and editing; Andrew Zelhof, Funding acquisition, Writing – review and editing; Stephanie E Mohr, Norbert Perrimon, Supervision, Funding acquisition, Writing – review and editing; Amanda Simcox, Conceptualization, Formal analysis, Supervision, Funding acquisition, Investigation, Methodology, Writing - original draft, Project administration, Writing – review and editing Author ORCIDs Daniel Mariyappa Molly Josifov Stephanie E Mohr Norbert Perrimon Amanda Simcox http://orcid.org/0000-0003-4775-1656 http://orcid.org/0000-0002-2899-7186 http://orcid.org/0000-0001-9639-7708 http://orcid.org/0000-0001-7542-472X http://orcid.org/0000-0002-5572-7042 Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.85814.sa1 Author response https://doi.org/10.7554/eLife.85814.sa2 Additional files Supplementary files • Supplementary file 1. Primary cultures and continuous lines produced from indicated genotypes. Glial primary cultures grew well at first and could be passaged several times; however, only one continuous line was produced. This line was cloned using single- cell dilution to produce three clonal derivatives. Tracheal lines were produced readily. Cloning the parental lines was not successful with either single- cell dilution or puro selection. Mesodermal lines were produced using expression of RasV12 with 24B- Gal4 but not Mef2- Gal4. Cloning of the continuous lines was done using single- cell dilution. Neuronal. Expression of RasV12 with neuronal Gal4 drivers (elav- Gal4 or scratch- Gal4) did not give rise to continuous lines. Cloning of lines generated by broad expression of RasV12 with Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 20 of 27 Cell Biology | Developmental Biology Tools and resources Act5C- GeneSwitch- Gal4 produced two clonal lines with neuronal characteristics and one with hemocyte characteristics. These were cloned using puro selection. • Supplementary file 2. Analysis of marker gene expression in parental lines and clones. Cell lines and their clonal derivatives were stained with antibodies against the indicated markers. The fraction of cells staining positive was determined. The intensity and cellular location of the signal are indicated in cases when there was variation. The clones and parental lines highlighted were analyzed by RNAseq. • Supplementary file 3. Fragments Per Kilobase of transcript per Million mapped reads (FPKM). FPKM values are shown for each of the clones and parental lines that were analyzed by RNAseq. • MDAR checklist Data availability Sequencing data have been deposited in GEO under accession code GSE219105. The following dataset was generated: Author(s) Mariyappa D, Luhur A, Zelhof A, Hu Y, Simcox A Year 2022 Dataset title Dataset URL Database and Identifier Continuous muscle, glial, epithelial, neuronal, and hemocyte cell lines for Drosophila research https://www. ncbi. nlm. nih. gov/ geo/ query/ acc. cgi? acc= GSE219105 NCBI Gene Expression Omnibus, GSE219105 References Ammeux N, Housden BE, Georgiadis A, Hu Y, Perrimon N. 2016. Mapping signaling pathway cross- talk in Drosophila cells. PNAS 113:9940–9945. DOI: https://doi.org/10.1073/pnas.1610432113, PMID: 27528688 Bateman JR, Lee AM, Wu C. 2006. Site- specific transformation of Drosophila via phiC31 integrase- mediated cassette exchange. Genetics 173:769–777. DOI: https://doi.org/10.1534/genetics.106.056945, PMID: 16547094 Baum B, Cherbas L. 2008. Drosophila cell lines as model systems and as an experimental tool. Methods in Molecular Biology 420:391–424. DOI: https://doi.org/10.1007/978-1-59745-583-1 Berger C, Renner S, Lüer K, Technau GM. 2007. The commonly used marker ELAV is transiently expressed in neuroblasts and glial cells in the Drosophila embryonic CNS. Developmental Dynamics 236:3562–3568. DOI: https://doi.org/10.1002/dvdy.21372, PMID: 17994541 Bour BA, O’Brien MA, Lockwood WL, Goldstein ES, Bodmer R, Taghert PH, Abmayr SM, Nguyen HT. 1995. Drosophila MEF2, a transcription factor that is essential for myogenesis. Genes & Development 9:730–741. DOI: https://doi.org/10.1101/gad.9.6.730, PMID: 7729689 Brand AH, Perrimon N. 1993. Targeted gene expression as a means of altering cell fates and generating dominant phenotypes. Development 118:401–415. DOI: https://doi.org/10.1242/dev.118.2.401 Buff E, Carmena A, Gisselbrecht S, Jiménez F, Michelson AM. 1998. Signalling by the Drosophila epidermal growth factor receptor is required for the specification and diversification of embryonic muscle progenitors . Development 125:2075–2086. DOI: https://doi.org/10.1242/dev.125.11.2075 Carmena A, Buff E, Halfon MS, Gisselbrecht S, Jiménez F, Baylies MK, Michelson AM. 2002. Reciprocal Regulatory Interactions between the Notch and Ras Signaling Pathways in the Drosophila Embryonic Mesoderm. Developmental Biology 244:226–242. DOI: https://doi.org/10.1006/dbio.2002.0606 Chen D, Roychowdhury- Sinha A, Prakash P, Lan X, Fan W, Goto A, Hoffmann JA. 2021. A time course transcriptomic analysis of host and injected oncogenic cells reveals new aspects of Drosophila immune defenses. PNAS 118:e2100825118. DOI: https://doi.org/10.1073/pnas.2100825118, PMID: 33737397 Cherbas L, Willingham A, Zhang D, Yang L, Zou Y, Eads BD, Carlson JW, Landolin JM, Kapranov P, Dumais J, Samsonova A, Choi J- H, Roberts J, Davis CA, Tang H, van Baren MJ, Ghosh S, Dobin A, Bell K, Lin W, et al. 2011. The transcriptional diversity of 25 Drosophila cell lines. Genome Research 21:301–314. DOI: https://doi. org/10.1101/gr.112961.110, PMID: 21177962 Cherbas L, Gong L. 2014. Cell lines. Methods 68:74–81. DOI: https://doi.org/10.1016/j.ymeth.2014.01.006 Debec A, Megraw TL, Guichet A. 2016. Methods to Establish Drosophila Cell Lines. Methods in Molecular Biology 1478:333–351. DOI: https://doi.org/10.1007/978-1-4939-6371-3 Dobin A, Davis CA, Schlesinger F, Drenkow J, Zaleski C, Jha S, Batut P, Chaisson M, Gingeras TR. 2013. STAR: ultrafast universal RNA- seq aligner. Bioinformatics 29:15–21. DOI: https://doi.org/10.1093/bioinformatics/ bts635, PMID: 23104886 Erickson MR, Galletta BJ, Abmayr SM. 1997. Drosophila myoblast city encodes a conserved protein that is essential for myoblast fusion, dorsal closure, and cytoskeletal organization. The Journal of Cell Biology 138:589–603. DOI: https://doi.org/10.1083/jcb.138.3.589, PMID: 9245788 Featherstone DE. 2011. Glial solute carrier transporters in Drosophila and mice. Glia 59:1351–1363. DOI: https://doi.org/10.1002/glia.21085, PMID: 21732427 Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 21 of 27 Cell Biology | Developmental Biology Tools and resources Goodman CL, Kang DS, Stanley D. 2021. Cell line platforms support research into arthropod immunity. Insects 12:738. DOI: https://doi.org/10.3390/insects12080738, PMID: 34442304 Gossett LA, Kelvin DJ, Sternberg EA, Olson EN. 1989. A new myocyte- specific enhancer- binding factor that recognizes A conserved element associated with multiple muscle- specific genes. Molecular and Cellular Biology 9:5022–5033. DOI: https://doi.org/10.1128/mcb.9.11.5022-5033.1989, PMID: 2601707 Halfon MS, Carmena A, Gisselbrecht S, Sackerson CM, Jiménez F, Baylies MK, Michelson AM. 2000. Ras pathway specificity is determined by the integration of multiple signal- activated and tissue- restricted transcription factors. Cell 103:63–74. DOI: https://doi.org/10.1016/S0092-8674(00)00105-7 Hockemeyer D, Jaenisch R. 2016. Induced pluripotent stem cells meet genome editing. Cell Stem Cell 18:573– 586. DOI: https://doi.org/10.1016/j.stem.2016.04.013 Hong M, Li T, Xue W, Zhang S, Cui L, Wang H, Zhang Y, Zhou L, Gu Y, Xia N, Li S. 2022. Genetic engineering of baculovirus- insect cell system to improve protein production. Frontiers in Bioengineering and Biotechnology 10:994743. DOI: https://doi.org/10.3389/fbioe.2022.994743, PMID: 35519615 Housden BE, Valvezan AJ, Kelley C, Sopko R, Hu Y, Roesel C, Lin S, Buckner M, Tao R, Yilmazel B, Mohr SE, Manning BD, Perrimon N. 2015. Identification of potential drug targets for tuberous sclerosis complex by synthetic screens combining CRISPR- based knockouts with RNAi. Science Signaling 8:rs9. DOI: https://doi.org/ 10.1126/scisignal.aab3729 Hu Y, Comjean A, Perrimon N, Mohr SE. 2017. The Drosophila Gene Expression Tool (DGET) for expression analyses. BMC Bioinformatics 18:98. DOI: https://doi.org/10.1186/s12859-017-1509-z Hu Y, Tattikota SG, Liu Y, Comjean A, Gao Y, Forman C, Kim G, Rodiger J, Papatheodorou I, dos Santos G, Mohr SE, Perrimon N. 2021. DRscDB: A single- cell RNA- seq resource for data mining and data comparison across species. Computational and Structural Biotechnology Journal 19:2018–2026. DOI: https://doi.org/10. 1016/j.csbj.2021.04.021 Hummel T, Krukkert K, Roos J, Davis G, Klämbt C. 2000. Drosophila Futsch/22C10 Is a MAP1B- like Protein Required for Dendritic and Axonal Development. Neuron 26:357–370. DOI: https://doi.org/10.1016/S0896- 6273(00)81169-1 Jenkins VK, Larkin A, Thurmond J, FlyBase Consortium. 2022. Using FlyBase: A Database of Drosophila Genes and Genetics. Methods in Molecular Biology 2540:1–34. DOI: https://doi.org/10.1007/978-1-0716-2541-5_1, PMID: 35980571 Kahn TG, Stenberg P, Pirrotta V, Schwartz YB. 2014. Combinatorial interactions are required for the efficient recruitment of pho repressive complex (PhoRC) to polycomb response elements. PLOS Genetics 10:e1004495. DOI: https://doi.org/10.1371/journal.pgen.1004495, PMID: 25010632 Klein SG, Steckbauer A, Alsolami SM, Arossa S, Parry AJ, Li M, Duarte CM. 2022. Toward best practices for controlling mammalian cell culture environments. Frontiers in Cell and Developmental Biology 10:788808. DOI: https://doi.org/10.3389/fcell.2022.788808, PMID: 35265608 Kurucz E, Zettervall C- J, Sinka R, Vilmos P, Pivarcsi A, Ekengren S, Hegedüs Z, Ando I, Hultmark D. 2003. Hemese, a hemocyte- specific transmembrane protein, affects the cellular immune response in Drosophila. PNAS 100:2622–2627. DOI: https://doi.org/10.1073/pnas.0436940100, PMID: 12598653 Lee H, McManus CJ, Cho D- Y, Eaton M, Renda F, Somma MP, Cherbas L, May G, Powell S, Zhang D, Zhan L, Resch A, Andrews J, Celniker SE, Cherbas P, Przytycka TM, Gatti M, Oliver B, Graveley B, MacAlpine D. 2014. DNA copy number evolution in Drosophila cell lines. Genome Biology 15:R70. DOI: https://doi.org/10.1186/ gb-2014-15-8-r70, PMID: 25262759 Lee H- G, Kahn TG, Simcox A, Schwartz YB, Pirrotta V. 2015. Genome- wide activities of Polycomb complexes control pervasive transcription. Genome Research 25:1170–1181. DOI: https://doi.org/10.1101/gr.188920.114 Leitão AB, Sucena É. 2015. Drosophila sessile hemocyte clusters are true hematopoietic tissues that regulate larval blood cell differentiation. eLife 4:e06166. DOI: https://doi.org/10.7554/eLife.06166, PMID: 25650737 Li H, Janssens J, De Waegeneer M, Kolluru SS, Davie K, Gardeux V, Saelens W, David FPA, Brbić M, Spanier K, Leskovec J, McLaughlin CN, Xie Q, Jones RC, Brueckner K, Shim J, Tattikota SG, Schnorrer F, Rust K, Nystul TG, et al. 2022. Fly Cell Atlas: A single- nucleus transcriptomic atlas of the adult fruit fly. Science 375:eabk2432. DOI: https://doi.org/10.1126/science.abk2432 Lilly B, Zhao B, Ranganayakulu G, Paterson BM, Schulz RA, Olson EN. 1995. Requirement of MADS Domain Transcription Factor D- MEF2 for Muscle Formation in Drosophila . Science 267:688–693. DOI: https://doi.org/ 10.1126/science.7839146 Lim MYT, Ng AWT, Chou Y, Lim TP, Simcox A, Tucker- Kellogg G, Okamura K. 2016. The Drosophila Dicer- 1 Partner Loquacious Enhances miRNA Processing from Hairpins with Unstable Structures at the Dicing Site. Cell Reports 15:1795–1808. DOI: https://doi.org/10.1016/j.celrep.2016.04.059 Love M, Anders S, Huber W. 2014. Deseq2. 3.4. . https://bioconductor.statistik.tu-dortmund.de/packages/3.4/ bioc/html/DESeq2.html Luan H, Peabody NC, Vinson CR, White BH. 2006. Refined spatial manipulation of neuronal function by combinatorial restriction of transgene expression. Neuron 52:425–436. DOI: https://doi.org/10.1016/j.neuron. 2006.08.028, PMID: 17088209 Luhur A, Klueg KM, Zelhof AC. 2019. Generating and working with Drosophila cell cultures: Current challenges and opportunities. Wiley Interdisciplinary Reviews. Developmental Biology 8:e339. DOI: https://doi.org/10. 1002/wdev.339, PMID: 30561900 Manivannan SN, Jacobsen TL, Lyon P, Selvaraj B, Halpin P, Simcox A. 2015. Targeted Integration of Single- Copy Transgenes in Drosophila melanogaster Tissue- Culture Cells Using Recombination- Mediated Cassette Exchange. Genetics 201:1319–1328. DOI: https://doi.org/10.1534/genetics.115.181230, PMID: 26500255 Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 22 of 27 Cell Biology | Developmental Biology Tools and resources Mao Y, Freeman M. 2009. Fasciclin 2, the Drosophila orthologue of neural cell- adhesion molecule, inhibits EGF receptor signalling. Development 136:473–481. DOI: https://doi.org/10.1242/dev.026054, PMID: 19141676 Márkus R, Laurinyecz B, Kurucz E, Honti V, Bajusz I, Sipos B, Somogyi K, Kronhamn J, Hultmark D, Andó I. 2009. Sessile hemocytes as a hematopoietic compartment in Drosophila melanogaster. PNAS 106:4805–4809. DOI: https://doi.org/10.1073/pnas.0801766106, PMID: 19261847 Martin RL, Maiorano J, Beitel GJ, Marko JF, McVicker G, Fondufe- Mittendorf YN. 2017. A comparison of nucleosome organization in Drosophila cell lines. PLOS ONE 12:e0178590. DOI: https://doi.org/10.1371/ journal.pone.0178590, PMID: 28570602 Mohr SE. 2014. RNAi screening in Drosophila cells and in vivo. Methods 68:82–88. DOI: https://doi.org/10.1016/ j.ymeth.2014.02.018, PMID: 24576618 Nakato E, Liu X, Eriksson I, Yamamoto M, Kinoshita- Toyoda A, Toyoda H, Kjellén L, Li J- P, Nakato H. 2019. Establishment and characterization of Drosophila cell lines mutant for heparan sulfate modifying enzymes. Glycobiology 29:479–489. DOI: https://doi.org/10.1093/glycob/cwz020, PMID: 30869121 Neal SJ, Dolezal D, Jusić N, Pignoni F. 2019. Drosophila ML- DmD17- c3 cells respond robustly to Dpp and exhibit complex transcriptional feedback on BMP signaling components. BMC Developmental Biology 19:1. DOI: https://doi.org/10.1186/s12861-019-0181-0, PMID: 30669963 Nicholson L, Singh GK, Osterwalder T, Roman GW, Davis RL, Keshishian H. 2008. Spatial and temporal control of gene expression in Drosophila using the inducible GeneSwitch GAL4 system I Screen for larval nervous system drivers. Genetics 178:215–234. DOI: https://doi.org/10.1534/genetics.107.081968, PMID: 18202369 Oda H, Uemura T, Harada Y, Iwai Y, Takeichi M. 1994. A Drosophila homolog of cadherin associated with armadillo and essential for embryonic cell- cell adhesion. Developmental Biology 165:716–726. DOI: https:// doi.org/10.1006/dbio.1994.1287, PMID: 7958432 Ogienko AA, Andreyeva EN, Omelina ES, Oshchepkova AL, Pindyurin AV. 2020. Molecular and cytological analysis of widely- used Gal4 driver lines for Drosophila neurobiology. BMC Genetics 21:96. DOI: https://doi. org/10.1186/s12863-020-00895-7, PMID: 33092520 Osterwalder T, Yoon KS, White BH, Keshishian H. 2001. A conditional tissue- specific transgene expression system using inducible GAL4. PNAS 98:12596–12601. DOI: https://doi.org/10.1073/pnas.221303298, PMID: 11675495 Ranganayakulu G, Zhao B, Dokidis A, Molkentin JD, Olson EN, Schulz RA. 1995. A series of mutations in the D- MEF2 transcription factor reveal multiple functions in larval and adult myogenesis in Drosophila. Developmental Biology 171:169–181. DOI: https://doi.org/10.1006/dbio.1995.1269, PMID: 7556894 Rhee DY, Cho D- Y, Zhai B, Slattery M, Ma L, Mintseris J, Wong CY, White KP, Celniker SE, Przytycka TM, Gygi SP, Obar RA, Artavanis- Tsakonas S. 2014. Transcription factor networks in Drosophila melanogaster. Cell Reports 8:2031–2043. DOI: https://doi.org/10.1016/j.celrep.2014.08.038, PMID: 25242320 Ritacco FV, Wu Y, Khetan A. 2018. Cell culture media for recombinant protein expression in Chinese hamster ovary (CHO) cells: History, key components, and optimization strategies. Biotechnology Progress 34:1407– 1426. DOI: https://doi.org/10.1002/btpr.2706 Robinow S, White K. 1991. Characterization and spatial distribution of the ELAV protein during Drosophila melanogaster development. Journal of Neurobiology 22:443–461. DOI: https://doi.org/10.1002/neu. 480220503, PMID: 1716300 Rossi G, Manfrin A, Lutolf MP. 2018. Progress and potential in organoid research. Nature Reviews. Genetics 19:671–687. DOI: https://doi.org/10.1038/s41576-018-0051-9, PMID: 30228295 Roth V. 2006. Doubling Time. Doubling Time Computing. http://www.doubling-time.com/compute.php Sato K, Siomi MC. 2020. The piRNA pathway in Drosophila ovarian germ and somatic cells. Proceedings of the Japan Academy. Series B, Physical and Biological Sciences 96:32–42. DOI: https://doi.org/10.2183/pjab.96. 003, PMID: 31932527 Sepp KJ, Schulte J, Auld VJ. 2001. Peripheral glia direct axon guidance across the CNS/PNS transition zone. Developmental Biology 238:47–63. DOI: https://doi.org/10.1006/dbio.2001.0411, PMID: 11783993 Shiga Y, Tanaka- Matakatsu M, Hayashi S. 1996. A nuclear GFP/beta- galactosidase fusion protein as A marker for morphogenesis in living Drosophila. Development, Growth and Differentiation 38:99–106. DOI: https://doi.org/ 10.1046/j.1440-169X.1996.00012.x Simcox A, Mitra S, Truesdell S, Paul L, Chen T, Butchar JP, Justiniano S. 2008a. Efficient genetic method for establishing Drosophila cell lines unlocks the potential to create lines of specific genotypes. PLOS Genetics 4:e1000142. DOI: https://doi.org/10.1371/journal.pgen.1000142, PMID: 18670627 Simcox AA, Austin CL, Jacobsen TL, Jafar- Nejad H. 2008b. Drosophila embryonic “fibroblasts”: extending mutant analysis in vitro. Fly 2:306–309. DOI: https://doi.org/10.4161/fly.7427, PMID: 19077546 Simcox A. 2013. Progress towards Drosophila epithelial cell culture. Methods in Molecular Biology 945:1–11. DOI: https://doi.org/10.1007/978-1-62703-125-7_1, PMID: 23097097 Smagghe G, Goodman CL, Stanley D. 2009. Insect cell culture and applications to research and pest management. In Vitro Cellular & Developmental Biology - Animal 45:93–105. DOI: https://doi.org/10.1007/ s11626-009-9181-x Stoiber M, Celniker S, Cherbas L, Brown B, Cherbas P. 2016. Diverse Hormone Response Networks in 41 Independent Drosophila Cell Lines. G3: Genes, Genomes, Genetics 6:683–694. DOI: https://doi.org/10.1534/ g3.115.023366, PMID: 26772746 Vanderploeg J, Vazquez Paz LL, MacMullin A, Jacobs JR. 2012. Integrins are required for cardioblast polarisation in Drosophila. BMC Developmental Biology 12:8. DOI: https://doi.org/10.1186/1471-213X-12-8 Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 23 of 27 Cell Biology | Developmental Biology Tools and resources Viswanatha R, Li Z, Hu Y, Perrimon N. 2018. Pooled genome- wide CRISPR screening for basal and context- specific fitness gene essentiality in Drosophila cells. eLife 7:e36333. DOI: https://doi.org/10.7554/eLife.36333, PMID: 30051818 Viswanatha R, Brathwaite R, Hu Y, Li Z, Rodiger J, Merckaert P, Chung V, Mohr SE, Perrimon N. 2019. Pooled CRISPR Screens in Drosophila Cells. Current Protocols in Molecular Biology 129:e111. DOI: https://doi.org/10. 1002/cpmb.111, PMID: 31763777 Wang YE, Kutnetsov L, Partensky A, Farid J, Quackenbush J. 2017. WebMeV: a cloud platform for analyzing and visualizing cancer genomic data. Cancer Research 77:e11–e14. DOI: https://doi.org/10.1158/0008-5472. CAN-17-0802, PMID: 29092929 Wen J, Mohammed J, Bortolamiol- Becet D, Tsai H, Robine N, Westholm JO, Ladewig E, Dai Q, Okamura K, Flynt AS, Zhang D, Andrews J, Cherbas L, Kaufman TC, Cherbas P, Siepel A, Lai EC. 2014. Diversity of miRNAs, siRNAs, and piRNAs across 25 Drosophila cell lines. Genome Research 24:1236–1250. DOI: https://doi.org/10. 1101/gr.161554.113, PMID: 24985917 Wilk R, Weizman I, Shilo BZ. 1996. trachealess encodes a bHLH- PAS protein that is an inducer of tracheal cell fates in Drosophila. Genes & Development 10:93–102. DOI: https://doi.org/10.1101/gad.10.1.93, PMID: 8557198 Xiong WC, Okano H, Patel NH, Blendy JA, Montell C. 1994. repo encodes a glial- specific homeo domain protein required in the Drosophila nervous system. Genes & Development 8:981–994. DOI: https://doi.org/10.1101/ gad.8.8.981, PMID: 7926782 Zaffran S, Astier M, and DG, Sémériva M. 1997. The held out wings (how) Drosophila gene encodes a putative RNA- binding protein involved in the control of muscular and cardiac activity . Development 124:2087–2098. DOI: https://doi.org/10.1242/dev.124.10.2087 Zirin J, Bosch J, Viswanatha R, Mohr SE, Perrimon N. 2022. State- of- the- art CRISPR for in vivo and cell- based studies in Drosophila. Trends in Genetics 38:437–453. DOI: https://doi.org/10.1016/j.tig.2021.11.006, PMID: 34933779 Zwarts L, Van Eijs F, Callaerts P. 2015. Glia in Drosophila behavior. Journal of Comparative Physiology. A, Neuroethology, Sensory, Neural, and Behavioral Physiology 201:879–893. DOI: https://doi.org/10.1007/ s00359-014-0952-9, PMID: 25336160 Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 24 of 27 Cell Biology | Developmental Biology Tools and resources Appendix 1 Appendix 1—key resources table Reagent type (species) or resource Genetic reagent (D. melanogaster) Designation Source or reference Identifiers Additional information 24B/how- Gal4 Bloomington Drosophila Stock Center Stock # 1767; FLYB:FBti0150063; RRID:BDSC_1767 FlyBase symbol: P(w[+ mW. hs]=GawB) how[24B] Genetic reagent (D. melanogaster) repo- Gal4 Bloomington Drosophila Stock Center Stock # 7415; FLYB:FBti0018692; RRID:BDSC_7415 FlyBase symbol: P(GAL4)repo Genetic reagent (D. melanogaster) btl- Gal4 Bloomington Drosophila Stock Center Stock # 78328; FLYB:FBti019793; RRID:BDSC_78328 FlyBase symbol: P(GAL4- btl.S)3–2 Genetic reagent (D. melanogaster) Act5C- GeneSwitch- Gal4 Bloomington Drosophila Stock Center Stock # 9431; FLYB:FBti0003040,FBti0076553; RRID:BDSC_9431 FlyBase symbol: P(UAS- GFP.S65T) Myo31DF[T2]; P(Act5C(- FRT)GAL4.Switch. PR)3 Genetic reagent (D. melanogaster) Genetic reagent (D. melanogaster) UAS- RasV12 (3) Bloomington Drosophila Stock Center Stock # 64195; FLYB:FBti0012505; RRID:BDSC_64195 FlyBase symbol: P(w[+mC]=UAS- Ras85D. V12)TL1 UAS- RasV12 (2) Bloomington Drosophila Stock Center Stock # 64196; FLYB:FBti0180323; RRID:BDSC_64196 FlyBase symbol: P(w[+mC]=UAS- Ras85D. V12)2 Genetic reagent (D. melanogaster) UAS- RasV12 with RMCE site (3) Bloomington Drosophila Stock Center Stock # 64197; FLYB: FBti0012505, FBti0102080; RRID:BDSC_64197 FlyBase symbol: P(w[+mC]=UAS- Ras85D. V12)TL1, P(w[+mC]=attP.w[+].attP)JB89B Genetic reagent (D. melanogaster) UAS- GFP nuclear Bloomington Drosophila Stock Center Stock # 4775; FLYB: FBti0012492; RRID:BDSC_4775 FlyBase symbol: P( UAS- GFP. nls) 14 Genetic reagent (D. melanogaster) bratdsRNA Bloomington Drosophila Stock Center Stock # 34646; FLYB:FBti0140815; RRID:BDSC_34646 FlyBase symbol: P(y[+t7.7] v[+t1.8]=TRiP. HMS01121)attP2 Genetic reagent (D. melanogaster) UAS- p35 baculovirus death inhibitor Bloomington Drosophila Stock Center Stock # 5072; FLYB:FBti0012594; RRID:BDSC_5072 FlyBase symbol: P(w[+mC]=UAS- p35.H) BH1 Genetic reagent (D. melanogaster) Gal80ts Bloomington Drosophila Stock Center Stock # 7019; FLYB:FBti0027796; RRID:BDSC_7019 FlyBase symbol: P(w[+mC]=tubP- GAL80[ts])20 Cell line (D. melanogaster) Cell line (D. melanogaster) Cell line (D. melanogaster) Cell line (D. melanogaster) Cell line (D. melanogaster) Cell line (D. melanogaster) Cell line (D. melanogaster) Cell line (D. melanogaster) Cell line (D. melanogaster) Cell line (D. melanogaster) Cell line (D. melanogaster) S2 Drosophila Genomics Resource Center Stock # 181; FLYB:FBtc0000181; RRID:CVCL_Z992 Cell line maintained in N. Perrimon lab; FlyBase symbol: S2- DRSC. 24B5- B8 Drosophila Genomics Resource Center Stock # 323; RRID:CVCL_C7G6 24B>Ras attP- L5- CloneB8 24BG1- G1 Drosophila Genomics Resource Center Stock # 324; RRID:CVCL_C7G7 24B>Ras attP- G1- CloneG1 24BG1- F3 Drosophila Genomics Resource Center Stock # 325; RRID:CVCL_C7G8 24B>Ras attP- G1- CloneF3 Rbr6- 2 Rbr6- 4 Drosophila Genomics Resource Center Stock # 326; RRID:CVCL_C7G9 repo>Ras bratdsRNA- L6- Clone2 Drosophila Genomics Resource Center Stock # 327; RRID:CVCL_C7GA repo>Ras bratdsRNA- L6- Clone4 Rbr6- F9 Drosophila Genomics Resource Center Stock # 328; RRID:CVCL_C7GB repo>Ras bratdsRNA- L6- CloneF9 ActGSI- 2 Drosophila Genomics Resource Center Stock # 329; RRID:CVCL_C7GC Act5C- GS>Ras attP- LB- Clone6 ActGSI- 2 Drosophila Genomics Resource Center Stock # 330; RRID:CVCL_C7GD Act5C- GS>Ras attP- GFP- LI- Clone2 ActGSI- 3 Drosophila Genomics Resource Center Stock # 331; RRID:CVCL_C7GE Act5C- GS>Ras attP- GFP- LI- Clone3 Btl3 Drosophila Genomics Resource Center Stock # 332; RRID:CVCL_B3N7 btl>Ras attP- L3 Appendix 1 Continued on next page Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 25 of 27 Cell Biology | Developmental Biology Tools and resources Appendix 1 Continued Reagent type (species) or resource Cell line (D. melanogaster) Cell line (D. melanogaster) Cell line (D. melanogaster) Cell line (D. melanogaster) Cell line (D. melanogaster) Cell line (D. melanogaster) Cell line (D. melanogaster) Designation Source or reference Identifiers Additional information OK6- 3 Rbr6 24BG1 24B5 Btl7 Btl8 Drosophila Genomics Resource Center Stock # 281; RRID:CVCL_XF56 OK6>Ras attP- L3 Drosophila Genomics Resource Center Stock # 282; RRID:CVCL_XF57 repo>Ras bratdsRNA- L6 Drosophila Genomics Resource Center Stock # 283; RRID:CVCL_XF51 24B>Ras attP GFP- L1 Drosophila Genomics Resource Center Stock # 284; RRID:CVCL_XF52 24B>Ras attP- L5 Drosophila Genomics Resource Center Stock # 285; RRID:CVCL_XF53 btl>Ras attP- L7 Drosophila Genomics Resource Center Stock # 286; RRID:CVCL_XF54 btl>Ras attP- L8 OK6- 2 Drosophila Genomics Resource Center Stock # 287; RRID:CVCL_XF55 OK6>Ras attP- L2 Subcloning efficiency DH5- alpha competent cells cell line (E. coli) DH5- alpha Thermo Fisher Cat. # 18265017 Transfected construct (D. melanogaster) Transfected construct (D. melanogaster) Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody pAc5.1B- EGFP Addgene Cat. # 21181; http://n2t.net/addgene: 21181; RRID:Addgene_21181 pAc5.1B- EGFP was a gift from Elisa Izaurralde pCoPURO Addgene Cat. # 17533; http://n2t.net/addgene: 17533; RRID:Addgene_17533 pCoPURO was a gift from Francis Castellino AffiniPure Rabbit Anti- Horseradish Peroxidase (Rabbit polyclonal) Jackson ImmunoResearch Cat. # 323- 005- 021; RRID: AB_2314648 Rabbit polyclonal; IF (1:500) 22C10 (mouse monoclonal) Developmental Studies Hybridoma Bank Cat. # 22C10 RRID: AB_528403. FBgn0259108 Rat- Elav- 7E8A10 anti- elav (rat monoclonal) Developmental Studies Hybridoma Bank Cat. # Rat- Elav- 7E8A10 anti- elav, RRID:AB_528218 8D12 anti- Repo (mouse monoclonal) Developmental Studies Hybridoma Bank Cat. # 8D12 anti- Repo, RRID:AB_528448 1D4 anti- Fasciclin II (mouse monoclonal) Developmental Studies Hybridoma Bank Cat. # 1D4 anti- Fasciclin II, RRID:AB_528235 Guinea pig anti- Twist (guinea pig polyclonal) M.Levine, UC Berkeley, CA 3E8- 3D3 (mouse monoclonal) Developmental Studies Hybridoma Bank Cat. # 3E8- 3D3, RRID:AB_2721944 22C10 was deposited to the DSHB by Benzer, S./Colley, N.; mouse monoclonal; IF (1:100) Rat- Elav- 7E8A10 anti- elav was deposited to the DSHB by Rubin, G.M.; rat monoclonal; IF (1:100) 8D12 anti- Repo was deposited to the DSHB by Goodman, C.; mouse monoclonal; IF (1:100) 1D4 anti- Fasciclin II was deposited to the DSHB by Goodman, C.; mouse monoclonal; IF (1:100) A gift from M. Levine, UC Berkeley, CA; guinea pig polyclonal; IF (1:500) 3E8- 3D3 was deposited to the DSHB by Saide, J.D.; mouse monoclonal; IF (1:100) DCAD2 (rat monoclonal) Developmental Studies Hybridoma Bank Cat. # DCAD2, RRID:AB_528120 DCAD2 was deposited to the DSHB by Uemura, T.; rat, monoclonal; IF (1:100) Rabbit anti- DMef2 (rabbit polyclonal) Mouse anti- H2 (mouse monoclonal) Cy3 AffiniPure Goat Anti- Mouse IgG (H+L) (Goat polyclonal) Cy3 AffiniPure Goat Anti- Rat IgG (H+L) (Goat polyclonal) doi:10.1101/gad.9.6.730 A gift from J. R. Jacobs; rabbit polyclonal; IF (1:500) doi:10.1073/pnas.0436940100 Kurucz et al., 2003; IF (1:10) Jackson ImmunoResearch Cat. # 115- 165- 003; RRID: AB_2338680 Goat polyclonal; IF (1:1000) Jackson ImmunoResearch Cat. # 112- 165- 003; RRID: AB_2338240 Goat polyclonal; IF (1:1000) Appendix 1 Continued on next page Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 26 of 27 Cell Biology | Developmental Biology Tools and resources Appendix 1 Continued Reagent type (species) or resource Antibody Antibody Antibody Designation Source or reference Identifiers Additional information Cy3 AffiniPure Goat Anti- Guinea Pig IgG (H+L) (Goat polyclonal) Jackson ImmunoResearch Cat. # 106- 165- 003; RRID: AB_2337423 Goat polyclonal; IF (1:1000) Cy3 AffiniPure Goat Anti- Rabbit IgG (H+L) (Goat polyclonal) Jackson ImmunoResearch Cat. # 111- 165- 045; RRID: AB_2338003 Goat polyclonal; IF (1:1000) Donkey anti- Rabbit IgG (H+L) Highly Cross- Adsorbed Secondary Antibody, Alexa Fluor 488 (donkey polyclonal) Thermo Fisher Cat. # A- 21206; RRID: AB_2535792 Donkey polyclonal; IF (1:1000) Commercial assay or kit Effectene Transfection Reagent QIAGEN Commercial assay or kit NucleoSpin Plasmid Kit (No Lid) Macherey- Nagel Commercial assay or kit DNeasy Blood & Tissue Kit QIAGEN Cat. # 301425 Cat. # 740499.250 Cat. # 69504 Chemical compound, drug KaryoMAX Colcemid Solution in PBS Chemical compound, drug Schneider′s Insect Medium Chemical compound, drug FBS Chemical compound, drug 0.05% Trypsin–EDTA (1×) Chemical compound, drug Penicillin–streptomycin (10,000 U/ml) Gibco Thermo Fisher Cat. # 15212–012 Sigma- Aldrich Cat. # S0146 Gibco Thermo Fisher Cat. # 26140–079 Gibco Thermo Fisher Cat. # 25300–120 Gibco Thermo Fisher Cat. # 15140122 Chemical compound, drug Chemical compound, drug Mifepristone Invitrogen Thermo Fisher Cat. # H11001 20- Hydroxyecdysone Sigma- Aldrich Cat. # H5142 Chemical compound, drug VECTASHIELD Antifade Mounting Medium With DAPI Software, algorithm GraphPad Prism version 9.5.1 Vector Laboratories Cat. # H1200 https://www.graphpad.com/ RRID:SCR_002798 Software, algorithm Fiji doi:10.1038/nmeth.2019 RRID:SCR_002285 Coleman- Gosser, Hu, Raghuvanshi et al. eLife 2023;12:e85814. DOI: https://doi.org/10.7554/eLife.85814 27 of 27 Cell Biology | Developmental Biology
10.1515_anona-2022-0288
Advances in Nonlinear Analysis 2023; 12: 20220288 Research Article Ali Taheri* and Vahideh Vahidifar Gradient estimates for nonlinear elliptic equations involving the Witten Laplacian on smooth metric measure spaces and implications https://doi.org/10.1515/anona-2022-0288 received September 19, 2022; accepted December 11, 2022 Abstract: This article presents new local and global gradient estimates of Li-Yau type for positive solutions to a class of nonlinear elliptic equations on smooth metric measure spaces involving the Witten Laplacian. The estimates are derived under natural lower bounds on the associated Bakry-Émery Ricci curvature tensor and find utility in proving fairly general Harnack inequalities and Liouville-type theorems to name a few. The results here unify, extend and improve various existing results in the literature for special nonlinea- rities already of huge interest and applications. Some consequences are presented and discussed. Keywords: smooth metric measure spaces, gradient estimates, nonlinear elliptic equations, Witten Laplacian, Harnack inequalities, Li-Yau estimates, Liouville-type theorems MSC 2020: 53C44, 58J60, 58J35, 60J60 1 Introduction ( μ , d ) is a smooth metric measure space by which it is meant that M is a complete Riemannian , Suppose M g manifold of dimension n , f is a smooth potential on M g, is the usual Riemannian metric and vd g is the Riemannain volume measure. In this article, we derive gradient estimates of local and global Li-Yau type along with Harnack inequalities and Liouville-type results for positive smooth solutions u to the nonlinear elliptic equation: 2≥ endowed with a weighted measure μ = − e df v g d Δ f u x ( ) + Σ , x u x [ ( )] = 0, Σ : M (cid:2) × → (cid:2) . (1.1) f f e u u Δ Δ = div f − e [ u ] ∇ = , f u − ⟨∇ ∇ ⟩ is the Witten Laplacian (also known as the weighted or drifting Here Laplacian, or occasionally to emphasise the choice of f , the f -Laplacian) where , div , and Δ are the usual gradient, divergence, and Laplace-Beltrami operators, respectively, associated with the metric g. The Witten Laplacian is a symmetric diffusion operator with respect to the weighted measure μ d v g and arises in many contexts ranging from probability theory, geometry and stochastic processes to quantum field theory and statistical mechanics [2,4,20]. It is the natural generalisation of the Laplace- Beltrami operator to the smooth metric measure space setting and it coincides with the latter precisely when the potential f is a constant. = − e df ∇  * Corresponding author: Ali Taheri, School of Mathematical and Physical Sciences, University of Sussex, Falmer, Brighton, United Kingdom, e-mail: a.taheri@sussex.ac.uk Vahideh Vahidifar: School of Mathematical and Physical Sciences, University of Sussex, Falmer, Brighton, United Kingdom, e-mail: v.vahidifar@sussex.ac.uk Open Access. © 2023 the author(s), published by De Gruyter. This work is licensed under the Creative Commons Attribution 4.0 International License. 2  Ali Taheri and Vahideh Vahidifar Σ = Σ ,[ in (1.1) is a sufficiently smooth multi-variable function depending on both The nonlinearity x u ] and the independent variable (solution) u. In what follows, in order to better the spatial variable x M∈ orient the reader and showcase the results, we discuss various examples of such nonlinearities arising from different contexts, e.g., from conformal geometry and mathematical physics, each presenting a different phenomenon whilst depicting a corresponding singular or regular behaviour. As for the curvature properties of the triple M g curvature tensor (see [2,3,30]), by writing, for m n≥ , ( , , d μ ), we introduce the generalised Bakry-Émery Ricci Ric m( ) g f = Ric g ( ) + Hess f ( ) − f f ∇ ⊗ ∇ m n − , (1.2) g( ) Ric denotes the Riemannain Ricci curvature tensor of g, Hess( ) stands for the Hessian of f , and where m n≥ is a fixed constant. For the sake of clarity, we point out that in the event m n= , by convention, f is n( ) , whilst, we also allow for m = ∞ in which case only allowed to be a constant, resulting in g g ( ) Ric f by formally passing to the limit in (1.2) we write g ( ) = Ric Ric According to the weighted Bochner-Weitzenböck formula, for every smooth function u on M, we have, = ∞Ric f Hess Ric f ( ) g ( ) g ( ) ≔ + f . f Δ f ∣ ∇ u 2 ∣ = Hess ∣ 2 u ( )∣ u + ⟨∇ ∇ , Δ f u ⟩ + Ric f ( , u ∇ ∇ u ) . (1.3) 1 2 Hence, by an application of the Cauchy-Schwartz inequality giving u Δ the identity − ⟨∇ ∇ ⟩, it is evident that u f = Δ Δ u u f , ≤ n Hess ∣ u ( )∣ and upon recalling Hess ∣ 2 u ( )∣ ≥ 2 ) , ( Δ u n 2 ) ( Δ u n + 2 , u f ⟨∇ ∇ ⟩ m n − ( ≥ Δ u f m 2 ) , and so it follows from (1.3) and (1.4) that 1 2 Δ f ∣ ∇ u 2 ∣ u − ⟨∇ ∇ , Δ f u ⟩ ≥ 1 m ( Δ f u ) 2 + Ric m ( f , u ∇ ∇ u ) . (1.4) (1.5) In particular, subject to a curvature lower bound curvature-dimension condition m Ric f ) (cf. [2–4,44]). CD ,k( m g ( ) ≥ g k , the operator L Δf= is seen to satisfy the Our principal objective in this article is to develop local and global gradient estimates of Li-Yau type and Harnack inequalities for positive smooth solutions to (1.1). It is well known that these estimates and inequalities form the basis for deriving various qualitative properties of solutions and are thus of great significance and utility (see, e.g., [13,22,26,28,35,49]). Such properties include (but are not restricted to) Hölder regularity and higher order differentiability, sharp spectral asymptotics and bounds, heat kernel bounds, Liouville-type results, and many more (cf. [1,6,8,18,20,29,33,37–39,42,43,53]). Whilst in proving gradient estimates, one typically works with an explicit nonlinearity with a specific structure (of singularity, regularity, decay, and growth), in this article, we keep the analysis and discussion on a fairly general level without confining to specific examples to firstly provide a unified treatment of the estimates and secondly to more clearly see how the structure and form of the nonlinearity influences the estimates and subsequent results. As such our approach and analysis largely unify, extend, and at places, improve various existing results in the literature for specific choices of nonlinearities (see the following for more). Gradient estimates for positive solutions to (1.1) in the special case of the nonlinearity being a super- position of a logarithmic and a linear term with variable coefficients: Δ f u x ( ) + p x u x ( ) ( ) log u x ( ) + q x u x ( ) ( ) = 0, (1.6) along with its parabolic counterpart have been the subject of extensive studies (see, e.g., Ma [31], Ruan [34], Wu [46], and Yang [50] and the references therein). The interest in such problems originates from its natural links with gradient Ricci solitons. Recall that a Riemannian manifold M g, ) is said to be a gradient Ricci soliton iff there exists a smooth function f on M and a constant λ (cid:2)∈ ( such that (cf. [11,15,30]) Ric g f ( ) = Ric g ( ) + Hess f ( ) = λg . (1.7) A gradient Ricci soliton can be expanding (λ 0< ). The notion is a generalisation of an Einstein manifold and has a fundamental role in the analysis of singularities of the 0= ) or shrinking (λ 0> ), steady (λ Gradient estimates for nonlinear elliptic equations  3 Ricci flow [23,53]. Taking trace from both sides of (1.7) and using the contracted Bianchi identity lead one to a simple form of (1.6) with constant coefficients: u Δ for suitable constant A0 and + u (see [31] for details). Other types of equations closely relating to (1.6) including: nλ u ) A 0( 2 λu log = − u e f= Δ f u x ( ) + p a x u x ( ) ( )∣ log u b x ∣ ( ) = 0, for real exponents a b, or more generally for a nonlinear function γ = γ s( ) on (cid:2): Δ f u x ( ) + p a x u x γ ( ) ( ( ) log u x )( ) + q b x u x ( ) ( ) = 0, (1.8) (1.9) have been studied in detail in [9,17,40,41,46,47]. x u ( ) = Δ Yamabe type equations u are also of form (1.1) with a power-like nonlinearity. 0 Bidaut-Véron and Véron [5] studied the equation u Δ on a compact manifold and under suitable conditions on the Ricci tensor, n and s, q showed that it only admits constant solutions. Gidas and Spruck [19] considered x u ( ) + = + + + 0 q q p u u s s Δ u x ( ) + p s x u x ( ) ( ) = 0, 1 s ≤ < ( n + 2 ) ( / n − 2 , ) (1.10) 0 Ric g ( ) ≥ g ( ) ≥ 3= in (1.10) is related to Yang-Mills equation (cf. Cafarelli et al. [9]) and the case s any non-negative solution to this equation must be zero. Yang [51] and showed that when 0 . showed that the same equation with constant Note that the case s 0< is related to the steady states of the thin films equation (cf. Guo and Wei [21]). For more related results, see Brandolini et al. [7], Li [25], Li et al. [27], and Zhang [52], and for a more detailed account on the Yamabe problem in geometry, see [24,32]. The natural form of Yamabe equation in the setting of smooth metric measure spaces is 0< admits no positive solution when 0 p > and s Ric Δ f u x ( ) + p s x u x ( ) ( ) + q x u x ( ) ( ) = 0. (1.11) For gradient estimates, Harnack inequalities, and other counterparts of the aforementioned results, we refer the reader to Case [12], Wu [47], and Zhang and Ma [54]. A more general form of Yamabe equation is the Einstein-scalar field Lichnerowicz equation (see, e.g., Choquet-Bruhat [14], Chow et al. [15], and Zhang [53]). When the underlying manifold has dimension n 3≥ , 2 2 0 this takes the form u − , n ) ( ) ( = / . The Einstein-scalar field Lichnero- 0 while when n q wicz equation in the setting of smooth metric measure spaces can be further generalised and written as follows: x u p ( ) Δ 2= , this takes the form u = with α = 2 u − x e r ( ) + + x u r ( ) 2 u x e ( ) n ( x ( ) and β + = 2 ) ( / x u ( ) + p 2 ) − + + + − Δ q n n 3 α β and Δ f u + p x u ( ) α + q x u ( ) β + r x u ( ) log u + h x u ( ) = 0, Δ f u + p x e ( ) 2 u + q 2 u − x e ( ) + r x ( ) = 0. (1.12) (1.13) For gradient estimates, Harnack inequalities, and Liouville-type results in this and related contexts, see Dung et al. [16], Song and Zhao [36], Taheri [40,41], and Wu [48] and the references therein. Let us end this introduction by briefly describing the plan of the article. In Section 2, we present the main results of the article, namely, a local and global gradient estimate of Li-Yau type for equation (1.1), followed by both local and global Harnack inequalities and a general Liouville-type result. Sections 3, 4, and 5 are then devoted to the detailed proofs. 1.1 Notation , we denote by d By fixing a base point p M∈ to the metric g and by r r xp( ) closed geodesic ball of radius R we often abbreviate and write d x( ), r x( ), or so s and s − with s s 0≤− 0≥+ s ++ = = . = d xp( ) the geodesic radial variable with origin at p. We denote by 0> the Riemannian distance between x and p with respect the centred at p. When the choice of the point p is clear from the context, and =+ R(cid:2) , respectively. We write s smax , 0( ) smin , 0( ) and s =− ) ⊂ R( M (cid:2) p 4  Ali Taheri and Vahideh Vahidifar For given multi-variable function Σ ,[ and we reserve the notation Σx for the function function of x; e.g., below we frequently use Σx∇ and Δ Σf = Σ x. , we denote its partial derivatives by subscripts, e.g., Σx, Σu, x u ] obtained by freezing the argument u and viewing it as a uΣ ,[ ] ⋅ For the sake of reader’s convenience, we recall that in local coordinates xi( ), we have the following formulae for the Laplace-Beltrami operator, Riemann, and Ricci curvature tensors, respectively: and and Note that here Δ = 1 g ∣ ∣ ∂ x ∂ i ⎛ ⎜ ⎝ g g ∣ ∣ ij ∂ x ∂ j , ⎞ ⎟ ⎠ [ Rm ℓ g ( )] = ijk Γ ∂ ℓ jk x ∂ i − ℓ Γ ∂ ik x ∂ j + p Γ Γ jk ℓ ip − p ℓ Γ Γ , jp ik [ Ric g ( )] = ij Γ ∂ k ij x ∂ k − Γ ∂ ℓ j ℓ x ∂ i + k Γ Γ ij ℓ k ℓ − k ℓ Γ Γ . ik j ℓ Γ k ij = k ℓ g 1 2 ℓ g ∂ j x ∂ i ⎛ ⎜ ⎝ + ℓ g ∂ i x ∂ j − g ∂ ij x ∂ ℓ , ⎞ ⎟ ⎠ (1.14) (1.15) (1.16) (1.17) are the Christoffel symbols and gij, g∣ ∣, and g ij the components of the inverse of the metric tensor g. = 1 − g ) ( ij are, respectively, the components, determinant, and 2 The main results In this section, we present the main results of the article. The proofs are delegated to the subsequent sections. We emphasise that throughout the article, the curvature lower bounds are expressed in the < ∞. As the estimates here are form of Li-Yau type, it is well-known that a lower bound on and m a suitable constant n m≤ Ric is not sufficient for the purpose. with k m( ) g f kg1 ) Ric 0≥ ≥ − gf ( ) m − ( 2.1 A local and a global Li-Yau type gradient estimate for (1.1) Theorem 2.1. Let M g , , d μ ( =(cid:2) m( ) kg1 g in ) ( ≥ − − Ric 2 R f R2(cid:2) . Then for every μ (1.1) in m (cid:2) p ) 2 ( R 1> and ε for suitable constants m n≥ and k ∈ and every x ∈ (cid:2) , R 0, 1 ) ( ) be a complete smooth metric measure space with μ d = − e v and assume that g 0≥ . Let u be a positive solution to df 2 ∣ 2 u ∣ ∇ μu + Σ , x u ] [ u ≤ mμ 2 2 R ⎡ ( ⎢ ⎣ c 2 + ( m − 1 c ) ( 1 1 + R k ) + 2 2 c 1 ) + 2 2 mc μ 1 4 μ ( − 1 ) ⎤ ⎥ ⎦ + + m 2 ⎡ ⎢ ( ⎣ 1 2 2 mμ A Σ ε μ )( − − 2 27 mμ 4 ε μ ( − 4 B Σ 2 1 ) + ⎡ ⎢ ⎣ 1 3 / ⎤ ⎥ ⎦ 2 1 ) + 2 μ C Σ 1 2 / ⎤ ⎥ ⎦ mμ 2 sup (cid:2) R2 ⎧ ⎨ ⎩ ( u Σ , x u ] [ u − u Σ , x u [ ]) ⎫ + ⎬ ⎭ , where the quantities ΣA , ΣB , and ΣC are given by 2 ( m − 1 ) ku ( + − Σ , x u ] [ A Σ = sup (cid:2) R2 ⎧ ⎨ ⎩ Σ , x u ] [ u − μu 2 Σ u + 2 u , x u [ uu ]) ⎫ + ⎬ ⎭ , (2.1) (2.2) Gradient estimates for nonlinear elliptic equations  5 B Σ = sup (cid:2) 2 R ⎧ ⎨ ⎩ Σ ∣ x , x u ] [ − Σ μu u , x u [ xu ]∣ ⎫ ⎬ ⎭ , C Σ = ( − Δ Σ f sup (cid:2) 2 R ⎧ ⎨ ⎩ x , x u [ u ]) ⎫ + ⎬ ⎭ . (2.3) The aforementioned local estimate has a global counterpart subject to the prescribed bounds in the theorem being global. The proof follows by passing to the limit R → ∞ in (2.1) and taking into account the vanishing of certain terms as a result of the bounds being global and the relevant constants being independent of R. The precise formulation of this is given in the following theorem. Theorem 2.2. Let M g m ( − − and ε , on M, where m n≥ 0, 1 and every x M∈ ) kg1 ) ( ∈ , d μ ( and k , ) be a complete smooth metric measure space with μ m( ) ≥ gf 0≥ . Assume u is a positive solution to (1.1) on M. Then for every μ 1> = − e Ric and df v g d u ∣ ∇ μu 2 ∣ 2 + Σ , x u ] [ u ≤ m 2 ⎡ ⎢ ( ⎣ 1 2 2 mμ A Σ ε μ )( − − 2 27 mμ 4 ε μ ( − 4 B Σ 2 1 ) + ⎡ ⎢ ⎣ 1 3 / ⎤ ⎥ ⎦ 2 1 ) + 2 μ C Σ 1 2 / ⎤ ⎥ ⎦ (2.4) + mμ 2 sup M ⎧ ⎨ ⎩ ( u Σ , x u ] [ u − u Σ , x u [ ]) ⎫ + ⎬ ⎭ . ΣA , ΣB , and ΣC are as in (2.2) and (2.3) in Theorem 2.1, except that now the supremums are taken over all Here, of M. 2.2 A local and a global elliptic Harnack inequality for (1.1) Theorem 2.3. Under the assumptions of Theorem 2.1 and the Bakry-Émery curvature bound ( − R2(cid:2) , for any positive solution u to (1.1), and for any x x, 2 ∈ (cid:2) , we have kg1 ) on m − R 1 The positive constant (cid:3) can be explicitly expressed in terms of the local bounds as follows: u x ( 2 ) ≤ 2 R e (cid:3) u x . ( ) 1 Ric m( ) ≥ gf (2.5) (cid:3) = 2 mμ c [( 2 + ( m + + 2 m μ mμ { / sup (cid:2) 2 R {( u 2 Σ mμ − 2 1 1 c ) ( 1 2 A [( / Σ , x u ] [ u 1 + R k ) + − − ε μ )( − Σ , x u [ ]) + ) 2 2 c 1 2 1 ) ] + 2 u / ( + 2 2 mc μ 1 2 mμ 27 [ 2 2 R ) 1 3 / − μ B 4 ( ( / 4 / Σ inf Σ , {( [ (cid:2) 4 ε μ ( ( x u 1 ))] ( / 2 1 ) )] , u } − / ]) − R )} − μ + 2 μ C } Σ 1 2 / (2.6) where ΣA , ΣB , and ΣC are as in (2.2) and (2.3) in Theorem 2.1. In particular, from (2.5), we have sup (cid:2) R u ≤ R2 e (cid:3) u . inf (cid:2) R (2.7) For the global version, we can use a similar argument utilising the global bounds in Theorem 2.2 and have now being its global version from (2.4). The the counterpart of (2.5) with d x x,1 2 precise formulation is given below. ) replacing R2 and 0(cid:3) > ( Theorem 2.4. Under the assumptions of Theorem 2.2 and the Bakry-Émery curvature bound ( − on M, for any positive solution u to (1.1), and any x x M , we have kg1 ) m − 2 ∈ ,1 u x ( ) 2 The positive constant (cid:3) can be explicitly expressed in terms of the global bounds as follows: u x . ( ) 1 ≤ e , d x x ( 1 2 (cid:3) ) (cid:3) = + 2 m μ mμ { / 2 sup M mμ {( 2 − 2 A Σ Σ 1 [( / , x u ] [ u u ε μ )( − Σ , x u [ − 2 1 ) ] ]) + / 27 [ + 2 u )} ( 2 mμ − μ 1 3 / B / 4 4 ε μ ( ( − Σ inf Σ , x u {( [ M 2 1 ) )] u } / , ]) − + 2 μ C } Σ 1 2 / Ric m( ) ≥ gf (2.8) (2.9) where ΣA , ΣB , and ΣC are as in Theorem 2.2. 6  Ali Taheri and Vahideh Vahidifar 2.3 A Liouville-type theorem and some applications ( Theorem 2.5. Let M g Let u be a positive solution to 0 u ≥ ) be a smooth metric measure space with μ 0 = − d e Σ uΣ [ ] ≥ , u = . Assume that 1> . Then u must be a constant. In particular, for some μ u [ ] u [ ] , d + + 0 Δ μ u Σ Σ Σ u u[ ] , f m( ) ≥ df v satisfying g Ric g f 2 0 Σ ≤ , and μu u u [ ] u[ ] − 0 uΣ [ ] = . 0 Σuu in M. u [ ] − The proof of this theorem follows from the gradient estimates established above and is presented in Section 5. We end this section by giving two stark applications of the aforementioned theorem. To this end, consider first a superposition of power-like nonlinearities with real coefficients jp and exponents aj for in the form 1 ≤ ≤ j N A direct calculation gives Σ N [ ] ∑= u 1 = j p j u a j . u Σ u [ ] u − Σ u [ ] = N ∑ 1 j = p ( a j j − 1 ) u a j , (2.10) (2.11) μu 2 Σ u [ ] uu − u Σ u [ ] u + Σ u [ ] = N ∑ 1 j = μ a a [ j j j p ( − 1 ) − p a j j + p j u ] a j = N ∑ 1 j = p [ ( a j j − 1 )( μa j − 1 )] u a j . (2.12) 2 uu Σ u [ ] and jp ≥ , we have Evidently for the range 1> suitably). Theorem 2.5 now leads to the following conclu- μu u u [ ] sion extending earlier results on Yamabe type problems to more general nonlinearities (cf. [16,19, 48,51]). Further applications and results in this direction will be discussed in a forthcoming article (see also [40,41]). ( ) ≥ , whilst subject to a uΣ ≥ (by choosing μ j ≤ , we have u u [ ] u u u[ ] 0 0 u [ ] − + − ≤ 0 0 Σ Σ Σ Σ 1 Theorem 2.6. Let M g , d Let u be a positive smooth solution to the equation: μ ( , ) be a complete smooth metric measure space with μ d If jp ≥ 0 and a 1 j ≤ for 1 ≤ ≤ , then u must be a constant. j N Δ f u N p∑+ 1 j = a j u j = 0. = − e df v g and Ric m( ) ≥ g f 0 . (2.13) Remark 2.7. Note that a constant solution to (1.1) must be a zero of Σ. Thus, if Σ has no positive zeros, then the above Liouville theorem becomes a non-existence result. In Theorem 2.6 and for (2.10), this happens when for at least one 1 ≤ ≤ . j N 0 jp > As another application, again relating to the discussions in Section 1, consider a superposition of a in the logarithmic and a power-like nonlinearity with real coefficients p and q, exponent s and γ form 2 (cid:2)( ∈ C ) A straightforward calculation then gives u ″ + p ] [( − μuγ q [ ( ′ + 1 )( uγ μs 1 ) − − μ s Σ u [ ] u s 1 u )] Σ = u [ ] 1 ) − − + . The following theorem now directly results from Theorem 2.5. and μu u [ ] u log u [ ] uγ q ( ( ′ − + p u u u Σ Σ uu s ) s 2 Σ u [ ] = p uγ ( log u ) + q u .s (2.14) Σ u [ ] = Theorem 2.8. Let M g , d Let u be a positive smooth solution to the equation: μ ( , ) be a complete smooth metric measure space with μ d = − e df v g and Ric m( ) ≥ g f 0 . Assume that some μ 1> (with , p q ≥ , s 1 μ 0 < Δ f u + p uγ ( log u ) + s q u = 0. (2.15) 1≤ and that along the solution u, we have γ 1 < / < < ). Then, u must be a constant. 0 if s s 1 0≥ , γ ′ ≤ , and μγ 0 ″ + ( μ − γ1 ) ′ ≥ 0 for Gradient estimates for nonlinear elliptic equations  7 3 Proof of the Li-Yau type gradient estimate in Theorem 2.1 This section is devoted to the proof of the main estimate for the positive solutions of (1.1) in its local form. In its global form, the estimate, as seen, then follows by passing to the limit R → 1. As this requires a number of technical lemmas and tools, we pause briefly, to present these necessary tools and results in the next subsection, before moving on to the proof of the local estimate in Theorem 2.1 in the following subsection. 3.1 Some technical lemmas and identities Lemma 3.1. Let u be a positive solution to equation (1.1) and let h log= u . Then h satisfies the equation: Δ f h ∣ + ∇ h 2 ∣ + h − e Σ , x e [ h ] = 0. Proof. An easy calculation gives h 2 Σ , x u u u ∣ ] [ − Δ ∇ = ∇ / and h u u ) giving the desired conclusion. ∣ / − ∇ / u ( 2 = ( Δ u u ) ∣ / − ∇ / u 2 ∣ 2 u . Hence, Δ f h = ( Δ f u u ) ∣ / − ∇ / u 2 ∣ Lemma 3.2. Let u be a positive solution to (1.1), h log= u , and let H be defined by where μ 1≥ is an arbitrary constant. Then H satisfies the equation H ∣ = ∇ h 2 ∣ + − μe h Σ , x e [ h ] , (3.1) 2 u = □ (3.2) Δ f H 2 2 ∣ = ∇ h 2 ∣ + 2 2 , h f ⟨∇ ∇ ⟩ m n − + μ Δ f h − e ( Σ , ) , − ⟨∇ ∇ ⟩ + h H 2 2 Ric m ( f ∇ ∇ − h h ) , 2 ( μ − 1 ) h − e Σ ∣ ∇ h 2 ∣ + 2 ( μ − 1 ) − e h , Σ ⟨∇ ∇ ⟩ h (3.3) where we have abbreviated the arguments of Σ and its derivatives. Proof. Referring to the formulation of H in (3.2) an application of Δf to both sides of the equation gives Δ f H = Δ f ∣ ∇ h 2 ∣ + μ Δ f h − e ( Σ , x e [ h ]) . Furthermore, referring to (3.1) and again to (3.2), it is evident that we have the relation: Δ f h (∣ = − ∇ h 2 ∣ + h − e Σ , x e [ h ]) = − ( H − ( μ − 1 ) h − e Σ , x e [ h ]) . (3.4) (3.5) Now as for the first term on the right-hand side of (3.4), by the generalised Bochner-Weitzenböck formula (as applied to h), we have Δ f ∣ ∇ h 2 ∣ 2 2 ∣ = ∇ h 2 ∣ + ⟨∇ ∇ , Δ h 2 h ⟩ + f 2 Ric m ( f ∇ ∇ + h h ) , 2 2 , h f ⟨∇ ∇ ⟩ m n − . (3.6) Hence, by substituting (3.6) into (3.4) and making note of (3.5), we have after a basic differentiation, Δ f H 2 2 ∣ = ∇ h 2 ∣ 2 2 ∣ = ∇ h 2 ∣ , + ⟨∇ ∇ − + H h 2 ( ( μ − 1 ) h − e Σ , x e [ h ]) ⟩ + 2 Ric m ( f ∇ ∇ + h h ) , + μ Δ f f − e ( Σ , x e [ h ]) , − ⟨∇ ∇ ⟩ − h H 2 2 ( μ − 1 ) h − e Σ , x e [ h ]∣ ∇ h 2 ∣ + 2 ( μ − 1 ) − e h h ⟨∇ ∇ ] ⟩ + 2 Ric m ( f , h ∇ ∇ h ) (3.7) + 2 2 , h f ⟨∇ ∇ ⟩ m n − + μ Δ f h − e ( Σ , x e [ h ]) , which upon a rearrangement of terms gives the desired identity. □ Lemma 3.3. Let u be a positive solution to (1.1), h , we have Ric kg1 ) ≥ − m m( ) g f − ( log= u and let H be as defined by (3.2). Then, if 2 2 , h f ⟨∇ ∇ ⟩ m n − h , Σ , x e [ 8  Ali Taheri and Vahideh Vahidifar Δ f H ≥ h ) ( Δ 2 f m 2 − ⟨∇ − , h e [ h Σ x − μ Σ ] ⟩ + xu − μe h x Δ Σ . f 2 , − ⟨∇ ∇ ⟩ + h H 2 h − e [ Σ − Σ H ] u + h − e [ Σ − Σ u + μe h Σ uu − 2 ( m − 1 k ) ]∣ ∇ h 2 ∣ (3.8) Proof. By virtue of the bound Ric m( ) g f ≥ − ( m − kg1 ) , it follows upon recalling the identity in Lemma 3.2 that ( Δ f H ≥ 2 + 2 Δ h ) f m − 2 μ e ⟨∇ h , Σ ∇ ⟩ + − μe h Δ Σ. f , − ⟨∇ ∇ ⟩ − h H 2 2 ( μ − 1 ) h − e Σ ∣ ∇ h 2 ∣ + 2 ( μ − 1 ) − e h , Σ ⟨∇ ∇ ⟩ − h 2 ( m − 1 k h ) ∣ ∇ 2 ∣ + μ ΣΔ f h − e (3.9) Note that in concluding the aforementioned inequality, specifically, the first term on the right-hand side, we have made use of the basic inequalities: 2 ∣ ∇ h 2 ∣ + 2 , h f ⟨∇ ∇ ⟩ m n − ≥ 2 ) ( Δ h n + 2 , h f ⟨∇ ∇ ⟩ m n − ( ≥ Δ h f m 2 ) . (3.10) Let us now proceed by attending to some useful and straightforward calculations relating to the non- linear term Σ = Σ , ( x e h ) . Evidently and thus by moving on to the Laplacian, we can write ∇ Σ , x e [ h ] = Σ x , x e [ h ] + h e Σ , x e [ u h ] ∇ h , ΔΣ , x e [ h ] = div Σ ( x , x e [ h ] + h e Σ , x e [ u h ] ∇ h ) . (3.11) (3.12) It is convenient to do the calculations in local coordinates, and so we proceed by writing ΔΣ , ( x e h ) = = n ∑ 1 i = n ∑ 1 i = ∂ x ∂ i ( Σ x i , x e [ h ] + e h Σ h x e h [ i u ] , ) { Σ x x i i , x e [ h ] + Σ , x e [ x u i h ]( e h ) i + + e h Σ h x e h [ ii u ] , } . h e h h Σ x e h [ i i u ] , + e h ( Σ , x e [ x u i h ] + Σ , x e [ uu h ]( e h ) i ) h i (3.13) Abbreviating the arguments x e, [ h ] of Σ for convenience and rewriting the above, we have ΔΣ ΔΣ = x = x ΔΣ + + e h Σ , ⟨ xu h 2 Σ , e ⟨ xu h ∇ ⟩ + h ∇ ⟩ + e h h ∣ ∇ h ∣ ∇ e 2 ∣ h Σ + u 2 Σ ∣ ( u e + h ⟨ e Σ , xu h Σ ) uu h ∇ ⟩ + h + e e 2 h ∣ ∇ Σ Δ . h u h 2 ∣ Σ uu + h e Σ Δ u h (3.14) As a result, the above on substitution gives Δ Σ ΔΣ = f − ⟨∇ ∇ ⟩ = , Σ f ΔΣ f − ⟨∇ , Σ ( x + h e Σ h ) ∇ ⟩ u = = = ΔΣ ΔΣ h f e Σ , Σ f − ⟨∇ x ⟩ − x x , Σ f − ⟨∇ ∇ ⟩ + h h ∇ ⟩ + ⟨ , h ⟨∇ ∇ ⟩ u h Σ , 2 h e ∇ ⟩ + ⟨ xu 2 h Σ e h ∣ ( ∣ ∇ Σ , xu 2 e + u x Δ Σ f e + h ∣ ∇ h Σ e h 2 ∣ ( ) uu Σ u + + h e h e Σ ) uu Σ Δ . h u f + h e Σ Δ u h − h e Σ , h ⟨∇ ∇ ⟩ u f Moreover, for the sake of future reference, we also note that Δ f h − e = = − Δ e h f − e h , − ⟨∇ ∇ ⟩ h , h ⟨∇ ∇ ⟩ h − , h f e ⟨∇ ∇ ⟩ − e − e 2 ∣ + h f h div ( − h − Δ e − e ( h = − = − Δ f h ) ∇ + h − e ∣ ∇ h ∣ − ∇ + h h 2 ∣ ) . (3.15) (3.16) Now returning to the inequality (3.9) and upon substituting from (3.11), (3.15), and (3.16), we obtain Gradient estimates for nonlinear elliptic equations  9 Δ f H ≥ 2 Δ ( f 2 h m / ) , − ⟨∇ ∇ ⟩ − h H 2 2 ( μ − 1 ) h − e Σ ∣ ∇ h 2 ∣ + 2 ( μ − 1 ) − e h ⟨∇ h , Σ x + h e Σ h ∇ ⟩ u − 2 ( m − 1 k h ) ∣ ∇ 2 ∣ − − μe h Σ Δ ( f h ∣ − ∇ h 2 ∣ ) − + − μe h Δ Σ [ f x + 2 e h ⟨ Σ , xu h ∇ ⟩ + h e ( Σ u + e h Σ − 2 μe h ⟨∇ h , Σ x + h e Σ h ∇ ⟩ u (3.17) )∣ ∇ h uu 2 ∣ + h e Σ Δ u f h ] , or upon rearranging Δ f H ≥ 2 Δ ( f 2 h m / ) 2 h H , − ⟨∇ ∇ ⟩ − h − − μ e ( 1 k h ) ∣ ∇ 2 ∣ 2 ( m − 2 ( μ − 1 ) h − e Σ ∣ ∇ h 2 ∣ + μ − 1 ) − e h ⟨∇ h , Σ ⟩ + x 2 ( μ − 1 Σ ) ∣ ∇ u h 2 ∣ Σ − Σ Δ ) u f h − μe h Σ − − 2 μe h ⟨∇ h , Σ ⟩ − x 2 Σ μ ∣ ∇ u h 2 ∣ (3.18) − + − μe h x Δ Σ f + 2 μ h ⟨∇ , Σ ⟩ + xu μ ( Σ u + e Σ )∣ ∇ uu 2 ( 2 ∣ h ∣ ∇ 2 ∣ h . + h Next by recalling (3.2) and (3.5), we can write Δ μ h f = − μ H ( − ( μ − 1 ) h − e Σ , x e [ h ]) = − H [ + ( μ − 1 )∣ ∇ h 2 ∣ ] , (3.19) and therefore by substituting the latter back in (3.18) and rearranging terms, it follows that Δ f H ≥ 2 Δ ( f 2 h m / ) , − ⟨∇ ∇ ⟩ + h H 2 h − e Σ ( − Σ )[ u H ( μ − 1 )∣ ∇ h 2 ∣ ] μ [ + − − 2 e 2 ( h ⟨∇ − − 1 ) h − e Σ − h , Σ ⟩ + x − μe 2Σ h − u Δ Σ f 2 ( x + 2 μ h ⟨∇ , Σ ⟩ xu . + 2 ∣ m − 1 k ) ]∣ ∇ h + − μe h Σ ( + μ Σ u + μe h Σ )∣ ∇ h uu 2 ∣ (3.20) Finally taking into account the necessary cancellations and by a further rearrangement of terms, we obtain Δ f H ≥ 2 Δ ( f 2 h m / ) , − ⟨∇ ∇ ⟩ + h H 2 h − e Σ ( − Σ ) u H + h − e [ Σ − Σ u + μe h Σ uu − 2 ( m − 1 k ) ]∣ ∇ h 2 − ⟨∇ − , h e h Σ x − μ Σ ⟩ + xu − μe h x Δ Σ , f which is the desired conclusion. 2 ∣ (3.21) □ The following lemma will also be used in the course of the proof of the local estimate in the next subsection. Lemma 3.4. Suppose a b z 0, 1 any ε ) , we have ( , , ∈ (cid:2)∈ , c y, 0> , and μ 1> are arbitrary constants such that y − μz > . Then for 0 ( y − 2 z ) − a y y ( − μz ) − by − c y ≥ ( y − μz ) 2 2 / μ − 2 2 a μ y ( − μz 8 ) [ ( / μ − 1 )] − ( 3 4 ) / c 4 3 / μ [ 2 / ( 4 ε μ ( − 2 1 ) )] 1 3 / (3.22) − 2 μ b ( 2 4 1 ) [ ( / − ε μ )( − 2 1 ) ] . Proof. Starting from the expression on the left-hand side in (3.22), we can write for any δ ε, considerations: by basic ( y − 2 z ) − a y y ( − μz ) − by − c y = ( 1 ε − − 2 δ y ) − ( 2 − εμ yz ) + 2 z + ( εy − a y )( y − μz ) + 2 δy − by − c y = ( 1 μ ε / − / 2 )( y − 2 μz ) + ( 1 + 2 δy − by − c y . ε − − − / + / μ δ ε 1 2 2 ) y + ( 1 − μ + 2 εμ / 2 z ) 2 + εy ( − a y )( y − μz ) (3.23) In particular, setting δ ε 2 and εμ / = 1 2 ε ) ( / − = 1 2 μ 0 μ μ 0 + − > , we can deduce from (3.23) that μ1 ( = / − 2 0 2 / = and ε 1 2 2 1 μ ( ) and so by making note of the inequality εy 2 = − / − μ / − 2 ( 1 ) 1 ) − = μ μ 2 / 2 2 gives a y − 1 ε a ≥ − / μ − − − / + with δ ε42 ( 1 ) ( y − 2 z ) − a y y ( − μz ) − by − c y ≥ ( y − μz ) 2 2 / μ − 2 2 a λ y ( − μz 8 ) [ ( / μ − 1 )] + ( μ − 1 ) 2 2 y μ / 2 − by Next, considering the last three terms only we can write, for any ε ∈ ( 0, 1 ) , − c y . (3.24) 10  Ali Taheri and Vahideh Vahidifar ( μ − 1 ) 2 2 y μ / 2 − by − c y ≥ ( μ − 1 ) − 1 ) 2 2 y μ / 2 − 2 μ b ( 2 4 1 ) [ ( / − ε μ )( − 2 1 ) ] − c y 2 2 2 y μ / 2 2 y μ / 1 ) − 2 ( 1 − ( ε μ )( 2 − 2 μ b ≥ ε μ ( − ( ≥ − / 3 4 ) c 4 3 / μ [ 2 / ( 4 ε μ ( − 4 1 ) [ ( / 2 1 ) )] − 1 3 / ε μ )( − ( 2 μ b − 2 2 1 ) ] − c y 4 1 ) [ ( / − ε μ )( − 2 1 ) ] , (3.25) where above we have made use of 2 3 4 4 μ ) − Substituting back into (3.24) gives the desired inequality. 1 ( − 2 1 ) )] ε μ )( 1 3 / ( ≥ − / 4 3 / μ [ − − to deduce the first and last 4 1 ) [ ( / c y ≥ − ε μ ( by 1 ) − − c / ( ( 2 2 2 y μ / 2 2 μ b 2 − 2 1 ε μ ) ] )( inequalities, and ε μ ( y1 2 2 )− / respectively. □ 3.2 Proof of Theorem 2.1 This is based on the estimate established in Lemma 3.3 and a localisation argument. To carry out the localisation and the relevant estimates, we proceed by first constructing a suitable cut-off function. Towards ¯( ) this end, we begin by introducing a function ψ ψ t (i) ψ¯ is of class 2[ . ¯ (ii) 1 ψ t ≤ for ( ) satisfying the following conditions: 0, )∞C t 0 ≤ < ∞ with = ≤ 0 ¯ ¯ ψ t ( ) = 1 0 ⎧ ⎨ ⎩ t t ≤ ≥ 1, 2. (3.26) (iii) ψ¯ is non-increasing (i.e., ψ¯ ′ ≤ ), and in addition, for suitable constants c c, 0 1 2 > , its first, and second- 0 order derivatives satisfy the bounds c − ≤ 1 ¯ ′ ψ ¯ ψ ≤ 0 and ¯ ψ ″ ≥ − c 2 . Next pick and fix a reference point p in M and with r = r xp( ) set ψ x ( ) = ¯ ⎛ ψ ⎝ r x ( ) ⎞ R ⎠ . (3.27) (3.28) ≤ It is evident that ψ 1≡ for when . Let us now consider the R2(cid:2) is the point, where ψH attains its localised function ψH supported on maximum. As for ψH 0≤ the estimate is trivially true, we can assume that ψH x ]( ) > . Furthermore, by an 1 argument of Calabi [10], we can assume that x1 is not in the cut locus of p. It then follows that at this point: (3.29) 0 R2(cid:2) . Let us also assume that x1 in for when r x and ψ 0≡ ( ) ≥ r x ( ) ψH ψH ψH R2 0, 0, 0. R ≤ 0 Δ Δ ( ) ) ( ) [ ( ∇ ≤ = ≤ f Now starting from the basic identity and making note of the relations (3.29) at the maximum point x1, we can write Δ f ( ψH ) = Δ H ψ f , ψ H + ⟨∇ ∇ ⟩ + 2 Δ ψ H f 0 ≥ Δ H ψ f , ψ H + ⟨∇ ∇ ⟩ + 2 Δ ψ H f ≥ Δ H ψ f + 2 ψ , ψ ( ⟨∇ ∇ ψH ) ⟩ − 2 2 ∣ ∣ ∇ ψ ψ ≥ Δ H ψ f − 2 2 ∣ ∣ ∇ ψ ψ Δ . H ψ H f + H ψ H Δ + f (3.30) (3.31) We proceed now by obtaining suitable lower bounds for each of the three individual terms on the right- hand side of this inequality. As for the first term, referring to (3.28), a straightforward calculation gives ψ ∇ = / . Subsequently from these, we have Δ ′ / ∇ and ψ ψ r = ″ ∇ / 2 ∣ ∣ ¯ Δ ψ r R + ′ ¯( ψ R r R ) ¯ 2 Gradient estimates for nonlinear elliptic equations  11 Δ f ψ = Δ ψ , f ψ − ⟨∇ ∇ ⟩ = ¯ ″ ψ R 2 r ∣ ∇ 2 ∣ + ¯ ψ ′ R Δ . r f (3.32) For the last term on the right here the Wei-Wiley Laplacian comparison theorem (cf. [45]) together with . Hence, by substituting back into (3.32) and noting m( ) g ) f 0 gives coth ≥ − k1 ) k r 1 ) m m − − ≤ Δ k r ( ( ( f ′ ≤ , we have: Ric ψ¯ Δ f ψ ≥ 1 ¯ ψ R 2 ″ + ( m 1 ) − R ¯ ψ k ′ coth ( k r ) . (3.33) Moreover upon noting ≤ (here we are using the monotonicity of coth k r ( ) k R coth coth ) ( coth and the bound x x and k k R ) ( coth x 1 ( ≤ 1≤ + + x k R R ) for x / , subject to R r ≤ ≤ 0> ), we deduce that R2 ( m − 1 ¯ ) ψ k ′ coth ( k r ) ≥ ( m − 1 1 )( + ¯ k R ψ R ) ′ / . Hence, substituting this back into (3.33) and making note of the assumptions on ψ¯ , specifically, and the lower bounds on ψ¯′, ψ¯″ in (3.27), it follows that Δ f ψ ≥ ( 1 ¯ ψ R 2 ″ + m − R − R 1 ) ⎛ ⎝ 1 ) ⎛ c 1 ⎝ 1 R 1 R + + ¯ ′ ⎞ k ψ ⎠ k ⎞ ⎠ ( m − c [ 2 + ( m − 1 c ) ( 1 1 + R k )] , ≥ − = − c 2 2 R 1 2 R (3.34) ¯ ψ0 ≤ ≤ 1 (3.35) which can be readily utilised to bound the first term on the right-hand side in (3.31). Referring next to the middle term in the same inequality, by using the imposed bounds on ψ¯′ (the first inequality in (3.27)), we have 2 ∣ ∣ ∇ ψ ψ = 2 ¯ ′ ∇ 2 ϱ ψ ∣ ∣ ¯ 2 ψ R = ¯ ψ ′ ¯ ψ 2 ⎞ ⎟ ⎠ ⎛ ⎜ ⎝ 2 ϱ ∣ ∣ ∇ 2 R ≤ 2 c 1 2 R . (3.36) Since for the third term on the right-hand side in (3.31), we already have the conclusion of Lemma 3.3 at our disposal, by substituting the aforementioned fragments back, we obtain at the maximum point x1 the inequality 0 ≥ Δ H ψ f − 2 (∣ ∇ ψ ψ H ψ H Δ + / ) 2 ∣ f ≥ − H c [ 2 ( m − 1 c ) ( 1 1 + − ψ e ( Σ − Σ u + μe + h + h Σ R k ) + − 2 ( m uu 2 ] / R 2 2 c 1 1 k ) )∣ − + − ψ e ( h Σ − Σ ) u H ψ + 2 Δ [ ( f 2 h m / ) ∇ h 2 ∣ − − 2 ψ h e ⟨∇ , h Σ x − μ Σ ⟩ + xu − ψμe 2 , h H − ⟨∇ ∇ ⟩ x Δ Σ . f h ] (3.37) Referring now to the aforementioned inequality, since we have H 0> that, and ψH ( ∇ ) = 0 at x1, it is easily seen , ψ h H ⟨∇ ∇ ⟩ = − ⟨∇ ∇ ⟩ ≤ , H h ψ H h ψ ∣ ∣∣ ∣ ∇ ∇ ≤ c 1 ψ R H h ∣ ∣ ∇ . Likewise by an application of the Cauchy-Schwarz inequality, we can write ⟨∇ − , h e h Σ x − μ Σ ⟩ ≤ ∇ ∣ xu − h e ∣∣ h Σ x − μ Σ . ∣ xu (3.38) (3.39) Therefore, by substituting (3.38) and (3.39) into (3.37), making note of (3.1) and multiplying through by ψ 0≥ , it follows that 0 ≥ − ψH c [ 2 + ( m − 1 c ) ( 1 1 + R k ) + 2 2 c 1 2 ] / R + 2 2 ψ m / ( )(∣ ∇ h 2 ∣ + h − e Σ ) 2 − 2 c ψ 1 3 2 / ∣ ∇ h H R / ∣ + 2 − ψ e [ h Σ − Σ u + μe h Σ uu − 2 ( m − + 2 − ψ μe h x Δ Σ . f 1 k ) ]∣ ∇ h 2 ∣ + 2 − ψ H e ( h Σ − Σ ) u − 2 − 2 ψ e ∣ h Σ x − μ Σ ∣∣ ∇ h ∣ xu (3.40) 12  Ali Taheri and Vahideh Vahidifar Now to obtain the desired bounds out of this, it is more efficient to proceed by setting In particular note, that y aforementioned in (3.40) thus gives z − = − ψ hΔf and y − μz ψH = by (3.1) and (3.2), respectively. Substituting the y ψ h ∣ ∇ = 2 ∣ , z = − Σ.h − ψe (3.41) 0 ≥ − ψH c [ 2 + ( m − m y m y − A Σ B Σ 1 c ) ( 1 1 2 / ] 1 + + R k ) 2 − ψ H e ( + h Σ ] 2 2 c 1 Σ − 2 / R ) u + 2 + ( / 2 − ψ μe m y )[( x h Δ Σ , f − 2 z ) − ( mc R y ) / 1 1 2 / ( y − μz ) − where A Σ = ( m − 1 ) k − inf (cid:2) R2 {( h − e Σ − Σ u + h μe Σ ) uu − / 2 , } B Σ = h − sup e ∣ (cid:2) R2 Σ x − μ Σ . ∣ xu Utilising Lemma 3.4 upon setting a mc R b m , = = / 1 ΣA , and c m ΣB= , it follows [see (3.22)] that ( y − 2 z ) − mc y 1 1 2 / ( y − μz R m y m y A Σ / − B − ) Σ 1 2 / ≥ 1 2 μ ( y − 2 μz ) − 2 2 2 m c μ 1 8 1 μ ) ( − R 2 ( y − μz ) − 2 2 2 m μ A Σ ε μ )( − − − 2 1 ) 4 1 ( Thus, from (3.42), it follows that 3 ⎡ ⎢ 4 4 ⎣ 4 2 m μ ε μ ( 4 B Σ 1 ) − 1 3 / ⎤ ⎥ ⎦ . 2 0 ≥ − ψH c [ 2 + ( m − 1 c ) ( 1 1 + R k ) + 2 2 c 1 2 ] / R + 2 ) − 2 m ⎡ ⎢ ⎣ ( ψH 2 μ 2 2 2 m c μ 1 1 8 μ ) ( − R 2 ( ψH ) − 2 2 2 m μ A Σ ε μ )( − − 2 1 ) 4 1 ( − 3 ⎡ ⎢ 4 4 ⎣ 4 2 m μ ε μ ( 4 B Σ 1 ) − 1 3 / ⎤ ⎥ ⎦ ⎤ ⎥ ⎦ 2 + 2 − ψ H e ( h Σ − Σ ) u + 2 − ψ μe h x Δ Σ , f or after basic considerations and a rearrangement of terms (3.42) (3.43) (3.44) (3.45) (3.46) 0 ≥ 2 mμ 2 ( ψH − ⎡ ⎢ ( 2 1 ⎣ 1 2 R 2 ) − ⎡ ⎢ ⎣ 2 mμ A Σ ε μ )( − 2 − c [ 2 + ( m − 1 c ) ( 1 1 + R k ) + 2 2 c 1 ] − inf (cid:2) 2 R {( h − e Σ − Σ ) } u − + + 2 1 ) 3 ⎡ ⎢ 2 4 ⎣ mμ ε μ ( 2 B − 4 Σ 1 ) 2 1 3 / ⎤ ⎥ ⎦ − μ inf (cid:2) 2 R {( h − e Δ Σ f x ) } − . ⎤ ⎥ ⎦ 2 2 mc μ 1 1 μ ) − R 2 4 ( ψH ⎤ ⎥ ⎦ (3.47) − Here, we have used ψH e ( h Σ − Σ ) u − ≤ 2 − ψ H e [ h Σ − Σ ] u − and μ e ( h x Δ Σ f ) ≤ − 2 − ψ μe h x Δ Σ f . Now upon setting c D [ = 2 + ( m − 1 c ) ( 1 1 + R k ) + 2 2 c 1 2 ] / R − inf (cid:2) R2 {( h − e Σ − Σ ) } u − + 2 2 mc μ [ 1 / 4 ( ( μ − 2 1 ) R )] , (3.48) and E = mμ 2 2 A / 2 1 ( ( − ε μ )( − 2 1 ) ) + / 3 2 mμ [ 2 4 B / 4 ( ε μ ( − 2 1 ) )] 1 3 −/ μ {( h − e Δ Σ f x ) } − , inf (cid:2) R2 and we can write (3.47) as follows: As a result, it follows from this inequality that 0 ≥ 2 ( ψH ) 2 / ( mμ 2 ) − ψH D ( ) − E . (3.49) (3.50) ψH ≤ ( mμ 2 ) / 4 D [ + 2 D + ( 8 E ) ( / 2 mμ ) ] ≤ ( mμ 2 ) / 4 2 [ D + ( 8 E ) ( / 2 mμ ) ] = ( mμ 2 / 2 D ) + μ m E / 2 . (3.51) Since ψ 1≡ on R(cid:2) and x1 is a maximum point of ψH on R2(cid:2) , we have H = sup (cid:2) R sup (cid:2) R ψH [ ] ≤ sup (cid:2) R2 ψH [ ] = ( ψH x .1 )( ) (3.52) Gradient estimates for nonlinear elliptic equations  13 Thus, it follows that H ≤ ( mμ 2 / 2 ) D + μ m E / 2 . sup (cid:2) R (3.53) Therefore, recalling (3.2), substituting for D and E from (3.48) and (3.49), we can write after multiplying both sides by μ1/ , that for every x ∈ (cid:2) : R μ 1 − ∣ ∇ h 2 ∣ + h − e Σ , ( x e h ) ≤ mμ D 2 / + ( m 1 2 / 2 E ) / ≤ mμ c [ 2 + ( m − 1 c ) ( 1 1 + R k ) + 2 2 c 1 ] ( / 2 R 2 ) − ( mμ 2 inf ) / (cid:2) 2 R {( h − e Σ − Σ ) } u − 8 ( ( / μ − 2 1 ) R )] + + 2 3 2 m c μ [ 1 { m mμ [ 2 2 4 1 A ( ( / Σ − ε μ )( − 2 1 ) )] ( + / 3 4 )[ mμ 2 B 4 Σ ( / 4 ε μ ( − 2 1 ) )] 1 3 / − ( μ 2 inf ) / (cid:2) 2 R {( h − e Δ Σ f x ) } − 1 2 / } . (3.54) Finally, reverting back to u upon noting the relation h log= u and rearranging terms ∣ 2 ∣ 2 u ∇ μu + Σ , x u ] [ u ≤ mμ c [ 2 + ( m − 1 c ) ( 1 1 + R k ) + 2 2 c 1 + 2 2 mc μ 1 4 ( ( / μ − 1 ))] ( / 2 R 2 ) − ( mμ 2 inf Σ ) / 2 R {( (cid:2) u / − Σ ) } u − + { m mμ [ 2 2 4 1 A ( ( / Σ − ε μ )( − 2 1 ) )] ( + / 3 4 )[ mμ 2 B 4 Σ ( / 4 ε μ ( − 2 1 ) )] 1 3 / − ( μ 2 inf Δ Σ {([ ) / 2 R (cid:2) f (3.55) x u ] / ) } − 1 2 / } , which is the desired estimate as in (2.1). The proof is thus complete. 4 Proof of the elliptic Harnack inequality in Theorem 2.3 Now to prove the Harnack inequality, we need to integrate the differential Harnack inequality along a R(cid:2) . Towards this end, let us begin by rewriting the local geodesic path γ joining the points x1 and x2 inside gradient estimate for (1.1) as follows: 2 ∣ u ∣ ∇ 2 u ≤ 2 mμ 2 2 R sup (cid:2) R ⎡ c [ 2 ⎢ ⎣ + ( m − 1 c ) ( 1 1 + R k ) + 2 2 c 1 ] + 2 2 mc μ 1 4 μ ( − 1 ) ⎤ ⎥ ⎦ + + m μ 2 ⎡ ⎢ ⎣ ( 1 − 2 2 mμ A ε μ )( − + ⎡ ⎣⎢ 2 1 ) 27 mμ 4 ε μ ( − 2 4 B 1 ) 2 1 3 / ⎤ ⎦⎥ − 2 inf μ (cid:2) 2 R ( Δ Σ f x u ⎧ ⎨ ⎩ 1 2 / ⎤ ⎥ ⎦ ) ⎫ − ⎬ ⎭ (4.1) 2 mμ 2 sup (cid:2) 2 R ⎧ ⎨ ⎩ ( u Σ , x u ] [ u − u Σ , x u [ ]) ⎫ + ⎬ ⎭ − μ inf (cid:2) R ⎧ ⎨ ⎩ Σ , x u ( [ u ]) ⎫ − ⎬ ⎭ ≔ (cid:3) . Here, we have denoted the expression on the right-hand side of (4.1) by (cid:3) which is a positive constant. ( ) = ), we have 1 ∣∇ / along a geodesic curve γ in R(cid:2) (with γ u u ∣ Now integrating the quantity and γ 0 ( ) = x 2 x 1 log u x ( 2 ) − log u x ( ) 1 = log u γ s ( ( )) d s d d s 1 ∫ 0 1 ∫ [ = ⟨ ∇ / u u γ s ]( ( )) , γ s ( ) ′ ⟩ d s (4.2) 0 ⎡ ⎢ ⎣ sup (cid:2) R ≤ ≤ , d x x 2 1 ( ) u ∣ ⎤ ∣ ∇ ⎥ u ⎦ (cid:3) 1 ∫ γ ∣ d ∣ ′ s 0 ≤ 2 R (cid:3) . 14  Ali Taheri and Vahideh Vahidifar Therefore, log u x [ ( 2 ) / u x ( )] 1 ≤ , d x x 1 ( 2 (cid:3) ) ≤ 2 R (cid:3) or after exponentiation u x ( 2 ) ≤ 2 R e (cid:3) u x ( ) 1 (4.3) giving the desired inequality. The remaining assertions are now straightforward consequences of this □ inequality. 5 Proof of the Liouville result in Theorem 2.5 Starting from (2.4) and noting that Δ Σ f x ≡ ) we obtain, after rearranging the inequality, 0 ΣB = 0 (as a result of Σ ∣ x − μu Σ ∣ xu ≡ ), k 0 0= and ΣC = 0 (as a result of 2 ∣ 2 u ∣ ∇ μu + Σ u ( ) u ≤ m 2 ⎡ ⎢ ( ⎣ 1 2 2 mμ A Σ ε μ )( − − 2 27 mμ 4 ε μ ( − 4 B Σ 2 1 ) + ⎡ ⎢ ⎣ 1 3 / ⎤ ⎥ ⎦ 2 1 ) + 2 μ C Σ 1 2 / ⎤ ⎥ ⎦ + mμ 2 sup M ⎧ ⎨ ⎩ ( u Σ u [ ] u − u Σ u [ ]) ⎫ + ⎬ ⎭ ≤ mμ ⎡ ⎢ ⎣ sup M ⎧ ⎨ ⎩ ( − Σ u [ ] + 4 ( u 1 Σ u [ ] − u ε μ )( − u [ ]) + 2 μu 1 ) Σ uu u − ⎫ ⎬ ⎭ + sup M ( u Σ u [ ] − u 2 u ⎧ ⎨ ⎩ Σ u [ ]) ⎫ + ⎬ ⎭ ⎤ ⎥ ⎦ . Next from the imposed assumptions on Σ and its derivatives, it is easily seen that ( − Σ u [ ] + u Σ u [ ] u − μu 2 Σ u [ ]) + uu ≡ 0, ( u Σ u [ ] u − Σ u [ ]) + ≡ 0. Hence, from (5.1), it follows that 2 ∣ 2 u ∣ ∇ μu + Σ u [ ] u ≡ 0, (5.1) (5.2) (5.3) 2 and so again from the assumptions imposed on Σ that u ∣∇ ∣ 2 / u ≡ . The conclusion now follows at once. □ 0 Acknowledgement: The authors gratefully acknowledge support from EPSRC. Author contributions: All authors have equal contribution in this research. Conflict of interest: The authors declare no conflict of interest. Data availability statement: The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request. References [1] E. Acerbi and R. Mingione, Gradient estimates for a class of parabolic systems, Duke Math. J. 136 (2007), 285–320. [2] D. Bakry, L’hypercontractivité et son utilisation en théorie des semigroupes, in: Lecture Notes in Mathematics, vol. 1581, Springer-Verlag, Berlin/New York, 1994, pp. 1–114. [3] D. Bakry and M. Émery, Diffusions hypercontractives, In: J. Azḿa, M. Yor, eds, Séminaire de Probabilités XIX 1983/84. Lecture Notes in Mathematics, vol. 1123, Springer, Berlin, Heidelberg. [4] D. Bakry, I. Gentil, and M. Ledoux, Analysis and geometry of Markov diffusion operators, A Series of Comprehensive Studies in Mathematics, vol. 348, Springer, 2012. [5] M. F. Biduat-Vèron and L. Veròn, Nonlinear elliptic equations on compact Riemannian manifolds and asymptotics of the Emden equations, Invent. Math. 106 (1991), 489–539. [6] V. Bögelein, F. Duzaar, and G. Mingione, The regularity of general parabolic systems with degenerate diffusion, Mem. Amer. Math. Soc. 221 (2013), no. 1041, vi+143 pp., AMS. Gradient estimates for nonlinear elliptic equations  15 [7] L. Brandolini, M. Rigoli, and A. G. Setti, Positive solutions of Yamabe type equations on complete manifolds and appli- cations, J. Funct. Anal. 160 (1998), 176–222. [8] K. Brighton and A Liouville-type theorem for smooth metric measure spaces, J. Geom. Anal. 23 (2013), 562–570. [9] L. A. Cafarreli, B. Gidas, and J. Spruck, Asymptotic symmetry and local behavior of semilinear elliptic equations with critical Sobolev growth, Comm. Pure Appl. Math. 42 (1989), 271–297. [10] E. Calabi, An extension of Hopf maximum principle with application to Riemannian geometry, Duke Math. J. 25 (1958), 45–56. [11] H. Cao, Recent progress on Ricci solitons, in: Recent Advances in Geometric Analysis, Advanced Lectures in Mathematics (ALM), Vol. 11, International Press, 2010, pp. 1–38. J. Case, A Yamabe-type problem on smooth metric measure spaces, J. Diff. Geom. 101 (2015), 467–505. [12] [13] S. Y. Cheng and S. T. Yau, Differential equations on Riemannian manifolds and their geometric applications, Comm. Pure Appl. Math. 28 (1975), 333–354. [14] Y. Choquet-Bruhat, General Relativity and the Einstein Equations, Oxford Mathematical Monographs, OUP, 2009. [15] B. Chow, P. Lu, and L. Nei, Hamiltonas Ricci Flow, Graduate Studies in Mathematics, vol. 77, AMS, 2006. [16] N. T. Dung, N. N. Khanh, and Q. A. Ngô, Gradient estimates for f -heat equations driven by Lichnerowiczas equation on complete smooth metric measure spaces, Manuscripta Math. 155 (2018), 471–501. [17] M. Ghergu, S. Kim, and H. Shahgholian, Exact behaviour around isolated singularity for semilinear elliptic equations with a log-type nonlinearity, Adv. Nonlinear Anal. 8 (2019), 995–1003. [18] M. Giaquinta, Multiple Integrals in the Calculus of Variations and Nonlinear Elliptic Systems, Annals of Mathematics Studies, Vol. 105, Princeton University Press, 1983. [19] B. Gidas and J. Spruck, Global and local behaviour of positive solutioins of nonlinear elliptic equations, Comm. Pure Appl. Math. 34 (1981), 525–598. [20] A. Grigor’yan, Heat kernel analysis on manifolds, Studies in Advanced Mathematics, AMS, 2013. [21] Z. M. Guo and J. C. Wei, Hausdoff dimension of ruptures for solutions of a semilinear equation with singular nonlinearity, Manuscr. Math. 120 (2006), 193–209. [22] R. Hamilton, A matrix Harnack estimate for heat equation, Comm. Anal. Geom. 1 (1993), 113–126. [23] R. Hamilton, The formation of singularities in the Ricci flow, Surv. Diff. Geom. 2 (1995), 7–136. [24] J. M. Lee and T. H. Parker, The Yamabe problem, Bull. Amer. Math. Soc. 17 (1987), 37–91. [25] J. Li, Gradient estimates and Harnack inequalities for nonlinear parabolic and nonlinear elliptic equations on Riemannian manifolds, J. Funct. Anal. 100 (1991), 233–256. [26] P. Li and S. T. Yau, On the parabolic kernel of Schrödinger operator, Acta Math. 156 (1986), 153–201. [27] P. Li, L. F. Tam, and D. G. Yang, On the elliptic equation = geometric applications, Trans. Amer. Math. Soc. 350 (1998), 1045–1078. Ku ku Δ − + 0 u p on complete Riemannian manifolds and their [28] P. Li, Geometric Analysis, Cambridge Studies in Advanced Mathematics, vol. 134, CUP, 2012. [29] X. D. Li, Liouville theorems for symmetric diffusion operators on complete Riemannian manifolds, J. Math. Pures Appl. 84 (2005), 1295–1361. [30] J. Lott, Some geometric properties of the Bakry-Émery Ricci tensor, Comment. Math. Helv. 78 (2003), 865–883. [31] L. Ma, Gradient estimates for a simple elliptic equation on complete noncompact Riemannian manifolds, J. Funct. Anal. 241 (2006), 374–382. [32] P. Mastrolia, M. Rigoli, and A. G. Setti, Yamabe Type Equations on Complete Non-compact Manifolds, Springer, Basel, 2012. [33] G. Morrison and A. Taheri, An infinite scale of incompressible twisting solutions to the nonlinear elliptic system ,L u; [ A B ] = ∇ P and discriminant Δ , h g ( ) , Nonlin. Anal. 173 (2018), 209–219. [34] Q. H. Ruan, Elliptic type gradient estimates for Schrödinger equations on noncompact manifolds, Bull. Lond. Math. Soc. 39 (2007), 982–988. [35] R. Schoen and S. T. Yau, Lectures on Differential Geometry, International Press, 1994. [36] X. F. Song and L. Zhao, Gradient estimates for the elliptic and parabolic Lichnerowicz equations on compact manifolds, Z. Angew. Math. Phys. 61 (2010), 655–662. [37] P. Souplet and Q. S. Zhang, Sharp gradient estimate and Yau’s Liouville theorem for the heat equation on noncompact manifolds, Bull. Lond. Math. Soc. 38 (2006), 1045–1053. [38] A. Taheri, Function Spaces and Partial Differential Equations, Vol. I, Oxford Lecture Series in Mathematics and its Applications, vol. 40, OUP, 2015. [39] A. Taheri, Function Spaces and Partial Differential Equations, Vol. II, Oxford Lecture Series in Mathematics and its Applications, vol. 41, OUP, 2015. [40] A. Taheri, Liouville theorems and elliptic gradient estimates for a nonlinear parabolic equation involving the Witten Laplacian, Published online in: Advances in Calculus of Variations, De Gruyter, 2021, doi: https://doi.org/10.1515/acv- 2020-0099. [41] A. Taheri, Gradient estimates for a weighted Γ-nonlinear parabolic equation coupled with a super Perelman-Ricci flow and implications, Published online in: Potential Analysis, Springer, 2021, doi: https://doi.org/10.1007/s11118-021-09969-2. 16  Ali Taheri and Vahideh Vahidifar [42] A. Taheri and V. Vahidifar, On multiple solutions to a family of nonlinear elliptic systems in divergence form combined with an incompressibility constraint, Nonlinear Anal. 221 (2022), 112889. [43] A. Taheri and V. Vahidifar, Gradient estimates for a nonlinear parabolic equation on smooth metric measure spaces with evolving metrics and potentials, To appear in: Nonlinear Analysis 2023. [44] C. Villani, Optimal Transport: Old and New, A Series of Comprehensive Studies in Mathematics, vol. 338, Springer, 2008. [45] G. Wei and W. Wylie, Comparison geometry for the Bakry-Émery Ricci tensor, J. Diff. Geom. 83 (2009), 377–405. [46] J. Y. Wu, Li-Yau type estimates for a nonlinear parabolic equation on complete manifolds, J. Math. Anal. Appl. 369 (2010), 400–407. [47] J. Y. Wu, Elliptic gradient estimates for a weighted heat equation and applications, Math. Z. 280 (2015), 451–468. [48] J. Y. Wu, Gradient estimates for a nonlinear parabolic equation and Liouville theorems, Manuscript Math. 159 (2019), 511–547. [49] S. T. Yau, Harmonic functions on complete Riemannian manifolds, Comm. Pure Appl. Math. 28 (1975), 201–228. [50] Y. Y. Yang, Gradient estimates for a nonlinear parabolic equation on Riemannian manifolds, Proc. Amer. Math. Soc. 136 (2008), 4095–4102. [51] Y. Y. Yang, Gradient estimates for the equation Δ u + cu α− = 0 on Riemannian manifolds, Acta Math. Sin. 26 (2010), 1177–1182. [52] Q. S. Zhang, Positive solutions of 0 [53] Q. S. Zhang, Sobolev Inequalities, Heat Kernels Under Ricci Flow and the Poincaré Conjecture, CRC Press, 2011. [54] J. Zhang and B. Ma, Gradient estimates for a nonlinear equation and its parabolic counterpart, Pac. J. Math. 213 (2004), 163–200. on complete noncompact manifolds, Vu Wu + Δ − u cu Δ = 0 α− u p = + f Comm. Math. 19 (2011), 73–84.
10.1080_02699931.2022.2157377
Cognition and Emotion ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/pcem20 Induced gratitude and hope, and experienced fear, but not experienced disgust, facilitate COVID-19 prevention Pascale Sophie Russell, Michal Frackowiak, Smadar Cohen-Chen, Patrice Rusconi & Fabio Fasoli To cite this article: Pascale Sophie Russell, Michal Frackowiak, Smadar Cohen-Chen, Patrice Rusconi & Fabio Fasoli (2023) Induced gratitude and hope, and experienced fear, but not experienced disgust, facilitate COVID-19 prevention, Cognition and Emotion, 37:2, 196-219, DOI: 10.1080/02699931.2022.2157377 To link to this article: https://doi.org/10.1080/02699931.2022.2157377 © 2023 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group Published online: 10 Jan 2023. Submit your article to this journal Article views: 718 View related articles View Crossmark data Full Terms & Conditions of access and use can be found at https://www.tandfonline.com/action/journalInformation?journalCode=pcem20 COGNITION AND EMOTION 2023, VOL. 37, NO. 2, 196–219 https://doi.org/10.1080/02699931.2022.2157377 Induced gratitude and hope, and experienced fear, but not experienced disgust, facilitate COVID-19 prevention Pascale Sophie Russella, Michal Frackowiak Fabio Fasoli b, Smadar Cohen-Chenc, Patrice Rusconi d and a aSchool of Psychology, Faculty of Health and Medical Sciences, University of Surrey, Guildford, United Kingdom; bInstitute of Psychology, University of Lausanne, Lausanne, Switzerland; cBusiness School, University of Sussex, Sussex, United Kingdom; dDepartment of Cognitive Sciences, Psychology, Education and Cultural Studies (COSPECS), University of Messina, Messina, Italy ABSTRACT Hope, gratitude, fear, and disgust may all be key to encouraging preventative action in the context of COVID-19. We pre-registered a longitudinal experiment, which involved monthly data collections from September 2020 to September 2021 and a six-month follow-up. We predicted that a hope recall task would reduce negative emotions and elicit higher intentions to engage in COVID-19 preventative behaviours. At the first time point, participants were randomly allocated to a recall task condition (gratitude, hope, or control). At each time point, we measured willingness to engage in COVID-19 preventative behaviours, as well as experienced hope, gratitude, fear, and disgust. We then conducted a separate, follow-up study in February 2022, to see if the effects replicated when COVID-19 restrictions were relaxed in the UK. In the main study, contrary to our pre-registered hypothesis, we found that a gratitude recall task elicited more willingness to engage in COVID-19 preventative behaviours in comparison to the neutral recall task. We also found that experienced gratitude, hope, and fear were positively related to preventative action, while disgust was negatively related. These results present advancement of knowledge of the role of specific emotions in the COVID-19 pandemic. ARTICLE HISTORY Received 29 April 2022 Revised 6 December 2022 Accepted 6 December 2022 KEYWORDS Hope; gratitude; fear; disgust; preventative behaviour One way to reduce COVID-19 is to engage in preventa- tive health behaviours, such as washing hands more frequently, engaging in social distancing, and receiv- ing vaccinations (e.g. Haug et al., 2003; Skegg et al., 2021). Preventative health behaviours refer to actions that create a benefit for the individual and, in turn, for the community or society. it has been found that intentions to engage in social distan- cing are related to actual social distancing behaviours (Gollwitzer et al., 2022). We also know from research on collective action that intentions often result in actual engagement (De Weerd & Klandermans, 1999; Webb & Sheeran, 2006). Thus, in the context of COVID-19, measuring intentions rather than actual behaviours Importantly, may be a good marker of whether individuals will engage in preventative health behaviours. Moving beyond previous studies, it is important to examine willingness to engage in action in the long term, not just at a single time point, as has been evidenced by COVID-19 burnout (Queen & Harding, 2020). This is essential, as prior evidence tells us that COVID-19 will have a long-term impact on societies (Rourke, 2020). experienced The emotions and participants’ willingness to engage in preventative health behaviours in a year-long longi- tudinal study and a separate, follow-up study in the In doing so, we examined whether recalling UK. hope and gratitude, as well as experienced emotions investigated current project CONTACT Pascale Sophie Russell © 2023 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivatives License (http://creativecommons. org/licenses/by-nc-nd/4.0/), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited, and is not altered, transformed, or built upon in any way. p.s.russell@surrey.ac.uk of gratitude, hope, fear, and disgust, are associated with preventive health behaviours. Emotions and COVID-19 context In the context of COVID-19, it is plausible that both positive and negative emotions can facilitate preven- tative action. Broadly, positive emotions can encou- rage attention, creative thinking, well-being, coping, resilience and reduce the effects of negative emotions (see Fredrickson, 2001 for a review). Research shows that not only positive emotions, such as hope, but also negative emotions, such as guilt, can sometimes facilitate positive outcomes, as even though the experience itself may be negative, negative emotions can also “do good”, e.g. guilt facilitating reparative action in interpersonal or intergroup relations (Cohen-Chen et al., 2014). It has been found that reap- praisal interventions can reduce negative emotions and increase positive emotions (both current and anticipated emotions), in the context of COVID-19 (Wang et al., 2021). However, the reappraisal manipu- lations in this research did not have an impact on pre- the ventative interventions did not elicit specific emotions. One way to impact preventative action is through eliciting the experience of specific emotions, as prior research has found that specific emotions are associated with unique appraisals, which can impact motivations and behaviours (Lerner et al., 2015; Roseman, 2011; Shaver et al., 1987; Smith & Ellsworth, 1985). behaviours, probably because According to the appraisal tendency framework (ATF), emotions differ by certain appraisals, such as certainty about the future, which impact motivations and behaviours (Lerner et al., 2015). One important appraisal in the context of COVID-19 is certainty, as it is likely to impact willingness to engage in preven- tative action. Prior research has found that disgust and anger are associated with feelings of certainty while fear is associated with feelings of uncertainty (Bachkirov, 2015; Polyportis et al., 2020; Tiedens & Linton, 2001). Fear and disgust are both associated with contagion concerns (Olatunji et al., 2009; Thorpe et al., 2011); thus, they are likely to play a Indeed, fear of key role in reactions to COVID-19. COVID-19 has been found to be associated with anti-COVID-19 behaviours (e.g. washing hands, wearing marks; Zingora et al., 2022). However, anger is triggered by appraisals of injustice and harm (see Lerner & Tiedens, 2006; Russell & Giner-Sorolla, 2013); thus, it is likely to have a more peripheral role COGNITION AND EMOTION 197 in the COVID-19 context in comparison to disgust and fear. For this reason, we focused on the latter two negative emotions in the present research. Fear, like hope, is a forward-focused emotion, based on appraisals of uncertainty, in which we consider the current situation to be ambiguous and the future uncertain (Baumgartner et al., 2008; Winterich & Haws, 2011). Fear and hope also trigger more in- depth processing and attention to the situation at hand (Polyportis et al., 2020). However, the appraisal that differs between fear and hope is coping poten- tial, with hope often triggering resourcefulness and perseverance (see Nabi & Myrick, 2022 for a review), making hope a good potential candidate emotion to be elicited long term to encourage COVID-19 preven- tative action. For this reason, we aimed at eliciting hope rather than fear in the present research. Beside looking at the effect of recalling hope versus grati- tude, it is also important to examine specific experi- enced emotions, i.e. hope, disgust, and fear, rather than general reappraisals, and how these specific emotions impact willingness to engage in preventa- tive action. Gratitude and hope in COVID-19 context Emotions have already been shown to play a role in the COVID-19 pandemic, influencing social outcomes and well-being. Prior research has identified that trait gratitude has been shown to be associated with more COVID-19 prosocial behaviours, such as adhering to public health guidelines (Syropoulos & Markowitz, 2021). Additionally, a gratitude intervention (i.e. to list three things that went well) was shown to have a positive impact on social connectedness in the context of COVID-19 (Dennis et al., 2022). In terms of state gratitude, it was found to be signifi- cantly higher in the gratitude condition in comparison to the control and Best Positive Self conditions, but levels of state gratitude did not differ between the nostalgia and gratitude conditions. In another study, a gratitude writing task was shown to reduce stress and negative affect, more so than an expressive writing task or control condition; however, levels of state gratitude remained relatively stable (Fekete & Deichert, 2022). Another study conducted in the context of COVID-19 found that, in comparison to a control condition, both a kindness and a gratitude manipulation (i.e. thinking of things they are grateful for) triggered more positive emotions; however, there was no difference between the two positive 198 P. S. RUSSELL ET AL. intervention effects (Datu et al., 2022). Additionally, the manipulations had no impact on life satisfaction, COVID-19 anxiety, and negative emotions; even though they did find that measures of gratitude and kindness were highest in respective conditions; thus, the manipulations elicited appropriate experiences. Another study has shown that a gratitude interven- tion based on weekly reflections (i.e. thinking and writing about up to five things in life to be grateful for 10 weeks improved university students’ for) mental well-being during the COVID-19 pandemic (Geier & Morris, 2022). Cumulatively, there is evidence that both trait (i.e. dispositional) and state (i.e. experi- enced) gratitude play a role in eliciting positive feel- ings and social outcomes in the COVID-19 context. However, the long-term impact of gratitude in the COVID-19 context, or how it compares to hope, which is also a positive emotion, is unknown. Accord- ingly, our research aims at addressing this gap in the literature. Hope has been linked to increased well-being during the COVID-19 pandemic; for example, trait hope mediated the relationship between COVID-19 stress and subjective well-being in a sample of under- graduate students in Turkey (Genç & Arslan, 2022). Another study with American adults found that trait hope predicted greater well-being, lower anxiety, and lower COVID-19-related perceived stress after a month via the mediation of perceived emotional control (Gallagher et al., 2021). These studies suggest that trait hope can counteract the negative effects during the pandemic and may facilitate positive social outcomes. These studies have focused on trait hope, but state hope (i.e. experienced in the moment or induced) may also play a role. This is important as state and trait emotions can both impact behaviour (Smith et al., 2018). Although trait hope is an ingrained and personality-based character- istic, while state hope is a discrete emotion arising from a specific appraisal within a specific context, the cognitive and affective mechanisms are similar, as with most relationships between traits and corre- sponding states (Fleeson, 2001). Specifically for hope, the similarities between trait hope and state hope have been documented in past literature (Luthans & Jensen, 2002; Ong et al., 2009; Snyder, 2000; Snyder et al., 1996; Yoshinobu et al., 1991). Based on this, initial understandings about possible attitudes and action tendencies arising from trait hope can indicate the role played by state hope. Also, we are not aware of any prior research that has examined the impact of hope recall tasks in the context of COVID-19 (i.e. eli- cited by an intervention or task), which may also have an impact on COVID-19 preventative action. Like gratitude, it is important to examine the long- term impact of recalling and experiencing hope, and whether hope is associated with COVID-19 preventa- tive action. Hope and gratitude both involve positive valence, meaning that they feel pleasant to the individual experiencing these emotions and therefore are mostly considered positive emotions (Cohen-Chen factors can et al., 2014). However, contextual influence the valence of emotions considered plea- sant and it has been shown that emotions can entail both positivity and negativity (An et al., 2017). Hope and gratitude are also relevant to daily social inter- actions, and in some instances morality (Haidt, 2003). However, hope and gratitude are specific dis- crete emotions, triggered by unique appraisals (Lazarus, 1991; Smith & Ellsworth, 1985). For example, when feeling gratitude, we may reflect on what others have done for us (Fox et al., 2015), while hope derives from thoughts regarding what our social world may be like in the future (Averill, 1994; Cohen-Chen et al., 2014; Stotland, 1969). Hope is both a forward-focused and energising emotion (Snyder, 2000). Experiencing hope involves a pleasant affective response to envisioning a future which is better than the current situation, and to which the experiencer attaches meaning and relevance (Stotland, 1969). Hope’s behavioural and attitudinal outcomes involve cognitively developing alternatives and plans to achieve future goals, which have been found to motivate behaviour (Yoshinobu et al., 1991). Particularly in collective contexts, when coupled with agency and a sense of efficacy, hope can induce collec- tive action intentions (Cohen-Chen & Van Zomeren, 2018), while efficacy had no such effect in situations In conflict, hope has been perceived as hopeless. found to induce conciliatory attitudes such as support for concession-making (Cohen-Chen et al., 2014, 2015). Cumulatively, it seems that hope can action behaviour conducive to promoting social change. On the other hand, gratitude is an emotion that focuses on reflections of what we have already received (Haidt, 2003), mostly from others (Algoe, 2020; Fox et al., 2015). This reflection on what we have received in the past then elicits the current state of gratitude. Thus, hope is a self-directed emotion focused on future possibilities, while grati- tude is an other-directed emotion focused on past gains and facilitating future reciprocal behaviours (Chang et al., 2020). For example, gratitude has been associated with numerous prosocial behaviours (Haidt, 2003), such as reducing selfishness, increasing forgiveness, helping (Bartlett & DeSteno, 2006; McCul- lough et al., 2008) and donation behaviour (Paramita et al., 2020). Gratitude also has a positive impact on emotional and social well-being (Jans-Beken et al., 2020). Lastly, gratitude has been found to lead to better relationships with others, repayment of bene- factor, and expanding our moral circle (Bartlett & DeSteno, 2006; McCullough et al., 2001). Thus, experi- enced gratitude also has the potential to facilitate positive behaviours, encouraging COVID-19 preventa- tive action. Fear and disgust in COVID-19 context However, as outlined before, sometimes negative emotions can also encourage positive outcomes through the actions they facilitate (Cohen-Chen et al., 2014), but it is also evident from prior research that interventions to date have not always had a large impact on negative experienced emotions (e.g. Datu et al., 2022). Thus, it is important to examine when negative emotions can have a positive impact encouraging COVID-19 preventative action. The COVID-19 pandemic has been a time of increased negative emotions (e.g. Xue et al., 2020), such as fear and disgust. A sample experiencing lockdown was found to have heightened disgust sensitivity in comparison to one that was not under lockdown (Ste- venson et al., 2021). However, disgust can be experi- enced as both a state and trait (i.e. disgust sensitivity), but to our knowledge prior research has not examined experienced disgust in the context of i.e. state disgust. Anxiety, a feeling like COVID-19, fear but experienced long term as a mood, has been shown to mediate the relationship between COVID- 19 conspiracy beliefs with placing more importance on governmental restrictions (Peitz et al., 2021). Fear for COVID-19 has been found to be related to anti COVID-19 behaviours (Zingora et al., 2022). Thus, both disgust and fear are likely to be heightened during the COVID-19 pandemic. Though prior research suggests that fear and disgust may have differential impacts on preventative action. Both fear and disgust are avoidance emotions (Ekman, 1999), related to the function of protecting from harm and disease (Rozin & Fallon, oneself should suggesting that 1987), these emotions COGNITION AND EMOTION 199 encourage preventative action. Furthermore, disgust often triggers the need to exclude others from our social circle (e.g. Dasborough et al., 2020; Greenbaum et al., 2020), thus, people may follow COVID-19 restric- tions to see themselves as superior and separate from others. Fear can also encourage individuals to flee the situation under certain circumstances when there is a threat from disease (Pliskin et al., 2015). preventive discourage On the other hand, disgust and fear have also been linked with defensive responses (Herek & McLemore, 2013; Liberman & Chaiken, 1992), suggesting these action. emotions may However, it has been found that fear can facilitate health preventative behaviours, when people are aware of what they need to do to cope with the situ- ation (Maddux & Rogers, 1983). Also, when the fear stimuli are not too threatening, fear can facilitate posi- tive health behaviours, i.e. behaviour change (Petty, 1995). It has also been shown that fear can turn into hope when efficacy is high (Nabi & Myrick, 2022) and this can then lead to behaviour change, i.e. increase sun protective behaviours. Interestingly, in this research hope was associated with behaviour change but not attitude change. Based on prior evi- dence, it is probable that experienced gratitude, hope, and fear are likely to have positive relationships with preventative behaviours, whilst experienced disgust is less likely to be related to preventative action. The present research We pre-registered a larger project on OSF, which involved hypotheses about the impact of gratitude and hope recall tasks on COVID-19 preventative beha- viours.1 We first conducted a pilot study in July 2020, recruiting 300 participants from Prolific, to inform the measures and materials used in the main study (pilot study data set can be found on OSF). Most of the measures were the same as the main study. In the main study, we wanted to test whether a simple recall task of hope or gratitude (vs. no emotion recall) would have a positive impact by increasing individuals’ willingness to engage in COVID-19 pre- ventative behaviours. Moving beyond other research looking at the role of emotions in COVID-19 preventa- tive behaviours, we wanted to test whether hope versus gratitude recall tasks would impact behaviour intentions over time, i.e. across multiple time points. Therefore, we conducted a longitudinal study (Study in which participants engaged with the same 1) 200 P. S. RUSSELL ET AL. recall task once a month for six months and then fol- lowed up with another time point six months later where participants did not engage in the recall task. The chosen time points for this study are based on previously used methods (e.g. Gilles et al., 2011), as it is necessary to look at the outcomes consistently across a six-month time frame (September 2020 to February 2021) and then conduct another 6 month follow up (August 2021) to see if there is sustained change in behaviours even when not exposed to the emotion manipulation, and, co-incidentally, this was a period where restrictions were being relaxed. The period of September 2020 to February 2021 included key events related to the pandemic in the UK, such as the introduction of the rule of six, tier lockdowns (see the Institute systems, and national for Government analysis/summary of events, 2021). Thus, to examine these time periods to compare the impact on the different experienced emotions (i.e. hope, gratitude, fear, and disgust), in comparison to a baseline control con- dition, which was necessary as it enabled us to monitor the impact of key events on these emotions. At each time point, after the recall task, participants rated how much gratitude, hope, fear, and disgust they were experiencing currently and when focusing on the future, as well as reported their willingness to engage in health preventative action. In addition to this longitudinal study, we also tested whether recalling hope and gratitude would have an impact on willingness to engage in preventative action (Study 2), even at a time when restrictions were lifted and engaging in COVID-19 preventative action was more of a personal choice than a government mandate. it was also useful Hypotheses The current paper will focus on hypotheses relevant to experienced emotions (i.e. state emotions) and preven- tative action. Based on the unique effects of hope (i.e. encouraging other health-protective behaviours and forward thinking), we predicted that recalling hope would be more effective in reducing fear (Hypothesis 1a) and disgust (Hypothesis 1b), and would encourage preventative action (Hypothesis 2) in comparison to a gratitude recall task and neutral task. Based on prior evidence suggesting that gratitude is key to the pandemic (Datu et al., 2022; Dennis et al., 2022; Fekete & Deichert, 2022; Syropoulos & Marko- likely candidate witz, 2021), gratitude is another emotion to facilitate outcomes; however, we pre- dicted that it would have a smaller impact in compari- son to recalling hope, due to hope’s motivating nature (Nabi & Myrick, 2022; Snyder, 2000). It is also useful to compare these two positive emotion recall conditions as prior research has not found a difference between the positive recall tasks (Datu et al., 2022), but has not compared specific emotion recall tasks; thus, this research will uncover when the two positive emotions can be associated with positive action, com- paring their effects on preventative action and nega- tive emotions. We will examine hope and gratitude’s impact as induced emotions (i.e. elicited by recall tasks) and experienced emotions (i.e. experienced cur- rently and when considering the future). Hypotheses 1 and 2 were pre-registered as part of a larger project with multiple variables and hypoth- trust, and efficacy eses, such as trait emotions, beliefs.2 We also conducted further multilevel model analyses to explore whether experienced emotions (hope, gratitude, fear, disgust) predict willingness to engage in preventative behaviours, and if these experienced emotions have an impact even when controlling for the impact of time and recall task con- dition. After our pre-registration and based on prior literature reviewed previously, we predicted that levels of experienced hope, gratitude, and fear would be associated with greater willingness to engage in preventative action, but disgust would be less likely to be related to preventative action. Study 1 Method Design This experiment utilised a 3 Emotion Recall Task Con- dition (Hope versus Gratitude versus Control, between-participants) x 7 Time (once a month for a period of 6 months and a 6 month follow up the emotion recall, within-participants) without mixed design. We examined the impact of time and emotion recall on preventative behaviours and experienced emotions by clustering mixed multilevel model analysis by participants (to facilitate the exploration of repeated measures nested within an individual). Participants Under the original analysis assumption, a G*Power 3.1 (Faul et al., 2009) a-priori power analysis indicated that an adequate sample size would be 247 (assuming an effect size of 0.20, with a power of 0.85 and α of .05, performing MANOVA analysis, repeated measures, with between-within interactions). Due to the longi- tudinal design and assuming attrition rates, we aimed to recruit 375 participants. We recruited from Prolific, participants were rewarded £2 at each time point. To take part in the research, participants had to be British, not having completed our pilot study, and a prior approval rate of 97% or above in prior studies. See Table 2 for full details of how many par- ticipants took part in each time point by emotion recall task condition. age range (Mage = 36.98, At time point 1, participants mostly identified as being female (71%).3 In terms of age, there was a variable SDage = 12.6, range: 18–88). The sample was predominantly White (86%). The majority of participants had a uni- versity degree or higher (63%). At time point 7, 247 in the participants (64%) follow-up questionnaire. Of the remaining sample, most identified as female (75%), White (87%), and received a university degree of higher (66%). The sample still had a variable age range (Mage = 39.53, SDage = 13.25, range: 19–89). returned to take part Materials and procedure Participants were presented with an information sheet and consent form at each time point. They then completed the measures/materials in the follow- ing order: Demographics. Participants first filled in the follow- ing demographic variables: age, gender, education, nationality, and ethnicity. recall: Emotion Participants were randomly assigned to recall five things that made them feel either hopeful or grateful (instructions below), or, for the neutral condition, they recalled five things that they planned to do the following Wednesday. Partici- pants were randomly assigned to one of the con- ditions at time point 1 and thereafter completed the same recall task in the remaining timepoints. Partici- pants completed the same emotion recall task at each of the 6 initial time points, but not at the 7th final time point. Please describe 5 events, situations, episodes, or objects that make you feel grateful/hopeful for what you currently have in the context of the COVID-19 pandemic. Describe in detail how these events, situations, episodes, or objects make you feel and why you feel this way? COGNITION AND EMOTION 201 Experienced emotions (current and future): Partici- pants then self-reported how much they were cur- rently experiencing several distinct emotions. They also rated the same list of discrete emotions but in relation to what our future may be like in the context of the COVID-19 outbreak. Specifically, for the current emotions block they were asked to indi- cate how much you feel the following emotions in relation to the current circumstances concerning the COVID-19 pandemic. Then, for the future emotions block they were asked to indicate how much you feel the following emotions in relation to what our future in the next month may be like because of the COVID-19 pandemic. The current and future emotions were presented in in a randomised order. The two separate blocks, emotion items were rated on a Likert scale from 1 (Not at all) to 7 (Extremely). For theoretical reasons our analysis focused on hope, gratitude, disgust, and fear as these emotions are associated with different appraisals and behavioural tendencies (see OSF for the full list of emotion items). Reliability was high (Cranford et al., 2006) at both within and between levels across the seven time points for all four of the emotions we analyzed. The reliability across the current hope, was high within par- ticipants (Rc = .75) and between participants (RKF = .97). Similarly, reliability across the current gratitude items was high within participants (Rc = .80) and high between participants (RKF = .97). For current fear, reliability was high within participants (Rc = .79) and between participants (RKF = .98). Finally, reliability across the current disgust items was moderate to high within participants (Rc = .73) and high between participants (RKF = .97). For future hope items, reliability was high within participants (Rc = .83) and between participants (RKF = .98). For future gratitude, reliability was high within participants (Rc = .82) and between participants (RKF = .97). The reliability across the future fear items was high within participants (Rc = .82) and high between participants (RKF = .98), and for future disgust, high within participants (Rc = .75) and between partici- pants (RKF = .98). Preventative behaviours. Participants were then asked about their willingness to engage in preventa- tive health behaviours (i.e. in the next month how willing would you be to … ), on a scale from 1 (Not at all likely) to 7 (Extremely likely). The instructions and scale items were partially adapted from Chuang (2015), and the scale comprised additional et al. 202 P. S. RUSSELL ET AL. items related specifically to COVID-19 (e.g. track-and- trace app, social distance, avoiding public transport, wear a mask, wash hands more frequently). The scale included 12 items in total (full list of items can be found on OSF). The reliability across the preventa- tive behaviours items was moderate to high within participants (Rc = .67) and high between participants (RKF = .98) across 7 time points. After completing all items, participants gave their consent to have their data submitted. All participants were fully debriefed after the final time point. Ethical approval for this study was obtained from the Univer- sity of Surrey. Data analysis strategy longitudinal The goal of this research was to examine the impact of the hope and gratitude recall tasks on experi- enced emotions and preventative action. We also explored the impact of experienced hope, gratitude, fear, and disgust on willingness to engage in pre- ventative behaviours, controlling for the impact of time and recall task. We decided to include partici- pants throughout the study, even if they missed a time point. To account for missing data, along with its and hierarchical nature, we decided to execute multilevel model analysis (MLM) instead of originally assumed MANOVA. We reasoned that because of the intensive longitudinal nature of our data, this type of analysis would allow us to distinguish between- and within-person level of data, nested within a person (Bolger & Lauren- ceau, 2013; Maas & Hox,2004 ). Clustering by partici- pants including participants who had missed certain time points, but it also provides random effects (intercepts and the slopes) missing data, one of the advantages of multilevel modelling is that it can handle intensive longitudi- nal data. MLM assumes that the data are missing at random and there is nothing systematic about the missing time points. Therefore, the model can handle participants who took part in the entire study as well as participants who took part only in certain time points. MLM computes the slope between those defined time points. in MLM allows not only for for each participant. In terms of The data were analyzed in R version 3.6.3 (R Core Team, 2018), using the “lmer” package (Bates et al., 2015), and “emmeans” package (Lenth et al., 2021) to compute simple effects across mixed models. To control for convergence of the executed MLM, we optimised the models for the non-linear parameter estimation using box-constrained optimisation L- BFGS-B from the “optimx” package (Nash, 2014). Complementary analyses were run using jamovi 2.2.2 (The jamovi project, 2021). (CurrentGratitudebetween), The predictor variables were disaggregated into a within-person and a between-person component. The between-person component was calculated based on an overall grand mean of the participant’s average score of each predictor variable, for example: Current hope (CurrentHopebetween), current gratitude fear (CurrentFearbetween), disgust (CurrentDisgustbetween). to compute the within-person component, the between-person com- ponent was subtracted from the uncentred individual score of each participant from the monthly values of, for example, current hope (CurrentHopewithin), current gratitude fear (CurrentFearwithin), disgust (CurrentDisgustwithin). (CurrentGratitudewithin), and In order current current current current and Apart testing from manipulation checks (e.g. whether gratitude and hope differed by condition), the hypotheses were tested with a series of MLM, including preventative behaviours as a dependent variable, and experienced current emotions as predic- tors, on between- and within-person level. The equation of the model, where the preventative beha- viours variable is predicted by current-oriented emotions is demonstrated below: Preventative Behaviorsit = g01(CurrentHopebetween) + g10(CurrentHopewithin) + g02(CurrentGratitudebetween) + g20(CurrentGratitudewithin) + g03(CurrentFearbetween) + g30(CurrentFearwithin) + g04(CurrentDisgustbetween) + g40(CurrentDisgustwithin) + time + condition + u0i + u1i (CurrentHopewithin) + u2i(CurrentGratitudewithin) + u3i (CurrentFearwithin) + u4i(CurrentDisgustwithin) + 1it In this model, i refers to individuals and t refers to time point, whereas g01 to g04 index present-oriented emotions on a between-person level. On the other hand, g10 to g40 describe all four model variables on a within-person level. u0i represents the random inter- cept, and terms from u1i to u4i represent the random COGNITION AND EMOTION 203 slopes for the model variables, respectively to the numbers allocated to between- and within-person effects. Finally, 1it stands for the regression residual for participant i on day t. We also repeated the ana- lyses with experienced future emotions entered in the model rather than current emotions; however, there were no differences (see Appendix 1). in the results Results Descriptive statistics Table 1 shows the overall means, standard deviations, correlations on between- and within-person level, along with intraclass correlations of the study vari- ables. Table 2 presents the means and standard devi- ations of the study variables by conditions and time points, along with the number of participants in each time point and condition. The descriptive stat- istics indicate that overall willingness to engage in preventative behaviours in our study was high (M = 6.01, SD = .92). The means in Table 2 indicate that par- ticipants in the gratitude recall condition were most likely to engage in preventative behaviours, whereas the scores were the lowest, although still very high, in the neutral recall condition. The values in the follow-up measure (time point 7) were consistently the lowest within each condition, but also had the highest a broader dispersion of values in the follow-up. standard deviation, which suggests Manipulation checks. First, to assess the success of our emotion recall tasks we examined whether experi- enced hope and gratitude were significantly different across conditions. We performed the analyses with both current and future-focused emotions; however, due to similar results future emotions analyses are presented in Appendix 1. In terms of current hope, the effect of the emotion recall condition was signifi- cant, t(553.07) = 2.74, p = .006. The values of experi- enced current hope were highest in the hope recall condition (M = 4.28, SD = 1.39), lower in gratitude recall condition (M = 4.10, SD = 1.38), and the lowest in the neutral recall condition (M = 3.76, SD = 1.51). The difference between neutral and hope recall con- ditions was significant (Est = -.51, SE = .14, p < .001), whereas the difference between hope and gratitude recall conditions was not significant (p = .28). Current hope was marginally different between the neutral and gratitude recall conditions (p = .067). The main effect of time was also significant for current hope, t . s t n o p i e m i t 7 s s o r c a l s e b a i r a v 1 y d u t s e h t n i s n o i t a e r r o c l n o s r e p - n h t i i w d n a n o s r e p - n e e w t e b d n a s c i t s i t a t s e v i t p i r c s e D . 1 e l b a T ) P B ( C C I ) 5 7 5 2 = n ( n o s r e p i n h t i W ) 1 8 3 = n ( n o s r e p n e e w t e B 6 6 . 4 8 . 6 4 . 3 5 . 2 5 . 9 5 . 9 4 . 0 5 . 6 5 . 9 5 . 4 4 . - - 0 5 . - - 0 1 4 1 − . 0 2 . 9 1 0 . 6 2 − . 1 1 − . 4 0 . 0 2 − . 3 0 − . 7 4 . 5 8 . 7 8 . 2 2 − . 5 1 − . 3 4 . 8 1 − . 6 0 − . 0 1 . 8 0 . 3 6 . 8 7 0 . 7 0 . 2 8 . 7 1 8 . 0 1 − . 4 1 − . 0 6 . 1 2 − . 1 2 − . 9 5 . - - 6 4 1 − . 6 1 . 5 1 0 . 3 2 − . 1 1 − . 4 0 . 4 2 − . 5 0 − . 4 4 . - - 4 1 1 . 4 0 . 1 5 . - 3 7 0 . 8 0 . - 2 0 1 − . - 1 - 0 1 7 1 − . 2 0 − . 5 2 − . 6 1 − . 4 5 . 1 5 . 5 2 − . 0 2 − . 3 2 . 9 1 0 . 3 2 − . 1 0 . 4 9 . 0 5 . 9 1 − . 6 0 − . 5 1 . 1 1 . 5 6 . 8 2 1 . 0 1 . 0 9 . 7 8 8 . 9 0 − . 6 1 − . 6 6 . 0 2 − . 2 2 − . 6 6 . - - 6 6 1 − . 2 0 − . 5 2 − . 2 1 − . 1 5 . - 9 1 . 5 0 0 . 6 2 − . 0 0 . - 4 7 1 . 5 0 . 4 5 . - 3 9 0 . 1 1 . - 2 2 1 − . - 1 – D S 2 9 0 . 8 3 1 . 1 5 1 . 2 6 1 . 1 6 1 . 0 5 1 . 4 4 1 . 9 5 1 . 7 5 1 . 6 4 1 . M 1 0 6 . 3 5 2 . 3 0 4 . 7 1 4 . 1 3 3 . 3 2 2 . 4 0 4 . 2 1 4 . 0 3 3 . 4 1 2 . i s r u o v a h e b . v e r P n o i t a n m i i r c s i D e d u t i t a r g t n e r r u C e p o h e r u t u F t s u g s i d t n e r r u C e p o h t n e r r u C e d u t i t a r g e r u t u F r a e f t n e r r u C r a e f e r u t u F . 1 . 2 . 3 . 4 . 5 . 6 . 7 . 8 . 9 t s u g s i d e r u t u F . 0 1 l e b a i r a V . ) d e l i a t - o w t , 5 0 . < p ( l d o b n i e r a s n o i t a e r r o c l t n a c fi n g S i i . s t n o p i e m i t n e v e s m o r f s t r o p e r i g n g a r e v a n o d e s a b e r a D S d n a M . 7 - 1 m o r f d e g n a r s e u a v l ’ l s e b a i r a v l l A . n o i t a e r r o c l s s a l c a r t n i = C C I : s e t o N 2 0 4 P . S . R U S S E L L E T A L . Table 2. Number of participants, means, and standard deviations of the study variables across 7 time points, by conditions for study 1. Condition Time Neutral Hope Gratitude 1 2 3 4 5 6 7 1 2 3 4 5 6 7 1 2 3 4 5 6 7 Preventative behaviours SD M Future hope SD M Current gratitude M SD Future fear SD M 5.94 5.91 5.94 5.95 6.17 6.07 5.34 5.99 5.98 6.08 6.04 6.16 6.17 5.45 6.17 6.20 6.18 6.14 6.19 6.25 5.65 0.74 0.86 0.91 0.81 0.74 0.87 1.14 0.93 0.99 0.83 0.89 0.96 1.05 1.18 0.73 0.67 0.78 0.94 0.94 0.85 1.04 3.28 3.17 3.90 3.94 3.82 4.26 4.36 3.93 3.61 4.64 4.47 4.24 4.65 4.39 3.68 3.62 4.35 4.31 4.12 4.45 4.49 1.37 1.41 1.52 1.54 1.61 1.53 1.56 1.43 1.31 1.32 1.47 1.43 1.47 1.41 1.41 1.53 1.39 1.46 1.46 1.42 1.47 3.78 3.67 3.89 4.01 4.01 4.13 4.39 3.92 3.84 4.52 4.38 4.28 4.36 4.32 4.53 4.12 4.38 4.37 4.37 4.42 4.22 1.65 1.66 1.65 1.72 1.70 1.61 1.62 1.64 1.64 1.62 1.58 1.66 1.69 1.51 1.55 1.54 1.57 1.51 1.60 1.60 1.57 3.86 3.86 3.26 3.18 3.79 3.05 3.00 3.80 3.75 3.10 3.20 3.30 2.59 2.62 3.67 3.72 3.02 3.08 3.30 2.56 2.71 1.64 1.66 1.54 1.54 1.63 1.57 1.51 1.72 1.40 1.56 1.78 1.70 1.42 1.43 1.53 1.51 1.59 1.59 1.52 1.50 1.32 Current disgust M 2.60 2.70 2.18 2.15 2.37 2.13 2.18 2.60 2.54 2.06 2.21 2.23 1.86 1.74 2.33 2.33 1.98 2.03 2.16 1.86 1.93 SD 1.62 1.63 1.34 1.44 1.62 1.46 1.39 1.59 1.61 1.38 1.57 1.76 1.41 1.19 1.49 1.49 1.33 1.40 1.49 1.31 1.29 Current hope SD M Future gratitude M SD Current fear SD M 3.46 3.09 3.79 4.07 3.74 4.10 4.38 4.03 3.69 4.62 4.62 4.27 4.39 4.56 3.89 3.59 4.21 4.36 4.03 4.38 4.48 1.35 1.35 1.44 1.52 1.56 1.57 1.47 1.38 1.32 1.19 1.42 1.43 1.47 1.26 1.37 1.34 1.33 1.29 1.38 1.40 1.39 3.57 3.57 3.91 4.03 4.08 4.14 4.42 3.80 3.68 4.42 4.40 4.32 4.46 4.21 4.15 4.11 4.41 4.36 4.40 4.42 4.33 1.70 1.68 1.62 1.73 1.71 1.60 1.57 1.48 1.53 1.45 1.47 1.45 1.52 1.57 1.55 1.49 1.57 1.51 1.61 1.54 1.58 3.76 3.78 3.22 3.06 3.86 3.21 2.89 3.75 3.73 3.03 3.07 3.37 2.59 2.59 3.47 3.52 3.12 3.35 3.46 2.81 2.66 1.58 1.58 1.45 1.56 1.66 1.57 1.51 1.56 1.40 1.47 1.63 1.65 1.48 1.42 1.50 1.51 1.56 1.61 1.57 1.46 1.38 Future disgust M 2.49 2.55 2.15 2.15 2.17 2.04 2.03 2.48 2.59 1.97 2.05 2.16 1.78 1.69 2.32 2.29 1.88 1.98 2.07 1.77 1.77 SD 1.54 1.69 1.44 1.42 1.48 1.43 1.25 1.40 1.64 1.45 1.63 1.70 1.36 1.06 1.49 1.46 1.26 1.34 1.43 1.25 1.21 N 131 115 104 100 99 97 85 120 102 98 82 93 82 73 130 114 108 90 97 90 87 (1782.86) = 5.91, p < .001; however, the interaction effect was not significant; thus, in general people became more hopeful across time. and the for current gratitude, time was significant p < .001, There was a significant effect of emotion recall condition for current gratitude, t(533.17) = 4.01, p < .001. The highest scores of experienced current gratitude were obtained in the gratitude recall con- dition (M = 4.36, SD = 1.56), lower in the hope recall condition (M = 4.21, SD = 1.61), and the lowest in the neutral recall condition (M = 3.96, SD = 1.67). The difference between neutral and gratitude recall con- ditions was significant (Est = -.44, SE = .16, p = .019), whereas the neutral-hope (p = .14) and hope-grati- tude contrasts were not (p = .68). The main effect of t interaction (1775.27) = 4.33, between time and recall condition was significant as well, t(1777.76) = −2.99, p = .002. At time point 1 there was a significant difference between current gratitude in the neutral and gratitude recall conditions, this pattern remained until time point 7 (see Figure 1 for interaction plots and Table 2 for means). There was no difference in current gratitude experienced between the grati- tude and hope recall conditions at time points 1 and 7, but between time points 3 through 6 people experienced more current gratitude in the gratitude recall condition than the hope recall con- dition. Further manipulation checks were also con- ducted, with the inclusion of the other positive emotions as a covariate, time, condition, and inter- action between time and condition factors (see the other positive Appendix 2). Controlling for reduced the effect of condition on emotions current hope, but the effect of recall condition on current gratitude remained significant. Figure 1. Interaction between time and condition for experienced current-oriented gratitude. COGNITION AND EMOTION 205 Effects of emotion recall tasks on experienced negative emotions for the effect of Next, we examined whether the hope recall triggered the greatest reduction in experienced fear (Hypoth- esis 1a) and disgust (Hypothesis 1b). For both nega- time was tive emotions, only t(1772) = −3.34, p < .001 significant, current disgust, and t(1770) = −4.54, p < .001 for current fear. The means suggested some fluctuation but, in general, people felt less fear and disgust at the final time point. Neither the main effect of emotion recall condition (current disgust: p = .12, current fear: p = .51) nor the interaction between emotion recall con- dition and time (current disgust: p = .81, current fear: p = .66) were significant. This disconfirms our first hypothesis that the hope recall would be better at reducing experienced negative emotions, in compari- son to the gratitude and neutral recall conditions. We performed these analyses for experienced future negative emotions as well, which resulted in similar effects (see Appendix 1). Effects of emotion recall tasks on preventative behaviours We next examined whether preventative behaviours were significantly different across the three conditions (Hypothesis 2) and time. When willingness to engage in preventative behaviours was the DV, in this model the effects of both time, t(1764) = −2.29, p < .022, and emotion recall condition, t(625.2) = 2.26, p = .024, were significant. However, the interaction between time and emotion recall condition was not significant, p = .72. The willingness to engage in preventative behaviours was the highest in the gratitude recall con- dition (M = 6.12, SD = .86), lower in the hope recall con- dition (M = 6.00, SD = .99), and the lowest in the neutral recall condition (M = 5.92, SD = .89). The differ- ence between neutral and gratitude recall conditions was statistically significant, Est = -.22, SE = .09, p = .036. The neutral vs hope, p = .62, and hope vs grati- tude, p = .40, contrasts were not significant. In terms of the effect of time, although scores were high in the first six time points, the results from the follow-up show that the willingness to engage in preventative behaviours dropped across all the emotion recall con- ditions (M = 5.48, SD = 1.12), but remained the highest in the gratitude recall condition, lower in the hope condition, and the lowest in the neutral recall con- dition (see Table 2 for means). 206 P. S. RUSSELL ET AL. The effect of experienced emotions on preventative behaviours (hope, gratitude, We ran multilevel models to explore the association between experienced emotions and preventative behaviors. We computed a MLM to explore if experi- enced current emotions fear, disgust) were associated with willingness to engage in preventative behaviours related to the COVID-19 pandemic, controlling for the effects of time and emotion recall task condition. Results with standar- dised and unstandardised coefficients can be found in Table 3. In this model, we found a significant nega- tive effect of time, B = -.05, p < .001, and a significant effect of emotion recall condition, B = .09, p = .043, even when entering the emotions into the model, indicating the same pattern of means across the recall tasks as the previous analyses. The results showed that experienced current grati- tude was not a significant predictor of preventative behaviours on either between- or within-person level. Experienced current hope, however, predicted preventative behaviours on the within-person level exclusively, B = .03, p = .039, which indicates that when participants reported higher current hope con- cerning the COVID-19 pandemic, they also reported in preventative higher willingness to engage behaviours. On the other hand, experienced current fear and current disgust predicted preventative beha- viours significantly on a between-person level: Fear positively, B = .025, p < .001, and disgust negatively, B = -.22, p < .001. This shows that participants who had general higher levels of experienced fear during the study, reported on average higher willingness to engage in preventative behaviours. However, higher levels of experienced disgust were associated with decreased willingness to engage in preventative action. Disgust was also a significant negative predic- tor on a within-person level, B = -.04, p = .005. The coefficient indicates that when participants reported higher feelings of disgust in the context of the pan- demic, they were less keen on engaging in preventa- tive behaviours. We also repeated these analyses with future emotions, again resulting in similar effects (see Appendix 1). Additionally, to account for the 6- month gap between time point 6 and 7, and because time point 7 did not involve a recall task, we have conducted the same analyses without the follow-up time point included in the original analyses. Results with standardised and unstandardised coeffi- cients can be found in Table 4. Most results remained similar, and they did not contradict our current the findings by condition and interpretation of Table 3. Parameter estimates for multilevel models of preventative behaviours as a function of current-oriented emotions (final time point included). Fixed effects (intercepts, slopes) Intercept Time Condition Level 1 (within-person) 26.98 −13.14 2.03 5.36 −0.05 0.09 6.04 −0.05 0.09 4.97 −0.06 0.00 5.76 −0.04 0.17 <.001 <.001 .043 .20 .004 .04 SE/SD CI UL CI LL β B p T Hope Gratitude Fear Disgust Level 2 (between-person) Hope Gratitude Fear Disgust Random effects Level 1 (within-person) Residual Level 2 (between-person) 0.03 0.01 0.02 −0.04 0.04 0.05 0.25 −0.22 0.22 .01 .01 .01 .01 .04 .04 .03 .03 .47 2.08 0.63 1.80 −2.86 0.90 1.21 7.60 −6.43 0.03 0.01 0.02 −0.04 0.04 0.06 0.32 −0.27 .039 .531 .073 .005 .368 .226 <.001 <.001 0.00 −0.02 −0.00 −0.07 −0.05 −0.03 0.18 −0.29 0.06 0.04 0.05 −0.01 0.12 0.12 0.31 −0.15 - 0.22 - 0.46 0.50 Intercept Hope Gratitude Fear Disgust 0.72 0.13 0.15 0.08 0.08 Notes: B = unstandardised estimates; SE = standard error; SD = standard deviation; β = standardised estimates; CI = 95% confidence interval; LL = lower limit; UL = upper limit; significant coefficients are in bold (p < .05, two-tailed). For fixed effects, SE have been reported. For random effects, SD have been reported. 0.61 0.03 0.07 0.00 −0.00 0.44 0.01 0.01 0.00 0.00 0.44 0.01 0.01 0.00 0.00 .66 .08 .10 .05 .05 - - - - - - - - - - COGNITION AND EMOTION 207 Table 4. Parameter estimates for multilevel models of preventative behaviours as a function of current-oriented emotions (final time point not included).6 Fixed effects (intercepts, slopes) Intercept Time Condition Level 1 (within-person) 4.84 0.01 −0.00 <.001 <.001 .078 26.50 3.76 1.76 5.25 0.02 0.08 5.64 0.04 0.16 5.85 0.02 0.08 .20 .01 .01 SE/SD CI UL CI LL β B p T Hope Gratitude Fear Disgust Level 2 (between-person) Hope Gratitude Fear Disgust Random effects Level 1 (within-person) Residual Level 2 (between-person) 0.02 0.02 0.02 −0.03 0.05 0.03 0.24 −0.22 0.17 .01 .01 .01 .01 .04 .04 .03 .03 .41 2.02 1.38 1.54 −2.25 1.04 0.79 7.44 −6.58 0.03 0.02 0.02 −0.03 0.05 0.04 0.31 −0.27 .044 .167 .123 .025 .301 .429 <.001 <.001 −0.00 −0.01 −0.00 −0.06 −0.04 −0.04 0.18 −0.28 0.05 0.04 0.04 −0.00 0.13 0.10 0.30 −0.16 - 0.17 - 0.40 0.44 Intercept Hope Fear Disgust 0.72 0.12 0.06 0.11 Notes: B = unstandardised estimates; SE = standard error; SD = standard deviation; β = standardised estimates; CI = 95% confidence interval; LL = lower limit; UL = upper limit; significant coefficients are in bold (p < .05, two-tailed). For fixed effects, SE have been reported. For random effects, SD have been reported. 0.61 0.04 −0.02 0.04 0.44 0.01 0.00 0.01 0.44 0.01 0.00 0.01 .67 .08 .04 .08 - - - - - - - - emotions. However, after removing the final time point, the effect of time was no longer a negative pre- dictor of willingness to engage in preventative beha- viours, but a positive predictor of willingness to engage in preventative behaviours. Discussion 2022; Syropoulos & Markowitz, In this longitudinal study that spanned a year, we found that a gratitude recall task increased willing- ness to engage in health preventative behaviours, more so than the baseline neutral recall task. This aligns with prior research showing the positive impact of gratitude in the context of COVID-19 (Datu et al., 2022; Dennis et al., 2022; Fekete & Dei- chert, 2021). However, the gratitude and hope recall tasks had similar impacts on preventative action, which aligns with prior research showing that gratitude has an equally positive impact as other positive manipula- tions, i.e. kindness (Datu et al., 2022). We found this result over the course of time points when the partici- pants engaged in the recall task, but not when inves- tigating the six-month follow-up which did not encompass a recall task (see Table 4). This may suggest that it is important to actively engage in a task to see an effect. gratitude or hope recall Another reason for this result may be that the final time point took place during a period where restric- tions were being relaxed; thus, the context itself may have decreased willingness to engage in preven- tative action. tasks With regards to the emotions participants reported experiencing (i.e. levels of gratitude, hope, disgust, and fear), we found that gratitude and hope recall tasks elicited respective emotions more so than the neutral condition; however, participants in the posi- tive emotion recall conditions experienced similar levels of gratitude and hope. Therefore, both positive emotion recall resulted in more positive emotional experiences. We also found that disgust and fear did not differ by recall conditions; thus, we did not find support for our hypothesis that a hope recall task would be the most efficacious at reducing Instead, both fear and disgust negative emotions. were significantly different across time and in general were lower at the final time point, suggesting that these emotions were influenced more by external and societal events than the recall task itself. We also explored whether experienced emotions (hope, gratitude, fear, disgust) were associated with willingness to engage in preventative behaviours related to the COVID-19 pandemic, whilst controlling for time and condition in the model. Overall, we found that the effects of time and condition remained sig- nificant even when including levels of experienced 208 P. S. RUSSELL ET AL. emotions in the model. The results suggested that on the within person level, hope but not gratitude was related to preventative action; thus, at times when an individual felt more hope they reported greater willingness to engage in COVID-19 preventative beha- viours. However, hope was not related to preventative action at the between person level, and the effects of the negative emotions were much stronger. At between person level, we found that disgust was related to less preventative action while fear was related to more preventative action. This suggests that negative emotions have a strong relationship with preventative action, even despite positive emotions experienced and elicited. These findings also support our assumption that fear is more likely to be a positive predictor of preventative action (e.g. Nabi & Myrick, 2022; Zingora et al., 2022), while disgust is more likely to be associated with defensive responses (Herek & McLemore, 2013). Thus, we did not find support for disgust promoting avoidance and preventative action (Dasborough et al., 2020; Ekman, 1999; Greenbaum et al., 2020; Rozin & Fallon, 1987). Study 2 To assess the effects of gratitude and hope recall tasks observed in the longitudinal study and address some limitations from this study, we conducted a follow-up study that involved a recall task and was conducted at a time (late February 2022) when restrictions were relaxed, and residents of the United Kingdom were the presence of COVID-19. To asked to accept extend Study 1 findings, we also examined whether the gratitude (vs. hope) recall needed to focus on aspects related to COVID-19 specifically, or whether feeling these emotions more generally would have an impact on willingness to engage in preventative action in the future. This seems important to test as incidental emotion effects, i.e. unrelated to the situ- ation, can be different from integral emotion effects, i.e. directly related to the situation (see Polyportis et al., 2020). Given the time in which the study was conducted, we decided to have all emotion recall tasks and measures focused on the future, which seemed rel- evant as we did not see large differences between experienced current and future emotions in Study 1. This also seemed beneficial as the neutral recalling task was also focused on the future. Like the longitudi- nal study, we also examined relationships between currently experienced hope, gratitude, fear, and disgust (felt in relation to what the future may be) with preventative action, to see if these emotions had similar relationships with preventative action during this time. Method Design This study utilised a 5 Emotion Recall Condition (Grati- tude COVID-19 vs. Gratitude General vs. Hope COVID- 19 vs. Hope General vs. Neutral) between-participants design. We examined the impact of the emotion recall tasks on preventative behaviours and emotions (grati- tude, hope, fear, and disgust) experienced in relation to what the future may be. Participants Based on the longitudinal study, we conducted a G*Power (Faul et al., 2009) a-priori power analysis assuming an effect size of 0.20, with a power of 0.95 and α of .05, performing ANOVA analysis. This indi- cated that an adequate sample size would be 470. We recruited from Prolific, participants had to be British and not having completed our pilot study or main longitudinal study. Participants were mostly female (86%).4 In terms of age, there was a wide age range (Mage = 37.67, SDage = 12.25, range: 18-75). Materials and procedure Participants were first presented with an information sheet and consent form. They then completed demo- graphic items for their age and gender. Participants were then randomly assigned to one of five emotion recall conditions. Two conditions asked them to recall five things that made them feel either hopeful or grateful for the future in relation to the COVID-19 pandemic. Alternatively, for the two other emotion conditions they were asked to recall five things that made them feel either hopeful or grateful for their future in general. For the neutral condition, they recalled five things that they planned to do the fol- like in Study 1. The emotion lowing Wednesday, recall wording for the COVID-19 specific focus and general focus were as follows: COVID Instructions: Please describe 5 events, situations, episodes or objects that make you feel grateful/hopeful for what the future of the COVID-19 pandemic may be like. Describe in detail how these events, situations, COGNITION AND EMOTION 209 episodes or objects make you feel and why you feel this way considering the context of the COVID-19 pandemic. Table 6. Correlations for all measures study 2. Preventative Behaviour Hope Gratitude Fear General Instructions: Please describe 5 events, situations, episodes or objects that make you feel grateful/hopeful for what the future may be like. Describe in detail how these events, situations, episodes or objects make you feel and why you feel this way. Hope Gratitude Fear Disgust *p < .05. ** p < .01. .02 .13** .34** -.08 .69** -.24** -.28** -.09* -.24 .40** After the recall task, participants self-reported how much they were currently experiencing several dis- tinct emotions in relation to what our future may be like because of the COVID-19 pandemic, specifically they were asked to indicate how much you currently feel the following emotions in relation to what our future may be like because of the COVID-19 pandemic. We used the same emotion terms for gratitude, hope, fear, and disgust as in Study 1. The emotion items were rated on a Likert scale from 1 (Not at all) to 7 (Extremely). Participants were also asked about their willingness to engage in ten preventative health behaviours in the future, on a scale from 1 (Not at all likely) to 7 (Extremely likely), using similar measures as those from Study 1 except we now used 10 items. The scale was found to be reliable, Cronbach α = .89. The wording of all measures can be found on OSF. Descriptive statistics Means and standard deviations for the emotion and preventative behaviours can be found in Table 5. We found that gratitude and fear were correlated with willingness to engage in preventative behaviour but hope and disgust were not significantly related to preventative behaviours, see Table 6. Effects of emotion recall tasks We conducted a MANOVA, entering experienced emotions as dependent variables (gratitude, hope, fear, and disgust) and emotion recall condition (Grati- tude COVID-19, Gratitude General, Hope COVID-19, Hope General, Neutral) as the independent variable. The main effect of emotion recall condition was found to have a significant impact on experienced emotions, Pillai V = .06, F(4, 495) = 1.77, p = .03, ɳp indicated significant 2 = .01. Univariate analysis effects on experienced gratitude, F(4, 495) = 3.32, p = .01, ɳp 2 = .02, and fear, F(4, 495) = 2.65, p = .03, ɳp 2 = .02; however, the effect of recall condition on experienced disgust statistically significant, F(4, 495) = 0.34, was not p = .85, ɳp 2 = .03, hope, F(4, 495) = 2.82, p = .03, ɳp 2 = .003. Post-hoc comparisons, indicated that levels of experienced hope were lower in the neutral condition in comparison to the gratitude COVID-19, p = .007, gratitude general, p = .007, and hope COVID-19 con- ditions, p = .04. Levels of hope were marginally different in the hope general condition than the grati- tude COVID-19, p = .08, and gratitude general con- ditions, p = .09, no other effects were statistically significant, all ps > .29. We found that experienced gratitude was higher in the gratitude COVID-19 con- dition in comparison to both the hope general, p = .04, and neutral conditions, p = .004. Gratitude was also higher in the gratitude general condition in com- parison to both the hope general, p = .04, and neutral conditions, p = .003. No other comparisons between conditions on levels of gratitude were found to be sig- nificant, all ps > .12. We found that levels of fear were lower in the gratitude COVID-19 condition in compari- son to both the gratitude general, p = .04, and hope general, p = .02, conditions. Fear was also lower in the hope COVID-19 condition in comparison to both the gratitude general, p = .03, and hope general, p = .02, conditions. However, the COVID-19 conditions did not lead to lower levels of fear in comparison to the neutral condition, all ps > .13, though the hope Table 5. Means and standard deviations for study 2. Dependent Variable Gratitude COVID Gratitude General Hope COVID Hope General Hope Gratitude Fear Disgust Preventative Behaviour 4.80 (1.24) 4.58 (1.50) 3.08 (1.27) 2.12 (1.39) 5.02 (1.16) 4.78 (1.19) 4.58 (1.68) 3.52 (1.61) 2.21 (1.31) 5.32 (1.19) 4.66 (1.30) 4.21 (1.68) 3.06 (1.56) 2.28 (1.57) 4.86 (1.46) 4.45 (1.59) 4.09 (1.70) 3.58 (1.55) 2.28 (1.40) 5.15 (1.36) Neutral 4.26 (1.58) 3.90 (1.75) 3.20 (1.52) 2.34 (1.46) 4.89 (1.36) 210 P. S. RUSSELL ET AL. general condition was marginally different, p = .08. For all means across recall conditions see Table 5. Next, we conducted an ANOVA with preventative behaviours as the dependent variable and emotion recall condition (Gratitude COVID, Gratitude General, Hope COVID, Hope General, Neutral) as the indepen- dent variable. The main effect of emotion recall con- dition was found to be only marginally significant, F (4, 495) = 2.18, p = .07, ɳp 2 = .02, with means suggesting that preventative behaviours were highest in the gratitude general condition.5 Emotions and preventative behaviours We tested across the sample whether any of the experienced emotions were associated with preventa- tive behaviours, since the emotion recall task only had a marginal impact on willingness to engage in preven- tative behaviours. Specifically, we conducted a mul- tiple regression analysis with measured emotions (hope, gratitude, fear, and disgust) as predictors of willingness to engage in preventative behaviours. The overall model was significant, R2 = .18, F(4, 495) = 26.99, p < .001. We found that experienced grati- tude, β = .14, t(495) = 2.45 p = .02, and fear, β = .43, t(495) = 9.48, p < .001, were associated with greater willingness to engage in preventative action. Whilst greater levels of disgust was related to less willingness to engage in preventative action, β = -.23, t(495) = −5.00, p < .001, and hope was not a significant predic- tor of preventative action, β = -.04, t(495) = −0.70, p = .48. Discussion We found the recalling tasks to have an overall mar- ginal effect on preventive behaviours, suggesting par- ticipants’ tendency to report the highest willingness to engage in preventive action in the gratitude general recall condition. The impact of the general gratitude condition on preventative action may have resulted because COVID-19 restrictions were relaxed at this point; however, this result should be inter- preted cautiously as the overall effect of condition was marginally significant. Collapsing across the experimental conditions, we found that levels of experienced gratitude and fear were associated with greater willingness to engage in preventative behav- iour, while disgust was again shown to promote less willingness to do so, and we found that hope, contrary to Study 1, was unrelated to preventative action in this context where restrictions had been lifted. Thus, during this time when restrictions were being relaxed in the UK, we found again that experienced gratitude and fear were positively related to preven- tive action. In terms of experienced positive emotions, we again found less differentiation in terms of the different recall tasks, though the positive emotions were always higher in the positive emotion conditions than the neutral condition. Interestingly, we found that levels of disgust did not differ by the recall tasks; however, fear was lower when participants recalled positive emotions specifically related to COVID-19, but fear was not lower when recalling posi- tive emotions in general. This may suggest that posi- tive emotions can be used to counteract fear, but they need to be specific in focus, i.e. elicited integrally. General discussion Contrary to our initial pre-registered hypotheses, we found that a gratitude recall task increased willing- ness to engage in COVID-19 preventative behaviours, more so than the baseline neutral recall task. We found that the gratitude recall task had this impact across the study time points, and even found that recalling gratitude had a small impact at a time when restrictions were relaxed. This research suggests that gratitude can be used in a positive way in the context of COVID-19, which aligns with prior research (e.g. Dennis et al., 2022), and can facilitate preventa- tive behaviours. This research extends past research demonstrating the multitude of ways that gratitude can be good for us, such as increasing well-being (Jans-Beken et al., 2020) and suggests that gratitude can have the potential to impact health preventative behaviours even outside of the context of COVID-19. However, we did not find that the gratitude and hope recall tasks had dissimilar effects, which aligns with other prior research showing that positive inter- ventions having similar effects, i.e. gratitude versus kindness (e.g. Datu et al., 2022). In both the longitudinal and follow-up study, we found less differentiation in terms of experienced state hope and gratitude (i.e. experienced or emotions). Participants generally felt more positive emotions in the positive emotions recall conditions than the neutral condition, and gratitude was slightly higher in respective conditions, but hope was high in both emotion recall conditions. However, when exam- ining associations between levels of the experienced emotions and preventative action we found gratitude to be a more consistent predictor of preventative action, as gratitude but not hope was found to be a significant predictor in the follow-up study. Addition- ally, in the longitudinal study we found that hope was related to preventative action only at the within person level. Thus, even though these experienced emotions were not that different by the positive emotion recall tasks, they showed unique relation- ships with COVID-19 preventative action in the different contexts. On the other hand, we did find disgust and fear to have different impacts in the case of the COVID-19 pandemic. We found that disgust was related to less preventative action, while fear was associated with more preventative action, and this was found in both the longitudinal and follow-up study. This may suggest that disgust triggers feelings of cer- tainty (Tiedens & Linton, 2001), and that people can It may also relax their health protective response. suggest that disgust is more closely linked to defen- sive responses (Herek & McLemore, 2013; Rozin & Fallon, 1987). In comparison, during the uncertain time of COVID-19, experienced fear increased willing- ness to engage in preventative action (see Zingora et al., 2022 for COVID-19 fear and anti-COVID-19 behaviours). This may suggest focused attention to the situation and promoted thinking cri- tically about what individuals need to do to reduce the risks from COVID-19 (Polyportis et al., 2020). Encouragingly, this pattern of results persisted even in the follow-up study which was a time when COVID-19 restrictions were relaxed. Across both studies, we also found that experienced disgust and fear were more strongly related to willingness to engage in preventative behaviours (though in opposite directions) than positive emotions were, i.e. gratitude and hope. We also found that levels of disgust and fear were not influenced by the emotion recall tasks. that fear Implications This research demonstrates that positive emotions were associated with preventative action, and there- fore can potentially be used to facilitate positive out- comes across multiple time points (that is, over time with longitudinal implications), which has both theor- etical and practical value, as their roles in facilitating preventive behaviours can be important when designing future interventions. As we found that COGNITION AND EMOTION 211 gratitude impacted willingness to engage in preven- tative behaviours, beyond a single time point, which is typical of positive emotion research (e.g. Bartoș et al., 2020). This is an important contribution to the field of emotion research, and it has wider impact as very little longitudinal research has been conducted in this domain. This research shows that simple emotion recall tasks on a monthly basis can be ben- eficial. Importantly, the recall task must be engaged in and can be both general or specific to the outcome itself, though the influence that general versus specific instructions have, are likely to differ by how much uncertainty and personal choice is present. Thus, gratitude recall can be effective at mul- tiple time points, but instructions may need to be adapted for the context. that The current results suggest the specific emotion of gratitude can be used to facilitate positive action. Therefore, gratitude is likely to be a useful tool in years to come, as we heal from the COVID-19 pan- demic. This research directly suggests that if individ- uals engage in frequent gratitude recall this can help them, even in such negative circumstances, results on individuals’ thus extending previous mental well-being (Geier & Morris, 2022) to beha- viours that can also benefit other individuals. Hope was also related to preventative action but only in the longitudinal study on the within-person level, and the effects of the hope recall task were not different than the neutral task in either study. The theoretical implications of these positive emotion effects suggests that gratitude may be a slightly better candidate for fostering social change in this context, but further research is needed to disentangle hope and gratitude’s unique effects on preventative action. It would also be useful to examine whether other factors besides the emotion recall tasks were contributing to willingness to engage in preventative action. Furthermore, these positive emotion recall tasks, further research should aim to these examine whether emotions, associated appraisals or motivational ten- dencies triggered by the recall tasks that have a posi- tive impact. For example, gratitude interventions have been shown to have an impact in numerous domains, such as wellbeing and prosocial/helping behaviour (Bartlett & DeSteno, 2006; Jans-Beken et al., 2020; McCullough et al., 2008; Paramita et al., 2020), but we do not know what about these tasks have an impact, i.e. emotional, cognitive, or behavioural factors. is the experience of in terms of it 212 P. S. RUSSELL ET AL. Focusing on negative emotions, this research shows that disgust may be counterproductive in facil- itating COVID-19 preventative action, whilst fear can be associated with an increase in positive action. This suggests that media campaigns and discourse should focus on reducing disgust. These findings also have theoretical value as they suggest that disgust and fear can be related to different behav- ioural responses. The findings also allude to the importance of the appraisal of certainty as suggested by the appraisal tendency framework, as fear and disgust differ on the appraisal of certainty, which in this context may have led to different relationships between disgust and fear with preventative action. Cumulatively, the results also stress the important influence that fear can have on our preventative beha- viours even independently of a positive emotion recall task. These results are in line with negativity effects highlighted by previous research, whereby negative emotions, as opposed to positive emotions, signal a problem and the need for action (e.g. Schwarz, 1990; Taylor, 1991). Thus, campaigns to change health behaviours, even outside of the COVID-19 pandemic, should examine how to impact levels of fear and disgust. This is important as fear and disgust are likely to have differential influences on health protective responses. Across both studies, it seems important that current gratitude and fear are channeled within health promotion campaigns along with informing individuals of what they can do to cope with the situation (Maddux & Rogers, 1983; Petty, 1995; Tannenbaum et al., 2015). This will lead to feelings of certainty and that change is poss- ible. However, we should still be cautious when focus- ing on fear, as other research has found that fear can have amplifying effects of the infectious disease and it can lead to stigma (e.g. Ahorsu et al., 2020). Limitations and future research Even though the proposed research has implications both in the context of COVID-19 and other societal implications, there are some limitations which need to be highlighted. First, we had a planned strategy and theoretical rationale for the time points used in this study. However, it would have been useful to examine these emotions and behaviours more closely whilst restrictions were being released, i.e. February 2021- August 2021. It may have also been useful to examine if more frequent recalls had a more positive impact, rather than having monthly time points. This seems to be important to test since we did not find large differences between our hope and gratitude recall tasks. In this research, we used recall tasks to have paral- lel instructions across conditions. However, recall tasks rely on people being able to recall certain emotions, which may have been more difficult during certain times of the research period, especially in the context of COVID-19 societal events and restric- tions. Additionally, even though they were recalling positive experiences some of the events may have been linked to different stressors (Mills & D’Mello, 2014). Another point to consider is that the control condition focused on recalling future plans, while in the main study the emotion recall tasks focused on current emotions and in the follow-up study on future emotions only. Thus, future research should endeavour to disentangle the effects of time and emotion more closely. Emotion recall tasks have been shown to be impactful in numerous contexts (see Mills & D’Mello, 2014 for a review) but may have been less effective in our studies. Hence, it would be useful if future research explores other ways to induce these emotions, such as videos or vignettes that could resemble campaigns delivered at the national level. This research also relied on self-reports, which can be problematic, as it can be susceptible to social desir- ability biases and people sometime struggle to accu- rately reflect on their emotions (Schwarz, 2012). Even though it has been demonstrated that intentions translate into actual action in the context of COVID- 19 (Gollwitzer et al., 2021), it would have still been beneficial to measure direct action. However, we used self-report measures to examine these outcomes in an appropriate sample size and across numerous it is important to further time points. Additionally, consider individual differences which may impact par- ticipants’ emotions and how they engaged in the recall task; for example, it has been found that narcis- sism impacts how message framing influences willing- ness to engage in preventative action (Otterbring et al., 2021). Specifically, they found that individuals high in narcissism responded more to a negative framing (than positive framing), resulting in more pre- ventative action. Additionally, contextual factors such as having more family and friends (i.e. more social support), may have impacted participants’ engage- ment with the task. Finally, our sample mostly con- sisted of women. Previous research (e.g. Brebner, 2003) has shown that gender plays a role in future studies experienced and reported emotions, such as the experience of gratitude (e.g. Kashdan et al., 2009). should consider a more Hence, gender balance sample and potentially compare gender differences. Though overall the research does suggest which certain emotions, induced gratitude, hope, and experienced fear, are likely to promote preventative action and when they are most likely to do so. i.e. Conclusion Contrary to our pre-registered hypothesis a simple gratitude recall task was found to encourage COVID- 19 preventative behaviours, more so than neutral events in our main study. We also found positive associations between experienced hope and grati- tude with preventative action in the different studies. Across both studies we found experienced fear was positively related with preventative action, while experienced disgust was negatively related to action. This research should hopefully encourage others to think of what they are currently grateful for as this can be a positive social tool, even when faced with adversity and change during the COVID- 19 pandemic. Notes 1. https://osf.io/upb2h/?view_only=6eb16e9373c642f7842 54972f42b449c 2. There were other measures included, which can be accessed through OSF but that are not considered in this manuscript, https://osf.io/upb2h/?view_only= 6eb16e9373c642f784254972f42b449c 3. Since there were more female participants in the sample, we tested the interaction between recall task condition and gender in the first time point on experienced grati- tude, and the interaction was not significant. We have tested the interaction between experimental condition and gender in the first time point, and the interaction was not significant (p = .25), and neither was the main effect of gender on preventative behaviors (Mmales = 6.08, Mfemales = 6.03, p = .61). 4. Gender and gratitude were not related in Study 2 either, as there was no difference in experienced gratitude by gender p = .15, and if gender was included in our preven- tative behavior analysis it did not have a large impact on the effect of recall task on preventative behaviors and the effect of the covariate of gender was p = .054. 5. Given that the effects in Study 1 were contrary to our original hypothesis, we also examined the post-hoc com- parisons, which indicated that the gratitude general con- dition elicited greater willingness to engage in preventative action in comparison to the neutral COGNITION AND EMOTION 213 condition p = .02, and hope COVID-19 condition, p = .01. No other comparisons were significant, all ps >.10. 6. After removing the final time point from the data set, the model failed to run, because the number of observations ( = 1852) was not enough to produce the number of random effects ( = 1935) for the random intercept and random slopes for the four predictors, clustered by par- ticipants. We have thus decided to exclude gratitude from random effects, because the fixed effects of this variable were not significant on either, within- or between-person level of the multilevel model. See analysis code for details. Disclosure statement No potential conflict of interest was reported by the author(s). Funding This work was supported by British Academy [grant number SRG20\201448]. Ethics statement All participants in our experiments were treated in accordance with the ethical standards of the British Psychological Society and the American Psychological Association. This research received ethical approval from the University of Surrey. ORCID Michal Frackowiak Patrice Rusconi Fabio Fasoli http://orcid.org/0000-0002-0995-8827 http://orcid.org/0000-0002-0095-4061 http://orcid.org/0000-0002-8043-1630 Data availability statement The data that support the findings of this research are openly available in the Open Science Framework (OSF) through this link https://osf.io/upb2h/?view_only = 6eb16e9373c642f78425 4972f42b449c. References Ahorsu, D. K., Lin, C. Y., Imani, V., Saffari, M., Griffiths, M., & Pakpour, A. H. (2020). The fear of COVID-19 scale: Development and initial validation. International Journal of Mental Health Addiction, 20, 1537–1545. doi:10.1007/ s11469-020-00270-8 Algoe, S. B. (2012). Find, remind, and bind: The functions of grati- tude in everyday relationships. Social and Personality Psychology Compass, 6(6), 455–469. doi:10.1111/j.1751-9004. 2012.00439.x An, S., Ji, L. J., Marks, M., & Zhang, Z. (2017). Two sides of emotion: Exploring positivity and negativity in six basic emotions across cultures. Frontiers in Psychology, 8, 610. doi:10.3389/fpsyg.2017.00610 214 P. S. RUSSELL ET AL. Averill, J. R. (1994). In the eyes of the beholder. In P. Ekman & R. J. Davidson (Eds.), The nature of emotion: Fundamental question (pp. 7–15). Oxford University Press. Bachkirov, A. A. (2015). Managerial decision making under specific emotions. Journal of Managerial Psychology, 30(7), 861–874. doi:10.1108/JMP-02-2013-0071 Bartlett, M. Y., & DeSteno, D. (2006). Gratitude and prosocial behavior helping when it costs you. Psychological Science, 17(4), 319–325. doi:10.1111/j.1467-9280.2006.01705.x Bartoș, S. E., Russell, P. S., & Hegarty, P. (2020). Heroes against homophobia: Does elevation uniquely block homophobia by inhibiting disgust? Cognition and Emotion, 34(6), 1123– 1142. doi:10.1080/02699931.2020.1726292 Bates, D., Mächler, M., Bolker, B., & Walker, S. (2015). Fitting linear Journal of Statistical mixed-effects models using lme4. Software, 67(1), 1–48. doi:10.18637/jss.v067.i01 Baumgartner, H., Pieters, R., & Bagozzi, R. P. (2008). Future oriented emotions: Conceptualization and behavioral effects. European Journal of Social Psychology, 38(4), 685– 696. doi:10.1002/ejsp.467 Bolger, N., & Laurenceau, J. P. (2013). Intensive longitudinal methods: An introduction to diary and experience sampling research. Guilford Press. Brebner, J. (2003). Gender and emotions. Personality and Individual Differences, 34(3), 387–394. doi:10.1016/S0191- 8869(02)00059-4 Chang, Y. P., Lin, Y. C., & Chen, L. H. (2012). Pay it forward: Gratitude in social networks. Journal of Happiness Studies, 13(5), 761–781. doi:10.1007/s10902-011-9289-z Chuang, Y. C., Huang, Y. L., Tseng, K. C., Yen, C. H., & Yang, L. H. (2015). Social capital and health-protective behavior inten- tions in an influenza pandemic. PLoS ONE, 10(4), e0122970– 15. doi:10.1371/journal.pone.0122970 Cohen-Chen, S., Crisp, R. J., & Halperin, E. (2015). Perceptions of a changing world induce hope and promotes peace in intract- able conflicts. Personality and Social Psychology Bulletin, 41(4), 498–512. doi:10.1177/0146167215573210 Cohen-Chen, S., Halperin, E., Crisp, R. J., & Gross, J. J. (2014). Hope in the Middle East: Malleability beliefs, hope, and the willing- ness to compromise for peace. Social Psychological and Personality doi:10.1177/ 5(1), 1948550613484499. 67–75. Science, Cohen-Chen, S., Pliskin, R., & Goldenberg, A. (2020). Feel good or do good? A valence–function framework for under- standing emotions. Current Directions in Psychological Science, 29(4), 388–393. doi:10.1177/0963721420924770 Cohen-Chen, S., & Van Zomeren, M. (2018). Yes we can? Group efficacy beliefs predict collective action, but only when hope is high. Journal of Experimental Social Psychology, 77, 50–59. doi:10.1016/j.jesp.2018.03.016 Cranford, J. A., Shrout, P. E., Iida, M., Rafaeli, E., Yip, T., & Bolger, N. (2006). A procedure for evaluating sensitivity to within- person change: Can mood measures in diary studies detect change reliably? Personality and Social Psychology Bulletin, 32(7), 917–929. doi:10.1177/0146167206287721 Dasborough, M. T., Hannah, S. T., & Zhu, W. (2020). The gener- ation and function of moral emotions in teams: An integra- tive review. Journal of Applied Psychology, 105(5), 433–452. doi:10.1037/apl0000443 satisfaction, positive emotions, negative emotions, and COVID-19 anxiety: An online pilot experimental study. Applied Psychology: Health and Well-Being, 14(2), 347–361. doi:10.1111/aphw.12306 Dennis, A., Ogden, J., & Hepper, E. G. (2022). Evaluating the impact of a time orientation intervention on well-being during the COVID-19 lockdown: past, present or future? The Journal of Positive Psychology, 17(3), 419–429. doi:10.1080/ 17439760.2020.1858335 De Weerd, M., & Klandermans, B. (1999). Group identification and political protest: Farmers’ protest in The Netherlands. European Journal of Social Psychology, 29(8), 1073–1095. doi:10.1002/(SICI)1099-0992(199912)29:8<1073::AID- EJSP986>3.0.CO;2-K Ekman, P. (1999). Basic emotions. In T. Dalgleish & M. Power (Eds.), Handbook of Cognition and Emotion (pp. 45–60). John Wiley & Sons. Faul, F., Erdfelder, E., Buchner, A., & Lang, A. G. (2009). Statistical power analyses using G* Power 3.1: Tests for correlation and regression analyses. Behavior Research Methods, 41(4), 1149– 1160. doi:10.3758/BRM.41.4.1149 Fekete, E. M., & Deichert, N. T. (2022). A brief gratitude writing intervention decreased stress and negative affect during Journal of Happiness Studies, the COVID-19 pandemic. 2427–2448. doi:10.1007/s10902-022-00505-6 Fleeson, W. (2001). Toward a structure- and process-integrated view of personality: Traits as density distributions of states. Journal of Personality and Social Psychology, 80(6), 1011– 1027. doi:10.1037/0022-3514.80.6.1011 Fox, G. R., Kaplan, J., Damasio, H., & Damasio, A. (2015). Neural correlates of gratitude. Frontiers in Psychology, 6, 1491. doi:10.3389/fpsyg.2015.01491 Fredrickson, B. L. (2001). The role of positive emotions in positive psychology: The broaden-and-build theory of positive emotions. American Psychologist, 56(3), 218–226. doi:10. 1037/0003-066X.56.3.218 Gallagher, M. W., Smith, L. J., Richardson, A. L., D’Souza, J. M., & Long, L. J. (2021). Examining the longitudinal effects and potential mechanisms of hope on COVID-19 stress, anxiety, and well-being. Cognitive Behaviour Therapy, 50(3), 234–245. doi:10.1080/16506073.2021.1877341 Geier, M. T., & Morris, J. (2022). The impact of a gratitude interven- tion on mental well-being during COVID-19: A quasi-exper- imental study of university students. Applied Psychology: Health and Well-Being, 14(3), 1–12. doi:10.1111/aphw.12359. Genç, E., & Arslan, G. (2021). Optimism and dispositional hope to promote college students’ subjective well-being in the context of Journal of Positive School Psychology, 5(2), 87–96. doi:10.47602/jpsp.v5i2.255 the COVID-19 pandemic. Gilles, I., Bangerter, A., Clémence, A., Green, E. G., Krings, F., Staerklé, C., & Wagner-Egger, P. (2011). Trust in medical organ- izations predicts pandemic (H1N1) 2009 vaccination behavior and perceived efficacy of protection measures in the Swiss public. European Journal of Epidemiology, 26(3), 203–210. Gollwitzer, A., McLoughlin, K., Martel, C., Marshall, J., Höhs, J. M., & Bargh, J. A. (2022). Linking self-reported social distancing to real-world behavior during the COVID-19 pandemic. Social Psychological and Personality Science, 13(2), 656–668. doi:10. 1177/19485506211018132 Datu, J. A. D., Valdez, J. P. M., McInerney, D. M., & Cayubit, R. F. (2022). The effects of gratitude and kindness on life Greenbaum, R., Bonner, J., Gray, T., & Mawritz, M. (2020). Moral emotions: A review and research agenda for management scholarship. Journal of Organizational Behavior, 41(2), 95–114. doi:10.1002/job.2367 Haidt, J. (2003). The moral emotions. In R. J. Davidson, K. R. Scherer, & H. H. Goldsmith (Eds.), Handbook of affective sciences. (pp. 852-870). Oxford University Press. Haug, N., Geyrhofer, L., Londei, A., Dervic, E., Desvars-Larrive, A., Loreto, V., Pinior, B., Thurner, S., & Klimek, P. (2020). Ranking the effectiveness of worldwide COVID-19 government inter- ventions. Nature Human Behaviour, 4(12), 1303–1312. doi:10.1038/s41562-020-01009-0 Herek, G. M., & McLemore, K. A. (2013). Sexual prejudice. Annual Review of Psychology, 64(1), 309–333. doi:10.1146/annurev- psych-113011-143826 Institute for Government. (2021). Timeline of UK coronavirus lock- https://www. downs, March instituteforgovernment.org.uk/sites/default/files/timeline- lockdown-web.pdf. to March 2021. 2020 Jans-Beken, L., Jacobs, N., Janssens, M., Peeters, S., Reijnders, J., Lechner, L., & Lataster, J. (2020). Gratitude and health: An updated review. The Journal of Positive Psychology, 15(6), 743–782. doi:10.1080/17439760.2019.1651888 Kashdan, T. B., Mishra, A., Breen, W. E., & Froh, J. J. (2009). Gender differences in gratitude: Examining appraisals, narratives, the willingness to express emotions, and changes in psychologi- cal needs. Journal of Personality, 77(3), 691–730. doi:10.1111/j. 1467-6494.2009.00562.x Lazarus, R. S. (1991). Emotion and adaptation. Oxford University Press. Lenth, R. V., Buerkner, P., Herve, M., Love, J., Riebl, H., & Singmann, H. (2021). Emmeans: Estimated Marginal Means, aka Least-Squares Means. R Package Version 1.7.0. https:// CRAN.R-project.org/package = emmeans. Lerner, J. S., Li, Y., Valdesolo, P., & Kassam, K. S. (2015). Emotion and decision making. Annual Review of Psychology, 66, 799– 823. doi:10.1146/annurev-psych-010213-115043 Lerner, J. S., & Tiedens, L. Z. (2006). Portrait of the angry decision maker: How appraisal tendencies shape anger’s influence on cognition. Journal of Behavioral Decision Making, 19(2), 115– 137. doi:10.1002/bdm.515 Liberman, A., & Chaiken, S. (1992). Defensive processing of per- sonally relevant health messages. Personality and Social Psychology Bulletin, 18(6), 669–679. doi:10.1177/014616 7292186002 Luthans, F., & Jensen, S. M. (2002). Hope: A new positive strength for human resource development. Human Resource Development Review, 1(3), 304–322. doi:10.1177/1534484302013003 Maas, C. J., & Hox, J. J. (2004). Robustness issues in multilevel regression analysis. Statistica Neerlandica, 58(2), 127–137. doi:10.1046/j.0039-0402.2003.00252.x Maddux, J. E., & Rogers, R. W. (1983). Protection motivation and self-efficacy: A revised theory of fear appeals and attitude change. Journal of Experimental Social Psychology, 19(5), 469–479. doi:10.1016/0022-1031(83)90023-9 McCullough, M. E., Kilpatrick, S. D., Emmons, R. A., & Larson, D. B. Is gratitude a moral affect? Psychological Bulletin, (2001). 127(2), 249–266. McCullough, M. E., Kimeldorf, M. B., & Cohen, A. D. (2008). An adap- tation for altruism: The social causes, social effects, and social evolution of gratitude. Current Directions in Psychological Science, 17(4), 281–285. doi:10.1111/j.1467-8721.2008.00590.x COGNITION AND EMOTION 215 Mills, C., & D’Mello, S. (2014). On the validity of the autobiogra- phical emotional memory task for emotion induction. PLOS ONE, 9(4), e95837. doi:10.1371/journal.pone.0095837 Nabi, R. L., & Myrick, J. G. (2019). Uplifting fear appeals: Considering the role of hope in fear-based persuasive mess- ages. Health Communication, 34(4), 463–474. doi:10.1080/ 10410236.2017.1422847 Nash, J. C. (2014). On best practice optimization methods in R. Journal of Statistical Software, 60(2), 1–14. doi:10.18637/jss. v060.i02 Olatunji, B. O., Moretz, M. W., McKay, D., Bjorklund, F., & de Jong, P. J. (2009). Confirming the three-factor structure of the disgust scale-revised in eight countries. Journal of Cross Cultural Psychology, 40(2), 234–255. doi:10.1177/002202210 832891828918. Ong, A. D., Edwards, L. M., & Bergeman, C. S. (2006). Hope as a source of resilience in later adulthood. Personality and Individual Differences, 41(7), 1263–1273. doi:10.1016/j.paid. 2006.03.028 Otterbring, T., Festila, A., & Folwarczny, M. (2021). Selfless or selfish? The impact of message framing and egoistic motivation on narcissists’ compliance with preventive health behaviors during COVID-19. Current Research in Ecological and Social Psychology, 2, 100023. doi:10.1016/j. cresp.2021.100023 Paramita, W., Septianto, F., & Tjiptono, F. (2020). The distinct effects of gratitude and pride on donation choice and amount. Journal of Retailing and Consumer Services, 53, 101972. doi:10.1016/j.jretconser.2019.101972 Peitz, L., Lalot, F., Douglas, K., Sutton, R., & Abrams, D. (2021). COVID-19 conspiracy theories and compliance with govern- mental restrictions: The mediating roles of anger, anxiety, and hope. 15, of 183449092110466. doi:10.1177/18344909211046646 Rim Psychology, Pacific Journal Petty, R. E. (1995). Attitude change. In ed. A. Tesser (Ed.), Advanced Social Psychology (pp. 195–255). McGraw-Hill. Pliskin, R., Sheppes, G., & Halperin, E. (2015). Running for your life, in context: Are rightists always less likely to consider fleeing their country when fearing future events? Journal of Experimental Social Psychology, 59, 90–95. doi:10.1016/j.jesp. 2015.04.001 Polyportis, A., Kokkinaki, F., Horváth, C., & Christopoulos, G. (2020). Incidental emotions and hedonic forecasting: The role of (un) certainty. Frontiers in Psychology, 11, 536376. doi:10.3389/fpsyg.2020.536376 Queen, D., & Harding, K. (2020). Societal pandemic burnout: A COVID legacy. International Wound Journal, 17(4), 873–874. doi:10.1111/iwj.13441 R Core Team. (2018). R: A language and environment for statistical computing. R Foundation for Statistical Computing. https:// www.R-project.org/. Roseman, I. J. (2011). Emotional behaviors, emotivational goals, emotion strategies: Multiple levels of organization integrate variable and consistent responses. Emotion Review, 3(4), 434–443. doi:10.1177/1754073911410744 Rourke, A. (2020). Global report: WHO says Covid-19 ’may never go away’ and warns of mental health crisis. The Guardian. https://www.theguardian.com/world/2020/may/14/global- report-who-says-covid-19-may-never-go-and-warns-of- mental-health-crisis 216 P. S. RUSSELL ET AL. Rozin, P., & Fallon, A. E. (1987). A perspective on disgust. Psychological Review, 94(1), 23–41. doi:10.1037/0033-295X. 94.1.23 Russell, P.S., & Giner-Sorolla, R. (2013). Bodily-moral disgust: What it is, how it is different from anger and why it is an unreasoned emotion. Psychological Bulletin, 139, 328–351. doi:10.1037/a0029319 Schwarz, N. (1990). Feelings as information: Informational and motivational functions of affective states. In R. Sorrentino & E. T. Higgins (Eds.), Handbook of motivation and cognition: (Vol. 2, pp. 527–561). Foundations of Guilford Press. social behavior Schwarz, N. (2012). Why researchers should think “real-time”: A (Eds.), cognitive rationale. Handbook of research methods for studying daily life (pp. 22–42). Guilford Press. In M. R. Mehl & T. Conner Shaver, P., Schwartz, J., Kirson, D., & O’Connor, C. (1987). Emotion knowledge: Further exploration of a prototype approach. Journal of Personality and Social Psychology, 52(6), 1061– 1086. doi:10.1037/0022-3514.52.6.1061 Skegg, D., Gluckman, P., Boulton, G., Hackmann, H., Karim, S. S. A., Piot, P., & Woopen, C. (2021). Future scenarios for the COVID-19 pandemic. The Lancet, 397(10276), 777–778. doi:10.1016/S0140-6736(21)00424-4 Smith, C. A., & Ellsworth, P. C. (1985). Patterns of cognitive appraisal in emotion. Journal of Personality and Social Psychology, 48(4), 813–838. doi:10.1037/0022-3514.48.4.813 Smith, R., Killgore, W. D., & Lane, R. D. (2018). The structure of emotional experience and its relation to trait emotional awareness: A theoretical review. Emotion, 18(5), 670–692. doi:10.1037/emo0000376 Snyder, C. R. (2000). The past and possible futures of hope. Journal of Social and Clinical Psychology, 19(1), 11–28. doi:10.1521/jscp.2000.19.1.11 Snyder, C. R., Sympson, S. C., Ybasco, F. C., Borders, T. F., Babyak, M. A., & Higgins, R. L. (1996). Development and validation of the state hope scale. Journal of Personality and Social Psychology, 70(2), 321–335. doi:10.1037/0022-3514.70.2.321 Stevenson, R. J., Saluja, S., & Case, T. I. (2021). The impact of the COVID-19 pandemic on disgust sensitivity. Frontiers in Psychology, 11. doi:10.3389/fpsyg.2020.600761 Stotland, E. (1969). The psychology of hope. Jossey-Bass. Syropoulos, S., & Markowitz, E. M. (2021). Prosocial responses to COVID-19: Examining the role of gratitude, fairness and legacy motives. Personality and Individual Differences, 171, 110488. doi:10.1016/j.paid.2020.110488 Tannenbaum, M. B., Hepler, J., Zimmerman, R. S., Saul, L., Jacobs, S., Wilson, K., & Albarracín, D. (2015). Appealing to fear: A fear appeal effectiveness and theories. meta-analysis of Psychological Bulletin, a0039729 141(6), 1178–1204. doi:10.1037/ Taylor, S. E. (1991). Asymmetrical effects of positive and negative hypothesis. events: Psychological Bulletin, 110(1), 67–85. doi:10.1037/0033-2909. 110.1.67 mobilization-minimization The The jamovi project. (2021). jamovi (Version 2.2.2) [Computer Software]. https://www.jamovi.org/. Thorpe, S. J., Barnett, J., Friend, K., & Nottingham, K. (2011). The mediating roles of disgust sensitivity and danger expectancy in relation to hand washing behaviour. Behavioural and 175–190. doi:10.1017/ Cognitive 39(2), S1352465810000676 Psychotherapy, Tiedens, L. Z., & Linton, S. (2001). Judgment under emotional certainty and uncertainty: The effects of specific emotions on information processing. Journal of Personality and Social Psychology, 81(6), 973. doi:10.1037/0022-3514.81.6. 973 Wang, K., Goldenberg, A., Dorison, C. A., Miller, J. K., Uusberg, A., Lerner, J. S., Gross, J. J., Agesin B. B., Bernardo, M., Campos, O., Eudave, L., Grzech, K., Ozery, D. H., Jackson, E. A., Garcia, E. O. L., Drexler, S. M., Jurković, A. P., Rana, K., Wilson, J. P., & Štrukelj, E. (2021). A multi-country test of brief reappraisal interventions on emotions during the COVID-19 pandemic. Nature Human Behaviour, 5(8), 1089–1110. doi:10.1038/ s41562-021-01173-x Webb, T. L., & Sheeran, P. (2006). Does changing behavioural intentions engender behavior change? A meta-analysis of the experimental evidence. Psychological Bulletin, 132(2), 249–268. doi:10.1037/0033-2909.132.2.249 Winterich, K. P., & Haws, K. L. (2011). Helpful hopefulness: The effect of future positive emotions on consumption. Journal of Consumer Research, 38(3), 505–524. doi:10.1086/ 659873 Xue, J., Chen, J., Hu, R., Chen, C., Zheng, C., & Su, Y. (2020). Twitter discussions and emotions about the COVID-19 pandemic: Machine of learning approach. Medical Internet Research, 22(11), e20550. doi:10.2196/ 20550 Journal Yoshinobu, L., Gibb, J., Langelle, C., & Harney, P. (1991). The will and the ways: Development of an individual-differences measure of hope. Journal of Personality and Social Psychology, 60(4), 570–585. doi:10.1037/0022-3514.60.4.570 Zingora, T., Birtel, M. D., Graf, S., Hrebickova, M., Lacko, D., Rupar, M., Tocik, J., & Voca, S. (2022). Change in anti-COVID-19 behaviour and anti-immigrant prejudice during the COVID- 19 pandemic: Longitudinal evidence from five European countries. European Journal of Social Psychology, 00, 1–19. doi:10.1002/ejsp.2928 Appendices Appendix 1 Future emotions Effects of emotion recall tasks on future positive emotions We assessed the success of our manipulation on future emotions examining whether experienced hope and gratitude were significantly different across conditions and time. For future hope, the effect of the emotion recall condition was sig- nificant, t(582.29) = 2.66, p = .008. The values of experienced future hope were highest in hope recall condition (M = 4.24, lower in the gratitude recall condition (M = 4.10, SD = 1.44), SD = 1.48), and lowest in the neutral recall condition (M = 3.77, SD = 1.55). Post-hoc comparisons with Tukey correction indi- cated that the difference between neutral and hope recall con- ditions was significant (Est = -.46, SE = .14, p = .003), whereas the difference between hope and gratitude recall conditions was not significant (p = .50). The difference between neutral and gratitude recall conditions was marginally significant (p = .08). The main effect of time was also significant for future hope; t(1793.44) = 6.29, p < .001, indicating that in general people became more hopeful across time. However, the interaction between time and condition was not significant. The effect of emotion recall condition was significant for future gratitude, t(545.29) = 3.55, p < .001. For future gratitude the highest scores were obtained in the gratitude recall con- dition (M = 4.30, SD = 1.55), lower in the hope recall condition (M = 4.15, SD = 1.52), and the lowest in the neutral recall con- dition (M = 3.93, SD = 1.68). Post-hoc differences (Tukey correc- tion) for future gratitude followed the same pattern as for current gratitude. The difference between neutral recall COGNITION AND EMOTION 217 condition and gratitude recall condition was significant (Est = -.39, SE = −16, p = .036), and as for the remaining contrasts, neither the neutral-hope comparison (p = .19) nor the hope- gratitude one (p = .68) was significant. The main effect of time was also significant for future gratitude, t(1783.15) = 5.23, p <.001, indicating that in general people became more grateful across time. The interaction between time and recall condition was significant for future gratitude as well, t(1785.34) = −2.55, p = .011, resulting in similar patterns as current gratitude (see Figure A1). Effects of emotion recall tasks on future negative emotions We examined whether the hope recall triggered the greatest reduction in experienced future disgust and future fear (Hypoth- esis 1); however, only the effect of time was significant, t(1769) = −3.39, p <.001 for future disgust, and t(1774.59) = −4.10, p <.001 for future fear. The means suggesting some fluctuation but in general people felt less fear and disgust at the final time point. Neither the main effect of emotion recall condition (future disgust: p = .28, future fear: p = .70) nor the interaction between recall condition and time (future disgust: p = .86, future fear: p = .12) were significant. The effect of experienced future emotions on preventative behaviours We ran multilevel models to explore the role of experienced future emotions on preventative behaviours (see all effects in Tables A1 and A2). Time had a negative effect on preventa- tive behaviours (B = -.05, p <.001), whereas emotion recall condition had a positive significant effect (B = .10, p = .023). In this model, the only significant predictor of preventative Table A1. Parameter estimates for multilevel models of preventative behaviours as a function of future-oriented emotions (final time point included). Fixed effects (intercepts, slopes) Intercept Time Condition Level 1 (within-person) 26.10 −13.43 2.28 5.39 −0.05 0.10 5.82 −0.04 0.19 6.02 −0.05 0.10 4.98 −0.06 0.01 <.001 <.001 .023 .21 .004 .04 SE/SD CI UL CI LL β B p t Hope Gratitude Fear Disgust Level 2 (between-person) Hope Gratitude Fear Disgust Random effects Level 1 (within-person) Residual 0.04 0.01 0.02 −0.03 0.02 0.05 0.25 −0.24 0.22 .01 .01 .01 .02 .04 .04 .03 .04 .47 2.73 0.50 1.25 −1.97 0.56 1.38 7.35 −6.65 0.04 0.01 0.02 −0.03 0.03 0.07 0.32 −0.29 .007 .618 .213 .051 .575 .167 <.001 <.001 0.01 −0.02 −0.01 −0.06 −0.06 −0.03 0.18 −0.31 0.07 0.03 0.04 0.00 0.11 0.13 0.31 −0.17 - 0.22 - 0.45 0.49 Level 2 (between-person) Intercept Hope Gratitude Fear Disgust 0.72 0.13 0.10 0.12 0.15 Notes: B = unstandardised estimates; SE = standard error; SD = standard deviation; β = standardised estimates; CI = 95% confidence interval; LL = lower limit; UL = upper limit; significant coefficients are in bold (p <.05, two-tailed). For fixed effects, SE have been reported. For random effects, SD have been reported. 0.44 0.01 0.00 0.01 0.01 0.62 0.04 0.01 0.03 0.05 0.44 0.01 0.00 0.01 0.01 .67 .09 .06 .08 .10 - - - - - - - - - - 218 P. S. RUSSELL ET AL. Table A2. Parameter estimates for multilevel models of preventative behaviours as a function of future-oriented emotions (final time point not included). Fixed effects (intercepts, slopes) Intercept Time Condition Level 1 (within-person) <.001 .001 .023 25.46 3.38 2.28 5.27 0.02 0.10 4.86 0.01 0.02 5.81 0.02 0.10 5.65 0.03 0.19 .21 .01 .04 SE/SD CI UL CI LL β B p t Hope Gratitude Fear Disgust Level 2 (between-person) Hope Gratitude Fear Disgust Random effects Level 1 (within-person) Residual Level 2 (between-person) 0.03 0.01 0.02 −0.02 0.03 0.04 0.24 −0.24 0.18 .01 .01 .01 .01 .04 .04 .03 .04 .42 2.78 0.82 1.71 −1.23 0.61 1.04 7.13 −6.69 0.03 0.01 0.02 −0.01 0.03 0.05 0.31 −0.29 .006 .412 .088 .222 .540 .300 <.001 <.001 0.01 −0.01 −0.00 −0.05 −0.05 −0.04 0.17 −0.30 0.06 0.03 0.04 0.01 0.11 0.11 0.30 −0.17 - 0.18 - 0.41 0.44 Intercept Hope Fear Disgust 0.72 0.07 0.05 0.13 Notes: B = unstandardised estimates; SE = standard error; SD = standard deviation; β = standardised estimates; CI = 95% confidence interval; LL = lower limit; UL = upper limit; significant coefficients are in bold (p <.05, two-tailed). For fixed effects, SE have been reported. For random effects, SD have been reported. 0.62 −0.02 −0.02 0.04 0.45 0.00 0.00 0.01 0.45 0.00 0.00 0.01 .67 .03 .03 .09 - - - - - - - - Figure A1. Interaction between time and condition for experienced future-oriented gratitude. behaviours on a within-person level was future hope (B = .04, p = .007). When participants had higher feelings of hope regarding the future of the COVID-19 pandemic, they also reported higher preventative behaviours. Future disgust and future fear predicted preventative behaviours on a between- person level: Fear positively (B = .25, p <.001); Disgust nega- tively (B = -.24, p <.001). Participants who had higher average levels of fear during the study run, reported on average higher willingness to engage in preventative beha- viours than participants who had lower average levels of fear. Disgust followed the opposite trajectory: Participants who on average reported higher levels of disgust (vs those who reported lower levels) were less willing to engage in pre- ventative behaviours. Appendix 2 Further manipulation checks with covariates Separate manipulation checks have been conducted to predict current- and future-oriented experienced emotions hope and gratitude, with covariates included in the models. Four additional multilevel models have been computed in total, pre- dicting the following: current hope, future hope, current grati- tude, and future gratitude. Each model contained time, condition, interaction between time and condition, and a covari- ate which was an experienced emotion in a different time frame than the predicted variable, i.e. when future hope was pre- dicted, current hope was a covariate, whereas when current gratitude was predicted, future gratitude served as a covariate, etc. The covariate was centred before adding to the model. Finally, we added an interaction between condition and a cov- ariate to explore whether its effect may differ by condition. Hope In the attempt to predict current hope, the overall effect of con- dition was significant (p = .001), and the post-hoc Bonferroni pairwise comparison revealed that only neutral condition differed significantly from hope condition, in that participants in the former reported lower experiences of current-oriented feelings of hope: t(386) = −3.72, Est = -.51, SE = .14, p <.001. Hope and gratitude conditions did not differ, although the difference between neutral and gratitude approached the margin of significance (p = .09). The overall effect of time was also significant (p <.001), with Bonferroni pairwise comparisons indicating significant differences between time points, most notably the significant decrease in experienced current-oriented hope between time points 1 and 2: t(1765) = 3.74, Est = .22, SE = .06, p = .004; and the increase between time point 1 and the follow-up measure: t(1782) = −3.05, Est = -.21, SE = .07, p = .049. The interaction effect between time and condition was not sig- nificant. The covariate future hope was a positive predictor of current hope, t(1700) = 32.33, p <.001. The effect of experienced future hope did not differ across conditions, but it did approach statistical significance (p = .093) In the next model, the dependent variable was future hope, whereas the covariate was current hope. The model’s findings reflect similar trends as the previous one. Here, the effect of con- (p = .007), with the only significant dition was significant COGNITION AND EMOTION 219 pairwise comparison found between neutral condition and hope, according to the Bonferroni post-hoc correction: t(385) = −3.10, Est = -.44, SE = .14, p = .006. The overall effect of time was also significant (p <.001), the notable pairwise com- parisons show that levels of experienced future hope were sig- nificantly higher in all the time points compared to time point 1 (p <.001), apart from time point 2 which did not differ signifi- cantly. The interaction between time and condition was not sig- nificant (p = .71). The covariate, current hope, was a significant predictor of future hope, t(276) = 31.56, p <.001. The effect of the covariate on the dependent variable, however, did not differ significantly across the conditions (p = .54). Gratitude In the next model, the outcome variable was current gratitude, whilst the covariate was future gratitude. The overall effect of experimental condition was significant (p = .015), but the only significant difference was observed between neutral and grati- tude condition, higher experiences of current gratitude being t(384) = −2.84, Est = -.45, SE = .16, observed in the later, p = .014. The overall effect of time was also significant (p < .001), yet the pairwise comparison did not reveal any notable differ- ences. The interaction between time and condition was not significant (p = .20). The covariate future gratitude was a positive predictor of current gratitude, t(304) = 26.64, p <.001. The effect of the covariate did not differ across the conditions (p = .58). The final model included future gratitude as the outcome variable, and current gratitude as the covariate. The overall effect of experimental condition was significant (p = .04), and Bonferroni pairwise comparison indicates that the experiences of future-oriented gratitude were statistically higher in gratitude condition, compared to the neutral, t(384) = −2.50, Est = -.38, SE = .15, p = .038. The overall effect of time was also significant (p <.001), with pairwise comparison revealing significantly higher scores of future gratitude in the final time point compared to time point 1, t(1734) = −4.48, Est = -.33, SE = −07, p <.001. The interaction between time and condition was not significant (p = .89). Current gratitude was a significant covariate, predicting the outcome variable positively, t(380) = 32.49, p <.001. The interaction between condition and covariate was not significant (p = .12), suggesting that its predictive trend remains consistent across conditions.
10.1017_err.2022.41
European Journal of Risk Regulation (2023), 14, 78–92 doi:10.1017/err.2022.41 A R T I C L E Reasons for Reinforcing the Regulation of Chemicals in Europe Erik Millstone1* and Peter Clausing2 1Science Policy Research Unit, Jubilee Building, University of Sussex, Brighton, UK and 2Pesticide Action Network Germany, Hamburg, Germany. *Corresponding author. Email: E.P.Millstone@sussex.ac.uk Abstract The European Commission’s 2020 draft Chemicals Strategy for Sustainability set the ambitious goal of achieving a “Toxic-Free Environment”. Those ambitions were harshly criticised by a team based in Germany’s Federal Institute for Risk Assessment (or BfR); they claimed that toxicological risks from chemicals had already been minimised and were optimally regulated. This paper outlines evidence to support the Commission’s implication that the European Union’s chemicals regulatory regime is sub- optimal. It also criticises the BfR team’s contentions by reference to empirical findings (eg concern- ing tumours, congenital anomalies and the toxicity of mixtures) and by disentangling their conceptual confusions. Keywords: Chemicals Strategy; European Commission; pesticides; sustainability I. Introduction: the policy context The European Commission’s October 2020 draft Chemicals Strategy for Sustainability (CSS) was an ambitious document; it was subtitled “Towards a Toxic-Free Environment”. In its introduction, the draft CSS correctly stated that the European Union (EU) “already has one of the most comprehensive and protective regulatory frameworks for chemicals”,1 yet it aspired to create a toxic-free environment. While the EU’s standards are amongst the high- est, when compared with other jurisdictions, the Commission implicitly acknowledged that the EU’s standards could and should be higher. While it is true that the EU’s standards are often relatively high, the implementation of the EU’s regimes in practice remains prob- lematic. The CSS aspires fully to “protect environment and human health, in particular that of vulnerable groups”. Nonetheless, some commentators criticised the Commission’s CSS as unnecessary. In Institute for Risk particular, senior members of the German government’s Federal Assessment (the Bundesinstitut für Risikobewertung or BfR) have argued that the implementation of a complex and interdependent network of regulations for chemical substances, including industrial chemicals, plant protection products, bio- cides, or chemicals in food and feed has minimised toxicological risks and has 1 European Commission, “Chemicals Strategy for Sustainability Towards a Toxic-Free Environment” COM(2020) 667 final <https://eur-lex.europa.eu/resource.html?uri=cellar:f815479a-0f01-11eb-bc07-01aa75ed71a1.0003.02/ DOC_1&format=PDF>. © The Author(s), 2023. Published by Cambridge University Press. This is an Open Access article, distributed under the terms of the Creative Commons Attribution licence (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted re-use, dis- tribution and reproduction, provided the original article is properly cited. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 1 4 2 2 0 2 . r r e / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h European Journal of Risk Regulation 79 continuously increased public health and wellbeing in the EU. Moreover, although this framework already provides one of the most advanced regulatory systems worldwide, it is con- stantly pressed for improvement : : : the respective regulatory processes are optimised with respect to both their efficiency and effectiveness.2 (emphasis added) The BfR team argued that EU regulations of chemical risks are already optimal, so there is no scope – and hence no need – for any significant tightening of the prevailing regime. The BfR team judged the status quo to be entirely acceptable. The CSS envisaged attaining its goal by incentivising the replacement of unsafe chem- icals with others that will be “safe by design”; it envisaged a completely fresh trajectory for chemicals research, development and innovation. The CSS argued: “Novel and cleaner indus- trial processes and technologies would help not only to lower the environmental footprint of chemicals production but also to reduce costs, improve market readiness and create new markets for the European sustainable chemicals industry” (p 7, emphasis in the original). We applaud the CSS’s ambitions, and this paper provides a critique of the BfR team’s objec- tions to the CSS and orthodox defences of the status quo. II. Our perspective and framing assumptions Before advancing our critique, several preliminaries should be made explicit. This paper is predicated on key well-grounded assumptions about the characteristics of the issues under discussion. The topic is located at the intersection of science and policymaking. Whenever expert advisory bodies provide “risk assessments” to policymakers, both scientific and pol- icy considerations are unavoidably involved and interconnected in important ways. Examples are provided below. Nonetheless, the influences of both sets of considerations and their interactions are often not explicitly acknowledged. While many scientific advi- sory bodies and the policymakers they advise claim that their judgments are based on – and only on – scientific considerations, this is almost always misleading. As academic phi- losophers frequently observe: you cannot derive an “ought” from an “is”. If scientific advice to policymakers was entirely derived just from scientific considerations, it should always take a plural and conditional form rather than a monolithic and prescriptive form.3 For example, it could indicate what is known and not known about the consequences of following – or failing to follow – a range of possible policy options that policymakers have enumerated. If, on the other hand, the advice recommends one particular course of action (eg “adopt and implement these measures”), value judgments must have influenced and con- tributed to that advice, even if those value judgments remain implicit and unacknowledged.4 Deciding, for example, what does and does not count as a relevant risk, which types of evidence are to be deemed relevant and how much of which types of evidence are deemed variously necessary and/or sufficient to sustain judgments in favour of accepting, restricting or banning some product or process all presuppose or imply value judgments; they are not empirical findings. Similarly, how the reliability of evidence of various sorts of studies is char- acterised and adjudicated depends on socially contestable value judgments. Those value judg- ments are increasingly – and informatively – coming to be termed “risk assessment policy” judgments.5 2 M Herzler et al, “The ‘EU Chemicals Strategy for Sustainability’ Questions Regulatory Toxicology as We Know It: Is It All Rooted in Sound Scientific Evidence?” (2021) 95 Archives of Toxicology 2589. 3 A Stirling, “‘Opening Up’ and ‘Closing Down’: Power, Participation, and Pluralism in the Social Appraisal of Technology” (2008) 33 Science, Technology, and Human Values 262. 4 E Millstone, “Can Food Safety Policy-Making Be Both Scientifically and Democratically Legitimated? If So, How?” (2007) 20 Journal of Agricultural and Environmental Ethics 483. 5 ibid. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 1 4 2 2 0 2 . r r e / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h 80 Erik Millstone and Peter Clausing The approach adopted in this paper will, therefore, involve making our own key value judgments explicit and indicating how they contribute to our understanding and interpre- tation of the science and to our policy judgments and recommendations. We shall, more- over, also endeavour to make explicit some of the value judgments that have been implicit in the advice and conclusions of the institutions that we shall be discussing, in particular the European Commission and its CSS and the BfR. Chemical risk assessment policy decision-making operates in a contestable and often contested space because different groups of stakeholders have different and often conflict- ing interests. Firms that manufacture, sell and use chemicals have commercial interests that depend on, for example, national governments and the European Commission cate- gorising their compounds as acceptably safe for sale and use. If, however, those products do or might pose risks to, for example, consumers, third parties or the natural environ- ment, then the organisations focused on protecting public and environmental health often favour tight restrictions or even bans on the production, sale and use of those materials. Biomedical sciences in general – and toxicology in particular – often have to deal with biological variability within and between species and populations, and therefore they often can provide only estimates of probabilities rather than certainty or precision. This is evi- dent, for example, whenever official bodies chose to use a “weight of evidence approach”. Such decisions inevitably reflect contestable value judgments when assigning the “weights”. In practice, when official bodies claim to have applied a “weight of evidence” approach, it has often involved assigning weights to different pieces of evidence in ways that were neither consistent (as, for example, between putative false positives and puta- tive false negatives) nor transparent or justified.6 Whenever risk assessors provide advice to risk managers that recommends one partic- ular course of action, those recommendations invariably presuppose some value judg- ments, which provoke questions concerning whether those judgments favour the protection of public and environmental health or the interests of commercial organisa- tions. Our values presume prioritising the protection of public and environmental health over commercial interests, or even over seeking to balance equally between those two conflicting interests. Those values inform our analysis and assessment of the status quo, which we judge to be inadequate, and consequently we interpret the CSS as a welcome step in the right direction. III. The scientific case for the Commission’s Chemical Strategy for Sustainability Herzler et al criticised the CSS by arguing that “without a detailed assessment of which risks are currently deemed to be insufficiently addressed, it is hard to establish whether additional regulation might be necessary or existing regulation might need to be improved, and for which part of the population”.7 But that approximates to arguing that, until a quantitative risk assessment has been completed, it would be premature to regulate on the basis of partial incomplete information, which implies that precaution should never be exercised. While the CSS did not provide a detailed assessment of the hazards or risks that it highlighted, it did not claim that it would or had. The CSS did not purport to provide a purely scientific argument for the directional policy shift that it was proposing. The CSS 6 P Clausing, C Robinson and H Burtscher-Schaden, “Pesticides and Public Health: An Analysis of the Regulatory Approach to Assessing the Carcinogenicity of Glyphosate in the European Union” (2018) 72 Journal of Epidemiology and Community Health 668; E Millstone and E Dawson, “EFSA’s Toxicological Assessment of Aspartame: Was It Even-Handedly Trying to Identify Possible Unreliable Positives and Unreliable Negatives?” (2019) 77 Archives of Public Health 34. 7 Herzler et al, supra, note 2. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 1 4 2 2 0 2 . r r e / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h European Journal of Risk Regulation 81 drew attention to the limitations of the available science; it complained about what it referred to politely as “data gaps” concerning, for example, levels of exposure and under- examined toxicological endpoints. The CSS was making a value judgment about the policy implications of such data gaps and scientific uncertainties. The approach adopted in the CSS is entirely consistent with Regulation (EC) No 1107/2009 that covers pesticidal prod- ucts, which states: “The precautionary principle should be applied and this Regulation should ensure that industry demonstrates that substances or products produced or placed on the market do not have any harmful effect on human or animal health or any unac- ceptable effects on the environment.”8 There are, moreover, extensive bodies of robust evidence indicating that the prevailing chemicals regulatory regime is not sufficiently protective of public health. The same is almost certainly true for the health of ecosystems, but this paper only focuses on threats to public health. An exhaustive review of the different types of harm that chemicals may be causing is beyond the scope of this paper, but several clear examples should be sufficient. 1. Carcinogens are inadequately regulated In relation to cancer, there is extensive evidence that fatalities due to cancer have steadily diminished, but that information falls far short of providing a complete picture. It ignores the fact that the incidence of several types of commonly occurring cancers have been ris- ing steadily since, for example, the 1970s. Herzler et al highlighted the fact that “life expec- tancy and healthy life years : : : while varying between EU Member States, have been rising more or less constantly for many years”.9 But while cancer mortality rates have fallen, this has been because of improved early detection of tumours and gains in thera- peutic efficacy rather than because of official regulatory measures. If the regulatory meas- ures were sufficient, incidences would have fallen, not risen. The occurrence and diagnosis of primary tumours, even in patients who survive, inflict immense sorrow, anxiety and costs on individuals and their families, as well as impacting health systems and economies. The long-run increases in the incidence of cases of many common types of cancers show that prevailing regulatory regimes are not just suboptimal but seriously inadequate. For example, in 2015, Haberland and Wolf reported a continuous increase in the age-corrected incidences of breast cancer (2.2-fold) and prostate cancer (2.7-fold) in Germany between 1970 and 2010.10 They also reported a modest decline in fatality rates of those two types of cancer. In the UK, similar patterns have occurred. Cancer Research UK (CRUK) gathers and anal- yses relevant data and publishes tables and graphic representations of those data on its website.11 CRUK provides, for example, graphics showing the “20 Most Common Cancers” in both women and men and indicating “Percentage Changes in Age-Standardised Three Year Average Incidence Rates, UK, 2006–2008 and 2016–2018”. They clearly show that the incidence rates of many types of tumours have been rising rather than falling. Similar patterns have emerged in many other jurisdictions. 8 European Parliament and Council, “Regulation (EC) No 1107/2009 of 21 October 2009 Concerning the Placing of Plant Protection Products on the Market and Repealing Council Directives 79/117/EEC and 91/414/EEC” (2009) the European Union 1 <http://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX: Official 32009R1107>. Journal of 9 Herzler et al, supra, note 2. 10 J Haberland and U Wolf, “Trendanalysen zur Inzidenz und Mortalität an Krebs in Deutschland seit 1970” (2015) 11 GMS Medizinische Informatik, Biometrie und Epidemiologie Doc03. 11 Cancer Research UK, “Cancer incidence for common cancers” <https://www.cancerresearchuk.org/health- professional/cancer-statistics/incidence/common-cancers-compared#heading-Three>. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 1 4 2 2 0 2 . r r e / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h 82 Erik Millstone and Peter Clausing Table 1. Prevalence of congenital anomalies per 10,000 births (five-year arithmetic means) in the European Union for the periods 1980–1984 and 2015–2019 for selected groups of anomalies. Anomalies All anomalies Congenital heart defects Respiratory anomalies Abdominal wall defects Genital anomalies 1980–1984 2015–2019 Increase 210 46.9 1.66 3.31 15.5 258 81.1 4.16 4.57 20.2 ×1.2 ×1.7 ×2.5 ×1.4 ×1.3 Note: We report five-year arithmetic means as indicators of prevailing trends. Source: <https://eu-rd-platform.jrc.ec.europa.eu/eurocat/eurocat-data/prevalence_en>, where more details can be found. 2. Endocrinological and reproductive toxicity remains insufficiently controlled In relation to reproductive toxicology, Herzler et al implied that the status quo was unproblematic because “the prevalence of anomalies in the newborn in the EU has been practically constant (at around 200/10,000) for the past 40 years”, although they acknowl- edged that they did not include genetic anomalies or adjust for parental age.12 An exami- nation of the database used by Herzler et al shows that their statement was misleading; the evidence for this claim is provided in Table 1. Firstly, the prevalence of all anomalies in new-borns in the EU has not been “practically constant”. On the contrary, it has increased slowly but continuously, with a 20% increase for the period of 2015–2019 (last year available) compared to 1980–1984. The data show even greater increases for four specific types of anomalies, although some declined over that interval. The reasons for those changes need to be elucidated, but the overall preva- lence increased rather than remaining practically constant. Furthermore, a major health concern with regard to reproductive toxicity has been – and remains – diminishing fertility rather than birth defects. In 1992, Carlsen et al pub- lished evidence showing that, in Europe, adult male sperm counts had fallen by approxi- mately 50% in 50 years.13 Skakkebaek et al recently reported a severe worldwide decline in fertility over the previous five decades. According to the authors, it remains to be eluci- dated whether this decline is due to exposure to environmental contaminants, changes in lifestyles or both, but increasing incidences of testicular cancer are widely considered to be a result of environmental exposures to contaminants.14 Furthermore, Levine et al con- ducted a meta-analysis of the results of 223 such studies and reported that sperm counts have been declining on all continents and that these rates of decline have accelerated.15 A report commissioned by the World Health Organization (WHO) and the United Nations Environment Programme (UNEP) articulated serious concerns about the rising incidence of many endocrine-related diseases and disorders and the widespread detection of endocrine- disrupting compounds (EDCs), as well as “significant uncertainties about the true extent of 12 Herzler et al, supra, note 2. 13 E Carlsen et al, “Evidence for Decreasing Quality of Semen During Past 50 Years” (1992) 305 British Medical Journal 609. 14 NE Skakkebaek et al, “Environmental Factors in Declining Human Fertility” (2022) 18 Nature Reviews Endocrinology 139; see also JJ Meeks, “Environmental Toxicology of Testicular Cancer” (2012) 30 Urologic Oncology 212; L de Toni, “Testicular Cancer: Genes, Environment, Hormones” (2019) 10 Frontiers of Endocrinology 408. 15 H Levine et al, “Temporal Trends in Sperm Count: A Systematic Review and Meta-Regression Analysis of Samples Collected Globally in the 20th and 21st Centuries” (2022) Human Reproduction Update dmac035 <https://doi.org/10.1093/humupd/dmac035>. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 1 4 2 2 0 2 . r r e / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h European Journal of Risk Regulation 83 risks from chemicals” due to the lack of data.16 It is sensible to assume that those problems have not diminished during the nine years that it took the European Commission to set out its “scientific criteria for the determination of endocrine disrupting properties”.17 The EU’s process of classifying particular compounds as EDCs and banning them only started recently, and only a very few chemicals have so far been banned, despite evidence that approximately 800 chemicals are known or suspected EDCs. Two studies that employed extensive Monte Carlo simulations estimated that health costs in the EU due to EDCs exceed 150 billion euros annually.18 3. Toxicity of combination effects from chemical mixtures Herzler et al provided a discussion of possible combination effects (ie the possibility of additive or even synergistic adverse effects of multiple exposures to chemicals). Their nar- rative implied, however, that everything that might deserve to be done was already being done because – according to them – “intentional/foreseeable mixtures are already explic- itly addressed in some of the EU’s major regulatory programs (e.g. pesticides and biocides)”.19 “Addressed” is not a particularly precise term. While the European Food Safety Authority (EFSA) has issued a Guidance Document on this matter,20 its practical application has so far been limited to additive effects from oral exposures of suspected pesticidal com- pounds on two endpoints, namely acute effects on the nervous system21 and chronic effects on the thyroid.22 Moreover, combination effects under different routes of exposure (eg by inhalation and/or dermal in addition to oral) and chemicals other than pesticides/ biocides have not yet been included. Moreover, the toxicological database is incomplete and often equivocal, and its reliabil- ity is problematic. Our knowledge of the scope and limits of the validity of extrapolative inferences from animal studies to human risks is insufficient. Available concordance anal- yses of newly developed pharmaceutical drugs offer promising opportunities for enriching the scientific basis of policymaking. In contrast to pesticides and industrial chemicals, can- didate pharmaceuticals, if reassuring data from animal studies are available, are deliber- ately given to people, and both animals and patients are carefully monitored. According to Clarke and Steger-Hartmann, the concordances between animal and human studies vary between more than 70% and less than 30%, depending on the organ system.23 More impor- tantly, “absence of toxicity in animals has very low predictivity for the lack of adverse 16 A Bergman et al, “State of the Science of Endocrine Disrupting Chemicals 2012” (United Nations Environment Programme and the World Health Organization, 2013) <https://www.unep.org/resources/report/state-science- endocrine-disrupting-chemicals>. 17 Commission Regulation (EU) 2018/605 of 19 April 2018 amending Annex II to Regulation (EC) No 1107/2009 by setting out scientific criteria for the determination of endocrine disrupting properties <https://eur-lex. europa.eu/legal-content/EN/TXT/PDF/?uri=CELEX:32018R0605&from=EN>. 18 L Trasande et al, “Estimating Burden and Disease Costs of Exposure to Endocrine-Disrupting Chemicals in the European Union” (2015) 100 Journal of Clinical Endocrinology & Metabolism 1245; L Trasande et al, “Burden of Disease and Costs of Exposure to Endocrine Disrupting Chemicals in the European Union: An Updated Analysis” (2016) 4 Andrology 565. 19 Herzler et al, supra, note 2. 20 EFSA, “Guidance Document on Scientific Criteria for Grouping Chemicals into Assessment Groups for Human Risk Assessment of Combined Exposure to Multiple Chemicals” (2021) 19 EFSA Journal 7033. 21 EFSA, “Cumulative Dietary Risk Characterisation of Pesticides That Have Acute Effects on the Nervous System” (2020) 18 EFSA Journal 6087. 22 EFSA, “Cumulative Dietary Risk Characterisation of Pesticides That Have Chronic Effects on the Thyroid” (2020) 18 EFSA Journal 6088. 23 M Clarke and T Steger-Hartmann, “A Big Data Approach to the Concordance of the Toxicity of Pharmaceuticals in Animals and Humans” (2018) 96 Regulatory Toxicology and Pharmacology 94. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 1 4 2 2 0 2 . r r e / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h 84 Erik Millstone and Peter Clausing events in humans”.24 Such evidence of uncertainty from human drug studies represents a profound challenge to Herzler et al’s complacency, especially regarding mixtures. That challenge is amplified by the findings of a recent systematic review of 1,220 studies of mixtures, which emphasised that two-thirds of the studies were solely based on mixtures of just two compounds and that “toxicity outcomes of relevance for human risk assess- ment (e.g. carcinogenicity, genotoxicity, reproductive toxicity, immunotoxicity, neurotox- icity) were rarely addressed”.25 IV. Herzler et al’s perspective versus precaution While Herzler et al’s concern that “the strong push” of the CSS for a stricter regulation of chemical mixtures “is not sufficiently data driven”, our interpretation is that the BfR team’s position is antithetical to the EU’s precautionary legislation. The precautionary principle is supposed to be used in case of scientific uncertainty26; the legislation mandates the introduction and/or maintenance of regulatory measures in the absence of demon- strated risks. Herzler et al simply misunderstand precaution. The BfR team implies that the reality of risk needs to be proven before any measures can be justified and that evidence that falls short of demonstrable proof should never be sufficient, except perhaps for carcinogenic, mutagenic or reprotoxic substances (CMRs) or EDCs. Firstly, that is self-evidently a value judgment, not an empirical finding, but no less importantly it is a value judgment that implies consistently awarding the benefits of all and any uncertainties to the commercial interests of the companies that produce, trade in and use the chemicals in question rather than to the protection of, for example, occupa- tional, public and environmental health. Herzler at al accepted that “precaution” could be appropriate in relation to CMRs or EDCs, but only if those measures are “proportional”. But Herzler et al assume that meas- ures should always be proportional to the magnitudes of particular risks. However, if the magnitudes of the possible risks are uncertain, then the requirement for proportionality cannot be satisfied. The arguments outlined and the evidence adduced in this section therefore support the European Commission’s judgment that the prevailing chemical regulatory regime may not sufficiently protect public health, and consequently that it could – and should – be strengthened. V. The Chemical Strategy for Sustainability’s strategy and tactics To attain the strategic objectives on which the CSS focused, the Commission proposed sev- eral tactical reforms, although some of them were under-described. One important change of tactics is the proposal to broaden the scope of EU chemical risk assessments to include wider ranges of potential impacts and outcomes than those currently included. That tac- tical change could be – and should be – acknowledged as a proposed change to a key EU risk assessment policy. The plan to widen the scope of risk assessments was evident when, for example, the CSS recommended developing “methodologies for chemical risk assessment that take into account the whole life cycle of substances, materials and products” rather than merely assessing the safety of the particular use for which it will be sold. While the CSS 24 ibid. 25 O Martin, “Ten Years of Research on Synergisms and Antagonisms in Chemical Mixtures: A Systematic Review and Quantitative Reappraisal of Mixture Studies” (2021) 146 Environment International Article 106206. 26 European Parliament and Council, supra, note 8. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 1 4 2 2 0 2 . r r e / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h European Journal of Risk Regulation 85 optimistically implied that the prevailing regime was already doing a good job in relation to “chemicals that cause cancers, gene mutations, affect the reproductive or the endocrine system, or are persistent and bioaccumulative”, it envisaged extending that “to further chemicals, including those affecting the immune, neurological or respiratory systems and chemicals toxic to a specific organ”.27 The CSS aspired to reduce morbidity as well as premature mortality. The CSS anticipated the EU’s chemicals regulatory regime becom- ing more precautionary than the prevailing regime, although the word “precaution” appeared nowhere in the document. The second tactical change focused in particular on EDCs, which interfere with hor- monal systems. The CSS said that EDCs “require specific attention”. The CSS proposed28 to: (cid:129) : : : establish legally binding hazard identification of endocrine disruptors : : : building on criteria already developed for pesticides and biocides, and apply it across all legislation; (cid:129) ensure that endocrine disruptors are banned in consumer products as soon as they are identified, allowing their use only where it is proven to be essential for society; (cid:129) strengthen workers’ protection by introducing endocrine disruptors as a category of substances of very high concern under REACH [the Registration, Evaluation, Authorisation and Restriction of Chemicals]; (cid:129) ensure that sufficient and appropriate information is made available to authorities to allow the identification of endocrine disruptors by reviewing and strength- ening information requirements across legislation; (cid:129) accelerate the development and uptake of methods to generate information on endocrine disruptors through screening and testing of substances. (emphases in the original) That wording acknowledged that prevailing definitions of endocrine-disrupting hazard- ous chemicals had not been legally binding and had been applied unevenly, and that, even when such disruptors had been identified, they had often not promptly been banned. That is one reason why the CSS also called specifically for the “full implementation of the EU rules on chemicals” (emphasis in the original).29 It also implicitly accepted that occupational risks from endocrine disrupters have not been adequately regulated, and that far too little information was available to EU authorities on the composition and toxicology of putative endocrine disrupters. This was partly because too few sufficiently sensitive and specific endocrinological studies have been conducted, but also because the Commission suspected that industry may have withheld some or all of the data they have regarding some of the studies that have been conducted. This is why the CSS explicitly articulated a new policy of: “ZERO TOLERANCE FOR NON-COMPLIANCE” (Section 2.3.2, p 18, emphasis in the original). It will “strengthen the principles of ‘no data, no market’ and the ‘polluter-pays’” (ibid). A third tactical change responded to the scale of the scarcity and inadequacies of the data. The CSS stated that the Commission will “amend : : : information requirements to enable identification of all carcinogenic substances manufactured or imported in the EU” (Section 2.4.1, p 20, emphasis in the original), and that there is much knowledge to be acquired by authorities on the intrinsic properties of a vast majority of chemicals : : : . Equally, knowledge on uses and exposure is frag- mented : : : as it relies on industry to provide accurate information. The sheer 27 European Commission, supra, note 1. 28 ibid. 29 ibid. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 1 4 2 2 0 2 . r r e / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h 86 Erik Millstone and Peter Clausing number of chemicals on the market represents an immense knowledge challenge, and the expected future rise in chemical production and use risks further widening the “unknown territory of chemical risks”. (Section 2.4, p 19) The fourth proposed tactical change responded to the fact that EU regulatory bodies have often relied mainly on studies designed and conducted by or on behalf of the chemi- cal industry, and that too little attention had been paid to studies conducted indepen- dently of corporate interests, as the CSS rather coyly observed that “academic studies are not sufficiently exploited” (Section 2.3.1, p 16, emphasis in the original). The fifth proposed tactical change responded to the fact that, in relation to pharma- ceutical products, pesticides and food ingredients, EU rules have recently been amended to stipulate that no study can be used in support of an application for market authorisation unless the protocol for that study had been notified to the Commission before the study commenced.30 This is designed to ensure that companies no longer have the discretion to decide – for themselves and confidentially – whether or not to provide the EU’s regulatory institutions with all data from all studies. The sixth proposed tactical change focused on additional risks potentially arising from the combination effects of chemicals. It stated: People and other living organisms are daily exposed to a wide mix of chemicals origi- nating from various sources. Significant progress has been made in recent years to close some knowledge gaps on the impact of the combination effect of those chemicals. However, the safety of chemicals in the EU is usually assessed through the evaluation of single substances, or in some cases of mixtures intentionally added for particular uses, without considering the combined exposure to multiple chemicals from differ- ent sources and over time : : : . Explicit requirements to take into account the impact of unintentional mixtures is generally lacking : : : . To adequately address the combina- tion effect of chemical mixtures, legal requirements need to be consistently in place to ensure that risks from simultaneous exposure to multiple chemicals are effectively and systematically taken into account across chemicals-related policy areas : : : sci- entific consensus is emerging that the effect of chemical mixtures needs to be taken into account and integrated more generally into chemical risk assessments. (Section 2.2.2, p 12, emphases in the original) The seventh proposed tactical reform is to strengthen the regime covering chemical pollutants in the natural environment, so that environmental health as well as public health are properly protected. The CSS provided particular comments on one type of pol- lutant, namely the per- and poly-fluoroalkyl substances (PFASs)31, which it said require special attention, considering the large number of cases of contamination of soil and water – including drinking water – in the EU and globally, the number of people affected with a full spectrum of illnesses and the related societal and economic costs. That is why the Commission proposes a comprehensive set of actions to address the use of and contamination with PFAS. (Section 2.2.3, p 13, emphasis in the original) Overall, the CSS recognised many of the achievements of the EU’s chemicals regulatory its limitations and shortcomings, and consequently it regimes, but also many of 30 “Decision Laying Down the Practical Arrangements on Pre-Submission Phase and Public Consultations” (EFSA, 11 January 2021) <https://www.efsa.europa.eu/sites/default/files/corporate_publications/files/210111- PAs-pre-submission-phase-and-public-consultations.pdf>. 31 See, eg, <https://echa.europa.eu/hot-topics/perfluoroalkyl-chemicals-pfas>. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 1 4 2 2 0 2 . r r e / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h European Journal of Risk Regulation 87 highlighted at least seven key respects in which the Commission judged that those regimes could and should be strengthened, not just in the sense of being more restrictive, but also in the sense of being more comprehensive, better informed and more accountable by being more transparent. VI. The BfR team’s response to the CSS Herzler et al’s critical response to the CSS provided a complacent narrative, asserting that toxicological risks from chemicals had already been minimised and that the prevailing regulatory regimes had been “optimised”.32 It implicitly assumed that all toxicological risks have been exhaustively identified, documented, quantified and subjected to controls that are comprehensive, effective and sufficient. That judgment combined both factual and normative considerations, many of which we do not share and are contesting. The BfR team argued that, while there might be a role for the Commission to consult with a broad range of social, environmental and commercial stakeholders, decisions about the composition and implementation of a chemicals sustainability strategy should be taken only by reference to scientific considerations. We reiterate, however, that science alone can never settle policy questions. The BfR team implied that its own narrative was an entirely scientific one; we shall, however, provide evidence showing that it was replete with unacknowledged value judgments. The BfR team asserted that institutions, like the German Federal Institute for Risk Assessment (BfR), have a decade-long record of both practically applying and further developing the principles and methods of the science of regulatory risk assessment for consumer health pro- tection in the EU : : : it is one of their foremost tasks to safeguard that scientific rig- our is applied as the cornerstone of chemical risk assessment : : : 33 VII. A critique of Herzler et al and of the status quo The above is a surprisingly bold assertion from the BfR given its profoundly flawed approach to producing a risk assessment of the controversial herbicide glyphosate. One would at least have expected that all applicable regulations and guidelines had been prop- erly followed by the BfR when it prepared for the EFSA and the European Commission its draft Renewal Assessment Report (RAR) of glyphosate in 2015. However, several flaws remained in the BfR’s RAR, though some of these were corrected in an Addendum,34 which rapidly emerged at the end of August 2015 in response to the International Agency for Research on Cancer (IARC) monograph on glyphosate that had been published a few weeks earlier. Other flaws in the BfR’s RAR, however, remained unaddressed. One of the flaws that was corrected was a reassessment of statistical significances using a proper statistical procedure. In the Addendum, BfR acknowledged that “initially, the RMS [Rapporteur Member State] [had] relied on the statistical evaluation provided with the study reports”.35 That wording reveals that the BfR failed to conduct an independent sta- tistical assessment of the data provided by the industry’s Glyphosate Taskforce until after 32 Herzler et al, supra, note 2. 33 ibid, page 2590. 34 Rapporteur Member State (RMS) Germany, “Glyphosate. Addendum 1 to RAR: Assessment of IARC Monographs Volume 112 (2015): Glyphosate”. 35 ibid. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 1 4 2 2 0 2 . r r e / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h 88 Erik Millstone and Peter Clausing it came under political pressure, once IARC asserted that glyphosate is “probably carcino- genic to humans”.36 Only then did the BfR conduct its own statistical evaluation, which used a statistical method appropriate for evaluating tumour incidences, namely a trend test, rather than the pairwise comparisons that were reported in Glyphosate Taskforce documents. The BfR also initially ignored available evidence of glyphosate exposure causing oxida- tive stress, which is a potentially relevant mechanism that could explain the development of tumours. That evidence had not been acknowledged in its original assessment. Only after IARC had identified “strong evidence that glyphosate : : : can act to induce oxidative stress”37 did the BfR acknowledge this in the RAR Addendum. Both the BfR and subsequently EFSA discounted evidence of glyphosate’s carcinogenic- ity by claiming that the dose levels used in some of the mouse studies exceeded what they alleged was a maximum “limit dose” of 1,000 mg/kg.38 But no such limit has been officially specified for carcinogenicity studies. The Organisation for Economic Co-operation and Development (OECD) guidance documents, to which the RAR referred, had recommended a maximum dose of 1,000 mg/kg for chronic toxicity studies but not for carcinogenicity studies.39 The spurious tactic of citing the OECD was invoked by the BfR in its 2015 report on glyphosate, by EFSA in its conclusions on glyphosate, in the Harmonised Classification and Labelling (CLH) report on glyphosate and again in a recent draft report on glyphosate of the EU’s Assessment Group on Glyphosate,40 even after the limit-dose argument had repeatedly been criticised in peer-reviewed journals.41 Instead, OECD guidance document No. 116 recommended avoiding dietary concentra- tions of test substances higher than 50,000 mg/kg in long-term bioassays.42 This concen- tration is well above the highest concentrations of 20,000 and 30,000 mg/kg used in two of the mouse carcinogenicity studies on glyphosate. This was, however, never acknowledged by the BfR, EFSA or the European Chemicals Agency (ECHA). VIII. Further problems with the status quo Another problematic feature of the status quo in official EU toxicological assessments of the carcinogenic potential of pesticides has been the tactic of invoking historical control data (HCD) as grounds for discounting evidence of adverse effects, especially increased tumour incidences. Often when corporate interests want to discount putative evidence of toxicity that emerges from comparing evidence of adverse effects in groups of test ani- mals with evidence from the concurrent control group, they argue that the tumour inci- dences seen in the concurrent control group were unusually low when compared to 36 International Agency for Research on Cancer (IARC), “Some Organophosphate Insecticides and Herbicides” (2017) 112 IARC Monographs on the Evaluation of Carcinogenic Risks to Humans. 37 ibid. 38 RMS Germany, “Renewal Assessment Report Glyphosate” (2015) Volume 3 Annex B.6. Toxicology and Metabolism; EFSA, “Conclusion on the Peer Review of the Pesticide Risk Assessment of the Active Substance Glyphosate” (2015) 13 EFSA Journal 4302. 39 Organisation for Economic Co-operation and Development (OECD), “Guidance document 116 on the Conduct and Design of Chronic Toxicity and Carcinogenicity Studies, Supporting Test Guidelines 451, 452 and 453, 2nd Edition” (2012) 116 Series on Testing and Assessment; OECD, “Combined Chronic Toxicity/Carcinogenicity Studies” (2009) 453 Guideline for the Testing of Chemicals. 40 RMS Germany, supra, note 38; EFSA, supra, note 38; Assessment Group on Glyphosate, “Combined Draft Renewal Assessment Report and CLH Report. Glyphosate” (2021) Volume 1. 41 CJ Portier and P Clausing, “Re: Tarazona et al. (2017): Glyphosate Toxicity and Carcinogenicity: A Review of the Scientific Basis of the European Union Assessment and Its Differences with IARC” (2017) 91 Archives of Toxicology 3195; Clausing et al, supra, note 6. 42 Organisation for Economic Co-operation and Development (OECD), supra, note 37. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 1 4 2 2 0 2 . r r e / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h European Journal of Risk Regulation 89 control groups examined previously in earlier studies. They argue that if the test animals were compared to historical control groups the differences would cease to be statistically and/or toxicologically significant. OECD guidance document No. 11643 and Guidance of the European Chemicals Agency44 both provided very similar recommendations for the conditions under which HCD can legitimately be used to interpret new data. While HCD on tumour incidences in earlier control groups may sometimes be relevant when assessing the findings of recent studies, OECD emphasised “that the concurrent control group is always the most important con- sideration in the testing for increased tumour rates”.45 Similarly, ECHA stipulated that “historical data must be from the same animal strain/species, and ideally, be from the same laboratory to minimise any potential confounding due to variations in laboratory conditions, study conditions, animal suppliers, husbandry”, and that historical data should be contemporary to the study being evaluated (eg within a period of up to around five years of the study).46 The BfR’s glyphosate RAR represented a flagrant example of violating those rules when citing HCD to discount adverse findings. In practice, the BfR used a compilation of HCD stretching over a timespan of twenty-two years from seven different laboratories, com- paring the results from a stressful housing condition (wire-bottom cages) of the historical controls with those of the glyphosate studies housed under less stressful conditions (eg solid-floor cages), while also neglecting the duration of studies (eighteen versus twenty-four months), which had an important influence on spontaneous tumour incidences. Other official advisory bodies in EU Member States have contributed to the mainte- nance of the status quo. For example, when acting as rapporteurs, they have often invoked HCDs to discount positive toxicological findings in ways that violated the OECD and ECHA criteria, and dismissed evidence of increased tumour incidences when compared with con- current controls. Specifically, such transgressions can be found in the carcinogenicity assessments in four out of nine other draft RARs reviewed by Clausing,47 namely those for the active ingredients dimoxystrobin, folpet, phosmet and pirimicarb. Official institu- tions have frequently acted in ways that are inconsistent with the rules with which they claimed to comply. Another important type of the failure in practically applying “the principles and meth- ods of the science of regulatory risk assessment”48 has been the willingness of European authorities to accept and rely on studies with high mortality rates during the course of those studies. The premature deaths of the animals entail, firstly, that many died too young to have provided sufficient data, and secondly, that the numbers of animals avail- able for detailed end-of-study post-mortem examination are too small to provide sufficient data from which statistically significant effects might be derivable. Carcinogenicity studies are long-term studies; they typically cover approximately 75% of the average life expectancy of laboratory rodents. The validity of the findings of such studies crucially depends on sufficient animals surviving to the end of the minimum study duration of eighteen or twenty-four months for mice and rats, respectively. Therefore, OECD guidance document No. 116 recommended: 43 ibid. 44 European Chemicals Agency (ECHA), “Guidance on the Application of the CLP Criteria” (2017, Version 5.0) <https://echa.europa.eu/documents/10162/2324906/clp_en.pdf>. 45 Organisation for Economic Co-operation and Development (OECD), supra, note 37. 46 European Chemicals Agency (ECHA), supra, note 44. 47 P Clausing, “Chronically Underrated? A Review of the European Carcinogenic Hazard Assessment of 10 Pesticides” (Pesticide Action Network Germany and Health and Environment Alliance, 2019 Report) <https:// www.env-health.org/wp-content/uploads/2019/10/October-2019-Chronically-Underrated-web.pdf>. 48 Herzler et al, supra, note 2. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 1 4 2 2 0 2 . r r e / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h 90 Erik Millstone and Peter Clausing For a negative result to be acceptable in a rat carcinogenicity bioassay, survival in the study should ideally be no less than 50% in all groups at 24 months : : : In a mouse study, survival in all groups in the study should be no less than 50% at 18 months.49 Details from two RARs serve to illustrate violations of that requirement.50 In the RAR of phosmet, the survival rate in a rat study was well below 50%, but it was nevertheless deemed acceptable and reliable by the authorities as indicating a lack of carcinogenicity. Survival rates at the end of the 108 weeks were 20%, 24%, 32% and 38% in males and 32%, 35%, 32% and 42% in females for the control, low-, mid- and high-dose groups, respectively. Similarly, a rat study on pirimicarb with survival rates in males of between 35% and 48% was accepted. Ironically, for the mouse study, the following was noted: “In this study there is evidence of a carcinogenic potential of pirimicarb but with the limitations in the his- torical control data and reduced survival this data alone is not sufficient to conclude on carcinogenicity.”51 In other words, in the rat study, where the Rapporteur Member State (the UK) should have taken low survival rates into account before accepting a putative negative finding, that criterion was ignored. However, in the mouse study, in a situation where survival rates were not significantly low, a positive finding was nonetheless dismissed by alleging reduced survival, even though reduced survival is not a legitimate criterion for exclusion in cases of positive findings of significantly increased tumour incidences. Herzler et al criticised the CSS for arguing that there is “a necessity for further improv- ing the protection of ‘vulnerable groups’”.52 Their view was that “without a detailed assess- ment of which risks are currently deemed to be insufficiently addressed, it is hard to establish whether additional regulation might be necessary or existing regulation might need to be improved, and for which part of the population” (ibid). But that approximates to arguing that, until a quantitative risk assessment has been completed, it would be prema- ture to regulate on the basis of partial incomplete information, which implies that precau- tion should never be exercised; but that is incompatible with EU legislation. In relation to EDCs, Herzler et al implied that, before the CSS’s approach could deserve to be adopted, and before any further particular measures are implemented, the Commission should comprehensively identify all toxicologically problematic chemicals; that approach is, however, directly antithetical to precaution. Herzler et al’s perspective also irredeemably involves a value judgment – and not a scientific one – about how much of which kinds of evidence should be deemed necessary and sufficient before specific measures are adopted. Their value judgment, moreover, accords with industrial interests rather than prioritising the protection of public and environmental health. IX. The BfR’s misunderstandings and misrepresentations of the Chemical Strategy for Sustainability Herzler et al complained that the CSS was biased, but they were seemingly unaware of their own biases, many of which are evident from the facts that they chose to emphasise and the indicators that they chose to highlight. For example, Herzler et al criticised the CSS for failing adequately to differentiate “risks” from “hazards”. Historically, chemicals were typically restricted or banned only on the basis of a demonstrable risk given anticipated levels of exposure, not just on the basis of their intrinsic hazardous properties without reference to particular levels or frequencies of exposure. Herzler et al said correctly that 49 Organisation for Economic Co-operation and Development (OECD), supra, note 37. 50 Clausing, supra, note 47. 51 ibid, p 21; BfR, Renewal Assessment Report. Glyphosate, Volume 3 B.6, p 107. 52 Herzler et al, supra, note 2. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 1 4 2 2 0 2 . r r e / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h European Journal of Risk Regulation 91 it is trivial yet important to clearly distinguish between hazard, exposure and risk, as only the latter provides information on whether something is harmful (i.e. actually causes harm) or not. Amalgamating the terms “hazard” and “risk” leads to conceptual misunderstandings, with the consequence of fostering muddled conclusions and perceptions.53 However, the BfR team was itself blurring this distinction by failing to engage properly with the fact that some hazards – under the explicit provisions of EU legislation – have been officially categorised as unacceptable. For example, EU Regulation 1107/2009 states that a genotoxic, carcinogenic or reprotoxic pesticides “shall only be approved : : : if it is not or has not to be classified : : : as category 1A or 1B” (for mutagenicity, carcinogenicity or reproductive toxicity). This pivotal piece of legislation focuses on hazard categories, referring to their intrinsic toxicological properties. The BfR itself risked disseminating “conceptual misunderstandings” by repeatedly stat- ing that, in the case of glyphosate, IARC “only” provided a hazard assessment, while imply- ing that the BfR had been more thorough by providing a risk assessment.54 In practice, moreover, the BfR’s judgment focused in particular on an important hazard characteristic when concluding that glyphosate is not carcinogenic. If the BfR had come to the same conclusion as IARC, under EU legislation no risk assessment would have been needed. The Commission did not fail to distinguish risks from hazards; rather, it articulated an evaluative policy judgment that both sets of considerations can – and under certain con- ditions should – be adequate grounds for introducing or maintaining regulatory measures. The BfR team simplistically suggested that policy decisions can be made solely on scientific grounds and assumed that their chosen version of “the science” was incontestable. Claiming or even just assuming that only risks (which the BfR team interpret as quantified proven damage but not evidence of hazards) should be sufficient grounds for regulatory controls is unambiguously a normative policy judgment rather than an empirical finding. X. Conclusion To characterise the CSS’s proposed changes to the prevailing regime as “unscientific” is to commit what academic philosophers helpfully call a “category error”. The CSS is not intended to be and should not be misrepresented as a purely scientific judgment. It was a critical and historically informed policy judgment about the inadequacies and short- comings of the prevailing regime and the incomplete, equivocal and contested corpus of scientific evidence on which it purports to be based. It advocated changes in risk assess- ment policies and provided a constructive strategic response to those shortcomings. Alleging that the consequences of implementing the Commission’s proposals would be “overprotective” entails making value judgments about how much protection is sufficient and about who should decide how much protection is enough. Questions of that sort inev- itably invite the response: “enough for whom?”. The BfR team had, in effect, claimed the right and authority to decide what should be deemed “sufficient”. Herzler et al accused the CSS of being “inherently arbitrary”, at the same time articulating their own arbitrary 53 ibid. 54 Bundesinstitut für Risikobewertung (BfR), “Mehr Sachlichkeit in der Diskussion um die EU-Wirkstoffprüfung von Glyphosat gefordert” (Mitteilung Nr. 25/2015 des BfR vom 28 Mai 2015) <https://www.bfr.bund.de/de/ presseinformation/2015/25/mehr_sachlichkeit_in_der_diskussion_um_die_eu_wirkstoffpruefung_von_glyphosat_ gefordert-195267.html>; Bundesinstitut für Risikobewertung (BfR), “Populäre Missverständnisse, Meinungen und Fragen im Zusammenhang mit der Risikobewertung des BfR zu Glyphosat” (Mitteilung Nr. 013/2016 des BfR vom 19 Mai 2016) <https://www.bfr.bund.de/cm/343/populaere-missverstaendnisse-meinungen-und-fragen-im- zusammenhang-mit-der-risikobewertung-des-bfr-zu-glyphosat.pdf>. s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 1 4 2 2 0 2 . r r e / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h 92 Erik Millstone and Peter Clausing policy judgments while pretending that they were legitimated solely by reference to sci- entific considerations. There are therefore good empirical grounds for concluding that current chemical reg- ulatory regimes in the EU are not just suboptimal but seriously inadequate. The European Commission’s CSS is a prudent response to the available evidence, although its full imple- mentation will be challenging. Herzler et al’s critique of the CSS and their attempted defence of the status quo were flawed in numerous conceptual and empirical respects. They misrepresented much of the scientific evidence and ignored lots more, and they mis- leadingly portrayed their position as if it were purely scientific when it was in fact replete with problematic normative assumptions and judgments. Competing interests. The authors declare none. Cite this article: E Millstone and P Clausing (2023). “Reasons for Reinforcing the Regulation of Chemicals in Europe”. European Journal of Risk Regulation 14, 78–92. https://doi.org/10.1017/err.2022.41 s s e r P y t i s r e v i n U e g d i r b m a C y b e n i l n o d e h s i l . b u P 1 4 2 2 0 2 . r r e / 7 1 0 1 0 1 / g r o . . i o d / / : s p t t h
10.1057_s41287-023-00576-y
The European Journal of Development Research (2023) 35:323–350 https://doi.org/10.1057/s41287-023-00576-y SPECIAL ISSUE ARTICLE Revealing the Relational Mechanisms of Research for Development Through Social Network Analysis  · Guillaume Fournie2 · Barbara Haesler2 · Grace Lyn Higdon1 · Marina Apgar1 Leah Kenny3 · Annalena Oppel3 · Evelyn Pauls3 · Matthew Smith4 · Mieke Snijder1 · Daan Vink2 · Mazeda Hossain3 Accepted: 6 December 2022 / Published online: 25 January 2023 © The Author(s) 2023 Abstract Achieving impact through research for development programmes (R4D) requires engagement with diverse stakeholders across the research, development and policy divides. Understanding how such programmes support the emergence of outcomes, therefore, requires a focus on the relational aspects of engagement and collaboration. Increasingly, evaluation of large research collaborations is employing social network analysis (SNA), making use of its relational view of causation. In this paper, we use three applications of SNA within similar large R4D programmes, through our work within evaluation of three Interidsiplinary Hubs of the Global Challenges Research Fund, to explore its potential as an evaluation method. Our comparative analysis shows that SNA can uncover the structural dimensions of interactions within R4D programmes and enable learning about how networks evolve through time. We reflect on common challenges across the cases including navigating different forms of bias that result from incomplete network data, multiple interpretations across scales, and the challenges of making causal inference and related ethical dilemmas. We conclude with lessons on the methodological and operational dimensions of using SNA within monitoring, evaluation and learning (MEL) systems that aim to support both learning and accountability. Keywords Social network analysis · Collaboration · Relational · Evaluation · Learning · Research for Development * Marina Apgar m.apgar@ids.ac.uk 1 Institute of Development Studies, University of Sussex, Library Road, Falmer, Brighton BN1 9RE, East Sussex, UK 2 Royal Veterinary College, 4 Royal College St, London NW1 0TU, UK 3 London School of Economics, Houghton St, London WC2A 2AE, UK 4 Edinburgh Napier University, Sighthill Campus, Sighthill Court, Edinburgh EH11 4BN, UK Vol.:(0123456789) 324 M. Apgar et al. Résumé Pour que les programmes de recherche pour le développement (R4D ou Research for Developmement en anglais) aient un impact, il faut un engagement entre diverses parties prenantes dans les domaines de la recherche, du développement et des poli- tiques. Il est nécessaire de se concentrer sur les aspects relationnels de l’engagement et de la collaboration si l’on souhaite comprendre la façon dont ce type de programme permet l’émergence de résultats. L’évaluation des grands consortia de recherche utilise de plus en plus fréquemment l’analyse des réseaux sociaux (SNA ou social network analysis en anglais) en appliquant sa vision relationnelle de la causalité. Dans cet article, en vue d’explorer son potentiel en tant que méthode d’évaluation, nous utilisons trois applications d’analyse des réseaux sociaux au sein de grands pro- grammes R4D similaires dans le cadre de notre travail d’évaluation de trois pôles interdisciplinaires du Fonds de recherche sur les défis mondiaux. Notre analyse comparative montre que l’analyse des réseaux sociaux peut révéler les dimensions structurelles des interactions au sein de ces programmes et permettre d’apprendre comment les réseaux évoluent dans le temps. Nous menons une réflexion quant aux défis communs qui émanent de ces cas, y compris la gestion de différentes formes de biais qui résultent de données de réseau incomplètes, de multiples interprétations sur des échelles différentes et les défis liés au fait d’établir une inférence causale et les dilemmes éthiques connexes. Nous concluons par des leçons sur les dimensions méthodologiques et opérationnelles de l’utilisation de l’analyse des réseaux sociaux dans les systèmes de suivi, d’évaluation et d’apprentissage (SEA) qui visent à soute- nir à la fois l’apprentissage et la redevabilité. Introduction Research for development programmes (R4D) aim to put research at the service of solving intractable development challenges, and often have a focus on improving livelihoods of marginalised or excluded populations. Achieving development out- comes for these populations requires engagement with diverse stakeholders across the research, development and policy divides. Relationships between partners within R4D networks are central mechanisms for shaping activities as well as engagement and impact strategies through collaboration (Temple et al. 2018). Outcomes emerge from these interactions, leading to uncertainty in their pathways to impact (Jacobi et al. 2020; Maru et al. 2018; Thornton et al. 2017). Understanding if and how R4D programmes support the emergence of outcomes, therefore, requires a focus on the relational aspects of engagement and collaboration. In the case of the Global Challenges Research Fund of the UK the funder, UKRI, set the ambition for the portfolio of R4D programmes well beyond the delivery of world class research, and included partnerships between the UK and the Global South as a desired outcome, requiring strong networks to be built between diverse stakeholders (Barr et al. 2019).1 The scale of the GCRF portfolio (initially proposed 1 https:// www. ukri. org/ our- work/ colla borat ing- inter natio nally/ global- chall enges- resea rch- fund/. 325 at £1.5billion) embracing R4D programme as network building initiatives, empha- sising collaboration and learning across the research and development sectors, cre- ated an unprecedented opportunity to deepen understanding of the relational mecha- nisms that contribute to achieving outcomes and impact. In this paper, we explore methodologies from within GCRF programmes used in evaluating them as network building initiatives. We focus on the use of social network analysis (SNA) as one tool in an R4D methodological repertoire. Although SNA alone is not sufficient to fully understand the contribution of these complex programmes to development outcomes and impact, our comparative analysis shows that it can uncover the structural dimensions of interactions within large R4D pro- grammes and enable learning about how networks evolve through time. We reflect on common challenges across the cases including navigating different forms of bias that result from incomplete network data, multiple interpretations across scales, the challenges of making causal inference and related ethical dilemmas. We con- clude with lessons on the methodological and operational dimensions of using SNA within monitoring, evaluation and learning (MEL) systems with dual aims of sup- porting both learning and accountability. SNA Within Complexity‑Aware Evaluation The uncertainty of impact pathways in R4D programmes, and the need to centre the network of social actors and their interactions throughout implementation call for evaluation designs that focus on explaining how change is unfolding in real time, often referred to as complexity-aware (Bamberger et al. 2015; Douthwaite and Hoffecker 2017; Gates and Fils‐Aime 2021; Patton 2010). These designs respond to understanding development programmes, policies and interventions as operat- ing under conditions of complexity, requiring multiple strategies and engagement with diverse actors within systems. Programme outcomes emerge from interac- tions between the parts (relationships between actors) rather than from what indi- vidual parts achieve alone (Hargreaves 2021; Walton 2016). This is even more evi- dent when programmes are working in conflict-affected contexts which are highly dynamic. Such programming requires non-linear evaluation designs to capture emergent outcomes through the interactions, as well as understanding achievement of intended outcomes, and emphasize iterative learning as change happens (Apgar et  al. 2020). These new approaches to evaluation offer opportunities for focussing on the interactions between actors in an R4D network. Within these broad designs, there is a need to zoom into the structural dimensions of collaboration in order to then explore causal relationships between networking and intended outcomes, along impact pathways. SNA is a recognised interdisciplinary methodological field within social sci- ence research, building on its sociological and mathematical (graph theory) roots (Freeman 2000). One of the central offerings of SNA is its relational view of causation, as Marin and Wellman (2011, p. 13) describe it “social net- work analysts argue that causation is not located in the individual, but in the social structure”. Using SNA as a method allows for intuitive visualisations of Revealing the Relational Mechanisms of Research for Development… 326 M. Apgar et al. relationships as well as tangible measures of “network quality” (Davies 2009). Analytical approaches for SNA are diversifying (including quantitative, qualita- tive, and mixed strategies), and combining structural and relational approaches to causation is leading to greater exploration of its use for evaluation. As a recent scoping review of the use of SNA in evaluation shows, there is a steady increase in its application since the turn of the century (Popelier 2018) increas- ing its potential to support evaluation of complex systems. A number of appli- cations are relevant to the R4D programming context (e.g. Aboelela et al. n.d.; Drew et  al. 2011; Haines et  al. 2011; Honeycutt and Strong 2012) and high- light both opportunities and challenges. In this paper, we add to this nascent field through comparative analysis of three experiences of SNA in the context of large R4D programmes. Methodology We use a case study methodology (Yin 1989) to learn within and across three applications of SNA in similar large interdisciplinary collaborations funded as Interdisciplinary Hubs by UKRI under the GCRF—we will refer to these R4D programme as ‘Hubs’. They have sufficient similarity in scale and approach to evaluation to support cross-case analysis, while each application is necessarily bespoke to its programme context and needs. Table  1 summarises each of the cases, showing that evaluation and research questions that drove the use of SNA in each Hub differ slightly, and consequently, the design of the data collection tools and analytical strategies also differ (justification for each analytical strat- egy can be found in the Online Technical Appendix). The Hubs experienced two major disruptions in the early phases of imple- mentation that influenced both the network formation processes and relatedly the application of the SNA method; (i) the COVID-19 pandemic required all Hubs to adapt to online collaboration and many network forming activities were no longer possible, and (ii) an unexpected and significant reduction in fund- ing (due to a reduction in overall UK government funding for ODA) led to loss of staff and reduced scope of monitoring activities for a 12-month period. Our focus in this paper, therefore, is necessarily on the initial phases of work. All three cases include a baseline application of SNA with the shared goal of assess- ing the way in which collaborations were shaped through the early phases of implementation, and where possible, how this was influenced by the disruptions experienced. We, the co-authors, are the designers and implementers of the SNA within the Hubs, involved as researchers, MEL specialists, data analysts and pro- gramme managers. The within-case analysis was carried out by each programme team independently, following its own strategy, and focussed on what the SNA revealed about the particular evaluation and learning goals. We were not exter- nal researchers using SNA to understand the programmes but active users of the method as a mechanism for programmatic learning through our positions within s e g n a h c g n i w o h s s i k r o w t e n g n i - g r e m e e h t r e h t e h w e t a g i t s e v n i d n a n o i t a s i l a r t n e c f o s m r e t n i n o i t a u l a v e f o t r a P . y h c r a r e i h y e k w o h e t a g i t s e v n i o t s m i a t a h t g n i n r a e l , n o i t a r o b a l l o c y l e m a n , g n i d l i u b y t i l i b a p a c f o s r a l l i p n i g n i v l o v e e r a , g n i r a h s d n a d n a s r e b m e m b u H g n o m a b u H e h t ) n o i t a r t s i n i m d a r e t f a e r e h t y l l a u n n a , n o i t s n o i t s e u q g n i n r a e l d n a n o i t a u l a v E s e t u b i r t t a l a n o i t a l e R s e t u b i r t t a e d o N y c n e u q e r f / e m a r f e m T i s b u H h c r a e s e r y r a n i l p i c s i d r e t n i d e d n u f F R C G n i A N S f o e s u f o s e s a c e e r h T 1 e l b a T b u H d e p a h s g n i e b e r a s n o i t a r o b a l , h c a e r t u o , h c r a e s e r . g . e ( s n o i t e n i l p i c s i d - p e c n i e c n i s s h t n o m 2 1 , ) e v i t - l o c h c i h w n i y a w e h t s s e s s a o T - c a r e t n i f o s e i r o g e t a c c fi i c e p S , y r t n u o c , e g a t s r e e r a c , r e d n e G - c e p s o r t e r ( n o i t p e c n i b u h e r o f e B b u H y r t l u o P h t l a e H e n O t c e j o r p h c r a e s e r r e d a o r b a f o t r a P n i g i r o , s n o i t c e n n o c f o h t g n e r t S , e g a t s r e e r a c , y h p a r g o e G - o r t e r ( n o i t p e c n i b u H e r o f e B y t i r u c e S d n a e c i t s u J , r e d n e G 327 p i h s n o i t a l e r e h t d n a t s r e d n u o t r e g r a l a d n a b u H e h t n e e w t e b d n a e c a e P , n e m o W f o k r o w t e n f o t r a P . s r e n o i t i t c a r p y t i r u c e S : 1 t c a p m I b u H f o n o i t a u l a v e y c a c o v d a d n a e g d e l w o n k w e N “ s e c i o v e h t y f i l p m a s k r o w t e n d e s i l a n i g r a m d n a n e m o w f o e g n a h c e s y l a t a c o t s p u o r g ” s e t i s d n a s e u s s i s s o r c a ) y t i v i t c a b u H c fi i c e p s , b u H , b u H - n o n ( s n o i t c e n n o c f o f o e p y t , ) e m e h t ( m a e r t s e c n i s s h t n o m 6 1 , ) e v i t c e p s n o i t u t i t s n i ) 0 2 0 2 e n u J ( n o i t p e c n i b u H Revealing the Relational Mechanisms of Research for Development… 328 M. Apgar et al. w o H : n o i t s e u q n o i t a u l a v e e h t o t k r o w t e n a g n i d l i u b b u H e h t s i d e g a g n e d n a y r a n i l p i c s i d r e t n i f o r e t a l n I ? n o i t c u d e r k s i r r e t s a s i d - u b i r t n o c e h t g n i t a u l a v e s e g a t s n a b r u n o d e s s u c o f s r e h c r a e s e r n i a m ( e c n a n r e v o g k s i r n a b r u r o f s a e r a t c a p m i d n a e m o c t u o n i s t f i h s o t b u H e h t f o n o i t ) b u H e h t e g a t s r e e r a c , n o i t u t i t s n i f o t r a p s a d n a , s t n e v e b u u H s w e i v e r l a u n n a s n o i t s e u q g n i n r a e l d n a n o i t a u l a v E s e t u b i r t t a l a n o i t a l e R s e t u b i r t t a e d o N y c n e u q e r f / e m a r f e m T i b u H g n i d n o p s e r s e g a t s l a i t i n I . n o i t a u ) l a m r o f n i d n a l a m r o f h t o b g n i f o e p y t r e d n e g , ) n o i t a i l ffi a r o j a m r e t f a y l t n e u q e s b u s d n a - l a v e d e s a b - y r o e h t f o t n e n o p m o C - r u t p a c ( s n o i t c a r e t n i f o h t g n e r t S y t i c ( y h p a r g o e g , e n i l p i c s i D n o i t a t n e m e l p m i f o s r a e y 2 r e t f A b u H s e i t i C s ’ w o r r o m o T ) d e u n i t n o c ( 1 e l b a T 329 each Hub. The diversity of roles we played has enabled analysis across methodo- logical, operational and strategic layers of use of SNA as an evaluation method. Learning from Use of SNA In this section, we summarise the application of SNA in each case and present the findings from within-case analysis. Full technical details of the SNA applications in each Hub are presented in the Online Technical Appendix and illustrate that analyti- cal strategies were specific to each case. In all cases, we reflect on whether the SNA findings primarily displayed aspects of project design (controlled) or social collabo- ration that occurs within the project (uncontrolled) in the early phases of programme implementation. One Health Poultry Hub The One Health Poultry Hub (OHPH) addresses zoonotic disease risks associated with poultry intensification, with a geographic focus on Bangladesh, India, Sri Lanka and Viet Nam. To address this challenge and ensure the safe and sustain- able production of poultry, it aims to promote interdisciplinary and cross-sectoral dialogue within a One Health environment. Indeed, given the cross-cutting nature of these issues, strengthening interdisciplinary research capability and competencies, collaboration and knowledge exchange are core activities. Assessment and monitor- ing of such attributes over the OHPH’s lifetime form part of the Hub’s MEL frame- work. Given the programme design we expected that (i) connections between study countries would be mainly mediated by a small number of UK partners in the early network, and (ii) the network structure would then become less centralised in the later study periods, with more direct connections between study country partners. A key principle driving the evaluation and so the demand by the UK management team was to produce learning to feed adaptive programme management and encour- age decentralised network growth. In this context, we applied SNA methods to investigate the evolution of the OHPH network, a dynamic partnership network consisting of approximately 120 named researchers from 27 institutions in 10 countries. Specific objectives were (i) to assess the way in which collaborations were being shaped among its members during the course of the project; (ii) to characterise the extent to which the emerg- ing network is dynamically changing across countries and research areas; and (iii) to investigate characteristics in the development of the OHPH network associated with factors such as career stage, scientific discipline and gender. Methods The SNA was conducted using data from two bespoke online surveys (see full tech- nical details in the Online Technical Appendix). The first was carried out in March Revealing the Relational Mechanisms of Research for Development… 330 M. Apgar et al. Fig. 1 Network diagrams showing the OHPH cohort networks (network of those who responded to all three time periods). Nodes are coloured according to the country in which they were based 2020, one year after the Hub’s launch, and the second in February 2021. All co- investigators and researchers engaged with the Hub, contracted research staff, post- graduate students and managerial staff were invited to respond (120 named research- ers). Respondents were asked to consider their collaborations and activities with all other Hub members over three periods: P0 (before the Hub’s inception), P1 (dur- ing the first year of the Hub) and P2 (during the second year of the Hub). In addi- tion, respondents were asked to indicate their primary scientific discipline or area of expertise, their primary role in the Hub, gender, and age category. Findings While some respondents filled the survey for all three periods, others provided information for only one or two. For each period, we, thus, considered two sets of nodes: all respondents who responded within each period (period-specific net- works), and respondents who completed all three questionnaires (cohort networks) for which changes in connection patterns over time among the same set of nodes can be assessed. The comparison of cohort and period-specific networks allows us to assess whether analytical results are affected by the composition of our sampled net- works (i.e. selection bias). All networks were undirected: if at least one respondent reported a collaboration with another respondent, an edge was constructed between them. The size of the period-specific networks ranged from 58 to 81 nodes (35 to 45% of Hub partners). The cohort networks had 37 nodes (see Online Technical Appendix for details). About two thirds of respondents were from the study coun- tries, and almost all others were based in the UK. Most respondents were male, bio- logical scientists, and at mid to late career stage. Each period-specific and cohort network showed a high small-world index which is indicative of high clustering and short path lengths between nodes (Humphries and Gurney 2008) See Fig. 1.  From P0 to P1, the proportion of connected dyads increased as the OHPH’s project activities started. This increase in connectedness was distributed among partners, reducing the extent to which a small number of 331 actors acted as mediators between most others. This small-world structure, and the evolution towards a reduction in centralisation would be expected to promote the diffusion of information and knowledge, and their equitable access by Hub mem- bers. Several face-to-face meetings were organised during P0 and P1, including a whole-Hub conference. Such meetings enabled partners from all disciplines and partner countries to meet and collaborate directly. This was likely a major driver in increasing the network connectedness and reducing its centralisation when com- pared to the pre-inception period. However, the COVID -19 pandemic effectively eliminated all such opportunities from the OHPH’s second year (which commenced in March 2020). Similar to other GCRF interdisciplinary Hubs, all OHPH-wide events, regular project coordination meetings, meetings of working groups leading the design and implementation of research, impact and learning activities, ad-hoc workshops, conferences, early career researcher group meetings and other opportu- nities for interaction were migrated to online platforms. Possibly as a result of this, we found that the network’s connectedness decreased from P1 to P2 as well as it becoming more centralised—that is, more connections were mediated by a small number of highly-connected nodes. This pattern was observed in the period-specific as well as cohort networks. We assessed whether the centrality of a node was associated with its attributes (country, discipline, career stage, gender) using multivariable permutation-based lin- ear models (as in Delabouglise et al. 2017). We considered two centrality measures: degree (the number of other nodes with which a node was connected) and between- ness (the extent to which a node lay on the shortest path between two others). For the period-specific network at P2, there was weak evidence that the average degree could be lower for study countries than for UK partners, and for women than for men. Active participation in P2 online events varied depending on internet access, bandwidth and quality. Not all participants had access to the required IT hardware when working from home, and it was apparent that online formats made participa- tion more challenging for partners for whom English is a second language. These associations were not, however, observed on the cohort network. Betweenness was not associated with any abovementioned node-level attributes. We assessed the possible influence of individual respondent factors (country, dis- cipline, career stage, gender) on the occurrence of edges between any two nodes. We found that, for all periods and network types, the likelihood of a connection increased if two partners were from the same country, but decreased if they were both from different study countries. By the second year of the Hub’s operation (P2), all UK and 82% of study country partners were engaged in connections between the UK and study countries, whereas connections across study countries only involved 47% of study country partners. The connectedness was higher among social sci- entists, mid and late career stage and male partners in the period-specific network. These associations were not, however, found in the cohort-specific networks. The design of the OHPH’s research programme, which was initiated in the second year, was to an extent replicated across the countries in which it was implemented (to enable standardisation and comparability of outcomes), but required modifications for each of the study sites to enable specific research questions to be addressed, as well as incorporating local differences. This required central coordination by the Revealing the Relational Mechanisms of Research for Development… 332 M. Apgar et al. core research and management team (based primarily in the UK) as well as intensive collaboration within the site teams. It is likely that this further contributed to cen- tralisation and compartmentalisation of the network over the P2 period. Moreover, online meetings might have made it more difficult for participants who may per- ceive themselves as being lower in the hierarchy of their institution (e.g. early career researchers) to express opinions or share knowledge. Reflections on Contributions of SNA for MEL While SNA has been useful to visualise and assess changes in connectivity and centrality of the OHPH’s network, care should be taken not to overinterpret these results. A major limitation in this analysis is the likely selection bias. Only a small proportion of the total hub partners (between 35 and 45%) took part in the study. The composition of the respondent groups was likely to be non-representative in the two surveys, as well as varying between the two surveys. Hence, depending on the factors which affected participation, some of the results may only be valid for the group of respondents and not the entire collaborative network, as suggested by the discrepancies in some analytical results between the cohort and period-specific networks. Nevertheless, the implementation of SNA using a repeated annual survey was found to be helpful to characterise the changing nature of the OHPH network over its lifetime, and provide insights into the dynamic processes and factors (some of which are external) which influence this. It allowed the Hub to assess whether the network structure was evolving towards the emergence of desirable characteris- tics, such as a reduced UK-focus centralisation, increased interactions across study countries and disciplines and reduced influence of one’s career stage and gender on node centrality. Although the limitations mentioned previously imply that the scope for SNA alone to explore or quantify such nuanced or complex issues is limited, it should be considered as a tool that can be applied alongside output-based indicator measurement approaches. Gender Justice and Security Hub Gendered political, economic and social injustices shape the outbreak and dynamics of conflict; war itself involves violence against women and girls as well as violations of other human rights; and redress from the gendered harms of war is intimately connected to the establishment of a lasting and just peace. In short, gender is a fun- damental web of social relations through which justice and security are mediated. The Gender, Justice and Security Hub (GJSH) responds to this challenge through bringing together 121 members across 42 partner organisations from a variety of 333 disciplinary perspectives, skill sets, career stages and geographic locations—with Afghanistan, Colombia, Kurdistan-Iraq, Lebanon, Sierra Leone, Sri Lanka and Uganda as focus countries and 17 more in which project work takes place. It aims to achieve the creation and growth of a network of academics, activists, practitioners and policymakers to advance progress towards gender justice and inclusive security in conflict-affected societies. A central component of the Hub’s MEL plan was tracking whether and how network collaborations and connections develop and change over time.. A quan- titative social network analysis was used to understand how collaborations were being shaped during the course of the project so that they could produce learning to inform programme adaptation. The study was to be applied at every year of the Hub’s lifetime (2020–2024); however, due to the ODA funding cuts, we only report on the first round of the survey in this paper. The objectives of the SNA were (i) to map the Hub network in order to visualise its overall structure, including density and strength of connections between members; (ii) to document how network con- nections change in number and strength over time; (iii) to identify patterns of con- nections within the network by attributes (by stream, career stage, geography, etc.) and (iv) to facilitate introductions among Hub members to improve the number and depth of collaborative relationships within the Hub. The expectation was that the Hub would be a fairly centralised network in its early stages given that the setting up of the Hub-level structure was run by a central Management Impact Commu- nications Administration (MICA) team based in the UK and the Executive Group, consisting of 12 Co-Directors leading six research streams.2 The Hub activities were designed to strengthen and expand connections over time between Hub members who are not in the core group during the initial setting up period, leading to a less centralised network structure. Methods Data were collected between June and December 2020. Our sampling framework included all individuals associated with the GJSH at the time of the study (121 individuals). This included individuals directly or indirectly involved with the Hub research and advocacy-related activities (e.g. administrative staff, research partners, management, activists and Hub Champions). Nearly half (45%) of GJSH members completed the full survey. (See further methodological details in the Online Techni- cal Appendix). Findings The analysis shows that one year after the establishment of the Hub, it was a mildly centralised network whereby a small number of actors at the centre of the network 2 The six research streams are thematic areas of work (Law and Policy Frameworks; Livelihood Land and Rights; Masculinities and Sexualities; Methodological Innovation; Migration and Displacement; Transformation and Empowerment). Revealing the Relational Mechanisms of Research for Development… 334 M. Apgar et al. Fig. 2 Network diagram showing all GJS Hub country networks Fig. 3 Network diagram showing all GJS Hub stream networks have a large number of connections across the Hub, and a majority have a small number of connections across the network. We further looked at whether actors with similar levels of connectivity are more likely to connect with each other or if it var- ies depending on their centralisation (e.g. the core group has a lot of ties to less well connected actors but they do not connect well with each other). We found that the central group is very well connected across the entire Hub but the individuals outside the core group are not well connected to each other. This matches the expectation 335 of how the setting up of the Hub would develop with the UK-based MICA team (see red dots in Fig. 2 and blue dots in Fig. 3) at the core. This small group of peo- ple (high degree actors), which also includes most members of the Executive Group (half of which are based in the UK), have connections to almost every individual in the network. We can see this confirmed in Fig. 2, which segments the network by country affiliation, clearly showing a core (red) group of UK-based Hub members. These findings confirm our initial design. Regarding connections within the different thematic streams of the Hub, mem- bers who responded to the survey had more connections external to their stream than within (see Fig. 3). This is most likely the result of streams having been established by the core group of the Hub, rather than being pre-existing research networks who joined the Hub as a whole. At the time of the survey, streams had only met in person once, at the Hub Convention in January 2020. Although we have only conducted one round of the survey so far, respondents could indicate if they had a connection with another member prior to the inception of the Hub and if so, how their relationship had changed over the last year of their joint Hub membership and what was driving this change. We found that 72% of tie changes were driven by the Hub (31% through the Convention and 41% through other Hub-related interactions). Roughly 60% of Hub members attended the in-per- son Convention in Sri Lanka in January 2020 and we would expect this to be the major event that introduced a lot of people to each other. In fact, 66% of respondents first met through a Hub-related activity. Apart from the Convention, members of the Hub mostly interact through their project work on the Hub, which most likely accounts for most of the ‘other Hub-related’ tie changes, but also, in the case of members of the management team and executive group, through regular meetings. Many members of the executive group were instrumental in putting together the ini- tial application for the Hub and setting up its operational structures so we would expect their relationships to have strengthened during the first year of the project. Reflections on Contribution of SNA for MEL The interpretation of SNA results is based on several assumptions from network theory. The most important one being that it is a good thing if individuals in a net- work have many ties and are well connected to other members of the network. This assumes that everyone wants or should want to connect, network and collaborate. It does, in the first instance, not take into account whether there might be disciplinary, gendered, or geographical differences, which might, for example, encourage work in small, closely knit teams—maybe because extremely sensitive work is taking place in conflict contexts—rather than across the entire Hub. Some of the projects e.g. in Colombia, Iraqi Kurdistan, Sierra Leone, or Uganda rely on extensive networks across the Hub for comparative work and policy influence nationally and with inter- national organisations—we would expect researchers and partners on these projects to be as connected as possible across the network. Other projects, e.g. in Afghani- stan or Sri Lanka, due to the sensitivity of the work on gender in Afghanistan (even before the Taliban takeover in August 2021), or on human rights violations and post- conflict issues in Sri Lanka, might look to connect extensively with communities Revealing the Relational Mechanisms of Research for Development… 336 M. Apgar et al. they are embedded and limit their interaction with foreign researchers and institu- tions as it might increase their exposure. There are two stages of growing this network over the course of the Hub. First, the Hub itself as a network had to be established by facilitating connections and collaborations between different members of the Hub who are from different geographies, disciplines and professions. In the early years of the Hub, we would likely expect a fairly centralised network structure with the Principal Investiga- tor (PI), Executive Group and the Management, Impact, Communication and Administration (MICA) team at its core. Those are the actors who were involved in the application stage and set up the governance structure of the Hub. The PI, MICA team and half of the Executive Group are Global North based and one aim of developing the Hub as a network is to decentralise it and facilitate more South-to-South connections. Hub members come from a variety of disciplinary perspectives, skill sets, career stages and geographic locations. The Hub model was designed to encompass feminist principles, which also includes an empha- sis on collaborative working and ensuring that opportunities, including network- ing opportunities, are equitably distributed among all Hub members. Since the inception of the Hub and especially in preparation for and during the in-person Convention in January 2020, almost all communication happened over Microsoft Teams, a cloud-based team collaboration software. It was used as a collabora- tive platform for whole-Hub interaction, including to communicate information from the central team, but also for stream- and project-level interaction with some projects using it as their main platform for data storage and virtual meet- ings. It also allowed for the quick creation of new communication channels when members expressed an interest, e.g. in arts-based approaches. Since Teams was already established as a platform for communication and all Hub members had access to it prior to the COVID-19 pandemic, the move to online-only interac- tion was fairly smooth. During the pandemic, Conventions were moved online and included cross-project collaborations and presentations. In the later stages of the Hub and after the SNA was completed, we have moved to co-creating outputs with Hub members from different projects, streams and countries such as books, papers, documentaries, and trainings. The SNA results allowed for the mapping of emerging relationships across the network and were designed to trace the development of relationships between and within groups including Global North and Global South partners, Early Career Researchers (ECRs) and more senior members, and practitioners and academics. Importantly, the SNA study aimed to provide tangible evidence on how networks like the GJS Hub can generate new insights through relationship development. The SNA would have been an important addition to surveys and anecdotal evidence in deciding where to target efforts of partnership building (as per objective iv). How- ever, in the absence of this longitudinal evidence due to the interrupted funding mid-way, the ability to do this in an adaptive and tailored way was significantly hampered. The second phase is to grow the Hub’s connections externally and link them to existing and emerging international networks and communities of practice on Women, Peace and Security, peacebuilding, International Law, and development. 337 It would have been possible to identify which Hub members would benefit from a closer connection with external networks—that could then be fostered—and which networks to tap into at a Hub level to create the most impact, especially on a policy level. The connections built with these other networks are ultimately what will facilitate local and global policy change and institutional reform to advance gender justice and sustainable peace, building on new knowledge and advocacy networks, which amplify marginalised voices across different conflict contexts. Tomorrow’s Cities Hub In a rapidly urbanising world, 60% of the area expected to be urban by 2030 remains to be built, opening a huge opportunity to build risk out of tomorrow’s cities. An initial assessment of the challenge showed that the disaster risk reduc- tion community (including scientists, policymakers, development practitioners and business leaders) is currently operating in disconnected communities of prac- tice and despite existing policy frameworks (e.g. Sendai Framework) approaches to disaster risk reduction are focussed on crisis management and not integrated into urban planning. The Tomorrow’s Cities Hub responds to this opportunity by working through research and development partnerships in four cities (Kath- mandu, Nairobi, Istanbul and Quito). Its aim is to co-produce interdisciplinary research on multi-hazard risk working with stakeholders in order to influence dis- aster risk reduction policy and practice. The co-produced research is implemented through a network consisting of 174 individuals from 54 partners including academic institutions, research centres, government departments and NGOs focussed in the four core cities. Given the aim of the Hub, strengthening capacity for interdisciplinary working as well as facilitating collaboration between researchers and stakeholders are critical to suc- cess. Monitoring how the hub network evolves through time, therefore, is a core component of the hubs MEL strategy. The drive for use of SNA as a tool for monitoring evolution of the hub network came from the evaluation and learning team, as part of a theory-based evaluation design. The intended users of the find- ings were the managers at central Hub level as well as in the city teams who were responsible for building an enabling environment for collaboration. SNA was used as a baseline to understand the extent to which cross-collabora- tion was occurring across different attributes of individuals including their gen- der, career level and location in the initial phases of implementation. The expec- tation based on the Hub design was for an initial centralised network given the central UK-based leadership team had built the proposal through their own rela- tionships with partners based in cities. Given the Hub’s commitment to equitable partnerships and to learn across the city contexts, we expected the network to evolve towards a less centralised structure through time. Intentionality in equita- ble partnerships and the potential of power asymmetries that funding structures across partners of the global North and South could reproduce the management Revealing the Relational Mechanisms of Research for Development… 338 M. Apgar et al. Fig. 4 Network diagram showing all TC Hub geographic location networks and research structures were built to enable work across UK and cities. Collabo- ration was encouraged through formation of cross-hub thematic research groups, as well as building and supporting a network of ECRs to co-produce research outputs. Regular all hub meetings were convened online and plans for the first in- person all hub conference in 2021 was curtailed by the pandemic. Methods Data were collected between February and March 2021, through two databases. An administrative database with generic information about collaborators in the hub and an online survey through the SumApp platform generated a specialised sur- vey to capture and visualise connections with other collaborators in the network. Our sampling framework included all individuals associated with the TC project at the time of the study. The total sample was 174 individuals and 53% made at least one connection (47% only show incoming ties and so we assume did not complete the survey) (for full details see Online Technical appendix). Self-identified female workers are slightly over-represented in the respondents (41% in respondents versus 23% non-respondents) which could account for more ties for women across the net- work, while ECRs are also slightly over-represented (47% respondents versus 37% non-respondents) which could also explain higher observed connections for ECRs overall. 339 Findings The analysis shows that after the establishment of the TC project, the network is fairly centralized. UK-based collaborators tend to be at the centre of the network and have a large number of connections across different locations—they are high degree actors. Yet, UK-based collaborators are also connected amongst themselves (purple nodes in Fig. 4). For example, connections realised (expressed as density) is twice as high among UK-based collaborators compared to those that work in LMIC coun- tries. This shows that overall collaborations are initiated by the UK as the central hub whereby the four city networks tend to be connected through UK-based and cross-city affiliated individuals (shown in Fig. 4). We then analysed collaborations between key attributes of location, career stage, gender and disciplines. We found that on average, individuals have 13 connections with those based at different locations. Yet only 9% (or an average of 1.2 connec- tions) of those occur between LMIC-based collaborators (South–South collabora- tion). Collaborations are higher between UK-based members than between LMIC located members (an average of 9.5 versus 4.2 connections, respectively). It is more likely for those located in the same location to collaborate if they have a common contact. These findings were expected given the project design. The UK represents the biggest segment with 80 collaborators who are linked to one or more cities and organized around disciplinary/thematically focussed groups that co-designed research within and across disciplines. A greater extent of shared contacts within members in similar locations may further reflect historical collaborations and con- textual knowledge held prior to establishing the new TC network. Collaboration also occurs among and across career stages. Peer-to-peer col- laboration is more frequent overall among those at mid and senior career stages (an average of 17) that among those at early career stage (an average of 12). This could be explained in part by the fact that most of the 77 ECRs were independently recruited new hires in the TC project and so did not have existing connections to each other. Yet when looking at location, this pattern is different. For those based in the UK, peer-to-peer connection is higher among mid and senior career mem- bers (16 connections on average) than among ECRs (6 connections on average). And cross-career level collaborations are 8 connections on average. In LMIC locations, however, peer-to-peer collaboration among ECRs is more common (7 connections on average), while peer-to-peer collaboration among mid and senior researchers is lower (average of 4). Cross-level collaborations are 7 connections on average. What this shows is that peer-to-peer collaboration of ECRs is driven largely by their loca- tion as is illustrated in Fig. 5. Lastly, regarding collaboration across genders, there are more connections among male collaborators overall (17 versus 12 among women). Yet, there is also col- laboration between men and women (an average of 13 connections) overall. Again we found the pattern differed depending on location. Men located in the UK are more likely to collaborate with other men as compared to men located in LMIC. For instance, collaboration between men located in the UK is higher (on average by 6 connections) than between women located in the UK. In LMIC, this is overall Revealing the Relational Mechanisms of Research for Development… 340 M. Apgar et al. Fig. 5 Network diagram showing collaboration among Early Career Researchers based outside of the UK for TC Hub more balanced. This suggests that gender influences how much members collaborate within their peer group. Reflections on Contribution of SNA for MEL This application of SNA was part of a staged and modular theory-based evalua- tion design that aimed to explore a core assumption of the hub’s theory of change around interdisciplinary working and equitable partnerships. The baseline appli- cation described the network at the outset in order: (i) to identify opportunities to enhance interdisciplinary and equitable partnerships as part of adaptive programme management; and, (ii) to inform impact evaluation design to assess the contribution of network collaborations to achieving intended shifts in urban planning (evaluating relationships beyond the hub). Visualization of the network overall and specific visual patterns in collaboration did enhance understanding of how the initial project design is reflected in the social fabric of collaboration. The fairly centralised initial network built confidence in the initial design with a large UK-based central team, and an explicit intent to build city- focussed research partnerships. While we did not have an ‘ideal network’ structure against which we intended to monitor the evolution of the network, we did anticipate that through time we would see greater collaboration between the non-UK-based members, in line with the intention to build equity in the partnership. 341 The analysis also enabled observation of unanticipated dynamics—such as gen- dered dynamics of collaboration between UK and non-UK-based individuals, and the marked difference between the way ECRs and more senior individuals are con- nected across the network. These structural patterns revealed  would need to be deepened through focus groups with hub collaborators to explore the drivers behind them and identify potential programme adaptations in line with the goal of moving towards a network structure with greater connections across more members. Visual- ising these patterns could also support existing conversations within the Hub around undertaking gender bias and power training as signalled by at least one of the city leadership teams as a priority. In this way, the application of SNA has shown poten- tial to produce learning that could influence the next phases of work to support col- laboration in the Hub. The application of these findings for decision-making processes, however, depends in large part on the quality of data and resulting findings. Data limitations are a common challenge in SNA (e.g. Wasserman and Faust 1994; Newman 2003) due to its exponential growth of observations and complexity. Further, SumApp required respondents to scroll through a list of all 174 people in the Hub which could result in biases, e.g. individuals mentioned towards the end could be less fre- quently selected due to response fatigue. This and other sources of response biases (which are common in SNA) influence the extent to which these findings constituted actionable learning for the Hub. Findings from Cross‑Case Analysis As R4D programmes, all three Hubs set out to intentionally build collaboration across disciplines, geographies and hierarchies. All cases share the dual (and inter- connected) objectives of (i) using SNA to monitor progress of their intended designs through describing and tracking how collaborative relationships change over time across significant attributes, and (ii) using the visualisations and resulting appre- ciation of the structure of the network to influence its development in intentional ways—to support ‘network weaving’ (Vance-Borland and Holley 2011). Our experi- ences are from the early phases of implementation. While we do not discuss result- ing adaptations, we do reflect on the opportunities for responding to learning that emerged when using SNA as a learning and an evaluation tool. Data Challenges and Respondent Bias The practical challenges of data collection and analysis in SNA are well described in the literature and relate, among other factors, to the extensive time required to respond to lengthy surveys leading to incomplete data sets with consequent implica- tions for rigorous understandings of whole networks (e.g. (Newman 2003; Penuel et  al. 2006; Popelier 2018). Particularly relevant to applications within evaluation are threats to construct validity that result from ambiguity in how relational attrib- utes are collected (how the type and strength of collaboration is described) and Revealing the Relational Mechanisms of Research for Development… 342 M. Apgar et al. relatedly, how these are interpreted (Popelier 2018). Incomplete datasets can lead to weak ties being under reported, influencing the overall validity of findings. Consequently, perhaps the most important step in the design of a SNA study in the context of evaluation is defining the ties, or connections, at the outset. This requires clarity of the aspects of collaboration that are of interest to the study, as well as knowledge of how these will be interpreted by respondents. All our cases are large networks (with over 100 members), and given the novelty of using SNA to explore relational aspects of research projects, choosing where to focus had to align with key evaluation and learning demands. As shown in Table  1, the relational attributes used in each case were driven by the specific and different evaluation objectives of each—OHPH mapped Hub-related collab- orations (including research, outreach, administrative activities), GJSH mapped both the strength (light, good, strong, none) and origin of connections (non-Hub, Hub, specific Hub activity), while TCH mapped both formal and informal inter- actions within the hub through strengths (in four categories). OHPH members were asked with whom they had worked on a range of activi- ties over pre-defined periods of time. However, the interpretation of the level of engagement which qualifies as “working together”, as well as the definition of a specific task is likely to vary among respondents. In the TCH case, four catego- ries or strengths of collaboration were used, but interpretations of each cannot be assumed to have been uniform. Given the size of the partnerships and the period of time covered by each survey, especially in the case of OHPH, recall bias cannot be excluded. Attempts were made to minimise them by including the list of all members in the questionnaire (and TCH and GJS included photos), so respondents were less likely to forget collaborators. In spite of these efforts, unsurprisingly, missing data was a challenge in all three cases. This was dealt with in different ways. OHPH avoided under estimat- ing connections by examining only partial networks, TCH used the reconstruc- tion method (Liu et al. 2019; Stork and Richards 1992) and assumed incoming ties are reciprocal for non-respondents in order to examine the full network, and GJSH examined both the full network (including non-respondents) and the par- tial network using the listwise approach (Pepinsky 2018). Complex model-based approaches, such as Baysean models, are gaining popularity, but are often dif- ficult to implement (requiring a complex model to be specified and estimated), and can result in introducing other forms of bias by imputing edges that over generalise the tendencies observed in other parts (i.e. information rich areas) of the network (Smith et al. 2022). In the context of SNA for learning, these more complex modelling strategies were not considered worth the additional time and effort. As we discuss later, there are inherent limits to what SNA can reveal on its own, and as part of broader MEL strategies, triangulation with other methods we posit is a better approach to mitigating the challenge of missing data. An obvious yet not insignificant response to overcoming the challenge of incomplete network data is to invest early in strategies to increase the proportion of Hub members participating in the survey. The TCH chose to use the SumApp survey in order to turn the SNA process into an explicit network weaving exer- cise, with the assumption that this would motivate hub members to respond. 343 SumApp creates a personal profile for each hub member with a unique URL and users can visualise the network real time as respondents update their con- nections (while the application is ‘live’). Feedback from ECR members sug- gests that this was indeed motivating for many of them because it aligned with their motivation to network within a large hub. Yet this did not necessarily hold true for other members of the Hub. Understanding what might motivate greater response, therefore, is an important step in planning SNA as a learning and net- work weaving tool. Challenges of Interpretation The challenge of confirmation bias in interpreting SNA findings in the context of programme evaluation is well described in the literature (e.g. (Popelier 2018)). Critics argue that limited ability to objectively interpret the results may lead to an interpretation which aligns with the investigators’ (and/or programme managers’) preconceived ideas rather than taking the data at face value, or indeed, pretend- ing that an ‘objective’ interpretation exists. This follows a gold standard view in evaluation that confirmation bias is to be avoided at all costs. In contrast, employ- ing complexity-aware evaluation designs, we were working within programmes implementing SNA as a participatory and learning-oriented evaluation method, working with (rather than controlling for) the experiences and aspirations of those involved in programme design (Apgar and Allen 2021). Interpretation of the find- ings, therefore, required the situated experiences of programme implementers and we embraced their interpretations (which are inherently biased) as an important explanatory device. As (Durland and Fredericks 2005) note the importance of specific network information should be seen as relative to programme needs at that particular time, embracing internal interpretation as the principle goal. In all three cases, the baseline application of SNA served as a useful empirical check on how the initial programme design was reflected in the social fabric of collaboration. The network structures that became visible at the initial phase were interpreted based on our expectations given the intentional designs. Table 2 provides a comparative view across the three cases. In all three cases, the network structures revealed in the early stages of the Hubs matched the expectation of centralised structures, based on their set up driven by the parameters of the funding set by UKRI. In all cases UK-based research leaders built the Hub networks initially through contracting partners and researchers based in countries of focus or operation. As others have shown, it is the repeated applications of SNA that enables a picture of evolution and change through time and brings it to life as a monitor- ing tool (Aboelela et al. n.d.; Provan et al. 2005). Yet as social networks are liv- ing and constantly evolving systems, we expect that they will organically shift in time and some of their dynamics will be unpredictable. This leads us to ask— how should we interpret the changes as part of monitoring the network? Some argue that lack of an ideal network structure means there cannot be standard Revealing the Relational Mechanisms of Research for Development… 344 M. Apgar et al. n a h t r e h t a r n i h t i w e v i t c a e r o m e r e w s n o i t a r o b a l l o c t a h t t n a e m y r t - n u o c y d u t s h c a e n i h t i w s e i t i v i t c a t c e j o r p f o n o i t a t n e m e l p m i e h t s e i r t n u o c y d u t s n e e w t e b . y r t n u o c e m a s e h t n i , t l u c ffi i d e r o m s u h t e r e w m u i t r o s n o c e h t s s o r c a s n o i t c a r e t n i e l i h W s r o t a r o b a l l o c n e e w t e b y l e k i l e r o m e b o t d n u o f s a w n o i t a r o b a l l o C s n o i t c a r e t n i . d e t c e p x e t o n s a w s d o i r e p e m i t d r i h t d n a d n o c e s e c a f - o t - e c a f d e t a t i l i c a f n i e g a g n e o t s r e b m e m k r o w t e n f o y t i l i b a n i e h t n e e w t e b s s e n d e t c e n n o c n i e s a e r c e d a e l i h w , d e t c e p x e s a w e h t o t d e l c i m e d n a p e h t h g u o r h t n o i t a r o b a l l o c o t n o i t p u r s i D s d o i r e p e m i t o w t t s r fi e h t s s o r c a s s e n d e t c e n n o c n i e s a e r c n i n A b u H y r t l u o P h t l a e H e n O s n o i t a r o b a l l o c g n i t s i x e - e r p d n u o r a d e m r o f g n i e b n a h t r e h t a r e m o s h t i w s r o t c a l l a o t d e t c e n n o c l l e w e r e w s r e b m e m k r o w t e n s r e b m e m d e s a b - K U y b s n o i t a r o b a l l o c w e n s a p u t e s e r e w s m a e r t S d e s i l a r t n e c , d e t c e p x e s a , n o i t a t n e m e l p m i f o r a e y e n o r e t f A b u H y t i r u c e S d n a e c i t s u J , r e d n e G s n o i t a n a l p x E e r u t c u r t s k r o w t e n d e t c e p x e s u s r e v e r u t c u r t s k r o w t e n d e v r e s b O s g n i d n fi A N S f o n o s i r a p m o C 2 e l b a T b u H h c a e f o e d i s t u o s r e b m e m h t i w d e h s i l b a t s e e r e w s n o i t a r o b a l l o c w e n , d e n o i s i v n e s A . s r e b m e m b u H l l a o t s n o i t c e n n o c g n i v a h . s e i t i v i t c a b u H o t e u d ) m a e r t s r o ( p u o r g c i t a m e h t f o e d i s t u o d n a n i h t i w s r e b m e m b u H h t i w s n o i t a r o b a l l o c r e t a e r G . d e t c e p x e n e e b e v a h d l u o w ) m a e r t s r o ( p u o r g c i t a m e h t h c a e - a r o b a l l o c f o s c i m a n y d d e r e d n e g e h t , b u H e h t n i s s e n e r a w a r e w o p c i t a m e h t r o y r a n i l p i c s i d n a h t r e h t a r d e t a c o l e r a y e h t e r e h w n e k a t r e d n e g o t s e h c a o r p p a t n e r e ff i d e h t y b d e n i a l p x e e b n a c n o i t - l o c h c u m w o h d e c n e u fl n i r e d n e g , r e h t r u F . s g n i p u o r g h c r a e s e r d n a r e d n e g n o n o i t c e fl e r h g u o r h t t l i u b g n i d n a t s r e d n u n o d e s a B y b y l e g r a l n e v i r d e b o t d n u o f s a w s R C E n e e w t e b n o i t a r o b a l l o C s n o i t a c o l d n a s e m e h t h c r a e s e r s s o r c a s r e d a e l y b t n e r e ff i d n i s m a e t d n a s p u o r g r e e p n i h t i w s n e p p a h n o i t a r o b a l g n i n n a l p t c a p m i . n g i s e d . s n o i t a c o l d n a n o i t a t n e m e l p m i d e t a r g e t n i r o f n o i t a r o b a l l o c n o i t a c o l n i h t i w l a i t i n i e h t g n i m r fi n o c , e r t n e c e h t t a s r o t a r o b a l l o c d e s a b - K U d e r i u q e r y l i r a s s e c e n b u H e h t f o n g i s e d h c r a e s e r d e s s u c o f - y t i c e h T h t i w , d e s i l a r t n e c s a w b u H e h t n o i t a t n e m e l p m i f o r a e y e n o r e t f A b u H s e i t i C s ’ w o r r o m o T 345 benchmarks for judging performance through use of SNA and this weakens its evaluation potential (e.g. (Haines et al. 2011)). Our intention was not to monitor progress against an ideal structure, but rather, to iteratively learn with and as the structure evolves. But assumptions from net- work theory are, often implicitly, applied. For example, assuming that a network is ‘stronger’ when individuals have many ties and are well connected to other members of the network. In large networks, such as R4D collaborations, however, it may in fact be that specialisation, or particular geographic clustering within the network is optimal for achievement of desired goals. Interpretation of the struc- ture, then, must follow the intention of the network, which in turn is always situ- ated in a particular moment and context. Our cases illustrate such a situated con- textual approach to using SNA. In all cases, we held assumptions about a desirable evolution of the hub struc- ture in line with the stated goals of equity across specific hierarchies of power (such as gender, Global North-Global South, career level) (see Table  2 for observed versus expected network structures). These assumptions are aligned with the theory of change of the GCRF funding mechanism overall and the pro- grammes specifically. The OHPH interpreted their findings based on an assumed evolution along a spectrum—a hypothetical star network (with the PI being connected to everyone and no link between others) at one end, and a complete network (with everyone connected to each other) at the other end. Indeed the SNA across time periods revealed an unexpected surge in centralisation in the last study period which suggested movement in the opposite direction. These observed trends in network development could be related to the move from face- to-face to online interactions due to COVID-19, adding some confidence that they are reflective of ‘actual’ processes. This finding can help to identify individuals between which collaborations should be fostered, and support thinking further about how the hub could recover post-pandemic to continue to build collabora- tion in the ways it intended. The GJSH captured connections prior to the inception of the Hub (retrospec- tively) and then 16 months into the project (time of the survey). This showed how the Hub was set up and how members started to connect. The importance of in- person all-Hub Conventions was underlined as it was both a key event where peo- ple first met each other as well as a key driver of strengthening connections. Over the course of the pandemic, similar to the OHPH, Conventions moved online and while this continued the opportunity to meet officially, it greatly reduced the chances for spontaneous interactions which are most likely to strengthen relationships. What our experiences suggest is that whether implicit or explicit, the programme designs were manifest through the baseline and subsequent application in a way that made them visible. This offered the opportunity to challenge earlier assumptions and to identify hidden or unexpected dimensions that warrant further exploration, as well as mechanisms to nudge the network in desirable directions. In the case of the TCH baseline, we saw that even in the early stages some unexpected dynam- ics were revealed—such as collaborations between female members based outside of the UK being greater than for female members based in the UK. The TCH had Revealing the Relational Mechanisms of Research for Development… 346 M. Apgar et al. procedures for reflecting on the ways in which power imbalances influenced internal team dynamics and how equitable decision making was. This was part of a broader intention to work in equitable ways (Snijder et al. this issue). Revealing the gendered dynamics of collaboration could add further weight to the requests of some city management teams to provide gender sensitivity training to all network members. It also allowed questioning an underpinning (positivist) premise of ‘more is better’. Starting with a deliberate project design as we did, network studies as an M&E tool offer the opportunity to re-evaluate and -classify measures from an aspect of appli- cability for the ‘ideal function’ rather than ‘ideal structure’ of a network. Strengthening Causal Inference The ways in which SNA can support causal inference are still debated within the evaluation literature, and many applications of SNA still struggle to determine causal relationships between internal network structure and external network out- comes. In the context of evaluating R4D programmes, being able to make this link is important to add weight to SNA as an evaluation method. The structural paradigm of SNA suggests that structural mechanisms influence how changes unfold, yet the relational aspects are so many, indeed potentially infinite, that establishing clear causal links is challenging (Doreian 2001; VanderWeele and An 2013). A response to this dilemma is to situate the structural analysis within a contextualized theory about how the causal inference is hypothesized (a causal theory of change). As discussed above, all three of our SNA cases were part of broader evalua- tion designs. While all three are focussed on visualising and describing the ‘inter- nal’ networks, in TCH and GJSH, the intention (prior to funding cuts) was to use understanding of how the internal network is working (and evolving through time), alongside other evaluation research on equitable partnerships to build con- tribution claims around how the network structure (and ways collaborations are taking shape across hierarchies and power structures) contributes to intended pro- gramme outcomes. In the case of TCH, the focus of external interactions is on influencing the co-production of risk informed urban planning, while in the case of GJSH, the focus is on the end goal of shifting patriarchal modes of knowledge production on sustainable peace. In this way, using SNA as a monitoring tool can not only build a picture of how the internal structure is evolving but also can offer data points to support theory-based evaluation. Additional methods are required to enhance interpretation and reveal the mean- ing given to relations identified and so to crystalise key causal  mechanisms for evaluation  to investigate further. Often quantitative SNA is complemented with qualitative approaches (Hopkins 2017; Kolleck 2016). All three Hubs have mul- tiple other sources of monitoring data that could be used to supplement the SNA. TCH, for example, developed outcome case studies for monitoring outcomes, and implemented a survey on interdisciplinary working to shed light on some of the patterns revealed. For example, an outcome case study on interdisciplinary work- ing in Quito evidences how intentional reflection on ways of working and building 347 of interdisciplinary capacities within the team has opened up opportunities for greater engagement with local partners for multi-hazard risk research. This quali- tative analysis can support causal claims around how collaboration between team members from different disciplines (visualised through SNA) which produce internal outcomes (such as openness to participatory methods) relates to move- ment along a desired pathway towards equitable outcomes. While SNA is not a causal methodology, it can provide evidence of collaboration and thus can con- firm or disconfirm achievement of early and internal desired outcomes in the way the programme (network) is set up to deliver impact. Ethical Dilemmas in SNA for Evaluation Making visible how individuals are interacting with colleagues or partners comes with ethical challenges and risks if the results are interpreted as judgements of indi- vidual performance (Kadushin 2005; Penuel et  al. 2006). The GCRF Hubs were funded to intentionally support collaboration across established hierarchies of power, turning collaborative behaviours into expectations. Anonymization, which is the standard approach to managing research ethics and minimizing risk to partici- pants, is often not possible and usually not desirable when using SNA to intention- ally support network weaving. The GJS Hub experienced a concern that members might alter how they reported a connection based on a perception of how socially acceptable that connection might be. Of course any reported connection is a subjective measure of that individual’s perception of a relationship but there is still likely to be systematic under- or over- reporting. For example, it is likely that an ECR might underreport connections with more senior scholars so as not to assume a ‘strong’ relationship when that percep- tion might not be reciprocal. Similarly, there might be cultural differences in percep- tions of relationship strength, where, for example, a US-based member of the Hub might consistently rate relationships as stronger than UK-based Hub members. As discussed above, the TCH used the SumApp online tool to support intentional network weaving which was experienced as motivating for ECRs. On the other hand, accessibility of all connections to everyone in the Hub might have deterred some from reporting for similar reasons described for GJSH. Visualising your own posi- tion in a network that aspires to be collaborative could, therefore, be experienced as positive or negative depending on how connected you are and how much you value connection. One advantage is that individual learning, and first person reflection, is made possible through seeing oneself as part of the network. But this requires net- work members to value this capability for self-reflections above any concerns they have about being judged based on their position in a network. Revealing the Relational Mechanisms of Research for Development… 348 M. Apgar et al. Lessons for Future Use of SNA in Evaluation of R4D Programmes Across the three cases of SNA in evaluation of large R4D programmes, we have illustrated that in spite of the challenges with data and interpretation (which are common to SNA), it is useful as a monitoring tool when used to reflect on under- lying assumptions about collaboration and resulting network structures. From our analysis we conclude with three lessons for future use of SNA within evaluation of R4D programmes. 1. The more explicit assumptions about collaboration are at the outset, the more useful the empirical view of collaboration revealed is to programme learning. A contextualized theory of collaboration could be created at the outset to guide the SNA study. This is in line with Davies’ (2009) call for a theory-based and deduc- tive approach to SNA in evaluation. 2. Combining SNA with other methods can enhance interpretation and reveal the meaning given to structural views. This can strengthen causal inference about relational causal mechanisms making SNA a necessary, but not sufficient method to evaluate R4D programmes. 3. Navigating the challenges of interpretation and ethical dilemmas requires careful consideration as well as an enabling institutional and political environment for use of SNA to support learning. Embedding the interpretation of SNA findings within participatory learning moments (such as after-action reviews) would strengthen the use of SNA findings in learning-oriented evaluation design, as suggested by others (Drew et al. 2011; Durland and Fredericks 2005) Supplementary Information The online version contains supplementary material available at https:// doi. org/ 10. 1057/ s41287- 023- 00576-y. Funding Funding was provided by Natural Environment Research Council. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Com- mons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/. References Aboelela, S.W., J.A.M. Rn, K.M. Carley, and E.L. Rn. 2007. Social Network Analysis to Evaluate an Interdisciplinary Research Center. Journal of Research Administration 38: 15. Apgar, M., and W. Allen. 2021. Participatory Monitoring, Evaluation and Learning: Taking Stock and Breaking New Ground. In SAGE Handbook of Participatory Research and Inquiry, ed. D. Burns, J. Howard, and S. Ospina. London: Sage Publications. 349 Apgar, Marina, K. Hernandez, and G. Ton. 2020. Contribution Analysis for Adaptive Management. 14. Bamberger, M., J. Vaessen, and E. Raimondo. 2015. Dealing with Complexity in Development Evalua- tion: A Practical Approach. London: Sage Publications. Barr, J., P. Simmonds, B. Bryan, and I. Vogel. 2019. Inception report—Global Challenge Research Fund (GCRF) Evaluation—Foundation Stage.https:// doi. org/ 10. 13140/ RG.2. 2. 31983. 89762 Davies, R. 2009. The Use of Social Network Analysis Tools in the Evaluation of Social Change Com- munications C.pdf. https:// mande. co. uk/ wp- conte nt/ uploa ds/ 2018/ 08/ The% 20Use% 20of% 20Soc ial% 20Net work% 20Ana lysis% 20Too ls% 20in% 20the% 20Eva luati on% 20of% 20Soc ial% 20Cha nge% 20Com munic ations% 20C. pdf Delabouglise, A., N. Antoine‐Moussiaux, D. Tatong, A. Chumkaeo, A. Binot, G. Fournié, E. Pilot, W. Phimpraphi, S. Kasemsuwan, M.C. Paul, and R. Duboz. 2017. Cultural practices shaping zoonotic diseases surveillance: the case of highly pathogenic avian influenza and Thailand native chicken farmers. Transboundary and emerging diseases 64 (4): 1294–1305. Doreian, P. 2001. Causality in Social Network Analysis. Sociological Methods & Research 30 (1): 81–114. https:// doi. org/ 10. 1177/ 00491 24101 03000 1005. Douthwaite, B., and E. Hoffecker. 2017. Towards a Complexity-Aware Theory of Change for Partici- patory Research Programs Working Within Agricultural Innovation Systems. Agricultural Systems 155: 88–102. https:// doi. org/ 10. 1016/j. agsy. 2017. 04. 002. Drew, R., P. Aggleton, P. Boyce, H. Chalmers, C. Maxwell, S. Pachauri, F. Thomas, I. Warwick, and K. Wood. 2011. Social Network Analysis to Evaluate Organisational Networks on Sexual Health and Rights. Development in Practice 21 (8): 1062–1079. https:// doi. org/ 10. 1080/ 09614 524. 2011. 590884. Durland, M.M., and K.A. Fredericks. 2005. An Introduction to Social Network Analysis. New Directions for Evaluation 2005 (107): 5–13. https:// doi. org/ 10. 1002/ ev. 157. Freeman, L.C. 2000. Social Network Analysis: Definition and History. New York: Oxford University Press. Gates, E.F., and F. Fils-Aime. 2021. System Change Evaluation: Insights from The Rippel Foundation and its ReThink Health Initiative. New Directions for Evaluation 2021 (170): 125–138. https:// doi. org/ 10. 1002/ ev. 20462. Haines, V.A., J. Godley, and P. Hawe. 2011. Understanding Interdisciplinary Collaborations as Social Networks. American Journal of Community Psychology 47 (1–2): 1–11. https:// doi. org/ 10. 1007/ s10464- 010- 9374-1. Hargreaves, M. 2021. Bricolage: A Pluralistic Approach to Evaluating Human Ecosystem Initiatives. New Directions for Evaluation 2021 (170): 113–124. https:// doi. org/ 10. 1002/ ev. 20460. Honeycutt, T.C., and D.A. Strong. 2012. Using Social Network Analysis to Predict Early Collaboration Within Health Advocacy Coalitions. American Journal of Evaluation 33 (2): 221–239. https:// doi. org/ 10. 1177/ 10982 14011 424201. Hopkins, M. 2017. A Review of Social Network Analysis and Education: Theory, Methods, and Applica- tions. Journal of Educational and Behavioral Statistics 42 (5): 639–646. https:// doi. org/ 10. 3102/ 10769 98617 698111. Humphries, M.D., and K. Gurney. 2008. Network ‘Small-World-Ness’: A Quantitative Method for Deter- mining Canonical Network Equivalence. PLoS ONE 3 (4): e0002051. https:// doi. org/ 10. 1371/ journ al. pone. 00020 51. Jacobi, J., A. Llanque, S. Bieri, E. Birachi, R. Cochard, N.D. Chauvin, C. Diebold, R. Eschen, E. Frossard, T. Guillaume, S. Jaquet, F. Kämpfen, M. Kenis, D.I. Kiba, H. Komarudin, J. Madrazo, G. Manoli, S.M. Mukhovi, V.T.H. Nguyen, and C. Robledo-Abad. 2020. Utilization of Research Knowledge in Sustainable Development Pathways: Insights from a Transdisciplinary Research-for- Development Programme. Environmental Science & Policy 103: 21–29. https:// doi. org/ 10. 1016/j. envsci. 2019. 10. 003. Kadushin, C. 2005. Who Benefits from Network Analysis: Ethics of Social Network Research. Social Networks 2 (27): 139–153. https:// doi. org/ 10. 1016/j. socnet. 2005. 01. 005. Kolleck, N. 2016. Uncovering Influence Through Social Network Analysis: The Role of Schools in Edu- cation for Sustainable Development. Journal of Education Policy 31 (3): 308–329. https:// doi. org/ 10. 1080/ 02680 939. 2015. 11193 15. Liu, X., L. Li, S. Wang, Z.-J. Zha, D. Meng, and Q. Huang. 2019. Adaptive Reconstruction Network for Weakly Supervised Referring Expression Grounding. IEEE/CVF International Conference on Com- puter Vision (ICCV) 2019: 2611–2620. https:// doi. org/ 10. 1109/ ICCV. 2019. 00270. Marin, A., and B. Wellman. 2011. SNA: An Introduction. Jhon Scott and Peter J. Carrington. Revealing the Relational Mechanisms of Research for Development… 350 M. Apgar et al. Maru, Y.T., A. Sparrow, J.R.A. Butler, O. Banerjee, R. Ison, A. Hall, and P. Carberry. 2018. Towards Appropriate Mainstreaming of “Theory of Change” Approaches into Agricultural Research for Development: Challenges and Opportunities. Agricultural Systems 165: 344–353. https:// doi. org/ 10. 1016/j. agsy. 2018. 04. 010. Newman, M.E.J. 2003. The Structure and Function of Complex Networks. SIAM Review 45 (2): 167– 256. https:// doi. org/ 10. 1137/ S0036 14450 342480. Patton, M.Q. 2010. Developmental Evaluation: Applying Complexity Concepts to Enhance Innovation and Use. New York: Guilford press. Penuel, W.R., W. Sussex, C. Korbak, and C. Hoadley. 2006. Investigating the Potential of Using Social Network Analysis in Educational Evaluation. American Journal of Evaluation 27 (4): 437–451. https:// doi. org/ 10. 1177/ 10982 14006 294307. Pepinsky, T.B. 2018. A Note on Listwise Deletion versus Multiple Imputation. Political Analysis 26 (4): 480–488. https:// doi. org/ 10. 1017/ pan. 2018. 18. Popelier, L. 2018. A Scoping Review on the Current and Potential Use of Social Network Analysis for Evaluation Purposes. Evaluation 24 (3): 325–352. https:// doi. org/ 10. 1177/ 13563 89018 782219. Provan, K.G., M.A. Veazie, L.K. Staten, and N.I. Teufel-Shone. 2005. The Use of Network Analysis to Strengthen Community Partnerships. Public Administration Review 65 (5): 603–613. https:// doi. org/ 10. 1111/j. 1540- 6210. 2005. 00487.x. Smith, J.A., J.H. Morgan, and J. Moody. 2022. Network Sampling Coverage III: Imputation of Missing Network Data Under Different Network and Missing Data Conditions. Social Networks 68: 148– 178. https:// doi. org/ 10. 1016/j. socnet. 2021. 05. 002. Stork, D., and W.D. Richards. 1992. Nonrespondents in Communication Network Studies: Problems and Possibilities. Group & Organization Management 17 (2): 193–209. https:// doi. org/ 10. 1177/ 10596 01192 172006. Temple, L., D. Barret, G. Blundo Canto, M.-H. Dabat, A. Devaux-Spatarakis, G. Faure, E. Hainzelin, S. Mathé, A. Toillier, and B. Triomphe. 2018. Assessing Impacts of Agricultural Research for Devel- opment: A Systemic Model Focusing on Outcomes. Research Evaluation 27 (2): 157–170. https:// doi. org/ 10. 1093/ resev al/ rvy005. Thornton, P., T. Schuetz, W. Förch, L. Cramer, D. Abreu, S. Vermeulen, and B. Campbell. 2017. Responding to Global Change: A Theory of Change Approach to Making Agricultural Research for Development Outcome-Based. Agricultural Systems 152: 145–153. https:// doi. org/ 10. 1016/j. agsy. 2017. 01. 005. Vance-Borland, K., and J. Holley. 2011. Conservation Stakeholder Network Mapping, Analysis, and Weaving: Conservation Stakeholder Networks. Conservation Letters 4 (4): 278–288. https:// doi. org/ 10. 1111/j. 1755- 263X. 2011. 00176.x. VanderWeele, T.J., and W. An. 2013. Social Networks and Causal Inference. In Handbook of Causal Analysis for Social Research, ed. S.L. Morgan, 353–374. Amsterdam: Springer. Walton, M. 2016. Expert Views on Applying Complexity Theory in Evaluation: Opportunities and Barri- ers. Evaluation 22 (4): 410–423. https:// doi. org/ 10. 1177/ 13563 89016 667890. Wasserman, S., and K. Faust. 1994. Social Network Analysis: Methods and Applications. Cambridge, U.K.: Cambridge University Press. Yin, R. K. 1989. Case Study Research: Design and Methods. https:// books. google. co. uk/ books? hl= en& lr= & id= FzawI AdilH kC& oi= fnd& pg= PR1& dq= Case+ study+ resea rch:+ Design+ and+ metho ds& ots=l_ 1X8ci W3x& sig= oQbz6 MCByW Q0zM0 kBHxq amutE n4#v= onepa ge&q= Case% 20stu dy% 20res earch% 3A% 20Des ign% 20and% 20met hods&f= false Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
10.7554_elife.86556
RESEARCH ARTICLE Annexin A6 mediates calcium- dependent exosome secretion during plasma membrane repair Justin Krish Williams1†, Jordan Matthew Ngo1†, Isabelle Madeline Lehman1, Randy Schekman2* 1Department of Molecular and Cell Biology, University of California, Berkeley, Berkeley, United States; 2Department of Molecular and Cell Biology, Howard Hughes Medical Institute, University of California, Berkeley, Berkeley, United States Abstract Exosomes are an extracellular vesicle (EV) subtype that is secreted upon the fusion of multivesicular bodies (MVBs) with the plasma membrane. Exosomes may participate in intercel- lular communication and have utility as disease biomarkers; however, little is known regarding the physiological stimuli that induce their secretion. Ca2+ influx promotes exosome secretion, raising the possibility that exosomes are secreted during the Ca2+- dependent plasma membrane repair of tissues damaged by mechanical stress in vivo. To determine whether exosomes are secreted upon plasma membrane damage, we developed sensitive assays to measure exosome secretion in intact and permeabilized cells. Our results suggest that exosome secretion is coupled to Ca2+- dependent plasma membrane repair. We find that annexin A6 (ANXA6), a well- known plasma membrane repair protein, is recruited to MVBs in the presence of Ca2+ and required for Ca2+- dependent exosome secretion, both in intact and in permeabilized cells. ANXA6 depletion stalls MVBs at the cell periphery, and ANXA6 truncations localize to different membranes, suggesting that ANXA6 may serve to tether MVBs to the plasma membrane. We find that cells secrete exosomes and other EVs upon plasma membrane damage and propose that repair- induced secretion may contribute to the pool of EVs present within biological fluids. Editor's evaluation This compelling study brings together two earlier observations: that Ca2+ influx can trigger exosome release from multivesicular bodies, and that plasma membrane repair after wounding requires Ca2+ and involves Ca2+- binding annexin proteins. This important work takes these earlier findings in an interesting new direction by showing that exosome release from MVBs is also triggered by Ca2+ influx during plasma membrane wounding and requires the annexin isoform ANX6. The study raises the interesting possibility that cell injury and repair may contribute to the release of exosomes into biological fluids. Introduction Protein secretion is an essential process that prokaryotes and eukaryotes utilize for intercellular communication. In eukaryotic cells, most secretory proteins are exported by the conventional secre- tory pathway which consists of co- translational signal peptide recognition by the signal recognition particle, polypeptide translocation through the Sec61 translocon, and anterograde transport from the endoplasmic reticulum (ER) to the Golgi apparatus via COPII vesicles (Park and Rapoport, 2012; Shan and Walter, 2005; Zanetti et al., 2011). However, eukaryotic cells have also developed alternative *For correspondence: schekman@berkeley.edu †These authors contributed equally to this work Competing interest: See page 20 Funding: See page 20 Preprinted: 29 December 2022 Received: 31 January 2023 Accepted: 18 May 2023 Published: 19 May 2023 Reviewing Editor: Suzanne R Pfeffer, Stanford University, United States Copyright Williams, Ngo et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 1 of 24 Research article secretory processes that bypass the conventional pathway (Malhotra, 2013). Some well- documented examples include the secretion of leaderless cargoes (proteins that lack a signal peptide), such as interleukin- 1β, and the egress of cytoplasmic proteins and RNAs within extracellular vesicles (EVs; Colombo et al., 2014; O’Brien et al., 2020; Zhang et al., 2020). EVs are membrane- enclosed compartments that are exported to the extracellular milieu of cells in culture and in vivo. Eukaryotic cells secrete EV subpopulations that can be classified broadly into two major categories with distinct biogenesis pathways: microvesicles and exosomes (Colombo et  al., 2014; van Niel et  al., 2018). Microvesicles form by outward budding from the plasma membrane (Cocucci et al., 2009; Tricarico et al., 2017). Exosomes are 30–150 nm vesicles formed by inward budding of the limiting membrane of late endosomes to produce multivesicular bodies (MVBs) that contain intraluminal vesicles (ILVs). Upon fusion of MVBs with the plasma membrane, ILVs are released to the extracellular space as exosomes (Harding et  al., 1983). Exosomes have potential utility as biomarkers for disease progression because they contain protein and small RNA molecules specific to their cell type of origin (Driedonks and Nolte-’t Hoen, 2018; Shurtleff et al., 2017; Upton et al., 2021). Little is known about the physiological circumstances that could elevate exosome secretion. Inter- estingly, elevation of cytosolic Ca2+ with ionophores has been demonstrated to induce rapid exosome secretion in a Rab27a, Munc13- 4- dependent manner (Messenger et al., 2018; Savina et al., 2003). These results raise the possibility that exosomes may be secreted as a byproduct of the Ca2+- dependent plasma membrane repair process that occurs after membrane disruption in vivo. Perforation of the plasma membrane results in an influx of extracellular Ca2+ into the cytoplasm and triggers a repair cascade that includes mobilization and fusion of lysosomes with the plasma membrane (Andrews and Corrotte, 2018). Given the similarity of lysosomes and MVBs, it is plausible to suggest that exosomes are secreted as a byproduct of Ca2+- dependent plasma membrane repair. However, some reports have suggested that, unlike lysosomes, MVBs do not undergo Ca2+- dependent exocytosis (Jaiswal et al., 2004; Jaiswal et al., 2002). To determine if exosomes are secreted during Ca2+- dependent plasma membrane repair, we developed an improved nanoluciferase (Nluc) reporter system to quantify exosome secretion that is sensitive, linear, and amenable to high throughput. Using this assay, we established that a Ca2+ ionophore, a pore- forming cytolysin, and physiological mechanical stress all stimulated exosome secretion in an extracellular Ca2+- dependent manner. We showed that annexin A6 (ANXA6), a well- characterized plasma membrane repair protein, is recruited to MVBs in the presence of Ca2+ and demonstrated that ANXA6 is required for Ca2+- dependent exosome secretion. We then observed that depletion of ANXA6 stalls MVBs at the periphery of cells treated with a Ca2+ ionophore. Next, we developed a streptolysin O (SLO)- permeabilized cell reaction that recapitulates Ca2+- and ANXA6- dependent exosome secretion. Finally, we demonstrated that the two annexin domains of ANXA6 become enriched at different membranes upon elevation of cytosolic Ca2+. Our results demonstrate that exosome secretion is coupled to Ca2+- dependent plasma membrane repair and that ANXA6 may serve as a potential tether for the recruitment of MVBs to the plasma membrane. Results Design and validation of an endogenous CD63-Nluc exosome secretion assay We sought to develop a cell- based exosome secretion assay that is quantitative, sensitive, amenable to high- throughput, and able to distinguish cell debris and bona fide exosomes. A previous study developed a luminescence- based assay to quantify exosome secretion by inserting Nluc into the endogenous locus of the tetraspanin protein CD63 (Hikita et al., 2018). We obtained the HCT116 CD63Nluc- KI #17 cell line (referred to herein as CD63- Nluc cells) from this group and built upon their assay. Sequestration of CD63 during ILV biogenesis results in the amino- and carboxyl- termini oriented to the exosome lumen. We leveraged this topology along with the relative membrane permeabili- ties of the substrate and an inhibitor of Nluc to develop a quantitative cellular assay to monitor the secretion of exosomes (Figure  1A). Furimazine, the substrate of Nluc, permeates through cellular membranes and enables luminescence from both cellular debris and intact CD63- Nluc exosomes Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 2 of 24 Cell Biology Research article Figure 1. Endogenous CD63- nanoluciferase (Nluc) is a faithful reporter of exosome secretion. (A) Schematic illustrating the topology of CD63- Nluc in different membranes and the cellular CD63- Nluc secretion assay. (B) Luminescence from the conditioned medium of CD63- Nluc cells with or without furimazine, the membrane- impermeable Nluc inhibitor, and 0.1% TX- 100 are shown. (C) Luminescence derived from the supernatant fraction of CD63- Nluc conditioned medium subjected to differential centrifugation (1k, 10k, and 100k) with or without 0.1% TX- 100 are shown. (D) Membrane- protected luminescence (in blue circles) and buoyant density (in black squares) of CD63- Nluc conditioned medium subjected to a high- resolution linear density gradient are shown. (E) Membrane- protected luminescence from CD63- Nluc conditioned medium collected over 24 hr with (blue circles) or without Figure 1 continued on next page Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 3 of 24 Cell Biology Research article Figure 1 continued 0.1% TX- 100 (black squares). Data plotted represent the means of three independent experiments, and error bars represent SDs. Note, for Figure 1D and E, the error bars are smaller than the dots in the image. present within the conditioned medium of cultured CD63- Nluc cells. The addition of a membrane- impermeable Nluc inhibitor eliminates luminescence derived from cellular debris where Nluc is readily accessible, while not affecting the luminescence derived from intact CD63- Nluc exosomes incubated with furimazine (Walker et al., 2017). Finally, solubilization of cellular membranes by the addition of the non- ionic detergent Triton X- 100 (TX- 100) exposes all CD63- Nluc to the inhibitor. The addition of the membrane- impermeable Nluc inhibitor depleted a significant portion of luminescence derived from the conditioned medium of CD63- Nluc cells (~30%) and membrane solubilization by TX- 100 reduced luminescence to background levels (Figure 1B). To further validate the localization of the CD63- Nluc luminescence signal to sedimentable vesicles, we subjected the conditioned medium fraction to differential centrifugation and found that CD63- Nluc luminescence was removed after high- speed sedimentation or upon the addition of TX- 100 (Figure 1C). Next, we obtained a partially clarified conditioned medium fraction by differential centrif- ugation and fractionated the material on a linear iodixanol gradient (Figure 1D). CD63- Nluc vesicles equilibrated at a density of 1.11  g/ml, a buoyancy similar to published reports (Jeppesen et  al., 2019). In a time course over 24 hr, we measured secreted CD63- Nluc luminescence in the presence of Nluc inhibitor alone or with TX- 100 and found the signal increased linearly, consistent with exosome secretion over time (Figure 1E). The luminescence signal in the presence of the Nluc inhibitor and TX- 100 remained constant, consistent with cell rupture/detachment during medium change. Thus, our modified CD63- Nluc assay faithfully reported the secretion of bona fide CD63- positive exosomes, and we have used this method to assess the molecular requirements for exosome secretion. Elevation of cytosolic Ca2+ during plasma membrane damage stimulates exosome secretion We investigated the relationship between elevation of cytosolic Ca2+ levels and exosome secretion. In agreement with previous reports (Messenger et al., 2018), influx of Ca2+ into the cytoplasm using the Ca2+ ionophore ionomycin (5 µM, 30 min) induced robust exosome secretion compared to the level of secretion of the vehicle control (Figure 2A). Exosome secretion initiated within 2 min of ionomycin treatment and was complete by 10 min (Figure 2B). An influx of extracellular Ca2+ triggers lysosome exocytosis as a means to repair plasma membrane damage (Andrews and Corrotte, 2018; Corrotte et al., 2015; Corrotte and Castro- Gomes, 2019; Demonbreun and McNally, 2016). A similar process is elicited in cells treated with a Ca2+ ionophore (Jaiswal et al., 2002). We reasoned that MVBs, like lysosomes, might fuse with the plasma membrane to facilitate membrane repair. Upon treatment with the pore- forming toxin, SLO, exosomes were secreted in a Ca2+- dependent manner (Figure 2C). Treatment with 400 ng/ml or 800 ng/ml of SLO significantly increased exosome production, dependent on the presence of extracellular CaCl2 (1.8 mM). Similar to ionomycin, SLO induced exosome secretion 2 min after initial application and was complete after 10–20 min (Figure 2D). In SLO time course experiments, pore formation was synchro- nized by pre- incubating cells with SLO on ice (Corrotte et al., 2015). We noted that SLO permea- bilized exosomes as well as cells. SLO treatment of conditioned medium decreased the CD63- Nluc signal when the membrane- impermeable Nluc inhibitor was present (Figure 2—figure supplement 1A). Thus, our assay likely underestimates SLO- induced exosome secretion. To test whether plasma membrane damage also elicits exosome secretion in a different cell line, we generated a HEK293T reporter cell line that expresses FLAG- Nluc- CD63 under expression of the weak L30 promoter. We observed an increase in FLAG- Nluc- CD63 exosome secretion from this reporter cell line in a SLO dose- and Ca2+- dependent manner (Figure 2—figure supplement 1B). Application of a physiological mechanical stress also induced exosome secretion in a Ca2+- dependent manner (Figure 2E). HCT116 CD63- Nluc cells were pumped slowly through a narrow- gauge needle to simulate a capillary. In the low mechanical stress regime, cells transiently experienced ~89 dyn/cm2 maximum fluid shear stress, whereas in the high- mechanical stress regime, cells transiently experienced ~178 dyn/cm2 maximum fluid shear stress (Barnes et al., 2012). This form of mechanical stress increased exosome production but only when CaCl2 was present in the conditioned medium. Endothelium and circulating lymphocytes Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 4 of 24 Cell Biology Research article Figure 2. Elevation of cytosolic Ca2+ levels promotes exosome secretion. (A) Normalized exosome production from CD63- nanoluciferase (Nluc) cells treated with 5 µM ionomycin or DMSO (ionomycin vehicle) are shown. (B) Relative rate of exosome secretion over time is shown. CD63- Nluc cells were treated with DMSO or 5 µM ionomycin for the indicated times, and the normalized exosome production index was calculated. (C) Normalized exosome production from CD63- Nluc cells treated for 30 min with increasing concentrations of streptolysin O (SLO), with or without 1.8 mM extracellular Ca2+. Note, error bars are smaller than the dots in the image. (D) Relative rate of exosome secretion over time is shown. CD63- Nluc cells were treated with PBS or 250 ng/ml SLO for the indicated time points, and the normalized exosome production index was calculated. (E) Normalized luminescence derived from CD63- Nluc cells treated with a high or low dose of mechanical stress, with or without 1.8 mM extracellular Ca2+. (F) Iodixanol gradient fractionation of conditioned medium 100k × g pellet fraction is shown. Conditioned medium from cells treated for 30 min with 5 µM ionomycin or 24 hr vehicle is compared. Line graphs show distribution of CD63- Nluc luminescence (with membrane- impermeable inhibitor added) across the linear gradient. Immunoblots show distribution of several extracellular vesicle (EV) markers across the linear gradient. (G) Iodixanol gradient fractionation of conditioned medium 10k × g supernatant fractions are shown. Conditioned medium from cells treated for 4 hr with 5 µM ionomycin or 4 hr vehicle is compared. Line graphs show distribution of CD63- Nluc luminescence (with membrane- impermeable inhibitor added) across the linear gradient. (H) Normalized exosome production from 30 min of 5 µM ionomycin or DMSO vehicle, co- treated with DMSO vehicle, 1 µM latrunculin A (LatA), or 10 µM nocodazole (Noco) is shown. Data plotted represent the means from three independent experiments, and error bars represent SDs. Statistical significance was performed using a Student’s T- test (**p<0.01). The online version of this article includes the following source data and figure supplement(s) for figure 2: Source data 1. Uncropped immunoblot images corresponding to Figure 2. Figure supplement 1. Elevation of cytosolic Ca2+ levels promotes exosome secretion. Figure supplement 1—source data 1. Uncropped immunoblot images corresponding to Figure 2—figure supplement 1. Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 5 of 24 Cell Biology Research article routinely experience mechanical stress of ~95 dyn/cm2 and transiently experience up to ~3000 dyn/ cm2 (Barnes et al., 2012). Next, we used linear iodixanol density gradient fractionation to assess if the Ca2+- dependent increase in extracellular CD63 was attributable to exosomes as opposed to the secretion of intact endosomes contained within plasma membrane- derived vesicles (Jeppesen et al., 2019). EVs from cells treated with ionomycin for 30  min or vehicle for 24  hr were separated on a high- resolution linear iodixanol gradient. A 30 min ionomycin treatment induced the secretion of low- buoyant density, ANXA2- and FLOT2- positive vesicles (Figure 2F, Fraction #8) compared to the 24 hr vehicle control. ANXA2 is a marker for a subpopulation of plasma membrane- derived EVs (Jeppesen et al., 2019). Alternatively, CD63 and Nluc signals equilibrated to higher buoyant density fractions corresponding to the position expected for exosomes (Figure 2F, Fraction #10). A comparison of ionomycin- and vehicle- treated samples showed drug treatment over equivalent times (4 hr) increased the Nluc lumi- nescence in putative exosome fractions (Figure 2G, Fraction #10). Treatment of cells with ionomycin for 4 hr did not induce apoptotic cell death (Figure 2—figure supplement 1C; Gil- Parrado et al., 2002). MVBs traffic toward the cell periphery on microtubules or actin filaments (Mittelbrunn et  al., 2015). We considered the possibility that Ca2+- dependent exosome secretion depended on MVB traffic on microtubules or actin filaments. Treatment with nocodazole, a microtubule polymerization inhibitor, but not latrunculin A, an actin polymerization inhibitor, reduced vehicle- and ionomycin- induced exosome secretion (Figure 2H, Figure 2—figure supplement 1D and E). We conclude that the influx of extracellular Ca2+ caused by several inducers of plasma membrane lesions triggers exosome secretion and that this secretion is dependent on the anterograde trafficking of MVBs on microtubules. Identification of candidate proteins involved in Ca2+-dependent exosome secretion Next, we sought to identify cytosolic proteins that are recruited in a Ca2+- dependent manner to MVBs. Such proteins may be involved in Ca2+- dependent exosome secretion and plasma membrane repair. CD63- Nluc cells were ruptured by homogenization, and a post- nuclear supernatant fraction was supplemented with 1 mM CaCl2 or 1 mM EGTA and mixed with immobilized Nluc antibody in order to immunoprecipitate (IP) MVBs and associated peripheral membrane proteins. After washing the immunoprecipitated MVBs to remove other organelles, Ca2+- dependent MVB binding proteins were eluted from the IP fraction using 2 mM EGTA. MVB proteins retained after the EGTA elution were solubilized using 0.2% TX- 100 (Figure 3A). The 0.2% TX- 100 elution fractions were significantly enriched in LAMP1 and diminished in GAPDH compared to the input, regardless of the presence of CaCl2 or EGTA in the elution (Figure 3B). A gel stained for total protein showed four intense protein bands unique to the EGTA- eluted sample (Figure  3C). These proteins were absent when the Nluc antibody was omitted or when the post- nuclear supernatant was treated with 1 mM EGTA instead of 1 mM CaCl2. Three sections of the gel were excised and analyzed by mass spectrometry (Figure 3D). We identified annexin A6 (ANXA6) and copine 3 (CPNE3) in the upper gel slice, annexin A2 (ANXA2) in the middle gel slice, and S100 Ca2+- binding protein A10 (S100A10) in the lower gel slice. ANXA6 depletion blocks Ca2+-dependent exosome secretion and stalls MVBs at the cell periphery We tested the effect of knocking- down genes encoding the Ca2+- dependent MVB- binding proteins identified in our proteomic analysis. We were unable to knock down ANXA2 using two different shRNAs. However, we were able to efficiently knock down ANXA6 and CPNE3 with two different shRNAs (Figure 4A and B), and in each case, Ca2+- dependent exosome secretion decreased relative to a GFP knockdown control (Figure 4C and D). One of the ANXA6 shRNAs also slightly decreased constitutive exosome secretion relative to the GFP control, although the effect was relatively small (Figure 4C). A polyclonal knockout of ANXA6 also decreased exosome secretion relative to a non- targeting control (Figure 4—figure supplement 1A and B). In order to assess how ANXA6 depletion affected exosome secretion, we visualized CD63- positive compartments after Ca2+ influx. CD63- Nluc cells were treated with ionomycin or vehicle control for 30  min, fixed, and analyzed using immunofluorescence microscopy. We observed an accumulation Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 6 of 24 Cell Biology Research article Figure 3. A targeted proteomics approach identifies genes important for Ca2+- dependent exosome secretion. (A) Schematic illustrating the isolation of Ca2+- dependent multivesicular body (MVB)/lysosome binding proteins (Ab: antibody; gray- beads, blue- MVB, green- mitochondria, orange-ER, red- proteins, and gold- Ca2+). (B) Immunoblot analysis of LAMP1 and GAPDH from the TX- 100 elutions, relative to the input, is shown. (C) Total protein gel (Sypro Ruby stained) of eluted fractions is shown. Red boxes indicate gel cuts sent for proteomic analysis. (D) Table depicting the top four proteomic hits from each gel cut are shown, excluding keratin family proteins. The online version of this article includes the following source data for figure 3: Source data 1. Uncropped immunoblot and gel images corresponding to Figure 3. Source data 2. Mass spectrometry analysis of proteins recruited to multivesicular bodies (MVBs) in the presence of Ca2+. Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 7 of 24 Cell Biology Research article Figure 4. ANXA6 depletion blocks ionomycin- mediated exosome secretion and stalls multivesicular bodies (MVBs) at the cell periphery. (A) Immunoblot analysis of ANXA6 and GAPDH expression from GFP, ANXA6- I, and ANXA6- II shRNA CD63- nanoluciferase (Nluc) cells is shown. (B) Immunoblot analysis of CPNE3 and GAPDH expression from GFP, CPNE3- I, and CPNE3- II shRNA CD63- Nluc cells are shown. (C) Normalized exosome production derived from GFP, ANXA6- I, ANXA6- II, CPNE3- I, and CPNE3- II shRNA CD63- Nluc cells grown in conditioned medium for 24 hr are shown. (D) Normalized exosome production derived from GFP, ANXA6- I, ANXA6- II, CPNE3- I, and CPNE3- II shRNA CD63- Nluc cells treated with 5 µM ionomycin for 30 min are shown. Data plotted represent the means from three independent experiments, and error bars represent SDs. Statistical significance was performed using an ANOVA (*p<0.05, ****p<0.0001, and ns = not significant). (E) CD63 immunofluorescence and phalloidin staining of GFP or ANXA6- I shRNA CD63- Nluc cells after 30 min of DMSO or 5 µM ionomycin treatment are shown. White arrows indicate peripheral CD63 puncta. Scale bars: 10 µm. The online version of this article includes the following source data and figure supplement(s) for figure 4: Source data 1. Uncropped immunoblot images corresponding to Figure 4. Figure 4 continued on next page Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 8 of 24 Cell Biology Research article Figure 4 continued Figure supplement 1. Exosome secretion from polyclonal ANXA6 KO cells. Figure supplement 1—source data 1. Uncropped immunoblot images corresponding to Figure 4—figure supplement 1. of CD63- positive vesicles at the cell periphery, particularly in ionomycin- treated ANXA6 knockdown cells (Figure 4E). We suggest that ANXA6 may be necessary for Ca2+- dependent exosome secretion, possibly at the point of docking between the MVB and the plasma membrane. Biochemical reconstitution of Ca2+- and ANXA6-dependent exosome secretion in permeabilized cells Many studies probing genes that may contribute to exosome secretion have relied on depletion or overexpression experiments in live cells. We sought to develop an assay in permeabilized cells to allow a direct assessment of the roles of different gene products in exosome secretion. SLO is a potent bacterial toxin secreted by group A streptococci that forms stable pores within cholesterol- containing biological membranes in a temperature- dependent manner (Alouf, 1980). SLO has been used to permeabilize cells to reconstitute various intracellular membrane trafficking and organelle exocytosis reactions because it forms stable pores that are large enough to allow the diffusion of large proteins, but not intact organelles, across the plasma membrane (Apodaca et al., 1996; Funato et al., 1997; Martys et al., 1995). We leveraged the cholesterol- and temperature- dependent properties of SLO to establish a permeabilized cell reaction that would allow us to study exosome secretion upon plasma membrane damage. Reaction components including an ATP regeneration system (ATPr), GTP, cytosol, and Ca2+ were mixed with SLO- permeabilized cells and incubated at 30°C and evaluated for exosome secretion just as in the cell- based exosome secretion assay (Figure 5A). We observed that addition of either rat liver cytosol or Ca2+ promoted exosome secretion slightly and that the addition of both rat liver cytosol and Ca2+ promoted exosome secretion ~12- fold over the control condition (Figure 5B). The addition of HCT116 WT cytosol and Ca2+ promoted exosome secretion ~15- fold over the control condition (Figure 5C). To assess the energy dependence of this reaction, we repeated the experiment using nucleotide- depleted cytosol in either the absence or presence of the ATPr and GTP (Figure 5D). We observed that the Ca2+- only reaction was stimulated by the addition of the ATPr and GTP (columns 3 and 7), whereas the stimulatory effect obtained with cytosol alone appeared to be nucleotide- independent (columns 2 and 6). The Ca2+ conditional and partial ATP dependence of the reaction may reflect distinct pools of MVBs, some perhaps already bound to the plasma membrane. The participation of ANXA6 in the reconstitution was assessed with blocking IgG antibodies targeting epitopes on either GFP or ANXA6. Addition of an anti- GFP antibody slightly decreased exosome secretion whereas the equivalent addition of an anti- ANXA6 antibody (whose specificity was validated by immunoblot analysis in Figure 4A and Figure 4—figure supplement 1A) decreased exosome secretion to a background level seen in a reaction without added cytosol (Figure 5E). These results suggest a direct role of ANXA6 in the docking of MVBs at the cell surface. ANXA6 truncations localize to different membranes upon ionomycin treatment After demonstrating that ANXA6 is required for Ca2+- dependent exosome secretion in intact and permeabilized cells, we sought to probe the membrane recruitment of distinct domains of ANXA6 upon Ca2+ influx. Previous studies have demonstrated that ANXA6 is recruited to a peripheral ‘repair cap’ that is formed at the site of plasma membrane lesions (Demonbreun et al., 2016). Unlike other members of the annexin protein family, ANXA6 contains two complete annexin domains, ANXA6(N) and ANXA6(C). To probe the roles of these domains, we generated fluorescent ANXA6 full length (FL), ANXA6(N), and ANXA6(C) fusion constructs and probed their localization in U- 2 OS cells by superresolution microscopy. We observed that ANXA6(FL) maintains a diffuse cytoplasmic distribution in unperturbed cells (Figure  6A). However, upon addition of ionomycin, ANXA6(FL) was recruited to both the plasma membrane (as indicated by wheat germ agglutinin [WGA] counterstain) and to intracellular vesi- cles, a portion of which were CD63- positive (Figure 6A). We also observed the budding of plasma Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 9 of 24 Cell Biology Research article Figure 5. Biochemical reconstitution of Ca2+- and ANXA6- dependent exosome secretion in permeabilized cells. (A) Schematic illustrating the permeabilized cell exosome secretion assay. (B) Permeabilized cell exosome secretion reactions with or without SLO, ATP regeneration system (ATPr)/ GTP, rat liver cytosol, and Ca2+ are indicated. (C) Permeabilized exosome secretion assays with or without HCT116 WT cytosol and Ca2+ are shown. (D) ATP requirements for the permeabilized exosome secretion assay. Reactions with or without nucleotide- depleted rat liver cytosol, Ca2+, and ATPr/ GTP are indicated. (E) Requirement of ANXA6 in the permeabilized exosome secretion assay. Reactions with or without HCT116 WT cytosol, an anti- GFP rabbit IgG antibody, and an anti- ANXA6 rabbit IgG antibody are depicted. Data plotted represent the means from three independent experiments, and error bars represent SDs. Statistical significance was performed using an ANOVA (*p<0.05, **p<0.01, ***p<0.001, ****p<0.0001, and ns = not significant). Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 10 of 24 Cell Biology Research article Figure 6. Localization of full- length and truncated ANXA6 constructs with or without ionomycin treatment. (A) Airyscan 3D projection of U- 2 OS cells expressing ANXA6(FL)- mNG with endogenous CD63 and wheat germ agglutinin (WGA) counterstain upon treatment with DMSO or 1 µM ionomycin for 10 min. Green: ANXA6(FL)- mNG; magenta: endogenous CD63 immunofluorescence; blue: WGA CF640R conjugate. Scale bars: 10 µm. (B) Airyscan 3D projection of U- 2 OS cells expressing ANXA6(N)- mNG and ANXA6(C)- mSc with WGA counterstain upon treatment with DMSO or 1 µM ionomycin for Figure 6 continued on next page Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 11 of 24 Cell Biology Research article Figure 6 continued 10 min. Green: ANXA6(N)- mNG; magenta: ANXA6(C)- mSc; blue: WGA CF640R conjugate. Scale bars: 10 µm. (C) Airyscan 3D projection of U- 2 OS cells expressing ANXA6(N)- mNG and ANXA6(C)- mSc with endogenous CD63 upon treatment with DMSO or 1 µM ionomycin for 10 min. Green: ANXA6(N)- mNG; magenta: ANXA6(C)- mSc; blue: endogenous CD63 immunofluorescence. Scale bars: 10 µm. The online version of this article includes the following video(s) for figure 6: Figure 6—video 1. Rotating 3D projection of DMSO- treated U- 2 OS merge in Figure 6A. https://elifesciences.org/articles/86556/figures#fig6video1 Figure 6—video 2. Rotating 3D projection of ionomycin- treated U- 2 OS merge in Figure 6A. https://elifesciences.org/articles/86556/figures#fig6video2 Figure 6—video 3. Rotating 3D projection of DMSO- treated U- 2 OS merge in Figure 6B. https://elifesciences.org/articles/86556/figures#fig6video3 Figure 6—video 4. Rotating 3D projection of ionomycin- treated U- 2 OS merge in Figure 6B. https://elifesciences.org/articles/86556/figures#fig6video4 Figure 6—video 5. Rotating 3D projection of DMSO- treated U- 2 OS merge in Figure 6C. https://elifesciences.org/articles/86556/figures#fig6video5 Figure 6—video 6. Rotating 3D projection of ionomycin- treated U- 2 OS merge in Figure 6C. https://elifesciences.org/articles/86556/figures#fig6video6 membrane- derived vesicles in the presence of ionomycin. These may correlate to the low- buoyant density vesicles marked by ANXA2 and FLOT2 that were released from ionomycin- treated cells (Figure  2F, Fraction #8). Next, we expressed fluorescently tagged ANXA6(N) and ANXA6(C) constructs and assessed their localization upon the addition of DMSO or ionomycin, relative to the plasma membrane and CD63, respectively (Figure 6B and C). Similar to ANXA6(FL), ANXA6(N) and ANXA6(C) displayed cytoplasmic localization in unstimulated cells. Upon ionomycin treatment, we observed recruitment of ANXA6(N), but not ANXA6(C) to the cell periphery, relative to the WGA counterstain (Figure 6B). In contrast to ANXA6(N), we observed that ANXA6(C) became enriched at intracellular membrane structures, a portion of which were positive for CD63 (Figure 6C). These two domains may serve to dock MVBs to the plasma membrane. Discussion Our results suggest that exosome secretion is coupled to Ca2+- dependent plasma membrane repair in HCT116 and HEK293T cells (Figure 7). We established cellular and biochemical exosome secretion assays to recapitulate and interrogate this process. Targeted proteomics was used to demonstrate that ANXA6 is recruited to MVBs in the presence of Ca2+. We show, both in intact and permeabilized cells, that ANXA6 is required for Ca2+- dependent exosome secretion. We then provide evidence that that ANXA6 depletion stalls MVBs at the cell periphery upon Ca2+ ionophore treatment. Finally, we demonstrate that truncations of ANXA6 are enriched at different cellular membranes upon cytosolic Ca2+ elevation. Exosomes are secreted upon damage to the plasma membrane We demonstrate that SLO- treated or mechanically stressed HCT116 CD63- Nluc and SLO- treated HEK293T FLAG- Nluc- CD63 cells secrete exosomes in response to Ca2+- dependent plasma membrane repair (Figure  2 and Figure  2—figure supplement 1B). Such treatments may mimic conditions of physiologic stress in vivo and lead to wound- induced EV secretion. This may contribute to the diver- sity of EVs present in biological fluids, including in samples used for liquid biopsy. Plasma membrane repair is common, especially in certain tissues. In rats, 6.5% of cells in the vascular endothelium of the aorta undergo plasma membrane repair at a given time (Yu and McNeil, 1992); 20% of muscle cells repair their membrane after muscle contractions (McNeil and Khakee, 1992). Motile cancer cells also have high rates of plasma membrane repair (Bouvet et al., 2020). Interestingly, Whitham, 2018 demonstrated that exercise stimulates the release of EV- associated proteins and myokines into circulating plasma (Whitham, 2018). In this study, a cohort of human participants had blood drawn from their femoral artery before and after 60 min of acute exercise on a cycle ergometer. Label- free quantitative proteomic analysis showed an increase in proteins such Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 12 of 24 Cell Biology Research article Figure 7. Schematic depicting the current model of Ca2+- and ANXA6- dependent exosome secretion. (A) Upon physiological mechanical stress or bacterial infection, plasma membrane lesions form. This results in the flow of Ca2+ from the extracellular space into the cytoplasm. (B) This influx of Ca2+ mediates the recruitment of ANXA6 to multivesicular bodies (MVBs), which are then transported on microtubules to a plasma membrane lesion. (C) The MVBs then dock at the plasma membrane (with ANXA6 serving as a putative membrane tether) and undergo fusion, resulting in plasma membrane repair and exosome secretion. as CD63, TSG101, CD9, and ANXA2 within EVs isolated from plasma by differential centrifugation postexercise. It is tempting to speculate that the increase in EV secretion is caused by increased plasma membrane repair. We postulate that exosome secretion during plasma membrane repair may allow for simple detection of tissue- specific damage during the routine clinical analysis of blood or urine samples. The biological functions of exosomes released during Ca2+- stimulated secretion remain unclear. Much attention has focused on the role of exosomes in intercellular communication (Meldolesi, 2018). A recent study from our laboratory established quantitative assays to measure the functional delivery of exosome cargo to recipient cells (Zhang and Schekman, 2023). Functional transfer of exosome cargo was inefficient, whereas functional intercellular transfer occurred quite effectively through the formation of open- ended membrane tubular connections. However, we do not exclude the possibility that exosomes and other EVs may functionally convey their cargo to recipient cells in specific phys- iological contexts (Ridder et al., 2014). Alternatively, a recent study showed that exosomes act as ‘decoys’ for pore- forming toxins such as α-toxin (Keller et al., 2020). During staphylococcus infection, cells upregulate exosome production. Instead of binding to the plasma membrane, α-toxin binds to exosomes. Correspondingly, we find that SLO binds exosomes at concentrations that may compete with plasma membrane binding (Figure 2—figure supplement 1A). Ca2+- dependent exosome secre- tion allows for the rapid secretion of high levels of exosomes and may serve as a defense against bacterial toxins that perforate the plasma membrane. Although our focus has been on exosome secretion, others have reported plasma membrane- derived vesicle secretion stimulated by plasma membrane disruptions (Horn and Jaiswal, 2018). Such shedded vesicles may correspond to the low- buoyant density vesicles marked by ANXA2 and FLOT2 which we find in the EV fraction produced by ionomycin treatment (Figure 2D). Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 13 of 24 Cell Biology Research article Do MVBs participate in plasma membrane repair? The egress of CD63- positive exosomes from HCT116 and HEK293T cells suggests that MVBs partic- ipate in the Ca2+- dependent repair of plasma membrane lesions. Microinjection studies in echino- derm eggs were the first to demonstrate that metazoan cells repair wounds to the plasma membrane and that this repair process is dependent upon the presence of extracellular Ca2+ (Chambers, 1917; Heilbrunn, 1930a; Heilbrunn, 1930b). Since then, a variety of mechanisms by which metazoan cells repair their plasma membrane (e.g. membrane shedding, organelle exocytosis, clot formation/wound constriction, and membrane internalization) has been proposed (Andrews and Corrotte, 2018; Cooper and McNeil, 2015; Demonbreun and McNally, 2016; Horn and Jaiswal, 2018). Jaiswal et al., 2002 used total internal reflection fluorescence microscopy to identify the primary membrane compartment responsible for Ca2+- dependent exocytosis in non- secretory cells. In their studies, the addition of a Ca2+ ionophore elicited the fusion of plasma membrane- proximal lysosomes with little or no evidence of involvement of early endosomes, late endosomes/MVBs, or vesicles derived from either the endoplasmic reticulum or Golgi apparatus. Our data that CD63- positive compart- ments fuse at the plasma membrane upon Ca2+ influx is consistent with their findings. However, our results also suggest that endosomes/MVBs can undergo Ca2+- dependent exocytosis. We conclude this based on the secretion of MVB cargo vesicles (exosomes) upon Ca2+ influx elicited by a Ca2+ iono- phore, a pore- forming toxin, or physiological levels of mechanical stress. Our results also suggest that both free and pre- docked late endosomes/MVBs participate in plasma membrane repair as exosome secretion induced by Ca2+ ionophore treatment is partially microtubule- dependent. The markers used by Jaiswal et al., 2002 to differentiate between late endosomes (Rab7) and lysosomes (CD63) have now been demonstrated to be present in both membrane compartments (Humphries et al., 2011). The combined data suggest that, in addition to lysosomes, late endosomes/MVBs can undergo Ca2+- dependent exocytosis. Our results are consistent with published studies conducted in sea urchin eggs and cytotoxic T lymphocytes (Keefe et al., 2005; Steinhardt et al., 1994). Ca2+- dependent plasma membrane repair of sea urchin eggs after micropuncture with glass micropipettes has been demonstrated to be both fusion- and microtubule- dependent as suggested by sensitivity to either botulinum toxins or anti- bodies that block kinesin- mediated microtubule transport (Ingold et al., 1988; Poulain et al., 1988; Schiavo et  al., 1992; Steinhardt et  al., 1994). Endosomes pre- loaded with Alexa- 488- conjugated transferrin traffic toward the plasma membrane upon treatment of cytotoxic T lymphocytes with perforin (Keefe et al., 2005). This result is consistent with our finding that MVBs undergo exocytosis in cells treated with SLO. Given our findings, we propose that both MVBs and lysosomes undergo Ca2+- dependent exocy- tosis during plasma membrane repair, resulting in the concurrent secretion of exosomes (Figure 7). Role of ANXA6 in exosome secretion We find that ANXA6 is required for Ca2+- dependent fusion of MVBs with the plasma membrane (Figure 4). Annexin proteins are known to be required for plasma membrane repair (Blazek et al., 2015; Bouter et al., 2015; Boye and Nylandsted, 2016; Croissant et al., 2021; Gerke and Moss, 2002; Koerdt et  al., 2019). Depletion of ANXA6 compromises plasma membrane repair, and the N- terminal domain of ANXA6 is insufficient for membrane repair (Croissant et  al., 2020; Potez et al., 2011; Swaggart et al., 2014). In addition to lysosome exocytosis, plasma membrane repair is accompanied by shedding of damaged membrane at the cell surface (Andrews and Corrotte, 2018). Annexins have been invoked in the capping and shedding of plasma membrane lesions (Demonbreun et al., 2016). Our results extend these conclusions to the fusion of MVBs at the cell surface. Our data favors a model where ANXA6 is directly involved in tethering an MVB to the plasma membrane during exocytosis. ANXA6 is recruited to CD63- positive compartments in a Ca2+- dependent manner (Figure  3). Additionally, CD63- positive membrane compartments stall near the plasma membrane in Ca2+- stimulated, ANXA6- depleted cells (Figure 4E). This conclusion is strength- ened by our observation that an anti- ANXA6 antibody blocks exosome secretion in permeabilized cells (Figure 5E). Unlike other members of the annexin family, ANXA6 contains two distinct annexin domains, possibly one for each membrane partner of a fusion pair. Upon ionomycin stimulation, we observed that the N- terminal annexin domain, but not the C- terminal annexin domain, was recruited to the plasma membrane and that the C- terminal annexin domain became enriched at intracellular Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 14 of 24 Cell Biology Research article membrane structures, a portion of which were CD63- positive (Figure 6B and C). ANXA6 has also been demonstrated to tether liposomes in vitro in the presence of Ca2+ (Buzhynskyy et al., 2009). Alter- native roles in docking, such as SNARE interactions or actin remodeling as seen for other annexins, remain possible (Gabel et  al., 2015; Gerelsaikhan et  al., 2012). Additionally, a recent study has demonstrated that the annexin isoform ANXA11 serves as a molecular tether that allows RNA gran- ules to ‘hitchhike’ on moving lysosomes for their transport (Liao et al., 2019). Although we focused on the role of ANXA6 in Ca2+- dependent exosome secretion, ANXA2 and CPNE3 were both recruited to MVBs and lysosomes in the presence of Ca2+ and knockdown of CPNE3 inhibited MVB exocytosis (Figures 3 and 4D). Thus, we hypothesize that these proteins could also be involved in the docking of MVBs and lysosomes at the plasma membrane during repair. Key resources table Materials and methods Reagent type (species) or resource Designation Source or reference Identifiers Additional information Gene (Homo sapiens) ANXA6 Addgene RRID: Addgene_29509 Gene (Homo sapiens) CPNE3 N/A N/A Abcam Abcam Catalog #: ab178677 Catalog #: ab201024 N/A N/A WB: (1:1000) WB: (1:1000), see Materials and methods for further details. Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Antibody Anti- ANXA2 (rabbit monoclonal) Anti- ANXA6 (rabbit monoclonal) Anti- CD9 (rabbit monoclonal) Anti- CD63 (mouse monoclonal) Anti- CPNE3 (rabbit polyclonal) Anti- Flotillin- 2 (mouse monoclonal) Anti- GAPDH (rabbit monoclonal) Anti- GFP (rabbit polyclonal) Anti- LAMP1 (rabbit monoclonal) Anti- Nluc (mouse monoclonal) Anti- Tubulin (mouse monoclonal) Anti- TSG101 (mouse monoclonal) Anti- Vinculin (rabbit monoclonal) Cell Signaling RRID: AB_2798139 WB: (1:1000) BD Biosciences RRID: AB_396297 WB: (1:1000), IF: (1:250) Sigma RRID: AB_10600703 WB: (1:1000) BD Biosciences RRID: AB_397766 WB: (1:1000) Cell Signaling RRID: AB_561053 WB: (1:1000) Torrey Pines Biolabs RRID: AB_10013661 See Materials and methods for further details. Cell Signaling RRID: AB_2687579 WB: (1:1000) Promega Catalog #: N7000 See Materials and methods for further details. Abcam RRID: AB_2241126 IF: (1:250) GeneTex RRID: AB_373239 WB: (1:1000) Abcam RRID: AB_11144129 WB: (1:1000) Cell line (Homo sapiens) Cell Line (Homo sapiens) Cell line (Homo sapiens) Cell line (Homo sapiens) Cell line (Homo sapiens) Continued on next page HCT116 CD63Nluc- KI #17 Hikita et al., 2018 N/A Dr. Chitose Oneyama (Aichi Cancer Center Research Institute) HCT116 HEK293T HEK293T FLAG- Nluc- CD63 Other Other This study U- 2 OS Other N/A N/A N/A N/A Cell culture facility at UC Berkeley Cell culture facility at UC Berkeley To assess exosome secretion in HEK293T cells Cell culture facility at UC Berkeley Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 15 of 24 Cell Biology Research article Continued Reagent type (species) or resource Designation Source or reference Identifiers Additional information Chemical compound and drug Ionomycin Cayman Chemical Catalog #: 10004974 See Materials and Methods for further details. Recombinant DNA reagent pLKO.1_Hygro GFP shRNA This study Recombinant DNA reagent pLKO.1_Hygro ANXA2- 1 shRNA This study Recombinant DNA reagent pLKO.1_Hygro ANXA2- 2 shRNA This study Recombinant DNA reagent pLKO.1_Hygro ANXA6- 1 shRNA This study Recombinant DNA reagent pLKO.1_Hygro ANXA6- 2 shRNA This study Recombinant DNA reagent pLKO.1_Hygro CPNE3- 1 shRNA This study Recombinant DNA reagent pLKO.1_Hygro CPNE3- 2 shRNA This study Recombinant DNA reagent lentiCRISPR v2- Blast sgNT (non- targeting) This study Recombinant DNA reagent lentiCRISPR v2- Blast sgANXA6 This study Recombinant DNA reagent pLJM1- L30- FLAG- Nluc- CD63- hPGK- BlastR This study Recombinant DNA reagent pN1- mNeonGreen (mNG) This study Recombinant DNA reagent pN1- mScarlet- i (mSc) This study Recombinant DNA reagent pN1_ANXA6(FL; aa1- 673)- mNG This study Recombinant DNA reagent pN1_ANXA6(N; aa1- 322)- mNG This study Recombinant DNA reagent pN1_ANXA6(C; aa323- 673)- mSc This study N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A shRNA target sequence: ACAA CAGC CACA ACGT CTAT shRNA target sequence: ACTT TAGA AACA CGTA CTTT G shRNA target sequence: TGAG GGTG ACGT TAGC ATTA C shRNA target sequence: AGTT GGAC ATGC TCGA CATT C shRNA target sequence: CGAA GACA CAAT CATC GATA T shRNA target sequence: ACTC TATG GACC AACT AATT T shRNA target sequence: AGCA TTCT TTCT AGGT TATT T Protospacer sequence: GCCC CGCC GCCC TCCC CTCC Protospacer sequence: AGCC TCCA GGTC CCGC TCGT To express FLAG- Nluc- CD63 in various cell lines under control of the low- expression L30 promoter To assess the localization of mNeonGreen- tagged proteins To assess the localization of mScarlet- i- tagged proteins To assess localization of full- length ANXA6 To assess localization of N- terminal truncation of ANXA6 To assess localization of C- terminal truncation of ANXA6 Software, algorithm Prism 9 GraphPad RRID:SCR_002798 N/A Peptide, recombinant protein Alexa Fluor 680 Phalloidin Invitrogen Peptide, recombinant protein CF640R Wheat Germ Agglutinin Conjugate Biotium N/A N/A IF: (1:400) IF: (5 µg/ml) Peptide, recombinant protein Streptolysin O (SLO) Sigma- Aldrich N/A See Materials and methods for further details. Cell lines, media, and general chemicals HCT116, HCT116 CD63- Nluc, HEK293T, HEK293T FLAG- Nluc- CD63, and U- 2 OS cells were cultured at 37°C in 5% CO2 and maintained in Dulbecco’s Modified Eagle’s Medium (DMEM) supplemented with 10% fetal bovine serum (FBS) (Thermo Fisher Scientific, Waltham, MA, USA). Cells were routinely tested (negative) for mycoplasma contamination using the MycoAlert Mycoplasma Detection Kit (Lonza Biosciences). HCT116, HEK293T, and U- 2 OS cells were authenticated using STR profiling at the UC Berkeley Cell Culture Facility. For the experiments detailed in Figure 2B and C and Figure 2— figure supplement 1B, we cultured HCT116 CD63- Nluc and HEK293T FLAG- Nluc- CD63 cells in Ca2+- free DMEM (Thermo Fisher Scientific). For Figure 2F, HCT116 CD63- Nluc cells were incubated in EV- depleted medium. EV- depleted medium was produced by ultracentrifugation at 186,000 × g (40,000 RPM) for 24 hr using a Type 45Ti rotor. Ionomycin was purchased from Cayman Chemicals. Unless otherwise noted, all other chemicals were purchased from Sigma- Aldrich (St. Louis, MO, USA). Lentivirus production and transduction HEK293T cells at 40% confluence within a 6- well plate were transfected with 165 ng of pMD2.G, 1.35 µg of psPAX2, and 1.5 µg of a pLKO.1- Hygro, lentiCRISPR v2- Blast, or pLJM1 plasmid using the Tran- sIT- LT1 Transfection Reagent (Mirus Bio) as per the manufacturer’s protocol. At 48 hr post- transfection, 1 ml of fresh DMEM supplemented with 10% FBS was added to each well. The lentivirus- containing medium was harvested 72 hr post- transfection by filtration through a 0.45 μm polyethersulfone (PES) filter (VWR Sciences). The filtered lentivirus was distributed in aliquots, snap- frozen in liquid nitrogen, and stored at –80°C. For lentiviral transductions, we infected HCT116 CD63- Nluc cells with filtered Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 16 of 24 Cell Biology Research article lentivirus in the presence of 8  μg/ml polybrene for 24  hr, and the medium was replaced. HCT116 CD63- Nluc cells were selected using 200 μg/ml hygromycin B or 4 μg/ml blasticidin S for 8 days and 6 days, respectively. The efficiency of each gene knockdown was assessed by immunoblot analysis. Immunoblotting Cells were lysed in PBS containing 1%  TX- 100 and a protease inhibitor cocktail (1  mM 4- aminobenzamidine dihydrochloride, 1 µg/ml antipain dihydrochloride, 1 µg/ml aprotinin, 1 µg/ml leupeptin, 1 µg/ml chymostatin, 1 mM phenylmethylsulfonyl fluoride, 50 µM N- tosyl- L- phenylalanine chloromethyl ketone, and 1 µg/ml pepstatin) and incubated on ice for 15 min. The whole cell lysate was centrifuged at 15,000 × g for 10 min at 4°C and the post- nuclear supernatant was diluted with 6× Laemmli buffer (without DTT) to a 1× final concentration. Samples were heated at 95°C for 5 min and proteins resolved on 4–20%  acrylamide Tris- glycine gradient gels (Life Technologies). Proteins were then transferred to polyvinylidene difluoride membranes (EMD Millipore, Darmstadt, Germany), blocked with 5% dry milk in TBS, washed 3× with TBS- T and incubated overnight with primary anti- bodies in 5% bovine serum albumin in TBS- T. The membranes were then washed again 3× with TBS- T, incubated for 1 hr at room temperature with 1:10,000 dilutions of anti- rabbit or anti- mouse secondary antibodies (GE Healthcare Life Sciences, Pittsburgh, PA, USA), washed 3× with TBS- T, washed once with TBS and then detected with ECL- 2 or PicoPLUS reagents (Thermo Fisher Scientific) for proteins from cell lysates or EV isolations, respectively. Immunofluorescence microscopy For immunofluorescence, we grew cells on Poly- D- Lysine (PDL)- coated coverslips which were then washed once with PBS, fixed in 4% EM- grade paraformaldehyde (PFA; Electron Microscopy Science Hatfield, PA, USA) for 15 min at room temperature, washed 3× with PBS and permeabilized/blocked in blocking buffer (2% BSA and 0.02% saponin in PBS) for 30 min at room temperature. Cells were then incubated with a 1:250 dilution of primary antibody and/or phalloidin stain overnight at 4°C, washed 3× with PBS, incubated with a 1:1000 dilution of fluorophore- conjugated secondary antibody for 1 hr at room temperature and washed 3× with PBS. Coverslips were then mounted overnight in ProLong- Gold antifade mountant with DAPI (Thermo Fisher Scientific) and sealed with clear nail polish before imaging. For microtubule visualization, cells were fixed using ice cold methanol and processed as above. For Figure  6, we transfected U- 2 OS cells grown on PDL- coated coverslips with the indicated plasmids using Lipofectamine 2000 as per the manufacturer’s protocol. The medium containing the transfection mixture was removed ~5 hr post- transfection and replaced with DMEM supplemented with 10% FBS. Cells were treated 16  hr post- transfection with DMSO or 1  µM ionomycin with or without 5 µg/ml CF640R WGA conjugate (Biotium) as indicated for 10 min at 37°C. Treated cells were then washed with PBS, fixed in 4% EM- grade PFA for 15 min at room temperature, and processed as above. The concentrations utilized for each plasmid transfection were as follows: pN1- ANXA6(FL)- mNeonGreen (200 ng/ml), pN1- ANXA6(N)- mNeonGreen (200 ng/ml), and pN1- ANXA6(C)- mScarlet (200 ng/ml). The images in Figure 2—figure supplement 1 and Figure 4 were acquired on an Echo Revolve Microscope using the 60× Apo Oil Phase, NA 1.42 objective. The images in Figure 6 were acquired using an LSM900 confocal microscope system (ZEISS) using Airyscan 2 superresolution mode and a 63× Plan- Apochromat, NA 1.40 objective. EV isolation and fractionation by iodixanol buoyant density gradient equilibration Fresh aliquots of 5, 7.5, 10, 12.5, 15, 17.5, 20, 22.5, and 25% (v/v) iodixanol solutions were prepared by mixing appropriate volumes of Solution B (0.25  M sucrose, 2  mM MgCl2, 1  mM EDTA, 20  mM Tris- HCl, and pH 7.4) and Solution D (41.7 mM sucrose, 2 mM MgCl2, 1 mM EDTA, 20 mM Tris- HCl, pH 7.4, and 50% (w/v) iodixanol). Iodixanol gradients were prepared by sequential 500 µl overlays of each iodixanol solution in a 5 ml SW55 tube, starting with the 25% iodixanol solution and finishing with the 5% iodixanol solution. After the addition of each iodixanol solution, the SW55 tube was flash frozen in liquid nitrogen. Complete iodixanol gradients were stored at –20°C and thawed at room temperature for 45 min prior to use. Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 17 of 24 Cell Biology Research article For the iodixanol gradients detailed in Figure  2F, conditioned medium (240  ml) was harvested from vehicle- or ionomycin- treated HCT116 CD63- Nluc cells. All subsequent manipulations were completed at 4°C. Cells and large debris were removed by low- speed sedimentation at 1,000 × g for 15  min in a Sorvall R6  +centrifuge (Thermo Fisher Scientific) followed by medium- speed sedi- mentation at 10,000 × g for 15 min using a fixed angle FIBERlite F14−6×500 y rotor (Thermo Fisher Scientific). The supernatant fraction was then centrifuged at 29,500 RPM for 1.25 hr in a SW32 rotor. The high- speed pellet fractions were resuspended in PBS, pooled, loaded at the top of a prepared iodixanol gradient, and centrifuged in a SW55 rotor at 36,500 RPM for 16 hr with minimum accelera- tion and no brake. Fractions (200 µl) were collected from top to bottom. An aliquot of each fraction was saved for luminescence analysis, and the rest was diluted in 6× Laemmli buffer (without DTT) for immunoblot analysis. Density measurements were taken using a refractometer. For the iodixanol gradients detailed in Figure 2G, we harvested conditioned medium from vehicle- or ionomycin- treated HCT116 CD63- Nluc cells grown in a 12- well plate. All subsequent manipulations were completed at 4°C. Cells and large debris were removed by low- speed sedimentation at 1000 × g for 15 min followed by medium- speed sedimentation at 10,000 × g for 15 min in an Eppendorf 5430 R centrifuge (Eppendorf, Hamburg, Germany). Aliquots (200 µl) of conditioned medium from the supernatant of the medium- speed centrifugation were loaded at the top of a prepared iodixanol gradient and centrifuged in a SW55 rotor at 36,500 RPM for 16 hr with minimum acceleration and no brake. Fractions (200 µl) were collected from top to bottom and analyzed for luminescence. Density measurements were taken using a refractometer. Immunoisolation of MVBs and Ca2+-dependent binding proteins HCT116 CD63- Nluc cells were grown to ~90% confluence in 7×150 mm dishes. Cells were scraped into 5 ml of cold PBS per plate and centrifuged at 200 × g for 5 min at 4°C. The cold PBS was aspi- rated, and the cell pellet was resuspended in two volumes of cold lysis buffer (136 mM KCl, 10 mM KH2PO4 pH 7.4, 1 mM DTT, protease inhibitor cocktail [see Immunoblotting section], and 6% Optiprep [w/v]). The cell slurry was passed 14 times through a 25- gauge syringe in a cold room, and the post- nuclear supernatant was prepared by centrifugation of lysed cells at 1000 × g for 10 min at 4°C. The post- nuclear supernatant was diluted 1:2 with lysis buffer. Beads from 300 µl of magnetic Protein G Dynabeads (Thermo Fisher Scientific) slurry were sedi- mented with a magnetic tube rack and resuspended in lysis buffer. The bead slurry was split evenly into three tubes and then re- centrifuged. The diluted post- nuclear supernatant was divided between the three tubes. Tube #1 also received 1 mM CaCl2 (Beads only control), tube #2 also received 1 mM CaCl2 and 5 µg anti- Nluc antibody (Ca2+- treated), and tube #3 also received 1 mM EGTA and 5 µg anti- Nluc antibody (EGTA treated). The reaction was incubated for 15  min at room temperature. Beads were washed three times with lysis buffer. A beads only control and Ca2+- treated samples both had 2 mM CaCl2 added to the wash buffer. The EGTA- treated sample was washed with lysis buffer containing 2 mM EGTA. The beads only control and Ca2+- treated samples were eluted with 50 µl lysis buffer containing 2 mM EGTA, and the EGTA- treated sample was eluted with lysis buffer containing 2 mM CaCl2. All three samples were eluted again with 50 µl lysis buffer containing 0.2% TX- 100. CD63-Nluc exosome secretion assay HCT116 CD63- Nluc cells were grown to ~80% confluence in 24- well plates. All subsequent manipu- lations were performed at 4°C. Conditioned medium (200 µl) was taken from the appropriate wells, added to a microcentrifuge tube, and centrifuged at 1000 × g for 15 min in an Eppendorf 5430 R centrifuge (Eppendorf, Hamburg, Germany) to remove intact cells. Supernatant fractions (150 µl) from the low- speed sedimentation were moved to a new microcentrifuge tube and centrifuged at 10,000 × g for 15 min to remove cellular debris. Supernatant fractions (50 µl) from this medium- speed centrifu- gation were then utilized to measure CD63- Nluc exosome luminescence. During these centrifugation steps, the cells were placed on ice, washed once with cold PBS, and lysed in 200 µl of PBS containing 1% TX- 100 and protease inhibitor cocktail. To measure CD63- Nluc exosome secretion, we prepared a master mix containing the membrane- permeable Nluc substrate and a membrane- impermeable Nluc inhibitor using a 1:1000 dilution of Extracellular NanoLuc Inhibitor and a 1:333 dilution of NanoBRET Nano- Glo Substrate into PBS (Promega, Madison, WI, USA). Aliquots of the Nluc substrate/inhibitor master mix (100 µl) were added Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 18 of 24 Cell Biology Research article to 50 µl of the supernatant fraction from the medium- speed centrifugation and vortexed briefly, and luminescence was measured using a Promega GlowMax 20/20 Luminometer (Promega, Madison, WI, USA). An aliquot (1.5  µl) of 10%  TX- 100 was then added to each reaction tube for a final concen- tration of 0.1%  TX- 100, and the sample was briefly vortexed before luminescence was measured again. For the intracellular normalization measurement, the luminescence of 50 µl of cell lysate was measured using the Nano- Glo Luciferase Assay kit (Promega, Madison, WI, USA) as per the manufac- turer’s protocol. The exosome production index (EPI) for each sample is calculated as follows: EPI = ([medium] – [medium + 0.1% TX- 100])/cell lysate. For the CD63- Nluc exosome secretion assays in Figure 2C and D and Figure 2—figure supple- ment 1B, cytosol was not depleted as in the biochemical reconstitution assays detailed in Figure 5. SLO time courses HCT116 CD63- Nluc cells were grown to ~80% confluence in 24- well PDL- coated plates. Cells were washed with 200 µl of cold PBS and incubated with 200 µl of cold (4°C) 250 ng/ml SLO in Ca2+- free DMEM with 1 mM EGTA. The cold, SLO- containing medium was aspirated, and 200 µl of pre- warmed (37°C) Ca2+- free DMEM with 1.8  mM CaCl2 was added. Cells were incubated at 37°C for the indi- cated times after which exosome secretion was measured as described in paragraph 2 of ‘CD63- Nluc exosome secretion assay’. Mechanical stress experiments Two 15 cm plates of HCT116 CD63- Nluc cells were harvested with 10 ml of Accutase and diluted with 40 ml of Ca2+- free DMEM. A cell slurry (16 ml) was added to three tubes and centrifuged at 300 × g for 5 min at room temperature. Cells were gently resuspended in either 0.5 ml Ca2+- free DMEM or Ca2+- free DMEM +2 mM Ca2+ (final). Cells were pumped through a 30- gauge needle at a flow rate of 3.5 µl/s (τ=4Qη/πR3 = ~89 dyn/cm2, where Q=0.0035 cm3/s, η=0.01 dyn*s/cm2, 30 G average internal radius = 7.94×10−3 cm) or twice the flow rate of 7 µl/s (τ=4Qη/πR3=178 dyn/cm2, where Q=0.0035 cm3/s, η=0.01 dyn*s/cm2, 30 G average internal radius = 7.94×10−3 cm) using a Harvard Apparatus syringe pump (Catalog No. 98–4730). Cells were incubated for 5 min at 37°C before being placed back on ice. The cell suspension was centrifuged at 300 × g for 5 min, and the supernatant fraction was then filtered through a 0.45  µm PES filter. Exosome secretion was measured as described in paragraph 2 of ‘CD63- Nluc exosome secretion assay,’ in this assay without a cell lysate measurement. Isolation of cytosol from cultured human cells HCT116 WT cells were grown to ~90% confluence in 20×150 mm dishes. All subsequent manipula- tions were performed at 4°C. Each 150 mm dish was washed once with 10 ml of cold PBS and then harvested by scraping into 5 ml of cold PBS. The 5 ml cell suspension was then used to harvest cells from four additional 150 mm dishes, and this process was repeated until all the cells were harvested. Cells were then collected by centrifugation at 200 × g for 5 min, and the supernatant fraction was discarded. The cell pellet was resuspended in 3 ml of cold hypotonic lysis buffer (20 mM HEPES, pH 7.4, 10 mM KCl, 1 mM EGTA, 1 mM DTT, and protease inhibitor cocktail) and placed on ice. After 15 min, the cell suspension was transferred to a pre- chilled 7 ml Dounce homogenizer, and the cells were mechanically lysed by ~80 strokes with a tight- fitting Dounce pestle. The lysed cells were centri- fuged at 1000 × g for 15 min to sediment intact cells and nuclei, and the post- nuclear supernatant was then centrifuged at 32,500 RPM (~128,000 × g) for 30 min in an Optima XE- 90 ultracentrifuge (Beckman Coulter). The supernatant (cytosol fraction) was collected conservatively without disturbing the pellet and then concentrated using a 4 ml Amicon–3 k concentrator to a final protein concentra- tion of ~40 mg/ml. The cytosol was distributed in aliquots, snap- frozen in liquid nitrogen, and stored at –80°C until use. The rat liver cytosol was prepared as described in Tang et al., 2020. Reconstitution of CD63-Nluc exosome secretion in permeabilized cells SLO was pre- activated in PBS containing 10 mM DTT at 37°C for 2 hr, distributed in aliquots into low- retention microcentrifuge tubes, snap- frozen in liquid nitrogen, and stored at –80°C until use. The protein concentration of each SLO batch was determined by a Bradford assay. HCT116 CD63- Nluc cells were grown to ~80% confluence in 24- well PDL- coated plates. The PDL plate was placed on ice, and the cells were washed once with PBS containing 1 mM EGTA. The PBS Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 19 of 24 Cell Biology Research article wash was aspirated, replaced with 200 µl of cold transport buffer (20 mM HEPES, pH 7.4, 250 mM D- sorbitol, 120 mM KCl, 10 mM NaCl, 2 mM MgCl2, 1.2 mM KH2PO4, 1 mM EGTA, and protease inhib- itor cocktail), supplemented with 0.6 µg/ml of pre- activated SLO, and incubated at 4°C for 15 min. Unbound SLO was aspirated, and the cells were washed once with cold transport buffer. Cells were then permeabilized by the addition of pre- warmed transport buffer containing 2 mM DTT followed by a 10 min incubation at 37°C. Permeabilized CD63- Nluc cells were then washed at 4°C in transport buffer, high- salt transport buffer (containing 1 M KOAc) and finally transport buffer (10 min each wash) to deplete cytosol. Complete permeabilized cell exosome secretion assays (200 µl) consisted of permeabilized CD63- Nluc cells, cytosol (4 mg/ml final concentration), 20 µl 10× ATPr (10 mM ATP, 400 mM creatine phos- phate, 2 mg/ml creatine phosphokinase, 20 mM HEPES, pH 7.2, 250 mM D- sorbitol, 150 mM KOAc, and 5 mM MgOAc), 3 µl of 10 mM GTP, 4 µl of 100 mM CaCl2 (2 mM final concentration), and cold transport buffer. The concentration of blocking IgG antibodies utilized in Figure 5E was 25 µg/ml. The assembled reaction mixes were added to the permeabilized CD63- Nluc cells for 5 min on ice prior to placing the entire 24- well PDL coated plate in a 30°C water bath for 2 min to stimulate exosome secretion. Permeabilized cells were then placed back on ice, and 100 µl of each reaction supernatant was loaded into a 0.4 µm AcroPrep filter plate (Pall Corporation) and centrifuged at 1500 × g for 1 min in an Eppendorf 5810 R centrifuge (Eppendorf, Hamburg, Germany) to collect CD63- Nluc exosomes. During this centrifugation step, 100 µl of cold transport buffer containing 2% TX- 100 and protease inhibitor cocktail was added to each well of the 24- well PDL coated plate to lyse the cells and bring the volume up to 200 µl. Filtrate aliquots (50 µl) and 50 µl of the cell lysate were used to measure exosome secretion and normalized to the number of cells per reaction, respectively, using the cell- based CD63- Nluc exosome secretion assay protocol detailed above. For the permeabilized cell exosome secretion assay detailed in Figure  5D, the nucleotide- depleted rat liver cytosol was generated by using a HiTrap Sephadex G- 25 Desalting Column (Cytiva Life Sciences) as per the manufacturer’s protocol. Acknowledgements We dedicate this work to Bob Lesch, our lab manager for the past several decades who was tragi- cally taken from us by an accident in 2021. We thank Criss Hartzell for reading and providing helpful comments on this manuscript. We would also like to thank the staff at the UC Berkeley shared facili- ties, the Cell Culture Facility (Alison Killilea), the Vincent J Coates Proteomics Facility (Lori Kohlstaedt), and the DNA Sequencing Facility. JMN is supported by a National Science Foundation Graduate Research Fellowship. RS is an Investigator of the Howard Hughes Medical Institute, a Senior Fellow of the UC Berkeley Miller Institute of Science, and Scientific Director of Aligning Science Across Parkin- son’s Disease (ASAP). Additional information Competing interests Randy Schekman: Founding Editor- in- Chief, eLife. The other authors declare that no competing inter- ests exist. Funding Funder Howard Hughes Medical Institute Sergey Brin Family Foundation Grant reference number Author Randy Schekman Randy Schekman The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 20 of 24 Cell Biology Research article Author contributions Justin Krish Williams, Jordan Matthew Ngo, Conceptualization, Formal analysis, Investigation, Meth- odology, Writing - original draft, Writing – review and editing; Isabelle Madeline Lehman, Formal analysis, Investigation; Randy Schekman, Conceptualization, Formal analysis, Supervision, Funding acquisition, Writing – review and editing Author ORCIDs Justin Krish Williams Jordan Matthew Ngo Isabelle Madeline Lehman Randy Schekman http://orcid.org/0000-0002-9447-5554 http://orcid.org/0000-0002-6566-3919 http://orcid.org/0009-0008-8667-401X http://orcid.org/0000-0001-8615-6409 Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.86556.sa1 Author response https://doi.org/10.7554/eLife.86556.sa2 Additional files Supplementary files • MDAR checklist Data availability All data generated or analyzed during this study are included in the manuscript and supporting source data files. Source Data files have been provided for Figure 2, Figure 2 - Figure supplement 1, Figure 3, Figure 4, and Figure 4 - Figure supplement 1. Video Files have been provided for Figure 6 - Videos 1- 6. References Alouf JE. 1980. Streptococcal toxins (streptolysin O, streptolysin S, erythrogenic toxin). Pharmacology & Therapeutics 11:661–717. DOI: https://doi.org/10.1016/0163-7258(80)90045-5, PMID: 7003609 Andrews NW, Corrotte M. 2018. Plasma membrane repair. Current Biology 28:R392–R397. DOI: https://doi.org/ 10.1016/j.cub.2017.12.034, PMID: 29689221 Apodaca G, Cardone MH, Whiteheart SW, DasGupta BR, Mostov KE. 1996. Reconstitution of Transcytosis in SLO- Permeabilized MDCK cells: existence of an NSF- dependent fusion mechanism with the apical surface of MDCK cells. The EMBO Journal 15:1471–1481. DOI: https://doi.org/10.1002/j.1460-2075.1996.tb00491.x, PMID: 8612570 Barnes JM, Nauseef JT, Henry MD. 2012. Resistance to fluid shear stress is a conserved biophysical property of malignant cells. PLOS ONE 7:e50973. DOI: https://doi.org/10.1371/journal.pone.0050973, PMID: 23226552 Blazek AD, Paleo BJ, Weisleder N. 2015. Plasma membrane repair: A central process for maintaining cellular homeostasis. Physiology 30:438–448. DOI: https://doi.org/10.1152/physiol.00019.2015, PMID: 26525343 Bouter A, Carmeille R, Gounou C, Bouvet F, Degrelle SA, Evain- Brion D, Brisson AR. 2015. Review: Annexin- A5 and cell membrane repair. Placenta 36:S43–S49. DOI: https://doi.org/10.1016/j.placenta.2015.01.193, PMID: 25701430 Bouvet F, Ros M, Bonedeau E, Croissant C, Frelin L, Saltel F, Moreau V, Bouter A. 2020. Defective membrane repair machinery impairs survival of invasive cancer cells. Scientific Reports 10:21821. DOI: https://doi.org/10. 1038/s41598-020-77902-5, PMID: 33311633 Boye TL, Nylandsted J. 2016. Annexins in plasma membrane repair. Biological Chemistry 397:961–969. DOI: https://doi.org/10.1515/hsz-2016-0171, PMID: 27341560 Buzhynskyy N, Golczak M, Lai- Kee- Him J, Lambert O, Tessier B, Gounou C, Bérat R, Simon A, Granier T, Chevalier JM, Mazères S, Bandorowicz- Pikula J, Pikula S, Brisson AR. 2009. Annexin- A6 presents two modes of association with phospholipid membranes. A combined QCM- D, AFM and Cryo- TEM study. Journal of Structural Biology 168:107–116. DOI: https://doi.org/10.1016/j.jsb.2009.03.007, PMID: 19306927 Chambers R. 1917. Microdissection studies I the visible structure of cell protoplasm and death changes. The American Journal of Physiology 43:1–12. DOI: https://doi.org/10.1152/ajplegacy.1917.43.1.1 Cocucci E, Racchetti G, Meldolesi J. 2009. Shedding microvesicles: Artefacts no more. Trends in Cell Biology 19:43–51. DOI: https://doi.org/10.1016/j.tcb.2008.11.003, PMID: 19144520 Colombo M, Raposo G, Théry C. 2014. Biogenesis, secretion, and Intercellular interactions of exosomes and other extracellular vesicles. Annual Review of Cell and Developmental Biology 30:255–289. DOI: https://doi. org/10.1146/annurev-cellbio-101512-122326, PMID: 25288114 Cooper ST, McNeil PL. 2015. Membrane repair: mechanisms and pathophysiology. Physiological Reviews 95:1205–1240. DOI: https://doi.org/10.1152/physrev.00037.2014, PMID: 26336031 Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 21 of 24 Cell Biology Research article Corrotte M, Castro- Gomes T, Koushik AB, Andrews NW. 2015. Approaches for Plasma Membrane Wounding and Assessment of Lysosome- Mediated Repair Responses, Methods in Cell Biology Elsevier Ltd. DOI: https:// doi.org/10.1016/bs.mcb.2014.11.009 Corrotte M, Castro- Gomes T. 2019. Lysosomes and plasma membrane repair. Current Topics in Membranes 84:1–16. DOI: https://doi.org/10.1016/bs.ctm.2019.08.001, PMID: 31610859 Croissant C, Gounou C, Bouvet F, Tan S, Bouter A. 2020. Annexin- A6 in membrane repair of human skeletal muscle cell: A role in the cap Subdomain. Cells 9:1742. DOI: https://doi.org/10.3390/cells9071742, PMID: 32708200 Croissant C, Carmeille R, Brévart C, Bouter A. 2021. Annexins and membrane repair dysfunctions in muscular Dystrophies. International Journal of Molecular Sciences 22:5276. DOI: https://doi.org/10.3390/ijms22105276, PMID: 34067866 Demonbreun AR, McNally EM. 2016. Plasma membrane repair in health and disease. Current Topics in Membranes 77:67–96. DOI: https://doi.org/10.1016/bs.ctm.2015.10.006, PMID: 26781830 Demonbreun AR, Quattrocelli M, Barefield DY, Allen MV, Swanson KE, McNally EM. 2016. An actin- dependent annexin complex mediates plasma membrane repair in muscle. The Journal of Cell Biology 213:705–718. DOI: https://doi.org/10.1083/jcb.201512022, PMID: 27298325 Driedonks TAP, Nolte-’t Hoen ENM. 2018. Circulating Y- RNAs in extracellular vesicles and ribonucleoprotein complexes; implications for the immune system. Frontiers in Immunology 9:3164. DOI: https://doi.org/10. 3389/fimmu.2018.03164, PMID: 30697216 Funato K, Beron W, Yang CZ, Mukhopadhyay A, Stahl PD. 1997. Reconstitution of phagosome- lysosome fusion in streptolysin O- permeabilized cells. The Journal of Biological Chemistry 272:16147–16151. DOI: https://doi. org/10.1074/jbc.272.26.16147, PMID: 9195911 Gabel M, Delavoie F, Demais V, Royer C, Bailly Y, Vitale N, Bader MF, Chasserot- Golaz S. 2015. Annexin A2- dependent actin bundling promotes secretory granule docking to the plasma membrane and exocytosis. The Journal of Cell Biology 210:785–800. DOI: https://doi.org/10.1083/jcb.201412030, PMID: 26323692 Gerelsaikhan T, Vasa PK, Chander A. 2012. Annexin A7 and Snap23 interactions in alveolar type II cells and in vitro: A role for Ca2+ and PKC. Biochimica et Biophysica Acta 1823:1796–1806. DOI: https://doi.org/10.1016/j. bbamcr.2012.06.010, PMID: 22713544 Gerke V, Moss SE. 2002. Annexins: from structure to function. Physiological Reviews 82:331–371. DOI: https:// doi.org/10.1152/physrev.00030.2001, PMID: 11917092 Gil- Parrado S, Fernández- Montalván A, Assfalg- Machleidt I, Popp O, Bestvater F, Holloschi A, Knoch TA, Auerswald EA, Welsh K, Reed JC, Fritz H, Fuentes- Prior P, Spiess E, Salvesen GS, Machleidt W. 2002. Ionomycin- activated calpain triggers apoptosis. A probable role for Bcl- 2 family members. The Journal of Biological Chemistry 277:27217–27226. DOI: https://doi.org/10.1074/jbc.M202945200, PMID: 12000759 Harding C, Heuser J, Stahl P. 1983. Receptor- mediated endocytosis of transferrin and recycling of the transferrin receptor in rat reticulocytes. The Journal of Cell Biology 97:329–339. DOI: https://doi.org/10.1083/jcb.97.2. 329, PMID: 6309857 Heilbrunn LV. 1930a. The surface precipitation reaction of living cells. Proceedings of the American Philosophical Society 69:295–301. Heilbrunn LV. 1930b. The action of various salts on the first stage of the surface precipitation reaction in Arbacia egg protoplasm. Protoplasma 11:558–573. DOI: https://doi.org/10.1007/BF01614367 Hikita T, Miyata M, Watanabe R, Oneyama C. 2018. Sensitive and rapid quantification of exosomes by fusing luciferase to exosome marker proteins. Scientific Reports 8:14035. DOI: https://doi.org/10.1038/s41598-018- 32535-7, PMID: 30232365 Horn A, Jaiswal JK. 2018. Cellular mechanisms and signals that coordinate plasma membrane repair. Cellular and Molecular Life Sciences 75:3751–3770. DOI: https://doi.org/10.1007/s00018-018-2888-7, PMID: 30051163 Humphries WH, Szymanski CJ, Payne CK. 2011. Endo- lysosomal vesicles positive for Rab7 and Lamp1 are terminal vesicles for the transport of dextran. PLOS ONE 6:e26626. DOI: https://doi.org/10.1371/journal.pone. 0026626, PMID: 22039519 Ingold AL, Cohn SA, Scholey JM. 1988. Inhibition of kinesin- driven microtubule motility by monoclonal antibodies to kinesin heavy chains. The Journal of Cell Biology 107:2657–2667. DOI: https://doi.org/10.1083/ jcb.107.6.2657, PMID: 2974459 Jaiswal JK, Andrews NW, Simon SM. 2002. Membrane proximal lysosomes are the major vesicles responsible for calcium- dependent exocytosis in nonsecretory cells. The Journal of Cell Biology 159:625–635. DOI: https://doi. org/10.1083/jcb.200208154, PMID: 12438417 Jaiswal JK, Chakrabarti S, Andrews NW, Simon SM. 2004. Synaptotagmin VII restricts fusion pore expansion during lysosomal exocytosis. PLOS Biology 2:E233. DOI: https://doi.org/10.1371/journal.pbio.0020233, PMID: 15226824 Jeppesen DK, Fenix AM, Franklin JL, Higginbotham JN, Zhang Q, Zimmerman LJ, Liebler DC, Ping J, Liu Q, Evans R, Fissell WH, Patton JG, Rome LH, Burnette DT, Coffey RJ. 2019. Reassessment of exosome composition. Cell 177:428–445. DOI: https://doi.org/10.1016/j.cell.2019.02.029, PMID: 30951670 Keefe D, Shi L, Feske S, Massol R, Navarro F, Kirchhausen T, Lieberman J. 2005. Perforin triggers a plasma membrane- repair response that facilitates CTL induction of apoptosis. Immunity 23:249–262. DOI: https://doi. org/10.1016/j.immuni.2005.08.001, PMID: 16169498 Keller MD, Ching KL, Liang FX, Dhabaria A, Tam K, Ueberheide BM, Unutmaz D, Torres VJ, Cadwell K. 2020. Decoy exosomes provide protection against bacterial toxins. Nature 579:260–264. DOI: https://doi.org/10. 1038/s41586-020-2066-6, PMID: 32132711 Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 22 of 24 Cell Biology Research article Koerdt SN, Ashraf APK, Gerke V. 2019. Annexins and Plasma Membrane Repair, 1st Ed, Current Topics in Membranes Elsevier Inc. DOI: https://doi.org/10.1016/bs.ctm.2019.07.006 Liao YC, Fernandopulle MS, Wang G, Choi H, Hao L, Drerup CM, Patel R, Qamar S, Nixon- Abell J, Shen Y, Meadows W, Vendruscolo M, Knowles TPJ, Nelson M, Czekalska MA, Musteikyte G, Gachechiladze MA, Stephens CA, Pasolli HA, Forrest LR, et al. 2019. RNA granules hitchhike on lysosomes for long- distance transport, using Annexin A11 as a molecular tether. Cell 179:147–164. DOI: https://doi.org/10.1016/j.cell.2019. 08.050, PMID: 31539493 Malhotra V. 2013. Unconventional protein secretion: an evolving mechanism. The EMBO Journal 32:1660–1664. DOI: https://doi.org/10.1038/emboj.2013.104, PMID: 23665917 Martys JL, Shevell T, McGraw TE. 1995. Studies of transferrin recycling reconstituted in streptolysin O permeabilized Chinese hamster ovary cells. The Journal of Biological Chemistry 270:25976–25984. DOI: https://doi.org/10.1074/jbc.270.43.25976, PMID: 7592788 McNeil PL, Khakee R. 1992. Disruptions of muscle fiber plasma membranes: role in exercise- induced damage. The American Journal of Pathology 140:1097–1109 PMID: 1374591. Meldolesi J. 2018. Exosomes and ectosomes in intercellular communication. Current Biology 28:R435–R444. DOI: https://doi.org/10.1016/j.cub.2018.01.059, PMID: 29689228 Messenger SW, Woo SS, Sun Z, Martin TFJ. 2018. A Ca2+-Stimulated exosome release pathway in cancer cells is regulated by Munc13- 4. The Journal of Cell Biology 217:2877–2890. DOI: https://doi.org/10.1083/jcb. 201710132, PMID: 29930202 Mittelbrunn M, Vicente Manzanares M, Sánchez- Madrid F. 2015. Organizing polarized delivery of exosomes at synapses. Traffic16:327–337. DOI: https://doi.org/10.1111/tra.12258, PMID: 25614958 O’Brien K, Breyne K, Ughetto S, Laurent LC, Breakefield XO. 2020. RNA delivery by extracellular vesicles in mammalian cells and its applications. Nature Reviews. Molecular Cell Biology 21:585–606. DOI: https://doi. org/10.1038/s41580-020-0251-y, PMID: 32457507 Park E, Rapoport TA. 2012. Mechanisms of Sec61Secy- mediated protein translocation across membranes. Annual Review of Biophysics 41:21–40. DOI: https://doi.org/10.1146/annurev-biophys-050511-102312, PMID: 22224601 Potez S, Luginbühl M, Monastyrskaya K, Hostettler A, Draeger A, Babiychuk EB. 2011. Tailored protection against plasmalemmal injury by annexins with different Ca2+ sensitivities. The Journal of Biological Chemistry 286:17982–17991. DOI: https://doi.org/10.1074/jbc.M110.187625, PMID: 21454475 Poulain B, Tauc L, Maisey EA, Wadsworth JDF, Mohan PM, Dolly JO. 1988. Neurotransmitter release is blocked Intracellularly by botulinum neurotoxin, and this requires uptake of both toxin polypeptides by a process mediated by the larger chain. PNAS 85:4090–4094. DOI: https://doi.org/10.1073/pnas.85.11.4090, PMID: 2897693 Ridder K, Keller S, Dams M, Rupp AK, Schlaudraff J, Del Turco D, Starmann J, Macas J, Karpova D, Devraj K, Depboylu C, Landfried B, Arnold B, Plate KH, Höglinger G, Sültmann H, Altevogt P, Momma S. 2014. Extracellular vesicle- mediated transfer of genetic information between the hematopoietic system and the brain in response to inflammation. PLOS Biology 12:e1001874. DOI: https://doi.org/10.1371/journal.pbio.1001874, PMID: 24893313 Savina A, Furlán M, Vidal M, Colombo MI. 2003. Exosome release is regulated by a calcium- dependent mechanism in K562 cells. The Journal of Biological Chemistry 278:20083–20090. DOI: https://doi.org/10.1074/ jbc.M301642200, PMID: 12639953 Schiavo G, Benfenati F, Poulain B, Rossetto O, Polverino de Laureto P, DasGupta BR, Montecucco C. 1992. Tetanus and botulinum- B neurotoxins block neurotransmitter release by proteolytic cleavage of synaptobrevin. Nature 359:832–835. DOI: https://doi.org/10.1038/359832a0, PMID: 1331807 Shan S, Walter P. 2005. Co- Translational protein targeting by the signal recognition particle. FEBS Letters 579:921–926. DOI: https://doi.org/10.1016/j.febslet.2004.11.049, PMID: 15680975 Shurtleff MJ, Yao J, Qin Y, Nottingham RM, Temoche- Diaz MM, Schekman R, Lambowitz AM. 2017. Broad role for Ybx1 in defining the small noncoding RNA composition of exosomes. PNAS 114:E8987–E8995. DOI: https://doi.org/10.1073/pnas.1712108114, PMID: 29073095 Steinhardt RA, Bi G, Alderton JM. 1994. Cell membrane resealing by a vesicular mechanism similar to neurotransmitter release. Science 263:390–393. DOI: https://doi.org/10.1126/science.7904084, PMID: 7904084 Swaggart KA, Demonbreun AR, Vo AH, Swanson KE, Kim EY, Fahrenbach JP, Holley- Cuthrell J, Eskin A, Chen Z, Squire K, Heydemann A, Palmer AA, Nelson SF, McNally EM. 2014. Annexin A6 modifies muscular dystrophy by mediating sarcolemmal repair. PNAS 111:6004–6009. DOI: https://doi.org/10.1073/pnas.1324242111, PMID: 24717843 Tang X, Yang F, Guo Y. 2020. Cell- free reconstitution of the packaging of cargo proteins into vesicles at the trans Golgi network. BIO- PROTOCOL 10:e3537. DOI: https://doi.org/10.21769/BioProtoc.3537, PMID: 33659511 Tricarico C, Clancy J, D’Souza- Schorey C. 2017. Biology and biogenesis of shed microvesicles. Small GTPases 8:220–232. DOI: https://doi.org/10.1080/21541248.2016.1215283, PMID: 27494381 Upton HE, Ferguson L, Temoche- Diaz MM, Liu XM, Pimentel SC, Ingolia NT, Schekman R, Collins K. 2021. Low- bias ncRNA libraries using ordered two- template relay: serial template jumping by a modified retroelement reverse transcriptase. PNAS 118:1–10. DOI: https://doi.org/10.1073/pnas.2107900118, PMID: 34649994 van Niel G, D’Angelo G, Raposo G. 2018. Shedding light on the cell biology of extracellular vesicles. Nature Reviews. Molecular Cell Biology 19:213–228. DOI: https://doi.org/10.1038/nrm.2017.125, PMID: 29339798 Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 23 of 24 Cell Biology Research article Walker JR, Hall MP, Zimprich CA, Robers MB, Duellman SJ, Machleidt T, Rodriguez J, Zhou W. 2017. Highly potent cell- permeable and impermeable NanoLuc luciferase inhibitors. ACS Chemical Biology 12:1028–1037. DOI: https://doi.org/10.1021/acschembio.6b01129, PMID: 28195704 Whitham M. 2018. Extracellular Vesicles provide a means for tissue crosstalk during exercise. Cell Metabolism 27:237–251. DOI: https://doi.org/10.1016/j.cmet.2017.12.001 Yu QC, McNeil PL. 1992. Transient disruptions of aortic endothelial cell plasma membranes. The American Journal of Pathology 141:1349–1360 PMID: 1466399. Zanetti G, Pahuja KB, Studer S, Shim S, Schekman R. 2011. COPII and the regulation of protein sorting in mammals. Nature Cell Biology 14:20–28. DOI: https://doi.org/10.1038/ncb2390, PMID: 22193160 Zhang M, Liu L, Lin X, Wang Y, Li Y, Guo Q, Li S, Sun Y, Tao X, Zhang D, Lv X, Zheng L, Ge L. 2020. A translocation pathway for vesicle- mediated unconventional protein secretion. Cell 181:637–652.. DOI: https:// doi.org/10.1016/j.cell.2020.03.031, PMID: 32272059 Zhang C, Schekman R. 2023. Syncytin- mediated open- ended membrane tubular connections facilitate the intercellular transfer of cargos including Cas9 protein. eLife 12:e84391. DOI: https://doi.org/10.7554/eLife. 84391, PMID: 36896791 Williams, Ngo et al. eLife 2023;12:e86556. DOI: https://doi.org/10.7554/eLife.86556 24 of 24 Cell Biology
10.1038_s41467-022-35532-7
Article https://doi.org/10.1038/s41467-022-35532-7 Supercooling of the A phase of 3He Received: 20 July 2022 Accepted: 7 December 2022 Y. Tian1, D. Lotnyk 1, A. Eyal1,2, K. Zhang3,4, N. Zhelev 1,5, T. S. Abhilash 1,6, A. Chavez1, E. N. Smith1, M. Hindmarsh3,4, J. Saunders7, E. Mueller J. M. Parpia 1 1 & Check for updates ; , : ) ( 0 9 8 7 6 5 4 3 2 1 ; , : ) ( 0 9 8 7 6 5 4 3 2 1 Because of the extreme purity, lack of disorder, and complex order parameter, the first-order superfluid 3He A–B transition is the leading model system for first order transitions in the early universe. Here we report on the path dependence of the supercooling of the A phase over a wide range of pressures below 29.3 bar at nearly zero magnetic field. The A phase can be cooled sig- nificantly below the thermodynamic A–B transition temperature. While the extent of supercooling is highly reproducible, it depends strongly upon the cooling trajectory: The metastability of the A phase is enhanced by transiting through regions where the A phase is more stable. We provide evidence that some of the additional supercooling is due to the elimination of B phase nucleation precursors formed upon passage through the superfluid transition. A greater understanding of the physics is essential before 3He can be exploited to model transitions in the early universe. The condensation of 3He pairs into a superfluid state occurs via a second-order phase transition at a pressure-dependent transition temperature, Tc, shown in Fig. 1. The anisotropic A phase is favored at high temperatures and pressures, while the isotropic B phase is the stable phase below the TAB(P) line1,2. In zero magnetic fields, the equilibrium phase diagram exhibits a polycritical point3 (PCP) at which the line of first-order transitions (TAB) intersects the line of second- order transitions (Tc) at 21.22 bar and 2.273 mK. The transition between the A and B phases is first order and thus subject to hysteresis. At the PCP, the bulk free energies of the A, B superfluid phases and the nor- mal state are equal. The A phase is highly metastable, and can persist down to extre- mely low temperatures for long times (≥1 day) at high pressures, providing surfaces of the container are smooth4,5. Standard homo- geneous nucleation theory6,7 argues that the transition from meta- stable A to stable B is mediated by thermal fluctuations that produce bubbles of characteristic size r. For small bubbles (size less than the critical radius, Rcrit), the interfacial energy cost (∝r2) is larger than the bulk free energy gain (∝r3), but for large bubbles the opposite holds. Thus, if thermal fluctuations create a bubble with r < Rcrit, it rapidly shrinks. Conversely, a bubble with r > Rcrit will grow. This model8, applied to 3He, leads to Rcrit ≈ 1.5 μm, and an activation energy that is many orders of magnitude above the thermal energy9–11 implying an unobservably small nucleation rate. Surface defects potentially alter the energetics (most surfaces favor the A phase12 and there is no clean explanation of how they would mediate the A–B transition). Despite extensive experimental4,13–20 (Fig. 1) and theoretical investigations21–26, the mechanism for B phase nucleation remains a mystery. Laboratory studies of the dynamics of first-order phase transitions have cosmological implications, as the statistical theories of the decay of a metastable state in condensed matter7 are non-relativistic analogs of the quantum field theories used in cosmological models27,28. Importantly, the possibility of a first-order electroweak symmetry- breaking phase transition29,30 in the early universe has been used to explain baryon asymmetry31. The same physics also produces gravita- tional waves32–35 whose detection are science targets for future space- based detectors such as Laser Interferometer Space Antenna (LISA)36,37. Experimental confirmation of the applicability of this model of first- order phase transitions to a laboratory system (whether in 3He or in cold atom systems38,39) would lend more weight to the calculations of gravitational wave production for LISA and other future probes of the early Universe. However, the theory of first-order phase transitions in 1Department of Physics, Cornell University, Ithaca, NY 14853, USA. 2Physics Department, Technion, Haifa, Israel. 3Department of Physics and Astronomy, University of Sussex, Falmer, Brighton BN1 9QH, UK. 4Department of Physics and Helsinki Institute of Physics, University of Helsinki, PL 64, FI-00014 Helsinki, Finland. 5Center for Applied Physics and Superconducting Technologies, Department of Physics and Astronomy, Northwestern University, Evanston, IL 60208, USA. 6VTT Technical Research Centre of Finland Ltd, Espoo, Finland. 7Department of Physics, Royal Holloway University of London, Egham, TW20 0EX Surrey, UK. e-mail: jmp9@cornell.edu Nature Communications | (2023) 14:148 1 Article https://doi.org/10.1038/s41467-022-35532-7 the early Universe27 is based on the same homogeneous nucleation theory which fails to explain the behavior of 3He in contact with “ordinary surfaces"; if there is a much more rapid intrinsic nucleation mechanism in operation, the gravitational wave signal could be ren- dered negligible35. Here we study the nucleation of B phase in a pair of chambers connected by a high aspect-ratio “letterbox" channel (Fig. 2). Both the geometry and the surface qualities are relevant: The smaller chamber Fig. 1 | Previous supercooling results. The equilibrium phase diagram for super- fluid 3He. The normal fluid (blue), stable A phase (yellow) and the B phase (green) are separated by a red line that marks Tc. The equilibrium A–B transition in zero magnetic field (black line) terminates at the polycritical point (PCP) where the A, B and Normal phases coincide. Centered on the PCP is the region investigated in ref. 20 (black rectangle) and inset where left-pointing triangles show supercooling extent (light yellow) under constant pressure, and downward pointing triangles, pressure decreased conditions (hatched yellow-green region), all in ≤0.1 mT. The region investigated in this paper is shown as a red box. Results from previous investigations in a variety of magnetic fields are shown as gray diamonds: 4.9 mT, 0.5 mT13; blue crosses: 56.9 mT, blue triangles: 28.4 mT14; orange triangle: 0 mT57; red circles: 0 mT, red diamonds: 10.0 mT, red square: 20.0 mT58; black squares, black diamonds 28.2 mT5. (denoted the Isolated Chamber (IC)) incorporates “ordinary" as- machined coin silver surfaces. It also houses a quartz fork whose resonant properties (frequency and quality factor, Q) allow us to infer the phase of the 3He in the chamber. The IC is separated from a larger chamber containing sintered silver by a micromachined channel con- struction consisting of a 1.1 μm tall × 3 mm wide × 100 μm long channel and 200 μm tall × 3 mm wide × 2.5 mm long lead-in channels on either side. This construction was nanofabricated in silicon and capped with glass40 (see also Supplementary Note 1). The silver sinter-containing chamber (denoted the Heat Exchange Chamber (HEC)) incorporates a quartz fork similar to that contained in the IC. The A phase is stabilized in confined spaces, and the narrow channel potentially prevents the propagation of an A–B phase boundary from one chamber to the other —allowing the transitions to be independent. In an earlier publication20, we reported initial observations of the reproducibility of B phase nucleation and an unexpected path dependence for the A phase’s stability. From those experiments, it was unclear whether the path dependence was limited to the region near the PCP and there were few clues about the microscopic origin of the phenomenon. Here, we have expanded the region investigated to include the highest pressure readily accessible to us (the pressure of the minimum of the 3He melting curve, 29.3 bar) and have designed a series of protocols that provide significantly more clarity about the phenomenon. As already emphasized, homogeneous nucleation theory is unable to explain the nucleation of the B phase from the A phase: There is a vanishing probability that thermal fluctuations produce a B phase bubble that is larger than the critical radius. The transition can be triggered4,17,41 by bringing a radioactive source near 3He—which is consistent with models where energetic particles (either deliberately introduced or due to Cosmic rays) are responsible for the observed A–B phase transition11,17,41. Those models cannot explain why when 3He is repeatedly cooled14,20, the transition consistently occurs along the same temperature and pressure line (dubbed the catastrophe line14). Alternative explanations have been sought. The theory of quantum tunneling of metastable states in field theory42 has been applied to the 3He system10, without substantially changing the mismatch in rates between theory and experiment. More complex field-theory-based models such as Q balls25 or Resonant Tunneling (RT)26 have been proposed; they are, however, not consistent with the path dependence Fig. 2 | Schematic of cell. The isolated chamber (IC) and heat exchange chamber (HEC) both contain quartz forks whose quality factor is monitored to determine the phase of the 3He superfluid. The chambers are separated by a 100 μm long channel with aperture 1.1 μm tall and 3 mm wide (the 3 mm width is hidden in the main view). As seen in the enlarged view (circled in red), this “letterbox" channel has two 200 μm tall 3 mm wide 2.5 mm long channels on either side, one opening to the IC, the second connecting to a 2-mm diameter cylindrical tube opening into the HEC. Nature Communications | (2023) 14:148 2 Article https://doi.org/10.1038/s41467-022-35532-7 seen in our experiments. The presence of topological defects such as vortices can enhance the nucleation rate21. Such vortices can be detected via calorimetry17, but our temperature measurements are insufficiently precise. Nonetheless, in our experiment, we do not expect to have a substantial number of vortices: Vortices are either shed during fluid flow or produced during rapid cooling43,44. Our flow and cooling rates are very low. In our previous experiment20, we found that the degree of supercooling was independent of the rate at which we passed through Tc—indicating that Kibble–Zurek vortices are not relevant. Moreover, vortex-induced enhancement of the nucleation rate is expected to be too small to explain our observations21. In an attempt to explain our observations, we note that the silver sinter contains a large number of chambers that are connected to the bulk fluid by narrow channels or constrictions. We hypothesize that, upon traversing Tc, the A phase is formed in bulk, but regions of dis- torted order parameters are formed in some of these chambers. They act as precursor “seeds" of the B phase. For computational expediency, we treat these chambers as if they are filled with B phase, and refer to them as “B phase seeds". However, confinement would result in a distorted order parameter quite different from the bulk B phase. Surface tension stabilizes the requisite A–B domain walls at sufficiently small constrictions. The size of the largest stable domain wall depends on pressure and temperature: In the A region of the phase diagram, the A phase will rush into any of the chambers whose opening is larger than this size. Conversely, the (path-dependent) catastrophe line will be determined by the size of the smallest constriction that connects to a B-filled chamber. This model is similar to the lobster-pot scenario which was proposed for understanding the nucleation of the A phase from B23. Cavities in the sinter are unable to explain all of our observations, and it is likely that some other mechanism is also at play. For example, the A phase order parameter (in standard experimental geometries) may contain complicated textures with highly frustrated points that may act as seeds for the B phase. Such seeds may involve B-inclusions, or just precursor regions where the A order parameter is strongly suppressed. While some of this structure forms spontaneously due to the Kibble–Zurek mechanism43,44, much of the spatial complexity is likely due to surface effects: surfaces constrain the components of the order parameter45 and surface corrugations or scratches can lead to complicated disgyrations and other structures46, perhaps containing precursor seeds of B phase. Similar to the cavity scenario, the observed A–B transition is set by the size of the “largest" seed, whose cata- strophe temperature is highest. These largest seeds are also the most fragile and may be eliminated by exposure to high pressures where the A phase is most stable. The key feature of the path-dependence in both scenarios is which seeds survive. Development of an understanding of this pressure dependence is essential if 3He is to be a useful model for phase transitions in the early Universe. We emphasize that the order parameter of helium is contained in a high dimensional space, and the paths connecting the A and B phases are strongly influenced by surfaces, textures, and distortions from confinement. Nucleation can occur through both thermal fluctuations and quantum tunneling, the latter of which can display interference effects that are particularly sensitive to such changes in the energy landscape42. Models of nucleation in inhomogeneous settings contain a multitude of complexities47. Results Experimental details The normal-superfluid and A–B transitions were detected using quartz forks located in the IC and HEC. The temperatures were obtained with reference to a 3He melting curve thermometer48 mounted on the cold plate of the nuclear demagnetization stage. For details of the operation of the forks and of the thermometry, we refer to the Methods section. Supercooling at constant pressure The first set of measurements was carried out while cooling at a con- stant rate (≤10 μK/h) and fixed pressure. Figure 3a shows the tem- peratures at which the A–B phase transition was detected in the HEC (pink triangles) and the IC (blue triangles). Below 23.8 bar, the HEC transition occurs at a substantially higher temperature than in the IC. In this regime, we believe that the A phase is stable in the channel: It acts as a plug, preventing the A–B wave-front from propagating from the HEC to the IC. The silver sinter in the HEC leads to more complicated variations of the order parameter, and it is reasonable that the HEC and IC contain different B-phase seeds with different catastrophe tem- peratures. Between 23.8 bar and 26 bar, there is a decrease in the separation between the two transitions, which suggests that the A–B wave-front is only weakly pinned by the channel. Above 26 bar, the two transitions happen simultaneously, and we conclude that in this regime, the channel is unable to sustain an A–B interface once the transition is initiated in the HEC (see the Discussion section and Sup- plementary Notes 2 and 3). Supercooling after decreasing pressure In our model, the largest degree of supercooling should occur for trajectories passing through the regions where the A phase is most stable (i.e., at high pressure). To explore this feature, we first cool at our highest accessible pressure (29.3 bar) followed by depressurizing and further cooling. In Fig. 3b we illustrate several such trajectories. The solid black lines show cooling trajectories where we maintained an approximately constant Q of the fork in the IC. This constant-Q con- dition yields a path that is roughly parallel to Tc. For these trajectories, we depressurized by 4-6 bar during the first day, followed by pro- ceeding at 1.3 bar per day, or less. During the rapid part of the ramp, but not during the slow part, there was some viscous heating observed in the IC. We found that the extent of supercooling was significantly greater than what we achieved while cooling at constant pressure (denoted by pink (HEC) and blue (IC) lines instead of data points in Fig. 3b). With the exception of the four lowest constant Q runs (closest to Tc), the A–B transitions occurred simultaneously in both chambers, and are depicted in Fig. 3b as coincident crosses and squares. The same sym- bols (crosses for IC and squares for HEC) are used to denote the observed T, P coordinates of the pressure-varied transitions for the four lowest points. A temperature correction is applied to the IC data to account for thermal offsets between the chambers. To further explore the path dependence, we considered the tra- jectories shown as dashed lines in Fig. 3b. These begin with constant- pressure cooling at 29.3 bar, followed by fixed temperature depres- surizations and fixed pressure cooling. In all cases, we observe sig- nificantly more supercooling than in Fig. 3a. Crucially, there appears to be a definitive locus of points in the T–P plane on which all of the trajectories fall. As illustrated by the dotted black and cyan paths (terminating at 25 and 27 bar), one finds the same A–B transition points when cooling after depressurizing or depressurizing at a constant temperature – as long as the trajectory passed through the A phase at 29.3 bar. In Supplementary Note 4 we present a detailed comparison between the 23 bar fixed pressure run, and one of the trajectories passing through 29.3 bar before cooling at 23 bar. We find that the only detectable difference is the temperature of the A–B transition. There are no signs of thermal gradients, viscous heating, or other systematic effects. Other supercooling results Figure 4 illustrates four additional runs, each of which involves cooling at 23 bar. The blue curve shows the quality factor of the quartz oscil- lator in the HEC during cooling. It jumps discontinuously at T = 2.12 mK (T − Tc = −0.113 mK), indicating the A–B phase transition. For the other Nature Communications | (2023) 14:148 3 Article https://doi.org/10.1038/s41467-022-35532-7 Fig. 3 | Constant-pressure and pressure-varied A–B transitions. a Constant pressure-cooled A–B transitions are shown with pink left-pointing symbols denoting transitions in the HEC, blue open left-pointing triangles denoting transi- tions in the IC. Above 25.8 bar, the transitions are coincident in time and are shown as nested triangles. Color coding follows that in Fig. 1. b A–B transitions observed while decreasing the pressure starting from 29.3 bar are shown along with their paths. Pink squares show transitions in the HEC, blue crosses show transitions in the IC. Where these transitions occur simultaneously, they are superposed. At low pressure, they separate with the A–B transition in the IC observed at a lower P, T than that in the HEC (see inset). Constant pressure-cooled transitions from panel (a) are shown as solid lines. Cyan, black and red lines each show two different paths terminating at the same (P, T) coordinates. The brown line shows the result of pressurization followed by further cooling at constant pressure. Hatched region marks enhanced path dependent supercooling. between the A and B phases are largest compared to the thermo- dynamic barriers. Note, the changes caused by these excursions are subtle enough that they do not appreciably change the quality factors (aside from shifting the A–B transition). The brown path in Fig. 3b illustrates the reverse effect. We tra- verse the stable A phase at 23 bar. We then increase the pressure to 26 bar before continuing to cool. We find that the A–B transition occurs at a higher temperature than if we simply cool at 26 bar. This path avoids the regions of the phase diagram where the A phase is most stable. To ensure that the supercooling is not significantly affected by the sweep rate, we repeated the experiment in Fig. 4, varying the rate of pressurizing and depressurizing during the jog from 23 bar to 27 bar and back. We varied this rate from 1.3 to 27.5 bar/day, finding no dif- ference in the degree of supercooling after completing the cooling at 10 μK/h. To verify the stability of the A phase obtained after depressur- ization, we selected a trajectory that terminated below the PCP from Fig. 3b. After cooling through Tc, starting from 29.3 bar and 2.15 mK, we depressurized (at fixed Q) to 20.5 bar and stopped at a point within 3 μK of the temperature where we previously observed the transition. We waited at this T, P for 1 day. We then slowly cooled at a rate of 0.5 μK/h until we observed the transition in the HEC approximately 2 μK below the previously observed result (open pink circle in Fig. 3b). Thus the supercooled A phase has a lifetime in excess of 24 h, and any dynamics which happen on this timescale do not appear to sig- nificantly influence the catastrophe line. Furthermore, the A–B transi- tion in the IC (×) occurred at a lower temperature than the transition in the HEC, consistent with the other depressurization runs terminating in this part of the phase diagram (see Fig. 3b and its inset). Discussion We analyze our data by considering the model from our introduction, where the B phase grows from seeds that are contained in small chambers with a distribution of narrow necks. While cooling through the A phase, the A phase intrudes on the chambers with the largest openings: the size of the remaining channels connecting to B seeds location of the observed A–B determines the path-dependent Fig. 4 | Comparison of constant-pressure and pressure-cycled runs—QHEC vsT cooled through Tc and TAB at 23 bar. The constant pressure-cooled experiment in the HEC is shown as a solid blue line. For each of the three pressure-cycled runs, after cooling through Tc, while the temperature was maintained at ≈2.2 mK, the 3He was pressurized to 24.5 bar (purple), 26 bar (green) and 27.5 bar (pink), then depressurized to 23 bar, and then cooled further at constant pressure till the A–B transition was observed. The HEC and IC transitions were simultaneous for the 26 and 27.5 bar runs. Arrows mark the positions of the various A–B transitions. The inset shows the A–B transitions and the paths in the P, T diagram. The hatched region in the inset is the same as in Fig. 3b. three runs, the helium is cooled to 2.2 mK, and then slowly pressurized to p max = 24.5 bar, 25 bar, or 27.5 bar. The pressure is then reduced back to 23 bar, and the temperature is reduced further. As expected from our model, the degree of supercooling is a monotonic function of p max: the B phase seeds are suppressed by excursions deep into the equilibrium A phase. In these regions the free energy differences Nature Communications | (2023) 14:148 4 Article https://doi.org/10.1038/s41467-022-35532-7 Fig. 5 | Curvature of stable domain walls. Contours show κA= sinðθAÞ and κB in the A and B portions of the phase diagram. The smallest curvature contour (between the two shades of green) corresponds to 0.078 μm−1, while each subsequent con- tour is a factor of 2 larger. κ is the mean curvature of a stable A–B domain wall, and θ is the contact angle of a domain wall with a surface. Here we assume minimal pairbreaking (specular scattering) boundary conditions. In the A phase, κA= sinðθAÞ quantifies the inverse size of orifice which can block the motion of an A–B domain wall, while κB represents the same quantity for the B phase. Under the assumption that the B phase is seeded from chambers with small openings, the largest A phase value of κA= sinðθAÞ will set the κB where the A → B transition is observed. Red line shows a contour, κ = 0.25 μm−1, which roughly corresponds to where domain walls pass freely through the 1.1 μm channel between the HEC and IC, corresponding to a contact angle of 74∘. To the left of this line, transitions in the two chambers always occur simultaneously. transition. Similar logic should apply to the cases where the B seeds are at the nodes of frustrated textures or distortions. In order to balance forces, an equilibrium domain wall between the A and B phases must be bowed with a mean curvature κ = ∣δf∣/(2σ), where σ is the surface tension and δf is the difference in free energy densities between the two phases. As detailed in Supplementary Note 3, a circular hole in a flat plate with diameter W will prevent the intrusion of the A phase if W < sinðθAÞ=κA, where the contact angle θA is determined by surface energies. Conversely, in the B region of the phase diagram, the same orifice will prevent the B phase from exiting if W < 1/κB—the contact angle does not appear in this expression because the B phase typically does not wet a surface. Note: these equations are sensitive to our modeling of the geometry of the orifice, and the contact angle depends on the surface properties. In Fig. 5 we show contours of constant κA= sinðθAÞ and constant κB, calculated using a Landau–Ginzburg theory and assuming minimal pairbreaking boundary conditions, corresponding to smooth surfaces. (See Supplementary Notes 2 and 3. Supplementary Note 5 deals with the results obtained for maximally pairbreaking boundary conditions. There is some ambiguity in the temperature dependence of the Landau–Ginzburg parameters, and Supplementary Note 6 discusses an alternative model.) As can be seen, κ vanishes at the equilibrium A–B transition, where the two phases have the same free energy. It also vanishes at Tc. The dark green regions show where it is small. The largest values of κA= sinðθAÞ are found at high pressure, and the largest values of κB are found at low temperature. Our model would predict that for a given cooling trajectory, κB at the A–B catastrophe line will coincide with the largest value of κA= sinðθAÞ encountered while cool- ing: i.e., Γ = κB sinðθAÞ=κA = 1. For example, a constant pressure-cooled trajectory at 23.25 bar will almost touch the contour between the light and dark green regions in the A phase. The A–B transition is therefore expected to be at the same contour in the B phase. All of the varied- pressure trajectories pass through the A phase close to the contour Fig. 6 | Pressure dependence of Γ = κB sinðθAÞ=κA. Ordinate, Γ is the ratio of the domain wall curvature at catastrophe point, κB, to the largest scaled curvature traversed in A phase, κA= sinðθAÞ. The contact angle θA depends on the boundary condition: here we use minimal pairbreaking. The abscissa shows the pressure at the observed A–B transition. The collapse of the data around unity above p = 24 bar suggests that trapped pockets of B-phase seeds are responsible for the observed A–B transition at high pressure. that separates the two lightest shades of yellow—and one therefore expects the catastrophe line to follow the corresponding B contour. To better quantify the data, in Fig. 6 we plot the ratio Γ = κB sinðθAÞ=κA vs. the pressure at which the A–B transition was observed. There is a remarkable data collapse for all pressures above 24 bar, despite the fact that the trajectories (pressure-varied or con- stant-pressure) are very different. The ratio is larger than the expected value of 1 (likely the result of the model’s assumptions) and is essen- tially constant. Variations in the geometry or boundary conditions could cause this ratio to be different from unity—for example, the contact angle could be slightly less than what is predicted by the theory. Non-circular interfaces or tapered channels could skew the ratio. Analyzing the data under the assumption of maximal pair- breaking conditions further increases the ratio (see Supplementary Figs. 6 and 7). The physically relevant boundary condition lies between minimal and maximal pairbreaking conditions49. Previous experiments have directly tested aspects of our model of the AB phase boundary50, including measuring equilibrium contact angles, surface tensions, and surface energies at low pressure. A number of theoretical works have also addressed the issue9,51,52. Below 24 bar the fixed pressure HEC and IC A–B transition data separate. Below 20.5 bar, a similar separation occurs in the pressure- varied runs. These features naturally correspond to when the channel connecting the HEC and IC can no longer support a domain wall. This feature is apparent in Fig. 5, where we draw a red line that corresponds to the contour with κ = 0.25 μm−1. To the right of this line the transitions in the IC and HEC occur independently, while to the left they occur simultaneously. The IC data points which follow this red line correspond to events where the pre-existing B phase in the HEC propagates through the channel, and do not represent independent nucleation events. This includes the points below the PCP accessed by depressurization and then cooling at constant pressure. As argued in Supplementary Note 3, the B phase can propagate into the channel when κ = cosðθÞ=W , where W = 1.1 μm is the height of the channel. Our inferred contact angle (θ ~ 75°) is larger than typical values predicted by the Landau–Ginzburg theory with minimal (θ ~ 30°) or maximal (θ ~ 60°) pairbreaking boundary conditions (see Supplementary Fig. 4). This may be a feature of the glass and silicon surfaces in the channel, or it may point toward limitations in the accuracy of our theoretical model. Nature Communications | (2023) 14:148 5 Article https://doi.org/10.1038/s41467-022-35532-7 Between 22.5 and 23.8 bar, the constant pressure-cooled transi- tions in the HEC continue to agree with our model, with Γ ~ 1 (pink triangles). Over the same range the IC data clusters near Γ ~ 2.5 (blue triangles), and it is likely that the transition is completely independent of the HEC. This clustering suggests that a similar model may apply there, but with different surface geometries and boundary conditions, or the involvement of different order parameter structures. The HEC contains sintered silver, while the IC incorporates as-machined coin silver surfaces, with no obvious cavities (and channels) which could be playing the role of the B-containing seeds in the sinter in the HEC. An additional potential mechanism for heterogeneous nucleation involves the presence of surface defects, or features, which favor a distorted order parameter. The simplest model would treat this as a B-phase seed, pinned at the surface with an associated A–B interface. The model of the catastrophe line would be analogous to the one we proposed for the sinter53. Given the multicomponent nature of the superfluid 3He order parameter (a complex 3 × 3 matrix), the nature of the spatially dependent order parameter of this seed region is com- plex. The path dependence could reflect evolution of the order para- meter structures that alter the energetics of the transformation from the A phase to the B phase, without the benefit of an actual interface that would be present if a “seed” of B phase were present. The curvature κA vanishes as one approaches the polycritical point from above, and hence Γ diverges near there for all of the constant-pressure data. At these pressures, the distribution of B seeds is likely determined by kinetic processes occurring during the normal- superfluid transition rather than details of the cooling trajectory. As emphasized in our previous work20, it is surprising that we form the A phase when cooling at pressures below the PCP even in magnetic fields below 0.1 mT. Perhaps, since superfluidity in bulk must be induced by the colder liquid in the sinter (where the order parameter is likely distorted by surfaces), the energy cost of an interface between B in bulk and a surface-induced A phase in the sinter is too great, and the A phase is nucleated in bulk. The same scenario could follow in the IC with the channel playing the role of the sinter. Below 24 bar, Γ falls for the pressure-varied data. This suggests that a separate mechanism is at play: The A–B transition occurs at a higher temperature than predicted by our model. Below 20.5 bar, the transition in the IC and HEC is separate. The ratio Γ for the HEC data continues to fall, further indicating a mechanism in the HEC which goes beyond our model. Between 20.5 and 19 bar, the transition in the IC is likely not an independent nucleation event, but rather due to the A–B domain wall breaking through the channel (corresponding to the red line in Fig. 5). This appears as a plateau Fig. 6. The cluster of IC transitions at 19 bar are likely independent nucleation events. Figure 6 contains two additional outliers. The green discs and black crosses show Γ for the pressure-cycled transitions depicted in Fig. 4. The trajectories that cycled to 26 bar and 27.5 bar agree very well with the rest of the data. The trajectory that cycled to 24.5 bar, how- ever, shows more supercooling than expected, and a surprisingly large value of Γ. While we do not understand why the HEC shows such a large degree of supercooling, the IC transition coincides with the red line in Fig. 5, and is likely due to the physics of the superfluid in the channel connecting the chambers. Similarly, the brown diamond corresponds to the trajectory in Fig. 3b which was cooled at p = 23 bar to 2.2 mK, pressurized to 25 bar, and then further cooled. It also lies on the red line in Fig. 5 and is presumed to correspond to the B phase being conveyed through the channel. The transition in the HEC for this tra- jectory (brown disc) agrees well with our model. Finally, we note that the domain wall between the A and B phases has a finite width, extending over a few temperature-dependent coherence lengths (see Eq. 5 in Supplementary Note 2 and Supple- mentary Fig. 2). Near Tc the temperature-dependent coherence length diverges, ξðTÞ ≈ ξ GLð1 (cid:2) T=T cÞ(cid:2)1=2. The resulting “thick" domain walls are likely to have different elastic properties and may not have a well- defined curvature. This feature may account for some of the decreases in the ratio Γ plotted in Fig. 6 for the pressure-varied runs that exten- ded to transitions near Tc. While we have developed a coherent picture, we emphasize that a number of mysteries still remain. First, we do not have a rigorous explanation for the appearance of the A phase upon cooling through Tc for a range of pressures below the tricritical point. This A region was not seen upon warming, and hence does not represent a stable phase. Superficially similar results were observed in parallel ringing experi- ments by Kleinberg et al. at 0.5 mT13. The primary difference is that in our experiment the extent of supercooling of the A phase in the IC cuts off very sharply below 20.9 bar20, while ref. 13 observed a much smoother termination. We believe that this difference implies that there is a distinct origin to the phenomenon. As illustrated by ref. 3, the stability of the A phase is very sensitive to magnetic fields, and the 0.5 mT field in ref. 13 was potentially responsible for their observa- tions. Our field is smaller. Second, we do not have a model for the nucleation of B seeds (or their exact nature) during the transition from the normal phase into the superfluid. Third, we have yet to establish the exact form of order parameter features that generate those seeds. This is particularly true in the IC, which lacks any natural cavities. In conclusion, we find that the supercooling of the A phase can be extended considerably by transiting through high-pressure regions where the A phase is more stable (measured by the ratio of the A–B free energy difference per coherence length to the surface tension). The path dependence observed here is remarkable, and is only possible because of the purity of 3He and the relatively large energy barriers between the superfluid phases. Importantly, the path dependence that we observe is not confined to the region of the PCP. Furthermore, we provide a quantitative model for much of the observed supercooling which can be ascribed to seeds of the B phase associated with struc- tures in the sinter and possibly with surface defects. This investigation has led to an improved understanding of heterogenous nucleation, but a quantitative explanation awaits more comprehensive modeling of the “seeds" and their connecting channels to the bulk. We note that the supercooled liquid is stable at pressures as low as 18.6 bar, which can be contrasted to the lowest stable pressure for bulk 3He, 21.23 bar. We found that the lifetime of the metastable fluid exceeded one day at 19.8 bar. For a significant part of the phase diagram the degree of super- cooling appears to be determined by the maximum value of κA= sinðθAÞ encountered – a quantity that corresponds to the inverse size of an aperture that can support an A–B domain wall. Despite these insights, aspects of the A–B transition remain enigmatic. In the superfluid 3He environment of this experiment, we have made a study of the systematics of heterogeneous nucleation by exploring a variety of trajectories in the pressure-temperature plane. We have shown that the surface energy of the A–B interface (strongly dependent on p and T), and the contact angle with surfaces play a central role in this nucleation process. On the other hand, in previous work on superfluid 3He confined in a nanofluidic cavity54 negligible supercooling was observed. The transformation from the A to the B phase involves a transit through a multi-dimensional landscape that could be hysteretic with pressure. To develop the A–B phase transition as a model for the first-order transitions in the early universe, identi- fication of all mechanisms is essential. As we observe in our analog system, it is possible that the early universe was not homogeneous, but may have contained structures such as topological defects or pri- mordial black holes, which could play a role in the nucleation of first- order phase transitions55,56. Further studies will include those of superfluid 3He confined in nano-structured environments, in which nucleation is studied in precisely engineered volumes, coupled to bulk liquid through “valves” which effectively isolate that volume from nucleation events in the bulk liquid and heat exchanger, and in which NMR or sound is used as a non-invasive probe. Similar structures might also be used to seed the non-equilibrium Polar phase and other phases Nature Communications | (2023) 14:148 6 Article https://doi.org/10.1038/s41467-022-35532-7 that are not by themselves stable or naturally occurring in bulk. In turn, analogs of these structures might provide insights into the under- pinnings of transitions in the early Universe. At the very least, the elimination of the possibility of a new rapid intrinsic nucleation mechanism will put the understanding of the generation of gravita- tional waves on a firmer footing, and allow LISA observations to be used to constrain—or discover—new physics at the electroweak scale. Methods Fork operation The two quartz forks were each driven at constant voltage (small enough so that no drive-dependent heating was observed) from a signal gen- erator. A current preamplifier was used as the first stage of amplification before the received signal was sent to a lock-in amplifier. The lock-in’s reference frequency was ported from the signal generator. By measuring and fitting the (complex) frequency-dependent non-resonant signal in the circuit, the (anti-symmetric) quadrature component of the received signal (after background subtraction) was used to infer the difference between the drive frequency and the resonant frequency, while the in- phase component was used to infer the “Q" or Quality factor of the fork. The forks were maintained within 10 Hz of the resonant frequency (near 32 kHz) with Q factors varying from ≈40 at Tc, to about 200 at the lowest temperatures at high pressure. In operation, the forks could track the Q well without attention. At the A–B transition, Fig. 4, the Q increased by ≈10 abruptly, providing a clear signature of the transition. Thermometry The temperature of the HEC detected at Tc was found to lag the tem- perature of the melting curve thermometer (mounted on the demag- netization stage) by only 1–2 μK providing the warming and cooling rates were less than 10 μK/h. We estimate the accuracy of our inferred A–B transition temperatures to be ±3 μK, as long as the cooling rate was held constant in a given A–B transition run. The cooling rate of the nuclear stage was controlled by adjusting the rate of decrease of the current in the magnet and could be reliably set to be a constant 10 μK/h or even held constant (±3 μK) for periods as long as a day. The tem- perature of the fork in the IC lagged that of the HEC by ≈15 μK (inferred by observing the differences in the observed Tc while cooling at con- stant temperature). In all graphs, the data have been adjusted for this lag. The inferred temperature of 3He in the IC were similarly adjusted appropriately while cooling at constant pressure if the supercooled transitions in the IC and HEC occurred at different times (i.e., below 24.5 bar see Fig. 3a). Pressure control The pressure was regulated using a temperature-controlled “bomb" consisting of a ≈10 cm3 volume in the form of a 9.5-mm diameter stainless steel tube. An insulated Nichrome wire was wound on this tube and was connected to a 0–25 W power source whose output was set by a digital proportional integral and differential controller. The bomb was semi-isolated from the lab environment by being contained in a large cylindrical tube 5 cm id × 25 cm long, open at both ends and mounted vertically. The pressure could be monitored by a digital Heise DXD 0-40 bar pressure gauge, allowing for high resolution with minimum volume in the system. A 0.3 cm3 volume filled with silver sinter was used as an additional heat sink to thermalize the 3He before it entered into the main HEC chamber. In this way, we were able to vary the pressure by as much as 5 bar/day without incurring significant heating, allowing large pres- sure changes to be effected quite rapidly. A–B transitions were observed while the pressure was constant or varied by 0.7–1.3 bar/day in order to avoid any issues with viscous heating in the channel. Due to the limited heat, we could apply to the bomb, pressure changes greater than 5 bar required multiple steps, where we would stop and reset the pressure in the external gas handling system. Data availability The P, T data generated in this study have been deposited in the Cor- nell University e-commons data repository database under accession code https://doi.org/10.7298/1sw8-f758. References 1. Greywall, D. 3He specific heat and thermometry at millikelvin tem- peratures. Phys. Rev. B 33, 7520–7538 (1986). 2. Wiman, J. J. & Sauls, J. A. Strong-coupling and the stripe phase of 3. 3-He. J. Low Temp. Phys. 184, 1054–1070 (2016). Paulson, D. N., Kojima, H. & Wheatley, J. C. Profound effect of a magnetic field on the phase diagram of superfluid 3He. Phys. Rev. Lett. 32, 1098–1101 (1974). 4. Schiffer, P., O’Keefe, M. T., Hildreth, M. D., Fukuyama, H. & Osheroff, D. D. Strong supercooling and stimulation of the A–B transition in superfluid 3He. Phys. Rev. Lett. 69, 120–123 (1992). Schiffer, P. & Osheroff, D. D. Nucleation of the AB transition in superfluid 3He: surface effects and baked alaska. Rev. Mod. Phys. 67, 491–501 (1995). 5. 6. Cahn, J. W. & Hilliard, J. E. Free energy of a nonuniform system. I. 7. Interfacial free energy. J. Chem. Phys. 28, 258–267 (1958). Langer, J. S. Statistical theory of the decay of metastable states. Annals Phys. 54, 258–275 (1969). 8. Osheroff, D. D. & Cross, M. C. Interfacial surface energy between the superfluid phases of 3He. Phys. Rev. Lett. 38, 905–909 (1977). Kaul, R. & Kleinert, H. Surface energy and textural boundary con- ditions between A and B phases of 3He. J. Low Temp. Phys. 38, 539–552 (1980). 9. 10. Bailin, D. & Love, A. Instantons in superfluid He-3. J. Phys. A 13, 11. L271 (1980). Leggett, A. J. Nucleation of 3He-B from the A phase: a cosmic-ray effect? Phys. Rev. Lett. 53, 1096–1099 (1984). 12. Ambegaokar, V., deGennes, P. G. & Rainer, D. Landau-Ginsburg equations for an anisotropic superfluid. Phys. Rev. A 9, 2676–2685 (1974). 13. Kleinberg, R. L., Paulson, D. N., Webb, R. A. & Wheatley, J. C. Super- cooling and superheating of the AB transition in superfluid 3He near the polycritical point. J. Low Temp. Phys. 17, 521–528 (1974). 14. Hakonen, P. J., Krusius, M., Salomaa, M. M. & Simola, J. T. Comment on “Nucleation of 3He-B from the A phase: a cosmic-ray effect?". Phys. Rev. Lett. 54, 245–245 (1985). 15. Buchanan, D. S., Swift, G. W. & Wheatley, J. C. Velocity of propa- gation of the 3HeA–B interface in hypercooled 3He–A. Phys. Rev. Lett. 57, 341–344 (1986). 16. Boyd, S. T. P. & Swift, G. W. New mode of growth of 3He-B in hypercooled 3He-A: evidence of a spin supercurrent. Phys. Rev. Lett. 64, 894–897 (1990). 17. Bäuerle, C., Bunkov, Yu. M., Fisher, S. N., Godfrin, H. & Pickett, G. R. Laboratory simulation of cosmic string formation in the early universe using superfluid 3He. Nature 382, 332–334 (1996). 18. Bunkov, Yu. M. & Timofeevskaya, O. D. "Cosmological” scenario for A–B phase transition in superfluid 3He. Phys. Rev. Lett. 80, 4927–4930 (1998). 19. Bartkowiak, M. et al. Primary and secondary nucleation of the transition between the A and B phases of superfluid 3He. Phys. Rev. Lett. 85, 4321–4324 (2000). 20. Lotnyk, D. et al. Path-dependent supercooling of the 3He superfluid A-B transition. Phys. Rev. Lett. 126, 215301 (2021). 21. Leggett, A. J. Leggett responds. Phys. Rev. Lett. 54, 246 (1985). 22. Yip, S. & Leggett, A. J. Dynamics of the 3He A-B phase boundary. Phys. Rev. Lett. 57, 345–348 (1986). 23. Leggett, A. J. & Yip, S. K. Nucleation and growth of 3He-B in the supercooled A-phase. In Helium Three, no. 26 in Modern Problems in Condensed Matter Sciences 3rd edn (eds Pitaevskii, L. P. & Hal- perin, W. P.) Ch. 8, 523–707 (Elsevier, Amsterdam, 1990). Nature Communications | (2023) 14:148 7 Article https://doi.org/10.1038/s41467-022-35532-7 24. Leggett, A. J. High-energy low-temperature physics: production of phase transitions and topological defects by energetic particles in superfluid 3He. J. Low Temp. Phys. 126, 775–804 (2002). 25. Ki Hong, D. Q-balls in superfluid 3He. J. Low Temp. Phys. 71, 51. Schopohl, N. Spatial dependence of the order parameter of superfluid 3He at the A-B phase boundary. Phys. Rev. Lett. 58, 1664–1667 (1987). 52. Thuneberg, E. V. A-B interface of superfluid 3He in a magnetic field. 483–494 (1988). Phys. Rev. B 44, 9685–9691 (1991). 26. Tye, S.-H. H. & Wohns, D. Resonant tunneling in superfluid 3He. Phys. Rev. B 84, 184518 (2011). 27. Linde, A. D. Decay of the false vacuum at finite temperature. Nucl.Phys. B216, 421 (1983). 28. Moore, G. D. & Rummukainen, K. Electroweak bubble nucleation, nonperturbatively. Phys. Rev. D 63, 045002 (2001). hep-ph/ 0009132. 29. Kirzhnits, D. A. Weinberg model in the hot universe. JETP Lett. 15, 529–531 (1972). [Pisma Zh. Eksp. Teor. Fiz.15,745(1972)]. 53. Balibar, S., Mizusaki, T. & Sasaki, Y. Comments on heterogeneous nucleation in helium. J. Low Temp. Phys. 120, 293–314 (2000). 54. Zhelev, N. et al. The A-B transition in superfluid helium-3 under confinement in a thin slab geometry. Nat. Commun. 8, 15963 (2017). 55. Hiscock, W. A. Nucleation of vacuum phase transitions by topolo- gical defects. Phys. Lett. B 366, 77–81 (1996). gr-qc/9510003. 56. Gregory, R., Moss, I. G. & Withers, B. Black holes as bubble nucleation sites. J. High Energy Phys. 03, 081 (2014). 1401.0017. 57. Fukuyama, H., Ishimoto, H., Tazaki, T. & Ogawa, S. 3He melting curve 30. Kirzhnits, D. A. & Linde, A. D. Macroscopic consequences of the below 15 mK. Phys. Rev. B 36, 8921–8924 (1987). Weinberg model. Phys. Lett. 42B, 471–474 (1972). 58. Swift, G. W. & Buchanan, D. S. Nucleation and growth of 3He-B in 31. Kuzmin, V. A., Rubakov, V. A. & Shaposhnikov, M. E. On the anom- alous electroweak Baryon number nonconservation in the early universe. Phys. Lett. B 155, 36 (1985). 32. Witten, E. Cosmic separation of phases. Phys.Rev. D30, 272–285 (1984). 33. Hogan, C. J. Gravitational radiation from cosmological phase tran- sitions. Mon. Notices Royal Astron. Soc. 218, 629–636 (1986). 34. Mazumdar, A. & White, G. Review of cosmic phase transitions: their significance and experimental signatures. Rept. Prog. Phys. 82, 076901 (2019). 1811.01948. 35. Hindmarsh, M. B., Lüben, M., Lumma, J. & Pauly, M. Phase transitions in the early universe. SciPost Phys. Lect. Notes 24, 1 (2021). 2008.09136. 36. Amaro-Seoane, P. et al. Laser interferometer space antenna. Pre- print at https://arxiv.org/abs/1702.00786 (2017). 37. Caprini, C. et al. Detecting gravitational waves from cosmological phase transitions with LISA: an update. J. Cosmol. Astropart. Phys. 03, 024 (2020). 1910.13125. 38. Fialko, O., Opanchuk, B., Sidorov, A. I., Drummond, P. D. & Brand, J. Fate of the false vacuum: towards realization with ultra-cold atoms. Europhys. Lett. 110, 56001 (2015). 39. Braden, J., Johnson, M. C., Peiris, H. V. & Weinfurtner, S. Towards the cold atom analog false vacuum. J. High Energy Phys. 2018, 14 (2018). 40. Zhelev, N. et al. Fabrication of microfluidic cavities using Si-to-glass anodic bonding. Rev. Sci. Instrum. 89, 073902 (2018). 41. Ruutu, V. M. H. et al. Vortex formation in neutron-irradiated super- fluid 3He as an analogue of cosmological defect formation. Nature 382, 334–336 (1996). 42. Callan, C. G. & Coleman, S. Fate of the false vacuum. II. First quantum corrections. Phys. Rev. D 16, 1762–1768 (1977). 43. Kibble, T. W. B. Topology of cosmic domains and strings. J. Phys. A: Math. Gen. 9, 1387–1398 (1976). 44. Zurek, W. H. Cosmological experiments in superfluid helium? Nat- ure 317, 505–508 (1985). 45. Levitin, L. V. et al. Phase diagram of the topological superfluid 3He 3He-A. Jpn. J. Appl. Phys. 26, 1828 (1987). Acknowledgements This work was supported at Cornell by the NSF under DMR-2002692 (J.M.P.), PHY-2110250 (E.M.), and in the UK by EPSRC under EP/J022004/ 1 and by STFC under ST/T00682X/1 (M.H., J.S., K.Z.). In addition, the research leading to these results has received funding from the Eur- opean Union’s Horizon 2020 Research and Innovation Programme, under Grant Agreement no 824109 (J.S.). Fabrication was carried out at the Cornell Nanoscale Science and Technology Facility (CNF) with assistance and advice from technical staff. The CNF is a member of the National Nanotechnology Coordinated Infrastructure (NNCI), which is supported by the National Science Foundation (Grant NNCI-1542081). Author contributions Experimental work and analysis was principally carried out by Y.T. with early contributions by D.L. and A.E. assisted by A.C. with further support from E.N.S. and J.M.P. Presentation of figures was the joint work of Y.T. and A.E. assisted by K.Z., M.H. and D.L.. N.Z. had established most of the routines for the phase locked loop operation of the quartz fork for earlier experiments. E.M. significantly contributed to exploration of the phase diagram and the writing of the manuscript, and N.Z. and T.S.A. estab- lished and carried out the nano-fabrication of the channel. M.H. and K.Z. in conjunction with E.M. explored the relationship of κ to σ and calcu- lated the contours of constant κ. J.M.P. supervised the work and J.M.P., E.M. and J.S. had leading roles in formulating the research and writing this paper. All authors contributed to revisions to the paper. Competing interests The authors declare no competing interests. Additional information Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41467-022-35532-7. confined in a nanoscale slab geometry. Science 340, 841–844 (2013). Correspondence and requests for materials should be addressed to J. M. Parpia. 46. Maki, K. Planar textures in superfluid 3He-A. J. Low Temp. Phys. 32, 1–17 (1978). 47. Chaikin, P. M. & Lubensky, T. C. Principles of Condensed Matter Physics (Cambridge University Press, 1995). 48. Greywall, D. Thermal conductivity of normal liquid 3He. Phys. Rev. B 29, 4933–4945 (1984). 49. Heikkinen, P. J. et al. Fragility of surface states in topological superfluid 3He. Nat. Commun. 12, 1574 (2021). 50. Bartkowiak, M. et al. Interfacial energy of the superfluid 3He A-B phase interface in the zero-temperature limit. Phys. Rev. Lett. 93, 045301 (2004). Peer review information Nature Communications thanks Viktor Tsepe- lin, and the other, anonymous, reviewers for their contribution to the peer review of this work. Peer reviewer reports are available. Reprints and permissions information is available at http://www.nature.com/reprints Publisher’s note Springer Nature remains neutral with regard to jur- isdictional claims in published maps and institutional affiliations. Nature Communications | (2023) 14:148 8 Article https://doi.org/10.1038/s41467-022-35532-7 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/ licenses/by/4.0/. © The Author(s) 2023 Nature Communications | (2023) 14:148 9
10.1016_j.esr.2022.101031
Contents lists available at ScienceDirect Energy Strategy Reviews journal homepage: www.elsevier.com/locate/esr The next climate war? Statecraft, security, and weaponization in the geopolitics of a low-carbon future Benjamin K. Sovacool a, b, c, *, Chad Baum a, Sean Low a a Aarhus University, Denmark b University of Sussex Business School, United Kingdom c Department of Earth and Environment, Boston University, USA A R T I C L E I N F O A B S T R A C T Keywords: Climate engineering Carbon dioxide removal Negative emissions technologies Solar radiation management Greenhouse gas removal Energy and geopolitics Climate change and international relations The impacts of global climate change on international security and geopolitics could be of historic proportion, challenging those of previous global threats such as nuclear weapons proliferation, the Great Depression, and terrorism. But while the evidence surrounding the security impacts of climate change is fairly well-understood and improving, less is known about the security risks to climate-technology deployment. In this study, we focus on the geopolitical, security, and military risks facing negative emissions and solar geoengineering options. Although controversial, these options could become the future backbone of a low-carbon or net-zero society, given that they avoid the need for coordinated or global action (and can be deployed by a smaller group of actors, even non-state actors), and that they can “buy time” for mitigation and other options to be scaled up. We utilize a large and diverse expert-interview exercise (N = 125) to critically examine the security risks associated with ten negative emission options (or greenhouse gas removal technologies) and ten solar geoengineering options (or solar radiation management technologies). We ask: What geopolitical considerations does deployment give rise to? What particular military applications exist? What risks do these options entail in terms of weaponization, misuse, and miscalculation? We examine such existing and prospective security risks across a novel conceptual framework envisioning their use as (i) diplomatic or military negotiating tools, (ii) objectives for building capacity, control, or deterrence, (iii) targets in ongoing conflicts, and (iv) causes of new conflicts. This enables us to capture a far broader spectrum of security concerns than those which exist in the extant literature and to go well beyond insights derived from climate modelling or game theory by drawing on a novel, rich, and original dataset of expert perceptions. 1. Introduction The impacts of global climate change on international security could be of historic proportion, challenging those of previous global threats such as nuclear weapons proliferation, the Great Depression, and terrorism [1]. For instance, global economic damages from natural ca- tastrophes, many of them climate related, have doubled every ten years and reached trillions of dollars per year in total damages over the past two decades [2]. The climate-change risks faced by developing countries are even more staggering in magnitude, including vulnerability to extreme weather, deteriorating national security, and degraded public health, among others [3,4]. More explicitly, melting glaciers could flood river valleys in Kashmir and Nepal, and reduced rainfall could aggravate water and food security so that 182 million people could die of disease epidemics and starvation attributable to climate change [5]. Under the most severe of these projections, if the Greenland Ice Sheet would melt, sea levels could rise by 6 m – enough to inundate almost all low lying island states as well as coastal areas from San Francisco and New York to Amsterdam and Tokyo [6]. Military analysts have therefore suggested that climate change acts as a “threat multiplier” to national-security concerns, something that takes existing problems and makes them worse, impinging on global stability [7]. But while the evidence base surrounding the security impacts of climate change are fairly well-understood and with such understanding ever improving, less is known about the security risks around the deployment of emergent climate technologies. In this study, we focus on the geopolitical, security [8], and military risks facing negative emis- sions and solar geoengineering options. We term this the “next climate * Corresponding author. Aarhus University, Denmark. E-mail address: benjaminso@hih.au.dk (B.K. Sovacool). https://doi.org/10.1016/j.esr.2022.101031 Received 28 May 2022; Received in revised form 15 November 2022; Accepted 7 December 2022 EnergyStrategyReviews45(2023)101031Availableonline14December20222211-467X/©2022TheAuthors.PublishedbyElsevierLtd.ThisisanopenaccessarticleundertheCCBYlicense(http://creativecommons.org/licenses/by/4.0/). B.K. Sovacool et al. war” with inspiration from Mann [9], who argued that the first “climate war” was a soft war about ideas and knowledge. That political war was waged over ideas and knowledge about climate change itself, and fuelled by a “thirty-year campaign to deflect blame and responsibility and delay action on climate change.” Our study, by contrast, points out how the next climate war, a second one, would not necessarily be a cold war, and could very well involve active military conflict or hybrid warfare, even a "hot war" or global nuclear conflaguration. Although controversial, negative emissions and geoengineering op- tions could become the future backbone of a low-carbon or net-zero society, given that they avoid the need for coordinated or global ac- tion (and can be deployed by a smaller group of actors, even non-state actors) [10]. Negative emissions technologies may already be neces- sary to account for committed emissions resulting from economic ac- tivity (and “locked in” global warming) and gaps in implementation facing the Paris Agreement [11]. For these reasons, some of the the climate-science literature suggests that negative emissions options “are essential” [12] for reaching climate targets, represent “an inevitable component” [13] of technology portfolios, and are “physically needed” to curtail global warming [14]. Solar geoengineering options are also to “buy being heralded as ways for mitigation and negative-emissions options to be scaled up [15], and to help address unknown and potentially dangerous tipping points in the climatic sys- tem [16]. Some suggest that solar geoengineering “must be considered” as a feasible option for supplementing carbon abatement [17]. time” In this study, we utilize a large and diverse expert interview exercise (N = 125) to critically examine the geopolitical, security, and military issues associated with ten negative emission options (or greenhouse gas removal technologies) and ten solar geoengineering options (or solar radiation management technologies). We ask: What geopolitical con- siderations does deployment give rise to? What particular military ap- plications exist? What risks do these options entail in terms of weaponization, misuse, and miscalculation? We examine such existing and prospective security risks by means of a novel conceptual frame- work that captures a far broader spectrum of security concerns than those which exist in the extant literature (see Section 2) and which go well beyond insights derived from climate modelling and game theory by drawing on a novel, rich, and original dataset of expert perceptions. In doing so, we aim to make four contributions. First, we expand the geopolitical discussion of climate protection beyond only references to the ability for deployment or research of geoengineering to cause con- flict [18]. Geoengineering as an actual weapon in war is plausible but currently deemed to be unlikely, while conflict to stop the deployment of geoengineering, or as a spill-over from its consideration and research, is more plausible [19,20]. Our study thus expands the depth and scope of connections between conflict and climate technology. Second, the risks posed by climate change are currently seen to be illustrative of indirect threat multiplication: states openly developing, initiating, reacting, or prohibiting the deployment of some climate technologies, makes the climate change’s diffuse problem structure more direct, and introduces a widespread “security logic” into climate governance. A key aspect is elevation out of “normal politics,” circum- venting usual systems of deliberation or checks and balances, with po- tential for systemic brinksmanship [21]. These options need not be deployed in order to for this to occur – active research and political platforms might suffice [22]. Third, we reveal governance spillovers and linkages – our inquiry extends a lens to potential consequences of negative emissions and solar geoengineering beyond climate security and securitization. Considering or deploying would have complicating effects within climate gover- nance (coordinated efforts to mitigate and adapt) as well as in adjacent global governance issues across economy and environment. Fourth and lastly, we build on recent conceptual advances on the new “green” geopolitics of energy, e.g., how the relative power of countries like Russia, India, China and the Middle East oil producers might be affected by low-carbon transitions, and what energy geopolitics may look like by mid-century. This focus aligns with recent calls for more politically informed and refined discussions of the future geopolitics of energy [23]. 2. Climate technology, energy resources, and conflict: towards a conceptual framework To both ground and contextualize our study and justify our con- ceptual framework, this section reviews key themes within the global political economy of energy and with a specific focus on conflict and climate change, conflict and energy resources, conflict and negative emissions technologies, and conflict and solar geoengineering. Notably, this does include earlier work on conventional energy systems as well as fossil fuels and energy resources. We contend that this literature had broader relevance for negative emissions and solar geoengineering technologies as a useful analogue to better understand how security and military concerns are treated among energy and climate scholars. Our review includes both empirical studies as well as conceptual ones. The section concludes with a synthetic conceptual framework. 2.1. Conflict and climate change The climate-security literature treats climate change as a “threat multiplier” for violent conflict and – more prominently – the varied socio-economic conditions of human welfare. Key elements include: systemic causality, situated (unique to actor) vulnerabilities and points of failure, and mismatches between academic assessment and security planning [24–27]. There is a hybrid understanding of state and human security in an environmental context [28,29]. Security rationales on climate are forward-looking but conservative: warning about, but ulti- mately coping with the implications of a warming climate for political, economic, and military activities [30,31], rather than advocating for a fundamental reformation of root economic causes [32]. 2.2. Conflict and energy resources The literature on conflict and energy resources is the most robust and established (compared to the more emergent literature on negative emissions or solar geoengineering). Here we summarize four distinct intellectual threads: oil embargoes and wars, manipulating cross-border energy flows, sanctions and the use of “energy weapons,” and work on energy resources and military conflict [33]. We first note the work on historical oil embargoes and wars, one of the most famous examples being the nationalization of the Suez Canal. In 1956, Egyptian President Gamel Abdel Nasser nationalized the Suez Canal, seizing a British asset and particularly salient symbol of its im- perial past. The crisis brought commercial oil flows through the canal to a halt, leading to severe oil shortages in Western Europe, a situation the press dubbed “Europe’s oil famine.” [34] It also triggered a military response from the UK, France, and Israel. US President Dwight Eisen- hower, wary that the conflict would play into Soviet hands, declined to send additional oil shipments to needy European countries until the British and French had withdrawn their troops from the area. The oil-starved British and French quickly succumbed to American pressure and, a year later, the canal was reopened [35]. The Yom Kippur War and oil embargo of 1973 is another classic example. In terms of manipulations of energy flows, another line of evidence suggests that decades of progressing globalization have brought us to an age of “connectivity” that creates opportunities for states to “weaponize interdependence” [36] and to pursue war by other, predominantly economic, means [37]. States have long used – or tried to use – energy resources and related technologies as instruments of foreign policy, a practice known as “energy statecraft.” [38]. These different techniques of energy statecraft can serve goals that are benign – for example, to foster peace and interdependence between countries – or less benign – to exert geopolitical leverage over other countries. States can attempt to EnergyStrategyReviews45(2023)1010312 B.K. Sovacool et al. manipulate cross-border flows of energy directly. They can do so through sanctions or boycotts, as in the iconic case of the 1973 Arab oil embargo. In the case of natural gas, the much-publicized Rus- sian–Ukrainian gas crises have often been interpreted as instances of the so-called Russian “gas weapon.” [39,40] Another opportunity to disrupt physical flows of energy lies not at the upstream point of production, but at the midstream segment – that is, transportation. Grid-based energy flows like pipelines or electricity transmission lines offer opportunities for sabotage or attacks. Long-distance pipelines that cross multiple countries can be disrupted by the transit countries [41]. While most of the time, these pipelines operate without any notable problem, politi- cally motivated disruptions have occurred [42]. Sanctions and boycotts are another prominent tool, often used by importers. In the mid-1980s, for example, Europe and the United States joined the efforts of a number of developing countries to place an oil embargo against Apartheid-era South Africa [43,44].Another high-profile case is the current oil sanctions against Iran, or the post-2014 energy sanctions against Russia over their annexation of Crimea [45]. Sometimes, countries can utilize their energy infrastruc- ture to exert geopolitical pressure or seek to stop sanctions. In November 2021, for example, the leader of Belarus threatened to cut off gas sup- plies to Europe, restricting access to their pipelines, if the European Union imposed sanctions for how Belarus was treating migrants wishing to enter Poland [46]. This shows how energy resources can be used as a tool to gain concessions over completely unrelated issues such as immigration, human rights and refugees. Finally, and most seriously, energy has been a key source of armed conflict, though probably in different ways than most people think. The dominant view is that oil and gas are scarce resources, which thereby provides direct fuel for most international conflicts and wars. The rapid depletion of conventional reserves coupled with the fast-growing energy hunger of countries like China and India is believed to trigger “resource wars” between major consumers. In reality, there have been few actual wars initiated primarily to territorially conquer oil and gas fields [47]. The first Gulf War may be an example, as well as Japan’s invasion of Indonesia during World War II, but it is hard to find other examples. Oil may have been a factor in several other wars – the US invasion of Iraq in 2003, for instance, or the Iran–Iraq War – but even in those instances it never was the sole or even primary casus belli. Conquering oil fields or destroying enemy oil installations and supply lines might explain some of the military developments in major conflicts – like, for instance, the battle of Stalingrad in World War II – but that is fundamentally different from causing the conflict itself. Table 1 Causal pathways between oil and international conflict. The logical consequence of the resource-war narrative, which guides how most people think about oil in international affairs, is that petros- tates are likely to be the target of an attack rather than act as the aggressor. However, the evidence points in the opposite direction. Jeff Colgan’s work has shown that “petrostates,” where revenues from oil exports constitute at least 10% of GDP, have an “above average pro- pensity to engage in militarized interstate disputes.” [48] He found that “petrostates” engaged in military conflict at a rate about 80% higher than non-petrostates over the period 1965–2001 – a phenomenon which he called “petro-aggression.” His explanation was that revolutionary leaders are able to rely on oil export revenues to consolidate power and provoke international conflict. Thus the international trade in oil as currently structured places large amounts of money into a political system ill-equipped to use it responsibly [49]. Examples of revolutionary petrostates are Iraq under Saddam Hussein (who invaded Iran in 1980 and Kuwait in 1990), Libya under Gaddafi (who engaged in four sepa- rate border wars with Chad, as well as a variety of militarized disputes with other countries such as Egypt, Tanzania, and the US), and Russia under Putin (who engaged militarily in Georgia, Ukraine, and Syria) [50]. Oil also impacts international security through its complicated links with terrorism. The world’s most celebrated international renegade, Osama bin Laden, cast the West’s consumption of Persian Gulf energy as a central part of a complicated narrative that features the plundering of the Middle East’s riches. Islamist insurgents appear to have taken these calls at least somewhat to heart and have mounted episodic attacks against energy targets. In February 2006, for example, the Saudis thwarted an attack on the oil-processing facility at Abqaiq [51]. Overall, however, the material effects of these (attempted) attacks have been limited. Recent research has also dispelled the myth that ISIS has been able to generate a large income from exploiting oil fields and refineries in Syria and northern Iraq [52]. Colgan also developed a typology of “causal pathways” between oil and international conflict, depicted in Table 1 [53]. It lists no less than eight mechanisms linking oil to war, including classic resource wars, petro-aggression, and terrorism. Sovacool and Walter looked explicitly at hydropower and noted five ways the literature suggests it can contribute to conflict: dams can be a military tool, can be targeted during military campaigns, can be attacked by non-state actors to pro- mote their agendas, can be used for political goals (jobs, poverty reduction), and can be a source of contention in political debates [54]. Other work has focused on typologizing or classifying forms of en- ergy conflict. Månsson notes that sometimes the end goal of a conflict is Dimension Pathway Causal mechanism Example(s) External and international: geopolitics and Resource wars Oil reserves raising the payoff of territorial conquest resources Iraq–Kuwait, 1990; Chaco War, Japan, 1941 Risk of market domination Threat of conquest to ally or key territory US–Iraq, 1991 Oil industry grievances Presence of foreign workers creates grievances for state or non- state actors Al-Qaida; Iran hostage crisis Internal and domestic: politics in producing Petro-aggression countries Oil reduces the accountability of leaders, lowering the risk of instigating wars Iraq–Iran; Libya–Chad; Egypt Petro-insurgency Oil income provides finances for actors to wage war Iran–Hezbollah; Saudi Arabia–Afghanistan Externalization of civil wars Oil creates conditions for civil war that then lead to foreign intervention or spillover Libya–NATO; Angola–Cuba; Sudan–Chad Internal and domestic: Access concerns in Transit route Efforts to secure transit routes create a security dilemma consuming countries Obstacle to multilateralism Importers attempt to curry favor with petrostates to prevent multilateral cooperation Sudan; South China Sea; Strait of Hormuz US–China friction over Iran; Sudan Source: Modified from [55]. EnergyStrategyReviews45(2023)1010313 B.K. Sovacool et al. are increasingly concerned that the same logics of exploitation and conflict more familiar in the former could be repeated [67,68]. 2.4. Conflict and solar geoengineering The conflict literature is built mainly around stratospheric aerosol injection (SAI) – currently thought to be the only cost-effective, plane- tary-scale, and high-leverage form of solar geoengineering. Studies suggest that SAI could germinate contestation and conflict since its deployment would spread globally—aerosol injection would circulate similarly to the sulfur eruptions from volcanoes [69]. Moreover, extensive and recurrent global cooperation would be needed to sustain solar geoengineering for decades to centuries, and without interruption from wars and economic stresses, including unprecedented liability claims [70,71]. As such, geoengineering would pose long term mainte- nance and liability concerns for those countries that do deploy it. Others suggest that SAI’s most resonant impacts may transpire ideologically, by fostering political brinksmanship [72] or by delaying decarbonization [73]. Dalby muses that security issues could make solar geoengineering the most important facet to “the new geopolitics of the Anthropocene.” [74]. The benefits and risks of deploying solar geoengineering on a global scale are prominently projected by earth system models. But these are increasingly constructed as optimized, best-case scenarios that assume century-long technical controllability and global coordination [75,76] - and the value of these scenarios for geopolitical risk analysis and decision-making has been called into question for exactly this reason [77–79]. Key advocates also reduce security to “weaponization” - the direct use of SAI to impact an adversary’s regional climate, weather patterns, and the systems they underpin. The implication is that SAI is too imprecise to directly weaponize, which may have merit – but a much more complex range of security issues is elided by this incorrect use of weaponization as a proxy [80]. SAI is less commonly imagined under non-ideal conditions, or as “unconventional” uses. These include deployments (or escalated research) conducted unilaterally, through proxy actors, via smaller co- alitions, or in a decentralized set of deployments, or where multiple competing schemes attempt to offset each other [81–84].Some uncon- ventional deployments have been assessed either as qualitative pieces of reasoning [85] or game-theoretic studies that calculate strategic state actions [86]. For some, unconventional scenarios are implausible and do not serve policy deliberation [87]. Others disagree - Corry argues that SAI research, or posturing – even without deployment – can serve as part of economic and diplomatic statecraft for achieving non-climate goals [88]. This is a valuable contribution; the goal of deploying SRM is usually assumed to be climate-related, rather than to (threaten to) affect climate as a proxy for other geopolitical aims. In this vein, Briggs and Matejova pose solar geoengineering as a potential kind of ‘hybrid’ conflict that combines eroding enemy infrastructure with technological, environmental, and economic dimensions [89]. 2.5. Towards a synthetic conceptual framework Drawing from these diverse strands of thought, we introduce a conceptual framework that captures a broader set of conditions and factors by which negative emissions and solar geoengineering options can shape geopolitics, statecraft, conflict, terrorism, and war (see Fig. 2). This framework explores four dimensions of how climate-technology deployment can interrelate with destabilizing politics, insecurity, and both internal and external conflict. The first is where climate technology can be utilized as a negotiating tool. In this category, climate geo- engineering systems are used as a means by an actor to impair the se- curity of other actors and achieve other, non-energy related, objectives. One example is geoengineering states deliberately using their power to get concessions on other things (trade, intellectual property). In the second dimension, climate technology is an objective for capacity, Fig. 1. A conceptual framework for energy resources and conflict. Source [57]: primarily for the participants to improve their own security by securing some part of the energy system, that is, energy is an objective in a conflict, such as the Iraqi invasion of Kuwait in the early 1990s (See Fig. 1) [56]. In a second category, an energy system is a means of initiating a conflict related to something else, such as Russia’s use of its fossil-fuel exports to get concessions in other political areas from countries such as Belarus and Ukraine. In a third situation, an energy system is partly the root cause of a conflict, as it has destabilized a so- ciety and thereby contributed to, or exacerbated, a conflict, such as rapid environmental degradation causing refugees or creating social move- ments that try to topple governments, such as those ongoing with indigenous people in North America and Asia. Månsson then goes on to describe a fair number of actual conflicts that meet his typology, with dozens of examples including major wars but also border disputes, suppliers such as OPEC using their “oil weapon,” and local attacks from terrorists and saboteurs in places such as Nigeria. 2.3. Conflict and negative emissions technologies Only recently has a first framework been constructed to elucidate the potential geopolitical dimensions of negative emissions technologies as a broad suite of large-scale energy production, resource usage, carbon storage, and land-use systems [58]. Direct air capture approaches rely on massive energy costs which could be coupled with either existing fossil-fuel or novel renewable infrastructures - possessing the potential to entrench or reorient the global carbon economy and its geopolitics [59,60]. Meanwhile, land-use approaches (large-scale forestry or agri- cultural management) by necessity entail heavy spatial and resource usage as well as pose inequities and trade-offs for the populations currently resident on or adjacent to the land [61]. Ocean based and marine carbon removal, and even the protection of coral reefs for ecosystem restoration, could also intersect with fisheries conflicts around the world [62]. This deliberately geopolitical focus on various aspects of negative emissions and carbon removal is nascent, but raises issues highlighted by antecedent conflicts in global food systems. These studies cite land- grabs and ownership conflicts, the food versus ethanol dilemma (e.g. the 2005 global food crisis), “phantom commodities”, the consequences of shifting prices in one-resource economies, and other issues and challenges confronting rural, smallholder communities – often accom- panied by the particular pressures experienced by indigenous pop- ulations, or in the global South [63–65]. Others cite extractive industries in energy and other natural resources as relevant antecedents, raising questions of hazardous siting and carbon infrastructure lock-in [66]. As carbon removal technologies and their related approaches are looking beyond the terrestrial and into coastal and oceanic environments, some EnergyStrategyReviews45(2023)1010314 B.K. Sovacool et al. Figure 2. A conceptual framework for the geopolitics, securitization, and weaponization of negative emissions and solar geoengineering technologies. Source: Authors. Although the qualitative analysis of our expert interview results (N = 125 participants) discusses numerous nuances and differences between the ability for various negative emissions and geoengineering technologies to create or contribute to conflict in the sections to come, we do approach them throughout the study from the same framework. control, or deterrence. The end goal is for an actor or government to improve their own security by securing some part of the geoengineering system. Otherwise, deployment could be used to protect or enhance the security of existing military assets. In the third dimension, climate technology is a target in an ongoing conflict. This is where systems can be threatened or destroyed during existing disputes, civil wars, or regional conflicts. Such threats can be both intentional and accidental, as well as executed physically or through cyberattacks. In the final category, a climate technology or geoengineering system could even represent, solely or partly, the cause of a new conflict, as it has destabilized a society and thereby contributed to, or exacerbated, a conflict. For example, the exploration, production, or use of geoengineering resources can have side-effects that cause destabilizing environmental stress. In other cases, states may deploy weather modification or geoengineering as a hostile, intentional act. We should note that these four broad dimensions operate at the level of ideal types. Some conflicts may be explained by several of the pro- posed categories or interactions between them. One example is a state that uses force to maintain control of a geoengineering system (first category), exploit such control to extort another state (second category), gain concessions during negotiation (third category) and reduce that state’s environmental quality or weather, perhaps as a way to demon- strate its capacity to do so (fourth category). Most of the examples pointed to in our data fit cleanly within one dimension, but some cut across multiple dimensions. Furthermore, there are differences and varying degrees to which particular negative emissions or geo- engineering technologies can engender conflict and instability. These include the much greater likelihood that a “switch on” or "switch off” of SRM would have a faster temperature response than carbon and greenhouse gas removal, the impact of the former on local weather patterns, the potential speed and ease of deployment of the former, and the relative lack of its buy-in in scientific and policy circles at this time. We will return to these themes in the Results and Conclusion. 3. Technology selection and research design With our conceptual framework in place, this section briefly explains our selection of twenty different negative emissions and solar geo- engineering options, elucidates our research design of original expert interviews, and discusses limitations with our approach. 3.1. Selecting a comprehensive portfolio of climate technologies One novelty to our study is that it explores a broad diversity of negative emissions and solar geoengineering options and pathways. As Table 2 summarizes, this includes ten distinct options (drawn from the literature) including nature-based or terrestrial negative emissions technologies such as soil management, ecosystem restoration, and forestry as well as engineered solutions such as direct air capture and enhanced weathering. We also examine ten solar geoengineering op- tions including various forms of albedo management, various ways of modifying clouds, the use of reflectors in space or the upper atmosphere, and stratospheric aerosol injection. 3.2. Research interviews To determine the geopolitical and security risks that may arise with these 20 options, we relied on a large pool of semi-structured interviews, the asking of semi-structured questions in this instance to experts on the topic. We proceeded to interview prominent experts who had high levels of knowledge about our 20 options as evidenced by publishing high- quality peer-reviewed papers on the topic (from 2011 to 2020) or who possessed patents and intellectual property concerning the technologies. Moreover, our recruitment and sampling of experts focused on a mix of advocates and critics of both negative emissions technologies and solar geoengineering pathways. We conducted 125 individual face-to-face interviews with experts closely associated with negative emissions and/or solar geoengineering research or commercialization over the course of May to August 2021. We explicitly asked, among other questions, “What are some of the governance or geopolitical risks with deployment?“, “What particular military applications exist?“, and “What risks do these options entail in terms of weaponization, misuse, and miscalculation?” Table 3 shows an overview of the demographics of our sample, and Annex I lists all 125 experts who participated (although it does not match them with their respondent numbers, to protect the anonymity of their statements). Although we did secure interviews with members of civil society and nongovernmental organizations as well those employed by governments EnergyStrategyReviews45(2023)1010315 B.K. Sovacool et al. Table 2 Exploring a portfolio of negative emissions and solar geoengineering technologies. Type Option Description Negative emissions Negative emissions Negative emissions Negative emissions Negative emissions Negative emissions Negative emissions Negative emissions Negative emissions Negative emissions Carbon capture and utilization and storage Employing technologies, processes or solvents that extract, capture, transport, utilize, and/or store carbon dioxide Afforestation and reforestation Planting trees or vegetation to absorb carbon dioxide Bioenergy with carbon capture and storage Harnessing specific energy crops (e.g., perennial grasses, or short-rotation coppicing) or increased forest biomass to replace fossil fuels, and capturing and storing consequent carbon dioxide Biochar Managing the thermal degradation of organic material in the absence of oxygen to increase soil carbon stocks and improve soil fertility Soil carbon sequestration or enrichment Growing cover crops, leaving crop residues to decay in the field, applying manure or compost, using low- or no-till systems, and employing other land management techniques to improve soil Ocean iron fertilization Utilizing planktonic algae and other microscopic plants to take up CO2 and convert it to organic matter, some of which sinks and is sequestered in ocean Enhanced weathering and ocean liming or alkalization Deploying physical or chemical mechanisms to accelerate the geochemical processes that naturally absorb CO2 at slow rates. Direct air capture Blue carbon and seagrass Ecosystem restoration Capturing carbon dioxide from the air via engineering or mechanical systems, and then using solvents or other techniques to store it safely Harnessing the ability for coastal mangrove forests, tidal marshes, and seagrass meadows to accelerate their uptake of carbon dioxide Managing the restoration of ecosystems (including wetlands, peatlands, and grasslands) to reverse environmental damage and increase their ability to absorb greenhouse gases Solar Space mirrors geoengineering Placing scatterers, reflectors, or spacecraft in outer space to reduce the amount of sunlight entering the Earth’s atmosphere Solar High altitude sunshades geoengineering Placing scatterers or reflectors in the upper atmosphere (e.g., stratosphere) to reduce the amount of sunlight entering the Earth Solar Stratospheric aerosol injection geoengineering Dispersing aerosol particles through high-altitude jets (e.g., sulfur) into the lower stratosphere, where they would reflect a small portion of incoming sunlight back to space, cooling temperatures. Solar Cirrus cloud thinning Reducing cirrus cloud cover to facilitate the release of outgoing radiation and lower temperature geoengineering Solar Marine sky or cloud brightening geoengineering Coordinating fleets of ships to spray sea water into the air below marine clouds, thereby increasing their reflectivity and longevity Solar geoengineering Albedo modification via human settlements Solar geoengineering Albedo modification via grasslands and crops Enhancing the reflectivity of buildings, roads, or other structures to cool the global temperature Enhancing the reflectivity of grasslands, crops, and land to cool the global temperature Solar Albedo modification via deserts Enhancing the reflectivity of deserts to cool the global temperature geoengineering Solar Albedo modification via clouds geoengineering Creating new clouds or reflecting more sunlight from the surface to increase the heating of the lower atmosphere, improving cloudy-sky shortwave climate forcing Solar Ice protection Protecting glaciers and ice sheets by either slowing their melting or reflecting solar radiation via tarpaulins geoengineering Source: Authors and commercial entities in the private sector, the sample is strongly concentrated towards experts at universities and research institutes (or who split their time between universities and one of these other sectors). That said, the sample does include scholars from more than 30 disci- plines as well as a dozen participants from the Global South. Given that interviewees were speaking on their own behalf, and also given the sensitivity of the topic, the data from these interviews is presented here as anonymous with a generic respondent number (e.g., R10 for respondent 10, or R110 for respondent 110). 3.3. Limitations Before outlining the results, and despite the many benefits of our large and diverse sample of expert interviews, e.g. by facilitating triangulation of different viewpoints and the range of perspectives engaged, we here highlight some shortcomings of our research design. One drawback to anonymity is that there is no guarantee this study can be replicated, because the authors cannot correlate the identity of respondents with interviewee statements. Another is that respondents tended to be more critical than positive; the results below, for example, are more serious and sober in their outlook. This is not because the au- thors were selective about comments, but perhaps because anonymity itself incentivizes people to be more forthcoming about problems and issues instead of strengths. Moreover, we took an ethnographic approach that did not correct or problematize responses, so we present the data unfiltered, even if our respondents may have had misperceptions. Lastly, we note that our conceptual framework (Section 2.5) was developed from the emerging research on the geopolitics of energy and resources. While this ensures that our discussion is well-grounded in an established literature, we note that negative emissions and solar geo- engineering options are not exact analogues. They do represent (i) ways to demonstrate one’s level of national commitment to tackling impending problems related to climate change but also (ii) the kinds of risk-limiting investments that are increasingly demanded to reduce the incidence and impact of climate disasters. Nevertheless, the importance of oil and other energy resources to national security stems from the fact EnergyStrategyReviews45(2023)1010316 B.K. Sovacool et al. Table 3 Summary of the demographics of experts who took part in our study. Summary information No. of experts No. of organizations represented No. of countries represented No. of academic disciplines represented Cumulative years spent in industry or the research community Average years spent in industry or the research community No. of experts whose current position falls into the following areas: Civil society and nongovernmental organizations Government and intergovernmental organizations Private sector and industrial associations Universities and research institutes No. of experts from the Global South Source: Authors. No. 125 104 21 34 881 7.8 12 8 11 94 12 that much of modern civilization has been established around and by means of them [90]. We do not claim that solar geoengineering or negative-emissions infrastructures are perfectly analogous to a new oil industry, nor that the next climate war will identically resemble wars of the recent past. Rather, the increasing reliance on these climate tech- nologies to mitigate and forestall climate disasters could unintentionally provide the invitation for them to become higher-priority targets as indispensable parts of energy supplies, resource extraction, food or se- curity systems, or as tools for negotiation and political leverage. 4. Results and discussion: insecurity and conflict in a carbon neutral and solar geoengineering world Our results from the interviews provide strong support for our con- ceptual framework. 4.1. Climate technology as a negotiating tool In this category, negative emissions technologies and geoengineering systems are used by an actor as a means to impair and draw attention to the security of other parties and thereby achieve other, non-energy related, objectives. Our data supports the existence of three distinct categories: Greenfinger, cartels and clubs, and concessions. 4.1.1. Greenfinger The term “Greenfinger” is meant to evoke the James Bond villain Goldfinger, highlighting in particular the threat of global blackmail or suspension of operation of climate systems as a negotiating tool, espe- cially if done by rogue states or terrorists, but also for more benign motivations such as a billionaire with a “green finger” deciding they want to save the world via pre-emptive deployment. R045 captured this latter scenario well when noting that: A billionaire megalomaniac like Richard Branson could deploy geoengineering quickly, some crazy man like Elon Musk, the kind of person that is more likely to have a kind of messiah complex these days. This could be a very dangerous thing. R085 also spoke about how “powerful people” and “hidden actors” with “no patience” could become convinced to deploy: At the moment, the most powerful set of actors who don’t get suf- ficient attention in terms of their potential to deploy geoengineering are the Silicon Valley finance folks. Venture-capital money. People who are impatient for climate action, who have resources to drive change, who bring a kind of libertarian worldview, and are comfortable bypassing political systems. They will do disruptive things, make it happen, bypass need for legitimacy or debate, or meaningful public input, transparency, all the things in principle government funding as opposed to private capital, unfettered and unhinged, could or should provide. “Private philanthropy” is a concern in this space at this moment, in this vacuum of clear governance and understanding of appropriate roles, which can lead to bad decisions, driven by powerful people with no patience for process. R081 agreed and added that “a capitalistic dictator like Bill Gates or Elon Musk could decide to deploy at a moment’s notice, dictating for all of humanity a pathway for climate protection” (emphasis added). R040 also identified blackmail as a real possibility with negative emissions options, as malign actors could “threaten to release all of the carbon from their reservoirs” unless their “demands were met.” R047 agreed and stated that: “It’s a nightmare, and it’s a totally understandable nightmare … where this kind of research leads you either to semi-rogue action from hostile states or private-sector rich dudes with money that launch a program from their couch in Vanuatu.” R109 surmised that “the special sauce of private sector is to be ruthless and quick and that’s not a good prescription for things that have, potentially, such large consequences.” R020 added that it could even be rogue nations or community groups, or even ordinary individuals, who act with green fingers: I really do envision a potential “greenfinger” scenario where a wealthy individual simply decides to do it [to deploy a massive climate geoengineering project]. It could also be a club of rogue nations or a group of countries in peril, think Small Island Devel- oping States in the Pacific that band together with Jack Ma, Jeff Bezos, and Bill Gates, there are quite a few billionaires to choose from. But it could also be ordinary people and communities. In the realm of social media, nothing is stopping a number of individuals jointly funding or crowdsourcing geoengineering without any governance. Through Facebook, I could see thousands of people buying balloons, letting these carry sulfuric acid into the strato- sphere, thousands and thousands of people doing this with the strength of social media, then you have deployment totally ungoverned. This scenario diffuses the risk beyond a single individual to any group of committed individuals able to pool their financial resources, triggering visions of the potential for such efforts to even be crowd- sourced. One study termed this “predatory geoengineering” to capture the potential for self-concerned, but ultimately reckless, actors to deploy their technology without concern for or full consideration of the con- sequences for others [91]. Wagner also identified the possibility for “greenfinger” action at the subnational level as well as other highly centralized forms of technology deployment for solar geoengineering (see Table 4) [92]. 4.1.2. Cartels and clubs Regional blocs of both negative-emissions and solar-geoengineering cartels could prospectively emerge, with the ability to control the pace of global climate change according to their own desires and geoclimatic situations as well as utilize that power to either dominate markets or create geopolitical obstacles to multilateralism, conflict resolution, and cooperation. R047 stated this explicitly: Solar geoengineering leads directly to cartels or clubs who can deploy the technology in question, moving from an experimental program to deployment. Geopolitical power relates here to those who can deploy, and deploy unilaterally, making them almost identical to strategic use of missile-based nuclear weapons … there are breakout risks and the tools that we use for studying strategic nuclear-weapons control between the United States and the Soviet Union could be applied to these [climate] technologies. EnergyStrategyReviews45(2023)1010317 B.K. Sovacool et al. Table 4 Number of actors and deployment pathways for geoengineering including a Greenfinger scenario. Number of actors or deployers 1 Form of governance Nonstate actors Unilateral “Greenfinger” Delivery mechanisms Newly designed aircraft Source: Modified from [93]. ~10 ~100 >~1000 Minilateral Moderately decentralized nonstate solar geoengineering A fleet of aircraft Multilateral Decentralized nonstate solar geoengineering Multiple fleets of aircraft Chaotic Highly decentralized nonstate solar geoengineering Multiple fleets of aircraft or small balloons This comment recasts the strategic value of climate technologies as equivalent to nuclear-weapons stockpiles. R047 went on to articulate that negative-emissions options also enable cartel-like behaviour given there will be small clubs of early deployers: For carbon dioxide removal, you have similar club risks. You can assume people who have large assets of land will be disproportion- ately powerful. All of the major options are highly dispersed - forestry potentially has bad above-ground governance, soil carbon near agriculture bad agricultural governance, underground injection into deep saline aquifers bad water governance, lots of storage or support space bad geologic governance. These would likely create broad-based coalitions of actors that deploy together. R040 even put this in terms of the emergence of a new “Green” Or- ganization of Petroleum Exporting Countries (OPEC): Some negative-emissions options could give rise to an OPEC situa- tion, or the natural gas situation in Europe with Russia. If bioenergy with carbon capture relies on bioenergy, to some degree you could think of a new ‘Green OPEC’ situation. If we’re going to make a big energy system with bioenergy, there are clearly producers and con- sumers. Western Europe would not be able to produce the amount of bioenergy in the way that it’s projected in the scenarios. And if it’s consuming these amounts of energy that are mentioned in the sce- narios, then it could not produce enough bioenergy itself. So, it would be importing bioenergy from Brazil. There are a few countries that will be exporters of biomass: Brazil, Russia, Canada potentially. And again, the same situation that we have for fossil fuels might emerge. Where a set of countries could, at some point, if there is such a reliance, come to dominate the markets. Such a “Green OPEC” could conceivably operate similar to the existing OPEC in terms of its attempts and mechanisms to control prices and access to resources, or even orchestrate embargoes. It could also emerge organically, as funding and policy coordination create coalitions of countries with similar interests, which could lead to “Big Green Deals” around the world and new regimes of cooperation [94]. 4.1.3. Concessions A final negotiating dimension to climate technology could be threats of weather modification or deployment to gain other concessions (e.g., trade, energy prices, changes in national policy) and meet other national objectives—in line with the “energy weapon” discussion in Section 2. R087 identified how hostile, manipulative use of solar geoengineering as a threat was the most likely scenario, noting: I feel conflicted, one of the most likely ways solar geoengineering will get used in a political context is as a political threat: tropical countries are vocal about negatively being impacted. They will say if you don’t offer us compensation, help us address the harm, give us aid or preferable trade deals, we will be forced to take matters in our own hands. R102 also emphasized the very real “potential for things to go wrong” and that there could be “a lot of room for diplomatic conflict” over the deployment of negative emissions and solar geoengineering options, since they could empower actors to make demands for concessions if they were in control of technology vital to stabilize the climate. R028 talks about how such concessions could potentially emerge through international negotiations: There will be people who are unhappy with any decision. What do you do about that? Some of the ways that international negotiations handle that is, they say, “Oh, you’re unhappy with that? F*** you, too bad.” Sometimes some people will pay them money. Sometimes there will be changes in markets that are allowed, like after the Gulf War, for example, gas prices just kind of looped up. Why? Because they needed to pay for the Gulf War, and that was a way to do it economically, and we just sort of said, “Yes, you can have it; that’s fine.” So, this sort of stuff, this kind of horse trading, it’s really interesting, and it doesn’t always go well. Adding more detail, R106 proposed that, given the difficulty and complexity of attribution, an “authoritarian leader who wants to show he is in control of things”, both to the world and his own population, need not even have the capability to engage in geoengineering; but rather: “you pretend to control things, even of course if you don’t control the secondary, tertiary effects. But at least you show you are moving the controls.” In this regard, one could imagine someone like Kim Jong Un attempting to demonstrate their capability in this area, similar to what is currently done for nuclear weapons. 4.2. Climate technology as an objective for capacity, control, or deterrence In this category, climate technology becomes an objective for an actor to enhance their own security by building their military capacity, securing or controlling some part of the geoengineering system, or de- terring others from attacking it. 4.2.1. Building military capacity States for example could utilize geoengineering technologies to build their military capacity, similar to the dual-use option of things like nu- clear technology (to make nuclear weapons or generate nuclear energy) or chemical manufacturing (which can manufacture chemical weapons). R070 spoke about these risks with solar-geoengineering options related to aerosols, sunshades, and space shields: Options that depend upon the expansion of an aerospace or space industry are a security risk, as you are bringing high-tech space in- dustries into countries that don’t normally have them. Players gain new capabilities that they didn’t have before, and these could spill- over into an arms race or new technology in the hands of new ac- tors could become military or hostile, that could be a possibility. If a rogue nation develops launch systems, that opens a door to their launching new satellites or defence systems or even missiles, creating tensions. R010 also spoke about how, while solar geoengineering techniques might not be realistic weapons themselves, they could become “coupled to weapons, by enhancing military capability, especially how much it would improve high-technology skills, skills that would be very useful for crossover impacts on military design.” R090 noted that the United States Air Force and Navy would likely benefit from having to design EnergyStrategyReviews45(2023)1010318 B.K. Sovacool et al. new planes for aerosol injection, or some aspects of cloud thinning, enabling them, and companies like “Boeing or Lockheed to take billions of dollars from taxpayers.” Some respondents also spoke about military capacity and negative- emissions technologies. R028 suggested that carbon dioxide removal could build military capability in research or engineering, noting that “any time you do anything this large-scale, the military industrial complex gets involved, it’s just totally unavoidable that they will benefit.” 4.2.2. Control Actors could also compete militarily or otherwise for resources critical to geoengineering as well as engage in resource wars over minerals or supply chains needed to manufacture climate technologies. One such resource would be the ocean, a resource necessary for large- scale enhanced weathering (as a basin for run-off materials and stor- age) or ocean iron fertilization. As R025 explained: The risk of conflict can be severe, risks are transboundary in nature, risks also transcend military security, can facilitate international conflict: e.g. the South China Sea and options such as ocean pro- tection there would be highly contested. What if the impacts are more negative in one country, more beneficial in another? R081 mentioned the potential for these new systems to usher in an era of new resource wars (conflicts over control of resources, or to control their consequences), stating that “we don’t want to have these new climate technologies because they’re hyper-centralised, they have enormous geological risks. They can produce what sociologists refer to as the resource curse.” R084 elaborated on this theme as well: That there is potential for some of these techniques – and I’m thinking particularly here of carbon dioxide-removal techniques – to operate in parts of the world which are generally less economically developed. So there is potential for use of large land areas, sparsely populated, which often correlates with a low development index. But there’s also risk of the resource conflicts where money flows to the elites in those societies, or corruption where benefits aren’t actually what they’re thought to be, because of poor governance. In addition to the potential for conflicts between nations or actors located in disputed territories, here the prospect is raised of provoking or heightening conflict between groups belonging to a particular society. 4.2.3. Deterrence Climate technology could enhance the protection of military sys- tems, augmenting potential deterrence—avoiding attacks because per- petrators believe they would not be successful, or would prompt strong retaliation. R103 commented that he believed early deployment would likely “involve the military” and could be used in military operations or extreme environments to protect installations—especially things like ice protection (for Arctic and Antarctic military bases) or ocean alkalin- isation or fertilization (for coastal naval bases). R011 also suggested that “the military will likely be an adopter, to protect its installations. One of the great threats of global warming is sea-level rise, and every naval base is at sea level, I can see them using geoengineering to build resilience.” R064 spoke about how there could also be direct military use for syn- fuels via direct air capture—meaning this could help various militaries reduce dependence on fossil fuels, especially oil, and provide an op- portunity to enhance power projection in the Middle East. R107 stated that he believed navies would deploy marine cloud-brightening, shading, or fogging options or could be called on to build installations or conduct ocean-geological activities. A corollary to this argument is that the military may be used for another form of deterrence, to deter noncompliance with meeting climate or negative-emissions targets. R064 spoke about how they believed “the military could be sent in to enforce compliance or ensure that afforestation or large-scale CDR projects are protected, the military could protect them to deter their destruction.” R063 also picked up on this theme: I could envision an entire shift in how navies or militaries operate. They no longer depend on kinetic power, as sources of influence, or killing people, but pivot to protecting people rather than harming them. I can see armies and navies getting sent in to protect forests or oceans, global public goods, to prevent them from being cut down. Stuart Candy has even envisioned—hypothetically, of course—how such military deployment could occur in the mid-2020s or 2030s with the creation of a U.S. Earth Force to serve alongside the Army, Navy, Air Force, Marines and Coast Guard, tasked with ensuring global climate security and enforcing compliance with international targets [95]. Fig. 3 shows what a futuristic advertising campaign might look like for such military deployment, in the name of deterrence of climate insecurity. Candy also asks one to imagine what could be accomplished if such military resources were repurposed towards climate-stability ends. 4.3. Climate technology as target in an ongoing conflict This category encompasses when or where climate technologies or geoengineering systems could be threatened or destroyed during ongoing disputes, civil wars, or regional conflicts, notably via physical attack, cyberattack, or accidental collateral damage. 4.3.1. Intentional physical attack The direct destruction of geoengineering systems could be of stra- tegic value during ongoing military campaigns or conflicts as it would induce localized climate-change effects or decrease the morale of enemy populations. R002 said that: Stratospheric aerosol injection, cirrus cloud thinning, and cloud brightening could be all be military targets themselves. Countries could target them as key infrastructure during ongoing conflicts, just like militaries target power lines or GPS satellites and critical infra- structure now. R064 added that “shooting down the planes doing aerosol injection, sinking the ships doing ocean protection, could be very plausible during conflicts.” R056 noted as well that “high-altitude sunshades would be prone to regional targeting in a conflict,” R099 mentioned the same vulnerability for “shooting down balloons” to stop aerosol injection. Such attacks do not have to be direct. R002 expressed concern about the connections between cloud brightening or cloud thinning and mili- tary deployments, noting that they could be used for light versions of weather modification that could help enhance the potential for military success: You could potentially use it for weather modification because you can vary the perturbation on the timescale of days and at as fine a spatial scale as your intervention is done. So, if you’ve got, whatever it is, 500 ships that are doing this deployment, you could turn half on, half off. Eventually, what you would have is you could run your weather forecast with or without your marine cloud brightening on and then you could pick which weather forecast you like better, whichever one suits your military deployment or targeting. R011 added that: I am not sure if it can be used as a weapon per se, but many of these technologies are linked with the longer history of climate and weather modification, where military uses of the technology are significant. The United States seeded clouds over Cuba to try to ruin the sugar harvest, and the government also did the same to try to make the Ho Chi Minh trail muddier in Vietnam … These technol- ogies enhance the opportunity for the CIA to possibly control the weather of other states. I am sure the military is thinking about it. Although seemingly innocuous, these sorts of changes could tip the EnergyStrategyReviews45(2023)1010319 B.K. Sovacool et al. We’ve weaponized the internet and software, so certainly something as physical as carbon dioxide removal or solar geoengineering has a very real security vulnerability. The control systems, the software systems would all be prone to cyberattack. R070 also discussed the risk of “terrorism” and “cyberterrorism” against the “control centers” for solar geoengineering, adding that “due to hacking, one could lose the system partially or completely.” R081 even spoke about how carbon dioxide-storage facilities could be prone to “hacking, the systems controlling them could be hacked.” As they went on to explain: Some inventive people might intentionally attack a system, there are some people who can turn everything into a weapon. I mean, res- ervoirs might be interesting for pirates and blackmailing, but I think the attack will not be physical. I think the attack will be mostly on software so the systems might be hacked, and you might blackmail whatever, a company or government who will say "if you are not paying me, whatever, ten billion Bitcoins, I’m going to release ten gigatons of carbon in an instantaneous impulse from your reservoir" … terroristic attacks or some kind of hacker attacks on the reservoirs are a risk. 4.3.3. Accidental collateral damage Unintended disruptions during other conflicts (e.g., civil wars, mil- itary campaigns) came up as a final concern. R055 said that “there is enough collateral damage already during conflicts, power plants and pipelines get attacked or destroyed all the time, I don’t see why geo- engineering infrastructure would be any different.” R100 added that “negative emissions technologies would need to be ubiquitous by 2050 or 2070 to the point where they would certainly be deployed among fragile states or within war zones, carbon-storage reservoirs would also invariably exist in some conflict zones, all of this would create a security risk.” 4.4. Climate technology as a cause of new conflict In this final category, climate technologies are solely or partly the cause of conflict, notably, by destabilizing a society and thereby contributing to a new conflict. Our evidence supports four different as- pects to this dimension: weaponization and hostile deployment, inse- curity and asymmetrical protection, miscalculation, and arms races along with counter-geoengineering. 4.4.1. Weaponization and hostile deployment Weaponization would be the most direct route to conflict, as it in- volves the hostile deployment of climate technology (most likely, though not limited to, weather modification) against another state or entity leading to a great power war. Some of these weapons could be quite powerful. R096 argued that deployment of some ocean technolo- gies could even be engineered to cause mass dead zones (affecting fishing and food supply): If governments cannot be trusted with human rights or fairly benign technologies, then they cannot be trusted with these technologies, the risks of them being used as weapons is too large. Ocean-based carbon dioxide removal or fertilization could be weaponized, with enormous consequences for fisheries. One could even devastate fishing areas by creating dead zones. R063 agreed and also noted that “ocean fertilization can create dead zones, so economic disruption to fishing stocks and the protein source for many coastal countries could be jeopardized.” Others spoke about the conversion of climate technologies into weapons that could affect rainfall, kill off agriculture of affect crops, degrade forests, or interfere with water security. R034 stated that: Fig. 3. A hypothetical deployment of a militarized U.S. “Earth Force” to enforce climate targets and deter carbon emissions. Source: Stuart Candy (used with permission). scales of future battlefields marginally towards victory. 4.3.2. Cyberattacks and information warfare Our participants mentioned various risks related to security and hacking for control, command, communications, and software, including ransomware. R034 put it this way: EnergyStrategyReviews45(2023)10103110 B.K. Sovacool et al. Many of the solar geoengineering options raise the same concerns about weather modification or rainfall modification. There is a strong risk of effects to crops and food security, because of changes in solar radiation. R047 even spoke about fears that “Chinese actors will use solar geoengineering to intentionally disrupt the monsoon to harm India.” R104 also noted that many negative-emissions and solar geoengineering options could give states “the ability to specifically target and disrupt other ecosystems, including forests.” As R104 went on to explain, centralized systems and large-scale systems would aggravate this particular security risk: Yes, so the ability to impact other people’s ecosystems in targeted or general ways goes up the more you do atmospheric management systems, whether it’s because you’re going to see things or you’re going to unsee things, so risks of conflict over rainfall, as terrestrial water systems get overbuilt, go up. So, there are a lot of built-in risks to the more centralised system, whether it is a source of disruption or a target. This statement also confirms that deployment could threaten corol- lary systems related to things like water supply. R007 added that “if stratospheric nanoparticles can direct sunlight into particular areas, this can become a very serious security issue or weapon.” R035 spoke about how recent research advances in “stratosphere-troposphere coupling” could enable a military to “change the positions of the weather patterns and the jet streams and the rest of it, by affecting the circulations of radiation or ozone in the stratosphere.” Other respondents spoke about why climate technologies would make optimal weapons, with arguments grounded in their ability to be rapidly deployed and cheaply produced. R002 suggested that some op- tions could be deployed in days rather than months: “cirrus cloud thinning and cloud brightening can change weather in a matter of days, you could equip 500 ships to do it very quickly.” R047 argued wea- ponization could be done cheaply as well, with a crude military program costing only $10 billion a year, so cheap “even Bangladesh or a Belarus can afford it.” R024 also expressed concern these options would make “fast, dirty, and cheap” weapons. In terms of governance and the likely repercussions of hostile use, R099 added that should such deployment occur, there is little in terms of governance or international control to stop it: In terms of weaponization, there is no mechanism to deal with such conflicts, should they arise. The United Nations General Assembly and climate convention are all very weak, majority-based, consensus-based approaches with no binding power. The Security Council is incompetent and has zero legitimacy where it matters. The international community is absolutely unfit to deal with these sorts of security challenges. The 1976 Convention on the Prohibition of Military or Any Other Hostile Use of Environmental Modification Techniques (ENMOD) expressly forbids weather modification as a weapon, but Horton and Reynolds write that it is unlikely to stop or prohibit hostile states from using it if they wanted to [96]. R106 expressed concern that “if you are going to suffer from climate change, and you have only one chance to improve your situation by weaponizing the technology, you won’t care if the international community accepts it.” Deployment among populist or authoritarian states is an added driver to this concern, as such regimes are unlikely to worry about international-relations repercussions and they would not be “contained” in the way many democratic states would be [97]. R106 expanded on this theme, and stated that: The Chinese regime could readily resort to solar geoengineering technologies, given their longstanding experience now with weather modification on smaller scales. They think that the Olympics in 2008 were a success, in terms of weather modification. It eventually boils down to the calculation of the leader, whether the direct risk of a climate change impact to himself or herself is higher than the un- certain risk that eventually will come. And of course, again, the question is, how ruthless is the leader? I am now getting into really historical [comparisons] that may be disturbing, but Hitler said, when the war was clearly lost, “Now I am willing to go down and also destroy the whole country.” Well, we have seen authoritarian leaders who are willing to tolerate a lot of collateral damage. And they may be willing to deploy solar geoengineering to do the same. R007 iterated how they believed “the Indians, Russians, and Chinese are thinking about weaponization of these options, and they are not covered by the same legal constraints as the United States or the ENMOD treaty … I am also sure a regime like North Korea would love it.” Some respondents spoke not about specific applications as weapons, but grander, larger security concerns such as shifting geopolitical trends and even the risk of great power wars. R045 said that: Deployment could lead to real conflict between nations. There are inherent political and strategic risks from deployment. There is the very profound risk of geopolitical conflict arising from any nation or group of nations being able to control the global thermostat or the climate system and placing control in their own hands and no one else. This is an enormous danger to the international world order. All sorts of conflicts could emerge between major powers, e.g. Russia spraying sulphate aerosols in the Arctic, the USA responds danger- ously. Or unexpected environmental impacts, like the shift of the Indian monsoon, another Cuban Missile crisis could emerge over sulphate aerosols around the world. That is a most profound risk. R071 even argued that these climate technologies are as dangerous as nuclear weapons and could even result in nuclear war, noting: “once you put your hand on the thermostat and can control the global ther- mostat, this can lead to the real risk of military conflict. It could lead to people fighting with nuclear weapons over who does it.” 4.4.2. Insecurity and asymmetrical protection This aspect captures insecurity via asymmetries in the levels of climate protection that are feasible or affordable across countries and which could thus propagate in internal or regional political instability. R113 spoke about how negative-emissions development would likely strain existing land-based systems for food, forests, and water, gener- ating internal insecurity: Yes, absolutely, there are negative impacts domestically. So for bioenergy with carbon capture and storage, you grow crops, you don’t have access to land for other purposes, you lose capacity for development, whatever it may be, as well as food, you affect water supplies, you have trade-offs with crop growth. Afforestation can lead to changes in precipitation patterns, and you have trade-offs then with decisions about location of carbon dioxide removal in- vestment infrastructure against changes in weather patterns locally, which is a really big problem. R111 also spoke about the risk of deployment causing “cascading impacts” to local ecosystems wherever options are deployed: If a country will hypothetically reduce temperature and deploy op- tions at large-scale, we don’t know what is going to be the impact on humans or ecosystem or animals or plants. If we are going to change rainfall patterns, we don’t know how it will impact agriculture. There might be a region where there is normal rainfall and we do sun shields or aerosol injection, and then it is going to change rainfall patterns … There are so many known unknowns. R112 agreed and also noted multiple ways that climate technologies could aggravate stress to land and water resources, with spill over im- pacts that are “huge:” EnergyStrategyReviews45(2023)10103111 B.K. Sovacool et al. The negative effects on food and food security, on acidification of water, on ecosystems, particularly rivers, on food supply, on fresh- water fish, could be serious … And if you imagine that something like that occurring in a geopolitical hotspot like the River Jordan, if the River Jordan is affected upstream, the implications politically in terms of security downstream are huge, absolutely huge. R028 noted how some governments are already deploying pilot projects to enhance the weather, pointing to “Sky River”: That’s China’s program to do weather modification to create rain over inland China. It has massive transboundary issues. When anyone builds a dam upstream, people downstream care about it, so you can imagine how much insecurity this program can create. R076 added that even nature-based efforts like afforestation pro- grams could present security risks when done at scale; they suggested that “things like the Trillion Trees initiative have a serious risk of ignoring the complexity of ecosystems and disrupting a host of natural systems all for the purpose of sinking carbon … it could be destabilizing to many communities.” Indeed, in their own assessment Horton and Reynolds concluded that “even if militarization and direct conflict are unlikely, the indirect effects of climate engineering could operate as drivers of existing conflicts.” [98]. 4.4.3. Miscalculation Hostile deployment may not need to be “real” in order for it to spawn conflict; it could instead result due to misperceptions of what has occurred or be prone to miscalculation. R001 articulated this extremely well under the term of “indirect weaponization” or “political weaponization:” The political risks of deployment are very hard to predict, it depends on who is doing it, how it is being done, and who is deciding. It might not even need to be done to be a risk, to procreate a geopolitical conflict. A country could accuse their neighbour of doing strategic aerosol injection if their rainfall patterns have changed, or they saw some suspicious aircraft activity, or the signal-to-noise ratio for un- derstanding atmospheric dynamics seemed different. It is hard to say with certainty and a short period of time whether one has done it, which means it could be used politically even without actually being used; a ploy, or an indirect or political weapon in that sense. It doesn’t have to be technically feasible to have a political effect as a signalling effect, to feed into a process of political weaponization. The fact that many options are believed to be cheap, fast, and easy only makes this political threat more credible. R047 framed the risk in terms of “inadvertent weaponization” or “psychological weaponization:” It’s possible that you could inadvertently weaponize these options. So you start a programme, and then there’s a rapid warming that happens when you stop doing that stuff. In the termination-risk world, you could accidentally weaponize something, because you stopped paying attention to it. You could accidentally allow your virus to sneak out of your lab, because you weren’t paying attention to that problem. So in that sense, it’s not so much weaponization as just being an idiot, and then accidentally causing massive harm to the world. The perception of weaponization, though, is more inter- esting. If we knew there was a fleet of Chinese or Russian aircraft flying around, modifying the climate, that’s psychological weapon- ization. It could make everyone feel vulnerable. These issues of weaponization or a termination shock have even permeated recent popular literature, with Neal Stephenson’s Termina- tion Shock and Gwynne Dwyer’s Climate Wars coming to mind. R007 added that misperception could involve not only deployment, but intent to deploy, or concentrated research in an area, all which could feed “conspiracy theories” that lead to uncertainty: What I tend to write about are the security risks, and I don’t mean security risks in terms of militaries necessarily going to war with one another over these technologies, but more in terms of the unpre- dictability of impacts, the difficulty in attributing changes in envi- ronmental systems to particular actions, and then the whole background of misinformation, conspiracies, and what not. Which means that any country that tries to deploy these technologies, and especially the more space-age technology it is, the more uncertainty there is going to be, at least in terms of social acceptability and conspiracies. Say you put up a sunshade or some sort of mirror, and then something else happens, somewhere in the world, and then people are going to attribute it to that; they’re going to draw a direct line between A and W, regardless of how many points are in between. R020 spoke about how such miscalculation could even lead to great power wars (having the same effect as weaponized deployment in 4.4.1, but without the intent): Imagine that stratospheric aerosol injection takes place, especially if it’s a unilateral kind and then the monsoons in India are failing, even if you can’t actually prove that is because of geoengineering, the perception is enough to create the Third World War, especially if it’s China doing the geoengineering and India is suffering. R064 similarly envisioned a scenario where: India decides to start solar geoengineering, a few years later major floods occur in Pakistan. Pakistani politicians blame these on India. Whether physically plausible or attributable doesn’t matter, geopo- litical rivals can use it in a way that sparks or intensifies conflict. In such situations, determining causation or responsibility wouldn’t matter, as “victimized states” could cast blame regardless of whether they actually believed someone was at fault, even if just as a means to distract the public and promote a “rally round the flag” response. And in response, they could still posture and demand compensation, retaliate with sanctions, or even attack soft targets [99]. 4.4.4. Escalation Our last mechanism for potential conflict via climate technology concerns the risk of escalation via arms races, counter-geoengineering, and proxy wars. R064 identified this risk as weaponization via “arms races” or “technology races," noting that “types of adversarial conflict could also emerge like an arms race of accelerated deployment.” R098 captured this risk eloquently: With geoengineering, countries could say “Others are creating a risky political climate, and then we have to respond that.” That’s the way we justified our massive infrastructure on biological and chemical weapons. We just said, “Oh, the Soviets are doing it. We have to prepare for that.” That’s why we responded with this kind of research … and a "let’s see what happens" attitude. Development or deployment could generate a competitive political dynamic that incentivizes countries to all match or even exceed rivals’ funding efforts for negative emissions or solar geoengineering, resulting in a “capacity race” [100]. An additional dynamic within this risk has been termed “counter- geoengineering,” whereby countries start to develop technical capabil- ities to modify the climate or deploy their own geoengineering, in order to counteract the unwanted activities of others and with the core stra- tegic intent of stopping it. R022 explained this logic as follows: The idea of counter-geoengineering is that even if you didn’t want a neighbour to do it, you develop the capabilities to do it. You might tell me: “Look, you start putting up sulphates and I’m going to release loads of difluoromethane." If you’re able to credibly threaten to counter-geoengineer, that gives you a veto over my deployment. Therefore, no-one can unilaterally do solar geoengineering until there’s an agreement amongst everyone who has the power to EnergyStrategyReviews45(2023)10103112 B.K. Sovacool et al. counter-geoengineer.… from a game-theoretic perspective it is plausible that, if you could create counter-geoengineering, then it would mean no unilateralism and enforced negotiation over its use. R071 similarly hypothesized how a country could threaten to release ozone-depleting substances to stop another state from deploying aerosols: Well, a state could counter aerosol deployment by responding with chlorofluorocarbons. Chlorofluorocarbons are cheap to produce and they’re really efficient in heating the system, but of course now you’re really setting up your own control on the climate system. Their intervention could be focused on climate forcing, a state could respond with counter-forcing: a logical response in my view. R074 spoke about how such actions will likely lead to multiple non- cooperative and detrimental games (in the game-theoretic sense) among state actors: The non-cooperative game potential is terrible, because you’re just deploying in two different directions and all the different side effects, they will just accumulate. And in the end, we will not see a lot of cooling or changes to the climate, but we will see a lot of military conflict, a lot of, maybe, environmental side effects, and so on, and so on. If it becomes a game purely of countries not talking to each other and being just adversaries, then counter-engineering would be pretty bad … It’s definitely the case that counter-engineering would in- crease the fragility of the whole system, because if cooperation then breaks down at some point in time, well, we don’t want to see that world happening. R091 also spoke about how counter-geoengineering not only creates destabilizing political effects but wastes resources and leads to sub- optimal outcomes as well, as “counter geoengineering sees a lot of money being spent with very limited effects on the temperature and a lot of other negative physical effects of these technologies being deployed and nothing happening.” They also noted that such counter-engineering could occur the moment any country initiates large-scale deployment: “the moment that you have deployment, especially without global agreement, all of a sudden counter-geoengineering becomes very rele- vant and very important.” As a final aspect of this risk, two respondents (R044, R081) mentioned how the revenues from geoengineering could enable actors to be aggressive, to instigate wars or finance the waging of war, with the proceeds or financial gains made from investing in climate technology. 5. Counterpoints: de-securitization, threat deflation, and permissive tolerance Much of the perspectives above discuss the threat of climate tech- nologies and reveal how climate action can be securitized or contribute to arms races and militarization. But there are also strong undercurrents throughout our interview data downplaying the threats posed by these technologies, seeking to de-securitize the discussion. Some data also elaborates on the likelihood that states will simultaneously support and oppose deployment in different constituencies and to different ends, depending on how such actions fit within their agendas. 5.1. De-securitization and threat deflation Some of our experts, rather than constructing climate technologies as a security threat or potential weapon, argued the opposite, believing that such technologies would make poor weapons and would contribute little to militarization. R001 spoke about how direct weaponization of solar geoengineering would depend on predictability, provability, pre- cision and control, while emphasizing how these would not be likely with negative emissions technologies: Direct weaponization is unlikely in my view, given the unpredict- ability of these options, and also provability, lack of precision, to show that it’s been done is difficult. In short: these technologies are not good enough to be used as reliable weapons. Which leads me to then ask: If the military cannot depend on them for a climate weapon, can we depend on them for climate protection? R023 added that: “I realize some people are concerned about mili- tarization, but I believe there are much easier ways of conducting war.” R042 also stated that: If I am a rogue state, and I want to do something bad to another one, I would use digital technologies, why would you even need negative emissions or solar geoengineering? There are so many available cheaper, more proven, cost-effective conventional weapons. I am not so worried about the geopolitical risks to these nascent and unproven climate technologies. R071 stated that: “I don’t want to think as a terrorist, but there are other ways to mess up society much faster than trying to geoengineer the climate; terrorists have much nastier things at their disposal.” R085 agreed and noted that “strategic aerosol injection makes for a lousy weapon.” R097 added that: I’m pretty firmly on the side of the fact that it’s very, very hard to weaponize. It’s not targetable, it’s not discriminative, you can’t aim it, as far as I understand. Now, that’s not to say that there wouldn’t be harms from one country to another. If India did it, it could have harms on China or vice versa, certainly. But as a weapon, it’s extremely crude. There are much better weapons to use, much more targeted weapons, if you wanted to harm somebody. Certainly, you can use the environment as a weapon. The US did it in Vietnam, and Saddam Hussein did it in the first Gulf War. You could break a dam. These would all make more effective weapons than climate technologies. One theme in our data was the difficulty of using negative emissions technologies as weapons. R047 noted that negative emissions technol- ogies “do not seem weaponizable; nobody is trembling in their boots.” R064 concurred when they noted that “carbon dioxide removal doesn’t have any weaponization issues, partially because it’s so slow, even if they were used in an India-China war, freezing CO2 out of the atmo- sphere: this would be a decade-long project, making it implausible as a weapon.” One of the key challenges to using solar geoengineering specifically as a weapon was lack of precision in its targeting and accuracy in use. R047 explained it this way: One has to distinguish between perception and reality. Reality is, it would be very hard to weaponize the technology, because solar geoengineering as we understand it is essentially a zonal play. So you’re putting particles into the stratosphere. Then, in the strato- sphere, they’re mixed zonally, so within a 5 or 10-degree latitude band, roughly at altitude. Then they slowly diffuse out of the zone, such that they cover an entire hemisphere in a year or so. Then they do inter-hemisphere transfer within 2 or 3 years. The problem is, by the time you’re getting to inter-hemisphere transfer, your particles are already raining out, so you need to continually refresh. What happens when you do this is, you overweight the injection of parti- cles into your zone. Wherever your aircraft have access, you have this 5 or 10-degree range of where they cause damage. This is a very imprecise weapon … So you could deploy these things, and then they would zonally mix. But that means that weaponizing the system is pretty hard, because the effects are zonally distributed, and you’re going to get hurt, or plausibly, as hurt. The last part of this statement is telling because it also indicates that weaponization of solar geoengineering could cause as much harm to the country initiating the attack as to the one receiving it. R064 also spoke EnergyStrategyReviews45(2023)10103113 B.K. Sovacool et al. about difficulty in targeting for such systems. In their view: Weaponization not plausible, they are too poorly targeted. The basic problem for stratospheric aerosol injection is that it is close to impossible to perfect the targeting, hemispheric or even global deployment. If you are a Great Power X, and you want to use it to harm geopolitical rival or enemy Y, maybe you get some sort of serious effect for them, but you will have all sorts of other effects on other countries, including allies, and even yourself. Militaries are moving to precision weapons: this isn’t that; this is a blunt weapon that would just widen any conflict. It may very well be less likely that negative emissions and solar geoengineering technologies will be used in warfare, but more likely that these technologies will be used to restrict the development of related industries in other countries. For example, in the future, after solar photovoltaic systems could come to occupy a large proportion of the power generation structure, the power system can be destroyed by destroying the photovoltaic system, and thus provide an advantage in warfare. Therefore, the role of these new climate technologies in warfare is perceived by some experts as more indirect than as a direct weapon. 5.2. Permissive tolerance, strategic ambiguity and brinksmanship A second complicating factor emerging from our data was that states would be more strategic about simultaneously supporting and restrict- ing the military applications of climate-engineering technologies, a sort of “permissive tolerance” where actors would complain if something went wrong but would openly support options that went right. R098 articulated that: The most realistic political response that I see involves a lot of permissive tolerance. This refers to an attitude of “Let others do stuff, and then just wait and see.” Officially, you can always complain but you’re kind of happy that somebody else is taking the blame. And let’s see what happens. The potential for permissive tolerance and brinkmanship compli- cates the security risks posed in Section 4, as it implies states could pursue a more complex pathway of strategic ambiguity. They could discretely pursue contradictory aims and ends with their technologies, or publicly assume different stances when it comes to the use and deployment of said technologies. 6. Conclusion Although the future impacts of climate change on global security remain uncertain, one certainty is that the deployment of climate- engineering options could have profound security implications for states, regions, and even the global political system. As summarized in Table 5, low-carbon technologies could be used as military negotiating tools, as mechanisms to build military capacity or secure resources and supply chains, as physical or cyber targets in ongoing conflicts, or as major causes of new conflicts arising from direct weaponization, direct and adverse impacts on insecurity, or even the risks of miscalculation or escalation via counter-geoengineering. Our expert-interview exercise with leading thinkers on the topic revealed how climate technologies can potentially propagate very different types of conflict at different scales and among diverse political actors. Conflict and war could be pursued intentionally (direct targeted deployment, especially weather-modification efforts targeting key re- sources such as fishing, agriculture, or forests) or result accidently (unintended collateral damage during existing conflicts or even owing to miscalculation). Conflict could be over material resources (mines or technology supply chains) or even immaterial resources (patents, soft- ware, control systems prone to hacking). The protagonists of conflict could be unilateral (a state, a populist leader, a billionaire) or multi- lateral in nature (via cartels and clubs, a new “Green OPEC”). Research and deployment could exacerbate ongoing instability and conflict, or cause and contribute to entirely new conflicts. Militarization could be over perceptions of unauthorized or destabilizing deployment (India worrying that China has utilized it to affect the monsoon cycle), or to enforce deployment or deter noncompliance (militaries sent in to protect carbon reservoirs or large-scale afforestation or ecosystem projects). Conflict potential could involve a catastrophic, one-off event such as a great power war or nuclear war, or instead a more chronic and recurring series of events, such as heightening tensions in the global political system to the point of miscalculation, counter-geoengineering, permis- sive tolerance and brinksmanship. Moreover, our findings point the way towards fruitful research di- rections. In this study, we interviewed a broad sample of experts about the military and security issues arising with negative emissions and geoengineering. But future work could more directly engage with mili- tary and security practitioners themselves. Many of the 20 technologies we examine within the GENIE project are still immature, and as a result the scale of their application will change greatly in the future, along with their potential impact on warfare. Future analysis should explore vary- ing pathways through which the different technologies could impact warfare, at different time stages, and refine those scenarios as more knowledge is accumulated and uncertainty diminishes. Nevertheless, the varied and compelling ways in which climate- technology deployment links to multiple dimensions of security and conflict strongly suggest that the topic deserves far greater attention within the political science, geopolitics, and international relations lit- eratures. Particularly, there is a strong need for research on the pro- spective geopolitics of solar geoengineering and negative emissions. But it also unveils the myriad ways in which a net-zero, carbon-neutral world could be more politically destabilized and geopolitically insecure than our current (already unstable) world order. States and actors will need to proceed even more cautiously in the future if they are to avoid making these predictions into reality, and more effective governance Table 5 Summary of the geopolitical and international security dimensions of negative emissions and solar geoengineering technologies. Negotiating tools Enhanced military capacity Targets in ongoing conflict Cause of conflict or weaponization Negative emissions technologies Formation of carbon dioxide removal clubs, risk of a “Green OPEC”, heightened diplomatic conflict Solar geoengineering techniques Pre-emptive deployment by a Greenfinger, formation of geoengineering clubs, use of “energy weapon” or “weather modification” to get concessions Source: Authors Coupling with military industrial complex, resource curse over forestry or oceans, military enforcement of climate targets Risk of carbon reservoirs being targeted by terrorists or geopolitical blackmail by countries storing carbon Use of ocean techniques to devastate fisheries, or land-based techniques to devastate forests or agriculture, ability to control global thermostat, strengthening of authoritarian or populist regimes, internal insecurity due to cascading impacts Augmenting aerospace or space capacity, crossover high- technology skills, protection of military bases Technology (aircraft, ships, balloons) and programs could be targeted as critical infrastructure, cyberattacks on control systems Interference with rainfall, monsoons, sunlight, and ecosystems, risk of miscalculation over deployment aims, cycles of counter-geoengineering EnergyStrategyReviews45(2023)10103114 B.K. Sovacool et al. architectures may be warranted to constrain rather than enable deployment, particularly in cases that might lead to spiralling, retalia- tory developments toward greater conflict. After all, to address the wicked problem of climate change while creating more pernicious po- litical problems that damage our collective security is a future we must avoid. Credit author statement All authors contributed equally to: Conceptualization; Data curation; Formal analysis; Funding acquisition; Investigation; Methodology; Project administration; Resources; Software; Supervision; Validation; Visualization; Roles/Writing - original draft; Writing - review & editing. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Data availability The data that has been used is confidential. Acknowledgements This project has received funding from the European Union’s Hori- zon 2020 research and innovation programme under the European Research Council (ERC) Grant Agreement No. 951542-GENIE-ERC- 2020-SyG, “GeoEngineering and NegatIve Emissions pathways in Europe” (GENIE). The content of this deliverable does not reflect the official opinion of the European Union. Responsibility for the informa- tion and views expressed herein lies entirely with the author(s). Appendix A. Supplementary data Supplementary data to this article can be found online at https://doi. org/10.1016/j.esr.2022.101031. References [1] Eileen Claussen, Janet Peace, Energy myth twelve – climate policy will bankrupt the U.S. Economy, in: B.K. Sovacool, M.A. Brown (Eds.), Energy and American Society, Springer Publishing Company, New York, 2007. [2] B. Sudhakara Reddy, Gaudenz B. Assenza, The great climate debate, Energy Pol. 37 (2009) 2997–3008. [3] Joshua W. Busby, Todd G. Smith, L. Kaiba, White, M. Shawn, Strange, “climate change and insecurity: mapping vulnerability in Africa, Int. Secur. 37 (No. 4) (2013) 132–172. [4] Economics of Climate Adaptation Working Group, Shaping Climate-Resilient Development: A Framework for Decision-Making, Climate Works Foundation, New York, 2009. [5] Bianca Jagger, "The threat of a global climate disaster is No longer up for debate," Testimony Before the House Select Committee for Energy Independence and Global Warming (March 6, 2008). [6] A.P. Sokolov, P.H. Stone, C.E. Forest, R. Prinn, M.C. Sarofim, M. Webster, S. Paltsev, C.A. Schlosser, D. Kicklighter, S. Dutkiewicz, J. Reilly, C. Wang, B. Felzer, J.M. Melillo, H.D. Jacoby, Probabilistic forecast for twenty-first-century climate based on uncertainties in emissions (without policy) and climate parameters, J. Clim. 22 (19) (2009) 5175–5204. [7] CNA Military Advisory Board, National Security and the Threat of Climate Change, 2007 (Available at: SecurityAndClimate.cna.org). [8] Elizabeth L. Chalecki, Lisa L. Ferrari, A new security framework for geoengineering, Strategic Studies Quarterly 12 (2) (2018) 82–106. [9] Michael E. Mann, The New Climate War: the Fight to Take Back Our Planet, Public Affairs, New York, 2021. [10] Jennie C. Stephens, David W. Keith, Assessing geochemical carbon management, Climatic Change 90 (2008), https://doi.org/10.1007/s10584-008-9440-y. [11] EASAC, Negative Emission Technologies: what Role in Meeting Paris Agreement Targets? EASAC Policy Report 35, February 2018. [12] Oscar Rueda, Jose M. Mogoll´on, Tukker Arnold, Laura Scherer, Negative- ◦ emissions technology portfolios to meet the 1.5 (2021), 102238. C target, Global Environ. Change [13] Felix Schenuit, Rebecca Colvin, Matthias Fridahl, Barry McMullin, Andy Reisinger, Daniel L. Sanchez, Stephen M. Smith, Asbjørn torvanger, anita wreford and oliver geden, “carbon dioxide removal policy in the making: assessing developments in 9 OECD cases, Frontiers in Climate 3 (2021), 638805. [14] Thomas Gasser, Celine Guivarch, Kaoru Tachiiri, Chris D. Jones, Philippe Ciais, Negative emissions physically needed to keep global warming below 2C, Nat. Commun. 6 (2015) 7958, https://doi.org/10.1038/ncomms8958. [15] Sean Low, Miranda Boettcher, Delaying Decarbonization, Climate governmentalities and sociotechnical strategies from Copenhagen to Paris, Earth System Governance 5 (2020), 100073. [16] Sikina Jinnah, Simon Nicholson, Introduction to the symposium on ‘geoengineering: governing solar radiation management, Environ. Polit. 28 (No. 3) (2019) 385–396. [17] Simon Nicholson, Sikina Jinnah, Alexander Gillespie, Solar radiation management: a proposal for immediate polycentric governance, Clim. Pol. 18 (3) (2018) 322–334. [18] Joshua B. Horton, Jesse L. Reynolds, The international politics of climate engineering: a review and prospectus for international relations, Int. Stud. Rev. 18 (2016) 438–461. [19] Achim Maas, Jürgen Scheffran, Climate Conflicts 2.0? Climate Engineering as a Challenge for International Peace and Security, Sicherheit und Frieden, 2012, pp. 193–200. [20] Chad M. Briggs, Is solar geoengineering a national security risk? in: Jason J. Blackstock, S. Sean Low (Eds.), Geoengineering Our Climate? Ethics, Politics and Governance Earthscan from Routledge, London, 2018, pp. 178–181. [21] Olaf Corry, The international politics of geoengineering: the feasibility of plan B for tackling climate change, Secur. Dialog. 48 (No. 4) (2017) 297–315. [22] McLaren Duncan, Olaf Corry, The Politics and Governance of Research into Solar Geoengineering, Wiley Interdisciplinary Reviews Climate Change, 2021, p. e707. [23] Morgan Bazilian, Model and manage the changing geopolitics of energy, Nature 569 (2019) 29–31. [24] Jon Barnett, “Security and climate change”, Global Environ. Change 13 (No. 1) (2003) 7–17. [25] Chad M Briggs, Climate security, risk assessment, and military planning, Int. Aff. 88 (No. 5) (2012) 1049–1064. [26] Jürgen Scheffran, Antonella Battagliani, Climate and conflicts: the security risks of global warming, Reg. Environ. Change 11 (No. 1) (2011) 27–39. [27] Katherine J. Mach, W. Neil Adger, Halvard Buhaug, Marshall Burke, D. James, Fearon, B. Christopher, Field et al, “Directions for Research on Climate and Conflict, Earth’s Future 8 (2020), e2020EF001532, https://doi.org/10.1029/ 2020EF001532. [28] Lester R. Brown, Redefining National Security, 1977. Worldwatch Paper No. 14. [29] Sharon Burke, “Natural Security”, Center for a New American Security (2009). [30] CNA Military Advisory Board, National Security and the Threat of Climate Change, 2007 (Available at: SecurityAndClimate.cna.org). [31] Kurt M. Campbell, Jay Gulledge, J.R. McNeill, John Podesta, Peter Ogden, Leon Fuerth, R. James Woolsey, T. Alexander, J. Lennon, Julianne Smith, Richard Weitz, Derek Mix, The Age of Consequences: the Foreign Policy and National Security Implications of Global Climate Change, Center for a New American Security, 2007. [32] Simon Dalby, Rethinking Geopolitics, Climate Security in the Anthropocene” Global Policy 5 (No. 1) (2014) 1–9. [33] Thijs van de Graaf, K. Benjamin, Sovacool. Global Energy Politics, Polity Press, Oxford, 2020. [34] Lisa Murkowski, John McCain, Bob Corker, “The US needs to end its ban on crude oil exports,” Foreign Pol. (April 14, 2015). [35] Rose McDermott, Risk-Taking in International Politics: Prospect Theory in American Foreign Policy, University of Michigan Press, Ann Arbor, MI, 1998. [36] Mark Leonard (Ed.), Connectivity Wars: Why Migration, Finance and Trade Are the Geo-Economic Battlegrounds of the Future, European Council on Foreign Relations, Berlin, 2016. [37] Robert D. Blackwill, Jennifer M. Harris, War by Other Means: Geoeconomics and Statecraft, Harvard University Press, Cambridge, MA, 2016. [38] Per H¨ogselius, Energy and Geopolitics, Routledge, New York, 2019. [39] Karen Smith Stegen, Deconstructing the “energy weapon”: Russia’s threat to Europe as case study, Energy Pol. 39 (10) (2011) 6505–6513. [40] James Henderson, Does Russia have a potent gas weapon? in: Thijs van de Graaf, et al. (Eds.), The Palgrave Handbook of the International Political Economy of Energy Palgrave Macmillan, London, 2016, pp. 461–486. [41] James Omonbude Ekpen, Cross-Border Oil and Gas Pipelines and the Role of the Transit Country: Economics, Challenges and Solutions, Palgrave Macmillan, Basingstoke, 2016. [42] Paul Stevens, Transit Troubles: Pipelines as a Source of Conflict, Chatham House, London, 2009. [43] Jaap Rodenburg, Richard Hengeveld, Embargo: Apartheid’s Oil Secrets Revealed, Amsterdam University Press, Amsterdam, 1995. [44] Neta C. Crawford, Oil sanctions against apartheid, in: Neta C. Crawford, Audie Klotz (Eds.), How Sanctions Work, Palgrave Macmillan, London, 1999, pp. 103–126. [45] Carlos Pascual, The New Geopolitics of Energy, Center on Global Energy Policy, Columbia University, 2015. [46] BBC World News. Belarus threatens to cut off gas to EU in border row. November 12, 2021. [47] Jeff D. Colgan, Fueling the fire: pathways from oil to war, Int. Secur. 38 (2) (2013), 147–80. EnergyStrategyReviews45(2023)10103115 B.K. Sovacool et al. [48] Jeff D. Colgan, Oil and revolutionary governments: fuel for international conflict, Int. Organ. 64 (4) (2010) 661–694. simulations of designing stratospheric sulfate aerosol geoengineering to meet multiple simultaneous climate objectives, J. Geophys. Res. Atmos. 122 (2017). [49] Jeff D. Colgan, Oil, domestic politics, and international conflict, Energy Res. [76] Ben Kravitz, Douglas G. MacMartin, Hailong Wang, Philip J. Rasch, Social Sci. 1 (2014) 198–205. [50] Jeff D. Colgan, Petro-Aggression: when Oil Causes War, Cambridge University Press, Cambridge, 2013. [51] Daniel Moran, James A. Russell (Eds.), Energy Security and Global Politics: the Militarization of Resource Management, Routledge, London, 2008, p. 10, 2008. [52] Quy-Toan Do, Jacob N. Shapiro, Christopher D. Elvidge, Mohamed Abdel-Jelil, Daniel P. Ahn, Kimberly Baugh, Jamie Hansen-Lewis, Mikhail Zhizhin, Morgan D. Bazilian, Terrorism, geopolitics, and oil security: using remote sensing to estimate oil production of the Islamic State, Energy Res. Social Sci. 44 (2018) 411–418. [53] Colgan, “Fueling the Fire: Pathways from Oil to War.”. [54] Benjamin K. Sovacool, G¨otz Walter, Internationalizing the political economy of hydroelectricity: security, development, and sustainability in hydropower states, Rev. Int. Polit. Econ. 26 (1) (2019) 49–79. Winter. [55] Colgan, “Fueling the Fire: Pathways from Oil to War.”. [56] Andr´e Månsson, Energy, conflict and war: towards a conceptual framework, Energy Res. Soc. Sci. 4 (2014) 106–116. [57] Andr´e Månsson, Energy, conflict and war: towards a conceptual framework, Energy Res. Soc. Sci. 4 (2014) 106–116. [58] Judith Kreuter and Markus Lederer, “The Geopolitics of Negative Emissions Technologies – Learning lessons from REDD+ and Renewable Energies for Afforestation, BECCS and Direct Air Capture,” Global Sustain. 4, e26, 1–14. https://doi.org/10.1017/sus.2021.24. [59] Felix Creutzig, Christian Breyer, Hilaire Jerome, Minx Jan, P. Glen, Peters and Robert Socolow, “The mutual dependence of negative emission technologies and energy systems, Energy Environ. Sci. 12 (2019) 1805–1817. [60] J. Fuhrman J, Haewon McJeon, Pralit Patel, C. Doney Scott, William M. Shobe, Andres F. Clarens, Food-energy-water implications of negative emissions technologies in a +1.5 C, Future Nat. Climate Change 10 (2020) 920–927. [61] Kreuter and Lederer, “The Geopolitics of Negative Emissions Technologies – ◦ Learning Lessons from REDD+ and Renewable Energies for Afforestation, BECCS and Direct Air Capture.”. [62] Jessica Spijkers, et al., Exploring the future of fishery conflict through narrative scenarios, One Earth 4 (2021) 386–396. Geoengineering as a design problem, Earth Syst. Dynam. 7 (2016) 469–497. [77] Corry, “The International Politics of Geoengineering”. [78] Sean Low, Matthias Honegger, A Precautionary Assessment of Systemic Projections and Promises from Sunlight Reflection and Carbon Removal Modeling, Risk Analysis, 2020, pp. 1–15, https://doi.org/10.1111/risa.13565. [79] Sean Low, Chad M. Baum, Benjamin K. Sovacool, Undone science in climate interventions: contrasting and contesting anticipatory assessments by expert networks, Environ. Sci. Pol. 137 (2022) 249–270. [80] David W. Keith, Towards constructive disagreement about geoengineering, Science 374 (No. 6569) (2021) 812–815. [81] Florian Rabitz, Going rogue? Scenarios for unilateral geoengineering, Futures 84 (2016) 98–107. [82] Andy Parker, Joshua B. Horton, David W. Keith, Stopping solar geoengineering through technical means: a preliminary assessment of counter geoengineering, Earth’s Future 6 (2018) 1058–1065. [83] Jesse L. Reynolds, Gernot Wagner, Highly Decentralized Solar Geoengineering, Environmental Politics, 2019, https://doi.org/10.1080/ 09644016.2019.1648169. [84] Axel Michaelowa, Solar radiation modification-A “silver bullet” climate policy for populist and authoritarian regimes? Global Policy 12 (2021) 119–128. [85] Robert N. Stavins, Robert C. Stowe (Eds.), Governance of the Deployment of Solar Geoengineering, Harvard Project on Climate Agreements, Cambridge, Mass, 2019. [86] Anthony Harding, Juan Moreno-Cruz, Solar geoengineering economics: from incredible to inevitable and back again, Earth’s Future 4 (No. 12) (2016) 569–577. [87] B. Joshua, Horton, ”evaluating solar geoengineering deployment scenarios, in: Robert N. Stavins, Robert C. Stowe (Eds.), Governance of the Deployment of Solar Geoengineering. (Cambridge, Mass, Harvard Project on Climate Agreements, 2019. [88] Corry, “The International Politics of Geoengineering”. [89] Chad M Briggs, Mirjam Matejova, Disaster Security: Using Intelligence and Military Planning for Energy and Environmental Risks, Cambridge University Press, Cambridge, 2019. [63] Holly J. Buck, Rapid scale-up of negative emissions technologies: social barriers [90] Timothy Mitchell, Carbon Democracy: Political Power in the Age of Oil, Verso, and social implications, Climatic Change 139 (2016) 155–167. London, 2011. [64] Wim Carton, Adeniyi Asiyanbi, Silke Beck, Holly J. Buck, Jens F. Lund, Negative Emissions and the Long History of Carbon Removal, 11 e671, Wiley Interdisciplinary Reviews Climate Change, 2020. [65] Peter Healey, Robert Scholes, Penehuro Lefale, Pius Yanda, Governing net zero carbon removals to avoid entrenching inequities, Frontiers in Climate 3 (2021) 1–6. Article 672357. [66] Andreas Malm, Wim Carton, Seize the means of carbon removal: the political economy of direct air capture, Hist. Mater. 28 (No. 1) (2020) 3–48. [67] Miranda Boettcher, Kerryn Brent, Holly J. Buck, Sean low, duncan McLaren and nadine mengis, “navigating potential hype and opportunity in governing marine carbon removal, Frontiers in Climate 3 (2021) 1–8. Article 664456. [68] Javier Lezaun, Hugging the shore: tackling marine carbon dioxide removal as a [91] Pablo Suarez, K. Maarten, van Aalst, Geoengineering: a humanitarian concern, Earth’s Future 5 (2017) 183–195, https://doi.org/10.1002/2016EF000464, 2017. [92] Gernot Wagner, The Gamble, Polity Press, Geoengineering, 2021. [93] Gernot Wagner, The Gamble, Polity Press, Geoengineering, 2021. [94] Bazilian Morgan, et al., Model and manage the changing geopolitics of energy, Nature 569 (2019) 29–31. [95] Stuart Candy, U.S. Earth force, April 22, https://medium. com/@futuryst/u-s-earth-force-58a2c1576d35#, 2020. [96] Joshua B. Horton, Jesse L. Reynolds, The international politics of climate engineering: a review and prospectus for international relations, Int. Stud. Rev. 18 (2016) 438–461. local governance problem, Frontiers in Climate 3 (2021) 1–6. Article 684063. [97] Axel Michaelowa, Solar radiation modification-A “silver bullet” climate policy for [69] Anthony E. Chavez, Using legal principles to guide geoengineering deployment, N. Y. Univ. Environ. Law J. 24 (No. 1) (2016) 59–110. [70] Stephen H. Schneider, The worst-case scenario, Nature 458 (2009) 1104–1105. [71] H. Stephen, Schneider, "Geoengineering: could we or should we make it work? Phil. Trans. R. Soc. A 366 (2008) 3843–3862. [72] Corry, “The International Politics of Geoengineering”. [73] Low and Boettcher, “Delaying Decarbonization”. [74] Simon Dalby, Geoengineering, The next era of geopolitics? Geography Compass 9/4 (2015) 190–201, 10.1111/gec3.12195. [75] Ben Kravitz, Douglas G. MacMartin, Michael J. Mills, Jadwiga H. Richter, Simone Tilmes, Jean-Francois Lamarque, Joseph J. Tribbia, Francis Vitt, First populist and authoritarian regimes? Global Policy 12 (2021) 119–128. [98] Joshua B. Horton, Jesse L. Reynolds, The international politics of climate engineering: a review and prospectus for international relations, Int. Stud. Rev. 18 (2016) 438–461. [99] Joshua B. Horton, Jesse L. Reynolds, The international politics of climate engineering: a review and prospectus for international relations, Int. Stud. Rev. 18 (2016) 438–461. [100] Joshua B. Horton, Jesse L. Reynolds, The international politics of climate engineering: a review and prospectus for international relations, Int. Stud. Rev. 18 (2016) 438–461. EnergyStrategyReviews45(2023)10103116
10.1080_14697688.2022.2159505
Assessing the accuracy of exponentially weighted moving average models Assessing the accuracy of exponentially weighted moving average models for value-at-risk and expected shortfall of crypto portfolios for value-at-risk and expected shortfall of crypto portfolios Carol Alexander, Michael Dakos-Mantoudis Publication date Publication date 10-06-2023 Licence Licence This work is made available under the CC BY 4.0 licence and should only be used in accordance with that licence. For more information on the specific terms, consult the repository record for this item. Document Version Document Version Published version Citation for this work (American Psychological Association 7th edition) Citation for this work (American Psychological Association 7th edition) Alexander, C., & Dakos-Mantoudis, M. (2023). Assessing the accuracy of exponentially weighted moving average models for value-at-risk and expected shortfall of crypto portfolios (Version 1). University of Sussex. https://hdl.handle.net/10779/uos.23494175.v1 Published in Published in Quantitative Finance Link to external publisher version Link to external publisher version https://doi.org/10.1080/14697688.2022.2159505 Copyright and reuse: Copyright and reuse: This work was downloaded from Sussex Research Open (SRO). This document is made available in line with publisher policy and may differ from the published version. Please cite the published version where possible. Copyright and all moral rights to the version of the paper presented here belong to the individual author(s) and/or other copyright owners unless otherwise stated. For more information on this work, SRO or to report an issue, you can contact the repository administrators at sro@sussex.ac.uk. Discover more of the University’s research at https://sussex.figshare.com/ Quantitative Finance ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/rquf20 Assessing the accuracy of exponentially weighted moving average models for Value-at-Risk and Expected Shortfall of crypto portfolios Carol Alexander & Michael Dakos To cite this article: Carol Alexander & Michael Dakos (2023): Assessing the accuracy of exponentially weighted moving average models for Value-at-Risk and Expected Shortfall of crypto portfolios, Quantitative Finance, DOI: 10.1080/14697688.2022.2159505 To link to this article: https://doi.org/10.1080/14697688.2022.2159505 © 2023 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group Published online: 20 Jan 2023. Submit your article to this journal Article views: 273 View related articles View Crossmark data Full Terms & Conditions of access and use can be found at https://www.tandfonline.com/action/journalInformation?journalCode=rquf20 Quantitative Finance, 2023 https://doi.org/10.1080/14697688.2022.2159505 Assessing the accuracy of exponentially weighted moving average models for Value-at-Risk and Expected Shortfall of crypto portfolios CAROL ALEXANDER * and MICHAEL DAKOS University of Sussex Business School, Falmer, BN1 6SL Sussex, UK (Received 6 April 2022; accepted 9 December 2022; published online 20 January 2023 ) A plethora of academic papers on generalized autoregressive conditional heteroscedasticity (GARCH) models for bitcoin and other cryptocurrencies have been published in academic journals. Yet few, if indeed any, of these are employed by practitioners. Previous academic studies produce results that are fragmented, confusing and conflicting, so there is no commercial incentive to drive an expensive implementation of complex multivariate GARCH models, which anyway would com- monly require more data for calibration than are available in the history of most cryptocurrencies, at least at the daily frequency. Consequently, this paper assesses the forecasting accuracy of simple parametric RiskMetricsTM type volatility and covariance models, with a focus on ad hoc parameter choice instead of a data-intensive calibration procedure. We provide extensive backtests of hourly and daily Value-at-Risk (VaR) and Expected Shortfall (ES) forecasts that are regarded as best prac- tice in the industry and commonly used for regulatory approval. Our results demonstrate that much simpler models in the exponentially weighted moving average (EWMA) class are just as accurate as GARCH models for VaR and ES forecasting, provided they capture an asymmetric volatility response and a heavy-tailed returns distribution. Moreover, on ranking each model’s variance and covariance forecasts using average scores generated from proper univariate and multivariate scoring rules, there is no evidence of superior performance of variance and covariance forecasts generated by GARCH models, using either daily or hourly data. Keywords: Volatility clustering; Conditional VaR; Continuous ranked probability score; Energy score; Traffic light tests JEL Classification: C22, C5, F31, G1, G2 1. Introduction The modelling and forecasting of volatility and quantile risk measures for cryptocurrencies is a fairly well-researched topic. Almost 350 papers have been published by academic journals and over 100 of these have appeared during the last 2 years.† This strand of research has become increasingly complex over time, examining numerous variants from the generalized autoregressive conditional heteroscedasticity (GARCH) family of models initially introduced by Boller- slev (1986), several models in the generalized autoregressive score (GAS) class introduced by Creal et al. (2013), as well as mixture and regime-switching specifications of both. A similar degree of variety and complexity exists in the distri- bution assumptions for cryptocurrency returns: while the nor- mal distribution is used by some authors, the most common choices are heavy-tailed distributions such as the Student-t. Many papers employ even more complex heavy-tailed and ∗Corresponding author. Email: c.alexander@sussex.ac.uk † A relevant Scopus search yields 342 papers published between 2015 and early 2022 in Economics, Econometrics, Finance, Busi- these ness, Management or Accounting journals and 131 of papers were published in 2021 or early 2022. These results are produced with the following Scopus search query: TITLE- ABS-KEY((‘bitcoin’ OR ‘Bitcoin’ OR ‘ethereum’ OR ‘ether’ OR ‘Ethereum’ OR ‘Ether’ OR ‘cryptocurrency’ OR ‘cryptocurren- cies’ OR ‘cryptoasset’ OR ‘crypto asset’ OR ‘crypto’ OR ‘dig- ital currency’ OR ‘digital asset’ OR ‘crypto currency’) AND (‘GARCH’ OR ‘EWMA’ OR ‘Value at Risk’ OR ‘VaR’ OR ‘Value-at-Risk’ OR ‘ES’ OR ‘Expected shortfall’ OR ‘volatility’ OR ‘covariance’ OR ‘variance’) AND (‘model*’ OR ‘forecast*’ OR ‘estimat*’)) AND (LIMIT-TO(DOCTYPE, ‘ar’)) AND (LIMIT- TO(SUBJAREA, ‘ECON’) OR LIMIT-TO(SUBJAREA, ‘BUSI’)). © 2023 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. 2 C. Alexander and M. Dakos skewed distributions, such as the generalized error distribu- tion (GED), the Weibull, Beta, generalized hyperbolic, inverse Gaussian and Johnson’s SU distribution. However, this complexity in modelling choices for cryp- tocurrency risk modelling in the academic literature is in stark contrast with current practice in cryptocurrency markets. It is quite common for investors to apply no form of risk analy- sis at all, with risk management strategies consisting at most of stop-loss limit orders placed at arbitrary price levels for open positions.† The few online sources that do discuss, use or provide forecasts of volatility, Value-at-Risk (VaR) and/or Expected Shortfall (ES) use equally-weighted methodologies and inappropriate assumptions. For instance, Cryptodatad- ownload, a cryptocurrency market data and analytics provider, produces daily 1% and 5% VaR and ES forecasts for several cryptocurrencies using a historical methodology over a two- year period, i.e. the percentage VaR is forecast as −1× the corresponding quantile of the empirical returns distribution and ES is −1× the average of the returns that are lower than the corresponding quantile. A blog from the cryptocurrency exchange OKEx presents a parametric VaR estimation for bit- coin, under the assumption that its one-minute returns follow a normal distribution; the 1% and 5% VaR are then forecast using the sample mean and standard deviation of one-minute returns over the past seven days.‡ Similarly, the daily ‘Bitcoin Volatility Index’ is calculated using the standard deviation of returns over the past 30 and 60 days; and the bitcoin Fear & Greed Index and a Forbes article (Bovaird 2021) reporting on bitcoin’s volatility both appear to be estimating volatility with a similar equally-weighted moving average. But there is a very well-known problem with any equally-weighted VaR or ES model. Even a single historical outlier, a large nega- tive return which may have occurred far in the past, will have exactly the same influence on the current value of the risk measure as if it happened just now.§ The calibration of GARCH and GAS models requires a large number of historical returns.¶ While some cryptocurren- cies such as bitcoin or ether have been trading for some time, the continuous emergence of new coins and tokens that gain investor attention often means that newer cryptocurrencies have insufficient data available to produce robust parameter estimates. For instance, at the time of writing, the list of top ten cryptocurrencies by market cap reported by Cryptocom- pare includes Avalanche, Solana and Terra which have only been trading for about two years. For such cryptocurrencies, volatility models that can be ‘jump-started’ and produce fore- casts without the need for a lengthy estimation period, such † Note that the above refers to relatively unsophisticated retail investors that maintain unhedged positions in cryptocurrencies; it does not apply e.g. to investors or market makers that partially or completely hedge their positions with derivatives. ‡ The OKEx blog even mentions that ‘VaR is useful for calculating the maximum expected loss on an investment’, which is a highly inaccurate and misleading interpretation. § These so-called ‘ghost features’ have been recognized in traditional financial markets for decades—see, for instance, Section IV.2.10.2 in Alexander (2008). ¶ The optimal sample size is highly dependent on the characteris- tics of the data. For cryptocurrency univariate volatility modelling, between one and two years of data have been a common choice. Much larger samples are required for multivariate model calibrations to be stable over rolling or expanding windows. as the RiskMetricsTM exponentially-weighted moving average (EWMA) model (Longerstaey and Spencer 1996), are ideal. EWMA models have the added advantage of allowing the use of ad hoc parameter values even when we include features such as an asymmetric volatility response and a heavy-tailed Student-t distribution assumption. However, before this paper we had little or no idea of the performance of EWMA models for bitcoin and other cryp- tocurrencies, relative to the more complex models that have been the focus of previous academic research. Indeed, a major limitation of the extant literature is the lack of consideration of simpler models, even though such models are most com- monly employed by practitioners. There are also numerous gaps in the extant literature on cryptocurrency risk metrics. For instance, there is a complete absence of the traffic lights for VaR and ES backtesting which have been standard prac- tice in the industry since Basel Committee (1996)—and there is just one single paper which uses scoring rules for density forecast evaluation. Likewise, only one other paper examines the VaR and ES of short positions on cryptocurrencies even though these are as easily traded as long positions on all the major exchanges. Furthermore, hardly any other papers exam- ine the forecasting accuracy of multivariate models, even though these should form the corner stone of cryptocurrency portfolio optimization techniques. And all previous academic studies employ data at the daily frequency, with samples that are often too small to yield robust and reliable results. None of them use hourly data even though these data are readily avail- able and there are distinct advantages of using hourly data: firstly for a 24-fold increase in sample size and hence a much larger data set for risk model calibration and backtesting; and secondly for a means to capture intraday volatility, which is especially important for cryptocurrencies because they have many more price jumps and short bursts of volatility than tra- ditional assets. One purpose of this paper is to fill all these gaps in the otherwise highly prolific literature. By contrast, the complex end of the modelling spectrum is over-researched, at least from the cryptocurrency practi- tioner’s perspective. Our tenet is that there is very limited scope for real-world applications of FIGARCH, ACGARCH, TGARCH, H-GARCH, ALL-GARCH, APARCH, MS- GARCH and several other varieties that have been explored in this strand of cryptocurrency research. By contrast, a class of EWMA models which extends the basic RiskMetricsTM methodology is ideally suited for risk-based applications of cryptocurrency portfolios—for two main reasons: first because the methodology is easy to understand, validate, and explain in a simple technical document; second, and perhaps most importantly, these models do not require large samples of historical data for parameter estimation, and so they can be backtested using the maximum amount of historical data available which, for some cryptocurrencies, is already rather small. This paper investigates the relative performance of dif- ferent types of EWMA model and a variety of GARCH models for capturing volatility clustering in USD prices of bitcoin, ether, ripple and litecoin. We use these coins because, unlike many other coins or tokens, they have the sufficiently long history that is needed for proper calibration and thor- ough backtesting of multivariate GARCH models. Bitcoin, ether and ripple are also among the largest coins by market Assessing the accuracy of exponentially weighted moving average models for VaR and ES of Crypto portfolios 3 capitalization, as litecoin also used to be. Our main pur- pose is to quantify the gains, if any, from using the complex GARCH models whose performance for bitcoin and a few other cryptocurrencies has already been extensively analysed in a burgeoning yet fragmented literature. First we present a concise and accessible summary of the crypto GARCH literature, reviewing its unifying themes and obvious gaps, and conclude that there is no consistent evidence to support the use of any model more complex than a simple asym- metric GARCH(1,1) with Student-t innovations. Empirical results are divided as to whether the exponential GARCH (EGARCH) model of Nelson (1991) or the GJR-GARCH model of Glosten et al. (1993) is better at capturing the nec- essary asymmetry—we find the EGARCH slightly better for major coins, but either would serve. Our benchmark volatility model is the sample standard deviation—a simple equally-weighted moving average of past squared returns—against which we assess the perfor- mance, in both univariate and multivariate systems, of sev- eral adapted EWMA models, with and without asymmetric volatility responses, and both symmetric and asymmetric GARCH models, all with Student-t innovations. Our applica- tions extend previous research in several ways: by analysing hourly as well as daily log returns; by backtesting one-step- ahead ES as well as standard VaR metrics; by studying both univariate and multivariate systems; and by further evaluat- ing the volatility and covariance forecasts using univariate and multivariate proper scoring rules. The daily data backtesting sample is from January 2017 to August 2021 and for the hourly data we produce fore- casts from 1 May 2021 to 1 July 2021. Because this research is targeted towards risk management professionals, we back- test VaR forecasts with the industry-standard traffic light and conditional coverage test of Christoffersen (1998); similarly we use a modified traffic light test for ES as well as the exceedance residual test of McNeil and Frey (2000). The accuracy of volatility forecasts is also assessed using the continuous ranked probability score of Gneiting and Ran- jan (2011), the energy score developed by Gneiting and Raftery (2007) is employed for evaluating covariance fore- casts, and we also assess forecasting accuracy using the univariate and multivariate negatively oriented logarithmic scoring rules, as mentioned by Gneiting and Ranjan (2011) and used by Catania et al. (2019). Overall, we conclude that EWMA models perform at least as well as GARCH models at all levels of coverage up to and including 99%, and sometimes they perform even better. Interestingly, we find that hourly forecasts are less accurate than daily forecasts in general, when examining the number of models that fail the VaR and ES backtesting in each case. Nevertheless, most EWMA models are sufficiently accurate to pass traffic light and coverage tests at all three tail quan- tiles, for both long and short positions. By contrast, the more sophisticated Student-t exponential GARCH models often fail to make accurate predictions at the hourly level. Their param- eter estimates are less stable than they are with a daily rolling- window re-calibration. At the hourly frequency it seems that GARCH models are fitting high-frequency fluctuations that appear irrelevant for forecasting the tails of one-hour-ahead distributions and it is better to use the stable, if ad hoc parameters of a EWMA model. For predicting the volatility and covariance structure and when assessing the results using proper scoring rules, all models (including the random walk benchmark) are equally (in)accurate. This is true for both univariate and multivariate density predictions and for one-day-ahead as well as one- hour-ahead forecasts. This finding supports a simple form of market efficiency, which is not surprising since the trad- ing volumes on large coins have grown very rapidly during the last few years, so by now the markets have become quite mature. Nevertheless it is worthwhile to have demon- strated this efficiency empirically, at the daily frequency since January 2017 and at the hourly frequency since 1 May 2021. In the following: Section 2 provides a critical survey of the extensive literature on cryptocurrency volatility, VaR, ES and covariance forecasting; Section 3 specifies the models used in our empirical study, as well as the backtesting of VaR and ES predictions and the use of proper scoring rules for assessing the accuracy of volatility and covariance forecasts; Section 4 provides an overview of the daily and hourly historical data used for the analysis; Section 5 presents our empirical results; and Section 6 summarizes and concludes. 2. State-of-the-art crypto risk models Here we summarize the burgeoning academic literature on cryptocurrency market risk modelling by focusing on papers which assess the in-sample and out-of-sample performance of parametric volatility and/or covariance models applied to cryptocurrency returns. For ease of reference, the main characteristics of the most relevant academic papers are sum- marized in Table 1. Table 1 reports the cryptocurrencies examined, the sam- the models employed and their distributional ple period, assumptions, and the performance criteria used to discrimi- nate between competing models. The cryptocurrencies most commonly examined are: bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC), dogecoin (DOGE), dash, mon- ero (XMR), maidsafecoin (MAID), stellar (XML), bytecoin (BCN), bitcoin cash (BCH), bitcoin gold (BTG), bitcoin diamond (BCD), bitcoin private (BTCP), and also an equally- weighted and a minimum variance portfolio.† A few authors examine a more expanded cryptocurrency universe, e.g. Cata- nia and Grassi (2021) analyse a total of 606 cryptocurrencies having at least 700 daily price observations until September 2019, but the majority of papers focus on bitcoin, ether, ripple and litecoin because these offer a historical period of at least five years and they are consistently amongst the largest coins by market capitalization. The sample frequency is almost invariably daily and the sample period used in each paper usually depends on the available historical data. For exam- ple, Katsiampa (2017) and Baur et al. (2018) only examine bitcoin, so their sample period begins in 2010. However, Fan- tazzini and Zimin (2020) use less than three years of data for both calibration and backtesting. This, like most of the † Historical data for the above cryptocurrencies are obtained from the following sources: blockchain.com, Binance, Bitstamp, Bloomberg, Brave New Coin, Coindesk, Coinmarketcap, Cryptocompare, Gem- ini and Kraken. Table 1. Key characteristics of the relevant academic papers that assess the forecasting performance of cryptocurrency volatility and covariance models. Author Assets Sample period Models Distributions In-sample Out-of-sample Bouoiyour and Selmi (2016) BTC 2011–2016 Chu et al. (2017) Katsiampa (2017) Baur et al. (2018) Bonello and Suda (2018) Ardia et al. (2019) Caporale and Zekokh (2019) Catania et al. (2019) Guesmi et al. (2019) Sosa et al. (2019) Tiwari et al. (2019) Trucíos (2019) Troster et al. (2019) Wang et al. (2019) BTC, XRP LTC, XMR DASH DOGE MAID BTC BTC BTC BTC BTC, ETH XRP, LTC BTC, ETH XRP, LTC BTC BTC BTC LTC BTC BTC BTC 2014–2017 2010–2016 2010–2015 2016–2018 2011–2018 2010–2018 2013–2018 2015–2018 2015–2017 2012–2018 2010–2019 2011–2018 2011–2017 2010–2018 2013–2018 GARCH, EGARCH APARCH, wCGARCH CMT GARCH GARCH, IGARCH GJR-GARCH, EGARCH APARCH, CGARCH, TGARCH AVGARCH, NGARCH, AGARCH ALL GARCH GARCH, APARCH CGARCH, ACGARCH EGARCH GARCH MS-GARCH GARCH, GJR-GARCH MS-GARCH MS-GJR-GARCH Single-regime/mix./MS GARCH, GJR-GARCH EGARCH, TGARCH EWMA TVP-VAR GARCH, GJR-GARCH EGARCH, FIGARCH, FIAPARCH DCC, ADCC, cDCC, cADCC GARCH, EGARCH TGARCH, APARCH CGARCH, ACGARCH GARCH, GJR-GARCH stochastic vol. GARCH, AVGARCH GAS, GARCH-MIDAS realized-GARCH robust-GARCH EGARCH, GJR-GARCH, APARCH TGARCH, CGARCH, NGARCH HGARCH, GAS GARCH, EGARCH CGARCH, ARJI Normal Normal Student-t GED, SU gen. hyperbolic inv. Gaussian Normal Normal Student-t Normal Student-t Normal Student-t Normal Student-t GED Normal Normal Normal GED Normal Student-t Normal, GED Student-t, SU gen. hyperbolic inv. Gaussian Normal Student-t GED, SU Normal AIC BIC HQ AIC cAIC corAIC BIC HQ AIC BIC HQ Parameters AIC BIC DIC Parameters UC, CC ER DIC CC, DQ UC, CC, DQ ER, ESR MCS MSE Log score, MCS AIC BIC LL AIC HQ MLR Parameters AIC BIC LL UC, CC, DQ MSE, QLIKE, RLF MCS UC, CC, DQ RMSE Regression test 4 C . A l e x a n d e r a n d M . D a k o s Acereda et al. (2020) Alexander and Dakos (2020) Author Bazán-Palomino (2020) Fantazzini and Zimin (2020) Hattori (2020) BTC, ETH XRP, LTC BTC Assets BTC, LTC BCH, BTG BCD, BTCP BTC, ETH XRP, LTC XLM, eq. w. portfolio BTC 2010–2018 2013–2018 2015–2018 2013–2019 Sample period 2013–2019 2017–2019 2018–2019 2016–2018 2016–2018 Köchling et al. (2020) BTC 2015–2018 Liu et al. (2020) until 2019 Nekhili and Sultan (2020) Segnon and Bekiros (2020) Catania and Grassi (2021) BTC ETH LTC BTC, XRP LTC, DASH XMR, XLM BCN BTC 606 large-cap coins 2014–2019 2013–2018 until 2019 GARCH, CGARCH NGARCH, TGARCH MS-GARCH MS-GJR-GARCH MS-EGARCH Models EWMA BEKK-GARCH DCC-GARCH GARCH DCC-GARCH copulas GARCH, IGARCH GJR-GARCH EGARCH, APARCH GARCH, IGARCH GJR-GARCH, EGARCH APARCH, CGARCH AVGARCH, TGARCH NGARCH, AGARCH score-driven EWMA EWMA TGARCH SVCJ GARCH, GJR-GARCH EGARCH, APARCH FIGARCH MS-GARCH GAS EGARCH Asymmetric Student-t Normal Student-t Distributions Normal Normal Student-t Normal Student-t Normal Student-t Normal Student-t Laplace gen. Pareto reflected Gamma Normal Student-t Normal Student-t gen. hyperbolic Beta-skew-t DIC IC Parameters In-sample Parameters Residuals LL AIC BIC Multi-level Out-of-sample UC, CC ER, Multi-level MCS MSE QLIKE MSE, MIX, QLIKE MCS UC, CC, DQ MCS CC, QL ER RMSE, MAE MCS LR DQ ER MSE, QLIKE CRPS (Continued) A s s e s s i n g t h e a c c u r a c y o f e x p o n e n t i a l l y w e i g h t e d m o v i n g a v e r a g e m o d e l s f o r V a R a n d E S o f C r y p t o p o r t f o l i o s 5 Table 1. Continued. Author Maciel (2021) Silahli et al. (2021) Assets Sample period Models Distributions In-sample Out-of-sample BTC, ETH XRP, LTC XMR, DASH BTC, XRP LTC, DASH min. var. portfolio 2013–2018 2014–2018 2015–2018 2014–2019 Single-regime/MS GARCH, EGARCH TGARCH Hist. VaR EQMA EWMA GARCH Normal Student-t GED Normal Weibull DIC CC, DQ, QL FZL joint DM UC, CC, DQ Notes: The columns indicate the author of each paper, the cryptocurrencies and the sample period, the models and distribution assumptions used, and the in- and out-of-sample analysis performed. In-sample diagnostics include the Akaike (AIC), Bayesian (BIC) and Hannan-Quinn (HQ) information criteria, and their modifications such as the consistent AIC (cAIC) and corrected AIC (corAIC). Other in-sample performance criteria include the direct comparison of the log likelihood (LL), the marginal likelihood ratio (MLR) and also, for models estimated via MCMC, the deviance information criterion (DIC) and the Bayesian predictive information criterion (IC); further analysis of the in-sample fit includes the examination of parameter estimates and residuals. Out-of- sample tests include the unconditional coverage (UC), conditional coverage (CC), dynamic quantile (DQ) tests for VaR and the exceedance residual (ER) and ESR tests for ES and also the multi-level ES approximation test via VaR of Kratz et al. (2018). Other forecast evaluation methods include the likelihood ratio (LR) test, the model confidence set (MCS) process and forecasting performance DM test of Diebold and Mariano (1995) and regression test of Mincer and Zarnowitz (1969), using loss functions such as the mean square forecasting error (MSE) or its square root (RMSE), mean absolute error (MAE), the robust family loss functions (RLF) of Patton (2011) including the MIX and QLIKE functions, the quantile loss (QL) function and the FZL joint VaR/ES loss function of Fissler and Ziegel (2016). Proper scoring rules include the continuous ranked probability score (CRPS) and log score. All GARCH models shown in the Models column refer to models of first order such as a GARCH(1,1). 6 C . A l e x a n d e r a n d M . D a k o s Assessing the accuracy of exponentially weighted moving average models for VaR and ES of Crypto portfolios 7 studies summarized in Table 1, would not pass the stringent Basel guidelines on historical data for market risk capital calculation.† 2.1. Survey of models employed First we summarize the models used not only in the papers summarized in Table 1 but also for numerous other applica- tions of GARCH models to cryptocurrencies returns. Regard- ing the literature summarized in Table 1, the most com- mon choices include the symmetric GARCH of Boller- slev (1986) and asymmetric models such as the GJR-GARCH of Glosten et al. (1993), the exponential GARCH (EGARCH) the threshold GARCH (TGARCH) of of Nelson (1991), Zakoian (1994), the asymmetric power ARCH (APARCH) of Ding et al. (1993) and, less often, the AGARCH of Engle and Ng (1993). These models are in some cases extended fur- ther with distribution mixture and Markov switching (MS) frameworks. Some authors use the component GARCH (CGARCH) of Engle and Lee (1999) and variants such as its asymmetric extension ACGARCH, the weighted compo- nent GARCH (wCGARCH) of Bauwens and Storti (2009) and the component with multiple threshold (CMT) GARCH of Bouoiyour and Selmi (2014). Still more complex volatil- ity model choices include the H-GARCH and ALL-GARCH of Hentschel (1995), the non-linear NGARCH of Higgins and Bera (1992), the AVGARCH of Schwert (1990), the robust GARCH model of Trucíos et al. (2017), the realized GARCH model of Hansen et al. (2012), the GARCH-MIDAS (mixed data sampling) model of Engle et al. (2013), and also an autoregressive jump intensity (ARJI) model and a stochastic volatility model with co-jumps (SVCJ). More sophisticated univariate models include the realized GARCH and stochastic volatility models which are discussed by Takahashi et al. (2016), Chen et al. (2021) and Takahashi et al. (2021). In the context of cryptocurrencies, stochastic volatility models have been used by Tiwari et al. (2019) and realized GARCH by Trucíos and Taylor (2022)—but only in the univeraite context. These papers also have mixed results, suggesting a possible need for further research. In a simiar vein, more complex distributional assumptions beyond the normal and Student-t and their skewed variants could be applied—including the generalized error distribution (GED), generalized hyperbolic, Weibull, Laplace, Beta-skew-t, gen- eralized Pareto, reflected Gamma, inverse Gaussian and John- son’s SU distribution. All of these GARCH variants have been explored in the voluminous research literature on univariate GARCH modelling, but their extension to large dimensional multivariate systems of returns presents a challenge. Con- sequently it is not surprizing that there is no evidence of widespread adoption of these complex models by financial risk practitioners, even for volatility modelling in traditional asset classes.‡ It may be that some of these state-of-the-art volatility models could produce superior results, for some individual cryptocurrencies, but in this paper our focus is on the widespread uptake of simpler, multivariate risk models by practitioners, specifically those that fall within an asymmetric extension of the RiskMetricsTM EWMA class. The vast majority of other papers are about the diver- sification or hedging effects of bitcoin, and these typically employ some variant of the GARCH class with normal or Student-t distributed innovations, and again all such models are GARCH(1,1).§ For instance: Dyhrberg (2016) compares bitcoin with gold and the dollar using both symmetric and exponential normal GARCH; Bouri et al. (2017) examine the hedging and safe-haven properties of bitcoin and use a symmetric model with innovations that follow a gener- alized error distribution (GED); Al-Khazali et al. (2018) compare the impact of macroeconomic news on bitcoin and gold and find that the best GARCH model is the exponen- tial GARCH with normally distributed error terms; Corbet et al. (2018) examine the applications of bitcoin futures and use a symmetric GARCH; Vidal-Tomás and Ibañez (2018) use a component GARCH to examine the efficiency of bitcoin traded prices; Al-Yahyaee et al. (2019) study the diversi- fication effects of bitcoin and gold for crude oil and S&P 500 investments and use several GARCH models including a fractionally integrated (FI) EGARCH model; and López- Cabarcos et al. (2020) analyse the effect of investor sentiment and S&P 500 and VIX returns on bitcoin’s volatility, using GARCH and EGARCH models. Due to its simplicity and ease of use, the RiskMetricsTM EWMA model of Longerstaey and Spencer (1996) is very popular in financial market applications, and some academic papers focus on assessing its forecasting accuracy using tradi- tional asset as well as cryptocurrency data. For instance, Pafka and Kondor (2001) examine its VaR forecasting ability for returns on the 30 constituent stocks of the DJIA index, argu- ing that it performs well at lower (e.g. 95%) coverage levels and for short-term risk horizons, but that its accuracy declines at 99% coverage and also for multi-period forecasts. Similar results are reported by McMillan and Kambouroudis (2009), now examining 31 stock market indices. Specifically in the cryptocurrency literature, there is some support for the use of integrated GARCH (IGARCH) models—and the EWMA model falls into the integrated volatility model class. For instance Chu et al. (2017) and Köchling et al. (2020) find that IGARCH provides the optimal in-sample fit for bitcoin and other cryptocurrencies; and Bouoiyour and Selmi (2016) and Baur et al. (2018) both find that bitcoin’s variance pro- cess is integrated. The forecasting performance of EWMA volatility models is assessed by Catania et al. (2019), Bazán- Palomino (2020), Nekhili and Sultan (2020) and Silahli et al. (2021). Silahli et al. (2021) also examine an even sim- pler equally-weighted moving average (EQMA) model as a benchmark, while Guesmi et al. (2019) and Segnon and Bekiros (2020) use fractionally integrated models such as the FIGARCH and FIAPARCH. Liu et al. (2020) consider several † See for example, [MAR31.12], [MAR31.23] and [MAR31.26] of Minimum capital requirements for market risk. ‡ Implementation of these models commonly employs maximum likelihood estimation (MLE), which is the simplest method of multi- variate optimization. Markov chain Monte Carlo (MCMC) has also been applied to regime-switching GARCH models, but only at the univariate level. We note that Tiwari et al. (2019) use the cross- entropy method of Rubinstein (1997) for calculating the marginal likelihood, again in a univariate model. § For this reason, in the following the term GARCH always means GARCH(1,1), unless otherwise stated. 8 C. Alexander and M. Dakos score-driven EWMA models based on the generalized autore- gressive score (GAS) model framework of Creal et al. (2013), and Trucíos (2019), Troster et al. (2019) and Catania and Grassi (2021) also use GAS models. The forecasting performance of multivariate covariance models has been only rarely studied, and in these few papers only in-sample performance has been assessed. Bouri et al. (2017) were the first to examine cryptocurrencies in a multivariate context, using a dynamic conditional correlation (DCC) model of Engle (2002) to test the hedge and safe- haven properties of bitcoin. The majority of other studies use the DCC model and only a few employ the earlier BEKK model of Engle and Kroner (1995). For instance, Bazán- Palomino (2020) considers the relationship between bitcoin and similarly structured cryptocurrencies using the multi- variate EWMA, BEKK-GARCH and DCC-GARCH, while Guesmi et al. (2019) use the DCC model to examine bitcoin as well as a number of traditional financial assets. Regard- ing applications of a multivariate EWMA model, Matkovskyy et al. (2020) use one to examine the interdependence between bitcoin, economic policy uncertainty and traditional financial assets, but none of the relevant papers assess its forecast- ing performance for VaR and ES of cryptocurrencies, nor do they evaluate the accuracy of covariance forecasts via scor- ing rules. Other covariance modelling choices reported in Table 1 include the asymmetric ADCC model of Cappiello et al. (2006), the modified cDCC and cADCC of Aielli (2013), multivariate extensions of the marginal densities using copula functions to model the correlation structure and time-varying parameter vector autoregression (TVP-VAR) models. 2.2. Survey of performance results Engle et al. (2012) provide a useful survey of the numer- ous papers that explore the best specification for univariate GARCH models on different types of financial data. To update this survey to include the recent research on cryptocurrencies is difficult because the results are often contradictory, suggest- ing that the best in-sample fit very much depends on both the cryptocurrencies chosen and the sample period, which vary considerably from study to study. Although, as noted above, at least all previous work employs data at the same, daily frequency. Katsiampa (2017) tests several parametric volatility mod- els for the best in-sample fit on bitcoin returns and all criteria indicate that the ACGARCH model is optimal; this is con- sistent with Bouoiyour and Selmi (2016) whose in-sample analysis also indicates a model with a transitory and a per- manent volatility component. The in-sample analysis of Baur et al. (2018) indicates superiority of the EGARCH model for bitcoin returns, and the authors note that using different asym- metric volatility models does not improve the in-sample fit. Tiwari et al. (2019) compare the fit of GARCH and stochas- tic volatility models for bitcoin and litecoin and find mixed results, for instance concluding that cryptocurrency returns do not exhibit any asymmetric volatility response, which is at odds with the previous findings. The findings of Sosa et al. (2019) suggest that an EGARCH model with GED innovations provides the best in-sample model fit for bitcoin. Troster et al. (2019) agree that a GED assumption instead of a normal significantly improves goodness-of-fit, but fur- ther conclude that the hyperbolic HGARCH model with GED innovations provides the best in-sample fit, which is again contrary to previous findings. In the class of regime-switching volatility models, Ardia et al. (2019) find that a two-state Markov switching skewed Student-t GJR-GARCH provides a better in-sample fit for bit- coin compared to both non-switching and three-state switch- ing models; the authors propose that the two-state model provides a better trade–off between fitting quality and model complexity and further show for three–regime models that fitting gains are only observed for the normal distribution. Alexander and Dakos (2020) also explore the in-sample fit of two-state Markov switching GARCH models for bitcoin returns and show that the best model depends on the exact source of data used. To sum up, the plethora of in-sample diagnostics applied to GARCH models of cryptocurrency volatility reveals a picture of numerous, but imprecise and highly contradictory conclu- sions, derived from the painstaking estimation of increasingly complex models which often use insufficient data to provide robust and accurate results. Yet, the state-of-the-art results on out-of-sample forecasting for cryptocurrency returns, to which we now turn, are even more confusing. Out-of-sample forecasting centres on VaR and Expected Shortfall backtests, usually focusing on the left tail of the returns’ distribution to assess the risk of downward price movements on long crypto asset positions. It is worth not- ing that the only study other than ours that assesses the performance of right-tail forecasts for losses made on short positions is that of Stavroyiannis (2018), who examines the GJR-GARCH model calibrated to bitcoin returns. The most common backtesting methodologies for VaR forecasts are the unconditional coverage (UC) test of Kupiec (1995), the con- ditional coverage (CC) test of Christoffersen (1998) and the dynamic quantile (DQ) test of Engle and Manganelli (2004); for ES, common backtesting methods include the exceedance residual (ER) of McNeil and Frey (2000), the regression- based ESR test of Bayer and Dimitriadis (2020) and the multi- level backtest approximation via VaR of Kratz et al. (2018). Other methods of analysis include the use of loss functions either in the model confidence set (MCS) process of Hansen et al. (2011) or also in hypothesis tests of equal forecasting performance such as the DM test of Diebold and Mari- ano (1995). Finally, the use of proper scoring rules to evaluate cryptocurrency returns density forecasts is much less com- mon, with Catania and Grassi (2021) using the continuous ranked probability score and Catania et al. (2019) using the log score. Also, and very much in the vein of our paper, we emphasize that the industry standard traffic light backtesting framework of the Basel Committee (1996), e.g. as described by Costanzino and Curran (2018), is overlooked by all these papers. One reason for the confusing conclusions drawn from out- of-sample results is that they depend not only on the models employed but also on the particular cryptocurrency returns studied, the sample period employed and the significance levels examined. For instance, Ardia et al. (2019) compare the VaR forecasting accuracy of single-regime and Markov Assessing the accuracy of exponentially weighted moving average models for VaR and ES of Crypto portfolios 9 switching models for bitcoin, concluding that only regime- switching models produce accurate VaR forecasts at the 1% significance level; however, it is worth noting that 5% daily VaR forecasts produced using the relatively simpler single- regime skewed Student-t GJR-GARCH model also succeed the CC test because we cannot reject the null hypothesis of no clustering in exceedances at the 5% significance level— and the DQ test also. Maciel (2021) compares the prediction performance of Markov switching GARCH against single- regime GARCH models for several crypto assets and is in favour of more complex models similar to Ardia et al. (2019), but the results for a similar set of single- and two-regime GARCH models, also applied to bitcoin, are somewhat mixed. Caporale and Zekokh (2019) also apply a variety of dif- ferent backtests to VaR and ES forecasts for bitcoin, ether, ripple and litecoin with an exhaustive set of mixture and regime switching model combinations, but again the results are inconclusive. It further transpires that even when very complex volatility models can produce accurate out-of-sample VaR and ES fore- casts, relatively simpler models can produce equally accurate results. For instance, Bonello and Suda (2018) compare VaR forecasts for bitcoin using single-regime and two-regime nor- mal and Student-t GARCH models, and find that all specifica- tions can produce accurate VaR forecasts at a 5% significance level. Troster et al. (2019) backtest daily 1% VaR forecasts for bitcoin and find that a Student-t standard GARCH model is on a par with several more complex GARCH and GAS models included in their study. Trucíos (2019) evaluates VaR forecasts for bitcoin between 2011 and 2017 using six compet- ing models, finding that only a robust bootstrap VaR method produces accurate forecasts at the 1% significance level. In fact, in the preliminary results of a subsequent working paper, Trucíos and Taylor (2022) use a more recent sample period and show that bitcoin and ether VaR forecasts based on sim- pler volatility models such as the standard GARCH may be considered accurate. Acereda et al. (2020) find that more complex model specifications do not outperform the simpler ones for bitcoin VaR, as long as heavy-tailed distributions are used instead of the standard normal. Silahli et al. (2021) also find that simple benchmark models succeed in various VaR backtests for several crypto assets. Contradictory results are even apparent when one considers EWMA models alone. For example, Silahli et al. (2021) claim that a normal EWMA volatility model produces accurate VaR forecasts for all cryptocurrencies, but Liu et al. (2020) find that a similar model fails VaR backtests. Nekhili and Sultan (2020) examine the out-of-sample performance of a benchmark RiskMetricsTM EWMA model and find that it pro- duces accurate VaR forecasts at the 5% level, but not at 1%; yet for ES forecasts of almost all cryptocurrencies examined, a EWMA produces accurate ES forecasts according to the ER test. Within the multivariate setting the results seem a lit- tle more consistent: Silahli et al. (2021) find that a EWMA covariance model used to produce VaR forecasts for a port- folio of bitcoin, litecoin, ripple and dash passes performance tests; and Catania et al. (2019) examine bitcoin, ether, rip- ple and litecoin, testing several complex multivariate models against a vector autoregression with EWMA variance and find that none significantly outperform this much simpler benchmark. Finally, Catania et al. (2019) and Catania and Grassi (2021) are the only applications of proper scoring rules specific to cryptocurrencies at the time of writing. Catania et al. (2019) produce multi-period point and density forecasts for bit- coin, litecoin, ripple and ether returns, employing the log score as a measure of forecast accuracy and conclude that most models outperform the EWMA benchmark. Catania and Grassi (2021) use the continuous ranked probability score (CRPS) to assess volatility forecasts from the GAS model versus EGARCH, concluding equal predictive ability as mea- sured by the DM test. They backtest VaR and ES forecasts for a total of 606 cryptocurrencies with at least 700 daily price observations until September 2019. The authors use the score- driven volatility model specifications that incorporate several stylized features such as leverage effects, long memory of the volatility process and time-varying higher order moments, with a generalized hyperbolic skewed Student-t distribution. These models are compared against a benchmark Beta-Skew- t-EGARCH, producing multi-period 1% and 5% VaR and ES forecasts. VaR and ES forecasts are backtested with the DQ and ER tests and the density forecasts are assessed using the CRPS. Score-driven specifications produce accurate 5% and 1% ES and 5% VaR forecasts more often than the Beta-Skew-t-EGARCH benchmark, but GAS models and the EGARCH benchmark are on par when backtesting 1% VaR. Regarding density forecast evaluation via CRPS the authors find that certain score-driven models outperform the bench- mark more often than they underperform it. However, even for these successful specifications, equal predictive ability is the most common outcome. For instance, when examining the uniformly-weighted CRPS of the one-day-ahead density forecast across all cryptocurrencies, equal predictive abil- ity occurs in 83% of cryptocurrencies examined, including bitcoin, ether, ripple and litecoin. While both Liu et al. (2020) and Catania and Grassi (2021) examine several volatility model specifications, the range of models examined is somewhat limited in both cases. Liu et al. (2020) focus specifically on EWMA-type models and do not test other more complex models such as GARCH spec- ifications, nor simpler model specifications that require no calibration such as an equally-weighted moving average or a EWMA with an ad-hoc value chosen for the decay parame- ter. Therefore, their results are not conclusive with respect to the overall suitability of EWMA-type models in forecasting cryptocurrency volatility compared to other more complex or simpler models. By comparison, Catania and Grassi (2021) focus on highly sophisticated GAS model specifications with a similarly sophisticated heavy-tailed distribution assump- tion and test these against an already complex benchmark Beta-skew-t-EGARCH model, often finding equal forecast- ing performance. It is important to note that, as discussed previously, the above finding also extends to VaR and ES forecasting, i.e. the VaR and ES forecasting performance of highly complex GARCH and GAS model specifications can be on par with relatively simpler models such as the standard GARCH. For instance, this is shown in the results of Bonello and Suda (2018), Troster et al. (2019), Acereda et al. (2020), 10 C. Alexander and M. Dakos Silahli et al. (2021) and also in the working paper results of Trucíos and Taylor (2022). 3. Methodology Our benchmark model is that returns are normally distributed with zero mean and variance estimated as an equally-weighted moving average of the past n squared returns. Except for the benchmark model, we make the universal assumption of Student-t innovations, again with zero mean returns.† This is because previous results, available on request, showed that none of the normal models outperformed their Student- t equivalent, for any cryptocurrency. On the other hand, using more complex distributional assumptions as in Chu et al. (2017), Trucíos (2019) and Liu et al. (2020) is tangential to the theme of this paper. It would obfuscate the motivation for this paper by providing too many details. Therefore, to retain our focus on the main story here—i.e. the relative effec- tiveness of using ad-hoc values for EWMA parameters—we only describe the models and report the results for Student-t innovations in all the EWMA and GARCH models. Our benchmark model assumes returns are normal with variance estimated by an n-period equally-weighted moving average of squared returns, we call it the random walk for short. Then we have a EWMA model as per the RiskMetricsTM technical document (Longerstaey and Spencer 1996) and our own asymmetric extension similar to the A-GARCH model of Engle and Ng (1993), a symmetric GARCH(1,1) model (Bollerslev 1986) and an asymmetric EGARCH(1,1) these models assume a model Student-t distribution. Joint density forecasts are produced via n-period equally-weighted moving average covariance matrix estimates, multivariate versions of the EWMA mod- els, and the GARCH and EGARCH models are combined with the dynamic conditional correlation (DCC) model of Engle (2002) and Tse and Tsui (2002) and also its asymmetric extension (ADCC) model of Cappiello et al. (2006). (Nelson 1991)—and all The basic econometric methodology consists of produc- ing one-period-ahead volatility and covariance forecasts on a daily or hourly rolling basis. These are then combined with parametric distribution assumptions to produce one-period- ahead VaR and ES forecasts at various quantiles, where each model has univariate versions for each cryptocurrency and a multivariate version. To assess the risk of both long and short positions we backtest quantiles at 1%, 2.5%, 5%, 95%, 97.5% and 99%. Then the accuracy of one-period ahead volatil- ity and covariance forecasts are evaluated via univariate and multivariate proper scoring rules, respectively. We test the performance of VaR and ES predictions using the traffic light backtests which have been the indus- try standard for more than two decades, (Basel Commit- tee 1996), along with the two standard tests for clustering of exceedances, i.e. the conditional coverage (CC) test of Christoffersen (1998) for VaR, and the (raw) exceedance † There is much support for the zero-mean assumption as already discussed in Section 2. For instance, Köchling et al. (2020) find that GARCH model specifications for bitcoin returns that have zero mean are very often included in the model confidence set. residual (ER) test of McNeil and Frey (2000) for ES.‡ Beyond quantile prediction backtesting, we also examine the accuracy of volatility forecasts using the continuous ranked probabil- ity score (CRPS) of Gneiting and Ranjan (2011) for uni- variate forecasts and the energy score from Gneiting and Raftery (2007) for covariance forecasts.§ Additionally, we employ the univariate and multivariate negatively oriented logarithmic scoring rule as described by Gneiting and Ran- jan (2011). Note that all models assume a zero mean so these scoring rules aim to examine the accuracy of one-period ahead volatility and covariance forecasts, over and above the specific quantile predictions previously assessed. 3.1. Variance and covariance models Denote the return on a single cryptocurrency at time t by rt and assume their mean is zero. In the random walk benchmark model we have: rt = σtεt, with εt ∼ N (0, 1), (1) where σ 2 is the average squared return over the most recent n t periods. In both the EWMA and GARCH models, returns are assumed to follow a zero-mean, location-scale transformed Student-t distribution: (cid:2) rt = σtεt with ν − 2 ν εt ∼ tν, (2) where tν denotes the standardized Student-t distribution with ν degrees of freedom, σt is the standard deviation of rt and the distribution of εt is defined such that εt has unit standard devi- ation. The variance under the standard EWMA model with decay parameter λ is calculated as: σ 2 t = (1 − λ)r2 t−1 + λσ 2 t−1. (3) Based on the AGARCH model of Engle and Ng (1993), we introduce the asymmetric EWMA model with a decay param- eter λ and an asymmetric volatility response parameter η. Under the AEWMA model, the variance is calculated as: σ 2 t = (1 − λ)(rt−1 − η)2 + λσ 2 t−1. (4) In the standard (symmetric) GARCH(1,1) model, the condi- tional variance is given by: σ 2 t = ω + αr2 t−1 + βσ 2 t−1. (5) Similarly, in the Student-t EGARCH(1,1) model, we have: (cid:4) (cid:3) σ 2 t ln = ω + g (εt−1) + βln (cid:4) (cid:3) σ 2 t−1 ‡ Note that the ‘raw’ ER test consists of using the raw residuals which are not divided by the estimated standard deviation. This ver- sion of the ER test is suggested by Bayer and Dimitriadis (2020), who argue that ‘the test using the standardized ER is in fact a joint backtest for the triple VaR, ES and volatility, whereas the test using the raw ER is a joint backtest for the pair VaR and ES’. § It would also be possible to assess the accuracy of volatility and covariance forecasts via a loss function such as the mean absolute error of point forecasts but, as mentioned by Gneiting and Ran- jan (2011), the CRPS may be regarded as an extension of the mean absolute error for density forecasts, instead of point forecasts. Assessing the accuracy of exponentially weighted moving average models for VaR and ES of Crypto portfolios 11 (cid:5) (cid:6) g(εt) = θ εt + γ |εt| − E[|εt|] . (6) Regarding volatility forecasts, the random walk, EWMA and AEWMA models described in Equations (1), (3) and (4) have a constant volatility term structure, so their volatility forecasts for period t + 1 are set equal to the corresponding volatility estimates at time t. For the GARCH and EGARCH models the one-period-ahead volatility forecasts ˆσt+1 are obtained by updating the conditional volatility Equations (5) and (6) using the estimated model parameters and the last of the in-sample estimates for ˆσt and ˆεt. In a multivariate setting, denote by rt the (m × 1) vector of the m cryptocurrency returns at time t. The multivariate random walk benchmark model assumes that rt follows a multivariate normal distribution: rt ∼ N (0, (cid:2)t), (7) where the covariance matrix (cid:2)t is estimated as the sample covariance matrix of returns over the past n days. The EWMA and GARCH models follow their univariate counterparts, so the vector of returns is assumed to follow a multivari- ate location-scale transformed Student-t distribution with ν degrees of freedom: (cid:7) rt ∼ tν 0, (cid:8) (cid:2)t , ν − 2 ν (8) where (cid:2)t is the covariance matrix of rt, so that ν−2 (cid:2)t is ν the distribution’s scale matrix. The covariance matrix in the multivariate EWMA model with parameter λ is given by: (cid:2)t = (1 − λ)rt−1r(cid:4) t−1 + λ(cid:2)t−1. (9) The covariance matrix of the asymmetric EWMA with param- eters λ and η is calculated as: (cid:2)t = (1 − λ)(rt−1 − η1)(rt−1 − η1)(cid:4) + λ(cid:2)t−1, (10) where 1 is an (m × 1) vector of ones. For the multivariate GARCH models, the covariance matrix is modelled as: (cid:2)t = DtCtDt Ct = diag(Qt )−1/2Qtdiag(Qt )−1/2, (11) where Dt is the diagonal matrix of variances estimated via the univariate GARCH or EGARCH model and Ct is the condi- tional correlation matrix, which is modelled indirectly via the Qt matrix to ensure that Ct is a proper, positive semi-definite correlation matrix. In the DCC model, Qt is given by: Qt = (1 − a − b) ¯Q + aεt−1ε(cid:4) t−1 + bQt−1. (12) Similarly, in the ADCC model Qt is calculated as: − where εt is the vector of standardized errors; ε− t are the zero- threshold errors defined as equal to εt when the corresponding elements are less than zero and equal to zero otherwise; and ¯Q and ¯Q are the unconditional covariance matrices of εt and ε− t . The one-period ahead covariance matrix forecasts are pro- duced similar to the volatility forecasts as described pre- viously. For the multivariate random walk, EWMA and AEWMA the 1-period-ahead covariance matrix forecast at time t is set equal to the estimate at time t − 1 and for the DCC and ADCC models it is obtained by updating the conditional covariance Equation (11). 3.2. Backtesting methods The forecasting accuracy of the volatility models presented in the previous section is assessed by producing rolling fore- casts and backtesting them against realized returns. For each of the two quantile risk measures, we use the industry stan- dard traffic light test of the Basel Committee (1996) and one academic standard test, i.e. the conditional coverage (CC) test of Christoffersen (1998) for VaR and the exceedance residual (ER) test of McNeil and Frey (2000) for ES. 3.2.1. Value-at-Risk. The VaR at a significance level α is defined as −1× the α-quantile of the one-period-ahead fore- cast Ft that is made at time t of returns’ distribution function. We set α = 1%, 2.5%, 5% for lower (left-tail) quantiles, using 1 − α for upper (right tail) quantiles, so: VaRt(α) = ⎧ ⎨ ⎩ −F−1 t F−1 t (α), (1 − α) for long positions (left-tail VaR) for short positions (right − tailVaR). (14) The traffic light approach of the Basel Committee (1996), as described in Costanzino and Curran (2018), is extended here to both left- and right-tail VaR. The exceedance indicator (α) of each 1-period-ahead left- and right-tail 100α%- X VaR t VaR forecast at times t = 1, . . . , N is defined as: X VaR t (α) = (cid:12) 1{rt≤−VaRt(α)}, 1{rt≥VaRt(α)} for long positions for short positions, (15) where 1{condition} denotes an indicator function which equals 1 if the condition is satisfied and 0 otherwise. The cumu- (α) over the entire lative number of VaR exceedances X VaR forecasting period t = 1, . . . , N is then calculated as: N X VaR N (α) = N(cid:13) t=1 X VaR t (α). (16) Under the null hypothesis that the VaR model is specified cor- rectly, the total number of VaR exceedances follows a bino- mial distribution with parameters N and α;† we approximate Qt = (1 − a − b) ¯Q − g ¯Q ε−(cid:4) t−1, + gε− t−1 − + aεt−1ε(cid:4) t−1 + bQt−1 † The null hypothesis that the VaR model is ‘specified correctly’ implies a joint hypothesis that the time series of VaR exceedance (α) is independent and identically distributed (i.i.d.) indicators X VaR and that the proportion of realized VaR exceedances is equal to the t (13) 12 C. Alexander and M. Dakos the binomial with a normal distribution as:† X VaR N (α) ∼ N (Nα, Nα (1 − α)) . (17) Let xVaR be the number of realized VaR exceedances over the forecasting period and let z be its standard normal transform. Denote the probability of obtaining xVaR or fewer exceedances as (cid:12)(z), where (cid:12) is the standard normal distribution func- tion.‡ The traffic light colour zones are then defined as: Green if (cid:12)(z) < 0.95; Yellow if 0.95 ≤ (cid:12)(z) < 0.9999; Red if (cid:12)(z) ≥ 0.9999. As described by the Basel Committee (1996), the three- zone approach is introduced to mitigate the statistical limita- tions of backtesting and balance the two error types: type I, i.e. the possibility that an accurate model is classified as inaccu- rate based on its backtesting results; type II, i.e. the possibility that an inaccurate model is not classified as such based on its backtesting results. In the green zone, the backtesting results are considered consistent with an accurate model and the probability of erroneously accepting an inaccurate model is low. In the red zone, the backtesting results are highly unlikely to have resulted from an accurate model, and the probabil- ity of erroneously rejecting an accurate is model is low. In the yellow zone, backtesting results could be consistent with either accurate or inaccurate models, so additional informa- tion is required to determine whether the model is specified correctly. The VaR forecasts are further backtested using the condi- tional coverage (CC) test of Christoffersen (1998), for which the likelihood ratio test statistic LRcc is: by an exceedance; n11 is the number of exceedances preceded by an exceedance.§ The asymptotic distribution of −2 ln LRcc under the null hypothesis is chi-squared with 2 degrees of freedom and the null hypothesis of the CC test for the true transition probabilities π01 and π11 is that π01 = π11 = α, sug- gesting that there is a correct probability of exceedances and no clustering in exceedances. 3.2.2. Expected Shortfall. Expected Shortfall is defined as the expected loss given that the corresponding VaR forecast is exceeded, i.e. (ES) ESt(α) = 1 α (cid:14) α 0 VaRt(p) dp. (19) Also called ‘expected tail loss’ or sometimes ‘conditional VaR’, ES addresses a limitation of VaR in that it cannot capture tail risk beyond the specified quantile of the returns distribution (Basel Committee 2012). A traffic light back- testing method for ES was introduced by Costanzino and Curran (2018) as a generalization of the VaR traffic light backtest of the Basel Committee (1996). Extending the idea of VaR exceedances, Costanzino and Curran (2018) intro- (α) ∈ [0, 1] duce the ES generalized exceedance indicator X ES by applying the definition of ES in Equation (19) to the (α) defined left- and right-tail VaR exceedance indicator X VaR (p)dp. We further in Equation (15), i.e. X ES t extend this definition to right-tail ES, which yields: (cid:15) α 0 X VaR t (α) = 1 α t t (cid:7) ⎧ ⎪⎪⎨ (cid:7) ⎪⎪⎩ (cid:8) 1 − Ft(rt) α 1 − 1 − Ft(rt) α (cid:8) 1{rt≤−VaRt(α)}, 1{rt≥VaRt(α)}, for long positions for short positions. LRcc = (cid:3) ˆπ n01 01 (cid:4) αn1 (1 − α)n0 (cid:3) n00 ˆπ n11 11 1 − ˆπ01 1 − ˆπ11 (cid:4) n10 , (18) (α) = X ES t ); n11 n10+n11 n01 n00+n01 ˆπ11 = ( where: α is the significance level used in the VaR model; ˆπ01 = ( ); n1 is the number of real- ized VaR exceedances; n0 = N − n1 is the number of real- ized returns that do not exceed the VaR forecast; n00 is the number of non-exceedances preceded by a non-exceedance; n01 is the number of exceedances preceded by a non- exceedance; n10 is the number of non-exceedances preceded VaR significance level α. Note that the above definition holds for both left- and right-tail VaR, as exceedances are defined respectively based on the 100α% left and right tail of the distribution. † Note that in the case of the VaR traffic light backtest, the bino- mial distribution is commonly used and is (almost) as easy to work with as the normal. We choose the normal approximation of the binomial distribution to ensure the easiest possible application of our ad hoc methodology, and also for consistency with the ES traf- fic light backtest where we use the asymptotic normal distribution approximation as suggested by Costanzino and Curran (2018). The approximation of the binomial distribution with the normal is consid- ered accurate based on the rule-of-thumb that both Nα and N(1 − α) should be greater than 5, which is the case for the analysis presented in Section 5, as we useN = 1, 704 in the daily frequency analysis, N = 1, 465 in the hourly frequency and α = 1%, 2.5% and 5%. ‡ As the number of realized VaR exceedances xVaR over the fore- (α) casting period is a realization of the random variable X VaR defined in Equations (16) and (17), the probability of obtaining (α) ≤ xVaR) = xVaR or fewer VaR exceedances is given as: P(X VaR N (cid:8) (cid:7) N P X VaR √ N (α)−Nα Nα(1−α) √ ≤ xVaR−Nα Nα(1−α) = (cid:12)(z), where z ∼ N (0, 1). (20) ) and (1 − 1−Ft(rt) The terms (1 − Ft(rt) ) capture the severity α of each VaR exceedance. Returns that exceed the VaR but not (α) is dom- the ES receive a relatively low weight and X ES inated by returns of greater magnitude that exceed both the VaR and ES. The cumulative ES generalized exceedance is then calculated as: α t (α) = X ES N N(cid:13) t=1 X ES t (α). (21) Under the null hypothesis that the ES model is specified cor- (α) is provided by Costanzino rectly, the distribution of X ES N and Curran (2018) based on the binomial and Irwin-Hall distributions;¶ the authors further note that the distribution tends asymptotically to a normal distribution for large fore- casting periods, based on the derivation of Costanzino and § As in the case of the traffic light backtest, the conditional coverage test definitions hold for both left- and right-tail VaR, as exceedances are defined respectively based on the 100α% left and right tail of the distribution. ¶ As noted previously for VaR, the null hypothesis that the ES model is ‘specified correctly’ implies a joint hypothesis that the (α) is i.i.d. time series of ES generalized exceedance indicators X ES and that for all p ∈ [0, α], the probability of VaR exceedances is P(rt ≤ −VaRt(p)) = p for the left tail and P(rt ≥ VaRt(p)) = p for the right tail. t Assessing the accuracy of exponentially weighted moving average models for VaR and ES of Crypto portfolios 13 Curran (2015):† (α) ∼ N X ES N (cid:7) 1 2 Nα, Nα (cid:8)(cid:8) (cid:7) 4 − 3α 12 . (22) Given the total realized ES generalized exceedances over the forecasting period xES, the probability of obtaining xES or fewer ES generalized exceedances is (cid:12)(z), where z is again derived from the standard normal transformation of xES. The traffic light colour zones are therefore again defined as: Green if (cid:12)(z) < 0.95; Yellow if 0.95 ≤ (cid:12)(z) < 0.9999; Red if (cid:12)(z) ≥ 0.9999. The ES forecasts are further analysed using the exceedance residual (ER) test of McNeil and Frey (2000) based on the raw residuals—i.e. not divided by the estimated standard deviation, as suggested by Bayer and Dimitriadis (2020): (cid:12) (−rt − ESt(α))1{rt≤−VaRt(α)}, (rt − ESt(α))1{rt≥VaRt(α)}, εt = for long positions for short positions. (23) The ER test statistic is then calculated as the sample mean of εt: can be used to rank the accuracy of a variance forecast in one simple number. We use the continuous ranked probability score (CRPS) for univariate distribution forecasts and its multivariate exten- sion, the energy score, for joint density forecast evaluation. Similarly, we use the negatively oriented logarithmic score (LogS) to evaluate the univariate and joint density fore- casts. The CRPS (Matheson and Winkler 1976 and Gneiting and Ranjan 2011) generalizes the mean absolute error of an observation y under a forecast distribution F: CRPS(F, y) = (cid:14) +∞ (cid:3) −∞ F(z) − 1{y≤z} (cid:4) 2 dz (25) According to Gneiting and Raftery (2007), the CRPS can also be expressed as: CRPS(F, y) = EF |X − y| − 1 2 EF (cid:18) (cid:18)X − X (cid:4) (cid:18) (cid:18) , (26) ⎧ ⎪⎪⎪⎨ ⎪⎪⎪⎩ (cid:17) (cid:17) ˆμ = (cid:17) N t=1 εt 1{rt≤−VaRt(α)} (cid:17) N εt t=1 1{rt≥VaRt(α)} , N t=1 N t=1 , for long positions for short positions. where X and X (cid:4) are independent random variables with sam- pling distribution F. This representation leads to the energy score extension which generalizes the CRPS for multivariate distributions and is defined (Gneiting and Raftery 2007) as: (24) The test statistic ˆμ does not have a standard distribution so we estimate it using a bootstrap simulation. In the results pre- sented in Section 5, the distribution of the ER test statistic ˆμ is simulated using 1000 bootstrapped replications. The null hypothesis is that E[εt] = 0; this is tested against a 1-sided alternative that E[εt] > 0, suggesting that ES is systemati- cally underestimated. 3.2.3. Score-based tests for variance. Scoring rules mea- sure the accuracy of probabilistic forecasts and allow for comparisons between competing prediction models. In the case of negatively oriented scoring rules, a lower score indi- cates a better forecast for the entire distribution, but the most important determinant of the score is the ability to predict an accurate expected value. Yet here we are setting all mod- els equal in that sense—every model simply assuming a zero mean return, because our focus is on the accuracy (or oth- erwise) of RiskMetricsTM type volatility forecasts. Therefore, the difference between scores in our study is entirely due to difference in accuracy of the variance forecast. We find these score-based tests useful, above and beyond the quantile pre- dictions relating to VaR and ES metrics, because our scores † Note that the derivation described in Costanzino and Curran (2015) yields the asymptotic distribution of our Equation (22), whereas the distribution’s variance in equation (17) of Costanzino and Cur- ran (2018) is mistakenly omitting a factor of N. Regarding the accuracy of the normal approximation, Clift et al. (2016) perform a simulation study and find that the approximation is accurate for a forecasting period of length N = 250; this confirms that the normal approximation is fit for use in our analysis, as we use forecasting periods that include over 1, 000 observations. ES(F, y) = EF ((cid:10)X − y(cid:10)) − 1 2 EF (cid:3)(cid:19) (cid:19)X − X(cid:4) (cid:4) (cid:19) (cid:19) , (27) where (cid:10) · (cid:10) denotes the Euclidian norm on Rn, X and X(cid:4) are independent (n × 1) random vectors from a multivariate distribution with CDF forecast F and y = (y1, . . . , yn) is a realized observation. Moreover, if F is given via m discrete (n- dimensional) samples X = (X1, . . . , Xn), then the energy score is calculated as: ES(F, y) = 1 m m(cid:13) i=1 (cid:10)Xi − y(cid:10) − 1 2m2 m(cid:13) m(cid:13) (cid:19) (cid:19)Xi − Xj (cid:19) (cid:19) . i=1 j=1 (28) Finally, the uniformly-weighted, negatively oriented logarith- mic score of an observation y from a univariate or multi- variate forecast distribution F is defined by Gneiting and Ranjan (2011) as: LogS (F, y) = −logF (y) . (29) Given the 1-period-ahead probability density function fore- casts ft, gt and their corresponding univariate or multivariate scores S(ft) and S(gt) produced on a rolling basis over the out-of-sample period t = 1, . . . , N, we compare the forecast- ing performance of f and g directly using their average scores over the out-of-sample period. Alternatively, we use the hypothesis test of equal performance described by Gneit- ing and Ranjan (2011). If the average scores of f and g over the out-of-sample period are ¯Sf N respectively, then the N and ¯Sg 14 C. Alexander and M. Dakos test of equal performance is based on the statistic: √ N tN = (cid:20) (cid:21) , ¯Sf N − ¯Sg ˆσN N where: ˆσ 2 N = 1 N N(cid:13) (cid:5) t=1 S(ft) − S(gt) (cid:6) 2 . (30) (31) The test statistic tN is asymptotically standard normal under the null hypothesis of vanishing expected score differentials; therefore in case of rejection, f is chosen if tN is negative and g is chosen if tN is positive. 4. Data Intraday volatility is much greater in cryptocurrency markets than in traditional financial markets, so it is worth analysing hourly data here.† Thus, we obtain both daily and hourly historical data on four of the largest cap cryptocurrencies as of 1 January 2021: bitcoin, ether, ripple and litecoin. Since then all but litecoin have remained in the top five cryp- tocurrencies by market cap. Nevertheless, we retain litecoin because so many of the papers reviewed earlier also apply their models and tests to litecoin. Historical price data are collected using the Cryptocompare API and are in the form of volume-weighted (VWAP) close prices, averaged across multiple USD-denominated exchange-traded prices for each crypto asset. For the daily frequency analysis, the sample period is between 20 August 2015 and 31 August 2021, with daily prices recorded at 00:00 UTC 365 days per year. The rolling estimation window length for the GARCH models is fixed at 500 days so that the forecasting period consists of 1,704 daily observations between 1 January 2017 and 31 August 2021. For the hourly frequency analysis, the sample period is between 1 January 2021 00:00 UTC and 1 July 2021 00:00 UTC, with an estimation window length of 4 months, i.e. 2,882 hourly returns observations; the forecasting period therefore consists of 1,465 hourly observations, between 1 May 2021 00:00 UTC and 1 July 2021 00:00 UTC. Figure 1 depicts time series of daily log returns for each cryptocurrency. Bitcoin appears to be considerably less volatile than the other currencies, except during the ‘Black Thursday’ crypto market crash on 12 March 2020, and com- mon volatility clusters are often observed simultaneously across all four cryptocurrencies. Figure 2 displays the time series of hourly log returns for each cryptocurrency over the entire sample period January to June 2021. All returns exhibit common volatility clustering and some extreme hourly returns above 10% or below − 10%, as also shown in the minimum and maximum returns in Table 3. Table 2 presents summary statistics for the daily log returns for each cryptocurrency over the sample period 20 August 2015 to 31 August 2021. All cryptocurrencies exhibit a † For instance, the live-streamed implied volatility index for bitcoin typically exceeds 100%, which is more than five times greater than the S&P 500 volatility index, VIX. relatively small mean, in line with our zero-mean assump- tion. The skewness is relatively small and can take either sign. The highly significant positive excess kurtosis justifies our assumption for heavy-tailed distributions, captured using Student-t innovations. Table 3 displays summary statistics for the hourly log returns for each cryptocurrency over the sam- ple period January to June 2021. The mean and skewness are again relatively small. As expected, hourly returns are more volatile than daily returns according to the annualized standard deviation, and excess kurtosis is again positive so a heavy-tailed distribution such as the Student-t should be preferable. We now ask whether the skewness observed in Tables 2 and 3 should warrant the use of the skewed Student-t dis- tribution, such as that defined by Aas and Haff (2006), for the GARCH model innovations. To this end, we esti- mate a univariate skewed Student-t EGARCH model for each cryptocurrency using the entire daily and hourly sam- ple periods and report the parameter estimates an p-values in Tables 4 and 5. Given the entire sample is very large, for both hourly and daily returns, the estimates of most param- eters, including the skewed-t distribution parameter τ , are significant. However, these conditional skewness estimates have a different sign to the unconditional sample skewness observed in Tables 2 and 3. Moreover, the estimates of τ are highly unstable when estimated on rolling sample win- dows, instead of the entire sample. This suggests that the use of a skewed Student-t distribution can produce a misspec- ified model, and because it is important to select a robust and easily-estimated GARCH model against which to judge the performance of the simpler, RiskMetricsTM-type mod- els with ad hoc parameter choices, we do not include the additional source of asymmetry afforded by skew Student-t innovations in our empirical study of asymmetric GARCH models. The asymmetric volatility response refers to a negative correlation between today’s return and tomorrow’s volatil- ity. Wu (2001), Avramov et al. (2006) and many others since show that this is a well-known feature of traditional finan- cial assets. Regarding cryptocurrency returns, the full-sample daily and hourly frequency Student-t EGARCH parameter estimates shown in Tables 6 and 7 indicate at least one sig- nificant volatility response parameter, i.e. θ and γ , as defined in Equation (6). For the daily data, Table 6 shows that the response size parameter θ is small and positive, and only significant for BTC, but the asymmetry parameter γ is signifi- cant for all four cryptocurrencies. For the hourly data, Table 7 shows that the asymmetry parameter γ is again significant for all four cryptocurrencies and the response size parameter θ is large and negative, except for XRP. Based on daily data in Figure 3 and hourly data in Figure 4 we present θ and γ parameter estimates based on a rolling window of 500 observations. Figure 3 shows that the daily frequency estimates for θ are relaively stable over time, but there is more variation in γ . For instance, during the March 2020 Covid crisis, both BTC and ETH daily returns exhib- ited an asymmetric volatility response which changed from day to day, due to extreme volatilty following the huge neg- ative returns of over 30%, on 12 March 2020. The most variable parameter estimates are for XRP, which we ascribe Assessing the accuracy of exponentially weighted moving average models for VaR and ES of Crypto portfolios 15 Figure 1. Daily log returns on bitcoin, ether, ripple and litecoin VWAP USD prices obtained from Cryptocompare. The sample period is 20 August 2015 to 31 August 2021. to a high-profile lawsuit between Ripple Labs and the SEC.† Figure 4 covers a different sample, starting only in May 2021. By this time the cryptocurrency markets had matured consid- erably, and the parameter estimates are much more stable than they are in Figure 3. Again, XRP stands out from the other currencies, with a much smaller θ but a greater γ . Another intuitive visualization of the volatility response uses the news impact curves defined by Engle and Ng (1993). These are depicted in Figure 5 for daily data and Figure 6 for hourly data. Figure 5 shows symmetric news impact curves at the daily frequency, the greatest impact being for XRP, which is to be expected given the extensive media coverage of the † For example, since December 2020 and thus during our entire sam- ple for hourly data, Ripple Labs has been engaged in a high-profile law suit, filed by the U.S. Securities and Exchange Commission (SEC), and still ongoing at the time of writing. There is now a major class action by XRP coin holders claiming that the SEC law suit has had a highly detrimental effect on XRP returns. SEC law suit. Figure 6 shows that negative hourly returns have a more pronounced news impact than positive returns of the same magnitude, except for XRP which is more sym- metric. We conclude that the cryptocurrency risk profile, as indicated by the shape of the news impact curve, varies con- siderably across different cryptocurrencies, as well as between daily and hourly returns. The differences observed here con- firm a highly idiosyncratic volatility response behaviour even within the four major cryptocurrencies. Because we use hourly data, we examine whether the cryptocurrencies in our sample exhibit intra-day volatil- ity periodicity which, if found, can be accounted for as discussed by Andersen and Bollerslev (1997), Stroud and Johannes (2014) and others. To this end, Figure 7 shows the per-hour average in-sample volatility estimates from a uni- variate Student-t EGARCH model, estimated using the entire sample of hourly data for each cryptocurrency. This shows that only slight variations are observed between different 16 C. Alexander and M. Dakos Figure 2. Hourly log returns on bitcoin, ether, ripple and litecoin VWAP USD prices obtained from Cryptocompare. The sample period is 1 January 2021 to 1 July 2021. Table 2. Summary statistics of daily log returns on bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC) VWAP USD prices obtained from Cryptocompare. Table 3. Summary statistics of hourly log returns on bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC) VWAP USD prices obtained from Cryptocompare. BTC ETH XRP LTC BTC ETH XRP LTC Mean (%) St. Dev. (% p.a.) Skewness Ex. Kurtosis Min. Max. 0.242% 77.1% − 0.83 12.07 − 48% 23% 0.360% 121.9% − 0.26 6.81 − 57% 38% 0.228% 143.9% 0.98 11.41 − 54% 62% 0.178% 109.1% 0.54 13.04 − 48% 55% Mean (%) St. Dev. (% p.a.) Skewness Ex. Kurtosis Min. (%) Max. (%) 0.004% 104.5% − 0.23 9.84 − 11% 12% 0.026% 134.3% − 0.80 8.78 − 17% 9% 0.028% 183.6% − 0.28 8.10 − 18% 13% 0.003% 148.4% − 0.57 5.67 − 13% 8% Notes: The sample period is 20 August 2015 to 31 August 2021. The mean, minimum and maximum are expressed in % and the daily standard deviation is annualized using 365. √ Notes: The sample period is 1 January 2021 to 1 July 2021. The mean, minimum and maximum are expressed in % and the hourly standard deviation is annualized using 24 × 365. √ hours-of-day which are very small relative to the overall average volatility levels. More specifically, based on hourly returns bitcoin exhibits an average annualized volatility of 99% but the range of the hour-of-day average volatility between maximum and minimum is only 3.1%. Similarly for ether, the average volatility is 126% and the range is Assessing the accuracy of exponentially weighted moving average models for VaR and ES of Crypto portfolios 17 Table 4. Parameter estimates and p-values (in parentheses) for each cryptocurrency obtained from the robust standard errors of the univariate skewed-Student-t EGARCH model estimated for the entire daily frequency sample period 20 August 2015—31 August 2021. BTC ETH XPR LTC ω β θ γ ν τ − 0.042 (0.44) 0.993 (0.00) 0.048 (0.02) 0.284 (0.00) 2.59 (0.00) 0.93 (0.00) − 0.300 (0.00) 0.946 (0.00) 0.005 (0.80) 0.339 (0.00) 3.28 (0.00) 1.01 (0.00) − 0.291 (0.00) 0.946 (0.00) 0.009 (0.74) 0.445 (0.00) 2.78 (0.00) 1.07 (0.00) − 0.033 (0.00) 0.995 (0.00) 0.033 (0.25) 0.241 (0.00) 2.68 (0.00) 1.03 (0.00) Notes: The first column denotes the model parameters as defined in Equation (6), with τ denoting the Student-t distribution’s skewness parameter, and the remaining columns denote the parameter estimates and p-values for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC). Table 5. Parameter estimates and p-values (in parentheses) for each cryptocurrency obtained from the robust standard errors of the univariate skewed-Student-t EGARCH model estimated for the entire hourly frequency sample period 1 January 2021—1 July 2021. BTC ETH XPR LTC ω β θ γ ν τ − 0.100 (0.00) 0.989 (0.00) − 0.064 (0.00) 0.121 (0.05) 4.40 (0.00) 0.94 (0.00) − 0.104 (0.00) 0.988 (0.00) − 0.043 (0.00) 0.164 (0.00) 4.82 (0.00) 0.93 (0.00) − 0.140 (0.00) 0.983 (0.00) 0.002 (0.87) 0.250 (0.00) 3.75 (0.00) 0.98 (0.00) − 0.148 (0.00) 0.983 (0.00) − 0.046 (0.00) 0.164 (0.00) 5.05 (0.00) 0.95 (0.00) Notes: The first column denotes the model parameters as defined in Equation (6), with τ denoting the Student-t distribution’s skewness parameter, and the remaining columns denote the parameter estimates and p-values for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC). Table 6. Parameter estimates and p-values (in parentheses) for each cryptocurrency obtained from the robust standard errors of the univariate Student-t EGARCH model estimated for the entire daily frequency sample period 20 August 2015—31 August 2021. BTC ETH XPR LTC ω β θ γ ν − 0.042 (0.31) 0.993 (0.00) 0.042 (0.02) 0.270 (0.00) 2.68 (0.00) − 0.302 (0.00) 0.945 (0.00) 0.006 (0.77) 0.341 (0.00) 3.28 (0.00) − 0.308 (0.01) 0.943 (0.00) 0.008 (0.76) 0.469 (0.00) 2.75 (0.00) − 0.033 (0.00) 0.994 (0.00) 0.034 (0.23) 0.242 (0.00) 2.67 (0.00) Note: The first column denotes the model parameters as defined in Equation (6) and the remaining columns denote the parameter estimates and p-values for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC). Table 7. Parameter estimates and p-values (in parentheses) for each cryptocurrency obtained from the robust standard errors of the univariate Student-t EGARCH model estimated for the entire hourly frequency sample period 1 January 2021—1 July 2021. BTC ETH XPR LTC ω β θ γ ν − 0.100 (0.00) 0.989 (0.00) − 0.065 (0.00) 0.121 (0.05) 4.39 (0.00) − 0.105 (0.00) 0.988 (0.00) − 0.045 (0.00) 0.163 (0.00) 4.90 (0.00) − 0.141 (0.00) 0.983 (0.00) 0.002 (0.88) 0.249 (0.00) 3.76 (0.00) − 0.147 (0.00) 0.983 (0.00) − 0.047 (0.00) 0.164 (0.00) 5.08 (0.00) Note: The first column denotes the model parameters as defined in Equation (6) and the remaining columns denote the parameter estimates and p-values for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC). 5.9%, and respectively for ripple 173% and 7.9% and for litecoin 141% and 6.4%. While it may be the case that state- of-the-art volatility models accounting for intra-day volatil- ity periodicity could produce superior results, such complex models may have a limited scope for application by many practitioners, who are likely to favour the simpler models with ad hoc parameter choices that are presented in this paper. 18 C. Alexander and M. Dakos Figure 3. Daily rolling asymmetry parameter estimates for θ (upper panel) and γ (lower panel) as defined in Equation (6) of the univariate Student-t EGARCH model for bitcoin (BTC, in blue), ether (ETH, in red), ripple (XRP, in green) and litecoin (LTC, in gray). The daily out-of-sample period is 1 January 2019—31 August 2021. Figure 4. Hourly rolling asymmetry parameter estimates for θ (upper panel) and γ (lower panel) as defined in Equation (6) of the univariate Student-t EGARCH model for bitcoin (BTC, in blue), ether (ETH, in red), ripple (XRP, in green) and litecoin (LTC, in gray). The hourly frequency out-of-sample period is 1 May 2021—1 July 2021. t (vertical axis) to previous-period returns εt−1 (horizontal axis) for (BTC, Figure 5. News impact curve denoting the volatility response σ 2 in blue), ether (ETH, in red), ripple (XRP, in green) and litecoin (LTC, in gray), obtained from the univariate Student-t EGARCH model estimated for the entire daily frequency sample period 20 August 2015—31 August 2021. Assessing the accuracy of exponentially weighted moving average models for VaR and ES of Crypto portfolios 19 Figure 6. News impact curve denoting the volatility response σ 2 t (vertical axis) to previous-period returns εt−1 (horizontal axis) for (BTC, in blue), ether (ETH, in red), ripple (XRP, in green) and litecoin (LTC, in gray), obtained from the univariate Student-t EGARCH model estimated for the entire hourly frequency sample period 1 January 2021—1 July 2021. Figure 7. Average in-sample volatility per hour-of-day obtained via the Student-t EGARCH model estimated for the entire hourly frequency 24 × 365 and averaged across sample period 1 January 2021—1 July 2021. In-sample estimated volatility is annualized with a factor of the sample for each hour-of-day. √ 5. Empirical results 5.1. Daily forecasts We generate out-of-sample, one-period-ahead analysis for bit- coin, ether, ripple and litecoin daily and hourly log returns, comparing the results of the EWMA-type models and more complex GARCH models against the random walk, first back- testing both left- and right-tail VaR and ES in a univariate setting, followed by the results on the univariate and multi- variate score-based tests of variance and covariance forecast accuracy. The hourly frequency analysis is then presented in the same order. We emphasize that the entire ‘raison d’être’ for this research is to set ad hoc values for the EWMA parame- ters. This is because we seek to argue for (or against) set- ting universal, or at least sector-specific, values for EWMA parameters which are the same for the tens, hundreds (or even thousands) of crypto assets included. This is an impor- tant question at this stage in the development of cryptocur- rency markets, where some type of extended RiskMetricsTM product needs thorough and independent examination. Insti- tutional in exploring investment opportunities in a wide variety of coins and tokens is increasing rapidly at the time of writing. However, our parameter selec- tions also need to be sensible. Setting a decay parame- ter less than about 0.9 would probably be unacceptable to regulators, and the asymmetry parameter should not be so large as to dominate the influence that the realized returns have on volatility. With these comments in mind, we describe how model parameters are selected in more detail below. interest Variance and covariance predictions are made daily between 1 January 2017 and 31 August 2021, i.e. for 1,704 daily obser- vations. The benchmark model employs an equally-weighted 30-day moving average. We select precisely 30 days because, although ad hoc, this time period is very commonly used by practitioners and is probably the best suited choice for a sin- gle benchmark model. EWMA and AEWMA volatilities and covariances are calculated using two possible values for λ, the RiskMetricsTM value 0.94 and a smaller value of 0.925; the AEWMA model further introduces an asymmetry parameter η which is set to 1%, 2% and 3% for left-tail (long posi- tions) VaR/ES forecasting and − 1%, − 3% and − 5% for the right tail (short positions). Results for other choices of η are presented in the Appendix. For instance, Table A5 explores the effect of η on right-tail daily VaR results for short posi- tions and Table A6 does the same, but for left-tail hourly VaR results, for long positions.† The univariate GARCH and mul- tivariate DCC model parameters are estimated using MLE on a fixed-size rolling window with 500 daily log returns, with model parameters are updated on a daily basis. It is important to note the trade-off between the length of the in-sample/rolling estimation period and the out-of- sample/forecasting period. A longer estimation period might improve the performance of GARCH models for which † As mentioned previously, note that the ad hoc value choices for the asymmetry parameter should be large enough to improve forecasting accuracy but also small enough to avoid a dominating influence of realized returns on volatility. 20 C. Alexander and M. Dakos Table 8. Backtesting results for one-day-ahead left-tail 1% and 2.5% VaR forecasts for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC), based on an out-of-sample period between 1 January 2017—31 August 2021. Table 9. Backtesting results for one-day-ahead right-tail 1% and 2.5% VaR forecasts for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC), based on an out-of-sample period between 1 January 2017—31 August 2021. Assessing the accuracy of exponentially weighted moving average models for VaR and ES of Crypto portfolios 21 Table 10. Backtesting results for one-day-ahead left-tail 1% and 2.5% ES forecasts for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC), based on an out-of-sample period between 1 January 2017—31 August 2021. parameters are estimated via MLE; however, this produces a significantly smaller out-of-sample period and allows for fewer sample points in the forecast accuracy tests. For instance, we have attempted an alternative specification for the GARCH models using a 1,240-day rolling estimation window, but this leaves fewer than 1,000 out-of-sample obser- vations between 1 January 2019 and 31 August 2021. The parameter estimates obtained using the longer estimation win- dow are slightly more stable, but the forecasting results are very similar to those reported here, which use a 500-day estimation window. Regarding the distribution assumptions, in the random walk benchmark model, crypto cryptocurrency returns are assumed to follow a zero-mean normal distribution. In the EWMA and AEWMA models, a zero-mean location-scale transformed Student-t distribution is used with ad hoc ν = 6 degrees of freedom, to produce a heavy-tailed distribution; similarly, a multivariate Student-t with ν = 6 is assumed for the joint distribution of bitcoin, ether, ripple and litecoin returns. The GARCH and DCC models also assume a univariate and multi- variate Student-t distribution, respectively, where the degrees of freedom parameter is estimated jointly with the model parameters based on the 500-day rolling estimation win- dow.† GARCH and DCC model estimations and forecasts † The daily returns exhibit some asymmetry in their sample statistics, but not always in the same direction, as shown in Table 2. But the asymmetric model specifications can still capture this characteristic to some extent even with the symmetric distribution assumption that we use. Also, the choice of a symmetric distribution simplifies the and also some of the VaR and ES backtesting methods are implemented using the rugarch and rmgarch R packages of Ghalanos (2019, 2020).‡ 5.1.1. VaR and ES backtests. We generate VaR and ES one- day-ahead forecasts for bitcoin, ether and ripple at the 1%, 2.5% and 5% significance levels for both the left and right tail of the returns distribution, to assess the ability of each model to capture risk that large negative daily returns present to long positions adequately—and also its ability to capture the risk that large positive daily returns present to short posi- tions. The accuracy of forecast generated by each model is assessed using the tests described in Section 3.2. The results for 1% and 2.5% are presented in Table 8 for the left tail (long positions) and in Table 9 for the right tail (short positions). The results for 5% are less important—at least from the point of view of market regulation—and so are left to the Appendix. These results are consistent with most of the findings in the relevant literature as discussed in Section 2. Cata- nia and Grassi (2021) examine GAS model specifications against an EGARCH benchmark for 606 cryptocurrencies process of modelling the joint density for the returns of bitcoin, ether, ripple and litecoin. ‡ The multivariate EWMA model is implemented using the RiskPortfolios R package of Ardia et al. (2017), and the remain- ing EWMA and AEWMA specifications and also the traffic light backtesting methodology are implemented using custom-written R code. 22 C. Alexander and M. Dakos Table 11. Backtesting results for one-day-ahead right-tail 1% and 2.5% ES forecasts for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC), based on an out-of-sample period between 1 January 2017—31 August 2021. Table 12. Average CRPS of one-day-ahead univariate density forecasts for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC) daily log returns, based on an out-of-sample period between 1 January 2017—31 August 2021. Benchmark tEWMA(0.925) tEWMA(0.94) tAEWMA(0.925, 1%) tAEWMA(0.925, 2%) tAEWMA(0.925, 3%) tAEWMA(0.925, − 1%) tAEWMA(0.925, − 3%) tAEWMA(0.925, − 5%) tAEWMA(0.94, 1%) tAEWMA(0.94, 2%) tAEWMA(0.94, 3%) tAEWMA(0.94, − 1%) tAEWMA(0.94, − 3%) tAEWMA(0.94, − 5%) tGARCH tEGARCH BTC 0.02226 98.92% 98.96% 99.18% 99.74% 100.89% 98.99% 101.17% 107.00% 99.23% 99.80% 100.97% 99.04% 101.22% 107.04% 98.85% 98.70% ETH 0.03010 98.79% 98.87% 98.87% 99.07% 99.50% 98.75% 99.52% 102.19% 98.96% 99.17% 99.62% 98.83% 99.61% 102.27% 97.99% 97.95% XRP 0.03436 98.78% 98.84% 98.76% 99.02% 99.62% 98.69% 99.39% 101.75% 98.80% 99.09% 99.70% 98.72% 99.45% 101.84% 97.57% 97.69% LTC Energy Score 0.03141 99.04% 99.06% 99.08% 99.27% 99.75% 99.04% 99.72% 102.10% 99.10% 99.32% 99.83% 99.06% 99.76% 102.15% 98.79% 98.60% 0.06727 98.47% 98.57% 98.93% 99.18% 99.66% 98.88% 99.57% 102.07% 99.02% 99.23% 99.73% 98.90% 99.62% 102.16% 98.35% 102.19% Notes: As the CRPS and energy score are negatively oriented and produce positive values, a relative score lower than 100% indicates outperformance of the benchmark. For the random walk benchmark model the average CRPS is reported outright and the average scores of the remaining models are expressed as a percentage of the benchmark’s score. The EWMA, AEWMA and GARCH models are based on a Student-t distribution assumption and the degrees of freedom for EWMA and AEWMA are set to the ad hoc ν = 6 and for GARCH the degrees of freedom are estimated via MLE. The right-most column reports the average energy scores for one-day-ahead joint density forecasts of all four coins. For the tGARCH and tEGARCH models, the energy scores refer to the ADCC multivariate model. The relative energy scores for the DCC multivariate model are 98.43% for tGARCH and 102.14% for tEGARCH. Assessing the accuracy of exponentially weighted moving average models for VaR and ES of Crypto portfolios 23 Table 13. Average log score of one-day-ahead univariate density forecasts for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC) daily log returns. Benchmark tEWMA(0.925) tEWMA(0.94) tAEWMA(0.925, 1%) tAEWMA(0.925, 2%) tAEWMA(0.925, 3%) tAEWMA(0.925, − 1%) tAEWMA(0.925, − 3%) tAEWMA(0.925, − 5%) tAEWMA(0.94, 1%) tAEWMA(0.94, 2%) tAEWMA(0.94, 3%) tAEWMA(0.94, − 1%) tAEWMA(0.94, − 3%) tAEWMA(0.94, − 5%) tGARCH tEGARCH BTC − 1.62193 114.55% 114.53% 113.92% 113.09% 111.41% 114.58% 111.13% 103.71% 113.87% 112.99% 111.28% 114.47% 111.00% 103.63% 115.03% 115.56% ETH − 1.42004 108.55% 108.46% 108.33% 108.12% 107.50% 108.81% 107.72% 103.33% 108.24% 107.96% 107.29% 108.71% 107.55% 103.18% 110.47% 110.54% XRP − 1.26143 117.21% 117.20% 117.47% 117.06% 115.72% 117.77% 116.67% 111.53% 117.46% 116.94% 115.53% 117.79% 116.46% 111.26% 121.21% 121.30% LTC − 1.32140 114.34% 114.35% 114.34% 114.12% 113.23% 114.51% 113.49% 109.16% 114.31% 113.98% 113.02% 114.50% 113.36% 109.01% 115.38% 116.13% MV LogS − 6.52492 124.45% 124.60% 123.11% 122.67% 121.90% 123.28% 122.01% 119.54% 123.42% 122.98% 122.22% 123.62% 122.36% 119.94% 125.15% 120.68% Notes: For the random walk benchmark model the average log score is reported outright and the average scores of the remaining models are expressed as a percentage of the benchmark’s score. As the negatively oriented version of the log score is used which produces negative values, a relative score higher than 100% indicates outperformance of the benchmark. The EWMA, AEWMA and GARCH models are based on a Student-t distribution assumption and the degrees of freedom for EWMA and AEWMA are set to the ad hoc ν = 6 and for GARCH the degrees of freedom are estimated via MLE. The right-most column reports the average multivariate log scores (MV LogS) for one-day-ahead joint density forecasts of all four coins. For the tGARCH and tEGARCH models, the multivariate log scores refer to the ADCC multivariate model. The relative multivariate log scores for the DCC multivariate model are 125.08% for tGARCH and 120.63% for tEGARCH. Table 14. Backtesting results for one-hour-ahead left-tail 1% and 2.5% VaR forecasts for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC). For each asset, the first column denotes the models used, where the EWMA, AEWMA and GARCH models are based on a Student-t distribution assumption and the degrees of freedom for EWMA and AEWMA are set to the ad hoc ν = 6 and for GARCH the degrees of freedom are estimated via MLE. 24 C. Alexander and M. Dakos Table 15. Backtesting results for one-day-ahead left-tail 1% and 2.5% VaR forecasts for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC), based on an out-of-sample period between 1 January 2017—31 August 2021. Table 16. Backtesting results for one-hour-ahead left-tail 1% and 2.5% ES forecasts for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC). Assessing the accuracy of exponentially weighted moving average models for VaR and ES of Crypto portfolios 25 Table 17. Backtesting results for one-hour-ahead right-tail 1% and 2.5% ES forecasts for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC). and find that the additional modelling complexity introduced by the GAS framework ‘pays off’ for 5% and 1% ES and 5% VaR with increased accuracy, but less so for 1% VaR. In that respect, the results presented in this section for one- day-ahead VaR and ES forecast backtesting are somewhat in agreement with Catania and Grassi (2021) in that intro- ducing additional modelling complexity may sometimes ‘pay off’ in increased forecasting accuracy, especially at lower significance levels; however, we often find that AEWMA specifications are on par with EGARCH in terms of VaR and ES forecasting accuracy even at the 1% significance level. 5.1.2. Score-based tests for variance. We now present the results on univariate and multivariate scores to rank the accu- racy of each model for forecasting the one-day-ahead volatil- ity and covariance of bitcoin, ether, ripple and litecoin log returns. The continuous ranked probability score (CRPS) and negatively oriented univariate log score are used to assess univariate density forecasts and joint distribution forecasts are evaluated with the energy score and again with the neg- atively oriented multivariate log score. Given the paramet- ric distribution assumptions in the models used, i.e. normal for the random walk model and Student-t for the EWMA and GARCH specifications, the one-day-ahead volatility and covariance forecasts fully define the one-day-ahead distribu- tion of log returns for each cryptocurrency and also their joint distribution, allowing for the scores’ calculation. All univariate and multivariate scores are calculated using the scoringRules R package of Jordan et al. (2019) combined with custom-written R code. The CRPS is calculated every day by comparing the real- ized return with the one-day-ahead log returns density forecast produced by each model and in order to rank the models we average their CRPS over the entire forecasting period. Table 12 reports the results expressed as a percentage of benchmark model’s average score which is given in the first row. Due to their negative orientation, relative scores below 100% suggest the possibility that the model is more accu- rate than the benchmark. We also use the individual scores for every day from each model to perform pair-wise comparisons of forecasting accuracy, by calculating the tN test statistic of Gneiting and Ranjan (2011) for the hypothesis test of equal forecasting performance as per Equation (30) in Section 3. As shown in Table 12, most models outperform the bench- mark although none of them achieve an average CRPS lower than 97% and the tN test for a significant difference between the highest and lowest average scores is always below 0.15, so the null hypothesis of equal forecasting performance is never rejected at 5%. This suggests that all models exam- ined, including our very simple benchmark, produce equally inaccurate one-day-ahead density forecasts for the returns of bitcoin, ether, ripple and litecoin. This result extends the findings of Catania and Grassi (2021) that equal forecast- ing performance between an EGARCH benchmark model and more complex GAS models is the most common outcome. The energy score is calculated by drawing 10,000 random samples from the forecast joint density of log returns pro- duced by each model, based on the corresponding realized 26 C. Alexander and M. Dakos Table 18. Average CRPS of one-hour-ahead univariate density forecasts for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC) hourly log returns. Benchmark tEWMA(0.925) tEWMA(0.94) tAEWMA(0.925, 0.2%) tAEWMA(0.925, 0.7%) tAEWMA(0.925, 0.8%) tAEWMA(0.925, 0.9%) tAEWMA(0.925, − 0.2%) tAEWMA(0.94, 0.2%) tAEWMA(0.94, 0.7%) tAEWMA(0.94, 0.8%) tAEWMA(0.94, 0.9%) tAEWMA(0.94, − 0.2%) tGARCH tEGARCH BTC ETH XRP LTC Energy Score 0.00613 0.00858 0.00920 0.00890 0.01788 99.25% 99.19% 99.15% 99.85% 100.32% 100.94% 99.36% 99.10% 99.87% 100.35% 100.97% 99.30% 98.91% 98.56% 99.13% 99.11% 99.04% 99.01% 99.12% 99.30% 99.20% 99.02% 99.04% 99.17% 99.35% 99.17% 98.82% 98.63% 98.44% 98.43% 98.40% 98.70% 98.89% 99.13% 98.49% 98.40% 98.74% 98.94% 99.19% 98.49% 98.04% 98.10% 99.25% 99.22% 99.16% 99.16% 99.28% 99.46% 99.32% 99.14% 99.20% 99.33% 99.52% 99.28% 98.89% 98.73% 98.67% 98.69% 98.90% 99.06% 99.27% 99.46% 99.05% 98.88% 99.14% 99.21% 99.50% 99.03% 98.68% 98.62% Notes: For the random walk benchmark model the average CRPS is reported outright and the average scores of the remaining models are expressed as a percentage of the benchmark’s score. As the CRPS and energy score are negatively oriented and produce positive values, a relative score lower than 100% indicates outperformance of the benchmark. The EWMA, AEWMA and GARCH models are based on a Student-t distribution assumption and the degrees of freedom for EWMA and AEWMA are set to the ad hoc ν = 6 and for GARCH the degrees of freedom are estimated via MLE. The right-most column reports the average energy scores for one-day-ahead joint density forecasts of all four coins. For the tGARCH and tEGARCH models, the energy scores refer to the ADCC multivariate model. The relative energy scores for the DCC multivariate model are 98.71% for tGARCH and 98.60% for tEGARCH. Table 19. Average log score of one-hour-ahead univariate density forecasts for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC) hourly log returns. Benchmark tEWMA(0.925) tEWMA(0.94) tAEWMA(0.925, 0.2%) tAEWMA(0.925, 0.7%) tAEWMA(0.925, 0.8%) tAEWMA(0.925, 0.9%) tAEWMA(0.925, − 0.2%) tAEWMA(0.94, 0.2%) tAEWMA(0.94, 0.7%) tAEWMA(0.94, 0.8%) tAEWMA(0.94, 0.9%) tAEWMA(0.94, − 0.2%) tGARCH tEGARCH BTC ETH XRP LTC MV LogS − 3.06639 102.49% 102.61% 102.67% 102.17% 101.76% 101.24% 102.42% 102.75% 102.15% 101.72% 101.20% 102.53% 103.08% 103.39% − 2.74713 102.85% 102.91% 102.99% 103.11% 102.97% 102.77% 102.80% 103.03% 103.05% 102.91% 102.70% 102.86% 103.28% 103.48% − 2.67047 104.16% 104.25% 104.27% 104.01% 103.77% 103.48% 104.13% 104.34% 103.97% 103.73% 103.42% 104.21% 104.80% 104.81% − 2.71165 − 13.18273 102.88% 102.94% 103.04% 103.10% 102.93% 102.70% 102.83% 103.07% 103.03% 102.85% 102.61% 102.89% 103.39% 103.57% 104.44% 104.68% 104.17% 103.97% 103.82% 103.65% 104.11% 104.41% 104.17% 104.03% 103.86% 104.37% 104.97% 104.91% Notes: For the random walk benchmark model the average log score is reported outright and the average scores of the remaining models are expressed as a percentage of the benchmark’s score. As the negatively oriented version of the log score is used which produces negative values, a relative score higher than 100% indicates outperformance of the benchmark. The EWMA, AEWMA and GARCH models are based on a Student-t distribution assumption and the degrees of freedom for EWMA and AEWMA are set to the ad hoc ν = 6 and for GARCH the degrees of freedom are estimated via MLE. The right-most column reports the average multivariate log scores (MV LogS) for one-day-ahead joint density forecasts of all four coins. For the tGARCH and tEGARCH models, the energy scores refer to the ADCC multivariate model. The relative log scores for the DCC multivariate model are 104.97% for tGARCH and 104.91% for tEGARCH. returns.† Again for comparison purposes, each score is aver- aged over the entire forecasting period. The right-hand col- umn of Table 12 reports the outright average energy score for the benchmark model and the average energy scores of all other models are again expressed relative to the benchmark’s † These are produced using the mvnfast R package of Fasiolo (2016). average score. Again, most models outperform the benchmark but the tN test statistic for equal forecasting performance can- not reject the null hypothesis even at 5% and even when test- ing the model having lowest energy score versus the bench- mark. So all models produce equally accurate one-day-ahead covariance forecasts—none of the multivariate parametric models perform any better than the simplest benchmark. Assessing the accuracy of exponentially weighted moving average models for VaR and ES of Crypto portfolios 27 As with the CRPS and energy scores, the univariate and multivariate log scores are calculated via the correspond- ing one-day-ahead predictive densities based on the realized returns and the volatility and covariance forecasts produced by each model. Again, the average log score is expressed outright for the benchmark in Table 13 and relative to the benchmark score for all other models. Table 13 shows that all models produce a higher relative score compared with the benchmark, contrary to the findings for the CRPS and energy score. However, this is because we employ the negatively ori- ented log score which produces negative average scores in all cases; consequently, a lower (i.e. more negative) log score is higher in magnitude but still indicates a better forecast. There- fore, the results shown in Table 13 are consistent with the findings for the CRPS and energy scores, in that all models outperform the benchmark. Moreover, the tN test statistic for equal forecasting performance applied to the log scores can- not reject the null hypothesis that the models produce the same scores, even at 5%. In other words, based on the logarithmic score results and well as the CRPS and energy score results, all models produce equally accurate one-day-ahead variance forecasts and similarly, none of the multivariate parametric models perform any better than the benchmark. 5.2. Hourly forecasts We now present the same analysis performed on a more recent sample of hourly bitcoin, ether, ripple and litecoin log returns between 1 May 2021 00:00 UTC and 1 July 2021 00:00 UTC. The benchmark model forecasts are based on an equally- weighted 72-hour moving average; the EWMA and AEWMA forecasts are again produced with λ set to 0.925 and 0.94 but the AEWMA η parameter is now set to lower values i.e. to 0.7%, 0.8% and 0.9% for long positions and − 0.2% and 0.2% for short positions, to be in line with smaller hourly returns.† GARCH models are calibrated using a rolling esti- mation window of 4 months, i.e. 2,882 hourly observations. The same distribution assumptions as for the daily frequency analysis are followed, i.e. we use normal innovations for the benchmark, a Student-t distribution with ν = 6 for the EWMA and AEWMA models, and for the GARCH models the degrees of freedom parameter is estimated with the other model parameters via MLE. 5.2.1. VaR and ES backtesting. The hourly 1% left-tail (long position) VaR forecast results show in Table 14, sug- gest that the AEWMA models are the only ones that produce fewer than the expected 14.65 exceedances for 1% VaR.‡ For these models the CC test’s null hypothesis of no clus- tering in exceedences is accepted even at the 10% level of significance. By comparison, the tEGARCH models are in the yellow zone and the standard tGARCH forecasts are in fact not considered accurate at all. The right-tail (short posi- tion) hourly 1% VaR backtesting results which are presented † Additional testing for other AEWMA η parameter choices is reported in the Appendix. ‡ The expected number of VaR exceedances for 1% VaR is calculated based on (17) using N = 1, 465 and α = 1%. in Table 15 appear to favour the simple model specifica- tions even more; for instance, the 1% right-tail VaR forecasts produced by the symmetric tEWMA models are considered accurate based on both the traffic light and CC tests, and even the simplest random walk benchmark model produces accu- rate 1% VaR forecasts for bitcoin and ripple, which is on a par with the more complex GARCH specifications. 5.2.2. Score-based tests for variance. The average CPRS based on the hourly volatility forecasts for each cryptocur- rency, and the energy score for the covariance forecasts, are presented in Table 18. Again, the average score for the bench- mark is given in the first row and the subsequent rows present each model’s average score relative to the benchmark score. Again, the benchmark yields a higher average score (indicat- ing a poorer performance) than almost every other model. The tEGARCH model yields the lowest average CRPS for all cryptocurrencies except ripple and the ADCC-tGARCH yields the lowest average energy score. Importantly, and simi- lar to the discussion of our daily frequency analysis, the test of equal forecasting performance does not distinguish between the models: the tN test statistic between the highest and low- est average scores is always below 0.45, so the null hypothesis of equal forecasting performance is accepted e.g. at the 5% and 10% significance levels based on both the CRPS for each cryptocurrency and on the energy score.§ This suggests that the equal forecasting performance identified previously based on the average scores of daily returns density forecasts also holds at the hourly frequency: i.e. all models examined, even the random walk benchmark, are considered equally accurate regarding one-hour-ahead volatility and covariance forecasts for the returns of bitcoin, ether, ripple and litecoin. Similarly, the average univariate and multivariate log scores are shown in Table 19. The average log scores of all models are greater in magnitude than the benchmark’s aver- age score which, as mentioned previously, indicates a better performance than the benchmark due to the negative sign of all average scores. As before, the test of equal forecasting per- formance indicates an equal performance across all models, for all assets and also for the multivariate log score, even at the 5% and 10% significance levels. It therefore appears that both the CPRS/energy score and univariate/multivariate log scores indicate an equal forecasting performance in all cases, for hourly data just as they do for daily data. 6. Conclusions The standard RiskMetricsTM EWMA methodology fails to provide the industry standard of 99% coverage for VaR and ES of cryptocurrencies, at least when the decay parameter λ is 0.94, this being equal to the value set for returns on tra- ditional financial assets. However, a lower value than 0.94 § As discussed in Section 3, the null hypothesis for the test of equal forecasting performance is that tN = 0, tested against the two-sided alternative that tN (cid:11)= 0, where tN ∼ N (0, 1). Therefore, if the null hypothesis is to be rejected e.g. at the 5% or 10% significance levels, then tN should be outside the 2.5% or 5% right- and left-tail critical values which are respectively ±1.96 and ±1.64. 28 C. Alexander and M. Dakos may be appropriate for cryptocurrencies because their volatil- ity tends to be more highly reactive to market shocks and also less persistent, and reducing λ has precisely this effect. Hav- ing said this, we have not been able to produce the required level of coverage with other values for λ, lower, or higher than 0.94, even when innovations are drawn from a Student-t distribution. What is important is that the EWMA model includes an asymmetry in volatility response. That is, volatility increases more following a negative return than it does following a pos- itive return of the same magnitude. It is simple to do this by introducing another parameter η to the EWMA model, and with this modification we refer to the model as AEWMA, or more precisely as the AEWMA(λ, η, ν) model, where ν is the degrees of freedom in the Student-t innovations. We explore the values of λ, η and ν that do provide 99% cover- age, with the same parameter values for all cryptocurrencies. This way, we may also use these values in a multivariate framework without concern for non-positive definiteness of the covariance matrix. Our empirical results show that this simple, constant parameter AEWMA model performs as well as the fully- calibrated Student-t EGARCH models, which require at least several hundred observations in-sample, before any back- testing begins—and even more observations are required for calibration of their multivariate equivalents if these are to be employed in portfolio optimization. Running several state-of-the-art VaR and Expected Shortfall backtests over a five-year period at the daily frequency, and using several thousands of observations at the hourly frequency, we find that a coverage level of 99% for VaR and ES is easy to achieve for long or short positions on all cryptocurrencies considered is possible, provided we have the right param- eter choices. Although a small improvement in coverage is obtainable using sophisticated GARCH models, the incre- mental returns to this effort are unlikely to be seen as attractive by industry practitioners. Also our univariate and multivariate score-driven tests for variance and covariance indicate no sig- nificant improvements in average scores which would warrant a ranking for GARCH models above our simple asymmetric extension for a RiskMetricsTM type EWMA. Disclosure statement No potential conflict of interest was reported by the author(s). ORCID Carol Alexander http://orcid.org/0000-0003-1247-0184 References Aas, K. and Haff, I.H., The generalized hyperbolic skew Student-t distribution. J. Financ. Econom., 2006, 4(2), 275–309. Acereda, B., Leon, A. and Mora, J., Estimating the expected shortfall of cryptocurrencies: An evaluation based on backtesting. Finance Res. Lett., 2020, 33, 101181. Aielli, G.P., Dynamic conditional correlation: On properties and estimation. J. Bus. Econ. Stat., 2013, 31(3), 282–299. Al-Khazali, O., Bouri, E. and Roubaud, D., The impact of positive and negative macroeconomic news surprises: Gold versus bitcoin. Econ. Bulletin, 2018, 38(1), 373–382. Al-Yahyaee, K., Mensi, W., Al-Jarrah, I., Hamdi, A. and Kang, S., Volatility forecasting, downside risk, and diversification benefits of bitcoin and oil and international commodity markets: A com- parative analysis with yellow metal. North American J. Econ. Finance, 2019, 49, 104–120. Alexander, C., Market Risk Analysis: Value-at-Risk Models, 2008 (John Wiley and Sons: Chichester). Alexander, C. and Dakos, M., A critical investigation of cryp- tocurrency data and analysis. Quant. Finance, 2020, 20(2), 173– 188. Andersen, T.G. and Bollerslev, T., Intraday periodicity and volatility persistence in financial markets. J. Empir. Finance, 1997, 4(2-3), 115–158. Ardia, D., Bluteau, K. and Rüede, M., Regime changes in bitcoin GARCH volatility dynamics. Finance Res. Lett., 2019, 29, 266– 271. Ardia, D., Boudt, K. and Gagnon-Fleury, J.-P., RiskPortfolios: Com- putation of risk-based portfolios in R. J. Open Source Softw., 2017, 2(10), 171. Avramov, D., Chordia, T. and Goyal, A., The impact of trades on daily volatility. Rev. Financ. Stud., 2006, 19(4), 1241–1277. Basel Committee, Supervisory framework for the use of backtesting in conjunction with the internal models approach to market risk capital requirements. Basel Committee on Banking Supervision, 1996. Available at: https://www.bis.org/publ/bcbs22.htm. Basel Committee, Fundamental review of the trading book: A revised market risk framework. Basel Committee on Banking Supervision, 2012. Available at: https://www.bis.org/publ/bcbs2 19.htm. Baur, D., Dimpfl, T. and Kuck, K., Bitcoin, gold and the US dollar: A replication and extension. Finance Res. Lett., 2018, 25, 103– 110. Bauwens, L. and Storti, G., A component GARCH model with time varying weights. Stud. Nonlinear Dyn. Econom., 2009, 13(2), 1– 33. Bayer, S. and Dimitriadis, T., Regression-based expected shortfall backtesting. J. Financ. Econom., 2020, 20, 437–471. Bazán-Palomino, W., Bitcoin and its offspring: A volatility risk approach. In Advanced Studies of Financial Technologies and Cryptocurrency Markets, edited by L. Pichl, C. Eom, E. Scalas and T. Kaizoji, pp. 233–256, 2020 (Springer: Singapore). Bollerslev, T., Generalized autoregressive conditional heteroskedas- ticity. J. Econom., 1986, 31(3), 307–327. Bonello, A. and Suda, D., Volatility regime analysis of bitcoin price dynamics using Markov-switching GARCH models. In Proceedings of the European Simulation and Modelling Confer- ence 2018, pp. 213–218, 2018 (Department of Industrial Sys- tems Engineering and Product Design, University of Gent: Gent). https://biblio.ugent.be/publication/8593929 Bouoiyour, J. and Selmi, R., Commodity price uncertainty and man- ufactured exports in Morocco and Tunisia: Some insights from a novel GARCH model. Econ. Bulletin, 2014, 34(1), 220–233. Bouoiyour, J. and Selmi, R., Bitcoin: A beginning of a new phase? Econ. Bulletin, 2016, 36(3), 1430–1440. Bouri, E., Molnar, P., Azzi, G., Roubaud, D. and Hagfors, L., On the hedge and safe haven properties of bitcoin: Is it really more than a diversifier? Finance Res. Lett., 2017, 20, 192–198. Bovaird, C., Bitcoin price volatility reached a 14-month high in June. Forbes, 2021. Available at: https://bit.ly/3zLZtXw. Caporale, G. and Zekokh, T., Modelling volatility of cryptocurrencies using Markov-switching GARCH models. Res. Int. Bus. Finance, 2019, 48, 143–155. Cappiello, L., Engle, R.F. and Sheppard, K., Asymmetric dynamics in the correlations of global equity and bond returns. J. Financ. Econom., 2006, 4(4), 537–572. Catania, L. and Grassi, S., Forecasting cryptocurrency volatility. Int. J. Forecast., 2021, 38, 878–894. Assessing the accuracy of exponentially weighted moving average models for VaR and ES of Crypto portfolios 29 Catania, L., Grassi, S. and Ravazzolo, F., Forecasting cryptocurren- cies under model and parameter instability. Int. J. Forecast., 2019, 35(2), 485–501. Chen, C.W., Watanabe, T. and Lin, E.M., Bayesian estimation of realized GARCH-type models with application to financial tail risk management. Econom. Statistics, 2021. Article in Press. https://www.sciencedirect.com/science/article/pii/S245230622 1000484 Christoffersen, P., Evaluating interval forecasts. Int. Econ. Rev. (Philadelphia), 1998, 39(4), 841–862. Chu, J., Chan, S., Nadarajah, S. and Osterrieder, J., GARCH mod- elling of cryptocurrencies. J. Risk Financ. Manage., 2017, 10(4), 17. Clift, S., Costanzino, N. and Curran, M., Empirical performance of backtesting methods for expected shortfall. SSRN Discussion Paper 2618345, 2016. Gneiting, T. and Ranjan, R., Comparing density forecasts using threshold- and quantile-weighted scoring rules. J. Bus. Econ. Stat., 2011, 29(3), 411–422. Guesmi, K., Saadi, S., Abid, I. and Ftiti, Z., Portfolio diversifica- tion with virtual currency: Evidence from bitcoin. Int. Rev. Financ. Anal., 2019, 63, 431–437. Hansen, P., Lunde, A. and Nason, J., The model confidence set. Econometrica, 2011, 79(2), 453–497. Hansen, P.R., Huang, Z. and Shek, H.H., Realized GARCH: A joint model for returns and realized measures of volatility. J. Appl. Econom., 2012, 27(6), 877–906. Hattori, T., A forecast comparison of volatility models using realized volatility: Evidence from the bitcoin market. Appl. Econ. Lett., 2020, 27(7), 591–595. Hentschel, L., All in the family: Nesting symmetric and asymmetric GARCH models. J. Financ. Econ., 1995, 39(1), 71–104. Corbet, S., Lucey, B., Peat, M. and Vigne, S., Bitcoin futures: What Higgins, M.L. and Bera, A.K., A class of non-linear ARCH models. use are they? Econ. Lett., 2018, 172, 23–27. Int. Econ. Rev. (Philadelphia), 1992, 33(1), 137–158. Costanzino, N. and Curran, M., Backtesting general spectral risk measures with application to expected shortfall. SSRN Discussion Paper 2514403, 2015. Jordan, A., Krüger, F. and Lerch, S., Evaluating probabilistic fore- casts with scoringRules. J. Stat. Softw., 2019, 90(12), 1–37. Katsiampa, P., Volatility estimation for bitcoin: A comparison of Costanzino, N. and Curran, M., A simple traffic light approach to GARCH models. Econ. Lett., 2017, 158, 3–6. backtesting expected shortfall. Risks, 2018, 6(1), 2. Köchling, G., Schmidtke, P. and Posch, P., Volatility forecasting Creal, D., Koopman, S.J. and Lucas, A., Generalized autoregressive score models with applications. J. Appl. Econom., 2013, 28(5), 777–795. Diebold, F.X. and Mariano, R.S., Comparing predictive accuracy. J. Bus. Econ. Stat., 1995, 13(3), 253–263. Ding, Z., Granger, C. and Engle, R., A long memory property of stock market returns and a new model. J. Empir. Finance, 1993, 1(1), 83–106. Dyhrberg, A., Bitcoin, gold and the dollar—A GARCH volatility analysis. Finance Res. Lett., 2016, 16, 85–92. Engle, R., Dynamic conditional correlation: A simple class of mul- tivariate generalized autoregressive conditional heteroskedasticity models. J. Bus. Econ. Stat., 2002, 20(3), 339–350. Engle, R., Ghysels, E. and Sohn, B., Stock market volatility and macroeconomic fundamentals. Rev. Econ. Stat., 2013, 95(3), 776– 797. Engle, R. and Manganelli, S., CAViaR: Conditional autoregressive value at risk by regression quantiles. J. Bus. Econ. Stat., 2004, 22(4), 367–381. Engle, R.F., Focardi, S.M. and Fabozzi, F.J., ARCH/GARCH models in applied financial econometrics. In Encyclopedia of Financial Models, edited by F.J. Fabozzi, 2012 (John Wiley and Sons). https://doi.org/10.1002/9781118182635.efm0062 Engle, R.F. and Kroner, K.F., Multivariate simultaneous generalized ARCH. Econ. Theory., 1995, 11(1), 122–150. Engle, R.F. and Lee, G., A Long-Run and Short-Run Component Model of Stock Return Volatility, 1999 (Oxford University Press: Oxford). Engle, R.F. and Ng, V.K., Measuring and testing the impact of news on volatility. J. Finance, 1993, 48(5), 1749–1778. Fantazzini, D. and Zimin, S., A multivariate approach for the simul- taneous modelling of market risk and credit risk for cryptocurren- cies. J. Ind. Bus. Econ., 2020, 47(1), 19–69. Fasiolo, M., An introduction to mvnfast. R package version 0.1.6., 2016. Fissler, T. and Ziegel, J.F., Higher order elicitability and Osband’s principle. Ann. Statist., 2016, 44(4), 1680–1707. Ghalanos, A., Rmgarch: Multivariate GARCH models. R package version 1.3-7., 2019. Ghalanos, A., Rugarch: Univariate GARCH models. R package version 1.4-4., 2020. Glosten, L.R., Jagannathan, R. and Runkle, D.E., On the rela- tion between the expected value and the volatility of the nom- inal excess return on stocks. J. Finance, 1993, 48(5), 1779– 1801. Gneiting, T. and Raftery, A.E., Strictly proper scoring rules, pre- diction, and estimation. J. Am. Stat. Assoc., 2007, 102(477), 359–378. accuracy for bitcoin. Econ. Lett., 2020, 191, 108836. Kratz, M., Lok, Y.H. and McNeil, A.J., Multinomial VaR backtests: A simple implicit approach to backtesting expected shortfall. J. Bank. Finance, 2018, 88, 393–407. Kupiec, P., Techniques for verifying the accuracy of risk measure- ment models. J. Derivatives, 1995, 3(2), 73–84. Liu, W., Semeyutin, A., Lau, C.K.M. and Gozgor, G., Forecasting value-at-risk of cryptocurrencies with riskmetrics-type models. Res. Int. Bus. Finance, 2020, 54, 101259. Longerstaey, J. and Spencer, M., RiskMetrics—Technical Docu- ment. J. P. Morgan/Reuters, 1996. López-Cabarcos, M., Pérez-Pico, A., Piñeiro Chousa, J. and Ševi´c, A., Bitcoin volatility, stock market and investor sentiment. Are they connected? Finance Res. Lett., 2020, 38, 101399. Maciel, L., Cryptocurrencies value-at-risk and expected shortfall: Do regime-switching volatility models improve forecasting? Int. J. Finance Econ., 2021, 26(3), 4840–4855. Matheson, J.E. and Winkler, R.L., Scoring rules for continuous probability distributions. Manage. Sci., 1976, 22(10), 1087– 1096. Matkovskyy, R., Jalan, A. and Dowling, M., Effects of economic policy uncertainty shocks on the interdependence between bitcoin and traditional financial markets. Q. Rev. Econ. Finance, 2020, 77, 150–155. McMillan, D.G. and Kambouroudis, D., Are riskmetrics forecasts good enough? Evidence from 31 stock markets. Int. Rev. Financ. Anal., 2009, 18(3), 117–124. McNeil, A. and Frey, R., Estimation of tail-related risk measures for heteroscedastic financial time series: An extreme value approach. J. Empir. Finance, 2000, 7(3-4), 271–300. Mincer, J.A. and Zarnowitz, V., The evaluation of economic forecasts. In Economic Forecasts and Expectations: Analysis of Forecasting Behavior and Performance, pp. 3–46, 1969 https://econpapers.repec.org/bookchap/nbrnberbk/ (NBER). minc69-1.htm Nekhili, R. and Sultan, J., Jump-driven risk model performance in cryptocurrency market. Int. J. Financ. Stud., 2020, 8(2), 1–18. Nelson, D.B., Conditional heteroskedasticity in asset returns: A new approach. Econometrica, 1991, 59(2), 347–370. Pafka, S. and Kondor, I., Evaluating the riskmetrics methodology in measuring volatility and value-at-risk in financial markets. Physica A: Statist. Mechan. Appl., 2001, 299(1-2), 305–310. Patton, A.J., Volatility forecast comparison using imperfect volatility proxies. J. Econom., 2011, 160(1), 246–256. Rubinstein, R.Y., Optimization of computer simulation models with rare events. Eur. J. Oper. Res., 1997, 99(1), 89–112. Schwert, G.W., Stock volatility and the crash of ’87. Rev. Financ. Stud., 1990, 3(1), 77–102. 30 C. Alexander and M. Dakos Segnon, M. and Bekiros, S., Forecasting volatility in bitcoin market. Ann. Finance, 2020, 16, 435–462. Silahli, B., Dingec, K., Cifter, A. and Aydin, N., Portfolio value- at-risk with two-sided Weibull distribution: Evidence from cryp- tocurrency markets. Finance Res. Lett., 2021, 38, 101425. Sosa, M., Ortiz, E. and Cabello, A., Bitcoin conditional volatility: GARCH extensions and Markov switching approach. Int. Finance Rev., 2019, 20, 201–219. Stavroyiannis, S., Value-at-risk and related measures for the bitcoin. J. Risk Finance, 2018, 19(2), 127–136. Stroud, J.R. and Johannes, M.S., Bayesian modeling and forecast- ing of 24-hour high-frequency volatility. J. Am. Stat. Assoc., 2014, 109(508), 1368–1384. Takahashi, M., Watanabe, T. and Omori, Y., Volatility and quantile forecasts by realized stochastic volatility models with generalized hyperbolic distribution. Int. J. Forecast., 2016, 32(2), 437–457. Takahashi, M., Watanabe, T. and Omori, Y., Forecasting daily volatility of stock price index using daily returns and real- ized volatility. Econom. Statist., 2021. Article in Press. https://www.sciencedirect.com/science/article/pii/S2452306 221000964 Tiwari, A., Kumar, S. and Pathak, R., Modelling the dynamics of bitcoin and litecoin: GARCH versus stochastic volatility models. Appl. Econ., 2019, 51(37), 4073–4082. Troster, V., Tiwari, A., Shahbaz, M. and Macedo, D., Bitcoin returns and risk: A general GARCH and GAS analysis. Finance Res. Lett., 2019, 30, 187–193. Trucíos, C., Forecasting bitcoin risk measures: A robust approach. Int. J. Forecast., 2019, 35(3), 836–847. Trucíos, C., Hotta, L.K. and Ruiz, E., Robust bootstrap forecast den- sities for GARCH returns and volatilities. J. Stat. Comput. Simul., 2017, 87(16), 3152–3174. Trucíos, C. and Taylor, J., A comparison of methods for fore- casting value-at-risk and expected shortfall of cryptocurrencies. ResearchGate Working Paper, 2022. Tse, Y.K. and Tsui, A.K.C., A multivariate generalized autore- gressive conditional heteroscedasticity model with time-varying correlations. 351– 362. J. Bus. Econ. 20(3), 2002, Stat., Vidal-Tomás, D. and Ibañez, A., Semi-strong efficiency of bitcoin. Finance Res. Lett., 2018, 27, 259–265. Wang, J.-N., Liu, H.-C., Chiang, S.-M. and Hsu, Y.-T., On the pre- dictive power of ARJI volatility forecasts for bitcoin. Appl. Econ., 2019, 51(44), 4849–4855. Wu, G., The determinants of asymmetric volatility. Rev. Financ. Stud., 2001, 14(3), 837–859. Zakoian, J.-M., Threshold heteroskedastic models. J. Econ. Dyn. Control, 1994, 18(5), 931–955. Assessing the accuracy of exponentially weighted moving average models for VaR and ES of Crypto portfolios 31 Table A1. Backtesting results for one-day-ahead left- (long position) and right-tail (short position) 5% VaR forecasts for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC), based on an out-of-sample period between 1 January 2017 and 31 August 2021. 32 C. Alexander and M. Dakos Table A2. Backtesting results for one-day-ahead left- (long position) and right-tail (short position) 5% ES forecasts for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC), based on an out-of-sample period between 1 January 2017 and 31 August 2021. Table A3. Backtesting results for one-hour-ahead one-day-ahead left- (long position) and right-tail (short position) 5% VaR forecasts for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC), based on an out-of-sample period between 1 May 2021 and 1 July 2021. Assessing the accuracy of exponentially weighted moving average models for VaR and ES of Crypto portfolios 33 Table A4. Backtesting results for one-hour-ahead one-day-ahead left- (long position) and right-tail (short position) 5% ES forecasts for bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC), based on an out-of-sample period between 1 May 2021 and 1 July 2021. Table A5. Backtesting of daily left-tail 1% VaR for a short position on bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC). 34 C. Alexander and M. Dakos Table A6. Backtesting of hourly left-tail 1% VaR on bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC). Assessing the accuracy of exponentially weighted moving average models for VaR and ES of Crypto portfolios 35 Table A7. Backtesting of hourly right-tail 1% VaR on bitcoin (BTC), ether (ETH), ripple (XRP) and litecoin (LTC).
10.1038_s41598-022-27193-9
OPEN A fully fiber‑integrated ion trap for portable quantum technologies Xavier Fernandez‑Gonzalvo * & Matthias Keller Trapped ions are a promising platform for the deployment of quantum technologies. However, traditional ion trap experiments tend to be bulky and environment‑sensitive due to the use of free‑ space optics. Here we present a single‑ion trap with integrated optical fibers directly embedded within the trap structure, to deliver laser light as well as to collect the ion’s fluorescence. This eliminates the need for optical windows. We characterise the system’s performance and measure the ion’s fluorescence with signal‑to‑background ratios on the order of 50, which allows us to perform internal state readout measurements with a fidelity over 99% in 600 µ s. We test the system’s resilience to thermal variations in the range between 22 and 53 and 300 Hz and find no effect on its performance. The combination of compactness and robustness of our fiber‑coupled trap makes it well suited for applications in, as well as outside, research laboratory environments, and in particular for highly compact portable quantum technologies, such as portable optical atomic clocks. While our system is designed for trapping 40Ca+ ions the fundamental design principles can be applied to other ion species. C, and the system’s vibration resilience at 34 Hz ◦ Trapped ions are a promising candidate for a wide range of quantum technologies. They are intrinsically repro- ducible systems, exhibiting long coherence and trapping lifetimes, and techniques to prepare, readout and manipulate their internal and external quantum states are well-developed. This makes them highly suitable to be used in quantum information processing1,2, precision spectroscopy3 and tests of fundamental physics4,5 amongst others. While there has been remarkable progress in the development and miniaturisation of novel ion trapping structures and associated vacuum systems6,7, the optical systems needed to manipulate and detect the state of the trapped ions are still mainly based on free-space optics. This leaves a compact ion trap surrounded by a large volume of optical components, which are often susceptible to drifts and vibrations, requiring regular realignment, since free-space optics can lead to beam-pointing instability and hence a deterioration of the sys- tem’s performance. While for laboratory-based research systems this can be acceptable, for operation outside research laboratories this poses a significant barrier. In particular, the susceptibility of the beam steering and detection optics to vibrations, temperature fluctuations and drifts hinders the use of trapped ions in fieldable metrology and sensor systems. In recent years there has been progress in integrating the fluorescence detection optics into the ion trap structure using optical fibers8–10. This eliminates the need for large numerical aperture lenses, which are prone to misalignment and drift and allows an easy connection to the photon detector. However, this comes with the disadvantage that the lack of spacial filtering results in a higher sensitivity to light scattered by the trap electrodes or the surrounding structures. Another approach is to use integrated superconducting single photon detectors11 and single-photon avalanche photodiodes12. While these offer great collection efficiencies they are best suited to planar ion traps as opposed to 3-dimensional trapping structures, the latter being preferred for atomic clock applications due to their lower heating rates and higher trapping efficiencies. Furthermore, the requirement to operate at cryogenic temperatures for superconducting devices prohibits their use in highly compact and portable systems. A third approach is to use in-vacuum integrated optics to maximise the collection of ionic fluorescence13–15, working in conjunction with out-of-vacuum optical elements. These solutions are well suited to planar ion traps, and are particularly interesting for multi-ion systems, but they still require a windowed vacuum chamber and careful alignment of external optical components. Progress has also been made in the integration of the delivery optics, using optical waveguides embedded into the substrate of surface ion traps16–19. Here, diffractive couplers are used to focus the beams onto the position of the ion. This leads to mechanically robust and realignment-free systems, and produces sufficiently small beam waists. However, aligning the input fibers with the embedded waveguides can be difficult, leading to low overall optical transmission efficiencies. Single-wavelength beam delivery using a single mode optical fiber integrated in Department of Physics and Astronomy, University of Sussex, Brighton BN1 9QH, UK. *email: x.fernandez- gonzalvo@sussex.ac.uk Scientific Reports | (2023) 13:523 | https://doi.org/10.1038/s41598-022-27193-9 1 Vol.:(0123456789)www.nature.com/scientificreports a surface trap has also been reported20, but so far total integration of all the delivery beams as well as fluorescence collection hasn’t been shown. In this article we present a fiber-integrated ion trap structure, eliminating the need for external free-space optics or optical access. An end-cap style ion trap based on10 has an optical multimode fiber integrated into one of the rf electrodes for fluorescence collection, and uses in-vacuum optical fibers and focusing optics to deliver the required laser light to the ion. This laser delivery structure facilitates the flexible alignment of the individual beam polarisations and angles during trap assembly. The geometrical arrangement of the multimode collection fiber, its close proximity to the ion and the good mode shape provided by the delivery optics allow us to measure the ion’s fluorescence with high signal-to-background ratios even without any spatial filtering. We have characterised the system under different temperature and vibration conditions, which allows us to show that its performance is unaffected by changes of the environmental condition. The compact size, robustness and flexibility of this trap design make it well suited for applications in single ion experiments outside the research lab, with a particular emphases in portable optical atomic clocks. The 40Ca+ ion and required wavelengths Our system is designed for trapping calcium ions. 40Ca+ is particularly well suited for applications in portable optical atomic clocks and sensors because all the wavelengths required for ionisation, cooling, repumping, quenching and spectroscopic interrogation of the clock transition are accessible through compact diode lasers. Moreover, all these wavelengths are compatible with fiber optic components, which is essential for the minia- turisation and ruggedisation of the setup. The relevant energy levels of 40Ca and 40Ca+ are shown in Fig. 1. In order to ionise neutral 40Ca we use a reso- nant transition at 423 nm and non-resonant light at 375 nm . We use the cooling transition in 40Ca+ at 397 nm , and repumping can be done with 866 nm light or a combination of 850 nm and 854 nm light. 40Ca+ has a clock transition at 729 nm . The 854 nm transition can also be used to quench the ion out of the D5/2 state following the clock interrogation readout step. The fiber‑integrated ion trap Trap geometry. The trap, schematically shown in Fig.  2, is an end-cap style trap, which provides three- dimensional rf confinement. It consists of two sets of cylindrical concentric electrodes facing each other, with the trap center being in the gap between the electrode assemblies. The inner electrodes are connected to the rf potential, while the outer ones are connected to ground. The inner rf electrodes are hollow, and house multi- mode fibers, which are used for fluorescence collection. The inner electrodes’ outer diameter is 500 µm and they protrude by 250 µm from the ground electrodes. The outer electrodes’ inner and outer diameters are 800 µm and 1.78 mm respectively, and they are tapered at 45 to increase the optical access angle and prevent clipping the laser beams. An alumina tube is used between the inner and the outer electrodes to electrically isolate them while maintaining concentricity. The electrodes and the alumina spacer are glued together using UVH compat- ible epoxy (EPO-TEK 353ND). The axial separation between rf electrodes is 500 µm . The inner electrodes are connected to the main rf source at the back of the electrodes. The outer electrodes are grounded by connecting them to the main body of the trap through a pair of capacitors. This allows them to be used as dc electrodes for micromotion compensation in the axial direction, while keeping them ac grounded. Two dc electrodes are utilised to supply micromotion compensation voltages in the radial plane. A resistively heated tantalum tube filled with calcium is mounted inside the copper body holding the trap, and serves as a calcium dispenser. Two pinholes collimate the calcium atomic beam to pass between the inner electrodes. ◦ Figure 1. Relevant energy levels for the ionisation of 40Ca and operation of a 40Ca+ atomic clock. In this work we cool the ion using the 397 nm transition, together with 850 nm and 854 nm repumpers. The clock transition in 40Ca+ is at 729 nm . Wavelengths have been grouped by colors (blue, orange or red) to represent the beams that can travel through the same type of optical fiber. Solid arrows denote the wavelengths used in this work. Scientific Reports | (2023) 13:523 | https://doi.org/10.1038/s41598-022-27193-9 2 Vol:.(1234567890)www.nature.com/scientificreports/ Figure 2. Schematic representation of the fiber-integrated ion trap. Temperature sensors, wiring and dc electrodes have been omitted. Left: overview of the trap showing the polarisation maintaining (PM) fibers used for light delivery and the multi-mode (MM) fiber used for fluorescence collection, as well as the fiber, dc and rf feedthroughs. Bottom right: zoom-in on the trapping structure showing the gradient-index (GRIN) lens collimators and the path followed by the delivery beams, as well as the rf decoupling capacitors. Top right: zoom-in and cross-section of the electrode structure, showing the MM fiber embedded inside the rf electrode. The position of the ion is represented with a light blue circle (not to scale). Fluorescence collection via multimode fibers. The integration of the fluorescence collection fiber into the electrode assembly removes the need for alignment, since the fiber is concentric with the rf electrodes and is therefore aligned with the expected position of the ion. The system is therefore insensitive to small misalign- ments of the fiber position, making it inherently robust to mechanical vibrations and thermal drifts. The multimode fiber used for fluorescence collection (Thorlabs FG200UEA) has a core diameter of 200 µm and a cladding diameter of 220 µm . The core is made of pure silica, and the cladding is made of fluorine-doped silica. The acrylate protective coating of the fiber was stripped and its end was tapered down to a diameter of 190 µm over 11 mm to provide a tight fit to the rf electrodes’ inner bore. The multimode fiber is retracted 90–100 µm with respect to the rf electrodes’ front surface. The fiber is glued at the back of the rf electrode using UHV compatible epoxy (EPO-TEK 301-2). Light collected in the multimode fiber is spectrally filtered using a narrow band-pass filter and a a photo- multiplier detector (PMT) is then used to measure the ion’s fluorescence. Based on the geometry of the system, the fraction of light captured by the fiber is approximately 1.2%, limited by its numerical aperture, meaning a total possible of about 2.4% if two fibers are used. In this work only one fiber was used, owing to an accidental breakage of the second one during the later stages of the assembly process. Optical losses between the ion and the PMT will comprise: reflection losses at the input and output faces of the MM fiber (3.6% on each surface, assuming a refractive index of 1.4721), propagation losses along the fiber (1% at 400 nm for a 1 m fiber) and transmission losses through the band-pass filter (7% at 397 nm ), leading to a total loss of 15%. With the PMT nominal photon detection efficiency at 400 nm of 30%, the overall fluorescence detection efficiency is about 0.3% (0.6% for both fibers). Beam delivery via polarisation maintaining fibers. To deliver the necessary laser beams for the ioni- sation of 40Ca and for the cooling and repumping of 40Ca+ ions we use different off-the-shelf optical fibers for different wavelength groups (refer to Fig. 1). They are all single mode polarisation maintaining fibers. We use an ultraviolet (UV) fiber (Thorlabs PM-S405-XP) to deliver the photoionisation lasers as well as the cooling beam, and a single infrared (IR) fiber (Thorlabs PM780-HP) to deliver the repumper beams at 850 nm and 854 nm . This IR fiber can also be used to deliver light at 866 nm . Additionally, the system is equipped with a second UV fiber for another cooling beam (not used in this work), and a dedicated fiber (Thorlabs PM630-HP) for the future clock laser. The fibers are fed into the vacuum system using optical fiber feedthroughs described in22, which were all independently tested to have a leak rate below our measurement limit of 1 ×10−9 mbar.l/s. Anti-reflection-coated gradient-index (GRIN) lenses with a design focal length of 10 mm are employed to focus the fiber outputs into the center of the trap. The delivery fibers sit in a ceramic ferrule just behind the GRIN lenses, with a fiber-to-lens separation of less than 100 µm . These laser delivery systems create close-to- diffraction-limited beams, with a measured beam waist w0 ( 1/e2 radius) of 5.71(6) µm and 5.43(2) µm for the 397 nm beams, 9.82(7) µm for the 729 nm beams and 11.1(1) µm for the 866 nm beams. As discussed below, we don’t fully exploit the small beam waists, but the good mode shape and absence of beam halos minimises background counts due to the beams scattering on the electrodes. As it will be shown later, this enables us to measure the ion’s fluorescence through the multimode fiber with high signal-to-background ratios without any Scientific Reports | (2023) 13:523 | https://doi.org/10.1038/s41598-022-27193-9 3 Vol.:(0123456789)www.nature.com/scientificreports/ spatial filtering. Note that the 729 nm beam is not used in this work, since probing the clock transition is out of the scope of this initial investigation. The beams are aligned to the geometrical center of the rf trap during assembly with a combination of a scat- tering screen placed between the inner electrodes and a pair of microscopes used to observe the laser beam positions. The alignment of the delivery assemblies was done using micro-positioning stages in three dimensions, and we estimate we were able to position the beam within 5 µm from the geometrical center of the trap. In order to increase the robustness against misalignment, the beam foci were positioned such that the beam ( 1/e2 ) radius was around 25 µm at the expected position of the ion. Once the alignment was optimised the lenses were glued ◦ to the main body of the trap using UHV compatible epoxy (EPO-TEK H21D). The epoxy was cured at 80 C for at least 4 hours, during which we manually fed back on the translation stages to keep the beams aligned. After the curing process the beams typically stayed aligned to the center of the trap to within 10 µm . We attribute the small alignment changes to stress accumulated in the epoxy during the curing process. Trap performance To characterise the trap we use a vacuum chamber with an optical window. This allows us to use an sCMOS camera (Andor Zyla) to observe the ion during characterisation, but this is not required to operate the trap. The system was pumped down to (cid:31)10 −10 mbar using a getter-ion combination pump (Saes NEXTorr D 100- 5). After bakeout and pumping, ions were trapped within the first two days of trying, since no optical alignment was necessary. Both atomic and ionic fluorescence could be observed through the multimode fiber by using an appropriate band-pass filter in front of the PMT. The trap is driven at a frequency of 13.7 MHz via a resonant transformer. The secular frequencies are kept between 0.6 MHz and 4.5 MHz in the axial direction and between 0.4 MHz and 2.0 MHz in the radial directions. Assuming the trap’s a-values to be negligible ( ax,y,z ≈ 0 ), the q-values are within the ranges qx,y = 0.08–0.41 and qz = 0.12–0.92. Excess micromotion due to external stray fields is compensated using a combination of the trap depth modu- lation method and the photon correlation method23. From loading to loading, the micromotion compensation voltage values only change by small amounts ( (cid:31)5%) and are otherwise stable. In contrast to the expected ion lifetime of hours, the ion lifetime within this trap is around 10 minutes. We attribute this to a virtual leak inside the electrode structure. Using a UHV compatible epoxy with a higher vis- cosity (e.g. EPO-TEK H21D) could have reduce the probability of gas pockets forming between the MM fibers and the rf electrodes due to capillary action. Cooling transition spectroscopy, signal to background ratio and beam‑ion positioning. With micromotion compensated, we measured the cooling transition spectral profile. These measurements are per- formed by scanning the frequency of the 397 nm laser using an acousto-optic modulator while recording the flu- orescence PMT counts at the output of the multimode fiber. Figure 3 shows a spectrum for a cooling laser power of 0.14 µW . Fitting a Lorentzian function to the data we can extract a half width at half maximum (HWHM) of 11.1(2) MHz (with the natural transition HWHM being 10.8 MHz24). Repeating this measurement for different powers shows that the main contribution to the line broadening is power broadening, with the HWHM at zero power converging to the natural half-width. The signal to background ratio SBR = (S − B)/B (where S is the count rate at the peak of the transition and B is the background count rate measured without an ion) will depend on the cooling laser power due to power broadening. The best values were obtained for powers under 0.2 µW , where power broadening is negligible, with a SBR on the order of 50. For the typical cooling powers used to operate the trap (between 3 and 4 µW ), the SBR is on the order of 10 to 20. Furthermore, we used a series of HWHM measurements at different laser powers to estimate the position of the cooling laser beam with respect to the ion. The laser intensity at the position of the ion can be inferred from Figure 3. Unsaturated cooling transition spectral profile measured at 0.14 µW . The orange solid line is a Lorentzian fit to the red-detuned data, showing a fitted linewidth close to the natural linewidth of the 40Ca+ cooling transition. The dashed orange line shows the count rate measured without an ion in the trap, i.e. the background scattering count rate. Scientific Reports | (2023) 13:523 | https://doi.org/10.1038/s41598-022-27193-9 4 Vol:.(1234567890)www.nature.com/scientificreports/ power broadening. Comparing this with the actual laser power and the beam waist at the position of the ion, we can calculate where the ion is located within the Gaussian profile of the beam. The ion-to-beam-center distance was found to be 10.8(1.1) µm , with the uncertainty being dominated by the measurement of the laser power at the position of the ion. With a beam waist of 25 µm the ion is well within the cooling laser beam. State detection fidelity. Next we characterise the state detection fidelity in the trap by preparing the ion in either a bright or a dark state, and comparing the photon counting statistics measured with the PMT. A bright state is obtained by keeping the ion in its cooling cycle, i.e. by keeping the cooling laser on, as well as the repumpers. A dark state is obtained by switching off the repumpers, shelving the ion into the D-states. In terms of determining the state readout fidelity this is equivalent to preparing the ion in either the S1/2 (bright) or the D5/2 (dark) state (replicating the shelving that will occur during clock interrogation of the 729 nm transition). The measurement sequence can be seen in Fig. 4b. Photons arriving to the PMT are counted for a time window of length τw for both a dark and a bright ion. The measurements are repeated multiple times, and two histograms are obtained. An example of these can be seen in Fig. 4a. In order to determine the state of an ion a threshold value nth is defined (along the horizontal axis in Fig. 4a), above which the ion will be considered to be bright, and below which the ion will be considered to be dark. For the bright state, the detection fidelity is given by: FB = (cid:31) ∞ hB(n) n=nth [hB(n) + hD(n)] (cid:31) ∞ n=nth , (1) with hB,D(n) being the bright and dark histograms as a function of the photon number n. Similarly, the detection fidelity for the dark state is given by: FD = (cid:31)nth−1 n=0 hD(n) n=0 [hB(n) + hD(n)] (cid:31)nth . (2) The state detection fidelity is then calculated as the average between the two, F = 1 2 (FB + FD). The optimal nth value depends on the detection window time, the cooling and repumper laser powers and their detunings with respect to the line centers. We measured the state detection fidelity for a range of detection window times and cooling laser powers, and we can achieve state detection fidelities better than 99% for detec- tion periods as short as 600µ s (example in Fig. 4). The state detection fidelities are calculated directly from the measured data, without correcting for finite state preparation fidelity, finite state lifetime or any other detrimental effects25, and that we haven’t made any assumptions about the statistical distribution of the measured histograms. Due to low clipping on the electrodes, the low PMT sensitivity to near-infrared light and the band-pass filter, there is no measurable scatter from the repump lasers. Figure 4. (a) State detection measurement for a measurement window τw = 600 µs . The orange (blue) histogram corresponds to an ion prepared in the dark (bright) state. The lines are Poisson fits to the data, for reference only. (b) Pulse sequence utilised for the state detection measurements. The cooling laser is always kept on, while the repumpers are switched on and off periodically to toggle the ion between the dark and bright states. The shaded areas represent the measurement window time during which counts are added to the bright and dark histograms. There is a 100 µ s delay between switching the repumpers off (on) and the measurement window, to ensure the ion has been shelved (de-shelved). Scientific Reports | (2023) 13:523 | https://doi.org/10.1038/s41598-022-27193-9 5 Vol.:(0123456789)www.nature.com/scientificreports/ Thermal stability. The stability of the fiber-integrated ion trap against temperature changes is an important factor for its use outside of research laboratory environments. In order to test the effects of changing tempera- tures in our trap we measure the optical pumping time τ� to the D-states (which is directly related to the laser intensity at the position of the ion) while raising the temperature of the trap. To do so we heat up the entire vacuum chamber using a resistive heating belt, and let the system thermalise for a few minutes. The temperature is measured using three PT100 temperature sensors mounted at different places directly on the trap structure (one on each block holding the electrodes and one on the main copper mount). In order to measure τ� we start by preparing the ion in the S1/2 state, and then switch on the cooling beam with the repumpers off. Fluorescence will be observed until the ion is shelved onto the D3/2 or the D5/2 state. Over many repetitions an exponential decay of the fluorescence will be observed (see Fig. 5). The time constant of this decay is τ� , which is directly related to the Rabi frequency of the cooling beam26. If the beam is misaligned the ion will be exposed to a different light intensity, which in turn will result in a different time constant τ� . The inset in Fig. 6 shows the dependence of τ� with the cooling beam power. In order to have high alignment sensitivity to the temperature dependence, measurements were taken using a cooling powers around 0.83(5) µW , avoiding saturation of the cooling transition while still having an acceptable count rate on the PMT. Figure 6 shows the ◦ measured τ� for a range of temperatures between 22 C. The variation with respect to the average is consistent with changes in laser power between (and during) the different measurements, which is the main contribution to the uncertainty of these measurements. With the beam center 10.8 µm away from the position C and 53 ◦ Figure 5. Fluorescence decay rate measurement at room temperature for a cooling laser power of 1.6 µW . The orange line is an exponential fit to the data, from which a decay time constant τ� can be extracted. Inset: pulse sequence used for the measurement of τ�. Figure 6. Fluorescence decay constant as a function of the trap temperature. The horizontal orange line is the averaged τ� between all the measurements. The horizontal error bars represent the statistical error on reading the temperature using the three different thermal sensors. The vertical error bars combine the statistical error in the fit for τ� and the error in determining the laser power P times the slope of the τ� vs. P curve. Inset: fluorescence decay constant as a function of laser power measured at 22 power range at which the data in the main figure were taken. C. The shaded blue area denotes the ◦ Scientific Reports | (2023) 13:523 | https://doi.org/10.1038/s41598-022-27193-9 6 Vol:.(1234567890)www.nature.com/scientificreports/ of the ion, a beam waist of 25 µm and a slope of at least 1.3 µs/µ W in the blue highlighted section of the inset plot in Fig. 6, and assuming the optical power level to be perfectly stable, the shift in the beam’s position is less than ±1 µm . This is an upper bound, and the actual shift is expected to be much lower, since the variation in τ� is fully consistent with the observed variation in laser power (on the order of 5%). This suggests that thermal expansion and contraction has a negligible effect on beam alignment within the range of temperatures explored. Another potential issue with changing temperatures is a change in the excess micromotion of the ion, caused by a changing geometry of the trap as it thermally expands or contracts. The micromotion compensation voltages were found to remain constant within 3% of the average value for all tested temperatures, compatible with the variation observed between different trap loading runs. Vibration resilience. Finally, we test the resilience to mechanical vibrations of the fiber-coupled ion trap. To do so, we attach two different sources of vibrations to the vacuum chamber containing the trap, and evalu- ate its performance. The first vibrating device generates vibrations at frequencies around 34 Hz and the second device at around 300 Hz. The sCMOS camera looking at the ion is mounted on a floated optical table, in a sta- tionary frame. The vacuum chamber rests on the same optical table but, in order to keep it mechanically isolated, it is loosely clamped to the bench. The result is a system where the vacuum chamber and its contents vibrate but the camera doesn’t. From the camera images (see Fig. 7), assuming the motion of the ion trap to be sinusoidal, the apparent peak accelerations can be calculated for each vibration device. These represent a lower bound to the actual peak acceleration felt by the trap, since the sCMOS camera is only able to capture the ion’s motion on a two dimen- sional plane. When using the first device at 34 Hz the apparent peak acceleration is 0.047(5) g. For the second device running at 300 Hz the apparent peak acceleration is 1.09(18) g. In either case, no significant difference is observed either in the ion’s fluorescence rate, the micromotion compensation voltages, the cooling transition spectroscopic profile or the fluorescence decay constant τ�. In order to determine the displacement sensitivity to vibrations, we use the atomic ion’s fluorescence. Because we cannot detect any change in the fluorescence level between the situations with and without vibrations, we assume that the fluorescence change due to vibrations is below 10% of the observed variations due to laser power fluctuations. By analysing how a sinusoidal oscillation of the ion’s position with respect to the laser beam influ- ences the ion’s average fluorescence level, we can derive an upper limit of the misalignment amplitude of 3.5 µm . However, we expect the actual amplitude to be considerably smaller. Conclusions In conclusion, we have presented a compact, fully fiber-integrated, single-ion trap, where optical fibers inside the vacuum chamber are used for beam delivery as well as ion fluorescence collection. The delivery beams are focused onto the expected position of the ion during assembly using GRIN lenses monolithically attached to the trap’s body. This makes the system robust against mechanical vibrations and thermal variations, and completely eliminates the need for beam realignment over time. The multimode collection fibers are housed directly inside the trap electrodes, allowing them to sit close to the ion, therefore granting a good solid angle capture and allowing us to measure the ion’s fluorescence with high signal to background ratios. We have performed a basic characterisation of the ion trap, including state detection fidelity measurements, and we have subjected the system to a range of temperatures and mechanical vibration conditions, showing no deterioration of its performance. We believe this is a step forward towards miniaturisation of ion traps for their use in compact and robust integrated systems for applications outside the research laboratory, and specifically for their use in portable opti- cal atomic clocks. Finally, while we use 40Ca+ as our ion of choice, the design principles presented here can be extended to other species by choosing fibers and lenses appropriate to the required laser wavelengths. Figure 7. Camera image comparison between having the vacuum chamber (a) at rest, (b) vibrating at 34 Hz with an apparent peak acceleration of 0.047(5) g and (c) vibrating at 300 Hz with an apparent peak acceleration of 1.09(18) g. Scientific Reports | (2023) 13:523 | https://doi.org/10.1038/s41598-022-27193-9 7 Vol.:(0123456789)www.nature.com/scientificreports/ Data availability The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Received: 15 March 2022; Accepted: 28 December 2022 References 1. Häffner, H., Roos, C. F. & Blatt, R. Quantum computing with trapped ions. Phys. Rep. 469, 155–203. https:// doi. org/ 10. 1016/j. physr ep. 2008. 09. 003 (2008). 2. Bruzewicz, C. D., Chiaverini, J., McConnell, R. & Sage, J. M. Trapped-ion quantum computing: Progress and challenges. Appl. Phys. Rev. 6, 021314. https:// doi. org/ 10. 1063/1. 50881 64 (2019). 3. Ludlow, A. D., Boyd, M. M., Ye, J., Peik, E. & Schmidt, P. O. Optical atomic clocks. Rev. Mod. Phys. 87, 637–701. https:// doi. org/ 10. 1103/ RevMo dPhys. 87. 637 (2015). 4. Cairncross, W. B. et al. Precision measurement of the electron’s electric dipole moment using trapped molecular ions. Phys. Rev. Lett. 119, 153001. https:// doi. org/ 10. 1103/ PhysR evLett. 119. 153001 (2017). 5. Shaniv, R. et al. New methods for testing Lorentz invariance with atomic systems. Phys. Rev. Lett. 120, 103202. https:// doi. org/ 10. 1103/ PhysR evLett. 120. 103202 (2018). 6. Choonee, K., Wilpers, G. & Sinclair, A. G. Silicon microfabricated linear segmented ion traps for quantum technologies. In: 19th international conference on solid-state sensors, actuators and microsystems (Transducers), 615–618. https:// doi. org/ 10. 1109/ TRANS DUCERS. 2017. 79941 24 (2017). 7. Schwindt, P. D. D. et al. A highly miniaturized vacuum package for a trapped ion atomic clock. Rev. Sci. Instrum. 87, 053112. https:// doi. org/ 10. 1063/1. 49487 39 (2016). 8. VanDevender, A. P., Colombe, Y., Amini, J., Leibfried, D. & Wineland, D. J. Efficient fiber optic detection of trapped ion fluores- cence. Phys. Rev. Lett. 105, 023001. https:// doi. org/ 10. 1103/ PhysR evLett. 105. 023001 (2010). 9. Brady, G. R. et al. Integration of fluorescence collection optics with a microfabricated surface electrode ion trap. Appl. Phys. B 103, 801–808. https:// doi. org/ 10. 1007/ s00340- 011- 4453-z (2011). 10. Takahashi, H. et al. An integrated fiber trap for single-ion photonics. New J. Phys. 15, 053011. https:// doi. org/ 10. 1088/ 1367- 2630/ 15/5/ 053011 (2013). 11. Todaro, S. L. et al. State readout of a trapped ion qubit using a trap-integrated superconducting photon detector. Phys. Rev. Lett. 126, 010501. https:// doi. org/ 10. 1103/ PhysR evLett. 126. 010501 (2021). 12. Setzer, W. J. et al. Fluorescence detection of a trapped ion with a monolithically integrated single-photon-counting avalanche diode. Appl. Phys. Lett. 119, 154002. https:// doi. org/ 10. 1063/5. 00559 99 (2021). 13. Streed, E. W., Norton, B. G., Jechow, A., Weinhold, T. J. & Kielpinski, D. Imaging of trapped ions with a microfabricated optic for quantum information processing. Phys. Rev. Lett. 106, 010502. https:// doi. org/ 10. 1103/ PhysR evLett. 106. 010502 (2011). 14. Merrill, J. T. et al. Demonstration of integrated microscale optics in surface-electrode ion traps. N. J. Phys. 13, 103005. https:// doi. org/ 10. 1088/ 1367- 2630/ 13/ 10/ 103005 (2011). Publisher: IOP Publishing. 15. Ghadimi, M. et al. Scalable ion-photon quantum interface based on integrated diffractive mirrors. NPJ Quant. Inf. 3, 1–4. https:// doi. org/ 10. 1038/ s41534- 017- 0006-6 (2017). 16. Niffenegger, R. J. et al. Integrated multi-wavelength control of an ion qubit. Nature 586, 538–542. https:// doi. org/ 10. 1038/ s41586- 020- 03104-8 (2020). 17. Mehta, K. K. et al. Integrated optical multi-ion quantum logic. Nature 586, 533–537. https:// doi. org/ 10. 1038/ s41586- 020- 2823-6 (2020). 18. Day, M. L. et al. A micro-optical module for multi-wavelength addressing of trapped ions. Quantum Sci. Technol. 6, 024007. https:// doi. org/ 10. 1088/ 2058- 9565/ abdf38 (2021). 19. Ivory, M. et al. Integrated optical addressing of a trapped ytterbium ion. Phys. Rev. X 11, 041033. https:// doi. org/ 10. 1103/ PhysR evX. 11. 041033 (2021). 20. Kim, T. H., Herskind, P. F. & Chuang, I. L. Surface-electrode ion trap with integrated light source. Appl. Phys. Lett. 98, 214103. https:// doi. org/ 10. 1063/1. 35934 96 (2011). 21. Malitson, I. H. Interspecimen comparison of the refractive index of fused silica. J. Opt. Soc. Am. 55, 1205–1209. https:// doi. org/ 10. 1364/ JOSA. 55. 001205 (1965). 22. Peerzada, A. R. et al. Versatile optical fiber feedthroughs for ultra-high vacuum applications. Vacuum 180, 109542. https:// doi. org/ 10. 1016/j. vacuum. 2020. 109542 (2020). 23. Berkeland, D. J., Miller, J. D., Bergquist, J. C., Itano, W. M. & Wineland, D. J. Minimization of ion micromotion in a Paul trap. Int. J. Appl. Phys. 83, 5025–5033. https:// doi. org/ 10. 1063/1. 367318 (1998). 24. Hettrich, M. et al. Measurement of dipole matrix elements with a single trapped ion. Phy. Rev. Lett. 115, 143003. https:// doi. org/ 10. 1103/ PhysR evLett. 115. 143003 (2015). 25. Burrell, A. H., Szwer, D. J., Webster, S. C. & Lucas, D. M. Scalable simultaneous multiqubit readout with 99.99% single-shot fidelity.. Phys. Rev. A 81, 040302. https:// doi. org/ 10. 1103/ PhysR evA. 81. 040302 (2010). 26. Takahashi, H., Kassa, E., Christoforou, C. & Keller, M. Cavity-induced anticorrelated photon-emission rates of a single ion. Phys. Rev. A 96, 023824. https:// doi. org/ 10. 1103/ PhysR evA. 96. 023824 (2017). Acknowledgements This research has received funding from EURAMET (EMPR SIB04-REG4) and the Engineering and Physical Sciences Research Council Quantum Technology Hub for Sensors and Metrology (EP/M013294/1). Author contributions X.F.G. and M.K. conceived the experiment and X.F.G. conducted the experiment and analysed the results. All authors reviewed the manuscript. Competing interests The authors declare no competing interests. Additional information Correspondence and requests for materials should be addressed to X.F.-G. Scientific Reports | (2023) 13:523 | https://doi.org/10.1038/s41598-022-27193-9 8 Vol:.(1234567890)www.nature.com/scientificreports/ Reprints and permissions information is available at www.nature.com/reprints. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/. © The Author(s) 2023 Scientific Reports | (2023) 13:523 | https://doi.org/10.1038/s41598-022-27193-9 9 Vol.:(0123456789)www.nature.com/scientificreports/
10.7554_elife.78558
ReSeaRCH aRTICLe Macrophages regulate gastrointestinal motility through complement component 1q Mihir Pendse1, Haley De Selle1, Nguyen Vo1, Gabriella Quinn1, Chaitanya Dende1, Yun Li1, Cristine N Salinas1, Tarun Srinivasan1, Daniel C Propheter1, Alexander A Crofts1, Eugene Koo1, Brian Hassell1, Kelly A Ruhn1, Prithvi Raj1, Yuuki Obata1*, Lora V Hooper1,2* 1Department of Immunology, The University of Texas Southwestern Medical Center, Dallas, United States; 2The Howard Hughes Medical Institute, The University of Texas Southwestern Medical Center, Dallas, United States Abstract Peristaltic movement of the intestine propels food down the length of the gastroin- testinal tract to promote nutrient absorption. Interactions between intestinal macrophages and the enteric nervous system regulate gastrointestinal motility, yet we have an incomplete understanding of the molecular mediators of this crosstalk. Here, we identify complement component 1q (C1q) as a macrophage product that regulates gut motility. Macrophages were the predominant source of C1q in the mouse intestine and most extraintestinal tissues. Although C1q mediates the complement- mediated killing of bacteria in the bloodstream, we found that C1q was not essential for the immune defense of the intestine. Instead, C1q- expressing macrophages were located in the intestinal submucosal and myenteric plexuses where they were closely associated with enteric neurons and expressed surface markers characteristic of nerve- adjacent macrophages in other tissues. Mice with a macrophage- specific deletion of C1qa showed changes in enteric neuronal gene expression, increased neurogenic activity of peristalsis, and accelerated intestinal transit. Our findings identify C1q as a key regulator of gastrointestinal motility and provide enhanced insight into the crosstalk between macrophages and the enteric nervous system. Editor's evaluation This study provides a fundamental finding that complement C1q produced by enteric macrophages shapes neuronal function and gut motility. The authors present convincing data showing that while macrophage- derived C1q is not necessary for defenses against enteric pathogens, it plays an important role in regulating neuronal gene expression and intestinal transit. These findings will be of interest to gastroenterologists, neuroscientists and immunologists in revealing a novel neuroimmune axis in gut homeostasis. Introduction Peristalsis is the physical force that propels food through the intestine, promoting digestion and nutrient absorption. The gastrointestinal motility that underlies peristalsis is a complex process that requires coordination of the activity of smooth muscle cells by enteric neurons (Rao and Gershon, 2016). Several studies have revealed that intestinal macrophages impact gastrointestinal motility by regulating the functions of enteric neurons and facilitating their interactions with smooth muscle cells (Muller et al., 2014; Matheis et al., 2020). *For correspondence: yuki.obata@utsouthwestern.edu (YO); lora.hooper@utsouthwestern. edu (LVH) Competing interest: The authors declare that no competing interests exist. Funding: See page 26 Preprinted: 28 January 2022 Received: 11 March 2022 Accepted: 17 April 2023 Published: 26 April 2023 Reviewing Editor: Isaac M Chiu, Harvard Medical School, United States Copyright Pendse et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 1 of 30 Research article Macrophages carry out diverse functions in the intestine that vary according to their anatomical location. For example, macrophages that localize to the tissue located directly underneath the gut epithelium — known as the lamina propria — contribute to immune defense against pathogenic bacteria (Gabanyi et  al., 2016). A distinct group of macrophages localizes to the tissues located beneath the lamina propria, between the circular and longitudinal muscle layers in the tissue region known as the muscularis externa. These muscularis macrophages express genes that are distinct from lamina propria macrophages (Gabanyi et  al., 2016). They directly regulate the activity of smooth muscle cells (Luo et al., 2018) and secrete soluble factors, such as bone morphogenetic protein 2 (BMP2), which interact with the enteric neurons that control smooth muscle activity (Muller et  al., 2014). Muscularis macrophages thus play a key role in regulating gut motility. However, we have a limited understanding of the molecular mechanisms by which these macrophages regulate intestinal neuromuscular activity and gut motility. C1q is a member of the defense collagen family that has distinct roles in immune defense and nervous system development and function (Bossi et al., 2014; Casals et al., 2019; Shah et al., 2015; Thielens et al., 2017). It is composed of six molecules each of C1qA, C1qB, and C1qC, forming a 410 kDa oligomer. C1q circulates in the bloodstream, where it participates in immune defense against infection by recognizing antibodies bound to invading bacteria. This binding interaction initiates the classical complement pathway, which entails the recruitment and proteolytic processing of other complement components that rupture the bacterial membrane and recruit phagocytic cells (Kishore and Reid, 2000; Noris and Remuzzi, 2013). C1q is also produced by microglia (brain- resident macrophage- like cells) in the brain where it promotes the pruning of neuronal synapses through an unclear mechanism (Hammond et al., 2020; Hong et al., 2016). Consequently, C1q deficiency results in heightened synaptic connectivity in the central nervous system which can lead to epilepsy (Chu et al., 2010). C1q is also produced at barrier sites, such as the intestine, where encounters with commensal and pathogenic microbes are frequent. However, little is known about the physiological role of C1q in barrier tissues. Liver immune cells, including macrophages and dendritic cells, produce serum C1q; however, the cellular source of C1q in barrier tissues including the intestine remains unclear (Petry et al., 2001). Here, we show that C1q is produced by macrophages of the mouse intestine. Intestinal C1q- expressing macrophages exhibit properties of neuromodulatory macrophages from other tissues and are located close to enteric neurons that have a known role in controlling gut motility. Accord- ingly, mice lacking macrophage C1q exhibit altered expression of enteric neuronal genes, increased neurogenic peristaltic contractions, and accelerated gastrointestinal motility. These findings identify C1q as a key mediator of a neuroimmune interaction that regulates gut motility. Results C1q is expressed by macrophages in the mouse small intestine Soluble defense collagens are an ancient, evolutionarily conserved family of antimicrobial proteins with shared structural features including a C- terminal globular head and a collagen- like region (Casals et al., 2019). Little is known about the function of defense collagens at mucosal barrier sites, where microbial encounter is frequent. Our initial goal in this study was to identify soluble defense collagens that are expressed by the mouse intestine and to assess their role in host defense. Therefore, we measured the expression of 18 defense collagen genes in the mouse small intestine and colon by RNA sequencing (RNA- seq). The most abundant soluble defense collagen transcripts in the small intestine and colon were those encoding C1qA, C1qB, and C1qC (Figure 1A; Figure 1—figure supplement 1). Serum C1q is produced by liver dendritic cells, monocytes, and macrophages (El- Shamy et  al., 2018). However, the cellular source(s) of C1q in peripheral tissues, including the intestine, is unknown. Quantitative PCR (qPCR) analysis of fluorescence- activated cell sorting (FACS)- sorted cell suspensions recovered from the small intestines of wild- type C57BL/6 mice revealed that C1qa, C1qb, and C1qc transcripts were most abundant in CD45+ cells, which include all immune cells, as compared to CD45- cells, which encompass epithelial cells and other non- immune cells (Figure  1B). Furthermore, C1q transcripts and protein were most abundant in CD45+ cells recovered from the subepithelial compart- ment, which includes both the lamina propria and muscularis, as compared to CD45+ cells recovered from the intraepithelial compartment of the small intestine (Figure 1C and D). Thus, C1q is expressed Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 2 of 30 Immunology and Inflammation Research article Mouse 1 2 3 A Adipoq C1qa C1qb C1qc **** **** **** C1qtnf1 C1qtnf2 C1qtnf3 C1qtnf4 C1qtnf6 C1qtnf7 C1qtnf9 Colec10 Colec11 Fcna Mbl2 Sftpa1 Sftpb Sftpc 0 10 20 30 40 Normalized read counts B i n o s s e r p x e e v i t a e R l C1qa *** C1qb ** C1qc p=0.06 2500 2000 1500 1000 500 10 5 0 6000 5000 4000 3000 2000 1000 10 5 0 1000 800 600 400 200 10 5 0 CD45- CD45+ Sm. int. cells C1qa ** C1qb ** C1qc ** C 100 80 60 40 20 0 100 i n o s s e r p x e e v i t a e R l 80 60 40 20 0 100 80 60 40 20 0 Intraepithelial Subepithelial Sm. int. cell source F l a l a i l e h t i p e - b u S i l e h t i p e - a r t n I D Sm. int. cell source: C1q Actin E 500K 400K C S S 300K 200K CD11b+ CD11b- 100K -104 500K 400K 104 105 106 0 CD11b C S S 300K 200K MHCII- MHCII+ 100K -104 0 500K 400K 104 105106107 MHCII C S S 300K 200K F4/80lo F4/80hi 100K -104 104 105 106 0 F4/80 40 30 20 10 0 40 30 20 10 0 25 i n o s s e r p x e e v i t a e R l C1qa **** ns C1qb **** ns C1qc **** 20 15 10 5 ns 0 CD11b+MHCII+F4/80lo CD11b+MHCII- CD11b- CD11b+MHCII+F4/80hi G Isotype control Anti-C1q CD11b- CD11b+ MHCII- CD11b+ MHCII+ F4/80lo CD11b+ MHCII+ F4/80hi x a m f o % 0 103 104 105 106 C1q C1q **** ns H y t i s n e t n i e c n e c s e r o u l f n a d e M i ) l o r t n o c e p y t o s i o t e v i t a e r ( l 80 60 40 5 4 3 2 1 0 CD11b+MHCII+F4/80lo CD11b+MHCII+F4/80hi CD11b+MHCII- CD11b- Figure 1. Complement component 1q (C1q) is expressed by macrophages in the mouse small intestine. (A) RNA- seq analysis of soluble defense collagen expression in the small intestines (ileum) of C57BL/6 mice. Data were adapted from a previously published RNA- seq analysis (Gattu et al., 2019). Data are available in the Gene Expression Omnibus repository under accession number GSE122471. Each column represents one mouse. (B) Quantitative PCR (qPCR) measurement of C1qa, C1qb, and C1qc transcript abundance in CD45+ and CD45- cells purified from mouse small intestines by flow cytometry. Each data point represents one mouse, and the results are representative of two independent experiments. (C) qPCR measurement of C1qa, C1qb, and C1qc transcript abundance in subepithelial and intraepithelial cells recovered from mouse small intestines. Each data point represents one mouse, and the results are representative of three independent experiments. (D) Representative immunoblot of subepithelial and intraepithelial cells recovered from mouse small intestines, with detection of C1q and actin (control). Each lane represents cells from one mouse and the immunoblot is representative of three independent experiments. (E) Flow cytometry gating strategy for analysis of mouse small intestinal cell suspensions in panels F, G, and H. Cells were pre- gated as live CD45+ cells. SSC, side- scatter; MHCII, major histocompatibility complex II. (F) qPCR measurement of C1qa, C1qb, and C1qc transcript abundance in cells isolated by flow cytometry from mouse small intestines as indicated in (E). Each data point represents cells pooled from three mice, and the results are representative of three independent experiments. (G) Flow cytometry analysis of intracellular C1q in small intestinal subepithelial cells identified as indicated in (E). (H) Quantitation of flow cytometry analysis in (G). Each data point represents one mouse, and the results are representative of two independent experiments. Sm. int., mouse small intestine; Error bars represent SEM. **p<0.01; ***p<0.001; ****p<0.0001; ns, not significant by one- way ANOVA (A,F) or two- tailed Student’s t- test (B,C,H). The online version of this article includes the following source data and figure supplement(s) for figure 1: Source data 1. Unedited, uncropped immunoblot for Figure 1D. Figure supplement 1. Complement component 1q (C1q) is expressed in the mouse colon. by immune cells located in the subepithelial compartment of the intestine and is largely absent from epithelial cells and intraepithelial immune cells. To identify intestinal immune cells that express C1q, we further analyzed the subepithelial CD45+ cell population by flow cytometry. Expression of C1q transcripts and protein was highest in CD11b+M- HCII+F4/80hi macrophages and was mostly absent from non- macrophage immune cells (Figure 1E–H). Thus, C1q is expressed by macrophages in the mouse small intestine. Macrophages are the primary source of C1q in the mouse gastrointestinal tract We next assessed whether macrophages are the primary source of C1q in the intestine by analyzing two mouse models. First, we depleted macrophages by injecting neutralizing antibodies directed against the receptor for colony- stimulating factor 1 (CSF1R)(Figure  2A), which is required for the development of a subset of lamina propria macrophages (Bogunovic et al., 2009) and all muscularis Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 3 of 30 Immunology and Inflammation Research article A intraperitoneal injection of anti-CSF1R 3 days macrophage depletion B s e g a h p o r c a m # ) 4 0 1 × ( d e r e v o c e r ) 5 0 1 × ( s l l e c B # d e r e v o c e r ) 5 0 1 × ( s l l e c T # d e r e v o c e r ** ns ns 6 4 2 0 5 4 3 2 1 0 4 3 2 1 C i n o s s e r p x e e v i t a e R l C1qa **** C1qb **** C1qc **** 2.0 1.5 1.0 0.5 0.0 2.0 1.5 1.0 0.5 0.0 2.0 1.5 1.0 0.5 0 Isotype control Anti-CSF1R 0.0 Isotype control Anti-CSF1R D E 0 8 / 4 F 106 105 104 0 10-4 C1qafl/fl × LysM-Cre G C1qa∆Mφ C1qafl/fl C1qaΔMφ 33.3% 3.73% i n o s s e r p x e a q 1 C e v i t a e R l *** **** 1.5 1.0 0.5 0.0 C1qafl/fl C1qaΔMφ C1qafl/fl C1qaΔMφ Sm. int. Colon 10-4 0 104 10510610-4 0 104 105106 C1q 15 10 ) 0 0 0 1 × y t i s n e t n F e c n e c s e r o u l f n a d e m i ( q 1 C C1qafl/fl C1qaΔMφ *** 5 ns ns ns i 0 CD11b+MHCII+F4/80lo CD11b+MHCII+F4/80hi CD11b+MHCII- CD11b- Figure 2. Macrophages are the primary source of complement component 1q (C1q) in the mouse gastrointestinal tract. (A) Macrophages were selectively depleted in C57BL/6 mice by intraperitoneal injection of anti- CSF1R antibody. Control mice were injected with isotype- matched non- specific antibodies. Mice were analyzed 72 hr after antibody injection. Panel was generated at Biorender.com. (B) Representative flow cytometry analysis of mouse small intestines after intraperitoneal injection of anti- CSF1R or isotype control antibody. All cells were gated as live CD45+. Macrophages were MHCII+ F4/80hi; B cells were CD19+; T cells were CD3+. Total small intestinal cell yields were 1.5 × 106 ± 4.9 × 105 cells. (C) Quantitative PCR (qPCR) measurement of C1qa, C1qb, and C1qc transcript abundance in mouse small intestines after intraperitoneal injection of anti- CSF1R or rat IgG2a (isotype control). Each data point represents one mouse and results are pooled from two independent experiments. (D) C1qafl/fl mice were crossed with LysM- Cre transgenic mice to generate mice having a macrophage- selective deletion of C1qa (C1qa∆Mφ mice). Panel was generated at Biorender.com. (E) Representative flow cytometry analysis of intracellular C1q expression in small intestinal macrophages from C1qafl/fl and C1qa∆Mφ mice. Mice were littermates from heterozygous crosses that remained co- housed. Cells were gated on live CD45+CD11b+MHCII+. (F) Quantitation of the flow cytometry analysis in (E). Each data point represents one mouse. Results are representative of two independent experiments. (G) qPCR measurement of C1qa transcript abundance in the small intestines (sm. int.) and colons of C1qafl/fl and C1qa∆Mφ littermates. Each data point represents one mouse. Error bars represent SEM. **p<0.01; ***p<0.001; ****p<0.0001; ns, not significant by the two- tailed Student’s t- test. The online version of this article includes the following figure supplement(s) for figure 2: Figure supplement 1. Complement component 1q (C1q) expression is lost systemically but preserved in the central nervous system of C1qa∆Mφ mice. macrophages (Muller et al., 2014). Antibody injection led to a >twofold reduction in the number of macrophages recovered from the small intestine (Figure 2B), and a corresponding reduction in small intestinal C1q gene expression (Figure 2C), suggesting that macrophages are the primary source of intestinal C1q. Second, we constructed a genetic model of C1q deficiency by crossing C1qafl/fl mice (Fonseca et al., 2017) to mice carrying the Lyz2- Cre transgene (LysM- Cre mice), which is selectively expressed in myeloid cells including macrophages (Figure 2D). These mice, hereafter designated as C1qaΔMϕ mice, lacked C1q expression in intestinal macrophages (Figure 2E and F). Importantly, C1qaΔMϕ mice had markedly lower C1q expression in both the small intestine and colon (Figure  2G), indicating that macrophages are the main source of C1q in the intestine. Unexpectedly, the C1qaΔMϕ mice also lost C1q gene expression in the lung, skin, kidney, and liver (but not the brain), and the C1q protein was undetectable in the serum (Figure  2—figure supplement 1). These findings indicate that macrophages are the primary source of C1q in the intestine and suggest that LysM+ macro- phages or macrophage- like cells are also the main sources of C1q in most extraintestinal tissues and the bloodstream. Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 4 of 30 Immunology and Inflammation Research article C1qaΔMφ mice do not show altered microbiota composition, barrier function, or resistance to enteric infection The classical complement pathway is a well- studied host defense system that protects against systemic pathogenic infection (Warren et al., 2002; Noris and Remuzzi, 2013). Circulating C1q activates the complement pathway by binding to antibody- antigen complexes or to bacterial cell surface mole- cules, and thus protects against systemic infection. Therefore, we assessed whether C1q promotes the immune defense of the intestine. We first determined whether C1q exhibits characteristics of known intestinal antimicrobial proteins, including induction by the intestinal microbiota and secretion into the gut lumen. C1qa was expressed at similar levels in the small intestines of germ- free and conventionally- raised mice (Figure  3A), suggesting that C1q expression is not induced by the gut microbiota. This contrasted with Reg3g, encoding the antimicrobial protein REG3G (Cash et al., 2006), which was expressed at a > twofold higher level in conventional as compared to germ- free mice (Figure 3A). Additionally, in contrast to REG3G, C1q was not detected in the gut lumen of either conventional or germ- free mice (Figure 3B). C1qa expression was also not markedly altered by a 24 hr oral infection with the intestinal pathogenic bacterial species Salmonella Typhimurium (Figure 3C). Although we cannot rule out the induction of C1q by longer- term pathogenic infections, these data indicate that C1q is not induced by the gut microbiota or by a 24  hr infection with S. typhimurium, in contrast to other intestinal antibacterial proteins. We next assessed whether C1q regulates the composition of the gut microbiota. 16 S rRNA gene sequencing analysis of the fecal microbiotas of C1qafl/fl and C1qaΔMϕ mice showed that the microbiota composition was not appreciably altered in the absence of macrophage C1q (Figure 3D). Analysis of 16 S rRNA gene copy number in mesenteric lymph nodes further indicated no statistically significant differences in translocation of the microbiota to the mesenteric lymph nodes (Figure 3E). We next challenged C1qafl/fl and C1qaΔMϕ mice with dextran sulfate sodium (DSS), which damages the colonic epithelium and exposes underlying tissues to the commensal microbiota. However, the sensitivity of the C1qaΔMϕ mice to DSS was similar to that of their C1qafl/fl littermates as assessed by change in body weight and histopathological analysis (Figure 3F; Figure 3—figure supplement 1). There was also no change in intestinal paracellular permeability in C1qaΔMϕ mice as measured by oral administration of FITC- dextran (Figure 3G). These results suggest that macrophage C1q does not substantially impact gut microbiota composition or intestinal epithelial barrier function. To determine whether C1q protects against enteric infection we conducted oral infection exper- iments with the enteric pathogen Citrobacter rodentium. We chose C. rodentium as our model organism for two reasons. First, C. rodentium is a non- disseminating pathogen, allowing us to test specifically for C1q’s role in intestinal infection. Second, C. rodentium clearance depends on immuno- globulins and complement component C3 (Belzer et al., 2011). Because C1q is bactericidal in concert with C3 and immunoglobulins, we predicted that C1qaΔMϕ mice would be more susceptible to C. rodentium infection. However, C1qaΔMϕ mice cleared C. rodentium similarly to their C1qafl/fl littermates (Figure 3H) and showed similar histopathology (Figure 3—figure supplement 2), indicating that C1q is dispensable for defense against C. rodentium infection. We also did not observe altered immunity in the absence of C1q. Measurement of transcripts encoding secreted immune effectors in the small intestines of C1qafl/fl and C1qaΔMϕ littermates revealed no statistically significant differences in cytokine expression (Figure 3I). Furthermore, there were no statistically significant differences in the percentages or absolute numbers of various T cell subsets, including Thelper1 (TH1), TH2, TH17, and regulatory T (Treg) cells between C1qafl/fl and C1qaΔMϕ mice (Figure 3J; Figure 3—figure supplement 3). Although total B cell numbers trended lower in C1qaΔMϕ mice, the difference was not statistically significant (Figure  3J; Figure  3—figure supplement 4). There were also no statistically significant differences in the percentages or absolute numbers of total plasma cells (Figure 3J; Figure 3—figure supplement 4), IgA+ plasma cells (Figure 3J; Figure 3— figure supplement 4), myeloid cells (Figure 3J; Figure 3—figure supplement 5), or innate lymphoid cells (Figure 3J; Figure 3—figure supplement 6) when comparing C1qafl/fl and C1qaΔMϕ mice. These results suggest that the absence of macrophage C1q has little impact on intestinal immunity. Alto- gether, our findings suggest that C1q does not participate substantially in intestinal immune defense and thus might have an intestinal function that is independent of its canonical role in activating the classical complement pathway. Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 5 of 30 Immunology and Inflammation Research article A i n o s s e r p x e e v i t a e R l E 2 1 0 ) e g n a h c l d o f 0 1 g o l ( -1 Sm. int. ** B C . t n i . Feces m GF CV S ns C1q REG3G CV CV GF Reg3g C1qa GF i n o s s e r p x e a q 1 C e v i t a e R l ns 1.5 1.0 0.5 0.0 S. Typhi- Uninf. murium D % ) ( e c n a d n u b a e v i t a e R l 100 80 60 40 20 0 Fecal microbiome C1qafl/fl C1qaΔMφ Bacterial class: Actinobacteria Coriobacteriia Bacteroidia Bacilli Clostridia Erysipelotrichi Betaproteobacteria Other 105 mLN ns F C1qafl/fl C1qaΔMφ 101 C1qaΔMφ C1qafl/fl S 6 1 I 110 100 i t h g e w y d o b l i a n g i r o % 90 80 70 C1qafl/fl C1qaΔMφ 0 1 2 3 4 5 6 7 3% DSS treatment (days) Small intestine s e c e f / g m U F C 0 1 g o l 7 6 5 4 3 2 1 0 -1 G n a r t x e d - C T F m u r e S I 250 200 C1qafl/fl C1qaΔMφ ns ) l m / g n ( 150 100 50 0 ns Control Indo- methacin H m u i t n e d o r . C ns C1qafl/fl C1qaΔMφ ns C1qafl/fl C1qaΔMφ 0 48 1 2 16 20 Days post-infection ns ns ns ns ns ns ns ns ns ns ns Ifng Il1 Tnfa Saa1 Saa2 Il4 Il5 Il6 Il10 Il13 Il17a Il17f Il23a T cells ns ns ns ns ns ns ns ns TH1 TH2 TH17 Treg B cells/plasma cells ns ns ns ns ns ns 40 30 20 10 s l l e c + 5 4 D C e v i l f o % d e r e v o c e r s l l e c # 0 ) 2.0 6 0 1 × ( 1.5 1.0 0.5 Myeloid cells ns 15 s l l e c ns ns 10 + 5 4 D C e v i l 5 ns f o % 0 ) 6 5 0 1 × ( ns 4 2 ns d e r e v o c e r s l l e c # 0.0 Plasma cells B cells IgA+ plasma cells 0 Macrophages Dendritic cells Monocytes 1.5 s l l e c 1.0 0.5 0.0 10 + 5 4 D C e v i l f o % ) 4 0 1 × ( d e r e v o c e r s l l e c # Innate lymphoid cells ns C1qafl/fl C1qaΔMφ ns ns ns 8 6 4 2 0 ns ns ILC1 ILC2 ILC3 i s e p o c e n e g A N R r e u s s i t g r e p 104 103 102 i n o s s e r p x e e v i t a e R l 2.5 2.0 1.5 1.0 0.5 0.0 J l l s20 e c + 4 D C 15 10 + 3 D C f o % ) 4 0 1 × ( d e r e v o c e r s l l e c # 5 0 10 8 6 4 2 0 Figure 3. C1qa∆Mφ mice do not show altered microbiota composition, barrier function, or resistance to enteric infection. (A) Small intestinal C1qa expression is not induced by the intestinal microbiota. Quantitative PCR (qPCR) measurement of Reg3g and C1qa transcript abundances in the small intestines of germ- free (GF) and conventional (CV) C57BL/6 mice. Each data point represents one mouse and the results are representative of two independent experiments. (B) C1q is not detected in the mouse intestinal lumen or feces. Representative immunoblot of an ammonium sulfate precipitation of intestinal luminal contents and feces from germ- free and conventional mice with detection of C1q. C1q in small intestinal tissue is shown for comparison at right. REG3G was analyzed as a control, as it is secreted into the intestinal lumen of conventional mice (Cash et al., 2006). Each lane represents multiple mice pooled (n=5 and 9 for germ- free and conventional, respectively) and the immunoblot is representative of three independent experiments. (C) C1q gene expression is not altered by acute enteric infection Figure 3 continued on next page Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 6 of 30 Immunology and Inflammation Research article Figure 3 continued with Salmonella typhimurium. qPCR measurement of C1qa transcript abundance in small intestinal tissue after oral inoculation of mice with 109 colony- forming units of S. typhimurium strain SL1344. Each data point represents one mouse, and the results are representative of two independent experiments. (D) Intestinal microbiota composition is not altered in C1qa∆Mφ mice. Phylogenetic analysis of 16 S rRNA gene sequences from fecal pellets collected from C1qafl/fl and C1qa∆Mφ littermates. Operational taxonomic units with an average of 100 reads and populations greater than or equal to 1% were included in the graphical analysis. Each bar represents one mouse. Data are available from the Sequence Read Archive under BioProject ID PRJNA793870. (E) C1qa∆Mφ mice do not show altered translocation of bacteria to mesenteric lymph nodes (mLN). 16 S rRNA gene copy numbers were measured by qPCR with reference to a standard curve. Each data point represents one mouse. (F) C1qa∆Mφ mice do not show altered susceptibility to dextran sulfate sodium (DSS)- induced colitis. Mice were provided with 3% DSS in drinking water and body weights were monitored for 7 days. n=4 and 6 for C1qafl/fl and C1qa∆Mφ littermates, respectively. Differences at each time point were not significant by the two- tailed Student’s t- test. (G) C1qa∆Mφ mice do not show altered intestinal permeability. To measure intestinal permeability, C1qafl/fl and C1qa∆Mφ littermates were gavaged with fluorescein isothiocyanate (FITC)- dextran (4 kDa), and serum FITC- dextran levels were determined by fluorescence microplate assay against a FITC- dextran standard curve. Indomethacin induces intestinal damage in mice and was used as a positive control. Each data point represents one mouse. (H) Time course of fecal Citrobacter rodentium burden following oral gavage of C1qafl/fl and C1qa∆Mφ mice with 5×108 colony forming units (CFU) of C. rodentium. n=5 and 5 for C1qafl/fl and C1qa∆Mφ littermates, respectively. Differences at each time point were not significant by the two- tailed Student’s t- test. (I) qPCR measurement of transcripts encoding secreted immune effectors in the small intestines of C1qafl/fl and C1qa∆Mφ littermates. Each data point represents one mouse. (J) Flow cytometry analysis of small intestinal immune cell subsets from C1qafl/fl and C1qa∆Mφ littermates. Gating strategies are shown in Figure 3—figure supplement 1 through 4. ILC, innate lymphoid cell. Total small intestinal cell yields were 8.8 × 106 ± 2.9 × 106 cells. Each data point represents one mouse. Sm. int., small intestine. Error bars represent SEM. **p<0.01; ns, not significant by the two- tailed Student’s t- test. The online version of this article includes the following source data and figure supplement(s) for figure 3: Source data 1. Unedited, uncropped immunoblot for Figure 3B. Figure supplement 1. Histological analysis of dextran sulfate sodium (DSS)- treated mice. Figure supplement 2. Colon histology of Citrobacter rodentium- infected mice. Figure supplement 3. Flow cytometry gating strategy for comparison of T cell populations in C1qafl/fl and C1qa∆Mφ mice. Figure supplement 4. Flow cytometry gating strategy for comparison of B cell and plasma cell populations in C1qafl/fl and C1qa∆Mφ mice. Figure supplement 5. Flow cytometry gating strategy for comparison of myeloid cell populations in C1qafl/fl and C1qa∆Mφ mice. Figure supplement 6. Flow cytometry gating strategy for comparison of innate lymphoid cell populations in C1qafl/fl and C1qa∆Mφ mice. C1q is expressed by muscularis macrophages that are located near enteric neurons Intestinal macrophages perform distinct functions depending on their anatomical location. Macro- phages in the lamina propria protect against invasion by pathogenic microbes and promote tissue repair (Grainger et al., 2017). In contrast, muscularis macrophages that reside in deeper intestinal tissues, such as the muscularis externa (Figure 4A), regulate enteric neurons and smooth muscle cells that drive gastrointestinal motility (De Schepper et al., 2018a; De Schepper et al., 2018b). Further- more, C1q has several well- described functions in regulating the development and activity of neurons of the central nervous system (Hammond et al., 2020; Hong et al., 2016), suggesting that intestinal C1q+ macrophages might interact with enteric neurons. These prior findings prompted us to charac- terize the anatomical localization of C1q+ macrophages within mouse intestinal tissues. The enteric nervous system is a network of neurons whose cell bodies are organized into two regions of the gastrointestinal tract: the submucosal plexus and the myenteric plexus (Figure  4A). Immunofluorescence microscopy revealed that C1q was localized close to submucosal plexus nerve fibers marked with βIII tubulins (TUBB3) in C1qafl/fl mice (Figure 4B and C). C1q was absent in C1qaΔMϕ mice despite the presence of similar overall numbers of CD169+ macrophages (Figure  4—figure supplement 1A). Although C1q immunoreactivity in the myenteric plexus was less pronounced, flow Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 7 of 30 Immunology and Inflammation e n i t s e t n i l l a m s n o o c l Research article A lumen epithelium (epi) lamina propria (LP) B epi C1qafl/fl lumen LP C1qaΔMφ C1qafl/fl (isotype control) neurons submucosal plexus (SP) SP/muscularis C1qafl/fl C1qaΔMφ circular muscle myenteric plexus longitudinal muscle l s i r a u c s u m P M M L C wild-type C57BL/6 epi LP Isotype control C1q CD169 (Mφ) DAPI (nuclei) C1qafl/fl C1qaΔMφ C1qafl/fl (isotype control) epi lumen SP/muscularis e n i t s e t n i l l a m s D l f / l f a q 1 C φ M Δ a q 1 C l f / l f a q 1 C φ M Δ a q 1 C SP/muscularis C1qafl/fl C1qaΔMφ C1q CD169 (Mφ) TUBB3 (neuron) DAPI (nuclei) C1q CD169 (Mφ) DAPI (nuclei) Longitudinal muscle-myenteric plexus (LMMP) C1q Csf1r (Mφ) HuC/D (neuron) Merge E 100 80 60 40 20 0 100 x a m f o % 80 60 40 20 0 100 80 60 40 20 e n i t s e t n i l l a m s n o o c l Small intestinal macrophages C1q- Mφ C1q+ Mφ CD169 ) 0 0 0 1 × ( y t i s n e t n i e c n e c s e r o u l f i n a d e M ARG1 TREM2 CD169 **** ARG1 ** TREM2 **** 30 25 20 15 10 5 0 20 15 10 5 0 6 5 4 3 2 1 0 0 -104 0 104 105 106 F4/80hi C1q- F4/80hi C1q+ Figure 4. Complement component 1q (C1q) is expressed by muscularis macrophages that are located near enteric neurons. (A) Graphic depicting the muscularis of the mouse small intestine. The lumen, epithelium (epi), lamina propria (LP), submucosal plexus (SP), and longitudinal muscle- myenteric plexus (LMMP) are indicated. Created at Biorender.com. (B) Immunofluorescence detection of C1q (violet) and macrophages marked with CD169 (green) in the small intestine and colon of C1qafl/fl and C1qa∆Mφ littermates. Nuclei were detected with 4’,6- diamidino- 2- phenylindole (DAPI; blue). Detection Figure 4 continued on next page Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 8 of 30 Immunology and Inflammation Research article Figure 4 continued with isotype control antibodies on C1qafl/fl small intestines is shown at right. Anti- rat IgG AlexaFluor 488 and streptavidin- Cy5 were used as secondary stains for CD169 and C1q, respectively. The intestinal surface is denoted with a red dotted line and the gut lumen, epithelium, and lamina propria are indicated. The approximate region encompassing the submucosal plexus and the muscularis is denoted with two white dotted lines. Examples of C1q+ areas are indicated with yellow arrows and examples of CD169+ macrophages are indicated with white arrowheads. Note that the violet staining near the bottom of the muscularis is non- specific, as indicated by its presence in the isotype control image. Images are representative of three independent experiments. Scale bars = 50 μm. (C) Immunofluorescence detection of C1q (violet), macrophages marked with CD169 (green), and neurons marked with TUBB3 (yellow) in the small intestines of wild- type C57BL/6 mice. Nuclei are detected with DAPI (blue). The epithelium and lamina propria are indicated. The approximate region encompassing the submucosal plexus and the muscularis is denoted with two white dotted lines. The expanded image area delineated by a yellow square shows an example of the close association between C1q and TUBB3+ neurons. Images are representative of images captured from three mice. Anti- rat IgG AlexaFluor 488, anti- rabbit IgG AlexaFluor 594, and streptavidin- Cy5 were used as secondary stains for CD169, TUBB3, and C1q, respectively, and an isotype control image is shown at upper right. Scale bars = 50 μm. (D) RNAscope detection of C1qa (green), muscularis macrophages marked by Csf1r (red), and immunofluorescence detection of enteric neuronal ganglia by HuC/D (blue) in LMMP wholemounts of small intestines and colons from C1qafl/fl and C1qa∆Mφ mice. The expanded area denoted by a yellow square shows a close association between C1qa- expresssing muscularis macrophages and enteric neurons. Images are representative of three independent experiments. Scale bars = 50 μm. (E) C1q+ intestinal macrophages express genes that are characteristic of nerve- adjacent macrophages. Flow cytometry analysis of CD169, Arginase 1, and TREM2 on C1q- and C1q+ macrophages recovered from the small intestines of wild- type C57BL/6 mice. Median fluorescence intensities from multiple mice are quantified in the panels at the right. Each data point represents one mouse (n=5–6 mice), and the results are representative of two independent experiments. Error bars represent SEM. **p<0.01; ****p<0.0001 by the two- tailed Student’s t- test. Epi, epithelium; LP, lamina propria; SP, submucosal plexus; Mφ, macrophage; DAPI, 4 ,6- diamidino- 2- phenylindole, LMMP, longitudinal muscle- myenteric plexus. Error bars represent SEM. ′ ns, not significant by the two- tailed Student’s t- test. The online version of this article includes the following figure supplement(s) for figure 4: Figure supplement 1. Flow cytometry analysis of complement component 1q (C1q) and CD169 expression on small intestinal macrophages. cytometry analysis indicated that C1q was expressed by macrophages recovered from the muscu- laris (Figure 4—figure supplement 1B), which encompasses the myenteric plexus. Supporting this finding, RNAscope analysis of longitudinal muscle- myenteric plexus (LMMP) wholemounts revealed C1qa- expressing macrophages next to HuC/D+ neurons (Figure 4D). Consistent with other validation data (Figure 2E–G), C1qa signals were mostly absent in muscularis macrophages of C1qaΔMϕ mice. Finally, C1q- expressing intestinal macrophages showed elevated expression of Arginase 1, CD169, and TREM2 (triggering receptor expressed on myeloid cells 2) (Figure  4E), which are enriched on macrophages with known neuromodulatory functions (Colonna, 2003; Paloneva et al., 2002; Ural et al., 2020). Thus, C1q- expressing intestinal macrophages are located near enteric neurons in the submucosal and myenteric plexuses and express proteins that are characteristic of nerve- adjacent macrophages in other tissues. Numbers of enteric neurons are similar in C1qafl/fl and C1qaΔMϕ mice Gut macrophages engage in crosstalk with the enteric nervous system and regulate functions, including gastrointestinal motility, that depend on the enteric nervous system (Muller et al., 2014). This cross- talk involves the exchange of specific proteins such as bone morphogenetic protein 2 (BMP2) (Muller et al., 2014). Furthermore, microglial C1q promotes central nervous system development while also regulating neuronal transcriptional programs (Benavente et al., 2020; Schafer et al., 2012; Stevens et al., 2007). Given that intestinal C1q+ macrophages phenotypically resemble peripheral neuromod- ulatory macrophages and reside near enteric neurons, we postulated that macrophage- derived C1q might also regulate enteric nervous system function. As an initial test of this idea, we compared the numbers of enteric neurons in C1qaΔMϕ and C1qafl/fl mice. Immunofluorescence analysis of LMMP wholemounts from the small intestine and colon revealed a similar number of HuC/D+ neurons and a similar density of TUBB3+ neuronal fibers (Figure 5A and B). There were also similar numbers of specific neuronal subsets, including excitatory (Chat+) and inhib- itory (Nos1+) neurons (Figure 5C and E), and a similar density of S100B+ enteric glial cells (Figure 5D and E). Thus, the anatomical features of the enteric nervous system are not appreciably altered in C1qaΔMφ mice. C1qaΔMϕ mice have altered gastrointestinal motility We next assessed whether C1qaΔMϕ mice show evidence of altered neuronal function. We performed RNAseq on the colonic LMMP from C1qafl/fl and C1qaΔMϕ littermates and then conducted unbiased Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 9 of 30 Immunology and Inflammation Research article Small intestine Colon B 5000 2 m m / s n o r u e N 4000 3000 2000 1000 ns C1qafl/fl C1qaΔMφ ns 0 Small intestine Colon HuC/D TUBB3 HuC/D nNOS Chat Merge Longitudinal muscle-myenteric plexus D Longitudinal muscle- myenteric plexus S100B E i n o s s e r p x e e v i t a e R l e n i t s e t n i l l a m S l n o o C l f / l f a q 1 C φ M Δ a q 1 C l f / l f a q 1 C φ M Δ a q 1 C e n i t s e t n i l l a m S l n o o C 2.5 2.0 1.5 1.0 0.5 0.0 3.0 2.5 2.0 1.5 1.0 0.5 0.0 4 3 2 1 0 Nos1 ns C1qafl/fl C1qaΔMφ ns Chat ns ns S100b ns ns Ileum Colon A l f / l f a q 1 C φ M Δ a q 1 C C l f / l f a q 1 C φ M Δ a q 1 C l f / l f a q 1 C φ M Δ a q 1 C Figure 5. Numbers of enteric neurons are similar in C1qafl/fl and C1qa∆Mφ mice. (A) Immunofluorescence analysis of enteric neuronal ganglia marked with HuC/D (red) and neuronal fibers marked with TUBB3 (green) in LMMP wholemounts of small intestines and colons from C1qafl/fl and C1qa∆Mφ mice. Anti- mouse IgG AlexaFluor 594 and anti- rabbit IgG AlexaFluor 488 were used as secondary stains for HuC/D and TUBB3, respectively. Images are representative of three independent experiments. Scale bars = 50 μm. (B) Quantification of total enteric neurons per unit area (mm2) from the images shown in panel (A). Data are pooled from two independent experiments. Each data point represents one mouse. (C) Visualization of specific neuronal subsets in the LMMP from C1qafl/fl and C1qa∆Mφ mice by RNAscope detection. Inhibitory neurons were marked by Nos1 (green) and excitatory neurons were marked by Chat (red). Neuronal nuclei marked by HuC/D (blue) were detected by immunofluorescence. Images are representative of two independent experiments. Scale bars = 50 μm. (D) Immunofluorescence detection of enteric glial cells marked by S100B (green) in LMMP wholemounts from the small intestines and colons of C1qafl/fl and C1qa∆Mφ mice. Images are representative of two independent experiments. Scale bars = 50 μm. (E) qPCR analysis of Nos1, Chat, and S100b in the LMMP of small intestines and colons from C1qafl/fl and C1qa∆Mφ mice. Each data point represents one mouse. Error bars represent SEM. ns, not significant by the two- tailed Student’s t- test. LMMP, longitudinal muscle- myenteric plexus. Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 10 of 30 Immunology and Inflammation Research article Gene Set Enrichment Analysis. Of the 22 biological pathways that were enriched in the LMMP of C1qaΔMϕ mice, 17 were related to neuronal development or function, including synapse organization, dendrite development, and neurotransmitter secretion (Figure  6A). Our analysis also identified 30 differentially expressed genes with known roles in regulating neuronal activity (e.g. Dusp26), synaptic transmission (e.g. Rasgrf2), and neuropeptide signaling (e.g. Tacr2) (Mao et al., 2017; Schwechter et al., 2013; Yang et al., 2017; Figure 6B). We also compared the genes differentially expressed in the C1qaΔMϕ mice to those differentially expressed in the TashT mouse line, which contains an inser- tional mutation that leads to dysregulated gut motility. The gut motility phenotypes in the TashT line are comparable to Hirschsprung’s disease, a human genetic disorder resulting in incomplete develop- ment of the enteric nervous system (Bergeron et al., 2015). A comparative analysis revealed a statis- tically significant overlap in the transcriptional changes in the colonic LMMP of C1qaΔMϕ mice and the neural crest cells of TashT mice (Figure 6B). These results suggested that macrophage C1q impacts enteric nervous system gene expression and function. Efficient coordination of gastrointestinal motility is necessary for proper digestion, nutrient absorp- tion, and excretion. Given that muscularis macrophages regulate enteric nervous system functions that govern gastrointestinal motility (Muller et  al., 2014), we assessed whether macrophage C1q impacts gut motility. We first tested this idea by measuring gut transit time using the nonabsorbable dye Carmine Red. C1qaΔMϕ and C1qafl/fl littermates were gavaged with the dye and the time to the first appearance of the dye in the feces was recorded. Transit times were decreased in C1qaΔMϕ mice rela- tive to their C1qafl/fl littermates, indicating accelerated gut motility (Figure 6C). This was not due to a change in the length of either the small intestine or the colon, which were unaltered in the C1qaΔMϕ mice (Figure 6D). By contrast, gut transit time was unchanged in C3-/- mice, suggesting that macro- phage C1q impacts gut motility independent of its canonical function in the classical complement pathway (Figure 6C). Accelerated transit was also observed in the small intestines of C1qaΔMϕ mice as assessed by rhodamine dye transit assay (Figure 6E). To assess colonic motility, we measured the expulsion time after intrarectal insertion of a glass bead and found that C1qaΔMϕ mice had accelerated colonic motility when compared to C1qafl/fl littermates (Figure 6F). Our results thus suggest that the absence of macrophage C1q results in defective enteric nervous system function and dysregulated gastrointestinal motility. A limitation of in vivo measures of gut motility is that they cannot distinguish between defects in ‘intrinsic’ enteric neurons and ‘extrinsic’ neurons that innervate the gastrointestinal tract (Berthoud et al., 2004; Uesaka et  al., 2016). We, therefore, used an ex vivo organ bath system to specifically assess enteric nervous system function by measuring the activity of colonic migrating motor complexes (CMMC; rhythmic peristaltic contractions that depend on the enteric nervous system) (Obata et al., 2020). Spatio- temporal mapping revealed that the colons of C1qaΔMϕ mice had increased total number, frequency, and velocity of CMMC as compared to C1qafl/fl littermates (Figure 6G and H; Figure 6—video 1; Figure 6— video 2). This indicated that the colons of C1qaΔMϕ mice maintained increased neurogenic peristaltic activity compared to their C1qafl/fl littermates even in the absence of gut- extrinsic signals. Thus, the absence of macrophage C1q increases enteric nervous system- dependent peristalsis and accelerates gut transit. Taken together, our findings reveal that macrophage C1q regulates gastrointestinal motility. Discussion Here, we have identified a role for C1q in regulating gastrointestinal motility. We discovered that macrophages are the primary source of C1q in the mouse intestine and that macrophage C1q regu- lates enteric neuronal gene expression and gastrointestinal transit time. Our findings reveal a previ- ously unappreciated function for C1q in the intestine and help to illuminate the molecular basis for macrophage- mediated control of gut motility. Our study identifies macrophages as the main source of C1q in the mouse small intestine and colon. Both transient antibody- mediated depletion of macrophages and in vivo deletion of the C1qa gene from macrophages led to a marked reduction in intestinal C1q expression. The C1qaΔMϕ mice also lacked C1q in the circulation, indicating that LysM+ macrophages or macrophage- like cells are the sources of circulating C1q in the absence of infection. This enhances findings from prior studies indicating that monocytes, macrophages, and immature dendritic cells are the main sources of C1q in the bloodstream (El- Shamy et al., 2018). Importantly, the C1qaΔMϕ mice retained C1q expression Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 11 of 30 Immunology and Inflammation Research article A GO analysis of C1qaΔMφ vs. C1qafl/fl (colonic LMMP) Synapse organization Synapse assembly Cell adhesion Trans-synaptic signaling Neuron projection development Dendrite development Axon development Synaptic vesicle transport Dendrite morphogenesis Neuron projection guidance Neurotransmitter secretion Cell-cell adhesion Glutamatergic synaptic transmission Synapse structure/activity Regulation of membrane potential T helper cell 1 differentiation Neurotransmitter transportation Regulation of secretory pathway(s) Regulation of exocytosis Postsynapse organization Nervous system development Cation channel activity 0 F genes per pathway 200 150 100 50 2 4 -log10(p value) 6 8 G B Clca1 Ang4 Fcgbp Mybpc2 Actn2 Six2 Isl2 Scin Aldh1a2 Pdzd2 Dusp26 9530036M11Rik Tacr2 Colec10 Mab21I2 Sdk1 Rasgrf2 Mettl7a1 Cpt2 Acox1 Rhou Pex11a Slc25a20 Mt1 Yam1 Smim24 Mgst1 Trp53i11 C2 Acaa1b ) e g n a h c l d o f ( 2 g o L ) l f / l f a q 1 C : φ M Δ a q 1 C ( 1 0 1 - C ) s r u o h ( e m i t t i s n a r t l a t o T D Carmine Red transit assay ns *** 4 3 2 1 0 C1qafl/fl C1qaΔMφ C3-/- Intestinal length Rhodamine B transit assay C1qafl/fl C1qaΔMφ ** E e c n e c s e r o u F % l d e r e v o c e r 50 40 30 20 10 0 C1qafl/fl ) m c ( h g n e L t 40 30 20 10 0 ns C1qafl/fl C1qaΔMφ C1qaΔMφ * * Intestinal segment: 2 Stomach 68 4 Sm. int. 1 0 12 14 Cecum Colon ns Sm. int. Colon Colonic bead expulsion assay ) s d n o c e s ( e m i i t n o s u p x E l ** 400 300 200 100 0 C1qaΔMφ C1qafl/fl ) s ( e m T i 0 500 1000 1500 2000 2500 0 0 500 1000 1500 2000 2500 0 Colonic migrating motor complexes (CMMC) C1qafl/fl 10 20 30 40 50 60 C1qaΔMφ 10 20 40 30 Gut length (mm) 50 H i ) n m / s t n u o c ( y c n e u q e r f C M M C Colonic migrating motor complexes (CMMC) 6 4 2 ) C M M C / n m i ( d o i r e p C M M C * 0.5 0.4 0.3 0.2 0.1 0.0 C1qaΔMφ C1qafl/fl * 1.5 ) s / m m ( *** l y t i c o e v C M M C 1.0 0.5 0.0 C1qaΔMφ C1qafl/fl 0 C1qaΔMφ C1qafl/fl Figure 6. C1qa∆Mφ mice have altered gastrointestinal motility. (A) RNA- seq was performed on colonic LMMP from C1qa∆Mφ and C1qafl/fl littermates. Annotated gene ontology (GO) biological processes were assigned to genes that were differentially expressed in C1qa∆Mφ mice when compared to their C1qafl/fl littermates. GO biological processes associated with neurons are in bold type. The dotted line indicates the cutoff for statistical significance. Five mice per group were analyzed as pooled biological replicates. Data are available from the Sequence Read Archive under BioProject ID PRJNA793870. (B) The colonic longitudinal muscle myenteric plexus of C1qa∆Mφ mice have a transcriptional profile like that of mice with a gastrointestinal motility disorder. RNA- seq was performed on the colonic longitudinal muscle- myenteric plexus from five C1qafl/fl and five C1qa∆Mφ littermates. Genes that were differentially expressed are represented in a heatmap that depicts log2(fold change). Genes that also showed altered expression in the TashT mouse line, which is a model of human Hirschsprung’s disease (Bergeron et al., 2015), are indicated in red. Statistical significance of the overlap between differentially expressed genes in C1qa∆Mφ and TashT mice was determined by Fisher’s exact test (p=0.0032). (C) Measurement of total intestinal transit time in C1qafl/fl and C1qa∆Mφ littermates and C3-/- mice. Mice were gavaged with 100 μl of Carmine Red [5% (w/v in 1.5% methylcellulose)]. Fecal pellets were collected every 15 min and transit time was recorded when the dye was first observed in the feces. Each data point represents one mouse and the results are pooled from five independent experiments. (D) Intestinal tract length is not altered in C1qa∆Mφ mice. Small intestines and colons from C1qafl/fl and C1qa∆Mφ littermates were excised and measured. Each data point represents one mouse. (E) Transit of rhodamine B- dextran through the intestines of C1qafl/fl and C1qa∆Mφ littermates. Mice were sacrificed 90 min after gavage with rhodamine B- dextran. The intestines were divided into 16 segments, the rhodamine B fluorescence was measured in each segment (top panel), and the geometric center of the fluorescence was determined for each mouse (bottom panel). Each data point represents one mouse and the results were pooled from four independent experiments. (F) Colonic motility was measured by determining the expulsion time of a glass bead inserted intrarectally into C1qafl/fl and C1qa∆Mφ littermates. Each data point represents one mouse and the results are representative of three independent experiments. (G) Representative spatiotemporal maps of colonic migrating motor complex (CMMC) formation in colons of C1qafl/fl and C1qa∆Mφ mice. Representative video recordings were captured in Figure 6—video 1 (C1qafl/fl mice) and Figure 6—video 2 (C1qa∆Mφ mice). Each map represents one mouse and is representative of two independent experiments. (H) Analysis of CMMC parameters in colons of C1qafl/fl and C1qa∆Mφ mice. Each data point represents one mouse (for CMMC frequency and CMMC period) Figure 6 continued on next page Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 12 of 30 Immunology and Inflammation Research article Figure 6 continued or one individual CMMC event (for velocity). Data are pooled from two independent experiments. LMMP, longitudinal muscle- myenteric plexus; sm. int., small intestine. Error bars represent SEM. *p<0.05; **p<0.01; ***p<0.001; ****p<0.0001; ns, not significant by the two- tailed Student’s t- test. The online version of this article includes the following video and figure supplement(s) for figure 6: Figure supplement 1. Single- cell RNA- seq analysis of intestinal macrophages from C1qa∆Mφ and C1qafl/fl littermates. Figure supplement 2. The gene encoding complement component 1q (C1q) receptor BAI1 (Adgrb1) is expressed by enteric neurons. Figure 6—video 1. Ex vivo recording of colonic peristalsis in C1qafl/fl mice. https://elifesciences.org/articles/78558/figures#fig6video1 Figure 6—video 2. Ex vivo recording of colonic peristalsis in C1qa∆Mφ mice. https://elifesciences.org/articles/78558/figures#fig6video2 in the brain, allowing us to analyze the effects of C1q deficiency without possible confounding effects on the central nervous system. C1q has two known physiological functions that are distinct and vary according to tissue context. C1q was originally discovered as having a role in the classical complement pathway, which tags and destroys invading microbes (Noris and Remuzzi, 2013; Schifferli et  al., 1986). Circulating C1q binds to invading microorganisms and recruits additional proteins that assemble into the membrane attack complex (MAC) (Kishore and Reid, 2000). C1q- mediated MAC formation has been described primarily in the bloodstream, where the necessary accessory proteins are present at high levels (Davis et al., 1979). However, even in the absence of infection, C1q is expressed in tissues such as the brain, where it regulates neuronal development and function (Kouser et al., 2015; van Schaarenburg et al., 2016). Our findings suggest that C1q does not play a central role in the immune defense of the intestine. First, we found that intestinal C1q expression was not induced by gut commensals or pathogens and was not deposited into the gut lumen. Second, C1q deficiency did not markedly alter gut microbiota composition or the course of disease after DSS treatment. There were also no major changes in cyto- kine expression or numbers and frequencies of intestinal immune cells that would indicate dysregu- lated interactions with the microbiota. Third, C1q was not required for clearance of C. rodentium, a non- disseminating enteric pathogen whose clearance requires antigen- specific IgG and complement component 3 (C3) (Belzer et  al., 2011). Although we cannot rule out a role for C1q in immune defense against other intestinal pathogens, or during chronic inflammation or infection, these findings suggest that C1q is not essential for intestinal immune defense in mice. Instead, our results indicate that C1q influences enteric nervous system function and regulates intestinal motility. First, C1q- expressing macrophages were present in the myenteric and submucosal plexuses and resided close to enteric neurons. Second C1q- expressing macrophages expressed cell surface markers like those expressed by nerve- adjacent C1q- expressing macrophages in the lung (Ural et al., 2020). Third, macrophage- specific deletion of C1qa altered enteric neuronal gene expres- sion. Finally, consistent with the altered neuronal gene expression, macrophage- specific C1qa dele- tion altered gastrointestinal motility in both the small and large intestines. Thus, our results suggest that the function of C1q in the intestine is similar to its function in the brain, where it regulates the development and function of neurons (Benoit and Tenner, 2011; Kouser et al., 2015; van Schaaren- burg et al., 2016). A function for macrophage C1q in intestinal motility adds to the growing understanding of how gut macrophages regulate intestinal peristalsis. Prior work has shown that CSF1R+ macrophages selectively localize to the muscularis of the mouse intestine (Muller et  al., 2014; Gabanyi et  al., 2016). These macrophages secrete BMP2, which activates enteric neurons that regulate colonic muscle contraction and thus colonic motility (Muller et al., 2014). We found that depletion of CSF1R+ macrophages reduces intestinal C1q expression and that macrophage- specific deletion of C1qa alters enteric neuronal gene expression and activity. Thus, our findings suggest that C1q is a key component of the macrophage- enteric nervous system axis. An important remaining question concerns the molecular mechanism by which C1q regulates gut motility. One possibility is that C1q shapes microbiota composition which, in turn, impacts gut motility. This idea is suggested by studies in zebrafish showing that a deficiency in intestinal macrophages leads to altered gut microbiota composition relative to wild- type zebrafish (Earley et al., 2018). Other Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 13 of 30 Immunology and Inflammation Research article studies in zebrafish and mice have shown that severe defects in enteric nervous system develop- ment produce changes in gut microbiota composition that are linked to dysregulated gut motility (Rolig et al., 2017; Johnson et al., 2018). However, we did not observe prominent changes in the composition of the gut microbiota in C1qaΔMϕ mice, arguing against a central role for the micro- biota in C1q- mediated regulation of gut motility. A second possibility is that the absence of C1q leads to immunological defects that alter gut transit time. This idea is consistent with studies showing that T- cell cytokines can influence gastrointestinal motility (Akiho et al., 2011). However, this seems unlikely given the lack of pronounced immunological abnormalities in the intestines of C1qaΔMϕ mice. A third possibility is that C1q changes the cell- intrinsic properties of the macrophages that express it, thus altering their interactions with neurons to influence gut motility. We explored this possibility by conducting single- cell RNA sequencing (scRNAseq) on macrophages isolated from small intestinal cell suspensions (Figure  6—figure supplement 1A). We identified 11 unique macrophage clusters and found that C1qaΔMϕ mice had alterations in at least three highly represented clusters (Figure 6— figure supplement 1B). Gene set enrichment analysis of the most significantly altered clusters did not reveal any pronounced functional differences (Figure 6—figure supplement 1C). However, anal- ysis of the differentially expressed genes across all macrophage clusters indicated lowered repre- sentation of transcripts that are linked to control of macrophage differentiation or functional states, such as Malat1, Neat1, and Etv3 (Cui et  al., 2019; Gao et  al., 2020; Villar et  al., 2023; Zhang et al., 2019; Figure 6—figure supplement 1D). Furthermore, a recent study identified a set of 13 ‘microglia- specific genes’ that represent a unique transcriptional overlap between microglia in the CNS and intestinal macrophages (Verheijden et al., 2015). In macrophages from C1qafl/fl mice, we observed the expression of eight ‘microglia- specific genes’ whose expression was lowered or lost in macrophages from C1qaΔMϕ mice (Figure 6—figure supplement 1E). Thus, it is possible that altered intestinal motility could arise in part from cell- intrinsic functional alterations in C1q- deficient intestinal macrophages. Such alterations could arise from a C1q autocrine signaling loop or C1q could imprint a neuronal function that feeds back to regulate macrophage gene expression as exemplified in Muller et al., 2014. A fourth possibility is that C1q+ macrophages engulf specific neurons. Indeed, macrophages restrain neurogenesis in the enteric nervous system through phagocytosis of apoptotic neurons, which is consistent with the ability of C1q to opsonize dying host cells (Kulkarni et al., 2017; Botto et al., 1998; Korb and Ahearn, 1997). However, we observed no marked differences in the overall numbers of enteric neurons or numbers of excitatory and inhibitory neurons when comparing C1qaΔMϕ and C1qafl/fl mice, which argues against this possibility. A fifth possibility is that C1q acts directly on enteric smooth muscle cells that regulate gut motility. Although we cannot rule out this possibility, our transcriptional profile of the colonic myenteric plexus of C1qaΔMϕ mice suggests that most of the transcriptional changes were associated with neuronal function and homeostasis. Given that the C1qaΔMϕ mice showed altered neuronal gene expression, a sixth possibility is that C1q interacts directly with enteric neurons or glial cells as a signaling molecule. Like macrophage- produced BMP2 (Muller et al., 2014), C1q might bind to specific receptors on neurons to regulate their activity. In support of this idea, we observed that mouse enteric neurons express Adgrb1, which encodes BAI1 (Figure 6—figure supplement 2A and B), a recently identified C1q receptor on human neural stem cells (Benavente et  al., 2020). These data suggest a possible signaling axis for C1q- mediated control of enteric nervous system function. Our findings on intestinal C1q have implications for human intestinal disease. Indeed, single- cell RNAseq analysis shows that macrophages recovered from the human intestinal muscularis selectively express C1q when compared to lamina propria macrophages (Domanska et al., 2022). Dysregulated peristalsis is a characteristic of irritable bowel syndrome (Vrees et al., 2002) and is present in a subset of inflammatory bowel disease patients (Bassotti et al., 2014). Our finding that macrophage C1q regulates gut motility could suggest new strategies to prevent or treat these diseases. Additionally, most humans with C1q deficiency develop systemic lupus erythematosus (SLE). Since C1q can target cellular debris for phagocytosis, it is thought that C1q deficiency results in increased exposure of self- antigen to the immune system, thereby reducing immune tolerance and causing autoimmune disease (Macedo and Isaac, 2016). Furthermore, roughly 42.5% of SLE patients report gastrointestinal symptoms that range from acute abdominal pain to chronic intes- tinal obstruction (Fawzy et al., 2016; Tian and Zhang, 2010). The exact cause of these symptoms Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 14 of 30 Immunology and Inflammation Research article is unclear. Given that C1q deficiency is strongly correlated with SLE in humans and alters gut motility in mice, we suggest that C1q could be a therapeutic target for SLE patients that present with chronic constipation or other forms of dysregulated intestinal motility. Key resources table Materials and methods Reagent type (species) or resource Strain, strain background (Mus musculus) Designation Source or reference Identifiers Additional information C1qafl/fl; B6(SJL)- C1qatm1c(EUCOMM) Wtsi/TennJ Jackson Laboratory; Fonseca et al., 2017 Stock #031261 Strain, strain background (Mus musculus) LysM- Cre; B6.129P2- Lyz2tm1(cre)Ifo/J Jackson Laboratory; Clausen et al., 1999 Stock #004781 Strain, strain background (Mus musculus) C1qa∆MΦ this paper Generated by crossing C1qafl/fl mice with LysM- Cre mice Strain, strain background (Mus musculus) C3-/-; B6.129S4- C3tm1Crr/J Jackson Laboratory; Wessels et al., 1995 Stock #029661 Strain, strain background (Mus musculus) Germ- free C57BL/6 J mice UT Southwestern Gnotobiotics Core Facility Strain, strain background (Salmonella enterica) Strain, strain background (Citrobacter rodentium) Salmonella enterica subsp. enterica serovar Typhimurium strain SL1344 Dr. Vanessa Sperandio; Eichelberg and Galán, 1999 Citrobacter rodentium strain DBS100 ATCC Strain# 51459 Antibody Anti- Actin HRP (rabbit monoclonal) Cell Signaling Clone: 13E5 Immunoblot (1:5000) Antibody Anti- ARG1 (sheep monoclonal) R&D Systems Clone: P05089 Flow (1:100) Antibody Anti- B220 (rat monoclonal) Thermo Fisher Clone: RA3- 6B2 Flow (1:500) Antibody Anti- C1q (rat monoclonal) Cedarlane Laboratories Clone: RmC7H8 Flow (1:50) Antibody Anti- C1q (rabbit polyclonal) Thermo Fisher Cat# PA5- 29586 Immunoblot (1:500) Antibody Anti- C1q- biotin (mouse monoclonal) Abcam Clone: JL1 ELISA (1:1000); Immunofluorescence (1:100) Antibody Anti- CD3 (rat monoclonal) Thermo Fisher Clone: 17A2 Flow (1:200) Antibody Anti- CD4 (rat monoclonal) BioLegend Clone: GK1.5 Flow (1:500) Antibody Anti- CD11b (rat monoclonal) Thermo Fisher Clone: M1/70 Flow (1:200) Antibody Anti- CD11c (Armenian hamster monoclonal) Thermo Fisher Clone: N418 Flow (1:500) Antibody Anti- CD16/32 (rat monoclonal) BioLegend Clone: 93 Fc receptor block (1:1000) Antibody Anti- CD19 (rat monoclonal) BioLegend Clone: 1D3 Flow (1:500) Antibody Anti- CD45 (rat monoclonal) BioLegend Clone: 30- F11 Flow (1:500) Antibody Anti- CD90.2 (rat monoclonal) BioLegend Clone: 30- H12 Flow (1:500) Antibody Anti- CD169 (rat monoclonal) BioLegend Clone: 3D6.112 Flow (1:200) Continued on next page Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 15 of 30 Immunology and Inflammation Research article Continued Reagent type (species) or resource Designation Source or reference Identifiers Additional information Antibody Anti- CD169 (rat monoclonal) Abcam Clone: 3D6.112 Immunofluorescence (1:200) Antibody Anti- CSF1R (rat monoclonal) Bio X Cell Antibody Anti- F4/80 (rat monoclonal) BioLegend Cat# AFS98 Clone: BM8 Macrophage depletion (100 mg/kg) Flow (1:100) Antibody Anti- FoxP3 (rat monoclonal) Thermo Fisher Clone: FJK- 16s Flow (1:50) Antibody Anti- GATA3 (mouse monoclonal) BD Biosciences Clone: L50- 823 Flow (1:50) Antibody Anti- IgA (rat monoclonal) Thermo Fisher Clone: 11- 44- 2 Flow (1:50) Antibody Anti- LY6C (rat monoclonal) BioLegend Clone: RB6- 8C5 Flow (1:500) Antibody Anti- MHCII (rat monoclonal) Thermo Clone: M5/114.15.2 Flow (1:500) Antibody Anti- REG3G antiserum (rabbit polyclonal) Cash et al., 2006; antiserum generated by Pacific Biosciences Immunoblot (1:1000) Antibody Anti- RORγt (rat monoclonal) Thermo Fisher Clone: AFKJS- 9 Flow (1:50) Antibody Anti- T- BET (mouse monoclonal) BioLegend Clone: 4B10 Flow (1:50) Antibody Anti- TREM2 (rat monoclonal) R&D Systems Clone: 237920 Flow (1:200) Antibody Anti- TUBB3 (rabbit polyclonal) Abcam Cat# ab18207 Immunofluorescence (1:200) Antibody Anti- S100β (rabbit polyclonal) Dako Cat# GA504 Immunofluorescence Antibody Anti- HuC/D (rabbit monoclonal) Abcam Cat# ab184267 Immunofluorescence (1:400) Antibody Goat anti- rabbit IgG HRP conjugate Antibody secondary antibodies – donkey polyclonal anti- rabbit/rat/mouse AlexaFluor 488/594/647 Antibody mouse IgG1 Abcam Cat# ab6721 Immunoblot (1:5000) Invitrogen Abcam Immunofluorescence (1:400) Cat# ab18443 ELISA (10 μg/ml) Antibody Rat IgG2a Thermo Fisher Clone: 2A3 Isotype control for anti- CSF1R macrophage depletion (100 mg/kg) Antibody Rat IgG1 PE isotype control Cedarlane Laboratories Cat# CLCR104 Flow (1:50) Sequence- based reagent mouse C1qa TaqMan assay Thermo Fisher Sequence- based reagent mouse C1qb TaqMan assay Thermo Fisher Sequence- based reagent mouse C1qc TaqMan assay Sequence- based reagent mouse Chat TaqMan assay Sequence- based reagent mouse Nos1 TaqMan assay Thermo Fisher Thermo Fisher Thermo Fisher Sequence- based reagent mouse S100b TaqMan assay Thermo Fisher Sequence- based reagent mouse Reg3g TaqMan assay Thermo Fisher Sequence- based reagent mouse Ifng TaqMan assay Sequence- based reagent mouse Il4 TaqMan assay Thermo Fisher Thermo Fisher Continued on next page Assay ID: Mm00432142_m1 Assay ID: Mm01179619_m1 Assay ID: Mm00776126_m1 Assay ID: Mm01221880_m1 Assay ID: Mm01208059_m1 Assay ID: Mm00485897_m1 Assay ID: Mm00441127_m1 Assay ID: Mm01168134_m1 Assay ID: Mm00445259_m1 Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 16 of 30 Immunology and Inflammation Research article Continued Reagent type (species) or resource Designation Source or reference Identifiers Additional information Sequence- based reagent mouse IL5 TaqMan assay Sequence- based reagent mouse Il10 TaqMan assay Sequence- based reagent mouse Il13 TaqMan assay Sequence- based reagent mouse Il17a TaqMan assay Sequence- based reagent mouse Il17f TaqMan assay Thermo Fisher Thermo Fisher Thermo Fisher Thermo Fisher Thermo Fisher Sequence- based reagent mouse 18 S gene TaqMan assay Thermo Fisher Sequence- based reagent bacterial 16 S universal rRNA forward primer Gift from Dr. Andrew Koh Sequence- based reagent Bacterial 16 S universal rRNA reverse primer Gift from Dr. Andrew Koh Assay ID: Mm00439646_m1 Assay ID: Mm01288386_m1 Assay ID: Mm00434204_m1 Assay ID: Mm00439618_m1 Assay ID: Mm00521423_m1 Assay ID: Mm03928990_g1 5’- ACTC CTAC GGGA GGCA GCAG T- 3 ’ 5’- ATTA CCGC GGCT GCTG GC- 3’ Sequence- based reagent bacterial 16 S V3 - rRNA gene forward primer Thermo Fisher; (Klindworth et al., 2013) 16 S rRNA gene sequencing 5'- TCGT CGGC AGCG TCAG ATGTGTA TAAG AGAC AGCC TACG GGNG GCWGCAG- 3′ 5′- GTCT CGTG GGCT CGGA GATGTGTA TAAG AGAC AGGA CTAC HVGG GTAT CTAATCC- 3′ Sequence- based reagent bacterial 16 S v4 - rRNA gene reverse primer Thermo Fisher; Klindworth et al., 2013 16 S rRNA gene sequencing Sequence- based reagent mouse C1qa RNAscope probe (C1) Advanced Cell Diagnostics Cat# 498241 Sequence- based reagent mouse C1qa RNAscope probe (C3) Advanced Cell Diagnostics Cat# 498241- C3 Sequence- based reagent mouse Chat RNAscope probe (C1) Advanced Cell Diagnostics Cat# 408731 Sequence- based reagent mouse Nos1 RNAscope probe (C2) Advanced Cell Diagnostics Cat# 437651- C2 Sequence- based reagent mouse Adgrb1 RNAscope probe (C1) Advanced Cell Diagnostics Cat# 317901 Sequence- based reagent mouse Csf1r RNAscope probe (C2) Advanced Cell Diagnostics Cat# 428191- C2 Peptide, recombinant protein recombinant mouse C1q Complementech Cat# M099 Commercial assay or kit Chromium Next GEM Single Cell 3’ Kit v3.1 10 x Genomics Cat# PN- 1000269 Commercial assay or kit Chromiium Next GEM Chip G Single Cel Kit 10 x Genomics Cat# PN- 1000127 Commercial assay or kit Commercial assay or kit Dual Index Kit TT Set A 10 x Genomics Cat# PN- 1000215 FOXP3/Transcription Factor Fixation/Permeabilization Buffer Set Thermo Fisher Cat# 00- 5523- 00 Continued on next page Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 17 of 30 Immunology and Inflammation Research article Continued Reagent type (species) or resource Commercial assay or kit Designation Source or reference Identifiers Additional information MMLV Reverse Transcriptase Kit Thermo Fisher Cat# 28025–021 Commercial assay or kit NextSeq 500/550 High Output Kit v2.5 Illumina Cat# 20024907 Commercial assay or kit PE300 (Paired end 300 bp) v3 kit Illumina Cat# MS- 102–3001 commercial assay or kit RNAscope Fluorescent Multiple Reagent Kit Advanced Cell Diagnostics Cat# 320850 Commercial assay or kit Commercial assay or kit Commercial assay or kit RNeasy Universal Mini Kit Qiagen Cat# 73404 DNEasy Blood & Tissue Kit Qiagen Cat# 69504 TaqMan Master Mix Thermo Fisher Cat# 4369542 Commercial assay or kit TruSeq RNA sample preparation kit Illumina Commercial assay or kit SsoAdvanced Universal SYBR Green Supermix BioRad Cat# RS- 122–2001 Cat# 1725270 Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Agencourt AmpureXP beads Beckman Coulter Genomics Cat# A63880 Carmine Red Sigma Cat# C1022- 25G Collagenase IV Sigma Cat# C5138- 1G Borosilicate glass beads (2 mm) Millipore Sigma Cat# Z273627- 1EA Dextran sulfate sodium Thomas Scientific Cat# 216011090 DNase I Sigma Cat# DN25 Dispase II Sigma Cat# D4693- 1G FITC- dextran (4000 Da) Sigma Cat# FD4- 1g Ghost 710 Tonbo Biosciences Cat# 13–0871 T100 Flow cytometry viability dye Methylcellulose Sigma Cat# M0262- 100G Nalidixic acid, sodium salt Research Products International Cat# N23100- 25.0 Continued on next page Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 18 of 30 Immunology and Inflammation Research article Continued Reagent type (species) or resource Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Software, algorithm Software, algorithm Software, algorithm Software, algorithm Software, algorithm Software, algorithm Software, algorithm Software, algorithm Designation Source or reference Identifiers Additional information Optimal Cutting Temperature Compound (OCT) Thermo Fisher Cat# 23- 730- 571 Percoll Plus GE Healthcare Cat# GE17- 0891- 09 4% Paraformaldehyde Solution Thermo Fisher Cat# J19943.K2 Normal donkey serum Southern Biotech Cat# 0030–01 Triton X- 100 Thermo Fisher Cat# A16046.AP Protease inhibitors Millipore Sigma Cat# 11836153001 Rhodamine B- dextran Thermo Fisher Cat# D1841 Streptavidin- Cy5 Thermo Fisher Cat# 434316 Streptavidin- HRP conjugate Abcam Cat# ab7403 ELISA Sylgard 184 Silicone Elastomer Fisher Scientific Cat# 4019862 VECTASHIELD Antifade Mounting Medium with 4′,6- diamidino- 2- phenylindole (DAPI) Vector Labs Cat# H- 1200–10 Cell Ranger Single- Cell Software Suite 10 X Genomics clusterProfiler Yu et al., 2012 CLC Genomics Workbench Qiagen CLC Bio microbial genomics module Qiagen FlowJo ggplot2 BD Biosciences Love et al., 2015 GraphPad PRISM GraphPad Software Version 7.0; RRID:SCR_002798 Gut Analysis Toolbox Sorensen et al., 2022 https://digitalinsights.qiagen.com/plugins/clc- microbial-genomics-module/ Continued on next page Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 19 of 30 Immunology and Inflammation Research article Continued Reagent type (species) or resource Software, algorithm Software, algorithm Software, algorithm Software, algorithm Software, algorithm Software, algorithm Software, algorithm Other Other Other Other Other Other Other Other Other Designation Source or reference Identifiers Additional information Igor Pro 9 WaveMetrics Illumina Nextera Protocol Illumina Part # 15044223 Rev. B ImageJ Limma National Institutes of Health Ritchie et al., 2015 NovoExpress Agilent Technologies PVCAM software Teledyne Photometrics Seurat V3 R Package Stuart et al., 2019 https://imagej.nih.gov/ij/ Agilent 2100 Bioanalyzer Agilent Technologies G2939A RNA integrity analysis Amicon Ultra centrifugal filters Millipore Cat #UFC900324 Fecal protein extraction BioRad ChemiDoc Touch System BioRad Cat# 1708370 Western blot imaging: Chromium Controller & Next GEM Accessory Kit 10 X Genomics Cat# PN- 120223 Single cell RNA sequencing library construction CMOS camera Teledyne Photometrics MOMENT Ex vivo peristalsis: Leica CM1950 (Cryostat) Leica Cryosectioning FACSAria BD Biosciences Flow cytometric cell sorting ORCA- Fusion sCMOS camera Hamamatsu Photonics C14440- 20UP Imaging Illumina MiSeq Illumina RRID:SCR_016379 16 S rRNA Other Illumina NextSeq 550 Illumina Bulk RNA sequencing and single cell RNA sequencing Other Other Other Other Other Other Other Keyence Fluorescence Microscope Keyence BZ- X800 Immunofluorescence NovoCyte 3005 Agilent Technologies Flow cytometry analysis Organ bath chamber Peristaltic pump Tokai Hit Gilson Ex vivo peristalsis MINIPULS3 Ex vivo peristalsis QuantStudio 7 Flex Real- Time PCR System Applied Biosystems Cat #4485701 qPCR analysis SpectraMax M5 plate reader Molecular Devices ELISA and small intestinal motility analysis Zeiss Axio Imager M1 Microscope Zeiss Immunofluorescence Mice Wild- type C57BL/6 J (Jackson Laboratory) and C3-/- mice (Jackson Laboratory; Wessels et al., 1995) were bred and maintained in the SPF barrier facility at the University of Texas Southwestern Medical Center. C1qaΔMϕ mice were generated by crossing C1qafl/fl mice (Jackson Laboratory; Fonseca et al., 2017) with a mouse expressing Cre recombinase controlled by the macrophage- specific mouse Lyz2 promoter (LysM- Cre mice; Jackson Laboratory; Clausen et  al., 1999). Mice that were 8–12  weeks of age were used for all experiments and cohoused littermates were used as controls (i.e. Cre+ and Cre- mice were from the same breeding pair). Both male and female mice were analyzed in experi- ments involving wild- type mice. Males were used for experiments involving C1qafl/fl and C1qaΔMϕ mice. Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 20 of 30 Immunology and Inflammation Research article Germ- free C57BL/6 J mice were bred and maintained in isolators at the University of Texas South- western Medical Center. All procedures were performed in accordance with protocols approved by the Institutional Animal Care and Use Committees (IACUC) of the UT Southwestern Medical Center. Quantitative polymerase chain reaction (qPCR) Tissue RNA was isolated using the RNeasy Universal Mini kit (Qiagen, Hilden, Germany). Cellular RNA was isolated using the RNAqueous Micro kit (Thermo Fisher). cDNA was generated from the puri- fied RNA using the M- MLV Reverse Transcriptase kit (Thermo Fisher). qPCR analysis was performed using TaqMan primer/probe sets and master mix (Thermo Fisher) on a Quant- Studio 7 Flex Real- Time PCR System (Applied Biosystems). Transcript abundances were normalized to 18 S rRNA abundance. TaqMan probe assay IDs are provided in the Key Resources table. Isolation and analysis of intestinal immune cells Lamina propria cells were isolated from the intestine using a published protocol (Yu et al., 2013; Yu et al., 2014). Briefly, intestines were dissected from mice and Peyer’s patches were removed. Intes- tines were cut into small pieces and thoroughly washed with ice- cold phosphate- buffered saline (PBS) containing 5% fetal bovine serum (PBS- FBS). Epithelial cells were removed by incubating intestinal tissues in Hank’s buffered salt solution (HBSS) supplemented with 2 mM EDTA, followed by extensive washing with PBS- FBS. Residual tissues were digested twice with Collagenase IV (Sigma), DNase I (Sigma), and Dispase (BD Biosciences) for 45 min at 37 °C with agitation. Cells were filtered through 70 μm cell strainers (Thermo Fisher) and applied onto a 40%:80% Percoll gradient (GE Healthcare). Subepithelial cell populations were recovered at the interface of the 40% and 80% fractions. For small intestinal cell suspensions, the epithelial fraction was kept and combined with enzymatically liber- ated subepithelial cells. Cells were washed with 2 mM EDTA/3% FBS in PBS and Fc receptors were blocked with anti- CD16/32 (93). Cells were then stained with the viability dye Ghost 710 (Tonbo Biosci- ences) followed by antibodies against cell surface markers including anti- CD45 (30- F11), anti- CD11b (M1/70), anti- MHCII (M5/114.15.2), anti- F4/80 (BM8), anti- CD3 (17A2), anti- CD4 (GK1.5), anti- CD19 (1D3), anti- B220 (RA3- 6B2), anti- CD11c (N418), anti- CD169 (3D6.112), anti- TREM2 (237920), and anti- LY6C (RB6- 8C5). Cells were fixed and permeabilized with the eBioscience FOXP3/Transcription Factor Fixation/Permeabilization buffer set (Thermo Fisher) and then subjected to intracellular staining with anti- C1Q (RmC7H8), anti- FOXP3 (FJK- 16s), anti- GATA3 (L50), anti- T- BET (4B10), anti- RORγ (AFKJS- 9), and anti- ARG1 (P05089). Cells were sorted using a FACSAria (BD Biosciences) or analyzed using a NovoCyte 3005 (Agilent Technologies). Data were processed with FlowJo software (BD Biosciences) or NovoExpress (Agilent Technologies). Macrophage depletion Anti- mouse CSF1R (Thermo Fisher; AFS98) and rat IgG2a isotype control (Thermo Fisher; 2A3) anti- bodies were administered intraperitoneally at a concentration of 100  mg/kg. Mice were sacrificed 72 hr post- injection and terminal ileum and colon were collected for qPCR analysis. Protein extraction from intestinal cells and feces To isolate proteins from intestinal cell suspensions, cell pellets were resuspended in 100 μl of RIPA Lysis Buffer (Thermo Fisher) supplemented with protease inhibitors (Millipore Sigma) and vortexed vigorously every 5 min for 20 min. Lysates were cleared of cellular debris by centrifugation at 13,000 g for 5  min. To isolate proteins from the intestinal lumen, the entire gastrointestinal tract (from the duodenum to distal colon) was recovered from five wild- types C57BL/6 J mice. The intestines were flushed with  ~50  ml cold PBS containing protease inhibitors (Millipore Sigma, 11836153001). The flushes and fecal pellets were homogenized by rotor and stator (TH Tissue Homogenizer; OMNI; TH01) and large particles were centrifuged at 100 g for 10 min at room temperature. The superna- tants were carefully decanted and centrifuged further at 3000 g for 20 min at room temperature. The clarified supernatants were precipitated with 40% ammonium sulfate overnight at 4 °C. Precipitated protein was centrifuged at 3000 g for 30 min at 4 °C, then resuspended in cold 40% ammonium sulfate and centrifuged again. The pellets were resuspended in room temperature PBS and allowed to mix for 10 min. Protein concentrations were determined by Bradford assay (BioRad). Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 21 of 30 Immunology and Inflammation Research article Immunoblot 50  μg of fecal protein or 25  μg of cellular protein was loaded onto a 4–20%  gradient SDS- PAGE and transferred to a PVDF membrane. Membranes were blocked in 5% nonfat dry milk in Tris- buffered saline (TBS) with 0.1% Tween- 20 and then incubated overnight with the following primary antibodies: anti- C1Q (PA5- 29586, Thermo Fisher) and anti- actin (13E5, Cell Signaling). REG3G was detected by incubating membranes with rabbit anti- REG3G antiserum (Cash et  al., 2006). After washing, membranes were incubated with goat anti- rabbit IgG HRP and then visualized with a BioRad ChemiDoc Touch system. Enzyme-linked immunosorbent assay (ELISA) Mouse C1q ELISA was performed as previously described (Petry et  al., 2001). Briefly, microtiter plates were coated overnight with mouse IgG1 and were then blocked with 5% BSA in PBS. Serum samples were diluted 1:50 and plated for 1  hr at room temperature. After washing with 0.05% Tween- 20 in PBS, bound C1q was incubated with a biotinylated anti- C1q antibody (JL1, Abcam). Biotinylated anti- C1q was detected with a streptavidin- HRP conjugate (Abcam). Optical density was measured using a wavelength of 492 nm. Plates were analyzed using a SpectraMax M5 microplate reader (Molecular Devices). Intestinal permeability assay Intestinal permeability assays were performed by treating mice with fluorescein isothiocyanate dextran (FITC- dextran; 4000 Da) by oral gavage. The non- steroidal anti- inflammatory drug (NSAID) indometh- acin was administered to mice as a positive control. For the experimental group, mice were treated with 190 μl 7% dimethyl sulfoxide (DMSO) in PBS by oral gavage. For the positive control group, mice were treated with 190 μl indomethacin (1.5 mg/ml in 7% DMSO in PBS) by oral gavage. After 1 hr, all mice were treated with 190 μl FITC- dextran (80 mg/ml in PBS) by oral gavage. Mice were sacrificed after 4 hr and sera were collected. Serum samples were centrifuged for 20 min at 4 °C at 800 g and supernatants were collected. Serum FITC- dextran levels were measured by a fluorescence microplate assay against a standard curve using a Spectramax plate reader (Molecular Devices). 16S rRNA gene quantification (absolute copy number) Age and sex- matched mice were sacrificed and mesenteric lymph nodes were harvested and weighed. Total DNA was extracted using the Qiagen DNEasy kit. Microbial genomic DNA was quantified against a standard curve by qPCR analysis using universal 16S rRNA gene primers and the SsoAdvanced SYBR Green Supermix (BioRad). Total copy numbers of bacterial 16S RNA genes were normalized to tissue weight. Dextran sulfate sodium (DSS) treatment Age and sex- matched mice were provided with 3% dextran sulfate sodium (weight/volume) in auto- claved drinking water for seven days. Animal weight and health were monitored in accordance with institutional IACUC guidelines. On day 7, animals were sacrificed and colon lengths were recorded. Terminal colon segments were fixed in Bouin’s fixative for 24 hr followed by washes in 70% ethanol. Tissues were paraffin- embedded and sectioned by the UT Southwestern Histopathology Core facility. Tissue specimens were scored by a pathologist who was blinded as to the mouse genotypes. Disease severity was scored using five different parameters on a scale of 0–4: inflammation severity, edema severity, epithelial cell loss severity, hyperplasia, and fibrosis. Scores for each individual parameter were added together to represent the overall histology score. Salmonella typhimurium infection To prepare bacteria for infection, Salmonella enterica serovar typhimurium (SL1344) was cultured in Luria- Bertani (LB) broth containing 50 μg/ml streptomycin in a shaking incubator at 37 °C (Eichelberg and Galán, 1999). The overnight culture was diluted the next day and grown to the mid- log phase (OD600 = 0.3–0.5). C1qafl/fl and C1qaΔMϕ littermates were inoculated intragastrically with 109 CFU. All mice were sacrificed 24 hr post- infection and small intestinal tissues were harvested for analysis. Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 22 of 30 Immunology and Inflammation Research article Citrobacter rodentium infection To prepare bacteria for infection, an overnight culture of C. rodentium (DBS100, ATCC) was grown in LB broth containing nalidixic acid (100 μg/ml) in a shaking incubator at 37 °C. The culture was diluted the next day and grown to the mid- log phase (OD600 = 0.4–0.6). Bacteria were pelleted, washed, and resuspended in PBS. Sex- matched littermates were inoculated intragastrically with 5 × 108 CFU. Fecal pellets were collected at a fixed time every 48 hr, homogenized in sterile PBS, diluted, and plated on LB agar with nalidixic acid (100 μg/ml). Immunofluorescence analysis of mouse intestines Mouse small intestines and colons were flushed with PBS and embedded with Optimal Cutting Temperature compound (OCT) (Thermo Fisher). Sections were fixed in ice- cold acetone, blocked with 1% BSA, 10% FBS, 1% Triton X- 100 in PBS, and then incubated overnight at 4 °C with the following antibodies: mouse anti- C1q biotin (JL- 1), rat anti- CD169 (3D6.112), and rabbit anti- TUBB3 (ab18207, Abcam). Slides were then washed with PBS containing 0.2% Tween- 20 (PBS- T) and incubated with donkey anti- rabbit AlexaFluor 488, donkey anti- rat AlexaFluor 594, and Streptavidin- Cy5 (Thermo Fisher) for 1 hr at room temperature in the dark. Slides were then washed in PBS- T and mounted with DAPI- Fluoromount- G (Southern Biotech). Mounted slides were cured overnight at 4 °C until imaging. For immunofluorescence analysis of longitudinal muscle- myenteric plexus wholemounts, intes- tines were prepared by first removing the adipose tissues and flushing the luminal contents. A 1 ml pipette tip was inserted into the intestinal lumen to fully extend the intestinal wall. The longitudinal muscle- myenteric plexus layer was then separated from the mucosa using cotton swabs as previously described (Ahrends et al., 2022; Obata et al., 2020). The longitudinal muscle- myenteric plexus layer was then stretched by pinning the tissues on a Sylgard- coated Petri dish (Fisher Scientific) containing cold PBS and fixed with 4% PFA overnight at 4 °C. The fixed tissues were rinsed five times with PBS at room temperature with shaking and then permeabilized and blocked with PBS containing 1% Triton X- 100 and 10% normal donkey serum (NDS) for 1 hr at room temperature. The tissues were incubated with primary antibodies in the same solution overnight at 4 °C. The tissues were then washed with PBS containing 1% Triton X- 100 and incubated with secondary antibodies in the blocking buffer for 2 hr at room temperature. Immunostained tissues were washed four times with PBS containing 1% Triton X- 100. After a final wash with PBS, tissues were mounted on Superfrost Microscope Slides using VECTASHIELD (Vector Laboratories). RNAscope analysis Fluorescence in situ hybridization on the longitudinal muscle- myenteric plexus was carried out using the Advanced Cell Diagnostics RNAscope Fluorescent Multiplex Kit according to the manufacturer’s instructions with some modifications as described previously (Obata et al., 2020; Obata et al., 2022). After hybridization, tissues were counterstained for neuronal nuclei as previously described and mounted on Superfrost Microscope Slides (Fisher Scientific) using VECTASHIELD (Vector Laboratories). Image processing Fluorescently labeled longitudinal muscle- myenteric plexus preparations were imaged by a spinning disk confocal microscope (Nikon) with a Hamamatsu Orca- Fusion sCMOS camera using the NIS- Elements Advanced Research software (Nikon). All image analyses were performed using the image- processing package Fiji and ImageJ. The number of HuC/D+ neurons in the myenteric plexus was quantified using a semi- automated image analysis pipeline Gut Analysis Toolbox (Sorensen et  al., 2022). RNA-seq analysis of colonic longitudinal muscle-myenteric plexus The colonic longitudinal muscle- myenteric plexus was collected from five age- matched male C1qafl/fl and C1qaΔMϕ mice by manual dissection using a 2 mm metal probe (Fisher Scientific). RNA was isolated using the RNeasy Mini kit according to the manufacturer’s protocol (Qiagen). Quantity and quality of RNA samples were assessed on a Bioanalyzer 2100 (Agilent Technologies). RNA- seq libraries were prepared using the TruSeq RNA sample preparation kit (Illumina) according to the manufacturer’s protocol. Libraries were validated on a Bioanalyzer 2100 (Agilent Technologies). Indexed libraries were sequenced on an Illumina NextSeq550 for single- end 75 bp length reads. CLC Genomics Workbench Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 23 of 30 Immunology and Inflammation Research article 7 was used for bioinformatics and statistical analysis of the sequencing data. The approach used by CLC Genomics Workbench is based on a method developed previously (Mortazavi et al., 2008). To identify differentially enriched biological pathways, all genes were ranked based on their log2fold- change, and pathway enrichment was identified using the R packages ‘clusterProfiler’ and ‘msigdbr.’ For analysis of differentially expressed genes, gene counts were analyzed using DESeq- 2, and differ- entially expressed genes were defined as having an adjusted p- value < 0.05. A Fisher’s Exact Test was conducted to assess the overlap between differentially expressed genes in C1qaΔMϕ mice and the TashT mouse (Bergeron et al., 2015). Single-cell RNA sequencing (scRNAseq) analysis Single- cell RNA sequencing was done in the Microbiome Research Laboratory at UT Southwestern Medical Center. Lamina propria cell suspensions were prepared as previously described (Yu et al., 2013; Yu et  al., 2014) from the small intestines of three C1qafl/fl and three C1qaΔMϕ littermates. Total small intestinal cells were pooled according to genotype and live CD45+CD11b+MHCI- I+F4/80hi macrophages were sorted using a FACSAria (BD Biosciences). 5000–10,000 macrophages from each genotype with a viability score of >70% were input into each library. A 10 X Genomics Chromium controller instrument was used for Gel Bead- in Emulsion (GEMs) preparation. Chro- mium Next GEM Single Cell 3' Kit v3.1 (PC- 1000269), Chromium Next GEM Chip G Single Cell Kit (PC- 1000127), and Dual Index Kit TT Set A Kit (PC- 1000215) were used for single- cell library prepa- ration. cDNA and final barcoded sequencing libraries were generated according to the manufac- turer’s specifications and their quality and concentration were assessed using a Bioanalyzer 2100 (Agilent Technologies) and qPCR, respectively. Single- cell libraries that passed the quality checks were sequenced on a NextSeq550 sequencer using a paired- end 75 bp High Output sequencing kit. About 20,000–30,000 sequencing reads were generated per single cell. Unique molecular identifier (UMI) counts for each cellular barcode were quantified and used to estimate the number of cells successfully captured and sequenced. The Cell Ranger Single- Cell Software suite (10  X Genomics) was used for demultiplexing, barcode processing, alignment, and initial clustering of the raw scRNAseq profiles. The Seurat V3 R package was used to filter and analyze the Cell Ranger output (Stuart et  al., 2019). Features that were in less than three cells and cells with less than 50 features were first filtered. To filter out dead or dying single cells, only cells that expressed more than 200 but less than 2500 features and cells in which mitochondrial transcripts accounted for less than five percent of all cell transcripts were used for further analysis. The single- cell data of these high- quality cells was then log- normalized and scaled. For further correction, the percentage of transcripts from mitochondria was regressed out. Dimension reduction was performed in Seurat and further differential gene expression was performed using limma (Ritchie et al., 2015). Pathway enrichment analysis was performed with Gene Set Enrichment Analysis (GSEA) via clusterProfiler (Yu et al., 2012). Visual representations of data were made using ggplot2 and Seurat R packages (Love et al., 2015). 16S rRNA gene sequencing and analysis The hypervariable regions V3 and V4 of the bacterial 16S rRNA gene were prepared using the Illu- mina Nextera protocol (Part # 15044223 Rev. B). An amplicon of 460 bp was amplified using the 16S Forward Primer and 16S Reverse Primer as described in the manufacturer’s protocol. Primer sequences are given in the Key Resources Table. The PCR product was purified using Agencourt AmpureXP beads (Beckman Coulter Genomics). Illumina adapter and barcode sequences were ligated to the amplicon to attach them to the MiSeqDx flow cell and for multiplexing. Quality and quantity of each sequencing library were assessed using Bioanalyzer (Agilent Technologies) and Picogreen (Thermo Fisher) measurements, respectively. Libraries were loaded onto a MiSeqDX flow cell and sequenced using the Paired End 300 (PE300) v3 kit. Raw fastq files were demultiplexed based on unique barcodes and assessed for quality. Samples with more than 50,000 quality control pass sequencing reads were used for downstream analysis. Taxonomic classification and operational taxonomic unit analysis were done using the CLC Microbial Genomics Module. Individual sample reads were annotated with the Greengene database and taxonomic features were assessed. Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 24 of 30 Immunology and Inflammation Research article Gastrointestinal motility assays Motility assays were adapted from previous studies (Luo et al., 2018; Maurer, 2016; Muller et al., 2014). To determine transit time through the entire gastrointestinal tract, age- matched male mice were fasted overnight and water was removed 1 hr prior to the start of the experiment. Mice were then singly housed for 1 hr and then gavaged with 100 μl of Carmine Red (5% weight/volume; Sigma) in 1.5% methylcellulose. Fecal pellets were collected every 15 min and transit time was recorded when the dye was first observed in the feces. For small intestinal motility measurements, age- matched male mice were fasted overnight and then gavaged with 100 μl of rhodamine B- dextran (5 mg/ml; Thermo Fisher) in 2% methylcellulose. After 90 min, mice were sacrificed and their stomachs, small intestines, ceca, and colons were collected. Small intestines were cut into eight segments of equal length and colons were cut into five segments of equal length. Segments were cut open lengthwise and vortexed in 1 ml PBS to release rhodamine B- dextran. Fluorescence was then measured on a SpectraMax M5 microplate reader (Molecular Devices). The geometric center of the dye was calculated as: GC = Σ (% of total fluorescent signal per segment × segment number). Relative fluorescence per segment was calculated as: (fluorescence signal in segment/total fluorescence recovered) × 100. To measure colonic motility, age- matched male mice were fasted overnight and lightly anesthe- tized with isoflurane. A 2 mm glass bead was inserted 2 cm intrarectally using a 2 mm surgical probe. Mice were then returned to empty cages and the time to reappearance of the bead was recorded. To account for potential circadian differences in gut motility, the time of day for the initiation of all experiments was held constant. Ex vivo peristaltic imaging Ex vivo video imaging and analysis of colonic peristalsis were carried out as described previously (Obata et al., 2020) on age- matched male mice. Colons were dissected, flushed with sterile PBS, and pinned into an organ bath chamber (Tokai Hit, Japan) filled with Dulbecco’s Modified Eagle Medium (DMEM). DMEM was oxygenated (95% O2 and 5% CO2), run through the chamber using a peristaltic pump (MINIPULS 3, Gilson), and kept at 37  °C. Colons were allowed to equilibrate to the organ chamber for 20 min before video recording. Time- lapse images of colonic peristalsis were captured with a camera (MOMENT, Teledyne photometrics) using PVCAM software (500 ms time- lapse delay) and recorded for 45 min. For analysis of colonic migrating motor complexes (CMMC), videos consisting of 5400 sequen- tial image frames were stitched together in Fiji and read into Igor Pro 9 (WaveMetrics) to generate spatiotemporal maps using a customized algorithm developed by the Pieter Vanden Berghe lab at the University of Leuven, Belgium (Roosen et al., 2012). The generated spatiotemporal maps were used to determine the frequency and period of CMMCs. Each CMMC on the spatiotemporal map was further projected onto the axes to obtain the distance traveled (millimeters) and the time for the CMMC to travel such distance (seconds), allowing us to calculate the velocity (millimeter/second) of CMMCs. Statistical analysis Graphed data are presented as means ± standard error of the mean (SEM). Statistics were determined with GraphPad Prism software. Statistical analyses were performed using a two- tailed Student’s t- test when comparing two groups, oneway ANOVA when comparing multiple groups, and Fisher’s exact test to assess overlap between groups of differentially expressed genes. The statistical tests used are indicated in the figure legends. *p<0.05; **p<0.01; ***p<0.001; ****p<0.0001; and ns, not significant (p>0.05). Acknowledgements We thank Shai Bel for assistance with immunofluorescence imaging experiments, the UT South- western Genomics Core for assistance with RNA sequencing experiments, the UT Southwestern Flow Cytometry Core for assistance with flow cytometry experiments, Bret Evers (UT Southwestern Histo Pathology Core) for pathology scoring, and the Quantitative Light Microscopy Core (QLMC), a Shared Resource of the Harold C Simmons Cancer Center. The QLMC is supported in part by the National Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 25 of 30 Immunology and Inflammation Research article Cancer Institute Cancer Center Support Grant P30 CA142543- 01 and NIH 1S10OD028630- 01. Citro- bacter rodentium strain DBS100 was a gift from Vanessa Sperandio (UT Southwestern). The laboratory of Pieter Vanden Berghe (University of Leuven, Belgium) provided the algorithm used to generate spatiotemporal maps of colonic migrating motor complexes. This work was supported by NIH grants R01 DK070855 (LVH), Welch Foundation Grant I- 1874 (LVH), the Walter M and Helen D Bader Center for Research on Arthritis and Autoimmune Diseases (LVH), and the Howard Hughes Medical Institute (LVH). MP was supported by NIH T32 AI005284. AAC was supported by NIH T32 AI005284 and NIH F32 DK132913. EK was supported by NIH F31 DK126391. YO is the Nancy Cain Marcus and Jeffrey A Marcus Scholar in Medical Research, in Honor of Dr. Bill S Vowell. Additional information Funding Funder National Institutes of Health Grant reference number Author R01 DK070855 Lora V Hooper Welch Foundation I-1874 Howard Hughes Medical Institute Lora V Hooper Lora V Hooper National Institutes of Health National Institutes of Health National Institutes of Health T32 AI005284 Mihir Pendse F32 DK132913 Alexander A Crofts F31 DK126391 Eugene Koo The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Author contributions Mihir Pendse, Conceptualization, Data curation, Formal analysis, Supervision, Investigation, Method- ology, Writing – original draft, Writing – review and editing; Haley De Selle, Nguyen Vo, Data cura- tion, Formal analysis, Investigation, Methodology; Gabriella Quinn, Alexander A Crofts, Data curation, Formal analysis; Chaitanya Dende, Daniel C Propheter, Investigation, Writing – review and editing; Yun Li, Cristine N Salinas, Tarun Srinivasan, Brian Hassell, Kelly A Ruhn, Investigation; Eugene Koo, Investigation, Methodology; Prithvi Raj, Data curation, Formal analysis, Investigation; Yuuki Obata, Investigation, Methodology, Writing – original draft, Writing – review and editing; Lora V Hooper, Conceptualization, Supervision, Funding acquisition, Writing – original draft, Project administration, Writing – review and editing Author ORCIDs Mihir Pendse Alexander A Crofts Yuuki Obata Lora V Hooper http://orcid.org/0000-0002-7810-6791 http://orcid.org/0000-0003-0811-9199 http://orcid.org/0000-0001-5461-3521 http://orcid.org/0000-0002-2759-4641 Ethics This study was performed in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. All of the animals were handled according to approved institutional animal care and use committee (IACUC) protocols (protocol #2015- 101212) of the University of Texas Southwestern Medical Center. Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.78558.sa1 Author response https://doi.org/10.7554/eLife.78558.sa2 Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 26 of 30 Immunology and Inflammation Research article Additional files Supplementary files •  MDAR checklist Data availability 16S rRNA gene sequencing data (Figure 3D) and RNA sequencing data (Figure 6A and B; Figure 1— figure supplement 1; Figure 6—figure supplement 1) are available from the Sequence Read Archive under BioProject ID PRJNA793870. All mouse strains used are available commercially. The following dataset was generated: Author(s) Pendse M, Raj P, Hooper LV Year 2022 Dataset title Dataset URL Database and Identifier Macrophages control gastrointestinal motility through complement component 1q https://www. ncbi. nlm. nih. gov/ bioproject/ PRJNA793870/ NCBI BioProject, PRJNA793870 The following previously published dataset was used: Year 2019 Author(s) Gattu S, Bang Y, Chara A, Harris T, Kuang Z, Ruhn K, Sockanathan S, Hooper LV Dataset title Dataset URL Database and Identifier Epithelial retinoic acid receptor beta regulates serum amyloid A expression and vitamin A- dependent intestinal immunity https://www. ncbi. nlm. nih. gov/ geo/ query/ acc. cgi? acc= GSE122471 NCBI Gene Expression Omnibus, GSE122471 References Ahrends T, Weiner M, Mucida D. 2022. Isolation of myenteric and submucosal plexus from mouse gastrointestinal tract and subsequent flow cytometry and immunofluorescence. STAR Protocols 3:101157. DOI: https://doi.org/10.1016/j.xpro.2022.101157, PMID: 35146454 Akiho H, Ihara E, Motomura Y, Nakamura K. 2011. Cytokine- Induced alterations of gastrointestinal motility in gastrointestinal disorders. World Journal of Gastrointestinal Pathophysiology 2:72–81. DOI: https://doi.org/10. 4291/wjgp.v2.i5.72, PMID: 22013552 Bassotti G, Antonelli E, Villanacci V, Salemme M, Coppola M, Annese V. 2014. Gastrointestinal motility disorders in inflammatory bowel diseases. World Journal of Gastroenterology 20:37. DOI: https://doi.org/10.3748/wjg. v20.i1.37 Belzer C, Liu Q, Carroll MC, Bry L. 2011. The role of specific IgG and complement in combating aprimary mucosal infection of the gut epithelium. European Journal of Microbiology and Immunology 1:311–318. DOI: https://doi.org/10.1556/EuJMI.1.2011.4.7 Benavente F, Piltti KM, Hooshmand MJ, Nava AA, Lakatos A, Feld BG, Creasman D, Gershon PD, Anderson A. 2020. Novel C1q receptor- mediated signaling controls neural stem cell behavior and neurorepair. eLife 9:e55732. DOI: https://doi.org/10.7554/eLife.55732, PMID: 32894219 Benoit ME, Tenner AJ. 2011. Complement protein C1q- mediated neuroprotection is correlated with regulation of neuronal gene and microRNA expression. The Journal of Neuroscience 31:3459–3469. DOI: https://doi.org/ 10.1523/JNEUROSCI.3932-10.2011, PMID: 21368058 Bergeron KF, Cardinal T, Touré AM, Béland M, Raiwet DL, Silversides DW, Pilon N. 2015. Male- Biased aganglionic megacolon in the tasht mouse line due to perturbation of silencer elements in a large gene desert of chromosome 10. PLOS Genetics 11:e1005093. DOI: https://doi.org/10.1371/journal.pgen.1005093, PMID: 25786024 Berthoud HR, Blackshaw LA, Brookes SJH, Grundy D. 2004. Neuroanatomy of extrinsic afferents supplying the gastrointestinal tract. Neurogastroenterology and Motility 16 Suppl 1:28–33. DOI: https://doi.org/10.1111/j. 1743-3150.2004.00471.x, PMID: 15066001 Bogunovic M, Ginhoux F, Helft J, Shang L, Hashimoto D, Greter M, Liu K, Jakubzick C, Ingersoll MA, Leboeuf M, Stanley ER, Nussenzweig M, Lira SA, Randolph GJ, Merad M. 2009. Origin of the lamina propria dendritic cell network. Immunity 31:513–525. DOI: https://doi.org/10.1016/j.immuni.2009.08.010, PMID: 19733489 Bossi F, Tripodo C, Rizzi L, Bulla R, Agostinis C, Guarnotta C, Munaut C, Baldassarre G, Papa G, Zorzet S, Ghebrehiwet B, Ling GS, Botto M, Tedesco F. 2014. C1Q as a unique player in angiogenesis with therapeutic implication in wound healing. PNAS 111:4209–4214. DOI: https://doi.org/10.1073/pnas.1311968111, PMID: 24591625 Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 27 of 30 Immunology and Inflammation Research article Botto M, Dell’Agnola C, Bygrave AE, Thompson EM, Cook HT, Petry F, Loos M, Pandolfi PP, Walport MJ. 1998. Homozygous C1q deficiency causes glomerulonephritis associated with multiple apoptotic bodies. Nature Genetics 19:56–59. DOI: https://doi.org/10.1038/ng0598-56, PMID: 9590289 Casals C, García- Fojeda B, Minutti CM. 2019. Soluble defense collagens: sweeping up immune threats. Molecular Immunology 112:291–304. DOI: https://doi.org/10.1016/j.molimm.2019.06.007, PMID: 31228661 Cash HL, Whitham CV, Behrendt CL, Hooper LV. 2006. Symbiotic bacteria direct expression of an intestinal bactericidal lectin. Science 313:1126–1130. DOI: https://doi.org/10.1126/science.1127119, PMID: 16931762 Chu Y, Jin X, Parada I, Pesic A, Stevens B, Barres B, Prince DA. 2010. Enhanced synaptic connectivity and epilepsy in C1q knockout mice. PNAS 107:7975–7980. DOI: https://doi.org/10.1073/pnas.0913449107, PMID: 20375278 Clausen BE, Burkhardt C, Reith W, Renkawitz R, Förster I. 1999. Conditional gene targeting in macrophages and granulocytes using lysmcre mice. Transgenic Research 8:265–277. DOI: https://doi.org/10.1023/a: 1008942828960, PMID: 10621974 Colonna M. 2003. TREMs in the immune system and beyond. Nature Reviews. Immunology 3:445–453. DOI: https://doi.org/10.1038/nri1106, PMID: 12776204 Cui H, Banerjee S, Guo S, Xie N, Ge J, Jiang D, Zörnig M, Thannickal VJ, Liu G. 2019. Long noncoding RNA MALAT1 regulates differential activation of macrophages and response to lung injury. JCI Insight 4:e124522. DOI: https://doi.org/10.1172/jci.insight.124522, PMID: 30676324 Davis CA, Vallota EH, Forristal J. 1979. Serum complement levels in infancy: age related changes. Pediatric Research 13:1043–1046. DOI: https://doi.org/10.1203/00006450-197909000-00019, PMID: 503656 De Schepper S, Stakenborg N, Matteoli G, Verheijden S, Boeckxstaens GE. 2018a. Muscularis macrophages: key players in intestinal homeostasis and disease. Cellular Immunology 330:142–150. DOI: https://doi.org/10.1016/ j.cellimm.2017.12.009, PMID: 29291892 De Schepper S, Verheijden S, Aguilera- Lizarraga J, Viola MF, Boesmans W, Stakenborg N, Voytyuk I, Schmidt I, Boeckx B, Dierckx de Casterlé I, Baekelandt V, Gonzalez Dominguez E, Mack M, Depoortere I, De Strooper B, Sprangers B, Himmelreich U, Soenen S, Guilliams M, Vanden Berghe P, et al. 2018b. Self- maintaining gut macrophages are essential for intestinal homeostasis. Cell 175:400–415.. DOI: https://doi.org/10.1016/j.cell. 2018.07.048, PMID: 30173915 Domanska D, Majid U, Karlsen VT, Merok MA, Beitnes A- CR, Yaqub S, Bækkevold ES, Jahnsen FL. 2022. Single- Cell transcriptomic analysis of human colonic macrophages reveals niche- specific subsets. The Journal of Experimental Medicine 219:e20211846. DOI: https://doi.org/10.1084/jem.20211846, PMID: 35139155 Earley AM, Graves CL, Shiau CE. 2018. Critical role for a subset of intestinal macrophages in shaping gut microbiota in adult zebrafish. Cell Reports 25:424–436. DOI: https://doi.org/10.1016/j.celrep.2018.09.025, PMID: 30304682 Eichelberg K, Galán JE. 1999. Differential regulation of Salmonella typhimurium type III secreted proteins by pathogenicity island 1 (Spi- 1) -encoded transcriptional activators invf and hila. Infection and Immunity 67:4099– 4105. DOI: https://doi.org/10.1128/IAI.67.8.4099-4105.1999, PMID: 10417179 El- Shamy A, Branch AD, Schiano TD, Gorevic PD. 2018. The complement system and C1q in chronic hepatitis C virus infection and mixed cryoglobulinemia. Frontiers in Immunology 9:1001. DOI: https://doi.org/10.3389/ fimmu.2018.01001, PMID: 29910796 Fawzy M, Edrees A, Okasha H, El Ashmaui A, Ragab G. 2016. Gastrointestinal manifestations in systemic lupus erythematosus. Lupus 25:1456–1462. DOI: https://doi.org/10.1177/0961203316642308, PMID: 27055518 Fonseca MI, Chu SH, Hernandez MX, Fang MJ, Modarresi L, Selvan P, MacGregor GR, Tenner AJ. 2017. Cell- Specific deletion of C1qA identifies microglia as the dominant source of C1q in mouse brain. Journal of Neuroinflammation 14:48. DOI: https://doi.org/10.1186/s12974-017-0814-9, PMID: 28264694 Gabanyi I, Muller PA, Feighery L, Oliveira TY, Costa- Pinto FA, Mucida D. 2016. Neuro- Immune interactions drive tissue programming in intestinal macrophages. Cell 164:378–391. DOI: https://doi.org/10.1016/j.cell.2015.12. 023, PMID: 26777404 Gao Y, Fang P, Li WJ, Zhang J, Wang GP, Jiang DF, Chen FP. 2020. Lncrna NEAT1 sponges miR- 214 to regulate M2 macrophage polarization by regulation of B7- H3 in multiple myeloma. Molecular Immunology 117:20–28. DOI: https://doi.org/10.1016/j.molimm.2019.10.026, PMID: 31731055 Gattu S, Bang YJ, Pendse M, Dende C, Chara AL, Harris TA, Wang Y, Ruhn KA, Kuang Z, Sockanathan S, Hooper LV. 2019. Epithelial retinoic acid receptor β regulates serum amyloid A expression and vitamin A- dependent intestinal immunity. PNAS 116:10911–10916. DOI: https://doi.org/10.1073/pnas.1812069116, PMID: 31097581 Grainger JR, Konkel JE, Zangerle- Murray T, Shaw TN. 2017. Macrophages in gastrointestinal homeostasis and inflammation. Pflugers Archiv 469:527–539. DOI: https://doi.org/10.1007/s00424-017-1958-2, PMID: 28283748 Hammond JW, Bellizzi MJ, Ware C, Qiu WQ, Saminathan P, Li H, Luo S, Ma SA, Li Y, Gelbard HA. 2020. Complement- Dependent synapse loss and microgliosis in a mouse model of multiple sclerosis. Brain, Behavior, and Immunity 87:739–750. DOI: https://doi.org/10.1016/j.bbi.2020.03.004, PMID: 32151684 Hong S, Beja- Glasser VF, Nfonoyim BM, Frouin A, Li S, Ramakrishnan S, Merry KM, Shi Q, Rosenthal A, Barres BA, Lemere CA, Selkoe DJ, Stevens B. 2016. Complement and microglia mediate early synapse loss in Alzheimer mouse models. Science 352:712–716. DOI: https://doi.org/10.1126/science.aad8373, PMID: 27033548 Johnson CD, Barlow- Anacker AJ, Pierre JF, Touw K, Erickson CS, Furness JB, Epstein ML, Gosain A. 2018. Deletion of choline acetyltransferase in enteric neurons results in postnatal intestinal dysmotility and dysbiosis. FASEB Journal 32:4744–4752. DOI: https://doi.org/10.1096/fj.201701474RR, PMID: 29570391 Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 28 of 30 Immunology and Inflammation Research article Kishore U, Reid KB. 2000. C1Q: Structure, function, and receptors. Immunopharmacology 49:159–170. DOI: https://doi.org/10.1016/s0162-3109(00)80301-x, PMID: 10904115 Klindworth A, Pruesse E, Schweer T, Peplies J, Quast C, Horn M, Glöckner FO. 2013. Evaluation of general 16S ribosomal RNA gene PCR primers for classical and next- generation sequencing- based diversity studies. Nucleic Acids Research 41:e1. DOI: https://doi.org/10.1093/nar/gks808, PMID: 22933715 Korb LC, Ahearn JM. 1997. C1Q binds directly and specifically to surface blebs of apoptotic human keratinocytes: complement deficiency and systemic lupus erythematosus revisited. Journal of Immunology 158:4525–4528 PMID: 9144462. Kouser L, Madhukaran SP, Shastri A, Saraon A, Ferluga J, Al- Mozaini M, Kishore U. 2015. Emerging and novel functions of complement protein C1q. Frontiers in Immunology 6:317. DOI: https://doi.org/10.3389/fimmu. 2015.00317, PMID: 26175731 Kulkarni S, Micci MA, Leser J, Shin C, Tang SC, Fu YY, Liu L, Li Q, Saha M, Li C, Enikolopov G, Becker L, Rakhilin N, Anderson M, Shen X, Dong X, Butte MJ, Song H, Southard- Smith EM, Kapur RP, et al. 2017. Adult enteric nervous system in health is maintained by a dynamic balance between neuronal apoptosis and neurogenesis. PNAS 114:E3709–E3718. DOI: https://doi.org/10.1073/pnas.1619406114, PMID: 28420791 Love MI, Anders S, Kim V, Huber W. 2015. Rna- Seq workflow: gene- level exploratory analysis and differential expression. F1000Research 4:1070. DOI: https://doi.org/10.12688/f1000research.7035.1, PMID: 26674615 Luo J, Qian A, Oetjen LK, Yu W, Yang P, Feng J, Xie Z, Liu S, Yin S, Dryn D, Cheng J, Riehl TE, Zholos AV, Stenson WF, Kim BS, Hu H. 2018. Trpv4 channel signaling in macrophages promotes gastrointestinal motility via direct effects on smooth muscle cells. Immunity 49:107–119. DOI: https://doi.org/10.1016/j.immuni.2018. 04.021, PMID: 29958798 Macedo ACL, Isaac L. 2016. Systemic lupus erythematosus and deficiencies of early components of the complement classical pathway. Frontiers in Immunology 7:55. DOI: https://doi.org/10.3389/fimmu.2016.00055, PMID: 26941740 Mao YL, Shen CL, Zhou T, Ma BT, Tang LY, Wu WT, Zhang HX, Lu HL, Xu WX, Wang ZG. 2017. Ablation of tacr2 in mice leads to gastric emptying disturbance. Neurogastroenterology and Motility 29:e13117. DOI: https:// doi.org/10.1111/nmo.13117, PMID: 28585346 Matheis F, Muller PA, Graves CL, Gabanyi I, Kerner ZJ, Costa- Borges D, Ahrends T, Rosenstiel P, Mucida D. 2020. Adrenergic signaling in muscularis macrophages limits infection- induced neuronal loss. Cell 180:64–78. DOI: https://doi.org/10.1016/j.cell.2019.12.002, PMID: 31923400 Maurer AH. 2016. Gastrointestinal motility, part 2: small- bowel and colon transit. Journal of Nuclear Medicine Technology 44:12–18. DOI: https://doi.org/10.2967/jnumed.113.134551, PMID: 26940448 Mortazavi A, Williams BA, McCue K, Schaeffer L, Wold B. 2008. Mapping and quantifying mammalian transcriptomes by RNA- seq. Nature Methods 5:621–628. DOI: https://doi.org/10.1038/nmeth.1226 Muller PA, Koscsó B, Rajani GM, Stevanovic K, Berres ML, Hashimoto D, Mortha A, Leboeuf M, Li XM, Mucida D, Stanley ER, Dahan S, Margolis KG, Gershon MD, Merad M, Bogunovic M. 2014. Crosstalk between muscularis macrophages and enteric neurons regulates gastrointestinal motility. Cell 158:1210. DOI: https://doi.org/10. 1016/j.cell.2014.08.002, PMID: 28917294 Noris M, Remuzzi G. 2013. Overview of complement activation and regulation. Seminars in Nephrology 33:479–492. DOI: https://doi.org/10.1016/j.semnephrol.2013.08.001, PMID: 24161035 Obata Y, Castaño Á, Boeing S, Bon- Frauches AC, Fung C, Fallesen T, de Agüero MG, Yilmaz B, Lopes R, Huseynova A, Horswell S, Maradana MR, Boesmans W, Vanden Berghe P, Murray AJ, Stockinger B, Macpherson AJ, Pachnis V. 2020. Neuronal programming by microbiota regulates intestinal physiology. Nature 578:284–289. DOI: https://doi.org/10.1038/s41586-020-1975-8, PMID: 32025031 Obata Y, Castaño Á, Fallesen TL, Bon- Frauches AC, Boeing S, Huseynova A, McCallum S, Lasrado R, Heanue TA, Pachnis V. 2022. Molecular profiling of enteric nervous system cell lineages. Nature Protocols 17:1789–1817. DOI: https://doi.org/10.1038/s41596-022-00697-4, PMID: 35676375 Paloneva J, Manninen T, Christman G, Hovanes K, Mandelin J, Adolfsson R, Bianchin M, Bird T, Miranda R, Salmaggi A, Tranebjaerg L, Konttinen Y, Peltonen L. 2002. Mutations in two genes encoding different subunits of a receptor signaling complex result in an identical disease phenotype. American Journal of Human Genetics 71:656–662. DOI: https://doi.org/10.1086/342259, PMID: 12080485 Petry F, Botto M, Holtappels R, Walport MJ, Loos M. 2001. Reconstitution of the complement function in c1q- deficient (c1qa-/-) mice with wild- type bone marrow cells. Journal of Immunology 167:4033–4037. DOI: https://doi.org/10.4049/jimmunol.167.7.4033, PMID: 11564823 Rao M, Gershon MD. 2016. The bowel and beyond: the enteric nervous system in neurological disorders. Nature Reviews. Gastroenterology & Hepatology 13:517–528. DOI: https://doi.org/10.1038/nrgastro.2016.107, PMID: 27435372 Ritchie ME, Phipson B, Wu D, Hu Y, Law CW, Shi W, Smyth GK. 2015. Limma powers differential expression analyses for RNA- sequencing and microarray studies. Nucleic Acids Research 43:e47. DOI: https://doi.org/10. 1093/nar/gkv007, PMID: 25605792 Rolig AS, Mittge EK, Ganz J, Troll JV, Melancon E, Wiles TJ, Alligood K, Stephens WZ, Eisen JS, Guillemin K. 2017. The enteric nervous system promotes intestinal health by constraining microbiota composition. PLOS Biology 15:e2000689. DOI: https://doi.org/10.1371/journal.pbio.2000689, PMID: 28207737 Roosen L, Boesmans W, Dondeyne M, Depoortere I, Tack J, Vanden Berghe P. 2012. Specific hunger- and satiety- induced tuning of guinea pig enteric nerve activity. The Journal of Physiology 590:4321–4333. DOI: https://doi.org/10.1113/jphysiol.2012.231134 Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 29 of 30 Immunology and Inflammation Research article Schafer DP, Lehrman EK, Kautzman AG, Koyama R, Mardinly AR, Yamasaki R, Ransohoff RM, Greenberg ME, Barres BA, Stevens B. 2012. Microglia sculpt postnatal neural circuits in an activity and complement- dependent manner. Neuron 74:691–705. DOI: https://doi.org/10.1016/j.neuron.2012.03.026, PMID: 22632727 Schifferli JA, Ng YC, Peters DK. 1986. The role of complement and its receptor in the elimination of immune complexes. The New England Journal of Medicine 315:488–495. DOI: https://doi.org/10.1056/ NEJM198608213150805, PMID: 2942776 Schwechter B, Rosenmund C, Tolias KF. 2013. Rasgrf2 Rac- GEF activity couples NMDA receptor calcium flux to enhanced synaptic transmission. PNAS 110:14462–14467. DOI: https://doi.org/10.1073/pnas.1304340110, PMID: 23940355 Shah D, Romero F, Zhu Y, Duong M, Sun J, Walsh K, Summer R. 2015. C1Q deficiency promotes pulmonary vascular inflammation and enhances the susceptibility of the lung endothelium to injury. The Journal of Biological Chemistry 290:29642–29651. DOI: https://doi.org/10.1074/jbc.M115.690784, PMID: 26487714 Sorensen L, Saito A, Poon S, Han MN, Humenick A, Mutunduwe K, Glennan C, Mahdavian N, Brookes SJH, McQuade RM, Foong JPP, Gómez- de- Mariscal E, Muñoz- Barrutia A, King SK, Haase R, Carbone S, Veldhuis NA, Poole DP, Rajasekhar P. 2022. Gut analysis toolbox. 1.0.0. Zenodo. https://doi.org/10.5281/zenodo.6399524 DOI: https://doi.org/10.5281/zenodo.6399524 Stevens B, Allen NJ, Vazquez LE, Howell GR, Christopherson KS, Nouri N, Micheva KD, Mehalow AK, Huberman AD, Stafford B, Sher A, Litke AM, Lambris JD, Smith SJ, John SWM, Barres BA. 2007. The classical complement cascade mediates CNS synapse elimination. Cell 131:1164–1178. DOI: https://doi.org/10.1016/j. cell.2007.10.036, PMID: 18083105 Stuart T, Butler A, Hoffman P, Hafemeister C, Papalexi E, Mauck WM III, Hao Y, Stoeckius M, Smibert P, Satija R. 2019. Comprehensive integration of single- cell data. Cell 177:1888–1902.. DOI: https://doi.org/10.1016/j.cell.2019.05.031 Thielens NM, Tedesco F, Bohlson SS, Gaboriaud C, Tenner AJ. 2017. C1Q: A fresh look upon an old molecule. Molecular Immunology 89:73–83. DOI: https://doi.org/10.1016/j.molimm.2017.05.025, PMID: 28601358 Tian XP, Zhang X. 2010. Gastrointestinal involvement in systemic lupus erythematosus: insight into pathogenesis, diagnosis and treatment. World Journal of Gastroenterology 16:2971–2977. DOI: https://doi.org/10.3748/wjg. v16.i24.2971, PMID: 20572299 Uesaka T, Young HM, Pachnis V, Enomoto H. 2016. Development of the intrinsic and extrinsic innervation of the gut. Developmental Biology 417:158–167. DOI: https://doi.org/10.1016/j.ydbio.2016.04.016, PMID: 27112528 Ural BB, Yeung ST, Damani- Yokota P, Devlin JC, de Vries M, Vera- Licona P, Samji T, Sawai CM, Jang G, Perez OA, Pham Q, Maher L, Loke P, Dittmann M, Reizis B, Khanna KM. 2020. Identification of a nerve- associated, lung- resident interstitial macrophage subset with distinct localization and immunoregulatory properties. Science Immunology 5:eaax8756. DOI: https://doi.org/10.1126/sciimmunol.aax8756, PMID: 32220976 van Schaarenburg RA, Suurmond J, Habets KLL, Brouwer MC, Wouters D, Kurreeman FAS, Huizinga TWJ, Toes REM, Trouw LA. 2016. The production and secretion of complement component C1q by human mast cells. Molecular Immunology 78:164–170. DOI: https://doi.org/10.1016/j.molimm.2016.09.001, PMID: 27648858 Verheijden S, De Schepper S, Boeckxstaens GE. 2015. Neuron- macrophage crosstalk in the intestine: a `` microglia’’ perspective. Frontiers in Cellular Neuroscience 9:403. DOI: https://doi.org/10.3389/fncel.2015. 00403, PMID: 26528133 Villar J, Cros A, De Juan A, Alaoui L, Bonte PE, Lau CM, Tiniakou I, Reizis B, Segura E. 2023. ETV3 and ETV6 enable monocyte differentiation into dendritic cells by repressing macrophage fate commitment. Nature Immunology 24:84–95. DOI: https://doi.org/10.1038/s41590-022-01374-0, PMID: 36543959 Vrees MD, Pricolo VE, Potenti FM, Cao W. 2002. Abnormal motility in patients with ulcerative colitis: The role of inflammatory cytokines. Archives of Surgery 137:439–445. DOI: https://doi.org/10.1001/archsurg.137.4.439, PMID: 11926949 Warren J, Mastroeni P, Dougan G, Noursadeghi M, Cohen J, Walport MJ, Botto M. 2002. Increased susceptibility of c1q- deficient mice to Salmonella enterica serovar typhimurium infection. Infection and Immunity 70:551– 557. DOI: https://doi.org/10.1128/IAI.70.2.551-557.2002, PMID: 11796582 Wessels MR, Butko P, Ma M, Warren HB, Lage AL, Carroll MC. 1995. Studies of group B streptococcal infection in mice deficient in complement component C3 or C4 demonstrate an essential role for complement in both innate and acquired immunity. PNAS 92:11490–11494. DOI: https://doi.org/10.1073/pnas.92.25.11490, PMID: 8524789 Yang CH, Yeh YJ, Wang JY, Liu YW, Chen YL, Cheng HW, Cheng CM, Chuang YJ, Yuh CH, Chen YR. 2017. NEAP/ DUSP26 suppresses receptor tyrosine kinases and regulates neuronal development in zebrafish. Scientific Reports 7:5241. DOI: https://doi.org/10.1038/s41598-017-05584-7, PMID: 28701747 Yu G, Wang LG, Han Y, He QY. 2012. ClusterProfiler: An R package for comparing biological themes among gene clusters. OMICS 16:284–287. DOI: https://doi.org/10.1089/omi.2011.0118, PMID: 22455463 Yu X, Rollins D, Ruhn KA, Stubblefield JJ, Green CB, Kashiwada M, Rothman PB, Takahashi JS, Hooper LV. 2013. Th17 cell differentiation is regulated by the circadian clock. Science 342:727–730. DOI: https://doi.org/10. 1126/science.1243884, PMID: 24202171 Yu X, Wang Y, Deng M, Li Y, Ruhn KA, Zhang CC, Hooper LV. 2014. The basic leucine zipper transcription factor NFIL3 directs the development of a common innate lymphoid cell precursor. eLife 3:e04406. DOI: https://doi. org/10.7554/eLife.04406, PMID: 25310240 Zhang P, Cao L, Zhou R, Yang X, Wu M. 2019. The lncRNA NEAT1 promotes activation of inflammasomes in macrophages. Nature Communications 10:1495. DOI: https://doi.org/10.1038/s41467-019-09482-6, PMID: 30940803 Pendse et al. eLife 2023;0:e78558. DOI: https://doi.org/10.7554/eLife.78558 30 of 30 Immunology and Inflammation
10.1007_s43076-022-00253-9
Trends in Psychology https://doi.org/10.1007/s43076-022-00253-9 ORIGINAL ARTICLE Uncertainty in Child Custody Cases After Parental Separation: Context and Decision‑Making Process Josimar Antônio de Alcântara Mendes1  · Thomas Ormerod1 Accepted: 10 November 2022 © The Author(s) 2023 Abstract Context factors (e.g. a family’s developmental crisis) can affect the child custody decision-making process and the child’s best interests after parental separation. But what are these context factors, and how can they vary across different cultures and legal systems? This paper reports a cross-cultural qualitative study funded by the Brazilian Ministry of Education and was carried out under a Naturalistic Decision- making approach. This study addresses context factors that impact the decision- making of experienced legal actors working in child custody cases. Interviews were conducted with 73 legal actors (judges, prosecutors, lawyers, psychologists, and social workers) in Brazil and England. The data gathered were analysed employing a reflexive thematic analysis that generated seven themes addressing how uncertainty is structured by context factors in child custody cases after parental separation. The themes generated encompassed three domains (‘family’, ‘family court’, and ‘legal- psychosocial’) that draw attention to the sources of uncertainty in child custody cases, especially to the role of contextual players (family and children) in the child custody decision-making process. Keyword Child custody · Decision-making · Divorce · Uncertainty · Thematic analysis Cases in which divorced parents cannot reach a settlement and therefore need to go to trial, are estimated to be about 5% of the total of divorces (Baker, 2012; Kelly, 2007; Wallace & Koerner, 2003). Despite being a small part of the total of divorces, * Josimar Antônio de Alcântara Mendes josimards@gmail.com 1 University of Sussex, Brighton, UK Vol.:(0123456789)1 3 Trends in Psychology these cases pose a challenge to family court professionals as such cases tend to be very complex and involve different factors that will impact the decision-making pro- cess and the child’s best interests.1 Extensive scholarship has addressed divorce-related factors that can affect the decision-making process. For instance, some studies addressed the application of ‘the best interests of the child’ standard (Eekelaar, 2015; Mendes & Ormerod, 2019), procedures for evaluation (Goldstein, 2016), judges’ attitudes (Stamps et al., 1996), psychologists’ and lawyers’ views (O’Neill et al., 2018) as well as ‘child and family features’ that can influence the judges’ decision-making (Wallace & Koerner, 2003). These issues reinforce the assumption that a decision-making process carried out in natural settings (i.e. in the real world) is affected by uncertainty (Klein et al., 1993; Lipshitz & Strauss, 1997; Lipshitz et al., 2001; Lipshitz, 1993a, b). However, there is still a lack of scholarship focused on how context factors can play a role in the decision-making process in child custody cases – especially factors that are not related to mental health issues, personality traits, intimate partner violence, child abuse and neglect. We understand ‘context factors’ as issues and/or dynamics regarding individual, organizational and system factors that can influence the decision-making process, especially by prompting uncertainty into this process. In general, two core domains constrain most of these factors’ variance: 1) type of legal system (e.g. laws, legal and technical guidance/practices); and 2) contextual issues (e.g. family’s developmental struggles) (Mendes & Bucher-Maluschke, 2017; Mendes & Ormerod, 2021). In this study, our exploration of context factors considers differences across two nations that differ according to their underlying legal systems: the common law approach of England, and the civil law approach of Brazil. The English family jus- tice system bounces between two contrasting practice approaches: (1) behaviour- focused and (2) outcome-focused (Eekelaar & Maclean, 2013). The former refers to the emphasis on settlements made by the parties through the modification of their expectations/behaviours rather than through proceedings and adjudication—in this scenario, whatever encompasses the settlement is less important than the parties’ agreement and closing the case. The latter approach refers to the idea that family justice works as an ‘impartial spectator’ that can provide fair outcomes throughout a fair process. In Brazil, since the enactment of the New Code of Civil Proceedings in 2015, the family justice system has specific routes that aim to promote consensual settle- ments or self-composition.2 However, the Brazilian legal system still is very liti- gation-driven and, in most cases, these routes are there just pro-forma (Mendes & Ormerod, 2021). In addition, the Brazilian family justice system has a child custody 1 Despite legal and definitional differences, ‘divorce’ and ‘parental separation’ will be referred to as the same thing throughout this paper: the relationship breakdown between two people that had a child together. 2 This is related to processes in which both parties (parents) find a functional way to communicate their differences, interests and goals regarding the matter under dispute and to thereby reach an agreement by themselves, without judicial mediation. 1 3 Trends in Psychology decision-making process that is ‘closed-ended’ as the law points out only two pos- sible outcomes: (1) joint custody (preferably); and (2) sole physical custody.3 Context factors, tend to define and frame "the space in which decision-making processes operate" (Jones et al., 2014, p. 203). In this sense, the task of understand- ing context factors that surround the process of making a decision is crucial (Lip- shitz, 1993a)—especially because uncertainty is the main impediment to an effec- tive decision-making process (Lipshitz & Strauss, 1997). This task is challenging for legal actors because family struggles are more related to psychosocial issues than legal ones, which leads to more uncertainty in such cases. Within the legal scholarship, the role of uncertainty is largely addressed as ‘legal uncertainty’ and it is seen as a consequence of generic legal standards that make it difficult to say, ex ante, if certain actions are legal and what legal officials might do (Lang, 2017). However, the legal literature neglects other factors that can lead to uncertainty within family justice and its decision-making processes. Some scholars have addressed how legal actors use heuristics to deal with uncer- tainty in child custody cases (e.g. Enosh & Bayer-Topilsky, 2015), noting that uncer- tainty is a key player in such cases. However, the literature in this field is lacking studies that investigate context factors in child custody cases that build and sustain the levels of uncertainty as well as the consequences of it. This is concerning as uncertainty “affects real-world decisions by interrupting ongoing action, delaying intended action, and guiding the development of new alternatives” (Lipshitz, 1993b, p. 173). Hence, family justice and its professionals should be aware of context fac- tors because they can lead to errors and biased judgments during the decision-mak- ing process, which can impair the quality of the decisions made, affecting the child’s best interests as well as the family’s well-being. In an attempt to draw attention to context factors (especially those not related to mental health issues, personality traits, intimate partner violence, child abuse and neglect and the like) and how they are structured within the child custody decision- making process, this study presents results from a qualitative inquiry that identified key context factors responsible for producing and sustaining uncertainty in child custody cases after parental separation.4 3 For further discussion regarding the Brazilian family justice and child custody after parental separa- tion, please see Mendes and Ormerod (2021). 4 These results are part of a larger research project that has identified cognitive strategies used by legal actors to cope with uncertainty prompted by context factors. The project had a naturalistic and cross-cul- tural design that approached legal and cultural issues in Brazil and England, and aimed to understand: 1) how the decision-making process is structured in terms of its contextual dynamics and constraints; 2) the role of legal actors in the decision-making process; 3) how ‘the best interests of the child’ is understood and applied; and 4) how the type of legal system (civil law in Brazil, common law in England) affects the decision-making process. 1 3 Trends in Psychology Method This study’s design incorporated a Naturalistic Decision-Making research meth- odology, which aims to understand and describe how individuals make their deci- sions in the real world. This approach highlights “how expert practitioners perform cognitively complex functions in demanding, real-world situations characterized by uncertainty, high stakes, and team and organizational constraints” (Patterson et al., 2016, p. 229). Instruments, Participants, and Procedures This study used semi-structured interviews with open-ended and closed questions – to check the interview questions, please see Online Resource 1.5 The first author conducted the interviews, which were held for 40 to 70 min, with an average inter- view time of 55 min. Seventy-three Brazilian and English participants (judges, pros- ecutors,6 lawyers, psychologists and social workers) took part in this study. The main inclusion criterion for all participants was to have at least two years of experience in child custody cases after parental separation. To check participants’ demographics, please see Online Resource 2. In both countries, we recruited participants in three ways: a) through the research- ers’ existing network; b) by sending participation invitations via email and mail; and c) through snowball recruitment7: each participant was asked if they knew some- one meeting the inclusion criteria, whom they could recommend to take part in the study. Access to English participants was difficult because applications to approach magistrates and social workers (from CAFCASS8) were not granted. Exclusively in England, we also reached participants via: i) LinkedIn; ii) inviting eligible lawyers by email invitation based on the list available at http:// www. resol ution. org. uk9; iii) inviting eligible psychologists by email (we used the list available at the British Psy- chological Society’s Directory of Expert Witnesses—https:// www. bps. org. uk/ lists/ EWT/ search); iv) emailing authors with papers published on child custody cases and/or the best interests of the child—they were asked if they would like to take part in the study or if they would nominate anyone else eligible. Nevertheless, due 5 The questions are based on prior studies that focused on: 1) law, procedures and judicial process regarding parental separation and child custody and contact/access in Brazil and England—see Mendes and Ormerod (2021); and 2) a systematic review on ‘the best interests of the child’ in English and Portu- guese—see Mendes and Ormerod (2019). 6 In Brazil and England, divorce and child custody are a private law matter. However, in Brazil, there are some cases in which the State is seen as an interested party and non-criminal prosecutors can be involved. For more clarification, see Mendes and Ormerod (2021). 7 See Sadler et al. (2010) for further information. 8 Stands for Children and Family Court Advisory and Support Service. It is the English “evaluation ser- vice” and they advise family courts about what is safe for children and what are the child’s best interests in child custody cases. 9 Resolution’ is an organisation promoting constructive resolution of family disputes and has over 6,500 members among family lawyers and other professionals. 1 3 Trends in Psychology to the circumstances described, the number of participants in England was smaller compared to Brazil, but as diverse as the Brazilian group.10 Informed consent was obtained from all individual participants included in the study and the interviews were conducted either in person, via Skype or by telephone in both countries, and recorded with a Sony ICDBX140 Digital Voice Recorder. The study and its materi- als (e.g. information sheet and consent form) were approved by the University of Sussex’s Social Sciences & Arts Research Ethics Committee under the Certificate of Approval number ER/JA454/2. The authors have no competing interests to declare that are relevant to the content of this article. Data Analysis This study adopted thematic analysis as its theoretical framework to understand and analyse the data gathered. A thematic analysis aims to search for patterns within qualitative data. Thematic analysis is a process that identifies, organises, and inter- prets these patterns, leading to analysis and final reporting on those patterns through the use of ‘themes’ (Boyatzis, 1998; Braun & Clarke, 2006, 2013). A theme can be seen as a ‘wall’ composed of a lot of ‘bricks’ (codes) connected by a strong ‘cement’ (meanings). Both ‘bricks’ and ‘cement’ are distinguished and understood by the researcher’s subjectivity and active role in the data analysis pro- cess, which is organic and interactive, going beyond the first round of coding, and extending throughout the whole process of analysis (Braun & Clarke, 2022a; Braun et al., 2019).11 We propose an Integrative Data-driven Thematic Analysis (IDDTA) that inte- grates inductive and abductive (theoretical) layers of analysis, revealing manifest and latent levels of content.12 IDDTA assumes that: (a) neither the data nor the meanings derived from it are given; both are detected and distinguished as such by 10 In Brazil, three cities were selected: 1) Brasília—it is Brazil’s capital and its court has a large and solid system for the evaluation of child custody cases; as such it is treated as a reference in Brazil; b) São Paulo—it is the biggest city in South America, has the biggest court in the world (considering the num- ber of cases per year) and also has the biggest family court in South America (where participants were recruited); and 3) Porto Alegre—it has a court known for launching case laws concerning family law that have spread to other courts, and has also inspired the enactment of acts in this field. Selecting these three cities enabled this study to economically but effectively achieve a representative sample of the ‘Brazilian child custody field’. We intended to take the same approach in England by selecting participants from London, Brighton (southern) and one northern city. However, gathering participants in England was a herculean task that took over eight months. Hence, we decided to recruit participants from all over Eng- land. 11 Thematic analysis is a highly flexible methodology, and does not prescribe procedures of data collec- tion, or limit the theoretical or epistemological perspectives possible within it (Braun & Clarke, 2006, 2013; Braun et al., 2019; Nowell et al., 2017). Boyatzis (1998, p. 1) refers to thematic analysis as a “way of seeing”, meaning that different people can see different things by looking at and analysing the same data. Moreover, different people can conceive and use thematic analysis in different ways (Braun et al., 2019). 12 A similar approach was proposed by Urquhart (2013) for Grounded Theory. She referred to the ‘mid- dle-range’ coding process in which the coding would emerge from inputs based on the raw data and on the literature, thus combining induction and abduction processes. 1 3 Trends in Psychology an observer13; (b) qualitative research is inevitably underpinned by the researcher’s subjectivity, hence no knowledge is neutral14; and (c) qualitative research is a pro- cess that analytically organises, interprets and reveals patterns of meanings within the data by means of analytic inputs and outputs that interact in a recursive way. Braun and Clarke (2022b) reflect on the key role of the researcher’s subjectivity dur- ing the data analysis process and how the researcher’s work should generate and report themes that go beyond a ‘topic summary’ by portraying ‘interpretative stories’ concerning consolidated meaning. We agree with this idea but we understand that it is also important to consider that: a) it is the researcher’s unique views, perspectives, experience and understanding (therefore, their subjectivity) that will guide them in the process of organising and describing the data. Hence, the researcher’s subjectiv- ity is present and is pivotal in the accomplishment of these tasks, even though these tasks’ outcomes might seem less complex and sophisticated than “interpretative sto- ries built around [a] uniting meaning” (Braun & Clarke, 2022b, p. 3); and b) qualita- tive research can be relevant for poly-making (Sale & Thielke, 2018; Tracy, 2010) and decision-making (Mendes, 2022). In this sense, whenever the outcomes of a qualitative study are aimed at or relevant for policy-makers and decision-makers, it is important to ensure this audience’s readership and grasping. Sometimes, this means providing results that are a little bit more ‘structured’ and descriptive. Taking these assumptions into account, and based on the assertions of Braun and Clarke (2022a) and Braun et al. (2019), IDDTA is a reflexive thematic analysis as it assumes and highlights the researcher’s active role in the process of outlining the generated themes; it also highlights the meaning rather than quantity of data. This study’s IDDTA had five phases inspired by and adapted from models in Braun and Clarke (2006, 2013, 2022a), Braun et  al. (2019) and Nowell et  al. (2017): Phase I—Familiarisation (before starting coding, the first author read the interview tran- scripts, intending to get ‘closer to the data’, its depth and breadth. This familiarisa- tion was an active process that looked for meanings and patterns by speed-reading the whole dataset before moving on to Phase II (open coding). During this initial phase, the first author used the memoing tool15); Phase II—First Level of Analysis: 13 In other words, we assume the assertion, given by Second-order Cybernetic theorists Maturana and Varela (1991) and Von Foerster (2003), that ‘things’ only become things when observed, distinguished and pointed out by an observer—i.e. it is the observer and their active perception that give meaning to things. Thus, reality and its contents (as meaningful constructs) emerge from an observer’s perspective. In IDDTA, this is set as an essential principle throughout the whole process that leads the observer to identify, interpret, classify and analyse codes and themes. 14 According to González Rey’s (2011) assertions, in qualitative data analysis, it is the researcher’s subjectivity, in a dialogic interaction with the data (for extension, with the research participants’ sub- jectivity too), that drives the process of interpretation (i.e., building up meanings and themes). Hence, no knowledge is produced outside of historical, social and cultural contexts; neither is it removed from the researcher’s subjectivity, previous knowledge or experiential framework. Therefore, no knowledge is totally neutral, pure or inductive. 15 The memoing tool was fundamental for this phase. It is used to take notes regarding any ideas, insights or interpretations that emerge during the process. This technique was applied throughout the whole analysis, and it was important to identify links that pointed out patterns and resulting themes. The notes were also important to embody the latent (interpretative) character of the process. 1 3 Trends in Psychology open coding16 (process aimed to organise, describe, sort and synthesise the dataset in a very open way, without restraints—this phase generated 62 codes (see Online Resource 3)17; Phase III—Second Level of Analysis: generating initial themes (analysis of initial codes to construct themes—this phase generated 12 candidate themes and 25 features; see Online Resource 4); Phase IV—Reviewing & Setting the Themes: definitions and relationships (refinement of candidate themes and features and trying to set them in a broader context alongside meaningful themes that also highlighted their connections—this phase generated 7 final themes and 22 features that will be presented in the next section. During the whole process, some themes were split or combined with others to compose other more meaningful themes and/ or features); and Phase V—Anchoring18 & Thematic Map (pointing out in which participants’ data themes and features were based (hence, ‘anchored’) on; a thematic map to showcase how themes and features are connected and interacting); Phase VI—Ensuring Trustworthiness: credibility and dependability (peer review/debrief- ing19 and reflexivity (see Online Resource 5)20). For the data analysis process of this IDDTA, the unit of coding (the basic seg- ment of raw data assessed that elicits meanings that help to identify patterns related to the studied phenomenon) was a sentence.21 Also, the unit of analysis (the entity considered as the information source upon which interpretation was focused) was the whole transcript concerning each interview. Figure 1 summarises the whole process of data analysis. This study was not preregistered. Also, due to the nature of this study, partici- pants did not agree for their data (whole transcripts) to be shared publicly. However, some supplemental material concerning the data analysis process will be available online. Results This study gathered data from 48 Brazilian and 25 English participants. Of these, 64% were female. The proportion of females and males in each country and within each category of legal actors was similar. The mean years of experience in Brazil was 14 (SD = 9.7) and 16.5 (SD = 8.9) in England. 16 Inspired by the conceptions of ‘open coding’ by Urquhart (2013) and ‘initial coding’ by Charmaz (2014). 17 This coding process was helped by the qualitative data analysis software NVivo 10 for Mac OS. 18 This strategy is just a tool used to provide the results’ confirmability. It should not be seen as a quan- titative measure in which ‘the larger the number of supporters (participants) pointed, the more significant that theme/feature is’. 19 Four expert practitioners and academics with expertise in child custody cases and/or qualitative research reviewed this study’s data analysis process and the themes generated. 20 To ensure the final results’ trustworthiness through ‘credibility’, ‘confirmability’ and ‘dependability’ as asserted by Creswell and Poth (2017), Darawsheh (2014), Flick et al. (2004) and Guest et al. (2012). 21 The level of analysis can be ‘line-by-line’, ‘sentence-by-sentence’, ‘paragraph-by-paragraph’ or ‘inci- dent-by-incident’. The researcher will choose the level of analysis according to their objectives and the data characteristics. 1 3 Trends in Psychology Fig. 1 Data analysis process The themes below are presented according to a hierarchy of attributes: a) a theme: generated according to meaningful content in the dataset; b) feature: signposts charac- teristics of the theme; and c) highlight: relevant issues arising within a feature. Each theme is illustrated with participants’ quotations that are linked to their ID, which pre- sents their country (‘BR’; ‘EN’) and category (‘Jd’ = Judge; ‘Lw’ = Lawyer; ‘Pr’ = Pros- ecutor; ‘Psy’ = psychologist; SW = Social Worker). In Brazil, participants also have their city pointed in their ID (BsB = Brasília; POA = Porto Alegre; SP = São Paulo). Table 1 presents the themes generated and their features (or subthemes). It also shows how these themes are anchored in the data. The thematic map presented in Fig. 2 showcases context factors present in child custody decision-making after parental separation. It shows how the seven themes are connected and interacting between and within each another. The map also shows their classification according to specific domains: 1) ‘family’; 2) ‘family court’; and 3) ‘legal-psychosocial’. Family Domain Themes that encompass the ‘family’ domain represent issues strictly related to the family interaction and dynamics after parental separation that can impact the 1 3 Trends in Psychology Table 1 Themes and features generated by the reflexive thematic analysis and their anchoring on the data Theme Data anchoring Theme CT1: Parental Separation: Crisis and Family Life Cycle (CT1.1) Dysfunctionally coping divorce: family crisis1 (CT 1.2) Misunderstanding and pathologisation of family interactions and coping strategies in the context of custody dispute: perspectives on parental alienation2 (CT 1.2.1) Tricks the decision-making (CT 1.2.2) Impairs the child’s role (CT 1.3) Parental separation as part of the family life cycle3 1 – P2, P8, P9, P11, P17, P20, P21, P24, P31, P35, P39, P42, P44, P45, P49, P55, P57, P58, P62, P67 2 – P1, P2, P3, P4, P5, P11, P16, P17, P22, P23, P24, P30, P32, P36, P40, P42, P43, P50, P54, P60, P62 3 – P1, P2, P12, P14, P18, P19, P24, P26 Theme CT2: Hindering the Best Interests of 4 – P1, P2, P3, P4, P5, P7, P11, P14, P15, P17, P18, the Child (CT 2.1) Conjugality Vs. Parenthood4 (CT 2.2) Detaching from the child and attaching to the litigation5 (CT 2.3) Lack of parenting skills6 (CT 2.4) “No ‘child maintenance’, no contact with the child”7 (CT 2.5) Misunderstanding joint custody8 (CT 2.6) Involving the child in parental conflict9 P22, P23, P24, P25, P26, P27, P34, P35, P36, P38, P41, P42, P43, P45, P50, P54, P56, P57, P58, P62, P63, P66, P67, P68, P70, P72, P73 5 – P1, P2, P3, P4, P5, P6, P7, P8, P12, P13, P15, P16, P17, P20, P21, P24, P25, P27, P28, P29, P30, P33, P34, P37, P44, P47, P49, P50, P51, P52, P53, P56, P58, P59, P62, P63, P64, P65, P68, P69, P70, P72, P73 6 – P1, P2, P3, P14, P15, P16, P36, P46 7 – P2, P3, P5, P27, P29, P31, P45 8 – P6, P9, P15, P16, P22, P25, P31, P34, P43, P44 9 – P1, P2, P3, P5, P8, P11, P12, P13, P14, P17, P24, P34, P35, P36, P37, P39, P40, P41, P42, P43, P44, P47, P50, P54, P56, P57, P60, P62, P66, P68, P69, P73 Theme CT3: The Judiciary’s Constraints & 10 – P2, P4, P11, P13, P14, P16, P21, P42, P44, P45, P48, P49 11 – P7, P8, P12, P18, P19, P20, P22, P25, P26, P29, P31, P34, P35, P42, P54, P57, P59, P71, P72, P73 12 – P9, P12, P36, P38 13 – P49, P51, P56, P61, P65, P68 Practices (CT 3.1) “The Law is powerless”: legal and epis- temological limitations of Law10 (CT 3.1.1) Limits of Law (CT 3.1.2) Litigious mindset (CT 3.2) Organisational issues11 (CT 3.2.1) Time & Workflow (CT 3.2.2) Staff & Workload (CT 3.2.3) Judges’ career & Courts (CT 3.2.4) Lack of training and knowledge (CT 3.3) Between fear and bravery: the psycholo- gists’ practice in Brazil12 (CT 3.4) An advocate in intractable cases: the psychologists’ practice in England13 Theme CT4: Applying The Best Interests of the 14 – P3, P5, P8, P9, P10, P20, P37, P41, P45, P46, Child Principle (CT 4.1) Indeterminacy14 (CT 4.2) Idiosyncrasy15 P47, P59, P62, P64, P69 15 – P3, P5, P6, P7, P14, P15, P17, P24, P27, P36, P39, P40, P42, P43, P44, P47, P51, P56, P57, P59, P63, P64, P71, P72 1 3 Table 1 (continued) Theme Theme CT5: Making the Decision-Making (CT 5.1) Misconduct, maltreatment and abuse Process Harder allegations16 (CT 5.2) Tied Parents: “I cannot pick one”17 (CT 5.3) Legal actors’ emotional struggles18 Theme CT6: Assessing the Child’s Best Interests in Child Custody Cases: Evaluation Services (CT 6.1) ‘Psychosocial Study’: the Brazilian model19 Trends in Psychology Data anchoring 16 – P2, P3, P6, P9, P13, P16, P18, P24, P25, P35, P36, P37, P38, P44, P45, P54, P56, P57, P59, P62, P63, P65, P66, P67, P71, P72 17 – P1, P27, P28, P44 18 – P16, P27, P34 19 – P2, P3, P4, P8, P10, P12, P13, P21, P22, P23, P24, P26, P35, P36, P39, P41, P42 20 – P49, P50, P52, P53, P54, P56, P57, P59, P60, P69 (CT 6.1.1) Family Firefighters: the role of psycho- social evaluation (CT 6.1.2) Interdisciplinarity (CT 6.1.3) Non-protocol-based practice (CT 6.2) ‘Children and Family Court Advisory and Support Service – CAFCASS’: the English model20 (CT 6.2.1) Protocol-based practice: Children Act’s Sect. 7 Report (CT 6.2.2) Non-evidence-based practice Theme CT7: Making a Child’s Arrangement 21 – P1, P15, P16, P21, P23, P27, P43, P44, P49, Decision Involving Adolescents (CT 7.1) “It’s quite impossible to go against their will”21 P50, P51, P52, P55, P66, P69, P71 22 – P2, P35, P42, P43, P44, P45, P47, P73 (CT 7.2) “They can play the game too”: getting into the litigating parents’ dynamic22 Fig. 2 Thematic map 1 3 Trends in Psychology decision-making process. For instance, this domain comprises issues concerning family life, family development, family member roles, parenting, co-parenting, liti- gation and coping strategies after the divorce. Theme CT1: Parental Separation: Crisis and Family Life Cycle The feature Dysfunctionally coping with divorce: family crisis (CT1.1) captures dys- functional strategies used by families to cope with times of hardship after parental separation: I understand that [parents are] going to court and asking the judge what are the best interests of the child is a dysfunctionality in the family itself (BR_ SP.Psy.01) Generally, what tends to happen is that there is a lot of heat when it comes to [parental] separation and that kind of tends to cloud a lot of the judgements when it comes to contact [with the child] (EN_Lw.03) Some legal actors see family dysfunctionality whenever a family goes to court for the purpose of delegating to a third party (the judge) the power to solve their prob- lems. This dynamic might be driven by multiple difficulties that the whole family endures during a separation. The intensification of these difficulties can lead a fam- ily—especially the parents—to become blind to the child’s interests and the family’s well-being. This process can be characterised as a family crisis moment: Everyone is very hurt, and there is no communication. Making a decision regarding child custody at this moment is very complicated (BR_Pr.01) [the parents need to] cope and overcome this moment of crisis so they will be able to see and care for their child again (BR_POA.Psy.01) The feature Misunderstanding and pathologisation of family interactions and coping strategies in the context of custody dispute: perspectives on parental aliena- tion (CT1.2) captures legal actors’ and the judiciary’s conceptions and understand- ings regarding the family crisis, which see some of the family dysfunctional coping strategies as examples of ‘parental alienation’. This is considered to be a frequent issue in judicial custody disputes. For some legal actors, its presence will make deci- sion-making more difficult and impair the child’s role within it, as it is likely that the child will be co-opted by one of the parents: I see as more difficult cases, those in which there is a clear Parental Alienation Syndrome already installed because we have the practice of alienation already installed (BR_BsB.Jd.01) Parental alienation [is a situation] in which the child is in service of the adult’s desire (BR_POA.Psy.04) Other legal actors do not rely on parental alienation assumptions or accept its rel- evance to the decision-making process, due to its broad definition and gratuitous use within child custody cases: 1 3 Trends in Psychology I don’t like to use the term ‘parental alienation’ because it has a number of connotations which don’t necessarily help (EN_Jd.02) I think that parental alienation has become fashionable, when in fact you have to value how this was built, how the other took part, and not whether or not there is parental alienation (BR_SP.Psy.02) The feature Parental separation as part of the family life cycle (CT1.3) cap- tures conceptions that see parental separation as part of the family’s developmental cycle, and that non-assertive behaviours might happen in such situations due to the moment of crisis typical in parental separation: It is a phase of life transition and that is how I see it. It is a phase of going through transitions, and sometimes they are very emotional and people, maybe, do not know how to deal with it in a positive way (BR_POA.SW.03) Some people sometimes ask me: Does divorce destroy families? It depends on the family; some get destroyed, others do not, and some [families] understand that it is something temporary and that time will heal those wounds and the children need to be protected (BR_BsB.Jd.01) Theme CT2: Hindering the Best Interests of the Child The feature Conjugality vs. Parenthood (CT2.1) captures a frequent issue faced by separated parents involved in high-level litigation: they cannot distinguish parental issues from conjugal ones: Well, quite frequently my experience is that when there’s still hostility between parents about why their marriage is broken down that can influence greatly influence their attitude towards either visiting contact… to be able to see the other parent, to be able to facilitate that (EN_Psy.09) I think that [separating parenting from conjugality issues] it is something that, many times, [must] pass through strong psychological support. I think the judi- ciary is not always prepared for that (BR_Pr.02) These excerpts highlight the risk of unsolved and problematic conjugal issues overlapping with parental performance, at which point the child’s well-being is jeop- ardised. Hence, for some interviewees, the acrimony between parents is based not on the child’s interests but rather on issues stemming from the broken relationship. The feature Detaching from the child and attaching to the litigation (CT2.2) cap- tures issues related to situations in which the parents are so involved in their own matters, and within which they keep up the conflict, that they can neglect and harm the child’s well-being: Parents go deep into the dispute and forget the child and the main aim, which is to protect and ensure a healthy development for the child and promote a positive familial coexistence (BR_POA.SW.01) It’s about winning a case and not about what is best for the child at all. You know, to the extent of completely ignoring what the child wants (EN_Lw.06) 1 3 Trends in Psychology The feature Lack of parenting skills (CT2.3) captures issues regarding parents who do not have the necessary parental skills to protect their child: I am going to call it the emotional immaturity of the parents, you know? This is when there is no pathology involved (BR_Pr.05) Sometimes a parent does not have the slightest ability to look after the child, for various reasons, people who have problems with drugs, with alcohol, so we have several cases like this (BR_BsB.Jd.01) The feature “No ‘child maintenance’, no contact with the child” (CT2.4) captures parents’ perspectives that misunderstand the best interests of the child by making the contact between the child and the non-custodial parent conditional upon receipt of maintenance payments: Those with lower-wage parents misunderstand a lot the issue of alimony and the issue of coexistence. So, if the father does not want to pay alimony, the mother says: ok, then I will also not let you see my child. The child becomes a bargaining chip (BR_BsB.Jd.02) They [parents] associate alimony with the right to have contact with the child. It happens especially amongst people who have very little education, this is rare in the middle class, but it happens there too (BR_BsB.Jd.03) Conflating child maintenance and the right to keep contact with both parents was seen only in Brazilian interviews, as in England child maintenance is not a judicial matter at first. This issue is commonly associated with low-income families in Brazil. The feature Misunderstanding joint custody (CT2.5) captures misunderstandings regarding this type of arrangement: Sometimes the person says: Ah, I want joint custody because I want to see my son every day. This is not joint custody. The joint custody is joint care, co- responsibility (BR_BsB.Lw.02) The parents see the joint custody as a kind of mystery, it is something that “everybody likes” but they do not have a clear notion about what this kind of arrangement really is (BR_SP.Lw.04) This issue was reported only by Brazilian participants, possibly because Brazilian law contributes to this misunderstanding: The law does not define well what this joint custody would be, because, you see, in truth, family power [i.e. parental responsibility] was already enshrined in the law beforehand (BR_Pr.02) The feature Involving the child in parental conflict (CT2.6) captures issues related to high-level litigation situations in which the parents involve the child in their con- flict, by either co-opting them to one side, forming alliances or neglecting the chil- dren who are forced to assume roles and functions more suited to adults or parents: [the parents can harm the child’s best interests when] putting pressure on the child, or, first of all, by exposing the children to the conflict, by negative talk about the other parent (EN_SW.01) 1 3 Trends in Psychology The child feels in the middle of it and is often put in a position of mediating this dispute between parents. It demands from the child a psychological basis and structure that are not there. I have seen cases in which the child ends up somatising these struggles (BR_SP.Psy.03) Some children become carers for parents who are facing a really difficult mar- riage breakdown. They take on too much responsibility, emotionally they’re not really ready for (EN_Psy.09) These excerpts highlight how reckless parental litigation can prove prejudicial towards children caught up in such situations, as they can either get triangulated within their parents’ conflicts (pushed to pick sides and form alliances) or be forced to assume parental roles and functions that they should not have to. Theme CT4: Applying the Best Interests of the Child Principle The feature Idiosyncrasy (CT4.2) captures characteristics that make the assurance of the best interests of the child principle (BIC) very idiosyncratic: It [BIC] will depend on the customs, moral and cultural values of each family, because we know that each family has its principles, its morality, and this will vary from family to family (BR_BsB.Lw.01) Therefore, I consider that [BIC] is extremely subjective from case to case because it varies so much, the way that the guidelines are interpreted (EN_ Psy.04) These idiosyncratic characteristics indicate that assuring the best interests of the child depends on moral and cultural variations between families, and consequently for each child in their respective circumstances. Therefore, this principle cannot be generalised for all cases. Theme CT5: Making the Decision‑Making Process Harder The feature Misconduct, maltreatment and abuse allegations (CT5.1) captures situ- ations in which there are allegations of abuse, violence or maltreatment against the child that make the custodial decision-making process even harder: They [hardest cases] are those in which there are allegations of violence of any kind (BR_SP.Psy.02) Cases involving allegations of sexual abuse [are the hardest]. Because they are almost impossible to prove always. It is very difficult to find pieces of evi- dence to support them because they sound more like made-up narratives (BR_ SP.Psy.04) Whether there are domestic violence allegations, true or not, whether there is a sexual abuse allegation or not… that causes problems, whether it’s true or not 1 3 Trends in Psychology because the court doesn’t know how to deal with it, only the parties know or only God knows whether that is true (EN_Lw.02) Cases with allegations of maltreatment and violence seem to be the most difficult because they bring into play two essential elements to consider during the decision- making: 1) jeopardy regarding the child’s physical and psycho-emotional integrity; and 2) allegations without proof. This can be a dilemma for decision-makers as, although they value safeguarding the child’s physical and psycho-emotional well- being, they are committed to making decisions based on concrete and provable facts. Theme CT7: Making a Custodial Arrangement Involving Adolescents The feature “They can play the game too”: getting into the litigating parents’ dynamic (CT7.2) captures legal actors’ perceptions that adolescents can consciously and intentionally involve themselves in the parental conflict: They tend to make alliances with one or the other according to their own inter- ests (BR_Pr.06) The chances of the child finding they can play one off against the other are massively enhanced and … that’s quite often the case that leads to the kind of private law proceedings in which I end up getting involved (EN_SW.05) Apparently, adolescents are not only more capable of expressing their voice and voting with their feet, they also get involved intentionally in their parents’ conflict to take advantage or to adjust themselves to the litigation dynamic within their family. Family Court Domain The ‘family court’ domain regards themes that comprise factors related to legal issues that constrain the decision-making process. These issues refer to the applica- tion of the law and its limits and procedural issues as well as how the court addresses the child during the decision-making process. Based on the participant’s account- ings regarding law limitations and legal mindset, we understand that these issues, alongside the family domain ones, are what most pressurize the decision-making process in child custody cases. Theme CT3: the Judiciary’s Constraints and Practices The feature “The Law is powerless”: legal and epistemological limitations of law (CT3.1) captures issues that the law cannot affect or control, such as domestic dynamics, parents’ behaviours outside the court, and daily routines involving the child. Also, law limitations would refer to the impossibility of preventing the child from suffering during parental separation: 1 3 Trends in Psychology I think in every divorce, or almost every divorce to some degree, the child suffers, that is my perception. But I think the law is powerless to solve this kind of problem (BR_BsB.Jd.04) We can make orders about what should happen to a child, but judges have no power to make sure it will happen (EN_Jd.01) The [family’s] reality often does not fit into legal guidelines (BR_BsB. SW.01) Another factor that constrains the legal work in child custody cases is the intrinsic adversarial modus operandi of law practice, which tends to lead parents into acrimonious litigation by encouraging a ‘litigious mindset’: If people want to fight, they will be able to and they will continue to fight whether the judgment has closed the case or not, because usually in a case like this, one parent wins and the other one loses (BR_Pr.03) Theme CT4: Applying the Best Interests of the Child Principle The feature Indeterminacy (CT4.1) captures legal and conceptual limitations that make ‘the best interests’ an unclear and vague construct: I have no way of giving you a definition [for BIC]. If you are going to look into the doctrine that underpins it, there is no specific definition for that principle (BR_BsB.Lw.01) I think it’s a very fluid concept, the best interests of the child. I think it’s open to interpretation (EN_Lw.07) Although the vagueness of ‘the best interests’ can be an issue for some legal actors, it seems a good thing for others: So it [BIC] being broad allows us to do this analysis case by case. […] If it was rigid, we would not be able to interpret it well. I prefer it to be open (BR_BsB.Jd.03) In this sense, the ‘best interests’ indeterminacy can highlight the legal actors’ discretionary power by allowing them to freely interpret what are the best inter- ests of the child according to each case. Theme CT5: Making the Decision‑Making Process Harder The feature Tied parents: “I cannot pick one” (CT5.2) captures perceptions regarding situations in which both parents present similar contexts: In situations where there is no clarity about who has the best conditions to protect or at least to take better care of the child [it is hard to make a deci- sion] (BR_BsB.Jd.01) 1 3 Trends in Psychology What is more difficult are those cases in which both parents want the cus- tody and both have similar conditions to be awarded the custody (BR_Pr.03) Theme CT7: Making a Custodial Arrangement Involving Adolescents The feature “It’s quite impossible to go against their will” (CT7.1) captures legal actors’ perceptions that it is impossible to force an adolescent to comply with a legal custody decision: The older the children, the judge becomes increasingly powerless (EN_ Jd.01) They [adolescents] are going to vote with their feet; in other words, the ado- lescent will go to live with whichever parent he or she wants to live with (EN_Jd.03) No judge or legal measure is capable of determining what an adolescent should do regarding their custody because, at the end of the day, they can do whatever they want once they leave the court. The older the adolescent, the weaker are legal custody measures. Legal‑Psychosocial Domain The ‘legal-psychosocial’ domain comprises themes that regard the evaluation ser- vices in Brazil and England. It also refers to some legal actors’ practices and their emotional struggles during the decision-making process. Theme CT3: the Judiciary’s Constraints and Practices The feature Between fear and bravery: the psychologists’ practice in Brazil (CT3.3) captures Brazilian psychologists’ perceptions on the edges of their work: It has happened to me that a lawyer questioned my competency and attached my résumé to the case transcripts in order to question my work. He had his own retained expert, then he used my résumé to claim that I was not good enough. […] This aspect, this characteristic of private family law cases makes us [staff] quite reluctant (BR_SP.Psy.04) In Brazil, the work of psychologists bounces between the fear of being targeted by the litigating dynamic (as pointed out by BR_SP.Psy.04) and the bravery to act as the child’s advocate. The feature An advocate in intractable cases: the psychologists’ practice in England (CT3.4) captures English psychologists’ commitment to safeguarding the child’s welfare in intractable cases: 1 3 Trends in Psychology [I see myself as] an advocate for the child. So, you are working for... If you’re working with the child you’re working for the child (EN_Psy.02) In England, the work of psychologists is required only on complex or intractable cases. This policy might be justified by the fact that the services of a psychologist in a child custody case tend to be more expensive than the services of social workers. None- theless, some psychologists see themselves as an advocate for the child in such cases. Theme CT6: Assessing the Best Interests of the Child in Child Custody Cases: Evaluation Services The feature ‘Psychosocial study’: the Brazilian model (CT6.1) captures the Brazilian evaluation process carried out by psychosocial staff, called a ‘psychosocial study’. It is similar to the idea of the ‘case study’ common within psychology and social work. However, understandings about the goals of such a study can vary amongst psychosocial staff: Whenever the case goes to psychosocial study, it is because the parental con- flict is very serious (BR_BsB.Jd.03) So not all cases go to a psychosocial evaluation. Only cases in which we notice a conflict; cases in which the parents agree do not go to psychosocial evalua- tion (BR_Pr.01) Judges and prosecutors tend to see psychosocial staff as ‘family firefighters’, the only solution for intractable cases. In turn, some psychosocial professionals see their role as a mediator: I think when I help adults to reflect on what is best for a child, on how the child will be better, I am doing something the judiciary should do, which is to protect the child. I think that protection should be present in all instances (BR_BsB.SW.01) [The psychosocial staff role] is to promote reflection, and intervention in some cases, where we perceive cases of vulnerability or risks that are spotted and referred to the support network (BR_BsB.SW.02) The lack of guidelines and protocol surrounding the evaluation is another charac- teristic of the Brazilian system: We do not have a standard, a rigid methodology (BR_POA.Psy.03) We do not use any protocol (BR_BsB.Psy.03) I think professional freedom is important, but I think it is also important to build a methodology of service, something that is consistent and incorporates some principles (BR_BsB.Psy.05) The feature ‘Children and Family Court Advisory and Support Service – CAF- CASS’: the English model (CT6.2) captures characteristics of the assessment carried out by English social workers from CAFCASS: 1 3 Trends in Psychology In most of those cases, there will be a report on section 7 of the Children Act, prepared either by a CAFCASS office or, if local authorities social services are involved, by a social worker.” EN_Jd.01 Unlike the Brazilian ‘psychosocial study’, the evaluation process in England is a more structured assessment with clear guidelines both from the Children Act 1989 (Sect.  7) and CAFCASS. However, there is a lack of evidence-based practice in England22: Reading through [the report], it was just absolute nonsense, it was just the CAFCASS officers’ views, it wasn’t based on facts, or logic or reasonableness (EN_Lw.02) I would say that a lot of the guidance we used to follow in CAFCASS was based on opinion, as opposed to hard research or based on evidence, and I think that could be a criticism that you might level at the system (EN_SW.01) Also, there is ‘risk-avoidance’23 related to the CAFCASS officers’ work: I do think that they are a very risk-averse organization. They certainly have become that. So, for instance, they will always take the safest route, safest route even if it means that a child potentially might suffer by not having a rela- tionship (EN_Lw.04) Discussion We understand that context factors displayed throughout the themes resemble what Wells (1978) called ‘estimator variables’ in eye-witness testimony within criminal justice. This type of variable affects the legal process but is not under its control. In the case of eye-witness testimony, they are part of the context in which the per- son witnessed a crime, and which consequently can influence a person’s testimony. Similarly, context factors constrain child custody cases and influence the decision- making process but they are not under the control of the legal system or decision- makers.24 Therefore, context factors produce uncertainty. 22 The CAFCASS website states that “practitioners use the Child Impact Assessment Framework (CIAF) when carrying out their analysis. The CIAF is a structured framework that sets out how children may experience parental separation and how this can be understood and assessed at Cafcass. It builds on our existing knowledge and guidance and follows a consistent and evidence-informed approach helping practitioners to find an outcome that is in the best interests of the children involved. The framework is informed by external research and our experience of supporting 140,000 children per year”. In regards to ‘risk-avoidance practice’, the CAFCASS website also outlines the process by which CAFCASS are asked to advise the court on what is best for the child, who are ultimately required to make a decision based on all of the information that is presented to them. 23 Idem 22. 24 Sometimes, the judiciary has the power to exercise control over these issues but it is impeded by micro or macro issues that limit powers or make them impossible to exercise. Examples are the number of cases that reach the judiciary, and financial limits on the number of legal civil servants available to tackle cases. 1 3 Trends in Psychology These estimator variables can impact the making of a decision in child custody cases as well as the child’s best interests. On one side, the family uses dysfunctional strategies to suppress the emotional distress caused by the divorce – these can blur the way legal actors perceive and understand the context in which the child’s inter- ests shall be safeguarded. On the other one, the laws and legal actors’ practices, shape how these interests will be understood and assured in such cases. Hence, the outcome for what is best for the child will depend on how both families and legal actors found themselves in each side as well as the quality of the interaction between them amongst those uncertainty factors. Every decision-making process that occurs in a natural setting will be surrounded by uncertainty (Klein et  al., 1993). In general, ‘uncertainty’ in real-life decision- making refers to the doubts generated by the perception of a certain problem and that struct and shape the search for a solution (Lipshitz & Strauss, 1997; Lipshitz, 1993b). We understand that the assembling of ‘estimator variables’, and interactions between and within them, is what structures the uncertainty in child custody cases after parental separation. However, we believe that context factors prompted by the family are the main source of uncertainty in such cases. ‘Family’: the Foremost Domain of Uncertainty in Child Custody Cases We believe that context factors in the family domain tend to produce most of the uncertainty in the decision-making process. The harder it is for the family to deal with the developmental crisis that parental separation prompts, the more uncertain the case shall be. That is because individuals and families going through a crisis are expected to act erratically, in a disorganised way, and usually employ non-asser- tive coping strategies (Mendes & Bucher-Maluschke, 2017; Sá et al., 2008). In this sense, it is possible that law professionals might have more difficulties in dealing with the families’ struggles than dealing with issues regarding the ‘family court’ and ‘legal-psychosocial’ domains because the family’s struggle relates more to psycho- social issues than legal ones. Features that encompass the family domain portray some interesting dynam- ics. For instance: a) family developmental crisis after parental separation (CT1.1; CT1.3); b) conjugal vs parental issues (CT2.1; CT2.2); c) triangulations and col- lusion inside the family (CT1.2; CT7); and d) maltreatment and abuse allegations (CT5.1). It is known that parental separation is linked to the family’s development, being part of its life cycle and representing a crisis moment to the family system (McGol- drick et al., 2014; Mendes & Bucher-Maluschke, 2017). The Family Life Cycle, in which parental separation occurs, is paced by developmental steps marked by uncer- tainty, instability and disorganisation, that push family interactions towards a change of patterns that will lead it to the next step of its development (Mendes & Bucher- Maluschke, 2017). However, a lot of families struggle with this transitional process and try to cope by means of dysfunctional and non-assertive strategies. This is a key point in the child custody decision-making process because this dynamic can shape 1 3 Trends in Psychology not only the parents’ attitudes and behaviours throughout proceedings but can also shape the characteristics of the information that shall be evaluated and taken into account to make decisions. Those non-assertive coping strategies displayed by the family can misdirect the decision-making and hinder the child’s role during a child custody dispute (CT1.2 [CT1.2.1; CT1.2.2]). An example is what some legal actors label as ‘parental alienation’. This is a very fragile concept if one considers its conceptual, scientific, ethical and technical dimensions (Barbosa et al., 2021; Barnett, 2020; Bruch, 2001; Mackenzie et al., 2020; Meier, 2020; Mendes & Bucher-Maluschke, 2017; Neilson, 2018; Pepiton et al., 2012; Shaw). ‘Parental alienation’ is a label that derives from the incomplete, imperfect, ambiguous and/or simplistic information available in child custody cases. Infor- mation in this scenario is fed and blurred by developmental struggles that the family display after parental separation. When legal actors are not aware of that, labelling can be an ‘easy way’ to go through a complex, erratic, multidetermined and dynamic scenario. This is a problem, as overly simplistic labels like ‘parental alienation’ can engender a ‘rebound effect’, as they tend to produce more of what they should tackle: uncertainty and litigation. That is because the type, amount and shape of uncertainty with which decision-makers must deal with will depend on the decision-making strategies they are applying (Lipshitz & Strauss, 1997). Hence, by applying over-simplistic uncertainty-coping strategies, legal actors might face even more uncertainty. Therefore, these labels can increase the fami- lies’ struggles (Barbosa et al., 2021; Mendes & Bucher-Maluschke, 2017), which might enhance the uncertainty and impair the child’s interests. In sum, what labels such as ‘parental alienation’ do is create a vicious cycle of uncertainty in child custody cases, as the uncertainty prompted by a family’s developmental struggles might lead to procedures and decisions that worsen the family’s devel- opmental struggles and, therefore, add more uncertainty to the decision-making process. Adolescents are significant players in the child custody scenario as they might be consciously involved in parental conflict (CT7.2). This triangulation on the part of the adolescent in the parents’ conflict shows that adolescents are not only active players in such cases but that they are also active in similar ways inside their family. Triangulation and collusion dynamics are common in child custody cases after parental separation. These dynamics are not necessarily dysfunctional or even permanent and they can be a way in which the family can go through and adjust itself to transitional developmental stages, especially very challenging ones (Emery, 2012; Juras & Costa, 2017). In this sense, some triangulations can even benefit the family. The problem is when the dynamic of a triangulation loses its transitional and adaptive character and becomes a long-lasting transactional structure, highlighting fixed and rigid oppositions that increase tension between family members. This can lead to coalitions, inflexible loyalties and triangulated conflicts that impede the family’s progress through its functional development (Barbosa et al., 2021; Juras & Costa, 2017; Mendes & Bucher-Maluschke, 2017). 1 3 Trends in Psychology The Main Difference Between Brazil and England We have observed interesting legal and cultural differences between Brazil and Eng- land that can impact the decision-making process.25 For instance, there is the way legal actors perceive divorce/parental separation. Families going through parental separation and child custody disputes seek judicial aid when they are facing a crisis moment (Mosten & Traum, 2017). However, only Brazilian participants acknowl- edged that and the dysfunctional dynamic it brings about. Only 11% of the total participants (Brazilian) referred to parental separation as part of the family life cycle. These frequencies yield that Brazilian legal actors might be more aware of the uncertainty caused by those context factors than English ones. Nevertheless, only a few of Brazilian legal actors see the separation as a potential phase for the family’s development. There are also differences regarding the way professional evaluation is carried out in each country. It tends to be non-protocol based in Brazil and non-evidence based in England. In the psychosocial evaluation, the safeguarding of the child’s interests can be weakened if one considers that the work carried out by psychologists and social workers in Brazil tends to be non-protocol-based (CT6.1[CT6.1.3]) and non- evidence based in England (CT6.2[CT6.2.2]). These results are surprising since we expected the Brazilian evaluation process to be stricter and structured due to its civil law system, which relies on written law rather than case law and customary practice. We also expected the English evaluation process to be more loose and marked by workarounds due to its common law system. However, we saw the opposite. Some Brazilian participants indicated that “the [family] reality often does not fit into legal guidelines” (BR_BsB.SW.01), so their practice needs to be more open and worka- rounds need to be applied so they can properly approach the case and cope with uncertainty. Even though English participants were working in a more open and cus- tomary system, they indicated that they rely heavily on protocols: “I tend, certainly, on a difficult case, to go through each element of the welfare checklist [from Chil- dren Act 1989] quite slavishly” (EN_Jd.01). Based on this, we understand that the nature of the legal system itself (civil or common law) is not what makes the tack- ling of uncertainty easier or harder for legal actors. In fact, this reinforces our belief that context issues, especially those regarding family developmental struggles are the greatest source of uncertainty in child custody cases. What to Make of These Results: a Preliminary Evaluation We understand that context factors are contingencies that impact legal actors’ per- formance throughout the decision-making process by influencing the cognitive strat- egies they choose to cope with uncertainty (Mendes & Ormerod, 2022). However, 25 Brazil is the most catholic country in the world. Hence, religious beliefs are likely to play a role in all matters concerning society, families and the justice system. However, religious beliefs were not pervasive or salient within the data. We believe further studies focused on legal actors’ religious issues are needed to properly investigate the role of these issues in child custody cases after parental separation. 1 3 Trends in Psychology context factors can also cue strategies that generate errors and biased judgements. Being aware of these factors, and properly interpreting them, might be the first step in assertively handling uncertainty in child custody cases as the understanding of contextual issues is an important part of the decision-making process (Ben-Haim, 2019). In a scenario of decision-making under uncertainty, any approach to tackling uncertainty is welcome, especially when ignoring uncertainty is more attractive and easier than recognising it and properly coping with it (Marchau et al., 2019). There are three typical strategies used to cope with uncertainty during a decision-making process (Lipshitz & Strauss, 1997): (1) reduce uncertainty; (2) acknowledge uncer- tainty; and (3) suppress uncertainty. The strategies to reduce uncertainty are mainly anchored in collecting additional information before making a decision. Whenever further information is not avail- able, the decision-maker can make some extrapolations based on the information available and then make a decision/take an action. In child custody cases, the strat- egy to reduce uncertainty would start with the collection of all relevant and avail- able information that might influence the decision-making process. This includes the information about the context factors presented in this paper. In principle, the themes presented in this paper can be used as an informal checklist by legal actors to ensure that they have considered all possible sources of uncertainty. Even though some of them might not be very novel for part of the readership, we believe that hav- ing them structured and organised and published, alongside pertinent discussions, is an important step for an informed and evidence-based practice within the family jus- tice system (Danser & Faith‐Slaker, 2019).26 Moreover, providing evidence is also important to provoke relevant changes and policy-making within organisations like the judiciary (Sanderson, 2002). Another strategy to handle uncertainty is to acknowledge and properly manage the sources of uncertainty. One cannot control or promote ‘harm reduction’ of what one does not know. Hence, legal actors cannot properly tackle uncertainty if they do not acknowledge it and how it can affect their decision-making process. In this sense, we believe this paper can promote awareness regarding the importance of acknowl- edging the uncertainty in child custody cases and, therefore, be able to select better courses of action that can avoid or handle risk factors (Lipshitz & Strauss, 1997), especially for the child’s interests and the family well-being. The ‘suppression strategy’ regards actions that either deny (e.g. ignoring or dis- torting information that is unwelcome) or rationalise the uncertainty within the deci- sion-making process. Our data suggest that this is a strategy invoked by some legal actors—e.g. CT1.2. We do not believe this is a good strategy to cope with uncer- tainty in child custody cases as this can lead to increasing uncertainty and, therefore, can put children and families in jeopardy. Instead, we believe that the best course of action is to acknowledge the sources of uncertainty (like the ones presented in this paper), map how they might affect the decision-making process in that specific case 26 Qualitative evidence is important for an evidence-based practice—See Sale and Thielke (2018). 1 3 Trends in Psychology and then, based on evidence-based practice, reduce uncertainty and make decisions that really are child-centred.27 Limitations and Future Directions This study’s design and the data gathered do not allow us to determine the optimal ways with which one can cope with uncertainty in child custody cases.28 They also do not allow us to properly approach the role of legal actors’ systems of beliefs in acknowledging and dealing with context factors and the uncertainty they produce. However, we believe this paper can help legal actors to understand how uncertainty in child custody cases constrains their performance and, thus, make them more aware of it—which is an important step in the tackling of uncertainty as mentioned before.29 Even though the results of this study make progress in understanding how con- text factors structure uncertainty in child custody cases, there are still processes that need to be investigated, such as how context factors are measured or weighed by legal actors when making a decision in a specific case and the role of ‘system of beliefs’—as mentioned. Also, future work should examine the strategies used to cope with uncertainty and whether there are optimal ways to cope with uncertainty in such cases, taking into account the child’s best interests. Final Considerations This paper allowed us, for the first time, under a ‘naturalistic decision-making’ approach, to identify and organise issues that shape uncertainty in child custody cases after parental separation. This is important, both to draw the attention of legal actors and academia to the role of the context in child custody cases and also to initi- ate research into ways of coping with uncertainty, aiming to avoid or diminish errors and biased judgments. We understand that the results presented in this paper not only further the knowledge in an underresearched field but they can also help legal 27 This is especially needed in Brazil, where family justice tends to adopt non-evidence based as well as ethically and scientifically questionable practices to mediate and solve conflicts/litigation within family courts—e.g. ‘systemic constellation work’ or ‘family constellation’: a mediumistic pseudo-psychother- apy imported from Germany without any sort of transcultural adaptation and/or scientific probe towards its efficacy within the family justice. 28 In the major study from which these results were extracted, we identified eight cognitive strategies used by legal actors to cope with uncertainty in child custody cases. Like context factors, we identified two domains for these strategies: (1) heuristics: strategic knowledge used to search the environment and set up shortcuts to make a decision; and (2) metacognition: referring to metacognitive knowledge that serves to monitor the decisions made and to make sure those decisions abide by the goal state. These domains are further explored by Mendes and Ormerod (2022). 29 The results from this study were also pivotal to helping us develop an experiment that might allow us further the discussion regarding ways to better cope with uncertainty and arrive at better decisions. It is a verbal protocol analysis based on a decision-making experiment with legal actors from Brazil and Eng- land. Currently, we are writing the results to then submit them for publication. 1 3 Trends in Psychology actors to be more aware of the sources of uncertainty in child custody cases that can impact their performance during the decision-making process. Supplementary Information The online version contains supplementary material available at https:// doi. org/ 10. 1007/ s43076- 022- 00253-9. Funding This study was funded by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior – Brasil (CAPES) – Finance Code 001, Ministry of Education, Brazil. Data Availability Data not available due to ethical restrictions.Due to the nature of this research, partici- pants of this study did not agree for their data (whole transcripts) to be shared publicly. However, some supporting material concerning the data analysis process will be available for both reviewers and readers. Declarations Ethics Approval This study was approved by University of Sussex’s Sciences & Technology C-REC under the Certificate of Approval ER/JA454/1. Informed Consent Informed consent was obtained from all individual participants included in the study. Consent for Publication All participants gave consent for their data to be used in publication. Conflict of Interest The authors declare no competing interests. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Com- mons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/. References Baker, K. K. (2012). Homogenous rules for heterogeneous families: The standardization of family law when there is no standard family. University of Illinois Law Review, 2012(2), 319–372. https:// doi. org/ 10. 2139/ ssrn. 17820 51 Barbosa, L. P. G., Mendes, J. A. A., & Juras, M. M. (2021). Dinâmicas disfuncionais, disputa de guarda e alegações de alienação parental: uma compreensão sistêmica. Nova Perspectiva Sistêmica, 30(69), 6–18. https:// doi. org/ 10. 38034/ nps. v30i69. 612 Barnett, A. (2020). A genealogy of hostility: Parental alienation in England and Wales. Journal of Social Welfare and Family Law, 42(1), 18–29. https:// doi. org/ 10. 1080/ 09649 069. 2019. 17019 21 Ben-Haim, Y. (2019). Info-Gap Decision Theory (IG). In V. A. Marchau, W. E. Walker, P. J. Bloemen, & S. W. Popper (Eds.), Decision making under deep uncertainty: From theory to practice (pp. 93–116). Springer Nature. Boyatzis, R. E. (1998). Transforming qualitative information: Thematic analysis and code development. Sage Publications. Braun, V., & Clarke, V. (2006). Using thematic analysis in psychology. Qualitative Research in Psychol- ogy, 3(2), 77–101. https:// doi. org/ 10. 1191/ 14780 88706 qp063 oa Braun, V., & Clarke, V. (2013). Successful qualitative research: A practical guide for beginners. Sage Publications. Braun, V., & Clarke, V. (2022a). Thematic Analysis: A practical guide. Sage. 1 3 Trends in Psychology Braun, V., & Clarke, V. (2022b). Toward good practice in thematic analysis: Avoiding common problems and be(com)ing a knowing researcher. International Journal of Transgender Health, 1–6. https:// doi. org/ 10. 1080/ 26895 269. 2022. 21295 97 Braun, V., Clarke, V., Hayfield, N., & Terry, G. (2019). Thematic analysis. In P. Liamputtong (Ed.), Handbook of research methods in health social sciences (pp. 843–860). Springer. Bruch, C. S. (2001). Parental alienation syndrome and parental alienation: Getting it wrong in child cus- tody cases. Family Law Quarterly, 35, 527. Charmaz, K. (2014). Constructing grounded theory. Sage Publications. Creswell, J. W., & Poth, C. N. (2017). Qualitative inquiry and research design: Choosing among five approaches. Sage Publications. Danser, R., & Faith-Slaker, A. (2019). Moving from evidence-light to evidence-based practice: Randomized control trials in family law. Family Court Review, 57(4), 491–500. https:// doi. org/ 10. 1111/ fcre. 12442 Darawsheh, W. (2014). Reflexivity in research: Promoting rigour, reliability and validity in qualitative research. International Journal of Therapy and Rehabilitation, 21(12), 560–568. https:// doi. org/ 10. 12968/ ijtr. 2014. 21. 12. 560 Eekelaar, J. (2015). The role of the best interests principle in decisions affecting children and decisions about children. International Journal of Children’s Rights, 23(1), 3–26. https:// doi. org/ 10. 1163/ 15718 182- 02301 003 Eekelaar, J., & Maclean, M. (2013). Family justice: The work of family judges in uncertain times. Blooms- bury Publishing. Emery, R. E. (2012). Renegotiating family relationships: Divorce, child custody, and mediation. The Guil- ford Press. Enosh, G., & Bayer-Topilsky, T. (2015). Reasoning and bias: Heuristics in safety assessment and placement decisions for children at risk. The British Journal of Social Work, 45(6), 1771–1787. https:// doi. org/ 10. 1093/ bjsw/ bct213 Flick, U., Kardoff, E., & Steinke, I. (2004). A companion to qualitative research. Sage Publications. Goldstein, M. (2016). Best interest factors in child custody evaluations. In M. Goldstein (Ed.), Handbook of Child Custody (pp. 11–15). Springer. https:// doi. org/ 10. 1007/ 978-3- 319- 13942-5_2 González Rey, F. L. (2011). Pesquisa qualitativa em Psicologia: Caminhos e desafios. Cengage Learning. Guest, G., MacQueen, K. M., & Namey, E. E. (2012). Applied thematic analysis. Sage Publications. https:// doi. org/ 10. 4135/ 97814 83384 436 Jones, R. N., Patwardhan, A., Cohen, S. J., Dessai, S., Lammel, A., Lempert, R. J., et al. (2014). Foundations for decision making. In C. B. Field, V. R. Barros, D. J. Dokken, K. J. Mach, M. D. Mastrandrea, & T. E. Bilir (Eds.), Climate Change 2014: Impacts, Adaptation, and Vulnerability. Part A: Global and Secto- ral Aspects. Contribution of Working Group II to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change (pp. 195–228). Cambridge University Press. https:// doi. org/ 10. 1017/ CBO97 81107 415379. 007 Juras, M. M., & Costa, L. F. (2017). He was neither a good father nor a good husband: Marital and parental roles in low-income separated families. Psicologia: Teoria e Pesquisa, 32(5), 1–9. https:// doi. org/ 10. 1590/ 0102- 3772e 32ne2 15 Kelly, J. B. (2007). Children’s living arrangements following separation and divorce: Insights from empirical and clinical research. Family Process, 46(1), 35–52. https:// doi. org/ 10. 1111/j. 1545- 5300. 2006. 00190.x Klein, G., Orasanu, J., Calderwood, R., & Zsambok, C. E. (Eds.). (1993). Decision-making in action: models and methods. Ablex Publishing Corporation. Lang, M. (2017). Legal uncertainty as a welfare enhancing screen. European Economic Review, 91, 274– 289. https:// doi. org/ 10. 1016/j. euroe corev. 2016. 10. 007 Lipshitz, R. (1993a). Converging themes in the study of decision making in realistic settings. In G. A. Klein, J. Orasanu, R. Calderwood, & C. E. Zsambok (Eds.), Decision-making in action: models and methods (pp. 103–137). Ablex Publishing Corporation. Lipshitz, R. (1993b). Decision making as argument-driven action. In G. A. Klein, J. Orasanu, R. Calder- wood, & C. E. Zsambok (Eds.), Decision-making in action: models and methods (pp. 172–181). Ablex Publishing Corporation. Lipshitz, R., & Strauss, O. (1997). Coping with uncertainty: A naturalistic decision-making analysis. Organi- zational Behavior and Human Decision Processes, 69(2), 149–163. https:// doi. org/ 10. 1006/ obhd. 1997. 2679 Lipshitz, R., Klein, G., Orasanu, J., & Salas, E. (2001). Taking stock of naturalistic decision-making. Journal of Behavioral Decision Making, 14(5), 331–352. https:// doi. org/ 10. 1002/ bdm. 381 1 3 Trends in Psychology Mackenzie, D., Herbert, R., & Robertson, N. (2020). ‘It’s not OK’, but ‘It’ never happened: Parental aliena- tion accusations undermine children’s safety in the New Zealand family court. Journal of Social Wel- fare and Family Law, 42(1), 106–117. https:// doi. org/ 10. 1080/ 09649 069. 2020. 17019 42 Marchau, V. A. W. J., Walker, W. E., Bloemen, P. J. T. M., & Popper, S. W. (2019). Introduction. In V. A. Marchau, W. E. Walker, P. J. Bloemen, & S. W. Popper (Eds.), Decision making under deep uncer- tainty: From theory to practice (pp. 1–20). Springer Nature. Maturana, H. R., & Varela, F. J. (1991). Autopoiesis and cognition: The realization of the living (Vol. 42). Springer Science & Business Media. McGoldrick, M., Carter, B. A., & Preto, N. A. G. (2014). The expanding family life cycle: Individual, family, and social perspectives. Pearson. Meier, J. S. (2020). US child custody outcomes in cases involving parental alienation and abuse allegations: What do the data show? Journal of Social Welfare and Family Law, 42(1), 92–105. https:// doi. org/ 10. 1080/ 09649 069. 2020. 17019 41 Mendes, J. A. A. (2022). The decision-making process in child custody cases after parental separation: a cross-cultural study between Brazil and England [unpublished doctoral dissertation]. University of Sussex. Mendes, J. A. A., & Bucher-Maluschke, J. S. N. F. (2017). Destructive divorce in the Family Life Cycle and its implications: criticisms of Parental Alienation. Psicologia: Teoria e Pesquisa, 33(2), 1–8. https:// doi. org/ 10. 1590/ 0102. 3772e 33423 Mendes, J. A. A., & Ormerod, T. (2019). The best interests of the child: An integrative review of English and Portuguese literatures. Psicologia em Estudo, 24, 1–22. https:// doi. org/ 10. 4025/ psico lestud. v24i0. 45021 Mendes, J. A. A., & Ormerod, T. (2021). A comparative look at divorce, laws and the best interests of the child after parental separation in Brazil and England. Revista da Faculdade de Direito UFPR, 66(2), 95–126. https:// doi. org/ 10. 5380/ rfduf pr. v66i2. 74001 Mendes, J. A. A., & Ormerod, T. (2022). Making sense out of uncertainty: cognitive strategies in child cus- tody decision-making process [Unpublished manuscript]. School of Psychology, University of Sussex. Mosten, F. S., & Traum, L. (2017). The family lawyer’s role in preventive legal and conflict wellness. Family Court Review, 55(1), 26–37. https:// doi. org/ 10. 1111/ fcre. 12260 Neilson, L. C. (2018). Parental alienation empirical analysis: child best interests or parental rights? FREDA Centre for Research on Violence Against Women and Children. Nowell, L. S., Norris, J. M., White, D. E., & Moules, N. J. (2017). Thematic analysis: Striving to meet the trustworthiness criteria. International Journal of Qualitative Methods, 16(1), 1–13. https:// doi. org/ 10. 1177/ 16094 06917 733847 O’Neill, A. T., Bussey, K., Lennings, C. J., & Seidler, K. M. (2018). The views of psychologists, lawyers, and judges on key components and the quality of child custody evaluations in Australia. Family Court Review, 56(1), 64–78. https:// doi. org/ 10. 1111/ fcre. 12323 Patterson, E. S., Militello, L. G., Su, G., & Sarkar, U. (2016). Characterizing a naturalistic decision-making phenomenon: Loss of system resilience associated with implementation of new technology. Journal of Cognitive Engineering and Decision Making, 10(3), 229–243. https:// doi. org/ 10. 1177/ 15553 43416 652524 Pepiton, M. B., Alvis, L. J., Allen, K., & Logid, G. (2012). Is parental alienation disorder a valid concept? Not according to scientific evidence. A review of parental alienation, DSM-5 and ICD-11 by William Bernet. Journal of Child Sexual Abuse, 21(2), 244–253. Sá, S. D., Werlang, B. S. G., & Paranhos, M. E. (2008). Intervenção em crise. Revista Brasileira de Terapias Cognitivas, 4(1), 1–10. Sadler, G. R., Lee, H. C., Lim, R. S. H., & Fullerton, J. (2010). Recruitment of hard-to-reach population sub- groups via adaptations of the snowball sampling strategy. Nursing & Health Sciences, 12(3), 369–374. Sale, J. E., & Thielke, S. (2018). Qualitative research is a fundamental scientific process. Journal of Clinical Epidemiology, 102, 129–133. https:// doi. org/ 10. 1016/j. jclin epi. 2018. 04. 024 Sanderson, I. (2002). Evaluation, policy learning and evidence-based policy making. Public Administration, 80(1), 1–22. https:// doi. org/ 10. 1111/ 1467- 9299. 00292 Stamps, L. E., Kunen, S., & Lawyer, R. (1996). Judicial attitudes regarding custody and visitation issues. Journal of Divorce & Remarriage, 25(1–2), 23–37. https:// doi. org/ 10. 1300/ J087v 25n01_ 02 Tracy, S. J. (2010). Qualitative quality: Eight “big-tent” criteria for excellent qualitative research. Qualitative Inquiry, 16(10), 837–851. https:// doi. org/ 10. 1177/ 10778 00410 383121 1 3 Trends in Psychology Urquhart, C. (2013). Grounded theory for qualitative research: A practical guide. Sage Publications. Von Foerster, H. (2003). Understanding understanding: Essays on cybernetics and cognition. Springer. Wallace, S. R., & Koerner, S. S. (2003). Influence of child and family factors on judicial decisions in con- tested custody cases. Family Relations, 52(2), 180–188. Wells, G. L. (1978). Applied eyewitness-testimony research: System variables and estimator variables. Jour- nal of Personality and Social Psychology, 36(12), 1546. Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. 1 3
10.1111_eva.13529
Received: 15 July 2022 |  Accepted: 27 December 2022 DOI: 10.1111/eva.13529 O R I G I N A L A R T I C L E Selecting for infectivity across metapopulations can increase virulence in the social microbe Bacillus thuringiensis Tatiana Dimitriu1  | Wided Souissi2 | Peter Morwool1 | Alistair Darby3  | Neil Crickmore2  | Ben Raymond1 1Centre for Ecology and Conservation, University of Exeter, Penryn, UK 2School of Life Sciences, University of Sussex, Brighton, UK 3Centre for Genomic Research, Institute of Integrative Biology, University of Liverpool, Liverpool, UK Correspondence Ben Raymond, Centre for Ecology and Conservation, University of Exeter, Penryn Campus, Treliever Road, Penryn TR10 9FE, UK. Email: b.raymond@exeter.ac.uk Funding information Biotechnology and Biological Sciences Research Council, Grant/Award Number: BB/S002928/1; Leverhulme Trust, Grant/ Award Number: RPG- 2014- 252 Abstract Passage experiments that sequentially infect hosts with parasites have long been used to manipulate virulence. However, for many invertebrate pathogens, passage has been applied naively without a full theoretical understanding of how best to select for increased virulence and this has led to very mixed results. Understanding the evolution of virulence is complex because selection on parasites occurs across multiple spatial scales with potentially different conflicts operating on parasites with different life histories. For example, in social microbes, strong selection on replication rate within hosts can lead to cheating and loss of virulence, because investment in public goods virulence reduces replication rate. In this study, we tested how varying mutation supply and selection for infectivity or pathogen yield (population size in hosts) affected the evolution of virulence against resistant hosts in the specialist insect pathogen Bacillus thuringiensis, aiming to optimize methods for strain improvement against a difficult to kill insect target. We show that selection for infectivity using competition between subpopulations in a metapopulation prevents social cheating, acts to retain key virulence plasmids, and facilitates increased virulence. Increased virulence was associated with reduced efficiency of sporulation, and possible loss of function in putative regulatory genes but not with altered expression of the primary virulence factors. Selection in a metapopulation provides a broadly applicable tool for improving the efficacy of biocontrol agents. Moreover, a structured host population can facilitate artificial selection on infectivity, while selection on life- history traits such as faster replication or larger population sizes can reduce virulence in social microbes. K E Y W O R D S Bacillus thuringiensis, directed evolution, evolution of virulence, mutators, public goods, social evolution This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. © 2023 The Authors. Evolutionary Applications published by John Wiley & Sons Ltd. Evolutionary Applications. 2023;16:705–720. wileyonlinelibrary.com/journal/eva  |  705 706 | 1  |  I NTRO D U C TI O N have low fitness in clonal infections (Granato et al., 2018; Pollitt et al., 2014; Rumbaugh et al., 2012). In addition, increased genetic Passage, the repeated infection and re- isolation of a microbe in a diversity within infections (low relatedness) will also tend to favor host, has been used as a tool for the manipulation of parasite vir- fast- replicating genotypes such as cheaters. These social biology ulence for decades, as well as a means of testing evolution of vir- concepts are relevant both for clinically important microbes and for ulence theory (Ebert, 1998; Raymond & Erdos, 2022). Passage has biocontrol agents. Experimental evolution with entomopathogenic been used as a means of producing attenuated live vaccines: se- nematodes, for instance, shows increasing opportunities for cheat- quential infection of animal tissue cultures can lead to loss of viru- ing can lead to attenuation and extinction and may explain historical lence in human hosts and has been used to produce polio and yellow problems with unstable virulence in laboratory culture of these par- fever vaccines among others (Barrett, 2017; Sabin & Boulger, 1973). asites (Shapiro- Ilan & Raymond, 2016). Conversely, adaptation of viruses to a particular host via passage can In this study, we aimed to apply social evolution theory to increase lead to increased virulence (Ebert, 1998). Although there are com- the virulence of the Gram- positive invertebrate pathogen, Bacillus plex relationships between virulence and fitness in natural popula- thuringiensis (Bt). Bt is a valuable model for testing novel passage tions (Alizon et al., 2009; Frank, 1996; Gandon et al., 2001), artificial regimes partly because its virulence factors are extremely well char- inoculation means that many of the negative consequences of high acterized (Adang et al., 2014) and because its social biology is well virulence in terms of reducing opportunities for transmission will studied (Cornforth et al., 2015; Raymond et al., 2012; van Leeuwen not operate in laboratory conditions and so it may be possible to et al., 2015; Zhou et al., 2014). Bt obligately requires pore- forming increase the virulence of naturally occurring pathogens via artificial toxins for infection. These are produced in the form of crystalline selection. bodies at sporulation in the cadaver but solubilized in the midgut Increasing pathogen virulence via passage has long been a goal after ingestion; the toxins disintegrate the midgut epithelium and in biocontrol research as researchers have sought to manipulate the allow these microbes access to the hemocoel (Adang et al., 2014). In efficacy and/or host range of microbes that are potential biocontrol some insect hosts, the production of quorum- regulated phospholi- agents (Raymond & Erdos, 2022). Naïve passage designs, in which pases and cellulases complements the action of the crystal toxins and the aim is simple infection and re- infection, without any other se- facilitates host invasion (Salamitou et al., 2000; Zhou et al., 2014). A lection pressure or modifications, can increase the virulence of range of other virulence factors, for example, vegetative insecticidal baculoviruses in terms of reducing doses required to kill 50% of in- proteins (Vips), are produced by most Bt genotypes, although their sects (LD50; Berling et al., 2009; Kolodny- Hirsch & Van Beek, 1997; Maleki- Milani, 1978). Simple passage of baculoviruses can increase contribution to pathogenicity is less clear (Chakroun et al., 2016). The entomocidal toxins of Bt have high selective potency against virulence partly because of the high genetic diversity found in nat- insect pests and this is the main reason why Bt dominates the mi- ural populations (Shapiro et al., 1992; Thézé et al., 2014). However, crobial biocontrol market and supplies the vast majority of insecti- another reason naïve passage can be effective is if reproductive rate cidal toxins for genetically modified crops (Bravo et al., 2011). There within hosts is positively correlated with virulence, as is assumed by is considerable interest in identifying mutants or proteins that can many classical models of virulence (Alizon et al., 2009; Frank, 1996; overcome host resistance in Bt (Badran et al., 2016). From a theo- Gandon et al., 2001). Even in the simplest passage design, there retical point of view, it may also make sense to target passage ex- will be competition between genetic variants within hosts, and we periments at hosts that are resistant but still vulnerable to some would expect that this within- host competition would favor higher level of infection. Based on fitness landscape theory, we expected replication rate, ignoring issues such as antagonism and evasion or that rapid, experimentally tractable evolution would be more likely manipulation of immunity (Massey et al., 2004; Raberg et al., 2006). in a pathogen of low fitness rather than one already at an adap- This means that simple isolation and re- infection can select for the tive peak (Poelwijk et al., 2007). We used a host species commonly genotypes which grow faster within hosts which can produce the targeted by Bt products: the diamondback moth Plutella xylostella, net result of an increase in virulence without the need for any addi- using a well- characterized genotype with a high level of resistance tional selection pressure. to B. thuringiensis kurstaki (Figure 1a). Moreover, for social microbes, Nevertheless, research on the social biology of microorganisms the selection pressure acting to maintain virulence does not come emphasizes that increased replication rate does not necessarily lead from within- host competition for rapid growth— it has to come from to increased virulence (Buckling & Brockhurst, 2008; Raymond & competition between populations in terms of total population size Erdos, 2022). Many bacterial pathogens, for instance, invest con- (yield; Griffin et al., 2004) or number of hosts infected (Raymond siderable resources in producing virulence factors such as toxins et al., 2012). This is because the fitness benefits (or public goods) or siderophores that are important for infecting hosts or accessing provided by cooperation in known social pathogens either increase host resources such as iron (Diard, Garcia, et al., 2014; Raymond the efficient use of host resource or, in the case of the crystal toxins, et al., 2012; West & Buckling, 2003). Investing in these resources provide access to host tissues. can slow replication rates, this means that intense within- host com- In this study, we conducted experimental evolution using cycles petition can select for cheaters which can outcompete virulent of Bt infection in resistant hosts alternating with spore production genotypes in mixed infections, although cheaters are expected to in vitro (Figure 1b). We tested passage regimes that would maximize DIMITRIU et al. (a) 1 0.8 0.6 0.4 0.2 y t i l a t r o m 0 1 susceptible resistant (c) (d) Yield selection Infectivity selection     |  707 100 dose (spores / 10000 l) 100000 Exclude low mortality groups (b) inoculation spore wash early infection ancestor in vivo selection in vitro control 5 passage rounds spore wash sporulation + antibiotic Pool all cadavers High mortality groups inoculate two groups F I G U R E 1 Experimental design to increase the virulence of Bacillus thuringiensis against resistant insects. (a) Bioassays using the Bt ancestor 7.1.o show substantial differences in mortality between our focal Bt- resistant insect VL- FR and its near- isogenic Bt susceptible counterpart VL- SS (F1,5 = 34.1, p = 0.0021). The LC50 of the susceptible insects was estimated at 7.4 CFU/μl (95% confidence limits 6.9– 7.9 CFU/μl), while the LC50 of the VL- FR resistant was 5 × 105 CFU/μl (Table S1). Selection experiments began with a dose that would kill 20%– 30% of VL- FR insects. (b) Selection cycles performed separately for strains with mutator or wild- type mutation rates. For in vivo selection treatments, cells selected from killed larvae were inoculated into sporulation medium. In vitro growth, sporulation, and purification steps were common to the in vivo selection and in vitro control. To prevent the invasion of cheaters not contributing to cooperative virulence, two in vivo selection treatments were compared. (c) In the yield selection treatment, all cadavers from a replicate lineage were pooled and inoculated together for spore production. (d) Under infectivity selection, half of the subpopulations, those causing the lowest mortality, are terminated at the end of each round of selection, while spores from the remaining subpopulations are divided and used to infect two subpopulations in the next round of infection. selection based on yield (population size within hosts) or infectivity. in terms of preserving population structure (Gardner & West, 2006; Selecting for yield involved combining all cadavers from a replicate Kümmerli et al., 2009). It should be emphasized that these selection into a single inoculum pool at each passage, so that genotypes with treatments are not exclusive or perfect; in other words, we cannot higher yield are better represented in the next round of infections prevent some selection for infectivity in the yield treatment (infec- (Griffin et al., 2004; Shapiro- Ilan & Raymond, 2016; Figure 1c). This tions have to happen) nor some selection for yield in the infectivity type of between- host competition has previously produced modest treatment (bacteria have to grow in cadavers), but we are attempt- increases in virulence in Bt (Garbutt et al., 2011). Pooling subpop- ing to maximize different types of selection with our experimental ulations also mimic a high pathogen dispersal rate within natural treatments. populations (Kümmerli et al., 2009), which is plausible for Bacillus Finally, we tested how increasing the mutation supply would (Pearce et al., 2009). The second selection regime involved directly affect virulence evolution, a common approach in directed evolu- selecting for infectivity. To do this, we imposed competition be- tion (Selifonova et al., 2001). Bacterial clones with elevated mu- tween subpopulations within a metapopulation (each metapopu- tation rates and defective proofreading genes, or mutators, are lation constituting an independent replicate), so that only the 50% prevalent in pathogenic species (Taddei et al., 1997), and we in- most infectious subpopulations are used to initiate the next round creased mutation supply by approximately 25- fold by carrying out of infection (Figure 1d). The most infectious subpopulations are selection in a mutator, made by disruption of the mutS gene, and divided and initiate two new pathogen subpopulations in the next in lineages using the wild- type ancestor. At the end of passage round of infection. The infectivity selection treatment, therefore, experiments, we bio- assayed for changes in virulence in evolved uses a form of budding dispersal and so may have additional benefits lineages (diverse independently evolved populations), as well as DIMITRIU et al. 708 | clones isolated from those lineages. We also measured changes in life history (spore production in vitro, competitive fitness) and was calculated with fluctuation assays. Thirty- two independent cultures per strain were inoculated into sporulation medium by 107- tested whether social conflicts might have affected experimental fold dilution from Luria Broth (LB) overnight cultures. After 24 h, outcomes. Finally, in order to explore the possible mechanisms for the observed increase in virulence in evolved clones, we explored cultures were plated on LB agar (LBA) plates containing 100 μg/ml rifampicin. Mutation rates and confidence intervals were calculated if changes in the expression of known virulence factors could ex- with the Ma- Sandri- Sarkar (MSS)- maximum likelihood method using plain results and conducted whole genome sequencing on a selec- FALCOR (Hall et al., 2009). tion of evolved clones. 2  |  M E TH O D S 2.1  |  Insects and insect rearing The Cry1Ac- resistant line “VL- FR” of the diamondback moth Spores and crystals of Bacillus thuringiensis (Bt) for use in se- lection experiments or bioassays were grown in sporulation media (HCO; Lecadet et al., 1980) containing polymyxin (100 IU/ml) (Oxoid) for 72 h at 30°C with shaking at 150 rpm. Selective antibiotics (10 μg/ml Erm for wt; 5 μg/ml Chl for mut strain) was used in all cul- tures except for the production of spores for bioassays which used antibiotic free culture. When grown from insect cadavers, 30 μg/ml Erm was used to inhibit growth of Gram- negative bacteria. Standard P. xylostella used for experiments was previously derived from a spore production conditions used 1 ml of HCO in 24- well growth cross of line NOQA- GE with the diet adapted susceptible line Vero plates. Rows of inoculated wells from the same line of selection were Beach: This line was reared and selected for resistance to Cry1Ac as alternated with empty rows to prevent and control for contamina- described previously (Zhou et al., 2018) and can survive exposure to ~104- fold higher doses of Bt compared to the susceptible line (Figure 1a). We established a population that was fixed for tion between lines. Initial cultures of ancestors at passage 1 used 100 ml cultures in 500 ml flasks. All other routine culture of Bt used overnight growth in LB or on LBA plates at 30°C with antibiotic se- resistance using a PCR screen of resistance alleles in the parental lection unless otherwise stated. population. An insect strain with similar genetic background (also derived from a NOQA- GE X Vero Beach cross) was established from F2 offspring but in this case fixed for susceptibility to Cry1Ac, this 2.3  |  Passage experiment line was denoted “VL- SS” (Zhou et al., 2018). Diamondback moth larvae used for selection were reared on autoclaved artificial diet Selection was performed with four replicate lineages per strain for (Baxter et al., 2011) without supplementary antibiotics until third each treatment combination, using a factorial design with two strains instar. All insects were reared in a controlled temperature facility at with diffferent mutation rates (wt and mut) and three treatments the University of Exeter's Cornwall campus. (yield, infectivity, and in vitro control), as detailed below. In total, five 2.2  |  Bacteria, plasmids, strain construction, and in vitro growth conditions rounds of selection were performed. For infection, sterile diet was cut into quarters in 55 ml dishes, and each quarter was inoculated with 90 μl of spore suspension containing 5 × 104 to 105 spores/μl (chosen to produce 25%– 30% mortality over 2– 4 days for resistant diamondback moth, Figure 1a) The ancestor Bt kurstaki strain was obtained by transforming the and dried. Ten third- instar VL- FR larvae were then added per 55 ml original Bt 7.1.o field isolate (Raymond et al., 2009) with the pHT315- dish. Insect death was recorded daily from 2 days after infection. gfp plasmid containing the ermB gene conferring erythromycin (Erm) When total mortality reached 25%– 30%, the earliest cadavers resistance (Zhou et al., 2014). The mutator (mut) strain was obtained were transferred to microcentrifuge tubes with sterile tooth- by inactivating the mutS gene in 7.1.o via disruption. The cat gene picks (Figure 1b). After 2 days at room temperature, 0.85% NaCl conferring chloramphenicol (Chl) resistance was digested from the solution (saline) was added to the tubes. Cadavers were homog- pAB2 plasmid using KpnI (Bravo et al., 1996); this fragment was enized with pellet pestles, briefly vortexed and the suspension inserted at position 1439 of the mutS- coding sequence. The sequence was transferred to sporulation media (Figure 1b). When growth between positions 601 and 1837 of mutS gene coding sequence with the cat insert was synthesized in pUC57 from GenScript after was complete, 200- μl spore aliquots were transferred to 96- well PCR plates with aluminum sealing film for quantification, infec- adding BamHI sites on both ends. This synthesized BamHI fragment tion, and storage. Plates were centrifuged at 3000 g for 15 min, was cloned into the thermosensitive vector pRN1501 (Lereclus the supernatant was removed, and spores were resuspended in et al., 1992). The resulting plasmid was introduced into Bt 7.1.o, and clones with recombination of the interrupted mutS sequence into 100 μl 0.85% NaCl. Plates containing spores to be used in next infection step were pasteurized (20 min, 65°C) and kept at 10°C the chromosome were obtained after growth at 42°C by screening until use, the others were stored at −20°C. Pasteurization pre- for Chl resistance and loss of Erm resistance (Lereclus et al., 1992). vents contamination from gut microbes and also selects against Insertion was confirmed by PCR and sequencing of mutS locus. The fast- growing asporogenic mutants. Spore density was estimated mutation rate of the ancestor with wild- type mutS (wt) and mut strain by measuring OD600 nm in a 96- well microplate reader of 10- fold DIMITRIU et al.     |  709 dilutions. A subset of the samples was plated at appropriate dilu- Evolved lineages are genetically diverse populations, which can tions, in order to visually check for contamination and calibrate present challenges when repeating experiments or when attempt- OD600 nm measurements. ing to link genotypes to phenotypes. Clones were therefore isolated Selection of cadavers and inoculation into sporulation medium from lineage samples by streaking single colonies three times on LBA depended on experimental treatment (Figure 1). The in vitro control and we conducted bioassays on both clones and lineage after selec- selection treatment culture and washing steps were performed as tion was complete. Clone nomenclature in figures and tables indi- above but without infection of insects (Figure 1b) and was similarly cates strain (W for wild- type mutation rate, M for mutator), selection performed on four lineages per strain (wt and mut) with five rounds of selection. Approximately 5 × 104 spores were used to inoculate 1 ml sporulation medium and cultured as above. After standard cul- treatment (y for yield, i for infectivity selected, v for in vitro control) with a letter for independent lineage and a number for clone, for example, Wia1, wild type, infectivity selected, lineage a, clone 1. ture, spores were centrifuged and resuspended in 0.85% NaCl, pas- The viable spore concentration of all cultures used in bioassays teurized and kept at 10°C until use. was measured by plating after pasteurization and measuring colony For the yield selection treatment, each lineage was split into six insect subpopulations using six dishes, which were pooled together at each round of selection. Cadavers from all six subpopulations in the pooled treatment were transferred to the same microfuge tube. forming units (CFUs). Standard virulence bioassays used at least three doses with 60 insects per dose (between 104 and 1.2 × 105 spores/μl for Bt kurstaki in resistant P. xylostella) and were repeated at least twice. The exception to this was the initial screens of clone Then, 300 μl saline was added and the suspension was distributed into six wells for inoculation. For replicate lineages with more than level variation of six clones per lineage which used 15– 30 insects per dose and two doses. Only strongly melanized cadavers were scored 30% death across insect populations at time of selection, the num- as Bt- killed insects. The dose response in bioassays of clones was ber of cadavers taken from subpopulations with most deaths was analyzed in terms of viable spores but also in terms of dilution factor reduced in order to select 30% cadavers overall and maintain com- of inocula relative to purified spore stocks as some clones displayed parable population sizes across treatments at inoculation (Figure 1c). variable germination rates. Clones of interest were further charac- For the infectivity treatment, each lineage was split into 12 in- terized in additional assays. sect subpopulations maintained separately in 12 dishes. The six sub- Viable spore counts of clones grown in sporulation media populations with highest larval mortality were selected and the six for bioassays were used in analyses of in vitro spore production. other subpopulations were discarded (Figure 1d). If the same mor- Assessment of bacterial development rate used three clones with tality was observed for several populations, the ones with earliest elevated virulence, three clones with virulence unchanged, as well observed deaths were selected. A maximum of three cadavers from the wild- type ancestor. The proportion of cells that were vegetative, selected subpopulations were transferred into separate tubes (6 the proportion of cells that had toxin crystals within the exosporium total), then 50 μl saline was added to each tube, and the suspension from each tube was inoculated into separate wells. After spore puri- and proportion of cells at the final stage of lysis of exosporium were assessed by phase- contrast microscopy after 72 and 144 h of growth fication, spores from each well were used to inoculate two subpop- in HCO. ulations in the next selection round (Figure 1d). 2.4  |  Phenotypic and genotypic characterization of evolved bacteria 2.4.1  |  Bioassays 2.5  |  Genetic analysis of Cry toxin gene complements The ancestor Bt 7.1.o wild type bears genes for several expressed toxins (Cry1Ac, Cry2Aa, Cry1Aa) on a large plasmid (homologous to CP009999, 317 kb) and one additional toxin gene (Cry1Ab) on After five rounds of selection, the evolved lineages were frozen at a smaller plasmid (homologous to CP010003, 69 kb; Tables S2 and −80°C (LB 20% glycerol) without pasteurization. These frozen stocks S3; Méric et al., 2018). Loss of each of those plasmids was detected were used to inoculate sporulation medium for assays of evolved in several clones with whole- genome sequencing. We extended lineages. The virulence of each independently evolved lineage those results by estimating Cry toxin plasmid loss for all clones from (Figure 2) was assayed by scoring proportional mortality (Figure 2) evolved lines by PCR (six clones from each insect- evolved lineage after 3 days (early mortality) and 5 days (late mortality). This analysis and two clones from each in vitro lineage). Primers were designed used data from two independent bioassays using independently and validated with the sequenced clones. Primers Cry2- F and Cry2- R grown spore stocks (minimizing effects due to variation in growth/ amplify a 1.1- kb product from the large toxin- encoding plasmid amplification in vitro). Analyses reported in Figure 2a used average (CP009999). Primers Cry1- F and Cry1- R amplify a 1.4- kb product mortality data that were normalized against the ancestor: The when any Cry1A gene is present; no amplification corresponds to average mortality (across the three doses) of each evolved lineage loss of both cry gene- bearing plasmids. replicate was divided by the average mortality measured in the PCR used HotStar Taq (Qiagen) with cycling parameters: 5 min at ancestor within the same assay. 95°C, 30 × (45 s at 94°C, 1 min at 50°C [Cry2 primers] or 53°C [Cry1 DIMITRIU et al. 710 | (a) d e z i i l a m r o n 3 y a d y t i l a t r o m d e z i i l a m r o n 5 y a d y t i l a t r o m 8 6 4 2 0 3 2 1 0 *** ** * strain wt mut infectivity yield control *** ** n o i t r o p o r p 1.0 0.5 0.0 n o i t r o p o r p 1.0 0.5 0.0 (b) ) t i g o l ( y t i l a t r o m 3 y a d 6 3 0 −3 −6 6 3 0 −3 −6 infectivity yield control n o i t r o p o r p 1.0 0.5 0.0 n o i t r o p o r p 1.0 0.5 0.0 a b c d lineage a b c d lineage a b c d lineage m u t n o i t r o p o r p 1.0 0.5 0.0 n o i t r o p o r p 1.0 0.5 0.0 a b c d lineage a b c d lineage a b c d lineage w t Cry toxin genes All Cry1Ab only none infectivity yield control selection treatment 4.0 4.4 4.8 5.2 4.4 4.0 5.2 dose (log10 CFU / µL) 4.8 4.0 4.4 4.8 5.2 F I G U R E 2 Mutation rate and infectivity selection shape evolution of virulence and retention of public good virulence factors (cry toxins). Data shown are from two separate bioassays performed with independent spore preparations of evolved lineages after five rounds of selection. (a) Proportional mortality (relative to the ancestor) at 3 and 5 days after inoculation for different treatments, control refers to in vitro passaged controls. The center value of the boxplots shows the median, boxes the first and third quartile, and whiskers represent 1.5 times the interquartile range; outliers are shown as dots (N = 8, 4 lineages per treatment in two assays). (b) Barplots show the proportion of clones containing no toxin genes (light gray), Cry1Ab only (dark gray), or both Cry1A and Cry2A (black) genes for each lineage (N = 6 clones per lineage). Scatter plots show logit- transformed mortality 3 days after inoculation is shown as a function of Bt dose; lines are fitted models for each lineage. For reference, the dotted black line shows the dose response for the ancestor, the replicate lineages within each treatment are color coded. See also Table S1 and Figures S1 and S2 for results of clone- level virulence assays. primers], 90 s at 72°C), 10 min at 72°C. To provide DNA templates, were sequenced. DNA was extracted using Qiagen DNeasy Blood fresh colonies were resuspended into 100 μl dH2O, frozen 20 min at −80°C, then boiled 10 min. Five microliters of supernatant was used and Tissue genomic extraction kits, after appropriate lysis for Gram- positive bacteria. Sequencing was performed on an Illumina in each 20 μl PCR mix. To exclude false- negative results, each nega- tive result was repeated three times from fresh colonies. HiSeq 2500 at a minimum of 30× coverage with 125 bp paired- end sequencing. 2.6  |  Whole- genome sequencing All Illumina reads were trimmed with Trimmomatic (Bolger et al., 2014). Unicycler version 0.4.7 was used to perform a hybrid assembly using PacBio self- corrected reads and Illumina reads of the mutator ancestor (Wick et al., 2017). Since most evolutionary The unmarked ancestral Bt kurstaki 7.1.o strain was sequenced change occurred in the mutator lineages, the mutator ancestor pro- with PacBio after DNA extraction using the Qiagen high- molecular vided a better reference for subsequent identification of mutations. weight kit. Data were produced using SMRTbell® Express Template This combination of read types resulted in a final assembly that Prep Kit 2.0 following the manufacturer's recommendations. This included small plasmids that were otherwise lost from the PacBio resulted in a 15- to 20- kb insert library which was sequenced on a data- only assemblies, due to the DNA fragment size section being PacBio Sequel system using one cell per library and 10- h sequencing larger than the size of the plasmids. Short read data were used to movie time. The data were processed to provide CCS and single pass validate and aid the comparison of plasmid genomes between a ref- data and assembled using Unicycler version 0.4.7 (Wick et al., 2017). erence (Bt HD- 1) and Bt 7.1.o (this study) by mapping the short reads For insect- evolved lineages, the two clones with highest viru- to the assembly (Li, 2012; Tables S2 and S3). lence based on a preliminary screening were chosen. At least two The final assembly was then checked against other Bacillus clones from each endpoint evolved lineage, in addition to the wt thuringiensis kurstaki strains using MAUVE (Darling et al., 2011) ancestor (7.1.o wt pHT315- gfp) and the mutator (7.1.o ΔmutS), to check for possible genome errors. The assembled genome was DIMITRIU et al.     |  711 then run though the software PROKKA (Seemann, 2014) providing acid was then added, the sample was vortexed and incubated a draft annotation. The PROKKA flag— use genus— genus Bacillus overnight at 4°C. The precipitate was spun down at 15,000 g for was used to improve taxon- specific gene annotations. The final 10 min, the supernatant removed, and the pellet was washed twice assembly was then checked against other Illumina whole- genome with ice cold 70% acetone then air- dried. Following the addition of shotgun data from the selection experiments. Sequenced clones and the final reference assembly were then analyzed with snippy 20 μl water, the mixture was sonicated for 2 min to resuspend the pellet. For immunodetection of Vip3Aa, samples were boiled with version 4.4.0 (https://github.com/tseem ann/snippy) to align reads SDS sample buffer and spotted onto a nitrocellulose membrane. against the newly assembled reference and used to call single- A Vip3A antibody (a kind gift from Professor Juan Ferré) was used nucleotide polymorphisms (SNPs). The Snippy SNP data were used in association with an anti- rabbit IgG HRP- linked antibody for to generate a Jukes– Cantor Neighbor- Joining (1000 boot strap enhanced chemiluminescent detection. replicates) phylogeny of isolates, implemented with Geneious Prime 2022.2.1. 2.9  |  Statistical analysis 2.7  |  Fitness measurements All statistical analyses were carried out in R (v4.0.4) (https://www.R- proje ct.org). Bioassay and mortality data were analyzed using one Measurements of relative fitness used a mutant of the 7.1.o isolate of two methods. For comparisons between independently evolved constructed by transformation with a pHT315- rfp plasmid containing lineages, we used generalized linear modeling (glms) with logit link the tetR gene conferring tetracycline resistance (Zhou et al., 2014). functions and quasibinomial errors to correct for overdispersion. This version of the ancestor, therefore, has a plasmid with a similar However, to test for treatment effects, or account for random backbone to the evolved wild- type clones but is distinguishable using effects associated with different lineages within treatments, we its distinct antibiotic resistance. Relative fitness was calculated from used generalized linear mixed models (glmer) in the package lme4 an estimate of relative reproductive rates (Malthusian parameter) of (Bates et al., 2015). Although we attempted to use glmer models the two genotypes as described previously (Zhou et al., 2020). with binomial errors, these commonly failed to converge, especially A 50/50 mix of spores of the two competitors was used to ini- tiate competition experiments and was produced under standard in more complex models with dose × treatment interactions. Instead, we used the conservative approach of arc- sine transforming conditions. Fitness was measured in conditions duplicating the se- proportional mortality and using normal errors. The glm and glmer lection experiment, that is, infection and transfer of cadaver to spor- approaches produce qualitatively similar results. Hypothesis testing ulating media, although insect cadavers were processed individually and HCO culture did not use antibiotics. After spore washing, mixes in glmers used likelihood ratio tests after model simplification. LC50s and their standard errors were calculated using the dose.p function were plated on antibiotic- free medium as well as tetracycline 10 μg/ ml (for the standard competitor ancestor). The density of evolved in the MASS package (Venables & Ripley, 2002) after fitting simple logit models using log10 dose and quasibinomial errors. clones in cadavers was calculated by subtracting competitor counts Analyses of viable spore production also used glmers with lin- from antibiotic- free plates total counts, as the cat gene present in eage fitted as a random effect, while the analysis of competitive fit- the mut strain proved unreliable at allowing growth on chloram- ness used evolved clone as a random effect and genotype (mutator phenicol on LBA and we wished to use a common method for both or wild type) and number of toxins as explanatory variables. Model wt and mut clones. 2.8  |  Assessment of toxin production assumptions (normality, heteroskedasticity) were checked with graphical analyses and qq plots. Raw experimental data are available from Zenodo (Dimitriu et al., 2022). Bt bacteria were cultured in sporulation medium as above. Crystal 3  |  R E S U LT S protein production was assessed by centrifuging 1 ml of culture and resuspending in 1 ml of water, followed by two rounds of sonication, 3.1  |  Characterization of the mut ancestor centrifugation, and washing to break open un- lysed cells. The final pellet was resuspended in 1 ml of water. Serial dilutions were plated out in order to calculate the concentration of viable spores as CFUs. SDS- PAGE analysis was then performed on equal numbers of CFUs to compare the production of crystal proteins. For the assessment of Wild- type (wt) mutation rate toward rifampicin resistance was 7 × 10−10 [95% CI 4.2 × 10−10 to 1.02 × 10−9] while for the ΔmutS strain, the mutation rate was 1.76 × 10−8 [95% CI 1.48 × 10−8 to 2.08 × 10−8], showing a 25- fold increase compared to the wt strain. The mutator, secreted Vip3A production, culture supernatant was passed through prior to passage, did not show any change in virulence relative to a 0.22- μm cellulose acetate filter and to 1 ml of the filtrate 50 μl of 2% sodium deoxycholate was added and incubated for 30 min on ice. One hundred and fifty microliters of 100% Trichloroacetic the ancestor (log10 LC50 = 4.9, test for difference from ancestor- effect size 0.21, SE = 1.22, p = 0.87). The competitive fitness of the mutator strain under the conditions of the passage experiment also DIMITRIU et al. 712 | did not differ from the wt ancestor (t = 0.49, p = 0.63, means ± SE of mut and wt are 1.09 ± 0.06 and 1.13 ± 0.06, respectively). bar plots) that correlated with loss of virulence. In particular, wild- type lineages lost more plasmids than mutator lineages (Figure 2b; Fisher's exact test two- tailed p = 0.006). Importantly, the infectivity selection regime was more effective at retaining these plasmids and 3.2  |  Changes in virulence in evolved lineages preventing invasion of putative cheaters (Figure 2b, Fisher's exact test two- tailed p = 0.00032). Duplicated endpoint assays of evolved lineages used total mortality, normalized to that of the wt ancestor in each experiment, to assess changes in virulence after five rounds of selection (ca. 60 3.4  |  Social cheating and Cry plasmid loss generations). These assays showed that increased mutation rate and infectivity selection between subpopulations resulted in increases Previously, we have seen that Cry toxin production is a public in normalized mortality relative to the in vitro controls (Figure 2a). good and loss of investment in Cry toxins can be driven by the Both strain (wild type or mutator) (F1,44 = 12.3, p = 0.001 at 3 days, F1,44 = 10.6, p = 0.002 at 5 days) and selection treatment (3 days, F2,44 = 21.4, p < 0.001; 5 days, F2,44 = 8, p = 0.0011) affected virulence of evolved lineages. Post hoc tests confirm increased virulence for selective advantage of increased competitive growth rate within hosts, or social cheating. In order to test whether mutants that had lost Cry- toxin plasmids were cheaters, we measured the fitness of one clone from each lineage in competition experiments that infectivity versus yield selected treatments and for mutators versus used a marked RFP mutant derived from the ancestor (Figure 3a). wild- type lineages (Tukey tests, all p < 0.01). Differences between treatments were larger when assessing early normalized mortality, suggesting changes in virulence affected timing and overall levels of mortality. We also used mixed models of arc- sine transformed mortality, which fitted independent lineage as a random effect. These gave qualitatively similar results for day 3 and day 5 mortality, Mutants carrying fewer Cry toxin genes had higher competi- tive fitness (mixed model likelihood ratio test (LRT) df = 1, like - lihood ratio (LR) = 9.05, p < 0.0026, Figure 3a). Mutants which had lost Cry toxins also had lower virulence, that is, higher LC50 (glm F1, 64 = 67.1, p < 0.0001, Figure 3b). Thus, these data were consistent with the hypothesis that low- virulence mutants were except in the mixed model analysis mutators and the infectivity free- loading on investment in Cry toxins by other variants in their treatment could be seen to increase virulence by causing more lineages (Raymond et al., 2012). mortality at higher doses (dose × selection treatment interactions and dose × strain interactions, all p < 0.01 day 3, and p < 0.001, day 5 mortality). Since the evolved lineages were not genetically homogeneous, we also conducted bioassays of six clones isolated from each lin- After accounting for toxin production, mutators evolved greater competitive fitness than wild- type lineages (Figure 3a; mixed model, df = 1, LR = 7.5, p = 0.006) and had increased virulence (i.e., lower LC50s; glm F1, 63 = 7.35, p < 0.0157, Figure 3b) suggesting that increas- ing mutation rate increased the supply of beneficial alleles without eage. Clonal assays of day 5 mortality show that mutation rate and reducing overall fitness. Since the original mutator mutant has in- infectivity selection increased virulence via interactions with dose distinguishable fitness from the wild- type ancestor (see Section 2, (mixed effect glm with lineage as random effect: dose × selection treatment interaction df = 3, LR = 16.95, p < 0.001; dose*strain interaction df = 1, LR = 12.59, p < 0.001; Figures S1 and S2). We identified a number of clones with substantial increases in virulence, Methods), the initial genetic background of these strains does not explain this pattern. quantified as a reduction of LC50 of more than an order of magni- 3.5  |  Spore production tude from endpoint assays (Table S1). The other major life- history trait that we explored was the efficiency of spore production in the sporulation media used in the selection 3.3  |  Patterns of Cry toxin plasmid carriage experiments. We saw lower spore production in the mutator and the infectivity- selected lineages, the treatments that produced We saw variation between independent evolved lineages in terms of higher virulence (Figure 4a, mixed models, selection treatment whether or not they increased or decreased virulence with respect to ancestors (Figure 2b, scatter plots). Low- virulence lineages clearly df = 3, LR = 11.3, p = 0.01, mutation rate df = 1, LR = 8.36, p < 0.01). Importantly, between lineages, lower spore production was as- had low mortality that did not increase with dose (glm lineage × dose interaction F1,17 = 5.92, p = 0.026). Cry toxins are encoded on two plasmids in Bt kurstaki— a mega- plasmid containing Cry1Aa, Cry1Ac, sociated with lower LC50, which corresponds to higher virulence (Figure 4b, F1,14 = 10.5, p < 0.01, adj. R2 = 0.39). We examined bacte- rial development and sporulation after 3 days (the typical growth pe- and Cry2Aa, and a circa 80 kb plasmid- carrying Cry1Ab (Méric riod in sporulation media) and after 6 days for clones with elevated et al., 2018). We used PCR primers to screen for loss of plasmids to test if loss of virulence in evolved lineages could be explained by changes in plasmid carriage. We observed numerous events of toxin virulence (n = 3), the wild- type ancestor and clones with unchanged virulence (n = 3). Elevated virulence was associated with delayed de- velopment. After 3 days, cultures of high- virulence clones retained gene loss in clones isolated from in vivo evolved lineages (Figure 2b 5%– 10% vegetative cells, while 85%– 95% cells had completed DIMITRIU et al.     |  713 (a) 1.4 s s e n t i f e v i t a e r l 1.2 1.0 0.8 strain ancestor mutator wild type (b) 1010 l μ / U F C 0 5 C L 108 106 104 0 1 2 3 number of Cry1A toxins 0 1 2 3 number of Cry1A toxins F I G U R E 3 Social cheating and fitness in experimentally evolved Bt lines. Evolved clones that had lost virulence plasmids encoding Cry genes had higher competitive fitness (a) and lower virulence (b) than the ancestor. Toxin complements in A (n = 16) are derived from Illumina sequencing data; LC50s are shown for all evolved clones with toxin complement data based on PCR (n = 70). Means ± SE of fitness for each clone are plotted as solid points, raw data are plotted at 50% transparency, the dashed line is a reference line for the ancestor while solid lines show fitted statistical models (glms fitted with strain and number of Cry toxins). development and undergone exosporium lysis. For clones with un- similarity to other Bt kurstaki genomes (Tables S2 and S3, Figure S4). changed virulence, these figures were 1% vegetative cells and 99% Resequencing of evolved mutants confirmed that none carried lysis. After 6 days, all clones showed 98%– 99% lysis indicating that mutations in their Cry genes or in regions immediately upstream. clones with elevated virulence could eventually undergo complete Comparison to the ancestral genome showed both small- scale sporulation. mutations and large- scale deletions associated with plasmid loss. This is biologically significant because toxin production is Particularly, toxin gene loss patterns detected previously by PCR can linked to sporulation in Bt, especially the Cry1A toxins that are be explained by the loss of one or both Cry plasmids. the dominant virulence factors in our experimental strains (Deng Phylogenetic analysis of our evolved clones shows that et al., 2014). In addition, we selected for sporulation by heat- there was considerably more genome evolution in the mutators treating preparations at the end of each round of selection to kill (Figure 5a). However, evolved clones with similar virulence phe- vegetative cells and possible contaminants. Reduced sporulation notypes did not group together, nor did clones from the same would therefore be expected to result in lower Cry toxin produc- treatments (Figure 5a). Mapping of mutations to the reference also tion and reduced virulence. Two classes of toxins are produced by suggests minimal convergent evolution in terms of shared changes Bt kurstaki earlier in the growth cycle than the Cry1A toxins and are in DNA (although there is phenotypic convergent evolution in terms expected to have activity against Cry1A- resistant insects: These of sporulation). We did identify a small list of genes that acquired are the Cry2 toxins and the vegetative virulence factor Vip3Aa mutations in more than one clone (Figure 5b); this list includes three (Carrière et al., 2015; Deng et al., 2014). However, we found no transcriptional regulators nprA (in clones Mia3 and Mia4), dagR, and significant differences in the ratio of Cry1A and Cry2 production sgrR with nonsynonymous mutations in insect- passaged lineage, but that correlated with changes in virulence. At the lineage level, nei- not in the in vitro controls. In general, missense or frameshift muta- ther mutators nor infectivity- selected replicates had elevated Cry2 tions were prevalent in genes encoding putative transcriptional reg- production (Figure 4c), nor did we see differences in Cry2 produc- ulators (Table S4). These included well- described regulators such tion between the ancestor and evolved clones with high virulence as NprA but also proteins with helix– turn– helix domains. Mutator (Figure S3). We did find substantial variation in secretion of Vip3Aa clones from the infectivity selection treatment accounted for 10 between evolved clones, but no association with increased viru- of the 16 loss of function mutations in putative transcriptional reg- lence was seen (Figure 4d). To seek mechanistic explanations for the reduced sporulation ulators. If we compared the virulence (LC50) of all the sequenced mutator clones, there was a significant increase in virulence in the and increased virulence, the genomes of two clones from each mutator clones with these loss of function mutations compared to lineage were sequenced and compared to a long- read assembly of our ancestor. The ancestral genome consisted of a 5.7- Mb chro- those with no mutations in regulatory genes (F1,16 = 5.91, p = 0.027, Figure 6), assuming that the data from these clones are independent mosome and 14 plasmids that resolved as single contigs with high observations. DIMITRIU et al. 714 | (a) s e g a e n i l t n e d n e p e d n i broth d broth c broth b broth a broth d broth c broth b broth a pool d pool c pool b pool a pool d pool c pool b pool a group d group c group b group a group d group c group b group a ancestor ancestor strain wild type mutator (b) l / µ U F C 0 1 g o l 0 5 C L 8 7 6 5 4 0e+00 2e+05 4e+05 spore production CFU/µl 6e+05 8e+05 1e+05 2e+05 4e+05 spore production CFU/µl 3e+05 5e+05 ancestor a infectivity lineages a b b c c d d (c) 250 kDa 180 kDa 130 kDa 95 kDa 72kDa 55 kDa (d) ancestor Wib2 Wyb3 Wic3 Mya5 Mic2 Myc6 Mid6 Myd2 Mya3 Mia3 d e g n a h c n u e c n e u r i v l e c n e u r i v l d e s a e r c n i F I G U R E 4 Increased virulence trade- offs against reduced spore production in vitro in evolved Bt lines but is not associated with increased Vip3 production. (a) Variation in spore production between evolved lineages; asterisks indicate the significance of contrasts between all evolved lineages and the ancestor in a glm; boxplots summarize the results from two assays of each clone. (b) Mean LC50 of each in vivo passaged lineage (averaged across clones) plotted against spore production, ancestors plotted as hollow circles. (c) SDS PAGE showing Cry toxin production for all infectivity- selected lineages, see Figure S3 for examples of individual clones. (d) Slot blot characterization of Vip3 production in the ancestral Bt clone and clones with increased virulence or virulence unchanged relative to the ancestor (see Table S1). See also Figures S3– S5 and Tables S2 and S3 for additional details on genomic and proteomic characterization of evolved clones. 4  |  D I S C U S S I O N Virulence factors can behave as public goods both in laboratory models and infections of insects with biocontrol agents such as Kin selection theory emphasizes that altered virulence can have entomopathogenic nematodes and Bt (Deng et al., 2015; Raymond different impacts on individual- and group- level fitness components et al., 2012; Shapiro- Ilan & Raymond, 2016; Zhou et al., 2014). In (Buckling & Brockhurst, 2008). Notably, investment in bacterial this experiment, it was clear that some clones within lineages had virulence factors requires resources to be diverted away from reduced virulence relative to the ancestor and so were putative growth of individual cells. If virulence factors are diffusible, they cheaters and we identified a link between the absence of Cry toxins may behave as “public goods” by conferring fitness benefits that and increases in competitive fitness. are shared among members of an infecting group (Diard, Sellin, This result mirrors previous experiments with Bt strains that et al., 2014; Harrison et al., 2006; Raymond et al., 2012). If cells have been cured of Cry toxin- encoding plasmids as well as com- reduce investment in public goods virulence, these “cheaters” petition between Bt and naturally Cry null B. cereus (Raymond can gain a reproductive advantage during growth in the host by et al., 2007, 2012). The key difference in this study is that loss freeloading on the products of others (Buckling & Brockhurst, 2008; of Cry toxin plasmids occurred spontaneously during experimen- Diggle et al., 2007; West & Buckling, 2003), although this may be tal evolution in multiple independent lineages and in different accompanied by a group- level cost such as reduced infectivity. ways, that is, by the loss of either or both of the large plasmids F I G U R E 5 Genomic analyses of evolved Bt clones. (a) Consensus neighbor- joining tree of sequenced evolved clones. For ease of interpretation, bootstrap of values >50 only has been displayed for the nodes. The scale bar indicates the number of substitutions per nucleotide. The reference refers to the hybrid PacBio/Illumina assembly, all other sequence information is derived from SNPs identified using Illumina data with reference to this assembly. (b) Histogram of nonsynonymous mutations identified in at least two evolved clones in our three selection regimes, infectivity selection, yield selection, and the in vitro passage controls. DIMITRIU et al. mvd1 94.6 88.2 90.1 70.9 71.8 60 (a) 86.5 710 mutator ancestor Hybrid assembly (reference) 48.1 89.8 86.5 96.3 90 Selection treatment (symbol key) infectivity yield in vitro control Clones with increased virulence 0.02 (b)     |  715 mia4 mya2 myd5 mib4 mib2 mib5 mia3 myc4 mic1 mid2 mid6 mib1 mic5 myc6 myd6 myb1 myb2 mva1 mya1 100 wib6 wic3 710 WT ancestor GFP wic2 wic2 wyc1 wya6 wvd1 wib3 wid5 wyd6 wva1 wia1 wic6 wyc3 wia5 wid4 wyb5 wic4 wic5 wya3 wic1 wyb1 wyd3 40 30 20 10 0 40 30 20 t n u o c 10 0 40 30 20 10 0 mutation rate mutator wild type i f n e c t i v i t y i y e d l i n v i t r o G) E 0(M e s orota dro Dihy e s ata h p s o h p oly p o x E A c o h p s n C u c k protein orter Bic n protein A ate tra n o arb Bic 5 4 e P m hro c yto C atio etoin utiliz at s e a h D 8 k 1 c A er n e o G er n e G A e Glc s a e e s a A A/Bip p Ty e kin erin s o m o H cr s n ra erm olate p Gly c nt e n o p m o orter ntip a( )/H( ) a N E n h ctive protein IntQ E stC 1 u b p P e protein P e protein P m u x p e efflu urin utative d P erm rt p s a e efe P e s a cle u n ore protein P p ble s r s a e erm ort p etic oth p hy al protein e IIC c erm n p a n e h Lic s a e prA ctivator N r erD e X s a bin m o c e re sin Tyro rter n al a criptio s n Tra cr s n Tra P p s n ate tra h p s o h P P DIMITRIU et al. 716 | 0 5 C L 0 1 g o l 5.0 4.5 4.0 3.5 3.0 (van Leeuwen et al., 2015), pooling cadavers can reduce relatedness relative to infectivity selection treatment. In other words, there is more scope for cheating and greater selection to increase competitive fitness in the yield treatment. An additional consideration is that the strain infectivity selection also imposed selection for gains in virulence at an ancestor mutator wild type additional spatial scale. In all treatments, there was selection for in- fectivity at a between- host level (successful infections had to occur in order for passage to proceed), but by using selection in a metapopula- tion, we imposed an additional level of competition between subpop- ulations within a metapopulation. The effects of this level of selection for infectivity in pathogens require further investigation, but we would hypothesize that competition between subpopulations is more important when investment in virulence has costs in terms of growth rate and population size within hosts (Raymond & Erdos, 2022). The primary aim of this study was to devise and test different methods of selection for maintaining and improving virulence, and so we will re- quire further work to tease out the precise evolutionary mechanisms behind the success of this particular treatment. The mutator lines showed clear differences in the rate of mo- lecular evolution as well as greater increases in virulence relative to the wild- type background. In contrast to previous work, we did not see an increase in levels of cheating under high mutation rate no yes regulatory mutations F I G U R E 6 Virulence of evolved mutators clones with and without loss of function mutations in putative transcriptional regulators. Virulence is expressed as log LC50 with evolved wild- type clones and the 7.1.o ancestor plotted as comparisons, after excluding data from sequenced clones that are Cry toxin cheaters (i.e., which carry 0 or 1 Cry toxin genes). Table S4 contains a list of mutations that encode these toxins. In addition, the selection pressure we (Harrison & Buckling, 2005; Racey et al., 2009). Mutagenesis is very imposed during experimental evolution affected how readily Cry commonly employed in strain improvement and directed evolu- toxins were lost: selection pressure to maximize yield (final bacte- tion with Bacillus sp. and other microbes (Lai et al., 2004; Raju & rial population size in cadavers) resulted in higher rates of plasmid Divakar, 2013). The evidence for the ability of mutators to accelerate loss than in the infectivity selection treatment. The loss of Cry increases in fitness is complex and has several unresolved questions. plasmids in response to selection on yield occurred because Cry Mutators can increase the supply of both beneficial and deleterious toxins impose costs on growth rate and total production of spores alleles and it is clear that there is no known direct benefit to having within hosts (Raymond et al., 2012). Both social evolution and evo- impaired proofreading: mutator alleles rise in frequency by hitchhik- lution of virulence theory commonly divide the selective forces ing on the fitness benefits of linked alleles (Chao & Cox, 1983; de into individual (within- host)- and group- level (between- hosts) Visser, 2002; Gentile et al., 2011; Raynes et al., 2013). Mutators are components; these are often in conflict in microbes (Cressler often seen at quite high proportions in pathogenic bacteria (LeClerc et al., 2016; Frank, 1997). Other bacterial virulence factors which et al., 1996; Matic et al., 1997; Oliver et al., 2000) and mutators can are public goods, such as siderophores, have a group- level benefit appear spontaneously in experimental evolution although evolving in terms of increasing yield within the host (Harrison et al., 2006; lineages often moderate mutation rates during long- term transfer West & Buckling, 2003). However, Cry toxins are produced at experiments (Good et al., 2017; Ho et al., 2021). sporulation in the cadaver, rather than in the early stages of in- One reason for the prevalence of mutators among pathogens, fection, and these Cry toxins require so much protein production and for their success in this study is that small populations and/or in- that Bt variants which invest in these toxins produce substantially fection bottlenecks may limit the supply of beneficial mutations (de fewer spores per unit resource than their Cry null counterparts Visser, 2002). While demographics and mutation supply are known (Deng et al., 2015; Raymond et al., 2012). In essence, Cry toxins to affect the long- term success of mutators, simulation and exper- are a different type of public good than siderophores and act to in- imental studies suggest that bottlenecks or small populations may crease infectivity rather than the quality of resources available for not favor high mutation rates (Ho et al., 2021; Raynes et al., 2018). growth within hosts. It remains to be seen whether other bacterial Nevertheless, mutators are more successful when large- effect ben- public goods can also reduce the population size of pathogens in eficial mutations are available for hitchhiking (Gentile et al., 2011; hosts but increase infectivity or transmission. Mao et al., 1997; Thompson et al., 2006) and the strong selection The infectivity selection was not only more effective at retaining imposed by new sporulation conditions and a novel host genotype Cry plasmids but also resulted in greatest gains in virulence relative in this study may account for the success of mutators in our exper- to the ancestor. There are several factors that could have contributed iments. Overall, the use of mutators to overcome host resistance to these results. First, budding dispersal in a metapopulation involved has a sensible biological basis. However, this may have had implica- less pooling of cadavers than the yield selection treatment. Since in- tions for the stability of virulence (Figure S2). Potentially more sta- dividual cadavers are likely to be dominated by different genotypes ble phenotypes can be produced by periodic rounds of chemically DIMITRIU et al.     |  717 induced mutagenesis or by using initial pathogen populations with example, cessation of growth can be cooperative and late switch- high standing genetic variation. ing to sporulation can provide growth advantages in competition It was not possible to identify a single simple cause for any gains (Gardner et al., 2007; Ratcliff et al., 2013). Efficient sporulation is in virulence in this study, although this was consistently related to essential for the production of both fungal and bacterial biocontrol decreased sporulation efficiency. Reduced sporulation means that agents and can be readily lost during laboratory selection. In gen- vegetative cells and their secreted proteins will be more prevalent eral, passage regimes that minimize social conflict and which can in final inocula. We observed that a proportion of mutants with maintain valuable cooperative traits could have broad importance reduced sporulation gave increased virulence per spore but not in across diverse groups of invertebrate pathogens, and here, we have terms of dilution of the broth in which they were cultured. This is po- shown the value of a metapopulation regime that can reduce social tentially because standardizing doses by spore means that a higher conflicts. Moreover, experimental evolution offers a mechanism- free concentration of broth- associated virulence factors are included in solution to overcoming pest resistance which is potentially applicable inocula. As yet, none of the classic virulence factors associated with to many pathogens and pests. In vivo selection experiment may help broth such as Vip3A appear to be responsible for this increase in us discover entirely novel virulence factors or virulence modification virulence. There are many virulence factors secreted into broth by responses, in contrast to directed evolution methods that require Bt and its relatives (Chitlaru et al., 2006; Guinebretière et al., 2002; well- understood protein– protein interactions (Badran et al., 2016). Stenfors Arnesen et al., 2008). While spores and crystals are by far the most significant contributors to virulence in Bt (Raymond AC K N OW L E D G M E N T S et al., 2010), broth- associated virulence may be more significant The authors thank Andy Matthews for support and technical advice. for Cry1A- resistant hosts and are known to be significant for some This study was supported by the Leverhulme Trust (RPG- 2014- 252) hosts such as Galleria mellonella (Salamitou et al., 2000). and BBSRC (BB/S002928/1). For example, secreted virulence factors are regulated by a quorum- sensing system (PlcR/PapR) that activates a suite of virulence factors C O N FL I C T O F I N T E R E S T involved in host invasion (Salamitou et al., 2000; Zhou et al., 2014). This work formed part of international patent application number Although these are typically produced at stationary phase, expression WO 2019/030529 A1 by BR & NC. of the PlcR- regulated genes is repressed in low- nutrient sporulation medium such as HCO (Lereclus et al., 2000). The stationary phase DATA AVA I L A B I L I T Y S TAT E M E N T secretome of Bt and B. anthracis in sporulation medium is mainly Sequence data are hosted at the SRA under BIOPROJECT composed of metalloproteases such as Inh1A and NprA and is much SUB10359598. Plasmids, ancestor, and mutator clones are avail- reduced in the diversity of proteins compared to nutrient- rich media able on request. Evolved mutants can be shared subject to Material (Chitlaru et al., 2006; Perchat et al., 2011). These metalloproteases Transfer Agreements. Experimental data are available at Zenodo are hypothesized to have a role in the digestion of cadavers in late- with a permanent doi: 10.5281/zenodo.7503995. stage infections (Perchat et al., 2016). Nevertheless, they are essen- tially lytic enzymes and could improve the ability of Bt to invade the O R C I D host at high concentration. Other possible virulence factors secreted Tatiana Dimitriu https://orcid.org/0000-0002-1604-2622 into sporulating media include chitinase- and chitin- binding proteins Alistair Darby https://orcid.org/0000-0002-3786-6209 (Chitlaru et al., 2006). Notably, we did identify mutations in nprA and Neil Crickmore https://orcid.org/0000-0002-8448-0763 other transcriptional regulators in several insect- passaged mutators. Ben Raymond https://orcid.org/0000-0002-3730-0985 Bt is an important pathogen for both organic horticulture and modern “biotech” crops, so the question of how to improve strains R E F E R E N C E S or discover new toxins is significant, especially given the need to respond to the evolution of resistance (Adang et al., 2014; Badran et al., 2016). However, these methods may be applicable to other par- asites, such as nematodes and fungi, which also produce costly ex- creted virulence factors (Shapiro- Ilan & Raymond, 2016). Moreover, in addition to providing a framework for increasing virulence, the combination of in vivo and in vitro selection used here could be ap- plied when the aim is adaptation to an alternative growth medium or the improvement of other phenotypes, while retaining insecticidal ef- ficacy. For some pathogens, for example, fungi, it is also worth bear- ing in mind that while virulence traits may not be social (there are no data yet), other important biocontrol traits such as sporulation are likely to be subject to social conflicts. The timing of a switch to any resting state (spore, persister, etc.) can be subject to cheating— for Adang, M. J., Crickmore, N., & Jurat- Fuentes, J. L. (2014). Diversity of Bacillus thuringiensis crystal toxins and mechanism of action. In T. S. Dhadialla & S. S. Gill (Eds.), Advances in insect physiology (pp. 39– 87). Academic Press. Alizon, S., Hurford, A., Mideo, N., & Van Baalen, M. (2009). Virulence evolution and the trade- off hypothesis: History, current state of affairs and the future. Journal of Evolutionary Biology, 22, 245– 259. https://doi.org/10.1111/j.1420- 9101.2008.01658.x Badran, A. H., Guzov, V. M., Huai, Q., Kemp, M. M., Vishwanath, P., Kain, W., Nance, A. M., Evdokimov, A., Moshiri, F., Turner, K. H., Wang, P., Malvar, T., & Liu, D. R. (2016). Continuous evolution of Bacillus thuringiensis toxins overcomes insect resistance. Nature, 533, 58– 63. https://doi.org/10.1038/natur e17938 Barrett, A. D. T. (2017). Yellow fever live attenuated vaccine: A very successful live attenuated vaccine but still we have problems controlling the disease. Vaccine, 35, 5951– 5955. https://doi. org/10.1016/j.vacci ne.2017.03.032 DIMITRIU et al. 718 | Bates, D., Machler, M., Bolker, B., & Walker, S. (2015). Fitting linear mixed- effect models using lme4. Journal of Statistical Software, 67, 1– 48. https://doi.org/10.18637/ jss.v067.i01 Baxter, S. W., Badenes- Pérez, F. R., Morrison, A., Vogel, H., Crickmore, N., Kain, W., Wang, P., Heckel, D. G., & Jiggins, C. D. (2011). Parallel evolution of Bacillus thuringiensis toxin resistance in lepidoptera. Genetics, 189, 675– 679. Berling, M., Blachere- Lopez, C., Soubabere, O., Lery, X., Bonhomme, A., Sauphanor, B., & Lopez- Ferber, M. (2009). Cydia pomonella granu- lovirus genotypes overcome virus resistance in the codling moth and improve virus efficiency by selection against resistant hosts. Applied and Environmental Microbiology, 75, 925– 930. https://doi. org/10.1128/AEM.01998 - 08 Bolger, A. M., Lohse, M., & Usadel, B. (2014). Trimmomatic: A flex- ible trimmer for Illumina sequence data. Bioinformatics (Oxford, England), 30, 2114– 2120. https://doi.org/10.1093/bioin forma tics/btu170 Bravo, A., Agaisse, H., Salamitou, S., & Lereclus, D. (1996). Analysis of cryIAa expression in sigE and sigK mutants of Bacillus thuringiensis. Molecular and General Genetics, 250, 734– 741. Bravo, A., Likitvivatanavong, S., Gill, S. S., & Soberon, M. (2011). Bacillus thuringiensis: A story of a successful bioinsecticide. Insect Biochemistry and Molecular Biology, 41, 423– 431. https://doi. org/10.1016/j.ibmb.2011.02.006 Buckling, A., & Brockhurst, M. A. (2008). Kin selection and the evolu- tion of virulence. Heredity, 100, 484– 488. https://doi.org/10.1038/ sj.hdy.6801093 Carrière, Y., Crickmore, N., & Tabashnik, B. E. (2015). Optimizing pyr- amided transgenic Bt crops for sustainable pest management. Nature Biotechnology, 33, 161– 168. https://doi.org/10.1038/ nbt.3099 Chakroun, M., Banyuls, N., Bel, Y., Escriche, B., & Ferré, J. (2016). Bacterial vegetative insecticidal proteins (Vip) from entomopatho- genic bacteria. Microbiology and Molecular Biology Reviews, 80, 329– 350. https://doi.org/10.1128/MMBR.00060 - 15 Chao, L., & Cox, E. C. (1983). Competition between high and low mu- tating strains of Escherichia coli. Evolution, 37, 125– 134. https://doi. org/10.2307/2408181 Chitlaru, T., Gat, O., Gozlan, Y., Ariel, N., & Shafferman, A. (2006). Differential proteomic analysis of the Bacillus anthracis secre- tome: Distinct plasmid and chromosome CO2- dependent cross talk mechanisms modulate extracellular proteolytic activities. Journal of Bacteriology, 188, 3551– 3571. https://doi.org/10.1128/ JB.188.10.3551- 3571.2006 Cornforth, D. M., Matthews, A., Brown, S. P., & Raymond, B. (2015). Bacterial cooperation causes systematic errors in pathogen risk assessment due to the failure of the independent action hypoth- esis. PLoS Pathogens, 11, e1004775. https://doi.org/10.1371/journ al.ppat.10047 75.s002 Diard, M., Garcia, V., Maier, L., Remus- Emsermann, M. N. P., Regoes, R. R., Ackermann, M., & Hardt, W.- D. (2014). Stabilization of coopera- tive virulence by the expression of an avirulent phenotype. Nature, 494, 353– 356. https://doi.org/10.1038/natur e11913 Diard, M., Sellin, M. E., Dolowschiak, T., Arnoldini, M., Ackermann, M., & Hardt, W.- D. (2014). Antibiotic treatment selects for cooperative virulence of Salmonella Typhimurium. Current Biology, 24, 2000– 2005. https://doi.org/10.1016/j.cub.2014.07.028 Diggle, S. P., Griffin, A. S., Campbell, G. S., & West, S. A. (2007). Cooperation and conflict in quorum- sensing bacterial populations. Nature, 450, 411– 414. https://doi.org/10.1038/natur e06279 Dimitriu, T., Soissi, W., Morwool, P., Darby, A., Crickmore, N., & Raymond, B. (2022). Selecting for infectivity across metapopulations can in- crease virulence in the social microbe Bt: Data set. Zenodo. https:// doi.org/10.5281/zenodo.7303995 Ebert, D. (1998). Experimental evolution of parasites. Science, 282, 1432– 1435. https://doi.org/10.1126/scien ce.282.5393.1432 Frank, S. A. (1996). Models of parasite virulence. Quarterly Review of Biology, 71, 37– 78. Frank, S. A. (1997). The Price equation, Fisher's fundamental theorem, kin selection, and causal analysis. Evolution, 51, 1712– 1729. Gandon, S., Jansen, V. A. A., & van Baalen, M. (2001). Host life history and the evolution of parasite virulence. Evolution, 55, 1056– 1062. Garbutt, J., Bonsall, M. B., Wright, D. J., & Raymond, B. (2011). in Bacillus https://doi. Antagonistic competition moderates virulence thuringiensis. 765– 772. 14, Ecology org/10.1111/j.1461- 0248.2011.01638.x Letters, Gardner, A., & West, S. A. (2006). Demography, altruism, and the ben- efits of budding. Journal of Evolutionary Biology, 19, 1707– 1716. https://doi.org/10.1111/j.1420- 9101.2006.01104.x Gardner, A., West, S. A., & Griffin, A. S. (2007). Is bacterial persistence a social trait? PLoS One, 2, e752. https://doi.org/10.1371/journ al.pone.0000752 Gentile, C. F., Yu, S. C., Serrano, S. A., Gerrish, P. J., & Sniegowski, P. D. (2011). Competition between high- and higher- mutating strains of Escherichia coli. Biology Letters, 7, 422– 424. https://doi. org/10.1098/rsbl.2010.1036 Good, B. H., McDonald, M. J., Barrick, J. E., Lenski, R. E., & Desai, M. M. (2017). The dynamics of molecular evolution over 60,000 generations. Nature, 551, 45– 50. https://doi.org/10.1038/natur e24287 Granato, E. T., Ziegenhain, C., Marvig, R. L., & Kümmerli, R. (2018). Low spatial structure and selection against secreted virulence factors attenuates pathogenicity in Pseudomonas aeruginosa. The ISME Journal, 12, 2907– 2918. https://doi.org/10.1038/s4139 6- 018- 0231- 9 Griffin, A. S., West, S. A., & Buckling, A. (2004). Cooperation and compe- tition in pathogenic bacteria. Nature, 430, 1024– 1027. https://doi. org/10.1038/natur e02744 Cressler, C. E., McLeod, D. V., Rozins, C., van den Hoogen, J., & Day, T. (2016). The adaptive evolution of virulence: A review of theoretical predictions and empirical tests. Parasitology, 143, 915– 930. https:// doi.org/10.1017/s0031 18201 500092x Guinebretière, M.- H. M., Broussolle, V. V., & Nguyen- The, C. C. (2002). Enterotoxigenic profiles of food- poisoning and food- borne Bacillus cereus strains. Journal of Clinical Microbiology, 40, 3053– 3056. https://doi.org/10.1128/JCM.40.8.3053- 3056.2002 Darling, A. E., Tritt, A., Eisen, J. A., & Facciotti, M. T. (2011). Mauve as- sembly metrics. Bioinformatics (Oxford, England), 27, 2756– 2757. https://doi.org/10.1093/bioin forma tics/btr451 de Visser, J. A. G. M. (2002). The fate of microbial mutators. Microbiology (Reading, Engl), 148, 1247– 1252. Deng, C., Peng, Q., Song, F., & Lereclus, D. (2014). Regulation of cry gene expression in Bacillus thuringiensis. Toxins, 6, 2194– 2209. https:// doi.org/10.3390/toxin s6072194 Deng, C., Slamti, L., Raymond, B., Liu, G., Lemy, C., Gominet, M., Yang, J., Wang, H., Peng, Q., Zhang, J., Lereclus, D., & Song, F. (2015). Division of labour and terminal differentiation in a novel Bacillus thuringiensis strain. The ISME Journal, 9, 286– 296. https://doi. org/10.1038/ismej.2014.122 Hall, B. M., Ma, C.- X., Liang, P., & Singh, K. K. (2009). Fluctuation AnaLysis CalculatOR: A web tool for the determination of mutation rate using Luria– Delbrück fluctuation analysis. Bioinformatics, 25, 1564– 1565. https://doi.org/10.1093/bioin forma tics/btp253 Harrison, F., Browning, L. E., Vos, M., & Buckling, A. (2006). Cooperation and virulence in acute Pseudomonas aeruginosa infections. BMC Biology, 4, 21. https://doi.org/10.1186/1741- 7007- 4- 21 Harrison, F., & Buckling, A. (2005). Hypermutability impedes coopera- tion in pathogenic bacteria. Current Biology, 15, 1968– 1971. https:// doi.org/10.1016/j.cub.2005.09.048 Ho, W. C., Behringer, M. G., Miller, S. F., Gonzales, J., Nguyen, A., Allahwerdy, M., Boyer, G. F., & Lynch, M. (2021). Evolutionary dy- namics of asexual hypermutators adapting to a novel environment. DIMITRIU et al. Genome Biology and Evolution, 13, 18. https://doi.org/10.1093/gbe/ evab257 Kolodny- Hirsch, D. M., & Van Beek, N. A. M. (1997). Selection of a mor- phological variant of Autographa californica nuclear polyhedrosis virus with increased virulence following serial passage in Plutella xylostella. Journal of Invertebrate Pathology, 69, 205– 211. https://doi. org/10.1006/jipa.1997.4659 Kümmerli, R., Gardner, A., West, S. A., & Griffin, A. S. (2009). Limited dispersal, budding dispersal, and cooperation: An experimental study. Evolution, 63, 939– 949. https://doi. org/10.1111/j.1558- 5646.2008.00548.x Lai, Y. P., Huang, J., Wang, L. F., Li, J., & Wu, Z. R. (2004). A new approach to random mutagenesis in vitro. Biotechnology and Bioengineering, 86, 622– 627. https://doi.org/10.1002/bit.20066 Lecadet, M. M., Blondel, M. O., & Ribier, J. (1980). Generalized transduc- tion in Bacillus thuringiensis var. berliner 1715 using bacteriophage CP- 54Ber. Journal of General Microbiology, 121, 203– 212. LeClerc, J. E., Li, B. G., Payne, W. L., & Cebula, T. A. (1996). High mu- tation frequencies among Escherichia coli and Salmonella patho- gens. Science, 274, 1208– 1211. https://doi.org/10.1126/scien ce.274.5290.1208 Lereclus, D., Agaisse, H., Grandvalet, C., Salamitou, S., & Gominet, M. (2000). Regulation of toxin and virulence gene transcription in Bacillus thuringiensis. International Journal of Medical Microbiology, 290, 295– 299. Lereclus, D., Vallade, M., Chaufaux, J., Arantes, O., & Rambaud, S. (1992). Expansion of insecticidal host range of Bacillus thuringiensis by in vivo genetic recombination. Biotechnology, 10, 418– 421. Li, H. (2012). Exploring single- sample SNP and INDEL calling with whole- genome de novo assembly. Bioinformatics (Oxford, England), 28, 1838– 1844. https://doi.org/10.1093/bioin forma tics/bts280 Maleki- Milani, H. (1978). Influence de passages répétés du virus de la polyèdrose nucléaire de Autographa californica chez Spodoptera littoralis [Lep.: Noctuidae]. Entomophaga, 23, 217– 224. https://doi. org/10.1007/BF023 73096 Mao, E. F., Lane, L., Lee, J., & Miller, J. H. (1997). Proliferation of mutators in a cell population. Journal of Bacteriology, 179, 417– 422. https:// doi.org/10.1128/jb.179.2.417- 422.1997 Massey, R. C., Buckling, A., & Ffrench- Constant, R. (2004). Interference competition and parasite virulence. Proceedings of the Royal Society of London Series B: Biological Sciences, 271, 785– 788. Matic, I., Radman, M., Taddei, F., Picard, B., Doit, C., Bingen, E., Denamur, E., & Elion, J. (1997). Highly variable mutation rates in commensal and pathogenic Escherichia coli. Science, 277, 1833– 1834. https:// doi.org/10.1126/scien ce.277.5333.1833 Méric, G., Mageiros, L., Pascoe, B., Woodcock, D. J., Mourkas, E., Lamble, S., Bowden, R., Jolley, K. A., Raymond, B., & Sheppard, S. K. (2018). Lineage- specific plasmid acquisition and the evolution of spe- cialized pathogens in Bacillus thuringiensis and the Bacillus cereus group. Molecular Ecology, 27, 1524– 1540. https://doi.org/10.1111/ mec.14546 Oliver, A., Canton, R., Campo, P., Baquero, F., & Blazquez, J. (2000). High frequency of hypermutable Pseudomonas aeruginosa in cys- tic fibrosis lung infection. Science, 288, 1251– 1253. https://doi. org/10.1126/scien ce.288.5469.1251 Pearce, D. A., Bridge, P. D., Hughes, K. A., Sattler, B., Psenner, R., & Russell, N. J. (2009). Microorganisms in the atmosphere over Antarctica. FEMS Microbiology Ecology, 69, 143– 157. https://doi. org/10.1111/j.1574- 6941.2009.00706.x Perchat, S., Dubois, T., Zouhir, S., Gominet, M., Poncet, S., Lemy, C., Aumont- Nicaise, M., Deutscher, J., Gohar, M., Nessler, S., & Lereclus, D. (2011). A cell– cell communication system regulates protease production during sporulation in bacteria of the Bacillus cereus group. Molecular Microbiology, 82, 619– 633. https://doi. org/10.1111/j.1365- 2958.2011.07839.x     |  719 Perchat, S., Talagas, A., Poncet, S., Lazar, N., Li de la Sierra- Gallay, I., Gohar, M., Lereclus, D., & Nessler, S. (2016). How quorum sens- ing connects sporulation to necrotrophism in Bacillus thuringien- sis. PLoS Pathogens, 12, e1005779. https://doi.org/10.1371/journ al.ppat.1005779 Poelwijk, F. J., Kiviet, D. J., Weinreich, D. M., & Tans, S. J. (2007). Empirical fitness landscapes reveal accessible evolutionary paths. Nature, 445, 383– 386. https://doi.org/10.1038/natur e05451 Pollitt, E. J. G., West, S. A., Crusz, S. A., Burton- Chellew, M. N., & Diggle, S. P. (2014). Cooperation, quorum sensing, and evolution of viru- lence in Staphylococcus aureus. Infection and Immunity, 82, 1045– 1051. https://doi.org/10.1128/IAI.01216 - 13 Raberg, L., de Roode, J. C., Bell, A. S., Stamou, P., Gray, D., & Read, A. F. (2006). The role of immune- mediated apparent competition in ge- netically diverse malaria infections. American Naturalist, 168, 41– 53. Racey, D., Inglis, R. F., Harrison, F., Oliver, A., & Buckling, A. (2009). The effect of elevated mutation rates on the evolution of cooperation and virulence of Pseudomonas aeruginosa. Evolution, 642, 515– 521. https://doi.org/10.1111/j.1558- 5646.2009.00821.x Raju, E., & Divakar, G. (2013). Bacillus cereus GD 55 strain improvement by physical and chemical mutagenesis for enhanced production of fibrinolytic protease. International Journal of Pharma Science and Research, 4, 81– 93. Ratcliff, W. C., Hoverman, M., Travisano, M., & Denison, R. F. (2013). Disentangling direct and indirect fitness effects of microbial dormancy. The American Naturalist, 182, 147– 156. https://doi. org/10.1086/670943 Raymond, B., Davis, D., & Bonsall, M. B. (2007). Competition and re- production in mixed infections of pathogenic and non- pathogenic Bacillus spp. Journal of Invertebrate Pathology, 96, 151– 155. Raymond, B., & Erdos, Z. (2022). Passage and the evolution of virulence in invertebrate pathogens: Fundamental and applied perspec- tives. Journal of Invertebrate Pathology, 187, 107692. https://doi. org/10.1016/j.jip.2021.107692 Raymond, B., Johnston, P. R., Nielsen- Leroux, C., Lereclus, D., & Crickmore, N. (2010). Bacillus thuringiensis: An impotent patho- gen? Trends in Microbiology, 18, 189– 194. https://doi.org/10.1016/j. tim.2010.02.006 Raymond, B., Johnston, P. R., Wright, D. J., Ellis, R. J., Crickmore, N., & Bonsall, M. B. (2009). A mid- gut microbiota is not required for the pathogenicity of Bacillus thuringiensis to diamondback moth larvae. Environmental Microbiology, 11, 2556– 2563. https://doi. org/10.1111/j.1462- 2920.2009.01980.x Raymond, B., West, S. A., Griffin, A. S., & Bonsall, M. B. (2012). The dy- namics of cooperative bacterial virulence in the field. Science, 337, 85– 88. https://doi.org/10.1126/scien ce.1218196 Raynes, Y., Halstead, A. L., & Sniegowski, P. D. (2013). The effect of population bottlenecks on mutation rate evolution in asexual pop- ulations. Journal of Evolutionary Biology, 27, 161– 169. https://doi. org/10.1111/jeb.12284 Raynes, Y., Wylie, C. S., Sniegowski, P. D., & Weinreich, D. M. (2018). Sign of selection on mutation rate modifiers depends on population size. Proceedings of the National Academy of Sciences of the United States of America, 115, 3422– 3427. https://doi.org/10.1073/pnas.17159 96115 Rumbaugh, K. P. K. P., Trivedi, U. U., Watters, C. C., Burton- Chellew, M. N. M. N., Diggle, S. P. S. P., & West, S. A. (2012). Kin selection, quorum sensing and virulence in pathogenic bacteria. Proceedings Biological Sciences/The Royal Society, 279, 3584– 3588. https://doi. org/10.1098/rspb.2012.0843 Sabin, A. B., & Boulger, L. R. (1973). History of Sabin attenuated poliovi- rus oral live vaccine strains. Journal of Biological Standardization, 1, 115– 118. https://doi.org/10.1016/0092- 1157(73)90048 - 6 Salamitou, S., Ramisse, F., Brehélin, M., Bourguet, D., Gilois, N., Gominet, M., Hernandez, E., & Lereclus, D. (2000). The plcR regulon is DIMITRIU et al. 720 | involved in the opportunistic properties of Bacillus thuringiensis and Bacillus cereus in mice and insects. Microbiology, 146, 2825– 2832. Society of London Series B: Biological Sciences, 270, 37– 44. https:// doi.org/10.1098/rspb.2002.2209 Seemann, T. (2014). Prokka: Rapid prokaryotic genome annotation. (Oxford, England), 30, 2068– 2069. https://doi. Bioinformatics org/10.1093/bioin forma tics/btu153 Selifonova, O., Valle, F., & Schellenberger, V. (2001). Rapid evolu- tion of novel traits in microorganisms. Applied and Environmental https://doi.org/10.1128/ Microbiology, AEM.67.8.3645- 3649.2001 3645– 3649. 67, Shapiro, M., Lynn, D. E., & Dougherty, E. M. (1992). More virulent biotype isolated from wild- type virus: Google Patents. Shapiro- Ilan, D., & Raymond, B. (2016). Limiting opportunities for cheat- ing stabilizes virulence in insect parasitic nematodes. Evolutionary Applications, 9, 462– 470. https://doi.org/10.1111/eva.12348 Stenfors Arnesen, L. P., Fagerlund, A., & Granum, P. E. (2008). From soil to gut: Bacillus cereus and its food poisoning tox- ins. FEMS Microbiology Reviews, 32, 579– 606. https://doi. org/10.1111/j.1574- 6976.2008.00112.x Taddei, F., Radman, M., Maynard- Smith, J., Toupance, B., Gouyon, P. H., & Godelle, B. (1997). Role of mutator alleles in adaptive evolution. Nature, 387, 700– 702. https://doi.org/10.1038/42696 Thézé, J., Cabodevilla, O., Palma, L., Williams, T., Caballero, P., & Herniou, E. A. (2014). Genomic diversity in European Spodoptera exigua mul- tiple nucleopolyhedrovirus isolates. Journal of General Virology, 95, 2297– 2309. https://doi.org/10.1099/vir.0.06476 6- 0 Thompson, D. A., Desai, M. M., & Murray, A. W. (2006). Ploidy controls the success of mutators and nature of mutations during bud- ding yeast evolution. Current Biology, 16, 1581– 1590. https://doi. org/10.1016/j.cub.2006.06.070 van Leeuwen, E., Neill, S. O. A., Matthews, A., & Raymond, B. (2015). Making pathogens sociable: The emergence of high relatedness through limited host invasibility. The ISME Journal, 9, 2315– 2323. https://doi.org/10.1038/ismej.2015.111 Venables, W. N., & Ripley, B. (2002). Modern applied statistics with S (4th ed.). Springer. West, S. A., & Buckling, A. (2003). Cooperation, virulence and sidero- phore production in bacterial parasites. Proceedings of the Royal Wick, R., Judd, L., Gorrie, C., & Holt, K. (2017). Unicycler: Resolving bacterial genome assemblies from short and long sequencing reads. PLoS Computational Biology, 13, e1005595. https://doi. org/10.1371/journ al.pcbi.1005595 Zhou, L., Alphey, N., Walker, A. S., Travers, L. M., Hasan, F., Morrison, N. I., Bonsall, M. B., & Raymond, B. (2018). Combining the high- dose/ refuge strategy and self- limiting transgenic insects in resistance management – A test in experimental mesocosms. Evolutionary Applications, 11, 727– 738. https://doi.org/10.1111/eva.12573 Zhou, L., Slamti, L., Lereclus, D., & Raymond, B. (2020). Optimal response to quorum- sensing signals varies in different host environments with different pathogen group size. mBio, 11, e00535- 20. https:// doi.org/10.1128/mBio.00535 - 20 Zhou, L., Slamti, L., Nielsen- Leroux, C., Lereclus, D., & Raymond, B. (2014). The social biology of quorum- sensing in a naturalistic host pathogen system. Current Biology, 24, 2417– 2422. https://doi. org/10.1016/j.cub.2014.08.049 S U P P O R T I N G I N FO R M AT I O N Additional supporting information can be found online in the Supporting Information section at the end of this article. How to cite this article: Dimitriu, T., Souissi, W., Morwool, P., Darby, A., Crickmore, N., & Raymond, B. (2023). Selecting for infectivity across metapopulations can increase virulence in the social microbe Bacillus thuringiensis. Evolutionary Applications, 16, 705–720. https://doi.org/10.1111/eva.13529 DIMITRIU et al.
10.1021_acssynbio.3c00148
pubs.acs.org/synthbio This article is licensed under CC-BY 4.0 Research Article Multiplicity of the Agrobacterium Infection of Nicotiana benthamiana for Transient DNA Delivery Erik D. Carlson, Jakub Rajniak, and Elizabeth S. Sattely* Cite This: ACS Synth. Biol. 2023, 12, 2329−2338 Read Online ACCESS Metrics & More Article Recommendations *sı Supporting Information ABSTRACT: Biological DNA transfer into plant cells mediated by Agrobacterium represents one of the most powerful tools for the engineering and study of plant systems. Transient expression of transfer DNA (T-DNA) in particular enables rapid testing of gene facile combinatorial products and has been harnessed for expression of multiple genes. In analogous mammalian cell-based gene expression systems, a clear sense of the multiplicity of infection (MOI) allows users to predict and control viral transfection frequencies for applications requiring single versus multiple transfection events per cell. Despite the value of Agrobacterium-mediated transient transformation of plants, MOI has not been quantified. Here, we analyze the Poisson probability distribution of the T-DNA transfer in leaf pavement cells to determine the MOI for the widely used model system Agrobacterium GV3101/Nicotiana benthamiana. These data delineate the relationship between an individual Agrobacterium strain infiltration OD600, plant cell perimeter, and leaf age, as well as plant cell coinfection rates. Our analysis establishes experimental regimes where the probability of near-simultaneous delivery of >20 unique T- DNAs to a given plant cell remains high throughout the leaf at infiltration OD600 above ∼0.2 for individual strains. In contrast, single-strain T-DNA delivery can be achieved at low strain infiltration OD600: at OD600 0.02, we observe that ∼40% of plant cells are infected, with 80% of those infected cells containing T-DNA product from just a single strain. We anticipate that these data will enable users to develop new approaches to in-leaf library development using Agrobacterium transient expression and reliable combinatorial assaying of multiple heterologous proteins in a single plant cell. KEYWORDS: Agrobacterium-mediated transient expression, Multiplicity of infection, Plant synthetic biology, Nicotiana benthamiana, GV3101, pEAQ expression vector ■ INTRODUCTION The stable expression of multiple transgenes is a powerful way to endow a plant with a new trait such as resistance to potato blight1 or the production of polyunsaturated fatty acids in soy.2 These strategies require the coordinated action of multiple for multigene transgenes. However, many challenges exist expression systems, including a lack of suitable genetic parts (promoters, terminators, etc.), rapid methods for integrating multiple genes, and lengthy design-build-test cycles required for optimizing pathways and gene sets. Agrobacterium-mediated DNA delivery is one of the most widely used tools in plant biotechnology. Agrobacterium transfers single-stranded DNA into plant cells as part of the infection process; this DNA is ultimately integrated into the termed “stable host genome for expression”. This process has been harnessed for the delivery of user-defined cargo and has become a major method for generating transgenic plant lines when DNA is delivered into undifferentiated tissue.3 While numerous species across the plant kingdom are known to be hosts for Agrobacterium sustained expression, infection, some are amenable to high infection rates with low symptom development. In a few of these plants, DNA can be expressed transiently in somatic plant tissues by infiltrating a suspension of bacterial cells into leaf tissue (Figure 1A). This strategy has been widely used for rapidly testing transgene function and even as a method for heterologous protein production without the need for generating stable lines.4 A particularly effective pairing is transgenic plant Agrobacterium strain GV3101(pMP90)5 with binary vector pEAQ6 in the plant host Nicotiana benthamiana (Figure 1A, Supporting Information (SI), Supplementary Figure 1A,B). Agrobacterium tumefaciens GV3101/pMP90 is a C58 lineage strain, with the T-DNA portion of the wild-type virulence Received: March 9, 2023 Published: August 9, 2023 © 2023 The Authors. Published by American Chemical Society 2329 https://doi.org/10.1021/acssynbio.3c00148 ACS Synth. Biol. 2023, 12, 2329−2338 Downloaded via 24.52.215.216 on March 7, 2025 at 19:03:40 (UTC).See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles. ACS Synthetic Biology pubs.acs.org/synthbio Research Article Figure 1. Transient expression of multiple genes in N. benthamiana leaves by coinfiltration of Agrobacterium expression strains. (A) Infiltration of Agrobacterium tumefaciens strain GV3101 into Nicotiana benthamiana leaves. (B) Cross-section diagram of N. benthamiana leaf based on microscope images of a typical leaf used in this study. Gray area indicates the air space. Blue and yellow coloring correspond to sections of the leaf in panel C. (C) Brightfield image of the abaxial (lower surface) side showing pavement cells where data are collected for this study. (D) Application of transient expression for reconstitution of multistep metabolic pathways in N. benthamiana leaves. Each gene is delivered in a separate Agrobacterium strain; infiltration of strain mixtures into leaves results in heterologous protein expression and metabolite production.9 (E) Strategy to quantify the number of unique Agrobacterium-derived T-DNA gene products that are present in a given plant cell. Genes for three orthogonal fluorescent proteins are each carried by the T-DNA in the pEAQ plasmid of individual strains. Strain mixtures are coinfiltrated into leaves at various ratios keeping total infiltration OD600 at 0.6. Confocal microscopy allows identification of coinfection events at a single cell level. Confocal microscopy image of 0.2 OD600 per strain, n = 3. White scale bar is 100 μm. plasmid pTiC58 replaced with a gentamycin resistance cassette to create pMP90.5 N. benthamiana grows relatively quickly, with ample leaf material for easy infiltration ∼4−6 weeks after planting. GV3101/pEAQ infiltration into N. benthamiana leaves leads to detectable T-DNA product expression in a matter of days and heterologous protein levels that have been reported to reach upward of 1.5 g kg−1 fresh weight (FW) of GFP and 325 mg kg−1 FW of human antibody 2G12.6 Notably, it has been found that coinfiltration of multiple Agrobacterium strains with unique expression constructs leads to simultaneous expression of the T-DNAs in plant leaves.7 This has enabled the rapid testing of multigene biosynthetic pathways.8−11 The modularity of this coinfiltration method and the fact that each gene can be driven by the same promoter (e.g., 35S) often make this approach for pathway reconstitution in N. benthamiana leaves more straightforward compared to traditional heterologous hosts such as E. coli and yeast, which also have limits with the complexity of heterologous proteins produced compared to a plant system. Furthermore, transient expression in plants has drastically accelerated the design− build−test cycle for to the generation of stable transgenic lines that can require months to years of development. testing plant pathways prior rapid, combinatorial this type of Multistep pathways (>10 unique gene products required, Figure 1D) are expressed routinely, and accumulation of the expected products is often observed. For example, coinfiltra- tion of 16 Agrobacterium GV3101(pMP90)/pEAQ strains delivering a 16 gene pathway transiently in N. benthamiana leaves produces (−)-deoxypodophyllotoxin up to 4.3 mg g−1 dry weight.10 Notably, in intermediates that would result “partial” pathways are typically not observed. This may be due to enzyme specificity (cells with fewer than all required genes do not make new metabolic products), metabolite sharing across cells, and high efficiency of T-DNA delivery, with enough cells receiving a complete set of unique gene constructs. In analogous experimental strategies for rapid genetic manipulation of cells, such as viral transfection of mammalian cells in culture, the multiplicity of infection (MOI) can be a critical experimental design parameter to help predict how many unique transfection events a given target cell is likely to undergo. MOI was developed with bacteriophage/bacteria systems12 and is now routinely used in viral/mammalian transformation systems.13 Despite the importance of Agro- bacterium-mediated transient expression as a widely used tool to investigate plant cell biology and pilot transgene constructs, function in a the process by which multiple T-DNAs coordinated way is not well understood, and no analogous MOI value has been determined for system. For biosynthetic pathway reconstitution using this approach, the observed robust pathway product formation by coinfiltration of multiple Agrobacterium strains in N. benthamiana leaves suggests high rates of coinfection at the plant cell level and/ or high rates of metabolite or nucleic acid sharing across plant cells, but this has not been extensively studied. this Here we develop an in planta coinfiltration fluorescence assay to quantify coinfection rates (Figure 1E) and a statistical model that allows us to predict the number of unique T-DNA products present in each plant pavement cell, an MOI metric for this system. We predict that a typical N. benthamiana plant pavement cell in the top three fully expanded leaves can receive 2330 https://doi.org/10.1021/acssynbio.3c00148 ACS Synth. Biol. 2023, 12, 2329−2338 ACS Synthetic Biology pubs.acs.org/synthbio Research Article Figure 2. Fluorescence microscopy of RFP-expressing Agrobacterium with GFP T-DNA load in N. benthamiana leaves. (A) A constitutive RFP cassette was added to the non-T-DNA portion of a pEAQ vector with GFP T-DNA. (B) Acetosyringone-induced bacterial cultures were infiltrated daily into different spots on the same leaves. (C) Representative 3, 4, and 5 days post-infiltration fluorescence microscopy images of leaf 11. The GFP channel shows characteristic cytosolic expression in the plant pavement cells, and the RFP channel marks the Agrobacterium location in the leaf. RFP channel images are shown as inverse negatives of a green false colored image, to display magenta on a white background for visual clarity. (D) GFP channel pixel intensity. (E) RFP channel pixel intensity counts. Each data point corresponds to the channel corresponding to a single image. on average 1−3 infection events by a given Agrobacterium strain infiltrated at 0.2 OD600. Our work suggests that a high rate of coinfection by coinfiltrated strains is a main driver for observed pathway reconstruction in this transient expression system. At low infiltration OD600s (<0.02), the majority of infection events are from a single strain, informing experimental design where at most 1 T-DNA type in a given cell is desired. This problem is analogous to MOI measure- ments for bacteriophage/bacteria and virus/mammalian cell systems that are commonly used for genetic screens. This is an essential first step toward the development of novel genetic screens in whole plant leaves, where at most 1 T-DNA type/ plant cell would be needed. More broadly, we anticipate that quantification and modeling of multigene delivery through Agrobacterium transient expression can contribute to accelerat- ing the design−build−test cycle for engineering plant traits. ■ RESULTS GV3101(pMP90)/pEAQ Strain Tracking and T-DNA Expression Timing. To capture early stages of infection and get a sense of timing for fluorescent protein expression, we first tracked Agrobacterium simultaneously with plant-cell expressed T-DNA after leaf infiltration, as in ref 14. We constructed a version of pEAQ-GFP (T-DNA containing green fluorescent protein) with an added constitutive red fluorescent protein (RFP) expression cassette15 on the non-T-DNA backbone, which turns the colony and cell pellet visibly pink (Figure 2A). This fluorescently tags Agrobacterium, which maintains the pEAQ plasmid with RFP, and GFP encoded in the T-DNA is present only once T-DNA is delivered to and expressed by a plant cell. This strategy enables us to track the strain during the infection time course and expression of T-DNA products by the infected plant cell. Once a day for 5 days and once on the morning of imaging, 0.6 OD600 acetosyringone-induced GV3101(pMP90)/pEAQ- GFP-RFP was infiltrated into different sites on the same top 3 fully expanded N. benthamiana leaves 9, 10, and 11 (Figure 2A,B and SI, Supplementary Figure 5A,B). Five days after the first infiltration, sites were imaged on the underside of the leaf using a Leica SP8 confocal microscope. Laser powers were set for the 5 days post-infiltration (dpi) spot on leaf 11 and kept constant for imaging all the remaining spots (SI, Supple- mentary Figure 5C,D). Representative images of the 3, 4, and 5 days post-infiltration spots show RFP-expressing Agrobacterium throughout the image with detectable GFP expression in the plant pavement cells starting at 3 dpi (Figure 2C and SI, Supplementary Figure 5C). The GFP channel is characteristic of cytosolic expression, showing a clear cell outline and nucleus, caused by the vacuole taking up the majority of the cell volume, thus pushing the cytosol to the cell edges and around the nucleus. Quantifying pixel intensity counts for GFP (Figure 2D) shows protein expression can be detected at 3 days postinfiltration, with a marked increase in fluorescent signal at 5 dpi. Agrobacterium RFP signal is observed in planta in all infiltration conditions (Figure 2E and SI, Supplementary Figure 5D). To check for plant defense priming effects from the first infiltration on subsequent infiltration spots, we used separate plants for each day post-infiltration time point (SI, Supplementary Figures 6 and 7). We again observed detectable T-DNA product formation at 3 dpi (SI, Supplementary Figure 6B), and stable Agrobacterium RFP signal in the experimental time frame (SI, Supplementary Figure 6C). We chose 5 dpi for coinfiltration experiments given the clear signal-to-noise ratio, 2331 https://doi.org/10.1021/acssynbio.3c00148 ACS Synth. Biol. 2023, 12, 2329−2338 ACS Synthetic Biology pubs.acs.org/synthbio Research Article Figure 3. Quantification of infected cells at different OD600s allows for determination of MOI. (A) One fluorescent stain (GV3101/pEAQ-CFP) used as a dilution strain, keeping total Agrobacterium infiltration OD600 constant at 0.6. Agrobacterium mixes were infiltrated into the top 3 fully expanded leaves of 6-week-old N. benthamiana plants (leaves 8, 9, and 10) and imaged after 5 days. Minimum of 3 images were collected for each leaf, with a minimum of 9 images total for each coinfiltration condition. (B) Selected image from Leaf 9, infiltration mix n = 60, Data set 2. (C, D) Multiplicity of infection analysis for Leaf 9 from Data set 2. (C) Fraction of infectable pavement cells positive for a given fluorescent protein (f FP+) is dependent on strain infiltration OD600. (D) Fraction of infectable pavement cells negative for a given fluorescent protein, f 0 = 1 − f FP+. Inset: fitting the reduced Poisson distribution model (eq 5) to calculate α values for mCherry and YFP. (E) Box and whisker plots show fraction of infected pavement cells that are empty (just CFP), single infected (just YFP or mCherry, with or without CFP), and double infected (YFP and mCherry, with or without CFP), at various indicator strain infiltration OD600, Ax. (F) Box and whisker plots for only YFP and/or mCherry positive plant pavement cells (PCs), single or double infected at various indicator strain infiltration OD600. Multiplicity of infection (MOI) values for various infiltration OD600 and α = 12 unique T-DNA products in a pavement cell per infiltration OD600. the early stage of infection, the lack of an apparent defense phenotype or tissue damage in the leaves, and the fact that biosynthetic pathway reconstitution experiments are typically performed in this time frame. Determination of Agrobacterium Infection Frequen- cies and the Relationship to Infiltration OD600. Multi- plicity of infection is classically defined as the ratio of infectious termed MOIinput. The more agents to infection targets, accurate and useful metric, MOIactual, is derived from experimental data and thus accounts for the dynamics of vector adsorption to the target cell and subsequent infection success. For the Agrobacterium-mediated transient expression system, we define an MOIactual metric based on the relationship of infiltration OD600 to coinfection frequencies. Specifically, we use a Poisson distribution model with parameter λ defined as the product of strain infiltration OD600 (Ax), and a fitted constant term α, defined as the mean number of unique T- DNA products delivered by an Agrobacterium strain to a given plant cell per infiltration OD600 for a given plant pavement cell (eqs 1 and 2 and Poisson Distribution Modeling Equations). the MOI should change, Accordingly, we anticipate that depending on the total number of Agrobacterium cells infiltrated into a leaf. To examine and quantify the transfer of unique T-DNAs in this type of combinatorial, coinfection process, we chose to use Agrobacterium strains each encoding a different fluorescent protein with an orthogonal emission signal distinguishable by confocal microscopy (Figure 3A and SI, Supplementary Figure 1C). Specifically, cyan, yellow, and red fluorescent proteins (CFP, YFP, mCherry, respectively, optimized for cytosolic expression in plant cells16) were cloned into the pEAQ-HT vector6 under the 35S promoter and transformed into Agrobacterium tumefaciens GV3101(pMP90). Imaging by confocal microscopy then allows us to detect and enumerate that contain different combinations of pavement cells fluorescent protein products after a specified time interval. To determine α, the infection frequency, as a function of Agrobacterium concentration, we varied inoculation OD600 and tracked expression of T-DNA fluorescent protein products. We noted that most but not all pavement cells express fluorescent 2332 https://doi.org/10.1021/acssynbio.3c00148 ACS Synth. Biol. 2023, 12, 2329−2338 ACS Synthetic Biology pubs.acs.org/synthbio Research Article Figure 4. Quantification of pavement cell coinfection rates of Agrobacterium at equal ODs and the effect of plant cell size. (A) Experimental setup and infiltration into leaves 6−13 of a 6 week old Nicotiana benthamiana plant. Three Agrobacterium strains, each containing a gene for a different fluorescent protein, were coinfiltrated at equal OD600 (0.2 OD600/strain, 0.6 OD600 total, 1:1:1). The 6-week old Nicotiana benthamiana plant structure infiltrates leaves 6−13 as counted up from the cotyledons (1 and 2). (B) Representative confocal microscopy images from leaves 13 and 6 of 1:1:1 coinfiltration condition of CFP/YFP/mCherry at 5 dpi. (C) Fraction of infected plant pavement cells that contains one, two, or three fluorescent proteins. Averages are four images each from a different region of the same plant leaf disc ± SD. (D) Box and whisker plots of plant pavement cell perimeters (μm) by leaf. Each data point is the perimeter of an individual plant pavement cell. Number of plant pavement cells fully in the four image frames per condition is noted by “n”. (E) Average pavement cell perimeters vs calculated multiplicity of infections for CFP, YFP, and mCherry. Error bars show the ± SD. protein when leaves are infiltrated with a saturating OD600 of Agrobacterium (SI, Supplementary Figure 2). We designated Agrobacterium containing either YFP- and mCherry-encoded T-DNA as “indicator” strains, and used CFP as a dilution strain representing a bulk population and as a marker for the total infectable plant pavement cells (Figure 3A,B). number of While we focused our analysis on pavement cells due to the ease of imaging, we noted that mesophyll cells in plant leaves are also infectable (SI, Supplementary Figure 4). Individual strain OD600s were varied as illustrated in Figure infiltration OD600 kept constant at 0.6. 3A with the total Induced strain mixes were hand infiltrated into the top 3 fully expanded leaves of 6-week-old N. benthamiana (leaves 8, 9, and 10, counting up from the cotyledons as 1 and 2), and imaged at 5 dpi (SI, Supplementary Figures 8C and 9A). At least three images were collected per leaf, with at least nine images total for each coinfiltration condition. Images were manually analyzed using ImageJ, to quantify the total number of cells expressing each of the three fluorescent proteins, as well as the combination of fluorescent protein signals in a given plant pavement cell (SI, Supplementary Tables 1 and 2). 2333 https://doi.org/10.1021/acssynbio.3c00148 ACS Synth. Biol. 2023, 12, 2329−2338 ACS Synthetic Biology pubs.acs.org/synthbio Research Article Figure 5. Poisson distribution modeling to estimate multiplicity of infection for GV3101 in N. benthamiana. (A) Expected number of infection events per plant pavement cell at various infection efficiencies based on Poisson distribution (eq 1). Curves with solid connection lines are based on the Poisson distribution form, with overlaid comparison of experimental data in a red square with dashed line. (B) Probability of a plant pavement cell being coinfected by all n coinfiltrated Agrobacterium strains, fixed total infiltration OD600, Atotal of 0.6, 1.6, 4, and 10. Representative range of α values used to generate curves. Points circled in black (α = 12) with black bars extending to upper and lower α values note the coinfiltration conditions with individual strain OD600 of 0.2 for a given total infiltration OD600. For the MOI calculation, we considered any cell that contained at least one fluorescent protein as “infectable” and counted the fraction of those cells that contained CFP, mCherry, and YFP ( f FP+) as a function of the strain infection OD600 (Figure 3C). As described in the Poisson distribution modeling equations experimental section, the reduced Poisson distribution of X = 0 can be applied to cells not expressing the f 0 = 1 − f FP+ (Figure 3D and SI, fluorescent protein, Supplementary Figures 8D−F and 9B−D), to fit for α and thus MOIactual. We found that MOI can vary from between ∼0.1− 3.6 for various individual strain OD600s (Figure 3F), revealing that at high infiltration OD600, multiple unique T-DNAs are expected to be delivered on average to a given plant cell. We also noted that the calculated α values changed depending on the leaf age. Older leaves, e.g., L8, were found to have a higher α (and corresponding MOI) relative to L10 (SI, Supple- mentary Figures 8D−F and 9B−D). A few assumptions we are making are as follows: first, does fluorescent protein detection in a cell correspond to Agro- bacterium infection and T-DNA transfer in that same cell? This is probably not always the case, as we know proteins of the size of RFP can travel from one plant cell to another via the plant’s plasmodesmata.17 Independence of infection events, meaning infection of a plant cell by one Agrobacterium cell does not affect the likelihood of a second Agrobacterium cell binding and infecting the same plant cell when coinfiltrated, is supported by Chi-square analysis of independence of infection frequencies of the two indicator strains, pEAQ-YFP and pEAQ-mCherry (SI, Supplementary Tables 1 and 2). Also of note, here we are not considering intensity of expression, which changes depending on the total number of T-DNA molecules delivered into a cell and the cell-to-cell variation in expression dynamics.18 In this work, these data reflect only binary yes/no presence of the fluorescent protein in a given plant pavement cell, which could result from one or more infection events with the same strain. Filtering the data to focus on indicator strains YFP and mCherry reveals useful experimental design spaces (Figure 3E,F). Above individual strain infiltration OD600 of ∼0.1, we noted that the majority of plant pavement cells are infected and positive for both YFP and mCherry. These data point to high rates of coinfection upon coinfiltration of multiple Agrobacterium strains encoding pathway enzymes as a main driver of the successful biosynthetic pathway reconstitution, where our lab typically uses an infiltration OD600 of 0.2−0.3 per Agrobacterium strain. The data also helps inform use cases such as genetic library delivery where, when plant cells are infected, the delivery of at most one T-DNA type per cell is desired. For these experiments, infiltration OD600s of less than ∼0.02 result in a majority of infected cells with only 1 of the indicator strains (Figure 3F). However, this is at a trade-off with total number of plant cells these conditions, e.g., ∼20% of infectable pavement cells for Ax = 0.005 contain YFP or mCherry (Figure 3E), with the majority of those cells with only one of the indicator strains (Figure 3F). infected under High Rates of Coinfection Correspond to Increasing Plant Pavement Cell Perimeters. With a method established for quantifying coinfection with three individual strains, we next examined which major variables influenced MOI across the plant body and specifically focused on leaf age. As before, we mixed and coinfiltrated three Agrobacterium strains, each bearing T-DNA for a different fluorescent protein but extended our analysis to include leaves 6−13. The three strains (encoding for CFP, YFP or mCherry) were mixed at equal volumes, for a final concentration of 0.2 OD600/strain, 0.6 OD600 total (Figure 4A and SI, Supplementary Figure 11A, 1:1:1). Separately, we also installed the Agrobacterium-labeling 2334 https://doi.org/10.1021/acssynbio.3c00148 ACS Synth. Biol. 2023, 12, 2329−2338 ACS Synthetic Biology pubs.acs.org/synthbio Research Article constitutive RFP cassette into the pEAQ-YFP vector backbone, creating pEAQ-YFP-RFP. This strain was induced and mixed with the pEAQ-CFP strain to visualize Agrobacterium and T- DNA expression at infiltration OD600 0.2 (SI, Supplementary Figure 11A, 2:1). Imaging the underside of the leaf at 5 dpi, we observed plant pavement cells expressing the cytosolic fluorescent proteins from T-DNA loads (Figure 4B and SI, Supplementary Figure 11C) and Agrobacterium tagged with RFP (SI, Supplementary Figure 11D). Representative images from leaves 13 (youngest) and 6 (oldest) show a robust expression of proteins from T-DNA constructs and that the majority of plant pavement cells are infected (Figure 3B). In each microscopy image for the 1:1:1 coinfiltration case, the perimeter and number of T-DNA products present was quantified for plant pavement cells fully in the microscope image (SI, Supplementary Figures 12 and 13A). Infected plant pavement cells were binned as single, double, and triple infected based on the total different fluorescent proteins observed in that cell (Figure 4C). In leaves 6−13, the majority of plant pavement cells are positive with at least one fluorescent protein (SI, Supplementary Figures 11C and leaves 6−7), we observed a 13A). For older leaves (e.g., majority of infected cells with all three fluorescent proteins (Figure 4C) that appeared to track with the increased perimeter of Indeed, quantifying the MOI, we observed good correlation with the average plant cell perimeter (Figure 4E). Data for cells >600 μm was not included for two (interrelated) reasons: (1) there are too few cells of this size that fit in a single image to give an accurate estimate, and (2) essentially all cells are triply infected in the older leaves, which also precludes getting an accurate estimate of MOI. Together, these data suggest a simple statistical model where the increase in cell perimeter increases the likelihood of Agrobacterium infection. Our results also highlight the importance of comparing leaves of comparable developmental stage for applications that rely on successful combinatorial expression of multiple distinct T-DNAs (e.g., biosynthetic pathway reconstitution). these pavement cells (Figure 4D). Exploring the Theoretical Coinfiltration Design Space. Based on the probability model determined by fitting the data to a Poisson distribution, we next explored possible experimental regimes for coinfection by n > 2 coinfiltrated strains. Toward this end, we generated plots based on the Poisson distribution analysis that represent likely experimental design for the GV3101(pMP90)/pEAQ in N. benthamiana system. Figure 5A plots various Poisson distribution curves (eq 1), with overlaid observed coinfection data for indicator strains YFP and mCherry, showing good agreement between the model and our observed data (Figure 5A and SI, Supple- mentary Figure 14). Figure 5B shows the number of coinfiltrated strains related to the probability of a given plant pavement cell being infected by all “n” strains. Curves of various total infiltration OD600 representing typical coinfiltra- tion experimental design spaces are shown (Atotal = 0.6, 1.6, 4, and 10 OD600), calculated with α = 7, 12, and 16 mean number of infection events in a given plant pavement cell per strain infiltration OD600. Reflective of typical coinfiltration experi- ments, noted on the graph are cases of individual strain infiltration OD600 of 0.2 for each total infiltration OD600 curve. The black bars extending up and down from the circled point show the possible range of complete coinfection by all n strains, depending on the MOIactual value for that given experiment. These data are in line with observed yields of 4.3 mg g−1 dry weight of a target medicinal compound, produced by coinfiltration of 16 unique Agrobacterium strains into N. benthamiana leaves at a total OD600 of 3;10 our model predicts that the probability of all 16 T-DNA products being present in a given plant cell approaches 50%, especially for infiltrations of older leaves. ■ DISCUSSION Agrobacterium-mediated transient expression in Nicotiana benthamiana is a powerful tool for combinatorial expression and discovery of heterologous pathway enzymes. Examples of reconstituted pathways include but are not limited to QS saponins,19 diosgenin,20 epipodophyllotoxin,9 forskolin,21 colchicine,9 and strychnine.22 Combinatorial expression is not only useful for pathway reconstitution, but it also enables pooled screens of candidate genes, as in,20 where batches of ∼30 enzymes were tested in combination to accelerate pathway discovery. This is especially valuable if multiple genes are required for a given enzymatic transformation, and the order of events is not known. Our data point to high rates of coinfection by coinfiltrated Agrobacterium strains to be a main driver of the platform’s utility in this space. By modulating the number and infiltration OD600 of Agrobacterium carrying different T-DNA loads, we show an ability to control key infection metrics like coinfection of a plant cell by different T-DNA loads. We have a clearer sense of this GV3101(pMP90)/pEAQ/N. benthamiana system for design- ing coinfiltration experiments. the accessible experimental landscape of Our data sets focus on plant pavement cells, due to the data collection limitations of confocal microscopy slide preparation and the number of infiltration conditions we targeted. The bulk of a Nicotiana benthamiana leaf tissue are mesophyll cells that do get infected upon GV3101/pEAQ infiltration (SI, Supplementary Figure 4) and are likely where most of the heterologous expression occurs. Generally, mesophyll cells are bigger than plant pavement cells and have more exposed surface area for Agrobacterium binding, so we postulate that coinfection rates would generally be higher in mesophyll cells compared to pavement cells. Therefore, MOI based on plant pavement cells would likely be an underestimate of MOI in mesophyll cells. to at Toward understanding the dynamics of heterologous pathway reconstitution in N. benthamiana, our statistical modeling predicts that coinfiltration of pathway enzymes by separate Agrobacterium leads least some level of coinfection into the same plant cell. However, we do not know about the possibility of enzyme or metabolite sharing between plant cells. If these scenarios do occur at appreciable levels, then this would increase only the likelihood of complete pathway reconstitution in the plant In designing Agrobacterium-mediated transient expression coinfiltration experiments, we show the main experimental variable determinants of coinfection rates are Agrobacterium infiltration OD600s, and leaf age (plant pavement cell size increasing with leaf age). Co-infection rates are positively correlated with both of these variables. leaf. Beyond reconstitution applications requiring high rates of coinfiltration, applications where at most 1 T-DNA type/plant cell is desired can be reached through these methods, e.g., for the delivery of CRISPR gRNA libraries to plant cells in an intact leaf. Single cell based genomewide knockout screens 2335 https://doi.org/10.1021/acssynbio.3c00148 ACS Synth. Biol. 2023, 12, 2329−2338 ACS Synthetic Biology pubs.acs.org/synthbio Research Article utilized in mammalian systems rely on transfection frequencies of at most 1 per cell (low MOI) in order to discover distinct loci involved in a particular cell phenotype. Furthermore, high MOIs have been used to explore combinatorial effects of perturbations.23 Exciting advances in translating these types of approaches to plants are being developed, e.g., in planta screening of enhancer libraries delivered by Agrobacterium.24 In this work, pooled Agrobacterium libraries were infiltrated at 1 OD600. Our work predicts a high MOI for this experiment, which is beneficial for rapid identification of rare functional regulatory sequences. Taken together, our work provides a quantitative evaluation of the MOI that we anticipate will be useful in the design and implementation of these types of methods, particularly where low MOIs are required. For example, by dialing down individual Agrobacterium infiltration OD600 to <0.02, the majority of infected plant cells are infected with only one T-DNA type. However, there is a trade-off with fraction of infected cells, with only ∼40% of plant pavement cells being infected at 0.02 OD600. We anticipate that measurements of MOI for the Agro- bacterium/N. benthamiana system will help with the design of future work requiring delivery of multiple unique T-DNA constructs and/or achieving at most 1 unique construct per cell. Furthermore, our investigation sheds light on the ease at which N. benthamiana can undergo transient infection with this Agrobacterium system and hopefully will help elucidate why equivalent DNA-delivery methods in e.g. monocots are more challenging. ■ EXPERIMENTAL SECTION Fluorescent Protein Cloning into pEAQ-HT Vector. Fluorescent proteins were cloned into the pEAQ-HT vector6 by standard cloning techniques, using the built in AgeI/XhoI cloning site. Primers 5′-TGAGATACCGGTATGGTGAG- CAAGGGCGAG (forward primer, added AgeI site in bold) and 5′-TGAGATCTCGAGTTAAGATCTGTACAGC- TCGTC (reverse primer, added XhoI site in bold) were used to amplify each fluorescent protein with polymerase chain reaction (PCR) with Phusion polymerase (NEB). Templates for PCRs were pAN579 (CFP), pAN581 (YFP) and pAN583 (mCherry), plasmids courtesy of the Nebenführ lab. Column- purified (Zymo) PCR products and purified plasmid pEAQ- HT were each digested with AgeI-HF and XhoI (NEB) for 1 h at 37 °C. After 30 min, CIP (NEB) was added to the pEAQ- HT digestion to dephosphorylate the backbone. Digestion products were column purified (Zymo), and eluted with nuclease-free water. Three ligation reactions were set up with T4 Ligase (NEB), 50 ng pEAQ-HT backbone, and 5-fold molar excess of fluorescent protein insert. Reactions were left at room temperature for 10 min and deactivated at 65 °C for 10 min, and 5 μL each were transformed into chemically competent E. coli DH5a (NEB). Following recovery in SOC at 37 °C for 1 h, cells were plated on LB-agar supplemented with 30 μg mL−1 kanamycin and incubated overnight at 37 °C. Colonies were picked into 5 mL of LB supplemented with 30 μg mL−1 kanamycin, grown overnight at 37 °C, and miniprepped (Zymo) for sequencing. Preparation of Competent GV3101/pMP90. GV3101/ pMP90 is grown at 30 °C, is resistant to gentamycin (15 μg mL−1) from pMP90 and can be grown in standard LB. To prepare chemically competent GV3101/pMP90, −80 °C stock was streaked out on LB-Agar plates supplemented with 15 μg mL−1 gentamycin and incubated at 30 °C for 2 days until clear colonies formed. A colony was picked to inoculate 5 mL of LB supplemented with 15 μg mL−1 of gentamycin and grown to saturation at 30 °C in a roller drum. The saturated culture was added to 1 L of LB supplemented with 15 μg mL−1 gentamycin in a baffled flask and incubated at 30 °C with 250 rpm shaking until an OD600 of ∼0.5 was reached. The culture flask was transferred to ice, shaken vigorously to rapidly cool the culture, and left on ice for 20 min with occasional shaking. Culture was pelleted in a precooled (4 °C) centrifuge for 5 min at 4000 × g. Supernatant was removed, and the pellet resuspended in 20 mL ice cold 20 mM CaCl2. The cell suspension was aliquoted (100 μL each) into 1.5 mL Eppendorf tubes, flash frozen in liquid nitrogen, and stored at −80 °C until use. Transformation of Competent GV3101/pMP90. Puri- fied pEAQ-HT plasmid DNA was added to an aliquot of the competent GV3101/pMP90 cells. Because of the low trans- formation efficiencies, ∼1 μg of DNA should be used (in this work, 5 μL was added of ∼200 ng μL−1 plasmid stock). The Eppendorf tube was placed in a 37 °C heat block for 5 min and then mixed well by flicking the tube 5−10 times. The cells were flash frozen on liquid nitrogen and thawed at 37 °C for 5 min. One mL of LB was added, and the tube incubated at 30 °C with rotation for 2 h. Cells were pelleted in a microcentrifuge for 4 min at 5000 × g. All but ∼100 μL of supernatant was removed, and cells resuspended. Concentrated cells were spread with sterile ∼4 mm glass beads on LB-agar plates supplemented with 15 μg mL−1 gentamycin and 30 μg mL−1 kanamycin, and incubated at 30 °C for 2 days until colonies appeared. A colony was then restreaked onto LB-agar agar plates supplemented with 15 μg mL−1 gentamycin and 30 μg mL−1 kanamycin and incubated at 30 °C for 1−2 days until colonies appeared. N. benthamiana Growth. Nicotiana benthamiana plants were grown indoors at room temperature under a 16/8 h light/ dark cycle. Plants were watered twice a week with 2 g L−1 of fertilizer (Peters Excel 15−5−15). Growth, Induction, and Infiltration of GV3101 Strains into N. benthamiana Leaves. A colony of GV3101/ pMP90/pEAQ-[CFP/YFP/mCherry] was picked with a sterile P10 pipet tip into 5 mL of LB supplemented with 15 μg mL−1 gentamicin and 30 μg mL−1 kanamycin and incubated for 24 h at 30 °C with shaking (250 rpm). Cell culture was transferred to a 15 mL falcon tube, pelleted at 5000 xg for 5 min, and resuspended in 4 mL of induction buffer (10 mM MES buffer pH 5.6; 10 mM MgCl2; 150 μM acetosyringone). Cell room temperature for 1−2 h, suspensions were left at occasionally inverting to mix. Following induction, the OD600 of the culture was taken (Thermo Scientific Genesys 20), and the culture OD600 was normalized to 0.6 OD600 with fresh induction buffer. Infiltration mixes were prepared with these induced cultures, depending on the experiment. Agrobacterium mixes were hand infiltrated into the underside of the leaf of 6- week old Nicotiana benthamiana using a 1 mL blunt-end syringe. Infiltrated plants were returned to an 18/6 light/dark cycle growth shelf for 5 days. Confocal Microscopy Imaging. Vacuum grease was applied near the edge of a square coverslip (VWR 48366− 223) with a 3 mL Luer lock syringe fitted with a blunt-end 18G needle, enough to form a continuous seal of ∼1 mm. Water (∼20 μL) was pipetted into the center of the coverslip. The leaf to be imaged was cut from the plant, and a leaf disk punched (1 cm diameter) from an infiltrated section of the leaf and placed underside (abaxial) side down on the water spot. A 2336 https://doi.org/10.1021/acssynbio.3c00148 ACS Synth. Biol. 2023, 12, 2329−2338 ACS Synthetic Biology pubs.acs.org/synthbio Research Article microscope slide (Fischer Scientific microscope slide 12−500- A3) was placed on top and pressed evenly and gently to form a seal with the coverslip and vacuum grease. A Leica TCS SP8 laser scanning confocal microscope in resonant scanning mode with LASX software was used to visualize fluorescent protein expression in leaves with a 20x dry objective. CFP fluorescence was imaged with an excitation of 440 nm/emission of 470/15 nm, GFP fluorescence with an excitation of 488 nm/emission of 525/50 nm, YFP fluorescence with an excitation of 515 nm/ emission of 550/30 nm, and RFP or mCherry fluorescence with an excitation of 580 nm/emission of 610/20 nm. Images for each fluorescent channel are taken sequentially. Emission filter gates were designed to avoid any overlap with emission of other fluorescent proteins, and was confirmed in practice in SI, Supplementary Figure 1C. Microscope images were processed using Fiji-ImageJ. Projection type of Max Intensity was used to flatten z-stacks, and Enhance Contrast with 0.3% saturated pixels and equalize histogram was used (except for Figure 2 and SI, Supplemental Figures 6 and 7 image processing, which were not contrast enhanced because of pixel intensity level quantification). Quantifying Plant Pavement Cell Perimeters. Micro- scope images were processed and printed, and then plant pavement cells fully within frame were traced onto blank paper by hand with a backlit LED light box. Traced images were scanned and transformed into a vector image using Adobe Illustrator pathfinder function (SI, Supplementary Figures 12 and 13A). Individual plant pavement cell perimeter was calculated using Adobe Illustrator; Window > Document Info, then select “Object” from top right pull-down menu in the new then select an object, and path length will be window, displayed. Co-infection events were manually counted using ImageJ, counting a cell positive for a given fluorescent protein if the key features of cell perimeter and nucleus could be clearly observed in that channel and cell. Poisson Distribution Modeling Equations. To model infection of plant cells by Agrobacterium, we assume that an Agrobacterium cell infiltrated into a leaf can lead to productive infection (i.e., infection that leads to gene expression) with infection events are some unknown probability and that mutually independent. If these conditions are satisfied, the number of productive Agrobacterium infection events per infectable plant cell (X) can be modeled using a Poisson distribution: (1) where k is a non-negative integer and λ is the average number of infection events per cell. Various factors such as leaf age, plant cell size, or nature of the T-DNA may affect λ, but under otherwise identical experimental conditions, λ will be directly to the total number of Agrobacterium cells proportional infiltrated, which can be quantified via an OD600 measurement. We therefore have (2) where A denotes the OD600 and α is a scaling factor. The probability that a cell is not infected (which we also denote as f 0) is then (3) 2337 while the probability that a cell is infected at least once is therefore (4) To determine α, we can vary A for an indicator strain carrying a fluorescent protein expression T-DNA and measure the fraction of plant cells that do not express this protein, i.e., f 0. The best fit value for α can be determined from experimental data by linear regression using a suitably transformed version of eq 3: (5) To model coinfection by multiple Agrobacterium strains, we assume that α is independent of the gene being expressed. Since infection events are mutually independent, we have for any pair of strains. Hence, for n different strains, the probability that a cell is infected by every strain at least once is given by (6) (7) where Ax is the OD600 of strain x. If each strain carries a biosynthetic pathway enzyme, for example, eq 7 gives the fraction of cells expected to express the entire pathway. For the case , where the OD600 of each infiltrated strain is the same, eq 7 simplifies to Finally, for a pair of strains with , we have (8) (9) (10) (11) Equations 10 and 11 can be used to estimate α in a data set where αAT is such that P(X1 = 0, X2 = 0) = e−αAT ≪ 1, but P(X1 > 0, X2 > 0) is reasonably small (between 0.1 and 0.8, say). ■ ASSOCIATED CONTENT Data Availability Statement All data is included in the Supporting Information and data file. *sı Supporting Information The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acssynbio.3c00148. Detailed experimental overviews, and raw data summary tables (PDF) Summary table (XLSX) setups, microscopy data set ■ AUTHOR INFORMATION Corresponding Author Elizabeth S. Sattely − Department of Chemical Engineering and Howard Hughes Medical Institute, Stanford University, Stanford, California 94305, United States; 0000-0002-7352-859X; Email: sattely@stanford.edu orcid.org/ https://doi.org/10.1021/acssynbio.3c00148 ACS Synth. Biol. 2023, 12, 2329−2338 PXkek()k==!A=fPXee(0)A0====PXPXe(0)1(0)1A>===fAln0=PXkXiPXkPXi(,)()()1212====×=PXXe(,...,0)(1)nxnA11x>==AxAnT=PXXe(,...,0)(1)nAnn1/T>=AAA122T==PXXe(0,0)A12T===PXXPXXee(0,0)(0,0)(1)AA1212/2/2TT=>=>==PXXe(0,0)(1)A12/22T>>= pubs.acs.org/synthbio Research Article ACS Synthetic Biology Authors Erik D. Carlson − Department of Chemical Engineering, Stanford University, Stanford, California 94305, United States Jakub Rajniak − Department of Bioengineering, Stanford University, Stanford, California 94305, United States Complete contact information is available at: https://pubs.acs.org/10.1021/acssynbio.3c00148 Author Contributions E.D.C. and E.S.S. conceived and designed the experiments. E.D.C. performed the experiments and microscopy work. E.D.C. and J.R. performed data analysis. E.D.C., J.R., and E.S.S. wrote the manuscript. Notes The authors declare no competing financial interest. ■ ACKNOWLEDGMENTS The authors would like to thank Heather Cartwright and Andrey Malkovskiy of the Carnegie Science Department of Plant Biology for training and use of the Leica SP8 Confocal Microscope, Alex Engel and KC Huang for helpful discussions, and Will B. Cody and Jack Liu for manuscript comments. ■ REFERENCES J. Transient production of in plants: evolution and (1) Ghislain, M.; Byarugaba, A. A.; Magembe, E.; Njoroge, A.; Rivera, C.; Román, M. L.; Tovar, J. C.; Gamboa, S.; Forbes, G. A.; Kreuze, J. F.; Barekye, A.; Kiggundu, A. Stacking three late blight resistance genes from wild species directly into African highland potato varieties confers complete field resistance to local blight races. Plant Biotechnol. J. 2019, 17, 1119−1129. (2) Qi, B.; Fraser, T.; Mugford, S.; Dobson, G.; Sayanova, O.; Butler, J.; Napier, J. A.; Stobart, A. K.; Lazarus, C. M. Production of very long chain polyunsaturated omega-3 and omega-6 fatty acids in plants. Nat. Biotechnol. 2004, 22, 739−745. (3) Lacroix, B.; Citovsky, V. Pathways of DNA Transfer to Plants from Agrobacterium tumefaciens and Related Bacterial Species. Annu. Rev. Phytopathol. 2019, 57, 231−251. (4) Kopertekh, L.; Schiemann, recombinant pharmaceutical proteins perspectives. Curr. Med. Chem. 2019, 26, 365−380. (5) Koncz, C.; Schell, J. The promoter of TL-DNA gene 5 controls the tissue-specific expression of chimaeric genes carried by a novel type of Agrobacterium binary vector. Mol. Gen. Genet. 1986, 204, 383−396. (6) Sainsbury, F.; Thuenemann, E. C.; Lomonossoff, G. P. pEAQ: versatile expression vectors for easy and quick transient expression of heterologous proteins in plants. Plant Biotechnol. J. 2009, 7, 682−693. (7) Kapila, J.; De Rycke, R.; Van Montagu, M.; Angenon, G. An Agrobacterium-mediated transient gene expression system for intact leaves. Plant Sci. 1997, 122, 101−108. (8) Wood, C. C.; Petrie, J. R.; Shrestha, P.; Mansour, M. P.; Nichols, P. D.; Green, A. G.; Singh, S. P. A leaf-based assay using interchangeable design principles to rapidly assemble multistep recombinant pathways. Plant Biotechnol. J. 2009, 7, 914−924. (9) Nett, R. S.; Lau, W.; Sattely, E. S. Discovery and engineering of colchicine alkaloid biosynthesis. Nature 2020, 584, 148−153. (10) Schultz, B. J.; Kim, S. Y.; Lau, W.; Sattely, E. S. Total Biosynthesis for Milligram-Scale Production of Etoposide Intermedi- ates in a Plant Chassis. J. Am. Chem. Soc. 2019, 141, 19231−19235. (11) Chuang, L.; Franke, J. Rapid Combinatorial Coexpression of Biosynthetic Genes by Transient Expression in the Plant Host Nicotiana benthamiana. Methods Mol. Biol. 2022, 2489, 395−420. (12) Ellis, E. L.; Delbrück, M. The growth of bacteriophage. J. Gen. Physiol. 1939, 22, 365−384. (13) Abedon, S. T.; Bartom, E. Multiplicity of Infection. Brenner’s Encyclopedia of Genetics; Elsevier, 2013; pp 509−510. (14) Li, X.; Pan, S. Q. Agrobacterium delivers VirE2 protein into host cells via clathrin-mediated endocytosis. Sci. Adv. 2017, 3, e1601528. (15) Schnabel, T.; Sattely, E. Engineering Posttranslational Regulation of Glutamine Synthetase for Controllable Ammonia Production in the Plant Symbiont Azospirillum brasilense. Appl. Environ. Microbiol. 2021, 87, No. e0058221. (16) Nelson, B. K.; Cai, X.; Nebenführ, A. A multicolored set of in vivo organelle markers for co-localization studies in Arabidopsis and other plants. Plant J. 2007, 51, 1126−1136. (17) Brunkard, J. O.; Runkel, A. M.; Zambryski, P. C. The cytosol must flow: intercellular transport through plasmodesmata. Curr. Opin. Cell Biol. 2015, 35, 13−20. (18) Saxena, P.; Thuenemann, E. C.; Sainsbury, F.; Lomonossoff, G. P. Virus-Derived Vectors for the Expression of Multiple Proteins in Plants. Methods Mol. Biol. 2016, 1385, 39−54. (19) Reed, J.; Osbourn, A. Engineering terpenoid production through transient expression in Nicotiana benthamiana. Plant Cell Rep. 2018, 37, 1431−1441. (20) Christ, B.; Xu, C.; Xu, M.; Li, F.-S.; Wada, N.; Mitchell, A. J.; Han, X.-L.; Wen, M.-L.; Fujita, M.; Weng, J.-K. Repeated evolution of cytochrome P450-mediated spiroketal steroid biosynthesis in plants. Nat. Commun. 2019, 10, 3206. (21) Pateraki, I.; Andersen-Ranberg, J.; Jensen, N. B.; Wubshet, S. G.; Heskes, A. M.; Forman, V.; Hallström, B.; Hamberger, B.; Motawia, M. S.; Olsen, C. E.; Staerk, D.; Hansen, J.; Mo̷ller, B. L.; Hamberger, B. Total biosynthesis of the cyclic AMP booster forskolin from Coleus forskohlii. eLife 2017, 6, na. (22) Hong, B.; Grzech, D.; Caputi, L.; Sonawane, P.; López, C. E. R.; Kamileen, M. O.; Hernández Lozada, N. J.; Grabe, V.; O’Connor, S. E. Biosynthesis of strychnine. Nature 2022, 607, 617−622. (23) Dixit, A.; Parnas, O.; Li, B.; Chen, J.; Fulco, C. P.; Jerby-Arnon, L.; Marjanovic, N. D.; Dionne, D.; Burks, T.; Raychowdhury, R.; Adamson, B.; Norman, T. M.; Lander, E. S.; Weissman, J. S.; Friedman, N.; Regev, A. Perturb-Seq: Dissecting Molecular Circuits with Scalable Single-Cell RNA Profiling of Pooled Genetic Screens. Cell 2016, 167, 1853−1866. (24) Jores, T.; Tonnies, J.; Dorrity, M. W.; Cuperus, J. T.; Fields, S.; Queitsch, C. Identification of Plant Enhancers and Their Constituent Elements by STARR-seq in Tobacco Leaves. Plant Cell 2020, 32, 2120−2131. 2338 https://doi.org/10.1021/acssynbio.3c00148 ACS Synth. Biol. 2023, 12, 2329−2338
10.1016_j.enpol.2022.113277
Community wealth building in an age of just transitions: exploring civil Community wealth building in an age of just transitions: exploring civil society approaches to net zero and future research synergies society approaches to net zero and future research synergies Max Lacey-Barnacle, Adrian Smith, Tim Foxon Publication date Publication date 01-01-2023 Licence Licence This work is made available under the CC BY 4.0 licence and should only be used in accordance with that licence. For more information on the specific terms, consult the repository record for this item. Document Version Document Version Published version Citation for this work (American Psychological Association 7th edition) Citation for this work (American Psychological Association 7th edition) Lacey-Barnacle, M., Smith, A., & Foxon, T. (2023). Community wealth building in an age of just transitions: exploring civil society approaches to net zero and future research synergies (Version 1). University of Sussex. https://hdl.handle.net/10779/uos.23492954.v1 Published in Published in Energy Policy Link to external publisher version Link to external publisher version https://doi.org/10.1016/j.enpol.2022.113277 Copyright and reuse: Copyright and reuse: This work was downloaded from Sussex Research Open (SRO). This document is made available in line with publisher policy and may differ from the published version. Please cite the published version where possible. Copyright and all moral rights to the version of the paper presented here belong to the individual author(s) and/or other copyright owners unless otherwise stated. For more information on this work, SRO or to report an issue, you can contact the repository administrators at sro@sussex.ac.uk. Discover more of the University’s research at https://sussex.figshare.com/ Contents lists available at ScienceDirect Energy Policy journal homepage: www.elsevier.com/locate/enpol Community wealth building in an age of just transitions: Exploring civil society approaches to net zero and future research synergies M. Lacey-Barnacle *, A. Smith , T.J. Foxon Science Policy Research Unit, University of Sussex, UK A R T I C L E I N F O A B S T R A C T Keywords: Community wealth building Grassroots innovations Transition pathways Community energy Just transitions Economic democracy Community Wealth Building (CWB) is a burgeoning international policy agenda for local economic development that seeks to enhance democratic ownership, retain the benefits of local economic activity and empower place- based economies and workers. Parallel to this, in the context of net zero transitions, there has been increasing interest in approaches to enhancing civil society and community ownership over local energy provision. How- ever, in academic and practitioner debates, there has been very little interaction between these two strands of thinking and action on the need for radical change in current energy provision, particularly as part of a wider transformative change away from dominant neoliberal economic thinking, policies and structures. In this Perspective, we explore the various ways in which synergies exist between CWB and energy transitions by considering two civil society approaches to transitions; namely, the Thousand Flowers transition pathway and research in Grassroots Innovations. We examine how community energy could be strengthened through CWB, by showing how the ideas within these two approaches respond to the five core principles of CWB. Promising future directions for research and practice are identified, including linking up CWB and just transitions strategies, a renewed focus on local financial innovation and the growing role of anchor institutions in supporting net zero transitions, particularly where CWB supports economic democracy transformations in new net zero economies. 1. Introduction Community Wealth Building (CWB) is a burgeoning international policy agenda for local economic development. CWB seeks to transform local-scale economies by repurposing and redirecting the procurement power of ‘anchor institutions’ towards local businesses and supply chains. Five principles for economic democratisation guide these de- velopments, diversifying ownership forms, retaining capital within lo- calities and strengthening worker involvement, security and rights. CWB arose as a counter to the dominance of neoliberal economic approaches that prioritise the privatisation, mobility and extraction of local wealth producing activities. Given growing interests in ‘just transitions’, we propose that CWB might offer practical ideas for decarbonising energy systems in which local economic empowerment, the democratisation of ownership and long-term social sustainability become more central. This is particularly important in a post-crisis ‘green recovery’. The current administrations in both the US and the UK are committed to ‘Build Back Better’, using responses to the global economic downturn caused by the Coronavirus pandemic to address long-standing (and worsened) social and economic challenges. This includes commit- ments to addressing regional economic inequalities (e.g. by ‘levelling up’ in the UK), as well as action to address climate change and promote a net zero transition. However, many scholars, practitioners and re- searchers are sceptical of mainstream approaches to addressing these challenges (e.g. Alperovitz and Dubb, 2014; Kelly and Howard, 2019; Guinan & O’Neill, 2020; Paul and Cumbers, 2021). There is criticism towards the track record of mainstream approaches and scepticism to- wards their future potential, which typically rely on measures to pro- mote inward investment to economically deprived areas and regions, alongside a focus on large-scale technology deployment and interna- tional competitiveness, often dominated by large multinational interests. As a result, communities across the US, Canada, Australia, the UK and Europe have been undertaking action to develop more bottom-up alternatives for both local economic development and civic and com- munity energy innovation. Burgeoning networks of policy, practice and research have grown in these two areas, though they have largely developed independently and therefore potential synergies may be Abbreviations: Community Wealth Building, CWB. * Corresponding author. E-mail address: m.lacey-barnacle@sussex.ac.uk (M. Lacey-Barnacle). https://doi.org/10.1016/j.enpol.2022.113277 Received 1 March 2022; Received in revised form 27 July 2022; Accepted 24 September 2022 EnergyPolicy172(2023)113277Availableonline1November20220301-4215/©2022TheAuthors.PublishedbyElsevierLtd.ThisisanopenaccessarticleundertheCCBYlicense(http://creativecommons.org/licenses/by/4.0/). M. Lacey-Barnacle et al. being missed. In this Perspective piece, we examine links and synergies between (1) Community Wealth Building as a new international move- ment for local economic development, and (2) civic/community energy and grassroots innovation approaches to energy decarbonisation. Through bringing these two fields together for the first time, we also aim to encourage others to advance research at this interface. 1.1. The rise of community wealth building CWB is a progressive policy, action and research movement that has grown in prominence and stature over the past decade. Having first emerged from the US before spreading to the UK (Hanna and Kelly, 2021; Guinan and O’Neill, 2020; Guinan and O’Neill, 2019), its trans- atlantic origins are now beginning to be transcended as projects circu- late across the globe, in locations as far flung as Australia (Fensham, 2020), Italy (Kohn, 2020), Tanzania (Collord, 2019) and closer to the US in Canada (Hanna, 2019). In both the UK and US, the locations for experimenting with CWB are numerous; with both Cleveland (US) and Preston (UK) being celebrated pioneers, whilst Oakland, Burlington, New York, Denver, Chicago and Detroit are just some of the US cities adopting a CWB approach, and local and regional governments in North Ayrshire, Newham, Islington, Sunderland, Stevenage, Oldham, Wigan, the North of Tyne, Sandwell, the Liverpool City Region, Lewes and Brighton & Hove are all adopting CWB in the UK. This is occurring alongside new commitments from the devolved governments of Wales and Scotland, with the Welsh govern- ment reforming national procurement policies and strategies and the Scottish government appointing a minister for community wealth and introducing a Community Wealth Building bill into the Scottish Parlia- ment (CLES, 2021). Given this substantial growth, we expect many more cities, regions and nations to emerge as key CWB actors internationally. CWB has to be understood through its origins as a direct response to the dominance of a global neoliberal political economy over the past four decades, which (under the guise of competition for inward invest- ment by capital) has seen privatisation, deregulation and liberalisation policies dominate the economic and political order of advanced liberal democracies, particularly the US and UK (Harvey, 2007). However, neoliberal approaches have largely failed to bring renewed prosperity to deindustrialised cities and regions in the US and the UK. It is therefore unsurprising that prominent CWB examples are emerging from these two countries, in numerous place-based economies. In resisting the neoliberal order, CWB proponents advocate five core principles in their innovative approach to local economic development: [1] Plural ownership of the economy [2] Making financial power work for local places [3] Fair employment and just labour markets [4] Progressive procurement of goods and services and [5] Socially productive use of land and property (Manley and Whyman, 2021). In addition, CWB has been defined as: ‘A local economic development strategy focused on building collaborative, inclusive, sustainable, and democratically controlled local economies. […] these include worker cooperatives, community land trusts, commu- nity development financial institutions, so-called ‘anchor institution’ procurement strategies, municipal and local public enterprises, and public and community banking’ (Guinan & O’Neill, 2020 p13-14.) CWB principles should be understood as resisting a globalised neoliberal economic system that has increasingly contracted out public services to (multinational) private companies, thereby reducing di- versity and ownership over local economic activity and empowering private financial and commercial institutions to own, manage and govern key public goods (Peters, 2012; Williams et al., 2014). This in- cludes energy and transport - key sectors for net zero transitions - that have also been privatised and liberalised in this way (Bayliss et al., 2021). Further, CWB principles can be connected to research advancing new conceptualisations of more progressive local economies, such as work on the ‘Foundational Economy’ (Heslop et al., 2019; Hansen, 2021), ’New Municpalism’ (Thompson, 2021) ‘Re-municipalisation’ (Cumbers, 2016; Paul and Cumbers, 2021), local policy responses to globalisation (Imbroscio et al., 2003) and critical work on the regressive impact of neoliberalism on localism (Catney et al., 2014; Davoudi and Madanipour, 2015). Finally, the intellectual influence of key US scholars is important here, particularly the influence of scholars such as Gar Alperovitz and Marjorie Kelly, who are both core members of the ‘Democracy Collab- orative’ think tank in the United States – a key advocate for CWB and a core actor at the heart of the successful ‘Cleveland model’ in the US (Lenihan, 2014). Going back half a century, Alperovltz (1972) coined the concept of a ‘pluralist commonwealth’ and continues to publish on the relevance of this concept for progressive economic reform today (Alperovitz and Dubb, 2014; Alperovitz, 2020). Seen as a precursor to CWB, the pluralist commonwealth is defined by four principles; the ‘democratisation of wealth’, ‘community as a guiding theme’, ‘decen- tralization’ and ‘democratic planning’ (Alperovitz, 2020). These prin- ciples demonstrate strong overlaps with Alperovitz and Dubb (2017), who drew upon these when mobilising for revitalising and regenerating Detroit, moving from theory to practice – a key hallmark of the CWB policy community. With reference to democratising ownership and the first principle of CWB, Kelly (2012), in ‘Owning our Future’, distinguishes between ‘extractive’ and ‘generative’ ownership forms. Extractive forms of ownership cater to an international shareholder class or ‘absentee membership’, where organisations - embedded in global capital markets and financialised networks - seek to move between a series of profit maximising opportunities in the short-term, above all other interests. The business generated by these investments are assumed to trickle-down to local actors. Generative ownership, in contrast, sees ‘rooted membership’ in local, public and civil society forms of organi- sation. Governance is controlled by those dedicated to a ‘social mission’ and organisations are constructed around both long-term and sustainability-oriented goals (Kelly, 2012). This generative/extractive distinction has influenced work on democratic economies (Kelly and Howard, 2019; Hanna and Kelly, 2021), whilst calls for more ‘genera- tive’ economies now appear amongst UK advocates for CWB (McInroy, 2020). Additionally, this distinction is also vital for civil society ap- proaches to net zero transitions and civic and community energy structures of organisation. Whilst clearly principled, at its core CWB is nevertheless deliberately pragmatic. Anchor institutions, such as universities, hospitals, schools, prisons, local government, housing associations, trade unions or large local companies/social enterprises, are all fixed in place and rooted to a locality or region by virtue of their organisational design, ’anchoring’ them to their local economies. Through pursuing a CWB approach, these anchor institutions seek to work in partnership with CWB organisations to switch their service contracts from multinational to local supply chains. Simultaneously, capacity is built up in local supply cooperatives and partnerships through coordinated facilitation and anchor networks. In Preston (UK), the promotion of a CWB approach by the local council has led to the percentage of total procurement spending in the city going up from 5% in 2013 to 18% in 2017, and from 39% in 2013 to 79% in 2017 across the Lancashire region (Jackson and McInroy, 2017). Un- employment has fallen from 6.5% in 2014 to 3.1% (O’Neill and Guinan, 2020). Preston has been named the most improved city in the UK ‘Good Growth for Cities 2018’ index and has moved from 143rd to 130th in the Social Mobility Commission Index. A further 4000 employees – including all council workers – now receive the Living Wage (Hadfield, 2019). In addition to this, a complex network of mutually supportive co-operatives and social enterprises has developed in Preston, under- pinned by the establishment of the Preston Cooperative Development Network, with the support of anchor institutions in the form of the local university and the local city council (Manley and Whyman, 2021). CWB thus positions itself as a pragmatically progressive form of bottom-up, locally-led social and economic development. And yet, there EnergyPolicy172(2023)1132772 M. Lacey-Barnacle et al. has seldom been investigations into how such achievements might connect to, complement or support the challenge confronting all local- ities – net zero transitions. We therefore ask the following two questions in this Perspective piece: 1. What synergies are there between CWB and civil society approaches to local net zero & sustainability transitions? 2. How can future research and policy support these synergies in practice? These questions are intended to open up lines of inquiry (and ac- tivity) into how CWB can engage in net zero transitions. In the next section, we argue that many of the CWB principles align well with emerging calls for just transitions in energy systems. 1.2. New frontiers: community wealth building and the green economy Whilst recent academic research into CWB analyses its potential in local economic development (Barnes et al., 2020; Manley and Whyman, 2021; Eder, 2021; Webster et al., 2021; Dubb, 2016), few of these studies explicitly address how CWB can engage with the green economy. Suc- cinct reviews of CWB for a general audience (O’Neill and Guinan, 2020), detailed essays on the history and future of CWB (Hanna and Kelly, 2021) and books devoted to engaging a wider audience in the history of the Preston Model (Brown and Jones, 2021) and deepening our aca- demic understanding of both the Preston Model and CWB (Manley and Whyman, 2021), all contain very little acknowledgement of its potential relevance to local net zero transitions. Furthermore, a practical 51-page toolkit designed to assist local councillors in implementing CWB con- tains only one mention of ‘net zero’ (Democracy Collaborative and Momentum, 2022). This is surprising, given that CWB is a trans- formational economic project and net zero transitions imply significant economic reorientations. Moreover, calls for just transitions to a net zero economy open an opportunity for CWB to enter into this terrain (Wang and Lo, 2021), alongside challenges to established notions of what constitutes a just transition, moving beyond a core focus on providing ‘green jobs’ in the face of a retracting fossil fuel industry (McCauley and Heffron, 2018), to understanding how justice, equity and inequality are constituted in new net zero economies (Morena et al., 2020) and how civil society and grassroots mobilisations for a just transition can be supported by the state (Routledge et al., 2018), specifically at the local level. There are signs that this disconnect between CWB and net zero transitions is beginning to be bridged, particularly in policy research. The think tank CommonWealth’s work on ‘Community Wealth Building for Economic and Environmental Justice’ (Brown et al., 2019) shows early signs of bringing together the two disconnected fields, arguing for anchor institutions to play pivotal roles in supporting local Green New Deals, green jobs and ‘green procurement policies’, whilst the Centre for Local Economic Strategies (CLES) report & toolkit on a ‘just energy transition through community wealth building’ (Radcliffe and Williams, 2021), alongside the Democracy Collaborative’s report on ‘Publicly owned and cooperative electric utilities as anchors for community wealth building and a just energy transition’ (Hanna et al., 2022) demonstrate the emergence of a new policy-research field. Turning briefly to these last two outputs, Radcliffe and Williams (2021) note a vital role for local authorities to intervene in energy transitions to advance CWB, where they play key roles in; (1) Acting as convenor (2) Creating demand (3) Direct delivery of transition projects (4) Encouraging the early adoption of zero carbon technology and (5) Funding the energy transition. The authors connect these roles to the five principles of CWB throughout the report, whilst also noting that anchor institutions have ‘a critical role in enabling cross-sector approaches to energy transition which build community wealth’ (Radcliffe and Wil- liams, 2021 p.14). This connects well to Hanna et al. (2022), who see ‘community utilities’ that are co-operatively and publicly owned as fundamental additions to the plethora of possible anchor institutions. The authors advance nine key policy recommendations for building community wealth in energy markets and transitions; (1) Block Priva- tisation (2) Deeper Democratic Governance (3) Renewable Energy Mandates (4) Renewable Energy Financial Incentives (5) Public Distributed Renewable Energy & Electrification (6) Procurement Pro- grams (7) Public Banking & Finance (8) Supporting Local Innovation and (9) Public finance for shifting ‘Investor Owned Utilities’ into Public and Co-operative Ownership. Both reports demonstrate renewed atten- tion being paid to critical connections between CWB and civil society-led energy transitions. Interestingly, older outputs from the Democracy Collaborative, such as Warren’s (2010) report entitled ‘Growing a Green Economy for All: from Green Jobs to Green Ownership’, pay attention to this juncture between CWB and green economy transitions, whilst the Cleveland model’s network of different organisations known as the ‘Evergreen Co- operatives’ supported local food growing, sustainable laundry and local solar PV deployment (Lenihan, 2014). As Sheffield (2017) reports, many of the Evergreen Cooperatives are now profitable, employing over 150 people locally, with plans to increase this number to 1000. In the case of the Evergreen Cooperative Laundry, for example: ‘After a six-month initial “probationary” period, employees begin to buy into the company through payroll deductions of 50 cents an hour over three years (for a total of $3,000). Employee-owners are likely to build up a $65,000 equity stake in the business over eight to nine years, a substantial amount of money in one of the hardest-hit urban neighbourhoods in the nation’ (Alperovitz et al., 2010 p.1). Indeed, this novel form of democratic ownership and governance – facilitated through laundry service contracts with anchor institutions (local hospitals and universities) - led Lenihan (2014) to describe the Cleveland model as; ‘The most robust ongoing American effort to enjoin the economic power of anchor institutions (and their growing ecological sensitivity) with the development goal of creating widely shared and more democratic asset and capital building in low-income neighborhoods’ (Lenihan, 2014 p.18 p.18) Despite this promising connection with sustainable transitions - both past and present - academic research seems to be severely lagging behind. We argue that the introduction of CWB research into the energy transitions terrain presents scope for facilitative links with established civil society approaches and theories in bottom-up and local energy transitions, such as civic energy sector transition pathways (Foxon, 2013) and grassroots innovations that seek to directly tackle the chal- lenges of sustainability transitions from the bottom-up (Smith and Seyfang, 2013; Smith et al., 2016). In the following section, we analyse more closely the links between CWB and these relevant approaches to civil society-led energy transitions. 2. Community wealth building and energy transitions: theoretical links In this section, we first explore the ways in which synergies already exist between CWB and two widely cited civil society approaches for local sustainability transitions: The Thousand Flowers transition pathway and its associated concept of a ‘Civic Energy Sector’ and the theory of Grassroots Innovations. A variety of related research fields could addi- tionally be explored, such as research on decentralised ownership and control over energy systems (Brisbois, 2019), polycentric governance (Bauwens, 2017), local community power (Kaye, 2020) and a rich his- tory of community energy research (Lacey-Barnacle, 2020; Creamer, 2018; Smith et al., 2016; Seyfang et al., 2013; Walker et al., 2007). All of these fields connect to both the Thousand Flowers and Grassroots In- novations literature; however, a review of more comprehensive links to CWB is beyond the scope of our paper. Our examination here, of the EnergyPolicy172(2023)1132773 M. Lacey-Barnacle et al. different ways in which two illustrative approaches in community-based energy developments and CWB share similar goals, values and ap- proaches, can inform future bridge-building research endeavours. ownership forms that advance direct community control, alongside providing inclusive energy tariff offers and energy efficiency services to vulnerable groups (Hoicka et al., 2021; Hanke et al., 2021). 2.1. Transition pathways and the civic energy sector Influenced by work on socio-technical transitions and the multi-level perspective on systems transformations (Verbong & Loorbach 2012; Geels 2002), the ‘Realising Transition Pathways’ research consortium, an 8-year multi-institution project spanning 2008–2016, produced considerable material and research outputs to assist UK government policymakers and academic research communities in grappling with the complexities of transitioning to a low-carbon energy system by 2050. The Pathways project developed detailed potential paths that would achieve this momentous transition. Associated outputs analysed how to ‘bring social structures and agency, including institutions and politics, into scenario […] studies of sustainable energy futures’ (Foxon, 2013 p.12). These scenarios enhanced understanding of the political and economic challenges and opportunities in UK low-carbon futures (compared to the technology-dominant scenarios in many energy scenarios and pathways studies). Different institutional and socio-technical configurations were explored for meeting the UK’s legally binding commitment (in the Climate Change Act 2008) to reduce GHG emissions by 80% by 2050 against a 1990 baseline. Three different transition pathways were con- trasted; Market Rules, Central Co-ordination and Thousand Flowers. Each pathway adheres to different governance logics in which power relations between market, state and civil society actors are varied (Foxon et al., 2010; Foxon, 2013; Barnacle et al., 2013; Chilvers et al., 2017). The Thousand Flowers pathway provides one of the few detailed ex- plorations of the greater role that civil society can play within future UK energy transitions. The pathway sees a ‘growing dominance of civil society in the governance of UK energy systems, which leads to an increase in di- versity of local bottom-up solutions for providing decentralised generation and energy conservation options’ (Barnacle et al., 2013 p.60). One outcome of this growing role for civil society in municipal and com- munity governance of energy, is the development of a ‘civic energy sector’, a scenario which delivers 50% of final electricity demand by 2050 (Hall et al., 2016). Central to this vision is a vibrant community energy sector, where community organisations take a leading role in purchasing, managing and governing local energy projects and infrastructures. A heavily researched field of both policy and practice, community energy has very often been seen by many researchers as particularly competent in meeting varying social, environmental and economic objectives at the local level (Zoellner et al., 2008; Warren, 2010; Musall & Kuik 2011; Seyfang et al., 2013). For example, community energy projects have encouraged and enabled the active participation of members of the local community in energy transition processes, while introducing behaviour change schemes and energy demand reduction into local communities. Secondly, many schemes have drawn upon local investment and tapped into local expertise and enthusiasm for renewable energy installations, raising the necessary capital and increasing local acceptance through direct community ownership. The wealth generated by newly-valuable renewable resources thereby circulates and multiplies more locally. Civic initiatives cultivate multi-actor partnerships working across mul- tiple scales to engage in and support transitions, and, using this multi-scalar collaboration, have been able to appropriately tailor local renewable energy deployment to the technological, political and eco- nomic specificities of a locality (Walker et al., 2007; Walker et al., 2007; Seyfang et al., 2013; Hargreaves et al., 2013; Bauwens et al., 2016). Importantly, the emergence of community and civic energy schemes is now influencing policy. For example, as part of the EU’s Clean Energy Package, ‘Energy Communities’ are now formally recognised as essential civil society entities which will aid the EU’s broader decarbonisation plans. Recent research also points towards their potential to contribute to a more just and democratic transition, particularly through novel Many of the above elements of local and community energy effec- tively align with CWB approaches, whilst also encouraging the broader empowerment of civil society actors. In a review of community energy projects in Europe, Hewitt et al. (2019) note that four aspects of com- munity energy projects underpin their potential for contributing to- wards social innovation; (1) Crises and opportunities; (2) the agency of civil society; (3) reconfiguration of social practices, institutions and networks; (4) new ways of working. All four of these aspects of com- munity energy schemes connect closely to CWB. The trigger for CWB in Preston, for example, was the collapse of a £700m inward-investment regeneration project in the wake of the global financial crisis, and therefore, the search for locally-resilient opportunities to develop the local economy resulted in a CWB approach (and inspired by the US Cleveland model) (Manley and Whyman, 2021). CWB typically seeks to enhance and empower the agency of civil society within multi-actor partnerships and to reconfigure institutions and networks, whilst the five principles of CWB foster new ways of forging those relations at the local scale. Importantly, community energy connects with CWB by seeking to localise and retain wealth and surplus revenue creation (Lacey-Barnacle, 2019; Stewart, 2021), democratise governance and engagement in local economies (Van Veelen, 2018) and experiment with novel social enterprise models and organisational structures (Becker et al., 2017). Forming a core part of the civic energy sector as outlined in the Thousand Flowers pathway, community energy schemes can be considered a vital part of local strategies to build ‘community wealth’. However, as we explore in subsection 2.2, this wealth is not always equitably shared and CWB may offer a point of strategic intervention to address more equitably some historic shortcomings in civic energy approaches. Whilst civil society is crucial, this does not negate roles for the state or market. Barton et al. (2015) note, through the prism of back-casting, that the Thousand Flowers pathway shifts the role of local government, as: ‘Local energy ownership became a focus of local government economic development […] as the scale of the opportunity became clear in terms of local value capture, net employment creation, and energy security’ (Barton et al., 2015 p.5 p.5) This ‘local value capture’ connects well to the redirection of pro- curement processes in CWB advocacy; whilst the focus on local energy ownership also demonstrates synergies with the ‘plural ownership of the economy’ principle. Indeed, the authors note that the ‘expansion of this sector would capture much of the value from energy production and con- sumption that currently leaks out of the local economy’ (Barton et al., 2015 p.27), demonstrating strong support for wealth retention within local economies. Furthermore, when anticipating how the Thousand Flowers pathway is achieved, the authors note that ‘local energy schemes devel- oped stable and familiar financial relationships with the local banking sector, which viewed civic power generation as a safe asset’ (Barton et al., 2015 p.5), connecting strongly to the CWB principle centred on making financial power work for local places. Drawing on the example of Germany as a ‘co-ordinated market economy’ (Hall and Soskice, 2001), Hall et al. (2016) show the impor- tance of the German local banking sector in facilitating civic ownership structures. This is in contrast to the UK neo-liberal economic model, in which financial institutions have a national and international focus and arguably are more motivated by short-term shareholder returns than long-term stable investment relationships with local partners. Interest- ingly, new bottom-up financial innovations, in the form of local municipal energy bonds, are now being developed in the UK (Davis, 2021; Green Finance Institute, 2021). These provide a simple, low-risk way to enable members of local communities to invest in local renew- able energy developments, by making use of the financial security of EnergyPolicy172(2023)1132774 M. Lacey-Barnacle et al. local municipal authorities. This approach could thus contribute to the second principle of CWB, whilst also allowing local financial innovation to be governed and managed by public institutions. Indeed, many of these crossovers between CWB and the Thousand Flowers pathway show that new local energy supply models have the potential to incorporate more complex value propositions, including economic, social and environmental values (Hall and Roelich, 2016). Intriguingly, the role of anchor institutions in leveraging procure- ment spending in support of local net zero innovation, projects and goals, has been understudied in civic energy research (Uyarra et al., 2016). The Thousand Flowers pathway does not conceive of anchor in- stitutions in its detailed scenarios. In identifying key anchor institutions, such as local hospitals, universities and local government, CWB brings another mechanism to civic and community energy that can facilitate novel contractual arrangements to support the growth of local net zero energy projects and supply chains: contracting energy co-operatives to provide energy consulting services, supplies of clean electricity, effi- ciency measures, and supporting community flexibility arrangements in smart local energy systems. Energy transitions could also form a more explicit part of what CWB scholars call the ‘anchor mission’ (Kelly et al., 2016), where their local economic power is used to strengthen local enterprise, with a focus on socially inclusive organisations. Here, through aligning anchor missions with net zero transitions, anchor in- stitutions can be used to offer preferential treatment to organisations that simultaneously pursue inclusive decarbonisation. 2.2. Grassroots innovations, local sustainability transitions and CWB In contrast to future scenario conditions under which empowered civic energy generation might become more widespread, research into grassroots innovation was borne of historical and contemporary analysis into innovative local sustainability initiatives. These often develop despite existing realities being unconducive to such initiatives (Seyfang and Smith, 2007; Fressoli et al., 2014). Local environmental initiative was reframed as grassroots innovation, in which networks of neigh- bours, activists, social entrepreneurs, community organisations, co- operatives, and others worked creatively and innovatively in generating and circulating bottom-up solutions for sustainability appropriate to the needs, aspirations and situations of those involved. In conceiving local environmental activity as innovative and gener- ative of wider change, so studies were able to adapt analytical resources in innovation studies and sustainability transitions. This enabled better understanding of how grassroots movements produce knowledge, reframe problems, form networks and attract resources, govern them- selves and challenge institutions, and thereby develop and diffuse ap- proaches and solutions for sustainability across localities in ways quite different to conventional market- and state-based institutions for inno- vation (Hess, 2007; Jamison, 2001; Smith and Stirling, 2018). Early research (in the 2010s) included studies of community energy, analysis of grassroots innovation in food, housing, manufacturing, mobility; as well as historical research into earlier movements for alternative tech- nology, socially useful production; and initiatives in the global South as well as global North (Smith et al., 2017; Pansera and Owen, 2017; Gupta, 2016). Theories about the development of ‘niche spaces’ for alternative innovation within the context of unfavourable incumbent energy regimes were used to explain the achievements and challenges confronting grassroots action (Smith, 2007). Grassroots Innovations can seek to change markets and prevailing market systems, despite sometimes being framed as an alternative to the market or as a more radical response to the failure of dominant and mainstream institutions on environmental issues (Feola and Nunes, 2014; Seyfang and Smith, 2007). They do this through the utilisation of a set of unique characteristics that set them apart from market and technology-oriented niche innovations (Fressoli et al., 2014). In the context of community energy, Hargreaves et al. (2013) identify these unique characteristics as: ‘Distinct organisational forms’; ‘Different resource bases’; ‘Divergent contextual situations’; ‘Alternative driving motivations’; and ‘the pursuit of qualitatively different kinds of sus- tainable development’ (Hargreaves et al., 2013). Indeed, prominent theorists of Grassroots Innovations suggest that, whilst it is particularly hard to correlate similarities across cases of local innovation that are by definition tailored to the specificities of a locality, many grassroots in- novations will draw upon social enterprise models or function more broadly within the social economy (Seyfang and Smith, 2007; Har- greaves et al., 2013; Smith, 2014). Thus, it is important to note that: ‘Grassroots innovation processes share a broadly similar vision and shared set of principles, regarding local inclusion and control in processes of technology development and innovative social organisation […] grassroots innovation movements confront similar fundamental chal- lenges, even though manifesting in particular ways in contrasting settings’ (Smith, 2014 p.115) Here, we can already see some strong connections to the five prin- ciples of CWB. Firstly, the use of ‘distinct organisational forms’ to sup- port grassroots innovations opens bridges to the demand for more ‘plural ownership of the economy’ by CWB advocates. Arguably, grassroots innovation has tended to gloss over questions of ownership and attended more to participation, so more explicit engagement with diversifying ownership, in line with CWB, can provide more depth here. Secondly, the reliance of CWB approaches on anchor institutions and the redi- rection of procurement processes to support local economies connects well to the reliance of grassroots innovations on ‘different resource bases’, which is supported further by the local financial innovation sought by CWB actors. Lastly, the desire that CWB advocates have for new models of local economic development that cater to the needs of different localities are reflective of the ‘alternative driving motivations’ and that underpin grassroots innovations. ‘divergent contextual situations’ While these multiple connections are important, there is one incon- sistency. Differences between CWB and grassroots innovations are found in the limited engagement of CWB literature in sustainability transitions and the importance of path-breaking innovations for future net zero economies, whilst grassroots innovations are often explicitly framed around contributing towards ‘different kinds of sustainable develop- ment’. Innovation and transformation as a goal and topic is not so prominent in CWB practice, where activity rests in carving out oppor- tunities within the given local economy. And yet, the five principles imply considerable organisational, business, process and product inno- vation, and even some changes to the contexts and purposes for tech- nological change which is the conventional focus of innovation. If CWB succeeds in bringing in a diversity of actors into local economic devel- opment (e.g. via anchor networks), the insights from grassroots inno- vation concerning how these alternative constellations can better approach innovation and transition could prove helpful. For example, a dilemma typical for many grassroots innovation movements seeking to scale-up, circulate more widely, and generally expand their niche innovations, is whether to align more closely with the logics of incumbent institutions for innovation (such as through commodification, intellectual property, and standardisation, thereby blunting their transformational potential) or to remain radical and continue pressing for radical reforms to powerful institutions. Such radical reforms ensure that innovation is conceived and supported using the participatory democratic norms and commons-based ownership models favoured by grassroots innovations. Dynamic tensions exist between ‘fit-and-conform’ versus ‘stretch- and-transform’ strategies for developing niche spaces: making them more palatable to prevailing institutions, or building power to transform those institutions (Smith and Raven, 2012). Analogous dilemmas might EnergyPolicy172(2023)1132775 M. Lacey-Barnacle et al. be evident in relations between local economic enterprises and anchor institutions who, no matter how sympathetic to worker control, say, or cooperative ownership, might be structurally constrained as to how far they can depart from norms of supply-chain and service-provision under capitalism as currently instituted (Smith, 2014). Without a better appreciation of the complexities of transformative innovation, there is a risk that CWB measures will tend towards safe, conservative economic activities or privilege the experimental designs of organisations with the resources to instigate them. That said, an enduring challenge is moving beyond creative prototypes and start-up organisational forms, to build enduring structures and institutions capable of enabling these novelties to succeed over the long-term (as seen with civic energy generation in the Thousand Flowers pathway). It might be that moving from innovation to diffusion, in ways that remain transformational and resist falling into conformity, is a challenge where insights from CWB can be helpful. CWB can help cultivate the capabilities, investment and work to develop innovation more consistently with motivating ideals: so, for example, community energy schemes remain locally democratic and accountable, rather than becoming increasingly utility-like. This is a goal that Hanna et al. (2022) say is fundamental to a CWB approach to energy transitions. However, accountability is not the only issue facing community en- ergy schemes. Community energy risks replicating issues around social inclusion. For example, researchers such as Catney et al. (2014) and Seyfang et al. (2013), when offering critical perspectives on community energy projects, note that much of the literature surrounding commu- nity energy focuses explicitly on the success stories of the sector, with little attention given to understanding which communities are unable to engage in these initiatives and why, leaving out considerations of how to bring about a more socially ‘just’ transition. Furthermore, Johnson et al. (2014) find that a decentralised energy system could risk reproducing, or even worsening, existing socio-economic inequalities within society. It is important to ask, therefore, whether CWB may encounter similar risks. Given the primacy of the local state and the need for representative political leadership to support CWB, we feel a CWB approach could avoid the pitfalls of a socially exclusive local economic development approach. Social justice and inclusion considerations are of vital importance to emerging CWB policy programmes and approaches, which we feel could be used to address and rectify some of the existing inequalities in access to community energy schemes and to advance a more inclusive just transition. 2.3. Comparing CWB, grassroots innovation and thousand flowers pathway approaches The above discussions suggest that CWB may be able to offer a normative direction to the kinds of transformation that many have argued are necessary for net zero transitions (e.g. more democratic, just, community-based, socially inclusive etc), addressing areas where soci- otechnical energy transitions research has been agnostic and lacking. CWB, aided by its five principles, also has the potential to inject local, municipal and community energy with stronger elements of democratic directionality, underpinned by a strong social justice ethos (Kelly et al., 2016). CWB might thus be a counter to the financialisation and extraction of local energy initiatives that comes with private institu- tional investment and corporate control, whilst – with the support of anchor institutions - offering stability and finance to develop more democratic and plural economic organisations. Linking back to our discussion of the Thousand Flowers pathway, there is clearly a key role for anchor institutions to play in a future where CWB becomes more closely aligned with the transition to a net zero economy. Drawing on our analysis above, we further summarise the critical overlaps and synergies between CWB and grassroots innovations and Thousand Flowers transition pathway in our table below (Table 1): Table 1 Overlaps between CWB principles and civil society approaches to transitions. CWB Principles Grassroots innovations Thousand Flowers pathway [1] Plural ownership of the economy Distinct organisational forms [2] Making financial power work for local places [3] Fair employment and just labour markets [4] Progressive procurement of goods and services Divergent contextual situations/Different resource bases Alternative driving motivations Different resource bases [5] Socially productive use of land and property Different kinds of sustainable development Dominance of civil society in the governance of UK energy systems Key financial relationships between the local banking sector & civic energy sector Net employment creation Local value capture/capture local value of energy production and consumption Local energy ownership a focus of local government economic development 3. Conclusion – The future of community wealth building and civil society-led just transitions CWB is emerging at a timely and critical juncture; given its recent expansion over the past decade, it is already demonstrating more dem- ocratic forms of local economic development, with potential for making more just a rapidly expanding net zero economy. This opens up new pathways and future scenarios for radical, diverse and pragmatic ap- plications of CWB to net zero economies. Our Perspective piece has outlined the need for CWB to see the transition to a green economy as a novel opportunity to expand its activities and scope, particularly as global trends towards decentralised net zero transitions and devolved governance continues apace (Rodríguez-Pose and Gill, 2003; Burger et al., 2020). This is particularly true for local and decentralised energy markets, where new innovations, technologies and organisations are hastening the shift from centralised to decentralised energy systems. Smart local energy systems, locational pricing, bespoke tariffs, peer-to-peer trading of local surplus electricity, flexibility markets and community opportunities for energy storage, alongside engagement in vehicle-to-grid markets, are just some of the net zero innovations that can be integrated into CWB strategies for engagement in energy markets. In this space, public and community ownership of novel technologies and platforms will be key to contributing to CWB synergies with net zero. Research on the financial benefits of community ownership shows that community-owned wind farms pay their communities 34 times more than their commercial (private) counterparts (Aquatera, 2021), whilst co-operative and community energy schemes are more effective in connecting the benefits of low-carbon technologies to deprived communities than individual, household models of deployment (Stew- art, 2021). Our Perspective thus has strong normative underpinnings; we see the presence of more plural, democratic, public and civil society forms of ownership and governance in net zero economies as constituting more ‘just’ forms of organisation, particularly when citizens and workers at the heart of local communities and economies are given greater auton- omy and agency in the face of historic corporate and state control over the expansion of the green economy. Understandings of a ‘just transi- tion’ cannot, therefore, be divorced from broader questions of owner- ship and governance in our economy and further explorations of civil society-led pathways to a just transition are vital. As we have high- lighted, anchor institutions and emerging anchor networks, supported by the local state, will be key in designing justice interventions and advancing social justice aims, where marginalised and deprived com- munities are placed at the forefront of local green recovery and regen- eration strategies. This is a vital area of future research for both the CWB and just transitions research communities. Moreover, CWB opens up wider discussions on the role of local EnergyPolicy172(2023)1132776 M. Lacey-Barnacle et al. economic democracy, its potential, its limits, and how it may effectively engage with net zero transitions and reshape our conception of a just transition. While both grassroots innovations and transition pathway literatures have acknowledged democratic ownership forms in sustain- ability transitions, often through the guise of a civic energy sector and community energy schemes, CWB has the potential to highlight how a local-state-backed form of economic democratisation can strengthen these endeavours, by drawing on successful examples and empirical analysis emerging across the world. However, cautioned by analysis of grassroots innovation alluded to above, CWB may, without due atten- tion, succumb to a ‘fit and conform’ strategy, where CWB is ultimately used to reinforce existing market-oriented power structures through using procurement to support local businesses and supply chains, which do not attend to decarbonisation and sustainability goals. In contrast, if CWB pursues a ‘stretch and transform’ approach, it can be used to move economic democratisation towards the heart of sustainable energy transitions, incorporating stronger social justice goals as outlined above. Despite this positive outlook, we do, however, exercise caution here; it is vital to not advocate for CWB uncritically. As Manley and Whyman (2021) point out, CWB has top-down tendencies whenever its goals and visions are set by local political leaders, rather than through local citi- zens and civil society deliberation. Ensuring economic democratisation in CWB is supported by appropriate development networks and educa- tional schemes has the potential to counter any technocratic tendency. This is vitally important, as research points towards a key role for eco- nomic democracy in enhancing both equality and sustainability in so- ciety (Power et al., 2016). Thus, whilst our paper has championed the possibilities of bridging with just sustainability transitions, we conclude by acknowledging three key challenges for CWB and just transitions that future action-research must address to assess its feasibility and potential to promote trans- formative change: (1) Linking up CWB and just transitions policies and strategies - Although CWB is receiving local and regional policy support across the globe, support for just transitions appears at multiple levels of governance globally and plays host to political support and buy-in at a broader scale than CWB. With broader top-down policy support and financing, CWB could arguably be more transformative and impactful. There is significant space in both research and policy to explore linkages, alignments and com- plementarities between CWB and just transitions to a net zero economy, particularly in ‘left behind’ areas, regions and com- munities that are seeking bold regeneration strategies after de- cades of deindustrialisation. (2) Local financial innovation - The role of finance is critical in decarbonising the economy and many of the financial mecha- nisms and innovations required for current net zero targets are beyond the reach of CWB, particularly in highly centralised financial systems. However, through the redirection of procure- ment practices that CWB advocates for, there is potential to redirect local spending towards climate goals and experiment with local financial innovation, such as Community Municipal Bonds (Davis, 2021), to support local decarbonisation and local wealth retention in net zero transitions. Research into how to unlock and access finance locally to support CWB approaches to new net zero economies will be vital in coming years, alongside exploring further how local financial innovations can support and link up to broader just transition concerns. (3) Anchor institutions supporting just net zero transitions – it is clear that anchor institutions, with their associated procurement power and natural embeddedness within place-based economies, will have a vital role to play in ensuring they use their local economic power to support grassroots innovations and civic en- ergy projects. The expansion of an ‘anchor mission’ – to include inclusive and sustainable local enterprise – will be fundamental to this challenge. Anchor institutions should ideally give prefer- ential treatment to democratic organisational structures in their economic developments. This would contribute towards the transformational potential that CWB promises to local economies across the world. These three challenges overlap. Local just transitions strategies, emboldened by CWB agendas, will need to ensure that key anchor in- stitutions support local financial innovation and leverage procurement spending to advance CWB approaches to net zero economies, whilst also supporting economic democratisation as part of reconceptualised just transitions. The novel insights offered in this Perspective suggest such an endeavour is worth embarking upon, in research, policy and practice. CRediT authorship contribution statement M. Lacey-Barnacle: Conceptualization, Supervision, Project administration, Writing – original draft, Writing – review & editing, Funding acquisition. A. Smith: Conceptualization, Writing – original draft, Writing – review & editing. T.J. Foxon: Conceptualization, Writing – original draft, Writing – review & editing. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Data availability No data was used for the research described in the article. Acknowledgements The authors wish to thank the reviewers for their time and their insightful comments which helped improve the paper. This research was supported by a Leverhulme Trust Early Career Research Fellowship ECF- 2021-191, of which the lead author is a recipient. References Alperovitz, G., 2020. A pluralist commonwealth and a community sustaining system. In: The New Systems Reader. Routledge, pp. 226–252. Alperovitz, G., Williamson, T., Howard, T., 2010. The Cleveland model. Nation 1 (1), 21–24. Alperovitz, G., Dubb, S., 2014. The possibility of a pluralist commonwealth and a community sustaining economy. In: From Capitalistic to Humanistic Business. Palgrave Macmillan, London, pp. 73–96. Alperovltz, G., 1972. Notes towards a pluralist commonwealth. Rev. Radic. Polit. Econ. 4 (2), 28–48. Alperowitz, G., Dubb, S., 2017. A community wealth-building vision for Detroit—and beyond. In: Reinventing Detroit. Routledge, pp. 183–200. Aquatera, 2021. A Comparison of the Financial Benefits Arising from Private and Community Owned Wind Farms. Report by Aquatera Ltd. Report for Point and Sandwick Trust. Barnes, M.C., Walker, C.D., Williamson, T., 2020. Community Wealth Building and the Reconstruction of American Democracy: Can We Make American Democracy Work? Edward Elgar Publishing. Barnacle, M., Robertson, E., Galloway, S., Barton, J., Ault, G., 2013. Modelling generation and infrastructure requirements for transition pathways. Energy Pol. 52, 60–75. Barton, J., Emmanuel-Yusuf, D., Hall, S., Johnson, V., Longhurst, N., O’Grady, A., Sherry- Brennan, F., 2015. Distributing Power: A Transition to a Civic Energy Future. Bauwens, T., 2017. Polycentric governance approaches for a low-carbon transition: the roles of community-based energy initiatives in enhancing the resilience of future energy systems. In: Complex Systems and Social Practices in Energy Transitions. Springer, Cham, pp. 119–145. Bauwens, T., Gotchev, B., Holstenkamp, L., 2016. What drives the development of community energy in Europe? The case of wind power cooperatives. Energy Res. Soc. Sci. 13, 136–147. Bayliss, K., Mattioli, G., Steinberger, J., 2021. Inequality, poverty and the privatization of essential services: a ‘systems of provision’ study of water, energy and local buses in the UK. Compet. Change 25 (3–4), 478–500. EnergyPolicy172(2023)1132777 M. Lacey-Barnacle et al. Becker, S., Kunze, C., Vancea, M., 2017. Community energy and social entrepreneurship: addressing purpose, organisation and embeddedness of renewable energy projects. J. Clean. Prod. 147, 25–36. Brisbois, M.C., 2019. Powershifts: a framework for assessing the growing impact of decentralized ownership of energy transitions on political decision-making. Energy Res. Social Sci. 50, 151–161. Harvey, D., 2007. A Brief History of Neoliberalism. Oxford University Press, USA. Hewitt, R.J., Bradley, N., Baggio Compagnucci, A., Barlagne, C., Ceglarz, A., Cremades, R., et al., 2019. Social innovation in community energy in Europe: a review of the evidence. Front. Energy Res. 7, 31. Heslop, J., Morgan, K., Tomaney, J., 2019. Debating the foundational economy. Renewal: J. Social Democracy 27. Brown, M., Jones, R., 2021. Paint Your Town Red: How Preston Took Back Control and Hess, D., 2007. Alternative Pathways in Science and Industry. The MIT Press, Cambridge, Your Town Can Too. Repeater Books. MA. Brown, G., Leibowitz, J., McInroy, N., 2019. Community Wealth Building for Economic Hoicka, C.E., Lowitzsch, J., Brisbois, M.C., Kumar, A., Camargo, L.R., 2021. and Environmental Justice. Centre for Local Economic Strategies. Burger, C., Froggatt, A., Mitchell, C., Weinmann, J., 2020. Decentralised Energy. Ubiquity press. Catney, P., MacGregor, S., Dobson, A., Hall, S.M., Royston, S., Robinson, Z., et al., 2014. Big society, little justice? Community renewable energy and the politics of localism. Local Environ. 19 (7), 715–730. Chilvers, J., Foxon, T.J., Galloway, S., Hammond, G.P., Infield, D., Leach, M., et al., 2017. Realising transition pathways for a more electric, low-carbon energy system in the United Kingdom: challenges, insights and opportunities. Proc. IME J. Power Energy 231 (6), 440–477. Collord, 2019. Building community wealth globally: the Kigoma-Preston collaboration. https://thenextsystem.org/learn/stories/building-community-wealth-globally -kigoma-preston-collaboration. Creamer, et al., 2018. Community Energy: Entanglements of Community, State & Private Implementing a just renewable energy transition: policy advice for transposing the new European rules for renewable energy communities. Energy Pol. 156, 112435. Imbroscio, D.L., Williamson, T., Alperovitz, G., 2003. Local policy responses to globalization: place-based ownership models of economic enterprise. Pol. Stud. J. 31 (1), 31–52. Jackson, M., McInroy, N., 2017. Community Wealth Building through Anchor Institutions. CLES. https://cles.org.uk/our-work/publications/community-wealth -building-throughanchor-institutions. Jamison, A., 2001. The Making of Green Knowledge. Environmental Politics and Cultural Transformation. Cambridge University Press, Cambridge. Johnson, V.C., Hall, S., Barton, J., Emanuel-Yusuf, D., Longhurst, N., O’Grady, ´ A., Robinson, E., 2014. Community energy and equity: The distributional implications of a transition to a decentralised electricity system. People Place Pol. 8 (3), 149–167. Kaye, S., 2020. Think Big, Act Small: Elinor Ostrom’s Radical Vision for Community Sector Geography Compass. Power (New Local). CLES, 2021. The Brave. Centre for Local Economic Strategies, UK. https://cles.org. Kelly, M., 2012. Owning Our Future: the Emerging Ownership Revolution. Berrett- uk/blog/the-brave/. Koehler Publishers. Cumbers, A., 2016. Remunicipalization, the low-carbon transition, and energy democracy. In: State of the World. Island Press, Washington, DC, pp. 275–289. Davis, M., 2021. Community municipal investments: accelerating the potential of local net zero strategies. https://eprints.whiterose.ac.uk/175824/1/2021_07_01_CMB- Innovate-UK-Report-Web-2.pdf. Davoudi, S., Madanipour, A. (Eds.), 2015. Reconsidering Localism. Routledge. Democracy Collaborative & Momentum, 2022. Community wealth building: a toolkit for local councillors. Prepared for People’s Momentum by The Democracy Collaborative. https://peoplesmomentum.com/wp-content/uploads/2022/04/C WB_MTM_11_April.pdf. Kelly, M., McKinley, S., Duncan, V., 2016. Politics of place/politics for places: community wealth building: America’s emerging asset-based approach to city economic development. Renewal: J. Labour Politics 24 (2), 51. Kelly, M., Howard, T., 2019. The Making of a Democratic Economy: Building Prosperity for the Many, Not Just the Few. Berrett-Koehler Publishers. Kohn, M., 2020. Community wealth building: lessons from Italy. In: Community Wealth Building and the Reconstruction of American Democracy. Edward Elgar Publishing. Lacey-Barnacle, M., 2019. Exploring Local Energy Justice in Times of Austerity: Civic Energy Sector Low-Carbon Transitions in Bristol City (Doctoral Dissertation. Cardiff University. Lacey-Barnacle, M., 2020. Proximities of energy justice: contesting community energy Dubb, S., 2016. Community wealth building forms: what they are and how to use them at and austerity in England. Energy Res. Social Sci. 69, 101713. the local level. Acad. Manag. Perspect. 30 (2), 141–152. Lenihan, R., 2014. Rust-belt recovery: the Cleveland model as economic development in Eder, J., 2021. Decreasing Dependency through Self-Reliance: Strengthening Local Economies an age of economic stagnation and climate change. Pepperdine Pol. Rev. 7 (1), 6. through Community Wealth Building (No. 124). ICAE Working Paper Series. Fensham, P., 2020. Community wealth building in Australia. https://cles.org.uk/publi cations/community-wealth-building-in-australia/. Manley, J., Whyman, P.B. (Eds.), 2021. The Preston Model and Community Wealth Building: Creating a Socio-Economic Democracy for the Future. Routledge. Morena, E., Krause, D., Stevis, D., 2020. Just Transitions. Social Justice in a Low-Carbon Feola, G., Nunes, R., 2014. Success and failure of grassroots innovations for addressing World. Pluto Press, London. climate change: The case of the Transition Movement. Glob. Environ. Change 24, 232–250. Foxon, T.J., 2013. Transition pathways for a UK low carbon electricity future. Energy Pol. 52, 10–24. Foxon, T.J., Hammond, G.P., Pearson, P.J., 2010. Developing transition pathways for a low carbon electricity system in the UK. Technol. Forecast. Soc. Change 77 (8), 1203–1213. Fressoli, M., et al., 2014. When grassroots innovation movements encounter mainstream institutions: implications for models of inclusive innovation. Innovat. Dev. 4 (2), 277–292. https://doi.org/10.1080/2157930X.2014.921354. Green Finance Institute, 2021. Local climate bonds. https://www.greenfinanceinstitute. co.uk/programmes/ceeb/lcbs/. Guinan, J., O’Neill, M., 2020. The Case for Community Wealth Building. Polity Press, Cambridge. McCauley, D., Heffron, R., 2018. Just transition: integrating climate, energy and environmental justice. Energy Pol. 119, 1–7. McInroy, N., 2020. Community is economic energy we must use it. https://www.lgcplus. com/politics/coronavirus/neil-mcinroy-community-is-economic-energy-we-must-us e-it-29-06-2020/. Paul, F.C., Cumbers, A., 2021. The return of the local state? Failing neoliberalism, remunicipalisation, and the role of the state in advanced capitalism. Environ. Plann.: Econ. Space, 0308518X211050407. Pansera, M., Owen, R., 2017. ‘Framing Inclusive Innovation within the Discourse of Development: Insights from Case Studies in India’, Research Policy, vol. 47. Elsevier, pp. 23–34. https://doi.org/10.1016/j.respol.2017.09.007 (1). Peters, J., 2012. Neoliberal convergence in North America and Western Europe: fiscal austerity, privatization, and public sector reform. Rev. Int. Polit. Econ. 19 (2), 208–235. Guinan, J., O’Neill, M., 2019. From community wealth-building to system change. IPPR Power, M., Wilkinson, R., Pickett, K., 2016. Inequality, Economic Democracy and Progressive Rev. 25 (4), 382–392. Gupta, A.K., 2016. Grassroots Innovation. Penguin, Gurgaon. Hadfield, M., 2019. What has the Preston Model achieved so far. Coop News. http Sustainability, vol. 37. World social science report, p. 160. Radcliffe, E., Williams, L., 2021. Toolkit: a Community Wealth Building Energy Transition. Centre for Local Economic Strategies. s://www.thenews.coop/143245/topic/economy/what-has-the-preston-model-achie ved-so-far/. Rodríguez-Pose, A., Gill, N., 2003. The global trend towards devolution and its implications. Environ. Plann. C Govern. Pol. 21 (3), 333–351. Hall, P.A., Soskice, D., 2001. Varieties of Capitalism: the Institutional Foundations of Routledge, P., Cumbers, A., Derickson, K.D., 2018. States of just transition: realising Comparative Advantage. Oxford University Press. Hall, S., Foxon, T.J., Bolton, R., 2016. Financing the civic energy sector: how financial institutions affect ownership models in Germany and the United Kingdom. Energy Res. Social Sci. 12, 5–15. Hall, S., Roelich, K., 2016. Business model innovation in electricity supply markets: the role of complex value in the United Kingdom. Energy Pol. 92, 286–298. Hanna, T.M., 2019. Community Wealth Building and Resilient Local Economies: the Role of Anchor Institutions. Public Finance for the Future We Want. Transnational Institute. Hanna, T.M., Kelly, M., 2021. Community Wealth Building: the Path towards a Democratic and Reparative Political Economic System. climate justice through and against the state. Geoforum 88, 78–86. Seyfang, G., Park, J.J., Smith, A., 2013. A thousand flowers blooming? An examination of community energy in the UK. Energy Pol. 61, 977–989. Seyfang, G., Smith, A., 2007. Grassroots innovations for sustainable development: towards a new research and policy agenda. Environ. Polit. 16 (4), 584–603. Available at: http://www.informaworld.com/10.1080/09644010701419121. Sheffield, H., 2017. The Preston model: UK takes lessons in recovery from rust-belt Cleveland. Guardian 11. Smith, A., 2014. Technology networks for socially useful production. J. Peer Product. (5) Available at: http://peerproduction.net/issues/issue-5-shared-machine-shops/ peer-reviewed-articles/technology-networks-for-socially-useful-production/. Hanna, T.M., Bozuwa, J., Rao, R., 2022. The Power of Community Utilities: Publicly Smith, A., 2007. ‘Translating Sustainabilities between Green Niches and Socio-Technical Owned and Cooperative Electric Utilities as Anchors for Community Wealth Building and a Just Energy Transition. The Democracy Collaborative and the Climate and Community Project, Washington, D.C. and Philadelphia. Hansen, T., 2021. The foundational economy and regional development. Reg. Stud. 1–10. Hanke, F., Guyet, R., Feenstra, M., 2021. Do renewable energy communities deliver energy justice? Exploring insights from 71 European cases. Energy Res. Social Sci. 80, 102244. Hargreaves, T., Hielscher, S., Seyfang, G., Smith, A., 2013. Grassroots innovations in community energy: The role of intermediaries in niche development. Glob. Environ. Change 23 (5), 868–880. regimes’, Technology Analysis & Strategic Management, vol. 19. Routledge, pp. 427–450 (4). Smith, A., et al., 2017. Grassroots Innovation Movements. Routledge, London. Smith, A., Seyfang, G., 2013. Constructing grassroots innovations for sustainability. Global Environ. Change 23 (5), 827–829. Smith, A., Hargreaves, T., Hielscher, S., Martiskainen, M., Seyfang, G., 2016. Making the most of community energies: three perspectives on grassroots innovation. Environ. Plann. 48 (2), 407–432. Smith, A., Stirling, A., 2018. Innovation, sustainability and democracy: an analysis of grassroots contributions. J. Self Govern. Manag. Econ. 6 (1), 64–97. EnergyPolicy172(2023)1132778 M. Lacey-Barnacle et al. Smith, A., Raven, R., 2012. What is protective space? Reconsidering niches in transitions to sustainability. Res. Pol. 41 (6) https://doi.org/10.1016/j.respol.2011.12.012. Stewart, F., 2021. All for sun, sun for all: can community energy help to overcome socioeconomic inequalities in low-carbon technology subsidies? Energy Pol. 157, 112512. Thompson, M., 2021. What’s so new about New Municipalism? Prog. Hum. Geogr. 45 (2), 317–342. Wang, X., Lo, K., 2021. ‘Just Transition: A Conceptual Review’, Energy Research & Social Science, vol. 82. Elsevier Ltd, 102291. https://doi.org/10.1016/j.erss.2021.102291. March. Walker, G., Hunter, S., Devine-Wright, P., Evans, B., Fay, H., 2007. Harnessing community energies: explaining and evaluating community-based localism in renewable energy policy in the UK. Global Environ. Polit. 7 (2), 64–82. Warren, D.B., 2010. Growing a Green Economy for All: from Green Jobs to Green Uyarra, E., Shapira, P., Harding, A., 2016. Low carbon innovation and enterprise growth Ownership. Democracy Collaborative at the University of Maryland. in the UK: challenges of a place-blind policy mix. Technol. Forecast. Soc. Change 103, 264–272. Van Veelen, B., 2018. Negotiating energy democracy in practice: governance processes in community energy projects. Environ. Polit. 27 (4), 644–665. Webster, A., Kuznetsova, O., Ross, C., Berranger, C., Booth, M., Eseonu, T., Golan, Y., 2021. Local regeneration and community wealth building–place making: co- operatives as agents of change. J. Place Manag. Dev. Williams, A., Goodwin, M., Cloke, P., 2014. Neoliberalism, big society, and progressive localism. Environ. Plann. 46 (12), 2798–2815. EnergyPolicy172(2023)1132779
10.3390_bs13010066
Article The Experience of Self-Transcendence in Social Activists Carol Barton 1 and Rona Hart 2,* 1 2 Previously School of Psychology, University of East London, Water Lane, London E15 4LZ, UK School of Psychology, University of Sussex, Falmer, Brighton BN1 9RH, UK * Correspondence: rona.hart@sussex.ac.uk Abstract: Every day the wellbeing of disadvantaged individuals and communities is being trans- formed through the activities of self-transcendent social activists. The positive contagion generated by their actions is felt globally through influence, replication, leadership training and education. These people are visionary, brave, and describe their lives as joyful, deeply fulfilled, and impactful. Seeking no personal recognition or accolade, born from a deep feeling of connectedness and a vision of how life could be better, participants describe the factors that influenced their decision to dedicate their lives to serving the greater good. Using Constructivist Grounded Theory, in-depth semi struc- tured interviews were carried out with eight participants who self-identified as self-transcendent social activists, who have initiated non-mandated and not-for-profit community action. Data was analyzed to explore each participant’s personal experiences of self-transcendence and how being self- transcendent has manifested their life choices. The findings present a definition of ‘self-transcendent social activism’ and a theoretical model that explains the development of participants’ activism: trigger, activate, maintain and sustain, resulting in an impact experienced at three levels - individual, community and global. Theoretical and practical implications are discussed. Keywords: self-transcendence; social activism; prosocial behavior 1. Introduction The course of history has been changed by many highly impactful self-transcendent social activists who committed their lives to bring about social transformation in the communities and countries in which they lived and served. Nobel Peace Prize winner (1964), Luther-King Jr., will long be remembered for his non-violent campaign against racism that resulted in his assassination and racial discrimination being declared illegal in southern US states. Nobel Peace Prize winner (1984) and former chairperson of the Truth and Reconciliation Commission, Tutu, was influential in his campaign against apartheid and for the peace negotiations in South Africa. Whilst the legacy of Gandhi, five times peace prize nominee, whose non-violent leadership led to his assassination and to independence for India, is celebrated annually through the award of the international Gandhi Peace Prize. A review of biographical literature reveals that these courageous, visionary, people of faith prioritized freedom, equality, and the eradication of poverty above self-interest [1–3]. From a position of feeling connected to others and a focus that extends beyond their own personal wellbeing, self-transcendent social activists are people who act to address global problems such as inequality, poverty, environmental issues and exploitation [4]. Social activism is defined as “instances in which individuals or groups of individuals who lack full access to institutionalized channels of influence engage in collective action to remedy a perceived social problem, or to promote or counter changes to the existing social order” [5] (p. 4). Social activists are therefore individuals or groups who engage in collective action to bring attention to and resolve social problems. They operate through groups or social movement organizations that are characterized by varying degrees of formal and informal structures [5]. Citation: Barton, C.; Hart, R. The Experience of Self-Transcendence in Social Activists. Behav. Sci. 2023, 13, 66. https://doi.org/10.3390/ bs13010066 Academic Editor: Andrew Soundy Received: 5 December 2022 Revised: 9 January 2023 Accepted: 9 January 2023 Published: 11 January 2023 Copyright: © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). Behav. Sci. 2023, 13, 66. https://doi.org/10.3390/bs13010066 https://www.mdpi.com/journal/behavsci behavioral sciences Behav. Sci. 2023, 13, 66 2 of 22 Self-transcendence is defined as an “increased awareness of dimensions greater than the self and expansions of personal boundaries within intrapersonal, interpersonal, transpersonal, and temporal domains” [6] (p. 179). It involves an endeavor to connect to a larger context with a prosocial intent to serve the greater good. As such, self-transcendence is a set of values and a state of mind that can prompt the motivation to engage with so- cial activism. However, to our knowledge there is no early research that examines the connection between the two concepts qualitatively. This study endeavors to contribute to the extant literature on the key motivations that drive social activism through the exploration of self-transcendence. Given the potential impact that activists have through the work they do in generating positive transformations in people, groups and entire societies, their goals, life choices and experiences of self- transcendence within the context of social activism is a worthy scientific undertaking. 1.1. Social Activism Social activism involves taking positive intentional action and mobilizing resources to bring about change in society. Activism, both peaceful and aggressive, is expressed in many forms from writing letters, lobbying, boycotts, protests, strikes, petitions, community led initiatives, and social media campaigns. Examples of topics that social activists may engage with include environmental issues, racial equality, gender equality, refugee and immigration policies, human rights, LGBTQ+ rights, religious freedom, poverty, housing, anti-war campaigns, welfare policies, and many other topics. Within the current Western neo-liberal social norms that emphasize individualistic pro- self goals, and independence rather interdependence, engaging in prosocial activism with a purpose of benefitting the greater good, might seem exceptional, especially since social activism is a costly endeavor, and that the chances of successful outcomes are uncertain [7]. The question of what motivates social activism, and whether it is motivated by pro-self or prosocial intents is particularly intriguing given the contrasting social norm setting. Given the numerous social causes that social activists are engaged with, motivations will likely vary in accordance with the goal being pursued, the context, and the ideologies that underlie the activity. Recent research on the motivations of social activists has therefore aimed to elicit overarching themes to explore the underlying motives of social activists. A repeated theme in the literature is that people might engage in activism because of injustices or deprivation that they suffered or witnessed or because they identify with the hardship of a particular group whose struggles coincide with their own experiences [8]. This suggests that critical life-events, needs, goals and interests may be key drivers of social activism [9]. An additional point raised in the literature is that negative emotions (such as pain, fear, anger or frustration) triggered by one’s sense of deprivation or injustice, or from a distressing life event, predict the willingness to engage in collective action, as well as the actual participation [10–12]. The perception that one’s group is negatively evaluated, disrespected, marginalized or discriminated against, can also induce willingness to engage in social action, both peaceful and violent [13–17]. Identity features strongly in the social activism literature as a motivating factor. Na- tional, professional, ethnic, racial, class or sexual orientation identities were found to be key drivers that motivate people to engage in social activism, and in turn, belonging to a social movement both intensifies the primacy of these identities, as well as generates a new identity that forms as a result of affiliating and identifying with the activist group [18–23]. Another motivation to engage in social action is because it renders activists personal, social or psychological benefits [24]. Gains accrued from activism include a sense of mean- ing and purpose, positive self-regard, belonging to a group or a community, and improved wellbeing [24–32]. Social activism was also associated with a feeling of personal signifi- cance [33], indicating that when people feel that they lack significance, they were more willing to pursue a political cause, at times involving violent actions, and making personal sacrifices [25,34]. Klar and Kasser [24] showed that activism was positively related to self-determination and meeting three basic needs: autonomy, competence, and related- Behav. Sci. 2023, 13, 66 3 of 22 ness. Another benefit from social activism comes in the form of positive emotions, such as exhilaration and awe, empowerment, pride, joy and sense of solidarity [11,24,35–37]. Values and ideologies often translate into visions and are also important motivating factors that can drive people to take social action often by eliciting a sense of social respon- sibility and urgency [38]. Prosocial values also play a central role in motivating action on behalf of a social cause [39–42]. Similarly, ideologies, moral convictions and religious beliefs are positively associated with activism [43–46], and acting on what one sees as core values engenders a sense of meaning in life, significance, and fulfillment which in turn elevate self-esteem [47,48]. In the context of political activism, moral convictions were found to be associated with pride [49]. Interestingly, people may adopt particular values and pursue social action due to guilt about one’s own privileges or for causing harm, or for not doing enough [50,51]. Another type of motivation that can drive social action is generativity: the desire to care about the welfare of future generations [22,52–54]. These are linked with other prosocial states such as empathy, perspective taking, compassion, accountability, and sympathy which have been shown as motivational factors that can prompt people into social action [55–59]. The brief review offered above of the factors that can motivate social activism suggests that it can be motivated by pro-self or prosocial goals, intents and attitudes, and these contrasting underlying mindsets, can impact both personal and community results. Pro-self motives can manifest in the desire to construct and maintain positive self-images of oneself as worthy and valuable, and might lead to displaying concern for others insomuch as this serves the need of the ego [60]. This can lead to the ‘white savior’ stereotype: imposing patronizing models or solutions on those in jeopardy, leading at times to the perpetuation of their condition [61,62]. In contrast, people who are motivated by prosocial motivations report empathetic identification with disadvantaged groups, experiencing acute awareness of issues that need to be changed and a belief that they can make a difference [63]. While they may sacrifice their personal time and resources, paradoxically their work may result in enhancing their own personal wellbeing [24], in addition to benefitting the greater good by attracting attention to social problems, creating solutions, and developing partnerships [64]. The link between motivation, intents and outcomes in social activism raises the ques- tion of self-transcendence as a driver of social activism. Next, we unpack the concept and briefly review the literature. 1.2. Self-Transcendence Frankl [65] (p. 115) maintained that “being human always points to something or someone greater than self . . . the more one forgets oneself—by giving himself to a cause to serve . . . the more human he is . . . ” Accordingly, within transpersonal psychology, self- transcendence involves serving a purpose greater than the self with a selfless intent [66,67]. Reed [68] (p. 397) defines self-transcendence as “the capacity to experience connectedness and expand self-boundaries in four dimensions: intra-personally by gaining more self- awareness, inter-personally by relating to others and nature, temporally by integrating past and future in a meaningful present, and trans-personally by connecting with spiritual dimensions of indiscernible world”. Other authors argued that the term signifies a devel- opmental process whereby one’s consciousness expands beyond personal, bounded, and self-directed ego, to include other people and concerns within that sense of expanded iden- tity [69]. Maslow’s [70] hierarchy of needs suggests that one of people’s top growth needs is the desire to reach self-actualization whereby one can realize his or her full potential. How- ever, it has been argued [71,72] that later in his life Maslow discovered that self-actualizing individuals were capable of even higher psychological development by transcending their own self-centered goals, and pursuing higher causes that are other-orientated. The most coherent and possibly the most cited description of self-transcendence emerged from Schwartz’s [73,74] values theory. Schwartz [73] conceptualizes values as beliefs about what is desirable, worthy and important. As such these beliefs shape one’s Behav. Sci. 2023, 13, 66 4 of 22 perceptions of oneself, others, and situations, guide one’s life goals, priorities, and decision- making, influencing related behaviors. Schwartz’s [74] values model offers a classification of two broad bi-polar dimensions, each of which incorporates several values: One axis ranges between ‘openness to change’ to ‘conservation’ values, and includes self-direction and stimulation on one side of the axis, and security, conformity and tradition on the other side. The second axis has ‘self-enhancement’ on one side and ‘self-transcendence’ on the other. It includes power and achievement on one side, and universalism, benevolence on the other side. Hedonism is placed across two dimensions: openness to change and self- enhancement. Self-transcendence values - benevolence and universalism, are characterized by a reduction in self-centeredness, and the capacity to transcend one’s own selfish needs, to care for the interest and welfare of others. As such they are considered other-focused, growth promoting values [73]. An alternative conceptualization of self-transcendence suggests that it is a personality trait linked to spirituality [75] whereby a person “identifies the self as part of the entire cosmos” [76] (p. 975), feeling a sense of connection to the universe, interdependence and responsibility. It is also seen as a core virtue within the VIA character strengths and virtues clas- sification, encompassing of character strengths of appreciation of beauty and excellence, gratitude, hope, spirituality and humor [77]. This trait has been associated with experi- encing elevation emotions such as awe, ecstasy, amazement, worship, and flow as well as meaning in life [77]. In terms of its development or emergence, self-transcendence seems to be expressed more strongly in people who confronted difficult life experiences such as loss or illness, hence seen as a sign of adversarial growth in terms of the capacity to transcend one’s own needs and experiences, taxing as they may be, to express universalism and prosociality [78–80]. Further exploration of the concept suggests that it can be active or passive in terms of its behavioral manifestation. Although self-transcendence is positively associated with taking action [81], it is indeed possible for self-transcendence to remain passive. Another finding is that self-transcendence values promote prosocial attitudes and states (such as empathy, trust, love, affection and compassion) and motivate a variety of prosocial behaviors (such as offering encouragement, care, or support) [82–85]. Some gender differences were detected, as women were found to attribute more importance to self-transcendence values while men attribute more importance to self-enhancement values [86]. There are indeed some self-benefits that people can gain from holding self-transcendence values. A positive association was found between self-transcendence and wellbeing, positive emotions, happiness, quality of life (in severely ill patients), healthy behaviors, meaning and purpose in life, self-esteem, hope, sense of coherence, mindfulness, flow, adaptive coping strategies and resilience [87–97]. 1.3. Self-Transcendent Social Activism While theoretically self-transcendence can become a strong motivator for social ac- tivism, there is little research that explores this point. In two cross-national studies [82], the authors concluded that people who hold self-transcendence values are more likely to be involved in political activism. Similarly, Gundelach and Toubøl [98] found that the values of self-transcendence were associated with activism in the context of refugee solidarity. Another study on environmental activism examined the relationship between activism and moral identity and concluded that self-transcendence positively predicts environmental activism, while self-interest values were associated with apathy leading to low environmen- tal activism [99]. In another correlational study Hackett [100] found that the association between self-transcendence values and activist behaviors was stronger when these values emerged from personal concerns. Behav. Sci. 2023, 13, 66 5 of 22 To our knowledge no further research has explored the association between self- transcendence and social activism, and there is no qualitative research which explores how self-transcendence is manifested in social activism. 1.4. The Current Study The aim of this study was to qualitatively explore the experiences of a mixed age and mixed faith group of activists, who self-identify as self-transcendent, in order to answer the question: ‘In what way does the experience of self-transcendence manifest in the work and lives of social activists?’ In exploring this topic qualitatively, the paper aims to address a gap in the literature and contribute to our understanding of the drivers of social activism. 2. Materials and Methods This study applied a Constructivist Grounded Theory (GT) approach to collect and analyze qualitative interview data [101], as a means to explore self-transcendence within the context of social activism. Grounded Theory is particularly useful for exploratory studies and its key strength is in facilitating the development of theoretical models emerging from the data. It has several distinctive features [102,103]: • Data collection and analysis cycles: Grounded Theory involves an iterative data collection and analysis process whereby early data collection and initial analyses inform subsequent decisions on the direction and focus of data to be collected and on sampling, while the analysis remains open to new emergent topics. Sampling aimed at theory generation: Sampling in Grounded Theory, is initially purposive (identifying and selecting participants who are knowledgeable about or experienced with the phenomenon of interest), and later it becomes theoretically driven (known as theoretical sampling), since sampling decisions draw on early analysis and reflect the ongoing theoretical development that occurs as a result of the data collection and analysis cycles. Developing a theory from data: Grounded theory is designed in a way that enables researchers to develop a theory/model from data [102,103]. As such, it requires the application of inductive reasoning (bottom-up) which enables researchers to extrapolate a theory from a set of individual cases. This involves moving from the particular case to the general, and from a detailed description to an abstract level [101]. Data analysis: Analyzing data in Grounded Theory involves applying several tech- niques. Initial or open coding involves analyzing the text by coding word-by-word and line-by-line and naming each segment of the data. This is often followed by focused coding which is aimed at generating conceptual codes [102]. Focused coding involves the use of some of the following techniques: • • • (cid:35) (cid:35) (cid:35) (cid:35) (cid:35) Axial coding: Relating categories to subcategories and making explicit connec- tions between them. Comparative coding: Constant comparisons between data in order to find similarities and differences and establish analytic distinctions. In-vivo coding: Preserving participants’ meaning in the coding. Selective coding: Distinguishing the core categories and connecting them to other categories. Core categories: Identifying the components of the model/theory (they are the ones that most frequently emerge from the data, they have identifiable properties and are linked to other categories). Theoretical coding: Specifying the relationships between categories and inte- grating categories to create a coherent depiction of a model or theory. (cid:35) Whilst the nature of the interviewer-imposed questions meant that it was impossible to totally eliminate researcher bias, interaction between researcher and participants took the form of clean language open questions and passive listening, enabling participants to speak openly and spontaneously of their life experiences and for theories to emerge Behav. Sci. 2023, 13, 66 6 of 22 from participants’ narratives [103]. The research outcomes, therefore, are a co-construction of a theoretical model based on the data and the interpretation, observations of the first author, who, for many years, has supported Africa-based social activists through coaching and consultancy. 2.1. Participants Criterion sampling (a sub-set of purposive sampling) was used in this study to define and invite the target participants. It involved searching for participants who meet a certain criteria. In this study, the key criteria was involvement in social-activism and experiencing self-transcendence. For the purpose of recruitment and self-selection of participants, the following definitions were used (see Table 1): Table 1. Definitions used for purpose of recruitment. Self-Transcendence • • • • A shift in focus from self (ego) to others; A shift in values and willingness to sacrifice self-interest to serve the greater good; An increase in moral concern and courage to act and take risks, aligned to moral compass Social Activism Non-mandated and not-for-profit practical action carried out by individuals or groups, to solve societal problems and bring about change for the good of others. The participants self-identified with the above statements and satisfied the following inclusion criteria: • • Feeling connected with something greater than oneself They had initiated a non-mandated not-for-profit community program to reduce poverty, injustice, homelessness; the program had been operational for at least two years and positive community impact can be evidenced. Potential social activists were identified through the first author’s personal networks, which included former colleagues, coaching and business clients. The researcher also invited former colleagues to recommend suitable participants. Prospective participants were initially contacted via an email that informed them that the researcher was seeking social activists who have experienced self-transcendence; the study aimed to explore their experience of self-transcendence and how this had manifested in their life choices. Eight social activists, six females and two males, of mixed nationalities and religions, aged between 35 and 60 committed to participate in the study. Table 2 details their back- ground and domain of social activism (pseudonym are used to protect their identities). No incentives were offered to encourage participation. Table 2. Participant demographics. Pseudonym Gender Nationality Country of Residence Religion Context/Projects Fiona Jemma F F Swazi Kenya Christian Kenyan Kenya Christian Pastor/Spiritual healer, laying the foundation for an international network of home educators Bringing hope to poor communities affected by HIV/AIDs by providing education, medical and social care Behav. Sci. 2023, 13, 66 7 of 22 Table 2. Cont. Pseudonym Gender Nationality Country of Residence Religion Context/Projects Tina Sam Todd Natalie Tandy Judith F M M F F F American Kenya Christian Kenyan Kenya Muslim Filipino Hawaii Christian British (Tobago origin) Chinese American UK Hindu USA/Kenya Christian American Kenya Christian Providing education, medical and social support services for children with disabilities and employment training and opportunities for their mothers. Eisenhower Fellow, developing local leaders, catalysing positive change, and alleviating poverty in the largest Kenyan slum Youth Pastor, building affordable housing units to support the homeless in Hawaii, Cambodia and Africa Teaching Meditation, peace circles and wellness programmes in US, India, UK and Virgin Islands Empowering teachers and transforming schools in Kenya through leadership training, instructional coaching and infrastructure support. Rescuing and equipping orphans and destitute children in Kenya and Romania 2.2. Data Collection Following receipt of ethical approval from the first author’s University, potential participants were contacted via email. Prior to the interviews, participants were provided with more detailed information about the purpose of the research including information about confidentiality and their right to withdraw. Written consent was obtained. A draft set of questions was provided prior to the interviews. Semi structured Grounded Theory interviews that lasted between 45 and 80 min were conducted online by the first author and recorded using Zoom. After reminding participants about the purpose of the research, interviews commenced by asking “what does self-transcendence mean to you?” The researcher used clean language [104], open questions to develop a conversation about their personal experience of becoming self- transcendent and the role that self-transcendence plays in decision making. Listening attentively for themes and insights, the researcher asked more probing follow up questions to stimulate deeper reflection about specific characteristics of self-transcendence and what factors strengthen or weaken their experience of self-transcendence. Example questions include: What does the term self-transcendence mean to you? Thinking about your own experience of becoming self-transcendent, how would you describe that? How has being self-transcendent influenced your life choices? How does being self-transcendent manifest itself in your social activism? In other areas of life? What are the benefits and challenges of being self-transcendent? Whilst one participant described in some detail her personal experience of becoming self-transcendent, other interviewees steered the interview in the direction of how being self- transcendent has motivated and influenced their life choices, and how this manifests in their pursuit of social activism. The resulting theory, therefore, represents a ‘self-transcendent’ infused model of social activism. The study followed Grounded Theory guidelines by conducting cycles of data collec- tion followed by initial analysis which entailed line by line coding [102]. This meant that between interviews, data were coded, and key themes identified for deeper exploration were introduced through focused questions in subsequent interviews. For example, in Behav. Sci. 2023, 13, 66 8 of 22 early interviews ‘courage’ and ‘empathy’, emerged as important themes leading to more exploratory questions in later interviews. Whilst time constraints meant that the number of participants and interviews was limited, the sample size was considered large enough for a robust theory to emerge [105], and data saturation was achieved within this sample size and interview framework. 2.3. Data Analysis Interviews were transcribed using a transcription service and manually checked to ensure verbatim accuracy. This enabled the researcher to gain an in depth understanding of the data. As noted, data collection and initial analysis (open coding) occurred simul- taneously. Once open coding was complete for all transcripts, several types of focused coding techniques were applied to create a more abstract analytical framework [102]. The first stage included sorting the numerous themes that emerged from the initial coding, to identify and focus on the most salient ones [102]. Then axial coding was applied as a means of linking between categories and their subcategories, some of which readily emerged from the text. Comparative coding followed and involved comparing categories across different segments of the data in order to find similarities and differences and to establish clearer distinctions between elements that initially seemed to be entangled together [103]. The next stage involved selective coding. At this stage it became clear that the focus of the model would be around the process of becoming self-transcendent social activists. This stage held the key to reducing the number of categories and focusing the analysis on the most significant ones which were eventually identified as the core categories [102,103]. The last stage involved theoretical coding - refining the categories, specifying the relationships between them, and integrating them into a coherent model [103]. In order to produce a visual representation of the emergent model, the data were then imported to NVIVO for further analysis. Earlier work by Bazeley [106] and oth- ers [107,108] demonstrated the usefulness of NVIVO in facilitating a grounded theory analysis. Hutchison, Johnston and Breckon [107] argued that the benefit of NVIVO is in providing a transparent account of the analysis process which enhances its rigor. Although NVIVO can be used to conduct all stages of the Grounded Theory analysis, in this study it was only used to help generate a clearer account of the model. The conceptualization of a theoretical model of ‘self-transcendent infused social ac- tivism’, enabled the researcher to refine, condense, and align the data to the final six themes which are described below. 3. Results What started off as an investigation into the experience of self-transcendence in the lives of social activists became a broader discourse about what motivated participants to commit their lives to activism, the impact this has had on their personal lives and the com- munities they serve and more globally. The analysis of data resulted in the emergence of: A definition of self-transcendence within this context 1. 2. A description of how self-transcendence activism has impacted the lives of partici- pants and the people they serve 3. A model comprising four continuous stages of activism - trigger, activate, maintain and sustain. These are summarized in Table 3. Behav. Sci. 2023, 13, 66 9 of 22 Table 3. Summary of Results. Feeling connected to something greater than oneself Self-awareness Definition Increased awareness of social justice issues Impact Triggers Activation Reduction in self-interest Desire to be of service Personal impact Community impact Global impact Early role models and exposure to social injustice Personal experience of tragedy Feeling ‘called’ or compelled Empathy, Compassion and Connection Courage and faith Having a vision Maintain Personal sacrifice and self-care A community of like-minded individuals for support Seeing possibilities and co-production Having a global perspective Sustain Growing leaders Teaching empathy, awareness and courage 3.1. Definition The definition domain describes how the participants responded to the question ‘what does self-transcendence mean to you?’. 3.1.1. Feeling Connected to Something Greater Than Oneself Without exception, Christian, Hindu and Muslim participants expressed the importance their faith, combined with a commitment to live a life aligned to their spiritual convictions: ‘It’s definitely, my Christian, commitment and wanting to walk and do things for others’ (Jemma). ‘I’m very strong in my faith, but . . . I don’t want to force that on other people. But I also make sure that I live my life in the values of my faith and that helps me in terms of how I walk and interact with the community’ (Sam). ‘When I walk in my calling, directed by God (Fiona)’. Connection to something greater also included the concept of seeing oneself as part of a bigger community, connected to all humanity: ‘A small cog in a large wheel’ (Sam), ‘As individuals we are not complete in our separateness’ (Natalie). ‘There’s another expression that says ‘you are because we are’ so you always understand that your life is connected to others . . . .’. (Fiona) 3.1.2. Self-Awareness Most respondents noted that self-awareness and self-care are precursors to self- transcendence and the process of becoming self-transcendent involves self-reflection, self- knowledge and healing. To help other people in a healthy, safe and benevolent way, first it is necessary to become a ‘safe person’: Behav. Sci. 2023, 13, 66 10 of 22 ‘ . . . in my process of transcendence, part of my journey was understanding who I am. I think you cannot transcend yourself if you haven’t taken care of yourself. So, there’s an element of understanding yourself, growing and knowing who you are’ (Fiona), ‘And then there’s your own growth as a human and your own sort of evolving identity that sort of interacts with that... it is a process because you have to continually answer the question of what is actually happening around me, how do I interpret it? How do I make meaning out of the things I’m seeing?’ (Tandy) 3.1.3. Increased Awareness of Social Justice Issues The majority of the participants reported a heightened awareness of inequality, poverty and other prejudices combined with a belief that the situation can be improved. Whereas other people might not be aware of injustices, participants reported both noticing and wanting to respond to inequitable access to resources and opportunities: ‘It makes you aware of other people’s lives, other people’s struggles. God put compassion and empathy in you, and you can’t limit that compassion and empathy to just a small group of people’ (Judith). ‘It’s how we view the world, how we value things. I cannot sit back and see somebody else being in total despair’ (Fiona). 3.1.4. Reduction in Self-Interest Whilst we all need validation, if affirmation, personal gain or enhanced self-esteem is the motivation; that is not self-transcendence, and this was noted by several participants. The participants also noted that in self-transcendence, the focus and concern are no longer on self but on the people being served. When self-gratification desires reduce there is a much greater sense of freedom: ‘You’re doing things not just for your ego, not to be noticed. You don’t need pub- lic acclamation. and you’re not doing it for personal gain. Doing it out of love and compasion—There is something deeper within you’ (Jemma). ‘So basically, it’s about putting others first rather than putting yourself first.’ (Tim) 3.1.5. Desire to Be of Service The act of serving others was mentioned by several participants who considered it much more satisfying and rewarding than doing things just for oneself. To serve others brings great personal blessings, to see the smile on the face of someone you’ve helped, or just to experience the privilege of serving others, counts for so much more than self-gratification: ‘There can be so much emptiness in just trying to self-gratify. There’s only so much you can do to self-gratify, but so much joy when you serve others and you see others happy’. (Jemma) 3.2. Impact 3.2.1. Personal Impact The work of an activist can be demanding and grueling, but participants overwhelm- ingly described their lives as joyful, fulfilled, aligned to calling, abundant and meaningful. Giving joy to others is described as contagious, great fun, extremely rewarding and this creates a desire to do more: ‘..it can just be exhausting. Honestly just to be so empty, you know . . . .as the social activist, learning to give your life away, and when you really look at what it definitely includes, bringing fulfilment, and when you are completely exhausted, exhausted for the social good.... it gives me a lot of joy. It’s grueling but it gives me joy’ (Jemma) ‘Yes, it does require personal sacrifice. But for me, I don’t see it as personal sacrifice because I enjoy doing what I do and I see it as an opportunity, I derive a lot of joy. So, for Behav. Sci. 2023, 13, 66 11 of 22 me I count it as a privilege . . . it makes you want to do it more because you get joy in other people’s joy. I think joy is contagious, and so, giving joy is just so much fun’ (Tina). ‘Fulfilment, deep fulfilment, challenging, rewarding.’ (Judith) 3.2.2. Community Impact Eight community programs are represented across four continents. Participants re- ported working with victims of HIV, disabled children and their families, the homeless and people living in slums to provide education, medical services, social care, adult skills and employment training, mediation, infrastructure support, leadership development, affordable housing and other initiatives to alleviate poverty and empower communities: ‘It began growing organically because when you support a woman she comes with the entire family. A woman comes with children, youth, adolescents, and she brings the community. And as a result, she also came with sickness and this affected the education of the children and became an issue. Socioeconomic empowerment is an issue we’ve been tackling initially as well. We wanted to see how we can support her to earn. You’re putting that wholesome completeness in that home. So, we began by offering economic empowerment, then education for the children, then the technical certificate for their older children. We were training women to do different skills and assessing their credit to start little businesses. So that’s how the whole project started . . . . . . .’. (Jemma) ‘We work with children with disabilities and their moms. There is no help in this country for families that are struggling with that. Every child that comes to our therapy center, comes with a mama and we provide each mama with employment . . . . . . Our heart is for people that are struggling with disabilities and their families. We work with a lot of HIV positive families and they’re just dealing with a lot of problems besides the disability. There’s so many other problems that come along when you live in poverty. But it’s always a thrill to be able to help somebody’. (Tina) 3.2.3. Global Impact Most participants talked about the ripple effect which has been created through developing international leaders, training others within existing programs, permitting replication (at no cost) of their community development model, extending their work internationally. One participant spoke of being invited to speak to UN representatives about his work to support the homeless: “God has been good in my life, putting me into this position where I can be influential to a lot of people as an affordable housing developer. I’m a newbie in this industry, but I’ve been recognized as the best affordable housing developer in town. And even United Nations got a hold of my story and my philosophy as a developer... So instead of just working on developing buildings for people for the money, I follow the need of people. So my focus is to work with people, find out the need. And that’s one of the reasons why I flew to Nairobi and I saw even greater need compared to Hawaii, because they’re in need of a half a million apartments for the 3 million people who live in slum . . . And besides being a developer, I created a non-profit organization. And we’re reaching out to Cambodia, to the Philippines. And now I’m thinking about reaching out to Tanzania’. (Todd) 3.3. Triggers This category refers to the life experiences that set participants on a course of taking action: 3.3.1. Early Role Models and Exposure to Social Injustice All participants described how the influence of early role models, and the environment in which they were raised, shaped their outlook and made them more sensitive to injustices and inequalities: ‘I grew up seeing my parents caring for other people, serving more than to be served and that’s how I grew to know life’. (Jemma) Behav. Sci. 2023, 13, 66 12 of 22 One respondent noted how her experience of a difficult childhood led to a sense of separation, fear and isolation which prompted a spiritual search for reconnection with some- thing greater, triggering a desire to help others (Natalie). Another recalled his experience of being raised in an institution: ‘It was shaped with my upbringing growing up in a children’s home which had more than 110 children. It’s not easy growing up in institutions - life was not easy. So that shaped my thinking about how I wanted to live my life’. (Sam). Exposure to social justice issues, such as homelessness, apartheid, refugees, triggered an early response and determination to take action: ‘We had refugees in our home, and you are meant to take care of them. I saw my dad bring one - he was an Ethiopian refugee when there was war in . . . and then when I was in the university myself, I brought in a refugee, I’ve always had that desire to reach out to people who are either homeless or suffering and to serve them’. (Jemma) 3.3.2. Personal Experience of Tragedy Experiencing personal tragedy, or seeing tragedy close up often triggered negative emotional and behavioral responses; however, for our participants experiencing trauma it triggered a motivation to help others: ‘Our son was born at 22 weeks. He survived many heart attacks and we saw him come back to life many times after having no heartbeat. And, he was a true miracle. and that was my baby . . . . . . and then God asked me to do a special needs ministry’. (Tina). ‘It was the first time I saw a mother and a child laying on the side of the street and I was in complete shock. Like I couldn’t believe that traffic wasn’t stopping, and people weren’t helping her. Like it was such a foreign concept to me. and that definitely was a trigger too.’. (Judith) 3.3.3. Feeling ‘Called’ or Compelled Six participants reported a sense of calling, feeling compelled, or hearing from God, to which, in spite of the personal sacrifices demanded and not knowing where resources might come from, triggered a conviction to respond. One participant reported seeking God’s will through prayer and reading the Bible: ‘God has called me to serve the very disadvantage, very poor, in the slum . . . . So out of obedience to God he called me to go into that community and walk alongside women like that’. (Jemma) ‘It’s a calling from God truly that he’s asked us to do this. I know that sounds, for some people kind of weird, but it is definitely what we feel called to do. Now, did I hear a voice when I say the word calling? No, but I spend a lot of time, praying and reading the Bible and asking God to keep directing us.’. (Tina). ‘There is the compelling and choosing not to ignore that compelling. God spoke to me and I know for certain that I heard the call and we responded’. (Judith) 3.4. Activation These themes moved participants from ‘making a decision’ to take action by the operationalization of that decision. 3.4.1. Empathy, Compassion and Connection Common themes running through all interviews were how compassion and empathy led to taking action. Empathy enables one to identify and connect with the community, as opposed to sympathy which can be seen as adopting a superior position and imposing solutions. Feeling compassionate often draws one into becoming deeply empathetic. ‘When you talk about transcendence, transcendence is not about sympathy. It must have empathy. If empathy is not in you then you’re totally missing the point. So, empathy Behav. Sci. 2023, 13, 66 13 of 22 enables you to identify and connect with the community. Whereas sympathy puts you on a higher position and you’ve got power. Sympathy is all about listening with your head. But empathy is about listening with your heart’. (Sam) ‘So, my job I believe is to inspire all these people that there is a choice that we can make to have compassion and empathy for other people who are less fortunate than them/me’. (Todd) 3.4.2. Courage and Faith Without courage, self-transcendence can remain passive. All participants spoke of the need to exercise courage, an internal quality that you carry on the inside–courage to admit one does not know all the answers, to be unpopular, to travel across the world and live in dangerous places, and courage to take risks. The notion of faith includes believing that resources will be provided, and things will work out whilst the path remains unclear: ‘For sure you can’t do what we do without courage. You need both self-courage and you just need overall courage. . . . .I want to learn the courage to say I’m not here to help. I’m here to walk with you and everything. and even the courage to have a brave face to go into hard places.’ (Sam). So you have to sacrifice something in order to be courageous and to step up and do, especially when you’re trying to help other people. You gotta have courage’. (Todd) 3.4.3. Having a Vision Participants reported observing patterns and seeing life through a lens of possibilities. Rather than looking at problems and what does not work, starting from the position of appreciating what works, seeing potential in others—what they are capable of becoming and having a visualization of what might be: ‘And I always say, because it is God’s work, he provides the resource, it’s his vision’. (Jemma). ‘I was primed to see things in a way that would make me want to do something about it. . . . . . . .. for several months prior to the vision trip that I took’ (Tandy). ‘So you have a vision. It’s challenging, but it’s also extremely rewarding, because I’ve been doing it for some time, like for instance our rescue centre in Romania, those kids are now grown’. (Judith) 3.5. Maintain The life of an activist can be demanding and exhausting. The resolve to remain committed is strengthened through several factors: 3.5.1. A Community of Like-Minded Individuals for Support Surrounding oneself with a supportive circle of encouraging, like-minded people who act as co-mentors increases motivation and provides opportunities to work collectively: ‘If you have healthy intimate relationships and strong connections with other people, there’s an exchange - you’re learning with other people, you’re serving with other people. I think that increases self-transcendence because you get the opportunity to watch other people being courageous’ (Fiona). For family members, the support of a partner and family to cheer you on is vital: ‘I don’t think that God’s going to call me one way and my husband another way because we are in this together as a married couple. and so, we make decisions together’. (Tina) 3.5.2. Personal Sacrifice and Self-Care The importance of exercising self-care, taking breaks and time out, spending time with family, spiritual connection and devotion were reported as being important to maintain good emotional, spiritual and physical health: Behav. Sci. 2023, 13, 66 14 of 22 ‘I made sacrifices thinking that I could withstand it, thinking that my marriage could withstand it. I’ve made a lot of sacrifices. I think first of all, money, it took me five years, before I launched xxx . . . ..And so that’s a very concrete data point around the financial cost’ (Tandy) ‘I want to have more time with my daughter. I kinda need to start being selfish myself. That’s called self-care and boundaries.’ (Sam). ‘ self-care is obviously very important. Having healthy boundaries is really important . . . . So, I have to go to the source, which is God, he has an abundance. So, if I’m not going to the source, it’s like not plugging my computer battery in. It’s not going to last very long’. (Judith) ‘And of course, in this kind of work, you really have to know how to take care of yourself. I’m here trying to recover. Cause the last week I was working so much, but I am happy.’ (Jemma) 3.5.3. Seeing Possibilities and Co-Production Co-production is when a community comes together to influence and design policies and services that benefit all, rather than becoming consumers of solutions supplied by non- community members. This approach creates a sense of interdependence and connectedness whereby people develop confidence to care for each other and co-create solutions. Co- production is seen as an essential factor in maintaining programs and accomplishing community empowerment. For many participants, this has involved exchanging western comforts to live in a Nairobi slum, to truly understand what this feels like on a day-to- day basis: ‘Once you start putting that community in a box and you’re not within that box you’re outside, then you’re not in the community, then that’s a problem. You’re not actually working with the community - you are working against the community. Or, you’re actually looking in terms of “how do I bring a fix” with me?’ (Sam) ‘I think that just being at the same level with everybody here is an important piece. Living with them, working side by side, shoulder to shoulder, trying to understand what they’re going through, even though ultimately I can never fully understand’ (Tina) 3.6. Sustain Participants expressed a desire to see the life of a self-transcendent activist become more commonplace, describing the possibility in terms of ‘heaven on earth’ or utopia, a world filled with more justice, equitable opportunities and resources, joy, compassion, gratitude and kindness. Poverty, oppression and greed would be reduced. Important factors that lead to sustaining impact and growth are identified below: 3.6.1. Having a Global Perspective Technology and media support a sense of connection with people all over the globe. Problems experienced by individuals, communities and countries are no longer viewed in isolation and participants reported how recognizing the interconnection of all things leads to the development of global solutions and co-operation that grows organically, often from something small to something that has global impact: ‘We seem to have embodied this ethos on a global scale because we have kids from all over the world’ (Fiona) ‘What I do - I offer up and create and hold space for entrepreneurs to also discover their own purpose and their own capacities and their own power’ (Sam) ‘Because the more people that are connected and understand this and are able to move outside of themselves, the better society is because then everybody, everybody becomes a resource but in a positive way, not in an exploitative way, but in a synergistic way, like in a way that that brings beauty to society’ (Fiona) Behav. Sci. 2023, 13, 66 15 of 22 3.6.2. Growing Leaders Leaving a legacy means training the next generation of leaders. Where this is ne- glected, the potential impact of initiatives is not sustainable. An example offered by one participant was of a situation where an influential community leader who had initiated many community programs, unexpectedly died before training successors. His death resulted in a fight for leadership and political chaos: ‘He was able to develop so many other things, but he failed in one thing. He failed in grooming leaders to take over from where he was. So indirectly you can say he was self-centered in his leadership because if he had intentionally groomed other leaders, we would not be having the chaos we are having with the political parties’ (Sam). ‘What I’ve done mostly I’ve chosen to mentor others then meet other people who are committed and have the same heart and the same calling. Increasingly, I’m investing my time doing that mentoring, coaching so that more people can develop that attitude.’. (Jemma) 3.6.3. Teaching Empathy, Awareness and Courage Without courage, self-transcendence can remain passive. According to Sam, ‘with- out empathy, you’re missing the point’. Self-reflection, self-knowledge and healing are necessary precursors to helping others. Awareness of social justice issues is a trigger for many activists. Embedding the concepts of empathy, self-awareness, awareness of social injustice and courage into the educational, mentoring and coaching methods deployed by participants and their organizations was reported to be a high priority: ‘There are others that are coming behind me that I need to teach and I need to teach them to be courageous’. (Fiona) ‘There needs to be a way in terms of how we start breaking those walls and start having conversations in terms of me and you, this is where I come from and where you come from. Not based on tribe ethnicity or your race or your religion - then we start developing empathy in a different way. So my priority now is I’m doing more in terms of one-to-one where people just want to have a conversation’. (Sam) The resultant model brings these themes together in a continuous process of self- transcendent infused social activation which results in individual, community and (in the case of participants) global impact. 4. Discussion In the midst of alarming news about escalating and urgent global problems, where every day millions live without adequate food, water and sanitation; children die from malnutrition, HIV kills thousands of people, increased carbon dioxide and other human- made emissions injure the planet and human activities create a wave of extinction of plants and animals, the lives of individuals and the well-being of disadvantaged communities is being transformed through the activities of impactful self-transcendent social activists. The positive contagion of their actions is felt globally through influence, replication, leadership training and education. Experiencing notable levels of eudemonic wellbeing [109] participants describe their lives as joyful, deeply fulfilled, privileged, spiritual and meaningful. Leading meaningful lives sensed as a calling, seeking no personal recognition or accolades, born from a deep feeling of connectedness and a vision of how life could be better, participants described what motivated them to ‘focus on what really matters’ (Jemma) by committing their lives to a self-transcendent purpose directed towards serving others [65]. What started off as an exploration into the experience of self-transcendence within the context of social activism, led to the emergence of a ‘self-transcendence infused’ values driven model (see Figure 1) of social activism that describes four key processes—trigger, activate, maintain and sustain. The model presents a continuous process of activism that Behav. Sci. 2023, 13, 66 16 of 22 generates personal joy fulfilment and meaning whilst creating a ripple effect of positive contagion that can be leveraged to address community and global issues. Figure 1. Self-transcendent social activism. 4.1. Self-Transcendent Social Activism A combination of early role models, exposure to social injustice, personal experience of tragedy and feeling ‘called’ triggered a resolve to help others; findings that are align to research carried out by Dutt and Grabe [110]. Empathy, compassion, a sense of connection, courage and faith moved participants from simply having a vison of how life could be better, to take action. Maintaining social activism requires sacrifice and is challenging; participants listed a number of factors that enhanced their commitment and motivation including being surrounded by a community of like-minded individuals for support [111], willingness to make personal sacrifices, self-care, seeing possibilities rather than problems and adopting an empathetic approach that empowers communities. Sustaining momentum, so that the ripple effect of their activism reaches new communities and future generations and becomes more universally contagious, involves having a global perspective, growing leaders, and embedding the concepts of empathy, awareness and courage into coaching, mentoring and educational organisation systems. 4.2. Context The study results have broader implications as shown in matrix below (Figure 2), which depicts comparative levels of activism and self-transcendence. Initiated in Hawaii, the approach taken to develop housing projects for the homeless has been extended to Cambodia and Kenya and is recognized by the UN. The approach that led to the creation and organic growth of a center of educational, medical and social support facilities located in a Kenya slum emaciated by HIV, is being multiplied through a ‘franchise’ type methodology and mentoring like-minded activists. A program which started many years ago in Romania, to rescue orphans, has led to a similar program being brought to Kenya. These are examples of how the influence of participant’s activism extends well beyond local communities. Participants self-identified as self-transcendent social activists thereby occupying quadrant B on the matrix above. High self-transcendence combined with high social activism has led to the development of sustainable co-produced community enterprises [60,64]. Here, a number of factors have coalesced, resulting in significant com- Behav. Sci. 2023, 13, 66 16 of 22 The resultant model brings these themes together in a continuous process of self-transcendent infused social activation which results in individual, community and (in the case of participants) global impact. 4. Discussion In the midst of alarming news about escalating and urgent global problems, where every day millions live without adequate food, water and sanitation; children die from malnutrition, HIV kills thousands of people, increased carbon dioxide and other human-made emissions injure the planet and human activities create a wave of extinction of plants and animals, the lives of individuals and the well-being of disadvantaged commu-nities is being transformed through the activities of impactful self-transcendent social ac-tivists. The positive contagion of their actions is felt globally through influence, replica-tion, leadership training and education. Experiencing notable levels of eudemonic wellbeing [109] participants describe their lives as joyful, deeply fulfilled, privileged, spiritual and meaningful. Leading meaningful lives sensed as a calling, seeking no personal recognition or accolades, born from a deep feeling of connectedness and a vision of how life could be better, participants described what motivated them to ‘focus on what really matters’ (Jemma) by committing their lives to a self-transcendent purpose directed towards serving others [65]. What started off as an exploration into the experience of self-transcendence within the context of social activism, led to the emergence of a ‘self-transcendence infused’ values driven model (see Figure 1) of social activism that describes four key processes—trigger, activate, maintain and sustain. The model presents a continuous process of activism that generates personal joy fulfilment and meaning whilst creating a ripple effect of positive contagion that can be leveraged to address community and global issues. Figure 1. Self-transcendent social activism. 4.1. Self-Transcendent Social Activism A combination of early role models, exposure to social injustice, personal experience of tragedy and feeling ‘called’ triggered a resolve to help others; findings that are align to research carried out by Dutt and Grabe [110]. Empathy, compassion, a sense of connec-tion, courage and faith moved participants from simply having a vison of how life could be better, to take action. Maintaining social activism requires sacrifice and is challenging; Behav. Sci. 2023, 13, 66 17 of 22 munity and global impact. By fully identifying with disadvantaged communities, working alongside them, contributing much needed resources and skills, empowering, training and co-producing sustainable initiatives, participants have delivered significant results. Figure 2. Self-transcendence + social-activism = impact. Quadrant A represents non-activated self-transcendence where the impact of a self- transcendent lifestyle remains individualistic. Feelings of connection to something greater than oneself and the motivation to do something meaningful are incubated before being activated. Life for research participants commenced in this space as self-awareness, aware- ness of injustice, and a desire to be of service increased. Feeling empathetic, compassionate and connected, believing they had a role to play in helping to improve the lives of others, exercising faith and bravery, overcoming challenges to pursue a goal or conviction [112]; participants moved from quadrant A to quadrant B by demonstrating commitment and a willingness to step out of comfort zones and confront challenge [113]. Quadrant C represents a form of activism that is not infused with self-transcendence values. Often more ego than eco driven, and sometimes driven by entrepreneurism and a desire to generate profit, frequently less impactful ‘solutions’ are imposed rather than co-created and are short-lived [60,114]. The research did not involve collecting Quadrant D data, which represents low self- transcendence and low activism; however, from spiritual literature [115], we may speculate that, for some, this is a lonely position, possibly with high levels of neuroticism and alienation [116] representing potential ground for further social activism. Self-transcendent social activism, which involves the integration of ego and eco goals is highly impactful. This form of activism leads to the development of co-produced sustainable initiatives and solutions that empower local communities and create positive contagion. In comparison, non-self-transcendent activism, often motivated by personal agendas, and the need for personal recognition leads to ‘outsider’ imposed, less sustainable, models and often causes resentment. Self-transcendent activism operates from a position of ‘empathy’. According to Sam, ‘empathy involves listening with the heart, whereas sympathy involves listening to the head.’ ‘If empathy is not in you then you’re totally missing the point’. Passive self-transcendence may benefit an individual; however, increasing societal impact involves transitioning from passive to active self-transcendence. Amongst other Behav. Sci. 2023, 13, 66 17 of 22 participants listed a number of factors that enhanced their commitment and motivation including being surrounded by a community of like-minded individuals for support [111], willingness to make personal sacrifices, self-care, seeing possibilities rather than problems and adopting an empathetic approach that empowers communities. Sustaining momen-tum, so that the ripple effect of their activism reaches new communities and future gen-erations and becomes more universally contagious, involves having a global perspective, growing leaders, and embedding the concepts of empathy, awareness and courage into coaching, mentoring and educational organisation systems. 4.2. Context The study results have broader implications as shown in matrix below (Figure 2), which depicts comparative levels of activism and self-transcendence. Figure 2. Self-transcendence + social-activism = impact. Initiated in Hawaii, the approach taken to develop housing projects for the homeless has been extended to Cambodia and Kenya and is recognized by the UN. The approach that led to the creation and organic growth of a center of educational, medical and social support facilities located in a Kenya slum emaciated by HIV, is being multiplied through a ‘franchise’ type methodology and mentoring like-minded activists. A program which started many years ago in Romania, to rescue orphans, has led to a similar program being brought to Kenya. These are examples of how the influence of participant’s activism ex-tends well beyond local communities. Participants self-identified as self-transcendent so-cial activists thereby occupying quadrant B on the matrix above. High self-transcendence combined with high social activism has led to the development of sustainable co-pro-duced community enterprises [60,64]. Here, a number of factors have coalesced, resulting in significant community and global impact. By fully identifying with disadvantaged communities, working alongside them, contributing much needed resources and skills, empowering, training and co-producing sustainable initiatives, participants have deliv-ered significant results. Quadrant A represents non-activated self-transcendence where the impact of a self-transcendent lifestyle remains individualistic. Feelings of connection to something greater than oneself and the motivation to do something meaningful are incubated before being activated. Life for research participants commenced in this space as self-awareness, Behav. Sci. 2023, 13, 66 18 of 22 things, moving from passive to active requires developing a vision of how life can be better [63], believing one can make a difference, and having courage. Courage can be taught [112]. The implications and application of this study are far reaching. The study suggests that teaching and modelling empathy, compassion and courage and embedding each stage of the ‘self-transcendent social activism model’, into coaching, mentoring and educational interventions will result in increased positive individual and community impact, generating a ripple effect of positive contagion which can be leveraged to address global challenges. 4.3. Limitations and Future Research A number of limitations of the current study should be considered when examining the results and conclusions. Findings were based on eight interviews with participants who self-identified as self-transcendent social activists. A limitation of the study was the predominance of female (6), and Christian (6), participants. Within the scope of the interviews, arguably, data saturation was achieved, and no new information emerged from latter interviews. However, given more time, it would be possible to increase the sample size and to extend the range of interview questions. The researcher has attempted to eliminate personal bias; however, a number of participants were known to her. Future research, deploying a quantitative methodology to evidence impact and the use of scales to measure the relationship between transcendence, activism and wellbeing would strengthen findings. Researching activism within the context of quadrant C—to include volunteerism, entrepreneurialism, and career activism would prove insightful. Furthermore, testing the model in terms of training, taking before and after mea- surements to evidence the effectiveness of interventions designed to develop empathy, compassion and courage, is suggested by the researcher. 5. Conclusions This study contributes to the extant of the literature by expanding our understanding of self-transcendence as a driver of social activism. It has resulted in the development of a new model of ‘self-transcendent social activism’ containing four key processes: trigger, activate, maintain and sustain engagement with social activism. Author Contributions: Conceptualization: C.B. and R.H.; methodology: C.B. and R.H.; software: Not relevant; validation: C.B.; formal analysis: C.B.; investigation: C.B.; resources: C.B. and R.H.; data curation: C.B.; writing—original draft preparation: C.B.; writing—review and editing: R.H. and C.B.; visualization: C.B. and R.H.; supervision: R.H.; project administration: C.B. and R.H.; funding acquisition: not relevant. All authors have read and agreed to the published version of the manuscript. Funding: This research received no external funding. Institutional Review Board Statement: The study was conducted in accordance with the Decla- ration of Helsinki, and approved by the University of East London Ethics Committee for studies involving humans. Informed Consent Statement: Informed consent was obtained from all subjects involved in the study. Data Availability Statement: Data supporting the reported results is kept by the first author. Acknowledgments: We would like to thank the visionary, courageous, resourceful research partici- pants, all of whom are engaged in initiatives to transform the lives of individuals and communities. They are the inspiration behind this study, and we want to showcase the great work they are doing. Conflicts of Interest: The authors declare no conflict of interest. However, we do note that the first author is a coach and consultant who has had some professional involvement with the organizations represented by several of the research participants. Behav. Sci. 2023, 13, 66 References 19 of 22 Luther King, M. The Autobiography Of Martin Luther King, Jr.; Warner Books: New York, NY, USA, 1998. 1. 2. Mandella, N. Long Walk to Freedom; Abacus: London, UK, 1995. 3. 4. 5. 6. McCarthy, V.L.; Ling, R.N.J.; Carini, R.M. The Role of Self-Transcendence: A Missing Variable in the Pursuit of Successful Aging? Tutu, D. No Future Without Forgiveness; Doubleday: New York, NY, USA, 2000. Barker, C.; Martin, B.; Zournazi, M. Emotional self-management for activists. Reflective Pract. 2008, 9, 423–435. [CrossRef] Briscoe, F.; Gupta, A. Social activism in and around organizations. Acad. Manag. Ann. 2016, 10, 671–727. [CrossRef] 7. 8. 9. Res. Gerontol. Nurs. 2013, 6, 178–186. [CrossRef] [PubMed] Jasko, K.; Szastok, M.; Grzymala-Moszczynska, J.; Maj, M.; Kruglanski, A.W. Rebel with a cause: Personal significance from political activism predicts willingness to self-sacrifice. J. Soc. Issues 2019, 75, 314–349. [CrossRef] Smith, H.J.; Pettigrew, T.F.; Pippin, G.M.; Bialosiewicz, S. Relative deprivation: A theoretical and meta-analytic review. Personal. Soc. Psychol. Rev. 2012, 16, 203–232. [CrossRef] Fielding, K.S.; McDonald, R.; Louis, W.R. Theory of planned behaviour, identity and intentions to engage in environmental activism. J. Environ. Psychol. 2008, 28, 318–326. [CrossRef] 10. Miller, D.A.; Cronin, T.; Garcia, A.L.; Branscombe, N.R. The relative impact of anger and efficacy on collective action is affected by feelings of fear. Group Process. Intergroup Relat. 2009, 12, 445–462. [CrossRef] 11. Tausch, N.; Becker, J.C. Emotional reactions to success and failure of collective action as predictors of future action intentions: A longitudinal investigation in the context of student protests in Germany. Br. J. Soc. Psychol. 2013, 52, 525–542. [CrossRef] [PubMed] 12. Van Zomeren, M. Four core social-psychological motivations to undertake collective action. Soc. Personal. Psychol. Compass 2013, 7, 378–388. [CrossRef] 13. Kteily, N.; Bruneau, E. Backlash: The politics and real-world consequences of minority group dehumanization. Personal. Soc. Psychol. Bull. 2017, 43, 87–104. [CrossRef] 14. De Zavala, A.G.; Cichocka, A.; Eidelson, R.; Jayawickreme, N. Collective narcissism and its social consequences. J. Personal. Soc. Psychol. 2009, 97, 1074–1081. [CrossRef] [PubMed] 15. Lyons-Padilla, S.; Gelfand, M.J.; Mirahmadi, H.; Farooq, M.; Van Egmond, M. Belonging nowhere: Marginalization & radicaliza- tion risk among Muslim immigrants. Behav. Sci. Policy 2015, 1, 1–12. 16. Piazza, J.A. Rooted in poverty? Terrorism, poor economic development, and social cleavages. Terror. Political Violence 2006, 18, 159–177. [CrossRef] 17. Piazza, J.A. Poverty, minority economic discrimination, and domestic terrorism. J. Peace Res. 2011, 48, 339–353. [CrossRef] 18. McAdam, D.; Friedman, D. Collective identity and activism. In Frontiers in Social Movement Theory; Yale University Press: New Haven, CT, USA, 1992; pp. 156–174. 19. Dono, J.; Webb, J.; Richardson, B. The relationship between environmental activism, pro-environmental behaviour and social 20. 21. identity. J. Environ. Psychol. 2010, 30, 178–186. [CrossRef] Schulte, M.; Bamberg, S.; Rees, J.; Rollin, P. Social identity as a key concept for connecting transformative societal change with individual environmental activism. J. Environ. Psychol. 2020, 72, 101525. [CrossRef] Foster, M.D.; Tassone, A.; Matheson, K. Tweeting about sexism motivates further activism: A social identity perspective. Br. J. Soc. Psychol. 2021, 60, 741–764. [CrossRef] [PubMed] 22. Matsuba, M.K.; Pratt, M.W.; Norris, J.E.; Mohle, E.; Alisat, S.; McAdams, D.P. Environmentalism as a context for expressing identity and generativity: Patterns among activists and uninvolved youth and midlife adults. J. Personal. 2012, 80, 1091–1115. [CrossRef] 23. Valocchi, S. Activism as a Career, Calling, and Way of Life. J. Contemp. Ethnogr. 2013, 42, 169–200. [CrossRef] 24. Klar, M.; Kasser, T. Some benefits of being an activist: Measuring activism and its role in psychological well-being. Political Psychol. 2009, 30, 755–777. [CrossRef] 25. Dugas, M.; Bélanger, J.J.; Moyano, M.; Schumpe, B.M.; Kruglanski, A.W.; Gelfand, M.J.; Touchton-Leonard, K.; Nociti, N. The quest for significance motivates self-sacrifice. Motiv. Sci. 2016, 2, 15–32. [CrossRef] 26. Venhoeven, L.A.; Bolderdijk, J.; Steg, L. Explaining the paradox: How proenvironmental behaviour can both thwart and foster well-being. Sustainability 2013, 5, 1372–1386. [CrossRef] 27. Becker, J.C.; Tausch, N.; Wagner, U. Emotional consequences of collective action participation: Differentiating self-directed and outgroup-directed emotions. Personal. Soc. Psychol. Bull. 2011, 37, 1587–1598. [CrossRef] 28. Tropp, L.R.; Brown, A.C. What benefits the group can also benefit the individual: Group-enhancing and individual-enhancing motives for collective action. Group Process. Intergroup Relat. 2004, 7, 267–282. [CrossRef] 29. Back, E.A.; Back, H.; Knapton, H.M. Group belongingness and collective action: Effects of need to belong and rejection sensitivity on willingness to participate in protest activities. Scand. J. Psychol. 2015, 56, 537–544. [CrossRef] 30. Wong, P.T.P. Self-Transcendence: A Paradoxical Way to Become Your Best. Int. J. Existent. Psychol. Psychother. 2016, 6, 9. 31. Sohr, S. Eco-activism and well-being: Between flow and burnout. In Life-Goals and Well-Being: Towards a Positive Psychology of Human Striving; Schmuck, P., Sheldon, K.M., Eds.; Hogrefe & Huber: Cambridge, WA, USA, 2001; pp. 202–215. Behav. Sci. 2023, 13, 66 20 of 22 32. Eigner, S. The relationship between “protecting the environment” as a dominant life goal and subjective well-being. In Life-Goals and Well-Being: Towards a Positive Psychology of Human Striving; Schmuck, P., Sheldon, K.M., Eds.; Hogrefe & Huber: Cambridge, WA, USA, 2001; pp. 182–201. 33. Kruglanski, A.W.; Jasko, K.; Chernikova, M.; Dugas, M.; Webber, D. To the fringe and back: Violent extremism and the psychology 34. of deviance. Am. Psychol. 2017, 72, 217–230. [CrossRef] Jasko, K.; LaFree, G.; Kruglanski, A. Quest for significance and violent extremism: The case of domestic radicalization. Political Psychol. 2017, 38, 815–831. [CrossRef] 35. Drury, J.; Reicher, S. Explaining enduring empowerment: A comparative study of collective action and psychological outcomes. Eur. J. Soc. Psychol. 2005, 35, 35–58. [CrossRef] 36. Evripidou, A.; Drury, J. This is the time of tension: Collective action and subjective power in the Greek anti-austerity movement. Contention: Multidiscip. J. Soc. Protest 2013, 1, 31–51. [CrossRef] 37. Apoifis, N. Anarchy in Athens: An Ethnography of Militancy, Emotions and Violence; Manchester University Press: Manchester, UK, 2016. 38. Lee, Y.; Kim, S.; Kim, M.; Choi, J. Antecedents and interrelationships of three types of pro-environmental behavior. J. Bus. Res. 2014, 67, 2097–2105. [CrossRef] 39. Liss, M.; Crawford, M.; Popp, D. Predictors and correlates of collective action. Sex Roles 2004, 50, 771–779. [CrossRef] 40. Miller, J.M.; Krosnick, J.A. Threat as a motivator of political activism: A field experiment. Political Psychol. 2004, 25, 507–523. [CrossRef] 41. Unsworth, K.L.; McNeill, I.M. Increasing pro-environmental behaviors by increasing self-concordance: Testing an intervention. J. Appl. Psychol. 2017, 102, 88–103. [CrossRef] 42. Van Zomeren, M.; Postmes, T.; Spears, R.; Bettache, K. Can moral convictions motivate the advantaged to challenge social inequality? Extending the social identity model of collective action. Group Process. Intergroup Relat. 2011, 14, 735–753. [CrossRef] 43. Den Hond, F.; De Bakker, F.G. Ideologically motivated activism: How activist groups influence corporate social change activities. 44. Acad. Manag. Rev. 2007, 32, 901–924. [CrossRef] Skitka, L.J.; Bauman, C.W.; Sargis, E.G. Moral conviction: Another contributor to attitude strength or something more? J. Personal. Soc. Psychol. 2005, 88, 895–917. [CrossRef] Skitka, L.J.; Bauman, C.W. Moral conviction and political engagement. Political Psychol. 2008, 29, 29–54. [CrossRef] 45. 46. Atran, S.; Ginges, J. Religious and sacred imperatives in human conflict. Science 2012, 336, 855–857. [CrossRef] 47. Molinario, E.; Kruglanski, A.W.; Bonaiuto, F.; Bonnes, M.; Cicero, L.; Fornara, F.; Scopelliti, M.; Admiraal, J.; Beringer, A.; Dedeurwaerdere, T.; et al. Motivations to Act for the Protection of Nature Biodiversity and the Environment: A Matter of “Significance”. Environ. Behav. 2020, 52, 1133–1163. [CrossRef] Sherman, D.K.; Cohen, G.L. The psychology of self-defense: Self-affirmation theory. Adv. Exp. Soc. Psychol. 2006, 38, 183–242. Skitka, L.J.; Hanson, B.E.; Wisneski, D.C. Utopian hopes or dystopian fears? Exploring the motivational underpinnings of moralized political engagement. Personal. Soc. Psychol. Bull. 2017, 43, 177–190. [CrossRef] [PubMed] 48. 49. 50. Kleres, J.; Wettergren, Å. Fear, hope, anger, and guilt in climate activism. Soc. Mov. Stud. 2017, 16, 507–519. [CrossRef] 51. Jacobsson, K.; Lindblom, J. Emotion work in animal rights activism: A moral-sociological perspective. Acta Sociol. 2013, 56, 55–68. [CrossRef] 52. Huo, H.; Qi, H.; Newton, N.J. Examining Identity and Generativity among Middle-aged Female Activists in Two Cultural 53. Contexts. J. Adult Dev. 2022, 29, 318–327. [CrossRef] Singer, J.A.; King, L.A.; Green, M.C.; Barr, S.C. Personal identity and civic responsibility: “Rising to the occasion” narratives and generativity in community action student interns. J. Soc. Issues 2002, 58, 535–556. [CrossRef] 54. Alisat, S.; Norris, J.E.; Pratt, M.W.; Matsuba, M.K.; McAdams, D.P. Caring for the earth: Generativity as a mediator for the prediction of environmenta narratives from identity among activists and nonactivists. Identity: Int. J. Theory Res. 2014, 14, 177–194. [CrossRef] Suk, J.; Abhishek, A.; Zhang, Y.; Ahn, S.Y.; Correa, T.; Garlough, C.; Shah, D.V. # MeToo, networked acknowledgment, and connective action: How “empowerment through empathy” launched a social movement. Soc. Sci. Comput. Rev. 2021, 39, 276–294. 56. Todorova, M. Interpreting for refugees: Empathy and activism. In Intercultural Crisis Communication; Bloomsbury Publishing: 55. London, UK, 2019; pp. 153–171. 57. Edrington, C.L. Social Movements and Identification: An Examination of How Black Lives Matter and March for Our Lives Use Identification Strategies on Twitter to Build Relationships. J. Mass Commun. Q. 2022, 99, 643–659. [CrossRef] 58. Elisha, O. Moral ambitions of grace: The paradox of compassion and accountability in Evangelical faith-based activism. Cult. Anthropol. 2008, 23, 154–189. [CrossRef] 59. Yong, A. All Mixed up: Multi/Racial Liberation and Compassion-Based Activism. Religions 2020, 11, 402. [CrossRef] 60. Crocker, J. From Egosystem to Ecosystem: Implications for Relationships, Learning and Well-Being. In Transcending Self-Interest; Wament, H., Bauer, J., Eds.; American Psychological Association: Washington, DC, USA, 2008; pp. 171–194. 61. McVeigh, K. Comic Relief to ditch White savior stereotype appeals. Guard 2018, 23, 29–37. 62. Kherbaoui, J.; Aronson, B. Bleeding through the band-aid: The white savior industrial complex. In Routledge Handbook of Critical Studies in Whiteness; Hunter, S., Van der Westhuizen, C., Eds.; Routledge: Oxon, UK, 2021; pp. 269–279. 63. Korsgaard, M.; Meglino, B. Beyond the Individualistic Self—A framework for pro-social motives and judgements. In Transcending Self-Interest; Wayment, H., Bauer, J., Eds.; APA: Washington, DC, USA, 2008; pp. 183–198. Behav. Sci. 2023, 13, 66 21 of 22 64. Cahn, E. No More Throw-Away People; Essential: Washington, DC, USA, 2000. 65. 66. Aron, A.; Aron, E.A.; Tudor, M.; Nelson, G. Close Relationships as Including Other in the Self. J. Personal. Soc. Psychol. 1991, Frankl, V.E. Self-Transcendence as a Human Phenomenon. J. Humanist. Psychol. 1966, 6, 97–106. [CrossRef] 60, 241–253. [CrossRef] 67. Hartman, D.; Zimberoff, D. Self-transcendence and Ego Surrender, a Quiet-enough Ego or an Ever-quieter enough Ego; Wellness Press: Washington, DC, USA, 2005. 68. Reed, P.G. Demystifying Self-Transcendence for Mental Health Nursing Practice and Research. Arch. Psychiatr. Nurs. 2009, 23, 397–400. [CrossRef] [PubMed] 69. McCrathy, V.L.; Bockweg, A. The role of transcendence in a holistic view of successful aging: A concept analysis and model of transcendence. J. Holist. Nurs. 2013, 31, 83–91. [CrossRef] [PubMed] 70. Maslow, A.H. Lessons from the peak experiences. J. Humanist. Psychol. 1962, 2, 9–18. [CrossRef] 71. Kaufman, S.B. Transcend: The New Science of Self-Actualization; Penguin: London, UK, 2021. 72. Koltko-Rivera, M.E. Rediscovering the later version of Maslow’s hierarchy of needs: Self-transcendence and opportunities for 73. 74. theory, research, and unification. Rev. Gen. Psychol. 2006, 10, 302–317. [CrossRef] Schwartz, S.H. An overview of the Schwartz theory of basic values. Online Read. Psychol. Cult. 2012, 2, 11–23. [CrossRef] Schwartz, S.H. Universals in the content and structure of values: Theoretical advances and empirical tests in 20 countries. Adv. Exp. Soc. Psychol. 1992, 25, 1–65. 75. Piedmont, R.L. Does spirituality represent the sixth factor of personality? Spiritual transcendence and the five-factor model. J. Personal. 1999, 67, 985–1013. [CrossRef] 76. Cloninger, C.R.; Svrakic, D.M.; Przybeck, T.R. A psychobiological model of temperament and character. Arch. Gen. Psychiatry 1993, 50, 975–990. [CrossRef] 77. Peterson, C.; Seligman, M. Character Strengths and Virtues; Oxford University Press: New York, NY, USA, 2004. 78. Ellermann, C.R.; Reed, P.G. Self-transcendence and depression in middle-age adults. West. J. Nurs. Res. 2001, 23, 698–713. [CrossRef] 79. Klaas, D. Testing two elements of spirituality in depressed and non-depressed elders. Int. J. Psychiatr. Nurs. Res. 1998, 4, 452–462. Stinson, C.K.; Kirk, E. Structured reminiscence: An intervention to decrease depression and increase self-transcendence in older 80. women. J. Clin. Nurs. 2006, 15, 208–218. [CrossRef] 81. Vecchione, M.; Schwartz, S.H.; Caprara, G.V.; Schoen, H.; Cieciuch, J.; Silvester, J.; Bain, P.; Bianchi, G.; Baslevent, C.; Mamali, C.; et al. Personal values and political activism: A cross-national study. Br. J. Psychol. 2015, 106, 84–106. [Cross- Ref] 82. Caprara, G.V.; Alessandri, G.; Eisenberg, N. Prosociality: The contribution of traits, values, and self-efficacy beliefs. J. Personal. Soc. Psychol. 2012, 102, 1289–1303. [CrossRef] 83. Ardenghi, S.; Rampoldi, G.; Bani, M.; Strepparava, M.G. Personal values as early predictors of emotional and cognitive empathy among medical students. Curr. Psychol. 2021, 1–9. [CrossRef] 84. Danioni, F.; Barni, D. The relations between adolescents’ personal values and prosocial and antisocial behaviours in team sports. Int. J. Sport Exerc. Psychol. 2019, 17, 459–476. [CrossRef] 85. Tamir, M.; Schwartz, S.H.; Cieciuch, J.; Riediger, M.; Torres, C.; Scollon, C.; Vishkin, A. Desired emotions across cultures: 86. 87. A value-based account. J. Personal. Soc. Psychol. 2016, 111, 67–82. [CrossRef] Schwartz, S.H.; Rubel, T. Sex differences in value priorities: Cross-cultural and multi-method studies. J. Personal. Soc. Psychol. 2005, 89, 1010–1028. [CrossRef] Sortheix, F.M.; Lönnqvist, J.E. Personal value priorities and life satisfaction in Europe: The moderating role of socioeconomic development. J. Cross-Cult. Psychol. 2014, 45, 282–299. [CrossRef] 88. Russo, C.; Barni, D.; Zagrean, I.; Lulli, M.A.; Vecchi, G.; Danioni, F. The resilient recovery from substance addiction: The role of 89. self-transcendence values and hope. Mediterr. J. Clin. Psychol. 2021, 9, 1–20. Fegg, M.J.; Wasner, M.; Neudert, C.; Borasio, G.D. Personal values and individual quality of life in palliative care patients. J. Pain Symptom Manag. 2005, 30, 154–159. [CrossRef] 90. Danioni, F.; Ranieri, S.; Villani, D. The role of personal values in gambling: A preliminary study with Italian adolescents. J. Genet. Psychol. 2020, 181, 413–426. [CrossRef] 91. Neill, J. Transcendence and transformation in the life patterns of women living with rheumatoid arthritis. Adv. Nurs. Sci. 2002, 24, 27–47. [CrossRef] 92. Coward, D.D. Self-transcendence and correlates in a healthy population. Nurs. Res. 1996, 45, 116–121. [CrossRef] 93. Coward, D.D. Facilitation of self-transcendence in a breast cancer support group II. Oncol. Nurs. Forum 2003, 30, 291–300. [CrossRef] 94. Nygren, B.; Aléx, L.; Jonsén, E.; Gustafson, Y.; Norberg, A.; Lundman, B. Resilience, sense of coherence, purpose in life and self-transcendence in relation to perceived physical and mental health among the oldest old. Aging Ment. Health 2005, 9, 354–362. [CrossRef] 95. Csikszentmihalyi, M. Flow: The Psychology of Optimal Experience; Harper & Row: New York, NY, USA, 1990. 96. Yaden, D.B.; Haidt, J.; Hood Jr, R.W.; Vago, D.R.; Newberg, A.B. The Varieties of Self-Transcendent Experience. Rev. Gen. Psychol. 2017, 17, 1089–2680. [CrossRef] Behav. Sci. 2023, 13, 66 22 of 22 97. Batthyany, A.; Russo-Netzer, P. Psychologies of meaning. In Meaning in Positive and Existential Psychology; Batthyany, A., Russo-Netzer, P., Eds.; Springer: Berlin/Heidelberg, Germany, 2014; pp. 3–22. 98. Gundelach, P.; Toubøl, J. High-and low-risk activism: Differential participation in a refugee solidarity movement. Mobilization Int. 99. Q. 2019, 24, 199–220. [CrossRef] Jia, F.; Soucie, K.; Alisat, S.; Curtin, D.; Pratt, M. Are environmental issues moral issues? Moral identity in relation to protecting the natural world. J. Environ. Psychol. 2017, 52, 104–113. [CrossRef] 100. Hackett, J.D. Values anchoring: Strengthening the link between values and activist behaviors. Soc. Influ. 2014, 9, 99–115. [CrossRef] 101. Charmaz, K. Grounded Theory Methods in Social Justice Research. In The Sage Handbook of Qualitative Research; Denzin, N.K., Lincoln, Y.S., Eds.; Sage: London, UK, 2011; pp. 359–380. 102. Charmaz, K. Constructing Grounded Theory; Sage: London, UK, 2006. 103. Glaser, B.G. Constructivist Grounded Theory? Glaser, Barney G. Constructivist Grounded Theory? [47 paragraphs]. Forum Qual. Soz./Forum Qual. Soc. Res. 2002, 3, 12. 104. Sullivan, W.; Rees, J. Clean Language: Revealing Metaphors and Opening Minds; Crown House Publishing: Carmarthen, Wales, 2008. 105. Padgett, D. Does the Glove Really Fit? Qualitative Research and Clinical Social Work Practice. Soc. Work. 1998, 43, 373–381. [CrossRef] 106. Bazeley, P. Qualitative Data Analysis with NVivo; Sage: London, UK, 2007. 107. Hutchison, A.J.; Johnston, L.H.; Breckon, J.D. Using QSR-NVivo to facilitate the development of a grounded theory project: An account of a worked example. Int. J. Soc. Res. Methodol. 2010, 13, 283–302. [CrossRef] 108. Bringer, J.D.; Johnston, L.H.; Brackenridge, C.H. Using computer assisted qualitative data analysis software to develop a grounded theory project. Field Methods 2006, 18, 245–266. [CrossRef] 109. Hart, R. Positive Psychology—The Basics; Routledge: London, UK, 2020. 110. Dutt, A.; Grabe, S. Lifetime activism, marginality, and psychology: Narratives of lifelong feminist activists committed to social change. Qual. Psychol. 2014, 1, 107–122. [CrossRef] 111. Bettencourt, B.A.; Dillmann, G.; Wollman, N. The intragroup dynamics of maintaining a successful grassroots organization: A case study. J. Soc. Issues 1996, 52, 169–186. [CrossRef] 112. Biswas-Diener, R. The Courage Quotient: How Science Can Make You Braver; Willey & Sons: Hoboken, NJ, USA, 2012. 113. Maddi, S.R. Hardiness: An Operationalization of Existential Courage. J. Humanist. Psychol. 2004, 44, 279–298. [CrossRef] 114. Bell, K.M. Raising Africa? Celebrity and the rhetoric of the white savior. Portal: J. Multidiscip. Int. Stud. 2013, 10, 1–24. 115. Rohr, R. The Divine Dance; Whitaker House: Pennsylvania, PA, USA, 2016. 116. Levenson, M.R.; Jennings, P.A.; Aldwin, C.M.; Shiraishi, R.W. Self-transcendence: Conceptualization and measurement. Int. J. Aging Hum. Dev. 2005, 60, 127–143. [CrossRef] [PubMed] Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
10.1515_culture-2022-0170
Open Cultural Studies 2023; 7: 20220170 Regular Article Can Kocak* How I Met Your Fans: A Comparative Textual Analysis of How I Met Your Mother and Its Reboots https://doi.org/10.1515/culture-2022-0170 received November 2, 2022; accepted March 1, 2023 Abstract: Since How I Met Your Mother (HIMYM) ended in 2014 with its ninth season, there have been multiple attempts to create a new series based on a similar premise – one parent telling their kid(s) how they met their other parent. How I Met Your Dad (HIMYD), which was planned to begin in 2014 right after HIMYM’s final, was not picked up by any network, while How I Met Your Father (HIMYF), which started airing in 2022, is continuing a successful [Upon its release, the show, whose premiere was watched in 420 K households, held the fifth place on the TV time charts and went on to stay on the list for four more weeks (The Entertainment Strategy Guy).] run, currently in its second season. This article compares the narrative and narration of the first episodes of these three series, including themes, recurring jokes, sound and visual effects, introductions of important characters and the “missing” parent, as well as the use of cameos. By doing this, it aims to delve into the reason behind why one of these reboot attempts was favoured over the other. The answer is provided as the difference between the amount of fan service in HIMYD and HIMYF, with the latter establishing more direct connections with HIMYM. In addition to becoming a rather common tendency by creatives, this article claims that fan service also seems to be used by networks and production companies as a way of ensuring audience engagement. Keywords: reboots, fandom, fan service, cameos, audience engagement Introduction TV Reboots help re-establish or revive the connection with an older TV show that is beloved by audiences. This means that in order for a reboot to come into life, there are two integral factors – one is a commercially successful, older TV show, and the other is a group of devout fans, whose presence on the other side of the screen help define this success. One noteworthy idea in early audience research was the “direct effects” model (McQuail 18), which spoke about the creators of TV shows – or any media product – setting the rules and the audiences – or fans – following along without any interference. The message was disseminated unidirectionally, in a linear fashion. Later on, ideas such as the “encoding/decoding model” (Hall), which gave interpretations as much power and authority as messages themselves, were introduced. TV Reboots take these models that attribute a certain autonomy towards audiences one step further, shifting the dynamic completely in favour of fans. In her article “American TV Series Revivals,” Loock writes about how reboots function, claiming that they “[…] rely on the televisual past to circulate new products through the crowded contemporary media landscape and […] seek to negotiate the televisual heritage of original series and feelings of generational  * Corresponding author: Can Kocak, Media and Cultural Studies, University of Sussex, Brighton, United Kingdom, e-mail: c.kocak@sussex.ac.uk Open Access. © 2023 the author(s), published by De Gruyter. This work is licensed under the Creative Commons Attribution 4.0 International License. 2  Can Kocak belonging, as well as notions of the past, present, and future in meaningful ways” (299). Therefore, the very fact that a reboot exists also suggests a negotiation of how the original series would be received, integrating the sense of nostalgia into the equation. However, this article will argue that in terms of content, some reboots steer this negotiation towards a more hegemonic reading (Hall) and idolise the original material. Throughout the past decade, the mainstream cultural milieu has started to be dominated by a phenom- enon referred to as “fan service.” Finding its roots in the Western love for Japanese manga, the phrase denotes creators of cultural products providing “narrative rewards” that cater to the fans’ desires (Beaty). Creators act sometimes as a response to tangible and specific demands by the fans from changing a character’s design (Power) to releasing the director’s cut of a global franchise (Dockterman), and other times on what they presume will be liked by the fans with “easter eggs,” on-screen hidden elements that require a certain investment in the narrative universe to notice (Beaty). As for reboots, due to an inextricable link with the original content, every endeavour could be inter- preted as fan service in one way or another – someone has to love something for it to come back, and reviving that very thing would be a service to those people. However, there are also examples where fans’ involvement is more noteworthy than others. One of them is the revival of Veronica Mars,¹ which was brought back as a feature film in 2014 through an effort of crowdfunding by the fans (Kelly)². Here, fans could be seen as partners or co-owners of the franchise, and their involvement in the content would have made sense from an economic point of view. However, as it dominates the way stories are told all around the current landscape, fan service also serves as a factor in the creation of reboots, arguably affecting content more than it ever has. Even though reboots did not start in the early twenty-first century, the recent increase in numbers has attracted the attention of many critics, some labelling the trend as “entertainment Groundhog Day” (Scott, “Why Are There”), others associating this with 90s kids being one of the most nostalgic generations (Rivero) or claiming that it is easier to convince producers for reboots rather than for original content (Porter, “We May Not Want”). Scholars have tied the use of familiar characters with triggering memory and evoking nostalgia (Lizardi 39) and noted how this familiarity provided comfort for audiences (Loock, “Whatever Happened to Predictability” 368). Building up on these, this article is interested in exploring how reboots function within contemporary television, as well as how fan service has affected the content itself. To illustrate this point, this article will look at two reboots that have attempted to follow the steps of How I Met Your Mother (HIMYM),³ a popular and critically acclaimed series that aired for nearly a decade from the mid-noughties to mid-2010s: How I Met Your Dad⁴ (HIMYD) and How I Met Your Father⁵ (HIMYF). As opposed to developing a new season that simply saw a continuation of the same characters’ stories – which would be referred to as a “revival” – both series tried to design new characters and used a similar premise – a story of the past being told to kids from the future. With constant jumps between the narrative’s present – audiences’ future – and the past – audiences’ present – time itself is used as a narrative and comedic device, reflecting the cultural sphere of the time it is aired in through its depiction of the past. The most notable difference between HIMYD and HIMYF is their relationship with the original material: HIMYD does not establish its ties with HIMYM as much and as directly as HIMYF, which, tying into the prominence of fan service in other cultural products as well, arguably is the reason why it never got picked up. As will be seen while comparing the three series, since HIMYF did more to evoke HIMYM, it was destined for success, while HIMYD did not feature enough fan service and was not given the chance to air. It is notable that the creators of HIMYM, Carter Bays and Craig Thomas, are also involved in the production of HIMYD, with Emily Spivey and Greta Gerwig also contributing to the writing. On the other hand, HIMYF is created by Isaac Aptaker and Elizabeth Berger, and aside from a writing credit in the pilot  1 Originally aired for three seasons between 2004 and 2007 on UPN, and later on The CW. 2 A fourth season of the series also aired on Hulu in 2019. 3 The series ran from 2005 to 2014 and aired on CBS. It is currently streaming on Disney +. 4 The series was expected to debut in 2014 but the pilot was never picked up by CBS. 5 The series began in 2022 on Hulu. A Comparative Textual Analysis of How I Met Your Mother and Its Reboots  3 episode – alongside Aptaker, Berger, and Spivey – Bays and Thomas do not seem to be that much involved in the creation of the series. This adds up to the fact that HIMYF works as a “fan letter” dedicated to HIMYM and legacy, while Carter and Bays’ previous attempt to revive the series, which saw them steer somewhat away from their original creation, was not given a chance by any network. Method In order to delve into the ideas around reboots and fan service, this article will use the first episode of each series: HIMYM’s “Pilot,” which aired on September 19, 2005, HIMYD’s pilot, which was shot in 2014, and HIMYF’s “Pilot,” which aired on January 18, 2022. To illustrate the emphasis on providing fans with what they supposedly want, embodied in the cameo of a beloved character from the original series, “Timing is Everything,” the first season finale of HIMYF, which aired on March 15, 2022, will also be included in the comparison. In line with the related terminology around television studies, this article refers to HIMYD and HIMYF as “reboots,” rather than “spin-offs” or “revivals.” This is because as opposed to a spin-off, which features a new story centring an existing character, or a revival, where the main cast of a series is brought back after a number of years for a new run (Rothman), HIMYD and HIMYF have a familiar premise but new characters. Since they do not repeat the exact same storyline as HIMYM, this article also refrains from referring to them as “remakes.” Textual analysis will be applied to focus on the narrative and narration of the series, including lan- guage, structure, patterns, and their relationship with the time in which they were released. In the context of a television series, this method refers to “an analysis of the text … that simply attempts to uncover its potential meaning through detailed close readings” (Creeber 26). For this article’s purposes, this will take the shape of looking at how voiceovers are used, to whom the story is told, how characters are established, how certain sound and visual effects are used for transitions and comedic effects, how the cultural milieu of the time is reflected onto the narrative, and how a similar idea of “the one” is reinforced through different theories around love and relationships. Going back to Creeber’s definition of textual analysis on television, with the rise of SVOD platforms, “television” has come to connote something larger than the device itself, implying a necessity for a shift in the area that studies it as well. This new paradigm has been referred to as TVIV, which took further the changes in terms of technological developments, audience behaviour, and industry, resulting in “an era of matrix media where viewing patterns, branding strategies, industrial structures, the way different media forms interact with each other or the various ways content is made available shift completely away from the television set” (Jenner 260). On the other hand, scholars have also pointed out the narrative differences between originals and reboots when the episodes of reboots’ seasons are all-in-one released, as opposed to the conventional model of weekly programming (Can 237). As opposed to continuing the same series with a new season, HIMYM’s two reboots feature new characters but use the same premise, also sticking with the “how I met your” phrase to establish their ties with the original series. Additionally, even though HIMYF is currently being aired on Hulu, it does not feature an all-in-one release, therefore not constituting a difference in viewing patterns. While this may still serve a different viewing experience for some audiences due to the availability of former episodes on the platform – those who did not start watching the series as it was aired can still binge it – neither HIYMD nor HIMYF will thus be considered a part of TVIV, as they do not feature the fluidity associated with the term and rather offer viewing experiences and branding strategies that one might find in conventional TV. 4  Can Kocak HIMYM: “Kids, I’m going to tell you an incredible story” The very first episode of HIMYM opens with a father declaring to his two kids that he would tell them a story. The year is established as 2030, the camera only shows the two kids sitting on a couch, and the father’s dialogues are only heard as a voiceover. In fact, throughout the series, this voiceover of the father from 2030 is used to provide exposition into the “past,” 2005, where the actual story takes place. The fact that the older father is voice-acted by Bob Saget, renowned for his role as the father in Full House, another beloved family comedy, is a self-reflexive and self-aware joke in and of itself, which requires from audiences to channel their extra-diegetic knowledge, much like how they would relate the content of a reboot to the original series. Full House was revived by Netflix as Fuller House in 2016 to evoke “the bygone days of TGIF-style programming of the 1980s and 1990s and reinvent family-friendly viewing in the present” (Loock, “Amer- ican TV Series Revivals” 307), turning into a nostalgic throwback for an older generation who watched Saget in Full House, and for young adult viewers who listened to his voice in HIMYM. Here, the comforting effect of a familiar voice (Loock, “Whatever Happened to Predictability?” 371) in the form of a renowned actor is juxtaposed with the father being an unreliable narrator due to a variety of factors in different instances, including his failed memory, his reluctance to share certain details with his kids, and his will- ingness to tease them (Terry 7). The father’s name is provided as Ted Mosby (Josh Radnor), and the audience sees 2005 Ted starting to feel sad that he has not figured out what to do with his life. This is especially instigated by the fact that his best friend Marshall (Jason Segel) decides to propose to his long-time girlfriend, Lily (Alyson Hannigan). The final member of their gang, Barney (Neil Patrick Harris), is introduced through his catch-phrases such as “suit up,” “wait for it,” and “legendary,” as well as his child-like obsession with laser tag. Another catchphrase and pastime activity of Barney’s also provides a gateway into the time’s cultural milieu. When Barney exclaims “This is going in my blog!,” one is reminded of a time where Internet blogs were a prominent form of communication – the number of blogs had risen from 50 in 1999 to around seventy million in 2007 (Mohammed). Ted is portrayed as desperate to find “the one,” the woman with whom he will fall in love and start a lifetime of happiness together. As the whole story is structured around the quest to find the person he married and had children with, the identity of Ted’s “the one” is built up from the get-go. This ties into the words used while introducing Robin (Cobie Smulders), such as “and there she was” and “this was no ordinary girl.” Occasionally, the excitement born out of this quest to find love forces Ted to say things that seem out of place, an example being his overly premature “I think I’m in love with you,” to Robin right after their first date, something which shocks her on the spot, but also Ted’s friends when they hear about what happened later that night, as well as his kids, when the same story is told by Future Ted. After the initial scene of the episode, this is the first moment where the audience is reminded of – and equated with – Ted’s kids. While Future Ted may be an unreliable narrator, Favard argues that “It is this unreliability, the erratic nature of Future Ted’s storytelling, that … enables both the kids and the viewer to catch proleptic glimpses of events to come” (Terry 7). Just like how the character learns from his own mistakes, Ted’s kids learn from the gaps or flaws in the narration, seeing through the framework of Ted’s story in the final episode, which Adeline Terry claims is the true cause of the narrator’s unreliability – this is not the story of how Ted met the kids’ mother, but of how he is “totally in love with Aunt Robin” (7). As spouted in nearly every episode, mostly by Barney, the gang believes there are certain theories that apply to interpersonal relationships, and for “The Pilot,” it is “The Olive Theory.” We are told that Lily loves olives and Marshall hates them, and this balances their relationship, making them a perfect couple. The fact that this theory is debunked in the very same episode perhaps is a message to the audience that suggests theories may not always work and that one does not require theories to find their “the one.” Much like Ted, who learns that he has to be patient to find his significant other, audiences have to wait until the last minute of the episode to learn Robin’s name, which happens when 2030 Ted claims “[…] and that’s how I met your Aunt Robin,” suggesting Robin’s not the mother. As her name comes at the end as a twist, one is also reminded of the several moments Robin does say her name but the camera cuts before the audience can hear. The most notable example is at the bar, when the gang is watching Robin report a news A Comparative Textual Analysis of How I Met Your Mother and Its Reboots  5 segment. With the way Robin is introduced, both the audience and Ted’s kids are content with having identified the mother. However, the sense of ease is quickly taken away by 2030 Ted, as he reminds everyone that this would be a long story. This choice also affects HIMYD and HIMYF, as both series attempt to slightly tweak the approach to when and how the other parent would be revealed. HIMYD: “Dear cricket, a wise person once told me[…]” HIMYD, which never came to life after its pilot episode, became accessible to the general public after a user named John Gillman posted the pilot episode on Vimeo (Tenreyro). While the series does follow the “parent tells kids the story of how they met their other parent” trope of HIMYM, there are creative choices where HIMYD takes a different direction than its predecessor, which, this article claims, has contributed to the fact that it was never picked up. The very fact that the series uses new terminology, “dad” and “mom,” instead of “father” and “mother,” suggests a shift in approach. While there is not much indication of a more serious and strict relationship between Ted and his kids in HIMYM, the word “father” is conventionally associated with a sense of authority and discipline, as opposed to “dad,” which suggests a much more mild-mannered and easy-going parent (Mitchell and Sugar). The child in HIMYD is also referred to by a nickname, “cricket,” rather than “kids,” which does not necessarily imply a tighter relationship in and of itself, but does show that the creators have attached a certain importance to names. While investigating how stories are told in contemporary television, Mittell identifies poetics as a method, particularly attaching importance to the language used in the process of creation (4). Seeing how the children in HIMYD and HIMYM are equated with or used as surrogates for the people watching the shows, one can argue that opting for words that imply further warmth and closeness is also a new way for creators to approach audiences. This time the story is told in 2044 from the perspective of a mom, Sally (Greta Gerwig), as the flashbacks take the audience to 2014. The voiceover narrates a letter written in a typewriter, which is referred to as an “ancient contraption” in order to establish that sense of future, while neither the 2044 mom – aside from a shot from her back – nor the child is seen on screen. This time the voiceover is done by Meg Ryan, an actor renowned for her roles in 90s’ romantic comedies who would be a figure beloved by audience members that are around the same age as the main character in 2014. The title is written with the same font as HIMYM, though there are no title sequences to compare, presumably because the creators were waiting to hear about the pilot’s approval before editing a title sequence. However, HIMYD does use the same freeze effects as HIMYM for the parts where the voiceover provides exposition. A notable difference with HIMYM comes when the voiceover gives a disclaimer for what she is about the tell, hence what the audience is about to see: “You’re going to learn a lot of weird stuff about your mom.” The fact that HIMYM does not have this arguably tells more about the creators’ idea of audience expecta- tions, rather than those of the characters’ respective kids. It may be argued that the creators were worried about the possibility of that supposedly “wild life” being scrutinised by the audience, simply because it is a woman rather than a man doing all these, and the protagonist losing her lovable qualities. Future Sally does not suffer from unreliability as Future Ted does, instead claiming that she will opt for honesty even if it does her disservice, which also is another indication of the will to establish a stronger relationship between the parent and the child, as well as the show and the audiences. In terms of representation and stylistic choices, there are some elements within HIMYD that tie it to other popular series of its time. One example is Sally’s brother, Andrew (Danny Javits), who falls within the stuffy and uptight gay stereotype (Scott, “Contested Kicks” 157). While both the cast and characters are much more diverse than the all-white and all-middle class HIMYM, there are also tropes these characters fall into, as illustrated by Andrew’s manners and physical qualities. The “running the tape back effect,” which sees the actual footage rolled back with a sound effect in order to compare what Andrew says with what Andrew actually wants to say is something previously unseen in the “how I met your” series, even 6  Can Kocak though it is a pretty standard convention used in many popular films and TV series. With this self-reflexive motion, sound is concretised and audiences are confronted with the fact that they are experiencing separate pieces of media – visual and aural – being brought together to create meaning (Kim-Cohen 224). The definitive features of characters appearing on screen in writing – as seen with words such as “Ken doll hair” and “obsessive need to constantly tell me what I was doing wrong 24 hours a day” that describe the physical and mental traits of Gavin (Anders Holm), Sally’s soon-to-be-ex fiancé – also evokes the visual techniques employed by Sherlock, another popular series of the time, which rely on the titular character’s deductions based on people’s appearances and behaviours. This narration technique, which is used in Sherlock in order to allow audiences’ access into the character’s mind (Burt), serves a similar purpose here, this time used for comedic purposes. The use of “me” as an object pronoun reminds audiences that this is still Sally’s story told with first-person narration, even when flashbacks are involved. Throughout the narrative, there are also a number of themes that resemble HIMYM. The most notable example would be the inciting incident that makes Sally concerned about her own life. While Ted’s desire to meet and talk to Robin is fuelled by Marshall’s proposal to Lily, Sally starts wondering about her own life after Andrew and his partner decide to adopt a child. Additionally, much like Marshall and Lily’s engage- ment celebration, popping a bottle of champagne is seen as a metaphorical initiation into adulthood. Finally, the series also introduces its own theory of “the one,” calling it “being the same tree.” This is elaborated as a metaphor for two people with differing interests – two separate trees – happily sharing a life – growing together as if they are the same tree. This narrative seems to attach more importance to this theory than HIMYM does to “The Olive Theory,” as “We are not the same tree,” becomes Sally’s reason to break up with Gavin. A big reveal at the end of the episode points out in a different direction than HIMYM, as Frank (Nicholas D’Agosto), who is randomly introduced to Sally by her best friend, turns out to be cricket’s dad. What creates that suspense and leaves audiences wondering as to how they ended up having a kid in the end is the episode’s version of 2014 ending when Sally and Frank decide to be only friends. Since the show got cancelled, the answer to this question is never revealed. HIMYF S01E01: “Hi sweetie, you look tired.” As soon as HIMYF’s pilot episode begins, one comes across a number of examples that tie the series’ conformity and allegiance to HIMYM. The first example is the decision to refer to the parent in question as “father,” rather than “dad.” The second is the theme song, which is essentially a cover version of the music that audiences have associated with HIMYM’s title sequence. So, from the first few moments of the series, ties are established with an ancestor, and the series openly acknowledges its status as a reboot. The mother appears in 2050 this time and tells the story to her son. However, as opposed to HIMYM’s voiceover that only shows the kids or HIMYD’s voiceover that shows a person typing from the back, the audience only sees the mother and not the kid, suggesting a shift of perspective. This further reinforces an identification between the kid listening to the story and the audiences watching the show; in addition to the amount of knowledge they possess, they are also visually equalised – both are not present on the screen and the narrator is looking at both. This can also be interpreted as a subtle message to the audience in terms of the narrator’s reliability – this time they are right there in front of the screen and hence can be trusted. The fact that the 2050 mother is portrayed by Kim Cattrall, arguably one of the two most famous actors of the show⁶ with Hilary Duff, adds another layer to the fact that she is not presented merely through a voiceover, while the way she is presented resembles a theatre actor making an appearance on stage.  6 Sex and the City, the late 90s and early 2000s phenomenon with which Cattrall rose to fame, was also revived in 2021 with the name And Just Like That…, but Cattrall herself was not among the cast members that re-joined the series. A Comparative Textual Analysis of How I Met Your Mother and Its Reboots  7 While the story is told from 2050, the story told takes place in 2022, and there are more examples that provide one with an idea of the cultural milieu than both HIMYM and HIMYD. The 2050 mother complains about the voice command system which “still does not work in 2050,” which is an obvious nod to the paradigms of the day and age audiences of 2022 are living in. Other examples include Tinder dates, using Uber to get around town, a person going viral and achieving a semi-famous state with the video of a failed marriage proposal, which all indicate specificity for 2020s. The fact that the series features quite a diverse cast – also seen in HIMYD – indicates that in addition to content, the cultural paradigms and conjuncture (Grossberg 4) of the times affect choices in production as well. While there are also some elements that separate HIMYF from its predecessors, the show mostly adheres to the rules established by HIMYM. An example of the former might be the references to pre- 2022 that are used for comedic effect throughout the episodes. While these are also seen in HIMYM and HIMYD, HIMYF lacks the distinct sound effect that the other two series use for flashbacks. Tying into Kim- Cohen’s claim that it is used to force the audience into confrontation (224), the fact that there is no effect here is the creators’ way of implying that there is also no room for confrontation in this show – it is solely for comfort viewing. Similarities to HIMYM, on the other hand, are much more common and easily identifiable. For instance, there is a lingering first kiss between Sophie (Hilary Duff) and Ian (Daniel Augustin), which recall the interaction between Ted and Robin after their first date. The “best friend gets their life together while protagonist wonders about their life” moment comes when Valentina (Francia Raisa) starts a long-term relationship with Charlie (Tom Ainsley), who is essentially a larger-than-life representation of a posh Londoner (“Smart People”; “Quintessential British Gentleman”). Finally, the soulmate theory of this series is established as “The Brooklyn Bridge Person” – someone worthy to walk across the Brooklyn Bridge with. The pilot’s biggest nod to HIMYM comes when one of the characters leases a redecorated version of Marshall and Lily’s – and formerly Ted’s – original flat. There is an additional hook for the audience when Wesleyan University, the school Ted, Marshall, and Lily went to, is namedropped to explain how HIMYF’s characters were made aware of the house’s availability. It is also noteworthy that in this scene, the focal point of the characters and the most important object in the eyes of the audience is a prop from a TV series that ended around 10 years ago, a set of swords used by Ted and Marshall in HIMYM. HIMYF’s pilot ends with a reveal that one of the four male characters introduced in the story – Ian, Jesse (Chris Lowell), Sid (Suraj Sharma), or Charlie – is the father, which suggests a different approach than both HIMYM, which does not reveal the mother until its final season, and HIMYD, which discloses the identity of the father right away. However, all three sitcoms rely on a similar type of relationship between comedy and disclosure – suspension based on knowledge and insight (Goodine 5). Goodine suggests that “laughter is frequently dependent upon fore-knowledge. … pleasure is not derived from surprise, but from the fulfillment of our expectations” (5). The audiences of these three shows have varying degrees of knowledge but what does not change is the fact that only the respective narrators know the full story. Additionally, while HIMYF’s choice to insinuate who the father might be demonstrates an attempt to balance new content with nostalgia, possibly as a way of attracting both old and new audiences (Loock, “American TV Series Revivals” 306), when examples of both are compared in terms of quantity and gravity, fan service outweighs additions. HIMYF S01E10: “You have got to stop with this random rich couple” Episode 10, which marks the final episode of the first season, features a few cameos from HIMYM characters. Conventionally, cameos use “celebrity images to reward fans eager to demonstrate their knowledge of favourite celebrities, allowing them to actively assume a role for themselves within mass culture” (Andersen 1). However, when reboots are concerned, the characters of the initial series are approached as celebrities in their own rights, which is based on an assumption that whoever’s watching the reboot is also familiar with the narrative of the former series and would take joy in seeing these characters make an appearance. Throughout nine seasons, there have been a number of celebrity cameos on HIMYM, including 8  Can Kocak Jennifer Lopez and Britney Spears. Then again, the cameos on HIMYF’s episode 10 are from four characters of HIMYM, the first one being George Van Smoot, or “The Captain” (Kyle MacLachlan) and Becky (Laura Bell Bundy), Robin’s former co-worker.⁷ Briefly appearing in the opening and closing sequences of the episode, The Captain and Becky eventually serve as the instigator of an inciting incident that see two lovers reunite. Since the episode abruptly begins with a mediation session between The Captain and Becky without any context or explanation – which is something also acknowledged by the 2050 mother following an objection from her kid – the audience is trusted to remember them and their subtle nods to certain incidents that took place within HIMYM. Here, one can observe the difference between Future Sophie’s kid and the show’s audience, as this is not a “random rich couple” for the latter. This also breaks down the conventional hierarchy between narrative events, as a satellite – a minor plot event – is treated as a kernel – a major plot event – (Chatman 53, 54) due to the fact that it recalls the events of another narrative. For instance, The Captain references “The Pineapple Incident” by reminiscing how he never found out who stole his pine- apple. This is something that took place in episode 10 of season 1 of HIMYM, where the audience watched Ted wake up with a pineapple on his side and tried to trace where it came from. The incident’s ties to The Captain are never confirmed in HIMYM, aside from a deleted scene from the ninth season. So, in a way, this reference in HIMYF also serves to confirm and “legitimise” the pineapple’s ties to The Captain, integrating this fact into the narrative universe. Another example of the duo’s references is Becky dramatically exclaiming she is after The Captain’s boats, which requires the audience the recall “The Boats Boats Boats Girl,” Becky’s nickname in HIMYM, something she earns after a commercial that sees her rise to fame. “Timing is Everything” also features a number of narrative beats that remind the audience of HIMYM’s plots. Jesse’s premature “I love you,” to Sophie is something that makes her worry about the seriousness and future of their relationship, while other couples argue over a range of topics that include costs of marriage, long-distance relationships, and wanting or not wanting children. The popular culture references that serve as a hook to the audience in 2022 continue with “Tiger King,” the figure who reached worldwide fame with the Netflix documentary series of the same name, being namedropped as a joke. The second cameo of the episode comes as Sophie goes to McLaren’s, the bar frequented by the HIMYM gang, and sees Carl the Bartender (Joe Nieves), who also is a recurring character in HIMYM. Since it is right under the flat HIMYF’s characters move into on episode 1, the bar’s rather late appearance in the series is addressed in another meta-joke by Sophie, when she says “Weird we never come in here,” to herself upon entering. In fact, the narrative audience (Phelan 135), Sophie’s kid, would not understand this joke, as it is rather targeted for the implied audience (Chatman 102), those who are fans of HIMYM and can recognise McLaren’s. On the other hand, this scene is mostly set up to prepare the audience for what is to come in a minute, which is an appearance by Robin. The way Robin is presented in this scene is as if her very presence is a dramatic beat – she makes a passing comment at Sophie, the camera cuts to her, the music rises, and the screen cuts to black. Here, the tension possibly comes from the acknowledgement that cameos can have a twofold effect: “If audiences greet the return of beloved characters with enthusiasm and excitement, there is usually also a sense of unease that cherished memories of the past might be overwritten by the new media texts” (Loock, “Amer- ican TV Series Revivals” 305). This is initially designed as a cameo by Robin, the beloved HIMYM character, and at first, she is seemingly there as a celebrity only known to the audience from another series they have watched. However, HIMYF adds another layer to this cameo as Sophie also recognises Robin as Robin Scherbatsky, the news anchor, further underlining how the two narratives are connected. In fact, up until the final episode of the first season, the narrative world that ties HIMYM and HIMYF is established in a way that resembles cinematic universes, attempting to provide a “depth of experience that motivates more consumption” (Jenkins 95, 96). For instance, in episode 9, one of HIMYF’s characters goes to a job interview at Goliath Corporation, which presumably is the company that owns Goliath National Bank, where Barney,  7 Both characters also make a brief cameo on episode 9 of the series, “Jay Street,” but their appearance is not contextualised until episode 10. A Comparative Textual Analysis of How I Met Your Mother and Its Reboots  9 Marshall, and Ted all worked at one point in their lives. It is safe to assume that these nods to HIMYM’s narrative universe will continue in HIMYF’s following seasons.⁸ From the moment she is seen in HIMYF, Robin is established as a character who has her life in order. She has a successful career, seems to be in a good place overall, and provides some advice to Sophie that turn out to be the defining theme of the whole episode. She starts off by telling Sophie that she would very much like to hear about her love life, using the line “Back in the day, my friends and I wasted years in this very bar.” While the line itself refers to HIMYM in its entirety, the camera also cuts to the booth Robin and her friends used to sit in, invoking a sense of belonging and nostalgia within the audience. As Sophie explains the premature “I love you,” she received, Robin takes Sophie through her own experience with Ted, using the line “I had a guy say ‘I love you’ on our first date.” Sophie believes this guy she is hearing of sounds like “a real piece of work,” but Robin clarifies that he is “a good piece of work” and laughs, reminding to the audience of the “will they won’t they” dynamic between Ted and Robin that goes on for multiple seasons, and hinting to the possibility that they are still together. She does not provide details of their relationship with Ted, because this is something the audience, who by that time have started to reminisce what went on between Robin and Ted in the span of nine seasons, already knows about. On the other hand, as she gives Sophie advice, it is apparent to those who know HIMYM’s storyline that she is also in a way summarising what she experienced herself, the line “Do not make decisions out of fear,” being a direct confirmation of this tie between what she is saying to Sophie and the decisions she made in HIMYM. In the end, another piece of advice by Robin, “Timing is everything,” turns into the main idea – and title – of the episode – as well as the whole season – as Sophie fails to make things work with Jessie but sees Ian return from his research trip in Australia. One could argue that storytelling is a way to reboot time itself because fiction offers endless possibi- lities. For instance, one could write a story in 2022 and include a character from 2050, who tells a story that takes place again in 2022. In fact, HIMYM, HIMYD, and HIMYF all use this feature of storytelling, both for comedic – the references to the cultural milieu – and dramatic – building up suspense before revealing who the parent is or how a certain character ended up as the parent – effect. The final episode of HIMYF’s first season also sees the idea of “timing is everything” permeating into the choice of editing, since it ties The Captain and Becky’s seemingly out-of-context marital dispute to Ian’s return. While the father is still not confirmed at the time of writing this article, it can be presumed that season 2 will see Sophie start a new relationship with Ian. When the time comes, every element in the story serves its purpose. Conclusion This article began with an inquiry on the very existence of reboots and intends to end on a question that is equally broad and difficult to answer: How do networks⁹ decide which TV shows to pick up? There are a number of factors at play here and most would go beyond the scope of this discussion. Perhaps, timing really is everything and that is what led to networks favouring HIMYF over HIMYD – 2014 was simply too soon for a HIMYM reboot. Perhaps, HIMYD would have been embraced by HIMYM’s audience had it been given a chance beyond its pilot episode. Then again, if one is to compare the choices HIMYM, HIMYD, and HIMYF make in narrative and narration, they can arrive at the conclusion that the two reboots’ respective networks have preferred more direct connections with and references to the original content, which serve as fan service and are used to get audiences further engaged with the new series. With audience studies acknowledging the importance of interpretation and engagement, fans have gained more power than they ever had in the past. Ensuring their satisfaction became so important for creators that acts of fan service, featuring direct connections with something else the fans are familiar with,  8 The season two premiere, which aired on 24 January 2023, also featured a cameo by Barney. 9 The word “network” is used interchangeably as the channels or streaming platforms that air these series. 10  Can Kocak became a cultural phenomenon, dominating the creative scene in both film and television. After carrying out a textual analysis of the pilot episode of three series, as well as the final episode of the first season of HIMYF, it can be seen that HIMYF is more invested in fan service than HIMYD and establishes more direct ties with HIMYM, much to the satisfaction of its network and fans. It should be noted that many elements within both HIMYD and HIMYF invoke HIMYM in one way or another. A number of examples can be given here, from the fact that they use the phrase “how I met your,” to their premise that features a parent telling their kids the story of how they met their significant other, or from how certain effects are used for transitions to how the cultural milieu can be deduced from the content, and to how an idea of “finding the one” is promoted by resorting to different metaphors. HIMYF takes these similarities a step further, placing its characters inside the same spaces once inhabited by HIMYM’s gang – the flat, Wesleyan University, the bar –, referring to objects that are central to HIMYM’s plots – a set of swords and a pineapple – and featuring cameos by HIMYM’s characters. Even though the audience has to wait for ten episodes for some of these to happen, which could have also been the case for HIMYD had it been picked up, the clear difference in approach to the original content and fans between HIMYD and HIMYF can also be seen in their respective pilot episodes. Not long after the first episode aired, HIMYF was renewed for a second season, which will be twice as long as the first, comprising 20 episodes (Porter, “‘How I Met Your Father’ Earns”). This suggests that both the network and fans are satisfied with it, which is also confirmed by Jordan Helman, head of scripted content for Hulu Originals, who has suggested that the show “has proven to be true appointment viewing that fans cannot get enough of week to week” (Porter). Critics with self-established ties to HIMYM have praised both the show’s pilot episode – “I realized that perhaps for the first time I am specifically the target audience of something” (Radulovic) – and the season finale – “Writers … took the opportunity to deliver a heavy dose of nostalgia by including a perfectly-timed mix of HIMYM cameos and callbacks” (Gallucci). As one of the managers in an SVOD platform, the fact that Helman refers to “appointment viewing,” which is typically associated with conventional TV and one-time TV programs such as sporting events (Conlin et al. 151), is quite noteworthy. Allocations – and manipulations – of time, discussed in the context of this article as a storytelling strategy employed by the creators and characters, are once again at play here, this time in terms of production and scheduling. While one would need a separate article to breakdown each ramification of this approach, it may be argued that streaming services are after varying levels of audience engagement: one that favours the amount of time spent on the platform, as seen in practices of binge-watching for all-in-one releases, and another that favours the quality of time spent on the platform, which shows a recurring commitment, as seen in the re-establishment of audience ties with the – original and rebooted – content for weekly releases. Abbreviations HIMYD How I Met Your Dad HIMYF How I Met Your Father HIMYM How I Met Your Mother Funding information: The author states no funding involved. Conflict of interest: The author states no conflict of interest. Ethical statement: The conducted research is not related to either human or animal use. Data availability statement: All data generated or analysed during this study are included in this published article. A Comparative Textual Analysis of How I Met Your Mother and Its Reboots  11 Works Cited Andersen, Joceline. “Stars, Nobodies, and Other Apparitions: Cameo Roles in Hollywood Film and Television,” PhD thesis, McGill University, Montreal, 2016. Beaty, Bart. “Superhero Fan Service: Audience Strategies in the Contemporary Interlinked Hollywood Blockbuster.” The Information Society, vol. 32, no. 5, 2016, pp. 318–25. Burt, Kayti. “How Paul McGuigan Crafted the Visual Language of Sherlock,” Den of Geek. August 2020, p. 5. https://www. denofgeek.com/tv/sherlock-visual-language-director-paul-mcguigan/. Can, Orcun. “The Bingeable Ms Gilmore: A Comparative Analysis of Narrative Structure in Broadcast TV Show ‘Gilmore Girls’ and Netflix Original Show ‘Gilmore Girls: A Year in the Life’.” Binge-Watching and Contemporary Television Research, edited by Mareike Jenner, Edinburgh University Press, 2021. Chatman, Seymour. Story and Discourse: Narrative Structure in Fiction and Film. Cornell University Press, 1978. Conlin, Lindsey, Andrew Billings, and Lauren Auverset. “Time-Shifting vs Appointment Viewing: The Role of Fear of Missing out within TV Consumption Behaviors.” Communication & Society, vol. 29, no. 4, 2016, pp. 151–64. Creeber, Glen. The Television Genre Book. Bloomsbury, 2006. Dockterman, Eliana. “The Snyder Cut Is a Better Version of Justice League. But It Sets a Dangerous Precedent.” Time, March 2021, p. 15. https://time.com/5946917/snyder-cut-justice-league-fans/. Gallucci, Nicole. “Nostalgic ‘How I Met Your Father’ finale hits all the right notes,” Mashable, March 2022, 15. https:// mashable.com/article/how-i-met-your-father-season-1-finale. Gatiss, Mark and Steven Moffat, creators. Sherlock. BBC and Hartswood Films, 2010-2017. Goodine, Jean. The Role of Suspense in Dramatic Comedy. PhD dissertation. University of New Hampshire, 1971, ProQuest. https://www.proquest.com/docview/302856920. Grossberg, Lawrence. Cultural Studies in the Future Tense. Duke University Press, 2010. Hall, Stuart. “Encoding and Decoding in the Television Discourse,” University of Birmingham, 1973. Jenner, Mareike. “Is this TVIV? On Netflix, TVIII and binge-watching.” New Media & Society, vol. 18, no. 2, 2016, pp. 257–73. Kelly, Stephen. “From Sonic the Hedgehog to Star Wars, are fans too entitled?” BBC Culture, February 2020, p. 14, https://www. bbc.com/culture/article/20200213-from-sonic-the-hedgehog-to-star-wars-are-fans-too-entitled. Kim-Cohen, Seth. In the Blink of an Ear: Toward a Non-Cochlear Sonic Art. Bloomsbury, 2009. Lizardi, Ryan. “The Nostalgic Revolution Will Be Televised.” Remake Television: Reboot, Re-Use, Recycle, edited by Carien Lavigne, Lexington Books, 2014, pp. 37–52. Loock, Kathleen. “American TV Series Revivals: Introduction,” Television & New Media, vol. 19, no. 4, 2018, pp. 299–309. Loock, Kathleen. “‘Whatever Happened to Predictability?’: Fuller House, (Post)Feminism, and the Revival of Family-Friendly Viewing,” Television & New Media, vol. 19, no. 4, 2018, pp. 361–378. McQuail, Denis. Audience Analysis. SAGE, 1997. Mitchell, Kathy and Marcy Sugar. “There’s a Big Difference Between What Makes a ‘Father’ and a ‘Dad’,” Daily Republic, June 2012, p. 17. https://www.dailyrepublic.com/all-dr-news/wires/features/theres-a-big-difference-between-what-makes-a- father-and-a-dad/. Mohammed, Farah. “The Rise and Fall of the Blog,” Jstor Daily, December 2017, p. 27. https://daily.jstor.org/the-rise-and-fall- of-the-blog/. Phelan, James. Narrative as Rhetoric: Technique, Audiences, Ethics, Ideology. Ohio State University Press, 1996. “Pilot.” How I Met Your Father, season 1, episode 1, 18 January 2022, Hulu, https://www.hulu.com/watch/23368e2c-b65a- 4706-9e76-cee4aa5db060. “Pilot.” How I Met Your Mother, season 1, episode 1, 19 September 2005, Disney +, https://www.disneyplus.com/series/how-i- met-your-mother/3kpBeRQiKjkq. Porter, Kathryn. “We May Not Want More Reboots, but New Streaming Services Depend on Them,” Paste. https://www. pastemagazine.com/tv/why-are-there-so-many-reboots-revivals-spinoffs/. Porter, Rick. “‘How I Met Your Father’ Earns Season 2 Renewal at Hulu,” Hollywood Reporter, February 2022, p. 15. https://www. hollywoodreporter.com/tv/tv-news/hulu-renews-how-i-met-your-father-season-2-1235093729/. Power, Tom. “Inside the Frantic Race to Fix Sonic,” Wired, February 2020, p. 15. https://www.wired.co.uk/article/sonic-movie- redesign. “Quintessential British Gentleman”, TVTropes, https://tvtropes.org/pmwiki/pmwiki.php/Main/ QuintessentialBritishGentleman. Radulovic, Petrana. “Unfortunately, I Loved the How I Met Your Father Premiere,” Polygon, January 2022, p. 18. https://www. polygon.com/22889690/how-i-met-your-father-review. Rivero, Kamila. “Why Do TV Networks Keep Rebooting Old Shows?” Showbiz CheatSheet, March 2019, 8. https://www. cheatsheet.com/entertainment/why-do-tv-networks-keep-rebooting-old-shows.html/. Rothman, Michael. “What’s Driving the Resurgence of Reboots, Remakes and Revivals in TV and Film,” ABC News, May 2017, 31. https://abcnews.go.com/Entertainment/driving-resurgence-reboots-remakes-revivals-tv-film/story?id=47645549. 12  Can Kocak Scott, D. Travers. “Contested Kicks: Sneakers and Gay Masculinity, 1964–2008,” Communication and Critical/Cultural Studies, vol. 8, no. 2, 2011, pp. 146–64. Scott, Daniela. “Why Are There so Many Film and TV Remakes Right Now?,” Cosmopolitan, November 2021, 4. https://www. cosmopolitan.com/uk/entertainment/a37787577/film-tv-reboots-why-so-many/. “Smart People Speak the Queen’s English”, TVTropes, https://tvtropes.org/pmwiki/pmwiki.php/Main/ SmartPeopleSpeakTheQueensEnglish. Tenreyro, Tatiana. “After Watching the Failed How I Met Your Dad Pilot, We Get Why it Didn’t Work.” The A.V. Club, July 2021, p. 22. https://www.avclub.com/after-watching-the-failed-how-i-met-your-dad-pilot-we-1847345513. Terry, Adeline. “‘Kids, You May be Wondering How Many of these Stories I’m Telling You Are True. It’s a Fair Question.’ – The Birth of Humour Through the Rebirth of the Narrative in How I Met Your Mother (CBS, 2005-2014)”, Études de stylistique anglaise [Online] vol. 10, February 2022, pp. 1–22. https://journals.openedition.org/esa/4651. The Entertainment Strategy Guy. “Hit or Not: How I Met Your Father, The Royal Treatment and Eternals Edition,” Entertainment Strategy Guy, February 2022, p. 21. https://entertainmentstrategyguy.com/2022/02/21/hit-or-not-how-i-met-your- father-the-royal-treatment-and-eternals-edition/. Thomas, Rob. Veronica Mars. Stu Segall Productions, Silver Pictures Television, Rob Thomas Productions, Spondoolie Productions and Warner Bros. Television. 2004-2019. “Timing is Everything.” How I Met Your Father, season 1, episode 10, Hulu, https://www.hulu.com/watch/5c6a2222-0981-4e8f- 85bc-94d3e1eaaaf1.
10.3390_ijerph20031722
Article Feasibility and Acceptability of a Remote Stepped Care Mental Health Programme for Adolescents during the COVID-19 Pandemic in India Kanika Malik 1,2 Preeti Chauhan 2, Pooja Nair 2, Vikram Patel 3,4 and Daniel Michelson 5,6,* , Tejaswi Shetty 2, Sonal Mathur 2, James E. Jose 2, Rhea Mathews 2, Manogya Sahay 2, 1 2 Jindal School of Psychology and Counselling, O.P. Jindal Global University, Sonipat 13100, India PRIDE Project, Sangath, New Delhi 110030, India 3 Department of Global Health and Social Medicine, Harvard Medical School, Boston, MA 02115, USA 4 Department of Global Health and Population, Harvard TH Chan School of Public Health, Boston, MA 02115, USA School of Psychology, University of Sussex, Brighton BN1 9RH, UK 5 6 Department of Child and Adolescent Psychiatry, Institute of Psychiatry, Psychology and Neuroscience, King’s College London, London SE5 8AF, UK * Correspondence: daniel.michelson@kcl.ac.uk Abstract: Remote mental health services were rapidly deployed during the COVID-19 pandemic, yet there is relatively little contemporaneous evidence on their feasibility and acceptability. This study as- sessed the feasibility and acceptability of a stepped care mental health programme delivered remotely by lay counsellors to adolescents in New Delhi, India, during a period of ‘lockdown’. The programme consisted of a brief problem-solving intervention (“Step 1”) followed by a tailored behavioural module (“Step 2”) for non-responders. We enrolled 34 participants (M age = 16.4 years) with a self-identified need for psychological support. Feasibility and acceptability were assessed through quantitative process indicators and qualitative interviews (n = 17 adolescents; n = 5 counsellors). Thirty-one (91%) adolescents started Step 1 and 16 (52%) completed the planned Step 1 protocol. Twelve (75%) of the Step 1 completers were non-responsive. Eight (67%) non-responsive cases started Step 2, all of whom met response criteria when reassessed at 12 weeks post-enrolment. Adolescents favoured voice-only sessions over video-calls due to privacy concerns and difficulties accessing suitable de- vices. Counsellors noted challenges of completing remote sessions within the allotted time while recognising the importance of supervision for developing competence in new ways of working. Both adolescents and counsellors discussed the importance of working collaboratively and flexibly to fit around individual preferences and circumstances. Disentangling pandemic-specific barriers from more routine challenges to remote delivery should be a focus of future research. Keywords: adolescents; mental health; remote intervention; stepped care; COVID-19; mixed meth- ods; India 1. Introduction The COVID-19 pandemic has led to severe disruptions and adverse physical and men- tal health outcomes for communities across the globe, with young people disproportionately affected by interrupted education, social isolation, and limited employment opportuni- ties [1,2]. These impacts have been especially pronounced in India, where COVID- related school closures persisted for almost two years in many states. Despite a shift towards online learning, a large proportion of students from disadvantaged backgrounds struggled to access this provision routinely [3,4]. Similar accessibility barriers have been reported for remote mental health services in India and other low-resource settings [5–8]. More contextualised evidence is required to address such barriers and shape ongoing initiatives to expand remote mental health care in India and beyond [7,9–12]. To address this need, Citation: Malik, K.; Shetty, T.; Mathur, S.; Jose, J.E.; Mathews, R.; Sahay, M.; Chauhan, P.; Nair, P.; Patel, V.; Michelson, D. Feasibility and Acceptability of a Remote Stepped Care Mental Health Programme for Adolescents during the COVID-19 Pandemic in India. Int. J. Environ. Res. Public Health 2023, 20, 1722. https://doi.org/10.3390/ ijerph20031722 Academic Editor: Paul B. Tchounwou Received: 8 December 2022 Revised: 12 January 2023 Accepted: 13 January 2023 Published: 17 January 2023 Copyright: © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). Int. J. Environ. Res. Public Health 2023, 20, 1722. https://doi.org/10.3390/ijerph20031722 https://www.mdpi.com/journal/ijerph International Journal ofEnvironmental Researchand Public Health Int. J. Environ. Res. Public Health 2023, 20, 1722 2 of 18 the current study piloted a pragmatic approach to reaching adolescents in need of mental health support during the pandemic. This involved adapting an existing school-based stepped care mental health programme, developed as part of the “Premium for Adoles- cents” (PRIDE) research programme [13,14], so that it could be accessed by adolescents outside of school premises during the pandemic related lockdown. The objectives and deliverables of the PRIDE programme (2016–22) were aligned with India’s national adolescent health programme, the Rashtriya Kishor Swasthya Karyakram (RKSK), which was launched in January 2014 with an ambition to reconfigure the prevailing clinic-based health system and instead focus on prevention and early intervention for adolescent health and developmental needs. RKSK specifically highlighted mental health as a public health priority and identified school-based mental health services as a key delivery platform [15]. Previous PRIDE research has shown the effectiveness of an in-person, lay counsellor-led version of a first line transdiagnostic problem-solving intervention (“Step 1”) provided in secondary schools for pupils with elevated mental health symptoms [16]. Non- responders are stepped up to a more tailored intervention (“Step 2”), in which behavioural modules are selected according to the main presenting problem [14,17]. Here, we present findings on the acceptability and feasibility of the PRIDE stepped care programme when delivered by lay counsellors in a remote format. The study was conducted among young people from low-income communities where smartphones and other internet-enabled devices, if available at all, are typically shared between family mem- bers in households with restricted mobile data and limited broadband access. Given these circumstances, the chosen approach emphasised flexibility such that students could choose between voice or video calls, with supplementary materials distributed electronically and by post. The intervention protocol also incorporated collaborative decision-making to deter- mine the need for stepping up. The findings from this study were intended to shape further innovations in remotely delivered interventions, during the pandemic period and beyond. 2. Materials and Methods 2.1. Design We used a mixed method design that assessed quantitative process indicators of feasibility and deployed qualitative semi-structured interviews to explore adolescents’ and counsellors’ experiences of intervention delivery. We also collected data on indicative outcomes using two adolescent-reported measures. The approvals for this study were obtained from local schools, the Institutional Review Boards of Sangath (the implementing organisation), and Harvard Medical School (the sponsor). 2.2. Participants The sampling frame consisted of 3780 students enrolled in grades 9–12 of four sec- ondary schools serving low-income communities in New Delhi, India. Two of these schools were government-aided public schools, and the other two were charity-aided private schools. There were more girls than boys in the sampling frame, reflecting the fact that two of four partner schools were all-girls schools. None of the schools had pre-existing counselling services, either face-to-face or remote. All four schools were closed to in-person teaching and offered an online curriculum for the duration of the study. Students within the sampling frame were eligible for the study if they had a self- identified need for psychological support (i.e., expressing a felt need, without being for- mally assessed for clinical severity), and had access to a voice-only telephone or internet- enabled device (smartphone, computer, or tablet). Participants were also required to be proficient in written and spoken Hindi or English in order to comprehend the intervention materials. Additionally, counsellors were invited to participate in a separate focus group discussion (FGD) at the end of the study. Int. J. Environ. Res. Public Health 2023, 20, 1722 3 of 18 2.3. Measures 2.3.1. Quantitative Feasibility Indicators We conceptualised feasibility as the extent to which an intervention can be success- fully used or conducted in a given setting [18]. Referral logs and electronic case records were used to obtain data on feasibility indicators related to demand (number of referred students); uptake (number of participants who were eligible for and who started each intervention step); completion (number of participants who received the recommended frequency of sessions for each intervention step); reasons for discontinuation; and dosing (the frequency/length of sessions and overall duration of each step). Additionally, the quality of intervention delivery was assessed for 20% of the audio- recorded sessions, selected randomly. These recordings were rated independently by fellow counsellors and a supervising psychologist during group supervision meetings, using a 20- item scale that rated therapeutic skills according to the level of demonstrated competence (1 = limited, 2 = basic, 3 = good, 4 = advanced) (see Appendix A for a copy of the rating scale). 2.3.2. Qualitative Individual Interviews and Focus Group Discussion Semi-structured individual interviews with intervention participants were used to explore the acceptability of intervention content, materials, and delivery processes. A topic guide was constructed in which acceptability was conceptualised as a multi-faceted construct that “reflects the extent to which people receiving a healthcare intervention consider it to be appropriate, based on anticipated or experienced cognitive and emotional responses to the intervention” [19]. A separate FGD was conducted with counsellors, focusing on barriers and facilitators to implementing the intervention. 2.3.3. Clinical Outcomes We used two adolescent-reported measures that have previously been translated into Hindi and used in the local context [20,21]. The Revised Children’s Anxiety and Depres- sion Scale (RCADS-25) is a widely used measure of internalising problems that includes both anxiety and depression subscales [22,23]. The Youth Top Problems (YTP) [24] is an idiographic measure that asks respondents to identify, rank, and rate the severity of three prioritised psychosocial problems. The RCADS-25 and YTP were administered remotely by research assistants at baseline (T0) and 12 weeks later (T2). Additional YTP ratings were obtained by counsellors for progress monitoring and to identify non-responders at the end of Step 1 who may require another Step 2 intervention (see Procedures for further details). Non-response was defined as achieving less than 50% improvement in the severity of the first-ranked priority problem (compared with baseline), and/or no improvement in any of the other two problem ratings. 2.4. Procedures Study referrals were generated between November 2020 and March 2021 using a variety of recruitment materials (see Table 1). Research assistants followed up with self-referred students through a voice call and assessed them individually for eligibility using a short screening proforma. The pro- forma included questions related to the school and grade in which they were enrolled, self-reported difficulties with reading and writing, access to a phone or digital device, and a single item adapted from the Strengths and Difficulties Questionnaire Impact sup- plement [26]. The latter assessed felt need for psychological support (“Overall, do you think that you have difficulties in any of the following areas for which you need support: emotions, concentration, behaviour or being able to get on with other people?”). Verbal assent was recorded for all eligible adolescents (or consent for adolescents aged 18 years and above), followed by recorded verbal parent/guardian consent for those aged less than 18 years. Int. J. Environ. Res. Public Health 2023, 20, 1722 4 of 18 Table 1. Recruitment activities and materials. Activity/Material Description Student webinar Teacher webinar Digital flyer YouTube video Adapted from an existing in-person classroom ‘sensitisation’ session [25]. Attendees were presented with an animated video and a PowerPoint presentation that contained age-appropriate information about common mental health problems and explained the rationale for the available intervention. This was followed by a counsellor-facilitated group discussion about the practicalities of remote delivery and procedures for enrolling in the study. Separate webinars were offered to each class. Structured around a similar PowerPoint presentation as the student webinar, followed by a group discussion concerned with how teachers could effectively disseminate information about the study and encourage students to participation in the intervention. Contained a brief description of the available intervention and how to enrol in the study. This was circulated by teachers to class WhatsApp groups. Students were able to begin the enrolment process either by calling or messaging (WhatsApp) a designated number without incurring any airtime charges. Contained an abbreviated version of the same information that was presented by counsellors during the webinar, in which a member of the study spoke to camera. It was intended for students who could not attend the live webinar due to technical issues or competing activities. Teachers circulated this to class WhatsApp groups along with the flyer. Following enrolment, participants were sent printed reference copies of the RCADS- 25 and YTP in their preferred language (English or Hindi) by post and electronic copies (non-editable PDF format) through WhatsApp. A research assistant subsequently made telephone contact to assist with administration. After receiving instructions from the research assistant, participants read aloud each item and their corresponding response. Categorical items/scores corresponding to the verbal responses were marked in real-time by the research assistants on the relevant electronic forms using a hand-held tablet device. Following the baseline assessment (T0), participants were offered up to five sessions of a problem-solving intervention (Step 1) delivered through a voice or video call. Sessions were intended to last for up to 30 min over a period of 3–4 weeks. The intervention structure was closely based on the in-person version of Step 1 [13], where session 1 introduced a three- part problem-solving technique based around the acronym POD (“Problem”, “Option”, “Do it”); sessions 2–4 supported the applied use of this approach to a prioritised problem; and an optional session 5 was used to consolidate learning and help with generalising the approach to other problems. Progress was monitored regularly through in-session ratings on the YTP. Step 1 was concluded at session 4 or 5 for those individuals who showed 50% or greater improve- ment on the main target problem, accompanied by a downward trend on at least one of their other YTP problem scores. Participants who did not meet response criteria on the YTP by the fourth session were considered for Step 2 through a shared decision-making protocol that was structured around the SHARE acronym [27]. This entailed “seeking” the participant’s involvement by explaining the meaning and purpose of shared decision- making, followed by “helping” the participant to review their progress and exploring the potential costs/benefits of concluding/continuing with counselling. Next, the participant’s preference and justification were “assessed”, after which the counsellor shared their rec- ommendation. The participant was encouraged to “reach” a decision about whether or not to step up. The participant was asked to “evaluate” their satisfaction (on a three-point scale from very satisfied-somewhat satisfied-not at all satisfied) with the final decision, followed by further discussion as needed. A similar shared decision-making process was used to decide on the need for early discharge during Step 1. If a participant achieved early response by session 2 or 3, they could then collaboratively decide with the counsellor to end the intervention at that point. Those progressing to Step 2 received up to 6 sessions over voice or video call, which were intended to last 30–35 min each and were scheduled over 4–6 weeks. Based on an existing intervention protocol [17], an introductory session on relaxation skills was Int. J. Environ. Res. Public Health 2023, 20, 1722 5 of 18 provided to all Step 2 participants, followed by one of three behavioural modules (2– 4 sessions) selected according to the main presenting problem, i.e., behavioural activation for depression, exposure for anxiety, or assertiveness and communication training for conduct problems. The final session focused on relapse prevention. In both steps, participants additionally received supplementary resource materials in both printed and electronic (editable PDF) formats. These materials included comic books, which used illustrated stories to describe common problems in the target age group, explain the essential concepts of the various behavioural skills, and offer practical suggestions (“quick tips”) for using these skills effectively. The stories were followed by suggested home practice exercises to develop skills further. Participants were also presented with summary posters at the end of each step. Both intervention steps were delivered by the same five lay counsellors, who were employed by Sangath. The counsellors had at least two years of experience in delivering Step 1 in a face-to-face format. They received a one-day “top up” office-based training from psychologists (SM, PN, and RM) with a particular emphasis on competencies needed to deliver Step 1 remotely, and a 6-day training period, focused specifically on shared decision-making, and Step 2 (which the counsellors had not previously delivered in any format). Additionally, the counsellors took part in weekly group supervision for the study duration, where audio-recorded sessions were reviewed and rated for quality by peers and a supervising psychologist. The format of group supervision is elaborated elsewhere [20]. For the current study, the training and supervision were conducted in a hybrid mode (partially online and partially in-person) in line with COVID-19 safety protocols. The YTP was administered by research assistants at enrolment (T0), during Step 1 by counsellors (T1), and again by researchers approximately 12 weeks from study enrolment (or immediately after Step 2 if this extended beyond 12 weeks) (T2). The RCADS was administered by researchers at T0 and T2 only. After the completion of outcome assess- ments, participants were invited to take part in semi-structured individual interviews. Those who agreed (n = 17) were interviewed within two weeks of completing the final outcome assessment. Additionally, a separate FGD (n = 5) was held with counsellors once the intervention delivery phase had ended. The assessments, interviews, and FGD were conducted by research staff who were not involved in intervention development or delivery. 2.5. Analysis Quantitative feasibility indicators were examined descriptively using frequencies, means and SDs. Descriptive analysis of clinical outcomes involved comparisons of mean YTP scores (based on the sum of the three problem severity scores) for participants who completed assessments across all three time points (T0, T1, and T2). The YTP score from the last attended Step 1 session was used for T1. On the RCADS, we noted descriptive trends over time in mean T-scores. The T values are standardised scores, with a mean of 50, and the standard deviation of 10. These were calculated from raw scores using spreadsheets available from the developer (link: https://www.childfirst.ucla.edu/resources/ (accessed on 15 November 2020)). Qualitative data were analysed using a thematic framework approach [28]. Two researchers (KM and TS) independently read all the transcripts in the original Hindi language. A preliminary hierarchical coding framework was then prepared deductively by drawing on the research questions and established conceptual definitions of feasibility and acceptability [18,19]. Initial deductive coding focused on thematic categories of intervention coherence (the extent to which participants understand the rationale for the intervention and what is required of them), affective attitude (how participants feel about an intervention and its elements), effectiveness (perceived changes in valued outcomes that may result from participation), self-efficacy (confidence that participants can do what is required of them), and burden (the perceived amount of effort to take part), and barriers to deliver remotely. Int. J. Environ. Res. Public Health 2023, 20, 1722 6 of 18 The preliminary framework was refined in discussion with the senior author (DM), then applied to a subset of transcripts. This was refined iteratively through data-driven coding, with granular codes grouped under higher-order themes and sub-themes. Coded data were organised into a spreadsheet matrix, with participants organised into rows, and theme and sub-themes organised as columns. Quotes from individual interview and FGD transcripts were placed in corresponding cells. Iterative revisions in the framework were reviewed with the senior author. The final stage involved operationalising themes and sub-themes and comparing them across participants to develop a narrative interpretation. Illustrative quotes were selected to supplement narratives and were translated from Hindi to English for the purpose of reporting. 3. Results 3.1. Feasibility Indicators A total of 755 students from 52 classes were sensitised remotely during the study period, and 56 (7%) were referred into the programme. As shown in Figure 1, among those referred, 47 (84%) were assessed for eligibility and 41 (87%) were eligible. Out of 41 eligible adolescents, 34 (83%) completed the assent/consent and baseline assessment procedures (T0). The demographic and clinical characteristics of these 34 participants were as follows: M age = 16.4 years, SD = 1.6; females n = 25 (74%); M RCADS Total T-score = 62.4, SD = 14.2; and M YTP score = 7.5, SD = 1.8. Figure 1. Participants flow in research and intervention activities. All 34 participants who enrolled in the study opted to have sessions through voice- only calls. From this group, 31 (91%) started Step 1, and 15 (48%) completed the intended number of Step 1 sessions (4 or 5). Additionally, one participant was offered early discharge (at session 3) and counted as completer, based on criteria mentioned above. Among those participants completing Step 1, 12 (75%) participants were non-responsive. Eight (67%) out of these 12 non-responders started Step 2, three non-responders opted out, and one was removed from the study and referred to a mental health professional for management of high suicidal risk. Demographic characteristics of Step 2 participants were as follows: M age = 16.0 years, SD = 1.4; females n = 6 (75%). All but one of the eight Step 2 participants completed the full course of Step 2. The reasons for discontinuing each intervention step are outlined in Figure 1. Int. J. Environ. Res. Public Health 2023, 20, 1722 7 of 19 Figure 1. Participants flow in research and intervention activities. All 34 participants who enrolled in the study opted to have sessions through voice-only calls. From this group, 31 (91%) started Step 1, and 15 (48%) completed the intended number of Step 1 sessions (4 or 5). Additionally, one participant was offered early dis-charge (at session 3) and counted as completer, based on criteria mentioned above. Among those participants completing Step 1, 12 (75%) participants were non-responsive. Eight (67%) out of these 12 non-responders started Step 2, three non-responders opted out, and one was removed from the study and referred to a mental health professional for man-agement of high suicidal risk. Demographic characteristics of Step 2 participants were as follows: M age = 16.0 years, SD = 1.4; females n = 6 (75%). All but one of the eight Step 2 participants completed the full course of Step 2. The reasons for discontinuing each inter-vention step are outlined in Figure 1. Step 1 completers attended, on average, 4.6 sessions (SD = 0.8), spread over 31.1 days (SD = 24.1). Step 2 completers attended an additional 5.1 sessions (SD = 0.7), spread over 54.4 days (SD = 18.5). The mean duration of Step 1 and Step 2 sessions was 45 min (SD= 11.0) and 59 min (SD = 11.20), respectively. Those who completed both intervention steps attended a mean of 9.1 sessions (SD = 0.7), spanning 84 days (SD = 23.0). Quality assess-ments of intervention delivery were in the “good” to “advanced” range for both steps and across both peer and supervisor scores (mean supervisor rating, Step 1 = 3.6, SD = 0.5; mean peer rating, Step 1= 3.8, SD = 0.2; mean supervisor rating, Step 2 = 3.4, SD = 0.3; mean peer rating, Step 2 = 3.8, SD = 0.2). Int. J. Environ. Res. Public Health 2023, 20, 1722 7 of 18 Step 1 completers attended, on average, 4.6 sessions (SD = 0.8), spread over 31.1 days (SD = 24.1). Step 2 completers attended an additional 5.1 sessions (SD = 0.7), spread over 54.4 days (SD = 18.5). The mean duration of Step 1 and Step 2 sessions was 45 min (SD= 11.0) and 59 min (SD = 11.20), respectively. Those who completed both intervention steps attended a mean of 9.1 sessions (SD = 0.7), spanning 84 days (SD = 23.0). Quality assessments of intervention delivery were in the “good” to “advanced” range for both steps and across both peer and supervisor scores (mean supervisor rating, Step 1 = 3.6, SD = 0.5; mean peer rating, Step 1= 3.8, SD = 0.2; mean supervisor rating, Step 2 = 3.4, SD = 0.3; mean peer rating, Step 2 = 3.8, SD = 0.2). 3.2. Qualitative Findings Seventeen adolescents took part in individual interviews (M age = 16.3 years, SD = 1.3; females n = 12 [71%]; received Step 1 only, n = 10 [59%]; received both Step 1 and 2, n = 7 [41%]). Five counsellors (M age = 31.2 years, SD = 4.1; females, n = 3 [60%]) took part in a separate FGD. Five overarching themes were developed: (i) coherence of the intervention steps; (ii) usefulness of counsellors’ guidance and supporting materials; (iii) balancing structure and adolescents’ needs in shared decision-making; (iv) valued outcomes and skill development; and (v) implementation of remote delivery methods. These are elaborated below, illustrated with relevant quotes. 3.2.1. Coherence of the Intervention Steps Problem solving was viewed by all adolescents as a useful skill. Interviewees, includ- ing some participants who dropped out early from Step 1, frequently recalled the “POD” acronym and described its importance in facilitating the process of problem solving, “Ma’am, I really liked ‘POD’. Through this, I learned recognising problems, then finding their options and learning to do it.” (Adolescent P9, 15 years, female, dropped out from Step 1) Adolescents who took part in Step 2 found it beneficial to learn additional comple- mentary behavioural skills to help with their persisting problems. While all acknowledged the benefit of Step 2 skills, there was mixed feedback about the appropriate placement of problem-solving skills in the intervention framework. Some interviewees acknowledged the benefit of learning broad problem-solving skills before the problem-specific Step 2 behavioural skills. “I prefer learning problem solving first. To me, being active [the Step 2 behavioural activation module] looked like one option of broad problem-solving framework. So, we should learn problem solving first and then gradually we can learn different options as part of this skill.” (Adolescent P6, 15 years, female, completed both steps) Others felt that personalised Step 2 skills (i.e., matched to specific problem types) should be taught earlier to accelerate therapeutic gains. “Deep breathing and overcoming fear [the Step 2 exposure module] have become my favourite [compared to POD]. Till the fourth or fifth session, [counsellor] sir taught us about POD and these other skills were taken up in the sessions after that. It would have been even better for our progress to learn things like deep breathing and overcoming fear much earlier.” (Adolescent P3, 16 years, male, completed both steps) 3.2.2. Usefulness of Counsellors’ Guidance and Supporting Handouts Counsellors were praised by adolescent interviewees as supportive guides who offered helpful suggestions for solving problems when participants felt stuck. “Madam [the counsellor] was very understanding and quickly understood everything I shared with her. She helped me a lot in finding alternative solutions to problems. Initially, I felt very nervous, but madam told us how to proceed and gave us step-by-step directions Int. J. Environ. Res. Public Health 2023, 20, 1722 8 of 18 and tips on finding solutions, which helped me greatly.” (Adolescent P7, 16 years, female, completed Step 1 and opted out of Step 2) Adolescents also appreciated counsellors’ willingness to adapt the pace of sessions and work collaboratively, especially when deciding about whether to continue beyond Step 1. “She [counsellor] neither give me any suggestions, nor did she pressurise me to continue or terminate. She had left it for me to decide, i.e., if I wanted to end, I could end it, and if I wanted to continue, then I could continue also; that was my choice. This was a good thing.” (Adolescent P6, 15 years, female, completed both steps) In line with adolescents’ feedback, counsellors also recognised the importance of pacing sessions according to individuals’ needs and preferences. “If we sense a student is reluctant to open up or talk about his difficulties, we try to take it slow. We talk about their interest areas, things they like and ask more specific questions to help them express themselves more concretely. We will provide them with step-by-step guidance if they struggle to open up or express their concerns. Usually, by the end of a couple of sessions, they feel more comfortable with the process. By the time they are in Step 2, they are pretty vocal and engaged in the process.” (Counsellor C2, female) Adolescents reported that the comic books and posters helpfully complemented counsellors’ inputs by keeping them engaged in between sessions, while also serving as a source of ready-made solutions for solving problems. Suggestions were also made about expanding the repertoire of stories to cover a wider array of problems and potential solutions. “The most helpful thing in both these booklets was that after reading their story, it seemed that when they [story characters] can [face their fears], why can’t I. The boy in the story feels scared of dogs and no one else feels that way; similarly, I feel scared of certain things, but others don’t. This is something we can learn and correct. Then, I did the right thing after reading the booklet.” (Adolescent P3, 17 years old, male, completed both steps) “The stories given were very good, as you can understand everything from the story. Then some ready-made solutions have been given, which are also very good. You people [programme developers] can write more stories... like something based on my problem. My main problem is overthinking; you can make a story about it and then you should give me that booklet.” (Adolescent P14, 15 years old, female, completed both steps) 3.2.3. Balancing Structure and Adolescents’ Needs in Shared Decision-Making In line with adolescents’ feedback, counsellors noted the importance of flexible pacing of sessions to accommodate individuals’ learning pace and time constraints to competing demands. At the same time, counsellors spoke positively about the structured intervention manual which detailed the sequence of in-session activities. Supervisory feedback was especially useful for implementing novel tasks from the updated manual. “Initially, it was challenging to do the shared decision-making. We struggled with how to say it, frame it in age-appropriate language, explain why [the adolescent] needs to get involved in it, when we should raise this topic, and what the next steps will be. To overcome these challenges, we used to read [the step-by-step manual], and discuss this thing with other counsellors. Then we also did role-play. Our supervisor also helped. It was a bit of a challenge initially, but all this made it easier as we moved ahead.” (Counsellor C5, female) While most adolescents were satisfied with how the shared decision-making process was structured by the counsellors, a few suggested that more information was needed. “Yes, it would have been good if she [counsellor] had explained what will happen next in the counselling journey. For example, had I known the total number of sessions [in Step Int. J. Environ. Res. Public Health 2023, 20, 1722 9 of 18 2], I would have taken the decision instantly rather than taking too much time to think about it. But I was confused about how long the counselling will go on. Also, she told me the benefits of continuing counselling but not the drawbacks of ending counselling. When you [counsellor] are helping me, explain both benefits and drawbacks so that I can be more confident in making the decision. Ultimately, it will be my decision only, but it will be good if I get more information and help from the counsellor.” (Adolescent P2, 16 years old, female, completed both steps) 3.2.4. Valued Outcomes and Skill Development Adolescent interviewees described improvements across a wide range of symptoms of anxiety, depression, and anger-related problems, and functional domains, such as academic performance, interpersonal difficulties with family and peers, and difficulties related to time management. Many of the adolescent participants described how their problems had benefited from the application of problem-focused coping skills from Step 1 and additional behavioural skills learned from Step 2. “During the sessions, the counsellor made me complete a graph that helped me understand how my problems were going. Then counsellor asked me to work on it. Based on our discussion, I worked on generating and applying alternative solutions, which helped reduce problems.” (Adolescent P4, 18 years old, male, completed Step 1 and opted out of Step 2) In addition to relief in problems for which they sought counselling, participants reported changes in their overall approach to managing problems, including realising the importance of making an active effort to solve problem and the need for continued practice. “Earlier I used to procrastinate but now I’m much faster. The counselling helped me learn that we have to plan and take steps to solve problems, troubles won’t go on their own.” (Adolescent P14, 18 years old, male, completed Step 1 and opted out of Step 2) “I made good improvements in counselling like my problem with overthinking were reduced, my fear subsided, and those negative bad thoughts were also gone down. At times, I still get bad thoughts, but now I can manage them. The most important thing is that I have learned to control, no matter what the situation is... Earlier, if I was afraid to go somewhere, I would not go, but now I try to control, deep breathe, plan to face [my fear] and go wherever I’m needed.” (Adolescent P3, 17 years old, male, completed both steps) Some participants recognised the limits of what could be achieved through the problem- solving approach, particularly with interpersonal issues in the context of pandemic. “We talked about [my problem with] trust issues, and came out with a few options, but I did not feel there was any suitable option, given the lockdown constraints. I concluded that leaving it as it is okay, and we do not need anything . . . Honestly, not much could be done about this but to leave it aside.” (Adolescent P9, 17 years old, female, completed both steps) 3.2.5. Implementation of Remote Delivery Methods Adolescents and counsellors held mixed opinions about attending counselling re- motely. Some adolescents felt more at ease with the format of telephone sessions, as opposed to the prospect of in-person or video meetings. “[Counselling] was very good on the phone because sometimes it happens with me that I cannot tell everything in detail face-to-face, but on the phone, it was easy to tell . . . If you are face-to-face, then I keep thinking about my and the counsellor’s facial expressions, which makes me feel a little hesitant.” (Adolescent P16, 18 years old, female, completed/responded to Step 1) Counsellors reported that remote counselling afforded students a larger time window to schedule sessions, thus providing options for timetabling outside of conventional school Int. J. Environ. Res. Public Health 2023, 20, 1722 10 of 18 hours. Counsellors also appreciated being able to use WhatsApp for sending reminders to students about forthcoming sessions, and for sharing electronic copies of intervention materials where needed. Notwithstanding these benefits of remote communication, there were also frequent reports of accessibility issues, such as students having to share tele- phones with family members, lack of privacy at home when taking part in sessions, and poor connectivity. These issues ruled out video calls for all participants and impacted the overall continuity and duration of sessions. Counsellors often needed to extend sessions to compensate for the extra time incurred in explaining and practising skills over the telephone amidst various disruptions. “Sometimes our sessions extend beyond an hour, but that is primarily due to network and connectivity issues. With multiple call drops, the counsellor had to wait 10–15 min before we could get into stable connectivity.” (Adolescent P2, 16 years old, female, completed both steps) “As everything was new [for students] in remote counselling, we had to set the ground rules and agenda, which took longer than we planned . . . A lot of our time was spent monitoring the progress, especially in the first session; we had to tell students how to do the progress monitoring for each problem. This was very difficult to explain on the phone, the child would be confused and since we could not see them, we just relied on whatever they said . . . Even other activities in booklets used to take more time, as we were remote, we had to explain to them go on this page, do this here and then go on that page, etc. This used to take half of the session, and then we will reach the main session agenda and to complete them well, it was necessary to spend time there too. So, this is why the session used to go on for longer than what was given in the manual.” (Counsellor C3, Male) Consequent to these accessibility issues, most young people favoured a face-to-face format for future rounds of counselling. “[Counselling] was good on the phone, as my problem is almost resolved. However, it would have been much better if she [counsellor] was offline. This entire year, whether it was counselling or the school classes, it was online, and we do not feel so comfortable with it. Talking face-to-face would have been much better than online. The counsellor will also find it much easier to work out solutions for our problems. Even from our body language, she can observe the smallest of things, which would help her understand us much better. [Even on the phone] she did understand everything, but it would have been better if it was offline.” (Adolescent P7, 16 years old, Female, completed/responded to Step 1) Notwithstanding feasibility concerns, counsellors remarked on the high motivation of many participants and the potential to overcome initial trepidation and establish positive therapeutic relationships at a distance. “We got one advantage [due to COVID restrictions] that 2–3 children out of every five children who came to us were either feeling bored at home or were in a situation where they could not talk to others and express themselves. So, when they came to us, we spent time building trust and genuinely showed concern for them during our session; they liked it a lot. They got engaged, thinking this was a platform where they could talk and express their viewpoints. They intrinsically felt motivated about the sessions and then we didn’t have to put in too much effort . . . Once the rapport is well-established, they easily open up and start talking about their problems.” (Counsellor C3, male) 3.3. Indicative Outcomes Complete outcome data on the YTP and RCADS were available for 27 (79%) out of 34 participants who enrolled in the study. As shown in Figure 2, “Category 1” participants who responded by the end of Step 1 experienced further improvements in problem scores up to 12 weeks, with all meeting the response criteria at T2. Among the participants who were non-responsive at T1, those who opted out of receiving any further intervention (“Category Int. J. Environ. Res. Public Health 2023, 20, 1722 11 of 18 2”), and those who continued to Step 2 (“Category 3”) followed similar trajectories of improvement in YTP scores. Except for one participant in each of Categories 2 and 3, all met response criteria at T2. This contrasted with the smaller reduction in YTP scores from T1 to T2 seen among “Category 4” participants who dropped out before completing Step 1. The majority (7 out of 12) of these participants did not meet response criteria at 12-week follow-up. Figure 2. Youth Top Problem (YTP) average scores across three time points. The mean baseline RCADS T-score for the 27 participants with paired pre-post scores was 62.4 (SD = 14.2). Follow-up RCADS scores were similar across each of the four categories: Category 1 (M = 44.8, SD = 9.5), Category 2 (M = 44.4, SD = 4.1), Category 3 (M = 44.6, SD = 9.8), and Category 4 (M = 44.8, SD = 15.8). 4. Discussion This study examined the feasibility and acceptability of a remotely delivered stepped care programme for common mental health problems among young people in India during the COVID-19 pandemic. We observed a limited uptake of the remote programme in the targeted population of secondary school pupils during the period of COVID-19 related school closures, with enrolment mostly restricted to older adolescents (mean age = 16.4 years). The non-completion rate for the first-line intervention was also higher than in earlier studies, conducted before the pandemic, in which we delivered a similar intervention on school premises [16,17]. Given a choice of different access options for the remote counselling, participants uniformly chose voice-only calls over video-calls, due to privacy concerns, and a lack of internet-enabled devices. Poor connectivity during calls also played a part in longer sessions than planned. Our findings align with other evidence on demographic and accessibility related barriers to remote mental health service delivery in low-resource settings, before and during the pandemic [29–34]. We sought to overcome anticipated feasibility issues by keeping technological requirements of the stepped care programme to a minimum (e.g., participants could access counsellors using a basic mobile phone), by offering free airtime, and by scheduling sessions flexibly to fit around online schooling. Counsellors’ transition to working remotely was facilitated by a structured intervention manual and regular Int. J. Environ. Res. Public Health 2023, 20, 1722 12 of 19 Figure 2. Youth Top Problem (YTP) average scores across three time points. The mean baseline RCADS T-score for the 27 participants with paired pre-post scores was 62.4 (SD = 14.2). Follow-up RCADS scores were similar across each of the four cate-gories: Category 1 (M = 44.8, SD = 9.5), Category 2 (M = 44.4, SD = 4.1), Category 3 (M = 44.6, SD = 9.8), and Category 4 (M = 44.8, SD = 15.8). 4. Discussion This study examined the feasibility and acceptability of a remotely delivered stepped care programme for common mental health problems among young people in India dur-ing the COVID-19 pandemic. We observed a limited uptake of the remote programme in the targeted population of secondary school pupils during the period of COVID-19 related school closures, with enrolment mostly restricted to older adolescents (mean age = 16.4 years). The non-completion rate for the first-line intervention was also higher than in ear-lier studies, conducted before the pandemic, in which we delivered a similar intervention on school premises [16,17]. Given a choice of different access options for the remote coun-selling, participants uniformly chose voice-only calls over video-calls, due to privacy con-cerns, and a lack of internet-enabled devices. Poor connectivity during calls also played a part in longer sessions than planned. Our findings align with other evidence on demographic and accessibility related bar-riers to remote mental health service delivery in low-resource settings, before and during the pandemic [29–34]. We sought to overcome anticipated feasibility issues by keeping technological requirements of the stepped care programme to a minimum (e.g., partici-pants could access counsellors using a basic mobile phone), by offering free airtime, and by scheduling sessions flexibly to fit around online schooling. Counsellors’ transition to working remotely was facilitated by a structured intervention manual and regular super-vision, and intervention quality ratings were generally at the higher end of the scale. Alt-hough a number of engagement challenges remained, qualitative interviews revealed that Int. J. Environ. Res. Public Health 2023, 20, 1722 12 of 18 supervision, and intervention quality ratings were generally at the higher end of the scale. Although a number of engagement challenges remained, qualitative interviews revealed that adolescents valued the available support from counsellors, and the behavioural focus was highly relevant to coping with stressors during the pandemic. We note the relatively low response rate (25%) for participants immediately after com- pleting Step 1, compared with response rates of around 50% in previous evaluations of Step 1 when delivered in-person [13,16,17]. Interpretation of outcomes is further complicated by differences in response criteria between studies, and the relatively high proportion of participants (48%) in the current study who had an unplanned ending to Step 1. Unlike the previous round of studies, we did not observe rapid or sustained improvement among participants who stopped attending Step 1 [13,16]. The lack of complete data on reasons for dropout, and potential demand characteristics affecting the data that was available, prevent any firm conclusions about why participants dropped out. However, it seems plausible that in many cases this was due to accessibility or acceptability reasons rather than a reduced need for psychological help. Outcome scores were available for eight participants who entered Step 2, and majority of them met the response criteria at the final follow-up assessment. The very small number of Step 1 non-responders who opted out from Step 2 prevents meaningful comparison of outcome trajectories, but our qualitative findings offer corroborative evidence for the incremental benefits of Step 2. Positive experiences of shared decision-making were also reflected in the qualitative data. This collaborative decision-making process may have contributed to a higher opt-in rate for Step 2 compared with an earlier evaluation of stepped care in which stepping up decisions were based solely on outcome scores [17]. More robust research designs will be needed to confirm this inference. This research sits alongside another study by our group [8], which highlighted the feasibility challenges of implementing a randomised controlled trial of an online problem- solving intervention for young people during the COVID-19 pandemic in India. The external validity of our findings is further strengthened by adopting a pragmatic approach to remote delivery and the triangulation of multiple data sources and perspectives, although we recognise that the small sample size and lack of a control group prevents conclusions about intervention effectiveness. We also recognise that many of the feasibility barriers described in this study will have been exacerbated by the pandemic, which limits wider applicability of findings. For example, access to telephones/computers would have been curtailed by demand from other household members in lockdown, while privacy would have been even more difficult to achieve in crowded homes. We note that the current study was initiated in the first year of the pandemic, prior to the inauguration of the government-funded National Tele-Mental Health Program (NTMHP) in 2022. Future research is needed to disentangle pandemic-specific barriers from more routine accessibility barriers affecting use of remote mental health services in large-scale, nationwide initiatives arising from NTMHP in India [11], and comparable developments in other low-resource settings [12]. Furthermore, studies should investigate infrastructure and resources that may reduce long-standing barriers. Examples could include subsidies and waivers for purchasing digital devices and airtime packages for vulnerable populations. Offering a wider choice of remote intervention formats (e.g., online, tele-counselling, and hybrid/blended approaches) may also help with feasibility and acceptability. In this direction, research should explore preferences and priorities of end users and other relevant stakeholders (parents, service providers, school authori- ties, and technology/communication companies) for remote options across a variety of therapeutic modalities, including those such as family therapy which conventionally use conjoint or group formats. Such evidence is important for ensuring that practice innova- tions arising from the pandemic do not perpetuate inequalities in access for historically disadvantaged groups. Int. J. Environ. Res. Public Health 2023, 20, 1722 13 of 18 5. Conclusions The current evaluation adds to the limited research literature on remote mental health interventions for young people in low- and middle-income countries. A stronger emphasis on evidence-informed, contextually appropriate innovations is needed to narrow existing gaps in accessibility and widen the proportion of young people who stand to gain from the use of communication technologies for mental health service delivery. In the meantime, the growing enthusiasm for a “digital first” approach to service delivery accelerated during the COVID-19 pandemic must be tempered by on-the-ground realities. Author Contributions: V.P. received the funding for this work. K.M., D.M. and V.P. designed the study. T.S., S.M., P.N. and R.M. were responsible for supervision, data collection and data checking. K.M. and J.E.J. were responsible for database design and management. K.M., T.S., R.M., M.S. and P.C. contributed to data coding and analysis. K.M. and D.M. drafted the report with critical inputs and revisions from S.M., T.S. and V.P. All authors have read and agreed to the published version of the manuscript. Funding: This work was supported by the Wellcome Trust, UK (Grant number 106919/Z/15/Z). The funder had no role in study design, data collection and analysis, writing of the report, or in the decision to submit the paper for publication. Institutional Review Board Statement: The study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board (or Ethics Committee) of the sponsoring organisation, Harvard Medical School (protocol code IRB20-0830 and date of approval: 23 October 2020) and implementing organisation, Sangath (protocol code VP_2015_017). Informed Consent Statement: Verbal informed assent/consent was obtained from all participants involved in the study. All consent were recorded and stored. Due to COVID-19 restrictions, there were no written consent. Data Availability Statement: The raw data and coding framework used in this study can be accessed on request from the corresponding author. Acknowledgments: We acknowledge the contributions of adolescents, their parents and guardians, school staff, research staff, and counsellors who made this work possible. We extend sincere thanks to Sai Priya Kumar, Niket Agrawal, Saleha Khatoon, and Anuj Kumar for their support in the preparing transcripts and compiling coded data. Conflicts of Interest: D.M. is a guest editor is a Special Issue of IJERPH. Other authors declare no conflict of interest. Appendix A In Table A1, each competency is rated on a 4-point scale, using the performance band given in Figure A1: Figure A1. Scoring Legend for Performance band *. * Not applicable is marked if the particular skill is not relevant for the particular session. This is to be used only for treatment specific skills. Int. J. Environ. Res. Public Health 2023, 20, 1722 14 of 19 emphasis on evidence-informed, contextually appropriate innovations is needed to nar-row existing gaps in accessibility and widen the proportion of young people who stand to gain from the use of communication technologies for mental health service delivery. In the meantime, the growing enthusiasm for a “digital first” approach to service delivery accelerated during the COVID-19 pandemic must be tempered by on-the-ground realities. Author Contributions: V.P. received the funding for this work. K.M., D.M. and V.P. designed the study. T.S., S.M., P.N. and R.M. were responsible for supervision, data collection and data checking. K.M. and J.E.J. were responsible for database design and management. K.M., T.S., R.M., M.S. and P.C. contributed to data coding and analysis. K.M. and D.M. drafted the report with critical inputs and revisions from S.M., T.S. and V.P. All authors have read and agreed to the published version of the manuscript. Funding: This work was supported by the Wellcome Trust, UK (Grant number 106919/Z/15/Z). The funder had no role in study design, data collection and analysis, writing of the report, or in the decision to submit the paper for publication. Institutional Review Board Statement: The study was conducted in accordance with the Declara-tion of Helsinki and approved by the Institutional Review Board (or Ethics Committee) of the spon-soring organisation, Harvard Medical School (protocol code IRB20-0830 and date of approval: 23 October 2020) and implementing organisation, Sangath (protocol code VP_2015_017). Informed Consent Statement: Verbal informed assent/ consent was obtained from all participants involved in the study. All consent were recorded and stored. Due to COVID-19 restrictions, there were no written consent. Data Availability Statement: The raw data and coding framework used in this study can be ac-cessed on request from the corresponding author. Acknowledgments: We acknowledge the contributions of adolescents, their parents and guardians, school staff, research staff, and counsellors who made this work possible. We extend sincere thanks to Sai Priya Kumar, Niket Agrawal, Saleha Khatoon, and Anuj Kumar for their support in the pre-paring transcripts and compiling coded data. Conflicts of Interest: D.M. is a guest editor is a Special Issue of IJERPH. Other authors declare no conflict of interest. Appendix A In Table A1, each competency is rated on a 4-point scale, using the performance band given in Figure A1: Figure A1. Scoring Legend for Performance band *. * Not applicable is marked if the particular skill is not relevant for the particular session. This is to be used only for treatment specific skills. 1 = Limited; skill not performed or inappropriate performance with major problems evident; skill delivery is not useful in session; majority of outlined features are missing 2 = Basic; skill performed is somewhat appropriate, some of the outlined feature is covered, however there are substantive problems and/or inconsistencies in therapist’s performance; 3 = Good; Skill performed appropriately; helpful to student; most of the outlined features covered; how-ever minimal problems and/or consistencies are evident in therapist’s performance 4 = Advanced; Skill is highly developed; helpful to student, consistently well-performed; all outlined fea-tures covered Int. J. Environ. Res. Public Health 2023, 20, 1722 14 of 18 Table A1. PRIDE Therapy Quality Rating Scale (P-TQRS). S.no Item Description Performance Band Assessing Progress: Assessing progress from first session in terms of achievement of counselling goals and assessment of any risk and suitable management of the same 1. Remote Delivery Ground Rules Agenda setting a. Progress monitoring Monitoring risk for harm (applicable if needed as per the safety protocol) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) Established well-defined ground rules for remote delivery before beginning the counselling Discussed all of the ground rules openly with the student Invited student to ask questions, provide additional ground rules, or any potential barriers Provided possible solutions for any barriers that came up Established a well-defined and specific agenda at start of the ses- sion, suited to stage of counselling/module being presented and/or student’s problems Established feasible agenda that ensured all items could be thor- oughly covered within the session duration Referred to counselling planner/map to conceptualize the treatment flow for the current session as well as the treatment process Asked student for any additions to the agenda from their side 1 2 3 4 NA 1 2 3 4 NA (cid:3) Monitoring tools were implemented in systematic and prescribed (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) order. Used sample YTP graph to explain the process of progress monitor- ing to the student Reviewed completed measures and used information gleaned for relevant clinical purposes Explained to the student about the purpose of review of measures, provided feedback on results of the measure and how to proceed further If relevant, thoroughly identified and explored barriers to monitor- ing and worked with the student to overcome the barriers Conducted safety check if required (as per RCADS score) and ex- plained the purpose of the same Allowed the student adequate time to share his/her experience Skilfully and sensitively probed to establish level of risk Followed risk management protocol step by step in case of a moder- ate/high risk 1 2 3 4 NA 1 2 3 4 NA Use of appropriate intervention: choosing and implementing the module/techniques most suitable to student’s Stage Selecting appropriate intervention HINT: to be rated for selection of steps within particular intervention such as problem solving, behavioral activation and choice of one of three elements (BA/Exposure/Communication) in Modular Behavioral Intervention (MBI) Rationale for intervention HINT: to be rated for educating about the counselling programme and steps within a particular intervention such as problem solving, behavioral activation and choice of one of three element in Modular Behavioral Intervention (MBI) Implementation of interventionHINT: to be rated for implementing steps within a particular intervention such as problem solving, behavioral activation and choice of one of three element in Modular Behavior Intervention (MBI) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) Choosing the intervention was based on student preferences and theory or evidence- based practice. Intervention that was chosen was a part of coherent treatment strat- egy that would be beneficial for the student’s treatment goals as decided earlier The intervention was introduced at appropriate time; when student could actively participate in the particular intervention strategy. Provided clear rationale for the intervention component being cov- ered in the session, linked with student’s problems and potential benefits (eg-relaxation in each session of MBI, being active for uplift- ing mood, problem solving skills to deal effectively with problems, assertiveness for better communication, exposure to face fears) Explained about concepts using culturally and developmentally appropriate concepts If there were any ambiguities or doubts expressed by student, ad- dressed them. Implemented the interventions from start to finish in the manner in which they were intended Implemented interventions with a high level of skill (flexible scheduling and physically distanced activities for behavioural acti- vation, creating hierarchy for exposure, phone-based role play on assertiveness) Able to tailor interventions to suit the student’s needs where neces- sary 1 2 3 4 NA 1 2 3 4 NA 1 2 3 4 NA 2. 3. b. 4. a. b. c. Int. J. Environ. Res. Public Health 2023, 20, 1722 15 of 18 Table A1. Cont. S.no Item Description Performance Band Reviewing InterventionHINT: to be rated for discussing the intervention done within the session and what were the results of implementing the intervention technique in one of three elements in Modular Behavioral Intervention (MBI)(Applicable for MBI sessions) (cid:3) (cid:3) (cid:3) Results of doing the intervention task, positive or negative, were discussed with the student Student was guided towards drawing conclusions about learning implications from the practice of intervention task Current learning was linked with target thoughts, feelings and behaviours as was decided in the beginning 1 2 3 4 NA Sharing Material and Home practice: Giving booklet/resource material with explanation and completion of written work and/or implementation of tasks given in real life) Sharing Material Rationale for home practice HINT: Covers rationale of reading and writing exercises in booklets and/or implementation of tasks given in booklets Reviewing home practice (Session 2 onwards)HINT: Covers review of reading and writing exercises in booklets and/or implementation of tasks given in booklets Choosing & planning suitable home practiceHINT: Covers planning of reading and writing exercises in booklets and/or implementation of tasks given in booklets (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) Guided the student towards checking that they have all the resource material before counselling and before each session Helped the student in identifying each of these resource materials Asked student to refer to the resource materials by using page numbers or colours to guide them Ensured student is referring to the correct resource material and page number by asking them to read out the headings Explained the purpose of resource material(s) to the student. Showed skills in facilitating the student’s understanding of the purpose of home practice tasks with regards to student’s treatment goals. Reviewed assigned home practice with the student Helped student understand the implications from completing/not completing home practice and to link this learning to their problem- solving goals. If required, counsellor worked with the student to identify relevant reasons for non-completion and helped them identify specific ways to overcome future blocks to completion. Home practice built upon important issues dealt with in-session Formulated a clear and detailed plan of exactly what home practice involved. Prepared the student well for practicalities of home practice (what, how, when, where) homework Identified any potential obstacles, fully discussed these obstacles and (where possible) identified ways to overcome them. Fostering Therapeutic Relationship 1 2 3 4 NA 1 2 3 4 NA 1 2 3 4 NA 1 2 3 4 NA Engagement (cid:3) (cid:3) Was non-judgmental, non-critical, warm, sensitive, and respectful Demonstrated a positive interpersonal style 1 2 3 4 NA Confidentiality Collaboration Listing Options (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) Explained that all discussions in counselling are confidential Encouraged student to choose a time and place that is private for the counselling sessions Explained about exception to confidentiality i.e., harm to self Explained the student purpose of recording the call If the student did not understand, or asked questions, fully dis- cussed the issues. Shared Decision Making Introduced counselling as a collaborative process and student as architect of his/her own journey Encouraged the student to take an active role in and to share respon- sibility for all aspects of the session Explained the purpose of working collaboratively to the student At all decision points explained the different options to student and guided him/her towards making a decision Informed student about having to choose from options Listed out all the possible options Discussed the advantages and disadvantages of each option Answered all questions raised by the student 1 2 3 4 NA 1 2 3 4 NA 1 2 3 4 NA d. 5. a. b. c. d. 6. a. b. 7. a. b. Int. J. Environ. Res. Public Health 2023, 20, 1722 16 of 18 Table A1. Cont. S.no Item Description Performance Band c. Exploring Student Understanding of Decision d. Respecting Student Decision 8. Summaries and Feedback 9. Effective use of Time (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) (cid:3) Asked specific questions to ensure that student has understood all information before making a decision Encouraged student to ask questions to get more information to make an informed decision Allowed enough time in sessions for student to respond to counsel- lor queries and ask questions If required, encouraged student to seek further help or advice for making decision Assured the student that counsellor can only give a recommendation and they have to make the final decision Ensured that student’s decision is based on their understanding and current needs If student chose an option that was different from the counsellor recommendation, respected student’s decision Encouraged the student to review and reflect on session content periodically Responded to student summaries and discussed by asking relevant questions and adjusts session delivery accordingly Reinforced student understanding and filled the gaps in understand- ing where required Provided clear and meaningful brief capsule summaries Maintained focus on session priorities (i.e., agenda items) Ability to pace the session in a manner which is well suited to agenda and student’s capacity If unanticipated issues arose, counsellor acknowledged these, skil- fully evaluated and showed appropriate flexibility 1 2 3 4 NA 1 2 3 4 NA 1 2 3 4 NA 1 2 3 4 NA Score: ____________________ Mean score: References 1. Moreno, C.; Wykes, T.; Galderisi, S.; Nordentoft, M.; Crossley, N.; Jones, N.; Cannon, M.; Correll, C.U.; Byrne, L.; Carr, S.; et al. How Mental Health Care Should Change as a Consequence of the COVID-19 Pandemic. Lancet Psychiatry 2020, 7, 813–824. [CrossRef] [PubMed] 2. Wu, T.; Jia, X.; Shi, H.; Niu, J.; Yin, X.; Xie, J.; Wang, X. Prevalence of Mental Health Problems during the COVID-19 Pandemic: A 3. 4. 5. 6. Status Report: Government and Private Schools during COVID-19. Systematic Review and Meta-Analysis. J. Affect. Disord. 2021, 281, 91–98. [CrossRef] [PubMed] Oxfam India. //www.oxfamindia.org/sites/default/files/2020-09/Status%20report%20Government%20and%20private%20schools% 20during%20COVID%20-%2019.pdf (accessed on 31 October 2022). United Nations Children’s Fund; United Nations Educational, Scientific and Cultural Organization; The World Bank. What Have We Learnt: Overview of Findings from a Survey of Ministries of Education on National Responses to COVID-19. 2020. Available online: https://data.unicef.org/resources/national-education-responses-to-covid19/ (accessed on 31 October 2022). Borghouts, J.; Eikey, E.; Mark, G.; Leon, C.D.; Schueller, S.M.; Schneider, M.; Stadnick, N.; Zheng, K.; Mukamel, D.; Sorkin, D.H. Barriers to and Facilitators of User Engagement With Digital Mental Health Interventions: Systematic Review. J. Med. Internet Res. 2021, 23, e24387. [CrossRef] [PubMed] James, K. Remote Mental Health Interventions for Young People: A Rapid Review of the Evidence. Youth Access. 2022. Available online: https://childhub.org/sites/default/files/library/attachments/remote-mental-health-interventions-for-young- people-rapid-review.pdf (accessed on 31 October 2022). 2020. Available online: https: 7. Mondal, I.; Anand, N.; Sharma, M.K.; Kohli, T.; Thakur, P.C.; Kande, J.S.; Venkateshan, S.; Sj, A. Telephonic Psychotherapy in 8. 9. India: A Reminder of Challenges in Times of COVID-19. Asian J. Psychiatry 2020, 53, 102432. [CrossRef] [PubMed] Gonsalves, P.; Sharma, R.; Jambhale, A.; Chodankar, B.; Verma, M.; Hodgson, E.S.; Weiss, H.A.; Laurent, B.; Cavanagh, K.; Fairburn, C.; et al. Pilot randomised controlled trial of a remotely delivered online intervention for adolescent mental health problems in India: Lessons learned about low acceptability and feasibility during the COVID-19 pandemic. BJPsych Open 2022, 9, e7. [CrossRef] Chew, A.M.K.; Ong, R.; Lei, H.-H.; Rajendram, M.; Grisan, K.V.; Verma, S.K.; Fung, D.S.S.; Leong, J.J.; Gunasekeran, D.V. Digital Health Solutions for Mental Health Disorders During COVID-19. Front. Psychiatry 2020, 11, 582007. [CrossRef] 10. Aboujaoude, E.; Gega, L.; Parish, M.B.; Hilty, D.M. Editorial: Digital Interventions in Mental Health: Current Status and Future Directions. Front. Psychiatry 2020, 11, 111. [CrossRef] Int. J. Environ. Res. Public Health 2023, 20, 1722 17 of 18 11. Sagar, R.; Singh, S. National Tele-Mental Health Program in India: A Step towards Mental Health Care for All? Indian J. Psychiatry 2022, 64, 117–119. [CrossRef] 12. Koly, K.N.; Saba, J.; Muzaffar, R.; Modasser, R.B.; M, T.H.; Colon-Cabrera, D.; Warren, N. Exploring the Potential of Delivering Mental Health Care Services Using Digital Technologies in Bangladesh: A Qualitative Analysis. Internet Interv. 2022, 29, 100544. [CrossRef] 13. Michelson, D.; Malik, K.; Krishna, M.; Sharma, R.; Mathur, S.; Bhat, B.; Parikh, R.; Roy, K.; Joshi, A.; Sahu, R.; et al. Development of a Transdiagnostic, Low-Intensity, Psychological Intervention for Common Adolescent Mental Health Problems in Indian Secondary Schools. Behav. Res. Ther. 2020, 130, 103439. [CrossRef] 14. Chorpita, B.F.; Daleiden, E.L.; Malik, K.; Gellatly, R.; Boustani, M.M.; Michelson, D.; Knudsen, K.; Mathur, S.; Patel, V.H. Design Process and Protocol Description for a Multi-Problem Mental Health Intervention within a Stepped Care Approach for Adolescents in India. Behav. Res. Ther. 2020, 133, 103698. [CrossRef] 15. Roy, K.; Shinde, S.; Sarkar, B.K.; Malik, K.; Parikh, R.; Patel, V. India’s response to adolescent mental health: A policy review and stakeholder analysis. Soc. Psychiatry Psychiatr. Epidemiol. 2019, 54, 405–414. [CrossRef] 16. Michelson, D.; Malik, K.; Parikh, R.; Weiss, H.A.; Doyle, A.M.; Bhat, B.; Sahu, R.; Chilhate, B.; Mathur, S.; Krishna, M.; et al. Effectiveness of a Brief Lay Counsellor-Delivered, Problem-Solving Intervention for Adolescent Mental Health Problems in Urban, Low-Income Schools in India: A Randomised Controlled Trial. Lancet Child Adolesc. Health 2020, 4, 571–582. [CrossRef] 17. Malik, K.; Ibrahim, M.; Mathur, S.; Jose, J.E.; Nair, P.; Sahu, R.; Krishna, M.; Jangra, D.; Mathews, R.; Cuijpers, P.; et al. Development of a Transdiagnostic Stepped Care Programme for Common Adolescent Mental Health Problems in Indian Secondary Schools: Lessons from a Pilot Study Examining Acceptability and Feasibility. Glob. Ment. Health 2022, 9, 521–525. [CrossRef] 18. Proctor, E.; Silmere, H.; Raghavan, R.; Hovmand, P.; Aarons, G.; Bunger, A.; Griffey, R.; Hensley, M. Outcomes for Implementation Research: Conceptual Distinctions, Measurement Challenges, and Research Agenda. Adm. Policy Ment. Health 2011, 38, 65–76. [CrossRef] Sekhon, M.; Cartwright, M.; Francis, J.J. Acceptability of Healthcare Interventions: An Overview of Reviews and Development of a Theoretical Framework. BMC Health Serv. Res. 2017, 17, 88. [CrossRef] 19. 20. Parikh, R.; Michelson, D.; Malik, K.; Shinde, S.; Weiss, H.A.; Hoogendoorn, A.; Ruwaard, J.; Krishna, M.; Sharma, R.; Bhat, B.; et al. The Effectiveness of a Low-Intensity Problem-Solving Intervention for Common Adolescent Mental Health Problems in New Delhi, India: Protocol for a School-Based, Individually Randomized Controlled Trial with an Embedded Stepped-Wedge, Cluster Randomized Controlled Recruitment Trial. Trials 2019, 20, 568. [CrossRef] 21. Mishra, P.; Udas, A.; Pahwa, S.; Sen, S.; Haldar, K. Adapting a Revised Child Anxiety and Depression Scale for Rural India: A Pilot, Amenable to Scale Up. PsyArXiv 2016. [CrossRef] 22. Ebesutani, C.; Reise, S.P.; Chorpita, B.F.; Ale, C.; Regan, J.; Young, J.; Higa-McMillan, C.; Weisz, J.R. The Revised Child Anxiety and Depression Scale-Short Version: Scale Reduction via Exploratory Bifactor Modeling of the Broad Anxiety Factor. Psychol. Assess. 2012, 24, 833–845. [CrossRef] 23. Krause, K.R.; Edbrooke-Childs, J.; Singleton, R.; Wolpert, M. Are We Comparing Apples with Oranges? Assessing Improvement Across Symptoms, Functioning, and Goal Progress for Adolescent Anxiety and Depression. Child Psychiatry Hum. Dev. 2022, 53, 737–753. [CrossRef] 24. Weisz, J.R.; Chorpita, B.F.; Frye, A.; Ng, M.Y.; Lau, N.; Bearman, S.K.; Ugueto, A.M.; Langer, D.A.; Hoagwood, K.E. Research Network on Youth Mental Health Youth Top Problems: Using Idiographic, Consumer-Guided Assessment to Identify Treatment Needs and to Track Change during Psychotherapy. J. Consult. Clin. Psychol. 2011, 79, 369–380. [CrossRef] [PubMed] 25. Parikh, R.; Hoogendoorn, A.; Michelson, D.; Ruwaard, J.; Sharma, R.; Bhat, B.; Malik, K.; Sahu, R.; Cuijpers, P.; Patel, V. Increasing Demand for School Counselling through a Lay Counsellor-Delivered Classroom Sensitisation Intervention: A Stepped-Wedge Cluster Randomised Controlled Trial in New Delhi, India. BMJ Glob. Health 2021, 6, e003902. [CrossRef] [PubMed] 26. Goodman, R. The Extended Version of the Strengths and Difficulties Questionnaire as a Guide to Child Psychiatric Caseness and Consequent Burden. J. Child Psychol. Psychiatry 1999, 40, 791–799. [CrossRef] 27. Makoul, G.; Clayman, M.L. An Integrative Model of Shared Decision Making in Medical Encounters. Patient Educ. Couns. 2006, 60, 301–312. [CrossRef] [PubMed] 28. Gale, N.K.; Heath, G.; Cameron, E.; Rashid, S.; Redwood, S. Using the Framework Method for the Analysis of Qualitative Data in Multi-Disciplinary Health Research. BMC Med. Res. Methodol. 2013, 13, 117. [CrossRef] 29. Pretorius, C.; Chambers, D.; Coyle, D. Young People’s Online Help-Seeking and Mental Health Difficulties: Systematic Narrative 30. Review. J. Med. Internet Res. 2019, 21, e13873. [CrossRef] Sweeney, G.M.; Donovan, C.L.; March, S.; Forbes, Y. Logging into Therapy: Adolescent Perceptions of Online Therapies for Mental Health Problems. Internet Interv. 2019, 15, 93–99. [CrossRef] 31. Galagali, P.M.; Ghosh, S.; Bhargav, H. The Role of Telemedicine in Child and Adolescent Healthcare in India. Curr. Pediatr. Rep. 2021, 9, 154–161. [CrossRef] 32. O’Connor, M. School Counselling during COVID-19: An Initial Examination of School Counselling Use during a 5-Week Remote Learning Period. Pastor. Care Educ. 2022, 40, 81–91. [CrossRef] Int. J. Environ. Res. Public Health 2023, 20, 1722 18 of 18 33. Madigan, S.; Racine, N.; Cooke, J.E.; Korczak, D.J. COVID-19 and Telemental Health: Benefits, Challenges, and Future Directions. Can. Psychol. 2021, 62, 5–11. [CrossRef] 34. Witteveen, A.B.; Young, S.; Cuijpers, P.; Ayuso-Mateos, J.L.; Barbui, C.; Bertolini, F.; Cabello, M.; Cadorin, C.; Downes, N.; Franzoi, D.; et al. Remote Mental Health Care Interventions during the COVID-19 Pandemic: An Umbrella Review. Behav. Res. Ther. 2022, 159, 104226. [CrossRef] Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
10.1016_j.isci.2022.104876
iScience ll OPEN ACCESS Article Scale space detector for analyzing spatiotemporal ventricular contractility and nuclear morphogenesis in zebrafish Tanveer Teranikar, Cameron Villarreal, Nabid Salehin, ..., Hung Cao, Cheng–Jen Chuong, Juhyun Lee juhyun.lee@uta.edu Highlights Cardiac defect genes in humans have corresponding zebrafish orthologs Light sheet modality is very effective for non- invasive, 4D modeling of zebrafish Hessian detector is robust to varying nuclei scales and geometric transformations Watershed filter is effective for separating fused cellular volumes Teranikar et al., iScience 25, 104876 September 16, 2022 https://doi.org/10.1016/ j.isci.2022.104876 iScience ll OPEN ACCESS Article Scale space detector for analyzing spatiotemporal ventricular contractility and nuclear morphogenesis in zebrafish Tanveer Teranikar,1 Cameron Villarreal,1 Nabid Salehin,1 Toluwani Ijaseun,1 Jessica Lim,1 Cynthia Dominguez,1 Vivian Nguyen,2 Hung Cao,3 Cheng–Jen Chuong,1 and Juhyun Lee1,4,5,* SUMMARY In vivo quantitative assessment of structural and functional biomarkers is essen- tial for characterizing the pathophysiology of congenital disorders. In this regard, fixed tissue analysis has offered revolutionary insights into the underlying cellular architecture. However, histological analysis faces major drawbacks with respect to lack of spatiotemporal sampling and tissue artifacts during sample preparation. This study demonstrates the potential of light sheet fluorescence microscopy (LSFM) as a non-invasive, 4D (3days + time) optical sectioning tool for revealing cardiac mechano-transduction in zebrafish. Furthermore, we have described the utility of a scale and size-invariant feature detector, for analyzing individual morphology of fused cardiomyocyte nuclei and characterizing zebra- fish ventricular contractility. INTRODUCTION Zebrafish (Danio rero) are emerging as potent vertebrate model’s for modeling human congenital heart disorders (CHD) (Kula-Alwar et al., 2021, p. 2; Lee et al., 2018; Miura and Yelon, 2011; Vedula et al., 2017a; Yu and Hwang, 2022; Zhao et al., 2020). This is due to numerous attractive traits such as embryonic optical transparency, high fecundity, and ease in genetic or biomechanical modulation for mimicking the human CHD pathophysiology (Choi et al., 2013; Lee et al., 2018; Miura and Yelon, 2011; Rafferty and Quinn, 2018; Tu and Chi, 2012). As a result, zebrafish enable access to phenotypic screening of dynamic biome- chanical stimuli such as contractility and blood flow, responsible for modulating heart maturation (Kula- Alwar et al., 2021; Lee et al., 2018; Miura and Yelon, 2011; Tu and Chi, 2012; Vedula et al., 2017a). Previously conducted zebrafish studies have observed conserved cardiomyocyte count proportional to atrial/ventricular mass or volume per developmental stage (Kula-Alwar et al., 2021, p. 2). Moreover, recent studies suggest cardiomyocyte shape hypertrophy across three distinct ventricular regions—atrio- ventricular (AV) canal, outer curvature (OC), and inner curvature (IC) regions—apart from distinct atrial cardiomyocyte morphology (Kula-Alwar et al., 2021; Miura and Yelon, 2011; Tu and Chi, 2012). In addition, biologists have questioned the implications of cardiac mechano-transduction on enlarged cardiomyocyte morphology in the OC region in front of AV canal, with respect to spherical (isotropic) cardiomyocytes in the IC (Tu and Chi, 2012; Zhao et al., 2020). However, the ability to observe cardiomyo- cyte morphogenesis in vivo is adversely affected by tissue birefringence, hindering characterization of beforementioned cardiovascular phenotypes (Bensley et al., 2016; Bray et al., 2010; Ghonim et al., 2017; Teranikar et al., 2020). In this regard, automated feature detectors are proving to be an indispens- able tool for segmenting cellular volumes without human intervention to avoid gross inconsistencies and produce refined datasets. (Bolo´ n-Canedo and Remeseiro, 2020; Sargent et al., 2009; Torres and Judson- Torres, 2019). Conventionally, invasive sectioning procedures have offered revolutionary insights into aberrant tissue up to the cellular scale (Bensley et al., 2016; Javaeed et al., 2021; Teranikar et al., 2022). However, histopath- ological analysis currently suffers from severe limitations, primarily disruption to tissue homeostasis (Klec- zek et al., 2020; Teranikar et al., 2022). With respect to these drawbacks, the optical sectioning modality light sheet fluorescence microscopy (LSFM) has proved instrumental in probing dynamic organogenesis several millimeters inside tissue.(Ding et al., 2017; Fei et al., 2019; Lee et al., 2016; Teranikar et al., 2020; 1Joint Department of Bioengineering, UT Arlington/UT Southwestern, Arlington, TX, USA 2Martin High School/ UT Arlington, Arlington, TX, USA 3Department of Electrical Engineering, UC Irvine, Irvine, CA, USA 4Department of Medical Education, TCU and UNTHSC School of Medicine, Fort Worth, TX 76107, USA 5Lead contact *Correspondence: juhyun.lee@uta.edu https://doi.org/10.1016/j.isci. 2022.104876 iScience 25, 104876, September 16, 2022 This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). 1 ll OPEN ACCESS iScience Article Vedula et al., 2017b). LSFM has tremendously benefitted embryologists to become cognizant of dynamic phenomena such as mechano-transduction and undifferentiated precursor cell signaling pathways. However, acquisition of dynamic organogenesis reported by endogenous fluorophores is a challenging task owing to anisotropic contrast across the field of view (FOV). This is due to photon propagation through heterogeneous tissue (Teranikar et al., 2020, 2021). Hence, precise orchestration of in vivo volumes requires high sensitivity with respect to dynamic tissue motion and differing scales. As a result, optical aberrations often induce redundancy in the imaging sample space, affecting interpretability of feature attributes. Furthermore, cell studies are largely restricted to manual boundary demarcation due to the limited avail- ability of binary classification methods impervious to heterogeneous contrast resolution (Astrakas and Ar- gyropoulou, 2010; Marsh et al., 2018; Rajasekaran et al., 2016; Yin et al., 2014). Traditionally intensity-based segmentation methods such as the Otsu’s method, adaptive thresholding, isodata thresholding, and entropy-based thresholding have been used for automated cell tracking for their simplicity and speed (Goh et al., 2018). However, these methods are incapable of separating attenuated objects in closeness, proximity into meaningful biological regions (Xu et al., 2020). Another conventionally favored approach for biomedical image segmentation is the watershed algorithm (Beucher and Mathmatique, 2000; Koyuncu et al., 2012; Rajasekaran et al., 2016; Veta et al., 2013). However, the technique often causes over-segmen- tation or false detection of non-existent objects in dense tissue (Rajasekaran et al., 2016; Xu et al., 2020). Hence, there is a clear need for feature detectors impervious to low signal-to-noise ratio bioimages, for aiding accurate cell segmentation within the tissue architecture. In this study, we propose the application of a scale space feature detector for isolating fused, myocardial nuclei blob morphology across distinct embryonic stages (Johnsen, 2000; Johnsen et al., 2011; Johnsen and Widder, 1999; Teranikar et al., 2020; Zhang et al., 2013). The proposed segmentation framework integrates Hessian difference of Gaussian (HDoG) feature detector with the watershed algorithm, for enhancing sensitivity of localizing individual nuclei. In combining these two algorithms, provides for seam- less and straightforward segmentation with respect to background noise (Bharodiya and Gonsai, 2019). The proposed algorithm enabled in vivo characterization of wild-type zebrafish ventricular contractility and morphological traits such as nuclei number, area, and sphericity and in the myocardium. RESULTS Isolating individual cardiomyocyte nuclei in dynamic zebrafish ventricular volumes Distinguishing fused nuclei boundaries by manual contour segmentation or determining the intensity threshold for distorted nuclei outside the light sheet confocal parameter (focus region) is a complex task due to varying pixel intensities of overlapping nuclei (Figures 1A–1D). Furthermore, autofluorescence and dynamic cardiac motion convolutes lateral and axial imaging planes (Figures 1I and 1J). To separate nuclei clusters into disjoint regions, we implemented the difference of Gaussian (DoG) filter to enhance cell boundaries followed by the watershed algorithm to separate merged boundaries affected by aniso- tropic contrast. As a result, we were able to successfully quantify nuclei across different scales in a moder- ately dense cell environment at 48 h postfertilization (hpf) (Figures 1E–1H). More importantly, integration of the DoG scale space detector with the watershed algorithm enabled us to split longitudinally merged nuclei (Figures 1K–1L). In this regard, photon travel through heterogeneous tissue and restrictions imposed on resolvable sample depth are prone to induce sample de-focus (Figure 2A). Integration of Hessian and difference of Gaussian (HDoG) to segment cardiomyocyte nuclei from dense environment Compared to pinhole-based microscopy techniques, a potential cause of concern for LSFM modality man- ifests in the form of background contrast between adjacent cardiomyocytes (Figures 2B and 2C). As nuclei move dynamically across the field of view (FOV), undesired fluorescence emitted from fluorophore-binding sites outside the optical section beam waist affects accurate volumetric reconstruction (Figure 2A). Furthermore, intensity attenuation caused by low numerical aperture (NA) objectives aggravates poor signal-to-noise ratio (SNR). Although we successfully separated longitudinally merged ventricular myocardial nuclei at 48 hpf using DoG feature detector, we encountered inaccurate nuclei number quantification beyond 72 hpf. We assume low pixel intensities produced by the DoG edge detector response prior binarization, resulted in under re- porting of nuclei (Figures 2D and 2E). To overcome this, we applied the Hessian difference of Gaussian 2 iScience 25, 104876, September 16, 2022 iScience Article ll OPEN ACCESS Figure 1. Isolating and segmenting cardiomyocyte nuclei from contracting heart using the DOG (Difference of Gaussian) filter in combination with the watershed algorithm (A–D) 48 h postfertilization zebrafish ventricular volume was reconstructed using light sheet microscopy, in order to visualize time-dependent motion of myocardial cardiomyocyte nuclei. Raw volume comprised of fused nuclei clusters (yellow highlighted boxes), exacerbated by tissue scattering (B–D) Zoomed in regions demonstrate fused contours of nuclei, adversely affecting individual nuclei analysis (E) Ventricular volume was processed using the difference of Gaussian (DoG) edge detector in conjunction with the watershed algorithm to distinguish individual nuclei from adjacent neighbors. (F–H) Zoomed in regions show successful separation of nuclei for aiding cell tracking and counting. (I–J) 2D lateral and axial views illustrate tissue birefringence resulting in merging of nuclei longitudinally (K-L) Segmented lateral and axial views were re- constructed for qualitative assessment of contour separation of overlapping nuclei. (scale bar = 100 microns), a: atrium, v: ventricle. detector (HDoG) in combination with the watershed algorithm, for accurate contour separation and assess- ing the morphology of wild-type myocardial nuclei in vivo. The hessian determinant was used to localize saddle points (Marsh et al., 2018). Saddle points can be defined as neither an intensity maximum nor min- imum, that represent connecting nuclei edges. This approach improved detection sensitivity in the pres- ence of multiple intensity peaks for a single biomarker (Figures 2F and 2G). The segmented labels were further used for investigating nuclei shape and ventricular contractility, apart from nuclei counting. Segmentation accuracy evaluation Nuclei were detected for each distinct developmental phase: 48 hpf (Figures 3A–3C), 72 hpf (Figures 3D– 3F), and 96 hpf (Figures 3G–3I), to compare segmentation robustness for sparse nuclei distribution at 48 hpf with respect to densely populated ventricle at 96 hpf. We used a segmentation ratio to evaluate segmen- tation accuracy, by comparing Hessian DoG nuclei images to manual nuclei segmentation, with respect to static 3D zebrafish heart confocal images. The segmentation ratio is the number of scale space segmented nuclei divided by the number of cardiomyocyte nuclei manually counted in the confocal images as ground truth. If the numerical value = 1, the segmentation is identical to the raw images. If the numerical value >1, there is over-segmentation in the segmented images. If the numerical value <1, there is under segmenta- tion in the segmented images (Figure S1). Our analysis found that the ideal segmentation was repeated across developmental stages using the Hessian scale space (Figure S2). Quantification of local contractility via tracking cardiomyocyte nuclei After we processed the images to visualize individual nuclei, we performed contractility analysis by tracking cardiomyocyte nuclei across 48 to 120 hpf to quantify the local cardiac contractility based on this novel seg- mentation approach (Figures 4A–4G). We investigated the stretch level change of developing zebrafish iScience 25, 104876, September 16, 2022 3 ll OPEN ACCESS iScience Article Figure 2. Isolating individual nuclei volumes among high-density cardiomyocyte clusters, at distinct phases of ventricular contraction cycle (A) Illustration depicting zebrafish ventricular myocardial nuclei sections, scanned by a Gaussian light sheet (blue solid line). There exists a tradeoff between the confocal parameter i.e. excitation lateral extent and beam waist (BW) i.e. light sheet axial resolution and hence, requires optimization of the Gaussian focus spot to effectively sample embryogenesis across different growth stages. The detection objective lens modulates effective field of view (FOV). Samples are scanned through the static optical section at discrete increments (dx) using mechanical transducers, to reconstruct complete in vivo 3days + time volumes from individual sections. Red arrow represents the blood flow direction of zebrafish heart. (B and C) Raw systolic and diastolic nuclei reconstruction at 96 h (about 4 days) post fertilization, consisted of closely packed nuclei blobs as compared to 48 h postfertilization. Inaccurate nuclei localization is further exacerbated by dynamic contraction and relaxation. (D and E) Application of difference of Gaussian (DoG) detector in conjunction with the watershed algorithm, exhibits reduced feature detection sensitivity leading to inaccurate reporting of nuclei number. (F and G) Hessian DoG feature detector exhibits improved sensitivity to local affine transformations experienced by nuclei pixel neighborhoods during image acquisition. (scale bar = 50 micron), av: atrioventricular canal, v: ventricle, ot: outflow tract. heart and normalized the temporally changing stretch values for the innermost and outermost curvatures at each developmental stage (Figure S3). In addition to stretch, we calculated area ratio comparison between innermost curvature and outermost curvature areas. The area ratio is a description of local deformation of the area inside of three markers’ 2D stretch ratio. We analyzed the area ratio as a function of time, using three cardiomyocytes as markers. We found area ratio of the outermost curvature area, where the opposite side of the atrioventricular canal receiving blood directly from the atrium, has a higher area ratio than the innermost curvature area of the ventricle (Figures 4H–4J). Quantifying zebrafish cardiomyocyte nuclei development The average values for number of nuclei in a developing zebrafish heart were 159 G 13, 222 G 17, 260 G 13, and 284 G 10 for 48, 72, 96, and 120 hpf, respectively (Figure 5A) (n = 15). We observed cardiomyocyte nuclei in outermost curvature had larger systolic and diastolic volumes to innermost curvature area (Figures 5B–5E, Table S1). Hence, we assume the outermost ventricular curvature experiences higher me- chanical deformation (Figures 4H and 4I) due to direct inflow of blood from AV canal (Figure 4G), resulting in larger nuclei volumes. Contractility effect on morphology of cardiomyocyte nuclei Apart from area characteristics, we also quantified the circularity of myocardial ventricular nuclei (Figure 6). Isotropic/spherical nuclei in the innermost curvature region (Figures 6A and 6B) were evaluated to have an average elongation index of 0.91, while more elongated/ellipsoid nuclei in the outer curvature region nuclei had an average elongation index of 0.71, suggesting structural anisotropy. Interestingly, nuclei exhibit distinct eccentricity (major/minor axis ratio) according to their ventricular location, despite dynamic expansion and contraction across the cardiac cycle (Figures 5B–5E). In this regard, we observed spatially 4 iScience 25, 104876, September 16, 2022 iScience Article ll OPEN ACCESS Figure 3. Visualizing cmlc:GFPnuc zebrafish ventricular nuclei deformation at distinct developmental stages (48 – 96 h postfertilization), across the cardiac contraction cycle (A–I) The Hessian DoG scale space representation was used for localizing cardiomyocyte nuclei ranging from different sizes, as a result of which we were able to assess ventricular contractility and complex nuclei morphology in vivo (scale bar for A-C = 100-micron, scale bar for D-I = 50 micron). A:atrium, v:ventricle. confined cardiomyocyte nuclei in the innermost curvature with shorter major and minor axis lengths, compared to larger outer curvature nuclei (Figure S4). Hence, we hypothesize that different cardiomyocyte shapes (Table S1) are modulated by varying contractility in different ventricular regions (Figures 4H and 4I). This is in accordance with elongated nuclei volumes for accommodating greater mechanical stress in the outer ventricular concave regions, as compared to smaller nuclei volumes in the inner convex region. DISCUSSION Scale space theory can be understood as a hierarchal set of 2D images produced for each optical section, ob- tained by blurring the image from fine to coarser scale (Lindeberg, 1993; Marsh et al., 2018; Witkin, 1983). This results in suppression of all image objects equal to the size of the Gaussian kernel. Each defocused image con- tains a distinct number of edges obtained by blurring unresolved pixel subsets to a coarser resolution. Hence, enabling multiscale edge visualization without any knowledge of nuclei sizes a priori (Lindeberg, 1999, 2013). As zebrafish myocardial nuclei length varies spatiotemporally across 2–6 microns (Figure S4), we propose the integration of scale space theory and watershed segmentation for robust scale-invariant edge iScience 25, 104876, September 16, 2022 5 ll OPEN ACCESS iScience Article Figure 4. Selected markers utilized area ratio analysis (A–F) represents the systolic reconstruction of ventricular myocytes at 48 hpf, 72 hpf, and 96 hpf, respectively, while (D–F) represents the diastolic reconstruction of myocytes at different developmental stages. (G and H) Schematic illustrating the nuclei region of interest. Blue windows represent light sheet sections. Zebrafish ventricular volumes were sampled to compare the innermost curvature contractility (green markers), with respect to the outermost curvature (red markers) (H) Area ratio for innermost curvature by tracking three cardiomyocytes highlighted green in the blue optical plane, which elucidate increasing contractility trend observed across distinct developmental stages. (I) The area ratio for outermost curvature calculated by tracking cardiomyocytes highlighted red in the blue optical plane, indicates the outermost curvature has higher contractility compared to the innermost curvature. (J) Outermost curvature has a significantly higher area ratio compared to innermost curvature after 72 hpf (n = 3, p = 0.05, one-tail t-test). detection. Utilizing the inherent de-focus adaptation ability of the HDoG blob detector, we successfully isolated individual centers of mass (Videos S1, Video S2) for tracking dynamic cardiomyocyte nuclei (Video S3, Video S4). As a result, we were able to successfully characterize ventricular myocardial stretch post AV valve specification to heart maturation (Kula-Alwar et al., 2021; Miura and Yelon, 2011). Although transpar- ency was induced in zebrafish using PTU, tissue birefringence (RI(cid:1)1.3–1.5 (Jing et al., 2018)) results in changes in optical path lengths of emitted photons (Johnsen, 2000; Johnsen et al., 2011; Johnsen and Wid- der, 1999; Teranikar et al., 2020). Consequently, the light scatter compromises the optical modality pene- tration capability, resulting in fusing of nuclei situated outside the confocal region (Figure 2A) (Teranikar et al., 2020). Taking this into consideration, we sought to design an automated blob detection framework that provides high sensitivity and repeatability for a singular Gaussian intensity peak detection correspond- ing to each nuclei centroid. Furthermore, the proposed framework enabled in vivo quantification of morphological descriptors such as nuclei volume, surface area, and shape. 6 iScience 25, 104876, September 16, 2022 iScience Article ll OPEN ACCESS Figure 5. Zebrafish cardiomyocyte nuclei analysis (A and B) We observed an increase in the number of ventricular cardiomyocyte nuclei for successive developmental stages. Asterisk denotes statistically significant difference with respect to previous time point. p % 0.05 (B) Systolic and diastolic nuclei volume expansion observed for the inner curvature. (C) Systolic and diastolic volume trends observed for the outer curvature. (D and E) Systolic and diastolic nuclei surface area growth observed for the inner curvature (E) Systolic and diastolic nuclei surface area observed for the outer curvature. n = 15. Wild-type Tg(cmlc2:egfp) zebrafish have been observed to exhibit cuboidal cardiomyocyte morphology in linear heart tube (24 hpf) and IC ventricular myocardium, with respect to elongated cardiomyocytes in the OC (Auman et al., 2007; Kula-Alwar et al., 2021; Miura and Yelon, 2011). Studies indicate regionally distinct cardiomyocyte phenotypes such as cell count, area, or sphericity are regulated by mechanical stimuli such as contractility or blood flow during heart maturation (Auman et al., 2007; Miura and Yelon, 2011). This has been validated through the mutation phenotype half-hearted (haf) mutation lacking ventricular contractility. The haf mutant exhibited elongated cardiomyocytes with increased surface area across different parts of the ventricle including IC, result- ing in a distended ventricle (Auman et al., 2007). Interestingly, cardiomyocyte count was observed to be consis- tent between the haf mutant and wild-type zebrafish ventricle, suggesting contractility is responsible for moder- ating the aberrant elongation of cardiomyocytes and not anomalous proliferation (Auman et al., 2007). Furthermore, previously performed studies (Auman et al., 2007) indicate cardiomyocyte number reflects a sigmoidal growth trend (Figure 5), subsequently plateauing at later stages (>96 hpf) thereby signaling specifica- tion into the myocardium. In this regard, the proposed feature detector and cell tracking algorithm can prove extremely beneficial for gaining insights into the effects of cardiac contraction on reducing proliferation and its secondary effects on cardiomyocyte morphology in zebrafish. Unfortunately, currently we cannot conclude that contraction is key to reducing the proliferation of cardiomyocytes due to lack of statistically significant data. However, the intensity of mechanical workload experienced by cardiomyocytes in different parts of the ventricle appears to be a regulatory mechanism for maturation into distinct shapes (Figures 4, Figure 6). Analyzing the cardiomyocyte motion (Figure S3), we quantified the outermost curvature has higher area ratio than the innermost curvature (Figures 4H and 4I), thereby experiencing greater mechanical workload. Furthermore, we observed elongated cardiomyocyte nuclei morphology in the OC with respect to spher- ical morphology in the IC. Although no phenotyping screening of nuclei morphology has been performed with respect to modulation of contractility, our data indicate that cardiomyocyte nuclei shape and size cor- responds to deformation experienced by distinct ventricular regions. Elongated nuclei (Figure 6) in the OC suggest larger surface area is required to accommodate greater OC mechanical intensity. On the other hand, IC consists of smaller, cuboidal nuclei due to lesser deformation compared to OC. In the develop- mental biology aspect, researchers primarily focus on the ventricular OC where trabeculae form, but there is lack of well-documented research regarding lack of trabeculae in the IC. Thus, the question remains how different biomechanical or molecular signaling engenders a trabeculated OC and smooth IC. Our study has iScience 25, 104876, September 16, 2022 7 ll OPEN ACCESS iScience Article Figure 6. Systole vs diastole circularity analysis (A) Inner curvature nuclei are observed to have a circular shape, (symmetric circle elongation = 1, ellipse <1) with slightly higher values observed for the diastole. (B) Outer curvature cardiomyocyte nuclei are observed to have an elongated shape with higher elongation observed in the diastole. (C and D) Volumetric reconstructions of the circular shape of inner curvature and elliptic shape of outer curvature myocytes were visually presented, respectively. In addition, the corresponding lateral and axial views are shown with binary images (scale bar = 15 um). the potential to elucidate ventricular development in zebrafish orthologs, and aid cardiac pathophysiology diagnosis or clinical translational of cardiac regeneration for pediatric population. However, further inves- tigations will be required to validate this assumption. In this regard, nuclear morphology observed in car- diomyocytes isolated from neonatal rat ventricles reports similar findings, regarding systolic and diastolic heterogeneous cross-sectional surface areas due to deformation experienced by the cardiac cycle (Bray et al., 2010). Hence, our novel study provides exciting avenues to characterize cell count, morphology, and intercellular forces that may be responsible for cardiomyopathy in humans. Future studies will involve modulation of contractility to characterize cardiomyocyte morphology in the IC and OC. In summary, we have presented a scale-invariant feature detector for quantifying individual morphological characteristics of merged nuclei and biomechanical analysis of the zebrafish ventricle. Our proposed blob detection and cell tracking approach will prove to be extremely beneficial for analyzing cell count, volume, area, sphericity, proliferation, or cardiac function for characterization of cardiomyopathy phenotypes. Conclusion In this report, we were able to successfully interrogate dynamic zebrafish cardiac tissue non-invasively using bona fide biomarkers such as cell elongation, volume, and surface area. Moreover, we quantified the num- ber of cells and the mechanical workload experienced by the ventricular inflow and outflow regions during the systole and diastole, respectively. Limitations of the study Although we successfully separated merged nuclei clusters across varying scales and densities, the reproduc- ibility of the Hessian DoG feature detector is highly dependent on appropriate identification of Gaussian blurring weights (Figure S1). Moreover, Hessian scale space detector followed by watershed postprocessing is more prone to over-segmentation with higher variability in nuclei count, in comparison to DoG feature detection 8 iScience 25, 104876, September 16, 2022 iScience Article ll OPEN ACCESS (Figure S1) if kernel weights are not selected appropriately. On the other hand, DoG scale space detector is inherently prone to erosion of boundaries due to bandpass operation, resulting in reduction of nuclei volumes and the object area affecting quantification. Other modality limitations include absence of peripheral nuclei dur- ing diastole that may be present during the systole, due to ballooning of ventricle outside the light sheet confocal region. As in vivo cardiomyocyte cell tracking and counting requires invariancy to sample translation without distortion in shape, the Hessian DoG operation was effectively used to localize individual nuclei based on pixel intensity gradients. STAR+METHODS Detailed methods are provided in the online version of this paper and include the following: d KEY RESOURCES TABLE d RESOURCE AVAILABILITY B Lead contact B Materials availability B Data and code availability d EXPERIMENTAL MODEL AND SUBJECT DETAILS d METHOD DETAILS B Light sheet microscope (LSFM) implementation B Preparation of zebrafish for assessing cardiac function B Image processing framework B Cell counting and area measurements B Cardiac myocyte nuclei tracking B Contractility analysis d QUANTIFICATION AND STATISTICAL ANALYSIS SUPPLEMENTAL INFORMATION Supplemental information can be found online at https://doi.org/10.1016/j.isci.2022.104876. ACKNOWLEDGMENTS The authors would like to express gratitude to Dr. Caroline Burns and Geoffrey Burns from Boston Chil- dren’s Hospital for providing Tg(cmlc:nucGFP) for imaging and analysis. This study was supported by grants from AHA 18CDA34110150 (J.L.) and NSF 1936519 (J.L.). AUTHOR CONTRIBUTIONS Methodology and visualization, T.T. and J.L. Conceptualization, investigation, software and validation, T.T., C.L., N.S., CJ-C, and J.L. Writing – Original draft, T.T., C.L., and J.L. Writing - Review and editing, T.T., T.I., J.L., C.D., V.N., H.C., CJ-C, and J.L. Supervision, T.T., H.C., CJ-C, and J.L. Funding acquisition, J.L. Received: May 13, 2021 Revised: April 1, 2022 Accepted: July 29, 2022 Published: September 16, 2022 REFERENCES Astrakas, L.G., and Argyropoulou, M.I. (2010). Shifting from region of interest (ROI) to voxel- based analysis in human brain mapping. Pediatr. Radiol. 40, 1857–1867. https://doi.org/10.1007/ s00247-010-1677-8. Auman, H.J., Coleman, H., Riley, H.E., Olale, F., Tsai, H.-J., and Yelon, D. (2007). Functional modulation of cardiac form through regionally confined cell shape changes. PLoS Biol. 5, e53. https://doi.org/10.1371/journal.pbio.0050053. Bensley, J.G., De Matteo, R., Harding, R., and Black, M.J. (2016). Three-dimensional direct measurement of cardiomyocyte volume, nuclearity, and ploidy in thick histological sections. Sci. Rep. 6, 23756. https://doi.org/10.1038/srep23756. Bolo´ n-Canedo, V., and Remeseiro, B. (2020). Feature selection in image analysis: a survey. Artif. Intell. Rev. 53, 2905–2931. https://doi.org/ 10.1007/s10462-019-09750-3. Beucher, S., and Mathmatique, C. (2000). The watershed transformation applied to image segmentation. Scanning Microsc. 6. Bharodiya, A.K., and Gonsai, A.M. (2019). An improved edge detection algorithm for X-Ray images based on the statistical range. Heliyon 5, e02743. https://doi.org/10.1016/j.heliyon.2019. e02743. Bray, M.-A.P., Adams, W.J., Geisse, N.A., Feinberg, A.W., Sheehy, S.P., and Parker, K.K. (2010). Nuclear morphology and deformation in engineered cardiac myocytes and tissues. Biomaterials 31, 5143–5150. https://doi.org/10. 1016/j.biomaterials.2010.03.028. Choi, W.-Y., Gemberling, M., Wang, J., Holdway, J.E., Shen, M.-C., Karlstrom, R.O., and Poss, K.D. iScience 25, 104876, September 16, 2022 9 iScience Article the ventricle. Dev. Biol. 470, 95–107. https://doi. org/10.1016/j.ydbio.2020.11.008. APL Bioeng. 4, 036103. https://doi.org/10.1063/1. 5144613. Lee, J., Fei, P., Packard, R.R.S., Kang, H., Xu, H., Baek, K.I., Jen, N., Chen, J., Yen, H., Kuo, C.C.J., et al. (2016). 4-Dimensional light-sheet microscopy to elucidate shear stress modulation of cardiac trabeculation. J. Clin. Invest. 126, 1679– 1690. https://doi.org/10.1172/JCI83496. Teranikar, T., Villarreal, C., Salehin, N., Lim, J., Ijaseun, T., Cao, H., Chuong, C., and Lee, J. (2021). Feature Detection to Segment Cardiomyocyte Nuclei for Investigating Cardiac Contractility. Preprint at bioRxiv. https://doi.org/ 10.1101/2021.03.03.433810. Lee, J., Vedula, V., Baek, K.I., Chen, J., Hsu, J.J., Ding, Y., Chang, C.-C., Kang, H., Small, A., Fei, P., et al. (2018). Spatial and temporal variations in hemodynamic forces initiate cardiac trabeculation. JCI Insight 3, e96672. https://doi. org/10.1172/jci.insight.96672. Lindeberg, T. (2013). Scale selection properties of generalized scale-space interest point detectors. J. Math. Imaging Vis. 46, 177–210. https://doi. org/10.1007/s10851-012-0378-3. Lindeberg, T. (1999). Principles for automatic scale selection. In Handbook on Computer Vision and Applications, B. Ja¨ hne, ed. (Academic Press), pp. 239–274. Lindeberg, T. (1993). Detecting salient blob-like image structures and their scales with a scale- space primal sketch: a method for focus-of- attention. Int. J. Comput. Vis. 11, 283–318. https://doi.org/10.1007/BF01469346. Marsh, B.P., Chada, N., Sanganna Gari, R.R., Sigdel, K.P., and King, G.M. (2018). The hessian blob algorithm: precise particle detection in atomic force microscopy imagery. Sci. Rep. 8, 978. https://doi.org/10. 1038/s41598-018-19379-x. Miura, G.I., and Yelon, D. (2011). Chapter 7 - a guide to analysis of cardiac phenotypes in the zebrafish embryo. In Methods in Cell Biology, H.W. Detrich, M. Westerfield, and L.I. Zon, eds. (Academic Press), pp. 161–180. https://doi.org/ 10.1016/B978-0-12-387036-0.00007-4. Rafferty, S.A., and Quinn, T.A. (2018). A beginner’s guide to understanding and implementing the genetic modification of zebrafish. Prog. Biophys. Mol. Biol. 138, 3–19. https://doi.org/10.1016/j. pbiomolbio.2018.07.005. Rajasekaran, B., Uriu, K., Valentin, G., Tinevez, J.-Y., and Oates, A.C. (2016). Object segmentation and ground truth in 3D embryonic imaging. PLoS One 11, e0150853. Sargent, D., Chen, C.-I., Tsai, C.-M., Wang, Y.-F., and Koppel, D. (2009). Feature detector and descriptor for medical images. In Medical Imaging 2009: Image Processing, J.P.W. Pluim and B.M. Dawant, eds. (SPIE), pp. 991–998. https://doi.org/10.1117/12.811210. Teranikar, T., Lim, J., Ijaseun, T., and Lee, J. (2022). Development of planar illumination strategies for solving mysteries in the sub-cellular realm. Int. J. Mol. Sci. 23, 1643. https://doi.org/10. 3390/ijms23031643. Teranikar, T., Messerschmidt, V., Lim, J., Bailey, Z., Chiao, J.-C., Cao, H., Liu, J., and Lee, J. (2020). Correcting anisotropic intensity in light sheet images using dehazing and image morphology. Torres, R., and Judson-Torres, R.L. (2019). Research techniques made simple: feature selection for biomarker discovery. J. Invest. Dermatol. 139, 2068–2074.e1. https://doi.org/10. 1016/j.jid.2019.07.682. Tu, S., and Chi, N.C. (2012). Zebrafish models in cardiac development and congenital heart birth defects. Differentiation. 84, 4–16. https://doi.org/ 10.1016/j.diff.2012.05.005. Vedula, V., Lee, J., Xu, H., Kuo, C.-C.J., Hsiai, T.K., and Marsden, A.L. (2017a). A method to quantify mechanobiologic forces during zebrafish cardiac development using 4-D light sheet imaging and computational modeling. PLoS Comput. Biol. 13, e1005828. https://doi.org/10.1371/journal.pcbi. 1005828. Vedula, V., Lee, J., Xu, H., Kuo, C.-C.J., Hsiai, T.K., and Marsden, A.L. (2017b). A method to quantify mechanobiologic forces during zebrafish cardiac development using 4-D light sheet imaging and computational modeling. PLoS Comput. Biol. 13, e1005828–24. https://doi.org/10.1371/journal. pcbi.1005828. Veta, M., van Diest, P.J., Kornegoor, R., Huisman, A., Viergever, M.A., and Pluim, J.P.W. (2013). Automatic nuclei segmentation in H&E stained breast cancer histopathology images. PLoS One 8, e70221. Witkin, A. (1983). Scale-Space Filtering (IJCAI). Xu, Y., Wu, T., Gao, F., Charlton, J.R., and Bennett, K.M. (2020). Improved small blob detection in 3D images using jointly constrained deep learning and Hessian analysis. Sci. Rep. 10, 326. https://doi. org/10.1038/s41598-019-57223-y. Yin, X., Ng, B.W.-H., He, J., Zhang, Y., and Abbott, D. (2014). Accurate image analysis of the retina using hessian matrix and binarisation of thresholded entropy with application of texture mapping. PLoS One 9. e95943–17. https://doi. org/10.1371/journal.pone.0095943. Yu, H.-L., and Hwang, S.-P.L. (2022). Zebrafish integrin a3b is required for cardiac contractility and cardiomyocyte proliferation. Biochem. Biophys. Res. Commun. 595, 89–95. https://doi. org/10.1016/j.bbrc.2022.01.083. Zhang, H., Li, Y., Chen, H., Yuan, D., and Sun, M. (2013). Perceptual contrast enhancement with dynamic range adjustment. Optik 124, 5906– 5913. https://doi.org/10.1016/j.ijleo.2013.04.046. Zhao, M.-T., Ye, S., Su, J., and Garg, V. (2020). Cardiomyocyte proliferation and maturation: two sides of the same coin for heart regeneration. Front. Cell Dev. Biol. 8, 594226. https://doi.org/ 10.3389/fcell.2020.594226. ll OPEN ACCESS (2013). In vivo monitoring of cardiomyocyte proliferation to identify chemical modifiers of heart regeneration. Development 140, 660–666. https://doi.org/10.1242/dev.088526. Ding, Y., Lee, J., Ma, J., Sung, K., Yokota, T., Singh, N., Dooraghi, M., Abiri, P., Wang, Y., Kulkarni, R.P., et al. (2017). Light-sheet fluorescence imaging to localize cardiac lineage and protein distribution. Sci. Rep. 7, 42209. https://doi.org/10.1038/srep42209. Fei, P., Nie, J., Lee, J., Ding, Y., Li, S., Zhang, H., Hagiwara, M., Yu, T., Segura, T., Ho, C.-M., et al. (2019). Subvoxel light-sheet microscopy for high-resolution high- throughput volumetric imaging of large biomedical specimens. Adv. Photonics 1, 1. Ghonim, S., Voges, I., Gatehouse, P.D., Keegan, J., Gatzoulis, M.A., Kilner, P.J., and Babu- Narayan, S.V. (2017). Myocardial architecture, mechanics, and fibrosis in congenital heart disease. Front. Cardiovasc. Med. 4, 30. Goh, T.Y., Basah, S.N., Yazid, H., Aziz Safar, M.J., and Ahmad Saad, F.S. (2018). Performance analysis of image thresholding: Otsu technique. Measurement 114, 298–307. https://doi.org/10. 1016/j.measurement.2017.09.052. Javaeed, A., Qamar, S., Ali, S., Mustafa, M.A.T., Nusrat, A., and Ghauri, S.K. (2021). Histological stains in the past, present, and future. Cureus 13, e18486. https://doi.org/10.7759/cureus.18486. Jing, D., Zhang, S., Luo, W., Gao, X., Men, Y., Ma, C., Liu, X., Yi, Y., Bugde, A., Zhou, B.O., et al. (2018). Tissue clearing of both hard and soft tissue organs with the PEGASOS method. Cell Res. 28, 803–818. https://doi.org/10.1038/s41422-018-0049-z. Johnsen, S. (2000). Transparent animals. Sci. Am. 282, 80–89. Johnsen, S., Marshall, N.J., and Widder, E.A. (2011). Polarization sensitivity as a contrast enhancer in pelagic predators: lessons from in situ polarization imaging of transparent zooplankton. Philos. Trans. R. Soc. Lond. B Biol. Sci. 366, 655–670. https://doi.org/10.1098/rstb.2010.0193. Johnsen, S., and Widder, E. (1999). The physical basis of transparency in biological tissue: ultrastructure and the minimization of light scattering. J. Theor. Biol. 199, 181–198. https:// doi.org/10.1006/jtbi.1999.0948. Kleczek, P., Jaworek-Korjakowska, J., and Gorgon, M. (2020). A novel method for tissue segmentation in high-resolution H&E-stained histopathological whole-slide images. Comput. Med. Imag. Graph. 79, 101686. https://doi.org/ 10.1016/j.compmedimag.2019.101686. Koyuncu, C.F., Arslan, S., Durmaz, I., Cetin-Atalay, R., and Gunduz-Demir, C. (2012). Smart markers for watershed-based cell segmentation. PLoS One 7, e48664. https://doi.org/10.1371/journal. pone.0048664. Kula-Alwar, D., Marber, M.S., Hughes, S.M., and Hinits, Y. (2021). Mef2c factors are required for early but not late addition of cardiomyocytes to 10 iScience 25, 104876, September 16, 2022 iScience Article STAR+METHODS KEY RESOURCES TABLE ll OPEN ACCESS REAGENT or RESOURCE SOURCE IDENTIFIER Chemicals, peptides, and recombinant proteins 0.0025% 1-phenyl 2-thoiurea Sigma-Aldrich 0.05% tricaine (MS 222) Sigma-Aldrich P7629 E10521 Experimental models: Organisms/strains Zebrafish: Tg(cmlc2:nucGFP) Software and algorithms Sharpe M et al. Gifted by Dr Barnes at Boston children’s hospital, Harvard Medical. ImageJ Schneider et al., 2012 https://imagej.nih.gov/ij/ Hessian Determinant plugin Sato, Y. et al. https://imagescience.org/meijering/software/featurej/ Other Cardiomyocyte nuclei tracking code This paper RESOURCE AVAILABILITY Lead contact Chuong CJ, Sacks MS, Templeton G, Schwiep F, Johnson RL Jr. Regional deformation, and contractile function in canine right ventricular free wall. Am J Physiol. 1991;260(4 Pt 2):H1224-H1235. https://doi.org/10.1152/ajpheart.1991.260.4.H1224 Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Dr Juhyun Lee (juhyun.lee@uta.edu). Materials availability This study did not generate new unique reagents. Data and code availability d Microscopy images published in this paper will be shared by the lead contact upon request. d Original code is uploaded in the supplementary documents and is publicly available as of the date of publication. Section 1: Data All data reported in this paper will be shared by the lead contact upon request. Section 2: Code All original code is available in this paper’s supplemental information. Section 3: Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. EXPERIMENTAL MODEL AND SUBJECT DETAILS The animal experiments were performed in agreement with the UT Arlington Institutional Animal Care and Use Committee (IACUC) protocol (#A17.014). The transgenic zebrafish line used in this particular study is the Tg(cmlc2:nucGFP), with the cardiomyocyte nuclei labeled with GFP (Green Fluorescent Protein). The zebrafish embryos were maintained at 28.5(cid:3)C in system water at the UT Arlington Aquatic Animal Core Facility. 0.0025% 1-phenyl 2-thoiurea was added to the embryo medium starting around 20–24 hpf to sup- press pigmentation. Prior to imaging, embryos were anesthetized in 0.05% tricaine (MS 222, E10521, Sigma-Aldrich, St-Louis, MO) to avoid sample movement. Sex determination and segregation of zebrafish was not performed in the embryonic stage (48–120 h postfertilization). iScience 25, 104876, September 16, 2022 11 ll OPEN ACCESS iScience Article METHOD DETAILS Light sheet microscope (LSFM) implementation Our home built light sheet microscope consists of a single-side excitation pathway and a custom water dipping lens (20x/0.5 NA UMPlanFL N, Olympus, Tokyo, Japan) detection. In the illumination pathway, a cylindrical lens (LJ1695RM, Thorlabs) coupled with a 4x objective lens (4X Plan Apochromat Plan N, light-sheet with (cid:1)4-5-micron thickness. Olympus, Tokyo, Japan), are used to collimate a cylindrical Furthermore, a mechanical slit aperture (VA100C, Thorlabs) is modulated across distinct developmental to accommodate ventricular circumferential extent across the stages (48-,72-,96- and 120 hpf), light sheet confocal region (Figure 2). A DC servo motor actuator (Z825B, Thorlabs) is used for sample translation in the axial direction (z-step velocity and acceleration = 0.005 mm/s). The optical detection pathway consisting of the water lens, infinity corrected tube lens (TTL 180-A, Thorlabs) and sCMOS camera (ORCA flash 4.0, Hamamatsu, Japan, camera pixel size = [6.5 (um)^2] = [6.5/20x = 0.325 um], camera exposure time = 30–50 ms), is used for non-gated 4D (3D + time) cardiac volume acquisition. As the zebrafish ventricle undergoes periodic deformation during peak systole to end diastole, optical sections were acquired at varying depths in the sample, covering 4–5 cardiac cycles20,23. Since triggering of image slices is not synchronized to a particular phase in the cardiac cycle, we performed volumetric reconstruction a posteriori to ensure alignment of adjacent optical sections. For this purpose, we estimated the period of each individual cycle by minimization of the least squares intensity difference criterion and calculated the relative period shift to ensure synchronization between independent cardiac cycles20,23 Preparation of zebrafish for assessing cardiac function The animal experiments were performed in agreement with the UT Arlington Institutional Animal Care and Use Committee (IACUC) protocol (#A17.014). The transgenic zebrafish line used in this particular study is the Tg(cmlc2:nucGFP), with the cardiomyocyte nuclei labeled with GFP (Green Fluorescent Protein)19. The zebrafish embryos were maintained at 28.5(cid:3)C in system water at the UT Arlington Aquatic Animal Core Facility. 0.0025% 1-phenyl 2-thoiurea was added to the embryo medium starting at 20–24 hpf 4,32 to suppress pigmentation. Prior to imaging, embryos were anesthetized in 0.05% tricaine (MS 222, E10521, Sigma-Aldrich, St-Louis, MO) to avoid sample movement. Upon administering the anes- thetic, alive embryos were embedded in 0.5% low-melt agarose gel inside a fluorinated ethylene propyl- ene (FEP) tube (1677L, IDEX, Chicago,IL). Furthermore, the FEP tube was suspended in water within a custom 3days printed ABS (Acrylonitrile Butadiene Styrene) cuvette (designed using solid works) housing the water dipping lens, to ensure near isotropic refractive index between the water dipping lens and sample inside the tube. (Refractive index of water = 1.33, refractive index of agarose and FEP tube = 1.34). Refractive index matching is necessary to avoid distortions and intensity attenuation in the optical sections13. Image processing framework Haze removal using the dark channel prior (DCP) method Introduction of haze by the ambient medium or scattering due to particulate matter, degrades the perfor- mance of computer vision tasks(Lee et al., 2016; Teranikar et al., 2020). A haze free image can be retrieved by using the image degradation model based on the Dark Channel Prior (DCP) algorithm. IðxÞ = JðxÞ:tðxÞ + Að1 (cid:4) tðxÞÞ (Equation 1) (Lee et al., 2016; Teranikar et al., 2020) where I(x) is the degraded image, J(x) is the original irradiance captured by the CMOS camera, t(x) repre- sents the scene depth and A is the scattering introduced by the ambient light. Using the dehazing algorithm, we estimated the intensity transmission map t(x) using the imreducehaze() MATLAB function (Teranikar et al., 2020). tðxÞ = e(cid:4) bdðxÞ (Equation 2) (Lee et al., 2016) 12 iScience 25, 104876, September 16, 2022 iScience Article ll OPEN ACCESS where b represents the scattering coefficient and d represents the scene depth. We used the estimated intensity transmission map as a preprocessing step before performing the DoG operation. By estimating the contrast attenuation with respect to distance, we were able to emphasize edges. Intensity maxima localization at nuclei centers using the difference of Gaussian (DoG) filter The DoG filter can be effectively used to enhance edge visualization for images suffering from poor contrast. In this study, the greyscale bandpass operation is performed by subtracting a blurred version of the transmission estimate from a lesser blurred version of itself, tðxÞ (cid:5) g1ðxÞ (cid:4) tðxÞ (cid:5) g2ðxÞ = tðxÞ (cid:5) ðg1ðxÞ (cid:4) g2ðxÞÞ; (Equation 3) where g1(x) and g2(x) are the Gaussian kernels having different standard deviations. Using the DoG filter, we were able to localize blobs to nuclei centers by isolating spatial frequencies correlating to the Gaussian illumination maxima. Precise contour delineation using the hessian scale space representation and watershed algorithm The hessian scale space representation can be described by the convolution: Dðx; y; tÞ = ½tðxÞ (cid:5) ðg1ðxÞ (cid:4) g2ðxÞÞ(cid:6) (cid:5) Gðx; y; tÞ; (Equation 4) (Marsh et al., 2018; Rajasekaran et al., 2016) Where D(x,y,t) represents the family of images, derived from the original image. t represents the degree of blurring. Hence, Equation (4) can be described as the convolution of the DoG blob maxima image with the hessian blob detector Gaussian blur kernel G(x,y) at different degrees of blur (t > 0). The blurring scale selection was based on the ratio t +1 = r *t (Marsh et al., 2018), where r is a constant. The workflow involved for the hessian blob involves (Marsh et al., 2018),(Rajasekaran et al., 2016), 1) Computing the absolute magnitude of the intensity gradient image obtained by convolving the DoG bandpass image with the derivative of Gaussian filter. 2) Computing the double derivative of the absolute magnitude image calculated in the previous step 3) Imposing boundary conditions on the hessian determinant value [det D(x,y,t) < 0] (Rajasekaran et al., 2016) at every pixel, for indicating saddle points. The image arithmetic operation (OR – operation) results in the union of the DoG localized intensity maxima and contour information from the Hessian blob, aiding the successful splitting of nuclei. Preprocessing strategies Images corrupted by noise or tissue scatter, were filtered by using a Gaussian kernel with an appropriate SD followed by the background subtract operation in ImageJ. In addition, image processing code is enclosed (Data S2, related to Figure 2). Cell counting and area measurements After converting raw optical images to binary images, we performed to count cardiomyocyte nuclei and their area analysis by using the 3D object counter plugin33 in ImageJ. The plugin can be accessed by: Im- ageJ – Analyze – 3D Object Counter. After cropping the ROI (ventricle in this case), we used the plugin to quantify number of object voxels (volume), surface voxels of individual nuclei volumes and the number of 3D nuclei objects in the ventricular stack. The plugin can also be used to retrieve the centroid geometric coordinates of object volumes. The user is required configure 2 important parameters namely, (a) intensity threshold to separate background and foreground pixel populations and (b) size threshold to exclude smaller objects from the analysis. The plugin allows user to configure object counting based on the presence or absence of touching edges. iScience 25, 104876, September 16, 2022 13 ll OPEN ACCESS iScience Article Cardiac myocyte nuclei tracking We utilized the segmented, processed, and time synchronized images to reconstruct three-dimensional volumes through time for a cardiac cycle to perform this tracking. We then passed these images through a custom MATLAB (Mathworks) code to perform for key steps (Data S3, related to Figure 4). This MATLAB code performed following 4 steps. 1 The code compiles images into easily searchable 4D matrices. 2 The code resolves the 4D matrices of segmented images into centers of mass based on high pixel concentration areas for each time step. 3 The user selects three markers to represent our plane for stretch calculations. 4 The code searches through the 4D stack of centers of mass to determine the closest center of mass in the next time step and stores these points in a matrix of position values. Each stored triplet value is the x, y, and z position of a particular nucleus at a particular time. This format is easily searchable and allows for a multitude of calculations. This code assumes that there can be no erratic motion of the nucleus with a high enough sampling frequency. The location at each time step depends on the prior location. Imaging with a high sampling frequency supply data that meets this assumption require- ment. Other works have utilized similar works, including Meijerling et al. (Meijering et al., 2009).Drawbacks of this method include the requirement for user interaction. To verify that the cell tracking occurs appro- priately, the user must analyze each vector to ensure the vector does not violate the small motion assump- tion. This process can become time-consuming and increases the chance of human error. Subsequent work can expand and refine this cell tracking method to include other parameters, including a probability net for machine learning applications and size and orientation to decrease ambiguity and reduce the user input requirement. Contractility analysis We selected and tracked three cardiomyocyte nuclei for both the innermost and outermost curvature. After tracking the location of three cardiomyocytes through each time instance, we utilized the following method to determine the deformation gradient with normalized one cardiac cycle as 0.5 s starting from ventricular end-systolic stage. We determined the stretch ratio at each time instance into principal stretch values re- ported as l1 and l2, or the longitudinal and circumferential principal direction followed by previous methods41,42. These principal stretch vectors correspond to the first and second principal strain directions. When viewed on Mohr’s circle, they correspond to the maximum and minimum normal strain values where the shear strain is resolved to zero (Figure S6). These values are represented in the Cartesian coordinate system as the x and y direction or in polar coordinates as the zero-degree rotation and 90-degree rotation. We established the area ratio by multiplying the two principal stretch values. Area ratio provides a description of the total in plane deformation from the initial undeformed state which was selected as the start of filling. QUANTIFICATION AND STATISTICAL ANALYSIS For statistical analysis, we performed ad hoc pairwise comparisons for three morphological parameters to characterize the maturity of the heart (p value = 0.05)10. We analyzed the number of visible nuclei, the total volume, and total surface area. We estimated each of these parameters using built in functions in ImageJ (NIH, Bethesda, MD) with n = 15. Additionally, we cleaned the data in excel utilizing Chauvenet’s criterion to determine which values were outliers and should be removed. After removing outliers and cleaning the datasets in excel to reduce the chance of error due to our sampling technique, we compared the data with one-way ANOVA. If we detected a statistically significant difference for any comparison, we performed Tukey’s test for multiple comparison of means. This test inherently compensates for multiple comparisons, which allowed us to use an alpha value of 0.5. All values herein are reported as mean +/(cid:4) standard devi- ation in the figures and respective figure legends. 14 iScience 25, 104876, September 16, 2022
10.18063_ijb.v9i2.668
Zn-doped chitosan/alginate multilayer coatings on porous hydroxyapatite Zn-doped chitosan/alginate multilayer coatings on porous hydroxyapatite scaffold with osteogenic and antibacterial properties scaffold with osteogenic and antibacterial properties Zhijing He, Chen Jiao, Junnan Wu, Jiasen Gu, Huixin Liang, Lida Shen, Youwen Yang, Zongjun Tian, Chang Wang, Qing Jiang Publication date Publication date 13-01-2023 Licence Licence This work is made available under the CC BY 4.0 licence and should only be used in accordance with that licence. For more information on the specific terms, consult the repository record for this item. Document Version Document Version Published version Citation for this work (American Psychological Association 7th edition) Citation for this work (American Psychological Association 7th edition) He, Z., Jiao, C., Wu, J., Gu, J., Liang, H., Shen, L., Yang, Y., Tian, Z., Wang, C., & Jiang, Q. (2023). Zn-doped chitosan/alginate multilayer coatings on porous hydroxyapatite scaffold with osteogenic and antibacterial properties (Version 1). University of Sussex. https://hdl.handle.net/10779/uos.23494505.v1 Published in Published in International Journal of Bioprinting Link to external publisher version Link to external publisher version https://doi.org/10.18063/ijb.v9i2.668 Copyright and reuse: Copyright and reuse: This work was downloaded from Sussex Research Open (SRO). This document is made available in line with publisher policy and may differ from the published version. Please cite the published version where possible. Copyright and all moral rights to the version of the paper presented here belong to the individual author(s) and/or other copyright owners unless otherwise stated. For more information on this work, SRO or to report an issue, you can contact the repository administrators at sro@sussex.ac.uk. Discover more of the University’s research at https://sussex.figshare.com/ International Journal of Bioprinting RESEARCH ARTICLE Zn-doped chitosan/alginate multilayer coatings on porous hydroxyapatite scaffold with osteogenic and antibacterial properties Zhijing He1,2†, Chen Jiao1,2†, Junnan Wu1,2, Jiasen Gu1,2, Huixin Liang3, Lida Shen1,2,4*, Youwen Yang4*, Zongjun Tian1,2,4, Changjiang Wang5, Qing Jiang6 1Institute of Additive Manufacturing (3D Printing), Nanjing University of Aeronautics and Astronautics, Nanjing, 210016, China 2College of Mechanical and Electrical Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing, 210016, China 3State Key Laboratory of Pharmaceutical Biotechnology, Nanjing University Medical School Affiliated Nanjing Drum Tower Hospital, Nanjing, 210008, China 4Institute of Additive Manufacturing, Jiangxi University of Science and Technology, Ganzhou, 341000, China 5Department of Engineering and Design, University of Sussex, Brighton, BN1 9RH, United Kingdom 6State Key Laboratory of Pharmaceutical Biotechnology, Division of Sports Medicine and Adult Reconstructive Surgery, Department of Orthopedic Surgery, Nanjing Drum Tower Hospital, The Affiliated Hospital of Nanjing University Medical School, Nanjing, 210008, China Abstract Porous hydroxyapatite (HA) scaffolds prepared by three-dimensional (3D) printing have wide application prospects owing to personalized structural design and excellent biocompatibility. However, the lack of antimicrobial properties limits its widespread use. In this study, a porous ceramic scaffold was fabricated by digital light processing (DLP) method. The multilayer chitosan/alginate composite coatings prepared by layer-by-layer method were applied to scaffolds and Zn2+ was doped into coatings in the form of ion crosslinking. The chemical composition and morphology of coatings were characterized by scanning electron microscope (SEM) and X-ray photoelectron spectroscopy (XPS). Energy dispersive spectroscopy (EDS) analysis demonstrated that Zn2+ was uniformly distributed in the coating. Besides, the compressive strength of coated scaffolds (11.52 ± 0.3 MPa) was slightly improved compared with that of bare scaffolds (10.42 ± 0.56 MPa). The result of soaking experiment indicated that coated scaffolds exhibited delayed degradation. In vitro experiments demonstrated that within the limits of concentration, a higher Zn content in the coating has a stronger capacity to promote cell adhesion, proliferation and differentiation. Although excessive release of Zn2+ led to cytotoxicity, it presented a stronger antibacterial effect against Escherichia coli (99.4%) and Staphylococcus aureus (93%). Keywords: Porous hydroxyapatite scaffold; Multi-layer polymer coating; Zn doping; Osteogenic property; Antibacterial property †These authors contributed equally to this work. *Corresponding authors: Lida Shen (ldshen@nuaa.edu.cn) Youwen Yang (yangyouwen@jxust.edu.cn) Citation: He Z, Jiao C, Wu J, et al., 2023, Zn-doped chitosan/ alginate multilayer coatings on porous hydroxyapatite scaffold with osteogenic and antibacterial properties. Int J Bioprint, 9(2): 668. https://doi.org/10.18063/ijb.v9i2.668 Received: September 7, 2022 Accepted: October 21, 2022 Published Online: January 13, 2023 Copyright: © 2023 Author(s). This is an Open Access article distributed under the terms of the Creative Commons Attribution License, permitting distribution, and reproduction in any medium, provided the original work is properly cited. Publisher’s Note: Whioce Publishing remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Volume 9 Issue 2 (2023) 292 https://doi.org/10.18063/ijb.v9i2.668 1. Introduction Skeleton supports human body and protects other vulnerable internal organs, and plays a key role in blood production and mineral storage[1-3]. To ensure a healthy skeletal system, bones are in a constant process of remodeling to adjust to mechanical injuries and tiny lesions. However, once the bone defects exceed the critical size that the bone can no longer repair itself, bone substitutes would be needed for healing purposes[4,5]. Since the traditional bone substitutes, including autologous bone and allogeneic bone, were unable to fulfill therapeutic requirements, synthetic bone substitutes have gradually become the focus of research[6]. Calcium phosphate (CaP) bioceramics were one of the earliest bone repair materials in relevant researches because of its biocompatibility and osteoblastic induction properties[7-9]. Nonetheless, pure CaP bioceramics showed poor mechanical and antimicrobial properties, which can be improved by modification[10,11]. In view of the increase of bone transplant failure cases caused by infection every year, antimicrobial property has become one of the most desired properties of bone substitutes[12,13]. Furthermore, due to the misuse of antibiotics for years, drug-resistant bacteria have also become another challenge[14,15]. Metal ions such as Ag+, Zn2+, and Cu2+ have shown the ability to kill bacteria and are unlikely to cause mutations in pathogenic bacteria[16,17]. Therefore, the doping of appropriate content of metal elements to scaffolds to kill bacteria in situ became the most adopted antibacterial strategy at present. In particular, it has been proven that Zn2+ ions not only possessed reliable antibacterial activity, but also could promote osteoblastogenesis[18,19]. However, once the release rate of Zn2+ ions exceeds the safety limits, the scaffolds will become toxic to normal cells such as osteocyte[20]. Some researchers gave priority to osteogenic activity at the expense of antibacterial effect[21,22]. The others studied the addition of other substances to offset the cytotoxicity of Zn2+[23,24]. Considering the fact that the perioperative period is associated with high incidence of infection[25,26], initial antibacterial property is more important than long- term antibacterial property. On that basis, a balanced strategy is proposed in this paper, which is to control the release rate of Zn2+ from scaffolds by preparing coatings with different Zn2+ contents layer by layer, so as to increase the killing rate of bacteria in the initial stage after implant surgery and to strengthen the osteogenic property in the later stage. Moreover, to obtain such multilayer coating, polymer is chosen as the most appropriate material. Chitosan and sodium alginate, as biodegradable natural polymers, have been widely developed in bone tissue engineering domains due to outstanding biocompatibility and bioabsorbability[27,28]. The positively charged amino group in chitosan can generate electrostatic interaction with the negatively charged carboxyl group in sodium alginate. Therefore, multilayer polymer coatings can be constructed by layer-by-layer method[29,30]. Besides, sodium alginate can be crosslinked with divalent metal ions to gelation, and the metal ions are stably loaded into the multilayer coatings[31]. In this study, porous hydroxyapatite (HA) bioceramic scaffolds was prepared by digital light processing (DLP) method, and chitosan/sodium alginate polymer coatings were added to scaffolds by layer-by-layer method. Next, Zn2+ ions were doped where the ion concentrations were changed layer by layer to achieve a stable and controllable release. Finally, the balance between antimicrobial properties and biological properties was achieved. 2. Materials and methods 2.1. Preparation of porous scaffolds The detailed preparation process of HA porous scaffolds is described as follows. The first step was to prepare the slurry. HA powders (Aladdin, China) and photosensitive resin (Shanghai, China) were mixed at a mass ratio of 1:1.1, with 3 wt.% of dispersant (Shanghai, China). All raw materials were put into a vacuum disperser, and the slurry was obtained after stirring at 1100 rpm for 30 min. Next, a self-developed DLP printer was used to fabricate the green bodies. The porous scaffold model was designed using Magics 23.0 software (the average pore size was ~1000 μm). According to the experimental requirements, the larger cuboid scaffolds (9 × 9 × 13.5 mm3) were prepared for mechanical test, the smaller cuboid scaffolds (9 × 9 × 4.5 mm3) were prepared for biological test, and the cylindrical scaffolds (Ф4.5 × 4.5 mm3) were prepared for inhibition zone test. After washing and drying, the ceramic green bodies were debinded and sintered at 1500°C for 3 h; detailed temperature curves were shown in the previous research[32]. The fabrication process is shown in Figure 1. 2.2. Preparation of coatings The chitosan (Aladdin, China) was dissolved in 0.2% acetic acid (Aladdin, China) solution with the concentration of 2  g/L. The sodium alginate (Aladdin, China) was dissolved in deionized water with the concentration of 2  g/L. Meanwhile, four different crosslinking solutions containing increasing concentrations of ZnCl2 solution (0,  0.25, 0.5, and 1  g/L) were prepared, and CaCl2 of variable concentration was added to adjust the total metal ion concentration to 2 g/L. Subsequently, the HA scaffolds were immersed in the alginate solution, chitosan solution, and one kind of crosslinking solution for 5 min. After each soaking process, the scaffolds were cleaned with distilled Volume 9 Issue 2 (2023) 293 https://doi.org/10.18063/ijb.v9i2.668 Zn-doped coatings with osteogenic and antibacterial propertiesInternational Journal of Bioprinting water to remove the uncrosslinked solution and then placed in a conventional oven at 60°C for 2 min. Then, the above-mentioned soaking process was repeated until all the coatings were prepared. The coating process is shown in Figure 2. Four kinds of scaffolds with different coatings were named in accordance with the content of Zn2+, namely, CHA-0, CHA-L, CHA-M, and CHA-H, respectively. Bare scaffold for comparison was named HA. Another specific group, denoted as CHA-G, with gradient coatings was also prepared, in which the Zn2+ concentrations of the inner two layers, the middle two layers and the outer two layers were 0.25 g/L, 0.5 g/L, and 1 g/L. The default number of coating layers was 6. Figure  1. Schematic diagram of porous ceramic scaffold digital light processing additive manufacturing. 2.3. Characterization of scaffolds The scanning electron microscope (SEM, SU5000, Hitachi Instruments, Japan) was used to analyze the surface and cross section morphologies of the samples. The energy dispersive X-ray (EDX, ULTIMATELYMAX40, Oxford, UK) was used to detect elemental distribution of the coatings. The SEM images of HA scaffold surfaces were statistically analyzed by ImageJ software, and the particle size of HA powders was analyzed with laser granularity analytical instrument (MS2000, Malvin, UK). The X-ray diffraction spectra (XRD) of HA powders and HA scaffolds were obtained on the SmartLab (Rigaku, Japan). The chemical composition and chemical state of the coatings were measured by X-ray photoelectron spectroscopy (XPS; ESCA Xi+, ThermoFischer, USA), and the Avantage software was used for data processing. A  universal tester (Zwick-Z250, Germany) was used to test the compressive strength of bare scaffolds and the coated scaffolds, and the crosshead loading speed was 0.1 mm/min. The compressive modulus was obtained from the stress and strain curves[33]. 2.4. In vitro degradation assays The HA, CHA-0, and CHA-H samples were immersed in simulated body fluid (SBF, pH=7.4; Scientific Phygene, China) at 37°C, and the ratio of scaffold mass to solution volume was 5 g/L. After 14 days, the apatite deposition on the scaffolds was observed by SEM. To evaluate the effect of coating on degradation property of scaffolds, the HA, Figure 2. Formation and reaction mechanism of multilayer coatings containing Zn2+. Volume 9 Issue 2 (2023) 294 https://doi.org/10.18063/ijb.v9i2.668 Zn-doped coatings with osteogenic and antibacterial propertiesInternational Journal of Bioprinting CHA-M, CHA-M (with 3-layer coating), and CHA-M (with 9-layer coating) were soaked in Tris-HCl buffer solution (pH 7.4; Scientific Phygene, China) with a scaffold mass to solution volume was 5 g/L at 37°C for 1, 2, 3, and 4 weeks. The solution was replaced with fresh tris-HCl solution every 7 days. After a predetermined soaking period, each sample was completely dried and weighed (Wb). The relative mass loss at time b was calculated as: Weight loss = (Wa-Wb)/Wa × 100%. Wa was the initial mass of the scaffold. A pH meter was used to measure the pH value in solution. 2.5. Bioactivity in vitro assessment line MC3T3-E1 Mouse calvarial pre-osteoblast cell (MC3T3-E1; Institute of Life Science Cell Culture Center, Shanghai, China) was used in the experiment. The cell growth medium was AMEM (Wisent, Canada) containing 10% fetal bovine serum (Gibco, USA) and 1% penicillin/ streptomycin (Wisent, Canada). After being resuscitated from frozen storage, the cells were cultured in incubator at 37°C with 5% CO2. The medium was replaced every 2 – 3 days. 2.5.1. Cytotoxicity test by live/dead staining Before the test, all samples were sterilized by autoclave boiler (Shenan, Shanghai, China). Then, pre-osteoblasts (5 × 104 cells/mL) were inoculated directly on the scaffolds. After 4 days of incubation, the scaffolds were stained with Live/Dead Double Staining Kit (Beyotime, China), and then, the staining results were observed under a confocal laser scanning microscope (LSM710, Zeiss, Germany). 2.5.2. Cell proliferation and attachment Cells (5 × 104 cells/well) were seeded onto scaffolds in 24-well plates. After culturing for 1, 4, and 7 days, the samples were washed three times with phosphate-buffered saline (PBS; Wisent, Canada), followed by 10% Cell Counting Kit-8 solution (CCK-8, diluted with medium; Beyotime, China), which was used for replacing the original medium and incubation for another 2  h. The cells grown in the blank well plate were defined as the control. Then cell viability was measured on a microplate reader (Multiskan FC, Thermo Fisher Scientific, USA) at 450  nm. To examine the cell morphology on scaffolds, the scaffolds were disposed by DAPI (4 ,6-diamidino-2-phenylindole, Beyotime, China) and Actin-Tracker Green (Beyotime, China) after 4  days of co-culture. Fluorescent images of cells on scaffolds were observed under a confocal laser scanning microscope. ′ 2.5.3. Cell differentiation The cells (104  cells/mL) were cultured with HA, CHA-0, CHA-L, CHA-M, CHA-H, and CHA-G in 12-well plate. After 14  days of culture, the alkaline phosphatase assay kit (Beyotime, China) was used to measure alkaline phosphatase (ALP) activity, and the optical density (OD) value was measured with a microplate reader at 405 nm. Cellular samples were stained with BCIP/NBT Alkaline Phosphatase Color Development Kit (Beyotime, China), after being fixed with 4% paraformaldehyde for 20 min. After 21 days of incubation, the samples were stained using the Alizarin Red S Staining Kit for Osteogenesis (Beyotime, China) for 30  min, followed by a rinse with distilled water. Then, an inverted optical microscope (CKX53, Olympus, Japan) was used to observe the stained cells. Afterwards, the 10% (w/v) cetylpyridinium chloride (diluted with PBS; Aladdin, China) was used to quantify the Alizarin red staining (ARS). The absorbance was measured at 562 nm. In addition, the amount of Zn2+ released from Zn-doped samples in the collected medium was detected using commercial kits (Zinc, China). 2.6. Antibacterial assay The antibacterial properties of HA, CHA-0, CHA-L, CHA- M, and CHA-H were assessed against Gram-negative Escherichia coli (E. coli, ATCC 25922) and Gram-positive Staphylococcus aureus (S. aureus, ATCC 29213). Before the test, the bacteria on logarithmic phase were dispersed into PBS, and the concentration of bacterial suspension was adjusted using a microplate reader (108 CFU/mL; OD600=0.1). Disk diffusion method was used to qualitatively evaluate antibacterial properties of samples. 100 μL of the bacterial suspension (107 CFU/mL) were equably spread over a 90 mm agar plate, followed by dispersedly placing the discoid samples onto the plate and culturing in an incubator with a humidity of about 90% at 37°C for 24 h. Afterward, the zone of inhibition was photographed. Plate colony-counting method was applied to quantificationally evaluate the antibacterial activities. The samples were exposed to bacterial suspension (106 CFU/mL) and then incubated on a shaker for 2 h and 8 h at 37°C. The group without scaffold was considered a control. After the appropriate proportion of dilution, the bacterial suspension was inoculated onto solid medium and then cultured until colonies were formed. The antibacterial efficiency was computed based on following equation: Antibacterial efficiency (%) = (Ac – As)/Ac × 100% Where Ac and As are the bacterial colony numbers of control and sample groups, respectively. 2.7. Statistical analysis The data were processed using Origin 2018 software (Originlab, USA). Values were expressed as means ± Volume 9 Issue 2 (2023) 295 https://doi.org/10.18063/ijb.v9i2.668 Zn-doped coatings with osteogenic and antibacterial propertiesInternational Journal of Bioprinting standard deviation (SD), and significance of data was determined by P < 0.05. Besides, the data with probability less than 0.05 (P < 0.05), 0.01 (P < 0.01), and 0.001 (P < 0.001) were represented by *, **, and ***, respectively. 3. Results and discussion 3.1. Characterization of scaffolds 3.1.1. Surface morphology The comparison of unsintered and sintered porous ceramic scaffold is shown in Figure 3A. After sintering, the photosensitive resin was removed completely, resulting in a considerable volume contraction of the porous scaffold, and the average linear shrinkage rate was 30%. The XRD patterns of the ceramic powder and bare scaffolds are displayed in Figure  3B. The main component of the scaffold was HA (# 70-0566), and no significant phase transition occurred during sintering. Compared with the size distribution of raw HA powders (Figure 3C), the particle size of HA scaffold surface mainly distributed from 10 to 15 μm (Figure  3D), which indicated that the grain grew normally. In addition, the SEM results of surface micrographs of HA scaffold showed that the grains were compacted (Figure 3D).”. As for the morphology of coating, the coating materials penetrated into the gaps between ceramic particles and a smoother surface was then obtained (Figure 3E). The cross-section morphology of the scaffold with coatings is shown at the top right corner of Figure 3E. A  multilayered structure was clearly displayed, and the total thickness of the coating was 6.7 μm, from which it could be inferred that the thickness of single coating was about 0.74 μm. In addition, the change of Zn concentration in the coating had no effect on its morphology. Finally, the EDS mapping of Ca and Zn elements in the coating of the CHA-H scaffold confirmed that Zn2+ was homogeneously distributed (Figure 3G and 3H). 3.1.2. XPS characterization To investigate the detailed chemical composition of the coatings, the XPS analyses were performed. Figure  4A shows the total XPS spectra of CHA-0 and CHA-H groups, which were typical among all groups. The elements contained in the raw material, such as C, N, O, Ca, and Zn, could be detected in CHA-H, and all of these elements A B C D E F G H Figure 3. (A) Images of unsintered and sintered hydroxyapatite (HA) scaffolds. (B) X-ray diffraction spectra of HA powder and scaffold. (C) Particle size distribution of ceramic powders. (D) Scanning electron microscope (SEM) image of HA scaffold; the inset image shows the size distribution of ceramic particle on the scaffold surface. (E) SEM image of HA scaffold with coatings; the inset image shows the cross section of coatings. (F) High-magnification SEM image of coatings. (G and H) EDS mapping of Ca and Zn on CHA-H sample. Volume 9 Issue 2 (2023) 296 https://doi.org/10.18063/ijb.v9i2.668 Zn-doped coatings with osteogenic and antibacterial propertiesInternational Journal of Bioprinting A D B E C F Figure 4. (A) The survey XPS spectra of CHA-0 and CHA-H. (B) O 1s spectra, A – D: CHA-0, CHA-L, CHA-M, and CHA-H, respectively. (C) C 1s and (D) N 1s spectra of CHA-0 and CHA-H, respectively. (E) Ca 2p spectra and (F) Zn 2p spectra of CHA-H. were also detected from CHA-0, except Zn. N  was only ascribed to chitosan in the coating, so the content was low and the spectral peak was not obvious. Furthermore, CaCl2 and ZnCl2 were employed as crosslinking agent, in which divalent metal ions interact strongly with carboxyl groups of sodium alginate. However, there were no groups in the coating that could interact with Cl-  ions, so its content was almost negligible. Similarly, the Na+ ions in sodium alginate were replaced by Ca2+ and Zn2+ ions, resulting in an insignificant spectral peak. The O 1s XPS spectra (Figure  4B) were deconvoluted into 4 peaks near 533.5, 532.7, 531.5, and 531 eV, corresponding to C=O, C-OH, C-O-C, and O…Zn/Ca[34]. The peak of O…Zn/Ca shifted rightward with the increase of Zn concentration in coatings caused by the different interaction forces of Zn2+ and Ca2+ on carboxyl groups, and the detailed values for CHA-0, CHA-L, CHA-M, and CHA-H groups were 531.1, 530.95, 530.85, and 531.7 eV, respectively. The C 1s spectra are shown in Figure  4C. They were deconvoluted into three principal peaks near 288.2, 286.6 and 284.8 eV, which were assigned O-C=O, C-O/C-N, and C-C bonds, respectively[35,36]. Figure  4D showed that the N 1s spectra and the peaks near 399.8 eV were attributed to amidogen groups (-NH3+)[37]. CHA-0 and CHA-H were selected because they were most likely to differ, and in fact all samples had similar C 1s and N 1s spectra. In Figure  4E and 4F, the binding energies of 351.4 and 347.8 eV were classified to the Ca 2p1/2 and Ca 2p3/2 orbitals of the Ca2+ in CHA-0 group, while 1045.1 and 1022 eV were classified to the Zn 2p1/2 and Zn 2p3/2 orbitals of the Zn2+ in CHA-H group[38]. The relative atomic ratios of Zn and Ca elements on the coating surface are given in Table  1. There were some acceptable differences between experimental and theoretical values caused by experimental errors. Therefore, the concentration of Zn ions in the coating could be adjusted by changing the proportion of Ca/Zn ions in the solution without changing the concentration of the crosslinking agent solution. 3.2. Mechanical properties of samples To evaluate the effect of coating on mechanical properties of scaffolds, the compressive strength and compressive modulus values of the bare scaffold and coated scaffold are compared. As shown in Figure 5, compressive strength of bare scaffold was 10.42 ± 0.56 MPa, while that of the coated scaffold was 11.52 ± 0.3 MPa. These results suggested that the coating could improve the mechanical properties of scaffolds, although the improvement was not significant. The reasons could be attributed to two components. On the one hand, although the mechanical strength of the polymer coating material was extremely poor, it could still play a certain supporting role under pressure[39]. On Volume 9 Issue 2 (2023) 297 https://doi.org/10.18063/ijb.v9i2.668 Zn-doped coatings with osteogenic and antibacterial propertiesInternational Journal of Bioprinting the other hand, it was inevitable that some defects would appear on the scaffold surface during the preparation process. The addition of the adhesive coating properly filled these defects, thus avoiding the premature collapse of scaffolds. 3.3. Degradation properties of samples Numerous studies have shown that HA scaffolds could induce apatite deposition on the surface in SBF, which was related to the bioactivity properties of scaffolds[40,41]. The HA, CHA-0, and CHA-H scaffolds were soaked in the SBF for 14  days at 37°C, and then the surface morphologies were observed by SEM. As showed in Figure 6A, coralloid deposits were formed on the surface of the HA scaffold. The EDS results (Figure  6B) confirmed that the deposits were calcium phosphate, and the calcium to phosphorus ratio was 1.37. Studies have suggested that calcium phosphate could be classified as HA when the Ca/P ratio was between 1.3 and 2.0[42]. As for the scaffold with coating (Figure  6C–F), the apatite deposits on the surface were less and appeared as spherical clusters, indicating that the addition of coating inhibited the deposition of apatite to a certain extent. In addition, the EDS results showed that the Ca/P ratio of the coated scaffold was significantly higher than that of the HA scaffold, which can be attributed to the fact that the penetration depth of EDS detection was too Table 1. Planned and experimentally obtained molar ratios of Ca:Zn in coatings. Sample Planned (in the crosslinking agent solution) XPS elemental analysis (in samples) CHA-0 CHA-H CHA-M CHA-L Mass ratio of CaCl2:ZnCl2 0:1 1:1 1:3 1:7 A Molar ratio of Ca2+:Zn2+ Molar ratio of Ca:Zn 0:1 1:1.23 1:3.68 1:8.58 0:1 1:1.31 1:3.48 1:6.73 B large to be interfered by the elements of the sample itself. In particular, it was interesting to note that the Ca/P ratio of CHA-H was higher than that of CHA-0, even though the calcium content in the coating of CHA-H group was lower than that of CHA-0 group. It indicated that the higher the Zn content in the coating, the higher the Ca/P ratio in the induced apatite deposition layer, which may be attributed to the complex nucleation mechanism of apatite[43]. Moreover, the surface morphology of CHA-0 and CHA-H samples also indicated that the coating with higher Zn concentration might have a stronger ability to induce apatite deposition. To evaluate the effect of coatings on the degradation of scaffolds, the bare HA scaffold and the CHA-M scaffolds with different number of coating layers (3, 6, and 9 layers, respectively) were immersed in Tris-HCl for 7, 14, 21, and 28 days. The degradation curves are shown in Figure 7A, The HA scaffold had the highest mass loss of 2.93% among all groups after 4  weeks. For coated scaffolds, the mass loss rate showed a dependence on the number of coating layers. As the number of coating layers increased, the weight loss rate of scaffolds was slower, showing a delayed effect. Consequently, the addition of coating inhibited the degradation of scaffolds. Figure  7B showed the variation of pH values of Tris- HCl solution after soaking the samples. During the whole soaking process, except for the CHA-M (9), the pH value of the other groups decreased first and then tended to be stable, but the reasons were not exactly the same. For the HA scaffold, its degradation products were weakly alkaline, while the deposition of apatite would consume the hydroxide ions in the solution, leading to a decrease of pH value[44,45]. Consequently, the pH value of the immersion solution showed a state of fluctuation. As for the coated scaffolds, in addition to the effect caused by apatite deposition, the decrease of the pH value of the solution might be due to the trace residue of acetic acid introduced Figure 5. (A) Compression stress-strain curves of the scaffolds. (B) Compressive strength and compressive modulus of the scaffolds. Volume 9 Issue 2 (2023) 298 https://doi.org/10.18063/ijb.v9i2.668 Zn-doped coatings with osteogenic and antibacterial propertiesInternational Journal of Bioprinting A B C D E F Figure 6. Scanning electron microscope images of mineral deposited on (A) hydroxyapatite (HA), (C) CHA-0, and (E) CHA-H scaffolds in SBF for 14 days. EDS analyses of the surface sediment of (B) HA, (D) CHA-0, and (F) CHA-H scaffolds are also shown. A B Figure 7. Weight loss (A) and pH value (B) of scaffolds after soaking in Tris-HCl for different durations (the solution refreshed every week). during the preparation of the coating. Although the pH value of the solution after soaking the bare HA scaffold was on average higher than that of coated scaffolds, these differences were not significant. Overall, the coating had little effect on the pH of body fluid environment. In particular, the pH values of all groups were always higher than 7, showing weak alkalinity, which has been proven to be beneficial to the growth of osteoblasts[46]. 3.4. In vitro biocompatibility properties Although the addition of the coating inhibited apatite formation, it was not the sole criterion for biological activity. The viability of MC3T3-E1 cells on days 1, 4, and 7 was evaluated by CCK-8 assay. As shown in Figure 8, the cell proliferation of all groups (except CHA-H) was better than that of blank control group in a time-dependent manner, indicating good biocompatibility of the scaffolds. Furthermore, at each detection node, their proliferative capacities in decreasing order were: CHA-M, CHA-L, HA, and CHA-0 groups. The low cell activity of CHA-H group was caused by high concentration of Zn2+ ion, which was toxic to cells. To further verify biocompatibility of scaffolds, the cell physiological status and survival was observed using fluorescence microscope on day 4. In Figure  9A, the cytoskeleton was displayed as green spots and the nucleus was shown as blue spots. The filopodia and pseudopodia extension of the cell could be conformably observed on all scaffolds except CHA-H, which indicate excellent growth of cells[47]. The cells on CHA-H scaffold surface were smaller and not extended enough. Furthermore, the quantity of cells in each group shown by fluorescence images was basically in accordance with the data of CCK-8 assay. To explore the damage on cells caused by high Zn2+ concentration, calcein and propidium iodide staining was performed. As shown in Figure 9B, living cells were green Volume 9 Issue 2 (2023) 299 https://doi.org/10.18063/ijb.v9i2.668 Zn-doped coatings with osteogenic and antibacterial propertiesInternational Journal of Bioprinting and dead cells were red. It was found that although the cells on CHA-H scaffold were less due to its low proliferative potential, only a few of them died. For other groups, the dead cells were almost invisible. 3.5. Osteogenesis differentiation Osteoblasts go through proliferation, differentiation and mineralization, and eventually become mature osteocytes[48]. Therefore, osteogenic differentiation is the key stage of bone regeneration. ALP activity was related to the early osteogenic differentiation[49], and the capacity of cell mineralization, which was evaluated by Alizarin red staining assays, was a marker of late osteogenic differentiation[50]. ALP staining results (purple stain) displayed that ALP was detected in all experimental groups (Figure 10A), and the results of ALP quantitative analysis are presented in Figure  10B. Among all groups, cells inoculated on the CHA-M samples showed the highest level of ALP expression. Interestingly, the ALP activity of CHA-G group was comparable to that of CHA-L group, second only to the CHA-M group. This result corroborated that Zn2+ ions released by CHA-H were not sufficient to kill the cells. Therefore, the cell viability was recovered with the decrease of Zn2+ ions concentration in the late culture stage. Furthermore, ARS staining and quantitative analysis (Figure 10C and 10D) showed a similar trend as noted in the ALP assay, confirming that Zn-coated scaffolds had better osteogenic potential than bare scaffolds. Finally, Zn2+ ions concentration in medium incubated with different samples were investigated, and the results are shown in Figure  11. CHA-H, CHA-M, and CHA-L groups could basically maintain steady release rate of Zn2+ ions over time, while CHA-G group showed a marked decreasing tendency, suggesting that multilayer coatings are feasible to control the release of Zn2+ ions. The results of biological experiments indicated that chitosan and sodium alginate coating had good biocompatibility and were non-toxic to cells. The composite polymer had properties similar to extracellular matrix, which was beneficial to cell adhesion. In addition, Zn2+ ions were doped into coatings in the form of interaction, ensuring the steady release of ions, and the release rate could be arbitrarily controlled by means of preparing the coating layer by layer. As a trace element in human body, Zn2+ ions were conducive to the proliferation and differentiation of osteoblasts at appropriate concentration. Although high Zn2+ concentration of ions could reduce cell viability, the fact that it could induce stronger antibacterial activity should warrant more investigation. 3.6. Antibacterial activity Figure 8. Cell viability of MC3T3-E1 cells after 1, 4, and 7 days of co-culture with samples. Bacterial infection represents one the major factors of transplant failure. Therefore, research on antibacterial A B Figure 9. Fluorescent images of (A) phalloidin/DAPI and (B) calcein/PI staining of cells incubated with different scaffolds for 4 days. Volume 9 Issue 2 (2023) 300 https://doi.org/10.18063/ijb.v9i2.668 Zn-doped coatings with osteogenic and antibacterial propertiesInternational Journal of Bioprinting Figure 10. (A) Alkaline phosphatase (ALP) staining of cells co-cultured with different scaffolds for 14 days, A1 – A7: control, HA, CHA-0, CHA-H, CHA- M, CHA-L, and CHA-G, respectively. (B) Quantitative analysis of ALP activity. (C) ARS staining of cells co-cultured with different scaffolds for 21 days, C1 – C7: control, HA, CHA-0, CHA-H, CHA-M, CHA-L, and CHA-G, respectively. (D) Quantitative analysis of mineralized rate. reactive oxygen species, thereby avoiding the deleterious effects of antimicrobial resistance[51]. E. coli and S. aureus, as the representative of two major categories of bacteria, were chosen to evaluate the antibacterial activity of scaffolds. As shown in Figure 12A, after being exposed to bacterial suspension for 2  h, the inhibition rate of Zn-doped scaffolds showed a dependence on the Zn2+ concentration, which were 87.2, 70.8, and 47.8% against E. coli, and 73.2, 59.6, and 39.6% against S. aureus (the colony count results are shown in Figure  11E and 11F). The HA and CHA-0 scaffold had almost no antibacterial properties, indicating that the antibacterial effect of scaffolds was mainly attributed to Zn2+ ions. Furthermore, bacterial mortality rates of all Zn-doped groups increased with the treatment time extended to 8  h, and the rate even reached 99.4% against E. coli and 93% against S. aureus in CHA-H group. For disk diffusion experiment (Figure  12C and 12D), obvious inhibition zone was observed around the CHA-H Figure 11. Concentration of Zn2+ ions released in cell culture media. scaffolds has been gaining popularity and attention in recent years. Zn2+ ions could kill bacteria by inducing Volume 9 Issue 2 (2023) 301 https://doi.org/10.18063/ijb.v9i2.668 Zn-doped coatings with osteogenic and antibacterial propertiesInternational Journal of Bioprinting A B C D E F Figure 12. The growth of bacteria on agar plates after being co-cultured with scaffolds for (A) 2 h and (B) 8 h. Inhibition zone of different samples to (C) Escherichia coli and (D) Staphylococcus aureus was determined by disk diffusion assay. The antibacterial rate against (E) E. coli and (F) S. aureus was obtained by counting the number of colonies. scaffold. To sum up, the scaffolds with Zn-doped coating showed great potential to inhibit bacterial infections in tissue engineering. 4. Conclusion In this paper, we propose a universal, convenient, and reliable preparation method of multilayer coating. The EDS and XPS results demonstrated that as a part of cross-linking agent, Zn2+ ions were stably and uniformly doped into the polymer coating, and the content of Zn2+ in the coating was subject to the ratio of Ca2+/Zn2+ in the crosslinking agent solution. Besides, the coating had little effect on the compressive strength of the scaffold but inhibited the degradation of scaffold to some extent. In vitro studies showed that CHA-M group with moderate Zn2+ release rate had the best biocompatibility and osteoinductivity. In contrast, the CHA-H group showed the worst biological activity, but excessive Zn2+ concentration did not kill the cells completely according to the results of live/dead cell staining. In addition, CHA-H scaffold exhibited outstanding antibacterial activity against E. coli and S. aureus. Therefore, coatings containing gradient Zn concentrations were specially prepared to reconcile the contradiction between the osteogenic and antimicrobial Volume 9 Issue 2 (2023) 302 https://doi.org/10.18063/ijb.v9i2.668 Zn-doped coatings with osteogenic and antibacterial propertiesInternational Journal of Bioprinting together, the multilayer coatings properties. Taken prepared by layer-by-layer method could realize the controllable release of Zn2+ and balance the osteogenic and antimicrobial activity, indicating its promising potential in bone tissue engineering. Acknowledgments bone tissue regeneration. Acta Biomater, 8: 3191–3200. https://doi.org/10.1016/j.actbio.2012.06.014 3. Arrigoni C, Gilardi M, Bersini S, et al., 2017, Bioprinting and organ-on-chip applications towards personalized medicine for bone diseases. Stem Cell Rev Rep, 13: 407–417. https://doi.org/10.1007/s12015-017-9741-5 The authors extend their sincere gratitude to those who contributed in instructions and experiments. 4. Koons GL, Diba M, Mikos AG, 2020, Materials design for bone-tissue engineering. Nat Rev Mater, 5: 584–603. Funding This work was supported by Jiangsu Provincial Key Research and Development Program (No. BE2019002). Conflict of interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Author contributions Conceptualization: Zhijing He, Lida Shen Data curation: Chen Jiao Formal analysis: Zongjun Tian Investigation: Changjiang Wang, Qing Jiang Methodology: Zhijing He, Junnan Wu, Jiasen Gu, Huixin Liang Supervision: Lida Shen, Zongjun Tian Project administration: Lida Shen Validation: Zhijing He, Chen Jiao, Youwen Yang, Changjiang Wang Visualization: Chen Jiao Writing – original draft: Zhijing He Writing – review & editing: Youwen Yang Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Availability of data The datasets used and analyzed during the present study can be obtained from the corresponding author on request. References 1. Cheng H, Chawla A, Yang Y, et al., 2017, Development of nanomaterials for bone-targeted drug delivery. Drug Discov Today, 22: 1336–1350. https://doi.org/10.1016/j.drudis.2017.04.021 2. Ferreira AM, Gentile P, Chiono V, et al., 2012, Collagen for https://doi.org/10.1038/s41578-020-0204-2 5. Garcia P, Histing T, Holstein JH, et al., 2013, Rodent animal models of delayed bone healing and non-union formation: A comprehensive review. Eur Cells Mater, 26: 1–12. https://doi.org/10.22203/ecm.v026a01 6. Turnbull G, Clarke J, Picard F, et al., 2018, 3D bioactive composite scaffolds for bone tissue engineering. Bioact Mater, 3: 278–314. https://doi.org/10.1016/j.bioactmat.2017.10.001 7. Aulakh TS, Jayasekera N, Kuiper JH, et al., 2009, Long- term clinical outcomes following the use of synthetic hydroxyapatite and bone graft in impaction in revision hip arthroplasty. Biomaterials, 30: 1732–1738. https://doi.org/10.1016/j.biomaterials.2008.12.035 8. Jeong J, Kim JH, Shim JH, et al., 2019, Bioactive calcium phosphate materials and applications in bone regeneration. Biomater Res, 23: 4. https://doi.org/10.1186/s40824-018-0149-3 9. Zeng JH, Liu SW, Xiong L, et al., 2018, Scaffolds for the repair of bone defects in clinical studies: A  systematic review. J Orthop Surg Res, 13: 33. https://doi.org/10.1186/s13018-018-0724-2 10. Mahanty A, Shikha D, 2022, Changes in the morphology, mechanical strength and biocompatibility of polymer and metal/polymer fabricated hydroxyapatite for orthopaedic implants: A review. J Polym Eng, 42: 298–322. https://doi.org/10.1515/polyeng-2021-0171 11. Vilardell AM, Cinca N, Concustell A, et al., 2015, Cold spray as an emerging technology for biocompatible and antibacterial coatings: State of art. J Mater Sci, 50: 4441–4462. https://doi.org/10.1007/s10853-015-9013-1 12. Przekora A, 2019, Current trends in fabrication of biomaterials for bone and cartilage regeneration: Materials modifications and biophysical stimulations. Int J Mol Sci, 20: 435. https://doi.org/10.3390/ijms20020435 13. Li F, Li S, Liu Y, et al., 2022, Current advances in the roles of doped bioactive metal in biodegradable polymer composite scaffolds for bone repair: A  mini review. Adv Eng Mater, Volume 9 Issue 2 (2023) 303 https://doi.org/10.18063/ijb.v9i2.668 Zn-doped coatings with osteogenic and antibacterial propertiesInternational Journal of Bioprinting 24: 2101510. https://doi.org/10.1002/adem.202101510 14. Sierra MA, Casarrubios L, de la Torre MC, 2019, Bio- organometallic derivatives of antibacterial drugs. Chemistry, 25: 7232–7242. https://doi.org/10.1002/chem.201805985 15. Li B, Webster TJ, 2018, Bacteria antibiotic resistance: New challenges and opportunities for implant-associated orthopedic infections. J Orthop Res, 36: 22–32. https://doi.org/10.1002/jor.23656 16. Godoy-Gallardo M, Eckhard U, Delgado LM, et al., 2021, Antibacterial approaches in tissue engineering using metal ions and nanoparticles: From mechanisms to applications. Bioact Mater, 6: 4470–4490. https://doi.org/10.1016/j.bioactmat.2021.04.033 17. Saidin S, Jumat MA, Amin NA, et al., 2021, Organic and inorganic antibacterial approaches in combating bacterial infection for biomedical application. Mater Sci Eng C Mater, 118: 111382. https://doi.org/10.1016/j.msec.2020.111382 18. Bhattacharjee A, Gupta A, Verma M, et al., 2019, Site-specific antibacterial efficacy and cyto/hemo-compatibility of zinc substituted hydroxyapatite. Ceram Int, 45: 12225–12233. https://doi.org/10.1016/j.ceramint.2019.03.132 19. Yang Y, Zan J, Shuai Y, et al., 2022, In Situ growth of a metal-organic framework on graphene oxide for the chemo- photothermal therapy of bacterial infection in bone repair. ACS Appl Mater Inter, 14: 21996–22005. https://doi.org/10.1021/acsami.2c0484121996 20. Wätjen W, Haase H, Biagioli M, et al., 2002, Induction of apoptosis in mammalian cells by cadmium and zinc. Environ Health Persp, 110: 865–867. https://doi.org/10.1289/ehp.110-1241262 21. Lu T, Yuan X, Zhang L, et al., 2021, High throughput synthesis and screening of zinc-doped biphasic calcium phosphate for bone regeneration. Appl Mater Today, 25: 101225. https://doi.org/10.1016/j.apmt.2021.101225 22. Shen J, Chen B, Zhai X, et al., 2021, Stepwise 3D-spatio-temporal magnesium cationic niche: Nanocomposite scaffold mediated microenvironment intramembranous for modulating ossification, Bioact Mater, 6: 503–519. https://doi.org/10.1016/j.bioactmat.2020.08.025 23. Ullah I, Siddiqui MA, Kolawole SK, et al., 2020, Synthesis, characterization and in vitro evaluation of zinc and strontium binary doped hydroxyapatite for biomedical application. Ceram Int, 46: 14448–14459. https://doi.org/10.1016/j.ceramint.2020.02.242 24. Wang Q, Tang P, Ge X, et al., 2018, Experimental and simulation studies of strontium/zinc-codoped hydroxyapatite porous scaffolds with excellent osteoinductivity and antibacterial activity. Appl Surf Sci, 462: 118–126. https://doi.org/10.1016/j.apsusc.2018.08.068 25. Campoccia D, Montanaro L, Arciola CR, 2006, The significance of infection related to orthopedic devices and issues of antibiotic resistance. Biomaterials, 27: 2331–2339. https://doi.org/10.1016/j.biomaterials.2005.11.044 26. Gasik M, 2017, Understanding biomaterial-tissue interface quality: Combined in vitro evaluation. Sci Technol Adv Mat, 18: 550–562. https://doi.org/10.1080/14686996.2017.1348872 27. Islam MM, Shahruzzaman M, Biswas S, et al., 2020, Chitosan based bioactive materials in tissue engineering applications-A review. Bioact Mater, 5: 164–183. https://doi.org/10.1016/j.bioactmat.2020.01.012 28. Aguero L, Alpdagtas S, Ilhan E, et al., 2021, Functional role of crosslinking in alginate scaffold for drug delivery and tissue engineering: A review. Eur Polym J, 160: 110807. https://doi.org/10.1016/j.eurpolymj.2021.110807 29. Huang C, Fang G, Zhao Y, et al., 2019, Bio-inspired layer-by-layer coating of chitosan/ nanocomposite by hyaluronic acid multilayers on a hard nanocellulose- hydroxyapatite matrix. Carbohyd Polym, 222: 115036. https://doi.org/10.1016/j.carbpol.2019.115036 30. Wang R, Sun L, Zhu X, et al., 2022, Carbon nanotube‐based strain sensors: Structures, fabrication, and applications. Adv Mater Technol, 2022: 2200855. https://doi.org/10.1002/admt.202200855 31. Hernandez-Gonzalez AC, Tellez-Jurado L, Rodriguez- Lorenzo LM, 2020, Alginate hydrogels for bone tissue engineering, from injectables to bioprinting: A  review. Carbohyd Polym, 229: 115514. https://doi.org/10.1016/j.carbpol.2019.115514 32. Jiao C, Xie D, He Z, et al., 2022, Additive manufacturing of Bio-inspired ceramic bone Scaffolds: Structural Design, mechanical properties and biocompatibility. Mater Design, 217: 110610. https://doi.org/10.1016/j.matdes.2022.110610 33. Liang H, Yang Y, Xie D, et al., 2019, Trabecular-like Ti-6Al-4V scaffolds for orthopedic: Fabrication by selective laser melting and in vitro biocompatibility. J  Mater Sci Technol, 35: 1284–1297. https://doi.org/10.1016/j.jmst.2019.01.012 34. He H, Lian J, Chen C, et al., 2022, Enabling multi- chemisorption sites on carbon nanofibers cathodes by an in-situ exfoliation strategy for high-performance Zn-Ion Volume 9 Issue 2 (2023) 304 https://doi.org/10.18063/ijb.v9i2.668 Zn-doped coatings with osteogenic and antibacterial propertiesInternational Journal of Bioprinting hybrid capacitors. Nanomicro Lett, 14: 106. https://doi.org/10.1007/s40820-022-00839-z 35. Khan MU, Razak SI, Ansari MN, et al., 2021, Development of biodegradable bio-based composite for bone tissue in vitro engineering: Synthesis, characterization and biocompatible evaluation. Polymers (Basel), 13: 3611. https://doi.org/10.3390/polym13213611 36. Li D, Liu P, Hao F, et al., 2022, Preparation and application of silver/chitosan-sepiolite materials with antimicrobial activities and low cytotoxicity. Int J Biol Macromol, 210: 337–349. https://doi.org/10.1016/j.ijbiomac.2022.05.015 37. Ding H, Peng X, Yu X, et al., 2022, The construction of a self- assembled coating with chitosan-grafted reduced graphene oxide on porous calcium polyphosphate scaffolds for bone tissue engineering. Biomed Mater, 17: 045016. https://doi.org/10.1088/1748-605X/ac6eab 38. Zhang R, Liu X, Xiong Z, et al., 2018, Novel micro/ nanostructured TiO2/ZnO coating with antibacterial capacity and cytocompatibility. Ceram Int, 44:  9711–9719. https://doi.org/10.1016/j.ceramint.2018.02.202 39. Aljohani W, Ullah MW, Zhang X, et al., 2018, Bioprinting and its applications in tissue engineering and regenerative medicine. Int J Biol Macromol, 107: 261–275. https://doi.org/10.1016/j.ijbiomac.2017.08.171 40. Bian T, Wang L, Xing H, 2021, Preparation and biological assessment of a ZrO2-based bone scaffold coated with hydroxyapatite and bioactive glass composite. Mater Chem Phys, 267: 124616. https://doi.org/10.1016/j.matchemphys.2021.124616 41. Khosrowshahi A, Khoshfetrat AB, Khosrowshahi YB, et al., 2021, Cobalt content modulates characteristics and osteogenic properties of cobalt-containing hydroxyapatite in in-vitro milieu. Mater Today Commun, 27: 102392. https://doi.org/10.1016/j.mtcomm.2021.102392 42. Wang S, Liu L, Zhou X, et al., 2019, Effect of strontium- containing on the properties of Mg-doped wollastonite bioceramic scaffolds. Biomed Eng Online, 18: 119. https://doi.org/10.1186/s12938-019-0739-x 43. Rajkumar P, Sarma BK, 2021, Role of Zn and Mg substitutions on the mechanical behaviour of biomimetic hydroxyapatite and insight of the emergence of hydroxyapatite-ZnO nanocomposite. Mater Charact, 176: 111107. https://doi.org/10.1016/j.matchar.2021.111107 44. Chen X, Zhang L, Liu X, et al., 2012, Preparation and properties of biodegradable composites derived from poly(lactide-co-glycolide), poly(L-lactide), and nanohydroxyapatite. J Macromol Sci B, 52: 462–475. https://doi.org/10.1080/00222348.2012.712007 45. Mohajernia S, Hejazi S, Eslami A, et al., 2015, Modified nanostructured hydroxyapatite coating to control the degradation of magnesium alloy AZ31 in simulated body fluid. Surf Coat Tech, 263: 54–60. https://doi.org/10.1016/j.surfcoat.2014.12.059 46. Shen Y, Liu W, Wen C, et al., 2012, Bone regeneration: local pH-strontium-doped borosilicate Importance of scaffold. J Mater Chem, 22: 8662–8670. https://doi.org/10.1039/c2jm16141a 47. Rokusek D, Davitt C, Bandyopadhyay A, et al., 2005, Interaction of human osteoblasts with bioinert and bioactive ceramic substrates. J Biomed Mater Res A, 75: 588–594. https://doi.org/10.1002/jbm.a.30459 48. Lian JB, Stein GS, 1992, Concepts of osteoblast growth and differentiation: basis for modulation of bone cell development and tissue formation. Crit Rev Oral Biol Med, 3: 269–305. https://doi.org/10.1177/10454411920030030501 49. Silva AS, Santos LF, Mendes MC, et al., 2020, Multi-layer pre-vascularized magnetic cell sheets for bone regeneration. Biomaterials, 231: 119664. https://doi.org/10.1016/j.biomaterials.2019.119664 50. Gregory CA, Gunn WG, Peister A, et al., 2004, An Alizarin red-based assay of mineralization by adherent cells in culture: Comparison with cetylpyridinium chloride extraction. Anal Biochem, 329: 77–84. https://doi.org/10.1016/j.ab.2004.02.002 51. Wang X, Liu S, Li M, et al., 2016, The synergistic antibacterial activity and mechanism of multicomponent metal ions- containing aqueous solutions against Staphylococcus aureus. J Inorg Biochem, 163: 214–220. https://doi.org/10.1016/j.jinorgbio.2016.07.019 Volume 9 Issue 2 (2023) 305 https://doi.org/10.18063/ijb.v9i2.668 Zn-doped coatings with osteogenic and antibacterial propertiesInternational Journal of Bioprinting
10.1016_j.jfs.2022.101072
Easy cleanups or forbearing improvements: the effect of CEO tenure on Easy cleanups or forbearing improvements: the effect of CEO tenure on successor's performance successor's performance Gonal Colak, Eva Liljeblom Publication date Publication date 01-12-2022 Licence Licence This work is made available under the CC BY 4.0 licence and should only be used in accordance with that licence. For more information on the specific terms, consult the repository record for this item. Document Version Document Version Published version Citation for this work (American Psychological Association 7th edition) Citation for this work (American Psychological Association 7th edition) Colak, G., & Liljeblom, E. (2022). Easy cleanups or forbearing improvements: the effect of CEO tenure on successor's performance (Version 1). University of Sussex. https://hdl.handle.net/10779/uos.23493284.v1 Published in Published in Journal of Financial Stability Link to external publisher version Link to external publisher version https://doi.org/10.1016/j.jfs.2022.101072 Copyright and reuse: Copyright and reuse: This work was downloaded from Sussex Research Open (SRO). This document is made available in line with publisher policy and may differ from the published version. Please cite the published version where possible. Copyright and all moral rights to the version of the paper presented here belong to the individual author(s) and/or other copyright owners unless otherwise stated. For more information on this work, SRO or to report an issue, you can contact the repository administrators at sro@sussex.ac.uk. Discover more of the University’s research at https://sussex.figshare.com/ Contents lists available at ScienceDirect Journal of Financial Stability journal homepage: www.elsevier.com/locate/jfstabil Easy cleanups or forbearing improvements: The effect of CEO tenure on successor’s performance ☆ Gonul Colak a, b, *, Eva Liljeblom b, c a University of Sussex, Brighton, United Kingdom b Hanken School of Economics, Helsinki, Finland c Lund University, Lund, Sweden A R T I C L E I N F O A B S T R A C T Long CEO tenure can harm firm performance even after the CEO is replaced. We analyze this issue by condi- tioning post-turnover firm performance on the length of the preceding CEO’s tenure. Identification comes from instrumenting sudden CEO deaths as an exogenous shock to tenure length. We find that when a successor takes over after a long-tenured CEO, operating performance and stock returns are significantly lower, restructuring costs are higher, “big baths” are larger, and firm recovery is slower. Weaker corporate governance and a long- tenured CEO with lower skills amplify these post-turnover effects. JEL classification: G3 G34 J24 M41 M43 Keywords: CEO tenure CEO term limits Restructuring costs Shareholder value Firm performance Hazard model 1. Introduction The typical goals of creating shareholder-friendly CEO remuneration contracts are to align managerial incentives with the interests of the firm’s shareholders and to engage the CEO with the firm through a longer-term contract1 (Holmstr¨om, 1982; Gibbons and Murphy, 1992). However, recent media debates and academic research (Whitehead, 2011; Henderson et al., 2006; Brochet et al., 2021) have put a spotlight on the issue of what is the optimal length of a CEO’s tenure. Given that managerial short-termism becomes a problem in the latter parts of their tenures (Prendergast and Stole, 1996; Edmans et al., 2012; Pan, Wang, and Weisbach, 2016), understanding whether long-tenured CEOs have long-lasting effects even after replacement is an economically relevant question. Studies have examined the relevance of CEO fixed effects on firm performance around CEO turnovers (Fee and Hadlock, 2004), which are usually mixed announcements events that also include information on the turnover reasons, as well as the successor’s characteristics (Jenter and Kanaan, 2015). The main goal of such studies is to provide evidence on whether the new CEO matters to firm performance (Huson, Mala- testa, and Parrino, 2004). In a theoretical study, Casamatta and Guembel (2010) model the dynamic relationship between a long-tenured CEO and ☆ We are grateful for the helpful suggestions by Francois Brochet, Ettore Croci, Claudia Custodio, Mats Ehrnrooth, Ruediger Fahlenbrach, Marc Gabarro, Jesper Haga, Dirk Jenter, Timo Korkeam¨aki, Anders L¨oflund, Pedro Matos, Maurizio Montone, Conny Overland, and Sami V¨ah¨amaa. We also thank the conference, seminar, and brownbag participants at the 20th Workshop on Corporate Governance and Investment in Oslo (2019), 11th Nordic Corporate Governance Network Conference in Olso (2019, Hanken School of Economics (2017), Lund University (2018), Universit`a Cattolica del Sacro Cuore in Milan (2018), and University of Vaasa (2017) for their comments. * Corresponding author at: University of Sussex, Brighton, United Kingdom. E-mail addresses: gonul.colak@hanken.fi (G. Colak), eva.liljeblom@hanken.fi (E. Liljeblom). 1 Such remuneration examples can be observed when some companies are extending the time period for paying bonuses into the future years. For instance, EU regulation for banks require a substantial part (at least 40%) of variable renumeration to be deferred over a period of at least 3 years. While this has been put in place mainly out of concerns for whether the managerial actions have been on a stable economic ground, it has the indirect effect of tying the management to the firm for extended periods. See e.g., https://eba.europa.eu/single-rule-book-qa/-/qna/view/publicId/2018_3815. https://doi.org/10.1016/j.jfs.2022.101072 Received 12 March 2022; Received in revised form 25 August 2022; Accepted 31 August 2022 JournalofFinancialStability63(2022)101072Availableonline8September20221572-3089/©2022TheAuthor(s).PublishedbyElsevierB.V.ThisisanopenaccessarticleundertheCCBYlicense(http://creativecommons.org/licenses/by/4.0/). G. Colak and E. Liljeblom her successor and claim that a manager’s decisions affect future firm performance, potentially beyond the manager’s own tenure. However, empirical evidence on this relationship is nonexistent, and whether and how the length of the CEO’s tenure affects decision-making and firm performance under the new CEO remains empirically a largely unex- plored topic. We focus on this legacy aspect of a long-tenured CEO. Focusing on predecessor’s tenure, rather than performance, has some advantages. The performance effects of poor prior management de- cisions may not be immediately observable; managers can temporarily mask bad performance through earnings management and other such tactics, facilitated by lower monitoring (Dikolli et al., 2014). Tenure, on the other hand, is easily observable and quantifiable. Thus, if a long tenure is indeed associated with bad performance (as suggested by Brochet et al., 2021), then focusing on its length is empirically more reliable in our paper’s context. We contribute to the literature on managerial tenure by providing the first comprehensive set of tests where we condition the subsequent firm performance on the length of the preceding CEO’s tenure. Using CEO turnovers events in US public firms, we test for a significant rela- tionship between the post-turnover performance of the firm and the tenure of the preceding CEO (“the predecessor CEO”). Since CEO-firm matching is endogenous (Jenter and Kanaan, 2015), the tenure length is also endogenously determined. Proper identification in such empirical settings is methodologically challenging (Jenter et al., 2016). However, when a CEO’s tenure is cut short by external factors, such as sudden death (Jenter et al., 2016; Quigley et al., 2017; Brochet et al., 2021), this endogenous relationship likely breaks, and identification becomes possible. We also use the age at which the predecessor CEO took office as a second instrument. If a person becomes CEO at an old age, that person is marginally less likely to have a long tenure. Thus, using sudden deaths and the age at which the predecessor CEO took office as two in- struments, we improve identification by implementing a self-selection (2SLS) estimation. Self-selection here refers to firms that, for various economic and governance reasons, self-selected to keep their CEOs for a long period of time and thus, created a group of long-tenured CEOs. We focus on this set of predecessor CEOs whose tenures lasted longer than seven years (the median tenure of a CEO in our sample) and compare them a difference-in-difference format, whereby the firm performance during the three years before is compared to the performance during the three years after a turnover (as in Huson et al., 2004). short-tenured ones. Our analyses are to in We find that during the post-turnover years, Tobin’s Q, the abnormal stock return, and ROA are significantly lower for the cases when the predecessor CEO had a tenure longer than the sample median. Using a variation of Cox’s (1972) hazard rate model, we formally establish that for these long-tenured cases, the above performance measures take much longer to recover their peak levels observed in the few years prior to the turnover event. Thus, for such cases, the recovery period from the damage done during the last years prior to the turnover (Brochet et al., 2021) is significantly extended. Furthermore, the size and intensity of the “cleanups” (restructuring costs and write-offs) after a long-tenured CEO are larger. In the first few years following replacement, the firms report significantly larger di- vestments and restructuring costs2 (Weisbach, 1995; Bens and Johnston, 2009; Barron et al., 2011) and higher incidences of “big baths” (Elliott 2 Several studies in the accounting literature have focused on the role of restructuring charges in managerial decision making. For example, Moehrle (2002) shows that restructuring charge reversals are routinely used to manage earnings. Lee (2014) finds a lesser association between restructuring charges and earnings management after SFAS 146. and Hanna, 1996; Hazarika et al., 2012).3 Adut et al., (2003) link restructuring charges to the cash compensation and tenures of CEOs; and thus, these costs are relevant in our context. Similarly, “big baths” represent severe forms of cleanups, and they do tend to clear the air (Haggard, Howe, and Lynch, 2015). Reportedly, most CEOs engage in some form of earnings management early in their career (Elliott and Shaw, 1988; Pourciau, 1993; Ali and Zhang, 2015). We show that the new CEOs that follow a long-tenured predecessor tend to conduct an economically larger and more severe form of earnings management; namely, asset write-offs. More drastic changes seem necessary to wipe the slate clean after a long-tenured CEO. Next, we focus on various cross-sectional analyses that could shed light on the firm dynamics that allow some CEOs to overstay much longer than optimal and, in the process, do damage to the firms’ pros- pects. We report that the effects of the post-turnover low performance are more severe and more frequent in firms with weak corporate governance (weaker boards and dual-role CEOs), which suggests that over the years the long-tenured CEOs have become very powerful and entrenched within the firm. They appear to “overstay their welcome,” and when departing they leave the firm in a shape that requires serious restructuring by their successor. This is in line with the predictions from the literature on employee psychology (Murphy, 1989; Hambrick and Fukutomi, 1991; Ackerman, 1992; Sturman, 2003) as well as the recent finding of an inverse U-shaped relationship between firm performance and tenure (Brochet et al., 2021). The post-turnover performance can also be lower because of the unusually good performance of predecessor CEOs (i.e., they were skil- led). In this case, we implicitly assume that the pre-turnover perfor- mance can at least be partly attributed to the abilities of the long- tenured CEOs (i.e., the firm was rational in keeping them for so long). We refer to such cases as high-skill CEOs and show that controlling for CEO ability (Demerjian et al., 2012) does not completely eliminate the aforementioned relationship between tenure and post-turnover perfor- mance. We find only weak evidence that the long-tenured CEO’s ability is the reason for post-turnover underperformance of the firm. Put differently, the corporate governance explanation seems more plausible than the CEO skill. The above findings may have policy implications for some executive boards’ practices of tying-up the CEOs for a longer time with their current firms as well as for succession planning. If indeed a CEO’s per- formance deteriorates when tenure is longer, then certain tenure-related corporate bylaws are called for. Such bylaws can be imposed internally through changing the firms’ charters or by openly formulating board polices of not keeping a CEO longer than a certain number of years. 2. Related literature While academic work on how current CEO’s tenure affects the suc- cessor’s managerial decisions and performance is scarce, the studies on how CEO tenure affect firm performance is quite voluminous. We review the relevant studies to gain insights on the possible mechanisms though which a preceding CEO’s lengthy tenure can affect the future perfor- mance and behavior of the incoming CEO. First, we report on the studies that analyze how tenure length impacts employee job performance. 3 The term “big bath” is generally used to describe a large loss, asset write- down, or other non-recurring charges. Elliott and Hanna (1996), Francis et al., (1996), and Haggard, Howe, and Lynch (2015) analyze the earnings-related motives behind major asset write-offs. We use incidences of large asset write-offs (larger than 1% of the total assets; see our Appendix) as an indication that the company implemented a “big bath” form of earnings man- agement. Moore (1973) argues that “taking a bath” alters the firm’s bench- marks (e.g., shrinks assets) and increases future earnings by removing future losses from future income statements. Thus, it should be considered a form of discretionary earnings management. JournalofFinancialStability63(2022)1010722 G. Colak and E. Liljeblom Then, we review how the nature of CEO turnovers affects the tenure- performance relationship. Finally, we develop our hypotheses on the long-tenured CEOs and their successor’s performance. 2.1. Tenure length and performance on the job A proper understanding of the relationship between CEO tenure and firm performance requires insights from many branches of academic scholarship. We reviewed several relevant strands of the literature that ultimately helped us understand the role of two separate managerial effects: the time-invariant style of each individual manager (managerial fixed effects such as talent) and the time-varying attributes of the same manager (age, tenure, etc.). On the one hand, most studies in financial accounting emphasize the time invariant features of the manager. For example, Bertrand and Schoar (2003) find that the managerial fixed effects (or managerial style) explain the unknown parts of firm performance and these effects remain persistent for the same manager across firms. The managerial style is also important in CEO compensation (Graham et al., 2012), cost of capital (Francis, Hasan, Zhu, 2020), and accounting practice (Dyreng, Hanlon, and Maydew, 2010; Bamber, Jiang and Wang, 2010). On the other hand, some papers from the management literature emphasize the time varying features of managerial performance and decision-making. The dynamic performance4 research has conducted studies on the predictability of employer performance (see Sturman, 2007 for a review). One strand of dynamic performance research aims to investigate within-person patterns of performance and to understand what affects these patterns. Useful frameworks for analyzing the factors that affect the temporal dynamics in job performance are given by the models of changing subjects and changing tasks. The changing subjects or changing-person model (Keil and Cortina, 2001) refers to changes in the personal characteristics of the employee, such as ability, knowledge accumulation, and motivation, all of which over time can influence performance. The changing tasks model (Alvares and Hulin, 1972; Deadrick and Madigan, 1990) predicts that performance changes, because job requirements change through the introduction of new technology. Many studies document an inverted U-shaped relation between job performance and temporal variables such as age (Rhodes, 1983; McEvoy and Cascio, 1989; Avolio et al., 1990). Sturman (2003) studies three temporal variables: age, job experience, and organizational tenure (i.e., an accumulation of work-related information that is conceptually distinct from job experience). While learning theory and organizational socialization posits that job experience and organizational tenure have a positive initial effect, the effect of age is generally assumed to be 4 The term dynamic performance refers to the idea that an individual’s job performance changes over time. Key theoretical developments within the dy- namic performance research include the employment stage models by Acker- man (1987), (1992) and Murphy (1989). Ackerman proposes a theory of skill acquisition, postulating that individuals proceed through three stages: 1) the cognitive phase requiring mental abilities during which progress is fast, 2) the associative phase involving refinement of the stimulus-response connections developed in stage one, and 3) the final stage where performance is automatic and can be completed even without the full attention of the performer. Task complexity and consistency naturally affects the process through these stages. Murphy (1989) extends Ackerman’s model from simple task performance to the job performance context by distinguishing between two stages, the transition and maintenance stages. Murphy’s model predicts that in the latter stage, cognitive ability is less important and personality and motivational factors play a more important role for job performance. increasingly negative over time.5 Sturman (2003) thus postulates an inverse U-shaped relationship between performance and time, as the negative effect of age should overtake the positive but declining effects of the other temporal variables. Indeed, he finds empirical support for a negative longer term slope between the temporal variables and job performance. Hambrick and Fukutomi (1991) model the progression of a lengthy CEO tenure that consists of five stages, whereby a negative time effect dominates in the last “dysfunctional” stage. Shen and Canella (2002) argue that the strong strategic inertia arising from long-tenured CEOs can create serious problems for the successor, since the strategies and structures created during the preceding CEO’s long tenure may no longer be appropriate. Consistent with these studies, Miller (1991) presents reasons (such as narrowed managerial perceptions, managerial power, and overconfidence) for why the strategies of the firm may deviate more from optimal when there is a long-tenured CEO in charge. Studies in the financial literature have addressed the temporal effect on performance in connection with the tenures of board members and CEOs. Huang and Hilary (2018) find support for a U-shaped inverse relationship between outside directors’ tenure and firm performance (Tobin’s Q), while controlling for many other factors. Their results show that an optimal tenure length is around 10 years. Similarly, Brochet et al. (2021) report an inverted U-shaped relation between CEO tenure and firm value as well as acquisition announcement returns. Their findings are the result of two interplaying forces. The positive but declining benefits from on-the-job learning, experience, and relationship building; and the negative effects of an increasing mismatch between the incumbent CEO and the firm (due to changes in its environment) along with the CEO’s increasing risk aversion and increasing reluctance to change (Holmstr¨om, 1982). Brochet et al. (2021) also discuss alternative explanations for their findings of which one is that CEOs with good performance may get recruited more easily to run bigger firms, while poorly performing CEOs have an incentive to stay on the job longer. Their specification tests indicate that the alternative explanations are not the main driver of their results. The average inflection point for the relationship between CEO tenure and firm value is around 11 years in their main model; however, in line with Hambrick and Fukutomi’s (1991) insights, the inflection point shifts to lower (higher) values for firms in more (less) dynamic settings. 2.2. Studies of CEO turnover The bulk of papers on CEO turnovers focus on managerial effects that are associated with the retiring CEO. Given the many reasons for turn- over (voluntary versus involuntary), conditioning on the reason for the replacement (retirement, new opportunities, sudden death, or involun- tary dismissal) is necessary to avoid negative and positive performance effects of a CEO change cancelling out in the aggregate. A major complication in studies of CEO succession is that the post-turnover ef- fects typically are mixed events (Combs et al., 2007) in which they refer to how the new CEO compares with the old one (is the change for better or worse). Therefore, the papers on CEO replacements tend to condition also on the characteristics of the successor (e.g., external or internal hire; Ertimur et al., 2018). CEO turnovers are associated with significant changes in asset di- vestments (Weisbach, 1995; Barron et al., 2011), top management team turnover (Hayes et al., 2006), earnings management (Elliott and Shaw, 1988; Hazarika et al., 2012; Ali and Zhang, 2015), “strategic noise” (Griffin et al., 2011), stock price reactions at the announcement (Weis- bach, 1988; Huson et al., 2004) as well as changes in the firm’s operating 5 The decrement theory of aging (Giniger et al., 1983) supports the claim of adverse effect of age on performance. Similarly, Verhaegen and Salthouse (1997) find strong support for a negative influence of age on cognitive abilities. Age also negatively interacts with ambition and overall career motivation (Rhodes, 1983). JournalofFinancialStability63(2022)1010723 G. Colak and E. Liljeblom performance (Denis and Denis, 1995; Huson et a, 2004). For example, Weisbach (1995) and Barron et al. (2011) study large asset sales in relation to CEO turnover. Weisbach reports an increased probability of divesting a major asset at a loss after a CEO change. Since there is apparent endogeneity (the divestment decision and the CEO change may be both due to the poor performance), Weisbach (1995) also studies normal retirements and finds no measurable difference in their effect on the likelihood of divestiture. Barron et al. (2011) condition the asset sales on successor type as well as top management team turnover. They find that a CEO departure is associated with discontinued operations only for succession by contenders (but not by followers).6 In the case of an outside CEO, the likelihood of discontinuation is significantly higher only when it coincides with the departure of other top management executives. Denis and Denis (1995) report that while normal CEO retirements are followed by small increases in operating income, forced turnovers are preceded by large and significant declines in operating performance and intense corporate control activity. Such forced CEO turnovers are also followed by large improvements in performance. Controlling for mean reversion in firm performance, Huson et al. (2004) finds that various accounting measures tend to deteriorate prior to CEO turnovers and improve afterwards, and this effect is stronger for forced turnovers. Using a control group, the authors find support for the improved man- agement hypothesis, rather than the scapegoat hypothesis which relies on mean reversion in firm performance. Since CEO dismissals are often presented by the firm as voluntary changes, many researchers have turned to studying sudden CEO deaths instead. These cases offer a cleaner sample of “random events” given that suicides are typically screened out. Such studies typically show rather insignificant market reactions in the total population (Johnson et al., 1985; Salas, 2010) but may find significant excess returns for certain subsets of CEOs based on either their or governance character- istics (Johnson et al., 1985; Worrell et al., 1986; Borokhovich et al., 2006). For example, Salas (2010) finds that while age and tenure indi- vidually correlate weakly with the stock market’s reaction to sudden deaths, the price reaction is strongly positive when the CEO’s tenure exceeds 10 years, and the abnormal stock returns over the last 3 years have been negative. Bennedsen et al. (2020), in turn, report that the hospitalization of a senior corporate executive is likely to cause a sig- nificant decline in firm profitability. Quigley et al. (2017) find a stronger reaction to sudden deaths in later years that indicates CEOs have become more influential over time. Finally, Jenter et al. (2016) have reported a very heterogeneous pattern with many specific significant effects such as negative abnormal returns to sudden deaths of a founder-CEO, but a positive effect for slow deaths or sudden deaths of old and long-tenured CEOs. Studies of the CEO-turnover-firm-performance relationship also offer some insights into the effect of CEO tenure. Dikolli et al. (2014) find that the turnover-performance relationship declines over the tenure and that longer tenured CEOs are associated with weaker governance structures. Specifically, they find that board independence and the number of board meetings were lower and that the board size was larger for tenured CEOs as compared to new ones. The causality of such a relationship remains largely ambiguous. The authors interpret their results as indicating that CEO tenure is related to past performance that reflects a form of “sta- tistical entrenchment” more than weak governance. However, they also conclude that they cannot rule out the possibility that weak governance permits the entrenchment performance-turnover relationship as tenure increases. This finding is relevant to our study as it indicates that in the long-tenured group of managerial weakens that 6 Barron et al. (2011) use the definitions of Shen and Canella (2002) who distinguish between three distinct types of CEO succession: outsider, follower (including an insider appointed after the retirement of the outgoing CEO), and contender (an insider appointed after the resignation of the outgoing CEO). CEOs, there may be greater heterogeneity, because of low monitoring CEOs with low performance may avoid being replaced. Interesting findings related to CEO tenure are also reported by Shen and Canella (2002) who claim that there is an inverted U-shaped relationship be- tween the tenure of the departing CEO and post-succession ROA of the firm. Griffin et al. (2011), on the other hand, find significantly more “strategic noise” (announcements of dividends, earnings, changes in key executives, etc.) around the appointment of the new CEO. One of the control variables used in the “strategic noise” regressions is the tenure of the preceding CEO. 2.3. Hypothesis development We base our hypotheses on the CEO-related dynamic performance theory (Miller, 1991; Hambrick and Fukutomi, 1991; Sturman, 2003; Brochet et al., 2021). According to this theory, CEO autonomy grows over tenure (connections to clients and suppliers strengthen, influence over the board and owners increases, etc.), and learning-on-the-job slows down. After a certain point, managers become fixated on their own ways and become resistance to change. Eventually their perception of their firm’s organizational structure narrows, and the need for reor- ganization goes unnoticed. A certain strategic inertia develops within the firm (Miller, 1991; Shen and Canella, 2002), and both the top managers and subordinates become reluctant to change anything. In a study of CEOs’ attitudes towards change, Musteen et al. (2005) report tenure to be the most important determinant of resistance to change. Tenure also reduces the effects of other attributes (e.g., age) as it gets longer. Jenter et al. (2016), among others, describe a simple frictionless competitive assignment framework to understand the matching between firms and CEOs and the replacement of the latter that also includes valuation consequences. Betzer et al. (2020) elaborate on some of the frictions that may distort the matching and replacement. In a frictionless world, the firm replaces the incumbent CEO as soon as its value under them becomes lower than its value under the next best alternative manager. However, according to Brochet et al. (2021), such frictions create the conditions under which some CEOs end up having unneces- sarily long tenures and that an inversed U-shaped relationship exists between those tenures and firm performance. In light of the findings by Brochet et al. (2021), we first postulate that long-tenured CEOs are damaging to firm performance during the last few years of their tenure. We then put focus on the performance of the incoming CEO. If indeed there is an inversed U-shaped relationship between CEO tenure and firm performance, then firm mismanagement should be greater in the case of very long-tenured CEOs as compared to the rest. Correcting for such mismanagement should require more effort from the successor and it should take more time. In such cases, the turnover of the CEO facilitates the correction of their “errors” —such as manager-specific investments (Shleifer and Vishny, 1989) or a lack of necessary (de)investments or restructurings (Boot, 1992; Prendergast and Stole, 1996)— and post-turnover actions taken by the successor may reflect this error correction.7 This is the basic idea that underlies all our tests, and it is new to the literature in the sense that no other study has systematically conditioned all post-turnover performance effects on the tenure of the preceding CEO. We focus on the tenure of the preceding CEO as it is easily observable and quantifiable as compared to, for example, past perfor- mance that can temporarily be masked especially under lower 7 According Eisfeldt and Kuhnen (2013) and Garrett and Pavan (2012), however, both CEO turnover and post-turnover actions can be responses to shocks that affect CEO/firm productivity. Such shocks are neither in the control of the predecessor CEO nor the board of directors. To control for such cases, we analyze whether the impact of long-tenured predecessor is different in dynamic industries (i.e., industries that are more prone to productivity shocks). JournalofFinancialStability63(2022)1010724 G. Colak and E. Liljeblom monitoring. Moreover, strategic misalignments may result in perfor- mance effects with long lags. We study operational performances (ROA, restructuring costs, and write-offs) as well as market returns and valu- ations (abnormal stock returns and Tobin’s Q). If indeed the latter years of a long-tenured CEO are damaging to firm performance (Huson et al., 2004; Brochet et al., 2021), then we predict that the new CEO will have a more difficult time turning the firm around. The relative under- performance of such firms would continue during the first few years of the new CEO as any correction will take some time before certain measures are put in place and the firm starts showing improvements. The need for major cleanups after long-tenured CEOs would also be stronger and deeper as the strategic inertia created by the predecessor would be much stronger in such cases (see e.g. Quigley and Hambrick, 2012). In short, we hypothesize that the relationship between the pre- ceding CEO’s tenure and firm performance will be negative. Similarly, the longer tenure of the preceding CEO is likely to increase the depth and the size of the cleanups (restructuring costs and asset write-offs). Hence, Hypothesis 1A. : There is a negative relationship between firm per- formance during the first few years of a new CEO and the tenure of the preceding CEO. : There is a positive relationship between the size of Hypothesis 1B. the clean-ups during the first few years of a new CEO and the tenure of the preceding CEO. How do long-tenured CEOs manage to stay in office and cause the mess that their successors have to clean up? We use our next hypothesis to analyze the mechanism that can explain the relationship described above. One of the factors that can modify this relationship is the corporate governance of the firm. We posit that firms with strong corporate governance will keep a CEO longer only when it is in the in- terest of the owners, that is, when the performance of the CEO justifies a longer tenure. However, long-tenured CEOs would also have more time to entrench themselves via investments (Shleifer and Vishny, 1989). Also, firms managed by such successful and charismatic CEOs may end up with weaker governance structures as suggested by Hermalin and Weisbach (1998), and they would be more reluctant to undo prior in- vestments even if a new productivity shock requires such change (Boot, 1992; Prendergast and Stole, 1996). On the other hand, if the CEO underperforms but again the firm’s corporate governance is weak (the CEO has more power), the firm may delay their firing. Regardless of the performance of the CEO, longer tenures are likely to be observed if a firm has weak corporate governance. Thus, we predict that the negative relationship between long-tenured CEOs and post-turnover performance will be more pronounced in weak corporate governance firms. Namely, Hypothesis 2. The negative relationship between firm performance during the first few years of a new CEO and the tenure of the preceding CEO is stronger when corporate governance is weak. Also, skill and motivational factors that affect the preceding CEO can modify the relationship described in Hypothesis 1. According to Murphy (1989), personality and motivational factors play a larger role in per- formance in the latter stages of a career. Hence, there should be some cross-sectional variation among the long-tenured CEOs’ performance based on their managerial ability. The post-succession negative effects should be stronger when predecessor CEO have stayed in power for a long time in spite of their being mediocre managers (i.e., long tenure was not justified by exceptional skills). Hypothesis 3. The negative relationship between firm performance during the first years after succession and the tenure of the preceding CEO is more pronounced if that CEO had less skill. 3. Data and sample selection In this section, we describe our data sources, our sample selection, and the construction of various measures that we use in this study. Detailed definitions of all the variables used in this study can be found in Appendix A. 3.1. CEO turnovers and long-tenured CEOs We start by identifying all the CEO turnover events between 1993 and 2013 from the Execucomp database. First, we retrieve the list of all the instances when the name and executive ID (EXECID) of the CEO of a given firm (GVKEY) changes from one year to another. There are about 4000 such cases between 1993 and 2013. This list constitutes our initial raw data for CEO turnovers. We then manually check each case by reading the relevant news from Factiva or Lexus-Nexus to verify that the turnover indeed occurred and when. We drop the cases when we cannot verify whether the turnover shown in Execucomp occurred within one year of the real turnover as mentioned in the news.8 At the end, we are left with 2428 CEO turnover events. We create a time frame around each turnover starting three years before and ending three years after the event (a la Huson et al., 2004). The three-year period leading up to the turnover (t = (cid:0) 3,(cid:0) 2,(cid:0) 3) is referred to as Before and the three-year period after (t = +1,+2,+3) is called After. The year of the turnover (t = 0) is dropped from our ana- lyses.9 For each of the years in this window we calculate various ac- counting and stock performance measures found in the Compustat, CRSP, and Execucomp databases. We assure that each turnover event has exactly six years associated with it.10 If a firm has two back-to-back CEO turnovers (say in 1997 and in 1999), then certain fiscal years for this firm appear twice in our data.11 Thus, we have a total of 14,568 (=2428 ×6) firm-years that correspond to 2428 CEO turnovers that involve 1475 unique firms (on average, 1.65 turnovers per firm). We also find some information about the preceding CEO’s tenure at the time of the turnover (Predecessor CEO Tenure). For example, the median Predecessor CEO Tenure in our sample is seven years that typi- cally indicates the preceding CEOs quit after seven years of service. Thus, using the Predecessor CEO Tenure measure, we classify each of our 2428 turnovers events as the ones that involve a long- or a short-tenured 8 Most of these dropped cases are related to restructuring events (takeovers or divestitures) whereby the CEO appears to change but in reality, they became the CEO of the divested company. 9 Prior literature recommends that this event year t = 0 to be excluded from the analyses (see Gertner et al., 2002; and Huson et al., 2004). However, in a series of robustness tests, we verify that including this event year t = 0 in our regression analysis does not affect our results. This issue is particularly relevant in the cases of write-offs (Haggard et al., 2015) since prior literature claims that the first year of a CEO turnover is very critical for earning management (Elliott and Shaw, 1988; Ali and Zhang, 2015). In untabulated tests we verify that there is a spike in write-offs in year t = 0 in our sample, as well. In a related robustness test, we again verify that including year t = 0 into our write-offs (clean-up) regressions does not alter our qualitative conclusions. 10 Our qualitative results do not change when we include only the cases where the full [(cid:0) 3;+ 3] time frame is observable (i.e., we remove the CEO turnovers for which the CEOs quit within 1–2 years after they started their jobs). 11 In the example above, the 1996 accounting variables will appear twice as belonging to the Before periods of both turnovers. The year 1998 will appear twice: once belonging to the After period of the first turnover and then belonging to the Before period of the second turnover. The year 2000 will also appear twice as belonging to the After periods of both events. The year 1997 will be dropped from first turnover’s time frame, but it will appear once as part of the Before period of the second turnover. Similarly, the year 1999 will appear once as part of the After period of the first turnover. JournalofFinancialStability63(2022)1010725 G. Colak and E. Liljeblom predecessor CEOs. The turnovers that involve CEOs who had a tenure longer than seven years are called Long-Tenured CEO turnovers.12 The rest are Short-Tenured CEO turnovers. 3.2. Data used for firm performance measures before and after a CEO turnover Our analysis of the cleanup efforts after a CEO turnover requires indicators of the activities that capture various dimensions of these ef- forts and the related firm performance measures. We use annual and quarterly accounting data from the Compustat database for the period from 1993 to 2013 when calculating the accounting performance mea- sures, such as return on equity (ROA) and restructuring charges. The stock-based performance measures, such as Tobin’s Q and long-run buy- and-hold (BHAR), are calculated using data on monthly dividend- adjusted stock returns from CRSP, along with the CRSP equally- weighted market returns. In our stock return analysis we include only firms with at least 12 months of available data on stock prices. The data on CEO characteristics is obtained from Execucomp, and the source for our firm and industry characteristics is Compustat’s annual files. 3.3. Descriptive statistics The descriptive statistics for our sample are reported in Table 1. The average age is quite high: 58.9 years with a median of 60. There is only a small difference between the averages and means for age between the short- and long-tenured CEOs (e.g., median ages of 59 and 61, respec- tively). Tobin’s Q is higher in the firms with long-tenured CEOs. More than half of our firms (54.5%) pay dividends, and 39% of the firms in our sample had a loss during at least one of the past three years. Average sales growth is 8% and is somewhat higher among the firms managed by long-tenured CEOs. In Table 2, we report the results from a univariate analysis of the differences in firm performance between the pre- and post-CEO turnover years (diff-in-diff). That is, we compare the development of each vari- able during the 3 years prior to the CEO change, and during the 3 years following the change (the CEO turnover year being excluded). We do that separately for two sub-groups of firms, formed on the basis of the predecessor CEO’s tenure. The differences are computed so that the prior year’s variable values have been deducted from the post-turnover year’s values (post minus pre). We find that the two subsamples (long- versus short-tenured) significantly differ from each other in many respects. The change in the average Tobin’s Q is much bigger in the group of long-tenured CEOs ((cid:0) 0.329) as compared to short-tenured CEOs ((cid:0) 0.156), and the differ- ence between the two groups is significant at the 1% level. Also, many other variables have a similar pattern: the restructuring charges and asset write-offs increase significantly more for the post-turnover period in the group of long-tenured CEOs, and the change in ROA is signifi- cantly lower in that group. For the short-tenured CEOs, the median stock price appreciates by about 2.7% over the 36 months following the turnover (BHAR36), but for long-tenured CEOs it deteriorates by about 11.8% (difference is statistically significant at the 1% level). On the other hand, an opposite pattern is observable for sales, assets, and the number of business segments: in dollar terms, they increase significantly more for long-tenured CEOs. Taken as a whole, the pattern points to efforts by successors of long- 12 Using a higher cut-off value (e.g., 10 years instead of 7) when defining Long- Tenured CEO, reduces the sample of treated observations. Nonetheless, our main results remain qualitatively the same. As we explain in Section 4, if instead of a binary variable Long-Tenured CEO we use a continuous tenure variable (Tenure_ Predecesor) and then interact it with After in a 2SLS estimation, does not alter our main conclusions either. The results from such a 2SLS are shown in our Online Appendix. tenured CEOs to increase sales even though the current productivity and performance of the firm is suffering. At the same time, there is evidence of restructuring the business and of reshuffling some segments in prep- aration for the future. Also, cleaning up for the future earnings by conducting write-offs (a la Haggard et al., 2015) is also observable in the univariate statistics. These findings provide preliminary evidence in support of Hypotheses 1 A and 1B. 4. Methodology In this section, we describe our setup for the difference-in-difference estimation. We also elaborate on our self-selection model and the in- struments used to achieve better identification. 4.1. Difference-in-difference estimation Throughout the study, we seek to understand the effects of longer tenures on various post-turnover economic outcomes. To better identify such effects, we use a form of a difference-in-difference (diff-in-diff) approach whereby we distinguish the effects of long-tenured CEOs from the effects of other CEOs. Specifically, we apply the form developed in Gertner et al. (2002) and Huson et al. (2004): Performit = β0i + β1iAfterit + β2iLTPi + β3t(Afterit ∗ LTPi) + γitXit + θt + ∅i + ϵit (1) where Performit is a performance variable for the firm-year (Tobin’s Qit, BHARit, ROAit, Restructuring Costsit, or major asset Write-offsit). The dummy Afterit indicates the post-turnover period for each firm i, and Long-Tenured Predecessor (LTPi) shows for which firms the predecessor CEOs had a long tenure (above seven years). The interaction variable (Afterit*LTPi) is of main interest to us, as it captures how well the new CEO performs conditional on their predecessor’s tenure. Xit is a vector of firm controls (Log(sales), Leverage, Cash Holdings, NetPPE, CAPX, RND, Dividend Dummy, Acquisitions, and Log (SEGN)). All our empirical setups include year fixed effects (θt). We also implement a tighter identification specification by applying firm fixed effects (∅i) in Eq.(1). As explained in the previous section, an average firm experiences about 1.65 (=2428 turnovers/1475 firms) turnovers during our sampling period that facil- itates the implementation of an identification-enhancing diff-in-diff regression with firm fixed effects. Moreover, to implement a proper before-and-after estimation we work with a symmetrical window that constitutes the three years before and after the turnover (see previous section for the definition of Afterit dummy). This estimation achieves a better identification through a difference-in-difference, whereby long-tenured CEOs are compared to short-tenured ones. However, for such a before-and-after estimation to work properly, we can only use firm-specific controls in Xit. The CEO- specific variables are different in the before and after periods, as the individuals occupying the CEO position change; thus, the CEO charac- teristics are not comparable across the two sides of the before-and-after estimation window. We control for any relevant CEO characteristics that may influence the relationship between the preceding CEO’s tenure and firm performance, such as power (corporate governance), by conducting subsample analyses for good versus weak corporate governance sub- samples. Other related robustness tests (e.g., whether the predecessor CEO was a founder of the company13) are conducted by removing such cases from our sample of turnovers and rerun our analyses (in Section 6). To verify the reliability of our difference-in-difference estimation, we conduct tests based on the parallel trends assumption. In the Online Appendix (Table A1), we implement our parallel trends tests following the method in Table IV of Giroud and Mueller (2015) who also work 13 Reportedly, the founder CEOs behave and perform differently than the other CEOs (Adams et al., 2009; Fahlenbrach, 2009). JournalofFinancialStability63(2022)1010726 G. Colak and E. Liljeblom Table 1 Summary Statistics: Long- vs Short-Tenured CEOs. VARIABLES N mean median N mean median N mean median Entire Sample Short-Tenured CEOs Long-Tenured CEOs 60 7 2428 2428 58.83 8.265 Panel A: Predecessor CEO characteristics at the time of the turnover event (year t = 0) Age (on turnover date) PredecessorTenure (on turnover date) Panel B: Firm Characteristics Tobin’s Q Tobin’s Q (industry adjusted) BHAR (36 months) CAR (36 months) ROA Restructuring Charges (REST) Writeoffs Idiosyncratic Risk Size (Log(Sales)) Revenues (Sales) Total Assets Leverage Cash Holdings Investments RND Dividend Dummy Acquisitions lnSEGN Cash Flows Net Income Profit Margin Institutional Ownership 1.975 0.459 0.0163 0.0286 0.0263 0.00308 0.0235 0.103 7.247 5126 5707 0.218 0.286 0.0570 0.0356 0.545 0.0241 0.787 0.287 291.0 0.138 0.742 1.529 0.0883 -0.0352 0.0152 0.0469 0 0 0.0890 7.191 1325 1319 0.207 0.0901 0.0423 0.00408 1 0 0.693 0.264 47.40 0.133 0.770 12,918 12,918 12,930 12,930 12,911 12,943 12,943 12,644 12,928 12,940 12,939 12,888 12,910 12,823 12,943 12,943 12,943 11,885 12,924 12,913 12,888 5752 1277 1277 6780 6780 6786 6786 6774 6793 6793 6659 6783 6791 6792 6761 6773 6716 6793 6793 6793 6288 6781 6774 6759 3017 58.00 3.955 59 4 1.871 0.350 -0.00406 0.0110 0.0174 0.00319 0.0221 0.102 7.266 4892 5281 0.222 0.280 0.0575 0.0353 0.569 0.0216 0.794 0.295 255.2 0.128 0.722 1.495 0.0474 -0.0511 0.00271 0.0424 0 0 0.0886 7.226 1371 1377 0.210 0.0876 0.0439 0.00654 1 0 0.693 0.274 43.20 0.125 0.748 1151 1151 6138 6138 6144 6144 6137 6150 6150 5985 6145 6149 6147 6127 6137 6107 6150 6150 6150 5597 6143 6139 6129 2735 59.75 ** 13.10 *** 61 ** 12 *** 2.089 *** 0.579 *** 0.0387 *** 0.0481 *** 0.0361 *** 0.00295 0.0250 * 0.103 7.226 * 5384 6177 0.214 *** 0.292 ** 0.0564 0.0359 0.519 *** 0.0268 *** 0.778 0.277 *** 330.5 ** 0.149 *** 0.764 *** 1.583 *** 0.138 *** -0.0186 *** 0.0275 *** 0.0509 *** 0 0 ** 0.0894 * 7.150 ** 1274 ** 1274 0.204 *** 0.0932 *** 0.0408 ** 0.000876 ** 1 *** 0 *** 1.099 0.253 *** 51.23 *** 0.141 *** 0.789 *** The table displays the summary statistics for the important variables used in this study. Panel A includes some CEO characteristics, and Panel B covers the firm and industry characteristics. We use a sample of CEO turnovers (2428 events) between 1993 and 2013. The firm years included in the sample correspond to the pooled firmyears that belong to the Before ([(cid:0) 3,(cid:0) 1]) and After ([+1,+3]) periods of each turnover event. The year 0 (the turnover year) is excluded from the sample. In our sample the median tenure of the predecessor CEO is 7 years, thus the sub-sample Long-Tenured Predecessor (Short-Tenured Predecessor) includes current CEO’s firm- years that had a predecessor with tenure longer (shorter) than 7 years. All the variables are defined in Appendix A. All the continuous variables are winsorized at 1% and 99% levels. The results from t-test (means) and Wilcoxon Ranksum test (medians) between “Short-Tenured CEOs” and “Long-Tenured CEOs” sub-samples are also shown. *** , **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. with subsamples in the context of a difference-in-difference analysis. We do not find evidence that the parallel trends assumption is violated in our sample. Thus, using the difference-in-difference specification in Eq. (1) is appropriate for our context. 4.2. Endogeneity of CEO tenure length (Heckman self-selection implement a Since many CEOs and firms are endogenously matched (Jenter and Kanaan, 2015), the extent to which a CEO stays with a firm is likely also endogenously determined. Achieving proper identification in such in- stances is difficult (Jenter et al., 2016). However, when the tenure of a CEO is cut short by external circumstances, such as sudden death (Brochet et al., 2021), this endogenous relationship is likely perturbated, and identification is possible. Thus, using sudden deaths as instruments, two-stage) model. we Self-selection here refers to a firm which, for various economic and governance reasons, self-selects to extend its CEO’s tenure for a longer period (i.e., LTPi is endogenously determined). As indicated by Eq. 1, we also implement a diff-in-diff method alongside the self-selection model. As explained in Gertner et al. (2002) and Huson et al. (2004), using a diff-in-diff method in such settings is preferable for identification pur- poses that are achievable with the help of a binary variable for tenure length, LTPi. However, we also implement a 2SLS estimation, whereby instead of a binary variable, we use the continuous tenure-length vari- able, PredecessorTenurei (see Table 1 for descriptive statistics). While such a 2SLS is no longer a diff-in-diff, it helps assure that the results are not driven by the dichotomization of the predecessor CEO’s tenure (a continuous variable). The results from this robustness analysis are dis- cussed in our Online Appendix. As explained in subsection 6.2 of Wooldridge (2002), when LTPi is endogenous, the interaction term Afterit*LTPi is also endogenous; thus, we need a separate first-stage regression for this term. The variables instrumenting for this interaction term can be themselves interactions of Afterit with the instruments for LTPi (see Wooldridge, 2002 for more detailed explanations). The same arguments apply when a 2SLS esti- mation is implemented with PredecessorTenurei instead of LTPi. Thus, our Heckman two-stage estimation can be specified as: LTPi = α0i + α1i Instruments for LTPi + α2i Zi + μi Afterit*LTPi = δ0i + δ1i Instruments for (Afterit*LTPi) + δ2i Zi + ξi (2) (3) Performit = β0i +β1iAfterit +β1iLTPi +β2i (Afterit*LTPi) +λ1iIMR1i +λ2iIMR2i +γitXit + єit (4) The variables that are included in Zi are the same control variables as in Xit. The inverse Mills ratios in Eqs. (2) and (3) are calculated as in Wooldridge (2002) and are referred to as IMR1 and IMR2, respec- tively.14 Again, the time and firm fixed effects are applied in the second stage. 14 Throughout the paper, we rely primarily on the IMR methodology to ac- count for the endogeneity bias as advocated by, among others, Colak and Whited (2007). However, for robustness purposes we consider alternative methodologies that are also known to reduce endogeneity bias. These meth- odologies are i) 2SLS with continuous tenure variable, ii) matching estimators (propensity score matching (Dehejia and Wahba, 2002) and bias-adjusted matching (Abadie and Imbens, 2006), and iii) entropy balancing method (implemented as in Hainmueller (2012)). The Online Appendix, Section B presents our main tests with these alternative methods. JournalofFinancialStability63(2022)1010727 G. Colak and E. Liljeblom Table 2 Diff-in-Diff Analyses of Long- and Short-Tenured CEOs. Performance Measure Difference (After minus Before) Tobin’s Q Tobin’s Q (ind. adjusted) BHAR (36 months) CAR (36 months) ROA Restructuring Charges Writeoffs Sales (in millions) Assets (in millions) Leverage Investments RND Acquisitions Number of Segments Cash Flows (CF) Net Income (in millions) Profit Margin (NI/Sales) Institutional Ownership Observations Short-Tenured Predecessor Long-Tenured Predecessor Mean Median Mean Median Short- Tenured minus Long- Tenured Diff-in-Diff -0.156 -0.037 -0.329 -0.113 0.173 *** -0.154 -0.037 -0.287 -0.084 0.133 ** -0.064 0.027 -0.345 -0.118 0.281 *** -0.028 -0.004 0.010 -0.002 -0.176 -0.021 -0.080 -0.011 0.148 *** 0.017 *** 0.001 -0.002 0.000 0.000 0.002 0.006 0.000 0.000 -0.001 *** -0.008 *** 606.464 125.115 1139.282 204.386 -532.818 ** 709.164 0.012 -0.008 0.001 0.000 131.367 0.000 -0.005 0.000 0.000 1477.706 0.012 -0.012 0.002 -0.003 241.390 0.000 -0.007 0.000 0.000 -768.542 *** 0.000 0.004 ** -0.001 0.003 0.155 0.036 0.000 0.021 0.305 0.013 0.000 0.008 -0.150 * 0.023 *** 79.083 12.453 73.809 9.490 5.273 -0.007 -0.002 -0.006 -0.005 -0.001 0.019 0.016 1166 0.047 1047 0.110 -0.028 The table shows univariate results that compare the new CEO’s performance conditioned on the predecessor’s tenure. The firm years included in the sample correspond to the years leading up to the turnover ([(cid:0) 3,(cid:0) 1]) and the years after the turnover event ([+1,+3]). The year 0 (the turnover year) is excluded from the sample. Before indicates the years before ([(cid:0) 3,(cid:0) 1]) and After indicates the years following the CEO turnover event (i.e., [+1,+3]). In our sample the me- dian tenure of the predecessor CEO is 7 years, thus the sub-sample Long-Tenured Predecessor (Short-Tenured Predecessor) includes current CEO’s firm-years that had a predecessor with tenure longer (shorter) than 7 years. The years after a CEO’s turnover are compared to the years leading up to the turnover event and then, this difference is compared across long- and short-tenured CEOs (univar- iate diff-in-diff analysis). Each performance measures are defined in the Ap- pendix. When the means of the sub-samples (diff-in-diff values) are significantly different at the 1%, 5%, or 10% significance level using t-test, they are indicated by *** , **, or * , respectively. 4.2.1. Sudden deaths as an instrument for the length of CEO tenure Building on the findings of Combs et al. (2007), Salas (2010), Nguyen and Nielsen (2014), Jenter et al. (2016), Quigley et al. (2017), and Brochet et al. (2021), we use sudden deaths of CEOs as another instru- mental variable that cuts short the planned tenure of the CEO. As argued in Brochet et al. (2021) and Jenter et al. (2016), this instrument should alleviate the identification concerns about the endogenous CEO-firm match and the resulting long tenure of a CEO. Sudden deaths occur randomly and exogenously to the current firm’s conditions (suicides are excluded from the sudden death sample). Since the CEO dies on the job, their tenure is cut shorter than intended and thus, its length is affected by these deaths (relevance condition of an instrument). Also, since such incidents are unexpected, they are unlikely to coincide with any other deliberate managerial changes that are independent of this event.15 long tenure and post-turnover Thus, the relationship between 15 All the top management changes that accompany the arrival of a new CEO are arguably occurring because of this sudden death of the predecessor. Thus, they can be considered as direct consequences of this event. performance is affected exclusively by the subsequent firm activities triggered by the sudden death of the CEO (i.e., the relationship is affected exclusively through this instrument). Another advantage of sudden deaths is that the new CEO is chosen independently of the preceding CEO’s opinion. Thus, the likelihood of long-tenured CEOs choosing their successors using different criteria than short-tenured ones is drastically reduced. In such cases, the incumbent CEOs have the power to comb through the new candidates and choose the one that fits their plans the most. External forces impose both the timing of the turnover and who would be the new CEO. Within our sample of 2428 turnovers, we manually identify the cases involving the sudden deaths of CEOs by closely following the method in Jenter et al. (2016) and Quigley et al. (2017).16 We find 37 cases that involve sudden CEO death.17 A dummy variable, CEODeathTurnoversi, is our second instrument to be used in the first-stage of the self-selection model.18 Of these sudden deaths, 19 belong to the short-tenured CEO subsample and 18 belong to long-tenured subsample. The difference between the mean CEODeathTurnoversi across the two subsamples is statistically insignificant that indicates the CEO deaths occur equally frequently in both the long- and short-subsamples. Similarly, while the correlation between age and sudden death is positive, its magnitude is only 0.07 (significant at the 10% statistical level) that means the sudden deaths do not necessarily have to occur at an older age; by definition, they are sudden. As suggested in Section 6 of Wooldridge (2002), to instrument for the interaction term, Afterit*LTPi, we create a new instrument from the interaction of Afterit and CEODeathTurnoversi. This interacted instrument (Afterit*CEODeathTurnoversi) is used together with our main instrument, CEODeathTurnoversi in the first stage of our model. 4.2.2. Age when the CEO took office: an instrument for tenure length The economic reasoning (relevance criteria) for constructing our second instrument is sound: if a manager became a CEO at a very old age, then biologically having a substantially long tenure is more diffi- cult. The decrement theory of aging (Giniger et al., 1983; Verhaegen and Salthouse, 1997) says that after a certain age, human cognitive abilities decline and performing their duties becomes very difficult for managers. Conversely, age also reduces career ambition and the overall motivation of managers (Rhodes, 1983). Thus, managers are marginally less likely to have very long tenures if they became a CEO at an older age. Such a variable should also satisfy the exclusivity criteria. Since in the post-turnover period the old CEO is no longer in power, it is unlikely for the age at which that person became CEO to affect the post-turnover performance through any other channel than the tenure of the prede- cessor CEO (i.e., LTPi). Specifically, when constructing such an instrument, we use the Execucomp database to determine the age at which the predecessor 16 We classify a CEO death as a “sudden death” if it is unexpected and not preceded by any indication of bad health. Sudden death can occur due to plane crashes, heart attacks, stroke, etc. We read the related news articles around the death of a CEO and textually search for the words “sudden” or “unexpected” (see Quigley et al., 2017). If these words are mentioned, then we carefully read all the preceding articles about that CEO’s health to verify that they do not mention any prior poor health in which case they are classified as slow deaths (see Jenter et al., 2016) and are excluded from our sample. 17 Jenter et al. (2016) identify 162 CEO sudden deaths between 1980 and 2012. Our sampling period is shorter, (1993–2013), and our sample covers contains large-capitalization firms covered by ExecuCOMP (the S&P 1500 firms). Thus, the low number of 37 sudden deaths between 1993 and 2013 is consistent with the rarity of such events. 18 Note that using several instruments for one endogenous variable is a desirable situation in 2SLS regressions, as they create over-identified condi- tions. Even if one of the instruments is deemed irrelevant, the remaining in- struments are sufficient for proper identification (see Wooldridge, 2002, subsection 5.1.2). JournalofFinancialStability63(2022)1010728 G. Colak and E. Liljeblom became the CEO of the focal company that experiences a turnover. We refer to this variable as Age Became CEOi.19 Again, to instrument for the interaction term, Afterit*LTPi, we create a new variable from the inter- action of Afterit and AgeBecameCEOi (denoted as Afterit*Age Became CEOi) to be used together with the other instruments. Thus, we use a total of four instruments in the first-stage estimation. To verify the quality of the instruments, we conduct several addi- tional tests. We implement the Stock and Yogo (2005) weak instrument test to find the potential bias introduced by our instruments, and then we apply the recommended cutoffs in that study. We complement this test with Staiger and Stock’s (1997) test for under-identification (in- struments are irrelevant). 4.3. Matching estimators and entropy balancing To further strengthen our identification, we also use alternative methods that are shown to reduce endogeneity bias. The first method involves matching estimators such as with propensity score matching (Dehejia and Wahba, 2002) or with bias-adjusted matching (Abadie and Imbens, 2006). Studies have successfully used these techniques before to improve identification in the cases when there are no reliable IV in- struments (Colak and Whited, 2007). Thus, our matching estimator tests should provide a robust alternative to our IMR and IV estimations. More recent studies in accounting (e.g., Shipman et al., 2017) have argued that traditional matching estimators, such as propensity score matching (PSM), are inferior to the more advanced entropy balanced matching approach of Hainmueller (2012). This entropy matching weighs each observation in the control sample such that the post-weighting distribution of each matching control variable (cova- riates) for the treatment and control samples are identically distributed. This is essentially a reweighting scheme of the control sample by applying new weights to each observation in that sample so that the distribution moments (mean, variance, skewness, etc.) of the covariates are identical across treatment and weighted control observations. This process is called covariate rebalancing, and studies have shown that it reduces endogeneity due to misspecification or due to a missing variable problem (see Chahine et al., 2020; McMullin and Schonberger, 2020). We essentially implement all three methods, the Abadie-Imbens matching, the PSM, and the entropy balancing, by using the LTP dummy to determine the treated and control samples, which are later used in the matching process. Thus, these matching techniques together with our IV regressions should substantially increase the quality of the identification in our analyses. Section B of our Online Appendix provides additional details on the implementation of these techniques. 5. Results 5.1. Post-turnover performance and predecessor’s tenure The first column of Table 3 presents the first-stage regression results from the self-selection model described in Section 4.2. Both of our in- struments (CEODeathTurnovers and AgeBecameCEO) negatively affect tenure length (LTP) that indicates the age and the sudden deaths indeed shorten their tenures (see first column under Stage 1 regressions). The interacted instruments (After*AgeBecameCEO and After*CEODeathTurn- overs) negatively affect the interaction term After*LTP (see second col- umn under Stage 1 regressions). The inverse Mills ratios (IMR1 and IMR2) are calculated using this first stage (Eqs. 2 and 3) and are added as controls in the second stage (Eq. 4). Table 3 also gives the results of the second-stage regression. The post-turnover Tobin’s Q is significantly negative, on average, for all CEO changes (coefficient for After). The unconditional value for the long-tenured predecessor dummy (LTP) is 19 A quick analysis discloses that the median age at which the predecessor became a CEO is 52 years old. significantly positive, that is, the more valuable firms keep CEOs for longer periods. However, the interaction variable After*LTP is negative and significant that indicates a substantial deterioration in the post- turnover valuation when LTP equals one. The relationship is strong despite the application of the diff-in-diff procedure, firm fixed effects, and the instrumental variable estimation. When BHAR is our performance variable, we again see that After*LTP is statistically and significantly negative that shows the firms under- perform after long-tenured CEOs’ are replaced. The coefficients for the other relevant variables (After and LTP) are similar in sign and signifi- cance to the Tobin’s Q regressions. A similar pattern is observed when ROA is used as a measure of firm performance, except that the coeffi- cient for After is significantly positive. This coefficient means that unlike Tobin’s Q and BHAR36, the ROA measure improves after a typical CEO turnover, which is consistent with Huson et al. (2004). Most importantly for us, the long-tenured CEOs are associated with a significantly more negative post-turnover stock performance (After*LTP). In the second-stage regressions, the coefficient for IMR1 (calculated from LTP’s probit regression) indicates how big the self-selection bias is without the first-stage adjustment. In some regressions the coefficient for LTP is downward biased (negative selection has occurred) and in some regressions it is positively (upward) biased. In the case of IMR2 (calculated from the interaction term’s probit, After*LTP), the selection bias on the interaction term, After*LTP, can be positive or negative depending on which dependent variable is used in the regression. To verify the quality of our instruments, we apply the Stock and Yogo (2005) weak instrument test. The worst case (largest bias) introduced by our instruments is less than 10% in all the regressions. The value statistic for the F-test for joint significance of the excluded instruments (in- struments are irrelevant) vary from 20.41 to 44.97, all of which are firmly above 10 (the rule-of-thumb cutoff proposed by Staiger and Stock, 1997). Thus, our tests do not suffer from a weak instrument. To further strengthen our identification, we implement several additional identification methods. First, we implement the entropy balancing technique of Hainmueller (2012), whereby the LTP dummy is used to determine the treated and control samples. These results are presented in Table B1 of the Online Appendix. Second, we implement PSM and Abadie-Imbens matching using the various CEO and firm characteristics as matching variables. These matching variables are the same controls used in Table 3. Table B2 presents these results. Third, we implement a 2SLS estimation with a continuous tenure variable, Prede- cessorTenurei (Online Appendix, Table B3). All of these alternative esti- mation methods yield results that are consistent with our conclusions from Table 3. 5.2. The need for cleanup and predecessor’s tenure; internal vs. external CEOs Next, we focus on whether new CEOs engage in serious cleanups if their predecessors had long tenures. Panel A of Table 4 has the estimates of the relationships between LTP and Restructuring Costs (in column 1) or Write-offs (in column 2). Again, a before-and-after approach is used together with firm and time fixed effects. Only the results from the second-stage regressions are reported. The first-stage regression is identical to the one shown in Table 3. The control variables are the firm characteristics that should affect a firm’s decision to conduct a major cleanup: log (Sales), Leverage, Cash Holdings, CAPX, and Ln (SEGN). In general, long-tenured CEOs experience fewer restructuring costs and asset write-offs in their last few years (coefficient for LTP is significantly negative in both columns). However, once these CEOs are replaced, the new CEO engages in relatively more serious cleanups; the coefficient for the interaction term, After*LTP, is positive and statically significant at the 5% level for both Restructuring Costs and Write-offs. Since all the estimations are conducted as a diff-in-diff and the endogenous firm-CEO match is taken into consideration, these results show that, ceteris par- ibus, the CEOs following a long-tenured predecessor implement deeper JournalofFinancialStability63(2022)1010729 G. Colak and E. Liljeblom Table 3 Long-Tenured Predecessors and Subsequent Firm Performance. VARIABLES After Long-Tenured Predec. (LTP) After * LTP Log(Sales) Leverage Cash Holdings ROA NetPPE CAPX RND Dividend Dummy Acquisitions Log(SEGN) Constant AgeBecameCEO CEODeathTurnovers After*AgeBecameCEO After*CEODeathTurnovers IMR1 (for LTP) IMR2 (for LTP*After) Observations R-squared Firm F.E. Year F.E. Stage 1 Probit1 (LTP) Probit2 (LTP*After) – – – 0.0075 [0.421] -0.0832 [0.301] 0.0890 ** [0.011] 0.7926 *** [0.000] 0.2681 *** [0.002] -0.8692 ** [0.013] 0.1435 [0.503] -0.1073 *** [0.000] 1.0438 *** [0.000] 0.0694 *** [0.000] 1.6994 *** [0.000] -0.0366 *** [0.000] -0.2359 * [0.072] 0.0000 [0.948] 0.0639 [0.733] – – 10,886 0.078 NO NO – – – 0.0173 [0.200] -0.2473 ** [0.030] 0.0207 [0.683] 0.2938 * [0.061] 0.3130 ** [0.012] -1.6994 *** [0.002] 0.2655 [0.359] -0.0810 ** [0.047] 0.5652 [0.105] 0.0760 *** [0.003] 1.7084 *** [0.000] -0.3156 *** [0.000] -3.6902 *** [0.000] -0.2790 *** [0.000] -3.5120 *** [0.000] – – 10,886 0.078 NO NO Stage 2 Tobin’s Q -0.2057 *** [0.001] 0.2283 *** [0.000] -0.0800 ** [0.013] -0.2546 *** [0.000] -0.4673 *** [0.000] 0.3955 *** [0.000] 1.9438 *** [0.000] -0.7737 *** [0.000] 4.0979 *** [0.000] 3.4033 *** [0.000] 0.0672 * [0.064] 0.1186 [0.465] -0.0568 ** [0.038] -0.3136 *** [0.000] – – – – -0.5900 *** [0.000] 0.3082 *** [0.000] 10,868 0.707 YES YES BHAR36 -0.1336 * [0.058] 0.1751 *** [0.000] -0.1688 *** [0.000] -0.1036 *** [0.002] -0.8767 *** [0.000] 0.2929 *** [0.000] 2.5514 *** [0.000] -2.3384 *** [0.000] 4.6860 *** [0.000] 0.2168 [0.605] -0.0110 [0.789] 1.3727 *** [0.000] -0.0422 [0.145] -0.1642 ** [0.049] – – – – -0.1783 [0.163] 0.2114 *** [0.007] 10,876 0.374 YES YES ROA 0.0185 ** [0.022] 0.0155 *** [0.000] -0.0142 *** [0.000] 0.0432 *** [0.000] -0.2391 *** [0.000] 0.0344 *** [0.000] – -0.1593 *** [0.000] 0.3017 *** [0.000] -0.6695 *** [0.000] 0.0004 [0.935] 0.0724 *** [0.000] -0.0140 *** [0.000] 0.0334 *** [0.003] – – – – 0.1060 *** [0.000] -0.0239 ** [0.012] 10,886 0.572 YES YES The table analyzes firm performance after a long-tenured CEO is replaced. The firm underperforms following a long-tenured predecessor’s replacement with a new CEO. Long-Tenured Predecessor (LTP) is a dummy equal to 1 if the predecessor CEO had a tenure longer than median tenure in our sample (which is 7 years). The firm years included in the sample correspond to the years leading up to the turnover ([(cid:0) 3,(cid:0) 1]) and the years after the turnover event ([+1,+3]). The year 0 (the turnover year) is excluded from the sample. After is a dummy indicating the years following the turnover event (i.e., [+1,+3]). The firm-CEO match is treated as endogeneous, and the sudden CEO deaths (CEODeathTurnovers) and the age the predecessor became a CEO (AgeBecameCEO) are used as instruments for LTP in the self-selection model. As suggested in Section 6.2. of Wooldridge (2002), we also use the interacted terms of these two instruments with After as instruments for the interaction term, After*LTP, and thus we have two Stage 1 equations (Probit) and four instruments. The Inverse Mills Ratios from first probit equation in Stage 1 (for LTP) is named IMR1 and the one from second probit equation (for After*LTP) is named IMR2. They represent the size of self-selection bias in OLS coefficients. Firm performance measures in Stage 2 equations are Tobin’s Q, ROA, and BHAR36. A sample of CEO turnovers (2428 cases) between 1993 and 2013 is used. The control variables are described in the text and in Appendix A. All continuous variables are winsorized at the 1st and 99th percentiles. Coefficients’ p-values shown in [square] brackets are calculated using robust standard errors. Year and Firm fixed effects are applied whenever indicated at the bottom. *** , **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. JournalofFinancialStability63(2022)10107210 G. Colak and E. Liljeblom Table 4 The Need for Clean-up: Long-Tenured CEOs, Restructurings, and Write-offs. Panel A: Entire Turnover Sample (1) (2) Self-selection Model - Stage2 Self-selection Model - Stage2 VARIABLES After Long-Tenured (LTP) After * LTP Log(Sales) Leverage Cash Holdings CAPX Log(SEGN) Constant IMR1 (for LTP) IMR2 (for LTP*After) Observations R-squared Firm F.E. Year F.E. Restructuring Costs -0.0005 * [0.093] -0.0142 *** [0.000] 0.0022 *** [0.007] -0.0018 *** [0.000] 0.0032 *** [0.001] 0.0007 [0.244] -0.0040 [0.165] 0.0005 ** [0.048] 0.0083 *** [0.000] -0.0008 * [0.062] 0.0166 *** [0.000] 11,706 0.395 YES YES Writeoffs -0.0032 [0.119] -0.0042 ** [0.021] 0.0075 *** [0.001] -0.0021 [0.149] 0.0373 *** [0.000] -0.0105 *** [0.000] 0.0137 [0.438] 0.0033 ** [0.015] -0.0097 * [0.099] 0.0010 [0.748] 0.0000 [0.999] 11,706 0.262 YES YES Panel B: Internal vs. External CEOs (Internal CEO=1 vs. Internal CEO=0) (1) Internal CEOs (2) External CEOs (3) Internal CEOs Self-selection Model - Stage2 Self-selection Model - Stage2 Self-selection Model - Stage2 VARIABLES After LTP After * LTP Log(Sales) Leverage Cash Holdings CAPX Log(SEGN) Constant IMR1 (for LTP) IMR2 (for LTP*After) Observations R-squared Firm F.E. Year F.E. Wald Test (F-stat, p-val) Restruct. Costs 0.0003 [0.484] -0.0112 *** [0.000] 0.0000 [0.979] -0.0024 *** [0.000] 0.0056 *** [0.000] 0.0000 [0.953] -0.0048 [0.137] 0.0004 [0.169] 0.0066 *** [0.000] 0.0001 [0.872] 0.0186 *** [0.000] 6334 0.413 YES YES 26.31 (0.0009)*** Restruct. Costs -0.0006 [0.210] -0.0235 *** [0.000] 0.0037 *** [0.009] -0.0013 *** [0.000] 0.0011 [0.375] 0.0012 ** [0.014] -0.0039 [0.364] 0.0001 [0.782] 0.0137 *** [0.000] -0.0015 * [0.052] 0.0095 [0.321] 5372 0.420 YES YES Writeoffs -0.0012 [0.631] -0.0006 [0.832] 0.0046 [0.115] -0.0013 [0.445] 0.0467 *** [0.000] -0.0053 * [0.068] 0.0008 [0.968] 0.0030 * [0.078] -0.0049 [0.419] -0.0005 [0.892] 0.0018 [0.934] 6334 0.279 YES YES 17.73 (0.0428)** (4) External CEOs Self-selection Model - Stage2 Writeoffs -0.0047 [0.116] -0.0083 *** [0.010] 0.0110 *** [0.005] -0.0025 [0.230] 0.0363 *** [0.000] -0.0130 *** [0.000] 0.0278 [0.274] 0.0029 [0.179] -0.0222 *** [0.001] 0.0038 [0.442] 0.0072 [0.900] 5372 0.274 YES YES The table shows the size of clean-ups (restructuring charges and asset write-offs) after a long-tenured CEO is replaced with a new CEO. The need for clean-up in the post-turnover period is higher when the predecessor is a long-tenured CEO. Panel A uses the entire sample of CEO turnovers (2428 cases) between 1993 and 2013 to analyze how predecessor’s tenure length affects the activities of the subsequent clean-up measures. Panel B partitions this turnover sample based on whether the new (incumbent) CEO is an outsider to the firm or s/he is one of the top executives of the firm (as listed in ExecuCOMP data in the year prior to the turnover). If the incumbent CEO is chosen from the internal talent pool of the firm, then Internal CEO = 1, otherwise Internal CEO = 0. There are 1312 (1116) turnover events that involve internal (external) CEOs. Panel A displays the estimation results from the second-stage of the self-selection model. The first-stage regression of this self- selection model is the same as in Table 3 and are not shown here for brevity. In Panel B, both regressions are from the second stage with the same specifications as in Panel A. A Long-Tenured Predecessor (LTP) is a dummy equal to 1 if the predecessor of the current CEO had tenure longer than median tenure in our sample (which JournalofFinancialStability63(2022)10107211 G. Colak and E. Liljeblom is 7 years). The firmyears included in the sample correspond to the years leading up to the turnover ([(cid:0) 3,(cid:0) 1]) and the years after the turnover event ([+1,+3]). The year 0 (the turnover year) is excluded from the sample. After is a dummy that indicates the years following the CEO turnover event (i.e., [+1,+3]). The dependent variables are Restructuring Costs and Writeoffs. The control variables are the possible firm characteristics that ar expected to affect a firm’s decision to conduct a major cleanup: log (Sales), Leverage, Cash Holdings, CAPX, and Ln (SEGN). All the variables are defined in Appendix A. The methodology is diff-in-diff regressions, whereby the years after a CEO’s turnover are compared to the years leading up to the turnover event and then, this difference is compared across long- and short-tenured CEOs. All the continuous variables are winsorized at the 1st and 99th percentiles. At the bottom of Panel B we present the results (F-stat and p-values) from the Wald Test for significant differences between the regression coefficients of the internal vs. external CEO subsamples. The coefficients’ p-values shown in [square] brackets are calculated using robust standard errors. Year and Firm fixed effects are also applied whenever indicated at the bottom of the column. *** , **, and * indicate sig- nificance at the 1%, 5%, and 10% levels, respectively. restructuring and more serious earnings management to “wipe the slate clean” relative to other CEOs.20 These findings are novel to the litera- ture, as they clarify that the reported restructurings (Weisbach, 1995; Barron et al., 2011) and earnings management (Pourciau, 1993; Elliott and Hanna, 1996; Hazarika et al., 2012; Ali and Zhang, 2015) that are implemented after CEO turnovers are primarily driven by the subsample of long-tenured CEOs. Again, these CEOs tend to do some operational damage to the firm in their last years of management (Huson et al., 2004; Brochet et al., 2021), and thus the new CEOs feel the need to take more drastic measures. Our tests show that the potential bias introduced by our instruments is less than 10% for the regressions in both columns (1) and (2). The F- statistic for joint significance of the excluded instruments is 25.67 in column (1) and 31.12 in column (2), both of which are well above the cutoff of 10 (Staiger and Stock, 1997). Thus, our instruments perform well in the cleanup regressions as well. In a related analysis, our incoming CEOs are sorted into internal versus external (see Panel B of Table 4). If a newly appointed CEO was listed among the top five executives of the focal firm at least once during this firm’s history in Execucomp, then this manager is considered an internal CEO (external otherwise).21 We conjecture that if an incoming CEO is appointed internally, they are more likely to be a follower of (influenced by) the departing CEO and thus, they are unlikely to take drastic measures that will ruin the legacy of the departing CEO. The externally enforced turnovers (e.g., sudden deaths) are less likely to have such disciple-mentor relationships, and thus we again use these instruments for better identification. We focus on Restructuring Costs and Write-offs as dependent variables,22 because these activities involve some degree of undoing and reversing of the policies of the departing CEOs. The results in Panel B of Table 4 show that indeed the external CEOs are associated with more significant restructurings and write-offs. The coefficient for the interaction term, After*LTP, is statistically significant 20 It is important to emphasize the reverse causality here. Lehn and Zhao (2006) find that CEOs who make bad acquisitions are 47% more likely to be replaced in the five years following the acquisition decision. Thus, the decisions of the CEO can shorten their tenure and at the same time the subsequent restructuring or divestment by the new CEO may have little to do with pre- decessor’s tenure, but it could simply be an artifact of undoing prior bad acquisition decisions. This issue –which is sometimes referred to as the endo- geneity of CEO tenure and subsequent restructuring decision– is resolved in our setting i) by using the dif-in-dif approach, ii) by using a 2SLS estimation, and iii) by using matching estimators (i.e., CEO tenure is matched across firms so that bad acquisition firms are matched with non-bad acquisition ones). 21 The exceptions are the cases when the incoming CEO was brought in the company a few months (i.e., less than 12 months) before the turnover date. These cases are usually considered as part of the succession plan (Ertimur et al., 2018) and are also treated as an external CEO as long as that incoming CEO was never a part of the company as a lover level manager. We rely on Execucomp to determine whether a manager was involved with the company in the past. 22 As discussed in Adut et al. (2003) and Ali and Zhang (2015), not all firms conduct restructuring and write-offs after the CEO turnover, however many of them do. In Table 4, we have 5601 observations that have non-zero REST and 4144 observations with non-zero write-offs. Thus, we work with quite a large sample of restructuring and write-offs. (at 5% or better levels) only for the externally appointed subsample of new CEOs. The results from the Wald test (at the bottom of the panel) show that there are significant differences between the regression co- efficients for the internal versus external CEO subsamples. This last result is important as it rules out the possibility that the firm’s post-turnover performance is driven by the change in the mana- gerial style of the new CEO and not by the replacement of the long- tenured predecessor. When the new CEO is internally appointed, the influence of their predecessor is likely very high, and the two executives share similar managerial styles. However, in the external CEO subsam- ple, the managerial styles almost certainly change after each turnover. Therefore, in Panel B of Table 4 we basically conduct an analysis of long- versus short-tenured CEOs within that subsample where all the turn- overs trigger a managerial style change. However, only the long-tenured ones implement a larger cleanup (restructuring costs and write-offs). Thus, a change in managerial style is not the only driver of the results reported in Table 4; the size of the cleanup depends also on the length of the predecessor’s tenure. Put differently, LTP*After has its own explan- atory power above and beyond the changing managerial style around the turnovers (i.e., our results still hold even when we condition on changes in managerial styles). To summarize, the results from our analyses so far suggest that the post-succession firm behavior is significantly related to the tenure of the preceding CEO. Specifically, when replacing a long-tenured CEO, there are significant drops in Tobin’s Q, ROA, and BHAR. The cleanups are also more serious and much larger as indicated by the accounting charges, such as restructuring costs and asset write-offs. These results are in line with Hypotheses 1 A and 1B. However, what our results cannot establish is whether the changes in post-turnover firm behavior are due to a) the successor being less skilled than the preceding CEO, or b) the firm in the latter years of the long-tenured CEO being left in a shape that requires more updating in terms of strategy and focus. The latter interpretation is in line with many theories in psychology and management (Murphy, 1989; Hambrick and Fukutomi, 1991; Acker- man, 1992; Sturman, 2003) as well as indirectly supported by some empirical results (Shen and Canella, 2002; Brochet et al., 2021; Jenter et al., 2016). 5.3. Duration of the recovery and predecessor CEO’s tenure Next, we run a hazard rate model to analyze the post-turnover per- formance of the firm conditioned on the tenure of the preceding CEO. We ask the question, does the firm recover after the turnover and at what speed? Is the speed of recovery related to the tenure of the predecessor? We essentially assess whether our results are either due to a) an above- average (very good) long-tenured CEO being replaced by an average one, or b) a CEO with deteriorating performance has to be replaced by a new one that would restructure the firm. Under alternative a), the lower performance might be the new equilibrium. Under b) there should be more restructuring needed, and the recovery would be slower in that case. We create a variable MaxPQ that is defined as the maximum observed annual Tobin’s Q under the preceding CEO during a five year pre-turnover window [(cid:0) 5,(cid:0) 1]. We, then, conduct a duration analysis (Cox Hazards Rate model) and test for the determinants of reaching that maximum Q during any of the five years after the turnover (excluding JournalofFinancialStability63(2022)10107212 G. Colak and E. Liljeblom Table 5 Long-Tenured Predecessors and Post-Succession Recovery Times. VARIABLES LTP Log(Sales) Leverage Cash Holdings ROA NetPPE CAPX RND Dividend Dummy Acquisitions Log(SEGN) Constant Industry F.E. Observations/ (1) Reached_PQ (3) Reached_PQ (2) Reached_PQ (4) Reached_PQ -0.1396 ** [0.044] -0.1476 ** [0.036] -0.1396 ** [0.044] NO -0.1476 ** [0.036] YES -0.2596 *** [0.000] -0.0502 * [0.065] 0.6762 *** [0.003] 0.0136 [0.899] 1.4767 *** [0.000] 0.0196 [0.934] 3.4181 *** [0.001] 1.6008 ** [0.050] 0.4688 *** [0.000] -2.0472 ** [0.022] -0.0846 [0.106] -0.2596 *** [0.000] NO -0.2760 *** [0.000] -0.0456 [0.135] 0.6634 *** [0.008] 0.0332 [0.764] 1.6154 *** [0.000] -0.2418 [0.472] 4.8141 *** [0.000] 2.7008 *** [0.005] 0.3725 *** [0.000] -2.0147 ** [0.028] -0.1100 * [0.057] -0.2760 *** [0.000] YES Numb. of Firms 2428 2428 1151 1663 1663 785 785 1151 Numb. of firms with LTP= 1 Percent of LTP= 1 firms that Reached_PQ= 1 Percent of LTP= 0 firms that Reached_PQ= 1 32.84% 32.84% 41.40% 41.40% 36.41% 36.41% 49.43% 49.43% The table analyzes how long it takes for the firm value (Tobin’s Q) to recover from the CEO turnover event. Table 3 reports that the firm valuations are on average lower after CEO turnover events. This table conducts duration analysis (using Cox Hazards Rate Model), whereby we calculate the time between the turnover year and the year when Tobin’s Q finally (if at all) reaches the pre- succession levels. The sample of turnovers is stratified using Long-Tenured Pre- decessor (LTP) dummy (=1 if the predecessor has tenure longer than the median tenure, which is 7 years). We use a sample of CEO turnovers (2428 cases) to conduct the duration analysis. To better estimate the duration times, we use a wider time window around the turnover events, [(cid:0) 5;+ 5]. The period before the CEO turnover is used to determine the maximum observed annual Tobin’s Q (MaxPQ) achieved under the predecessor during the five year pre-turnover window [(cid:0) 5;(cid:0) 1]. Once this maximum value is determined, the pre-turnover years are dropped from the estimation sample. Thus, we focus on the firm- years that are in the post-succession period [+ 1;+ 5]. Then, for each of the 2428 turnover events we determine during which of the post-turnover years the current Tobin’s Q reaches MaxPQ (indicated by the dummy Reached_PQ = 1). If, for example, this level is reached in year t = +3 (i.e., the hazard event occurs in year +3), then for that turnover event the duration time is 3 years and Reach- ed_PQ takes a value of 1. If for a given turnover event MaxPQ level is never reached, then we assume Reached_PQ = 0, and the duration time is censored at t = +5. The Cox Hazard Rate Models are estimated using the dependent vari- able, Reached_PQ, and the control variables are as in Table 3. All the variables are defined in Appendix A. The shown values are the coefficient estimates (i.e., not the hazard rates). The p-values shown in [square] brackets under the coefficients are for the z-statistics. Industry (FF48) fixed effects are applied whenever indi- cated. *** , **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. At the bottom of the table, we provide some univariate statistics for each subsample (stratified by LTP dummy) related to the number and proportion of the firms that have Reached_PQ = 1 (i.e., percent of firms for which “the hazard” occurred). the event year). In such a model, we use Reached PQ equals one if “the hazard event” occurred during any of the post-turnover years [+ 1,+ 5], and we record the year when the MaxPQ was reached (e.g., in year t = +3) as the duration period for that turnover. If for a given turnover the MaxPQ level is never reached, then we assume Reached_PQ equals zero, and the duration time is censored at t = +5. As the explanatory variables in our hazard rate model, we use various control variables plus the LTP dummy that is our variable of interest.23 The results of our duration analysis are reported in Table 5. The LTP dummy is significant in all model specifications at the 5% level and has a negative sign that indicates recovery from long-tenured CEOs is significantly slower. This result gives some support to the alternative explanation b) above. The results remain qualitatively the same when we use ROA instead of Tobin’s Q.24 6. Testing for alternative explanations We further analyze the explanation for firm underperformance in the period after the succession of the long-tenured CEOs. We focus on the two main explanations that relate to Hypotheses 2 and 3. The first explanation is that the firm’s corporate governance was too weak, the corporate board was too ineffective, and consequently the CEO became too powerful. This lack of governance allowed the CEO to stay in power for a longer time that in turn, exacerbated the agency problems within the firm. In such cases, the cleanup would likely take longer, as there may be deeper issues to solve, and the observed post- turnover underperformance would reflect the major agency problems that resurfaced once the CEO left office. Support for Hypothesis 2 con- cerning a stronger effect of post-succession negative performance within weakly governed firms would support the alternative explanation b) in which the firm in the latter years of the long-tenured CEO was left in a shape that requires more updating in terms of strategy and focus. The second explanation is based on the predecessor’s merit. Under the alternative explanation a) in which the successor is less skilled than the preceding CEO, we would expect a negative relationship between the preceding CEO’s skill and post-succession performance (highly skilled, worse post-succession performance). If the preceding CEO was indeed very successful in managing the firm, it is natural that the corporate board would want to use the skills and the expertise of this CEO for as long as possible. This success would be reflected in the clear over- performance of the firm during the tenure of this “superstar” CEO, while the observed underperformance in the post-turnover period would be a mere reflection that the new CEO (recruited from a limited talent pool) would not be able to “fill the shoes” of their predecessor. If a positive relationship (low-skilled predecessor, low post-succession performance) is found, we obtain support for the alternative explanation b); that is, more restructuring is needed after a longer time of mismanagement after a poor CEO. A positive relationship would in this way be in line with Hypothesis 3 and with Murphy’s (1989) learning model. That is, there is cross-sectional variation related to personal and motivational factors, and a better performing preceding CEO is likely to leave the firm in better shape. Under this circumstance, a mediocre successor would not meet the preceding performance standards, and the firm would appear to underperform after the succession. The results from these two tests are reported in subsections 6.1 and 6.2. 23 Since only the post-turnover years are considered to conduct this duration analysis, the After dummy and the interaction term, After*LTP, are naturally dropped. 24 The economic interpretation of BHAR36, being a month-over-month accu- mulated stock return, makes it economically difficult to interpret the hazard rate estimation using this variable. So, we did not conduct such an estimation. JournalofFinancialStability63(2022)10107213 G. Colak and E. Liljeblom Table 6 Testing for the Corporate Governance Explanation. 6.1. Corporate governance and agency costs explanation F-stat (p- value) 27.23 (0.26) 26.13 (0.02)** 37.42 (0.01) *** 19.51 (0.05)** 22.54 (0.04)** Variables Corporate Governance (CG) Strong CG Self- selection - Stage2 Weak CG Self- selection - Stage2 Tobin’s Q After LTP After * LTP BHAR36 After LTP After * LTP ROA After Restruct. Costs LTP After * LTP After LTP After * LTP Writeoffs After LTP After * LTP Numb. of Obs. Controls Firm F. E. Year F. E. -0.2812 ** [0.025] 0.2778 *** [0.000] -0.1506 * [0.090] -0.1481 [0.327] 0.1240 [0.150] -0.0977 [0.165] 0.0328 ** [0.045] 0.0029 [0.754] -0.0062 [0.414] 0.0028 * [0.056] -0.0024 *** [0.004] 0.0018 [0.108] -0.0102 [0.502] 0.0087 [0.315] 0.0107 [0.129] 3256 YES YES YES -0.2354 *** [0.008] 0.1910 *** [0.001] -0.0638 * [0.086] -0.2043 * [0.063] 0.1230 * [0.077] -0.2349 *** [0.000] 0.0063 [0.472] 0.0155 *** [0.005] -0.0217 *** [0.000] -0.0013 [0.121] -0.0014 ** [0.011] 0.0010 ** [0.038] -0.0105 [0.148] -0.0102 ** [0.026] 0.0093 ** [0.018] 4573 YES YES YES The table analyses two subsamples formed on the basis of corporate governance quality. A firm’s corporate governance score is measured as the sum of three dummy variables capturing different aspects of corporate governance: Corporate Governance Score=Dual Role for CEO + Ineffective Board + Low Instit. Ownership. CEOs with dual role (i.e., CEO is also a board chairman; Chair=1) are arguably more powerful and thus more likely to stay longer in their position. We average this Chair dummy during the last three years of the predecessor CEO and compare it to the sample median. If it is above (below) the sample median, we assume the Dual Role for CEO= 1 (=0). The dummy Independent Board is equal to one if the firm’s board is constituted mostly (>50%) of independent directors during that year; zero otherwise. Again, we find the average of this variable over the last three years of the predecessor’s tenure, and if this average is above (below) the sample median, we assume the dummy Ineffective Board = 1 (=0). A firm is considered to have a low institutional ownership (Low Instit. Own- ership=1) if, on average, it had below median institutional ownership within its two-digit industry during the last three years of predecessor’s tenure; zero otherwise. A firm is classified as Strong Governance if its predecessor CEO’s governance score for a given year is equal to zero. The rest of the sample is classified as Weak Governance firms, i.e., the cases when either the CEO is too powerful, the board is too ineffective, and the external monitoring by the in- stitutions is poor. The dependent variables are Tobin’s Q, ROA, BHAR36, Restructuring Costs, and Writeoffs. The estimation is the self-selection model as in Eqs. (2), (3), and (4). The first stage regressions are the same as in Table 3 and are suppressed to save space. The controls are the same as in Tables 3 through 5 and are not reported for brevity. The last column reports the Wald test’s F-sta- tistics that coefficients are identical between the Strong CG and the Weak CG subsamples. All regressions use firm and year fixed effects. All the variables are defined in Appendix A. The p-values are reported in parentheses below the co- efficients. Statistical significance at the 1%, 5%, and 10% levels are marked with *** , **, and * , respectively. The role of weak corporate governance in exacerbating the harmful effects of a long tenure is the focus of our next test. In line with Hy- pothesis 2, we expect that firms with weak corporate governance are more likely to allow their CEOs to stay longer that in turn would lead to a higher need for restructuring and a more pronounced negative effect on post-turnover firm performance. We construct our governance measure using three key elements of corporate governance: CEO power (Finkelstein and D’aveni, 1994), board effectiveness (internal monitoring; Huson et al., 2001), and external monitoring by institutional investors (Parrino et al., 2003; Khan et al., 2017). These are among the most important governance issues that might affect the likelihood of a CEO having a tenure beyond the optimal length. CEOs with a dual role (i.e., CEOs that also serve as the head of the corporate board, Chair, equals one) are arguably more powerful and thus more likely to stay longer in their position. We average this dummy during the last three years of a long-tenured CEO and compare it to the sample median for all the 2428 turnover cases. If it is above (below) the sample median, we assume the variable Dual Role for CEO equals one and zero otherwise. Similarly, we consider a board to be ineffective if it is constituted mostly of dependent directors (>50%) during that year. Again, we find the average of this variable over the last three years of the predecessor’s tenure, and if this average is below (above) the sample median, we create a dummy Ineffective Board that equals one and zero otherwise. Next, a firm is considered to have weak external monitoring if it has little institutional ownership. Essentially, if on average, a firm has institutional ownership below the median within its two-digit industry during the last three years of the long-tenured CEO, then our dummy Low Instit. Ownership equals one and zero other- wise. Using these three indicators, we construct a governance score as follows: Governance Score = Dual Role for CEO + Ineffective Board + Low Instit⋅Ownership (5) Using this formula, each turnover in our sample is assigned a score depending on the governance characteristics during the predecessor’s last three years of tenure. A score of zero indicates the cases with the strongest corporate governance, while a score of three indicates that the firm has a powerful CEO, an ineffective board, and limited external monitoring. We sort all the turnovers based on their governance score. A firm is classified as having Strong Governance if its predecessor CEO’s governance score for a given year is equal to zero. The rest of the sample is classified as Weak Governance firms, that is, the cases when either the CEO is too powerful, the board is too ineffective, or the outside investors cannot properly monitor the CEO’s actions.25 Then, each of the analyses in Tables 3 and 4 are rerun again for the two subsamples: weak and strong governance subsamples. Table 6 presents the results from such a subsample analysis. For brevity, we show only the coefficients for After, LTP, and After*LTP. We also report the Wald test’s F-statistics that the regression coefficients are the same between the weak and strong governance subsamples. In the strong corporate governance subsample, the length of the preceding CEO’s tenure is not significantly related to reduced performance in the post-succession period (i.e., the interaction term After*LTP is not nega- tively significant with the possible exception being the result for Tobin’s 25 In a robustness test reported in our Online Appendix, we define Strong Governance firms as those who have an above median fraction of independent directors (i.e., Independent Board = 1). The results confirm that post-turnover underperformance is not observed among the firms with more independent boards. Similarly, in untabulated results we use the GIndex of Gompers, Ishii, and Metrick (2003) and the EIndex of Bebchuk, Cohen, and Ferrell (2009). These variables are missing for many of our observations, but nonetheless we find qualitatively similar results as in Table 6. JournalofFinancialStability63(2022)10107214 G. Colak and E. Liljeblom Table 7 Testing for the Merit-Based Explanation. 6.2. Predecessor CEO’s skills F-stat (p- value) 32.75 (0.05)** 343.78 (0.07)* 55.30 (0.02)** 27.76 (0.01)*** 19.12 (0.04)** Variables Predecessor’s Merit High Merit Self- selection - Stage2 Low Merit Self- selection - Stage2 Tobin’s Q After LTP After * LTP BHAR36 After LTP After * LTP ROA After Restruct. Costs LTP After * LTP After LTP After * LTP Writeoffs After LTP After * LTP Numb. of Obs. Controls Firm F. E. Year F. E. -0.2514 ** [0.030] 0.1841 *** [0.005] -0.0569 [0.176] -0.0984 [0.525] 0.1256 [0.151] -0.1312 [0.118] 0.0014 [0.917] 0.0071 [0.355] -0.0046 [0.528] 0.0004 [0.746] -0.0004 [0.520] -0.0003 [0.611] -0.0080 [0.386] 0.0025 [0.631] 0.0014 [0.776] 1861 YES YES YES -0.1646 ** [0.018] 0.2228 *** [0.000] -0.0722 ** [0.042] -0.1753 ** [0.036] 0.2274 *** [0.000] -0.2069 *** [0.000] 0.0182 ** [0.027] 0.0186 *** [0.000] -0.0163 *** [0.000] -0.0008 [0.231] -0.0014 *** [0.000] 0.0012 *** [0.001] -0.0103 [0.169] -0.0065 [0.106] 0.0109 *** [0.006] 8997 YES YES YES The table conducts regression analyses on two subsamples formed on the basis of predecessor CEO’s merit. A CEO could have long tenure due to exceptionally high skills (High Merit CEO). Each year the firms in COMPUSTAT database are ranked into deciles within their industry (SIC2) based on their cash flows scaled by assets (decile 10 has the firms with the highest cash flows and decile 1 the ones with the lowest). A predecessor CEO’s merit score is created using the performance of her company as reflected in the average cash flow decile of the firm in the past five year. If this merit score is higher than or equal to decile 9, then this firm had an exceptional performance during the tenure of the prede- cessor CEO and it is classified as High Merit CEO; otherwise the observation belongs to a Low Merit CEO. The dependent variables are Tobin’s Q, ROA, BHAR36, Restructuring Costs, and Writeoffs. The estimation is the self-selection model as in Eqs. (2), (3), and (4). The first stage regressions are the same as in Table 3 and are suppressed to save space. The controls are the same as in Ta- bles 3 through 5 and are not reported for brevity. The last column reports the Wald test’s F-statistics that coefficients are identical between the High Merit and the Low Merit subsamples. All regressions use firm and year fixed effects. All the variables are defined in Appendix A. The numbers in parentheses below the coefficients are the p-values. Statistical significance at the 1%, 5%, and 10% levels are marked with *** , **, and * , respectively. Q). There is no significant need to restructure or write-off assets either. However, in the weak governance subsample, the interaction term After*LTP is strongly significant that means the results we found in the previous section are primarily driven by firms with weak governance. The firms who are in the weak governance subsample and had a long- tenured CEO experience much lower post-turnover performance, and they have much greater restructuring and higher asset write-offs costs. Put differently, the post-turnover underperformance appears to be concentrated primarily in the firms with long-tenured CEOs with un- checked powers, which is in line with Hypothesis 2. Cleaning up after such CEOs seems harder too. Next, we analyze the effect of CEO skill as an alternative explanation for the aforementioned underperformance of the firms in the post- succession period. Could the post-succession effects reported in the previous subsection be related to the incumbent CEO being replaced by an average CEO? Or, alternatively, are these effects related to the departing CEO leaving the firm in better shape? We sort all observations into two subsamples based on the preceding CEO’s skill score. Each year the firms in the Compustat database are ranked into deciles within their industry (SIC2) based on their cash flows scaled by assets (deciles 1 and 10 have the observations with, respec- tively, the lowest and the highest cash flows). A predecessor CEO’s skill score is created using the performance of their firm as reflected in the average cash flow decile of the firm during the last five years. If this skill score is higher than or equal to decile 9, then this firm had exceptional performance during the tenure of the predecessor. Such turnover cases are classified as High Merit CEO and the rest of the observation belong to the Low Merit CEO subsample.26 A CEO could have a long tenure due to exceptionally high skills (High Merit CEO). We do not expect that such long-tenured CEOs do major damage to their firms during the latter years of their tenure. Conse- quently, in this subsample, there should not be major differences be- tween the long- and short-tenured CEOs in terms of the need for major “cleanups” and the post-succession underperformance. Instead, these differences will be more pronounced for the subsample in which the preceding CEO was not so successful (Low Merit CEO). Subpar CEOs who end up having longer tenures are more likely to create conditions that require major cleanups, and in such cases the new CEO will have a harder time turning the business around. Table 7 shows the regression analyses on the two subsamples formed on the basis of predecessor CEO’s skill when using the same setup as in Table 6. Again, we focus on the coefficients for After, LTP, and After*LTP. There is some evidence that the differences between the long- and short- tenured CEOs are more pronounced for the Low Merit CEO subsample. In that subsample, the length of the preceding CEO’s tenure is related to the reduced performance in the post-succession period (i.e., the interaction term After*LTP is significantly negative). There are significant costs associated with cleanup measures (restructurings and write-offs) as well. We consider these results as consistent with the claim that if the pre- decessor CEO had a skill-based reason for having a longer tenure, then the need for a cleanup afterwards is significantly less (Hypothesis 3 is true). The post-turnover underperformance appears to be concentrated mostly in firms where the predecessors end up having long tenures despite being average or subpar managers. 6.3. CEO ownership and firm founders In our next analysis, we consider the complications that arise when the preceding CEO is a major shareholder and/or the founder of the firm. In such special cases, many known and unknown considerations confuse the identification of the tenure’s effect on post-turnover performance (Morck et al., 1988; McConnell and Servaes, 1990; Himmelberg et al., 26 As a robustness test, we use Demerjian et al. (2012) measure of managerial ability that is based on firm efficiency measures rather than cash flows. In particular, we use their industry and year ranked (into deciles) version of the Managerial Ability measure, and again consider the CEOs in the highest two deciles in terms of ability as High Merit CEO. Our results, which are reported in the Online Appendix, are qualitatively the same as in Table 7: there is some support of Hypothesis 3. We also use the General Ability Index (GAI) of Cus- todio et al. (2013); however, this index (retrieved from the author’s website) is only available from 1993 to 2007 and our sampling period is 1993–2013. Nonetheless, our qualitative conclusions still hold using the GAI index for the period between 1993 and 2007. JournalofFinancialStability63(2022)10107215 G. Colak and E. Liljeblom Table 8 Firm Founders and High CEO Ownership. Panel A: Post-turnover performance (Tobin’s Q, BHAR36, and ROA) VARIABLES After Long-Tenured Predec. (LTP) After * LTP AgeBecameCEO CEODeathTurnovers After*AgeBecameCEO After*CEODeathTurnovers IMR1 (for LTP) IMR2 (for LTP*After) Controls Observations R-squared Firm F.E. Year F.E. Stage 1 Probit (LTP) – – – … -0.0378 *** [0.000] 0.1137 [0.437] -0.0001 [0.913] -0.0498 [0.809] – – YES 8720 0.078 NO NO Panel B: Post-turnover clean-ups (Restructuring Costs and Writeoffs) Probit (LTP*After) – – – … -0.3014 *** [0.000] 3.1560 *** [0.000] 0.2633 *** [0.000] -3.0775 *** [0.000] – – YES 8720 0.078 NO NO Stage 2 Tobin’s Q -0.1568 ** [0.015] 0.1829 *** [0.000] -0.0726 ** [0.043] … – – – – -0.2297 *** [0.003] 0.1267 [0.154] YES 8706 0.714 YES YES BHAR36 ROA 0.0590 [0.457] 0.1946 *** [0.000] -0.1478 *** [0.001] … – – – – -0.1174 [0.221] -0.1192 [0.277] YES 8712 0.367 YES YES 0.0275 *** [0.001] 0.0038 [0.423] -0.0070 * [0.090] … – – – – 0.0510 *** [0.000] -0.0384 *** [0.000] YES 8720 0.578 YES YES VARIABLES After Long-Tenured (LTP) After * LTP IMR1 (for LTP) IMR2 (for LTP*After) Controls Observations R-squared Firm F.E. Year F.E. Self-selection Model - Stage2 Restructuring Costs -0.0009 [0.199] -0.0011 *** [0.005] 0.0011 *** [0.004] … -0.0017 ** [0.035] 0.0008 [0.410] YES 8720 0.403 YES YES Self-selection Model - Stage2 Writeoffs -0.0128 * [0.054] 0.0001 [0.975] 0.0059 * [0.080] … -0.0258 *** [0.001] 0.0188 ** [0.041] YES 8720 0.218 YES YES The table analyzes the relationship between prolonged tenure and post-turnover performance after controlling for firm founders and high CEO ownership effects. To adjust for complications that arise when the CEO is a founder (and thus difficult to replace) or when the CEO is a major shareholder (higher than 2% ownership of total shares), we remove the turnover cases when the predecessor CEO is either a founder or a major shareholder. There are 463 turnover cases that involve such “special” CEOs (55 of these are founders). We focus on the remaining 1965 turnovers (11,790 firm-years) and run the same analyses as in Table 3&4. Panel A presents the results for the post-turnover performance (Tobin’s Q, BHAR36, ROA), and Panel B for the clean-up measures (Restructuring Costs and Writeoffs). The estimation set-up is the same as in Section 5: a self-selection model with the four instruments described earlier and with the same control variables as in Table 3 or Table 4. All the variables are defined in Appendix A. All the continuous variables are winsorized at the 1st and 99th percentiles. To save space, we suppress reporting the coefficients of the control variables. Coefficients’ p-values shown in [square] brackets are calculated using robust standard errors. Year and Firm fixed effects are applied wherever indicated. *** , **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. JournalofFinancialStability63(2022)10107216 G. Colak and E. Liljeblom 1999; Guay, 1999; Adams et al., 2009; Fahlenbrach, 2009). Thus, to account for such complicated cases, we remove them from our sample of turnovers and rerun the analyses in Tables 3 and 4. We manually collect the information on whether the preceding CEO is a founder (see Appendix A for further details on our data collection process). We closely follow the procedure described in other studies (Adams et al., 2009; Fahlenbrach, 2009; Jenter et al., 2016) and identify 55 incidences when the preceding CEO was the founder of the firm. Similarly, using the CEO ownership information in Execucomp and the total shares outstanding from CRSP, we determined the average CEO ownership in the three years leading to the turnover. We consider any CEO that owns more than 2% of the firm’s shares as a large ownership CEO.2728 There are 431 turnovers involving large ownership CEOs (some of them overlap with the founders). We drop from our turnovers sample the cases that involve either a founding CEO or a large ownership CEO. This exclusion leaves us with a sample of 1965 turnovers. Using this new sample of 11,790 firm-year observations (=1965 ×6), we re-estimate the self-selection models described in Tables 3 and 4 with identical setups. The new results are reported in Panels A and B of Table 8. The interaction term, After*LTP, preserves its sign and signifi- cance from Tables 3 and 4. Therefore, firm founders do not appear to be the main drivers of the results in Section 5. The Online Appendix presents some additional analyses related to Demerjian et al.’s (2012) measure of CEO ability, the firm’s industry characteristics, and the size of the external CEO talent pool (within in- dustries as in Gao et al., 2015 or within big cities as in Francis et al., 2016). 6.4. Forced (exogenous) turnovers Next, we conduct a robustness test that focuses only on the turnovers that are arguably exogenous. One of our instruments (CEODeathTurn- overs) is related to arguably the most exogenous turnover, sudden death, but the number of such turnovers is only 37. Thus, as an additional layer of robustness, we focus only on exogenous or forced turnovers. Our sample of forced turnovers is constructed using the approach in Eisfeldt and Kuhnen (2013) and Peters and Wagner (2014). We start with the sample of exogenous turnovers as provided by the authors of Eisfeldt and Kuhnen (2013), and we supplement it with the forced turnover sample of Peters and Wagner (2014).29 Thus, our sample covers both exogenous and forced turnovers. Of our sample of 2428 turnovers, 707 are forced. Using this smaller sample (707 turnovers and 4242 firm-year ob- servations), we conduct the same estimation as in Table 3. The results are presented in Table F1 of the Online Appendix. While the statistical significance decreases a bit, our main conclusions remain intact; the long-tenured cases are associated with relatively worse firm perfor- mance after the turnover. 6.5. Dynamic industries and the inherent business of the firm We also investigate whether our results are more pronounced for firms that are in more dynamic industries whereby the business condi- tions change more frequently, and exogenous shocks are more common a la Eisfeldt and Kuhnen (2013). In such industries, the damage from subpar performance during the unnecessarily long tenure could be more severe (Hambrick and Fukutomi, 1991; Henderson et al., 2006). We classify an industry as dynamic if during a given year the mean inno- vation activity (measured as mean 5-year average of RND spending up to that year) in that industry (within the two-digit SIC) is above the me- dian. The results presented in the Online Appendix (Table G1) show that indeed the interaction effect of After*LTP*DynamicIndustry is statisti- cally and significantly negative that indicates a much more pronounced underperformance due to LTP in the dynamic industries. 7. Conclusion Using data from the manually collected CEO turnovers during the time period from 1993–2013, we study the effect of the preceding CEO’s tenure on corporate performance post-turnover period. We report several interesting findings. First, we show that more highly valued firms keep their CEOs for a longer time (as indicated by a significant and positive relationship between tenure and Tobin’s Q), and the stock price development as well as ROA during such firm-years are significantly better (and the restructuring costs significantly smaller) as compared to other firms. Second, we find that the post-succession firm performance is significantly related to the tenure of the preceding CEO. It seems to be more challenging to take over after a long-tenured CEO: there are sig- nificant post-turnover drops in Tobin’s Q, ROA, and BHAR as well as significant increases in restructuring costs and asset write-offs. Also, the firm’s recovery to prior maximum valuation levels takes longer. Our additional tests for firms with different qualities of corporate governance during the preceding CEO’s tenure show that the post- turnover effects are mainly present only for unchecked CEO power in firms (weak governance). This finding gives some support to the claim that weak corporate governance and the resulting agency problems during the preceding CEO’s tenure may explain the substantial under- performance of the firms after the departure of a long-tenured CEO. We also find that these long- versus short-tenure effects are mostly present in firms that had a preceding CEO with a low skill score; that is, when firms keep an average CEO for too long, it is harder for the successor CEO to turn around the firm’s performance. Appendix A. Variable construction Panel A. Performance measure and proxies of “clean-up” intensity Variable: Definition. BHAR: The variable measures the buy-and-hold abnormal returns using the adjusted monthly stock returns (or abnormal returns AR) over a certain period (12 or 36 months). A stock’s abnormal return is calculated as ARi,t = Ri,t – Rm,t, where (continued on next page) 27 Using a 5% ownership cutoff (there are 236 such cases) to define a high-ownership CEO does not yield qualitatively different results. Results are available on request. 28 Our sample median (mean) of the average ownership of the predecessor CEO for all the 2428 turnover cases is around 0.3% (2.01%) of the total shares outstanding. As expected, long-tenured CEOs (LTP=1) tend to have higher ownership. The median (mean) ownership for the subsample of long-tenured CEOs is 0.0057 (0.0286); and for the subsample of short-tenured CEOs, it is 0.0017 (0.0132). 29 The exogenous turnover sample of Eisfeldt and Kuhnen (2013) is obtained from Andrea Eisfeldt’s website. The forced turnover sample of Peters and Wagner (2014) is extended by the authors until after 2016 and thus, it covers our entire sampling period of 1993–2016. The data are available through Florian Peters’ website. We thank the authors of both papers for making their sample available to other researchers. JournalofFinancialStability63(2022)10107217 G. Colak and E. Liljeblom (continued ) Ri,t is the stock i’s return for month t (with dividends) and Rm,t is the return on the CRSP equally-weighted market return during month t (with dividends). Then, buy-and-hold return (BHAR) is calculated as BHARj,T = ∏T (1 + Rj,t) (cid:0) t=1 ∏T (1 + t=1 Rm,t). Data source is CRSP monthly data files. When matching with the COMPUSTAT annual data, we use the values corresponding to the month when the fiscal year ends. CAR: The variable measures the sum of the market adjusted monthly stock returns (or abnormal returns AR) over a certain period (12 or 36 months). A stock’s abnormal return is calculated as ARi,t = Ri,t – Rm,t, where Ri,t is the stock i’s return for month t (with dividends) and Rm,t is the return on the CRSP equally-weighted market return during month t (with dividends). Data source is CRSP monthly data files. When matching with the COMPUSTAT annual data, we use the values corresponding to the month when the fiscal year ends. Restructuring Charges (REST): It is a one-time cost that must be paid by a company when it reorganizes. A restructuring charge might be incurred in the process of furloughing or laying off employees, closing manufacturing plants, shifting production to a new location or writing off assets. When a company restructures, it is usually experiencing significant problems and restructuring is an attempt to improve the business and recover financially. COMPUSTAT Annual Dataset’s after-tax restructuring costs (RCA) are reported with minus sign, and we take this sign into consideration when creating our REST variable. We essentially use (-RCA) scaled by the same year’s Sales (REST = (-RCA / Sales)). Return on Assets (ROA): It is calculated as net income over total assets (NI /AT). Data from COMPUSTAT annual files. Tobin’s Q: Calculated as Assets minus Common Equity plus Market Value of Equity scaled by Assets = (AT – CEQ + MVE)/ AT. MVE is calculated at the end of the fiscal year. Data is from COMPUSTAT annual files and CRSP monthly files. Writeoffs: For a negative special item (SPI), when the ratio of its absolute value to total assets at the beginning of the year exceeds one percent, then the variable takes the value of the ratio, otherwise the variable equals zero. Data obtained from COMPUSTAT annual files. Panel B. Firm and CEO Characteristics Variable: Definition. Acquisitions: The total acquisition spending of the firm during the year (AQC) scaled by total assets (AT). Data obtained from COMPUSTAT annual data. Age: The age of the CEO as reported in ExecuComp. Assets: The total assets (AT) of the firm. Data retrieved from COMPUSTAT annual files. Blockholders Percent: The percentage of shares held by large shareholders (blockholders). A blockholder is defined as a shareholder that owns at least 5% of the company’s shares. Data from 13F files. Founder: We determine a CEO’s founder status following a manual data collection process that also relies on some information from ISS Lab’s database and the data used in Adams et al. (2009). Renee Adams graciously shared with us her founders’ data for Fortune 500 companies for the period 1992–2011. From ISS Lab database we check whether the variable Rolecode indicates that the executive is the founder. We also manually verify the cases identified through ISS database by reading the CEO’s background and following a procedure similar toAdams et al. (2009), Fahlenbrach (2009), and Jenter et al. (2016). Out of 2428 turnover case in ours ample, we determine that 55 of these cases involve a founder CEO being replaced by a new CEO. Independent Board: A dummy that equals one if a firm’s ratio of independent directors to the total number of directors is above the sample median, and zero otherwise. Data is from RiskMetrics Governance database. Investments (CAPX): The total level of capital expenditures (CAPX) made by the firm for each fiscal year, scaled by total assets (AT). Data is from COMPUSTAT annual files. Cash Holdings (Cash): It is defined as the cash and cash equivalents (CHE) of the firm for each fiscal year, scaled by total assets (AT). Data is from COMPUSTAT annual files. Chair: an indicator variable taking a value of 1 if the CEO is also the chairman of the executive board, and 0 otherwise. The raw data is from ExecuComp, but manually checked for accuracy. CEO Ownership: Measures the fraction of the shares outstanding owned by CEO. Data is from 13F files. Dividend (DIV): Represents the common dividend (DVC) paid during the fiscal year, scaled by total assets (AT). Data is from COMPUSTAT annual files. Dividend Dummy: Takes a value of one when the firm pays a common dividend (DVC) during the fiscal apyear. Data is from COMPUSTAT annual files. Leverage: it is calculated as Short-Term Debt (DLC) + Long Term Debt (DLTT) divided by assets (AT). We use annual COMPUSTAT data. Market-to-Book (MTB): The market-to-book ratio is measured as the market value of equity (MVE) divided by the book value of equity (CEQ); data is from COMPUSTAT annual and CRSP monthly data. Market Value of Equity (MVE): It is equal to stock price as of the end of the fiscal year multiplied with the shares outstanding; both values are from CRSP monthly data. Net Income (NI): It is calculated as firm’s net income (NI) scaled by its total assets (AT). Data is from COMPUSTAT Annual Dataset. Net PPE: Calculated as Net Property, Plant, and Equipment (PPENT) divided by total assets (AT). COMPUSTAT annual data is used. Research and Development (RND): The total level of Research and Development (XRD) reported by the firm for each fiscal year, scaled by firm’s assets (AT). If the observation for the XRD is missing, we assume it is 0. Data is from COMPUSTAT annual files. Sales: Firm’s total sales. Data is from COMPUSTAT annual files. SEGN: The number of business segments a firm operates with in a given year. Data is from COMPUSTAT Segment data (annually). Tenure: The number of years since the CEO has been in his position. We use the variable BECAMECEO from ExecuComp data. JournalofFinancialStability63(2022)10107218 G. Colak and E. Liljeblom References Abadie, A., Imbens, G.W., 2006. Large sample properties of matching estimators for average treatment effects. Econometrica 74, 235–267. Ackerman, P.L., 1987. Individual differences in skill learning: an integration of psychometrics and information processing perspectives. Psychol. Bull. 102, 3–27. Ackerman, P.L., 1992. Predicting individual differences in complex skill acquisition: dynamics of ability determinants. J. Appl. Psychol. 77, 360–364. Adams, R., Almeida, H., Ferreira, D., 2009. Understanding the relationship between founder-CEOs and firm performance. J. Empir. Financ. 16, 136–150. Adut, D., Cready, W.H., Lopez, T.J., 2003. Restructuring charges and CEO cash compensation: a reexamination. Account. Rev. 78 (1), 169–192. Ali, A., Zhang, W., 2015. CEO tenure and earnings management. J. Account. Econ. 59, 60–79. Alvares, K.M., Hulin, C.L., 1972. Two explanations of temporal changes in ability-skill relationships: a literature review and theoretical analysis. Hum. Factors 14, 295–308. Avolio, B.J., Waldman, D.A., McDaniel, M.A., 1990. Age and work performance in nonmanagerial jobs: the effects of experience and occupational type. Acad. Manag. J. 33, 407–422. Bamber, L.S., Jiang, J., Wang, I.Y., 2010. What’s my style? The influence of top managers on voluntary corporate financial disclosure. Account. Rev. 85 (4), 1131–1162. Barron, J.M., Chulkov, D.V., Waddell, G.R., 2011. Top management team turnover, CEO succession type, and strategic change. J. Bus. Res. 64 (8), 904–910. Bebchuk, L., Cohen, A., Ferrell, A., 2009. What matters in corporate governance? Rev. Financ. Stud. 22, 783–827. Bennedsen, M., Perez-Gonzalez, F., Wolfenzon, D., 2020. Do CEOs matter? Evidence from hospitalization events. J. Financ. 75, 1877–1911. Bens, D.A., Johnston, R., 2009. Accounting discretion: use or abuse? An analysis of restructuring charges surrounding regulator action. Contemp. Account. Res. 26 (3), 673–699. Bertrand, M., Schoar, A., 2003. Managing with style: The effect of managers on firm policies. Q. J. Econ. 118 (4), 1169–1208. Betzer, A., Lee, H.S., Limbach, P., Salas, J., 2020. Are generalists beneficial to corporate shareholders? Evidence from exogenous executive turnovers. J. Financ. Quant. Anal. 55 (581–619), 2020. Boot, A., 1992. Why hang on to losers? Divestitures and takeovers. J. Financ. 47, 1401–1423. Borokhovich, K.A., Brunarski, K.R., Donahue, M.S., Harman, Y.S., 2006. The importance of board quality in the event of CEO death. Financ. Rev. 41, 307–337. Brochet, F., Limbach, P., Schmid, M., Scholz-Daneshgari, M., 2021. CEO tenure and firm value. Account. Rev. 96 (6), 47–71. Casamatta, C., Guembel, A., 2010. Managerial legacies, entrenchment and strategic inertia. J. Financ. 65 (6), 2403–2436. Chahine, S., Colak, G., Hasan, I., Mazboudi, M., 2020. Investor relations and IPO performance. Rev. Account. Stud. 25 (2), 474–512. Colak, G., Whited, T., 2007. Spin-offs, divestitures, and conglomerate investment. Rev. Financ. Stud. 20, 557–595. Combs, J.G., Ketchen Jr, D.J., Perryman, A.A., Donahue, M.S., 2007. The moderating effect of CEO power on the board composition–firm performance relationship. J. Manag. Stud. 44 (8), 1299–1323. Cox, D.R., 1972. Regression models and life-tables. J. R. Stat. Soc. Ser. B 34 (2), 187–220. Custodio, C., Ferreira, M.A., Matos, P., 2013. Generalists versus specialists: lifetime work experience and chief executive officer pay. J. Financ. Econ. 108, 471–492. Deadrick, D.L., Madigan, R.M., 1990. Dynamic criteria revisited: a longitudinal study of performance stability and predictive validity. Pers. Psychol. 43, 717–744. Dehejia, R.H., Wahba, S., 2002. Propensity score-matching methods for nonexperimental causal studies. Rev. Econ. Stat. 84, 151–161. Demerjian, P., Lev, B., McVay, S., 2012. Quantifying managerial ability: a new measure and validity tests. Manag. Sci. 58 (7), 1229–1248. Denis, D.J., Denis, D.K., 1995. Performance changes following top management dismissals. J. Financ. 50, 1029–1057. Dikolli, S.S., Mayew, W.J., Nanda, D., 2014. CEO tenure and the performance-turnover relation. Rev. Account. Stud. 19, 218–327. Dyreng, S.D., Hanlon, M., Maydew, E., 2010. The effects of executives on corporate tax avoidance. Account. Rev. 85 (4), 1163–1189. Edmans, A., Gabaix, X., Sadzik, T., Sannikov, Y., 2012. Dynamic CEO compensation. J. Financ. 67 (5), 1603–1647. Eisfeldt, A.L., Kuhnen, C.M., 2013. CEO turnover in a competitive assignment framework. J. Financ. Econ. 109, 351–372. Elliott, J., Hanna, D., 1996. Repeated accounting write-offs and the information content of earnings. J. Account. Res. 34 (Suppl), S135–S155. Elliott, J., Shaw, W., 1988. Write-Offs as accounting procedures to manage perceptions. J. Account. Res. 26 (Suppl), S91–S119. Ertimur, Y., Rawson, C., Rogers, J., Zechman, S., 2018. Bridging the gap: evidence from externally hired CEOs. J. Account. Res. 56, 521–579. Fahlenbrach, R., 2009. Founder-CEOs, investment decisions, and stock market performance. J. Financ. Quant. Anal. 44, 439–466. Fee, C.E., Hadlock, C.J., 2004. Management turnover across the corporate hierarchy. J. Account. Econ. 37, 3–38. Finkelstein, S., D’aveni, R., 1994. CEO duality as a double-edged sword: how boards of directors balance entrenchment avoidance and unity of command. Acad. Manag. J. 37 (5), 1079–1108. Francis, B.B., Hasan, I., Kose, J., Waisman, M., 2016. Urban agglomeration and CEO compensation. J. Financ. Quant. Anal. 51, 1925–1953. Francis, B.B., Hasan, I., Zhu, Y., 2020. Managerial effect or firm effect: evidence from the private debt market. Financ. Rev. 55 (1), 25–59. Francis, J., Hanna, J.D., Vincent, L., 1996. Causes and effects of discretionary asset write- offs. J. Account. Res. 34, 117–134. Gao, H., Luo, J., Tang, T., 2015. Effects of managerial labor market on executive compensation: evidence from job-hopping. J. Account. Econ. 59, 203–220. Garrett, D., Pavan, A., 2012. Managerial turnover in a changing world. J. Political Econ. 120, 879–925. Gertner, R., Powers, E., Scharfstein, D., 2002. Learning about internal capital markets from corporate spin-offs. J. Financ. 57, 2479–2506. Gibbons, R., Murphy, K.J., 1992. Optimal incentive contracts in the presence of career concerns: theory and evidence. J. Political Econ. 100, 468–505. Giniger, S., Dispenzieri, A., Eisenberg, J., 1983. Age, experience, and performance on speed and skill jobs in an applied setting. J. Appl. Psychol. 68, 469–475. Giroud, X., Mueller, H.M., 2015. Capital and labor reallocation within firms. J. Financ. 70 (4), 1767–1804. Gompers, P., Ishii, J., Metrick, A., 2003. Corporate governance and equity prices. Q. J. Econ. 118, 107–155. Graham, J.R., Li, S., Qiu, J., 2012. Managerial attributes and executive compensation. Rev. Financ. Stud. 25 (1), 144–186. Griffin, S.D., Carpenter, M.A., Boivie, S., 2011. What’s all that (strategic) noise? Anticipatory impression management in CEO succession. Strateg. Manag. J. 32, 748–770. Guay, W.R., 1999. The sensitivity of CEO wealth to equity risk: an analysis of the magnitude and determinants. J. Financ. Econ. 53, 43–71. Haggard, K.S., Howe, J.S., Lynch, A.A., 2015. Do baths muddy the waters or clear the air? J. Account. Econ. 59 (1), 105–117. Hainmueller, J., 2012. Entropy balancing for causal effects: a multivariate reweighting method to produce balanced samples in observational studies. Political Anal. 20 (1), 25–46. Hambrick, D., Fukutomi, G., 1991. The season’s of a CEO’s tenure. Acad. Manag. Rev. 16, 719–742. Hayes, R., Oyer, P., Schaefer, S., 2006. Co-worker complementarity and the stability of top-management teams. J. Law, Econ., Organ. 22, 184–212. Hazarika, S., Karpoff, J.M., Nahata, R., 2012. Internal corporate governance, CEO turnover, and earnings management. J. Financ. Econ. 104, 44–69. Henderson, A.D., Miller, D., Hambrick, D.C., 2006. How quickly do CEOs become obsolete? Industry dynamism, CEO tenure, and company performance. Strateg. Manag. J. 27, 447–460. Hermalin, B., Weisbach, M., 1998. Endogenously chosen boards of directors and their monitoring of the CEO. Am. Econ. Rev. 88, 96–118. Himmelberg, C., Hubbard, R., Palia, D., 1999. Understanding the determinants of managerial ownership and the link between ownership and performance. J. Financ. Econ. 53, 353–384. Holmstr¨om, B., 1982. Managerial incentive problems. In: Essays in Economics and Management in Honor of Lars Wahlbeck. Swedish School of Economics, Helsinki, pp. 209–230. Huang, S., Hilary, G., 2018. Zombie board: board tenure and firm performance. J. Account. Res. 56, 1285–1329. Huson, M.R., Parrino, R., Starks, L.T., 2001. Internal monitoring mechanisms and CEO turnover: a long-term perspective. J. Financ. 56, 2265–2297. Huson, M.R., Malatesta, P.H., Parrino, R., 2004. Managerial succession and firm performance. J. Financ. Econ. 74, 237–275. Jenter, D., Kanaan, F., 2015. CEO turnover and relative performance evaluation. J. Financ. 70, 2155–2183. Jenter, D., Matveyev, E., Roth, L., 2016. Good and Bad CEO, Working Paper. London School of Economics and University of Alberta. Johnson, B.W., Magee, R., Nagarajan, N., Newman, H., 1985. An analysis of the stock price reaction to sudden executive death: Implications for the management labor market. J. Account. Econ. 7, 151–174. Keil, C.T., Cortina, J.M., 2001. Degradation of validity over time: a test and extension of Ackerman’s model. Psychol. Bull. 127, 673–697. Khan, M.N., Srinivasan, S., S, Tan, L., 2017. Institutional ownership and corporate tax avoidance: new evidence. Account. Rev. 92 (2), 101–122. Lee, Y.G., 2014. An examination of restructuring charges surrounding the implementation of SFAS 146. Rev. Account. Stud. 19, 539–572. Lehn, K., Zhao, M., 2006. CEO turnover after acquisitions: are bad bidders fired? J. Financ. 61 (4), 1759–1811. McConnell, J., Servaes, H., 1990. Additional evidence on equity ownership and corporate value. J. Financ. Econ. 27, 595–612. McEvoy, G.M., Cascio, W.F., 1989. Cumulative evidence of the relationship between employee age and job performance. J. Appl. Psychol. 74, 11–17. McMullin, J.L., Schonberger, B., 2020. Entropy-balanced accruals. Rev. Account. Stud. 25 (1), 84–119. Miller, D., 1991. Stable in the saddle: CEO tenure and the match between the organization and the environment. Manag. Sci. 37 (1), 34–52. Moehrle, S., 2002. Do firms use restructuring charge reversals to meet earnings targets? Account. Rev. 77 (2), 397–413. Morck, R., Shleifer, A., Vishny, R.W., 1988. Management ownership and market valuation: an empirical analysis. J. Financ. Econ. 20, 293–316. Murphy, K.R., 1989. Is the relationship between cognitive ability and job performance stable over time? Hum. Perform. 2, 183–200. Musteen, M., Barker III, V.L., Baeten, V.L., 2005. CEO attributes associated with attitude towards change: the direct and moderating effects of CEO tenure. J. Bus. Res. 59, 602–612. JournalofFinancialStability63(2022)10107219 G. Colak and E. Liljeblom Nguyen, B.D., Nielsen, K.M., 2014. What death can tell: are executives paid for their Shleifer, A., Vishny, R., 1989. Management entrenchment: the case of manager-specific contributions to firm value? Manag. Sci. 60, 2994–3010. investments. J. Financ. Econ. 25, 123–139. Pan, Y., Wang, T., Weisbach, M., 2016. CEO investment cycles. Rev. Financ. Stud. 29, Staiger, D., Stock, J.H., 1997. Instrumental variables regression with weak instruments. 2955–2999. Parrino, R., Sias, R.W., Starks, L.T., 2003. Voting with their feet: institutional ownership changes around forced CEO turnover. J. Financ. Econ. 68, 3–46. Peters, F.S., Wagner, A.F., 2014. The executive turnover risk premium. J. Financ. 69, 1529–1563. Econometrica 65 (3), 557–586. Stock, J.H., Yogo, M., 2005. Testing for weak instruments in linear IV regression. In: Andrews, D.W.K., Stock Editions, J.H. (Eds.), Identification and Inference for Econometric Models: Essays in honor of Thomas Rothenberg. Cambridge University Press, New-York, pp. 80–108. Pourciau, S., 1993. Earnings management and non-routine executive changes. Sturman, M.C., 2003. Searching for the inverted U-shaped relationship between time and J. Account. Econ. 16, 317–336. Prendergast, C., Stole, L., 1996. Impetuous youngsters and jaded old-timers: acquiring a reputation for learning. J. Political Econ. 104, 1105–1134. Quigley, T.J., Hambrick, D.C., 2012. When the former CEO stays as board chair: effects on successor discretion, strategic change, and performance. Strateg. Manag. J. 33, 834–859. Quigley, T.J., Crossland, C., Campbell, R.J., 2017. Shareholder perceptions of the performance: meta-analyses of the experience/performance, tenure/performance, and age/performance relationships. J. Manag. 29, 609–640. Sturman, M.C., 2007. The past, present, and future of performance research, Research in Personnel and Human Resource Management 26, 49–110. Verhaegen, P., Salthouse, T.A., 1997. Meta-analyses of age-cognitive relations in adulthood: estimates of linear and nonlinear age effects and structural models. Psychol. Bull. 122, 231–249. changing impact of CEOs: market reactions to unexpected CEO death, 1950–2009. Strateg. Manag. J. 38, 939–949. Weisbach, M.S., 1988. Outside directors and CEO turnover. J. Financ. Econ. 20, 431–460. Weisbach, M.S., 1995. CEO turnover and the firm’s investment decisions. J. Financ. Rhodes, S.R., 1983. Age-related differences in work attitudes and behavior: a review and Econ. 37, 159–188. conceptual analysis. Psychol. Bull. 93, 328–367. Whitehead, C.K., 2011. Why not a CEO term limit? Boston Univ. Law Rev. 91, Salas, J., 2010. Entrenchment, governance, and the stock price reaction to sudden 1263–2011. executive deaths. J. Bank. Financ. 34, 656–666. Wooldridge, J.M., 2002. Econometric analysis of cross section and panel data. MIT Press, Shen, W., Canella, A.A., 2002. Revisiting the performance consequences of CEO Cambridge, MA. succession: the impact of successor type, post succession senior executive turnover, and departing CEO tenure. Acad. Manag. J. 45, 717–733. Shipman, J., Swanquist, Q., Whited, R., 2017. Propensity score matching in accounting Worrell, D.L., Davidson, W.N., Chandy, P.R., Garrison, S.L., 1986. Management turnover through deaths of key executives: effects on investor wealth. Acad. Manag. J. 29, 674–694. research. Account. Rev. 92 (1), 213–244. JournalofFinancialStability63(2022)10107220
10.1016_j.indmarman.2022.12.003
B2B Brand positioning in emerging markets: exploring positioning signals via B2B Brand positioning in emerging markets: exploring positioning signals via websites and managerial tensions in top-performing African B2B service websites and managerial tensions in top-performing African B2B service brands brands Emmanuel Mogaji, Maria Restuccia, Zoe Lee, Nguyen Phong Nguyen Publication date Publication date 01-01-2023 Licence Licence This work is made available under the CC BY 4.0 licence and should only be used in accordance with that licence. For more information on the specific terms, consult the repository record for this item. Document Version Document Version Published version Citation for this work (American Psychological Association 7th edition) Citation for this work (American Psychological Association 7th edition) Mogaji, E., Restuccia, M., Lee, Z., & Nguyen, N. P. (2023). B2B Brand positioning in emerging markets: exploring positioning signals via websites and managerial tensions in top-performing African B2B service brands (Version 1). University of Sussex. https://hdl.handle.net/10779/uos.23494082.v1 Published in Published in Industrial Marketing Management Link to external publisher version Link to external publisher version https://doi.org/10.1016/j.indmarman.2022.12.003 Copyright and reuse: Copyright and reuse: This work was downloaded from Sussex Research Open (SRO). This document is made available in line with publisher policy and may differ from the published version. Please cite the published version where possible. Copyright and all moral rights to the version of the paper presented here belong to the individual author(s) and/or other copyright owners unless otherwise stated. For more information on this work, SRO or to report an issue, you can contact the repository administrators at sro@sussex.ac.uk. Discover more of the University’s research at https://sussex.figshare.com/ Contents lists available at ScienceDirect Industrial Marketing Management journal homepage: www.elsevier.com/locate/indmarman B2B brand positioning in emerging markets: Exploring positioning signals via websites and managerial tensions in top-performing African B2B service brands Emmanuel Mogaji a,*, Mariachiara Restuccia b, Zoe Lee c, Nguyen Phong Nguyen d a University of Greenwich, London SE10 9LS, UK b University of Sussex Business School, Falmer, UK c Cardiff Business School, Cardiff University, Wales, UK d University of Economics Ho Chi Minh City, Ho Chi Minh City, Viet Nam A R T I C L E I N F O A B S T R A C T Keywords: Business-to-business branding Service brands Positioning Africa Emerging economies This research explored important questions concerning how top-performing African B2B service brands position their brands in this increasingly globalised, technology-driven and competitive digital ecosystem. In Study 1, we performed a content analysis of the homepages of the websites of 140 top-performing African B2B service brands to explore the most frequently used positioning signals. In Study 2, we conducted semi-structured interviews with 32 managers to understand the processes, challenges and, ultimately, the tension involved in managing their brands. Our analysis revealed tensions around curating a sense of professionalism, building a sense of trustworthiness and being proud of African roots. The research contributes to the literature on B2B brand positioning to show how the alignment between the signal on delivered positioning on brand websites and managers’ perception in developing brand positioning may shape different positioning strategies. In addition, the study offers practical implications for managers of B2B service brands in Africa on how to develop their brands, decide the possible signals to include in their websites and manage the branding tension within their business operations in the best way. The African market is recognised for its dynamism and creativity (Leke & Yeboah-Amankwah, 2018) and is fast growing, with a popular phrase such as ‘Africa rising’ coined for it (Amankwah-Amoah, Boso, & Debrah, 2018; Kolk & Rivera-Santos, 2018; Taylor, 2014). Yet, top- performing African business-to-business (B2B) service brands are continually facing intense competition from local and international firms within the digital services ecosystem (Blankson, Iyer, Owusu- Frimpong, Nwankwo, & Hinson, 2020; Omokaro-Romanus, Anchor, & Konara, 2019) in various forms, such as the trend to work for tech de- velopers outside Africa, the growing competition from other service providers within the African countries and the emerging gig economy and freelancers, who operate at lower costs due to low overheads (Mogaji et al., 2021). The spotlight has, therefore, fallen on B2B man- agers to face the uncertainty around competition, asses the debate be- tween localness and globalness (Davvetas, Diamantopoulos, & Halkias, 2016; Mohan, Brown, Sichtmann, & Schoefer, 2018) and investigate how they can sustain their brand’s competitive positioning to attract prospective customers, especially the customers who use their websites. In a fast-paced, resource-constrained, digitally driven and techno- logically connected environment, a website presents an integral part of communicating B2B brand positioning (Alkire & Hammedi, 2021). Yet, firm websites are often underutilised, especially in this consumer-centric age (Cartwright, Liu, & Raddats, 2021; Iankova, Davies, Archer-Brown, Marder, & Yau, 2019). Brands use websites as an information ‘dump’ and miss the opportunity to engage with wider stakeholders and directly communicate their real benefits. These B2B brands can craft narratives and manage content and information on their websites (S´anchez-Cha- parro, Soler-Vic´en, & G´omez-Frías, 2022). It is important to recognise that different website features and elements often reflect the strategic decisions made by B2B managers. Yet, the interplay between commu- nicated B2B brand positioning and managers’ decision processes is little understood. And any confusion over communicated B2B positioning on * Corresponding author. E-mail addresses: E.O.Mogaji@greenwich.ac.uk (E. Mogaji), m.restuccia@sussex.ac.uk (M. Restuccia), leesh4@cardiff.ac.uk (Z. Lee), nguyenphongnguyen@ueh. edu.vn (N.P. Nguyen). https://doi.org/10.1016/j.indmarman.2022.12.003 Received 27 July 2021; Received in revised form 6 August 2022; Accepted 4 December 2022 IndustrialMarketingManagement108(2023)237–250Availableonline12December20220019-8501/©2022TheAuthors.PublishedbyElsevierInc.ThisisanopenaccessarticleundertheCCBYlicense(http://creativecommons.org/licenses/by/4.0/). E. Mogaji et al. websites can cost B2B firms. directions. Previous studies on B2B brand positioning have mostly focused on strategies highlighting specific and effective competitive positions and attributes (Blankson & Kalafatis, 2019; Iyer, Davari, Zolfagharian, & Paswan, 2019). However, little research has been conducted on the strategic approaches managers take, especially from an African perspective, as previous studies have often focused on B2C and B2B service brands in Europe and North America, highlighting a gap in the knowledge about how and why African B2B service brands position themselves in the way they do. Moreover, despite the growing digital- isation, few studies have examined the value of websites as a digital brand identity in conveying trust, information and signal capabilities, instead often focusing on social media platforms. The few studies exploring B2B service brand positioning through websites, including the studies conducted by Williams, Lueg, Hancock, and Goffnett (2019) and Panda, Paswan, and Mishra (2019), focus on only developed countries. Important questions concerning how African B2B companies position their brands in this increasingly globalised, technology-driven and competitive context remain unanswered. Specifically, this research project aims to answer the following research questions: RQ1. How is African B2B brand positioning communicated in an on- line environment, especially on firm websites? RQ2. How do African B2B brand managers perceive their positioning strategies, and what are the tensions they face in developing their intended brand positioning? To answer these questions, this research adopted a multi-study approach. In Study 1, we performed a content analysis (CA) of the homepages of the websites of the 140 top-performing B2B service brands featured in the 2020 Clutch list to explore the signals most frequently used by African B2B service brands and as the main vehicle for the front- end communication of the corporate brand promise. In Study 2, we conducted semi-structured interviews with 32 managers from the brands mentioned in the 2020 Clutch list to learn more about and un- derstand the processes, challenges and tensions involved in managing their brands, especially in the context of using websites as digital gate- ways for prospective clients and the localisation and globalisation challenge. Our analysis revealed tensions around curating a sense of professionalism, building a sense of trustworthiness and being proud of African roots. The present study is theoretically positioned on and contributes to studies on B2B brand positioning literature by using signalling theory to show how alignment between the signal on delivered positioning on firm websites and managers’ perception of developing brand positioning may shape different positioning strategies. We also found space for considering both external positioning signals (e.g., via websites) and internal perception of positioning strategies (through managers’ perspective) as a way to deal with the sales-led culture of many B2B firms. Superficially, we showed the paradoxical tension around African roots positioning that can both elevate and hinder B2B firms’ desires and intentions to reach multiple customers. In addition, the study offers practical implications for B2B service brand managers in Africa on how to develop their brands, decide on the possible signals to include on their website and manage the branding tensions within their business oper- ations in the best way. This paragraph discusses the structure of this paper. We begin by reviewing three relevant literature strands: B2B brand positioning in the service context, the role of websites and the emerging economies. Based on the literature review, we identified a significant gap in the posi- tioning process for B2B services, which is particularly relevant in the African context due to Africa’s unique competitive landscape. The methodology section details Study 1 (CAs of the homepages) and Study 2 (semi-structured interviews with managers). Following the results sec- tion, we discuss the theoretical and managerial implications of the study before concluding with the study’s limitations and further research 1. Literature review 1.1. Brand positioning in the B2B service context Brand positioning is defined as the ‘act of designing the company’s offering and image to occupy a distinctive place in the mind of the target market’ (Kotler & Keller, 2003, p. 867). The main dilemma B2B mar- keters face is that many B2B firms are sales-oriented and focus on operation improvement, thus ignoring the consideration of brand posi- tioning as meaningful (Kalafatis, Blankson, Luxly Boatswain, & Tsogas, 2020; Iyer et al., 2019). B2B firms market and brand a wide range of offerings, such as products, services, hybrid offerings and solutions (Hutt & Speh, 2017), as well as ‘ingredient’ products and services (Helm & ¨ Ozergin, 2015). This can be partly understood from customer percep- tions that most industrial markets are viewed as commodity markets and that the customers have a clear idea about the product category and competition (Kotler & Pfoertsch, 2006). Yet, some studies have identi- fied a shift in the understanding, such that brand positioning is being regarded as an increasingly important strategic decision (Leek & Christodoulides, 2011a; Blankson & Kalafatis, 2019). This shift has largely been caused by increased competition, similarity in offering and a higher level of imitation from traditional as well as non-traditional competitors. Therefore, a distinctive B2B brand positioning can enable a firm to convey its differentiation (Beverland, Napoli, & Yakimova, 2007; Jalkala & Ker¨anen, 2014) and project a credible image of itself to relevant stakeholders. Leek and Christodoulides (2011) have identified several benefits provided by B2B brands to buyers and suppliers, such as ‘higher confi- dence, uncertainty reduction, increased satisfaction, greater comfort and identification’ (p. 831) for buyers and ‘quality, differentiation, higher demand, premium price, brand extensions, distribution power, barrier to entry, goodwill, loyal customers, customer satisfaction and referrals’ (p. 831) for suppliers. Yet, the unique characteristics of ser- vices such as intangibility, heterogeneity, inseparability, perishability and unobservability can present further challenges to B2B service pro- viders’ positioning activities (De Chernatony & Riley, 1999; Guenther & Guenther, 2019; Williams et al., 2019). Buyers typically perceive a higher level of complexity when purchasing services rather than prod- ucts (Smeltzer & Ogden, 2002). According to the taxonomy given by Fitzsimmons, Noh, and Thies (1998), not all service purchases are equally complex, as complexity depends on the interplay between a service’s focus (property, people or process) and its value to the buyers (low or high). As Åhlstr¨om and Nordin (2006) underlined, purchasing services often entails difficulties in ‘defining, pricing, evaluating and controlling the service delivery’ (p. 78), which significantly influences the perception of quality of such B2B offerings. Such complexity in services means that B2B marketers need a creative brand positioning and messaging to stand out. Numerous calls have been made to advance the understanding of brand positioning in a B2B context (Blankson & Kalafatis, 2019; Leek & Christodoulides, 2011), with significant gaps in the research, especially in the context of B2B services (Ker¨anen, Piirainen, & Salminen, 2012; Williams et al., 2019). Iyer et al. (2019) took the first step to under- standing the different types of brand positioning strategies. Some ex- amples of brand positioning services are - using intangible elements such as trust and expertise as sources of differentiation (Beverland et al., 2007), focusing on service quality and relationship (Roberts & Merri- lees, 2007), developing brand personality elements (Herbst & Merz, 2011), enhancing service quality capabilities (Jalkala & Ker¨anen, 2014), giving importance to human values (He, Huang, & Wu, 2018) and ensuring sustainability (Casidy & Yan, 2022). By focusing on internal processes and behaviours within an organisation, compared to the image-led approach of positioning, Iyer et al. (2019) further highlighted that B2B researchers need to go deeper into the organisational process IndustrialMarketingManagement108(2023)237–250238 E. Mogaji et al. viewpoint of positioning strategies. Interestingly, few studies have explored the congruence between both the front end and the backstage of the brand positioning process. With an exception, Blankson, Kalafatis, Coffie, and Tsogas (2014a, 2014b) examined the congruence between managers’ intended positioning strategies (Level 1), the firm’s actual or delivered positioning strategies (Level 2) and consumers’ perception of the positioning strategies (Level 3). They recognise feedback loops among the different positioning levels and underline the importance of maintaining congruence between them. In addition, the shift to digital marketplaces has allowed B2B firms to use websites to transmit effective signals and shape customers’ percep- tions of their intended brand positioning (Williams et al., 2019). The literature offers many anecdotal suggestions for B2B marketers in uti- lising websites, including acting as a face for the firm (Coulter, 2012), providing a direct link to customers (Williams et al., 2019), promoting credibility (Virtsonis & Harridge-March, 2009), reaching potential cus- tomers cost-effectively (S´anchez-Chaparro et al., 2022), influencing customers’ perceptions of service quality (Cretu & Brodie, 2007) and promoting user interactivity (Ellinger, Lynch, Andzulis, & Smith, 2003). Yet, the complexity of the value customers seek means that different segments need different positioning signals and that how B2B managers choose brand positioning strategies for communicating on their websites remains a challenge, especially in the multi-faceted context in Africa (Connelly, Certo, Ireland, & Reutzel, 2011). 1.2. Positioning B2B service brands in Africa Investigation of B2B marketing issues in an international context has a long history (e.g., Korkmaz & Messner, 2008; Spyropoulou, Skarmeas, & Katsikeas, 2010). However, B2B marketing has only recently broken ‘the frontiers of traditional [developed] markets, reaching emerging economies’ (Cortez & Johnston, 2017, p. 92). With their rapid economic development and government policies directed towards economic lib- eralisation (Arnold & Quelch, 1998; Sheth & Sinha, 2015), emerging economies in South America, the Middle East and Africa confront mar- keters with new challenges. Mirroring this market evolution, the research on branding and positioning initiatives in Africa has witnessed major growth. Brand positioning strategies developed in Western contexts apply to the Afri- can context too, with Indigenous African companies likely to replicate foreign firms’ positioning tactics (Blankson et al., 2020). As per Koch and Gyrd-Jones (2019a, 2019b), most studies on branding and posi- tioning in Africa appear to have focused on positioning strategies (Blankson et al., 2020; Blankson, Ketron, & Darmoe, 2017). On the other hand, the positioning process has received less attention, with notable exceptions found in some studies about managers’ views on positioning practices (Coffie, 2016; Coffie & Owusu-Frimpong, 2014). Significant emphasis has been devoted to the retail sector (Adokou & Kyere- Diabour, 2017; Blankson, Ketron, & Coffie, 2017), while positioning studies looking at the broad service context do not appear to have clearly distinguished between B2C and B2B companies (Coffie, 2016; Coffie & Owusu-Frimpong, 2014). In B2B domains, such as architecture, do- mestic African brands appear to have been affected by a negative perception due to their origin (Verster, Petzer, & Cunningham, 2019), while brand orientation has been found to enhance brand loyalty in a B2B industrial holding context (Dludla & Dlamini, 2018). Most studies have a single-country focus, especially on Ghana (Blankson et al., 2020; Coffie, Blankson, & Dadzie, 2018; Coffie & Owusu-Frimpong, 2014) and South Africa (Dludla & Dlamini, 2018; Verster et al., 2019). To the best of our knowledge, no multi-country study has been conducted on B2B service positioning, which could reveal potential similarities and dif- ferences among B2B service brands in different African countries. 1.3. Signalling theory Signalling theory is becoming increasingly popular among marketing scholars (e.g., Connelly et al., 2011; Helm & ¨ Ozergin, 2015; Kirmani & Rao, 2000), as the theory directs attention to the core problems man- agers face, namely, how to use signals to reduce the uncertainty asso- ciated with making a selection among a choice set in situations that have incomplete and asymmetrically distributed information (Spence, 1973). For example, the decision around B2B positioning strategies via websites (Williams et al., 2019) may be wrapped in incomplete information and can create uncertainty for both providers and buyers. Service buyers value reputation and experience when purchasing B2B services (Fitz- simmons et al., 1998). Hence, B2B service providers are expected to convey such signals, alongside cues to service quality, in their commu- nicated actual positioning (Mccoll, Truong, & Rocca, 2019; Williams et al., 2019). Firms avail of different tools to signal their actual posi- tioning in physical and digital spaces. Advertising has long been considered a prime tool in both developed and emerging economies (Alden, Steenkamp, & Batra, 1999; Blankson et al., 2014a, 2014b, 2020; Ries & Trout, 1986). In an increasingly digital global marketplace, positioning efforts, for example, are now moving online through brands’ websites and social media presence (Magno & Cassia, 2020; Williams et al., 2019). Developed by firms and/or according to their guidelines, these websites provide a reliable indicator of the key messages trans- mitted to prospective and current buyers. Successful brand positioning involves ‘a competitive frame of refer- ence in terms of the target market and the nature of competition’ (Keller, 2005, p. 705). Hence, our investigation explores managers’ perception of competition, which can be defined in multiple ways (e.g., industry-, firm-, manager- or customer-focused; Gur & Greckhamer, 2019) and at different geographical levels (e.g., domestic, regional or international level; Hutt & Speh, 2017; Omokaro-Romanus et al., 2019). Given brand positioning’s increasing level in emerging economies (Alkire & Ham- medi, 2021; Arnold & Quelch, 1998; Coffie & Owusu-Frimpong, 2014), it is crucial to understand how managers of African B2B service brands incorporate competitive considerations into their brand positioning process. Thus, in this study, we use signalling theory to explore how and why African B2B firms communicate their brand positioning signals via websites in the way they do. 2. Methodology Considering the exploratory nature of our research, we deemed an inductive multi-study approach appropriate to understand the intended and actual positioning of B2B service brands in Africa through their websites as digital dimensions for reinforcing their brand identities. The first stage of the research design involved a data-driven examination of the homepages of the websites of the top-performing B2B service brands in the Clutch, 2020 list. The second stage utilised semi-structured in- terviews with 32 managers from the top-performing B2B service brands mentioned in the Clutch, 2020 ranking to understand the process behind their intended positioning through their digital dimensions. Fig. 1 pro- vides a graphical overview of the research design and the following sections provide details on each study. 2.1. Study 1 – website content and thematic analysis The first study aimed to identify the visible features used on the homepages of the websites of B2B service brands as key messages illustrating top-performing B2B service brands in a digital environment. Websites have been used as B2B brand positioning vehicles by Williams et al. (2019) in their study of B2B carriers’ service and by Panda et al. (2019) in the context of franchisors. The homepages were chosen as the unit of analysis because they are the gateways to information on brands. As customers land on the homepage of a brand’s website, they are likely to associate the homepage’s features with the corporate brand. For the analysis, 140 top-performing African B2B service brands published by Clutch (2020) were used as a sample. Clutch is a global B2B research organisation that has been covering Africa since 2019. Clutch IndustrialMarketingManagement108(2023)237–250239 E. Mogaji et al. Fig. 1. Research Design. Adapted from Braun & Clarke (2006), Badejo, Charles, & Millisits (2022) and Kaur, Mogaji, Wadera, & Gupta (2022). IndustrialMarketingManagement108(2023)237–250240 E. Mogaji et al. analyses each brand in terms of the services offered, awards received, online presence and reviews on Clutch. Clutch’s selection methodology ensures that the winners represent the ‘best of the best’ (Clutch, 2020). Although this sample has limited temporal coverage (i.e., only for the year 2020), it ensures the study’s consistency, validity and replicability. The latest cohort of the leading B2B brands in Africa was presented by Clutch in August 2020 and included 140 brands from four African countries: Kenya (n = 28, 20%), Nigeria (n = 54, 38.6%), South Africa (n = 49, 35%) and Tanzania (n = 9, 6.4%). The CA provided insights into the frequency of the key features on the websites, while the thematic analysis grouped the key features into significant relevant themes. Following standard CA procedures, an emerging coding scheme was developed (Landrum & Ohsowski, 2018; Panda et al., 2019), starting with the definition of our objective: to gain an understanding of the brands’ actual positioning via their websites and of the sample to be analysed, the top-performing brands in the 2020 Clutch ranking of Af- rican B2B service brands. A coding sheet was developed to facilitate data collection and analysis, starting with a sample of two randomly selected homepages of the B2B service brands mentioned in the Clutch, 2020 ranking for each country, adding to eight homepages. Iteratively, fea- tures were extracted from the website content (e.g., contact and blog post), and the process stopped when no new features emerged. The final coding sheet identified 18 types of objective features that could be present on the homepage of the website of a B2B service brand in Africa (Appendix 1). As per Camprubí and Coromina (2016), ‘CA will perform better when studies use more than one judge or when judges are inde- pendent from authors’ (p. 139). This approach, which was recently used in B2B positioning research (Panda et al., 2019), was retained in our study, with the recruitment and training of two independent coders (Czarnecka & Mogaji, 2020). Along with the coding sheet, the coders were provided with a coding book that included a description and actual examples of different website features. A pre-test was conducted to establish intercoder reliability, with a coefficient ranging between 0.98 and 1, indicating an acceptable level of reliability (Landis & Koch, 1977). Following the pre-test, all 140 websites were coded by the two coders. Based on the coding, we determined the occurrence of each feature on the brands’ homepages. Through an inductive thematic analysis, the research team then generated overarching themes by grouping the features coded by the independent coders and found to be conveying a similar signal. 2.2. Study 2: managers’ perception of intended positioning strategies and processes While Study 1 is more about the delivered positioning strategy on the websites, Study 2 focuses on the processes underlying intended brand positioning activities by engaging with high-ranking management team members of the top-performing B2B service brands on the Clutch (2020) list. We emailed all 140 brands on this list to introduce our project and invited their representatives for interviews. Following clarifications, 32 of the 58 companies that expressed interest confirmed their participa- tion. Consistent with previous studies (Iyer et al., 2019; Mogaji, Czar- necka, & Danbury, 2018), our data collection used a key informant approach. The companies’ designated representatives (hereafter referred to as participants) to interact with the research team and the interview appointments were then arranged. As shown in Table 1, the companies in our sample provided mostly professional services, and the participants included chief executive officers (CEOs), chief operating officers (COOs) and marketing managers. Nigeria (n = 15, 47%) was the most represented country, followed by South Africa (n = 9, 28%), Kenya (n = 5, 16%) and Tanzania (n = 3, 9%). To structure the data collection, we developed an interview guide and protocol (see Appendix 2). Drawing from Koch and Gyrd-Jones (2019a, 2019b), we created an initial set of questions to discern the actors, activities, context and drivers of corporate brand positioning for African B2B service brands. A pilot study was run to (a) better Table 1 Profile of participants for the semistructured interviews. Label Country Gender Position Service provided P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 P11 P12 P13 P14 P15 P16 P17 P18 P19 P20 P21 P22 P23 P24 P25 P26 P27 P28 P29 P30 P31 P32 Kenya Kenya Kenya Kenya Kenya Nigeria Nigeria Nigeria Nigeria Nigeria Nigeria Nigeria Nigeria Nigeria Nigeria Nigeria Nigeria Nigeria Nigeria Nigeria South Africa South Africa South Africa South Africa South Africa South Africa South Africa South Africa South Africa Tanzania Tanzania Tanzania Digital Marketing Chief Executive Officer Chief Marketing Officer Branding Agency Chief Marketing Officer Marketing Agency Chief Operating Officer App Developer Chief Operating Officer Software Developer Digital Marketing Chief Operating Officer Chief Operating Officer Website Developer Female Female Male Male Male Female Female Female Managing Director Female Female Male Male Male Male Co-Founder Head of Marketing Managing Director Chief Operating Officer Chief Executive Officer Chief Executive Officer App Developer Digital Marketing Branding Agency Marketing Agency Marketing Agency Video Production Social Media Marketing Digital Marketing Branding Agency Marketing Agency Website Developer Web Hosting IT Security Marketing Agency Marketing Agency Marketing Agency Marketing Agency Chief Marketing Officer Marketing Director Chief Executive Officer Co-Founder Chief Executive Officer Chief Consultant Male Male Male Male Male Male Female Marketing and Sales Director Chief Technology Officer Human Resources Manager Chief Executive Officer Female Female Female Male Chief Operating Officer Marketing Agency Male Marketing Manager Digital Agency Male Creative Director Advertising Agency Male Creative Director Branding Agency Male Co-Founder Digital Marketing Male Male Male Chief Executive Officer Co-Founder Co-Founder App Developer Website Developer Digital Marketing comprehend and capture the context of the participants; (b) refine the wording and the order of the questions and (c) check the duration of a typical interview. Five managers participated in the pilot, and they were excluded from the final sample of the participants (n = 32). After the pilot study, the updated interview guide included 12 open-ended ques- tions, beginning with background information about the company and the participant and progressing to actual corporate brand development and positioning activities on the websites. The first author conducted all interviews between April and May 2021. The interview sessions were designed to be highly convenient for the participants. Due to the multi-country nature of this research, the in- terviews were conducted via Zoom or Skype software. In line with the ethical approval obtained at the first author’s institution, the partici- pants were assured of anonymity and protection of personal and business-sensitive information. The interviews lasted between 48 and 64 min (median = 54). The audio recordings from the interviews were transcribed by a professional transcriber (132 single-spaced pages in total) and imported into Nvivo, a qualitative analysis software, for data analysis. Due to the exploratory nature of this study, our inductive data analysis followed Braun and Clarke’s (2006) six phases of thematic analysis (Fig. 1). The first phase involved immersion in the data through repeated reading of the transcripts. Inductive coding followed it, with the growing number of codes being collated and assigned to a relevant overarching theme (Farinloye, Mogaji, Aririguzoh, & Kieu, 2019; Mogaji & Nguyen, 2022). 18 sub-themes emerged and finally merged into three IndustrialMarketingManagement108(2023)237–250241 E. Mogaji et al. main themes (aggregate dimension; fifth phase of a thematic analysis), which will be discussed in the manuscript with relevant participant quotes. See Fig. 1 for detailed data analysis. To ensure the study’s credibility and authenticity, several measures were put in place. Before the coding, a member check (Merriam & Tis- dell, 2015) was performed: the transcribed interviews were sent to all the participants for verification and returned with their approval. Following Miles, Huberman, and Saldana (2013), we employed a rigorous peer debriefing process, code-checking and ongoing compari- son of the researchers’ findings. In addition, interview quotes were used to illustrate each point, in line with Lincoln and Guba (1985) recom- mendation to provide a ‘thick description of the sending context, so that someone in a potential receiving context may assess the similarity be- tween them and the study’ (p. 125). Finally, an audit trail (Shenton, 2004) documented this study’s methods, procedures and decision points. As evidenced by the audit trail, the data were not used arbi- trarily, and the researchers’ positions did not overpower the partici- pants’ voices. 3. Findings 3.1. Study 1 – signals used on firm websites and positioning strategies The website analysis goal was to gain insights into the actual Table 2 Summary of content analysis findings, with totals and subtotals by country (Study 1). Theme Website feature(s) Description Country subtotals Nigeria (n = 54) South Africa (n = 49) Tanzania (n = 9) Kenya (n = 28) 3 (11%) 4 (7%) 10 (36%) 15 (28%) 10 (20%) 15 (31%) s/ n 1 2 3 4 5 6 7 8 9 10 ‘We can do it’ Staff Team Why Choose Us Services Facts and Figure Proficiency Levels Blog Post We did it before Client list Client Testimonial (Images) Client Testimonial (Video) Case Studies 11 What they say about us Accreditation and Partnership 12 13 Press Coverage Reviews 14 We are here for you Contact 15 16 Physical addresses Social Media 17 Chatbot and virtual assistant 18 We are proud of our roots This category includes information about staff members within the company, with their pictures and job title. It provides information about the team’s composition, proficiency and diversity. This category covers key features about the company’s business operations and provides reasons why they should be hired. The rationale for the choice could be grounded in their working process, the services they offer, their team, level of expertise and awards received. This category provide summary about the services of the company. They are often represented with relevant icons, and prospective customers are expected to click for more information about those services. Facts and figures are presented on the website to provide insights into the firm’s operations. These can be objective (e.g., number of completed projects, number of clients, and staff) or more humorous (e.g., cups of coffee made in a year). This category includes information presented in slide bars or percentages highlighting the firm’s proficiency with specific technology, tools or business operations. Blog posts are provided on the home page or a separate page. Companies use this form of content creation to educate their customers and share knowledge and expertise. Information about past customers is presented in the list format, with the number and type of clients. Often the logo of these clients is displayed. This includes verbatim quotes from a past satisfied customer. The person’s image, name, job title, and company might be provided. This includes a video-recorded testimonial with feedback about the service provider. The video might contain their names and relevant information. Detailed information about a specific project brought to completion. Information in the case study might include the client, the brief they worked on and the final product. Information about recognitions, accreditations and partnerships are often displayed at the bottom/footer of the website. For digital marketing, they might indicate Google, Microsoft and Facebook partnership. Awards and certifications from third parties are also included. Brands might present their press coverage on their website. Information includes the logo of the press and news organisations, both foreign and domestic. Customer reviews from third-party platforms (e.g., Google, Trustpilot or Clutch) might be integrated into the brand’s website through widgets. The reviews provided on these external partners are also shown, often at the footer/bottom of the page. Brands might use a different way to engage with prospective clients and invite them to keep in touch, including invitations to: subscribe to the newsletter, submit a quote or a free consultation. Street addresses of the physical officers are provided on the homepage, often at the footer of the webpage. In addition, it sometimes includes telephone numbers and email addresses. Social media icons are presented at the footer of the website. These icons link to the profile of the company on different social media, including: Facebook, Twitter, LinkedIn, Instagram, YouTube, Whatsapp, GitHub, Pinterest, Vimeo, Dribble, Medium, Behance, Skype and Tumblr, and Google +. Aligned with growing trends of digital transformation, this feature might include a chatbot or virtual assistant. This category includes features connecting the brand to its African roots. These features might include country-specific headlines on the websites and the use of African languages. 11 (48%) 23 (43%) 24 (49%) 4 (14%) 8 (15%) 4 (8%) 3 (11%) 4 (7%) 2 (4%) 10 (36%) 7 (13%) 23 (82%) 9 (32%) 4 (14%) 19 (68%) 37 (68%) 28 (52%) 0 37 (68%) 7 (25%) 14 (26%) 1 (3%) 1 (3%) 0 2 (4%) 9 (17%) 1 (2%) 10 (36%) 23 (43%) 20 (71%) 41 (76%) 16 (33%) 36 (73%) 23 (47%) 1 (2%) 36 (73%) 20 (41%) 2 (4%) 3 (8%) 7 (14%) 21 (43%) 32 (65%) 12 (43%) 6 (21%) 19 (35%) 9 (17%) 13 (26%) 3 (6%) Total (n = 140) 18 (13%) 40 (29%) 61 (44%) 19 (13%) 11 (8%) 34 (24%) 103 (73%) 64 (46%) 6 (4%) 99 (70%) 43 (31%) 5 (4%) 13 (9%) 9 (6%) 58 (41%) 98 (70%) 47 (34%) 18 (13%) 1 (11%) 0 3 (33%) 3 (33%) 2 (22%) 1 (11%) 7 (78%) 4 (44%) 1 (11%) 7 (78%) 2 (22%) 0 0 1 (11%) 4 (44%) 5 (56%) 3 (33%) 0 IndustrialMarketingManagement108(2023)237–250242 E. Mogaji et al. positioning of B2B service brands, and this section presents five features African B2B brands use to communicate their brand position via web- sites as follows: • we can do it – competence; • we did it before – past customer experience; • what they say about us – reputation-related features; • we are here for you – availability to customers and • we are proud of our roots – connection to African roots. Table 2 reveals the similarities and differences among the B2B ser- vice brands operating in the four countries included in the scope of this study in their use of various website features in conveying actual brand positioning. (See Table 3.) 3.1.1. Theme 1: we can do it This theme is about demonstrating competence. B2B service brands reassure prospective clients that they have the expertise to take on and deliver projects, both in terms of resources and capability. These brands flaunt their staff base, skills level and technology proficiency and buttress them with some facts and figures. For example, Blue Matrix from Nigeria boldly presented on their website that they employ cutting- edge JavaScript Technology and 63 WebStudio used facts and figures to indicate that they have ‘300+ projects completed and 20,500+ hours spent on projects’, while South Africa’s Mango5 indicated that they have ‘14+ years in business, 25m+ interactions, 100+ satisfied clients and 250+ consultants’. Several B2B service brands further highlighted key aspects of their business operations and provided reasons why they should be hired, such as their working process, team, level of expertise and awards received, as further evidence of their credibility and capa- bilities. To illustrate, Blue Light, a Nigerian software development agency, stated that customers should choose them because they ‘provide a dedicated project manager to oversee [client’s] project from beginning to deployment’. 3.1.2. Theme 2: we did it before This theme signals customer experiences, allowing brands to flaunt their past achievements as an indication of their ability to deliver the services they advertise. The brands showcase lists of past and present clients, testimonials and feedback from clients. For example, the Tan- zanian company EvMak displayed their clients’ logos under the caption ‘we have built solutions for...’. In South Africa, Mango5 and Galactic Digital used the headings ‘industries served’ and ‘friends we have made, doing what we love’, respectively, to showcase their client lists. Homepages prominently feature ‘case studies’ (n = 99) and ‘client tes- timonials’ (n = 64), including the clients’ company names, along with their key personnel’s images, names and job titles. Some brands went further by providing video recordings of their customers’ testimonials (n = 6). 3.1.3. Theme 3: what they say about us Many African B2B service brands exhibit their recognition, press coverage and partnerships to highlight what external stakeholders say about them. They use these elements to get a form of external validation for their brand, possibly to convince prospective clients that other people find them credible. Reviews are provided on third-party plat- forms, such as Google, Trustpilot and Clutch, and integrated into the brands’ websites via widgets. B2B service brands also use press releases and press coverages, including ‘as seen on’ features and lists of different foreign and domestic media companies. These companies also provide information about their accreditations and partnerships. Collaborations with global tech companies, such as Google, Microsoft and Facebook, and accreditations with various professional bodies are displayed, possibly to demonstrate that the brand is well connected to significant global players in the industry and can capitalise on such connections. Table 3 Summary of main themes and sample quotes of participants. Main theme Sub themes Curating a sense of professionalism • Being authentic or driving the reach for the brand. • Website development • Website management and • Content creation • Evaluating internal capabilities • Integrating physical elements into the digital dimensions Building a sense of trustworthiness • Addressing clients’ expectations • Emphasising credentials and capabilities • Building trust from a global perspective. • Fictitious information on website (testimonials, portfolios, and case studies) Being proud of • Globalisation versus African roots. localisation approach. • Globalisation and competing with global clients • Reinforcing huge credibility within the local markets. • Continent of origin shaping buyers’ perceptions. • Recognising their African roots. • Negative perception of African B2B service brands, • Testimonials of a very popular African businesspeople Sample quotes from participants ‘We know we needed a new website but we do not have the in-house capabilities to develop this, so we had to look for a web designer to help us’. ‘People think because you develop an app, you should be able to develop your website. Yes, I tried this when we started, but now we have outsourced the design’. ‘I designed the first website of the company and we have changed it over many years, but it is still done in house.’ We had to get everyone dressed in the branded T-shirt of the agency and take a group photograph for the website. We wanted to communicate our strength as a team for people visiting our website’. ‘We are in competition with developers in Europe and those freelancers on fiverr, everyone now thinks they can design a website, so it’s important for us to update our website with our certifications, partnership, and client testimonials’. ‘We believe our reputation will grow and people can trust us more when we get featured in Forbes. You know Nigerians have a different reputation, but we believe having the Forbes feature on our website will help our brand’ ‘I haven’t had a client tell me seeing us on Forbes makes them trust us, there is pressure on everyone to appear trustworthy but it’s not always a guarantee’. ‘We have junior designers and developers working on projects for a dummy company and we upload these designs in our portfolio on the website, but we often declare that it’s a sample work and not for a real client’. People don’t believe something good can come out of Africa. Coupled with the fact that you are competing with outside brands, the persisting notion about Africa and Africans often poses a challenge and a hurdle’. ‘Even our brothers and sisters here don’t appreciate our work. They do not see value in what we have to offer, and they want to go to Europe or Asia to get their app developed’. ‘Having him on our website showcase that we are provide outstanding service even though we are based in Nigeria, we want to be proudly (continued on next page) IndustrialMarketingManagement108(2023)237–250243 E. Mogaji et al. Table 3 (continued ) Main theme Sub themes • Influx of large international service providers Sample quotes from participants associated with the big Nigeria clients we have worked with’. ‘In Kenya, I can say we are the best app developers, we have done many projects and displayed this on our portfolio on the website, but I don’t think we are ready to go global now, allow us conquer Africa’. ‘We are seeing an influx of large agencies coming in the form of network agencies and this has presented a huge concern for us on how to reinforce our relationship with our existing local brands, we need to protect our market share’. 3.1.4. Theme 4: we are here for you Considering the impending competition within the industry and the lack of humanness in B2B service, many of these top-performing brands provide elements on their website to further reassure prospective clients of their availability to offer the service. They provide both virtual and physical contact means and offer various channels of communication to their customers. In the study, 47 B2B service brands provided chatbots and virtual assistants on their homepages, conveyed availability and aligned with recent and rapidly developing digital transformation trends. Prospective clients were also invited by the brands to explore their presence on various social media platforms and use the contact form on their websites. These brands also provided physical office ad- dresses to indicate their availability to meet their customers on their premises, possibly to build trust and connections, especially as they compete with online freelancers and those in the gig economy, who might not have a physical contact office. Some even invite prospective clients to visit their offices to meet the team, explore their resources and learn more about their offerings. Some companies, such as Galactic Digital—a digital agency based in Cape Town—display professional pictures of their open-plan offices on their websites. 3.1.5. Theme 5: we are proud of our roots The top-performing B2B brands in Africa are recognising their Afri- can roots and using them as signals on their websites. While some signals may be expected and usual for digital service providers, the African root signals offer a unique approach to the brand positioning for these service brands. Although only 18 B2B brands used these signals in our study, it illustrated the brands’ desires to be associated and connected with Af- rica, anticipating that this might convince other pan-African clients to work with them, considering their connection to the African roots. On their websites, the brands made notable claims, such as being a ‘leading provider of technology solution from a South African base’, being ‘proudly South African’ or ‘providing software solutions for businesses in Nigeria’. Alternatively, some companies used words from African languages, such as Jambo, which means ‘hello’ in Swahili. Moreover, 63 WebStudio indicated that they are a digital agency that builds website and mobile app solutions, nurtures African talents and collaborates with both African and international clients. Furthermore, Albanny Technol- ogies claimed that they are ‘Nigeria’s most reliable web design com- pany’, while Madavi Agency used the slogan ‘Homegrown Service. Global Thinking’. 3.2. Study 2 – interview insights into intended positioning The interviews with 32 managers from the 140 B2B service brands mentioned in the Clutch (2020) ranking led to further insights into the development of a B2B service brand’s intended positioning. This section highlights three themes with regard to their approach to managing their websites: curating a sense of professionalism, building a sense of trust- worthiness and being proud of African roots. These themes have been discussed in the following sections, and the representative participant quotes have been used to expand on them. 3.2.1. Curating a sense of professionalism For many managers in our sample, curating a sense of professional- ism for their brand through their website creates a state of tension that they deal with. A challenge for a B2B service brand manager is being authentic or driving the reach for the brand. Especially around website development, website management and content creation, managers recognise that they have to manage considerable tension as part of their business operations. In developing their brands’ websites, managers evaluate their internal capabilities and explore if they should design the website themselves or outsource it to another agency, especially for brands providing branding, digital marketing or app development and which are not primary website design companies. There is the tension of developing the websites, to show their authenticity even though it might not be visually appealing compared to what the website would be if they outsourced the website development. Most of the participants (93.7%, n = 30) recognised the need for developing a clear message when posi- tioning their B2B brands through their websites, but this can be a challenge when they explore their options. P3, the chief marketing of- ficer of a digital marketing agency in Kenya, shared his experience, saying: ‘We know we needed a new website but we do not have the in-house capabilities to develop this, so we had to look for a web designer to help us’. This was also corroborated by P4, the COO of an app developing com- pany, who said: ‘People think because you develop an app, you should be able to develop your website. Yes, I tried this when we started, but now we have outsourced the design’. The experience of P18, the co-founder of a web development company in Nigeria, was different, however: ‘I designed the first website of the company and we have changed it over many years, but it is still done in house.’ Beyond the designs, the content creation strategy was also another area of tension in the managers’ perspectives, as they evaluated their options for curating a sense of professionalism. Deciding the website’s content to convey a unique brand positioning could pose a challenge. In some situations, the managers explained how they try to integrate some physical elements into the digital dimensions but further questioned if it was necessary for their websites. P10, the head of marketing at a branding agency in Nigeria, shared how they had decided to put staff images on their website, saying: ‘We had to get everyone dressed in the branded T-shirt of the agency and take a group photograph for the website. We wanted to communicate our strength as a team for people visiting our website’. However, P21, the marketing and sales director of a marketing agency in South Africa, was sceptical about showing the graffiti on the wall and the colourful interior of their offices on their website, feeling it could be distracting and possibly not worth it for prospective consumers visiting their website. 3.2.2. Building a sense of trustworthiness Considering the competition these top-performing service brands face, their managers recognise the considerable tension involved in building a sense of trustworthiness. Twenty participants noted the increasing number of freelancers entering the market and offering lower-cost services. The participants recognised that this alters client expectations and affects their brands’ operations, as clients may prefer to engage with freelancers due to the lower overhead costs. As a result, many of the participants often had to emphasise their credentials, pro- fessionalism and expertise when positioning their brand for prospective clients. P31, the co-founder of a website development company in Tanzania, said: ‘We are in competition with developers in Europe and those freelancers on fiverr, everyone now thinks they can design a website, so it’s IndustrialMarketingManagement108(2023)237–250244 E. Mogaji et al. important for us to update our website with our certifications, partnership, and client testimonials’. While building trust on the home front, managers also contemplate building trust from a global perspective. Although they are not certain it enhances their brand, they feel exploring different options is important. A Nigerian manager shared his experience in trying to get featured in Forbes. He said: ‘We believe our reputation will grow and people can trust us more when we get featured in Forbes. You know Nigerians have a different reputation, but we believe having the Forbes feature on our website will help our brand’. P25, however, had a different opinion and doubted if the prospective clients care about those elements on the website, saying: ‘I haven’t had a client tell me seeing us on Forbes makes them trust us, there is pressure on everyone to appear trustworthy but it’s not always a guarantee’. In further exploring the tension in building a sense of trustworthi- ness, many managers reiterated the idea of putting fictitious information like testimonials, portfolios and case studies on their brands’ websites to make their brands appear hardworking and productive, even though they may not have many clients. Some participants replied that it was, citing the aphorism ‘fake it till you make it’. Two participants, however, stated unequivocally that they preferred to use a real client, giving the customer’s firm the name and a link to their website to verify the claim. Another manager said it is possible to show works for imaginary brands, which can be uploaded on the website to make it look like current works. He said: ‘We have junior designers and developers working on projects for a dummy company and we upload these designs in our portfolio on the website, but we often declare that it’s a sample work and not for a real client’. 3.2.3. Being proud of African roots The managers of these African B2B brands struggle with some ten- sion around their globalisation versus localisation approach. This ten- sion mainly emerges around the topic of recognising their African roots. The managers are aware of the global and digitally driven nature of their services, but they feel challenged about their approach towards glob- alisation and competing with global clients or reinforcing their huge credibility within the local markets. The managers recognised that they have the certifications and partnerships (with Amazon, Google and Facebook) that allow them to compete and pitch for projects on a global level. Seven of the participants stated that they are keen to showcase their African roots on their websites and that they express pride in being from an African country and displaying their African visual identity and design. This feeling of pride was most prominent among the South Af- rican participants, who often included this information on their web- sites. Two Nigerian participants stated that they incorporate the slogan ‘Proudly Nigerian’ in all of their presentations and customer-facing communications. Along with their pride in their African roots, several of the partici- pants shared the challenges they face because of operating from Africa and how their continent of origin shapes their buyers’ perceptions. Many of them noted an inherent difficulty with the concept of an ‘African B2B brand’, stating that several clients were often hesitant to work with them. According to several participants, customers frequently have a negative perception of African B2B service brands, expecting them as cheap, and negotiate ‘ridiculous’ pricing. As a result, numerous partic- ipants acknowledged the struggles they encounter and highlighted the importance of building a strong B2B firm to transcend these limiting features of the African brand. P8 said: ‘People don’t believe something good can come out of Africa. Coupled with the fact that you are competing with outside brands, the persisting notion about Africa and Africans often poses a challenge and a hurdle’. This notion was corroborated by P30, who feels that the lack of patronage from Africa makes them question the value of their African roots, saying: ‘Even our brothers and sisters here don’t appreciate our work. They do not see value in what we have to offer, and they want to go to Europe or Asia to get their app developed’. The tension managers face regarding this issue is in evaluating the right approach to take – to compete with other service providers from Europe, Asia or America or to intensify their stronghold in Africa. One of the participants cited how they had to use the testimonials of a very popular businessman in Nigeria on their website. She stated that: ‘Having him on our website showcase that we are provide outstanding service even though we are based in Nigeria, we want to be proudly associated with the big Nigeria clients we have worked with’. Considering these are the top- performing brands in Africa, the managers feel they have an edge over the competition in Africa, but they think that success is not guaranteed outside Africa for them. This was corroborated by a Kenyan manager, who said: ‘In Kenya, I can say we are the best app developers, we have done many projects and displayed this on our portfolio on the website, but I don’t think we are ready to go global now, allow us conquer Africa’. According to several participants, the intense competition in the African technology industry presents a tension for them on how to strategically position their brand. The participants were aware of the influx of large international service providers from Europe and North America vying for African clients. This influx often challenges them to strengthen relationships with their customers, highlight their African connection and reiterate that they also provide good service, if not better than the large international service providers. P16, the marketing di- rector of a branding agency in Nigeria, stated: ‘We are seeing an influx of large agencies coming in the form of network agencies and this has presented a huge concern for us on how to reinforce our relationship with our existing local brands, we need to protect our market share’. In summary, managers of B2B Service Brands in Africa recognise the need to curate a sense of professionalism and build a sense of trust- worthiness but also face considerable tension in managing these polar- ising decisions. The managers recognise that their brands’ websites offer global platforms to showcase their services yet there is a significant amount of tension involved in managing the branding process, espe- cially in the global and local context, the impending competition from global and local brands and the strategic direction, content creation and management strategy of their websites. Table 2 presents a summary of the themes and additional partici- pants’ quotes. 4. Discussion B2B brand positioning is largely focused on the relevance of posi- tioning strategies (Blankson & Kalafatis, 2019; Iyer et al., 2019; Leek & Christodoulides, 2011), with little consideration of the interplay be- tween communicated or delivered brand positioning via websites with the role of managers in shaping the market narratives (Williams et al., 2019; Beverland & Cankurtaran, 2022). We showed that signalling theory is a useful lens that weaves its way through many of its communication practices and intended brand positioning. Study 1 included a CA of the 140 top-performing African B2B service brands from Clutch (2020) list. We found that the African B2B service brands use five common signals on their websites to express their positioning strategies around ‘professionalism’, ‘trust’ and ‘African roots pride’. These common signals include competence (i.e., ‘we can do it’), past customer experience (i.e., ‘we did it before’), reputation (i.e., ‘what they say about us’) and availability (i.e., ‘we are here for you’). In study 2, we turned to managerial insights by identifying the tensions that emerge in selecting these relevant brand positioning strategies and brand posi- tioning processes for their firms. See Table 2 and Fig. 2. We contribute to the B2B brand positioning literature by showing how the alignment between the signal on delivered positioning on brands’ websites and managers’ perception in developing brand posi- tioning may shape different positioning strategies. Specifically, we showed that B2B brand positioning may not always be generic, as sug- gested by prior research (Blankson et al., 2020; Iyer et al., 2019), but may be distinctive when it is perceived as authentic and linked to IndustrialMarketingManagement108(2023)237–250245 E. Mogaji et al. Fig. 2. Summary of _ndings from Study 1 and Study 2 with key themes and tensions. organisation-level attributes. Brodie & Benson-Rea (2016) suggested that geographical regions can create a sustained strategic advantage for all network actors, yet we found that such positioning needs to deal with the challenge of ‘reterritorialising’ the brand in the placelessness of globalisation (Beverland & Cankurtaran, 2022). In our study, the par- ticipants were proud of their African roots but were also aware that many prospective buyers could associate these roots with substandard quality or cheaper prices. This is consistent with the provenance paradox, where companies from emerging countries are unable to command a fair price because consumers and buyers associate certain geographies with the best product or services. Thus, developing market firms can’t command fair prices even when the quality of their services is on par or superior to that of established players (Deshpand´e, 2010). Indeed, for many companies in our sample, their African roots appear to pose an inherent challenge, especially in the technology domain. While many B2B brands derive value from their country of origin (e.g., ‘Proudly Nigerian’), this could be overshadowed by the buyers’ commonly negative and undermining perceptions of African products and services, as Verster et al. (2019) found for architectural services in South Africa. Based on our findings, the positioning of an African B2B service brand appears to be intertwined with the existing perception of Africa as their continent of origin and their home country as their country of origin. This seems to be a promising topic to explore in other emerging contexts (e.g., South America or the Middle East), where B2B companies experience a tension between brand localness and globalness (Mohan et al., 2018). In other words, our findings suggest that a ‘continent of origin’ effect might exist alongside the well- established ‘country of origin’, which is an important cue for purchas- ing decisions (Balabanis & Siamagka, 2017; Koch & Gyrd-Jones, 2019a, 2019b). This finding also underscores that the so-called ‘liability of origin’ (Ramachandran & Pant, 2010) or of ‘emergingness’ (Madhok & Keyhani, 2012) might affect not only multinational firms engaged in internationalisation activities in emerging markets in developed coun- tries but also the companies of different sizes engaged in domestic and regional marketing initiatives within their same continent of origin. Unlike the findings of the study of Iyer et al. (2019) and the wider B2B brand positioning literature (Blankson & Kalafatis, 2019), we also found space for considering both the external positioning signals (e.g., via websites) and the internal perception of positioning strategies as a way to deal with the sales-led culture of many B2B firms. This approach is consistent with the call for an alignment process, especially in linking the front end of positioning (e.g., websites) with the backstage (e.g., managerial role in determining a brand positioning; Hatch & Schultz, 2008, 2010). In doing so, we argue that B2B positioning, especially in the African context, needs to consider wider stakeholders’ interaction in articulating the meaning of their competitive brand positioning (Gus- tafson & Pomirleanu, 2021; Iglesias, Landgraf, Ind, Markovic, & Koporcic, 2020). Aware of the fierce competition and the negative aura surrounding African products and services, the managers in our sample converged on the significance of developing clear and multiple messages through their website. In line with Gur and Greckhamer (2019), the managers in our study appeared to use a customer-centred approach to competitor identification, as they considered alternative ways for cus- tomers to obtain a service as a starting point (e.g., company versus freelancers). Aside from the challenge these African B2B brands face in competing with large international companies, Anwar and Graham (2020) acknowledge that freelancing has a disruptive impact on the African service industry and, hence, would be a useful actor to engage with. These diverse insights lead to three theoretical contributions. 4.1. Theoretical contributions This study carries several theoretical contributions. First, given the limited research on B2B brand positioning in Africa compared to that on North America, Europe or Asia (Blankson et al., 2020; Cortez & John- ston, 2017; Odoom, Narteh, & Boateng, 2017), our research contributes to the literature on B2B brand positioning in emerging markets by exploring the relationship between communicated positioning signals in an online environment (e.g., firm websites) and managers’ perception of positioning strategies and activities. Previously, it tended to focus on a single African country or a specific service industry (Blankson, Ketron, & Coffie, 2017; Coffie & Owusu-Frimpong, 2014). Our study specifically uncovers the interplay between positioning signals on firm websites and managers’ perceptions and related tensions. Despite some differences in IndustrialMarketingManagement108(2023)237–250246 E. Mogaji et al. specific features, the results of Study 1 indicate that many domestic B2B service providers in Kenya, Nigeria, South Africa and Tanzania send similar signals to prospective service buyers to reduce the buyers’ perceived uncertainty about the brands. Although sustainability does not appear to be a prominent positioning feature in our sample, our findings support Coffie and Owusu-Frimpong (2014) conclusion that African brands often use case studies, proficiency levels and endorse- ments to highlight their service professionalism, which is consistent with the findings of Panda et al. (2019) and Williams et al. (2019) in devel- oped markets. However, some African B2B service brands do not limit themselves to following the generic positioning strategies used in developed markets. They can, and often do, engage in different positioning strategies salient to the geographical proximity to the audience and situational influences. We show that some firms engage with their African roots in their posi- tioning which is consistent with the findings of Beverland and Can- kurtaran (2022). The sense of pride in their African roots goes deep into the values of their audience, even in the placelessness of globalisation. However, we question the extent to which African roots and the brands’ sense of pride can continue to be relevant, especially in the technology domain. When these African B2B brands innovate and explore new products and markets, they need to inform and educate buyers in a specific way. Our data suggest that some African B2B brands may not have sufficient resources to promote and change stakeholders’ stigma towards African roots. Second, this research contributes new insights into the dynamic process underpinning the intended positioning for B2B service brands via either website (Blankson et al., 2014a; Iglesias et al., 2020). Previous research indicates that successful B2B brand positioning follows a managerial-driven perspective to ensure consistency in delivering brand promises (Leek & Christodoulides, 2011; S´anchez-Chaparro et al., 2022). But the current study indicates that having stable and generic positioning strategies as a basis for identification and differentiation may no longer be sufficient to differentiate African B2B brands from competitors or attract multiple stakeholders (e.g., global clients, share- holders, freelancers, local providers) that are of utmost importance to the survival of African B2B service brands. Given the recent improve- ments in the technological landscape, African B2B service brands need to more adequately embrace their brand positioning as a dynamic social interactive process involving a multiplicity of stakeholders (Guenther & Guenther, 2019; Koch & Gyrd-Jones, 2019a, 2019b). Gustafson and Pomirleanu (2021) highlight that involving actors outside the brand can contribute to the brand’s legitimacy in their social roles as language interpreters and storytellers. In reaching the prospective clients, web- sites offer a great tool to initiate conversations. The current research adds that tensions around African roots and connections with other stakeholders can be explored in greater detail via storytelling to enhance these brands’ positioning. 4.2. Managerial implications Our multi-method exploratory research results provide actionable recommendations for managers of B2B service brands operating in the African context. First, managers can use our findings to get information on website-related decisions. We identified specific groups of website features signalling competence (i.e., ‘we can do it’), past customer experience (i.e., ‘we did it before’), reputation (i.e., ‘what they say about us’) and availability (i.e., ‘we are here for you’). This approach draws managers’ attention to how each website feature conveys a signal about the company’s actual positioning. We produced a mock-up design of a homepage (Fig. 3) to illustrate these signals to B2B service brand man- agers in emerging contexts who are creating or updating their website content. This mock-up is an exemplar based on the best practice insights we have gathered together from evaluating the websites of other rele- vant African businesses. the activities involved in the definition of their intended positioning – whether they are developing a clear message with prospective buyers across their websites or are building their credibility through their customers, brands or third parties. Moreover, our findings underscore how the evolution of competition pushes managers to continuously redefine their offerings and brand positioning efforts. The way B2B service brands manage their service portfolios (e.g., restructuring their offerings) and considering new forms of competitors (e.g., freelancers) has important implications. In managing the tension around curating a sense of professionalism, managers must evaluate their business operations. While it may be ex- pected that a web development company may design their websites on their own, an app development or brand agency company should consider their in-house capabilities for designing and managing the website or possibly outsourcing it. As a company grows and many staff members get on board, its managers need to consider their continuity plans, especially allocating responsibilities for managing the websites. Although the founder may have been responsible for designing the website, there should be plans for continuity in the absence of the manager. Content creation strategies in terms of frequency of update, the accuracy of content, the type of content to be uploaded and the people responsible for all this need to be well documented. Last, but not least, our two studies draw managers’ attention to the paradoxical nature of African roots as a source of pride for African B2B service providers, an added value for some buyers and a basis of scep- ticism for other clients. This finding invites managers to consider both the challenges and opportunities arising from defining and communi- cating an African-based positioning. 5. Conclusion This exploratory study addressed critical questions about the posi- tioning of domestic African B2B service brands. Our research high- lighted the challenges and opportunities associated with brand positioning that B2B service managers face by uncovering the para- doxical nature of their African roots. Given the growth of the service sector in African emerging economies, this empirical study sheds light on an important topic and provides a more solid basis for managerial action and future research in this area. The research comes with some limitations, which could provide promising avenues for future research. First, the empirical data for this study does not yield a longitudinal understanding of the brand posi- tioning process, and insights from consumers were also not explored. Therefore, it cannot truly lay claim to the branding process view and co- creation of brand image and identity. To assess the congruence of all brand positioning levels (Blankson et al., 2014b), data gathered from customers would be necessary to tap into the perceived brand posi- tioning. Future studies can explore consumers’ perspectives to gain a better understanding of the co-creation and cooperation between a brand and its various stakeholders. Second, the study’s sample came from a pre-existing list of top- performing brands compiled by an international B2B research com- pany, Clutch. As the list is based on a self-nomination process, our sample may include only companies proactively engaged in marketing activities and seeking external recognition for their efforts. Although this sample composition is coherent with the exploratory nature of our study, the Clutch-based sample may portray a limited view of Africa’s B2B service sector. Moreover, the Clutch ranking featured only four countries and professional B2B services, which are highly knowledge- intensive and more complex than other types of services (Fitzsimmons et al., 1998; Guenther & Guenther, 2019). Such constraints may affect our ability to generalise our findings. Hence, future replications with a broader set of B2B companies from a wider range of service sectors and more countries (e.g., obtained from scraping the web) would strengthen this study’s conclusions. Second, managers could also use the interview outcomes to analyse Finally, our research employed a cross-sectional approach, capturing IndustrialMarketingManagement108(2023)237–250247 E. Mogaji et al. Fig. 3. Examplar website mock-up based on the best practice insights gathered together from evaluating the websites of other relevant African businesses. IndustrialMarketingManagement108(2023)237–250248 E. Mogaji et al. website information at a specific point in time and retrospectively interrogating managers on their positioning-related activities. Posi- tioning is a dynamic process that involves changes and repositioning over time. This is mirrored in the dynamic nature of websites, which often alter over time (Williams et al., 2019). A longitudinal research approach that tracks and compares changes in managers’ accounts of their positioning efforts and website content over time could be a fruitful research path to better understanding the dynamic nature of positioning. Appendix A. Supplementary data Supplementary data to this article can be found online at https://doi. org/10.1016/j.indmarman.2022.12.003. References Adokou, F. A., & Kyere-Diabour, E. (2017). Positioning strategies of retail firms in Ghana. Journal of African Business, 18(2), 221–237. Åhlstr¨om, P., & Nordin, F. (2006). Problems of establishing service supply relationships: Evidence from a high-tech manufacturing company. Journal of Purchasing and Supply Management, 12(2), 75–89. Alden, D. L., Steenkamp, J. B. E. M., & Batra, R. (1999). Brand positioning through advertising in Asia, North America, and Europe: The role of global consumer culture. Journal of Marketing, 63(1), 75–87. Alkire, L., & Hammedi, W. (2021). Guest editorial: Expanding service research in the MEA (Middle East and Africa) region. Journal of Services Marketing, 35(7), 849–856. Amankwah-Amoah, J., Boso, N., & Debrah, Y. A. (2018). Africa rising in an emerging world: An international marketing perspective. International Marketing Review, 35(4), 550–559. Anwar, M. A., & Graham, M. (2020). Hidden transcripts of the gig economy: Labour agency and the new art of resistance among African gig workers. Environment and Planning A: Economy and Space, 52(7), 1269–1291. Arnold, D. J., & Quelch, J. A. (1998). New strategies in emerging markets. MIT Sloan Management Review, 40(1), 7–23. Badejo, F. A., Charles, S., & Millisits, J. (2022). To build my career or build my brand? Exploring the prospects, challenges and opportunities for sportswomen as human brand. European Sport Management Quarterly, 22(3), 379–397. Balabanis, G., & Siamagka, N. T. (2017). Inconsistencies in the behavioural effects of consumer ethnocentrism: The role of brand, product category and country of origin. International Marketing Review, 34(2), 166–182. Beverland, M., Napoli, J., & Yakimova, R. (2007). Branding the business marketing offer: Exploring brand attributes in business markets. The Journal of Business and Industrial Marketing, 22(6), 394–399. Beverland, M. B., & Cankurtaran, P. (2022). Closing corporate branding gaps through authentic internal brand strategies. In The Routledge companion to corporate branding (pp. 378–394). Routledge. Blankson, C., Iyer, P., Owusu-Frimpong, N., Nwankwo, S., & Hinson, R. (2020). Positioning strategies of foreign and indigenous firms in an African cultural milieu. Journal of Business Research, 119, 627–638. Blankson, C., & Kalafatis, S. P. (2019). Positioning strategies in industrial and B2B markets. Industrial Marketing Management, 81, 14–15. Blankson, C., Kalafatis, S. P., Coffie, S., & Tsogas, M. H. (2014a). Comparisons of media types and congruence in positioning of service brands. The Journal of Product and Brand Management, 23(3), 162–179. Blankson, C., Kalafatis, S. P., Coffie, S., & Tsogas, M. H. (2014b). Comparisons of media types and congruence in positioning of service brands. The Journal of Product and Brand Management, 23(3), 162–179. Blankson, C., Ketron, S., & Coffie, S. (2017). Positioning strategies by foreign retailers at the Accra mall in Ghana: A case study approach. Journal of Managerial Issues, 29(3), 294–314. Blankson, C., Ketron, S., & Darmoe, J. (2017). The role of positioning in the retail banking industry of sub-Saharan Africa. The International Journal of Bank Marketing, 35(4), 685–713. Braun, V., & Clarke, V. (2006). Using thematic analysis in psychology. Qualitative research in psychology, 3(2), 77–101. Brodie, R. J., & Benson-Rea, M. (2016). Country of origin branding: an integrative perspective. Journal of Product & Brand Management, 25(4), 322–336. https://doi. org/10.1108/JPBM-04-2016-1138 Camprubí, R., & Coromina, L. (2016). Content analysis in tourism research. Tourism Management Perspectives, 18, 134–140. Cartwright, S., Liu, H., & Raddats, C. (2021). Strategic use of social media within business-to-business (B2B) marketing: A systematic literature review. Industrial Marketing Management, 97, 35–58. Casidy, R., & Yan, L. (2022). The effects of supplier B2B sustainability positioning on buyer performance: The role of trust. Industrial Marketing Management, 102, 311–323. Clutch. (2020). Clutch celebrates the highly-rated B2B companies in Kenya, Nigeria, South Africa & Tanzania for 2020. https://clutch.co/press-releases/celebrates -highly-rated-b2b-companies-kenya-nigeria-south-africa-tanzania-2020. Coffie, S. (2016). An exploration of managerial views on positioning Services in Ghana. Journal of African Business, 17(3), 360–376. Coffie, S., Blankson, C., & Dadzie, S. A. (2018). A review and evaluation of market orientation research in an emerging African economy. Journal of Strategic Marketing, 28(7), 565–582. Coffie, S., & Owusu-Frimpong, N. (2014). Alternative positioning strategies for services in Ghana. Thunderbird International Business Review, 56(6), 531–546. Connelly, B. L., Certo, S. T., Ireland, R. D., & Reutzel, C. R. (2011). Signaling theory: A review and assessment. Journal of Management, 37(1), 39–67. Cortez, R. M., & Johnston, W. J. (2017). The future of B2B marketing theory: A historical and prospective analysis. Industrial Marketing Management, 66, 90–102. Coulter, K. S. (2012). Special issue on “Marketing and social media”. Management Research Review, 35(9), 1–5. Cretu, A. E., & Brodie, R. J. (2007). The influence of brand image and company reputation where manufacturers market to small firms: A customer value perspective. Industrial Marketing Management, 36(2), 230–240. Czarnecka, B., & Mogaji, E. (2020). How are we tempted into debt? Emotional appeals in loan advertisements in UK newspapers. The International Journal of Bank Marketing, 38(3), 756–776. Davvetas, V., Diamantopoulos, A., & Halkias, G. (2016). ‘Turning the tables’: Building strong local brands to compete against global rivals. In 2016 proceedings annual conference of the European marketing academy (pp. 24–27). Oslo: EMA. De Chernatony, L., & Riley, F. D. O. (1999). Experts’ views about defining services brands and the principles of services branding. Journal of Business Research, 46(2), 181–192. Deshpand´e, R. (2010). Why you aren’t buying Venezuelan chocolate. Harvard Business Review, 88(12), 25–27. Dludla, G. M., & Dlamini, S. (2018). Does brand orientation lead to brand loyalty among senior and top management in a south African business-to-business organisation? South African Journal of Business Management, 49(1), 1–10. Ellinger, A. E., Lynch, D. F., Andzulis, J. K., & Smith, R. J. (2003). B-to-B E-commerce: A content analytical assessment of motor carrier websites. Journal of Business Logistics, 24(1), 199–220. Farinloye, T., Mogaji, E., Aririguzoh, S., & Kieu, T. A. (2019). Qualitatively exploring the effect of change in the residential environment on travel behaviour. Travel Behaviour and Society, 17, 26–35. Fitzsimmons, J. A., Noh, J., & Thies, E. (1998). Purchasing business services. The Journal of Business and Industrial Marketing, 13(5), 370–380. Guenther, M., & Guenther, P. (2019). The value of branding for B2B service firms-the shareholders’ perspective. Industrial Marketing Management, 81, 88–101. Gur, F. A., & Greckhamer, T. (2019). Know thy enemy: A review and agenda for research on competitor identification. Journal of Management, 45(5), 2072–2100. Gustafson, B. M., & Pomirleanu, N. (2021). A discursive framework of B2B brand legitimacy. Industrial Marketing Management, 93, 22–31. Hatch, M. J., & Schultz, M. (2008). Taking brand initiative: How companies can align strategy, culture, and identity through corporate branding. John Wiley & Sons. Hatch, M. J., & Schultz, M. (2010). Toward a theory of brand co-creation with implications for brand governance. Journal of Brand Management, 17(8), 590–604. He, J., Huang, H., & Wu, W. (2018). Influence of interfirm brand values congruence on relationship qualities in B2B contexts. Industrial Marketing Management, 72, 161–173. Helm, S. V., & ¨ Ozergin, B. (2015). Service inside: The impact of ingredient service branding on quality perceptions and behavioral intentions. Industrial Marketing Management, 50, 142–149. Herbst, U., & Merz, M. A. (2011). The industrial brand personality scale: Building strong business-to-business brands. Industrial Marketing Management, 40(7), 1072–1081. Hutt, M. D., & Speh, T. W. (2017). Business marketing management: B2B (2nd ed.). Cengage Leaning. Iankova, S., Davies, I., Archer-Brown, C., Marder, B., & Yau, A. (2019). A comparison of social media marketing between B2B, B2C and mixed business models. Industrial Marketing Management, 81, 169–179. Iglesias, O., Landgraf, P., Ind, N., Markovic, S., & Koporcic, N. (2020). Corporate brand identity co-creation in business-to-business contexts. Industrial Marketing Management, 85, 32–43. Iyer, P., Davari, A., Zolfagharian, M., & Paswan, A. (2019). Market orientation, positioning strategy and brand performance. Industrial Marketing Management, 81, 16–29. Jalkala, A. M., & Ker¨anen, J. (2014). Brand positioning strategies for industrial firms providing customer solutions. The Journal of Business and Industrial Marketing, 29(3), 253–264. Kalafatis, S. P., Blankson, C., Luxly Boatswain, M., & Tsogas, M. H. (2020). Preference for action: Regulatory mode in B2B positioning decision-making. Journal of Business & Industrial Marketing, 35(12), 2111–2125. Kaur, J., Mogaji, E., Wadera, D., & Gupta, S. (2022). Sustainable consumption practices in Indian households: a saga of environment management linked to Indian ethos and generational differences. Society and Business Review, 17(3), 441–468. https://doi. org/10.1108/SBR-08-2021-0132 Keller, K. L. (2005). Strategic brand management process. Wiesbaden: Gabler Verlag. Ker¨anen, J., Piirainen, K. A., & Salminen, R. T. (2012). Systematic review on B2B branding: Research issues and avenues for future research. Journal of Product and Brand Management, 21(6), 404–417. Kirmani, A., & Rao, A. R. (2000). No pain, no gain: A critical review of the literature on signaling unobservable product quality. Journal of Marketing, 64(2), 66–79. Koch, C. H., & Gyrd-Jones, R. I. (2019a). Corporate brand positioning in complex industrial firms: Introducing a dynamic, process approach to positioning. Industrial Marketing Management, 81, 40–53. Koch, C. H., & Gyrd-Jones, R. I. (2019b). Corporate brand positioning in complex industrial firms: Introducing a dynamic process approach to positioning. Industrial Marketing Management, 81, 40–53. IndustrialMarketingManagement108(2023)237–250249 E. Mogaji et al. Kolk, A., & Rivera-Santos, M. (2018). The state of research on Africa in business and management: Insights from a systematic review of key international journals. Business & Society, 57(3), 415–436. Mohan, M., Brown, B. P., Sichtmann, C., & Schoefer, K. (2018). Perceived globalness and localness in B2B brands: A co-branding perspective. Industrial Marketing Management, 72, 59–70. Korkmaz, S., & Messner, J. I. (2008). Competitive positioning and continuity of Odoom, R., Narteh, B., & Boateng, R. (2017). Branding in small-and medium-sized construction firms in international markets. Journal of Management in Engineering, 24 (4), 207–216. enterprises (SMEs): Current issues and research avenues. Qualitative Market Research: An International Journal, 20(1), 68–89. Kotler, P., & Keller, K. (2003). Marketing management (international version). Englewood Cliffs, NJ: Prentice Hall. Kotler, P., & Pfoertsch, W. (2006). Being known or being one of many. Brand Management, B2B, 1–14. Landis, J. R., & Koch, G. G. (1977). An application of hierarchical kappa-type statistics in the assessment of majority agreement among multiple observers. Biometrics, 33(2), 363–374. Landrum, N. E., & Ohsowski, B. (2018). Identifying worldviews on corporate sustainability: A content analysis of corporate sustainability reports. Business Strategy and the Environment, 27(1), 128–151. Leek, S., & Christodoulides, G. (2011). A literature review and future agenda for B2B branding: Challenges of branding in a B2B context. Industrial Marketing Management, 40(6), 830–837. Leke, A., & Yeboah-Amankwah, S. (2018). Africa: A crucible for creativity. Harvard Business Review, 96(6), 116–126. Lincoln, Y. S., & Guba, E. G. (1985). Naturalistic inquiry. Sage. Madhok, A., & Keyhani, M. (2012). Acquisitions as entrepreneurship: Asymmetries, opportunities, and the internationalization of multinationals from emerging economies. Global Strategy Journal, 2(1), 26–40. Magno, F., & Cassia, F. (2020). Establishing thought leadership through social media in B2B settings: Effects on customer relationship performance. The Journal of Business and Industrial Marketing, 35(3), 437–446. Mccoll, R., Truong, Y., & Rocca, A. L. (2019). Service guarantees as a base for positioning in B2B. Industrial Marketing Management, 81, 78–86. Merriam, S. B., & Tisdell, E. J. (2015). Qualitative research: A guide to design and implementation. John Wiley & Sons. Miles, B., Huberman, M., & Saldana, J. (2013). Qualitative data analysis: A methods sourcebook. Thousand Oaks, CA: SAGE. Mogaji, E., Adeola, O., Hinson, R. E., Nguyen, N. P., Nwoba, A. C., & Soetan, T. O. (2021). Marketing bank services to financially vulnerable customers: Evidence from an emerging economy. The International Journal of Bank Marketing, 39(3), 402–428. Mogaji, E., Czarnecka, B., & Danbury, A. (2018). Emotional appeals in UK business-to- business financial services advertisements. The International Journal of Bank Marketing, 36(1), 208–227. Mogaji, E., & Nguyen, N. P. (2022). Managers’ understanding of artificial intelligence in relation to marketing financial services: insights from a cross-country study. International Journal of Bank Marketing, 40(6), 1272–1298. https://doi.org/10.1108/ IJBM-09-2021-0440 Omokaro-Romanus, C., Anchor, J. R., & Konara, P. (2019). The internationalization of Nigerian firms: Motivations and location patterns. Thunderbird International Business Review, 61(1), 75–88. Panda, S., Paswan, A. K., & Mishra, S. P. (2019). Impact of positioning strategies on franchise fee structure. Industrial Marketing Management, 81, 30–39. Ramachandran, J., & Pant, A. (2010). The liabilities of origin: An emerging economy perspective on the costs of doing business abroad. In The past, present and future of international business & management. Emerald Group Publishing Limited. Ries, A., & Trout, J. (1986). Marketing warfare. Journal of Consumer Marketing, 3(4), 77–82. Roberts, J., & Merrilees, B. (2007). Multiple roles of brands in business-to-business services. The Journal of Business and Industrial Marketing, 22(6), 410–417. S´anchez-Chaparro, T., Soler-Vic´en, M. ´ A., & G´omez-Frías, V. (2022). Be good and look good: Communicating the triple bottom line through corporate websites. Journal of Business Research, 144, 136–145. Shenton, A. K. (2004). Strategies for ensuring trustworthiness in qualitative research projects. Education for information, 22(2), 63–75. Sheth, J. N., & Sinha, M. (2015). B2B branding in emerging markets: A sustainability perspective. Industrial Marketing Management, 51, 79–88. Smeltzer, L. R., & Ogden, J. A. (2002). Purchasing professionals’ perceived differences between purchasing materials and purchasing services. Journal of Supply Chain Management, 38(1), 54–70. Spence, M. (1973). Job market signaling. Quarterly Journal of Economics, 87(3), 355–374. Spyropoulou, S., Skarmeas, D., & Katsikeas, C. S. (2010). The role of corporate image in business-to-business export ventures: A resource-based approach. Industrial Marketing Management, 39(5), 752–760. Taylor, I. (2014). Is Africa rising? The Brown Journal of World Affairs, 21(1), 143–161. Verster, A., Petzer, D. J., & Cunningham, N. (2019). Using brand identity to build brand equity: A comparison between the south African and Dutch business-to-business architectural industry. South African Journal of Business Management, 50(1). Virtsonis, N., & Harridge-March, S. (2009). Brand positioning in the B2B online environment: A case from the UK print industry. Journal of Brand Management, 16(8), 556–570. Williams, Z., Lueg, J. E., Hancock, T., & Goffnett, S. P. (2019). Positioning through B2B carrier signals: Understanding how service quality is communicated via websites. Industrial Marketing Management, 81, 54–64. IndustrialMarketingManagement108(2023)237–250250
10.1177_09579265221148691
1148691 DAS0010.1177/09579265221148691Discourse & SocietyWei research-article2023 Article Kindness in British communities: Discursive practices of promoting kindness during the Covid pandemic Discourse & Society 2023, Vol. 34(4) 502 –520 © The Author(s) 2023 Article reuse guidelines: sagepub.com/journals-permissions https://doi.org/10.1177/09579265221148691 DOI: 10.1177/09579265221148691 journals.sagepub.com/home/das Jilan Wei Changzhou Institute of Technology, China University of Sussex, UK Abstract This research adopts Critical Discourse Analysis as a perspective to explore how kindness was expressed and promoted in university communities and city communities from January to March in 2020 when the Covid pandemic broke out in the UK and provide a window on British culture in which kindness was expressed and promoted through discourse. It combines a qualitative method with a corpus-based quantitative method. It is found that kindness was meant for providing support and showing compassion and inclusion to community members and that strategies in lexis, syntax and metaphor can reproduce or resist the expression and promotion of kindness in communities. During the pandemic, the intentional kindness expressed by community authorities was respect of diversity rather than inclusion of different values or ethnicity and no substantial support was provided to vulnerable members even though authorities were trying to impress the public by claiming that they were making constant efforts to support the community. Case studies revealed that we should caution against the use of passivation and the pronouns like they. Keywords Community, corpus, critical discourse analysis, inclusion, kindness, pandemic, support Introduction This research adopts Critical Discourse Analysis to describe how community authorities intended to express and promote kindness through discourse to the target audience such as international students (e.g. students from China) and East Asian communities at the Corresponding author: Jilan Wei, Changzhou Institute of Technology, Changzhou, Jiangsu 213022, China; School of Media, Arts and Humanities, University of Sussex, Arts B348, Brighton, England BN1 9RH, UK. Email: jw715@sussex.ac.uk Wei 503 initial stage of the Covid pandemic. It concentrates on the linguistic strategies adopted by authorities for building an inclusive and supportive community among people with diverse backgrounds and concerns in light of the dynamics of the pandemic. This work can therefore reveal how the value of kindness is deployed by community authorities when public health and safety is being challenged. Community, traditionally, was defined based on social relations and interaction in a locality. With modernisation, urbanisation and globalisation, ‘community without pro- pinquity’ (Webber, 1964) came into being and community is formed when individuals pursue high levels of interaction, common interests, identity and shared values (Mannarini and Fedi, 2009). It is suggested that community be defined along three dimensions: eco- logical (space, time), social structural (networks of institutions and interaction) and sym- bolic cultural (identity, norms and values) dimensions (Hunter, 2018). In view of the heterogeneity of community, it is argued that community is defined not as a group of co-occurring populations, but as groups of co-occurring individuals of dif- ferent populations (Looijen and Van Andel, 1999: 218). In this view, a group of individu- als may be labelled as members of different communities. For example, international students may be part of the local (i.e. a city/town/village) community if they live off- campus as well as retaining their university community membership. Likewise, students from China in the UK not only belong to their host university community, but may also obtain membership in the East Asian community in that area. The varied interpretations of community consistently point to the significance of cul- tivating sense of belonging, for which the concept of kindness is of great use. Kindness describes any voluntary act that protects or benefits others and is not driven by external rewards or punishments (Eisenberg et al., 2006) but is motivated by compassion or con- cern (Peterson and Seligman, 2004). Acts of kindness are ‘essentially unobligated, often emotionally complex and always deeply social’ (Anderson and Brownlie, 2019: 12). Recent studies show that adolescents think kind acts involve 10 themes: emotional support, proactive support, social inclusion, positive sociality, complimenting, helping, expressing forgiveness, honesty, generosity, formal kindness such as fundraising and volunteering (Cotney and Banerjee, 2019); and that university students instantiate kind- ness by being polite, showing care and concern, being selfless, being self-aware or hav- ing a positive attitude, helping, improving other’s lives, being inclusive, nice/friendly, showing respect and following the ‘golden rule’ (Binfet et al., 2022: 448). Both groups include helping and being inclusive (social inclusion) as kind acts and believe being kind means being generous or selfless. However, it is claimed that kindness is not ‘a selfless helping’, but how help is offered, that is, ‘with gentleness, respect, amiability and con- cern’ (Faust, 2009: 297). Kindness is lubricative (as a ‘social lubricant’, Binfet et al., 2022: 444) as well as reciprocal and contagious (susceptible to ‘social contagion’, Tsvetkova and Macy, 2014). The beneficiaries of kind acts are likely to give a helping hand to others (Fowler and Christakis, 2010) out of gratitude (Bartlett and DeSteno, 2006) and elevation (an emo- tional response to witnessing acts of virtue or moral beauty) (Algoe and Haidt, 2009: 106). In this way, kindness promotes altruism and contributes to building social connec- tion and cultivating sense of belonging. 504 Discourse & Society 34(4) Kindness became a catchphrase of British community communication during the coronavirus outbreak when the public health was challenged and so many people suf- fered. In this research, what is focused on is the kindness expressed and described through public discourse during the pandemic. Background of the study The Covid-19 pandemic has been one of the biggest challenges to human survival. China was the first country suffering greatly from the Covid-19 coronavirus in the world. In the UK, March witnessed the rapid spread of the virus, therefore the first national lockdown was put into effect on 23 March 2020 to mitigate the worsening epidemic situation. In view of the highly infectious virus, all of the students, including international students, were advised to go back home because in-person teaching was terminated and they were put at great risk of infection in their journey. What’s worse, international students from China and other residents with East Asian backgrounds who stayed in the UK were sub- ject to racism and discrimination during the first wave of the disease. Some were assaulted and hurt physically and mentally, especially when they were wearing face masks (Lim et al., 2022: 20). Authorities in university and local communities recognised that these vulnerable members were in need of particular care and support. Data and methodology During the Covid-19 crisis, through emails and statements, community authorities have disseminated great amounts of pandemic information to members in the name of kind- ness, with the intention of protecting the public health and safety. Based on official emails and public statements collected from four universities and three city councils in England between 23 January (first day of Wuhan lockdown) and 23 March 2020 (first day of UK lockdown), two corpora (University Corpus and City Corpus) were built (see Table 1). These emails and statements were official in that they were released from a vice-chancellor, a provost and/or a registrar or a council leader to the general public, that is, the students and staff in a university or municipal community. The four British universities were selected for this research because they have a high percentage of international students: University College London (UCL; 54.3%), Table 1. Corpus details. Corpora University Corpus Words 63,115 City Corpus 19, 557 Communities Emails Statements UCL LU UY US LCC YCC BHCC 0 5 4 8 0 0 0 8 3 10 8 4 24 8 Wei 505 Lancaster University (LU; 29.8%), University of York (UY; 14.1%) and University of Sussex (US; 26.7%). Moreover, UY and US are the universities with confirmed cases of Covid found in February 2020. Among the four universities, UY and LU are located in the north, while US and UCL in the south. The three city councils chosen for the research are Lancaster City Council, York City Council and Brighton and Hove City Council because they are the communities to which three of the universities belong. Camden Council, to which UCL is affiliated, is absent because no pertinent public statements or emails were available either on its website or on that of London Councils. Despite the differences in scales, members, locations and goals, the chosen communities have mem- bers from different ethnic groups, with different cultural values or beliefs. In this context it is meaningful to examine how kindness was adopted by authorities for exercising control and protecting public safety and wellbeing during this health crisis. The perspective adopted for the research is Critical Discourse Analysis (known as CDA). According to Van Dijk (1993: 250), CDA studies ‘the relations between discourse structures and power structures more or less directly’. Between discourse and domi- nance, social cognition should be taken as the necessary theoretical (and empirical) ‘interface’. Ideologies are the ‘fundamental social cognitions that reflect the basic aims, interests and values of groups’ (Van Dijk, 1993: 258), ‘such as their identity, tasks, goals, norms, values, position and resources’ (Van Dijk, 1995: 18). They ‘allow members of a group to organise (admission to) their group, coordinate their social actions and goals, to protect their (privileged) resources, or, conversely, to gain access to such resources in the case of dissident or oppositional groups’ (Van Dijk, 1995: 19). In order to enact power, social actors who have privileged access to the discourse may manifest the dominance through context control as well as in some subtle and unintentional ways, such as lexical or syntactic style, rhetorical devices and local semantics. CDA therefore focuses on the discursive strategies used for legitimating control, or otherwise naturalising the social order, and especially relations of inequality (Fairclough, 1985). Situated in public health communication, CDA is of great value in showcasing how community authorities exer- cise control through discourse and how public discourse maintains or resists the relations of dominance and inequality. The discourse analysis is corpus-based. With the help of Sketch Engine, corpus data is examined in terms of frequency, collocation, wordlist and concordance (including CQL, Corpus Query Language). These corpus analysis techniques are adopted for dis- closing the contextual meaning of kindness and the discursive strategies for expressing or promoting kindness in university and municipal communities in times of health crisis. Furthermore, case studies are conducted with regard to wearing masks and returning home in an attempt to demonstrate what linguistic strategies better express and promote kindness within communities. Meaning of kindness in community Language is viewed as an ideological phenomenon, because ideologies can be encoded and communicated through language (Volosinov, 1973). Van Dijk (2012) echoed this by asserting that ‘discourse plays a fundamental role in daily expression and reproduction of ideologies’ (p. 3). Discourse can either contribute to the ‘enactment of dominance in 506 Discourse & Society 34(4) text and talk in specific contexts’ or have impact on ‘the minds of others’ (Van Dijk, 1993: 279). It is suggested that in doing discourse analysis, primary attention should be paid to the discursive ‘properties that express or signal the opinions, perspective, posi- tion, interests or other properties of groups’ (Van Dijk, 1995: 22), because all of these are included in ideology which any group may develop in order to cultivate ‘loyalty, cohe- sion, interaction and operation of its members’ (Van Dijk, 2011: 380). As part of group ideologies, kindness is explicitly or implicitly expressed in community communication. What type of kindness is stressed and pursued, and how effectively kindness is expressed and promoted can be tracked in detailed discourse analysis. Kindness has two occurrences in University Corpus and one in Community Corpus, as shown below in Examples 1–3. Coordination is adopted for the specification and emphasis of kindness in the context. According to Lang (1984: 19), a coordinate struc- ture is formed based on the hypothesis that conjuncts meet a set of conditions on their structural homogeneity. Conjuncts have two possible relations: compatibility (i.e. seman- tic non-distinctness, semantic inclusion and mutual independence) and incompatibility (contrariness and contradictoriness). Obviously, coordination used here illustrates the semantic inclusion between kindness and the three words support, compassion and inclusion. 1. Cllr Keith Aspden, Leader of City of York Council, said: It’s now five days since the Covid- 19 COBRA meeting was called, and we have been overwhelmed by the kindness and offers of support from York’s residents and businesses. (City Corpus) In Example 1, using coordination, authorities have marked the importance of support from residents and businesses as an act of kindness in community. In addition, exclusive we and passive voice are used to highlight authorities’ recognition of the acts of kindness and the great difference the kind support has made in community. We represents a group or an establishment with the speaker as a central or defining member (Wortham, 1996: 333) and it is divided into inclusive we and exclusive we (Daniel, 2005). Inclusive we signifies the speaker, the group or organisation he or she speaks on behalf of and the audience, while exclusive we doesn’t refer to the audience. The purposeful use of exclu- sive we displays their authoritative stance in recognising and advocating kindness in communities. 2. It is especially important that we all show kindness and compassion to our students who are from China, who will be finding this a difficult start to the year – and we are particularly concerned at making sure they are OK. (University Corpus) In Example 2, kindness means compassion for Chinese students in the community. The sentence pattern it is important that. . . allows authorities to adopt an objective per- spective to make a request. What should be noted here is that compassion is different from kindness because compassion prioritises the experience of those who witness suf- fering whereas kindness is defined largely by the beholder or recipient rather than the giver. Moreover, compassion grows out of privilege and reaffirming hierarchies and inequalities (Spelman, 2001). Here, compassion is advocated as a particular way to show Wei 507 kindness, but it reveals the inequality between members and highlights the superiority of the we-group. Additionally, in contrast to the inclusive we (all) and our which contribute to solidar- ity, the pronoun they reveals that Chinese students have been excluded from the univer- sity community. Van Dijk (1995) points out that ‘representation of ideologies are often articulated along an us vs them dimension’ (p. 22) and the use of us and them entails the positive and negative evaluations (Van Dijk, 1993: 264). Obviously, they is a sign of conscious or subconscious distancing, though students from China are marked as our group to whom great concern is expressed in this instance. Authorities othered the Chinese group of members when they were calling on members to show compassion. 3. The time has never been more important to demonstrate our values of kindness and inclusion – and to take action if we see anyone being treated in a harmful way. (University Corpus) Example 3 shows that inclusion is advocated as the third specification of kindness in community. Equating kindness with inclusion, authorities intended to deliver the guid- ance on stopping racism and discrimination within the community. Similar to Example 2, the first-person pronouns we and our are purposefully used here. As the modifier of the noun values, our, which is related to inclusive we, that is, the audience is included, ena- bles authorities to align members to the shared values and cultivate a sense of commu- nity. While exclusive we is employed to depict authorities’ agency in spotting the discriminatory behaviours, authorities’ agency hasn’t been explicitly represented in the specific response to racism and discrimination. Instead, the inanimate time is topicalised so as to collocate with the infinitive to take action. The pattern the time has never been more important to do . . .is used to emphasise the urgent importance of showing inclu- sion at that point of time. It conveys authorities’ stronger desire to advocate inclusion in community as compared to the pattern it is important to do. . .. In these three examples, coordination helps define kindness as offers of support com- passion and inclusion respectively. Example 1 shows the support from residents and businesses in city community has won acclaim from authorities as an act of kindness. Examples 2 and 3 indicate that compassion and inclusion have been highlighted in view of the needs of the highly vulnerable groups in communities: students from China and Asian communities. Expression and promotion of kindness in community Having analysed what kindness has meant in community during the pandemic, we now concentrate on how kindness was promoted when anybody could pose risks to public safety due to the highly contagious virus and when racism and abuse took place in communities. Showing inclusion To better understand kindness in community, we can pay our first attention to how com- munity is defined. Adjectives have the inherent function of characterising a person or an 508 Discourse & Society 34(4) object (Dixon, 1999). As attributive adjectives serve as pre-head internal dependent (i.e. part of a nominal) in the structure of the NP (Pullum and Huddleston, 2002: 528) and they can add some evaluative or descriptive information to the modified nouns, we examine the attributive adjectives of community to illustrate how community was described and evaluated by authorities. As Table 2 shows, wide, whole, South/East Asian, diverse and international (high- lighted) appear in both corpora. This suggests that in the surveyed communities diversity and solidarity are the core pursuit and South/East Asian communities have been brought in focus in community communication. In University Corpus, wide and university are used more frequently while in City Corpus global and Asian have higher frequency, as shown in Figure 1 (where the darker and larger a word is, the more frequently it appears in the corpus). This indicates both university communities and city communities paid close attention to diversity and inclusion. 4. The University is a global community, with one in every four people on our campus (both staff and students), coming from outside the UK. We value, celebrate and are proud of our inclusive and diverse community. (University Corpus) Table 2. Attributive adjectives of community. Noun University Corpus City Corpus Community wide, whole, York, university, remarkable, PG Research, IPC, South/ East Asian, vibrant, PGR, scientific, diverse, Sussex, campus, international global, South/East-Asian, Asian, international, tolerant, diverse, whole, voluntary, different, wide Figure 1. Frequencies of attributive adjectives of community. Wei 509 Concordance analysis can provide us with a wider view about what is advocated within communities. Example 4 emphasises the importance of inclusion and diversity in university communities by collocating community with global, diverse and inclusive. Additionally, the inclusive we and its variant our were intended for uniting members to show inclusion. 5. I’ve spoken to people from Asian communities about concerns around the origin of the virus being in China and how this could impact negatively on their community and businesses. I want to reassure our international communities that you’re a welcome and valued part of our city’s rich fabric. Brighton & Hove has long prided itself on being a fair and inclusive place to live, work and visit, for people across Britain and beyond. (City Corpus) In Example 5, using the pronoun you and the verb reassure, the city authority was trying to show kindness directly to international communities. The impersonal reference of you embodies a sense of informal camaraderie and a sense of universality (Kitagawa and Lehrer, 1990: 742). What is worth noting here is the adoption of the metaphor in which city is likened to fabric to stress the importance of diversity in the community. In order to promote kindness within the city, the authority mentioned the concerns expressed to Asian communities about the negative impact of the virus on them. By contrast, inclu- sion was shown to international communities which may include but is not limited to Asian communities. This subtle shift is likely to help authorities articulate their vague stance in supporting Asian communities and avoid any challenge from local communi- ties who may have bias towards Asian communities. Although authorities claimed that communities were inclusive and diverse, racism, abuse and discrimination did occur during the pandemic and it had severe social impacts. At the initial stage of the pandemic, in the UK, ‘maskaphobia’ (dubbed by The Guardian) prevailed and wearing a surgical mask in public, especially if you looked East Asian, could even invite racist attacks (Weale, 2020). 6. We have been saddened by reports of racism from a minority of the public. This type of behaviour has no place in York or anywhere else. As a city of sanctuary, we respect and welcome cultures and communities from across the world. (City Corpus) When racism and discrimination took place in British communities, authorities dis- played unfavourable attitudes in different ways. As illustrated in Examples 6–8, city authorities have shown their different stances. In Example 6, authorities expressed their sadness about reports of racism and announced the ban for racism in the community. Nonetheless, a passive structure is used for emotional expression and an active voice is adopted for announcing the ban. Although passivation implicates authorities’ negative attitudes towards racism, the ban with this type of behaviour as the topicalised subject doesn’t show any specific actions would be taken to stop racism. Instead, what has been advocated here is just formalistic respect of cultural diversity. 7. The Vice-Chancellor of Lancaster University, Professor Steve Bradley, said: ‘I have made it clear to our students that if they are subject to any form of racism or discrimination, they must report it immediately and we will take action. It’s very important at this difficult time that they feel they have our support and understanding’. (City Corpus) 510 Discourse & Society 34(4) In Example 7, authorities are trying to encourage students to report racism and dis- crimination by quoting a professor. Coordination used in this instance implies that no action would be taken if students who were subject to any form of racism or discrimina- tion didn’t report. Moreover, the frequent use of we, our, they clearly shows the distinc- tion between in-groups and out-groups in the community. 8. Councillor Dr Erica Lewis, leader of Lancaster City Council, said: ‘Our district is stronger for being a diverse and tolerant community. I am deeply concerned to hear that any residents has been subject to racist abuse, and condemn this behaviour. At a time when people are worried about friends and family, it is important that we show our support for those members of our community, and our concern for everyone infected and affected, particularly communities in China and South East Asia where the greatest impacts are being felt’. (City Corpus) In Example 8, authorities don’t use the pronoun they or any of its variants to refer to the people affected by racism. Conversely, any residents, those members of our commu- nity and everyone are adopted to include the vulnerable group in the community. In this way, support and inclusion are well expressed and advocated. 9. London is a diverse city which is home to a vibrant community from around the world. But like any major city, deplorable incidents like this can occur. During the current outbreak of coronavirus, we are very proud of the way in which our community has responded calmly and is supporting each other. We want to make clear that abuse, racism and hate speech have no place at UCL. (University Corpus) Examples 9–11 can depict how university authorities positioned themselves in deal- ing with racist abuse and hate speech during the pandemic. Similar to Example 6, Example 9 topicalises abuse, racism and hate speech to inform the members of a rele- vant ban. What’s interesting here is authorities’ explanation for the happening of the saddening incidents: London is a diverse community where the incidents are bound to happen. This naturalises the occurrence of racism, abuse and hate speech in London though it claims the rejection of the deplorable incidents at UCL. 10. There have been some cases in the news of people experiencing bullying and abuse related to coronavirus. We would like to remind everyone that UCL and Students’ Union UCL do not tolerate bullying or abuse of any kind, including racial discrimination and hate speech. (University Corpus) 11. Unfortunately, there have been some shocking media reports about abuse and racism at other universities linked to the current coronavirus situation. This behaviour will not be tolerated at Lancaster University. We encourage anyone experiencing or witnessing abuse or inappropriate behaviour to report it to the University. (University Corpus) Contrary to Example 9, Example 10 adopts an active voice to show an institutional stance in dealing with bullying or abuse: not tolerate. Example 11 demonstrates univer- sity authorities’ unfavourable attitudes towards virus-related racial discrimination. On the one hand, they displayed their intolerance of abuse and racism implicitly through Wei 511 passivation and topicalisation (This behaviour will not be tolerated). On the other hand, they actively encouraged reporting. The examples from the two types of communities consistently show that authorities suppressed their agency in tackling virus-related abuse and discrimination and they encouraged or requested members to report for getting support. In community commu- nication, what we can see is the promotion of support, diversity and inclusion, but we can’t see description of any specific action to prevent the happening of racism and dis- crimination. In addition, when delivering regulations, university and city authorities were aware of their domination in the public communication and they favoured the pat- terns such as it is important that. . . and we want/would like to remind/make clear that. . .. Further examination tells us these two patterns have different semantic proso- dies. Subordination in the pattern It is important that. . .carries positive prosody and expresses the shared values in communities, while what is subordinated to we would like to remind/make clear that. . .connotates negative profiles and conveys disagreement. Despite the different semantic prosodies, the two types of sentence patterns were intended for promoting kindness, respect and inclusion within communities, whereby authorita- tive and dominant position got guaranteed and maintained. Offering support As a word, support has two parts of speech. As a verb, in University Corpus it occurs 106 times (frequency per million words: 1500.63), while in City Corpus it has 61 occurrences (frequency per million words: 2751.84). As a noun, it has 161 instances (frequency per million words: 2279.26) in University Corpus and 55 instances (frequency per million words: 2481.17) in City Corpus. In University Corpus it is used more frequently as a noun, while in City Corpus it has higher frequency as a verb. Whether it is a noun or a verb, support has higher relative frequency in City Corpus than in University Corpus. However, in both corpora it ranks 22nd or higher (its rankings are highlighted in bold) in the word frequency list as each part of speech (see Table 3). In order to figure out how community authorities offered their support and help to members, we can take a close look at the use of modal verbs, as they carry important information about the sender’s attitude to the message and other interpersonal meanings (McCarthy, 1991). 12. We will let you know more details as soon as we can, but please note that this may take several weeks. (University Corpus) 13. We are grateful for all the support that so many people at Sussex have offered to our students and staff from China. Please know that the University is here to help and support you however we can. (University Corpus) 14. It is abundantly clear that we are living in unprecedented times and dealing with exceptional circumstances. As a city, we must find our resolve and do everything we can to support those who need our help the most. (City Corpus) 512 Discourse & Society 34(4) Table 3. Use of support in the corpora. Parts of speech Ranking Top list Verb Noun Top 17 in University Corpus Top 7 in City Corpus Top 13 in University Corpus Top 22 in City Corpus be, have, do, please, take, follow, need, work, continue, return, contact, provide, advise, include, make, use, support be, have, work, include, do, continue, support student, advice, coronavirus, university, staff, UCL, health, information, travel, home, UK, Covid-19, support health, city, advice, council, people, public, York, coronavirus, information, England, case, service, government, Covid-19, business, child, NHS, school, contact, community, home, support 15. Council staff have been adjusting and re-focusing their efforts to support our communities over the coming months, particularly to ensure our frontline services can continue to operate to keep the city moving. I would like to thank the fantastic response of our staff during this crisis, and we will continue to do what we can to support residents, businesses and communities across the city. (City corpus) Can is a modal verb employed to express ability, possibility and permission. It is used for making offers and requests. Collocating with the pronoun we, it reflects that authori- ties were able or likely to support and serve the members during the health crisis. Such expressions as everything we can, however we can and all we can are constantly used to demonstrate authorities’ ability and resolve to support and protect community members. In Example 12, we can appears in an adverbial clause and displays authorities’ active role in information service. The second conjunct in the example further reveals that several weeks was marked ‘soon’ and authorities were actually not so capable of providing more information. Example 13 also depicts the use of can in an adverbial for demonstrating authorities’ sincerity and determination involved in helping and supporting members. On the other hand, can which appears in everything we can and what we can may contribute to constructing authorities’ active image in community support, illustrated in Examples 14 and 15. We can has 64 and 21 cases in University Corpus and City Corpus respectively. As Table 4 shows, in total, the subordinated we can accounts for 36.0% and 42.9% respec- tively in the corpora. This means that can was favoured by authorities for expression of attitudes rather than description of actions in response to the pandemic. 16. Please refer to this page for accurate and updated advice about the coronavirus and UCL’s response. We will update this page regularly with more information as it becomes available. (City Corpus) Will enables authorities to demonstrate their determination and capability in respond- ing to the health crisis. It can point to either volition or future prediction, or both. Wei 513 Table 4. Use of we can in subordination. Patterns University Corpus City Corpus Frequency Percentage (%) Frequency Percentage (%) as. . .as we can everything/all we can however we can what we can Total 6 11 3 3 23 9.4 17.2 4.7 4.7 36.0 1 5 3 9 4.8 23.8 14.3 42.9 Normally it is used with a human subject and it has volitional element though it can also be a future marker (Aijmer, 1985). Will mainly collocates with the animate subjects we and you. In Example 16, we will signals a volitional action in the future. Just because of its vagueness, will becomes preferable to other modal verbs. It has 744 occurrences in University Corpus and 118 in City Corpus, collocating with we 148 and 33 times respec- tively. It is followed by be and continue more frequently in both corpora. Next, we turn to the application of tense or aspect to showing support. Semantically, aspect involves the boundedness of events (Radden and Dirven, 2007). Progressive aspect marks an event which is conceptualised as unbounded from ‘inside’ the situation where it is occurring. In addition, progressive aspect contributes to expressing ‘greater immediacy’ and implicating the potential continuation of a situation unless impeded (Radden and Dirven, 2007: 178). In contrast, perfect aspect is used for an event which is bounded from the ‘outside’. Present perfect aspect means that an event has an endpoint which is the conceptualiser’s now. Different from present perfect aspect, present perfect progressive aspect highlights the continuous or interrupted duration of a situation and it has inferential (describing ‘the inferred state following an anterior unbounded event’) and continuative (stressing ‘the durational phase of a situation’) functions (Radden and Dirven, 2007: 217). 17. We will be providing a remote and reduced operating model for accessing services and facilities on campus. (University Corpus) 18. We are working closely with partners about how to best support those at risk at food poverty and who claim free school meals, and further guidance will be issued. (City Corpus) 19. Since the outbreak of the Coronavirus we have been making every effort to safeguard our staff and students around the world. (University Corpus) In Example 17, a declarative sentence adopts the future progressive tense and helps university authorities to display their agency in responding to the pandemic in the future. In Example 18, a declarative sentence in the present progressive tense describes the event where city authorities make continuous efforts in providing support for members. Example 19 uses the present perfect progressive tense to show the durational phase of authorities’ efforts in protecting the community. 514 Discourse & Society 34(4) Table 5. Frequencies of three patterns with we. Patterns University Corpus City Corpus “we” “be” [tag= “V.*”] “we” “will” “be” [tag= “V.*”] “we” “have” “be” [tag= “V.*”] 80 14 14 41 6 13 No matter in what aspects, progressive tense effectively reflects authorities’ ongoing attempts to support and serve members when public health is under threat. And fre- quency counts show that we are doing (“we” “be” [tag= “V.*”]) enjoys the most popular- ity among the three patterns, illustrated in Table 5. This suggests that authorities were trying to impress the public as an organisation that was making constant efforts to sup- port members and protect the public health and safety. Two case studies on kindness expression Having explored how kindness is expressed and promoted through discourse in commu- nity, we need to check if there is any discursive evidence for expression of and deviation from kindness with respect to specific themes. Wearing masks With regard to wearing masks in the pandemic, opinions and practices may vary from member to member in community. In public statements they is constantly used as the synonym of out-groups and negative attitudes are attached to it. In Example 20, face- mask group has been marked as out-group by means of the pronoun their. The initiative was intended to advocate respect and solidarity within community, but it negated the protective practice which is conventional and acceptable in Asian communities by using the evidential structure there is little evidence that. . .. Consequently, the supportive and kind proposal fails to meet the psychological needs of face-mask group, let alone culti- vating their sense of community. 20. Should I wear a protective face mask? Advice from the NHS and Public Health England states that following hygiene precautions such as thoroughly washing hands with soap and water, covering your mouth and nose with a tissue if you cough or sneeze, and keeping surfaces clean, are the best ways to avoid catching or spreading the virus. There is little evidence that masks are effective in preventing the spread of the virus. However, we know that some people may choose to wear face masks to protect themselves and others from possible infection. It is important that we recognise their decision and we encourage our community to be as supportive of each other as possible. (University Corpus) Wei 515 In addition, Example 20 exploits the sentence pattern it is important that. . . for deliv- ering the authoritative or persuasive message in community. This pattern, together with its variant it is important to do. . ., is often adopted for formal suggestion or persuasion because it enables authorities to distance themselves from the current situation so as to make a suggestion or guidance as objectively as possible. 21. Should I wear a mask? Some of our community choose to wear face masks for cultural, social and personal reasons. The wearing of a face mask is not necessarily a sign that the wearer is ill with a cold, flu or any other virus. Public Health England do not stress the need to wear a face mask but instead issue this advice to prevent sharing germs. As such, the University will not be asking anyone to wear masks or supplying them, but please remember that some people choose to. (University Corpus) Conversely, Example 21 can be taken as a better choice for authorities to resolve the dispute of wearing masks in communities. By using the phrase some of our community, authorities have skilfully included face-mask group in community and, to some degree, successfully nurtured their sense of community. Then adopting the nominalised subject (the wearing of a face mask) and the neutral signifier (the wearer), authorities were try- ing to eliminate misunderstanding about mask-wearing in community. By quoting Public Health England for the use of face masks, authorities eventually displayed their mediat- ing stance: mask wearing is not an obligation but a personal choice and it should be respected. What is worth mentioning here is that when Covid-19 situation became serious in the UK, more communities started to use I instead of you to refer to the audience (mainly students and staff) so as to give the contextually dependent guidance directly and effec- tively. As can be seen in Examples 20 and 21, the pronoun I signifies a perspective change in providing specific guidance and moral support. It subtly converts a formal dialogue to an intimate soliloquy so as to facilitate the information delivery and promote kindness expression. Returning home Regarding public guidance or advice, authorities tended to adopt passive voice to indi- cate its formality and objectivity, as shown in Example 22. 22. I want to go home because of the coronavirus outbreak. What is UCL's policy on this? On 17 March, UCL announced that most university buildings will be closed by Friday 20 March and students are strongly advised to return home wherever possible. (University Corpus) However, passivation implies little attention is paid to the emotional needs or mental health of members at this unprecedented time. The advice on returning home cannot 516 Discourse & Society 34(4) gloss over the unequal power relations between university authorities and students. It also effectively suppresses the agentic role of authorities in urging students to leave the campus. In this way, no one should take on the responsibility of publicising the guidance. 23. We strongly recommend that it is in your interest to make your travel plans as soon as practically possible if you do wish to return home. (University Corpus) In public discourse, personal pronoun you contributes to defining the target audience and constructing a dialogic communication model. As shown in Example 23, using the pronouns you and its variant your to refer to students, authorities stated straightforwardly their advice about returning home. Kindness and concern were intended to express by means of the sentence pattern it is in your interest to do. . .. Nevertheless, activation increases the level of imposition and deviates a little from authorities’ kind intention. It seems to be incompatible with the British culture which highlights personal autonomy and negative politeness. However, it is likely to demonstrate an affective stance and ren- der the persuasion emotional and appealing. Discussion and conclusion Kindness is not only a personal virtue, but also a social lubricant and a catalyst for altru- ism. By expressing and promoting kindness through discourse, community authorities intended to cultivate members’ sense of community and maintain their dominant posi- tion. This research explores how kindness was expressed verbally to the vulnerable group of people in community, that is, students from China and East Asian communities, during the health crisis. Detailed discourse analysis has uncovered what kindness was meant and promoted by authorities in British university communities and city communities during the Covid pandemic. It is found that kindness was instantiated by providing support to community members and showing compassion and inclusion to members with different cultural backgrounds. This justifies young people’s interpretations of kind acts as helping and inclusion (Binfet et al., 2022; Cotney and Banerjee, 2019). Kindness, especially adolescent kindness, falls into three categories: random or reac- tionary kindness, intentional kindness and quiet kindness (Binfet and Enns, 2018). Evidently, what community authorities expressed and promoted was intentional kind- ness, which involved ‘planning, gathering resources, identifying recipients, scheduling, and execution’ (Binfet and Enns, 2018: 35). The intentional kindness was oriented towards Chinese students or East Asian communities and aimed at exercising control on community members and maintaining the public safety at this particular time. Carefully choosing adjective modifiers of community as well as the pronouns such as inclusive we and impersonal you, authorities strived to emphasise diversity and inclusion and unite members to show kindness and respect diversity. Comparing city to fabric, authorities intended to picture metaphorically a diverse and inclusive community and advocate community diversity and inclusion. Wei 517 However, we have to differentiate inclusion from diversity. Diversity and inclusion are often loosely conflated, but diversity does not generate inclusion automatically (Johnson, 2011; Sherbin and Rashid, 2017). As the functional partner to diversity, inclu- sion requires the shift from performative diversity (highlighting quotas and statistics as a measure of diversity) to cognitive diversity (emphasising creating space to diverse opinions and perspectives) (Brix et al., 2022: 267). During the pandemic, mask wearing has become a touchstone of community inclusion. In using the pronoun they and its vari- ants their and them, authorities consciously or accidentally excluded some groups (e.g. students from China or other Asian countries) from a community (e.g. university com- munity) and negate the values or beliefs (e.g. wearing masks is protective) of some mem- bers so much so that kindness (i.e. respect and inclusion) couldn’t be expressed or promoted effectively within communities. When discrimination and racism did take place, community authorities, in most times, suppressed their agency to protect the vul- nerable groups. They adopted passivation and topicalisation to subtly distance them- selves from virus-related racial discrimination and take an objective stance in dealing with the unkind and inappropriate behaviours so as to mediate the complex relationships among members in a safe position. Overall, their authoritative stance in dealing with rac- ism and discrimination was featured by reminders or warnings other than action plans. It is true that this may be interpreted as a respectful speech act which corresponds to the no-imposition British culture, but this fails to satisfy the emotional and safety needs of the target audience. In addition, in the metaphor of fabric, what was highlighted was diversity rather than inclusion because what authorities welcomed was not Asian com- munities, but international communities. Kindness can also be seen in providing support. Using declarative sentences in pro- gressive tense with different aspects, authorities underlined their continuous efforts in supporting members. Additionally, modal verbs such as can and will enable authorities to demonstrate their active role in providing support and expressing kindness. The con- stant use of we can in subordination (e.g. everything we can, as soon as we can) is con- ducive to displaying authorities’ volition and sincerity involved in giving support to people. However, it cannot show what specific or substantial actions authorities were able or likely to take to help the members. More evidence of attitudinal display other than action-oriented narrative can be found in the subordinated request, advice or warning in such sentence patterns as it is important that. . ., it is important to do. . ., we would like to remind that. . . and we want to make clear that. . .. The verbal kindness were intended to give support to the vulnerable group in the community. However, this British style of persuading values kindness and respect to every member, not just the vulnerable groups of people in urgent need of help and support. Therefore, it may be taken as a kind of lip service and impressed as ritualised kindness. Just as what Jones (2021) has said, discourse analysis can ‘provide us with frame- works to notice when and how meaning is created, and sometimes to productively inter- vene’ (Jones, 2021: 3). Based on the two case studies, we can figure out how kindness is likely to be better expressed in discursive practice. It is suggested that we caution against the use of the third-person plural pronoun they and the passive structure in expressing and advocating kindness so as to well protect members’ safety and well-being during the pandemic. 518 Discourse & Society 34(4) This research is a tentative attempt to explore the expression and promotion of kind- ness within British university communities and city communities at the beginning of the pandemic from the perspective of CDA. Kindness expression and promotion has been contextualised and targeted towards some groups of members in the community com- munication during the Covid pandemic. This research has examined a subset of the lin- guistic ways to convey or advocate kindness. The next step could be to survey the attitudes of the target audience to the official emails and public statements and examine the effectiveness of kind expression in community in terms of the beholders and recipients. Acknowledgements I would like to express my heartfelt gratitude to Professor M. Lynne Murphy, Professor Teun A. van Dijk and Dr Roberta Piazza for their careful guidance and timely encouragement. In addition, I am very grateful to the two anonymous reviewers for their insightful and valuable comments on earlier drafts of the article. This work also benefited a lot from discussions with Fatimah Yahya M. Fagehi. Special thanks should go to Kindness UK for supporting me to share a relevant research at ICLASP 17. Declaration of conflicting interests The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article. Funding The author(s) received no financial support for the research, authorship, and/or publication of this article. ORCID iD Jilan Wei https://orcid.org/0000-0001-6474-3012 References Aijmer K (1985) The semantic development of will. In: Fisiak J (ed.) Historical Semantics- Historical-Word-Formation. Amsterdam, Netherlands: Mouton, pp.11–21. Algoe SB and Haidt J (2009) Witnessing excellence in action: The ‘other-praising’ emotions of elevation, gratitude, and admiration. The Journal of Positive Psychology 4(2): 105–127. Anderson S and Brownlie J (2019) Public Policy and the Infrastructure of Kindness in Scotland. Dunfermline: Carnegie UK Trust. Bartlett MY and DeSteno D (2006) Gratitude: Helping when it really costs you. Psychological Science 17(4): 319–325. Binfet JT and Enns C (2018) Quiet kindness in school: Socially and emotionally sophisticated kindness flying beneath the radar of parents and educators. Journal of Childhood Studies 42(2): 31–45. Binfet JT, Willis-Stewart S, Lauze A, et al. (2022) Understanding university students’ conceptual- izations and perceptions of kindness: A mixed methods study. Journal of Further and Higher Education 46(4): 441–460. Brix KA, Lee OA and Stalla ASG (2022) Understanding inclusion. BioScience 72(3): 267–275. Wei 519 Cotney JL and Banerjee R (2019) Adolescents’ conceptualisations of kindness and its links with well-being: A focus group study. Journal of Social and Personal Relationships 36(2): 599– 617. Daniel M (2005) Understanding inclusives. In: Filimonova E (ed.) Clusivity: Typology and Case Studies of Inclusive-Exclusive Distinction. Amsterdam, Netherlands; Philadelphia, PA: John Benjamins Publishing Company, pp.3–48. Dixon RMW (1999) Adjectives. In: Brown K and Miller T (eds) Concise Encyclopedia of Grammatical Categories. Amsterdam, Netherlands: Elsevier, pp.1–8. Eisenberg N, Fabes R and Spinrad T (2006) Prosocial development. In: Damon W and Lerner RM (eds) Handbook of Child Psychology Volume 3: Social, Emotional and Personality Development, 6th edn. Hoboken, NJ: John Wiley and Sons, Inc, pp.646–718. Fairclough NL (1985) Critical and descriptive goals in discourse analysis. Journal of Pragmatics 9(6): 739–763. Faust HS (2009) Kindness, not compassion, in healthcare. Cambridge Quarterly of Healthcare Ethics 18(3): 287–299. Fowler JH and Christakis NA (2010) Cooperative behavior cascades in human social networks. Proceedings of the National Academy of Sciences 107(2): 5334–5338. Hunter A (2018) Conceptualising community. In: Cnaan RA and Milofsky C (eds) Handbook of Community Movements and Local Organizations in the 21st Century. Cham, Switzerland: Springer International, pp.3–23. Johnson DR (2011) Women of color in science, technology, engineering, and mathematics (STEM). New Directions for Institutional Research 2011(152): 75–85. Jones RH (ed.) (2021) Viral Discourse. Elements in Applied Linguistics. Cambridge: Cambridge University Press. Kitagawa C and Lehrer A (1990) Impersonal use of personal pronouns. Journal of Pragmatics 14(5): 739–759. Lang E (1984) The Semantics of Coordination. Amsterdam, Netherlands: John Benjamins. Lim MA, Li H, Yu J, et al. (2022) COVID 19 incidents against Chinese and East Asian students in the UK: Safety security and communication. HERE@Manchester. Looijen RC and Van Andel J (1999) Ecological communities: Conceptual problems and defini- tions. Perspectives in Plant Ecology, Evolution and Systematics 2(2): 210–222. McCarthy M (1991) Discourse Analysis for Language Teachers. Cambridge: Cambridge University Press. Mannarini T and Fedi A (2009) Multiple senses of community: The experience and meaning of community. Journal of Community Psychology 37(2): 211–227. Peterson C and Seligman MEP (2004) Character Strengths and Virtues: A Handbook and Classification. Oxford: Oxford University Press. Pullum GK and Huddleston R (2002) The Cambridge Grammar of the English Language. Cambridge: Cambridge University Press. Radden G and Dirven R (2007) Cognitive English Grammar. Amsterdam, Netherlands: John Benjamins. Sherbin L and Rashid R (2017) Diversity doesn’t stick without inclusion. Harvard Business Review, 1 February 2017. Available at: https://hbr.org/2017/02/diversity-doesnt-stick-with- out-inclusion (accessed 10 October 2022). Spelman E (2001) Fruits of Sorrow: Framing our Attention to Suffering. Boston, MA: Beacon Press. Tsvetkova M and Macy MW (2014) The social contagion of generosity. PLoS One 9(2): e87275. Van Dijk TA (1993) Principles of critical discourse analysis. Discourse and Society 4(2): 249–283. 520 Discourse & Society 34(4) Van Dijk TA (1995) Discourse analysis as ideology analysis. In: Schaffne C and Wenden AL (eds) Language and Pace. London: Routledge, pp. 17–33. Van Dijk TA (2011) Discourse and ideology. In: Van Dijk TA (ed.) Discourse Studies: A Multidisciplinary Introduction. London: SAGE, pp. 379–408. Van Dijk TA (2012) Ideology and Discourse: A Multidisciplinary Introduction. Available at: https://discourses.org/wp-content/uploads/2022/06/Teun-A.-van-Dijk-2012-Ideology-And- Discourse.pdf (accessed 1 July 2022). Volosinov VN (1973) Marxism and the Philosophy of Language: Studies in Language (trans. L Matejka and IR Titunik). London: Seminar Press Limited. Weale S (2020) Chinese students flee UK after ‘maskaphobia’ triggered racist attacks. The Guardian. Available at: https://www.theguardian.com/education/2020/mar/17/chinese-stu- dents-flee-uk-after-maskaphobia-triggered-racist-attacks (accessed 25 April 2020). Webber M (1964) Urban place and the nonplace realm. In: Webber MM, Dyckman JW, Fole y DL, et al. (eds) Explorations Into Urban Structure. Philadelphia, PA: University of Pennsylvania Press, pp.79–153. Wortham SEF (1996) Mapping participant deictics: A technique for discovering speakers’ footing. Journal of Pragmatics 25(3): 331–348. Author biography Jilan Wei is a lecturer of Linguistics at Changzhou Institute of Technology, China. Now she is studying for her PhD at the University of Sussex, UK. Her PhD thesis is concerned with a corpus- based analysis of evidentiality and emotionality in UK government daily briefings during the Covid pandemic.
10.1016_j.molcel.2023.02.026
Article Stepwise activities of mSWI/SNF family chromatin remodeling complexes direct T cell activation and exhaustion Graphical abstract Authors Elena Battistello, Kimberlee A. Hixon, Dawn E. Comstock, ..., Fabiana Perna, Iannis Aifantis, Cigall Kadoch Correspondence ioannis.aifantis@nyulangone.org (I.A.), cigall_kadoch@dfci.harvard.edu (C.K.) In brief mSWI/SNF ATP-dependent chromatin remodeling complexes direct chromatin accessibility and gene expression during T cell activation and exhaustion. Targeting mSWI/SNF complexes with clinically relevant small-molecule inhibitors and degraders attenuates T cell exhaustion and increases T cell persistence and anti-tumor activity in cell culture systems and in vivo. Highlights d mSWI/SNF targeting and activity is specific to T cell activation and exhaustion states d Genetic and chemical disruption of cBAF complexes enhances T cell persistence d cBAF complex activities facilitate T cell exhaustion and prevent memory features d mSWI/SNF pharmacologic disruption improves CAR-T expansion and anti-tumor control Battistello et al., 2023, Molecular Cell 83, 1216–1236 April 20, 2023 ª 2023 The Author(s). Published by Elsevier Inc. https://doi.org/10.1016/j.molcel.2023.02.026 ll ll OPEN ACCESS Article Stepwise activities of mSWI/SNF family chromatin remodeling complexes direct T cell activation and exhaustion Elena Battistello,1,11 Kimberlee A. Hixon,2,3,4,11 Dawn E. Comstock,2,3,5,11 Clayton K. Collings,2,3 Xufeng Chen,1 Javier Rodriguez Hernaez,1 Soobeom Lee,1 Kasey S. Cervantes,2,3 Madeline M. Hinkley,2,3 Konstantinos Ntatsoulis,1 Annamaria Cesarano,6 Kathryn Hockemeyer,1 W. Nicholas Haining,2 Matthew T. Witkowski,7 Jun Qi,8 Aristotelis Tsirigos,1,9 Fabiana Perna,6 Iannis Aifantis,1,12,* and Cigall Kadoch2,3,10,12,13,* 1Department of Pathology and Laura and Isaac Perlmutter Cancer Center, NYU Grossman School of Medicine, New York, NY 10016, USA 2Department of Pediatric Oncology, Dana-Farber Cancer Institute and Harvard Medical School, Boston, MA 02215, USA 3Broad Institute of MIT and Harvard, Cambridge, MA 02142, USA 4Biological and Biomedical Sciences Program, Harvard Medical School, Boston, MA 02115, USA 5Program in Immunology, Harvard Medical School, Boston, MA 02115, USA 6Department of Medicine, Division of Hematology/Oncology, Indiana University School of Medicine, Indianapolis, IN 46202, USA 7Department of Pediatrics-HemeOnc and Bone Marrow Transplantation, University of Colorado Anschutz Medical Campus, Aurora, CO 80045, USA 8Department of Cancer Biology, Dana-Farber Cancer Institute and Harvard Medical School, Cambridge, MA, USA 9Applied Bioinformatics Laboratories, Office of Science & Research, NYU Grossman School of Medicine, New York, NY, USA 10Howard Hughes Medical Institute, Chevy Chase, MD, USA 11These authors contributed equally 12Senior author 13Lead contact *Correspondence: ioannis.aifantis@nyulangone.org (I.A.), cigall_kadoch@dfci.harvard.edu (C.K.) https://doi.org/10.1016/j.molcel.2023.02.026 SUMMARY Highly coordinated changes in gene expression underlie T cell activation and exhaustion. However, the mechanisms by which such programs are regulated and how these may be targeted for therapeutic benefit remain poorly understood. Here, we comprehensively profile the genomic occupancy of mSWI/SNF chro- matin remodeling complexes throughout acute and chronic T cell stimulation, finding that stepwise changes in localization over transcription factor binding sites direct site-specific chromatin accessibility and gene activation leading to distinct phenotypes. Notably, perturbation of mSWI/SNF complexes using genetic and clinically relevant chemical strategies enhances the persistence of T cells with attenuated exhaustion hallmarks and increased memory features in vitro and in vivo. Finally, pharmacologic mSWI/SNF inhibition improves CAR-T expansion and results in improved anti-tumor control in vivo. These findings reveal the central role of mSWI/SNF complexes in the coordination of T cell activation and exhaustion and nominate small-molecule-based strategies for the improvement of current immunotherapy protocols. INTRODUCTION T cells undergo dynamic morphologic and gene regulatory changes upon acute or sustained antigen exposure.1–5 Impor- tantly, chronic antigen stimulation causes T cells to enter a dysfunctional state known as T cell exhaustion characterized by poor effector function, reduced proliferative capacity, expres- sion of inhibitory receptors, and altered cytokine production.6 Targeting T cell exhaustion has formed the basis for numerous studies in the context of both chimeric antigen receptor (CAR)-T cell generation and checkpoint blockade.7–13 However, the molecular mechanisms governing T cell activation and exhaustion as well as the factors directing the expression of key state-specific biomarkers remain poorly understood. Uncov- ering such mechanisms may enable improvements to current immunotherapeutic approaches. Studies in both mouse and human settings have defined chro- matin changes during T cell activation and exhaustion, including linking locus-specific accessibility changes with anti-tumor re- sponses.14–22 Species-level differences have also been identi- fied owing to the fact that a number of gene regulatory networks differ between mouse and human cells, highlighting the need for increased understanding of human T cells.18 Studies aiming to identify strategies to prevent or reverse exhaustion and to define 1216 Molecular Cell 83, 1216–1236, April 20, 2023 ª 2023 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Article A CD3/CD28 Dynabeads CHRONIC STIMULATION B 0h 3h 24h 48h 72h D6 D9-Chr ACTIVATION TRANSIENT STIMULATION D9-Tr -mSWI/SNF complex occupancy (SMARCA4, SS18, ARID1A, PBRM1 C&T) -Histone mark occupancy (H3K27Ac C&T) -Chromatin accessibility (ATAC-seq) -Gene expression (RNA-seq) -Immunoprofiling (FACS) and proliferation SMARCA4 SS18 % 7 0 . 2 1 : 2 C P 6 . 0 4 . 0 2 . 0 0 . 0 2 . 0 − 4 . 0 − % 8 . 4 1 : 2 C P 6 . 0 4 . 0 2 . 0 0 . 0 2 . 0 − 4 . 0 − 6 . 0 − Human CD8+ T cells % 3 8 . 1 1 : 2 C P 4 . 0 2 . 0 0 . 0 2 . 0 − 4 . 0 − C D ll OPEN ACCESS 3h 0 98 0 3 10 2 4 10 D6 39 12 24 5 10 4 10 3 10 0 3 -10 5 10 4 10 3 10 0 3 -10 24h 5 25 69 48h 10 17 72 5 10 4 10 3 10 0 3 -10 0 3 10 4 10 0 3 10 4 10 D9-Ch 49 9 25 D9-Tr 0 96 1 5 10 4 10 3 10 0 3 -10 5 10 4 10 3 10 0 3 -10 5 10 4 10 3 10 0 3 -10 EXH ACT NAIVE/ MEMORY 3 M T I PD1 3 M T I 3 M T I 5 10 4 10 3 10 0 3 -10 5 10 4 10 3 10 0 3 -10 0h 0 98 0 3 10 2 4 10 PD1 72h 32 18 47 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 PD1 H3K27ac ATAC-seq % 8 9 . 8 1 : 2 C P 4 . 0 2 . 0 0 . 0 2 . 0 − 4 . 0 − 6 . 0 − 0h 3h 24h 48h 72h D6 D9-Ch D9-Tr Donor1 Donor2 −0.4 −0.2 0.0 0.2 PC1: 22.24% 0.4 −0.4 −0.2 0.0 0.2 0.4 0.6 −0.6 −0.4 −0.2 0.0 0.2 0.4 0.6 PC1: 22.68% PC1: 24.08% −0.4 −0.2 0.0 0.2 0.4 0.6 PC1: 37.98% SMARCA4 Z-Score SS18 Z-Score H3K27Ac Z-Score ATAC-seq Z-Score E 2 1 0 -1 -2 2 1 0 -1 -2 2 1 0 -1 -2 2 1 0 -1 -2 Cluster 1 (n=11167) Pan-state targeted/-accessible Cluster 2 (n=4439) Early activation Cluster 3 (n=2508) Activation/Exhaustion Cluster 4 (n=6998) Activation/Exhaustion Cluster 5 (n=4894) Late activation Cluster 6 (n=2759) Exhaustion Cluster 7 (n=3085) Naive/memory (strongest in D9-Tr) Cluster 8 (n=10163) Naive/memory (reduced upon activation) Cluster 9 (n=27495) Naive/memory (lower accessibility) 0 h 3 h 2 4 h 4 8 h 7 2 h D 6 D 9 - C h D 9 - T r 0 h 3 h 2 4 h 4 8 h 7 2 h D 6 D 9 - C h D 9 - T r 0 h 3 h 2 4 h 4 8 h 7 2 h D 6 D 9 - C h D 9 - T r 0 h 3 h 2 4 h 4 8 h 7 2 h D 6 D 9 - C h D 9 - T r Time Point Activation (Merged 24h, 48h, 72h) 35415 1962 16755 763 601 321 5712 Exhaustion (Merged D6, D9-Ch) 28752 139 498 6658 8420 432 267 Memory (Day9-Tr) 33568 179 292 3780 1797 171 78 ATAC-Seq Merged Peaks H3K27ac Merged Peaks SMARCA4 & SS18 Merged Peaks Figure 1. Stepwise changes in mSWI/SNF complex targeting and chromatin accessibility during CD8+ T cell activation and exhaustion (A) Schematic for CD3/CD28 bead-based stimulation of human CD8+ T cells. (B) FACS-based profiling of PD1 and TIM3 markers indicating naive/memory, activated, and exhausted populations. (C) PCA for mSWI/SNF subunit and H3K27Ac CUT&Tag and ATAC-seq profiles across the time course (donors 1 and 2, see KRT for donor information). (legend continued on next page) Molecular Cell 83, 1216–1236, April 20, 2023 1217 ll OPEN ACCESS which T cell populations (i.e., terminally versus progenitor exhausted) should be targeted represent active areas of investigation.23–25 Several critical transcription factors (TFs) that contribute to distinct stages of T cell activation and exhaustion have been iden- tified,3,24 including NFAT, NF-kB, AP-1, GATA3, and t-BET, as mediators of activation, and elevated T cell receptor (TCR)- responsive TFs such as TOX, NFATC1, IRF4, BATF, MYB, and others in exhausted states.6,25–33 Efforts to target such factors directly are challenged by high-affinity TF-DNA interactions and functional redundancy between multiple TFs, making the inhibi- tion or depletion of a single TF often insufficient. Networks that encompass multiple TFs have been identified, yet strategies to prioritize those that play ‘‘master regulatory’’ roles and that are secondary remain challenging. With few exceptions, the role of chromatin regulatory complexes and epigenetic modifiers in T cell dysfunction remains less clear.8,15,34 Recently, several CRISPR screens in mouse T cells revealed chromatin regulatory proteins and protein complexes as mediators of exhaustion pro- grams.35,36 In particular, genome-wide CRISPR screens have identified components of the mammalian SWI/SNF (mSWI/SNF) family of ATP-dependent chromatin remodeling complexes as mediators of regulatory T cell and exhausted states.37,38 Our group and others have shown that a wide range of human TFs interacts transiently with mSWI/SNF complexes resulting in their site-specific targeting genome wide.39–42 mSWI/SNF com- plexes are heterogeneous, multi-subunit entities that alter DNA- nucleosome contacts, generating chromatin accessibility and coordinating the timely binding of transcriptional machinery required for gene expression.43–47 mSWI/SNF complexes exist in three forms, termed canonical BAF (cBAF), polybromo-associated BAF (PBAF), and non-canonical BAF (ncBAF), each demarcated by the incorporation of distinct subunits and unique association with chromatin features.43,48,49The genes encoding the 29 total subunits are collectively mutated in over 20% of human cancers, in some cases representing hallmark drivers44,50 and have been implicated in both cellular differentiation and cell-state changes. Although the importance of mSWI/SNF complexes as regulators of chromatin accessibility in tumor-intrinsic settings represents an active area of investigation, their role in governing immune cell function remains less clear.51–59 Understanding the role for mSWI/SNF-directed chromatin remodeling across T cell states has uniquely high significance, given that such cells often need to undergo dynamic changes to carry out specific functions and to orchestrate anti-tumor immune responses. RESULTS Stepwise rewiring of mSWI/SNF complex occupancy and chromatin accessibility during acute and chronic human T cell stimulation To study chromatin-level changes across T cell states, we devel- oped a system for antigen-independent TCR stimulation of Article human CD8+ T cells (Figure 1A). T cells were profiled without stimulation (0 h) or at 3, 24, 48, and 72 h (3 days) following stim- ulation to capture different stages of early T cell activation. To profile changes related to chronic antigen stimulation, cells were replated in presence of new beads on days 3, 6, and 9 following initial activation (‘‘chronic,’’ [Ch] condition). In addition, a ‘‘transient’’ (Tr) stimulation condition in which beads were removed after the initial 3-day incubation was used to mimic a memory phenotype. We monitored T cell proliferation, immuno- phenotype, and functionality across these stages (Figures 1A, S1A, and S1B). Stimulated T cells displayed sustained prolifera- tion during the first 6 days after activation and, as expected, lost proliferative capability upon transient stimulation. Chronically stimulated T cells also lost proliferative capacity by day 9, indi- cating the acquisition of an exhausted-like state (Figure S1A). Fluorescence-activated cell sorting (FACS) analyses performed at the early activation (3 and 24 h), activation (48 and 72 h), exhaustion-like (Day6-Day9-Ch), or memory (Day9-Tr) stages re- vealed the acquisition of markers consistent with these pheno- types (Figure S1B), such as the CD25 activation marker at 24 h, sustained the upregulation of the PD1 immune checkpoint and elevated terminal exhaustion markers, TIM-3 and CD39, at days 6 and 9 of the chronic but not of the transiently stimulated condition (Figures 1B and S1B). Chronically activated T cells ex- pressed low levels of CCR7 and CD45RA and displayed reduced IFNg, TNFa, and GZMb staining upon re-stimulation, indicating dysfunction (Figure S1B). In addition, we established a similar system in mouse CD8+ T cells, with highly consistent prolifera- tion and immunophenotyping results (Figures S1C and S1D). Owing to the long-held concept that T cells undergo major rearrangements in their nuclear architecture and chromatin acces- sibility during activation and exhaustion,60 we next sought to char- acterize the chromatin occupancy of the mSWI/SNF complexes, which are known to play directive roles in generating DNA acces- sibility.56,58,60–62 These complexes have been implicated in various T cell contexts, from dysfunction to the generation of T regulatory cells; however, comprehensive profiling of their genome-wide oc- cupancy and activities across T cell populations has not been examined.35,37,60,63 To achieve this, we performed cleavage under targets and tagmentation (CUT&Tag) experiments using anti- bodies targeting the ATPase subunit, SMARCA4, which is present in all three subcomplexes of the mSWI/SNF family (cBAF, PBAF, and ncBAF).43,64,65 Further, we profiled SS18 (a member of cBAF and ncBAF), as well as ARID1A (cBAF-specific) and PBRM1 (PBAF-specific), in parallel with H3K27Ac, a marker of active chro- matin. We observed striking increases in total mSWI/SNF and his- tone peak numbers at activation, consistently in T cells from two independent donors (Figures S2A–S2C). Principal-component analyses (PCAs) revealed concordant time point- and cell state- specific changes in mSWI/SNF as well as H3K27Ac occupancy from both human donors across the course (Figures 1C and S2D). In parallel, we characterized chromatin accessibility using assay for transposase-accessible chromatin using sequencing (D) K-means clustering for SMARCA4, SS18, H3K37Ac, and ATAC-seq performed over merged SMARCA4, SS18, H3K27ac, and ATAC-seq peaks; heatmap intensity depicts quantile-normalized log2-transformed RPKM values transformed into Z scores. (E) Venn diagrams showing the overlap between SMARCA4/SS18 merged, H3K27Ac CUT&Tag peaks with ATAC-seq peaks across time points shown. See also Figures S1 and S2. 1218 Molecular Cell 83, 1216–1236, April 20, 2023 Article (ATAC-seq),66,67 which revealed similar changes and directionality upon PCA analysis (Figures 1C, S2E, and S2F). We next merged the mSWI/SNF complex occupancy and accessibility data and performed k-means clustering analyses to reveal changes in complex localization and accessibility across the time course. The combination of SMARCA4-SS18- H3K27Ac-ATAC-seq-merged peaks (quantile-normalized, log2- transformed RPKM values transformed into Z scores) revealed 9 distinct clusters of mSWI/SNF occupancy and DNA accessi- bility (Figures 1D and S2G). Integrating these data with the im- mune profiling results (Figure S1B), we found that early activation was highlighted in cluster 2 (C2), activation and exhaustion in clusters 3 and 4 (C3 and C4), late activation in cluster 5 (C5), exhaustion in cluster 6 (C6), and memory phenotype in clusters 7 and 8 (C7 and C8) (Figures 1D, S1B, and S2G). Cluster 1 (C1) sites contained TSS-proximal targets of varied targeting and accessibility across conditions and cluster 9 (C9) sites encom- passed TSS-distal sites of lower accessibility and with the high- est mSWI/SNF signal prior to and at early stimulation (Figures 1D and S2H). Across all time points, mSWI/SNF complex occu- pancy, H3K27Ac signal, and accessibility overlapped substan- tially genome wide, with mSWI/SNF complex-bound sites representing a fraction of the total accessible sites (Figures 1E and S2I). Further, mSWI/SNF-bound sites represented increas- ingly lower fractions of total accessible sites across activation to exhaustion, suggesting an increasingly specific group of sites directing the chromatin accessibility and gene regulatory pro- grams hallmark to these states (Figure 1E). Finally, we obtained similar results in mouse settings, albeit with fewer differences between Day9-Ch and Day9-Tr conditions (Figure S2J). Taken together, these studies establish mSWI/SNF complex binding and chromatin accessibility profiles throughout T cell activation and exhaustion, enabling the dissection of their roles in medi- ating state-specific transcriptional networks. Differential, state-specific targeting and activity of mSWI/SNF complexes over TF binding sites during T cell activation and exhaustion We next performed motif analyses using HOMER and archetype- based calling across the 9 clusters of mSWI/SNF-bound, acces- sible sites (Figures 2A, S3A, and S3B). Similarly, locus overlap analysis (LOLA) performed across the activation-exhaustion time course revealed state-specific TF binding enrichment over mSWI/SNF-occupied and accessible sites (Figure 2B). Sites cor- responding to early activation were enriched in AP-1 (JUN/FOS), BATF, NF-kB, and NFAT motifs, corresponding to TFs known to play critical roles in early T cell activation (Figures 2A, 2B, S3A, and S3B).68–71 TF motifs enriched in sites at the middle to late activation stages included those for ATF3, BCL6, CREB1, JUND, and TBX1 (Figures 2A and 2B). Motifs in the exhaus- tion-associated C6 (Day 6 and Day9-Ch time points) included those for MYB/MYBL1, TCF7 (TCF1), which have been impli- cated in T cell stemness and/or exhaustion, as well as CUX2, POU5F1, and SOX3/10 TFs, which are less well characterized in T cells but have been suggested to interact with mSWI/SNF complexes (Figures 2A, 2B, and S3B).6,16,72–77 Intriguingly, we identified a particularly significant enrichment of mSWI/SNF complex occupancy over motifs corresponding to hepatocyte ll OPEN ACCESS nuclear factor-1-beta (HNF1B) not previously implicated in T cell biology (Figures 2A, 2B, S3A, and S3B). Further, TF motifs identified under mSWI/SNF-occupied sites specifically at memory-like/naive clusters (C7/C8) included ETS factors (ETS1, ERG), RUNX1/2, GATA3, STAT factors, and others (Figures 2A and 2B). Several of these factors have been reported by our group and others to interact with mSWI/SNF complexes in other contexts.39,40,78,79 Finally, motifs enriched at mSWI/SNF target sites prior to and at early stimulation, as well as at Day 6, Day9-Tr, and Day9-Ch (C9) time points were nearly exclusively those for CTCF, targeted uniquely by ncBAF (Figures 2A and S3B).49,51 To define putative direct mSWI/SNF genomic targets as well as downstream accessible sites, we next used ATAC-seq data in isolation to identify TF motifs under mSWI/SNF-bound and -unbound accessible regions (Figures 2C and S3C). In addition to the strong enrichment of TF motifs shared with those identified at mSWI-SNF-bound sites (Figure 2C, bold), we also identified enrichment of TF motifs at sites of gained accessibility lacking mSWI/SNF occupancy, including ZNF and DMRT in early activa- tion, and AIRE, MEF2, HLTF, and ZIM3 in exhaustion (Figure 2C). These data suggest that specific sites are made accessible following initial mSWI/SNF targeting, which in turn generates secondarily accessible regions that amplify state-specific programs. We next integrated gene expression profiling by RNA sequencing (RNA-seq) with mSWI/SNF complex binding and accessibility. PCA analyses of RNA-seq data indicated stepwise changes in expression profiles, with greatest shifts between 0 h/3 h and 24 h early activation (PC1) and exhaustion (PC2) (Fig- ures 2D, 2E, and S3D). Given that TFs are considered as the main directive factors governing transcriptional programs in T cells, we first focused on the impact of mSWI/SNF complex occu- pancy and accessibility generation over the expression of TF genes themselves. By integrating gene expression with frac- tional mSWI/SNF complex occupancy (i.e., the enrichment of mSWI/SNF binding over a given TF motif genome wide, relative to others), we identified AP-1 TF genes FOS, FOSB, FOSL1, ATF3, and BATF3 as upregulated mSWI/SNF targets during early activation and with reduced gene activation but retained fractional mSWI/SNF occupancy at late activation time points (Figures 2F and S3E). Interestingly, in the exhausted state, we identified uniquely high mSWI/SNF occupancy at HNF1B motifs and significant changes in expression of HNF1B (>3 log2FC). In the memory-like state, we identified more subtle changes in TF gene expression, of RUNX1, STAT4, STAT6, TCF7L1, and IRF1 genes coupled with slight changes in mSWI/SNF fractional enrichment (Figures 2F and S3F). Lastly, the visualization of mSWI/SNF complexes (SMARCA4, SS18), H3K27ac, and ATAC-seq signal over key genes for activation (IFNG) and early and late exhaustion (CXCL13 and ENTPD1, encoding CD39) confirmed concomitant regulatory region and/or promoter mSWI/SNF binding and chromatin opening (Figure 2G). BAF complexes are bound and active over HNF1B TF binding sites genome wide in exhausted T cells Notably, examining the top 10% differentially upregulated genes late activation, across activation, intermediate activation, Molecular Cell 83, 1216–1236, April 20, 2023 1219 ll OPEN ACCESS A TF motif enrichment B A C H 2 B A T F B A T F 3 C E B P A F O S F O S L 1 F O S L 2 I I R F 7 R F 9 J U N J D P 2 J U N B L I N 5 4 M A F F N F A T 5 M E F 2 C N F A T C 1 P O U 2 F 2 S T A T 1 / 2 T E A D 1 I R F 4 N F K B 2 R E L B c 6 l R E L A A T F 3 J U N D T B X 1 C R E B 1 T E A D 4 B A R X 1 I A R D 3 A C U X 2 F O X J 3 F O X P 1 G a t a 1 H N F 1 A H N F 1 B I I R F 1 R F 5 M E F 2 B M Y B M Y B L 1 N R 4 A 2 S O X 1 0 P O U 5 F 1 S o x 3 T C F 7 B C 1 1 A T C F 7 L 1 E L F 1 E R G E T S 1 E T V 1 F O X O 1 F O X O 3 G A T A 3 K L F 4 L E F 1 O L I G 2 P 5 3 R U N X 1 R U N X 2 S P I 1 S P B I S T A 5 B S T A T 4 S T A T 6 T B X 2 1 C T C F C1 C2 C3 C4 C5 C6 C7 C8 C9 1 0.5 0 −0.5 −1 Fractional enrichment l ) e u a v − P j d a ( 0 1 g o L − f n I 0 0 3 0 0 2 0 0 1 0 Article B TF enrichment (LOLA) Early/intermediate activation Late activation Exhaustion Naive/memory C2 C3 C4 C5 C6 C7 C8 C9 Cluster 24h vs 0h C s D I f i t o M F T AP1.1 BATF NFAT.1 NFAT.2 BCL6.1 NFAT.3 POU.1 HD.3 NFKB.2 ZBTB14 NFKB.3 SOX.3 NFKB.1 IRF.1 IRF.3 NFAT.4 ZNF652 DMRT3 MEF2 LIN54 ZNF384.1 HD.14 CREB.ATF.2 CREB.ATF.3 NR.6 SPI ZNF436 FOX.5 ZBTB7A MYB.2 HD.9 FOX.2 SNAI2 MYB.4 HOMEZ ETS.1 HINFP1.3 HINFP1.2 PAX.hs HNF1 AP1.1 BATF BCL6.1 CREB.ATF.3 ZNF652 NFKB.2 NFAT.1 NFKB.1 NFAT.2 NR.9 CUX.4 CREB.ATF.2 STAT.2 NR.6 IRF.3 DMRT3 NFKB.3 TATA POU.1 HD.13 SOX.3 SOX.8 MEF2 NR.7 REL.halfsite ZNF24 HD.7 BCL6.2 YY1 ARI5A HINFP1.2 E2F.4 HD.9 HOMEZ ETS.1 HSFY2 SPDEF.1 PAX.hs HNF1 MYB.4 Accessibility UP DOWN 72h vs 0h D9-Ch vs 0h HNF1 BATF BCL6.1 ZNF652 IRF.3 HD.13 ZNF232 IRF.1 NFKB.1 TCF.LEF NR.6 HD.25 FOX.2 RUNX.1 NFKB.3 NFAT.2 NR.9 AP1.1 AIRE MEF2 HLTF SOX.4 POU.1 SOX.3 ZIM3 HD.7 CPEB1 TBX.4 ZNF436 HINFP1.3 HSFY2 HD.1 HINFP1.2 HIF GFI SPDEF.1 HIC.1 HD.9 PAX.hs SNAI2 Accessibility UP DOWN IRF.3 STAT.2 IRF.1 RUNX.1 AIRE BCL6.2 EVI1.MECOM TCF.LEF ZNF418 BCL6.1 ZNF28 ETS.1 HD.22 ZNF652 SNAI2 NR.11 ZIM3 NR.5 ZNF136 SPI NFI.1 HNF1 HSFY2 THAP1 NFI.2 NFAT.2 TBX.4 CUX.3 TEAD MFZ1 RBPJ NFKB.1 GMEB2.3 NFAT.4 EGR PAX.hs ZNF384.1 NFKB.3 NFAT.1 AP1.1 Accessibility UP DOWN −0.4 −0.2 0 0.2 0.4 −0.50 −0.25 0 0.25 0.50 Model Coefficient Model Coefficient −0.3 0 Model Coefficient 0.3 D % 7 4 . 4 2 : 2 C P 6 . 0 4 . 0 2 0 . 0 0 . 2 . 0 − 4 . 0 − . 6 0 − D9-Tr vs 0h Accessibility UP DOWN −0.1 0 0.1 Model Coefficient RNA-seq −0.6 −0.4 −0.2 0.0 0.2 0.4 0.6 PC1: 33.42% BATF FOS JUN IRF4 JUND NFATC1 RELA HNF1B BCL6 FOXP3 GATA3 RUNX1 SPI1 STAT5B CTCF 0h 3h 24h 48h 72h D6 D9-Ch D9-Tr Donor1 Donor2 E RNA-seq 3 2 1 0 -1 -2 -3 Z- score: FOS IL1B IL2 IFNG EGR1 IL4 CCL3 XCL1 IL10 IL1A IL13 FOXO6 NFATC4 FOXP3 NTRK1 PPARG MYB STAT1 FGFR1 SOX4 HNF1B CD86 BMP2 NTRK2 CD96 CXCR1 0h 3h 24h 48h 72h D6 D9 -Ch Donor1 Donor2 D9 -Tr F 2 . 1 0 . 1 8 . 0 6 . 0 4 . 0 2 . 0 t n e m h c i r n E f i t o M 3 , 2 r e t s u C l Early activation JUND JDP2 FOSL2 JUNB JUN ATF3 FOSL1 FOSB BATF3 FOS NFATC1 NFATC3 HNF1B NFAT5 POU2F2 IRF7 MAF ATF4 REL MYBL2 IRF4 BCL6B FOXP3 XBP1 TBX21 NFKB1 NFKB2 IRF8 TF Motif Archetype AP1.1/2 BCL6.1 + CREB CCAAT.CEBP Ebox.CATATG FOX.9 HD.4 MAF MIES NR.19 OVOL1 SOX.5 TBX.1 Other 2 . 1 0 . 1 8 . 0 6 . 0 4 . 0 2 . 0 0 . 0 t n e m h c i r n E f i t o M 5 r e t s u C l Late activation FOSL2 JUND JUNB BATF ATF3 JUN BCL6B TBX19 TGIF1 NFIL3 RORA RORB ISL2 SOX18 FOXB1 BARX1 TF Motif Archetype AP1.1 AP1.2 BATF+BCL6.1 CCAAT.CEBP + CPEB1 CREB.ATF.2/3 Ebox.CATATG +OVOL1 FOX.4/5 HD.2/10 IRF.1/3/4+MAF +MYB.3 NFAT.2/4 NFKB.1/2/3 POU.1/2+SOX.3 +TCF.LEF TBX.4+TEAD Other 0 5 3H vs. 0H + 24H vs. 0H + 24H vs. 3H TF Gene LogFC 10 4 3 2 Exhaustion HNF1B BATF3 ARID5A FOXB1 FOXO4 1 MYBL2 MYBL1 SOX5 CUX1 LEF1 0 IRF9 TCF7 IRF7 GATA6 STAT1, STAT2 2 t n e m h c i r n E f i t o M 6 r e t s u C l TF Motif Archetype ARID5A+BATF +CUX.3 CCAAT.CEBP Ebox.CATATG FOX.2/3/4/5/6 GATA HD.2/10 IRF.1/2/3 MEF2+MYB.3 NFAT.2 SOX.4/5 TCF.LEF t n e m h c i r n E f i t o M 7 FOXP2 BHLHE22 r e t s u C l 0 4 D6 vs. 72HR + D9-CH vs. 72H TF Gene LogFC 8 6 0 5 2 . 0 0 2 . 0 5 1 0 . 0 1 0 . 1 4 5 2 3 72h vs. 48h TF Gene LogFC 6 7 Memory RUNX1 STAT6 ZFP28 TP53 STAT4 TCF7L1 IRF1 SOX7 FOXP3 TF Motif Archetype FOX.5 IRF.2 P53−like.3 RUNX.1 SOX.6 STAT.1/2 TCF.LEF ZNF28 Other 0.0 0.5 1.0 1.5 D9-Tr.vs.D9-Ch TF Gene LogFC G 4 A C R A M S 8 1 S S c A 7 2 K 3 H q e s - C A T A 0h 3h 24h 48h 72h D6 D9-Ch D9-Tr 0h 3h 24h 48h 72h D6 D9-Ch D9-Tr 0h 3h 24h 48h 72h D6 D9-Ch D9-Tr 0h 3h 24h 48h 72h D6 D9-Ch D9-Tr chr12:68,538,626-68,562,420 chr4:78,498,312-78,522,161 chr10:97,584,909-97,598,382 IFNG CXCL13 ENTPD1 (CD39) Figure 2. State-specific transcription factor motif enrichment of mSWI/SNF complex occupancy and activity during T cell activation and exhaustion (A) Fractional motif enrichment in clusters C1–C9 (relative to all sites). (B) LOLA enrichment of 15 selected TFs across C2–C9. (C) Differential motif accessibility between time points indicated (top 40 coefficients of logistic regression models). (D) PCA performed on RNA-seq datasets from T cells isolated from 2 independent donors at each time point. (E) Z scored heatmap reflecting the top 25% most variable genes across the time course, partitioned into 8 groups by K-means clustering with selected genes labeled. (F) Plots representing state (cluster(s)-specific) TF fractional motif enrichment (y axis) and gene expression (x axis). (G) Representative SMARCA4, SS18, H3K27ac C&T, and ATAC-seq tracks over the IFNG, CXCL13, and ENTPD1 loci. See also Figure S3. 1220 Molecular Cell 83, 1216–1236, April 20, 2023 ll OPEN ACCESS Naive/Memory Activation/Exhaustion Article A C C6 enrichment in scRNAseq exhaustion signatures Tirosh Zheng Zhang Guo Sade- Feldman 0 5 10 15 20 -Log10 (p value) Early activation Int. activation C2+C3 23% C4 14% Late activation C5 29% Exhaustion C6 40% Memory C7 24% DEGs 3h−24h DEGs 24h−48h DEGs 72h DEGs D6-D9-Ch DEGs D9-Tr E = 2.5 p = 9.9e-37 E = 1.4 p = 5.4e-11 E = 4.3 p = 2e-42 E = 10.5 p = 5.8e-162 E = 5.7 p = 5.2e-67 B 0 1 5 0 ) i M P C g o L ( n o s s e r p x E e n e G Satpathy et al. (BCC) −7.5 −5.0 −2.5 0.0 2.5 UMAP 1 Kourtis et al. (ccRCC) D 5 0 2 P A M U −5 E 10 5 0 2 P A M U −5 Cell type Naive CD8 T Effector CD8 T Early TEx Intermediate TEx Terminal TEx Memory CD8 T Cell type Effector CD8 T Early TEx TEx 5 0 2 P A M U −5 10 5 0 2 P A M U −5 HNF1B motif enrichment 4 3 2 1 0 −7.5 −5.0 −2.5 UMAP 1 0.0 2.5 F HNF1B motif enrichment 4 3 2 1 0 −10 −10 −5 0 UMAP 1 5 −10 −10 −5 0 UMAP 1 5 R 7 L I 7 F C T L L E S 9 6 D C G N F I 2 L I B M Z G 3 G A L N Y A L 1 D C D P 3 1 L C X C Y L N G 1 D P T N E HNF1B D9-Ch 8 0 . M P R m r o N Q 0 . 0 0h 3h 24h 48h 72h D6 D9-Ch D9-Tr 8 3 D C X A G T I 1 0 1 D C 4 A L T C 1 M D R P 2 R C V A H X O T 2 red bold= A mSWI/SNF G T targets I I T G T I C2 C3 C4 C5 C6 C7 C8 C9 G chr10:97,584,909-97,598,382 H Cluster 6 HNF1B Sites 4 A C R A M S 8 1 S S c A 7 2 K 3 H B 1 F N H q e s - C A T A q e s - A N R D9-Ch D9-Tr D9-Ch D9-Tr D9-Ch D9-Tr D9-Ch D9-Tr D9-Ch D9-Tr D9-Ch D9-Tr FOX.5 HD.2 HD.18 FOX.4 SOX.1 ZNF384.2 POU.2 TCF.LEF POU.3 FOX.3 NFAC.2 POU.1 HD.4 SOX.5 IRF.1 HD.8 OCT4+SOX2 HD.16 FOX.6 BATF SOX.2 HD.5 AP1.1 CCAAT.CEBP PRDM1 HD.14 SOX.4 LEF1 ZNF354 SOX.3 0.0 0.6 Motif Density Difference 0.4 0.2 Cluster 6 HNF1B Sites Arid3a ZNF384 PRDM6 Sox5 Hoxd8 Lhx3 PDX1 Zfp652 ZIM3 STAT1 ZFP28 IRF1 FOXJ3 PRDM1 NR2E3 ANDR Foxj3 BATF3 STAT1+STAT2 FOXJ3 IRF1 NR4A1 FOXO4 FEZF1 NR4A2 ENTPD1 0.0 0.3 Motif Density Difference 0.2 0.1 sgCTRL sgHNF1B 75 Kda I K HNF1B Actin −1500 0 1500 J 5 10 4 10 3 10 0 3 -10 3 M T I sgCTRL sgHNF1B 89 0 0 11 4 10 3 10 5 10 4 10 3 10 0 3 -10 76 0 0 24 4 10 3 10 PD1 RNAseq D9-Ch sgHNF1B vs sgCTRL L SASH1 RHOU IL3 l e u a v - p j d a 0 1 g o L - CCL22 AFAPIL2 APOD GLYN CSF2RB TIE1 KLRC2 NR4A3 NR4A1 NR4A2 PTPN14 C6 sites with predicted HNF1B binding (1269 sites) RHO GTPase cycle VEGFA-VEGFR2 signaling pathway Dephosphorylation Regulation of Wnt signaling pathway Signaling by Receptor Tyrosine Kinases Immune system development MAPK signaling pathway Actin cytoskeleton organization Negative regulation of cellular comp org. Protein phosphorylation Positive regulation of cytokine production Neuron projection development Response to lipopolysaccharide MAPK cascade Adipogenesis Metabolism of lipids Negative regulation of phosphorylation Brain development Regulation of cell-cell adhesion Cellular response to insulin stimulus 0 2 4 8 6 -log10(P) 10 12 14 mSWI/SNF-HNF1B-dually bound accessible sites (727 sites) Log2 FC 0 0 2 2 4 4 6 8 -log10(P) 8 6 Cell Cycle, Mitotic DNA metabolic process mRNA metabolic process regulation of DNA metabolic process Diseases of sig.trans. by GF rec, 2nd mess Lymphocyte activation Pathways in cancer Chromatin organization Mitotic cell cycle Retinoblastoma gene in cancer Regulation of chromosome organization G2/M Transition Positive regulation of organelle organization Regulation of hemopoiesis Transcriptional Regulation by TP53 Overlap signal transduction in LMNA laminopathies VEGFA-VEGFR2 signaling pathway DNA-templated transcription Regulation of cell cycle process Regulation of cellular catabolic process 10 12 14 10 12 14 Figure 3. Exhaustion-associated gene expression and chromatin targeting is partially mediated by the HNF1B transcription factor (A) Pie charts representing fractions of the top 10% differentially expressed genes near sites within clusters indicated. (B) Lollipop plots representing the gene expression (LogCPM) of marker genes for the different states throughout the time course with mSWI/SNF-bound genes indicated. (C) Enrichment of C6-associated genes across exhaustion signatures from published scRNA-seq datasets. (D) UMAP projections of 12,643 CD8+ T cells from basal cell carcinoma (BCC) tumor biopsies, clustered by phenotype (left) or colored by HNF1B motif enrichment (right). (E) UMAP projection of 13,613 CD8+ T cells from clear cell renal cell carcinoma (ccRCC) tumor biopsies, clustered by phenotype (left) or colored by HNF1B motif enrichment (right). (F) Enrichment of HNF1B (CUT&Tag performed on Day9-Ch T cells) across clusters. (G) Representative tracks over the ENTPD1 locus. (legend continued on next page) Molecular Cell 83, 1216–1236, April 20, 2023 1221 ll OPEN ACCESS Article exhaustion, and memory states, we found that 23%, 14%, 29%, 40%, and 24% of upregulated gene loci, respectively, were occupied by mSWI/SNF complexes (Figures 3A and S3F). Of note, mSWI/SNF occupancy was present over the greatest per- centage (40%) of loci corresponding to genes upregulated at exhaustion (Figure 3A), suggesting a heightened role for mSWI/ SNF complexes in the establishment and maintenance of the exhaustion transcriptional signature. Next, we monitored expression changes of key hallmark genes of naive, activated, and exhausted T cells (Figure 3B). Genes known to be expressed in naive T cells (IL7R, TCF7, SELL) had the highest expression at the no stimulation (0 h) time point, whereas levels of hallmark activation genes such as IFNG, IL2, PDCD1, CXCL13, GZMB, and LAG3 were most elevated at the 3–72 h time points (Figures 3B, S3D, and S3F). Importantly, key exhaustion hall- mark genes (TOX, ENTPD1, ITGA2, and TIGIT) were upregulated in the exhaustion-like state (Day 6, Day9-Ch) and bound by mSWI/SNF. Additional key exhaustion hallmarks (HAVCR2, PRDM1, CTLA4) were upregulated as secondary (non-mSWI/ SNF bound) target genes (Figure 3B). These studies inform the chromatin landscape and gene regulatory signatures across T cell states, highlighting potential mSWI/SNF-directed as well as downstream changes. Given the uniquely abundant collection of mSWI/SNF target genes among exhaustion hallmarks (Figures 3A and 3B), we next compared the observed changes in gene expression with exhaustion and memory signatures derived from single-cell hu- transcriptomic datasets80–84 man tumor microenvironment (Table S1). Across 5 scRNA-seq datasets evaluated, exhaustion signatures were the most highly enriched over our Day 6 and Day9-Ch time points, whereas memory signatures were en- riched in the unstimulated and Day9-Tr time points (Figure S3G). Remarkably, cluster 6 genes, characterized by mSWI/SNF occu- pancy exhaustion time points and enriched in HNF1B motifs, displayed significant enrichment in all exhaustion signatures, indicating that mSWI/SNF directly regulate a collection of genes hallmarking exhaustion (Figures 3C, S3G, and S3H). Further, we extracted scATAC-seq profiles of intra-tumoral T cells from hu- man basal cell carcinoma (BCC, Satpathy et al.) and clear cell renal cell carcinoma (ccRCC, Kourtis et al.) samples. UMAP pro- jections identified distinct T cell subsets (Figures 3D and 3E) and ChromVar analyses revealed enrichment of the HNF1B motif specifically in exhausted T cell states (Figures 3D, 3E, S3I, and S3J).22,60 Owing to these findings, we next profiled the occupancy of HNF1B using CUT&Tag at Day9-Ch and Day9-Tr time points, which exhibit high and low expression of HNF1B, respectively (Figure S3K). Indeed, HNF1B occupancy from Day9-Ch T cells was most enriched over C6 (exhausted) BAF-bound accessible sites (Figures 3F and S3L), exemplified at the ENTPD1 locus (Figure 3G). Archetype and non-archetype motif enrichment an- alyses performed over HNF1B target sites and over HNF1B sites within C6 revealed the co-enrichment of other exhaustion-asso- ciated TFs such as BATF, NR4A1/2, SOX4, PRDM1, and others (Figures 3H and S3M). In line with this, sgRNA-mediated knockout of HNF1B in human CD8+ T cells resulted in reduced TIM3+PD1+ exhausted cells (Figures 3I and 3J). However, HNF1B KO failed to generate a proliferative advantage (Fig- ure S3N). Further, RNA-seq analyses performed on Day9-Ch wild-type (WT) and HNF1B KO T cells revealed substantial downregulation of exhaustion-associated NR4A1-3 genes and the differential expression of cytokine genes such as GLNY (Fig- ure 3K). Other genes less characterized in T cell biology such as SASH1 and RHOU were strongly downregulated. Finally, the analysis of genes near C6 sites with HNF1B motif density and those containing HNF1B binding exhibiting reduced expression included those involved in metabolism, MAPK signaling, and im- mune system development (Figures 3L and S3O). Lastly, we identified TF motifs with accessibility changes across the time course in mouse CD8+ T cells. Although the ma- jority of TF motifs enriched under sites of mSWI/SNF occupancy were enriched similarly in human and mouse settings, few notable exceptions included the exhaustion state (C6)-enriched HNF1B (homeodomain-containing Hd.10 factors) motifs, which were specific to human cells (Figures S3P and S3Q). Taken together, these data define the targeting specificities of mSWI/ SNF complexes across different T cell states, suggesting their key roles in orchestrating the exhaustion expression signature. Chromatin-focused CRISPR-Cas9 screens identify cBAF components regulators of T cell exhaustion We next aimed to unbiasedly characterize the roles for chromatin regulatory factors in modulating the exhaustion hallmarks of chronically stimulated T cells.85–89 We generated lentiviral con- structs expressing RFP and a custom sgRNA library targeting 310 known epigenetic regulators, containing 6 sgRNAs per gene, non-targeting sgRNAs as negative controls, and sgRNAs targeting PD1 and HAVCR2 (TIM3) as positive controls (total: 1,928 sgRNAs) (Table S2). We then performed a CRISPR-Cas9 screen in mouse CD8+ T cells purified from Rosa26Cas9-EGFP mice (Figure 4A). Cells were activated for 24 h, transduced with the sgRNAs library, ensuring a proper library expression by day 3 of the activation protocol, then chronically stimulated until day 9. At Day9-Ch, sgRNA-RFP+, PD1+TIM3+ T cells were sorted and sgRNA library representation was compared with the initial library representation. Quality control analyses of deep-sequenced sgRNAs libraries confirmed efficient capture of all sgRNAs (97.6%–99% of all sgRNAs) and an even distribu- tion of sgRNA sequences (Gini indexes: 0.05–0.09) (Figure S4A). We then analyzed the genes whose sgRNAs were enriched or (H) Motif enrichment over HNF1B target sites in (top) cluster 6 HNF1B target sites and (bottom) all HNF1B target sites. (I) Western blot for HNF1B and beta-actin in Day9-Ch sgCTRL and sgHNF1B T cells (donor 7). (J) PD1 and TIM3 immunoprofiling on sgCTRL and sgHNF1B T cells at Day9-Ch. (K) Volcano plot depicting differential gene expression (RNA-seq) in sgCTRL and sgHNF1B T cells. (L) Metascape analysis performed over (top) C6 sites with predicted HNF1B binding (>2 motifs) and (bottom) C6 sites with CUT&Tag HNF1B binding, mSWI/SNF occupancy and accessibility. See also Figure S3. 1222 Molecular Cell 83, 1216–1236, April 20, 2023 Article A CD3/CD28 Dynabeads 0h Mouse CD8+ T cells (Cas9-GFP mice) chromatin sgRNA library 307 chromatin regulators 6 sgRNAs/gene Sorting PD1+ TIM3+ 3 M T I - P F R A N R g s 24h 72h D6 D9-Ch Cas9-GFP PD1 CHRONIC STIMULATION sgRNAs representation by deep sequencing ll OPEN ACCESS D 3 M T I 3 M T I 3 M T I TIM3/PD1 immunoprofiling (Day 9) sgSmarca4 sgROSA sgSmarcc1 71 51 62 4 10 3 10 0 3 -10 18 19 16 104 103 0 -10 3 11 18 3 10 4 10 5 10 0 3 10 4 10 5 10 0 3 10 4 10 5 10 4 10 3 10 0 3 -10 PD1 4 0 sgArid1a 50 sgArid1b 67 sgDpf2 55 4 10 3 10 0 3 -10 6 0 4 10 3 10 0 3 -10 17 13 25 3 10 4 10 5 10 0 3 10 4 10 5 10 4 10 3 10 0 3 -10 PD1 4 10 3 10 0 3 -10 PD1 14 31 0 3 10 4 10 5 10 sgArid2 72 5 0 0 9 3 10 4 10 5 10 Ints12 Kat6a Enriched in PD1+ TIM3+ C Smarca4 Smarcc1 Smarcc2 Arid1a Arid1b Dpf2 Arid2 Phf10 Pbrm1 Brd7 Brd9 PAN- mSWI/ SNF cBAF pBAF ncBAF 0 −9 Log2(Fold change) −6 −3 Depleted in PD1+ TIM3+ B 2 0 −2 C F 2 g o L −4 −6 Ezh2 Arid1b Arid2 Ep300 Smarca4 Smarcc1 Dpf2 Pbrm1 Brd2 Brd4 Hdac3 Pdcd1 Kmt2d Havcr2 Arid1a Phf5a 0 100 200 300 Rank 400 300 200 100 E ) 3 y a D f o % ( s l l e c + P F R F ) 3 y a D f o % ( s l l e c + P F R 400 300 200 100 0 Day: CHRONIC OT-1 T cells + B16-OVA ) 3 y a D f o % ( s l l e c + P F R TRANSIENT OT-1 T cells + B16 400 300 200 100 3 5 7 9 11 3 5 7 9 11 3 5 7 9 11 3 5 7 9 11 sgRosa sgSmarca4 sgSmarcc1 sgDpf2 0 Day: 3 5 7 9 11 3 5 7 9 11 3 5 7 9 11 3 5 7 9 11 sgRosa sgSmarca4 sgSmarcc1 sgDpf2 0 Day: 3 6 913 3 6 913 3 6 913 3 6 913 3 6 913 3 6 913 3 6 913 3 6 913 3 6 913 3 6 913 sgRosa sgSmarca4 sgSmarcc1 sgArid1a-1 sgArid1a-2 sgArid1b-1 sgArid1b-2 sgDpf2 sgArid2-1 sgArid2-2 Pan mSWI/SNF cBAF PBAF Pan mSWI/SNF cBAF Pan mSWI/SNF cBAF G 5 10 4 10 3 10 0 3 -10 3 M T I Human CD8+ T cells sgCTRL sgSMARCA4 H 89 0 0 11 3 10 4 10 5 10 4 10 3 10 0 3 -10 3 - M T I : : - A - 7 y C C P A - p m o C 71 0 0 28 3 10 4 10 PD1 ) l m / n o i l i m ( r e b m u n l l e C 120 80 40 0 ** ** CTRL sgSMARCA4 ** ** 5 3 8 12 15 Days after electroporation Figure 4. Chromatin-focused CRISPR-Cas9 screens identify cBAF components as regulators of T cell exhaustion (A) Schematic for CD8+ PD1+/TIM3+ T cell screening using a custom sgRNA library of chromatin regulators. (B) Rank plot depicting log2FC scores (average of n = 6 guides) targeting chromatin regulator genes and negative/positive controls. Depleted genes are high- lighted in black; positive controls are highlighted in gray; mSWI/SNF complex genes are highlighted in orange. (C) Log2FC values for n = 6 independent guides in PD1+TIM3+ cells. (D) FACS plots depicting PD1+/TIM3+ T cell populations in control and mSWI/SNF subunit KO conditions. (E) Bar graph depicting RFP+ cells (% cells of day 3) for control, pan-mSWI/SNF, cBAF, and PBAF genes. (F) Bar graph depicting RFP+ cells (% cells of day 3) for chronic (+B16+OVA) and transient (+B16) stimulation of OT-1 T cells in each sgRNA condition. (G) FACS plots depicting PD1+/TIM3+ T cell populations in control and sgSRMARCA4 KO conditions in human CD8+ T cells. Donor 7 was used for this experiment, which was conducted together with the experiment in Figure 3J, thus the same control was used. (H) Bar graph depicting cell proliferation of human sgCTRL or sgSMARCA4 CD8+ T cells. Error bars represent mean ± SD of 3 technical replicates. **p < 0.01. See also Figure S4. Molecular Cell 83, 1216–1236, April 20, 2023 1223 ll OPEN ACCESS depleted in the PD1+TIM3+ population (Figure 4B; Table S3). Interestingly, most of the significant hits (abs log2FC > 1, false discover rate [FDR] < 0.05) were depleted (n = 27), while only 2 hits in total were enriched, highlighting the diverse chromatin- level contributions to T cell dysfunction. In addition to positive controls Pd1 and Havcr2, depleted hits included genes encoding the mSWI/SNF complexes (Arid1a, Dpf2, Smarcc1, Smarca4), genes involved in histone acetylation (Kat5, Kat8, Ep300, Hdac3), methylation (Kmt2d, Prmt5, Ezh2, Kdm1a, Kdm6a), and other processes, highlighting a range of epigenetic mecha- nisms playing potential roles in regulating exhaustion immune checkpoints (Figures 4B and S4B). Intriguingly, genes encoding mSWI/SNF complex subunits were the top-scoring genes depleted in the PD1+/TIM3+ population (Arid1a = rank 2, Dpf2 = rank 6, Smarcc1 = rank 10, and Smarca4 = rank 15), spe- the cBAF complex43 (Figures 4B, 4C, and S4B; cifically, Table S3). Arid1b, the paralog for Arid1a, was moderately depleted but to a lesser extent given its lower expression and hence likely lower stoichiometric abundance in cBAF complexes (Figure S4C). PBAF- and ncBAF-specific subunits such as Arid2, Pbrm1, and Brd9 were also depleted but to a minimal extent. Taken together, these data highlight unbiasedly the role of the mSWI/SNF complexes, specifically, Arid1a- and Dpf2-contain- ing cBAF complexes, as among the most significant determi- nants of exhausted state. To functionally validate top hits, we transduced two indepen- dent sgRNA-RFP plasmids into Cas9-EGFP T cells and evaluated for PD1+ and TIM3+ populations. We observed a reduction in the percentage of PD1+TIM3+ cells upon the depletion of all pan- mSWI/SNF and cBAF-specific components at day 9. By contrast, the depletion of Arid2 had a minimal effect (Figures 4D and S4D). Although no significant proliferation differences were identified before day 9, pan-mSWI/SNF and cBAF subunit knockout cells (but not PBAF or ncBAF KO cells) out-competed WT cells at day 9, indicating increased T cell persistence (Figure 4E). We also monitored a slight increase in the percentage of cells with a central memory (CM) phenotype (CD44+ CD62L+) (Figure S4E) and a decrease in the killing capacity in the B16-OVA/OT-1 sys- tem at increasing target-effector (T-E) ratios, confirming a mem- ory-like phenotype (Figure S4F). Finally, to confirm these findings, we performed a second CRISPR screen in mouse CD8+ T cells in which we sorted for the top and bottom 15% TIM-3-expressing cells (Figures S4G and S4H), which revealed similar results (Figures S4I–S4K; Table S4). We next sought to validate these findings using an indepen- dent system, based on the co-culture of mouse OT-1 CD8+ T cells with B16 cells expressing the model antigen OVA (chronic stimulation) or B16 WT cells as control (transient stimulation) (Figure S4L). Splenocytes from Rosa26Cas9-EGFP-OT-1 mice were incubated with the MHC-I-specific OVA peptide SIINFEKL for 48 h to achieve T cell activation, then CD8+ T cells were purified, infected with sgRNAs, and co-cultured with B16 or B16-OVA cells (Figure S4L). Here again, we observed a decrease in T cell proliferation at day 9 of stimulation (Fig- ure S4M), upregulated PD1 and TIM3, and an effector phenotype (CD44+ CD62L(cid:1)), indicative of exhaustion-like features (Fig- ure S4N). Consistent with our previous findings, the depletion of Smarca4, Smarcc1, and Dpf2 led to a decreased percentage 1224 Molecular Cell 83, 1216–1236, April 20, 2023 Article of PD1+TIM3+ cells and an increase in the CM population (Fig- ure S4O). Also in this model, we detected an increased persis- tence of chronically, but not transiently, stimulated T cells upon the sgRNA-mediated depletion of Smarcc1, Smarca4, and Dpf2 (Figure 4F). Finally, to validate these findings in human CD8+ T cells, we electroporated the Cas9 ribonucleoprotein and either control (CTRL) or SMARCA4-targeting sgRNAs in hu- man CD8+ T cells. In the setting of SMARCA4 KO (Figure S4P), we observed a decrease in the percentage of PD1+ TIM3+ T cells (Figure 4G) and increased persistence of human CD8+ T cells (Figure 4H). Taken together, these studies highlight the in unique role of mSWI/SNF complexes, particularly cBAF, T cell exhaustion in concordance with the high degree of mSWI/SNF targeting and activity over exhaustion genes. Diverse mSWI/SNF ATPase-specific inhibitors and degraders attenuate T cell exhaustion and increase memory phenotypes We next sought to evaluate the impact of pharmacologic mSWI/ SNF perturbation, using both small-molecule allosteric inhibitors of SMARCA4/2 ATPase activity, CMP14 and FHT-1015, as well as degraders of the SMARCA4/2 ATPase protein subunits, ACBI1 and AU-15330, which result in the degradation of the entire 5-subunit ATPase module of mSWI/SNF com- plexes.43,56,90–93 Of note, an analog of FHT-1015, FHD-286, recently entered Phase I clinical trials for hematologic and solid tumors (NCT04891757 and NCT04879017). resulted in statistically significant We activated human CD8+ T cells with CD3/CD28 beads for 3 days and added either DMSO (control) or one of the four mSWI/SNF SMARCA4/2-targeting compounds at two different concentrations (50 and 100 nM) for an additional 6 days coupled with chronic stimulation (Figures 5A and S5A). At day 9, treat- ment with all compounds (ACBI1, AU-15330, CMP14, and FHT-1015) reductions of PD1+/TIM3+ exhausted populations at 51.5%, 51.5%, 20%, and 50%, respectively, compared with DMSO-treated condi- tions, consistently across donors (Figures 5B, 5C, and S5B). This was coupled with statistically significant increases in acti- vated/progenitor-exhausted T cells (PD1+/TIM3(cid:1)) and lower levels of CD39, an additional marker of terminal exhaustion (Figures 5B, 5C, S5B, and S5C). The profiling of CD45RA and CCR7 indicated a decrease in the effector T cell pool and an in- crease in both effector memory (EM) and CM cells upon treat- ment (Figures S5D–S5F). We also observed decreases in the expression of cytokines (IFNg, TNFa, or both) of T cells treated with all compounds (Figure S6A). Remarkably, although treat- ment did not significantly impact cell proliferation or viability until day 9, we observed an increase in persistence in treated cells from multiple donors of different ages and sex across further stimulations (fold differences: ACBI1: 6- to 20-fold; AU-15330: 5- to 13-fold; CMP14: 5- to 17-fold; FHT-1015: 2- to 5-fold) and a decrease in the percentage of Annexin-positive apoptotic cells in the treated conditions (Figures 5D, S6B, and S6C). To test whether mSWI/SNF PROTACs or inhibitors could not only prevent the onset of exhaustion but also revert it, we acti- vated human T cells and treated them with mSWI/SNF inhibitors or degraders starting at days 3, 6, or 9. Intriguingly, although treatment starting at day 3 attenuated the onset of exhaustion, Article A S N HO HO O NH N O O O O NH F OH N N N N NH2 N N N O NH2 N OH N H N O O NH N S F N HO O H N O S N N H O NS N H N H F F O N O S S ACBI1 AU-15330 CMP14 FHT-1015 ll OPEN ACCESS B 5 10 4 10 3 10 0 3 -10 N 3 M T I DMSO 68 5 17 ACBI1 50nM ACBI1 100nM AU-15330 50nM AU-15330 100nM 5 10 4 10 3 10 0 3 -10 51 5 38 5 10 4 10 3 10 0 3 -10 35 5 58 5 10 4 10 3 10 0 3 -10 44 5 47 5 10 4 10 3 10 0 3 -10 35 4 58 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 SMARCA4/2 degraders SMARCA4/2 inhibitors PD1 CD3/CD28 Dynabeads 5 10 4 10 3 10 0 3 -10 3 M T I DMSO 50 6 38 CMP14 50nM CMP14 100nM FHT-1015 50nM FHT-1015 100nM 5 10 4 10 3 10 0 3 -10 24 2 73 5 10 4 10 3 10 0 3 -10 10 2 87 5 10 4 10 3 10 0 3 -10 46 3 48 5 10 4 10 3 10 0 3 -10 25 2 71 0h 72h D6 D9-Ch PD1 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 CHRONIC STIMULATION + Compounds ** * ** * ** * ** * C + 8 D C f o % 80 60 40 20 0 DMSO ACBI1 50nM ACBI1 100nM AU-15330 50nM AU-15330 100nM 100 80 60 40 20 0 + 8 D C f o % *** * *** ** ** *** * DMSO CMP14 50nM CMP14 100nM FHT-1015 50nM FHT-1015 100nM PD1- TIM3- PD1+ TIM3- PD1+ TIM3+ PD1- TIM3- PD1+ TIM3- PD1+ TIM3+ D 400 Donor 3 ) 0 y a D n o n o i l l i m 1 r e p ( s l l e c n o i l l i M 200 0 120 Donor 4 80 40 0 1200 Donor 5 800 400 0 DMSO ACBI1 50nM ACBI1 100nM AU-15330 50nM AU-15330 100nM Donor 1 Donor 2 300 200 100 0 800 600 400 200 0 1000 Donor 6 800 600 400 200 0 DMSO CMP14 50nM CMP14 100nM FHT-1015 50nM FHT-1015 100nM ) 0 y a D n o n o i l l i m 1 r e p ( s l l e c n o i l l i M N°of stimulations Days 1 3 2 6 3 9 4 13 5 16 N°of stimulations Days 1 3 2 6 3 9 4 12 5 15 Figure 5. Pharmacologic disruption of mSWI/SNF complexes attenuates human T cell exhaustion (A) Schematic for inhibitor and degrader experiments with compounds added at day 3 and refreshed (with stimulation) every 3 days. (B) FACS plots depicting PD1/TIM3 populations in CD8+ T cells at day 9 treated with 50 nM and 100 nM of SMARCA4/2 degraders (donor 3) and inhibitors (donor 1). (C) Bar graph depicting % of CD8+ T cells in PD1(cid:1)/TIM3(cid:1), PD1+TIM3(cid:1), and PD1+TIM3+ populations in DMSO, ACBI1, and AU-15330 (Left) or DMSO, CMP14 and FHT-1015 (right) conditions. Error bars represent mean ± SD of 3 or 4 independent CD8+ T cell donors. Statistical analysis was performed using an unpaired t test. (D) Bar graphs depicting cell number upon treatment with ACBI1/AU-15330 (left, donors 3, 4, and 5) or CMP14/ FHT-1015 (right, donors 1, 2, and 6). See KRT for donor information. Error bars represent mean ± SD of 3 technical replicates per donor. *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001. See also Figures S5 and S6. as assessed by FACS and cell proliferation, the impact was reduced or absent if the treatment was initiated at day 6 or day 9, respectively (Figures S6D and S6E), suggesting a narrowing window amenable to chromatin landscape modification during the progenitor exhausted-like (Day 6) to terminal exhaustion- like (Day9-Ch) states. Finally, we performed similar experiments using mouse CD8+ T cells. Also in the mouse T cell context, treatment with CMP14 Molecular Cell 83, 1216–1236, April 20, 2023 1225 ll OPEN ACCESS Article ATAC-seq (Day9-Ch) B C ACBI1 vs CTRL-1 AU-15330 vs CTRL-1 % 2 1 . 5 2 : 2 C P 6 . 0 4 . 0 2 . 0 0 . 0 2 . 0 − 4 . 0 − 6 . 0 − A D CTRL-1 ACBI1 AU−15330 CTRL-2 CMP14 FHT-1015 Donor1 Donor2 Donor3 Donor4 ACBI1 389 650 699 AU−15330 340 281 1543 1680 531 783 2656 1397 FHT-1015 6918 CMP14 −0.6 −0.4 −0.2 0.0 0.2 0.4 0.6 PC1: 27.67% C1 (n=11167) Pan-accessible C2 (n=4439) Early activation C3 (n=2506) Activation/Exhaustion C4 (n=6998) Activation/Exhaustion C5 (n=4894) Late activation C6 (n=2759) Exhaustion C7 (n=3085) Naive/memory (strongest in D9-Tr) C8 (n=10163) Naive/memory (reduced upon activation) C9 (n=27495) Naive/memory (low accessibility) ATAC-seq Log2FC Degraders Inhibitors 3 3 log2FC (RPKM) 2 1 0 -1 -2 -3 2 1 0 -1 -2 -3 CXCL13 ENTPD1 A C B I 1 v s C T R L - 1 A U - 1 5 3 3 0 v s C T R L - 1 C M P 1 4 v s C T R L - 2 F H T - 1 0 1 5 v s C T R L - 2 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0 E s M K P R 2 g o L d e z i l a m r o N e l i t n a u Q F Clusters 3 & 4 1.5e-89 2.1e-57 0 1.3e-36 Cluster 6 1.1e-119 7.5e-110 1.1e-109 0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0 . 1 R H C I 1 B C A 0 3 3 5 1 − U A . 2 R H C 4 1 P M C 5 1 0 1 - T H F . 1 R H C I 1 B C A 0 3 3 5 1 − U A . 2 R H C 4 1 P M C 5 1 0 1 - T H F ACBI1 AU-15330 ATAC lost ATAC lost C6 C6 DEGs Down DEGs-Down CMP14 ATAC lost FHT-1015 ATAC lost C6 C6 DEGs Down DEGs Down s D I f i t o M F T s D I f i t o M F T ZBED1 ARI5A SNAI2 GMEB2.3 HD.9 HSFY2 IRF.1 HOMEZ SPDEF.1 E2F.1 NFAT.1 HINFP1.1 E2F.4 MYB.2 NFY HIF CTCF AHR MYB.1 GMEB2.2 PAX.halfsite YY1 NFKB.2 P53.like.3 ZNF146 CREB.ATF.3 CREB.ATF.2 SPI NR.14 ZNF652 BATF FOX.2 FOX.1 NFKB.3 AP1.1 AIRE NFKB.1 TCF.LEF RUNX.1 HNF1 −0.50 ZBED1 SNAI2 HOMEZ GMEB2.3 HSFY2 ARI5A HINFP1.1 HIF SPDEF.1 AHR E2F.4 NFY PROX1 CENBP CTCF YY1 IRF.1 E2F.1 NR.5 ZNF490 SMARCA1 ZNF232 ZNF652 MEF2 ZIM3 CREB.ATF.2 ZNF250 NFKB.2 NR.14 CREB.ATF.3 SPI NFKB.1 BATF NFKB.3 AIRE AP1.1 FOX.1 TCF.LEF RUNX.1 HNF1 −0.25 0 Model Coefficient 0.25 0.50 Accessibility: UP DOWN −0.25 0 Model Coefficient 0.25 CMP14 vs CTRL-2 FHT-1015 vs CTRL-2 SNAI2 SPDEF.1 CENBP HD.9 ARI5A GMEB2.3 PROX1 E2F.4 HINFP1.3 CTCF CUX.1 HIC.1 AHR E2F.1 ZBTB7A MYB.4 HSFY2 PAX.halfsite GMEB2.2 NFY MYB.5 BCL6.1 HD.13 AIRE NFKB.2 NR.6 FOX.2 CUX.4 ZNF652 FOX.1 CREB.ATF.3 MEF2 CREB.ATF.2 NFKB.3 NFKB.1 TCF.LEF RUNX.1 AP1.1 BATF HNF1 SNAI2 GMEB2.3 ARI5A AHR HOMEZ SPDEF.1 GMEB2.2 HSFY2 MYB.2 E2F.1 ZNF384.1 E2F.4 HIC.1 CTCF HD.9 HIF HINFP1.1 CUX.1 YY1 CENBP IRF.1 ZBTB7A NR.14 NFKB.2 POU.1 ZNF652 TCF.LEF FOX.2 HD.13 NFKB.1 FOX.1 AIRE CREB.ATF.2 NFKB.3 CREB.ATF.3 MEF2 RUNX.1 BATF AP1.1 HNF1 Accessibility: UP DOWN −0.50 −0.25 0 Model Coefficient 0.25 0.50 −0.50 −0.25 0 0.25 0.50 Model Coefficient G Degraders Inhibitors scRNAseq signatures: Exhaustion - Tirosh Exhaustion - Zheng Exhaustion - Zhang Exhaustion - Guo Exhaustion - Sade-Feldman Memory - Sade-Feldman Normalized enrichment score 3 0 −3 A C B I 1 v s C T R L - 1 A U - 1 5 3 3 0 v s C T R L - 1 C M P 1 4 v s C T R L - 2 F H T - 1 0 1 5 v s C T R L - 2 H chr12:68,545,746-68,556,467 chr4:78,498,312-78,522,161 chr10:97,494,009-97,631,689 CTRL-1 [0 - 0.53] CTRL-1 [0 - 2.64] CTRL-1 [0 - 2.62] ACBI1 AU-15330 CTRL-2 CMP14 FHT-1015 ACBI1 AU-15330 CTRL-2 CMP14 FHT-1015 ACBI1 AU-15330 CTRL-2 CMP14 FHT-1015 IFNG 0 0 . 5 1 . 0 1 . 5 RNA (relative cpm) CXCL13 0 0 . 5 1 . 0 1 . 5 RNA (relative cpm) 1226 Molecular Cell 83, 1216–1236, April 20, 2023 ENTPD1 0 0 . 5 1 . 0 1 . 5 ENTPD1-AS1 RNA (relative cpm) ENTPD1 ENTPD1 ENTPD1 (legend on next page) Article and FHT-1015 SMARCA4/2 ATPase inhibitors led to a reduction of PD1+TIM3+ cells and in cytokines expression (Figure S6F). Taken together, these results indicate that pharmacologic tar- geting of the SWI/SNF complex attenuates the onset of exhaus- tion hallmarks and promotes increased persistence and memory phenotypic features of both human and mouse T cells. Pharmacologic disruption of mSWI/SNF complex activity alters accessibility over TF motif sites and inhibits T cell exhaustion To dissect the mechanistic basis for the observed phenotypes, we performed ATAC-seq upon treatment with SMARCA4/2 ATPase degraders and inhibitors during chronic stimulation in human CD8+ T cells. Notably, PCA analyses revealed dramatic differences in chromatin accessibility between control and treated cells at Day9-Ch, consistent between independent do- nors, with PC1 capturing the impact of SMARCA4/2 pharmaco- logic perturbation (Figures 6A and S7A). As expected from previous studies, treatment with these compounds resulted in genome-wide decreases in accessibility, with sites affected be- ing highly consistently impacted across all four compound treat- ments relative to control (Figures 6B and S7A–S7C). Notably, sites decreased in accessibility were enriched for motifs corre- sponding to the HNF1B TF, underscoring the function of mSWI/SNF complexes in mediating accessibility over HNF1B binding motifs (Figure 6C). In addition, motifs corresponding to BATF, AP1, NF-kB, and FOX were also enriched over sites exhib- iting decreased accessibility following treatment (Figure 6C). Chromatin accessibility changes upon all treatments compared with control were highly concordant across the 9 clus- ters identified (Figures 1, 6D, and 6E). Importantly, C6 exhaus- tion-associated sites exhibited decreases in the accessibility of the highest significance and magnitude, whereas C3 and C4 sites, which included sites broadly accessible during activation fold changes and exhaustion, decreased with (Figures 6D and 6E), indicating that mSWI/SNF ATPase inhibition most strongly suppressed the accessibility over regions en- riched for exhaustion-associated genes as well as selected acti- vation-associated genes (Figure 6F). Interestingly, mSWI/SNF disruption moderately impacted accessibility (both increases and decreases) over naive/memory-associated sites (C7 and C8), consistent with the fact that these cells gain some but not all memory-like features. Reflecting this, genes near sites with increased accessibility included TCF7, ID3, and KLF2, and those lower ll OPEN ACCESS near sites with reduced accessibility included SELL and BHLHE40 genes (Figures 6D–6F). Intriguingly, accessibility was markedly increased over genes within C1 and C9 clusters, in agreement with previous findings that cBAF complex perturba- tions result in enhanced abundance and function of ncBAF com- plexes over CTCF motifs, which we demonstrated were strongly enriched in the C9 cluster (Figures 2A, 2B, and S3B).56 across treatment expressed genes To further investigate the impact of mSWI/SNF pharmacologic inhibition on the attenuation of exhaustion-like states, we next performed RNA-seq analyses to reveal changes in gene expres- sion programs (Figures S7D and S7E). PCA analysis revealed similarly changed profiles and high concordance between differ- entially conditions (Figures S7D, S7F, and S7G). Downregulated genes were en- riched for those involved in immune-related pathways, key acti- vation and exhaustion TFs, IFNg response, and TNFa signaling via NF-kB (Figure S7H). A subset of genes downregulated upon mSWI/SNF inhibition was contained within genomic re- gions losing DNA accessibility upon compound treatment (30.5%–49.6%), a subset of which (8%–16.3% of total downre- gulated genes) were genes bound by mSWI/SNF in the ex- hausted state (C6) (Figure 6F). Examples included ENTPD1, TI- GIT, IFNG, and GZMB, which demonstrated decreased accessibility and gene expression upon compound treatment (Figure S7I). Interestingly, we also observed chromatin opening and enhanced gene expression at a smaller set of sites (Figures S7A–S7C and S7E–S7I), exemplified over genes such as IRF1 (Figure S7J), which may indicate a set of factors that facilitate the increased T cell persistence.94 Finally, we applied the C6 mSWI/SNF perturbation signature to differentially accessible and differentially expressed genes with the exhaustion and memory signatures derived from human scRNA-seq datasets.80–84 Remarkably, treatment with all four distinct mSWI/SNF-disrupting compounds showed negative enrichment exhaustion signature genes, whereas memory genes increased in expression (Figures 6G and 6H). These data support a mechanism whereby mSWI/SNF inhibition or degradation at- tenuates the activation status of T cells, preventing them from undergoing exhaustion, and enables the maintenance of the memory-like phenotype with sustained proliferation capability over time. Finally, we performed ATAC-seq in mouse CD8+ T cells treated with SMARCA4/2 ATPase inhibitors and found that sites reduced in accessibility were enriched for similar motifs, such as Figure 6. mSWI/SNF pharmacological disruption alters chromatin accessibility and TFs recruitment at key T cell activation and exhaus- tion sites (A) PCA of ATAC-seq profiles of control (CTRL) and ACBI1, AU-15330, CMP14, or FHT-1015-treated human CD8+ T cells (100 nM), at day 9. CTRL-1 and CTRL-2 are the controls for the ACBI1/AU-15330 (donors 3 and 4) and CMP14/FHT-1015 (donors 1 and 2) experiments, respectively. (B) Venn diagram showing the overlap in sites with decreased accessibility (logFC < (cid:1)1) upon treatment with ACBI1, AU-15330, CMP14, or FHT-1015 (100 nM). (C) Top 40 coefficients of logistic regression models fitting motif counts across all sites to changes in accessibility for indicated comparisons. (D) Heatmap showing the log2 fold-change of accessibility upon ACBI1, AU-15330, CMP14, or FHT-1015 treatment compared with control in the 9 clusters identified in Figure 1. (E) Quantification of chromatin accessibility (quantile-normalized log2 RPKM) at sites within clusters 3 and 4 and cluster 6. CTRL-1 and CTRL-2 are the controls for the ACBI1/AU-15330 and CMP14/FHT-1015 experiments, respectively. P values were computed using standard t tests. (F) Pie charts representing percentages of downregulated genes after treatment near cluster 6 (C6) sites with strong decreases in accessibility (log2FC < (cid:1)1). (G) Gene set enrichment analysis (GSEA) analysis of exhaustion and memory signatures derived from scRNA-seq datasets in the selected comparisons. (H) Representative SMARCA4, SS18, H3K27ac C&T, and ATAC-seq tracks over the IFNG, CXCL13, and ENTPD1 loci. See also Figure S7. Molecular Cell 83, 1216–1236, April 20, 2023 1227 ll OPEN ACCESS A Human CD4+ T cells DMSO ACBI1 100nM AU15330 100nM CMP14 100nM FHT-1015 100nM 37 4 10 3 10 0 18 4 10 3 10 0 21 4 10 3 10 0 28 4 10 3 10 0 29 4 10 3 10 0 3 -10 11 44 3 -10 8 71 3 -10 9 66 3 -10 3 67 3 -10 1 69 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 3 M T I Article PD1 B DMSO ACBI1 AU-15330 CMP14 FHT-1015 C ) 1 y a D t a M 1 m o r f ( s l l e c n o i l l i M ) 4 ^ 0 1 x ( 5 4 3 2 1 0 DMSO ACBI1 100nM AU-15330 100nM CMP14 100nM FHT-1015 100nM 0 10 3 CD39 10 4 10 5 N°of stimulations Days 1 3 2 6 3 9 4 13 5 16 CD19-CAR-T construct PROTACs (ACBI1/AU-15330) D CD3/CD28 Dynabeads Human CD3+ T cells E - A C S F 250K 200K 150K 100K 50K 0 FACS profiling Proliferation 0h 24h 72h D7 D10-Ch ACTIVATION TRANSDUCTION EXPANSION + Compounds G CAR-T-GFP + 52.3 3 -10 0 3 10 4 10 5 10 FITC 3 M T I CD19 - CAR-T DMSO 82 ACBI1 100nM AU-15330 100nM 48 63 4 10 3 10 0 3 -10 PD1 0 0 6 10 5 10 4 10 3 10 2 10 1 10 0 10 4 10 3 10 0 3 -10 17 1 50 4 10 3 10 0 3 -10 3 10 4 10 0 3 10 4 10 1 0 35 3 10 CAR-T Donor 10 DMSO ACBI1 100nM AU-15330 100nM s l l e c n o i l l i M ) 0 y a D n o n o i l l i m 1 r e p ( CAR-T Donor 9 6 10 5 10 4 10 3 10 2 10 1 10 0 10 N°of stimulations Days 1 4 2 7 3 10 4 13 5 16 1 4 2 7 3 10 4 13 5 16 AU-15330 (100nM) FHT-1015 (100nM) 500 400 300 200 100 0 DMSO Treatment from Day3 Treatment from Day3 - Released at Day9 I ) 3 m m ( e m u o v l r o m u T 3 6 9 Days 13 16 3 6 9 Days 13 16 3 6 9 Days 13 16 Control of B16-OVA tumor growth with CD8+ OT-1 T cells -/+ mSWI/SNF inhibition DMSO (n=5) FHT-1015 (n=5) 3000 ) 3 m m ( e m u o v l r o m u T 10 19 13 7 Days after tumor engraftment 16 2000 1000 0 7 10 13 16 19 Days after tumor engraftment 3000 2000 1000 0 1228 Molecular Cell 83, 1216–1236, April 20, 2023 (legend on next page) F H s l l e c n o i l l i M DMSO ACBI1 AU-15330 2000 1500 1000 500 0 0 3 10 4 10 0 4 10 5 10 LAG-3 CD39 ACBI1 (100nM) 2000 1500 1000 500 0 ) 0 y a D n o n o i l l i m 1 r e p ( Article ll OPEN ACCESS those corresponding to BATF, MYB, E2F, and NFY factors, but not for HNF1B, confirming its specificity to human cells. Impor- tantly, examining the gene expression changes across clusters of mSWI/SNF occupancy and accessibility in mouse cells (Fig- ure S2J), we identified the most strongly downregulated genes were again those in C6, corresponding to the exhaustion state (Figures S7K and S7L). Pharmacologic disruption of mSWI/SNF increases in vitro persistence during CAR-T cell generation and enhances T cell-mediated anti-tumor efficacy in vivo CAR-T infusion products displaying decreased exhaustion and increased memory hallmarks have been demonstrated to have increased efficacy in vivo.89,95–100 With the above results, we reasoned that pharmacologic mSWI/SNF complex perturbation may represent a viable approach to improve T cell fitness and pre- vent exhaustion during CAR-T cell manufacturing.101 CAR-Ts are routinely generated from both CD4+ and CD8+ T cells at variable ratios and expanded using beads similar to those used in our in vitro exhaustion experiments.102,103 As CD4+ T cells have been implicated to be longer-lasting relative to CD8+ cells, detected even decades after tumor remission, we sought to understand whether CD4+ T cells displayed similar increased persistence as CD8+ T cells in response to mSWI/SNF inhibi- tion.101,104,105 Treatment with the mSWI/SNF ATPase inhibitor and degrader compounds led to a significant reduction in PD1+TIM3+ and CD39+ populations and increased CD4+ T cell persistence (Figures 7A–7C), indicating that these treatments have the same outcome in both CD8+ and CD4+ lineages. We then generated CD19-CAR-T cells using total human CD3+ T cells and engineered to express a CAR-T construct targeting the CD19 antigen (Figures 7D and S7M). Cells were expanded for 2 additional days, then treated with either DMSO or SMARCA4/2 degraders, and analyzed at day 10 (Figures 7D– 7F). Remarkably, the treatment of CAR-T cells with mSWI/SNF ATPase degraders resulted in the marked depletion of PD1+TIM3+, LAG3+, and CD39+ populations (Figures 7D–7F) and increased persistence across two independent human T cell donors (Figure 7G). These results suggest potential approaches in which CAR-T cells are expanded in the presence of SMARCA4/2 degraders or inhibitors, prior to infusion into patients. In an attempt to assess the potential duration of proliferative advantage of treated T cells once injected in vivo (where no inhibitor would be present), we released treatment on day 9 and monitored proliferation. Cells retained a significant proliferation advantage compared with DMSO-treated cells for at least 1 week after treatment release (Figure 7H). Finally, to assess the anti-tumor functionality of mSWI/SNF-in- hibited T cells, we implemented an OVA antigen-expressing mel- anoma model (B16) with a subsequent infusion of CD8+ T cells from OT-1 TCR transgenic mice. Intriguingly, the pre-treatment of CD8+ OT-1 T cells with the FHT-1015 SMARCA4/2 ATPase in- hibitor resulted in decreased levels of day 9 in vitro cell killing at 24 and 48 h time points across a range of target:effector ratios relative to control T cells (Figure S7N), consistent with results us- ing cBAF subunit genetic depletion experiments. Importantly, OT-1 T cells pretreated with FHT-1015 attenuated B16-OVA tu- mor growth in vivo relative to DMSO control-treated cells (Fig- these experiments demonstrate that ure 7I). Collectively, T cells subjected to mSWI/SNF inhibition exhibit increased persistence resulting in enhanced anti-tumor activity. DISCUSSION In this study, we dissected the contribution of mSWI/SNF chromatin remodeling complexes targeting and activity to gene expression programs and functional phenotypes during T cell activation and exhaustion. Our efforts to probe the mech- anistic contributions of mSWI/SNF complexes were inspired by the known extensive changes in chromatin accessibility dur- ing T cell activation and exhaustion and in CAR-T cell populations,14–22 coupled with work by our group and others establishing mSWI/SNF complexes as major mediators of the establishment and maintenance of tissue-specific chromatin accessibility.56,58,62,106,107 recent CRISPR-based Further, screening studies begun to reveal roles for mSWI/SNF com- plexes in T cell exhaustion,35,36 presenting opportunities to define their functional contributions. Our studies describe the chromatin, gene regulatory, pheno- typic, and in vitro and in vivo functional impact of four independent inhibitors and degraders that have been biochemically and struc- turally confirmed to specifically target the mSWI/SNF SMARCA2/ 4 ATPases. Underscoring the clinical relevance of our findings, the FHT-1015 compound is an analog of the phase I compound, FHD- 286, currently being evaluated in human acute myeloid leukemia (AML), myelodysplastic syndrome (MDS), and uveal melanoma.93 Figure 7. mSWI/SNF targeting improves T cell-based cancer immunotherapy approaches (A) FACS plots depicting PD1/TIM3 populations in DMSO, ACBI1, AU-15330, CMP14, and FHT-1015 conditions (100 nM), at day 9, for one human CD4+ T cell donor (donor 8). (B) FACS plot depicting the profiling of CD39 in human CD4+ T cells treated with DMSO, ACBI1, AU-15330, CMP14, or FHT-1015 (100 nM), at day 9. (C) Bar graph depicting human CD4+ T cell number upon treatment with DMSO, ACBI1, AU-15330, CMP14, or FHT-1015 (100 nM). Error bars represent the mean ± SD of 3 technical replicates of one donor. (D) Schematic for CD19-CAR-T cell generation, stimulation, and treatments. (E) FACS plots depicting CD19-CAR-T-GFP cells identification and PD1/TIM3 populations, in cells treated with ACBI1 or AU-15330 (donors 9 and 10). (F) FACS histograms of LAG-3 and CD39 expression in CAR-T cells treated with DMSO, ACBI1, or AU-15330. (G) Bar graphs depicting CAR-T cell number upon treatment with DMSO, ACBI1, or AU-15330 (100 nM). Error bars represent the mean for each donor. (H) Bar graphs depicting cell number upon treatment with ACBI1, AU-15330, or FHT-1015 (all 100 nM) at day 3 onward or at day 3 with treatment washout at day 9. Error bars represent mean ± SD of 3 technical replicates of one donor. (I) In vivo B16 melanoma tumor growth curves in mice injected with DMSO or FHT-1015-treated CD8+ OT-1 T cells. Two-way ANOVA generated p value for comparison between the two groups’ means at day 19: p < 0.05. See also Figure S7. Molecular Cell 83, 1216–1236, April 20, 2023 1229 ll OPEN ACCESS With these agents, we find highly consistent chromatin accessi- bility and gene regulatory impacts, suggesting that mSWI/SNF catalytic activity is equivalent to assembly and function of the entire ATPase module. The degradation of SMARCA4/2 prevents assembly of ACTL6A, beta-actin, SS18, and BCL7 family subunits on to mSWI/SNF family complexes.43,108,109 Based on a range of studies by our group and others, we sug- gest a model where the activities of mSWI/SNF complexes are controlled by the repertoire of TFs expressed in a given cell type, which collectively guides their positioning. Indeed, the pharmaco- logic inhibition of mSWI/SNF complexes results in the closing of previously accessible, complex-targeted regions enriched in DNA motifs corresponding to highly specific TFs (Figures 6C and 6D). In addition, small-molecule inhibition resulted in minimal cell viability impacts, underscoring the preferential role for mSWI/SNF complexes over sites important for the exhaustion program rather than those supporting cell homeostatic programs (Figures 5D, 7D–7H, S6B, and S6C). These data, coupled with immunophenotyping indicating decreased T cell exhaustion, sug- gest the favorable utility of mSWI/SNF inhibitors in the setting of ex vivo-manipulated CAR-T cells. The use of small-molecule inhib- itors represents a more facile, lower-cost, and safer approach to improve CAR-T cells fitness, contrary to the challenges and safety concerns associated with genetic strategies such as DNMT3A deletion.8,110,111 Intriguingly, in both of our screens, Dnmt3a, which has been previously shown to contribute to T cell exhaustion, was not depleted relative to mSWI/SNF genes, indicating relative con- tributions of these distinct epigenetic regulators (Figures 4 and S4; Tables S3 and S4)8,34,112,113 Of note, our studies suggest that mSWI/SNF inhibition can pre- vent T cell exhaustion but may not be able to revert it (Figures S6D and S6E). This indicates that mSWI/SNF complexes may facilitate the stable binding of important TFs to exhaustion sites—once the exhaustion program has been triggered, mSWI/SNF inhibition may not be sufficient to displace those TFs. Additional studies em- ploying different time points at which mSWI/SNF inhibitors are introduced will be needed to comprehensively address whether mSWI/SNF perturbation can affect already exhausted cells and reinvigorate tumor-infiltrating T cells in vivo. We identify here several previously unknown genomic features of the exhausted T cell state as well as determinants of exhaus- tion-specific mSWI/SNF targeting, such as the connection be- tween mSWI/SNF occupancy and the strong enrichment of HNF1B sites. HNF1B is a heterodimeric TF (heterodimerizes with HNF1A) and was originally characterized for its functions in the development of the pancreas and liver and in controlling insulin production.114 Of note, we found that HNF1B is not ex- pressed in mouse T cells (Figure S3K), which may reconcile why HNF1B was not previously identified in any CRISPR screen, all of which have been performed in mouse T cell settings. For HNF1B in particular, mice with heterozygous mutations in HNF1B show no phenotype relative to that seen in humans.114 In solid tumors, HNF1B regulates glucose uptake, glucose function.102,103,115–118 Ex- metabolism, and mitochondrial hausted T cells are known to have rewired glucose metabolism, a higher rate of glycolysis and impaired oxidative phosphoryla- tion.103 Gene ontology (GO) analysis of HNF1B targets identified the MAPK signaling pathway and metabolic pathways as 1230 Molecular Cell 83, 1216–1236, April 20, 2023 Article enriched processes (Figure 3L), pointing toward a connection between HNF1B and T cell metabolism that could be further explored. In summary, our study presents a comprehensive dissection of the mechanisms by which mSWI/SNF complexes regulate T cell activation and exhaustion, serving as a valuable resource for the field and advancing new clinically relevant strategies for immunotherapy improvement. Limitations of the study Although our results suggest that cBAF complexes specifically mediate T cell exhaustion, SMARCA4/2 inhibitors and degraders target all mSWI/SNF family sub-complexes, thus we cannot exclude the contributions of ncBAF and PBAF activities. Further, although we have accomplished an extensive repertoire of genomic studies in primary T cells, we were unable to biochem- ically identify and characterize TF-BAF complex protein-protein interactions owing to limitations in cell numbers. Additionally, the experiments presented here were performed across donors with varied naive:memory T cell ratios—further work will be required to compare the impact of cBAF disruption in human naive versus memory T cell populations. Finally, in vivo studies assessing the impacts of mSWI/SNF inhibitors and degraders may pose chal- lenges in uncoupling tumor cell intrinsic and T cell microenviron- mental changes. STAR+METHODS Detailed methods are provided in the online version of this paper and include the following: d KEY RESOURCES TABLE d RESOURCE AVAILABILITY B Lead contact B Materials availability B Data and code availability d EXPERIMENTAL MODEL AND SUBJECT DETAILS B Primary T cells d METHOD DETAILS B Human CD8+ and CD4+ T cell isolation B Mouse CD8+ T cell isolation B In vitro T cell activation and exhaustion B FACS staining B Chromatin-focused CRISPR screen B Mouse CRISPR sgRNAs knock-out experiments B Human CRISPR sgRNAs knock-out experiments B B16-OVA in vitro exhaustion model B Western blots B Cell killing assays B CD19-CAR-T cell experiments B In vivo B16-OVA killing assays B SMARCA4/SMARCA2 Inhibitor and Degrader Small Molecule Treatment Studies B RNA-seq B ATAC-seq B Cut & Tag d QUANTIFICATION AND STATISTICAL ANALYSIS B NGS Data Processing Article B RNA-seq data analysis B Cut & Tag and ATAC-seq data analysis and integration B Transcription factor motif and archetype analyses B CRISPR screen data analysis B scRNAseq and scATACseq datasets and analyses SUPPLEMENTAL INFORMATION Supplemental molcel.2023.02.026. information can be found online at https://doi.org/10.1016/j. ACKNOWLEDGMENTS We are grateful to all members of the Kadoch and Aifantis laboratories for their helpful discussions during the preparation of this study. We are also grateful to Drs. Arlene Sharpe, Sarah Weiss, Debamatta Sen, and Hsaio-Wei Tsao for their helpful insights and for providing mouse T cell protocols and reagents, as well as Z. Herbert and M. Sullivan of the Molecular Biology Core Facility (MBCF) at the Dana-Farber Cancer Institute. We would also like to thank the NYU Genome Technology Center for sequencing experiments. This work was supported in part by NIH 1F31CA271427-01 (K.H.), NIH 5F30CA239317 (D.E.C.), T32GM007753 and T32GM144273 (D.E.C.), and 1DP2CA195762 (C.K.). C.K. was also supported by the Mark Foundation for Cancer Research Emerging Leader Award. E.B. was supported by the Swiss National Science Foundation (SNSF) and the Lymphoma Research Foundation (LRF). I.A. was 5R01CA228135, 5P01CA229086, 5R01CA242020, 1R01CA243001, and 1R01CA252239) and the Vogelstein Foundation. NYU Langone’s Genome Technology Center (RRID: SCR_017929) is partially supported by the Cancer Center Support grant P30CA016087 at the Laura and Isaac Perlmutter Cancer Center. (5R01CA173636, supported by the NIH/NCI AUTHOR CONTRIBUTIONS E.B., K.A.H., D.E.C., I.A., and C.K. conceived of and directed the study. E.B., K.A.H., and D.E.C. performed all cell-based and in vivo experiments, with help from M.M.H. All computational and statistical analyses were performed by C.K.C., with help from J.R.H., K.S.C., and S.L. X.C. performed in vivo experi- ments. D.E.C. was jointly supervised by W.N.H. I.A. and C.K. funded the study. E.B., K.A.H., I.A., and C.K. wrote the manuscript, and all authors critically re- viewed and edited the manuscript. DECLARATION OF INTERESTS C.K. is the scientific founder, scientific advisor to the board of directors, scien- tific advisory board member, shareholder, and consultant for Foghorn Thera- peutics, Inc. (Cambridge, MA). C.K. is also a member of the scientific advisory board and is a shareholder of Nested Therapeutics and Nereid Therapeutics, serves on the scientific advisory board for Fibrogen, Inc. and on the Molecular Cell editorial board, and is a consultant for Cell Signaling Technologies and Google Ventures. I.A. is a scientific consultant for Foresite Labs and receives research funding from AstraZeneca Inc. F.P. is an inventor on patents related to adoptive cell therapies, held by MSKCC (some licensed to Takeda), serves as a consultant for AstraZeneca, and receives research support from Lonza and NGMBio. A.T. is a scientific advisor to Intelligencia AI. INCLUSION AND DIVERSITY We support inclusive, diverse, and equitable conduct of research. One or more of the authors of this paper self-identifies as an underrepresented ethnic mi- nority in their field of research or within their geographical location. Received: November 4, 2022 Revised: January 11, 2023 Accepted: February 23, 2023 Published: March 20, 2023 ll OPEN ACCESS REFERENCES 1. Kaech, S.M., and Cui, W. (2012). Transcriptional control of effector and memory CD8+ T cell differentiation. Nat. Rev. Immunol. 12, 749–761. https://doi.org/10.1038/nri3307. 2. Schietinger, A., Philip, M., Krisnawan, V.E., Chiu, E.Y., Delrow, J.J., Basom, R.S., Lauer, P., Brockstedt, D.G., Knoblaugh, S.E., H€ammerling, G.J., et al. (2016). Tumor-specific T cell dysfunction is a dy- namic antigen-driven differentiation program initiated early during tumor- igenesis. Immunity 45, 389–401. https://doi.org/10.1016/j.immuni.2016. 07.011. 3. Seo, W., Jerin, C., and Nishikawa, H. (2021). Transcriptional regulatory network for the establishment of CD8+ T cell exhaustion. Exp. Mol. Med. 53, 202–209. https://doi.org/10.1038/s12276-021-00568-0. 4. Zhang, N., and Bevan, M.J. (2011). CD8+ T cells: foot soldiers of the im- mune system. Immunity 35, 161–168. https://doi.org/10.1016/j.immuni. 2011.07.010. 5. Chang, C.-H., Qiu, J., O’Sullivan, D., Buck, M.D., Noguchi, T., Curtis, J.D., Chen, Q., Gindin, M., Gubin, M.M., van der Windt, G.J.W., et al. (2015). Metabolic competition in the tumor microenvironment is a driver of cancer progression. Cell 162, 1229–1241. https://doi.org/10.1016/j. cell.2015.08.016. 6. Wherry, E.J., and Kurachi, M. (2015). Molecular and cellular insights into T cell exhaustion. Nat. Rev. Immunol. 15, 486–499. https://doi.org/10. 1038/nri3862. 7. Pardoll, D.M. (2012). The blockade of immune checkpoints in cancer immunotherapy. Nat. Rev. Cancer 12, 252–264. https://doi.org/10. 1038/nrc3239. 8. Prinzing, B., Zebley, C.C., Petersen, C.T., Fan, Y., Anido, A.A., Yi, Z., Nguyen, P., Houke, H., Bell, M., Haydar, D., et al. (2021). Deleting DNMT3A in CAR T cells prevents exhaustion and enhances antitumor ac- tivity. Sci. Transl. Med. 13, eabh0272. https://doi.org/10.1126/sci- translmed.abh0272. 9. Pauken, K.E., and Wherry, E.J. (2015). Overcoming T cell exhaustion in infection and cancer. Trends Immunol. 36, 265–276. https://doi.org/10. 1016/j.it.2015.02.008. 10. Wang, B., Zhang, W., Jankovic, V., Golubov, J., Poon, P., Oswald, E.M., Gurer, C., Wei, J., Ramos, I., Wu, Q., et al. (2018). Combination cancer immunotherapy targeting PD-1 and GITR can rescue CD8+ T cell dysfunction and maintain memory phenotype. Sci. Immunol. 3, eaat7061. https://doi.org/10.1126/sciimmunol.aat7061. 11. Klein Geltink, R.I., Edwards-Hicks, J., Apostolova, P., O’Sullivan, D., Sanin, D.E., Patterson, A.E., Puleston, D.J., Ligthart, N.A.M., Buescher, J.M., Grzes, K.M., et al. (2020). Metabolic conditioning of CD8+ effector T cells for adoptive cell therapy. Nat. Metab. 2, 703–716. https://doi.org/ 10.1038/s42255-020-0256-z. 12. Chowell, D., Yoo, S.-K., Valero, C., Pastore, A., Krishna, C., Lee, M., Hoen, D., Shi, H., Kelly, D.W., Patel, N., et al. (2022). Improved prediction of im- mune checkpoint blockade efficacy across multiple cancer types. Nat. Biotechnol. 40, 499–506. https://doi.org/10.1038/s41587-021-01070-8. 13. Snyder, A., Makarov, V., Merghoub, T., Yuan, J., Zaretsky, J.M., Desrichard, A., Walsh, L.A., Postow, M.A., Wong, P., Ho, T.S., et al. (2014). Genetic basis for clinical response to CTLA-4 blockade in mel- anoma. N. Engl. J. Med. 371, 2189–2199. https://doi.org/10.1056/ NEJMoa1406498. 14. Li, P., and Leonard, W.J. (2018). Chromatin accessibility and interactions in the transcriptional regulation of T cells. Front. Immunol. 9, 2738. https://doi.org/10.3389/fimmu.2018.02738. 15. Philip, M., Fairchild, L., Sun, L., Horste, E.L., Camara, S., Shakiba, M., Scott, A.C., Viale, A., Lauer, P., Merghoub, T., et al. (2017). Chromatin states define tumour-specific T cell dysfunction and reprogramming. Nature 545, 452–456. https://doi.org/10.1038/nature22367. Molecular Cell 83, 1216–1236, April 20, 2023 1231 ll OPEN ACCESS 16. Henning, A.N., Roychoudhuri, R., and Restifo, N.P. (2018). Epigenetic control of CD8+ T cell differentiation. Nat. Rev. Immunol. 18, 340–356. https://doi.org/10.1038/nri.2017.146. 17. Sen, D.R., Kaminski, J., Barnitz, R.A., Kurachi, M., Gerdemann, U., Yates, K.B., Tsao, H.-W., Godec, J., LaFleur, M.W., Brown, F.D., et al. (2016). The epigenetic landscape of T cell exhaustion. Science 354, 1165–1169. https://doi.org/10.1126/science.aae0491. 18. Gennert, D.G., Lynn, R.C., Granja, J.M., Weber, E.W., Mumbach, M.R., Zhao, Y., Duren, Z., Sotillo, E., Greenleaf, W.J., Wong, W.H., et al. (2021). Dynamic chromatin regulatory landscape of human CAR T cell exhaustion. Proc. Natl. Acad. Sci. USA 118. e2104758118. https://doi. org/10.1073/pnas.2104758118. 19. Jiang, P., Zhang, Z., Hu, Y., Liang, Z., Han, Y., Li, X., Zeng, X., Zhang, H., Zhu, M., Dong, J., et al. (2022). Single-cell ATAC-seq maps the compre- hensive and dynamic chromatin accessibility landscape of CAR-T cell dysfunction. Leukemia 36, 2656–2668. https://doi.org/10.1038/s41375- 022-01676-0. 20. Wang, W., Fasolino, M., Cattau, B., Goldman, N., Kong, W., Frederick, M.A., McCright, S.J., Kiani, K., Fraietta, J.A., and Vahedi, G. (2020). Joint profiling of chromatin accessibility and CAR-T integration site anal- ysis at population and single-cell levels. Proc. Natl. Acad. Sci. USA 117, 5442–5452. https://doi.org/10.1073/pnas.1919259117. 21. Chen, G.M., Chen, C., Das, R.K., Gao, P., Chen, C.-H., Bandyopadhyay, S., Ding, Y.-Y., Uzun, Y., Yu, W., Zhu, Q., et al. (2021). Integrative bulk and single-cell profiling of premanufacture T-cell populations reveals factors mediating long-term persistence of CAR T-cell therapy. Cancer Discov. 11, 2186–2199. https://doi.org/10.1158/2159-8290.CD-20-1677. 22. Kourtis, N., Wang, Q., Wang, B., Oswald, E., Adler, C., Cherravuru, S., Malahias, E., Zhang, L., Golubov, J., Wei, Q., et al. (2022). A single-cell map of dynamic chromatin landscapes of immune cells in renal cell car- cinoma. Nat. Cancer 3, 885–898. https://doi.org/10.1038/s43018-022- 00391-0. 23. van der Leun, A.M., Thommen, D.S., and Schumacher, T.N. (2020). CD8+ T cell states in human cancer: insights from single-cell analysis. Nat. Rev. Cancer 20, 218–232. https://doi.org/10.1038/s41568-019-0235-4. 24. Utzschneider, D.T., Gabriel, S.S., Chisanga, D., Gloury, R., Gubser, P.M., Vasanthakumar, A., Shi, W., and Kallies, A. (2020). Early precursor T cells establish and propagate T cell exhaustion in chronic infection. Nat. Immunol. 21, 1256–1266. https://doi.org/10.1038/s41590-020-0760-z. 25. Tsui, C., Kretschmer, L., Rapelius, S., Gabriel, S.S., Chisanga, D., Kno¨ pper, K., Utzschneider, D.T., N€ussing, S., Liao, Y., Mason, T., et al. (2022). MYB orchestrates T cell exhaustion and response to checkpoint inhibition. Nature 609, 354–360. https://doi.org/10.1038/s41586-022-05105-1. 26. Macian, F. (2005). NFAT proteins: key regulators of T-cell development and function. Nat. Rev. Immunol. 5, 472–484. https://doi.org/10.1038/ nri1632. 27. Alfei, F., Kanev, K., Hofmann, M., Wu, M., Ghoneim, H.E., Roelli, P., Utzschneider, D.T., von Hoesslin, M., Cullen, J.G., Fan, Y., et al. (2019). TOX reinforces the phenotype and longevity of exhausted T cells in chronic viral infection. Nature 571, 265–269. https://doi.org/10.1038/ s41586-019-1326-9. 28. Scott, A.C., D€undar, F., Zumbo, P., Chandran, S.S., Klebanoff, C.A., Shakiba, M., Trivedi, P., Menocal, L., Appleby, H., Camara, S., et al. (2019). TOX is a critical regulator of tumour-specific T cell differentiation. Nature 571, 270–274. https://doi.org/10.1038/s41586-019-1324-y. 29. Khan, O., Giles, J.R., McDonald, S., Manne, S., Ngiow, S.F., Patel, K.P., Werner, M.T., Huang, A.C., Alexander, K.A., Wu, J.E., et al. (2019). TOX transcriptionally and epigenetically programs CD8+ T cell exhaustion. Nature 571, 211–218. https://doi.org/10.1038/s41586-019-1325-x. 30. Yao, C., Sun, H.-W., Lacey, N.E., Ji, Y., Moseman, E.A., Shih, H.-Y., Heuston, E.F., Kirby, M., Anderson, S., Cheng, J., et al. (2019). Single- cell RNA-seq reveals TOX as a key regulator of CD8+ T cell persistence in chronic infection. Nat. Immunol. 20, 890–901. https://doi.org/10.1038/ s41590-019-0403-4. 1232 Molecular Cell 83, 1216–1236, April 20, 2023 Article 31. Seo, H., Chen, J., Gonza´ lez-Avalos, E., Samaniego-Castruita, D., Das, A., Wang, Y.H., Lo´ pez-Moyado, I.F., Georges, R.O., Zhang, W., Onodera, A., et al. (2019). TOX and TOX2 transcription factors cooperate with NR4A transcription factors to impose CD8+ T cell exhaustion. Proc. Natl. Acad. Sci. USA 116, 12410–12415. https://doi.org/10.1073/pnas. 1905675116. 32. Man, K., Gabriel, S.S., Liao, Y., Gloury, R., Preston, S., Henstridge, D.C., Pellegrini, M., Zehn, D., Berberich-Siebelt, F., Febbraio, M.A., et al. (2017). Transcription factor IRF4 promotes CD8+ T cell exhaustion and limits the development of memory-like T cells during chronic Infection. Immunity 47, 1129–1141.e5. https://doi.org/10.1016/j.immuni.2017.11.021. 33. Martinez, G.J., Pereira, R.M., A¨ ijo¨ , T., Kim, E.Y., Marangoni, F., Pipkin, M.E., Togher, S., Heissmeyer, V., Zhang, Y.C., Crotty, S., et al. (2015). The tran- scription factor NFAT promotes exhaustion of activated CD8+ T cells. Immunity 42, 265–278. https://doi.org/10.1016/j.immuni.2015.01.006. 34. Ghoneim, H.E., Fan, Y., Moustaki, A., Abdelsamed, H.A., Dash, P., Dogra, P., Carter, R., Awad, W., Neale, G., Thomas, P.G., et al. (2017). De novo epigenetic programs inhibit PD-1 blockade-mediated T cell rejuvenation. Cell 170, 142–157.e19. https://doi.org/10.1016/j.cell.2017.06.007. 35. Guo, A., Huang, H., Zhu, Z., Chen, M.J., Shi, H., Yuan, S., Sharma, P., Connelly, J.P., Liedmann, S., Dhungana, Y., et al. (2022). cBAF complex components and MYC cooperate early in CD8+ T cell fate. Nature 607, 135–141. https://doi.org/10.1038/s41586-022-04849-0. 36. Belk, J.A., Yao, W., Ly, N., Freitas, K.A., Chen, Y.-T., Shi, Q., Valencia, A.M., Shifrut, E., Kale, N., Yost, K.E., et al. (2022). Genome-wide CRISPR screens of T cell exhaustion identify chromatin remodeling fac- tors that limit T cell persistence. Cancer Cell 40, 768–786.e7. https://doi. org/10.1016/j.ccell.2022.06.001. 37. Loo, C.-S., Gatchalian, J., Liang, Y., Leblanc, M., Xie, M., Ho, J., Venkatraghavan, B., Hargreaves, D.C., and Zheng, Y. (2020). A genome-wide CRISPR screen reveals a role for the non-canonical nucle- osome-remodeling BAF complex in Foxp3 expression and regulatory T cell function. Immunity 53, 143–157.e8. https://doi.org/10.1016/j.im- muni.2020.06.011. 38. Zhou, M., Yuan, J., Deng, Y., Fan, X., and Shen, J. (2021). Emerging role of SWI/SNF complex deficiency as a target of immune checkpoint blockade in human cancers. Oncogenesis 10, 3. https://doi.org/10. 1038/s41389-020-00296-6. 39. Sandoval, G.J., Pulice, J.L., Pakula, H., Schenone, M., Takeda, D.Y., Pop, M., Boulay, G., Williamson, K.E., McBride, M.J., Pan, J., et al. (2018). Binding of TMPRSS2-ERG to BAF chromatin remodeling com- plexes mediates prostate oncogenesis. Mol. Cell 71, 554–566.e7. https://doi.org/10.1016/j.molcel.2018.06.040. 40. Boulay, G., Sandoval, G.J., Riggi, N., Iyer, S., Buisson, R., Naigles, B., Awad, M.E., Rengarajan, S., Volorio, A., McBride, M.J., et al. (2017). Cancer-specific retargeting of BAF complexes by a prion-like domain. Cell 171, 163–178.e19. https://doi.org/10.1016/j.cell.2017.07.036. 41. Alver, B.H., Kim, K.H., Lu, P., Wang, X., Manchester, H.E., Wang, W., Haswell, J.R., Park, P.J., and Roberts, C.W. (2017). The SWI/SNF chromatin remodelling complex is required for maintenance of lineage specific enhancers. Nat. Commun. 8, 14648. https://doi.org/10.1038/ ncomms14648. 42. Barisic, D., Stadler, M.B., (2019). Mammalian ISWI and SWI/SNF selectively mediate binding of distinct transcription factors. Nature 569, 136–140. https://doi.org/10.1038/ s41586-019-1115-5. Iurlaro, M., and Sch€ubeler, D. 43. Mashtalir, N., D’Avino, A.R., Michel, B.C., Luo, J., Pan, J., Otto, J.E., Zullow, H.J., McKenzie, Z.M., Kubiak, R.L., St Pierre, R., et al. (2018). Modular organization and assembly of SWI/SNF family chromatin re- modeling complexes. Cell 175, 1272–1288.e20. https://doi.org/10. 1016/j.cell.2018.09.032. 44. Mittal, P., and Roberts, C.W.M. (2020). The SWI/SNF complex in cancer — biology, biomarkers and therapy. Nat. Rev. Clin. Oncol. 17, 435–448. https://doi.org/10.1038/s41571-020-0357-3. Article 45. Centore, R.C., Sandoval, G.J., Soares, L.M.M., Kadoch, C., and Chan, H.M. (2020). Mammalian SWI/SNF chromatin remodeling complexes: emerging mechanisms and therapeutic strategies. Trends Genet. 36, 936–950. https://doi.org/10.1016/j.tig.2020.07.011. 46. Kadoch, C., and Crabtree, G.R. (2015). Mammalian SWI/SNF chromatin remodeling complexes and cancer: mechanistic insights gained from hu- man genomics. Sci. Adv. 1, e1500447. https://doi.org/10.1126/sciadv. 1500447. 47. Cenik, B.K., and Shilatifard, A. (2021). COMPASS and SWI/SNF com- plexes in development and disease. Nat. Rev. Genet. 22, 38–58. https://doi.org/10.1038/s41576-020-0278-0. 48. Mashtalir, N., Dao, H.T., Sankar, A., Liu, H., Corin, A.J., Bagert, J.D., Ge, E.J., D’Avino, A.R., Filipovski, M., Michel, B.C., et al. (2021). Chromatin landscape signals differentially dictate the activities of mSWI/SNF family complexes. Science 373, 306–315. https://doi.org/10.1126/science. abf8705. 49. Michel, B.C., D’Avino, A.R., Cassel, S.H., Mashtalir, N., McKenzie, Z.M., McBride, M.J., Valencia, A.M., Zhou, Q., Bocker, M., Soares, L.M.M., et al. (2018). A non-canonical SWI/SNF complex is a synthetic lethal target in cancers driven by BAF complex perturbation. Nat. Cell Biol. 20, 1410–1420. https://doi.org/10.1038/s41556-018-0221-1. 50. Kadoch, C., Hargreaves, D.C., Hodges, C., Elias, L., Ho, L., Ranish, J., and Crabtree, G.R. (2013). Proteomic and bioinformatic analysis of mammalian SWI/SNF complexes identifies extensive roles in human ma- lignancy. Nat. Genet. 45, 592–601. https://doi.org/10.1038/ng.2628. 51. St Pierre, R., Collings, C.K., Same´ Guerra, D.D., Widmer, C.J., Bolonduro, O., Mashtalir, N., Sankar, A., Liang, Y., Bi, W.L., Gerkes, E.H., et al. (2022). SMARCE1 deficiency generates a targetable mSWI/ SNF dependency in clear cell meningioma. Nat. Genet. 54, 861–873. https://doi.org/10.1038/s41588-022-01077-0. 52. Zullow, H.J., Sankar, A., Ingram, D.R., Same´ Guerra, D.D., D’Avino, A.R., Collings, C.K., Lazcano, R., Wang, W.-L., Liang, Y., Qi, J., et al. (2022). The FUS::DDIT3 fusion oncoprotein inhibits BAF complex targeting and activity in myxoid liposarcoma. Mol. Cell 82, 1737–1750.e8. https://doi. org/10.1016/j.molcel.2022.03.019. 53. Mathur, R., Alver, B.H., San Roman, A.K., Wilson, B.G., Wang, X., Agoston, A.T., Park, P.J., Shivdasani, R.A., and Roberts, C.W.M. (2017). ARID1A loss impairs enhancer-mediated gene regulation and drives colon cancer in mice. Nat. Genet. 49, 296–302. https://doi.org/ 10.1038/ng.3744. 54. Helming, K.C., Wang, X., Wilson, B.G., Vazquez, F., Haswell, J.R., Manchester, H.E., Kim, Y., Kryukov, G.V., Ghandi, M., Aguirre, A.J., et al. (2014). ARID1B is a specific vulnerability in ARID1A-mutant can- cers. Nat. Med. 20, 251–254. https://doi.org/10.1038/nm.3480. 55. Xu, G., Chhangawala, S., Cocco, E., Razavi, P., Cai, Y., Otto, J.E., Ferrando, L., Selenica, P., Ladewig, E., Chan, C., et al. (2020). ARID1A determines luminal identity and therapeutic response in estrogen-recep- tor-positive breast cancer. Nat. Genet. 52, 198–207. https://doi.org/10. 1038/s41588-019-0554-0. 56. Pan, J., McKenzie, Z.M., D’Avino, A.R., Mashtalir, N., Lareau, C.A., St Pierre, R., Wang, L., Shilatifard, A., and Kadoch, C. (2019). The ATPase module of mammalian SWI/SNF family complexes mediates subcomplex identity and catalytic activity–independent genomic targeting. Nat. Genet. 51, 618–626. https://doi.org/10.1038/s41588-019-0363-5. 57. McBride, M.J., Pulice, J.L., Beird, H.C., Ingram, D.R., D’Avino, A.R., Shern, J.F., Charville, G.W., Hornick, J.L., Nakayama, R.T., Garcia- Rivera, E.M., et al. (2018). The SS18-SSX fusion oncoprotein hijacks BAF complex targeting and function to drive synovial sarcoma. Cancer Cell 33, 1128–1141.e7. https://doi.org/10.1016/j.ccell.2018.05.002. 58. Nakayama, R.T., Pulice, J.L., Valencia, A.M., McBride, M.J., McKenzie, Z.M., Gillespie, M.A., Ku, W.L., Teng, M., Cui, K., Williams, R.T., et al. (2017). SMARCB1 is required for widespread BAF complex–mediated activation of enhancers and bivalent promoters. Nat. Genet. 49, 1613– 1623. https://doi.org/10.1038/ng.3958. ll OPEN ACCESS 59. Agarwal, R., Chan, Y.-C., Tam, C.S., Hunter, T., Vassiliadis, D., Teh, C.E., Thijssen, R., Yeh, P., Wong, S.Q., Ftouni, S., et al. (2019). Dynamic mo- lecular monitoring reveals that SWI–SNF mutations mediate resistance to ibrutinib plus venetoclax in mantle cell lymphoma. Nat. Med. 25, 119–129. https://doi.org/10.1038/s41591-018-0243-z. 60. Satpathy, A.T., Granja, J.M., Yost, K.E., Qi, Y., Meschi, F., McDermott, G.P., Olsen, B.N., Mumbach, M.R., Pierce, S.E., Corces, M.R., et al. (2019). Massively parallel single-cell chromatin landscapes of human im- mune cell development and intratumoral T cell exhaustion. Nat. Biotechnol. 37, 925–936. https://doi.org/10.1038/s41587-019-0206-z. 61. Valencia, A.M., Collings, C.K., Dao, H.T., St Pierre, R., Cheng, Y.-C., Huang, J., Sun, Z.-Y., Seo, H.-S., Mashtalir, N., Comstock, D.E., et al. (2019). Recurrent SMARCB1 mutations reveal a nucleosome acidic patch interaction site that potentiates mSWI/SNF complex chromatin remodel- ing. Cell 179, 1342–1356.e23. https://doi.org/10.1016/j.cell.2019.10.044. 62. Iurlaro, M., Stadler, M.B., Masoni, F., Jagani, Z., Galli, G.G., and Sch€ubeler, D. (2021). Mammalian SWI/SNF continuously restores local accessibility to chromatin. Nat. Genet. 53, 279–287. https://doi.org/10. 1038/s41588-020-00768-w. 63. Cortez, J.T., Montauti, E., Shifrut, E., Gatchalian, J., Zhang, Y., Shaked, O., Xu, Y., Roth, T.L., Simeonov, D.R., Zhang, Y., et al. (2020). CRISPR screen in regulatory T cells reveals modulators of Foxp3. Nature 582, 416–420. https://doi.org/10.1038/s41586-020-2246-4. 64. Henikoff, S., Henikoff, J.G., Kaya-Okur, H.S., and Ahmad, K. (2020). Efficient chromatin accessibility mapping in situ by nucleosome-tethered tagmentation. eLife 9, e63274. https://doi.org/10.7554/eLife.63274. 65. Kaya-Okur, H.S., Wu, S.J., Codomo, C.A., Pledger, E.S., Bryson, T.D., Henikoff, J.G., Ahmad, K., and Henikoff, S. (2019). CUT&Tag for efficient epigenomic profiling of small samples and single cells. Nat. Commun. 10, 1930. https://doi.org/10.1038/s41467-019-09982-5. 66. Buenrostro, J.D., Giresi, P.G., Zaba, L.C., Chang, H.Y., and Greenleaf, W.J. (2013). Transposition of native chromatin for fast and sensitive epi- genomic profiling of open chromatin, DNA-binding proteins and nucleo- some position. Nat. Methods 10, 1213–1218. https://doi.org/10.1038/ nmeth.2688. 67. Corces, M.R., Trevino, A.E., Hamilton, E.G., Greenside, P.G., Sinnott- Armstrong, N.A., Vesuna, S., Satpathy, A.T., Rubin, A.J., Montine, K.S., Wu, B., et al. (2017). An improved ATAC-seq protocol reduces back- ground and enables interrogation of frozen tissues. Nat. Methods 14, 959–962. https://doi.org/10.1038/nmeth.4396. 68. Kurachi, M., Barnitz, R.A., Yosef, N., Odorizzi, P.M., DiIorio, M.A., Lemieux, M.E., Yates, K., Godec, J., Klatt, M.G., Regev, A., et al. (2014). The transcription factor BATF operates as an essential differenti- ation checkpoint in early effector CD8+ T cells. Nat. Immunol. 15, 373–383. https://doi.org/10.1038/ni.2834. 69. Seo, H., Gonza´ lez-Avalos, E., Zhang, W., Ramchandani, P., Yang, C., Lio, C.-W.J., Rao, A., and Hogan, P.G. (2021). BATF and IRF4 cooperate to counter exhaustion in tumor-infiltrating CAR T cells. Nat. Immunol. 22, 983–995. https://doi.org/10.1038/s41590-021-00964-8. 70. Kane, L.P., Lin, J., and Weiss, A. (2002). It’s all Rel-ative: NF-kB and CD28 costimulation of T-cell activation. Trends Immunol. 23, 413–420. https://doi.org/10.1016/S1471-4906(02)02264-0. 71. Klein-Hessling, S., Muhammad, K., Klein, M., Pusch, T., Rudolf, R., Flo¨ ter, J., Qureischi, M., Beilhack, A., Vaeth, M., Kummerow, C., et al. (2017). NFATc1 controls the cytotoxicity of CD8+ T cells. Nat. Commun. 8, 511. https://doi.org/10.1038/s41467-017-00612-6. 72. Gautam, S., Fioravanti, J., Zhu, W., Le Gall, J.B., Brohawn, P., Lacey, N.E., Hu, J., Hocker, J.D., Hawk, N.V., Kapoor, V., et al. (2019). The tran- scription factor c-Myb regulates CD8+ T cell stemness and antitumor im- munity. Nat. Immunol. 20, 337–349. https://doi.org/10.1038/s41590- 018-0311-z. 73. Miller, B.C., Sen, D.R., Al Abosy, R., Bi, K., Virkud, Y.V., LaFleur, M.W., Yates, K.B., Lako, A., Felt, K., Naik, G.S., et al. (2019). Subsets of ex- hausted CD8+ T cells differentially mediate tumor control and respond Molecular Cell 83, 1216–1236, April 20, 2023 1233 ll OPEN ACCESS Article to checkpoint blockade. Nat. Immunol. 20, 326–336. https://doi.org/10. 1038/s41590-019-0312-6. stores functionality in exhausted CAR-T cells through epigenetic remodel- ing. Science 372, eaba1786. https://doi.org/10.1126/science.aba1786. 74. Abdelsamed, H.A., Moustaki, A., Fan, Y., Dogra, P., Ghoneim, H.E., Zebley, C.C., Triplett, B.M., Sekaly, R.-P., and Youngblood, B. (2017). Human memory CD8 T cell effector potential is epigenetically preserved during in vivo homeostasis. J. Exp. Med. 214, 1593–1606. https://doi. org/10.1084/jem.20161760. 88. Pauken, K.E., Sammons, M.A., Odorizzi, P.M., Manne, S., Godec, J., Khan, O., Drake, A.M., Chen, Z., Sen, D.R., Kurachi, M., et al. (2016). Epigenetic stability of exhausted T cells limits durability of reinvigoration by PD-1 blockade. Science 354, 1160–1165. https://doi.org/10.1126/ science.aaf2807. 75. King, H.W., and Klose, R.J. (2017). The pioneer factor OCT4 requires the chromatin remodeller BRG1 to support gene regulatory element function in mouse embryonic stem cells. eLife 6, e22631. https://doi.org/10.7554/ eLife.22631. 76. Huang, X., Park, K.M., Gontarz, P., Zhang, B., Pan, J., McKenzie, Z., Fischer, L.A., Dong, C., Dietmann, S., Xing, X., et al. (2021). OCT4 coop- erates with distinct ATP-dependent chromatin remodelers in naı¨ve and primed pluripotent states in human. Nat. Commun. 12, 5123. https:// doi.org/10.1038/s41467-021-25107-3. 77. Weider, M., K€uspert, M., Bischof, M., Vogl, M.R., Hornig, J., Loy, K., Kosian, T., M€uller, J., Hillg€artner, S., Tamm, E.R., et al. (2012). Chromatin-remodeling factor Brg1 is required for Schwann cell differen- tiation and myelination. Dev. Cell 23, 193–201. https://doi.org/10.1016/j. devcel.2012.05.017. 78. Wang, Z., Wang, P., Li, Y., Peng, H., Zhu, Y., Mohandas, N., and Liu, J. (2021). Interplay between cofactors and transcription factors in hemato- poiesis and hematological malignancies. Signal Transduct. Target. Ther. 6, 24. https://doi.org/10.1038/s41392-020-00422-1. 79. Takaku, M., Grimm, S.A., Shimbo, T., Perera, L., Menafra, R., Stunnenberg, H.G., Archer, T.K., Machida, S., Kurumizaka, H., and Wade, P.A. (2016). GATA3-dependent cellular reprogramming requires activation-domain dependent recruitment of a chromatin remodeler. Genome Biol. 17, 36. https://doi.org/10.1186/s13059-016-0897-0. 80. Tirosh, I., Izar, B., Prakadan, S.M., Wadsworth, M.H., Treacy, D., Trombetta, J.J., Rotem, A., Rodman, C., Lian, C., Murphy, G., et al. (2016). Dissecting the multicellular ecosystem of metastatic melanoma by single-cell RNA-seq. Science 352, 189–196. https://doi.org/10. 1126/science.aad0501. 81. Zheng, C., Zheng, L., Yoo, J.-K., Guo, H., Zhang, Y., Guo, X., Kang, B., Hu, R., Huang, J.Y., Zhang, Q., et al. (2017). Landscape of infiltrating T cells in liver cancer revealed by single-cell sequencing. Cell 169, 1342–1356.e16. https://doi.org/10.1016/j.cell.2017.05.035. 82. Zhang, L., Yu, X., Zheng, L., Zhang, Y., Li, Y., Fang, Q., Gao, R., Kang, B., Zhang, Q., Huang, J.Y., et al. (2018). Lineage tracking reveals dynamic relationships of T cells in colorectal cancer. Nature 564, 268–272. https://doi.org/10.1038/s41586-018-0694-x. 83. Guo, X., Zhang, Y., Zheng, L., Zheng, C., Song, J., Zhang, Q., Kang, B., Liu, Z., Jin, L., Xing, R., et al. (2018). Global characterization of T cells in non-small-cell lung cancer by single-cell sequencing. Nat. Med. 24, 978–985. https://doi.org/10.1038/s41591-018-0045-3. 84. Sade-Feldman, M., Yizhak, K., Bjorgaard, S.L., Ray, J.P., de Boer, C.G., Jenkins, R.W., Lieb, D.J., Chen, J.H., Frederick, D.T., Barzily-Rokni, M., et al. (2018). Defining T cell states associated with response to check- point immunotherapy in melanoma. Cell 175, 998–1013.e20. https:// doi.org/10.1016/j.cell.2018.10.038. 85. Fraietta, J.A., Lacey, S.F., Orlando, E.J., Pruteanu-Malinici, I., Gohil, M., Lundh, S., Boesteanu, A.C., Wang, Y., O’Connor, R.S., Hwang, W.-T., et al. (2018). Determinants of response and resistance to CD19 chimeric antigen receptor (CAR) T cell therapy of chronic lymphocytic leukemia. Nat. Med. 24, 563–571. https://doi.org/10.1038/s41591-018-0010-1. 86. Ahmadzadeh, M., Johnson, L.A., Heemskerk, B., Wunderlich, J.R., Dudley, M.E., White, D.E., and Rosenberg, S.A. (2009). Tumor antigen– specific CD8 T cells infiltrating the tumor express high levels of PD-1 and are functionally impaired. Blood 114, 1537–1544. https://doi.org/ 10.1182/blood-2008-12-195792. 87. Weber, E.W., Parker, K.R., Sotillo, E., Lynn, R.C., Anbunathan, H., Lattin, J., Good, Z., Belk, J.A., Daniel, B., Klysz, D., et al. (2021). Transient rest re- 1234 Molecular Cell 83, 1216–1236, April 20, 2023 89. Deng, Q., Han, G., Puebla-Osorio, N., Ma, M.C.J., Strati, P., Chasen, B., Dai, E., Dang, M., Jain, N., Yang, H., et al. (2020). Characteristics of anti- CD19 CAR T cell infusion products associated with efficacy and toxicity in patients with large B cell lymphomas. Nat. Med. 26, 1878–1887. https://doi.org/10.1038/s41591-020-1061-7. 90. Papillon, J.P.N., Nakajima, K., Adair, C.D., Hempel, J., Jouk, A.O., Karki, R.G., Mathieu, S., Mo¨ bitz, H., Ntaganda, R., Smith, T., et al. (2018). Discovery of orally active inhibitors of Brahma homolog (BRM)/ SMARCA2 ATPase activity for the treatment of Brahma related gene 1 (BRG1)/SMARCA4-mutant cancers. J. Med. Chem. 61, 10155–10172. https://doi.org/10.1021/acs.jmedchem.8b01318. 91. Farnaby, W., Koegl, M., Roy, M.J., Whitworth, C., Diers, E., Trainor, N., Zollman, D., Steurer, S., Karolyi-Oezguer, J., Riedmueller, C., et al. (2019). BAF complex vulnerabilities in cancer demonstrated via struc- ture-based Protac design. Nat. Chem. Biol. 15, 672–680. https://doi. org/10.1038/s41589-019-0294-6. 92. Xiao, L., Parolia, A., Qiao, Y., Bawa, P., Eyunni, S., Mannan, R., Carson, S.E., Chang, Y., Wang, X., Zhang, Y., et al. (2022). Targeting SWI/SNF ATPases in enhancer-addicted prostate cancer. Nature 601, 434–439. https://doi.org/10.1038/s41586-021-04246-z. 93. Vaswani, R.G., J.A., N., Millan, D.S., Schiller, S.E.R., Wilson, K.J., and Huang, D.S. (2020). Compounds and Uses Thereof. 94. Huang, Z., Zak, J., Pratumchai, I., Shaabani, N., Vartabedian, V.F., Nguyen, N., Wu, T., Xiao, C., and Teijaro, J.R. (2019). IL-27 promotes the expansion of self-renewing CD8(+) T cells in persistent viral infection. J. Exp. Med. 216, 1791–1808. https://doi.org/10.1084/jem.20190173. 95. Yee, C. (2014). The use of endogenous T cells for adoptive transfer. Immunol. Rev. 257, 250–263. https://doi.org/10.1111/imr.12134. 96. Brentjens, R.J., Davila, M.L., Riviere, I., Park, J., Wang, X., Cowell, L.G., Bartido, S., Stefanski, J., Taylor, C., Olszewska, M., et al. (2013). CD19- targeted T cells rapidly induce molecular remissions in adults with chemotherapy-refractory acute lymphoblastic leukemia. Sci. Transl. Med. 5, 177ra38. https://doi.org/10.1126/scitranslmed.3005930. 97. Sadelain, M., Brentjens, R., and Rivie` re, I. (2013). The basic principles of chimeric antigen receptor design. Cancer Discov. 3, 388–398. https:// doi.org/10.1158/2159-8290.CD-12-0548. 98. Hinrichs, C.S., Borman, Z.A., Gattinoni, L., Yu, Z., Burns, W.R., Huang, J., Klebanoff, C.A., Johnson, L.A., Kerkar, S.P., Yang, S., et al. (2011). Human effector CD8+ T cells derived from naive rather than memory sub- sets possess superior traits for adoptive immunotherapy. Blood 117, 808–814. https://doi.org/10.1182/blood-2010-05-286286. 99. Berger, C., Jensen, M.C., Lansdorp, P.M., Gough, M., Elliott, C., and Riddell, S.R. (2008). Adoptive transfer of effector CD8+ T cells derived from central memory cells establishes persistent T cell memory in pri- mates. J. Clin. Invest. 118, 294–305. https://doi.org/10.1172/JCI32103. 100. Gattinoni, L., Lugli, E., Ji, Y., Pos, Z., Paulos, C.M., Quigley, M.F., Almeida, J.R., Gostick, E., Yu, Z., Carpenito, C., et al. (2011). A human memory T cell subset with stem cell–like properties. Nat. Med. 17, 1290–1297. https://doi.org/10.1038/nm.2446. 101. Melenhorst, J.J., Chen, G.M., Wang, M., Porter, D.L., Chen, C., Collins, M.A., Gao, P., Bandyopadhyay, S., Sun, H., Zhao, Z., et al. (2022). Decade-long leukaemia remissions with persistence of CD4+ CAR T cells. Nature 602, 503–509. https://doi.org/10.1038/s41586-021-04390-6. 102. Scharping, N.E., Rivadeneira, D.B., Menk, A.V., Vignali, P.D.A., Ford, B.R., Rittenhouse, N.L., Peralta, R., Wang, Y., Wang, Y., DePeaux, K., et al. (2021). Mitochondrial stress induced by continuous stimulation Article ll OPEN ACCESS under hypoxia rapidly drives T cell exhaustion. Nat. 205–215. https://doi.org/10.1038/s41590-020-00834-9. Immunol. 22, glycolytic activity in ovarian clear cell carcinoma. Mol. Carcinog. 54, 35–49. https://doi.org/10.1002/mc.22072. 103. Vardhana, S.A., Hwee, M.A., Berisa, M., Wells, D.K., Yost, K.E., King, B., Smith, M., Herrera, P.S., Chang, H.Y., Satpathy, A.T., et al. (2020). Impaired mitochondrial oxidative phosphorylation limits the self-renewal of T cells exposed to persistent antigen. Nat. Immunol. 21, 1022–1033. https://doi.org/10.1038/s41590-020-0725-2. 117. Scharping, N.E., Menk, A.V., Moreci, R.S., Whetstone, R.D., Dadey, R.E., Watkins, S.C., Ferris, R.L., and Delgoffe, G.M. (2016). The tumor micro- environment represses T cell mitochondrial biogenesis to drive intratu- moral T cell metabolic insufficiency and dysfunction. Immunity 45, 374–388. https://doi.org/10.1016/j.immuni.2016.07.009. 104. Crawford, A., Angelosanto, J.M., Kao, C., Doering, T.A., Odorizzi, P.M., Barnett, B.E., and Wherry, E.J. (2014). Molecular and transcriptional ba- sis of CD4+ T cell dysfunction during chronic. Immunity 40, 289–302. https://doi.org/10.1016/j.immuni.2014.01.005. 105. Miggelbrink, A.M., Jackson, J.D., Lorrey, S.J., Srinivasan, E.S., Waibl- Polania, J., Wilkinson, D.S., and Fecci, P.E. (2021). CD4 T-cell exhaus- tion: does it exist and what are its roles in cancer? Clin. Cancer Res. 27, 5742–5752. https://doi.org/10.1158/1078-0432.CCR-21-0206. 106. Kelso, T.W.R., Porter, D.K., Amaral, M.L., Shokhirev, M.N., Benner, C., and Hargreaves, D.C. (2017). Chromatin accessibility underlies synthetic lethality of SWI/SNF subunits in ARID1A-mutant cancers. eLife 6, e30506. https://doi.org/10.7554/eLife.30506. 107. Wang, X., Lee, R.S., Alver, B.H., Haswell, J.R., Wang, S., Mieczkowski, (2017). J., Drier, Y., Gillespie, S.M., Archer, T.C., Wu, J.N., et al. SMARCB1-mediated SWI/SNF complex function is essential for enhancer regulation. Nat. Genet. 49, 289–295. https://doi.org/10.1038/ ng.3746. 108. Mashtalir, N., Suzuki, H., Farrell, D.P., Sankar, A., Luo, J., Filipovski, M., D’Avino, A.R., St Pierre, R., Valencia, A.M., Onikubo, T., et al. (2020). A structural model of the endogenous human BAF complex informs dis- ease mechanisms. Cell 183, 802–817.e24. https://doi.org/10.1016/j. cell.2020.09.051. 109. Pan, J., Meyers, R.M., Michel, B.C., Mashtalir, N., Sizemore, A.E., Wells, J.N., Cassel, S.H., Vazquez, F., Weir, B.A., Hahn, W.C., et al. (2018). Interrogation of mammalian protein complex structure, function, and membership using genome-scale fitness screens. Cell Syst. 6, 555– 568.e7. https://doi.org/10.1016/j.cels.2018.04.011. 110. Zhang, Q., Hresko, M.E., Picton, L.K., Su, L., Hollander, M.J., Nunez- Cruz, S., Zhang, Z., Assenmacher, C.A., Sockolosky, J.T., Garcia, K.C., et al. (2021). A human orthogonal IL-2 and IL-2Rb system enhances CAR T cell expansion and antitumor activity in a murine model of leuke- mia. Sci. Transl. Med. 13, eabg6986. https://doi.org/10.1126/sci- translmed.abg6986. 111. Aspuria, P.-J., Vivona, S., Bauer, M., Semana, M., Ratti, N., McCauley, S., Riener, R., de Waal Malefyt, R., Rokkam, D., Emmerich, J., et al. (2021). An orthogonal IL-2 and IL-2Rb system drives persistence and activation of CAR T cells and clearance of bulky lymphoma. Sci. Transl. Med. 13, eabg7565. https://doi.org/10.1126/scitranslmed.abg7565. 112. Zebley, C.C., Brown, C., Mi, T., Fan, Y., Alli, S., Boi, S., Galletti, G., Lugli, E., Langfitt, D., Metais, J.-Y., et al. (2021). CD19-CAR T cells undergo exhaustion DNA methylation programming in patients with acute lymphoblastic leukemia. Cell Rep. 37, 110079. https://doi.org/10.1016/ j.celrep.2021.110079. 113. Zhang, Z., Liu, S., Zhang, B., Qiao, L., Zhang, Y., and Zhang, Y. (2020). T cell dysfunction and exhaustion in cancer. Front. Cell Dev. Biol. 8, 17. https://doi.org/10.3389/fcell.2020.00017. 114. El-Khairi, R., and Vallier, L. (2016). The role of hepatocyte nuclear factor 1b in disease and development. Diabetes Obes. Metab. 18, 23–32. https://doi.org/10.1111/dom.12715. 115. Casemayou, A., Fournel, A., Bagattin, A., Schanstra, J., Belliere, J., Decramer, S., Marsal, D., Gillet, M., Chassaing, N., Huart, A., et al. (2017). Hepatocyte nuclear Factor-1B controls mitochondrial respiration in renal tubular cells. J. Am. Soc. Nephrol. 28, 3205–3217. https://doi. org/10.1681/ASN.2016050508. 118. Yu, Y.R., Imrichova, H., Wang, H., Chao, T., Xiao, Z., Gao, M., Rincon- Restrepo, M., Franco, F., Genolet, R., Cheng, W.C., et al. (2020). Disturbed mitochondrial dynamics in CD8+ TILs reinforce T cell exhaus- tion. Nat. Immunol. 21, 1540–1551. https://doi.org/10.1038/s41590-020- 0793-3. 119. Wang, B., Wang, M., Zhang, W., Xiao, T., Chen, C.H., Wu, A., Wu, F., Traugh, N., Wang, X., Li, Z., et al. (2019). Integrative analysis of pooled CRISPR genetic screens using MAGeCKFlute. Nat. Protoc. 14, 756–780. https://doi.org/10.1038/s41596-018-0113-7. 120. Dobin, A., Davis, C.A., Schlesinger, F., Drenkow, J., Zaleski, C., Jha, S., Batut, P., Chaisson, M., and Gingeras, T.R. (2013). STAR: ultrafast uni- versal RNA-seq aligner. Bioinformatics 29, 15–21. https://doi.org/10. 1093/bioinformatics/bts635. 121. Ramı´rez, F., Ryan, D.P., Gr€uning, B., Bhardwaj, V., Kilpert, F., Richter, A.S., Heyne, S., D€undar, F., and Manke, T. (2016). deepTools2: a next generation web server for deep-sequencing data analysis. Nucleic Acids Res. 44, W160–W165. https://doi.org/10.1093/nar/gkw257. 122. Bolger, A.M., Lohse, M., and Usadel, B. (2014). Trimmomatic: a flexible trimmer for Illumina sequence data. Bioinformatics 30, 2114–2120. https://doi.org/10.1093/bioinformatics/btu170. 123. Langmead, B., and Salzberg, S.L. (2012). Fast gapped-read alignment with Bowtie 2. Nat. Methods 9, 357–359. https://doi.org/10.1038/nmeth.1923. 124. Broad Institute. Picard Toolkit. Broad Institute, GitHub repository https:// github.com/broadinstitute/picard 125. Li, H., Handsaker, B., Wysoker, A., Fennell, T., Ruan, J., Homer, N., Marth, G., Abecasis, G., and Durbin, R.; 1000 Genome Project Data Processing Subgroup (2009). The sequence alignment/map format and SAMtools. Bioinformatics 25, 2078–2079. https://doi.org/10.1093/bioin- formatics/btp352. 126. Zhang, Y., Liu, T., Meyer, C.A., Eeckhoute, J., Johnson, D.S., Bernstein, B.E., Nusbaum, C., Myers, R.M., Brown, M., Li, W., et al. (2008). Model- based analysis of ChIP-seq (MACS). Genome Biol. 9, R137. https://doi. org/10.1186/gb-2008-9-9-r137. 127. Kuhn, R.M., Haussler, D., and Kent, W.J. (2013). The UCSC genome browser and associated tools. Brief. Bioinform. 14, 144–161. https:// doi.org/10.1093/bib/bbs038. 128. Zhu, Q., Liu, N., Orkin, S.H., and Yuan, G.-C. (2019). CUT&RUNTools: a flexible pipeline for CUT&RUN processing and footprint analysis. Genome Biol. 20, 192. https://doi.org/10.1186/s13059-019-1802-4. 129. Robinson, M.D., McCarthy, D.J., and Smyth, G.K. (2010). edgeR: a Bioconductor package for differential expression analysis of digital gene expression data. Bioinformatics 26, 139–140. https://doi.org/10. 1093/bioinformatics/btp616. 130. Subramanian, A., Tamayo, P., Mootha, V.K., Mukherjee, S., Ebert, B.L., Gillette, M.A., Paulovich, A., Pomeroy, S.L., Golub, T.R., Lander, E.S., and Mesirov, J.P. (2005). Gene set enrichment analysis: A knowledge- based approach for interpreting genome-wide expression profiles. Proc. Natl. Acad. Sci. USA 102, 15545–15550. https://doi.org/10.1073/ pnas.0506580102. 131. Love, M.I., Huber, W., and Anders, S. (2014). Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550. https://doi.org/10.1186/s13059-014-0550-8. 116. Okamoto, T., Mandai, M., Matsumura, N., Yamaguchi, K., Kondoh, H., Amano, Y., Baba, T., Hamanishi, J., Abiko, K., Kosaka, K., et al. (2015). Hepatocyte nuclear factor-1b (HNF-1b) promotes glucose uptake and 132. Opgen-Rhein, R., and Strimmer, K. (2007). Accurate ranking of differentially expressed genes by a distribution-free shrinkage approach. Stat. Appl. Genet. Mol. Biol. 6, Article9. https://doi.org/10.2202/1544-6115.1252. Molecular Cell 83, 1216–1236, April 20, 2023 1235 ll OPEN ACCESS Article 133. Sch€afer, J., and Strimmer, K. (2005). A shrinkage approach to large-scale covariance matrix estimation and implications for functional genomics. Stat. Appl. Genet. Mol. Biol. 4, Article32. https://doi.org/10.2202/1544- 6115.1175. 134. Quinlan, A.R., and Hall, I.M. (2010). BEDTools: a flexible suite of utilities for comparing genomic features. Bioinformatics 26, 841–842. https:// doi.org/10.1093/bioinformatics/btq033. 135. Zhu, L.J., Gazin, C., Lawson, N.D., Page` s, H., Lin, S.M., Lapointe, D.S., and Green, M.R. (2010). ChIPpeakAnno: a Bioconductor package to annotate ChIP-seq and ChIP-chip data. BMC Bioinformatics 11, 237. https://doi.org/10.1186/1471-2105-11-237. 136. Sheffield, N.C., and Bock, C. (2016). LOLA: enrichment analysis for genomic region sets and regulatory elements in R and Bioconductor. Bioinformatics 32, 587–589. https://doi.org/10.1093/bioinformatics/ btv612. 137. Heinz, S., Benner, C., Spann, N., Bertolino, E., Lin, Y.C., Laslo, P., Cheng, J.X., Murre, C., Singh, H., and Glass, C.K. (2010). Simple combinations of lineage-determining transcription factors prime cis-regulatory elements required for macrophage and B cell identities. Mol. Cell 38, 576–589. https://doi.org/10.1016/j.molcel.2010.05.004. 139. Friedman, J., Hastie, T., and Tibshirani, R. (2010). Regularization paths for generalized linear models via coordinate descent. J. Stat. Softw. 33, 1–22. 140. Schep, A.N., Wu, B., Buenrostro, J.D., and Greenleaf, W.J. (2017). chromVAR: inferring transcription-factor-associated accessibility from single-cell epigenomic data. Nat. Methods 14, 975–978. https://doi. org/10.1038/nmeth.4401. 141. Feinberg, A.P., Koldobskiy, M.A., and Go¨ ndo¨ r, A. (2016). Epigenetic modulators, modifiers and mediators in cancer aetiology and progres- sion. Nat. Rev. Genet. 17, 284–299. https://doi.org/10.1038/nrg.2016.13. 142. Hogg, S.J., Beavis, P.A., Dawson, M.A., and Johnstone, R.W. (2020). Targeting the epigenetic regulation of antitumour immunity. Nat. Rev. Drug Discov. 19, 776–800. https://doi.org/10.1038/s41573-020-0077-5. 143. Chen, X., Glytsou, C., Zhou, H., Narang, S., Reyna, D.E., Lopez, A., Sakellaropoulos, T., Gong, Y., Kloetgen, A., Yap, Y.S., et al. (2019). Targeting mitochondrial structure sensitizes acute myeloid leukemia to venetoclax treatment. Cancer Discov. 9, 890–909. https://doi.org/10. 1158/2159-8290.CD-19-0117. 138. Vierstra, J., Lazar, J., Sandstrom, R., Halow, J., Lee, K., Bates, D., Diegel, M., Dunn, D., Neri, F., Haugen, E., et al. (2020). Global reference mapping of human transcription factor footprints. Nature 583, 729–736. https:// doi.org/10.1038/s41586-020-2528-x. 144. Buenrostro, J.D., Wu, B., Litzenburger, U.M., Ruff, D., Gonzales, M.L., Snyder, M.P., Chang, H.Y., and Greenleaf, W.J. (2015). Single-cell chro- matin accessibility reveals principles of regulatory variation. Nature 523, 486–490. https://doi.org/10.1038/nature14590. 1236 Molecular Cell 83, 1216–1236, April 20, 2023 Article STAR+METHODS KEY RESOURCES TABLE ll OPEN ACCESS REAGENT or RESOURCE SOURCE IDENTIFIER Antibodies Brilliant Violet 605(cid:2) anti-human CD279 (PD-1) APC/Cyanine7 anti-human CD366 (Tim-3) Brilliant Violet 605(cid:2) anti-human CD197 (CCR7) APC/Cyanine7 anti-human CD45RA APC anti-human CD39 APC/Cyanine7 anti-mouse CD279 (PD-1) PerCP/Cyanine5.5 anti-mouse CD366 (Tim-3) APC anti-mouse CD62L PerCP/Cyanine5.5 anti-mouse/human CD44 PE anti-human TNF-a Alexa Fluor(cid:3) 700 anti-human IFN-g APC anti-mouse TNF-a PE/Cyanine7 anti-mouse IFN-g PE/Cyanine7 anti-human CD3 APC/Cyanine7 anti-mouse CD8a FITC anti-human CD3 PE/Cy7 anti-human CD8 Alexa Fluor(cid:3) 700 anti-human CD4 Mouse Anti-Actin (clone C4) BioLegend BioLegend BioLegend BioLegend BioLegend BioLegend BioLegend BioLegend BioLegend BioLegend BioLegend BioLegend BioLegend BioLegend BioLegend Cat# 329923; RRID: AB_11124107 Cat# 345025; RRID: AB_2565716 Cat# 353223; RRID: AB_11124325 Cat# 304127; RRID: AB_10708419 Cat# 328209; RRID: AB_1953233 Cat# 135223; RRID: AB_2563522 Cat# 119717; RRID: AB_2571934 Cat# 104411; RRID: AB_313098 Cat# 103032; RRID: AB_2076204 Cat# 502908; RRID: AB_315260 Cat# 506515; RRID: AB_961351 Cat# 506307; RRID: AB_315429 Cat# 505825; RRID: AB_1595591 Cat# 317333; RRID: AB_2561451 Cat# 100714; RRID: AB_312753 ThermoFisher Scientific Cat#11-0038-42; RRID: AB_2043831 Biolegend Biolegend Millipore Cat# 344712; RRID: AB_2044007 Cat# 317426; RRID: AB_571942 Cat# MAB1501; RRID: AB_2223041 PE/Cyanine7 anti-human CD25 BioLegend Cat# 302612; RRID: AB_314282 Rabbit Anti-IgG Donkey Polyclonal Antibody (HRP) Cytiva Cat# NA934-1ML; RRID: AB_772206 Normal Rabbit IgG Cell Signaling Technology Ca# 2729S; RRID: AB_1031062 Anti-Histone H3 (acetyl K27) Abcam Cat# ab4729; RRID: AB_2118291 Rabbit Anti-SMARCA4 (Brg1) (D1Q7F) Cell Signaling Technology Cat# 49360S; RRID: AB_2728743 Rabbit Anti-SS18 (D614Z) Rabbit Anti-ARID1A (D2A8U) Rabbit Anti-PBRM1 (91894) Rabbit Anti-HNF1B Bacterial and virus strains Cell Signaling Technology Cat# 21792S; RRID: AB_2728667 Cell Signaling Technology Cat# 12354S; RRID: AB_2637010 Cell Signaling Technology Cat# 91894S; RRID: AB_2800173 Proteintech Cat# 12533-1-AP; RRID: AB_2116758 MegaX DH10B electro-competent bacteria Life Technologies Cat# C640003 Biological samples Primary Human CD8+ T cells isolated from PBMCs (Donor 1: 24y/o, Female; Donor 2: 35y/o, Male; Donor 5: 24y/o, Male) Primary Human CD8+/CD4+ T cells isolated from PBMCs (Donor 3: 28y/o, Female; Donor 4: 19y/o, Female; Donor 6: 28y/o, Female; Donor7: <30y/o, N/A; Donor8: <30y/o, Male) Primary Human CD3+ T cells isolated from PBMCs: (Donor 9: 43y/o, Male; Donor 10: 42y/o, Female) StemCell Technologies Inc Cat# 70027 New York Blood Center https://www.nybc.org/ Indiana University N/A (Continued on next page) Molecular Cell 83, 1216–1236.e1–e12, April 20, 2023 e1 ll OPEN ACCESS Continued Article REAGENT or RESOURCE SOURCE IDENTIFIER Chemicals, peptides, and recombinant proteins Zombie Aqua(cid:2) Fixable Viability Kit Zombie Violet(cid:2) Fixable Viability Kit eBioscience Foxp3/Transcription Factor Staining Buffer Set kit BsmBI-v2 Anza Esp1 enzyme T4 DNA ligase ATP BioLegend BioLegend Cat# 423101 Cat# 423113 Life Technologies Cat# 00-5523-00 New England Biolabs Cat# R0739L Life Technologies New England Biolabs New England Biolabs Cat# IVGN0136 Cat# M0202L Cat# PO756S Plasmid safe ATP-dependent DNase Thermo Fisher Scientific Cat# NC9046399 TaKaRa Ex Taq DNA Polymerase Takara Bio Cat# RR001B Polyethylenimine (PEI) Polybrene Thermo Fisher Scientific Cat# NC1014320; CAS: 9002-98-6 Santa Cruz Biotechnology Cat# SC-134220; CAS: 28728-55-4 TaKaRa Ex Taq DNA Polymerase Takara Bio Cat# RR001B 10mM 2-Mercaptoethanol Life Technologies Cat# 21985023; CAS: 60-24-2 T4 PNK T4 ligase SIINKEFL peptide (OVA 257-264) ACBI1 AU-15330 New England Biolabs Cat# VWR #101228-174 New England Biolabs Invivogen SelleckChem Cat# M0202M Cat# vac-sin Cat# S9612; CAS: 2375564-55-7 MedChem Express LLC Cat #HY-145388; CAS: 2380274-50-8 Cas9-GFP ribonucleoprotein Integrated DNA Technologies Cat#10008100 CMP14 FHT-1015 Recombinant Human IL-2 Recombinant Human IL-7 Recombinant Murine IL-2 Recombinant Murine IL-7 Recombinant Murine IFNg APC Annexin V Critical commercial assays QIAquick Gel Extraction Kit Phusion High-Fidelity PCR Kit MinElute Reaction Cleanup Kit PureLink(cid:2) HiPure Plasmid Filter Maxiprep Kit QIAamp DNA Mini Kit PureLink(cid:2) HiPure Plasmid Filter Maxiprep Kit RNeasy Plus Mini Kit Nebnext Poly(A) mRNA Magnetic Isolation Module NEXTFLEX(cid:3) Poly(A) Beads 2.0 Nebnext Ultra II Directional RNA Library Prep Kit NEXTFLEX(cid:3) Rapid Directional RNA-Seq Kit 2.0 Qiagen MinElute Reaction clean up kit NextSeq(cid:2) 500/550 High output flow cell kit CellTiterGlo Deposited data Cut&Tag-seq, RNA-seq, and ATAC-seq scATAC-seq scATAC-seq Unprocessed blot images Courtesy of Jun Qi; Papillon et al.90 MedChemExpress(cid:3) Peprotech, Inc. Peprotech, Inc. Peprotech, Inc. Peprotech, Inc. Peprotech, Inc. BioLegend N/A Cat# HY-144896; CAS: 2368903-18-6 Cat# 200-02-1MG Cat# 200-07-50UG Cat# 212-12-50UG Cat# 212-17-50UG Cat# 315-05 Cat# 640920 Qiagen Life Technologies Qiagen Thermo Fisher Scientific Qiagen Cat# 28706X4 Cat# F553S Cat# 28204 Cat#K21001 Cat# 51304 Thermo Fisher Scientific Cat# K21001 Qiagen New England Biolabs Cat# 74136 Cat# E7490 Perkin Elmer Cat# NOVA-512991 New England Biolabs Cat# E7760 Perkin Elmer Cat# NOVA-5198-01 Qiagen Illumina Promega This study Sathpathy et al.60 Kourtis et al.22 This study Cat# 28206 Cat# 20024906 Cat# G7571 GSE212357 GSE181062 GSE181064 Mendeley data: https://doi.org/10.17632/7msw2d62r6.1 (Continued on next page) e2 Molecular Cell 83, 1216–1236.e1–e12, April 20, 2023 Article Continued ll OPEN ACCESS REAGENT or RESOURCE SOURCE IDENTIFIER Experimental models: Organisms/strains Mice (C57BL/6J) The Jackson Laboratory Jackson strain #000664; RRID: IMSR_JAX:000664 OT-1 mice (C57BL/6-Tg(TcraTcrb)1100Mjb/J) The Jackson Laboratory Jackson strain #003831; RRID: IMSR_JAX:003831 Cas9 Mice (Gt(ROSA)26Sortm1.1 (CAG-cas9*,-EGFP)Fezh/J) The Jackson Laboratory Jackson strain #024858; RRID: IMSR_JAX:024858 Rag1KO mice B6.129S7-Rag1<tm1Mom>/J The Jackson Laboratory Jackson strain #002216; RRID: IMSR_JAX:002216 Twist Bioscience https://www.twistbioscience.com/ products/oligopools Oligonucleotides Mouse sgRNAs oligo pool targeting epigenetic genes sgROSA (mouse) 5’-AACGGCTCCA CCACGCTCGG-3’ sgSMARCA4-1 (mouse, sg1553 in library) 5’- ATAATGGCCTACAAGATGT- 3’ sgSMARCA4-2 (mouse, sg1554 in library) 5’- ATTGCCCGACCACCTGCAGA- 3’ sgSMARCC1-1 (mouse, sg1563 in library) 5’- TGCCCCAAGAATGTGACACA- 3’ sgSMARCC1-2 (mouse, sg1564 in library) 5’- TTGGTGACATGCTTCCCAA- 3’ sgARID1A-1 (mouse, sg1844 in library) 5’- TGCACCACAAGCACCCAGAG- 3’ sgARID1A-2 (mouse, sg1845 in library) 5’- GTCACTGAGGAAGCGCACCA- 3’ sgARID1B-1 (mouse, sg1851 in library) 5’- GCTCAGCACCCCGTACCCCG- 3’ sgARID1B-2 (mouse, sg1853 in library) 5’- CAGCAGAGCAGCCCATACCC- 3’ sgDPF2-1 (mouse, sg0385 in library) 5’- TCAGCAGATCCAGACACAGG- 3’ sgDPF2-2 (mouse, sg0386 in library) 5’- AAGGAGTGAGACAGTACATG- 3’ sgARID2-1 (mouse, sg1854 in library) 5’- GCCGTTTAAGAAGATCCCTG- 3’ sgARID2-2 (mouse, sg1858 in library) 5’- ACCAGAGTCACTACTTTAGG- 3’ SMARCA4 sgRNAs cocktail (human) 5’-ACUCCAGACCCACCCCUGGG-3’5’- CCCUAGCCCGGGUCCCUCGC-3’5’- GUCCUGCUGAGGGCGGCCCU-3’ HNF1B sgRNAs cocktail (human) 5’-AGCCCUCGUCGCCGGACAAG-3’5’- GGCCGAGCCCGACACCAAGC-3’5’- CGGGGUCACCAAGGAGGUGC-3’ Recombinant DNA pLKO5.sgRNA.EFS.tRFP plasmid psPAX2 plasmid pMD2G plasmid Epigenetic mouse CRISPR library cloned in pLKO5.sgRNA.EFS.tRFP plasmid This study This study This study This study This study This study This study This study This study This study This study This study This study Synthego Synthego Addgene Addgene Addgene This study Anti-CD19-FMC63scFV-41BB-GFP This study N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A https://www.synthego.com/products/ crispr-kits/gene-knockout-kit https://www.synthego.com/products/ crispr-kits/gene-knockout-kit Cat #57823; RRID: Addgene_57823 Cat#12260; RRID: Addgene_12260 Cat#12259; RRID: Addgene_12259 N/A N/A (Continued on next page) Molecular Cell 83, 1216–1236.e1–e12, April 20, 2023 e3 ll OPEN ACCESS Continued REAGENT or RESOURCE Software and algorithms MAGECK STAR v2.5.2b deepTools v2.5.3 Trimmomatic v0.36 Bowtie2 v2.29 Picard v2.8.0 SAMtools v 0.1.19 MACS2 v2.1.1 software UCSC utilities CutRunTools edgeR v3.12.1 GSEA DESeq2 v1.30.1 corpcor Bedtools multiIntersectBed and mergeBed functions ChIPpeakAnno v3.17.0 LOLA v1.12.0 HOMER v4.9 GLMnet ChromVAR Other Lymphoprep(cid:2) Pan T Cell Isolation kit, Human CD8+ T Cell Isolation Kit, Human CD4+ T Cell Isolation Kit, Human CD8a+ T Cell Isolation Kit, Mouse Article SOURCE IDENTIFIER https://sourceforge.net/p/mageck/wiki/Home/. https://github.com/alexdobin/STAR; RRID: SCR_004463 https://deeptools.readthedocs.io/ en/develop/; RRID: SCR_016366 http://www.usadellab.org/cms/? page=trimmomatic; RRID: SCR_011848 https://bowtie-bio.sourceforge.net/ index.shtml; RRID: SCR_016368 https://github.com/broadinstitute/ picard; RRID:SCR_006525 https://samtools.sourceforge.net; RRID: SCR_002105 https://pypi.org/project/MACS2/; RRID: SCR_013291 https://genome.ucsc.edu/util.html; RRID: SCR_005780 https://zenodo.org/record/ 3374112#.Y22KTy2B3xs https://bioconductor.org/packages/ release/bioc/html/edgeR.html; RRID: SCR_012802 https://www.gsea-msigdb.org/ gsea/index.jsp; RRID: SCR_003199 https://bioconductor.org/packages/ release/bioc/html/DESeq2.html; RRID: SCR_015687 https://cran.r-project.org/web/ packages/corpcor/index.html https://bedtools.readthedocs.io/en/latest/ https://bioconductor.org/packages/ release/bioc/html/ChIPpeakAnno.html; RRID: SCR_012828 https://bioconductor.org/packages/ release/bioc/html/LOLA.html; RRID: SCR_006912 http://homer.ucsd.edu/homer/; RRID: SCR_010881 https://www.vierstra.org/resources/ motif_clustering https://glmnet.stanford.edu/index.html; RRID: SCR_015505 https://greenleaflab.github.io/ chromVAR/index.html MAGeCKFlute package119 Dobin et al.120 Ramı´rez et al.121 Bolger et al.122 Langmead and Salzberg.123 Picard.124 Li et al.125 Zhang et al.126 Kuhn et al.127 Zhu et al.128 Robinson et al.129 Subramanian et al.130 Love et al.131 Opgen-Rhein and Strimmer132; Sch€afer and Strimmer.133 Quinlan and Hall134 Zhu et al.135 Sheffield and Bock136 Heinz et al.137 Friedman et al.139 Schep et al.140 Miltenyi Biotec Miltenyi Biotec Miltenyi Biotec Miltenyi Biotech non-redundant archetype consensus motifs Vierstra et al.138 STEMCELL Technologies Inc Cat# 07811 Cat# 130-096-535 Cat# 130-096-495 Cat# 130-096-533 Cat #130-104-075 Cat# 11132D Cat# 11453D (Continued on next page) Human T-Activator CD3/CD28 Dynabeads Thermo Fisher Scientific Mouse T-Activator CD3/CD28 Dynabeads Thermo Fisher Scientific e4 Molecular Cell 83, 1216–1236.e1–e12, April 20, 2023 Article Continued REAGENT or RESOURCE SOURCE IDENTIFIER ll OPEN ACCESS ACK lysis buffer 1X GlutaMAX 1X Non-essential Amino Acids 1X Sodium Pyruvate Cell Stimulation Cocktail Human TruStain FcX(cid:2) Ampure beads Concanavalin A CUTANA pAG-Tn5 4%-12% Bis-Tris polyacrylamide gels Life Technologies Immobilon PVDF Transfer Membrane Millipore eBioscience Brefeldin A Solution SuperSignal West Femto Maximum Sensitivity Substrate Life Technologies Life Technologies RESOURCE AVAILABILITY Quality Biological Inc Cat# 118156101C Life Technologies Life Technologies Cat# 35050061 Cat# 11140050 Thermo Fisher Scientific Cat# MT25000CI Affymetrix Human TruStain FcX(cid:2) Thermo Fisher Scientific BioMag(cid:3)Plus Epicycpher Cat# 00-4970-93 Cat# 422302 Cat# NC9933872 Cat# 86057 Cat# 15-1117 Cat# NP0336BOX Cat# IPVH00010 Cat# 00-4506-51 Cat# 34095 Lead contact Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Cigall Kadoch (cigall_kadoch@dfci.harvard.edu). Materials availability This study did not generate new unique reagents. Data and code availability d Genomics data (CUT&Tag, ATAC-seq, and RNA-seq) have been deposited at GEO under accession number GSE212357 and are publicly available as of the date of publication. This paper also analyzes existing, publicly available data (GSE181062 and GSE181064 key resources table). Unprocessed blot images have been deposited in Mendeley (Mendeley data: https://doi.org/ 10.17632/7msw2d62r6.1). See Key Resources Table for details. d This paper does not report original code. d Any additional information required to reanalyze the data reported in this paper is available from the lead, Cigall Kadoch, upon request. EXPERIMENTAL MODEL AND SUBJECT DETAILS Primary T cells Human peripheral blood mononuclear cells (PBMCs) isolated from 19- to 43-year-old male and female healthy donors were obtained through the New York Blood Center (NYBC). These de-identified human PBMC samples were collected under an IRB-exempt pro- tocol with donors providing written consent for banking and research of their specimens. For selected experiments in which high cell numbers were needed (Figures 1, 5, and 6), CD8+ T cells isolated from healthy donors PBCMCs were purchased from StemCell (StemCell Technologies Inc, Cat# 70027). Detailed donor information is reported in the key resources table. METHOD DETAILS Human CD8+ and CD4+ T cell isolation For T cell isolation, blood samples were diluted with 1 volume of PBS 2% FBS, then the diluted blood was added dropwise to 1 vol- ume of Lymphoprep (Stemcell Technologies Inc, Cat# 07811) at room temperature. Samples were centrifuged at 800g for 20 minutes at 22 (cid:3)C, then the PBMCs layer was harvested, washed two times in PBS 2% FBS, and resuspended in PBS 2% FBS. CD8+ T cells were then purified by two subsequent rounds of isolation: first, T cells were enriched using the Pan T Cell Isolation kit (Miltenyi Biotec, Cat# 130-096-535); then, negative CD8 T cell isolation was performed with the CD8+ T Cell Isolation Kit, human (Miltenyi Biotec, Cat# 130-096-495) or negative CD4 T cell isolation was performed with the CD4+ T Cell Isolation Kit, human (Miltenyi Biotec, Cat# 130- 096-533). Both enrichment steps were performed using an AutoMACS machine. For both mouse and human cells, CD8 T cell purity Molecular Cell 83, 1216–1236.e1–e12, April 20, 2023 e5 ll OPEN ACCESS Article was assessed by FACS staining using mouse or human anti-CD3 and anti-CD8 antibodies at 1:100 dilution (PE/Cyanine7 anti-human CD3, BioLegend, # Cat317333; APC/Cyanine7 anti-mouse CD8a, BioLegend, Cat# 100714; FITC anti-human CD3, ThermoFisher Scientific, Cat#11-0038-42; PE/Cy7 anti-human CD8, Biolegend, Cat# 344712). Human CD4+ T cell purity was assessed using the A700 anti-human CD4 antibody, Biolegend, Cat# 317426 at 1:100 dilution. Mouse CD8+ T cell isolation Mouse CD8+ T cells were isolated from spleens and lymph nodes of male and female 8-12 weeks old C57BL/6J mice (Jackson strain #000664), C57BL/6-Tg(TcraTcrb)1100Mjb/J (OT-1 mice) (Jackson strain #003831), Gt(ROSA)26Sortm1.1(CAG-cas9*,-EGFP)Fezh/J (Cas9 mice) (Jackson strain #024858). Cas9-OT-1 mice were obtained by breeding OT-1 and Cas9 strains, and both Cas9 homozy- gous and heterozygous mice were used for experiments. All animals used were bred and maintained at NYU School of Medicine and all experiments were performed in accordance with the Guidelines for the Care and Use of Laboratory Animals and approved by the Institutional Animal Care and Use Committees at NYU. For T cell isolation, organs were harvested and single cell suspensions were obtained by smashing and filtering through a 40mM strainer. In spleen samples, red blood cell lysis was performed by incubation in ACK lysis buffer (Quality Biological Inc, Cat# 118156101C) for 1 minute, followed by resuspension in FACS buffer (PBS 2% FBS). Negative CD8+ T cell isolation was then performed with the CD8a+ T Cell Isolation Kit, mouse (Miltenyi Biotech, Cat #130-104- 075), following the manufacturer’s instructions, utilizing an AutoMACS machine. In vitro T cell activation and exhaustion Mouse or human CD8+ T cells were cultured in RPMI media supplemented with 10% FBS, 1% Pen/Strep, 1X GlutaMAX (Life Tech- nologies Cat# 35050061), 1X Non-essential Amino Acids (Life Technologies Cat# 11140050), 1X Sodium Pyruvate (Thermo Fisher (Life Technologies Cat# 21985023). Non-activated cells were Scientific Cat# MT25000CI) and 10mM 2-Mercaptoethanol maintained in culture for maximum 3 days in presence of 1ng/ml mouse or human IL7 (Murine IL-7, Peprotech, Inc., Cat# 217-17-50UG; Recombinant Human IL-7, Peprotech, Inc., Cat# 200-07-50UG), while activated cells were supplemented with 30U/ml mouse or human IL2 (Recombinant Murine IL-2, Peprotech, Inc., Cat# 212-12-50UG; Recombinant Human IL-2, Pepro- tech, Inc., Cat# 200-02-1MG). For in vitro T cell activation and exhaustion experiments, cells were thawed and plated at 1 million/ml in presence of Mouse T-Activator CD3/CD28 Dynabeads (Thermo Fisher Scientific, Cat# 11453D) or Human T-Activator CD3/CD28 Dynabeads (Thermo Fisher Scientific, Cat# 11132D) at 1:1 beads-to-cells ratio. After 2 days, cells were split 1:2 by adding fresh media. Beads were removed on day 3 and cells were replated at 0.5M/ml in presence of new beads at 1:1 ratio. Cells were split 1:3 on day 4 and 1:2 on day. 5. Beads were removed again on day 6 and cells were replated at 0.5 million/ml in presence of new beads at 1:1 ratio. Cells were split 1:2 on day 7 and collected on day 9. For RNA-seq, ATAC-seq and C&T profiling, cells were harvested at 0h, 3h, 24h, 48h, 72h, 6 Days and 9 Days along this protocol. For long-term experiments in the presence of inhibitors or with genetic KO lines, cells were replated at 1M/ml on day 9 in presence of new beads at 1:1 ratio, then split 1:2 on day 10. Beads were removed on day 12 or 13 and cells were replated at 1M/ml in presence of new beads at 1:1 ratio. Cells were then harvested on day 15 or 16. At every time point, alive and dead cells were counted by diluting 10ml of cell suspension with 10ul of Trypan blue and analyzed on a Countess machine (Thermo Fisher Scientific). FACS staining For surface FACS staining, cells were harvested, washed in PBS 2% FBS (FACS buffer), incubated in Fc Block Solution (Human TruStain FcX(cid:2), BioLegend Cat# 422302) for 5 minutes, then incubated for 30 minutes at 4(cid:3)C in FACS buffer with antibodies targeting the proteins of interest. Cells were washed two times in FACS buffer and diluted in FACS buffer supplemented with DAPI for live/dead cell exclusion. The following antibodies were used at 1:100 dilution unless stated otherwise: Brilliant Violet 605(cid:2) anti-human CD279 (PD-1) (BioLegend, Cat# 329923), APC/Cyanine7 anti-human CD366 (Tim-3) (BioLegend, Cat# 345025), Brilliant Violet 605(cid:2) anti-hu- man CD197 (CCR7) (BioLegend, Cat# 353223), APC/Cyanine7 anti-human CD45RA (BioLegend, Cat# 304127), APC anti-human CD39 (BioLegend, Cat# 328209), APC/Cyanine7 anti-mouse CD279 (PD-1) (BioLegend, Cat# 135223, 1:200 dilution), PerCP/ Cyanine5.5 anti-mouse CD366 (Tim-3) (BioLegend, Cat# 119717), APC anti-mouse CD62L (BioLegend, Cat# 104411, 1:400 dilution), PerCP/Cyanine5.5 anti-mouse/human CD44 (BioLegend, Cat# 103032). For intracellular cytokine profiling, cells were stimulated with Cell Stimulation Cocktail (Affymetrix, Cat# 00-4970-93) and supple- mented with Brefeldin A (eBioscience Brefeldin A Solution, Life Technologies, Cat# 00-4506-51) for 3h at 37(cid:3)C to block cytokine secretion. Cells were then harvested and stained with Zombie dyes for live/dead cell discrimination (Zombie Aqua(cid:2) Fixable Viability Kit, BioLegend, Cat# 423101 or Zombie Violet(cid:2) Fixable Viability Kit, BioLegend, Cat# 423113), following the manufacturer’s instruc- tions. Fixation and permeabilization were then performed using the eBioscience Foxp3/Transcription Factor Staining Buffer Set kit (Life Technologies, Cat# 00-5523-00), according to the manufacturer’s instructions. The following antibodies for intracellular FACS analyses were used at 1:100 dilution: PE anti-human TNF-a (BioLegend, Cat# 502908), Alexa Fluor(cid:3) 700 anti-human IFN-g (BioLegend, Cat# 506515), APC anti-mouse TNF-a (BioLegend, Cat# 506307), PE/Cyanine7 anti-mouse IFN-g (BioLegend, Cat# 505825). Annexin staining was performed with APC Annexin V (BioLegend, Cat# 640920). Samples were analyzed using a Fortessa cytometer. e6 Molecular Cell 83, 1216–1236.e1–e12, April 20, 2023 Article ll OPEN ACCESS Chromatin-focused CRISPR screen Library design For designing a chromatin-focused CRISPR library, a list of epigenetic modifiers was first compiled based on literature search.141,142 Domain-focused sgRNA sequences were then designed using the Sanjana lab software, accessible through http://guides. sanjanalab.org/#/, with the option to target protein domains selected, and expression data and average data from all tissues were used to pick and define exons. A target of 6 sgRNAs were generated per gene. 60 non-targeting sgRNAs, as well as Pdcd1 and Havcr2 sgRNAs were added as controls. A ‘G’ was added at the 5’ of every sgRNA, if not already present. The following overhangs were then added: AGGCACTTGCTCGTACGACGCGTCTCACACC – (sgRNA 20 nt) – GTTTCGAGACGATGTGGGCCCGGCACC TTAA. The final library consisted of 1928 sgRNAs targeting 310 protein-coding genes. Library cloning sgRNA sequences were cloned in the pLKO5.sgRNA.EFS.tRFP plasmid (Addgene, Cat #57823) The plasmid was a gift from Bejamin Ebert. Briefly, plasmid restriction was performed with BsmBI-v2 (NEB, Cat# R0739L) for 2h at 55(cid:3)C, then the digested plasmid was run on a 1% agarose gel and purified using the QIAquick Gel Extraction Kit (Qiagen, Cat# 28706X4). The sgRNA library was diluted to 1ng/ml in H20, then PCR amplified using the Phusion High-Fidelity PCR Kit (Life Technologies, Cat# F553S), with the following primers: Forward primer: AGGCACTTGCTCGTACGACG, Reverse primer: ATGTGGGCCCGGCACCTTAA. Two ng per reaction were used and five total (50ml) reactions were performed to ensure the maintenance of library representation. PCR conditions were the following: 30 seconds at 98 (cid:3)C, then 10 seconds at 98 (cid:3)C, 30 seconds at 53 (cid:3)C, 30 seconds at 72 (cid:3)C, for 24 cycles, then 5 minutes at 72 (cid:3)C. The PCR product was then run on a 1% agarose gel and purified using the QIAquick Gel Extraction Kit (Qiagen, Cat# 28706X4). Cloning into the library vector was then performed using Golden Gate cloning, with the following protocol: 5mg di- gested vector, 500 ng PCR insert, 5ml Anza Esp1 enzyme (Life Technologies, Cat# IVGN0136), 5ml T4 DNA ligase (New England Bio- labs, Cat# M0202L), 20ul Anza Buffer, 20 ml 10mM ATP (New England Biolabs, Cat# PO756S), and H20 to 200ul final volume. The reaction was incubated for 30 minutes at 37 (cid:3)C, then 30 minutes at 16 (cid:3)C for 25 cycles. Samples were incubated with 1ml of Plasmid safe ATP-dependent Dnase (Thermo Fisher Scientific, Cat# E3101K) and incubated at 37 (cid:3)C for 15 minutes. Reaction cleanup was then performed using the MinElute Reaction Cleanup Kit (Qiagen, Cat# 28204), and the elution product was electroporated into MegaX DH10B electro-competent bacteria (Life Technologies, Cat# C640003) using a BioRad Gene Pulser II Electroporation system. Following incubation at 37 (cid:3)C for 1h, bacteria were plated in 4x24cm square LB plates containing Ampicillin and grown at 30(cid:3)C for (cid:4)20h. The next day, bacteria were harvested from the plates and grown for 2 hours in 500ml liquid LB media with Ampicillin. Plasmid DNA was harvested using the PureLink(cid:2) HiPure Plasmid Filter Maxiprep Kit (Thermo Fisher Scientific, Cat#K21001). Library repre- sentation was checked by amplifying 200ng of library using the TaKaRa Ex Taq DNA Polymerase (Takara Bio, Cat# RR001B) for 15 cycles, and sequencing 10 million reads on a MiSeq 2 instrument, followed by alignment and QC using MAGECK. CRISPR screen For lentivirus production, HEK293T cells were plated in five 15 cm dishes (9 million cells each), in DMEM media supplemented with 10% FBS, 1% Pen/Strep and 1X Glutamax (Life Technologies, Cat# 35050061). The next day, cells were transfected using PEI (Poly- ethylenimine, Linear, Thermo Fisher Scientific, Cat# NC1014320) with the following plasmids (ug per plate): 15mg psPAX2 (Addgene, Cat#12260), 10mg pMD2G (Addgene, Cat#12259), and 20mg library plasmid. Media was changed 6h after transfection with HEK293T media, and again 24h after transfection with T cell media. 48h after transfection, the viral supernatant was filtered through a 0.45mM filter and added dropwise to mouse CD8+ T cells pre-activated for 24h, in the presence of 5 mg/ml Polybrene (Santa Cruz Biotech- nology, Cat# SC-134220). The viral supernatant from each 15 cm dish was pooled and used to transduce 16 million T cells. Spin infection was performed by centrifuging at 1500g, for 60 minutes at 32(cid:3)C. A second viral collection was performed 72h after trans- fection and a second round of spin infection was performed. After 2 days, cells were harvested, beads were removed, cells were washed two times in PBS and resuspended in media supplemented with DAPI. RFP+ cells were sorted using a SY3200 Cell Sorter. One million cells were then harvested to assess initial library representation (coverage = (cid:4)500x). Cells were then cultured as described in the ‘In vitro T cell activation and exhaustion’ section. At Day 9, cells were harvested and stained with PD1 and TIM3 antibodies and PD1 high TIM3 high or TIM3 low/high populations were sorted (1 to 2 million cells = coverage (cid:4)500-1000x). Genomic DNA was purified using the QIAamp DNA Mini Kit (Qiagen, Cat# 51304) and sgRNA sequences were amplified from genomic DNA using the TaKaRa Ex Taq DNA Polymerase (Takara Bio, Cat# RR001B). Five reactions per condition, each containing 1mg of genomic DNA, were performed to maintain library representation. The primers used were: P5 primers: equimolar mix of: For_01:AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGACGCTCTTCCGATCT CTTGTGGAAAGGACGAAACACC, For_02:AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGACGCTCTTCCGATC TACTTGTGGAAAGGACGAAACACC, For_03:AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGACGCTCTTCCGAT CTGACTTGTGGAAAGGACGAAACACC, For_04:AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGACGCTCTTCC GATCTCGACTTGTGGAAAGGACGAAACACC, For_05:AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGACGCTC TTCCGATCTACGACTTGTGGAAAGGACGAAACACC; P7 primers (indexed): IDX1:CAAGCAGAAGACGGCATACGAGATTCGCCTTGGTGACTGGAGTTCAGACGTGTGCTCTTCCGATCTCTCTACTATTCTTTC CCCTGCACTGT, IDX2:CAAGCAGAAGACGGCATACGAGATATAGCGTCGTGACTGGAGTTCAGACGTGTGCTCTTCCGATCTCTCTACTATTCTTTC CCCTGCACTGT, Molecular Cell 83, 1216–1236.e1–e12, April 20, 2023 e7 ll OPEN ACCESS Article IDX4:CAAGCAGAAGACGGCATACGAGATATTCTAGGGTGACTGGAGTTCAGACGTGTGCTCTTCCGATCTCTCTACTATTCTTTC CCCTGCACTGT, IDX8:CAAGCAGAAGACGGCATACGAGATTTGAATAGGTGACTGGAGTTCAGACGTGTGCTCTTCCGATCTCTCTACTATTCTTTC CCCTGCACTGT. PCR conditions were the following: 1 minute at 95 (cid:3)C, then 30 seconds at 95 (cid:3)C, 30 seconds at 52 (cid:3)C, 10 minutes at 72 (cid:3)C, for 22 cycles, then 10 minutes at 72 (cid:3)C. PCR product purification and size selection were performed using Ampure beads (Thermo Fisher Scientific, Cat# NC9933872), with right selection using beads at 0.4x ratio and left selection with beads at 0.6x ratio. Samples were sequenced at 10 million single-end reads each on a NextSeq500 instrument. Mouse CRISPR sgRNAs knock-out experiments For single gRNAs validations, two sgRNA sequences with the greatest efficiency per gene were identified within the CRISPR library, and individually cloned in the pLKO5.sgRNA.EFS.tRFP plasmid (Addgene, Cat #57823). Briefly, the plasmid was digested with BsmBI-v2 (NEB, Cat# R0739L) for 2h at 55(cid:3)C, then run on a 1% agarose gel, followed by isolation from the gel band using the QIA- quick Gel Extraction Kit (Qiagen, Cat# 28706X4). sgRNA oligos were phosphorylated and annealed by incubation in T4 PNK (New England Biolabs, Cat# VWR #101228-174) and T4 ligase buffer at 37(cid:3)C for 30 minutes. Temperature was then gradually decreased (-1(cid:3)C/minute) until room temperature. Oligos were diluted 1:200 in H20 and 1ml of diluted oligos were ligated with 25ng of digested vector for 1h at room temperature with T4 ligase (New England Biolabs, Cat# M0202M). Following transformation in Stbl3 cells, single colonies were analyzed by Sanger sequencing using a primer targeting the U6 promoter (sequence: GACTATCATATGCTTACCGT), expanded and purified using the PureLink(cid:2) HiPure Plasmid Filter Maxiprep Kit (Thermo Fisher Scientific, Cat#K21001). Lentiviral transduction was performed by transfecting HEK293T cells with 3.75mg psPAX2 (Addgene, Cat#12260), 2mg pMD2G (Addgene, Cat#12259), and 5mg library plasmid. Viral supernatant was collected 48h and 72h after transfection, filtered through a 0.45mM filter and used to spin-infect pre-activated mouse CD8+ T cells at 1500g, for 60 minutes at 32(cid:3)C. RFP% was assessed using Fortessa machine at 3 day intervals post-activation in parallel with the in vitro exhaustion protocol. Human CRISPR sgRNAs knock-out experiments For sgSMARCA4 and sgHNF1B KO experiments in human cells, three CTRL or SMARCA4- or HNF1B-targeting sgRNAs were de- signed and synthetized by Synthego (https://www.synthego.com/products/crispr-kits/gene-knockout-kit). The Synthego-optimized multi-sgRNA approach was used, where the three different sgRNAs were co-electroporated in one reaction to increase knock-out efficiency. The SMARCA4-targeting sgRNA sequences were the following: ACUCCAGACCCACCCCUGGG, CCCUAGCCCGG GUCCCUCGC, GUCCUGCUGAGGGCGGCCCU. The HNF1B-targeting sgRNA sequences were the following: AGCCCUCGUCGC CGGACAAG, GGCCGAGCCCGACACCAAGC, CGGGGUCACCAAGGAGGUGC. Human CD8+ T cells were activated for 48h, then beads were removed, and 200,000 T cells were electroporated with a mix con- sisting of 1.5 ug of Cas9-GFP ribonucleoprotein (Integrated DNA Technologies, Cat#10008100) and 1ug of sgRNAs, using the Neon transfection system (1,200V, Width=40, 1 pulse). After electroporation cells were plated at 1 million/ml in antibiotics-free media. After 4 hours, cells were harvested and GFP+ cells were sorted. Cells were then replated at 1 million/ml with activation beads, and expanded through the in vitro exhaustion protocol. B16-OVA in vitro exhaustion model B16-F10 and B16-F10-OVA cell lines used for co-culture mediated T cell exhaustion were a gift of Dr. Weber’s lab. For B16-T cells co- cultures, splenocytes were harvested from OT-1/Cas9 mice and cultured at a concentration of 10 million/ml in T cell media in the presence of 1mM SIINKEFL peptide (OVA 257-264, Invivogen # vac-sin). After 48h, CD8+ T cells were purified with the CD8a+ T Cell Isolation Kit, mouse (Miltenyi Biotech, Cat #130-104-075), following the manufacturer’s instructions. Purified T cells were plated on 6-well plates containing B16 or B16-OVA cells, pretreated for 24h with 1ng/ml IFNg to promote MHCI expression. T cells were passaged on new B16 or B16-OVA plates, pre-treated with IFNg, every 48 hours. For CRISPR KO experiments in this model, cells were transduced as previously described following T cell purification, then cultured on B16 or B16-OVA plates and profiled 9 days after activation. Western blots Western blots were performed as described previously.143 Briefly, proteins were isolated in RIPA buffer, quantified and loaded on 4%-12% Bis-Tris polyacrylamide gels (Thermo Fisher Scientific). Proteins were then transferred onto PVDF membranes (Millipore) and probed using the SMARCA4 antibody (Cell Signaling Technology Cat# 49360T, 1:1000 dilution), the anti-HNF1B antibody Pro- teintech Cat# 12533-1-AP, 1:1000 dilution) or the anti-Actin (Millipore Cat# MAB1501, 1:10000 dilution). Following incubation with horseradish peroxidase-conjugated secondary antibodies (GE Healthcare), chemiluminescence was assessed with ECL (Life Technologies). Cell killing assays Killing assays were performed in 96-well plates, by mixing 50000 B16 or B16-OVA cells (Target) and serial dilutions of OT-1 T cells (Effector) at Day9 of the chronic stimulation protocol, in 200 ul of RPMI media supplemented with 10% FBS, 1% Pen/Strep, e8 Molecular Cell 83, 1216–1236.e1–e12, April 20, 2023 Article ll OPEN ACCESS 1X GlutaMAX (Life Technologies Cat# 35050061), 1X Non-essential Amino Acids (Life Technologies Cat# 11140050), 1X Sodium Py- ruvate (Thermo Fisher Scientific Cat# MT25000CI) and 10mM 2-Mercaptoethanol (Life Technologies Cat# 21985023). Four replicates were seeded per condition, and controls containing B16 or B16-OVA without T cells were included. After 24h or 48h incubation at 37(cid:3)C, media was removed, wells were washed twice in 200ml of PBS, then 100ul of PBS per well were added. CellTiterGlo (Promega Cat# G7571) was then added (100ml/well), plates were incubated for 10 minutes protected from light, and measured luminescence. CD19-CAR-T cell experiments For CAR-T experiments, we generated an anti-CD19 CAR lentiviral vector incorporating a 41BB co-stimulatory domain and CD3z activation domain. The single-chain variable fragment (scFv) was derived from the murine FMC63 anti-human antibody, that has high affinity and specificity for CD19 and it is utilized in clinical trials. The complete CAR construct is driven by an EF1a promoter and contains the internal ribosome entry site (IRES)-GFP signal for cell selection. The CAR-T vector was cloned in house in the Perna Lab (Indiana University). For CAR-Ts preparation, peripheral blood was obtained from de-identified healthy human volunteers under IRB-exempt protocol with written consent for banking and research of their specimens given for each donor. Peripheral blood mono- nuclear cells (PBMCs) were isolated by density gradient centrifugation, purified using the Human Pan T Cell Isolation Kit (Miltenyi Biotec, Cat#130-096-535), stimulated with CD3/CD28 T cell activator Dynabeads (Thermo Fisher Scientific, Cat# 11132D) for 2 days and cultured in X-VIVO-15 media (Lonza Cat#BE02-060Q) supplemented with human serum (5%) and IL-2 (200U/ml). After 24 hours of activation, cells were transduced with lentiviruses encoding anti-CD19 CAR and GFP genes, in presence of polybrene. Transduction efficiencies were assessed by FACS and were ranging from 20 to 50%. Then, cells were counted and plated (0.5 million/ ml) in the presence of beads, and DMSO or PROTACS, ACBI1 or AU-15330 (100nM). The same process was repeated at 3-4 days intervals. At Day 10, beads were removed and immunophenotype was analyzed by flow-cytometry using the following markers: CCR7, CD45RA, PD1, TIM3, LAG-3, CD39 (vendors and catalogs previously stated in methods). To assess in vitro persistence, cells were incubated with fresh beads at 3-days intervals and counted at Day 13 and 16. In vivo B16-OVA killing assays For in vivo experiments, 8-12 weeks-old Rag1KO mice B6.129S7-Rag1<tm1Mom>/J (Jackson # 002216), were injected subcutane- ously with 0.5 million B16-OVA cells in 100ml PBS. The same day, mouse CD8+ OT-1 T cells were purified from 8-12 week old OT-1 mice, and activated in vitro with CD3/CD28 Dynabeads at 1:1 ratio. After 72h, beads were removed, cells were counted and plated with fresh beads at 1:1 ratio in presence of DMSO or FHT-1015 at 100nM. Cells were split 1:2 or 1:3 every day by adding fresh media and DMSO or FHT-1015. At Day 7, T cells were washed in PBS, counted and 2 million cells were injected intravenously into tumor bearing mice. Tumor growth was assessed by caliper measurement every 3 days. SMARCA4/SMARCA2 Inhibitor and Degrader Small Molecule Treatment Studies For inhibitor and small-molecule PROTAC treatments, mouse or human T cells were activated for 3 days as previously described. At Day 3, cells were counted and plated at 0.5 million /ml in presence of DMSO, 50nM or 100nM of inhibitor or PROTAC. At every sub- sequent time point (Days 6, 9, 12, 16), beads were removed and cells were replated at 0.5M/ml in presence of inhibitors or PROTACs and fresh beads at 1:1 ratio. Cells were diluted 1:2 or 1:3 in between time points by adding media containing the corresponding con- centration of inhibitor or PROTAC. The drugs used were: ACBI1 (SelleckChem, Cat#S9612), AU-15330 (MedChem Express Llc, Cat #HY-145388), CMP14 (synthesized), FHT-1015 (synthesized). RNA-seq For RNAseq experiments, 50.000 to 0.5 million cells were harvested on ice and washed in cold PBS. RNA extraction was then per- formed using the Rneasy Plus Mini Kit (Qiagen Cat#74136), following the manufacturer’s instructions. Poly-A selection was performed using the the Nebnext Poly(A) mRNA Magnetic Isolation Module (New England Biolabs Cat#E7490) for all RNAseq exper- iments, except the RNAseq upon PROTAC treatment experiment, where NEXTFLEX(cid:3) Poly(A) Beads 2.0, (Perkin Elmer Cat#NOVA- 512991) were used. Library preparation was performed using the Nebnext Ultra II Directional RNA Library Prep Kit (New England Biolabs Cat# E7760), or NEXTFLEX(cid:3) Rapid Directional RNA-Seq Kit 2.0 (Perkin Elmer Cat#NOVA-5198-01) for PROTAC experi- ments. For all libraries, quality was assessed by Tapestation. Samples were sequenced on NovaSeq6000 and NextSeq500 machines (Illumina) at sequencing depth of 30 million reads per sample. ATAC-seq Cells were harvested at 0h, 3h, 24h, 48h, 72h, Day6 and Day9. ATAC-seq experiments were completed and samples were prepared into libraries using the previously described methodology.66,67,144 Cells (50,000) were collected in media and washed in cold PBS. Cells were spun at 500rcf for 5 minutes to form a pellet and PBS was removed. Cold lysis buffer was added and cells were gently resuspended by pipetting. Resuspended cells were incubated on ice for three minutes. Lysis was quenched by adding wash buffer and mixing by inverting the tube three times. Lysed material was pelleted at 400 rcf for 10 minutes, and supernatant was discarded. The pelleted DNA was resuspended in transposition reaction buffer and the transposition reaction was carried out for 30 minutes at 37(cid:3)C with gentle shaking at 1,000 rpm on a thermomixer. The resultant tagmented DNA was purified using Qiagen MinElute Reaction clean up kit (Qiagen Cat# 28206) and eluted in dH20. Tagmented DNA libraries were amplified with 7 total cycles using a standard Molecular Cell 83, 1216–1236.e1–e12, April 20, 2023 e9 ll OPEN ACCESS Article ATAC-seq amplification protocol and custom PCR primers. ATAC-seq libraries were sequenced on the Illumina NextSeq500 with 35 base-pair paired end sequencing parameters and using the NextSeq(cid:2) 500/550 High output flow cell kit (Illumina Cat# 20024906). Cut & Tag The epicypher protocol for Cleavage under targets and tagmentation was used with slight modifications.65 Concanavalin A (ConA, BioMag(cid:3)Plus Cat# 86057) beads were activated with bead activation buffer and stored on ice until further use. Cells (100,000) were collected and washed with cold PBS. Cells were spun at 300rcf for 5 minutes at 4(cid:3)C and PBS supernatant was removed from the cell pellet. Nuclear extraction buffer was added to the tube and the pellet was gently resuspended by pipetting to lyse cells and extract nuclei. Activated ConA beads and nuceli were incubated were mixed and incubated at room temperature for 10 minutes. The nuclei- conjugated bead complexes were resuspended in antibody binding buffer and add primary antibody rotating on a nutator overnight at 4(cid:3)C (Added 2.5, 0.5, 0.25, 0.25, 0.67 and 0.25 ug of IgG, kH3K27Ac, Brg1, SS18, ARID1A, and PBRM1 respectively). Primary anti- body mixture was removed, and nuclei-bead complexes were incubated with 0.5mg secondary antibody in digitonin 150 buffer for 1 hour at room temperature on the nutator. After secondary antibody incubation, samples were washed with digitonin 150 buffer and resuspended in digitonin 300 buffer supplemented with 2 microliters of CUTANA pAG-Tn5 (Epicycpher Cat#15-1117) added per sample. Samples were incubated with Tn5 for 1 hour at room temperature on the nutator. Digitonin 300 buffer was added two times to remove excess enzyme from samples. Targeted chromatin tagmentation was completed following the epicypher protocol. Li- braries were amplified with 14 PCR cycles and purified by single sided 1.3x AMPure bead purification. The NextSeq500 and 35 base-pair paired end sequencing parameters were used for library sequencing. QUANTIFICATION AND STATISTICAL ANALYSIS NGS Data Processing Cut and Tag, ATAC-Seq, and RNA-Seq samples were sequenced with the Illumina technology, and output data were demultiplexed using the bcl2fastq software tool. RNA-Seq reads were aligned to the hg19 genome with STAR v2.5.2b,120 and tracks were generated using the deepTools v2.5.3 bamCoverage function.121 For ATAC-Seq data, quality read trimming was carried out by Trimmomatic v0.36,122 followed by alignment, duplicate read removal, and read quality filtering using Bowtie2 v2.29,123 Picard v2.8.0,124 and SAM- tools v 0.1.19,125 respectively, and ATAC-seq broad peaks were called with the MACS2 v2.1.1 software126 using the BAMPE option and a broad peak cutoff of 0.001. For ATAC-Seq track generation, output BAM files were converted into BigWig files using MACS2 and UCSC utilities127 in order to display coverage throughout the genome in RPM values. For Cut and Tag libraries, the CutRunTools pipeline was leveraged to perform read trimming, quality filtering, alignment, peak calling, and track building using default parame- ters.128 All sequencing data analyzed in this study have been deposited at NCBI’s Gene Expression Omnibus under accession num- ber GSE212357. RNA-seq data analysis For RNA-seq data, output gene count tables from STAR based on alignments to the hg19 reflat annotation were used as input into edgeR v3.12.1129 to obtain normalized log CPM values and to evaluate differential gene expression. Log2 fold change values from edgeR were used as input into GSEA,130 and the GseaPreranked tool was run with default settings to measure gene set enrichment. For day 9 treatment comparisons, the DESeq2 v1.30.1 R software package was used to evaluate differential gene expression.131 For heatmaps displaying differentially expressed genes, log CPMs were transformed into Z-scores followed by hierarchical or K-means clustering. In order analyze gene set enrichment for select subsets of genes, hypergeometric tests were performed on overlaps with various MSIGDB gene sets and select enriched gene sets were displayed. Principal component analysis was performed using the wt.scale and fast.svd functions from the corpcor R package on RPKM values,132,133 which were quantified using median length iso- forms and total mapped read counts computed by the Samtools idxstats function. Cut & Tag and ATAC-seq data analysis and integration The Bedtools multiIntersectBed and mergeBed functions were used for peak merging,134 and the R package, ChIPpeakAnno v3.17.0,135 was used to visualize peak overlaps. Distance-to-TSS peak distributions were computed utilizing Ensembl protein-cod- ing gene coordinates. To generate the heatmap in Figure 1D, which served as a platform for several downstream analyses, first, 32 sets of peaks derived from the SS18 and SMARCA4 samples (called by the CutRunTools pipeline) from both donors and from all time points were merged with the Bedtools multiIntersectBed function. Since there wsere two donors for every time point, we removed any peaks or parts of peaks that did not overlap with at least one peak to remove outlier peaks and outlier peak segments. This over- lap information from the 32 sets of peaks was provided by the output of the Bedtools multiIntersectBed function. After the outlier peaks and outlier peak segments were removed, the Bedtools mergeBed function was used to merge the filtered peaks. In an iden- tical manner, the peaks of the H3K27ac and ATAC-seq samples were separately merged. Second, these three sets of merged peaks from all time points were overlapped and merged to generate the Venn Diagram in Figure S2I, which represent the sites in the Fig- ure 1D heatmap. Third, the Cut & Tag and ATAC-seq RPKM data from the merged peaks were log2 transformed, followed by the separate quantile normalization of each BAF subunit, H3K27ac, and ATAC-seq data across time points. Finally, K-means clustering was applied in a semi-unsupervised manor to partition the SMARCA4, SS18, H3K27ac and ATAC-seq data into the 9 groups or e10 Molecular Cell 83, 1216–1236.e1–e12, April 20, 2023 Article ll OPEN ACCESS clusters, which are exhibited in Figure 1D, followed by transformation into Z-scores across timepoints to highlight the differences within clusters among time points. IgG control subtraction and data normalization CUT&TAG datasets from two independent human T cell donors were merged and RPKM values were computed for SMARCA4, SS18, H3K27Ac and IgG samples for each timepoint across the T cell activation and exhaustion time course (as above). The RPKM values for each mark were log2-transformed and individually subjected to quantile normalization. Quantile-normalized IgG sig- nals were then subtracted from the SMARCA4, SS18, and H3K27Ac quantile-normalized signals. From here, for selected figure panels, quantile-normalized, IgG-subtracted signal values were separately transformed into Z-scores for all timepoints. Finally, IgG control peaks that were present and overlapped SMARCA4, SS18 and H3K27Ac merged sites were removed, while maintaining the order of the non-overlapping sites. Heatmaps were generated from the resultant non-overlapping and IgG-subtracted sites. Principal component analyses (PCA) were also performed on these quantile-normalized log2-transformed RPKM values. For day 9 treatment analyses, DESeq2 was used to evaluate differential accessibility, and quantile-normalized log2 RPKM values were trans- formed into Z-scores following by hierarchical clustering to display differentially accessible sites. Transcription factor motif and archetype analyses Transcription factor enrichment and motif analyses were carried out by the LOLA v1.12.0136 and HOMER v4.9137 software packages, respectively. In addition to using HOMER to analyze motif enrichment, for several motif enrichment analyses conducted in this study, we determined the number of motif occurrences for 286 non-redundant archetype consensus motifs138 within +/- 250 base pairs of peak centers for each peak within given peak sets. The coordinates of these archetype motifs as well as non-archetype motifs across the entire human and mouse genomes can be downloaded from the following resource https://www.vierstra.org/resources/ motif_clustering#downloads. Average archetype and non-archetype motif occurrences and densities within all sites and within clusters of sites were also deter- mined, and fractional enrichment values relative to all sites were displayed for select motifs with high occurrence and variability. The following formulas were used to compute motif densities and enrichment. 1: Site Motif ‘X0 Occurrence = # of Motif X Counts within + = (cid:1) 250 base pairs within center of Site 2: Total Motif ‘X0 Density = ½Total Sum of Motif X Counts at All Sites(cid:5) = ½Total # of All Sites(cid:5) 3: Cluster ‘Y0 Motif ‘X0 Density = ½Total Sum of Motif X Counts at Cluster Y Sites(cid:5) = ½Total # of Cluster ‘Y0 Sites(cid:5) 4: Cluster ‘Y0 Motif ‘X0 Density Difference = = ½Cluster ‘Y0 Motif ‘X0 Density(cid:5) -- ½Total Motif ‘X0 Density(cid:5) 5: Cluster ‘Y0 Motif ‘X0 Fractional Enrichment = ½Cluster ‘Y0 Motif ‘X0 Density Difference(cid:5) = ½Total Motif ‘X0 Density(cid:5) These archetype motif fraction enrichment values in clusters were also plotted against corresponding TF gene log fold change values for several stepwise comparisons across the T-cell activation and exhaustion time course. For logistic regression analyses Molecular Cell 83, 1216–1236.e1–e12, April 20, 2023 e11 ll OPEN ACCESS Article on archetype motifs, matrices of motif counts for given merged peaks were generated, and the R software program GLMnet139 was used to produce logistic regression models to fit the motif counts to binomial vectors where 1’s represented sites with BAF or ATACseq log2 fold changes > 0. Divergent barplots were used to display large coefficients in output models to estimate the influence of motifs on BAF occupancy or accessibility in terms of magnitude, directionality and predictability. CRISPR screen data analysis Data analysis was performed using MAGECK,119 following the standard analysis pipeline reported in https://sourceforge.net/p/ mageck/wiki/Home/. Data visualization was performed using the Bioconductor MAGeCKFlute package. scRNAseq and scATACseq datasets and analyses scRNA-seq scRNAseq signatures consisted of the marker genes identified in exhausted or memory cells from the several literature sources.80–84 These gene lists were used as ‘‘gene set’’ inputs into GSEA along with log2 fold change values from edgeR for all expressed genes for several given comparisons, and the GseaPreranked tool was run with default settings to measure gene set enrichment.130 A positive score indicates an enrichment of genes within a given gene set that have increasing expression, while a negative score indicates an enrichment of genes within a given gene set that have decreasing expression. GSEA output normalized enrichment scores or regular enrichment scores were displayed in heatmaps. Negative log base 10 p-values from select the indicated gene sets were displayed in barplots. scATAC-seq To assess the TF activity toward SWI/SNF bound regions in tumor infiltrating human CD8+ T cells, we used publicly available scATAC-seq datasets (GSE181062, GSE181064).22,60 From the raw data, we used read counts that fall into the peak coordinate defined as SWI/SNF bound regions in this study. We then used the Viestra et al. motif database as an input of the ‘motifmatchr’ func- tion of ChromVAR and followed the ChromVAR workflow with default settings.140 ‘deviationScores’ is used for visualization of TF activity and cell type annotations in the original papers were used. e12 Molecular Cell 83, 1216–1236.e1–e12, April 20, 2023 Molecular Cell, Volume 83 Supplemental information Stepwise activities of mSWI/SNF family chromatin remodeling complexes direct T cell activation and exhaustion Elena Battistello, Kimberlee A. Hixon, Dawn E. Comstock, Clayton K. Collings, Xufeng Chen, Javier Rodriguez Hernaez, Soobeom Lee, Kasey S. Cervantes, Madeline M. Hinkley, Konstantinos Ntatsoulis, Annamaria Cesarano, Kathryn Hockemeyer, W. Nicholas Haining, Matthew T. Witkowski, Jun Qi, Aristotelis Tsirigos, Fabiana Perna, Iannis Aifantis, and Cigall Kadoch A B Human CD8+ T cells 10 TRANSIENT 10 CHRONIC Figure S1 i n o s n a p x e d o F l 8 6 4 2 i n o s n a p x e d o F l 8 6 4 2 0 Days: 0 3 6 9 0 Days: 0 3 6 9 3h 0 0h 0 CD25+ 250K 200K 150K 100K 50K 0 24h 94 250K 200K 150K 100K 50K 0 250K 200K 150K 100K 50K 0 48h 98 250K 200K 150K 100K 50K 0 72h 95 6 Days 9 Days Chr 9 Days Tr 99 250K 200K 150K 100K 50K 0 94 250K 200K 150K 100K 50K 0 70 250K 200K 150K 100K 50K 0 3 -10 0 3 10 4 10 5 10 3 -10 0 3 10 4 10 5 10 3 -10 0 3 10 4 10 5 10 3 -10 0 3 10 4 10 5 10 3 -10 0 3 10 4 10 5 10 3 -10 0 3 10 4 10 5 10 3 -10 0 3 10 4 10 5 10 0 3 10 4 10 5 10 250K 200K 150K 100K 50K 0 - A C S F CD25 0h 3h 24h 48h 72h 6 Days 9 Days Chr 9 Days Tr CM N 1 68 1 75 8 81 17 73 20 63 90 17 0 32 48 4 10 3 10 0 4 10 3 10 0 4 10 3 10 0 4 10 3 10 0 4 10 3 10 0 4 10 3 10 0 22 9 12 12 3 7 4 5 9 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 0 4 10 8 5 10 6 4 0 4 10 3 10 0 82 4 10 3 10 0 1 5 10 15 5 0 4 10 5 10 0 4 10 5 10 0 4 10 CD45RA 250K 200K 150K 100K 50K 0 0h 1 CD39+ 250K 200K 150K 100K 50K 0 3h 1 24h 2 250K 200K 150K 100K 50K 0 250K 200K 150K 100K 50K 0 48h 7 250K 200K 150K 100K 50K 0 72h 10 6 Days 9 Days Chr 9 Days Tr 90 250K 200K 150K 100K 50K 0 98 250K 200K 150K 100K 50K 0 3 250K 200K 150K 100K 50K 0 0 3 10 4 10 5 10 0 3 10 4 10 5 10 0 3 10 4 10 5 10 0 3 10 4 10 5 10 0 3 10 4 10 5 10 0 3 10 4 10 5 10 0 3 10 4 10 5 10 0 3 10 4 10 5 10 CD39 5 10 4 10 3 10 0 3 -10 5 87 0 IFNγ 250K 200K 150K 100K 50K 0 - A C S F 0h 3h 24h 48h 72h 6 Days 9 Days Chr 9 Days Tr 7 1 4 10 5 10 0h 0 GZMb+ 4 87 0 5 10 4 10 3 10 0 3 -10 250K 200K 150K 100K 50K 0 7 2 4 10 5 10 3h 0 30 17 47 0 6 4 10 5 10 48h 74 41 21 34 0 4 4 10 5 10 24h 55 5 10 4 10 3 10 0 3 -10 250K 200K 150K 100K 50K 0 5 10 4 10 3 10 0 3 -10 250K 200K 150K 100K 50K 0 6 3 77 0 13 4 10 5 10 6 Days 49 28 10 57 0 5 4 10 5 10 72h 70 5 10 4 10 3 10 0 3 -10 250K 200K 150K 100K 50K 0 5 10 4 10 3 10 0 3 -10 250K 200K 150K 100K 50K 0 5 10 4 10 3 10 0 3 -10 250K 200K 150K 100K 50K 0 12 8 9 74 5 10 4 10 3 10 0 3 -10 9 8 69 0 10 4 10 5 10 0 4 10 5 10 9 Days Chr 69 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 7 R C C EM EMRA CD45RA 7 R C C - A C S F α F N T GZMb C Mouse CD8+ T cells TRANSIENT 5 4 3 2 1 i n o s n a p x e d o F l i n o s n a p x e d o F l 5 4 3 2 1 0 Days: 0 3 6 9 0 Days: 0 3 6 9 CHRONIC D Mouse CD8+ T cells 0h 0 2 5 10 4 10 3 10 0 3 -10 72h 0 71 29 D9-chr 81 1 8 5 10 4 10 3 10 0 3 -10 5 10 4 10 3 10 0 3 -10 D9-tr 4 72 16 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 5 10 4 10 3 10 0 3 -10 3 M T I PD1 98 Figure S1. Establishment and characterization of human and mouse CD8+ T cell activation and exhaustion using cell culture systems, Related to Figure 1. A. Fold expansion for human CD8+ T cells in the chronic or transient stimulation conditions across the Day 0, 3, 6, and 9 time points. Bar graphs represent mean ± SEM from 3-4 independent CD8+ T cell donors. B. FACS plots depicting the profiling of CD25, CD45RA and CCR7, CD39, IFNγ, TNFα and GZMb across the activation and exhaustion time course. C. Fold expansion for mouse CD8+ T cells in the chronic or transient stimulation conditions across the no stim, and Day 3, 6, and 9 time points. Bar graphs represent mean ± SEM from 3-5 independent mice. D. FACS-based profiling of PD1 and TIM3 markers indicating putative naïve/memory, activated, and exhausted T cell populations in mouse CD8+ T cell studies. A 20000 10000 t n u o c k a e P 0 B Antibody SMARCA4 SS18 ARID1A PBRM1 H3K27ac H 0 . 1 D H 0 . 2 D H 3 . 1 D H 3 . 2 D H 4 2 . 1 D H 4 2 . 2 D H 8 4 . 1 D H 8 4 . 2 D H 2 7 . 1 D H 2 7 . 2 D 6 y a D . 1 D 6 y a D . 2 D h c 9 y a D . 1 D h c 9 y a D . 2 D r t 9 y a D . 1 D r t 9 y a D . 2 D H 0 . 1 D H 0 . 2 D H 3 . 1 D H 3 . 2 D H 4 2 . 1 D H 4 2 . 2 D H 8 4 . 1 D H 8 4 . 2 D H 2 7 . 1 D H 2 7 . 2 D 6 y a D . 1 D 6 y a D . 2 D h c 9 y a D . 1 D h c 9 y a D . 2 D r t 9 y a D . 1 D r t 9 y a D . 2 D H 0 . 1 D H 0 . 2 D H 3 . 1 D H 3 . 2 D H 4 2 . 1 D H 4 2 . 2 D H 8 4 . 1 D H 8 4 . 2 D H 2 7 . 1 D H 2 7 . 2 D H 0 . 1 D 6 y a D . 1 D 6 y a D . 2 D h c 9 y a D . 1 D h c 9 y a D . 2 D r t 9 y a D . 1 D r t 9 y a D . 2 D . . 1 p e R H 0 . 2 D 2 p e R H 0 . 2 D H 3 . 1 D H 3 . 2 D H 4 2 . 1 D H 4 2 . 2 D H 8 4 . 1 D H 8 4 . 2 D H 2 7 . 1 D H 2 7 . 2 D 6 y a D . 1 D 6 y a D . 2 D h c 9 y a D . 1 D h c 9 y a D . 2 D r t 9 y a D . 1 D r t 9 y a D . 2 D H 0 . 1 D H 0 . 2 D H 3 . 1 D H 3 . 2 D H 4 2 . 1 D H 4 2 . 2 D H 8 4 . 1 D H 8 4 . 2 D H 2 7 . 1 D H 2 7 . 2 D 6 y a D . 1 D 6 y a D . 2 D h c 9 y a D . 1 D h c 9 y a D . 2 D r t 9 y a D . 1 D r t 9 y a D . 2 D C SMARCA4 R-value 1 0.8 0.6 0.4 0.2 0 SS18 R-value 1 0.8 0.6 0.4 0.2 0 ARID1A R-value 1 0.8 0.6 0.4 0.2 0 D Sample 1 0.57 0.51 0.56 0.51 0.66 0.6 0.63 0.64 0.6 0.47 0.42 0.53 0.48 0.56 0.5 0.57 1 0.54 0.54 0.49 0.55 0.57 0.6 0.59 0.61 0.38 0.37 0.43 0.43 0.5 0.48 0.51 0.54 1 0.51 0.51 0.45 0.44 0.48 0.46 0.52 0.25 0.26 0.33 0.32 0.43 0.4 0.56 0.54 0.51 1 0.53 0.51 0.45 0.48 0.53 0.5 0.31 0.25 0.38 0.32 0.46 0.38 0.51 0.49 0.51 0.53 1 0.46 0.38 0.42 0.47 0.46 0.25 0.2 0.33 0.25 0.44 0.34 0.66 0.55 0.45 0.51 0.46 1 0.75 0.75 0.72 0.64 0.66 0.61 0.71 0.67 0.65 0.61 0.6 0.57 0.44 0.45 0.38 0.75 1 0.85 0.74 0.72 0.68 0.69 0.7 0.76 0.6 0.67 0.63 0.6 0.48 0.48 0.42 0.75 0.85 1 0.75 0.74 0.65 0.66 0.68 0.73 0.6 0.65 0.64 0.59 0.46 0.53 0.47 0.72 0.74 0.75 1 0.71 0.58 0.52 0.63 0.61 0.6 0.56 0.6 0.61 0.52 0.5 0.46 0.64 0.72 0.74 0.71 1 0.5 0.5 0.55 0.58 0.56 0.58 0.47 0.38 0.25 0.31 0.25 0.66 0.68 0.65 0.58 0.5 1 0.83 0.78 0.8 0.66 0.72 0.42 0.37 0.26 0.25 0.2 0.61 0.69 0.66 0.52 0.5 0.83 1 0.76 0.85 0.62 0.74 0.53 0.43 0.33 0.38 0.33 0.71 0.7 0.68 0.63 0.55 0.78 0.76 1 0.81 0.64 0.64 0.48 0.43 0.32 0.32 0.25 0.67 0.76 0.73 0.61 0.58 0.8 0.85 0.81 1 0.6 0.71 0.56 0.5 0.43 0.46 0.44 0.65 0.6 0.6 0.6 0.56 0.66 0.62 0.64 0.6 1 0.67 0.5 0.48 0.4 0.38 0.34 0.61 0.67 0.65 0.56 0.58 0.72 0.74 0.64 0.71 0.67 1 Day9ch.D1 0H.D2 3H.D2 0H.D1 3H.D1 Day6.D1 Day6.D2 Day9ch.D2 Day9tr.D1 Day9tr.D2 48H.D1 48H.D2 72H.D1 72H.D2 24H.D1 24H.D2 1 0.55 0.51 0.51 0.35 0.24 0.25 0.3 0.5 0.47 0.55 0.48 0.45 0.4 0.49 0.5 0.55 1 0.57 0.57 0.44 0.34 0.34 0.37 0.54 0.63 0.67 0.52 0.49 0.53 0.6 0.63 0.51 0.57 1 0.55 0.33 0.27 0.27 0.27 0.52 0.55 0.56 0.48 0.4 0.47 0.5 0.5 0.51 0.57 0.55 1 0.37 0.32 0.32 0.31 0.54 0.57 0.57 0.53 0.45 0.5 0.52 0.52 0.35 0.44 0.33 0.37 1 0.84 0.8 0.83 0.47 0.55 0.51 0.62 0.7 0.73 0.66 0.66 0.24 0.34 0.27 0.32 0.84 1 0.86 0.81 0.44 0.51 0.41 0.6 0.63 0.74 0.56 0.57 0.25 0.34 0.27 0.32 0.8 0.86 1 0.86 0.43 0.52 0.41 0.69 0.72 0.73 0.56 0.58 0.3 0.37 0.27 0.31 0.83 0.81 0.86 1 0.42 0.48 0.43 0.65 0.76 0.69 0.59 0.59 0.5 0.54 0.52 0.54 0.47 0.44 0.43 0.42 1 0.61 0.59 0.57 0.5 0.63 0.6 0.61 0.47 0.63 0.55 0.57 0.55 0.51 0.52 0.48 0.61 1 0.7 0.59 0.54 0.71 0.69 0.72 0.55 0.67 0.56 0.57 0.51 0.41 0.41 0.43 0.59 0.7 1 0.56 0.54 0.61 0.68 0.71 0.48 0.52 0.48 0.53 0.62 0.6 0.69 0.65 0.57 0.59 0.56 1 0.73 0.67 0.59 0.6 0.45 0.49 0.4 0.45 0.7 0.63 0.72 0.76 0.5 0.54 0.54 0.73 1 0.63 0.62 0.61 0.4 0.53 0.47 0.5 0.73 0.74 0.73 0.69 0.63 0.71 0.61 0.67 0.63 1 0.74 0.76 0.49 0.6 0.5 0.52 0.66 0.56 0.56 0.59 0.6 0.69 0.68 0.59 0.62 0.74 1 0.76 0.5 0.63 0.5 0.52 0.66 0.57 0.58 0.59 0.61 0.72 0.71 0.6 0.61 0.76 0.76 1 3H.D2 0H.D2 0H.D1 3H.D1 72H.D2 72H.D1 48H.D1 48H.D2 Day9ch.D1 Day9tr.D1 Day9tr.D2 24H.D1 24H.D2 Day6.D1 Day6.D2 Day9ch.D2 D a y 9 c h . D 1 0 H D 2 . 3 H D 2 . 0 H D 1 . 3 H D 1 . D a y 6 . D 1 D a y 6 . D 2 D a y 9 t r . D a y 9 t r . D 1 D 2 D a y 9 c h . D 2 4 8 H D 1 . 4 8 H D 2 . 7 2 H D 1 . 7 2 H D 2 . 2 4 H D 1 . 2 4 H D 2 . 3 H D 2 . 0 H D 2 . 0 H D 1 . 3 H D 1 . 7 2 H D 2 . 7 2 H D 1 . 4 8 H D 1 . 4 8 H D 2 . D a y 9 t r . D a y 9 t r . D 1 D 2 D a y 9 c h . D 1 2 4 H D 1 . 2 4 H D 2 . D a y 6 . D 1 D a y 6 . D 2 D a y 9 c h . D 2 H3K27ac 1 0.52 0.48 0.47 0.38 0.43 0.28 0.3 0.26 0.48 0.45 0.45 0.51 0.44 0.49 0.49 0.52 1 0.48 0.47 0.36 0.41 0.26 0.26 0.23 0.48 0.45 0.42 0.48 0.41 0.47 0.45 0.48 0.48 1 0.51 0.33 0.38 0.27 0.29 0.2 0.5 0.47 0.46 0.47 0.44 0.44 0.45 0.47 0.47 0.51 1 0.3 0.35 0.23 0.25 0.18 0.48 0.45 0.42 0.44 0.41 0.41 0.4 0.38 0.36 0.33 0.3 1 0.7 0.6 0.75 0.76 0.47 0.62 0.52 0.55 0.59 0.58 0.61 0.43 0.41 0.38 0.35 0.7 1 0.65 0.73 0.74 0.51 0.58 0.57 0.58 0.66 0.65 0.65 0.28 0.26 0.27 0.23 0.6 0.65 1 0.7 0.68 0.42 0.5 0.49 0.45 0.58 0.52 0.54 0.3 0.26 0.29 0.25 0.75 0.73 0.7 1 0.83 0.46 0.6 0.54 0.5 0.66 0.57 0.6 0.26 0.23 0.2 0.18 0.76 0.74 0.68 0.83 1 0.39 0.53 0.48 0.47 0.6 0.55 0.59 0.48 0.48 0.5 0.48 0.47 0.51 0.42 0.46 0.39 1 0.54 0.55 0.56 0.59 0.57 0.57 0.45 0.45 0.47 0.45 0.62 0.58 0.5 0.6 0.53 0.54 1 0.56 0.55 0.59 0.57 0.57 0.45 0.42 0.46 0.42 0.52 0.57 0.49 0.54 0.48 0.55 0.56 1 0.65 0.63 0.62 0.67 0.51 0.48 0.47 0.44 0.55 0.58 0.45 0.5 0.47 0.56 0.55 0.65 1 0.6 0.66 0.69 0.44 0.41 0.44 0.41 0.59 0.66 0.58 0.66 0.6 0.59 0.59 0.63 0.6 1 0.65 0.68 0.49 0.47 0.44 0.41 0.58 0.65 0.52 0.57 0.55 0.57 0.57 0.62 0.66 0.65 1 0.71 0.49 0.45 0.45 0.4 0.61 0.65 0.54 0.6 0.59 0.57 0.57 0.67 0.69 0.68 0.71 1 0H.D2 3H.D2 0H.D1 3H.D1 24H.D2 72H.D2 72H.D1 48H.D1 48H.D2 Day9ch.D1 24H.D1 Day9tr.D1 Day9tr.D2 Day6.D1 Day6.D2 Day9ch.D2 0 H D 2 . 3 H D 2 . 0 H D 1 . 3 H D 1 . 2 4 H D 2 . 7 2 H D 2 . 7 2 H D 1 . 4 8 H D 1 . 4 8 H D 2 . D a y 9 c h . D 1 2 4 H D 1 . D a y 9 t r . D a y 9 t r . D 1 D 2 D a y 6 . D 1 D a y 6 . D 2 D a y 9 c h . D 2 R-value 1 0.8 0.6 0.4 0.2 0 PBRM1 R-value 1 0.8 0.6 0.4 0.2 0 1 0.47 0.49 0.44 0.45 0.47 0.46 0.5 0.37 0.35 0.48 0.47 0.41 0.42 0.5 0.53 0.54 0.47 1 0.42 0.34 0.37 0.39 0.36 0.41 0.28 0.27 0.41 0.42 0.34 0.37 0.49 0.48 0.47 0.49 0.42 1 0.71 0.7 0.69 0.69 0.68 0.6 0.62 0.64 0.66 0.67 0.55 0.52 0.57 0.63 0.44 0.34 0.71 1 0.79 0.72 0.76 0.73 0.7 0.74 0.69 0.71 0.77 0.57 0.47 0.53 0.63 0.45 0.37 0.7 0.79 1 0.77 0.78 0.8 0.71 0.74 0.66 0.68 0.73 0.6 0.53 0.6 0.69 0.47 0.39 0.69 0.72 0.77 1 0.72 0.76 0.65 0.67 0.63 0.64 0.66 0.57 0.53 0.6 0.68 0.46 0.36 0.69 0.76 0.78 0.72 1 0.81 0.68 0.67 0.67 0.65 0.69 0.53 0.49 0.59 0.7 0.5 0.41 0.68 0.73 0.8 0.76 0.81 1 0.68 0.67 0.66 0.64 0.66 0.57 0.57 0.65 0.75 0.37 0.28 0.6 0.7 0.71 0.65 0.68 0.68 1 0.81 0.72 0.69 0.72 0.57 0.47 0.5 0.62 0.35 0.27 0.62 0.74 0.74 0.67 0.67 0.67 0.81 1 0.69 0.71 0.78 0.61 0.44 0.48 0.57 0.48 0.41 0.64 0.69 0.66 0.63 0.67 0.66 0.72 0.69 1 0.71 0.7 0.55 0.53 0.56 0.65 0.47 0.42 0.66 0.71 0.68 0.64 0.65 0.64 0.69 0.71 0.71 1 0.74 0.59 0.51 0.54 0.6 0.41 0.34 0.67 0.77 0.73 0.66 0.69 0.66 0.72 0.78 0.7 0.74 1 0.61 0.46 0.5 0.58 0.42 0.37 0.55 0.57 0.6 0.57 0.53 0.57 0.57 0.61 0.55 0.59 0.61 1 0.49 0.5 0.53 0.5 0.49 0.52 0.47 0.53 0.53 0.49 0.57 0.47 0.44 0.53 0.51 0.46 0.49 1 0.61 0.64 0.53 0.48 0.57 0.53 0.6 0.6 0.59 0.65 0.5 0.48 0.56 0.54 0.5 0.5 0.61 1 0.71 0.54 0.47 0.63 0.63 0.69 0.68 0.7 0.75 0.62 0.57 0.65 0.6 0.58 0.53 0.64 0.71 1 0H.D1 3H.D1 Day9ch.D1 Day6.D1 Day6.D2 Day9ch.D2 Day9tr.D1 Day9tr.D2 24H.D2 48H.D2 24H.D1 48H.D1 72H.D1 72H.D2 3H.D2 0H.D2.Rep1 0H.D2.Rep2 0 H D 1 . 3 H D 1 . D a y 6 . D 1 D a y 6 . D 2 D a y 9 c h . D 1 D a y 9 t r . D a y 9 t r . D 1 D 2 D a y 9 c h . D 2 2 4 H D 2 . 4 8 H D 2 . 2 4 H D 1 . 4 8 H D 1 . 7 2 H D 1 . 7 2 H D 2 . 3 H D 2 . . . 0 H D 2 . R e p 1 0 H D 2 . R e p 2 1 0.38 0.46 0.5 0.48 0.42 0.35 0.27 0.22 0.26 0.18 0.33 0.31 0.36 0.48 0.44 0.38 1 0.45 0.45 0.41 0.5 0.45 0.52 0.44 0.45 0.4 0.64 0.57 0.62 0.59 0.52 0.46 0.45 1 0.56 0.62 0.55 0.53 0.39 0.38 0.37 0.35 0.43 0.45 0.5 0.56 0.55 0.5 0.45 0.56 1 0.7 0.51 0.43 0.39 0.33 0.35 0.28 0.47 0.45 0.51 0.64 0.58 0.48 0.41 0.62 0.7 1 0.58 0.55 0.46 0.45 0.43 0.41 0.52 0.57 0.64 0.68 0.7 0.42 0.5 0.55 0.51 0.58 1 0.79 0.72 0.67 0.74 0.71 0.65 0.62 0.63 0.62 0.6 0.35 0.45 0.53 0.43 0.55 0.79 1 0.69 0.71 0.73 0.76 0.62 0.65 0.65 0.55 0.61 0.27 0.52 0.39 0.39 0.46 0.72 0.69 1 0.9 0.83 0.86 0.81 0.75 0.72 0.57 0.57 0.22 0.44 0.38 0.33 0.45 0.67 0.71 0.9 1 0.8 0.89 0.74 0.77 0.72 0.51 0.57 0.26 0.45 0.37 0.35 0.43 0.74 0.73 0.83 0.8 1 0.85 0.7 0.68 0.64 0.51 0.54 0.18 0.4 0.35 0.28 0.41 0.71 0.76 0.86 0.89 0.85 1 0.7 0.73 0.68 0.46 0.52 0.33 0.64 0.43 0.47 0.52 0.65 0.62 0.81 0.74 0.7 0.7 1 0.82 0.83 0.68 0.65 0.31 0.57 0.45 0.45 0.57 0.62 0.65 0.75 0.77 0.68 0.73 0.82 1 0.88 0.66 0.7 0.36 0.62 0.5 0.51 0.64 0.63 0.65 0.72 0.72 0.64 0.68 0.83 0.88 1 0.72 0.76 0.48 0.59 0.56 0.64 0.68 0.62 0.55 0.57 0.51 0.51 0.46 0.68 0.66 0.72 1 0.76 0.44 0.52 0.55 0.58 0.7 0.6 0.61 0.57 0.57 0.54 0.52 0.65 0.7 0.76 0.76 1 3H.D1 Day9ch.D1 3H.D2 0H.D1 0H.D2 24H.D1 24H.D2 72H.D1 72H.D2 48H.D1 48H.D2 Day6.D1 Day6.D2 Day9ch.D2 Day9tr.D1 Day9tr.D2 3 H D 2 . 0 H D 1 . 0 H D 2 . 2 4 H D 1 . 2 4 H D 2 . 7 2 H D 1 . 7 2 H D 2 . 4 8 H D 1 . 4 8 H D 2 . 3 H D 1 . D a y 9 c h . D 1 D a y 6 . D 1 D a y 6 . D 2 D a y 9 t r . D a y 9 t r . D 1 D 2 D a y 9 c h . D 2 SS18.3HR.d1 SS18.3HR.d1 SS18.0HR.d1 SS18.3HR.d1 813 283 1077 778 571 8066 SS18.24HR.d1 1896 6707 13495 SS18.24HR.d1 SS18.48HR.d1 Figure S2 SS18.48HR.d1 SS18.72HR.d1 SS18.Day6.d1 SS18.Day6.d1 4602 15577 SS18.72HR.d1 SS18.Day9ch.d1 537 5813 11816 9652 1628 10345 971 SS18.Day9tr.d1 SS18.Day9ch.d1 SS18.Day9ch.d1 SS18.0HR.d1 SS18.Day9tr.d1 SS18.0HR.d1 696 819 3798 1167 351 739 4198 424 667 ARID1A PBRM1 0h 3h 24h 48h 72h D6 D9-Ch D9-Tr Donor1 Donor2 % 5 4 . 1 1 : 2 C P 4 . 0 2 . 0 0 . 0 2 . 0 − 4 . 0 − 6 . 0 − 0h 3h 24h 48h 72h D6 D9-Ch D9-Tr Donor1 Donor2 % 8 3 . 1 1 : 2 C P 4 . 0 2 . 0 0 . 0 2 . 0 − 4 . 0 − 6 . 0 − E −0.6 −0.4 −0.2 0.0 0.2 0.4 PC1: 23.54% ATAC-seq F −0.6 −0.4 −0.2 0.0 0.2 0.4 PC1: 20.64% ATAC-seq R-value 1 0.8 0.6 0.4 0.2 0 60000 40000 20000 t n u o c k a e P 0 1 0.95 0.79 0.83 0.86 0.86 0.8 0.8 0.72 0.74 0.65 0.6 0.57 0.56 0.55 0.58 0.95 1 0.79 0.8 0.82 0.86 0.81 0.81 0.65 0.69 0.59 0.58 0.5 0.54 0.48 0.55 0.79 0.79 1 0.93 0.9 0.91 0.75 0.74 0.68 0.72 0.8 0.79 0.66 0.7 0.59 0.67 0.83 0.8 0.93 1 0.94 0.92 0.73 0.72 0.68 0.72 0.76 0.7 0.61 0.62 0.55 0.6 0.86 0.82 0.9 0.94 1 0.95 0.71 0.71 0.66 0.69 0.72 0.66 0.57 0.57 0.52 0.56 0.86 0.86 0.91 0.92 0.95 1 0.73 0.73 0.63 0.67 0.69 0.67 0.55 0.58 0.49 0.57 0.8 0.81 0.75 0.73 0.71 0.73 1 0.95 0.88 0.89 0.69 0.7 0.7 0.74 0.7 0.76 0.8 0.81 0.74 0.72 0.71 0.73 0.95 1 0.85 0.89 0.67 0.67 0.66 0.69 0.65 0.7 0.72 0.65 0.68 0.68 0.66 0.63 0.88 0.85 1 0.92 0.74 0.71 0.79 0.77 0.81 0.8 0.74 0.69 0.72 0.72 0.69 0.67 0.89 0.89 0.92 1 0.75 0.72 0.77 0.75 0.76 0.76 0.65 0.59 0.8 0.76 0.72 0.69 0.69 0.67 0.74 0.75 1 0.93 0.89 0.85 0.79 0.79 0.6 0.58 0.79 0.7 0.66 0.67 0.7 0.67 0.71 0.72 0.93 1 0.89 0.91 0.78 0.83 0.57 0.5 0.66 0.61 0.57 0.55 0.7 0.66 0.79 0.77 0.89 0.89 1 0.93 0.92 0.89 0.56 0.54 0.7 0.62 0.57 0.58 0.74 0.69 0.77 0.75 0.85 0.91 0.93 1 0.89 0.92 0.55 0.48 0.59 0.55 0.52 0.49 0.7 0.65 0.81 0.76 0.79 0.78 0.92 0.89 1 0.94 0.58 0.55 0.67 0.6 0.56 0.57 0.76 0.7 0.8 0.76 0.79 0.83 0.89 0.92 0.94 1 Day9tr.D1 Day9tr.D2 3H.D2 3H.D1 0H.D1 0H.D2 Day6.D2 Day9ch.D2 Day6.D1 Day9ch.D1 24H.D1 24H.D2 48H.D1 48H.D2 72H.D1 72H.D2 . 1 D H 0 . 2 D H 0 . 1 D H 3 . 2 D H 3 . 1 D H 4 2 . 2 D H 4 2 . 1 D H 8 4 . 2 D H 8 4 . 1 D H 2 7 . 2 D H 2 7 1 D . 6 y a D 2 D . 6 y a D 1 D . h c 9 y a D 2 D . h c 9 y a D 1 D 2 D . r t 9 y a D . r t 9 y a D 3 H D 2 . 3 H D 1 . 0 H D 1 . 0 H D 2 . D a y 9 t r . D a y 9 t r . D 1 D 2 D a y 6 . D 1 D a y 6 . D 2 D a y 9 c h . D 2 D a y 9 c h . D 1 2 4 H D 1 . 2 4 H D 2 . 4 8 H D 1 . 4 8 H D 2 . 7 2 H D 1 . 7 2 H D 2 . G ARID1A Z-Score PBRM1 Z-Score J Sample Cluster 1 (n=11167) Pan-state targeted/ accessible Cluster 2 (n=4439) Early activation Cluster 3 (n=2508) Activation/Exhaustion Cluster 4 (n=6998) Activation/Exhaustion Cluster 5 (n=4894) Late activation Cluster 6 (n=2759) Exhaustion Cluster 7 (n=3085) Naive/memory (strongest in D9-Tr) Cluster 8 (n=10163) Naive/memory (reduced upon activation) Cluster 9 (n=27495) Naive/memory (low accessibility) H n o i t c a r F e t i S 0 h 3 h 2 4 h 4 8 h 7 2 h D 6 D 9 - C h D 9 - T r 0 h 3 h 2 4 h 4 8 h 7 2 h D 6 D 9 - C h D 9 - T r I Distance To TSS > 100000 10000 - 100000 1000 - 10000 100 - 1000 0 - 100 All sites 45585 1753 17430 7267 679 417 375 1.00 0.75 0.50 0.25 0.00 C1C2C3C4C5C6C7C8C9 Cluster Smarca4 log2- transformed RPKM Ss18 log2- transformed RPKM H3K27Ac log2- transformed RPKM ATAC-seq log2- transformed RPKM Mouse CD8+ T cells C1 (n=24188) Low accessibility C2 (n=3100) Naive C3 (n=7928) Early activation C4 (n=3900) Activation/ D9 C5 (n=5982) Activation C6 (n=4672) Day9 0 h 4 8 h 7 2 h D 9 - C h D 9 - T r 0 h 4 8 h 7 2 h D 9 - C h D 9 - T r 0 h 4 8 h 7 2 h D 9 - C h D 9 - T r 0 h 4 8 h 7 2 h D 9 - C h D 9 - T r ATAC-Seq Merged Peaks H3K27ac Merged Peaks SMARCA4 & SS18 Merged Peaks Figure S2. Chromatin occupancy and accessibility profiling of human CD8+ T cells during activation and exhaustion, Related to Figure 1. A. Raw peak counts for SMARCA4, SS18, ARID1A, PBRM1, and H3K27Ac CUT&Tag experiments performed across time course, for 2 independent human donors (denoted as D1,D2). B. Venn diagrams reflecting pairwise comparisons of peak numbers for SS18 mSWI/SNF complex subunit CUT&Tag across time points (Donor 1 shown). C. Pairwise correlations between all samples for SMARCA4, SS18, ARID1A, PBRM1, and H3K27Ac CUT&Tag occupancy levels within merged peaks. D. Principal component analyses (PCA) for ARID1A (cBAF) and PBRM1 (PBAF) CUT&Tag profiles throughout the activation/exhaustion time course. E. Peak counts for ATAC-seq experiments performed across time course, for 2 independent human donors. F. Pairwise correlations between all samples for ATAC-seq experiments. G. K-means clustering for ARID1A and PBRM1 peaks as in Figure 1D; Quantile-normalized Log2-Transformed RPKMs are presented in the heatmap as Z-scores. H. Distance-to- TSS plot for Clusters 1-9 from Figure 1D. I. Venn diagram reflecting overlap between ATAC-seq accessible peaks, mSWI/SNF complexes (SS18/SMARCA4 merged) and H3K27Ac. J. Heatmaps reflecting K-means clustering of quantile-normalized log2-transformed RPKMs from Smarca4, Ss18, H3K27ac, and ATAC-seq merged peaks from mouse CUT&Tag and ATAC-seq experiments. C1 C2 C3 C4 C5 C6 C7 C8 C9 1 0.5 0 −0.5 −1 Fractional enrichment B Elk4 YY1 Elk1 ELF1 ETS NFY NRF NRF1 Fli1 Ronin GFY−Staf GABPA Sp1 Sp5 GFX ETV1 GFY KLF3 ETS1 E2F4 Cluster 1 Cluster 2 Cluster 3 Cluster 4 Cluster 5 Cluster 6 Cluster 7 Cluster 8 Cluster 9 HOMER analysis Atf3 BATF AP−1 JunB Fra1 Fra2 Fosl2 Jun−AP1 NFAT:AP1 NFAT Bach2 NFkB−p50,p52 Tbet MafK NFkB−p65 Bach1 Nrf2 MafB NF−E2 BMAL1 BATF Atf3 AP−1 JunB Fra1 Fra2 Fosl2 Jun−AP1 Bach2 NFAT:AP1 NFAT Egr1 MafK NFkB−p65 NFkB−p50.52 Pdx1 MafB Egr2 NF−E2 MafA NFkB−p65 BATF Fra1 JunB Atf3 Fra2 AP−1 Fosl2 Jun−AP1 NFkB−p65−Rel Egr1 Egr2 NFAT NFkB−p50.52 IRF:BATF bZIP:IRF Bach2 Nur77 NFAT:AP1 IRF4 Atf3 Fra1 BATF AP−1 Fra2 JunB Fosl2 Jun−AP1 Bach2 Chop Atf1 STAT5 MafK Bach1 Atf7 c−Jun−CRE ZEB1 Atf4 Atf2 STAT1 HNF1b Hnf1 MYB NFkB−p65 AMYB NFkB−p65−Rel BMYB Tcf4 Nkx6.1 RUNX1 RUNX2 Nanog FOXK1 Lhx1 RUNX RUNX−AML FOXA1 Foxo3 Foxf1 Foxa2 ETS1 Fli1 ETV1 ERG Etv2 GABPA ETS:RUNX EWS:FLI1 Elk1 EWS:ERG Elk4 SPDEF EHF Ets1−distal ELF1 PU.1 ELF3 ETS RUNX ELF5 ETS1 ETV1 Etv2 Fli1 ERG GABPA Ets1−distal EWS:FLI1 EWS:ERG ETS:RUNX EHF ELF1 SPDEF Elk1 ETS ELF3 PU.1 Elk4 ELF5 RUNX1 CTCF BORIS CTCF−Sat Erra Zic3 Slug E2A Unk−ESC NeuroD1 TEAD COUP−TFII LRF REST−NRSF ZEB1 IRF1 TEAD4 IRF2 ISRE Ptf1a Ascl1 Figure S3 0 400 0 600 0 400 0 200 D 0 −Log (P-value) 800 0 400 0 200 0 1000 0 6000 A Transcription Factor Archetype Motif Enichment Analysis (50 Select Enriched & Variable Motifs) I Z C I Z C . 2 P A X . 2 R U N X . 2 T F A P 2 . 1 N R F 1 E b o x . C A C G T G . 2 C T C F E 2 F. 2 E G R E T S . 1 E T S . 2 I H N F P 1 . 1 Pan- targets A P 1 . 1 B A T F N F A C . 2 N F A T . 1 N F A T . 2 P O U . 1 P O U . 2 P O U . 3 S O X . 3 S O X . 4 S O X . 5 M E F 2 N F K B . 1 N F K B . 2 N F K B . 3 A P 1 . 2 C R E B A T F. 1 . M A F T B X . 1 T B X . 4 H D . 1 0 C U X . 1 F O X . 4 F O X . 5 T C F. L E F L E F 1 I I R F. 1 R F. 2 B C L 6 . 2 R U N X . 1 S P I S T A T . 1 S T A T . 2 Z N F 3 8 4 . 2 C R E B A T F. 2 . C R E B A T F. 3 . O C T 4 + S O X 2 Early/intermediate activation Late activation Exhaustion Naive/memory 3h vs 0h 48h vs 0h D6 vs 0h D9-Tr vs D9-Ch AP1.1 BATF NFAT.1 NFAT.2 BCL6.1 ZNF652 NFKB.2 IRF.3 NFKB.1 NFAT.4 CREB.ATF.3 POU.1 NFAT.3 MEF2 CUX.4 NFKB.3 IRF.1 LIN54 HLTF NR.6 STAT.2 DMRT3 CREB.ATF.2 REL.halfsite SOX.3 ZNF146 HD.11 ZNF431 ARI5B GMEB2.1 SPDEF.1 HIC.1 YY1 ZBED1 FOX.5 HINFP1.2 ETS.1 MYB.4 HD.10 PAX.halfsite HD.10 BATF BCL6.1 ZNF652 AP1.1 IRF.3 NFKB.1 ZNF232 NR.6 CREB.ATF.3 TCF.LEF AIRE IRF.1 STAT.2 NR.9 RUNX.1 NFKB.3 POU.1 HD.25 NFAT.2 HLTF NR.10 CUX.3 Ebox.CATATG BCL6.2 MEF2 SOX.5 REL.halfsite FOX.5 ZBTB7A HD.1 IRF.4 HIF ARI5A SNAI2 HD.9 HIC.1 HINFP1.2 SPDEF.1 PAX.halfsite SNAI2 SPDEF.1 ETS.1 GFI HIC.1 SMARCA1 SPI ZNF431 MEF2 NR.9 HD.8 HLTF FOX.1 IRF.1 POU.1 ZNF306 ZNF384.1 MBD2 HD.14 FOX.2 NFAT.3 IRF.3 NR.6 ZNF232 SOX.7 MYB.5 NFKB.2 BCL6.1 NFAT.1 CUX.3 CUX.4 NFAT.2 HD.25 ZNF652 NFKB.1 NFKB.3 HD.3 AP1.1 BATF HD.10 Accessibility UP DOWN Accessibility UP DOWN Accessibility UP DOWN Accessibility UP DOWN C s D I f i t o M F T BCL6.1 AP1.1 ZBED1 HINFP1.2 GMEB2.3 HD.3 IRF.1 NFAT.3 NFKB.1 NFAT.1 HD.13 ZNF384.1 NFAT.2 BATF NFAT.4 CREB.ATF.1 AHR PAX.halfsite IRF.3 ZBTB49 HINFP1.3 CREB.ATF.3 NFKB.3 FOX.3 Ebox.CACCTG SIX.2 TCF.LEF ZNF53 ETS.1 DMRT1 AIRE ZNF431 RUNX.1 DDIT3.CEBPA ZIM3 GFI SPI ZNF547 ZNF490 HD.1 F −0.2−0.1 0 0.1 0.2 Model Coefficient 0 1 5 0 ) i M P C g o L ( n o s s e r p x E e n e G F T 7 F C T 2 F L K 3 F L K 2 D 1 R N S E M O E 0 2 B T B Z 1 X N U R 1 F 6 U O P 2 C 3 R N 1 D 1 R N 1 T A T S 1 F R 4 F R I I C Y M L E R 8 F R I 2 R G E 3 R G E 1 A 4 R N 5 T A F N R H A B L E R 1 R G E 3 A 4 R N 2 L S O F 2 A 4 R N 1 C T A F N F A M B S O F F F A M 1 L S O F 3 F T A 3 F T A B S O F I X F N 7 V T E 5 F T A 7 6 3 F N Z 4 0 7 F N Z 6 F 2 R N 1 P D F T 1 A G M H 3 L 2 E F N 2 1 X O S D N U J 2 D E B Z 2 3 B T B Z 0 4 E H L H B 3 L I F N A R X R 3 P X O F 1 X R A B 1 S A P E 1 J X O F 6 O X O F 2 L B Y M 1 M X O F 4 C T A F N 1 F 2 E B 1 F N H A P N E C 5 F R I 7 F 2 E B Y M R R H A 7 F R I 4 X O S 3 X H F Z 2 X B C 2 A 5 R N 1 4 5 F N Z A 3 D R A I 1 F A 2 U O P 3 6 P T 2 F 2 E 8 F 2 E 1 T E T 3 D X M 3 I A N S 3 L 3 B E R C 0.4 0.2 0.0 −0.2 L H e r u t a n g s i 6 C g a T & T U C 0h 3h 24h 48h 72h D6 D9-Ch D9-Tr G J 5 0 n o i t a v e D i P scRNAseq signatures: Exhaustion - Tirosh Exhaustion - Zheng Exhaustion - Zhang Exhaustion - Guo Exhaustion - Sade-Feldman Memory - Sade-Feldman 6.3e−10 p < 2.22e−16 Cell type 6.4e−06 Mouse CD8+ T cells K 150 100 M P C Effector CD8 T Early TEx TEx 48h vs 0h Accessibility UP DOWN BATF IRF.3 AP1.1 NFY BCL6.1 NFKB.2 STAT.2 BCL6.2 IRF.1 AHR HIF ZNF140 TBX.4 STAT.1 ZNF431 PRDM16 REL.halfsite ZNF436 ZNF586 SPI HD.16 CREB3.XBP1 ZNF57 SCRT1 RFX.1 RUNX.1 ARI5B HD.14 HD.24 CREB.ATF.1 HD.18 ZNF146 SOX.6 FOX.5 CUX.4 YY1 ETS.1 HD.25 ZNF410 HD.5 s D I f i t o M F T HNF1B expression 6.90E-03 ** 4.77E-01 ns 2.59E-04 *** NA NA NA Human Mouse 72h vs 0h Accessibility UP DOWN TCF.LEF FOX.1 TBX.4 ETS.1 TATA IRF.1 ZBTB14 SMARCA1 STAT.1 MEF2 SNAI2 NFKB.2 HD.10 SPI SOX.2 ZNF547 HD.24 SIX.2 HD.8 POU.1 HD.18 MECP2 ZNF57 GATA NFI.2 KLF.SP.1 HD.25 ZBED1 GLI CREB3.XBP1 MFZ1 ZNF24 HIF ZNF140 ZNF410 YY1 AIRE MYB.4 CREB.ATF.1 BATF 50 0 NFKB.2 NFY BATF TBX.4 IRF.3 HLTF HIF RUNX.1 AP1.1 ZNF431 REL.halfsite ZNF449 HD.21 PRDM14 ZNF436 ZNF708 ZNF140 BCL6.1 HD.18 GFI FOX.5 ZNF384.1 E2F.1 TATA HINFP1.1 CREB3.XBP1 GMEB2.2 HD.8 ZNF57 SOX.2 TEAD HD.14 LIN54 MYB.4 SOX.6 HD.25 HD.10 NFAT.2 DMRT3 HD.5 −0.4−0.2 0 0.2 0.4 Model Coefficient −0.3 0 0.3 Model Coefficient −0.6−0.3 0 0.3 0.6 Model Coefficient TF Gene Expression during T−Cell Activation/Exhaustion 0HR 3HR 24HR 48HR 72HR D6 D9ch D9tr E 3 2 1 0 −1 −2 −3 0 h 3 h 2 4 h 4 8 h 7 2 h D 6 D 9 - C h D 9 - T r Normalized enrichment score Tirosh et al. melanoma n=1176 R = 0.46, p<2.2e-16 Zheng et al. HCC n=696 R = 0.72, p<2.2e-16 0.50 0.25 0.00 −0.25 −0.50 0.6 0.4 0.2 0.0 −0.2 −0.4 −2.5 0.0 2.5 5.0 0 1 2 3 0 1 2 Zhang et al. CRC n=1693 Guo et al. NSCLC n=1959 Sade-Feldman et al. melanoma n=2438 R = 0.63, p<2.2e-16 0.50 R = 0.57, p<2.2e-16 0.75 R = 0.5, p<2.2e-16 0.25 0.00 −0.25 −0.50 0.50 0.25 0.00 −0.25 0 1 2 3 −1 0 1 2 8 . 0 t n e m h c i r n E f i t o M 4 r e t s u C l 6 . 0 4 . 0 2 . 0 0 0 . Intermediate activation JUN ATF3 JDP2 JUNB HNF1B POU4F1 SOX4 IRF7 POU5F1B SOX18 IRF5 IRF2 TCF7L1 BCL6B FOXD1 FOXP3 FOXM1 FOXH1 BARX1 TF Motif Archetype AP1.1 BCL6.1 Ebox.CATATG FOX.1/4/5 HD.5 HD.10 IRF.1/3/4 OVOL1 POU.1/2 SOX.3/4 TCF.LEF Other 0 2 4 6 48H vs. 24H TF Gene LogFC 8 0.0054 3.1e−05 p < 2.22e−16 Cell type Naive CD8 T Effector CD8 T Early TEx Intermediate TEx Terminal TEx Memory CD8 T 181 22 727 SMARCA4-SS18 CUT&Tag (D6, D9-Ch merged) 923 1167 6060 36016 I 5 0 n o i t i a v e D −5 O HNF1B CUT&Tag (D9-Ch) HNF1B D9-Ch N ) l / m n o i l l i m ( r e b m u n l l e C CTRL sgHNF1B p<0.05 * 80 60 40 20 0 exhaustion signature HNF1B CUT&Tag M Sample D9-Ch D9-Tr HNF1B D9-Ch 7500 t n u o C . k a e P 5000 2500 0 Q D9-Tr vs D9-Ch HNF1b Sp1 Hnf1 NFY KLF3 Klf9 Sp5 NRF Ronin NRF1 GFY−Staf KLF5 Klf4 KLF6 KLF10 GFY KLF14 GFX Pbx3 ZBTB33 s e t i i S d e n a G n a m u H 0 . 1 5 . 0 0 . 0 5 . 0 − 0 . 1 − 0 100 300 200 −Log (P-value) 400 48HR vs. 0HR R = 0.965 AP1.1 NFAT.1 NFAT.2 3 5 8 12 Days after electroporation 15 ATAC-seq (D6,D9-Ch merged) 72HR vs. 0HR Day9ch vs. 0HR Day9tr vs. 0HR s e t i i S d e n a G n a m u H 0 . 1 5 . 0 0 . 0 5 . 0 − 0 . 1 − R = 0.832 AP1.1 BATF ZNF24 HD.5 ZIC.2 KLF.SP.1 s e t i i S d e n a G n a m u H 0 . 1 5 . 0 0 . 0 5 . 0 − 0 . 1 − R = 0.865 HD.10 (HNF1) SOX.2 IRF.1 IRF.3 SOX.6 POU.1 AP1.1 NFKB.2 NFKB.1 NFKB.3 AP1.2 R = 0.868 STAT.2 BCL6.2 IRF.3 ZIM3 ZNF384.2 IRF.1 TBX.4 BATF POU.2 NR.13 HD.10 AP1.2 NFAT.1 AP1.1 s e t i i S d e n a G n a m u H 5 . 0 0 . 0 5 . 0 − Accessibility UP DOWN −1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0 −0.5 0.0 0.5 Mouse Gained Sites Mouse Gained Sites Mouse Gained Sites Mouse Gained Sites 72HR vs. 48HR Day9ch vs. 72HR Day9tr vs. Day9ch s e t i i S d e n a G n a m u H 0 . 1 5 . 0 0 . 0 5 . 0 − 0 . 1 − R = 0.820AP1.1 CREB.ATF.2 CREB.ATF.1 CREB.ATF.3 RUNX.1 BATF AP1.2 ETS.1 s e t i i S d e n a G n a m u H 0 . 2 5 . 1 0 . 1 5 . 0 0 . 0 5 . 0 − 0 . 1 − R = 0.794 HD.10 (HNF1) IRF.3 IRF.1 SOX.2 FOX.5 TCF.LEF RUNX.1 CREB.ATF.3 BATF AP1.1 CREB.ATF.2 CREB.ATF.1 AP1.2 R = 0.840 s e t i i S d e n a G n a m u H 0 . 1 5 . 0 0 . 0 5 . 0 − 0 . 1 − ZNF384.2 NR.14 TBX.4 EVI1.MECOM SOX.6 NR.2 NFKB.3 SOX.2 AP1.1 −0.2 Model Coefficient 0.0 0.2 −0.1 0.0 0.1 Model Coefficient −0.2 −0.1 0.0 0.1 0.2 Model Coefficient −1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0 1.5 2.0 −1.0 −0.5 0.0 0.5 1.0 Mouse Gained Sites Mouse Gained Sites Mouse Gained Sites Figure S3. mSWI/SNF targeting and accessibility over TF target genes during human T cell activation and exhaustion, Related to Figure 2,3. A. Fractional motif enrichment in clusters (relative to all sites) for selected archetype motifs with high occurrence and variability. B. HOMER motif enrichment analysis across indicated clusters. C. Differential motif accessibility between time points indicated (top 40 coefficients of logistic regression models). D. Volcano plots of changes in gene expression from n=2 independent donors throughout the time course. Significantly up-regulated and down-regulated genes (Abs Log2FC > 1, FDR < 0.01) are colored in red and blue, respec- tively. The top 20 differentially expressed genes in every comparison are labeled. E. Plot representing state-specific TF fractional motif enrichment (y-axis) and gene expres- sion (x-axis) at the intermediate activation (C4) state. F. Gene expression levels (logCPMs) of 80 select TF genes with high expression and variability during T-cell activation and exhaustion. G. Enrichment (-log10(p-value)) of mSWI/SNF-bound genes at different time points across signatures derived from scRNAseq datasets. H. Correlations between expression changes for genes in C6 (Cluster 6) and exhaustion signatures for each published study. R correlation values and p values are indicated. I. HNF1B motif enrichment across cell types in the Satpathy et al. scATAC-seq dataset. J. HNF1B motif enrichment across cell types in the Kourtis et al. scATAC-seq dataset. K. HNF1B gene expression (CPM) in human and mouse T cells across the activation/exhaustion time course. L. HNF1B CUT&Tag raw peak numbers in Day9-Ch and Day9-Tr conditions. M. Motif enrichment analysis performed on HNF1B CUT&Tag in Day9-Chr condition. N. T cell proliferation in sgCTRL and sgHNF1B conditions. O. Venn diagram reflecting overlap between SMARCA4/SS18 CUT&TAG, HNF1B CUT&Tag and ATAC-seq peaks in the Day9-Ch condition. P. Top 40 coefficients of logistic regression models fitting motif counts across all sites to changes in accessibility for selected time point comparisons in mouse T-cells during activation and exhaustion. Q. Archetype motif fractional enrichment for sites of gained accessibility (LogFC > 0,) relative to all sites in human and mouse settings. Selected motifs are labeled in red. + 4 4 D C + L 2 6 D C 2.5 2.0 1.5 1.0 0.5 0.0 ) - - P F R A N R g s o t d e z i l a m r o n ( A E H K N 15 10 5 0 ) 5 ^ 0 1 x ( s d a e r d e p p a m f o r e b m u N all PD1+ TIM3+ x e d n i i i n G ) s d a e r A N R g s f o s s e n n e v e ( 0.08 0.06 0.04 0.02 0.00 all PD1+ TIM3+ B W r i t e r s E r a s e r s R e a d e r s HATs HMTs HDACs HDMs SWI/SNF BET SPs other E/EM CM 4 4 D C CD62L N 0 2 4 6 8 Number of significantly depleted hits (FDR < 0.05) sgRNA 1 sgRNA 2 F C 600 Mouse RNA Expression * p = 0.0212 D ) - M P C 400 200 0 ARID1A ARID1B Human RNA Expression * p = 0.0167 300 I + 3 M T + 1 D P Figure S4 sgRNA 1 sgRNA 2 - + - + - + - + - + - + - + 1.5 1.0 - P F R A N R g s o t 0.5 d e z i l 0.0 a m r o n ( sgRosa sgSmarca4 sgSmarcc1 sgArid1a sgArid1b sgDpf2 sgArid2 M P C 200 100 0 ARID1A ARID1B sgRosa sgSmarca4 sgArid1a sgSmarcc1 sgDpf2 g n i l l i k % 0.6 0.4 0.2 0.0 - + - + - + - + - + - + - + sgRosa sgSmarca4 sgSmarcc1 sgArid1a sgArid1b sgDpf2 sgArid2 15 10 5 0 ) 5 ^ 0 1 x ( s d a e r d e p p a m f o r e b m u N TIM3 low TIM3 high x e d n i i i n G ) s d a e r A N R g s f o s s e n n e v e ( 0.08 0.06 0.04 0.02 0.00 2 4 8 16 32 64 128 256 T:E KAT7 DOT1L KAT5 Enriched in TIM3+ HDAC3 HDAC1 PRMT1 SETD1A EZH2 PBRM1 ARID2 PDCD1 SMARCC2 ARID1B SMARCA4 EP300 DPF2 PRDM1 KDM2B, KMT2C SMARCC1 I 2.5 0 C F 2 g o L −2.5 ARID1A HAVCR2 −5 0 Depleted in TIM3+ 100 Rank 200 300 L SIINFEKL sgRNA CHRONIC STIMULATION B16-OVA B16-OVA B16-OVA 0h Mouse splenocytes (Cas9 - OT-1 mice) 48h CD8+ T cell isolation D5-Ch D7-Ch D9-Ch B16 B16 B16 TIM3 low TIM3 high Pan- mSWI/ SNF cBAF pBAF Smarca4 Smarcc1 Smarcc2 Arid1a Arid1b Dpf2 Arid2 Phf10 Pbrm1 Brd7 Brd9 −4 TIM-3 HIGH vs LOW low 15% high 15% - P F R A N R g s Cas9-GFP W r i t e r s E r a s e r s R e a d e r s HATs HMTs HDACs HDMs SWI/SNF BET SPs other 1 0 2 5 Number of significantly depleted hits (FDR < 0.05) 3 4 G J M TRANSIENT 15 10 5 i n o s n a p x e d o F l i n o s n a p x e d o F l CHRONIC 15 10 5 0 D3 D5 D7 D9 D3 D5 D7 D9 −2 −3 0 Log2(Fold change) −1 ncBAF 1 ACTIVATION D5-Tr D7-Tr D9-Tr TRANSIENT STIMULATION 0 Days: D9-Tr 35 3 0 61 3 10 4 10 D9-Ch 69 4 10 3 10 0 0 0 30 3 10 4 10 4 10 3 10 0 3 M T I PD1 69 D9-Tr D9-Ch 31 84 16 4 4 D C 0 CD62L 0 0 0 O 3 M T I 4 10 3 10 0 0 0 PD1 87 4 4 D C 0 CD62L sgROSA sgSmarca4 sgSmarcc1 78 69 4 10 3 10 0 4 10 3 10 0 22 3 10 4 10 0 0 30 3 10 4 10 55 0 0 45 3 10 4 10 4 10 3 10 0 sgDpf2 73 P 0 0 26 3 10 4 10 sgROSA sgSmarca4 sgSmarcc1 sgDpf2 13 67 33 75 24 78 21 sgSMARCA4 sgCTRL SMARCA4 250 kDa Actin 55 kDa 0 0 0 1 0 1 0 Figure S4. Contributions of mSWI/SNF (cBAF) complexes to T cell exhaustion in two independent CRISPR-Cas9-based screens in mouse CD8+ T cells, Related to Figure 4. A. Number of mapped reads (right) and Gini index representing the evenness of sgRNA reads (left) for the PD1+TIM3+ CRISPR screen. B. Number of signifi- cantly depleted hits (Log2FC < -1, FDR<0.05) within the indicated classed of chromatin writers, erasers or readers. C. Bar graph indicating the expression levels (RPKM) or ARID1A and ARID1B in mouse and human CD8+ T cells. D. Quantification of the PD1+TIM3+ population in cells infected with sgRNAs targeting the indicated genes, normalized to the sgRNA-negative population in the same culture, at Day9 of chronic stimulation. Different sgRNAs for the same gene are labeled with different shapes. Bar graphs represent mean ± SEM from 2-5 independent biological replicates. E. Quantification of the CD62L+ CD44+ population in cells infected with sgRNAs targeting the indicated genes, normalized to the sgRNA-negative population in the same culture, at Day9 of chronic stimulation. Different sgRNAs for the same gene are labeled with different shapes. Bar graphs represent mean ± SEM from 2-4 independent biological replicates. F. In vitro killing efficiency of B16-OVA cells by OT-1 T cells transduced with the indicated sgRNAs, at Day9 of chronic stimulation. B16-OVA cells and OT-1 T cells were co-incubated for 48 hours at the indicated Target-to-Effector (T:E) ratios. Means ± SEM from 4 technical replicates are shown. G. Sorting strategy for the TIM3 High vs Low screen using a custom sgRNA library of chromatin regulators. H. Number of mapped reads (right) and Gini index representing the evenness of sgRNA reads (left) for the TIM3 High vs Low CRISPR screen. I. Rank plot depicting Log2FC scores (average of n=6 guides) targeting all chromatin regulator genes and negative/positive controls. Depleted genes are highlighted in black; positive controls are highlighted in gray; mSWI/SNF complex genes are highlighted in orange. J. Number of significantly depleted hits (Log2FC < -1, FDR<0.05) within the indicated classed of chromatin writers, erasers or readers in the TIM3 High vs Low CRISPR screen. K. Log2FC values for n=6 independent guides in the TIM3 High vs Low CRISPR screen. L. Schematic for stimulation of mouse OT-1 CD8+ T cells based on co-culture with B16 or B16-OVA cells to profile early activation, transient stimulation, and chronic stimulation/exhaus- tion states. M. Fold expansion for mouse OT-1 CD8+ T cells in the chronic or transient stimulation conditions across the Day 3, 5, 5, 9 time points. Bar graphs represent mean ± SEM from 9 independent mice. N. FACS-based profiling of PD1, TIM3, CD62L and CD44 at Day9 of transient or chronic stimulation in mouse OT-1 CD8+ T cell studies. O. FACS-based profiling of PD1, TIM3, CD62L and CD44 at Day9 of chronic stimulation in control and mSWI/SNF subunit gene KO conditions. P. Western blot analysis of SMARCA4 levels in sgCTRL or sgSMARCA4 human CD8+ T cells, profiled at Day 9 of the chronic stimulation protocol, compared to actin as loading control. A ACBI1 AU-15330 nM: 0 50 100 50 100 SMARCA4 Actin C Donor 3 DMSO ACBI 50nM ACBI 100nM AU-15330 100nM 0 4 10 5 10 Comp-APC-A :: CD39 ) 3 ^ 0 1 x ( I F M 9 3 D C 8 6 4 2 0 L R T C M n 0 5 1 I B C A M n 0 0 1 1 I B C A M n 0 5 0 3 3 5 1 - U A M n 0 0 1 0 3 3 5 1 - U A 250 kDa 55 kDa Donor 4 DMSO ACBI 50nM ACBI 100nM AU-15330 100nM ) 3 ^ 0 1 x ( I F M 9 3 D C 8 6 4 2 0 0 4 10 5 10 Comp-APC-A :: CD39 Donor 1 Donor 2 L R T C M n 0 5 1 I B C A M n 0 0 1 1 I B C A M n 0 5 0 3 3 5 1 - U A M n 0 0 1 0 3 3 5 1 - U A ) 3 ^ 0 1 x ( I F M 9 3 D C 5 4 3 2 1 0 0 3 10 Comp-APC-A :: CD39 L R T C M n 0 0 1 4 1 P M C M n 0 0 1 5 1 0 1 - T H F 1.5 ) 3 ^ 0 1 x ( I 1 0.5 F M 9 3 D C 0 L R T C M n 0 0 1 4 1 P M C M n 0 0 1 5 1 0 1 - T H F 0 3 10 Comp-APC-A :: CD39 D B Donor 4 DMSO ACBI1 50nM ACBI1 100nM AU-15330 50nM AU-15330100nM 5 10 4 10 3 10 0 3 -10 64 5 19 5 10 4 10 3 10 0 3 -10 50 7 36 5 10 4 10 3 10 0 3 -10 40 7 49 5 10 4 10 3 10 0 3 -10 54 6 34 5 10 4 10 3 10 0 3 -10 41 7 47 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 Donor 5 DMSO ACBI1 50nM ACBI1 100nM Figure S5 5 10 4 10 3 10 0 3 -10 66 42 5 10 4 10 3 10 0 5 10 4 10 3 10 0 31 3 0 26 2 3 -10 55 2 3 -10 65 3 10 4 10 5 10 0 3 10 4 10 5 10 0 3 10 4 10 5 10 Donor 2 DMSO CMP14 50nM CMP14 100nM FHT-1015 50nM FHT-1015 100nM 5 10 4 10 3 10 0 3 -10 44 10 41 5 10 4 10 3 10 0 3 -10 29 3 66 5 10 4 10 3 10 0 3 -10 11 4 85 5 10 4 10 3 10 0 3 -10 41 8 49 5 10 4 10 3 10 0 3 -10 32 6 60 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 Donor 6 DMSO CMP14 50nM CMP14 100nM FHT-1015 50nM FHT-1015 100nM 5 10 4 10 3 10 0 3 -10 3 M T I 49 9 25 5 10 4 10 3 10 0 3 -10 8 12 79 5 10 4 10 3 10 0 3 -10 4 15 81 5 10 4 10 3 10 0 3 -10 28 10 58 5 10 4 10 3 10 0 3 -10 21 11 65 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 0 3 10 4 10 PD1 E F + 8 D C f o % + 8 D C f o % 80 60 40 20 0 80 60 40 20 0 *** *** *** *** * * * * ** * DMSO ACBI1 50nM ACBI1 100nM AU-15330 50nM AU-15330 100nM N N (CCR7+ CD45RA+) E E (CCR7- CD45RA+) EM EM (CCR7- CD45RA-) CM CM (CCR7+ CD45RA-) * * N N (CCR7+ CD45RA+) E E (CCR7- CD45RA+) EM EM (CCR7- CD45RA-) CM CM (CCR7+ CD45RA-) DMSO CMP14 50nM CMP14 100nM FHT-1015 50nM FHT-1015 100nM Donor 3 8 DMSO 17 ACBI1 50nM 16 13 ACBI1 100nM 21 10 AU-15330 50nM 10 15 AU-15330100nM 19 9 4 10 3 10 0 4 10 3 10 0 4 10 3 10 0 4 10 3 10 0 4 10 3 10 0 29 46 45 26 50 18 51 24 54 18 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 Donor 4 10 DMSO 13 ACBI1 50nM 17 14 ACBI1 100nM 21 17 AU-15330 50nM 15 18 AU-15330100nM 19 16 4 10 3 10 0 4 10 3 10 0 4 10 3 10 0 4 10 3 10 0 4 10 3 10 0 33 44 44 25 43 18 44 23 46 20 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 Donor 5 6 DMSO 18 ACBI1 50nM 18 18 ACBI1 100nM 23 24 4 10 3 10 0 4 10 3 10 0 4 10 3 10 0 26 50 37 27 32 21 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 Donor 1 DMSO CMP14 50nM CMP14 100nM FHT-1015 50nM FHT-1015 100nM 7 12 13 12 18 16 8 11 9 8 4 10 3 10 0 4 10 3 10 0 4 10 3 10 0 4 10 3 10 0 4 10 3 10 0 45 36 56 19 47 19 50 31 62 21 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 Donor 2 DMSO CMP14 50nM CMP14 100nM FHT-1015 50nM 8 20 21 22 20 22 12 24 FHT-1015 100nM 18 16 4 10 3 10 0 4 10 3 10 0 4 10 3 10 0 4 10 3 10 0 4 10 3 10 0 29 43 36 21 34 24 31 33 45 21 0 4 10 5 10 Donor 6 DMSO 0 4 10 5 10 CMP14 50nM 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 CMP14 100nM FHT-1015 50nM FHT-1015 100nM 5 10 4 10 3 10 2 10 0 2 -10 7 R C C 16 15 41 28 5 10 4 10 3 10 2 10 0 2 -10 15 7 60 18 5 10 4 10 3 10 2 10 0 2 -10 12 8 55 25 5 10 4 10 3 10 2 10 0 2 -10 21 13 45 21 5 10 4 10 3 10 2 10 0 2 -10 23 11 48 18 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 0 4 10 5 10 CD45RA Figure S5. Evaluation of mSWI/SNF small molecule inhibitors and degraders in human T cells, Related to Figure 5. A. Western blot analysis of SMARCA4 levels in control, ACBI- or AU-15330- treated human CD8+ T cells at the indicated concentrations, profiled at Day 9 of the chronic stimulation protocol, compared to actin as loading control. B. FACS plots depicting PD1/TIM3 populations in CD8+ T cells from the indicated donors at Day 9 of chronic stimulation, treated with 50nm and 100nM of SMAR- CA4/2 degraders and inhibitors. C. FACS plots and associated MFI quantifications of CD39 surface levels in CD8+ T cells from the indicated donors at Day 9 of chronic stimulation, treated with 50nm and 100nM of SMARCA4/2 degraders (top) and inhibitors (bottom). D. FACS plots depicting CD45RA/CCR7 populations in CD8+ T cells from the indicated donors at Day 9 of chronic stimulation, treated with 50nm and 100nM of SMARCA4/2 degraders and inhibitors. E. Bar graph depicting % of CD8+ T cells in Naïve (N), Effector (E), Effector Memory (EM) and Central Memory (CM) populations based on CD45RA and CCR7 surface levels, in DMSO, ACBI1, and AU-15330 conditions. Error bars represent mean ± SEM of 3 independent CD8+ T cell donors. F. Bar graph depicting % of CD8+ T cells in Naïve (N), Effector (E), Effector Memory (EM) and Central Memory (CM) populations based on CD45RA and CCR7 surface levels, in DMSO, CMP14 and FHT1015 conditions. Error bars represent mean ± SEM of 3 independent CD8+ T cell donors. Error bars represent mean ± SEM of 3 technical replicates per donor. *p < 0.05, **p < 0.01, ***p < 0.001. ****p < 0.0001. AU-1533050nMAU-1533050nM A Donor 3 DMSO ACBI1 50nM ACBI1 100nM AU-15330 50nM AU-15330100nM Donor 2 DMSO CMP14 50nM CMP14 100nM FHT-1015 50nM FHT-1015 100nM Figure S6 8 14 12 12 5 10 10 23 11 15 67 7 5 10 4 10 3 10 0 3 -10 11 16 67 6 5 10 4 10 3 10 0 3 -10 3 -10 0 3 10 4 10 5 10 3 -10 0 3 10 4 10 5 10 3 -10 0 3 10 4 10 70 8 5 10 5 10 4 10 3 10 0 3 -10 4 10 3 10 0 α F N T 70 6 3 -10 0 3 10 4 10 5 10 15 14 11 13 5 10 4 10 3 10 0 5 10 4 10 3 10 0 5 10 4 10 3 10 0 8 18 10 8 5 10 4 10 3 10 0 41 26 51 25 48 23 39 35 61 20 0 3 10 4 10 5 10 0 3 10 4 10 5 10 0 3 10 4 10 5 10 0 3 10 4 10 5 10 0 3 10 4 10 5 10 4 10 3 10 0 3 -10 17 12 60 3 -10 0 3 10 4 10 11 5 10 Donor 4 DMSO 5 10 10 22 4 10 3 10 0 3 -10 α F N T 5 10 4 10 3 10 0 3 -10 5 10 4 10 3 10 0 3 -10 ACBI1 50nM 10 9 49 19 64 3 -10 0 3 10 4 10 5 10 3 -10 0 3 10 4 10 17 5 10 ACBI1 100nM 6 8 71 3 -10 0 3 10 4 10 15 5 10 5 10 4 10 3 10 0 3 -10 5 10 4 10 3 10 0 3 AU-15330 50nM AU-15330100nM 8 11 11 9 10 5 IFNγ 4 10 3 10 0 3 -10 63 17 3 -10 0 3 10 4 10 5 10 64 3 -10 0 3 10 4 10 17 5 10 DMSO ACBI1 50nM ACBI1 100nM AU-15330 50nM AU-15330 100nM 100 Donor 1 C DMSO CMP14 50nM CMP14 100nM FHT-1015 50nM FHT-1015 100nM IFNγ B s l l e c e v i l A % Donor 3 Donor 4 Donor 5 100 50 0 100 50 0 100 50 0 D ) 1 y a D t a M 1 m o r f ( s l l e c n o i l l i M 2000 1500 1000 500 0 1.2 0.8 0.4 0.0 s l l e c d e t s u a h x e + 3 M T + 1 D P I l o r t n o c O S M D f o % E F 3 M T I s l l e c e v i l A % Donor 2 Donor 6 50 0 100 50 0 100 50 0 i ) + V n x e n n A ( s l l e c c i t o t p o p a % FHT-1015 DMSO ACBI1 50nM ACBI1 100nM AU-15330 50nM AU-15330 100nM 100 80 60 40 20 0 Donor 3 4 5 DMSO Treatment from Day3 Treatment from Day6 Treatment from Day9 N°of stimulations Days 1 3 2 6 3 9 4 13 5 16 N°of stimulations Days 1 3 2 6 3 9 4 12 5 15 ACBI1 2000 1500 1000 500 0 AU-15330 500 400 300 200 100 0 3 6 9 13 16 3 6 9 Days 13 16 3 6 9 13 16 Profiled at Day13 DMSO ACBI1 AU-15330 FHT-1015 s l l e c d e t s u a h x e + 9 3 D C l o r t n o c O S M D f o % 1.2 0.8 0.4 0.0 Profiled at Day13 Treatment from Day3 Treatment from Day6 Treatment from Day9 DMSO ACBI1 AU-15330 FHT-1015 Mouse CD8+ T cells DMSO CMP14 50nM CMP14 100nM FHT-1015 50nM FHT-1015 100nM DMSO FHT-1015 100nM 5 10 4 10 3 10 0 3 -10 PD1 1 0 3 10 5 10 4 10 3 10 0 3 -10 83 5 4 10 5 10 4 10 3 10 0 3 -10 55 5 4 10 12 0 3 10 5 10 4 10 3 10 0 3 -10 46 5 4 10 2 0 3 10 5 10 4 10 3 10 0 3 -10 79 5 4 10 3 0 3 10 68 5 4 10 8 0 3 10 5 5 10 4 10 3 10 0 4 α F N T 91 0 5 10 4 10 3 10 0 0 3 10 4 10 5 10 15 81 4 0 0 3 10 4 10 5 10 IFNγ Figure S6. mSWI/SNF pharmacologic inhibition attenuates exhaustion of human and mouse T cells, Related to Figure 5,6. A. FACS plots depicting IFNγ/TNFα populations in CD8+ T cells from the indicated donors at Day 9 of chronic stimulation, treated with 50nm and 100nM of SMARCA4/2 degraders and inhibitors. B. Bar graphs depicting the percentage of alive cells at Days 3, 6, 9, 13 and 16 of chronic stimulation, for the indicated donors, treated with 50nm and 100nM of SMARCA4/2 degraders (left) and inhibitors (right). Error bars represent mean ± SEM of three technical replicates. C. Quantification of the percentage of apoptotic cells indicated by AnnexinV staining, at Day 16 of the chronic stimulation protocol, for the indicated donors treated with 50nm and 100nM of SMARCA4/2 degraders. D. Bar graphs depicting cell number upon treatment with ACBI1, AU-15330 or FHT-1015 (10^6 cells/10^ 6 cells at Day 0), initiated at the indicated time points, in one human CD8+ T cell donor. E. Quantification of the PD1+TIM3+ population in human CD8+ T cells upon treatment with ACBI1, AU-15330 or FHT-1015, initiated at the indicated time points. Data from one CD8+ T cell donor are represented. F. FACS plots depicting PD1/TIM3 (right) and IFNγ/TNFα (left) populations in mouse CD8+ T cells at Day 9 of chronic stimulation, treated with the indicated concentrations of SMARCA4/2 inhibitors. A B Z-score 3 2 1 0 −1 −2 −3 Z-score 3 2 1 0 −1 −2 −3 Z-score 3 2 1 0 −1 −2 −3 C Z-score 3 2 1 0 −1 −2 −3 ATAC-seq union sites Abs LogFc >1 Figure S7 Log2FC (RPKM) RNA-seq (Day9-Ch) E CTRL-1 ACBI1 AU−15330 CTRL-2 CMP14 FHT-1015 Donor1 Donor2 Donor3 Donor4 0 5 2 0 0 2 0 5 1 0 0 1 0 5 0 l ) e u a v − P . j d a ( 0 1 g o L − −0.4 −0.2 0.0 0.2 0.4 0.6 PC1: 36.87% D 4 0 . % 4 6 3 2 . : 2 C P 2 . 0 0 0 . 2 . 0 − . 4 0 − 6 . 0 − G D 3 . A C B I 1 D 3 . A C B I 1 D 4 . A C B I 1 D 4 . A C B I 1 D 3 . C T R L - 1 D 3 . C T R L - 1 D 4 . C T R L - 1 D 4 . C T R L - 1 D 3 . A U 1 5 3 3 0 D 3 . A U 1 5 3 3 0 D 4 . A U 1 5 3 3 0 D 4 . A U 1 5 3 3 0 D 3 . C T R L - 1 D 3 . C T R L - 1 D 4 . C T R L - 1 D 4 . C T R L - 1 D 1 . C M P 1 4 D 2 . C M P 1 4 D 1 . C T R L - 2 D 2 . C T R L - 2 ACBI1 vs. CTRL-1 AU15330 vs. CTRL-1 CMP14 vs. CTRL-2 FHT-1015 vs. CTRL-2 RNA-seq D 1 . C T R L - 2 D 2 . C T R L - 2 D 1 . F H T - 1 0 1 5 D 2 . F H T - 1 0 1 5 F RNA-seq (Day9-Ch) 0 5 1 0 0 1 0 5 0 l ) e u a v − P . j d a ( 0 1 g o L − 0 4 0 3 0 2 0 1 0 l ) e u a v − P . j d a ( 0 1 g o L − l ) e u a v − P . j d a ( 0 1 g o L − 0 1 8 6 4 2 0 I A C B 1 v s C T R L - 1 A U - 1 5 3 3 0 v s C T R L - 1 C M P 1 4 v s C T R L - 2 F H T - 1 0 1 5 v s C T R L - 2 AU−15330 166 207 ACBI1 64 48 217 139 FHT-1015 580 70 142 362 97 528 CMP14 −5 0 Log2 Fold Change 5 −5 0 Log2 Fold Change 5 −5 0 Log2 Fold Change 5 −5 0 Log2 Fold Change 5 Day 9 Treatment Downregulated Genes (N = 580) Z-score 3 2 1 0 −1 −2 −3 Z-score 3 2 1 0 −1 −2 −3 Z-score Z-score 3 2 1 0 −1 −2 −3 3 2 1 0 −1 −2 −3 D 4 . A C B I 1 D 4 . A C B I 1 D 3 . A C B I 1 D 3 . A C B I 1 D 3 . C T R L - 1 D 3 . C T R L - 1 D 4 . C T R L - 1 D 4 . C T R L - 1 D 4 . A U 1 5 3 3 0 D 4 . A U 1 5 3 3 0 D 3 . A U 1 5 3 3 0 D 3 . A U 1 5 3 3 0 D 3 . C T R L - 1 D 3 . C T R L - 1 D 4 . C T R L - 1 D 4 . C T R L - 1 D 1 . C T R L - 2 D 2 . C T R L - 2 D 1 . C M P 1 4 D 2 . C M P 1 4 D 1 . C H R - 2 D 2 . C T R L - 2 D 1 . F H T - 1 0 1 5 D 2 . F H T - 1 0 1 5 chr5:131,809,559-131,836,313 [0 - 4.76] K Mouse CD8+ T cells H GO_DEFENSE_RESPONSE HALLMARK_INTERFERON_GAMMA_RESPONSE HALLMARK_INTERFERON_ALPHA_RESPONSE REACTOME_CYTOKINE_SIGNALING_IN_IMMUNE_SYSTEM GO_REGULATION_OF_CELL_POPULATION_PROLIFERATION KEGG_CYTOKINE_CYTOKINE_RECEPTOR_INTERACTION HALLMARK_INFLAMMATORY_RESPONSE GO_RESPONSE_TO_CYTOKINE GO_CYTOKINE_MEDIATED_SIGNALING_PATHWAY GO_INFLAMMATORY_RESPONSE HALLMARK_TNFA_SIGNALING_VIA_NFKB GO_IMMUNE_EFFECTOR_PROCESS GO_REGULATION_OF_IMMUNE_SYSTEM_PROCESS GO_POSITIVE_REGULATION_OF_IMMUNE_SYSTEM_PROCESS HALLMARK_IL2_STAT5_SIGNALING 0 5 10 -Log10 (P-value) 15 20 I Z-score 0 1 −1 −2 −3 3 2 Degraders Inhibitors JUN GZMB IFNG HAVCR2 TIGIT ENTPD1 PRDM1 PDCD1 . D 3 C T R L - 1 . D 3 C T R L - 1 . D 4 C T R L - 1 . D 4 C T R L - 1 . D 3 A C B 1 _ 1 I . D 3 A C B 1 _ 2 I . D 4 A C B 1 _ 1 I . D 4 A C B 1 _ 2 I . D 3 A U 1 5 3 3 0 _ 1 . D 3 A U 1 5 3 3 0 _ 2 . D 4 A U 1 5 3 3 0 _ 1 . D 4 A U 1 5 3 3 0 _ 2 . D 1 C T R L - 2 . D 2 C T R L - 2 . D 1 C M P 1 4 . D 2 C M P 1 4 . D 1 F H T - 1 0 1 5 . D 2 F H T - 1 0 1 5 J CTRL-1 ACBI1 AU-15330 CTRL-2 CMP14 FHT-1015 M N 100 80 60 40 20 i y t i c x o t o t y c % s D I f i t o M F T 0 . 0 0 . 5 1 . 0 1 . 5 2 . 0 RNA (relative cpm) BATF E2F.1 BCL6.1 NFY AHR MYB.4 MYB.3 IRF.1 STAT.2 IRF.3 ARI5A DDIT3.CEBPA CENBP LIN54 GMEB2.2 E2F.4 ZBTB14 ZNF384.1 SNAI2 CTCF SOX.2 HINFP1.1 FOX.1 MEF2 Ebox.CATATG MIES TBX.4 CREB.ATF.1 NFI.1 CREB.ATF.3 ZNF410 SPI TCF.LEF HD.15 HIC.1 AP1.1 HD.25 ETS.1 CREB.ATF.2 RUNX.1 CMP14 vs CTRL FHT-1015 vs CTRL BATF NFY E2F.1 PROX1 MYB.3 IRF.1 E2F.4 ARI5A AHR ZBTB14 MYB.4 BCL6.1 CUX.2 HINFP1.1 HD.14 CTCF DDIT3.CEBPA ZNF490 IRF.3 GFI POU.1 CCAAT.CEBP HIF CUX.1 TBX.3 ZNF317 HAND1 TCF.LEF NFI.1 Ebox.CATATG MEF2 HD.12 HIC.1 TBX.1 TBX.4 ETS.1 AP1.1 HD.15 RUNX.1 CREB.ATF.2 Accessibility UP DOWN Accessibility UP DOWN −0.2 0.2 0.0 Model Coefficient −0.2 0.0 0.2 Model Coefficient B16-OVA in vitro killing (OT-1 T cells) 100 i y t i c x o t o t y c % 80 60 40 20 0 48h DMSO FHT-1015 32 64 128 2 4 8 32 64 128 16 T:E 0 24h 2 4 8 16 T:E L ATAC-seq Log2FC 3 2 1 0 -1 -2 -3 log2FC (RPKM) Mouse CD8+ T cells C1 (n=24188) Low accessibility C2 (n=3100) Naive C3 (n=7928) Early activation C4 (n=3900) Activation/ D9 C5 (n=5982) Activation C6 (n=4672) Day9 C M P 1 4 v s C T R L F H T - 1 0 1 5 v s C T R L Figure S7. Chromatin accessibility, gene expression and T cell effector functional profiling in human and mouse T cells treated with mSWI/SNF inhibitors and degraders, Related to Figure 6,7. A. Volcano plots of changes in accessibility upon mSWI/SNF disruption (100nM); inhibitors and degraders are grouped. Significantly up-regulated and down-regulated genes (Abs Log2FC> 1, adjP< 0.05) are colored in red and blue, respectively. Differential accessibility was calculated from n=2 indepen- dent human CD8+ T cell donors. B. Hierarchical-clustered heatmap of ATAC-seq sites changes upon treatment with the indicated compounds (100nM). The top 5% most significant differentially accessible peaks are shown. C. K-means clustering (n=2) of ATAC-seq log2 fold changes upon treatment with the four inhibitors/degraders relative to DMSO control (averaged independent donors). D. Principal component analyses (PCA) of RNA-seq profiles of Control (CHR), and ACBI1, AU-15330, CMP14 or FHT-1015-treated human CD8+ T cells (100nM), at Day9. CHR.1 and CHR.2 are the controls for the ACBI1/AU-15330 and CMP14/FHT-1015 experiments, respectively. E. Volcano plots of changes in gene expression upon treatment with the indicated compounds (100nM). Significantly up-regulated and down-regulated genes (Abs Log2FC > 1, adjp< 0.05) are colored in red and blue, respectively. Differential expression was calculated from n=2 independent human CD8+ T cell donors. F. Venn diagram showing the overlap in down-regulated genes (LogFC <-1) upon treatment with ACBI1, AU-15330, CMP14 or FHT-1015 (100nM). G. Hierarchical-clustered heatmaps of gene expression changes upon treatment with the indicated compounds (100nM). The top 1000 differentially expressed genes are shown. H. Gene ontology analysis of the 580 down-regulated genes shared among treatment with ACBI1, AU-15330, CMP14 and FHT-1015. I. Z-scored heatmap reflecting the expression of selected genes following mSWI/SNF inhibitor or degrader treatments. J. Representative ATAC-seq tracks of untreated (CTRL) or treated T cells over the IRF1 locus with corresponding RNAseq gene expression levels. K. Top 40 coefficients of logistic regression models fitting motif counts across all sites to changes in accessibility upon treatment with CMP14 or FHT-1015 in mouse CD8+ T cells. L. Heatmap displaying ATAC-Seq log2 fold change values upon treatment with CMP14 or FHT-1015, with clusters from Fig. S3a indicat- ed. M. Schematic of anti-CD19 CAR-T construct used for CAR-T cell studies. N. In vitro killing efficiency of B16-OVA cells by OT-1 T cells treated with FHT-1015 (100nM), at Day9 of chronic stimulation. B16-OVA cells and OT-1 T cells were co-incubated for 24 hours (left) or 48 hours (right) at the indicated Target-to-Effector (T:E) ratios. Means ± SEM from 4 technical replicates are shown. IRF1IRF1IRF1IRF1IRF1
10.1007_s10796-023-10369-7
Information Systems Frontiers https://doi.org/10.1007/s10796-023-10369-7 Snakes and Ladders: Unpacking the Personalisation‑Privacy Paradox in the Context of AI‑Enabled Personalisation in the Physical Retail Environment Ana Isabel Canhoto1  · Brendan James Keegan2 · Maria Ryzhikh3 Accepted: 4 January 2023 © The Author(s) 2023 Abstract Artificial intelligence (AI) is expected to bring to the physical retail environment the kind of mass personalisation that is already common in online commerce, delivering offers that are targeted to each customer, and that adapt to changes in the customer’s context. However, factors related to the in-store environment, the small screen where the offer is delivered, and privacy concerns, create uncertainty regarding how customers might react to highly personalised offers that are delivered to their smartphones while they are in a store. To investigate how customers exposed to this type of AI-enabled, personalised offer, perceive it and respond to it, we use the personalisation-privacy paradox lens. Case study data focused on UK based, female, fashion retail shoppers exposed to such offers reveal that they seek discounts on desired items and improvement of the in-store experience; they resent interruptions and generic offers; express a strong desire for autonomy; and attempt to control access to private information and to improve the recommendations that they receive. Our analysis also exposes contradictions in customers’ expectations of personalisation that requires location tracking. We conclude by drawing an anal- ogy to the popular Snakes and Ladders game, to illustrate the delicate balance between drivers and barriers to acceptance of AI-enabled, highly personalised offers delivered to customers’ smartphones while they are in-store. Keywords Artificial intelligence · Personalisation · Privacy · Personalisation-privacy paradox · Retail · Geo-location 1 Introduction Artificial intelligence (AI) is expected to transform business practice in in-store retailing (Davenport et  al., 2020), by bringing to the physical retail environment the kind of mass personalisation that is already common in online commerce (Kumar et  al., 2017). Personalisation benefits retailers because targeted messages get noticed amid the noise of other communications (Balan & Mathew, 2020), increase sales, and support customer intimacy, involvement with the brand * Ana Isabel Canhoto a.i.canhoto@sussex.ac.uk Brendan James Keegan brendan.keegan@mu.ie Maria Ryzhikh maria.ryzhikh@gmail.com 1 University of Sussex, Sussex, UK 2 Maynooth University, Maynooth, Ireland 3 Weber-Stephen Products EMEA GmbH, Berlin, Germany (Gardino et al., 2021) and customer loyalty (Pappas et al., 2018). Moreover, campaign response can be monitored directly and corrective action can be taken promptly, thus improving conversion rate (Chou & Shao, 2021). In the physical retail environment, personalisation is typically provided by the salesperson, which has several limitations. On the supply side, sales staff have access to limited customer data in-store which constrains their ability to adapt their recommendations (van de Sanden et al., 2019). On the demand side, increasingly, customers do not want to interact with a salesperson, particularly in the wake of the Covid-19 pandemic (Mondada et al., 2020; Yoganathan et al., 2021). Where technology is used for in-store recommendations, but not drawing on AI, these are based on customer segmentation rather than individual behaviours. Moreover, such recommendations tend not to reflect real time changes in the context, such as the customer’s location, the store’s inventory levels or the level of crowding in specific area. AI can overcome these limitations of in-store personalisation, due to its ability to integrate multiple sources of information, and create data-driven offers (Kietzmann et al., 2018). Moreover, given that many retail customers use their Vol.:(0123456789)1 3 mobile phones while shopping (Rippé et al., 2017), retailers can deliver the AI-created, targeted messages to customers’ mobile devices while they are in—or near – their store. We refer to this type of targeted offer, which has been personalised by artificial intelligence technology and is delivered to individual shoppers’ phones, in the physical retail environment as “artificial intelligence enabled personalisation” (hereafter referred to as AI-EP). While there is a rich body of work examining consumer experiences of personalisation in the online environment (see Boerman et al., 2017 for a review), this has not been replicated for physical retail (van de Sanden et al., 2019). However, attitudes towards personalisation vary signifi- cantly with the context in which it takes place (Aguirre et al., 2016). First, as consumers’ motivations vary for online vs in- store retail (Haridasan & Fernando, 2018), their perceptions and evaluation of personalisation in the physical environ- ment may differ from those identified in the extant literature on personalisation. Second, the interface through which the message is delivered influences the perception of the extent to which the message has been personalised, with high qual- ity interfaces increasing the perception of personalisation (Ameen et al., 2022). The small screen of mobile phones may impact negatively on consumers’ involvement with the message (Grewal et al., 2016), offsetting their suitability as targeting devices. Third, privacy concerns negatively impact consumers’ evaluation of personalisation in online shopping environments (Li et al., 2017). However, paradoxically, this effect was not detected in Ameen et al (2022)’s study of consumer interactions with smart technologies in shop- ping malls. In summary, while, from a technical perspec- tive, AI-EP may be similar to online personalisation, factors related to the context of message delivery (in-store), the for- mat of message delivery (small screen) and the salience of privacy concerns in different media suggest that consumer acceptance of personalisation may vary significantly across the two environments. This uncertainty represents a limita- tion in the current conceptual understanding of consumer acceptance of personalisation and is also a key barrier to adoption AI by businesses (Bughin et al., 2017). That is why Ameen et al (2022), Riegger et al (2021) and van de Sanden et al (2019), among others, have called for empiri- cal research examining consumers’ attitudes towards AI-EP. This paper aims to advance the conceptual understanding of AI-EP by investigating the following research question: “How do consumers experience and respond to AI-EP?”. To frame this investigation, we draw on the personali- sation-privacy paradox, particularly Sutanto et al’s (2013) research on smartphone users. This lens allows us to go beyond understanding whether consumers accept or reject AI-EP, to identify the reasons for their behaviour, as well as how they manage any tensions that may arise while inter- acting with AI-EP, as urged by Riegger et al (2021). We Information Systems Frontiers investigate these dynamics empirically by focusing on a UK fashion retail personalisation app. We focused on one spe- cific app in order to develop an holistic understanding of the usage climate of this technology, as recommended by Wang et al (2015). We chose fashion retail because this is a highly dynamic industry, which benefits from targeted, location- based communication with customers (Kumar et al., 2017); and because this is one of the most promising sectors for AI applications (Davenport et al., 2020). Finally, we chose the UK because it is at the forefront of the digital retailing revolution (Ameen et al., 2022). Given that AI-EP is a relatively unexplored phenomenon (Riegger et al., 2021), and the paradoxical findings that are beginning to emerge (e.g., Ameen et al., 2022), we opted for an exploratory approach. Specifically, a qualitative case study which included in-depth interviews with 18 female, millennial fashion retail shoppers, who had been exposed to a personalised advert. The paper makes three contributions. First, we show that customers welcome this innovative way of interacting with them in the retail environment. However, their experiences with online personalisation create very high expectations of the extent of AI-EP, as well as additional services such as creation of wish lists or the ability to edit their prefer- ences. These findings can guide practitioners’ investment in AI-EP. Second, we provide empirical evidence of how the impact of the context of message delivery, the format of message delivery and the salience of privacy concerns differs for AI-EP vs online personalisation. This can guide the application of findings from extant research, and guide future research efforts. Third, we identify the content and process gratifications derived from AI-EP, extending Sutanto et al (2013)’s work on the manifestation of the personalisation-privacy paradox among smartphone users. The paper is organized as follows. Section 2 considers the emerging literature on the opportunities and challenges for AI use in physical retail. Section 3 presents the theoretical background. Section 4 articulates the approach to data col- lection and analysis. Section 5 communicates the empirical findings. Section 6 discusses the findings, and uses the motif of the Snakes and Ladders game to capture the factors that support or prevent acceptance of AI-EP, Finally, Sect. 7 cap- tures the contributions of this empirical investigation to the advancement of theory and practice of AI deployment for personalisation in physical retail environments. 2 Research Background 2.1 Prior Studies in AI in Retail AI studies have seen a significant amount of attention in recent years from many different disciplines, and applied 1 3 Information Systems Frontiers to many different settings, including retail (Dwivedi et al., 2021). Several studies propose that AI can help retailers develop new and innovative applications from the various datasets available to them (e.g., Davenport et al., 2020), and in doing so, achieve competitive advantage. However, they tend to lack empirical evidence, and to overlook the customer perspective. There is also a growing a body of work focusing on the obsta- cles to effective use of AI (e.g., Boratto et al., 2018). Authors mention the risk of consumer backlash and of negative impact for firms. Though, the lack of customer focused research results in insufficient understanding of consumers’ perceptions of AI use in retail. In turn, the literature on digital personalisation (e.g., Boerman et al, 2017) suggests that AI-EP could enhance but also frustrate customers. Yet, except for Ameen et al (2022), these studies examine personalisation in controlled experiments rather than actual in-store experience. Finally, the effectiveness of personalisation efforts tends to be limited by customers’ privacy concerns (e.g., Aguirre et al, 2016). While some of these studies focus on smartphones (e.g., Sutanto et al., 2013), they provided limited insight into how customers manage the tensions arising. Table 1 summarises the notable themes identified in the stream of literature related to AI and its use for personalisa- tion. The right-hand column emphasises the research gaps. 2.2 Personalisation‑Privacy Paradox The review of the literature revealed a lack of customer focused, evidenced based understanding of how AI-EP benefits retail customers, and which factors may create resistance to acceptance of AI-EP or destroy value for customers. While personalisation can bring benefits to consumers, they may resist personalisation if they deem that the collection and use of personal data that underpin personalisation is too invasive (Moore et  al., 2015). This tension has been termed the Personalisation-Privacy paradox.1 To unpack the conditions under which the personalisation-privacy paradox manifests in each context, it is necessary to identify the gratifications that users derive from interacting with the medium through which personalisation is delivered, as well as their desires and concerns about information privacy (Sutanto et al., 2013). 2.2.1 Gratifications from Personalisation Sutanto et al (2013) put forward two types of gratification arising from personalisation: content gratification, referring 1 The term “personalisation-privacy paradox” is also, sometimes, used to refer to the disparity between users’ privacy protection inten- tions and their privacy protection behaviours (e.g., Norberg, Horne & Horne 2007). to the enjoyment derived from the personalised message itself; and process gratification, referring to the enjoyment derived from the medium in which the personalised offer is delivered. The personalisation literature identifies various content related gratifications such as receiving offers that reflect cus- tomers’ preferences (Krishnaraju et al., 2016; Pappas et al, 2017) and context (Xu et al., 2011), reducing the effort or time required to complete the purchase (Tam & Ho, 2006), and enabling cost savings and other financial gains (Schmidt et al., 2020). However, personalised messages can also stir negative emotions such as irritation (Haghirian et al., 2005) or anger (Pappas et al., 2018), thus rendering personalisation efforts ineffective (Demoulin & Willems, 2019). Customers are likely to resist offers that are seen as a threat to their freedom of choice (Brehm & Brehm, 2013). AI-EP may be perceived as restricting the options available to customers, which may result in customers rejecting the AI offer, in order to reaffirm their autonomy (André et al., 2018). In turn, process gratification arises from the ability to control how messages are received (Brusilovsky & Tasso, 2004), such as being able to filter out certain messages, or to control when and how they are displayed (Sutanto et al., 2013). Research has also shown that being able to control which information is collected and how it is used increases message effectiveness (Tucker, 2014), while lack of trans- parency from firms has the opposite effect (Aguirre et al., 2015). AI algorithms are, typically, opaque (Burrell, 2016), preventing customers to see – and influence – how they produced a specific recommendation, which may result in resistance to AI-EP. While Sutanto et al (2013) found, in the context of smart- phones, that personalisation gives users process gratification but not content gratification, by and large, the personalisa- tion literature focuses on the latter (Boerman et al., 2017). 2.2.2 Privacy Concerns The effectiveness of personalisation efforts may be offset by users’ concerns over the privacy of their personal infor- mation (Awad & Krishnan, 2006). For instance, online ads that closely match customers’ browsing history reduce pur- chase intentions, because they raise concerns over firms’ surveillance practices (Aguirre et al., 2016). Customers set boundaries – psychological or physical – around their per- sonal data (Stanton & Stam, 2003), and attempts to cross those boundaries raise concerns, and are met with resistance (Xu et al., 2008). Customers manage information boundaries by selectively sharing or withholding information (Sutanto et al., 2013). In addition, they may purposefully provide false information, such as using a false name or birth date (Miltgen & Smith, 2019), when firms attempt to collect per- sonal data that they deem private. 1 3 Information Systems Frontiers d a o r b a y b d e c n e u fl n i y l l a c i t a m a r d e b n a c d n a , e l b 9 1 - d i v o C e h t g n i r u d d e s s e n t i w s a , s r o t c a f f o e g n a r - a t s n u d n a t l u c ffi i d n o i t c i d e r p e k a m s d n e r t l a n o s a e S c i m e d n a p - a c i l p p a m o r f s e m o c t u o e v i t i s o p y l t r e v o e c u d e d o t s e l a s , s c i t y l a n a e v i t c i d e r p , n o i t a t n e m g e s t e k r a m r e m o t s u c e h t g n i k o o l r e v o , s r e l i a t e r o t I A f o s n o i t n o i t a s i l a n o s r e p d n a , g n i t s a c e r o f e v i t c e p s r e p - e g a n a m a t a d r e m o t s u c s a h c u s s e s s e c o r p d n e - k c a B d e c n a h n e n e e b e v a h s i s y l a n a t e k s a b s e l a s d n a t n e m I A y b r a e p p a s e i d u t s t n a t x E . e c n e d i v e l a c i r i p m e t u o h t i w d e v o r p m i : s a h c u s s n o i t a v o n n i r o f r e w o p g n i s s e c o r p 8 1 0 2 , a m r a h S & m a y S ; 8 1 0 2 , . l a t e n n a m , l i a t e r n i I A f o l a i t n e t o p d e g a s i v n e e h t n o s u c o f s e i d u t S a t a d f o s m r e t n i l a i t n e t o p e v i s s e r p m i r e ff o n a c I A - z t e i K ; 1 2 0 2 , t s u R & g n a u H ; 0 2 0 2 , . l a t e t r o p n e v a D l i a t e R n i l a i t n e t o P I A d e fi i t n e d I p a G s m i a l C y e K e c r u o S l i a t e R n i I A n o s e i d u t s t n e c e r d e t c e l e S 1 e l b a T e m e h T , g n i t s a l g n o l d n a , t n a c fi i n g i s e v a h d l u o c h c i h w , g n i p i h s r e d a e l d n a s l l i k s f o k c a l , y t i l i b a l i a v a a t a d r o o p s e i d u t s y n a m t o n , t e Y . s m r fi r o f t c a p m i e v i t a g e n - a l u g e r d n a l a c i h t e d n a t n e m y o l p e d f o t s o c , n i - y u b s n o i t p e c r e p r e m u s n o c n o s u c o f s n o i t c i r t s e r y r o t 1 2 0 2 , . l a t e a v i r G - h s i n r a t n o i t a t u p e r d n a h s a l k c a b r e m o t s u c n i t l u s e r s a h c u s , y l e v i t c e ff e d n a y l t n e i c ffi e a t a d s s e c o r p ; 1 2 0 2 , . l a t e o n i d r a G ; 1 2 0 2 , . l a t e i d e v i w D ; 9 1 0 2 d l u o c y t i l a e r d n a e s i m o r p I A e h t n e e w t e b p a g e h T o t I A f o t n e m y o l p e d r o f t s i x e s e g n e l l a h c y n a M , . l a t e k c i r C ; 0 2 0 2 , . l a t e o l l i t s a C ; 8 1 0 2 , . l a t e o t t a r o B l i a t e R n i I A f o s e g n e l l a h C s s e c o r p e h t h t i w y s a e n u y l e t a m i t l u e r a t u b , s r e ff o f o w o h g n i d n a t s r e d n u f o t n e t x e e h t o t e n o g t o n e v a h t u b s r e m o t s u C . e v i t c e ff e e b o t n o i t a m r o f n i r e m o t s u c ; 3 1 0 2 , . l a t e o t n a t u S ; 1 2 0 2 , . l a t e r e g g e i R ; 6 1 0 2 , s e n o h p t r a m s r i e h t o t d e r e v i l e d s r e ff o d e s i l a n o s r e p s d n a m e d t i s a , I A h t i w g n i t c a r e t n i , o t y l e v i t a g e n , . l a t e l a w e r G ; 0 2 0 2 , . l a t e n y u r B e D ; 9 1 0 2 , . l a t e e s i r a y a m t a h t s n o i s n e t y n a e g a n a m s r e m o t s u c e g a t n a v d a e k a t o t n o i t a m r o f n i t i m b u s o t g n i l l i w e r a 1 2 0 2 , . l a t e n a h t a n a g o Y d n a , t n e m n o r i v n e g n i p p o h s e h t f o e d i s t u o , s t n e m , m e h t o t e u q i n u s r e ff o g n i k e e s e r a o h w , s r e m o t s u c n e d n a S e d n a v ; 3 1 0 2 , . l a t e o t n a t u S ; 1 2 0 2 e c n e i r e p x e e r o t s - n i e h t e n i m a x e o t t e y e v a h I A y b d e v i r e d s a 8 1 0 2 , . l a t e z t r i W ; 9 1 0 2 , . l a , . l a t e t e - i r e p x e d e l l o r t n o c n i n o i t a s i l a n o s r e p e n i m a x e s e i d u t S e t a r t s u r f s a l l e w s a s s e r p m i n a c n o i t a s i l a n o s r e P r e g g e i R ; 7 1 0 2 , . l a t e n a m r e o B ; 2 2 0 2 , . l a t e n e e m A n o i t a s i l a n o s r e P l a t i g i D d e l b a n e - I A o t d e t a l e r s n o i s n e t y f i t n e d i e t a d o t s e i d u t S t c a e r r o , t u o b a e l b a t r o f m o c n u l e e f s r e m o t s u C o l e t s a C ; 6 0 0 2 , n a n h s i r K & d a w A ; 6 1 0 2 , l a t e e r r i u g A y c a v i r P 1 3 Information Systems Frontiers The literature indicates that customers may be comfort- able disclosing information deemed to be relevant for the intended outcome (Xu et al., 2011), when access to the ser- vice is time critical (Hubert et al., 2017), and where the information is routinely requested in that context (Stanton & Stam, 2003). However, customers resist sharing information that is deemed sensitive, such as their health status (Sutanto et al., 2013); or which could be used for discrimination (Stanton & Stam, 2003). They also resist sharing informa- tion when they feel that they lack control over what data are collected, how data are used, and with whom they are shared (Liu et al., 2019; Schmidt et al., 2020). However, informa- tion boundaries vary across individuals and are dynamic. Namely, those customers that value information transpar- ency are also most likely to resist the data collection that underpins personalisation (Awad & Krishnan, 2006). Cus- tomers also change whether they share information depend- ing on the perceived gains or losses of each situation (Kar, 2020). The perception of being under surveillance is particu- larly prevalent in online interactions and in smart services (Bues et al., 2017). Therefore, in addition to providing privacy features (Awad & Krishnan, 2006), firms also need to identify which information customers are comfortable to share, and what trade-offs they are prepared to make in order not to break their personal information boundaries (Pentina et al., 2016). This is particularly relevant for AI-EP, given the need for large volumes of data to support the development of targeted offers (Davenport et al., 2020). 3 Research Design The aim of our study was to advance the conceptual under- standing of AI-EP by investigating the following research question: “How do consumers experience and respond to AI-EP?”. Hence, a qualitative, exploratory case study methodology (Sarker et al., 2018) was adopted. The unit of analysis was shoppers’ interactions with an AI-enabled smartphone application, in the context of fashion retail. This methodology offered an opportunity to collect primary data from customers in situ experiencing the AI-enabled per- sonalisation offer, guided by key studies in the field (e.g., Ameen et al., 2022; Riegger et al., 2021). It also offered the unique opportunity to collect rich and diverse perspectives from participants, as they reflected upon the hybrid digital- physical experience of AI-EP, extending previous works in the area, particularly Sutanto et al. (2013). In doing so, the method adopted allowed us to understand and analyse a broad range of participant views, and to theorise and concep- tualise (Eisenhardt, 1989), in line with other case studies that have examined the impact of technology upon personalisa- tion (e.g., Griva et al., 2021). 3.1 The Selected App The mobile app selected as the focus for this case study was the Regent Street App. The app was first launched in 2012 to enhance the shopping experience of visitors to this famous shopping district, in London (UK). As shown in Fig. 1, the app included the option to receive personalised offers while shopping in the area. To create and deliver these offers, the app combined “two technologies: geofencing beacons that use location aware to offer content to users within a specified proximity to the store and cloud-based artificial intelligence (AI) to ensure personal relevancy of offers” (Lemmon, 2017). Circa 80% of the stores in this shopping district joined the scheme, implementing the associated technology in their premises, such as beacons around the store and microchips in the items on sale (Scott, 2014), in addition to artificial intelligence programme to personalise the offers. Moreover, 98.6% of app users created a personal profile and signed up to receive personalised content (Lemmon, 2017). The AI-EP messages are delivered when app users are in the vicinity of the stores that signed-up to the app (Demp- sey, 2015), resulting in a 7.4% increase in response rate for AI-EP vs. untargeted offers (Lemmon, 2017). 3.2 Data Collection To gather customer experiences, we used in-depth, semi- structured interviews, to allow participants to articulate their actions and intentions towards the AI-EP, as well as implica- tions for their personal data. In order to recruit participants, one of the authors (who conducted all the interviews) positioned themselves outside a specific fashion store in Regent Street, which was known to use the Regent Street App for the delivery of AI-enabled personalised offers. As shoppers walked past the store, the interviewer approached them, showed them the advert in Fig. 2, and invited them to participate in an interview. This approach is in line with Kar (2020)’s recommendation that research on customer perceptions of digital technology should take place immediately after encounter with that technology. Some interviews took place outside the store, others at a nearby café. No financial incentives were offered to the interview participants. The interview protocol (Table 2) reflected the key themes identified in the extant literature. The questions focused on perceptions of the message rather than the technology underpinning it, as customers don’t always understand the technology behind personalisation. This approach allowed us to move beyond a simplistic view of positive vs. negative attitudes, and to understand the black box of the customers’ response (Belk, 2017). As resistance to AI-EP may depend on customer char- acteristics (Yoganathan et  al., 2021), we recruited an 1 3 Information Systems Frontiers Fig. 1 Case study App. Image source: http:// okosv aros. lechn erkoz pont. hu/ en/ node/ 558 homogeneous sample via purposive sampling (Bryman & Bell, 2015), to give direction to the data collected in support of the case study (Yin, 2012). We focused on female shop- pers aged 18 to 30 years old, because, in the UK, this demo- graphic group cares the most about looking trendy (YouGov, 2020). Women in this age group are twice as likely than men to agree that they spend a lot on clothes and to value immediate access to fashion items; and they are also more likely than men and then older women to shop at multiple retailers (YouGov, 2020). Consequently, this demographic group are a key target for high street fashion retailers’ pro- motional efforts. This demographic group are also more open than others to sharing their personal data with firms, given that they grew up in a digital world (Liu et al., 2019). However, women may resist AI, especially when outcomes are consequential (Castelo et al., 2019). We conducted 18 audio-recorded interviews, each lasting between 30 min and one hour. Each recording was transcribed with an average of 9,000 words, equating to just over 160,000 words in the final dataset. The data was checked for accuracy and prepared for analysis. 3.3 Data Analysis The interview data was analysed using NVIVO and fol- lowing Krippendorff's (2004) systematic approach to thematic analysis. As is customary of exploratory case studies in the information systems discipline (see Sarker et al., 2018), the theory was used to guide the design of the study and to set the general direction of data analysis. In practice, this meant that a preliminary coding book was developed based on the themes identified in the lit- erature, and this was used in stage 1 of data analysis to deductively code the transcripts into a) gratifications from personalisation, b) privacy concerns and c) reaction to AI-EP. Subsequently, in stage 2, for each of the themes 1 3 Information Systems Frontiers Fig. 2 Interview prompt Table 2 Interview protocol Section Stimulus AI-EP – Gratifications AI-EP – Outcomes AI-EP – Privacy concerns Question Participant receives targeted prompt. Upon opening the screen, the participant learns that the offer is exclusive to users of the Regent Street mobile app walking past that store, and who have bought in that store, previously 1. What do you think of this offer? 2. This offer has been personalised based on your location and shopping preferences. Is this offer useful? 3. How does it enhance your shopping experience? 4. When the brand sends real-time, relevant offers to your mobile phone, are they mostly trying to sell more, or try- ing to build relationships with customers like you, by serving your specific needs? 5. Do you think that the company will always make the best offer specifically for you? 6. Why do you suppose that? 7. Do personalised offers help you develop bonds with this brand? 8. Would receiving this type of offer discourage you from switching to another brand? Why? 9. Which personal information would enhance your experience with this retailer? [Probe for location and behavioural data] 10. Are you willing to share that information with the company, so that they can develop offers specifically for you? 11. Where should the limit be? 12. What are the benefits of letting the company access your personal data? 13. What are the risks of letting the company access your personal data? 14. Through the app, the company can track your movements not only in-store but also in the proximity of stores on Regent Street? How does that make you feel? in the code book, the analysis of the data proceeded in an inductive fashion, with subsequent codes emerging from the data. The final set of codes is depicted in Table 3. The findings emerging from this analytical process are presented in the next section, following a polyphonic account. This approach presents the range of perspectives offered by the research participants in order to develop a layered account of the phenomenon being investigated (Travers, 2001), as is customary of interpretive research. This is in contrast with identifying the dominant narrative or single shared reality typical of positivist approaches to data analysis (Sarker et al., 2018). 1 3 Table 3 Coding structure Aggregation 1st order 2nd order Illustrative quotes Information Systems Frontiers Gratifications Content Relevancy of the offer that is better than humans Time saving attributes to the customer experi- ence Financial benefits that are attractive to modern customer base through appetite for discounts Other benefits Process Message Delivery Information collection process Information use processes enhancing value to the in-store experience Privacy concerns Information boundaries Boundary management practices Information—Willingness to share Information—Desire to protect Acceptance of AI-EP Perceptions Positive Negative Behaviour Acceptance of AI-EP and Customers are willing to share information to receive personalisation offer Rejection due to irritation from notifications, interruptions, lack of control I mean if you can choose what you like and then they will remember it that would be so much easier to go and shop there and maybe you would buy a bit even more It’s really useful because you get to know what is there I would value it a lot. There is nothing to lose for customers and it is not like we are com- mitting to a sale of any sort or a purchase of any sort But also the things like pretty macarons or lemonade If I would receive an offer from a store I really like and I already have a 10% offer, I would definitely go inside and check out the stuff they have I would rather have a setting in application— right now I am shopping for my dad. Rather than registering it under me. Or buying gifts for him or for her rather but still that the information being given It would definitely help because I can make a profile of things I like. It is an amazing tool definitely I only share information about fashion. Only information where I know it can create value for me if somebody wants to track me down they can do it, they have (the data), anyway… but on the other hand, it does not really matter what they are going to do because they can have it anyway I want to know if it’s going to be used for more than just trying to fulfil my needs within the shop Sometimes I just want to have something which I already have, which is different from what I already have. So, personalizing is useful for me in terms of fashion If the company has bad intentions, there may be some downside in sharing the information Telling them about your style, so they would know what specific things to target to you, and maybe saying your age group and gender, because that might help them to target you towards particular things as well If I am not shopping I would not want that sort of notifications or if I am doing something else I do not know. If you end up passing there every day it could be quite annoying 1 3 Information Systems Frontiers 4 Findings 4.1 Gratifications from Personalisation Our participants were very positive about using the app while shopping in Regent Street and receiving personalised recommendations on their phones: “You are going to (Regent Street) in your free time, and want to have a nice day, and, through, the app it might be even nicer.” (Interviewee 3). Contrary to the participants in Sutanto et  al (2013)’s experiment, for whom personalisation via smartphone apps delivered process but not content gratification, our partici- pants identified both types of personalisation. The analysis of the data (Table 4) showed that the participants experienced relevance, time savings and financial gratifications, in line with the literature on online personalisation. However, for most of our interviewees, discounts seemed to be the main benefit expected from AI-EP, undermining the promise that this form of personalisation can increase basket variety and improve retailer profitability (Kumar et al., 2017). For those interviewees, discounts might be complemented by other benefits, such as time savings, but not replaced by them. As for process gratifications (Table 5), some deemed receiving personalised notifications on their phones as supe- rior to relying on shop window information to gain informa- tion about new products or about deals (e.g., participant 11); Table 4 Content Gratifications or receiving offers via e-mail (e.g., participant 16). However, many more commented that, at times, the volume of notifica- tions became a nuisance. This is particularly relevant for the Regent Street app, as this is a central London location, next to theatres, cafés and other leisure venues, as well as a com- muting route, as mentioned by participant 4. A high volume of notifications could result in information overload (e.g., participant 13), intrude in relaxation time (e.g., participant 6), as well as drain the phone’s battery (e.g., participant 18). While most participants mentioned the option of switching off their Bluetooth to stop notifications, this was an unsat- isfactory solution for many. Instead, many expressed the desire to control effortlessly when to receive notifications and what type, which is in line with literature on the role of customer autonomy in technology interactions (e.g., André et al., 2018). Opinions varied as to whether the app was an effective way of collecting and using information for AI-EP. Some, like participant 14, were happy with the data collection process. However, others felt that the app should integrate with other data sources (e.g., participant 7). Still, others, like interviewee 5, lamented the lack of ability to edit pur- chase histories, or to select when not to collect data (e.g., for gifts and other one-off purchases). Because AI doesn’t understand the reasons behind a purchase (Woo Kim & Duhachek, 2020), the research participants predicted that 1 3 Table 5 Process Gratifications Information Systems Frontiers one-off purchases would be added to their purchase his- tory, undermining the quality of future recommendations. Two interviewees (4 and 17) indicated an explicit desire to understand why they had received specific recommenda- tions. In Sutanto et al (2013)’s examination of app users’ willingness to share personal information, the process benefits referred to the experience with the medium itself (namely, navigation of the app). However, interviewees 3 and 17 also seemed to value process benefits at the level of the in-store experience broadly, emphasising the hybrid nature of AI-EP. 4.2 Privacy Concerns In terms of boundary management behaviours, as detailed in Table 6, we found various instances of selective information disclosure to tap into benefits. For instance, the interview- ees were willing to provide information directly into the app or via surveys (e.g., participant 9) to improve the accuracy and relevance of the resulting recommendations. They also engaged in redemptive behaviours (Stanton & Stam, 2003), whereby they shared information to reduce the losses gen- erated by irrelevant recommendations, as illustrated by 1 3 Information Systems Frontiers Table 6 Privacy Concerns Interviewee 5. The interviewees were also keen to engage in information withdrawal. In particular, they wanted to remove records of one-off purchases, as well as historical information that was no longer relevant (e.g., participant 6), corresponding to Stanton and Stam (2003)’s political and protective behaviours. However, those options were seen to be unavailable or too difficult to access. Finally, we did not find evidence of interviewees disclosing fake information to 1 3 Information Systems Frontiers manage the benefits and risks of AI-EP, contrary to what was reported in Miltgen and Smith (2019). The interviewees were aware that, by using the app, a range of companies could access their personal data, including the mobile phone operator, the app developer, and the fashion brand. This situation was seen as the default in the digital era, as illustrated by interviewee 6’s quote. As shown in Table 6, an in line with extant literature (e.g., Miltgen & Smith, 2019), the interviewees were willing to share information such as clothes’ size, specific body measures, preferred styles, or favourite colours, to obtain relevant recommendations. As Interviewee 13 said: “The use of these (types of) data does not bother me because I think it is a win–win situation”. In contrast, and in line with Sutanto et al. (2013), most were unwilling to share personal information which they did not deem essential for the task at hand, or which could leave them vulnerable to manipu- lation, nuisance, or fraud (e.g., Interviewee 4). The topics of location and social media data divided opin- ions, however. Regarding the former, interviewees 3, 12 and 15 expressed the view that sharing geo-tracking was a natu- ral extension of what already happens on other media and was useful to develop targeted offers. However, the others expressed reservation towards various aspects of the track- ing of their location. They described this activity as “creepy” (e.g., interviewee 11) and, in line with Schmidt et al. (2020), they expressed a strong desire to limit the app’s ability to track their movements (e.g., interviewee 2). Regarding social media data, interviewees 1, 5 and 15 were in favour. But the remaining felt that these data should be off limits to the app. Some, like interviewee 6, rejected this because they felt that the data would be too revealing; others, like interviewee 7, because social media data were deemed irrelevant. Two key nuances emerged regarding privacy concerns associated with AI-EP. The first nuance relates to control over access to personal information. Specifically, inter- viewees would be willing to share more information if they could be in control of what data was collected and when (e.g., interviewee 1), or if they were reassured that the app provider would not take advantage of the situation to access other areas of their phones (e.g., Interviewee 3). The second nuance refers to trusted parties. The app provider was, implicitly, a trusted party, but this sentiment did not necessarily extend to specific stores on the app, particularly smaller ones (see Interviewee 17) due to concerns of the lat- ter’s ability to fend off security attacks. On the other hand, there were other parties that the interviewees trusted more than the app provider – namely, Apple (as mentioned by interviewee 10). Table 7 Perceptions of AI-EP 1 3 Information Systems Frontiers 4.3 Acceptance of AI‑EP The literature’s enthusiasm for AI-EP (e.g., Bues et al., 2017) was mirrored in our research participants’ reactions. The analysis of the findings (Table 7) reveals that some participants found this type of offers interesting (e.g., participant 8), exciting (e.g., participant 1) and useful (e.g., participant 17). Many felt valued by the company behind the offers (e.g., participant 10) and, as a result, developed a positive attitude towards the company (e.g., participant 3), which indicates the potential of AI-EP for relational benefits (Liu et al., 2019). Having said that, 10 out of the 18 participants could not see any relational benefits. They expressed scepticism about the intentions behind AI-EP offers, seeing them as mostly an attempt to get customers to increase their expenditure (e.g., participant 12). Participant 4 also expressed scepticism about AI-EP’s ability to meet her needs, due to limitations of the technology, as well as the associated cost. Other negative emotions reported were annoyance (e.g., participant 16), and creepiness or the feeling of being stalked (e.g., participant 9). Some participants also reported a feeling of intrusion in what is meant to be a leisurely, relaxing activity, with interviewee 8 describing it as thus: “It’s like a sales assistant running out into the street and grabbing me.” In addition, interviewee 15 reported a fear of over-spending as a result of AI-EP, while participant 18’s comment that “You would be less aware of what is going on. Table 8 Behavioural Outcomes You would be in a loop” echoes the perceived threat to freedom of choice identified by Brehm and Brehm (2013). The positive sentiments translated in willingness to act on the offers delivered via AI-EP, particularly if they came in the form of exclusive, time-limited discounts, for their favourite stores, as exemplified by participant 3’s quote (Table 8). While participants 13 and 17 said that AI-EP might lead them to try new stores, most ignored offers from stores that they did not usually shop at, or which they were not familiar with. That is, it seems that AI-EP works better for customer retention than for customer acquisition, and for the pre-approach stage of the sales process, which contradicts claims that AI can add value at any stage of the sales process (e.g., Syam & Sharma, 2018). However, participants have very high expectations of AI-EP. While some are willing to accept some trial and error (participant 4), in general, they expect extremely tar- geted and unique offers (e.g., participant 8). This expectation might reflect the participants’ experience with personalisa- tion in the online environment, where users typically receive very unique recommendations (Griva et al., 2021). Failing to meet such expectations seems to result in disappointment with the app (e.g., participant 5, Table 8), rather than with the brand (e.g., participant 4, Table 7). This reaction is in contrast with extant literature on online personalisation (e.g., Baek & Morimoto, 2012), but aligned with literature on mobile shopping apps (e.g., Shankar et al., 2016). 1 3 5 Discussion The extant literature argues that fashion retailers may enhance the customer experience through the use of AI-EP by harnessing company-owned as well as external datasets to create highly individualised offers (van de Sanden et al., 2019). Though, the broader personalisation literature implies that the effectiveness of AI-EP may be compromised by pri- vacy concerns (Aguirre et al, 2016), and that AI-EP may even result in customer dissatisfaction, due to inflated expec- tations or negative experiences (e.g., Karumur et al., 2018). Our focus on the customer perspective, and the exploration of an actual in-store experience, provides empirical evidence of the tensions in place, and how customers navigate them, as discussed next. 5.1 The Personalisation‑Privacy Paradox in the AI‑EP Context Our findings are aligned with those from research on per- sonalisation in the online environment, which established that personalised offers may deliver content gratification in the form of relevance (Krishnaraju et al., 2016), plus time (Tam & Ho, 2006), and cost savings (Schmidt et al., 2020). Though, in AI-EP, the opportunity for cost savings seems to dominate over the other forms of content gratification mentioned in the online personalisation literature. The limited importance of relevance in AI-EP might reflect the nature of shopping in the physical environment where, typically, there are fewer options on display than in online shopping (Kumar et al., 2017). Therefore, custom- ers may feel less overwhelmed by choice in the physical environment. Moreover, some of our interviewees seemed to associate fashion shopping in the physical environment with an hedonic experience (Gardino et al., 2021), rather than a functional one. The pleasant nature of in-store shopping may explain the reduced importance of time savings in AI-EP vs. online personalisation. The resistance to suggestions by the AI could also indicate that customers do not trust that AI has the skill to make such recommendations (Woo Kim & Duhachek, 2020), given that fashion shopping is a task rich in intuition and subjectivity (Castelo et al., 2019). This familiarity with personalisation in online fashion retail suggests a compelling path for future adoption by retailers. However, the emphasis on discounts contradicts the predic- tion that AI-EP will generate additional sales opportunities and improve retailer profitability (e.g., Kumar et al., 2017) by prompting customers to consider complementary items, or generating impulse purchases (e.g., Griva et al., 2021). Our findings also show the need for a careful approach to the process of delivering the personalised offer. In line with previous research on personalisation online (e.g., Information Systems Frontiers Brusilovsky & Tasso, 2004) and on smartphones (Sutanto et al., 2013), many participants expressed a strong desire for controlling notifications and other aspects of message delivery. Moreover, we observed intricate interactions between the receipt of notifications and various contextual factors such as phone battery depletion, the purpose of visit (e.g., shopping vs meeting friends) or additional information provided. Customers also expressed a strong desire to be in control of the information held in the system and used to create personalised recommendations, which is line with findings from online personalisation research (e.g., Aguirre et al., 2015; Tucker, 2014). Moreover, customers wanted the ability to edit information held by the retailers and which they perceived to be undermining the quality of the AI-EP. However, it is not clear that enabling customers to engage in such boundary management behaviours (Stanton & Stam, 2003) would deliver the results sought by retailers. As shown in the context of online personalisation, messages need to be persuasive in order to be successful (Pappas et al, 2017); and fashion retailers need access to large and stable datasets about customers and their context (Ameen et al., 2022) in order for the AI to create high quality, persuasive messages. Some customers also expressed a desired to understand why they received specific recommendations. It will be difficult for retailers to meet this particular customer expectation because algorithms are opaque, and it is difficult to trace exactly which data inputs are generating which outputs (Burrell, 2016). As a result, some customers may reject the AI-EP offer to reaffirm their autonomy (André et al., 2018). Exposure to widespread collection of personal data in the online environment may have influenced our respondents’ willingness to share data for AI-EP (Stanton & Stam, 2003). Many also showed willingness to participate in ad-hoc data collection initiatives, as they saw these as an opportunity to improve their shopping experience. However, there were noticeable nuances in terms of comfort with disclosing certain types of personal data, which require a very careful approach from retailers in order not to violate customers’ personal information boundaries (Pentina et  al., 2016). Mobile apps are useful tools to collect data such as unique customer identifier and transaction history, due to the high penetration of mobile phones, and because they can be linked to individual users (Shankar et al., 2016). However, customers need to perceive a link between the information requested and the resulting offer (Xu et al., 2011). Moreover, firms need to avoid collecting information which customers deem likely to be misused, or to leave them in a vulnerable position. Some participants also opposed the collection of social media activity. Another data input that is essential for instore AI-EP is location (Schmidt et al., 2020). This can either be individual 1 3 Information Systems Frontiers data such as the customer’s whereabouts, or contextual data such as the weather or crowd levels (Verhoef et al., 2017). However, the emotionally charged descriptors used by some of our participants, indicate that customers intensely dislike extensive tracking in the physical environment. This presents a challenge for fashion retailers: one the one hand, location data enables them to take full advantage of AI’s capabilities for personalisation; on the other hand, customers may see this as an invasion of privacy (Xu et al., 2008), which may result in negative attitudes towards AI-EP and, ultimately, its rejection (Shankar et al., 2016). 5.2 Effectiveness of AI‑EP Based on our findings, attempts to use AI-EP for customer acquisition may be ineffective (Demoulin & Willems, 2019), or even detrimental (Baek & Morimoto, 2012) for the brand. This finding was somehow surprising given that the app considered in this case study was provided by a trusted party which offered discounts to a variety of stores in a given shopping district. Trust has been shown to impact the perception of a personalised offer (Aguirre et al., 2016) and, as such, familiarity with the Regent Street app might lead customers to be receptive to AI-EP attempts from new brands (Chen & Dibb, 2010). Furthermore, we found that customers expressed a strong desire for autonomy and freedom of choice, as reported in the context of online personalisation (Balan & Mathew, 2020). Though, while previous research focused on choice and agency in relation to the content of the message, we witnessed a willingness to control message delivery, too. Granting this flexibility might return a sense of control to customers (Brehm & Brehm, 2013), but may increase the complexity of the app (e.g., in terms of navigation), which will negatively impact the user experience (Shankar et al., 2016). Moreover, it reduces the retailers’ ability to collect data and deliver targeted messages (Chou & Shao, 2021). While AI can integrate multiple sources of customer, contextual and transactional data, our study exposes limitations to the extent of in-store personalisation (Ameen et al., 2022; Boratto et al., 2018). Namely, in contrast with the online environment, where personalisation may influence the search and evaluation stages (Davenport et al., 2020), AI-EP was revealed to be most valued at point of purchase stage, albeit not for payment purposes. Furthermore, whilst algorithms underpinning AI-EP need to be rigorously tested (Sutanto et al., 2013), our findings indicate that fashion shoppers have low tolerance for such trial and error. As in the online environment, consumer trust and positive emotions are essential for successful personalisation (e.g., Pappas, 2018). As with personalisation in the online environment (e.g., Pappas et al, 2017), customers have high expectations of AI-EP. The inflated expectations and the low tolerance for mistakes, are likely to result in disappointment and app abandonment (Riegger et al., 2021; Shankar et al., 2016), represents a waste of resources, and inability to continue collecting data about customers. Figure 3 presents an overarching view of how in-store AI-EP can enhance customer experiences, capturing both the enabling factors from content and process gratifications, and the detracting factors related to unmet process gratification expectations and from privacy concerns. We represent the AI- EP journey consisting of opportunities and threats for retailers, as encapsulated in the well-known game of Snakes and Ladders. This model highlights the potential as well as the risk for brands about to embark upon such an endeavour. Moreover, from our review of personalisation in both retail and digital spheres, this is the first such conceptual framework of its kind representing the user-end perspective of such innovations in technology. The game begins from the moment a user/player is within proximity of the store. The player is then faced with two options, either an enabling force (indicated by a ladder) moving them higher up the personalisation journey, or a detractor (indicated by a snake) preventing progress on the board. Each factor is described with key attributes as generated from the findings of the study. We envisage that AI-EP is not a one shoe fits all experience for users, and that it may take a circuitous route. As retailers continue to innovate, the blank squares represent the stages of the journey not relevant to AI-EP. The final goal is where the AI-EP has delivered a positive in-store experience and created value for customers and retailers. 6 Conclusion The deployment of AI technology for personalisation promises to address some of the business challenges faced by high-street retailers (Kumar et al., 2017), such as increased competition, heightened price sensitivity or the emergence of the show-rooming phenomenon. AI-EP apps, such as the one analysed in this paper, enable the creation of offers that draw on individual behaviours and contextual information, as opposed to aggregate segment information (as in the case of non-AI, automated personalisation) or intuition (as in the case of sales staff personalisation). As a result, AI-EP offers can be more relevant, granular and timely than either of those alternatives. However, factors related to the context of message delivery, the format of message delivery, and the salience of privacy concerns may impact the relevance of extant research on technology- enabled personalisation—mostly performed in the online environment—to help us understand consumers’ acceptance of AI-EP. Therefore, we responded to calls by Ameen et al (2022), Riegger et al (2021) and van de Sanden et al (2019), 1 3 Information Systems Frontiers Fig. 3 The Snakes and Ladders of AI-Enabled Personalisation among others, for empirical research on how consumers experience and respond to AI-EP. The qualitative investigation of consumers’ interac- tion with AI-DP in a shopping district with London, UK, through the lens of the personalisation-privacy paradox enabled us to identify the perceived content and process benefits derived from AI-EP, as well as how privacy con- cerns undermine these benefits and inform the custom- ers’ boundary management tactics. Together, these factors result in a carefully orchestrated process whereby cus- tomers either accept or reject the artificial intelligence- derived personalisation offer, but with a high degree of control over their interaction with the offer, and in par- ticular the use of their personal information. 6.1 Theoretical Contributions Our study makes the following three contributions. First, we showed that customers welcome this innovative way of interacting with them in the retail environment as posited by Davenport et al. (2020) and others, which should give con- fidence to practitioners considering adoption of AI (Bughin et al., 2017). However, we found that customers’ experiences with online personalisation create very high expectations of the extent of personalisation possible via AI-EP, address- ing the gaps identified in Table 1. Those high expectations may be difficult to meet, given not only the technological restrictions of AI-EP but also consumers’ discomfort with location tracking as well as the safeguarding of data which is essential for the efficacy of the offer, which can create customer backlashes and reputation damage (Castillo et al., 2020). Customers’ online experiences also shape their desire for additional services and functionalities, such as the crea- tion of wish lists or the ability to edit their preferences. This desire presents unique challenges from the point of view of interface design which have not been reported, yet. We rep- resented the range of factors impacting positively vs nega- tively on customers’ experiences with – and assessment of – AI-EP via the motif of Snakes and Ladders boardgame. Second, we provided empirical evidence of how the impact of the context of message delivery, the format of message delivery and the salience of privacy concerns dif- fers for AI-EP vs online personalisation. Specifically, regard- ing the impact of the different motivations for online vs. in-store retail on customers’ perception and evaluation of personalisation efforts (Haridasan & Fernando, 2018), we found that customers may be in a particular physical location for reasons other than shopping, and that this may result in 1 3 Information Systems Frontiers heightened irritation from app notifications. Moreover, cus- tomers seem more sensitive to evidence of tracking of past purchase behaviour in the physical environment than online, and more likely to resist the tracking of location and shelf- browsing behaviour than online browsing. In terms of the impact of message delivery interface, our findings confirm that the small screen of mobile phones impact negatively on consumers’ involvement with the message (Grewal et al., 2016), and that there is a need for attention-grabbing subject lines to make shoppers want to check the message, imme- diately. Future research could test the effectiveness of the same message delivered online vs via AI-EP, to quantify the effect of delivery interface on the effectiveness of personali- sation campaigns. Another factor that could limit the impact of AI-EP was the high number of notifications that mobile phone users typically receive on their devices, not just from direct messages from other users, but also from social media apps, calendar apps and others. Having said that, AI-EP could be more effective than e-mail offers, possibly because of the relative novelty of this form of personalisation, but also because of the volume of traffic that e-mail may attract (including spam content). Finally, regarding the impact of privacy concerns on consumers’ evaluation of AI-EP, our findings – like Ameen et al (2022)’s study of consumer inter- actions with smart technologies in shopping malls – seem to contradict Li et al. (2017). Unlike studies of personalisation in the online environment (e.g., Pappas, 2018), customers do not seem too concerned with the firm’s access to their personal information, in principle. This could be because the collection of such information is now seen as a condition for accessing services in the digital era. However, it could also be because of the particular type of app used in our case study. Like Ameen et al (2022)’s app, ours was valid for a shopping area, rather than a specific retailer. This fact may decrease the customers’ perception of surveillance, and increase their trust in the firm behind the AI-EP. Further research is needed to separate the effect of type of app (i.e., retailer vs location specific) from the overall privacy con- cerns with AI-EP. However, customers did express concerns over access to information which they did not deem essential for the task at hand, and access by unfamiliar retailers. Our findings thus assist in contextualising extant literature on AI-enabled personalisation online vs in-store. Third, we identified the specific content and process gratifications derived from AI-EP, and how they enhance or detract from the value of AI-EP for retail customers. Con- tent gratifications included discounts, time savings and rel- evance of offers, with the first one seemingly dominating the others. Receiving notifications on the phone was a process gratification for some but detracted from the overall benefit for others. Likewise, opinions were divided on the process gratification derived from how this app collected and used information for AI-EP. Our findings, thus, extend Sutanto et al (2013)’s work on the manifestation of the personali- sation-privacy paradox among smartphone users, in hybrid (physical-digital) environments. 6.2 Practical Contributions Collectively, these findings mean that the use of AI tech- nology for personalisation in the physical environment can address some of the business challenges faced by high-street retailers as suggested in Davenport et al. (2020), but with significant differences vis a vis personalisation in the online environment. Specifically, our findings have the following managerial implications. First, AI-EP is more suitable for customer retention efforts, than for customer acquisition. This is both because of the type of dataset required to deliver on customer expecta- tions of AI-EP and avoid the risk of customer backlash, and because of customers’ intense negative reaction to receiving personalised offers from brands that they usually do not buy from. A better way to acquire customers in this demographic group might be through the use of dynamic, entertaining adverts on social media; or by including their items in cloth- ing subscription services (YouGov, 2020). Second, to attract customers, retailers should offer entic- ing discounts on desired items. This is because, contrary to the online environment and to what is suggested in the litera- ture (e.g., Kietzmann et al., 2018), we found that customers weren’t driven by hedonic offers, and that there was limited scope for shopping basket expansion. Third, retailers should focus on providing information about items’ features, availability and other attributes that are important in the pre-purchase stage. This is because, while shoppers may interact with their smartphones across all stages of the purchase process (e.g., Syam & Sharma, 2018), they seemed most receptive to AI-EP offers in the lead-up to the purchase, rather than during the purchase (e.g., payment options) or afterwards (e.g., asking for feedback). Fourth, retailers need to test various aspects of offer deliv- ery, in order to minimise the concerns and irritants detected in our study. These include the number of notifications, to address shoppers' concerns with battery depletion and the fact that customers may be in the store’s neighbourhood for different reasons; and the wording of the message, to assuage customers’ desire to understand why they got a specific offer. It is also important for retailers to unpack which personal- ised offers are rejected because customers want to reaffirm their autonomy vs the AI (André et al., 2018), rather than because the offer itself was not persuasive. Fifth, retailers need to approach data collection and use, carefully. Our study revealed that the use of location and social media data, which is accepted in the online context, caused intense negative reactions among some customers. 1 3 Conversely, the relative novelty of in-store AI-EP means that customers may be willing to participate in ad-hoc data collection initiatives, if they perceive a link between the information requested and improvements in their shopping experience. 6.3 Research Limitations and Further Research It is important to recognise the limitations resulting from the focus and characteristics of our approach. The focus on fashion retail, on a multi-store app, and on the UK may limit the transferability of our findings to other research contexts. Research into other empirical settings is needed before claims can be made about consumer perceptions and experiences of AI-EP, generally. Likewise, young female consumers exhibit distinct attitudes to fashion shopping, sharing digital data and interacting with AI, meaning that our findings may not be directly applicable to older female shoppers, or to male shoppers of similar age. Findings from personalisation in the online environment indicate that perception of personalisation benefits is a key a factor in acceptance of personalisation (Pappas et  al., 2017). Therefore, it is important to identify which messages most clearly communicate the desired content gratification valued by different types of customers and/or different contexts. Moreover, by adopting a qualitative approach, we were able to identify a range of issues relevant for fashion retail customers. However, we are not able to quantify their absolute or relative importance. Further research employing quantitative approaches, namely natural experiments (e.g., Tag et al, 2021), is needed before claims can be made about the salience of specific gratifications and privacy concerns, or about the magnitude of their impact on consumer acceptance of AI-EP. Likewise, the use of methodologies such as fuzzy-set qualitative comparative analysis (see Pappas, 2018) would enable the identification of how the different factors identified in this study combine to amplify – or not – purchase intention when exposed to AI-EP. Furthermore, our focus on consumers overlooks the retailers’ perspective of AI-EP, which is a worthy area of further study. In particular, an avenue of further study that would advance our findings, as well as the work of Yoga- nathan et al. (2021), is to examine the relationship between AI-EP and access to onsite retail staff, homing in on the digital-physical customer experience dynamic. Given the practical nature of such an investigation, and the need for close collaboration with the organisation deploying the AI-EP solution, it would be beneficial to adopt the clinical inquiry approach method (see Schein, 2008). In this meth- odological approach, academic researchers and practitioners work together to shape the project, with the explicit goal of improving practice. Clinical inquiry is particularly useful for Information Systems Frontiers instigating digital innovation from within the organisation, as demonstrated in Vassilakopoulou et al (2022)’s analysis of the potential for creating hybrid human/AI service teams. Declarations Conflicts of Interest The authors have no relevant financial or non- financial interests to disclose. Open Access This article is licensed under a Creative Commons Attri- bution 4.0 International License, which permits use, sharing, adapta- tion, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/. References Aguirre, E., Mahr, D., Grewal, D., de Ruyter, K., & Wetzels, M. (2015). Unraveling the Personalization Paradox: The Effect of Information Collection and Trust-Building Strategies on Online Advertisement Effectiveness. Journal of Retailing., 91(1), 34–49. Aguirre, E., Roggeveen, A. L., Grewal, D., & Wetzels, M. (2016). The personalization-privacy paradox: Implications for new media. Journal of Consumer Marketing, 33(2), 98–110. Ameen, N., Hosany, S., & Paul, J. (2022). The personalisation-privacy paradox: Consumer interaction with smart technologies and shop- ping mall loyalty. Computers in Human Behavior, 126(January), 106976. https:// doi. org/ 10. 1016/j. chb. 2021. 106976. André, Q., Carmon, Z., Wertenbroch, K., Crum, A., Frank, D., Gold- stein, W., … & Yang, H. (2018). Consumer choice and autonomy in the age of artificial intelligence and big data. Customer Needs and Solutions, 5(1), 28–37. Awad, N. F., & Krishnan, M. S. (2006). The Personalization Privacy Paradox: An Empirical Evaluation of Information Transparency and the Willingness to Be Profiled Online for Personalization. MIS Quarterly, 30(1), 13–28. https:// doi. org/ 10. 2307/ 25148 715 Baek, T. H., & Morimoto, M. (2012). Stay Away from Me: Examin- ing the Determinants of Consumer Avoidance of Personalized Advertising. Journal of Advertising, 41(1), 59–76. Balan, U. M., & Mathew, S. K. (2020). Personalize, Summarize or Let them Read? A Study on Online Word of Mouth Strategies and Consumer Decision Process. Information Systems Frontiers. https:// doi. org/ 10. 1007/ s10796- 020- 09980-9. Belk, R. W. (2017). Qualitative Research in Advertising. Journal of Advertising, 46(1), 36–47. Boerman, S. C., Willemsen, L. M., & Van Der Aa, E. P. (2017). “This post is sponsored”: Effects of sponsorship disclosure on persua- sion knowledge and electronic word of mouth in the context of Facebook. Journal of Interactive Marketing, 38, 82–92. Boratto, L., Carta, S., Kaltenbrunner, A., & Manca, M. (2018). Guest Editorial: Behavioral-Data Mining in Information Systems and the Big Data Era. Information Systems Frontiers, 20, 1153–1156. Brehm, S. S., & Brehm, J. W. (2013). Psychological reactance: A theory of freedom and control. Academic Press. 1 3 Information Systems Frontiers Brusilovsky, P., & Tasso, C. (2004). Preface to special issue on user modeling for web information retrieval. User Modeling and User- Adapted Interaction, 14(2), 147–157. Bryman, A., & Bell, E. (2015). Business Research Methods, 4th edi- tion. OUP Oxford. Bues, M., Steiner, M., Stafflage, M., & Krafft, M. (2017). How mobile in-store advertising influences purchase intention: Value drivers and mediating effects from a consumer perspective. Psychology & Marketing, 34(2), 157–174. Bughin, J., Hazan, E., Ramaswamy, S., Chui, M., Allas, T., Dahlstrom, P., Henke, N., & Trench, M. (2017). Artificial Intelligence - The next digital frontier? McKinsey Global Institute. https:// www. mckin sey. com/ ~/ media/ mckin sey/ indus tries/ advan ced% 20ele ctron ics/ our% 20ins ights/ how% 20art ifici al% 20int ellig ence% 20can% 20del iver% 20real% 20val ue% 20to% 20com panies/ mgiar tific ial- intel ligen ce- discu ssion- paper. pdf. Accessed 13 Jan 2023 Burrell, J. (2016). How the machine ‘thinks’: Understanding opacity in machine learning algorithms. Big Data & Society, 3(1), 1–12. Castelo, N., Schmitt, B., & Sarvary, M. (2019). Human or Robot? Con- sumer Responses to Radical Cognitive Enhancement Products. Journal of the Association for Consumer Research, 4(3), 217–230. https:// doi. org/ 10. 1086/ 703462 Castillo, D., Canhoto, A. I., & Said, E. (2020). The Dark Side of AI-pow- ered Service Interactions: Exploring the process of co-destruction from the customer perspective. The Service Industries Journal, 41(13–14), 900–925. https:// doi. org/ 10. 1080/ 02642 069. 2020. 17879 93 Chen, J., & Dibb, S. (2010). Consumer trust in the online retail con- text: Exploring the antecedents and consequences. Psychology & Marketing, 27(4), 323–346. Chou, Y. C., & Shao, B. B. M. (2021). Adoption and performance of mobile sales channel for e-Retailers: Fit with M-Retail character- istics and dependency on e-Retailing. Information Systems Fron- tiers, 23, 681–694. https:// doi. org/ 10. 1007/ s10796- 020- 09989-0 Davenport, T., Guha, A., Grewal, D., & Bressgott, T. (2020). How artificial intelligence will change the future of marketing. Journal of the Academy of Marketing Science, 48(1), 24–42. Demoulin, N., & Willems, K. (2019). Servicescape irritants and cus- tomer satisfaction: The moderating role of shopping motives and involvement. Journal of Business Research, 104, 295–306. https:// doi. org/ 10. 1016/j. jbusr es. 2019. 07. 004 Dempsey, P. (2015). The teardown one plus two smartphone. Engineer- ing & Technology, 10(9), 88–89. Dwivedi, Y. K., Hughes, L., Ismagilova, E., Aarts, G., Coombs, C., Crick, T., ... & Williams, M. D. (2021). Artificial Intelligence (AI): Multidisciplinary perspectives on emerging challenges, opportunities, and agenda for research, practice and policy. Inter- national Journal of Information Management, 57, 101994. Eisenhardt, K. M. (1989). Building theories from case study research. Academy of Management Review, 14(4), 532–550. Gardino, G. B., Meo, R., & Craparotta, G. (2021). (2021) Multi-view Latent Learning Applied to Fashion Industry. Information Systems Frontiers, 23, 53–69. https:// doi. org/ 10. 1007/ s10796- 020- 10005-8 Grewal, D., Yakov, B., Spann, M., & Zubcsek, P. P. (2016). Mobile Advertising: A Framework and Research Agenda. Journal of Interactive Marketing, 34, 3–14. https:// doi. org/ 10. 1016/j. intmar. 2016. 03. 003 Griva, A., Bardaki, C., Pramatari, K., & Doukidis, G. (2021). Factors affecting customer analytics: evidence from three retail cases. Information Systems Frontiers, 1–24. https:// doi. org/ 10. 1007/ s10796- 020- 10098-1. Haghirian, P., Madlberger, M., & Tanuskova, A. (2005). Increasing advertising value of mobile marketing-an empirical study of ante- cedents. In Proceedings of the 38th annual Hawaii international conference on system sciences, pp 32c–32c. https:// doi. org/ 10. 1109/ HICSS. 2005. 311 Haridasan, A. C., & Fernando, A. G. (2018). Online or in-store: Unrav- elling consumer’s channel choice motives. Journal of Research in Interactive Marketing, 12(2), 215–230. https:// doi. org/ 10. 1108/ JRIM- 07- 2017- 0060 Hubert, M., Blut, M., Brock, C., Backhaus, C., & Eberhardt, T. (2017). Acceptance of smartphone-based mobile shopping: Mobile ben- efits, customer characteristics, perceived risks, and the impact of application context. Psychology & Marketing, 34(2), 175–194. Kar, A. K. (2020). What affects usage satisfaction in mobile payments? modelling user generated content to develop the “digital service usage satisfaction model”. Information Systems Frontiers, 23, 1341–1361. https:// doi. org/ 10. 1007/ s10796- 020- 10045-0 Karumur, R. P., Nguyen, T. T., & Konstan, J. A. (2018). Personality, User Preferences and Behavior in Recommender systems. Infor- mation Systems Frontiers, 20, 1241–1265. Kietzmann, J., Paschen, J., & Treen, E. (2018). Artificial intelligence in advertising: How marketers can leverage artificial intelligence along the consumer journey. Journal of Advertising Research, 58(3), 263–267. Kim, T. W. & Duhachek, A. (2020). Artificial intelligence and persua- sion: A construal-level account. Psychological Science, 31(4), 363-380. https://doi.org/10.1177/0956797620904985 Krippendorff, K. (2004). Reliability in content analysis: Some common misconceptions and recommendations. Human Communication Research, 30(3), 411–433. Krishnaraju, V., Mathew, S. K., & Sugumaran, V. (2016). Web person- alization for user acceptance of technology: An empirical investi- gation of E-government services. Information Systems Frontiers, 18(3), 579–595. Kumar, V., Anand, A., & Song, H. (2017). Future of Retailer Profit- ability: An Organizing Framework. Journal of Retailing, 93(1), 96–119. Lemmon, C. (2017). Regent Street boosts engagement with app. Retail Systems, 14/3/2017 https:// www. retail- syste ms. com/ rs/ Regent_ Street_ autoG raph_ Mobile_ App. php last accessed 14 January 2022. Li, H., Luo, X. R., Zhang, J., & Xu, H. (2017). Resolving the privacy paradox: Toward a cognitive appraisal and emotion approach to online privacy behaviors. Information & Management, 54(8), 1012–1022. Liu, Z., Wang, X., & Liu, J. (2019). How digital natives make their self- disclosure decisions: A cross-cultural comparison. Information, Technology & People, 32(3), 538–558. Miltgen, C. L., & Smith, H. J. (2019). Falsifying and withholding: Exposing individuals’’ contextual privacy-related decision-mak- ing. Information & Management, 56, 696–717. Mondada, L., Bänninger, J., Bouaouina, S. A., Gauthier, G., Hänggi, P., Koda, M., Svensson, H., & Tekin, B. S. (2020). Doing paying dur- ing the Covid-19 pandemic. Discourse Studies, 22(6), 720–752. Moore, R. S., Moore, M. L., Shanahan, K. J., & Mack, B. (2015). Creepy Marketing: Three Dimensions of Perceived Excessive Online Privacy Violation. Marketing Management, 25(1), 42–53. Pappas, I. O. (2018). User experience in personalized online shopping: A fuzzy-set analysis. European Journal of Marketing, 52(7/8), 1679–1703. https:// doi. org/ 10. 1108/ EJM- 10- 2017- 0707 Pappas, I. O., Kourouthanassis, P. E., Giannakos, M. N., & Chris- sikopoulos, V. (2017). Sense and sensibility in personalized e-commerce: How emotions rebalance the purchase intentions of persuaded customers. Psychology & Marketing, 34(10), 972–986. Pappas, I. O., Mikalef, P., Giannakos, M. N., Krogstie, J., & Lekakos, G. (2018). Big data and business analytics ecosystems: Paving the way towards digital transformation and sustainable societies. Information Systems and e-Business Management, 16, 479–491. Pentina, I., Zhang, L., Bata, H., & Chen, Y. (2016). Exploring pri- vacy paradox in information-sensitive mobile app adoption: A cross-cultural comparison. Computers in Human Behavior, 65, 409–419. 1 3 Riegger, A. S., Klein, J. F., Merfeld, K., & Henkel, S. (2021). Technol- ogy-enabled personalization in retail stores: Understanding drivers and barriers. Journal of Business Research, 123, 140–155. Rippé, C. B., Weisfeld-Spolter, S., Yurova, Y., Dubinsky, A. J., & Hale, D. (2017). Under the sway of a mobile device during an in-store shopping experience. Psychology & Marketing, 34(7), 733–752. Sarker, S., Xiao, X., Beaulieu, T., & Lee, A. S. (2018). Learning from First-Generation Qualitative Approaches in the IS Discipline: An Evolutionary View and Some Implications for Authors and Evaluators (PART 1/2). Journal of the Association for Information Systems, 19(8), 752–774. Schein, E. H. (2008). Clinical inquiry/research. In P. Reason & H. Bradbury (Eds.), Handbook of action research (pp. 266–279). SAGE Publications. Schmidt, L., Bornschein, R., & Maier, E. (2020). The effect of privacy choice in cookie notices on consumers’ perceived fairness of fre- quent price changes. Psychology & Marketing, 37(9), 1263–1276. https:// doi. org/ 10. 1002/ mar. 21356 Scott, M. (2014). At store after store, a pitch by phone. The New York Times. https:// www. nytim es. com/ 2014/ 12/ 02/ fashi on/ regent- street- london- uses- app- and- beaco ns- to- reach shopp ers. html. Accessed 13 Jan 2023 Shankar, V., Kleijnen, M., Ramanathan, S., Rizley, R., Holland, S., & Morrissey, S. (2016). Mobile Shopper Marketing: Key Issues, Current Insights, and Future Research Avenues. Journal of Inter- active Marketing, 34(C), 37–48. Stanton, J. M., & Stam, K. (2003). Information Technology, Privacy, and Power Within Organizations: A View from Boundary Theory and Social Exchange Perspectives. Surveillance and Society, 1(2), 152–190. Sutanto, J., Palme, E., Tan, C. H., & Phang, C. W. (2013). Addressing the Personalization-Privacy Paradox: An Empirical Assessment from a Field Experiment on Smartphone Users. MIS Quarterly, 37(4), 1141–1164. Syam, N., & Sharma, A. (2018). Waiting for a sales renaissance in the fourth industrial revolution: Machine learning and artificial intelligence in sales research and practice. Industrial Marketing Management, 69, 135–146. Tag, B., Goncalves, J., Webber, S., Koval, P., & Kostakos, V. (2021). A retrospective and a look forward: Lessons learned from research- ing emotions in-the-wild. IEEE Pervasive Computing, 21(1), 28–36. Tam, K. Y., & Ho, S. Y. (2006). Understanding the impact of web personalization on user information processing and decision out- comes. MIS Quarterly, 30(4), 865–890. https:// doi. org/ 10. 2307/ 25148 757 Travers, M. (2001). Qualitative Research Through Case Studies. SAGE Publications Ltd ISBN: 9780761968061. Tucker, C. E. (2014). Social Networks, Personalized Advertising, and Privacy Controls. Journal of Marketing Research, 51(1), 546–562. van de Sanden, S., Willems, K., & Brengman, M. (2019). In-store location-based marketing with beacons: From inflated expecta- tions to smart use in retailing. Journal of Marketing Management, 35(15–16), 1514–1541. https:// doi. org/ 10. 1080/ 02672 57X. 2019. 16891 54 Vassilakopoulou, P., Haug, A., Salvesen, L. M., & Pappas, I. O. (2022). Developing human/AI interactions for chat-based customer ser- vices: Lessons learned from the Norwegian government. Euro- pean Journal of Information Systems. https:// doi. org/ 10. 1080/ 09600 85X. 2022. 20964 90 Information Systems Frontiers Verhoef, P. C., Stephen, A. T., Kannan, P. K., Luo, X., Abishek, V., Andrews, M., … & Zhang, Y. (2017). Consumer connectivity in a complex, technology-enabled, and mobile-oriented world with smart products. Journal of Interactive Marketing. 40, 1–8. Wang, Y., Yuan, Y., Turel, O., & Tu, Z. (2015). Understanding the Development and Diffusion of Mobile Commerce Technologies in China: A Biographical Study with an Actor-Network Theory Perspective. International Journal of Electronic Commerce, 19(4), 47–76. https:// doi. org/ 10. 1080/ 10864 415. 2015. 10293 58 Xu, D. J., Liao, S. S., & Li, Q. (2008). Combining empirical experi- mentation and modeling techniques: A design research approach for personalized mobile advertising applications. Decision Sup- port Systems, 44(3), 710–724. https:// doi. org/ 10. 1016/j. dss. 2007. 10. 002 Xu, H., Luo, X. R., Carroll, J. M., & Rosson, M. B. (2011). The per- sonalization privacy paradox: An exploratory study of decision making process for location-aware marketing. Decision Support Systems, 51(1), 42–52. Yin, R. K. (2012). Case study methods. Yoganathan, V., Osburg, V.-S., H. Kunz, W., & Toporowski, W. (2021). Check-in at the Robo-desk: Effects of automated social presence on social cognition and service implications. Tourism Manage- ment, 85, 104309. https:// doi. org/ 10. 1016/j. tourm an. 2021. 104309. YouGov. (2020). The Fashion Industry in Great Britain. YouGov. https:// yougov. co. uk/ topics/ consu mer/ artic les- repor ts/ 2020/ 02/ 25/ fashi on- indus try- great- brita in [Last accessed 22 July 2022]. Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Ana Isabel Canhoto is Professor of Digital Business at the University of Sussex, UK. Her research focuses on the role of digital technology (including Artificial Intelligence Big Data and the  Metaverse) in interactions between firms and their customers. She examines drivers of adoption, user experiences, consequences of adoption, and the role of context. She is also committed to the pedagogical use of digital technology, including using machine learning to support pupil performance, creating quasi-simulations for experiential learning, and training early career researchers to use social media to develop and disseminate their work. Brendan James Keegan is an Assistant Professor in Marketing at Maynooth University, Ireland. His current research activities are within the areas of business to business relationships, artificial intelligence and machine learning applications in digital marketing, and digital placemaking. His work is published in Information Systems Frontiers, Industrial Marketing Management, European Journal of Marketing, European Management Review. He is the Principal Investigator for the ongoing digital placemaking research within the H2020 funded GoGreenRoutes Project. Maria Ryzhikh is the E-Commerce Manager for the EMEA region at Weber-Stephen Products EMEA GmbH. She completed the MSc Marketing program at Oxford Brookes University in 2016  and then continued to pursue her career in digital marketing in various technologically-driven companies in the UK and Germany. Her particular areas of interests lie in performance marketing,  digital marketing analytics, online consumer behaviour and psychology, and conversion rate optimization. 1 3
10.3390_biomedicines11020375
Article Effect of pH, Ionic Strength and Agitation Rate on Dissolution Behaviour of 3D-Printed Tablets, Tablets Prepared from Ground Hot-Melt Extruded Filaments and Physical Mixtures Nour Nashed 1 , Stephanie Chan 1, Matthew Lam 2, Taravat Ghafourian 3 and Ali Nokhodchi 1,4,* 1 Pharmaceutics Research Laboratory, Arundel Building, School of Life Sciences, University of Sussex, Brighton BN1 9QJ, UK 2 Department of Chemical and Pharmaceutical Sciences, School of Human Sciences, London Metropolitan University, 166-220 Holloway Road, London N7 8DB, UK 3 Department of Pharmaceutical Sciences, College of Pharmacy, Nova Southeastern University, 3200 South University Drive, Fort Lauderdale, FL 33328, USA Lupin Inhalation Research Center, 4006 NW 124th Ave., Coral Springs, FL 33065, USA 4 * Correspondence: alinokhodchi@lupin.com Abstract: With the current focus on 3D-printing technologies, it is essential to understand the pro- cesses involved in such printing methods and approaches to minimize the variability in dissolution behaviour to achieve better quality control outcomes. For this purpose, two formulations of theo- phylline tablets were prepared using hydroxypropyl cellulose (HPC) and ethyl cellulose (EC). Among the two types of tablets, three different methods (physical mixture (PM), hot-melt extrusion (HME) and 3D-printing fused deposition modelling (FDM)) were applied and their dissolution behaviours were studied under various conditions using a biodissolution tester. This was carried out at pH values of 1.2, 2.2, 5.8, 6.8, 7.2 and 7.5, mimicking the medium in the gastrointestinal tract. Dissolution tests under two dipping rates (10 dpm and 20 dpm) and two ionic strengths (0.2 M and 0.4 M) were conducted to mimic fed and fasting conditions. The dissolution efficiency (DE%), release rate, similar- ity factor (f 2) and difference factor (f 1) were calculated. When comparing the DE%, the formulation containing EC showed less sensitivity to changes in the dipping rate and ionic strength compared to the HPC formulation. As for the manufacturing method, 3D-printing FDM could improve the robustness of the dissolution behaviour of both formulations to dipping rate changes. However, for ionic strength changes, the effect of the manufacturing method was dependent on the formulation composition. For example, the 3D-printed tablets of the HPC formulation were more sensitive to changes in ionic strength compared to the EC-containing formulation. The release mechanism also changed after the thermal process, where n values in the Korsmeyer–Peppas model were much higher in the printing and HME methods compared to the PM. Based on the formulation composition, the 3D-printing method could be a good candidate method for tablets with a robust dissolution behaviour in the GI tract. Compared to HPC polymers, using hydrophobic EC polymers in printable formulations can result in a more robust dissolution behaviour in fed and fasting states. Keywords: 3D printing; hot-melt extrusion; physical mixture; biodissolution; ionic strength; agitation rate Citation: Nashed, N.; Chan, S.; Lam, M.; Ghafourian, T.; Nokhodchi, A. Effect of pH, Ionic Strength and Agitation Rate on Dissolution Behaviour of 3D-Printed Tablets, Tablets Prepared from Ground Hot-Melt Extruded Filaments and Physical Mixtures. Biomedicines 2023, 11, 375. https://doi.org/10.3390/ biomedicines11020375 Academic Editor: M. R. Mozafari Received: 30 December 2022 Revised: 23 January 2023 Accepted: 25 January 2023 Published: 27 January 2023 Copyright: © 2023 by the authors. 1. Introduction Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). The dissolution behaviour of tablets in the gastrointestinal (GI) tract is a critical step in their absorption and biological response. Many factors can influence the dissolution of tablets, including those related to GI tract conditions and those controlled by drug properties, tablet formulation and manufacture characteristics. For example, the wide range of pH values in the GI tract (between 1.2 and 7.5) can affect drug solubility and polymer behaviour, hence effectively influencing drug release [1]. In addition, based on the Biomedicines 2023, 11, 375. https://doi.org/10.3390/biomedicines11020375 https://www.mdpi.com/journal/biomedicines biomedicines Biomedicines 2023, 11, 375 2 of 17 health condition of the GI tract and the food intake, hydrodynamic forces and ionic strength may change, which can affect drug release [1,2]. Furthermore, an increase in the motility of the stomach and intestines can promote the disintegration of tablets and boost the drug release. The increased secretion of ions when consuming food can cause salting in/out of polymers and change the drug release behaviour of tablets [3,4]. Such GI factors can be simulated in vitro by adjusting the agitation speeds and dissolution media components during dissolution testing, in order to achieve a better prediction of the tablet dissolution behaviour in vivo. The dissolution test is an in vitro method used as a quality control tool which, based on the applied conditions, could be sensitive to and detect changes in the formulation composition and critical product attributes [5,6]. It is also needed for bioequivalence studies and in vitro–in vivo correlation (IVIVC) studies when developing new formulations. An effective dissolution test should closely simulate the drug release in the GI tract. Therefore, for more predictive in vitro results, it is preferable to use a series of conditions that mimic the environment of the GI tract in various physiological states [7,8]. A comparison of the release profiles obtained under various dissolution conditions can be used as a robustness indicator to show if the tablet will behave similarly under various physiological conditions. In the case of highly soluble drugs such as BCS I and III, simplified dissolution conditions could be used; however, a more sophisticated test condition is needed for poorly water- soluble drugs [7]. Although it has been suggested that a biodissolution apparatus (USP 3) is preferable for studying the dissolution behaviour for extended-release oral formulations under fed and fasted conditions by manipulating the pH and ionic strength of the media and changing the agitation rate [9], it is still not as widely used as a paddle or basket dissolution apparatus in dissolution studies. The hydrodynamic stress in USP 3 is superior to that in USP 1 and 2, as the 5 dips per minute (dpm) speed in USP 3 matches a 50 rpm paddle speed in USP 2 and a 100 rpm basket speed in USP 1 [10]. In IVIVC studies, USP 3 facilitates the application of in vivo-like conditions, including a variety of physicochemical and hydrodynamic conditions such as pH gradients, ionic strengths and agitation speeds, giving more comprehensive and predictive in vitro results [11,12]. Many methodologies have been proposed for USP 3 with regards to studying dissolution profiles in in vivo-like conditions [13–15]. The manufacturing method is one of the critical product attributes that can affect a tablet’s characteristics, including its dissolution performance, hence the drug bioavailabil- ity [5]. For example, 3D printing using the FDM method involves the melting/softening of materials at high temperatures and extrusion, the products of which are hardened on cooling. This is very different from various conventional compaction tabletting methods. With the current increasing interest in 3D-printing technology for tablet manufacture, it is crucial to investigate the mechanisms involved in the process that impact the dissolution profile of such tablets. Dissolution profile changes after 3D printing may be associated with changes in the porosity, density and hardness, depending on the formulation and printing conditions [16,17]. Thus, a discriminatory dissolution testing method is needed to provide a better understanding of how the printing method can change the release behaviour of a given formulation, which will consequently influence the bioavailability. This work investigates how the printing method can change the release profile under harsh dissolution conditions such as high agitation rates and high ionic strengths. In the authors’ previous work [16], dissolution studies were carried out in USP 2, while in this paper, more harsh dissolution conditions were applied using a USP 3 apparatus to predict the in vivo behaviour of printed tablets. To the best of the authors’ knowledge, this is the first time a USP 3 apparatus has been utilized for a comparison study of the dissolution profile of printed tablets with tablets prepared by conventional manufacturing methods, either from physical mixtures of the formulation powders (PM) or from filaments obtained through the hot-melt extrusion of the same formulation (HME). For all these manufacturing methods, two formulations were prepared using HPC or EC as the polymer and theophylline as Biomedicines 2023, 11, 375 3 of 17 the active pharmaceutical ingredient. The tablets were manufactured by three different methods: PM, HME and 3D printing via FDM. The data from the drug dissolution tests of the mentioned production methods were compared amongst each other to evaluate the impact of variable dissolution conditions on the drug release behaviour of the tablets. The test was carried out in the biological pH range from 1.2 to 7.5, with different agitation speeds (dipping rate; 10 and 20 dpm) and ionic strengths (0.2 M and 0.4 M) to simulate fed and fasted states in the GI tract. 2. Materials and Methods 2.1. Materials Theophylline anhydrous with a purity of >99% was purchased from Fisher Scientific (Loughborough, UK). Two grades of hydroxypropyl cellulose (HPC), i.e., klucelTM EF and klucelTM JF, were provided by Ashland Inc. (Rotterdam, Netherlands). Ethyl cellulose (EC, Ethocel 10 FP) was obtained from Colorcon Ltd. (Dartford, UK). Dibutyl sebacate (DBS) was supplied by Sigma-Aldrich (St. Louis, MO, USA). Dissolution buffer solutions were prepared using the following materials: hydrochloric acid (HCl), potassium chloride (KCl), sodium hydroxide (NaOH) and potassium monobasic phosphate (KH2PO4), which were supplied by Fisher Scientific (Loughborough, UK), and deionised water. All materials were used as received. 2.2. Tablet Formulations Two formulations were prepared (Table 1) and were selected for biodissolution studies. The theophylline percentage was fixed at 30% in both formulations, and the remainder was polymers (e.g., EC and HPC), as shown in Table 1. Two viscosity grades of HPC were used, EF (low viscosity grade) and JF (high viscosity grade). The selection of the HPC grade was based on the printability of filaments. EC was plasticised with 5% DBS, which was left overnight for better sorption of the plasticizer into the polymer matrix. HPC, EC and theophylline were selected because they are thermally stable materials under the used conditions [16]. The formulations were converted to tablets by three different manufacturing methods, which are described in Section 2.3 below. Table 1. Composition of studied formulations. Formulation ** Theophylline (% w/w) HPC EF (% w/w) HPC JF (% w/w) EC (% w/w) F1 F2 — 70% ** Tablet weight was 333.33 mg, containing 100 mg of theophylline. * Plasticised with 5% w/w of DBS overnight. 35% * — 35% — 30% 30% 2.3. Preparation Methods of Tablets The two formulations were prepared via three different methods: compression of a physical mixture of powders (PM), compression of hot-melt extruded powders (HME), and 3D printing of the powders (FDM). The average weight of the tablets prepared by any of the three methods was 333.33 mg, where theophylline accounted for 100 mg. The tablets were white in colour, round and cylindrical. They had the same surface area-to-volume ratio (SA/V) of 0.8 mm−1 to minimize its effect on the dissolution behaviour. 2.3.1. Preparing Physical Mixture (PM) Tablets Based on the composition mentioned in Table 1, the powders were thoroughly mixed using a mortar and pestle (for approximately 5 min). The physical mixture powder blend was compressed into a tablet (333.33 mg) with a manual tableting press (Model MTCM-I, Globe Pharma, USA) equipped with 10 mm-diameter concave punches. The employed pressure was 150 bar and the dwelling time was adjusted to 10 s. The manufactured tablets were stored in enclosed vials in a cabinet at room temperature (22 ± 2 ◦C) and tested for dissolution within a week. Biomedicines 2023, 11, 375 4 of 17 2.3.2. Preparing HME Tablets The powders were mixed manually and fed to a 10 mm twin-screw extruder L/D 20 (assembled by Point 1 Control Systems Ltd., Stoke-on-Trent, UK) at a screw speed of 50 rpm. The temperature of the extruder was set to over the Tg of the used polymers (around 120–130 ◦C) so that they were softened and mixed properly with the theophylline to obtain smooth and cylindrical filaments. The temperature of the feeding zone was 110 ◦C and that of the other sections (including the die) was 150 ◦C. The filaments were collected using a winder manually and were ground after being cooled down using a ball mill (PM 100, Retsch GmbH Germany) at 400 rpm for 4 min. The resulting powder was finally compressed into tablets using a manual tableting press, as with the PM tablets (see Section 2.4 for details) at a pressure of 150 bar and 10 s of dwelling time. The tablets produced were stored in enclosed vials in a cabinet at room temperature and tested for dissolution within a week. 2.3.3. Preparing 3D-Printed Tablets Filaments produced by HME (see Section 2.3.2 for details) were used as feedstock for 3D printing, with a diameter between 1.6 and 1.8 mm to fit the printing nozzle (1.75 mm). The tablets were printed with a height of 5.45 mm and a diameter of 8.8 mm. These specific dimensions of the 3D-printed tablets matched the weight (333.33 mg) as well as the SA/V of the tablets prepared by HME and PM, which increased the accuracy when comparing their drug-release profiles. The printing settings were as follows: a layer height of 0.2 mm, an infill density of 100%, a number of shells of 2, a printing speed of 90 mm/s, a bed temperature of 50 ◦C and a nozzle temperature of 220 ◦C. The produced tablets were collected and stored in enclosed vials in a cabinet at room temperature and tested for dissolution within a week. 2.4. Biodissolution Studies An automated USP 3, Bio-Dis (Varian, US), was used to carry out the dissolution tests. The dissolution test was performed in a varied range of pH values, covering the pH values in the GI tract, using buffered solutions, as shown in Table 2. The dissolution program was employed as reported in other studies [18,19]. In brief, tablets were kept for 60 min at a pH of 1.2, followed by 60 min at a pH of 2.2, then 10 min at a pH of 5.8, 120 min at a pH of 6.8, 30 min at a pH of 7.2 and 30 min at a pH of 7.5. The total duration was 310 min. The absorbance of the released theophylline was measured at a wavelength of 271 nm using a UV/Visible spectrophotometer (Varian, Cary 50). Two dissolution variables were tested, the dipping rate and the ionic strength. The dissolution data were analysed to explore which manufactured method could produce tablets that were robust enough in terms of drug release when subjected to fasted and fed states in the GI tract. For this purpose, the dissolution efficiency (DE%), difference factor (f 1) and similarity factor (f 2) were calculated for a quantitative comparison. The DE% is the area under a dissolution curve between defined time points. f1 = ∑n (cid:12) (cid:12) (cid:12) (cid:12)Rj − Tj j=1 Rj j=1 ∑n × 100 f2 = 50 × log   (cid:34)  1 + (1/n) (cid:12) (cid:12)Rj − Tj (cid:12) (cid:12) 2 n ∑ j=1 (cid:35)−0.5 × 100    Biomedicines 2023, 11, 375 5 of 17 Table 2. Composition, pH values and transit time of each buffer used in biodissolution test. Mimicked GI Tract Segment Stomach Stomach Duodenum Jejunum Proximal ileum Distal ileum Buffer Composition * 0.05 M KCl and 0.073 M HCl in DW 0.05 M KCl and 0.007 M HCl in DW 0.004 M NaOH and 0.05 M KH2PO4 in DW 0.022 M NaOH and 0.05 M KH2PO4 in DW 0.035 M NaOH and 0.05 M KH2PO4 in DW 0.04 M NaOH and 0.05 M KH2PO4 in DW Ionic Strength ** pH Value Transit Time (min) 0.12 M 0.06 M 0.28 M 0.3 M 0.3 M 0.32 M 1.2 2.2 5.8 6.8 7.5 7.8 60 60 10 120 30 30 min * DW = distilled water. ** These ionic strengths are the initial ionic strength of the dissolution medium without adding more ionic strength to investigate the effect of ionic strength on dissolution behaviour. In the above equations, Rj and Tj are the mean percentages of the drug dissolved at each time point for the reference and test products, respectively. n denotes the number of samples taken during the dissolution run. When f 2 was above 50% and f 1 was less than 15%, there was no difference between the two dissolution profiles compared. 2.4.1. Dipping Rate The dissolution test was run according to the proposed program at a low speed (10 dpm). The tablets were also tested for dissolution at a higher dipping rate of 20 dpm. The buffer solutions with pH values ranging between 1.2 and 7.5 were used as the dissolution medium for both of the mentioned dipping rates. During the study on the dipping rate, the ionic strength was fixed to the characteristic ionic strength values for each pH value as per Table 2 (0 M added of KCl). Changes in the dipping rate mimicked the changes in GI motility under various physiological states and during fed and fasted states. 2.4.2. Ionic Strength To the various buffers listed in Table 2, KCl was added in two concentrations of 0.2 M and 0.4 M to study how ionic strength changes can affect drug release. This simulated the ionic strength changes as the result of food presence in the GI tract [3,14,18,19]. Note that each buffer in Table 2 has its own pH-specific ionic strength, but the focus here will be on the effect of the further added ionic strength (by adding KCl) to each buffer solution. The dissolution results in buffers without any additional ionic strength were considered as the reference to compare with the results after adding 0.2 M and 0.4 M ionic strength. During the study of ionic strength, the dipping rate was fixed at 10 dpm. The ionic strength was calculated using the equation below: I = 1 2 j ∑ i CiZ2 i In the above equation, I is the ionic strength, j is the number of species of ions in the solution, Ci is the molar concentration of ion i to j and Zi is the charge number of ion i. 2.5. Release Kinetics The release kinetics for all the tablets prepared by the three methods were analysed using DDSolver software (an add-in program in Microsoft Excel) [20]. Data were fitted to the Korsmeyer–Peppas model, while special attention was paid to the n values of this equation in order to explore any change in the release mechanisms when the tablets were subjected to different dipping rates and ionic strengths. Below is the Korsmeyer–Peppas Biomedicines 2023, 11, 375 6 of 17 model, where Mt/M∞ is the fraction of the drug released at time t, K is the drug release rate constant and n is the release exponent: Mt/M∞ = Ktn The values of the diffusional exponent, n, differ based on the geometry of the tablet. For 60% of the release profile of a cylindrical-shaped tablet, the drug release is governed by diffusion (Fickian model) when n approaches 0.45. However, when n approaches 0.85, swelling (polymer chain relaxation) is the main release mechanism, where the drug release follows zero-order kinetics (non-Fickian model, case II). For 0.45 < n < 0.85, the drug release is governed by both diffusion and swelling (anomalous transport). Values of the diffusional exponent, n, exceeding 0.85 mean that the drug release is controlled by polymer relaxation (non-Fickian model, super case II) [21]. 3. Results and Discussion 3.1. Effects of Dipping Rate and Formulation Composition Dissolution tests were carried out in three replicates for all PM, HME and 3D-printed tablets of both formulations under different dipping rates (10 dpm and 20 dpm) and added ionic strengths (0.2 M and 0.4 M). The dissolution profiles are depicted in Figure 1. The dissolution results at 10 dpm in different pH media were considered as the reference, which was used to investigate the effects of a higher agitation rate (20 dpm) and added ionic strength (0.2 and 0.4 M) on the dissolution rate of the tablets manufactured via three different methods. The DE% values were used for the overall comparison of the dissolution behaviour, and the release rate calculated for each pH value was used for an in-depth understanding of the release behaviour and also to determine how the formulation composition and manufacturing methods affected the dissolution behaviour of tablets. The release rate is usually calculated in two different ways, either from the release% vs. time plot or from the release% vs. square root of the time plot [6,22]. As the release% versus time showed higher R2 values compared to when the release% was plotted versus the square root of time, the former was used to calculate the release rate at each pH value. Figure 1 shows that tablets prepared by PM and HME had higher SD values compared to the printed tablets, implying that the printing method could improve the reproducibility of the dissolution behaviour, as shown by the reduced variations from batch to batch. The reason behind this is not clear, but the increased hardness of the printed tablets compared to the HME and PM tablets, as proven in our previous work, could have had an impact [16]. Looking at the dissolution graphs in Figure 1, it is clear that the F1 formulation had a much slower release pattern compared to F2, regardless of the manufacturing method and dissolution conditions. Under the 10 dpm dissolution testing condition, the DE% values of F1 were 16.57%, 33.76% and 13.12% for the PM, HME and printed tablets, respectively, while F2 showed a faster release with DE% values of 45.32%, 53.91% and 57.12% for the PM, HME and printed tablets, respectively (Table 3). A similar pattern of a faster F2 compared to F1 release was seen under a 20 dpm dipping rate and the 0.2 M and 0.4 M added ionic strengths. This shows that the presence of the insoluble EC polymer in F1 made the tablet more hydrophobic than the F2 tablets, which may have been responsible for the slow drug release. Biomedicines 2023, 11, 375 7 of 17 Figure 1. Dissolution profiles of F1 and F2 tablets prepared by the three manufacturing methods (PM, HME and 3D-printed) measured under dipping rates of 10 dpm and 20 dpm, and added ionic strengths of 0.2 M and 0.4 M. Error bars indicate SD values. For the clarity of figures, error bars were drawn from just one side. Table 3. Parameters extracted from dissolution profiles, including DE% values and similarity (f 2) and difference (f 1) factors of the dissolution profiles of the various tablets dissolved at different dipping rates. Formulation/Method DE% * 10 dpm 20 dpm F1 PM tablets F1 HME tablets F1 printed tablets F2 PM tablets F2 HME tablets F2 printed tablets 16.57 33.76 13.12 45.32 53.91 57.12 19.82 39.14 13.33 65.35 68.90 69.41 f 2 f 1 Inference 72.27 62.10 99.08 34.27 40.36 44.71 16.45 13.67 1.93 29.96 21.52 17.53 Similar Similar Similar Different Different Different * No added KCl in dipping rate studies. Ionic strength was from the buffer composition (baseline ionic strength). From Figure 1, it can be deduced that thermal treatment plays a major role in the release of theophylline. This is obvious, since the dissolution profiles of both the heat- Biomedicines 2023, 11, x FOR PEER REVIEW 6 of 17 2.5. Release Kinetics The release kinetics for all the tablets prepared by the three methods were analysed using DDSolver software (an add-in program in Microsoft Excel) [20]. Data were fitted to the Korsmeyer–Peppas model, while special attention was paid to the n values of this equation in order to explore any change in the release mechanisms when the tablets were subjected to different dipping rates and ionic strengths. Below is the Korsmeyer–Peppas model, where Mt/M∞ is the fraction of the drug released at time t, K is the drug release rate constant and n is the release exponent: Mt/M∞ = Ktn The values of the diffusional exponent, n, differ based on the geometry of the tablet. For 60% of the release profile of a cylindrical-shaped tablet, the drug release is governed by diffusion (Fickian model) when n approaches 0.45. However, when n approaches 0.85, swelling (polymer chain relaxation) is the main release mechanism, where the drug re-lease follows zero-order kinetics (non-Fickian model, case II). For 0.45 < n < 0.85, the drug release is governed by both diffusion and swelling (anomalous transport). Values of the diffusional exponent, n, exceeding 0.85 mean that the drug release is controlled by poly-mer relaxation (non-Fickian model, super case II) [21]. 3. Results and Discussion 3.1. Effects of Dipping Rate and Formulation Composition Dissolution tests were carried out in three replicates for all PM, HME and 3D-printed tablets of both formulations under different dipping rates (10 dpm and 20 dpm) and added ionic strengths (0.2 M and 0.4 M). The dissolution profiles are depicted in Figure 1. The dissolution results at 10 dpm in different pH media were considered as the reference, which was used to investigate the effects of a higher agitation rate (20 dpm) and added ionic strength (0.2 and 0.4 M) on the dissolution rate of the tablets manufactured via three different methods. The DE% values were used for the overall comparison of the dissolu-tion behaviour, and the release rate calculated for each pH value was used for an in-depth understanding of the release behaviour and also to determine how the formulation com-position and manufacturing methods affected the dissolution behaviour of tablets. The release rate is usually calculated in two different ways, either from the release% vs. time plot or from the release% vs. square root of the time plot [6,22]. As the release% versus time showed higher R2 values compared to when the release% was plotted versus the square root of time, the former was used to calculate the release rate at each pH value. Figure 1 shows that tablets prepared by PM and HME had higher SD values com-pared to the printed tablets, implying that the printing method could improve the repro-ducibility of the dissolution behaviour, as shown by the reduced variations from batch to batch. The reason behind this is not clear, but the increased hardness of the printed tablets compared to the HME and PM tablets, as proven in our previous work, could have had an impact [16]. Biomedicines 2023, 11, x FOR PEER REVIEW 7 of 17 Figure 1. Dissolution profiles of F1 and F2 tablets prepared by the three manufacturing methods (PM, HME and 3D-printed) measured under dipping rates of 10 dpm and 20 dpm, and added ionic strengths of 0.2 M and 0.4 M. Error bars indicate SD values. For the clarity of figures, error bars were drawn from just one side. Looking at the dissolution graphs in Figure 1, it is clear that the F1 formulation had a much slower release pattern compared to F2, regardless of the manufacturing method and dissolution conditions. Under the 10 dpm dissolution testing condition, the DE% val-ues of F1 were 16.57%, 33.76% and 13.12% for the PM, HME and printed tablets, respec-tively, while F2 showed a faster release with DE% values of 45.32%, 53.91% and 57.12% for the PM, HME and printed tablets, respectively (Table 3). A similar pattern of a faster F2 compared to F1 release was seen under a 20 dpm dipping rate and the 0.2 M and 0.4 M added ionic strengths. This shows that the presence of the insoluble EC polymer in F1 made the tablet more hydrophobic than the F2 tablets, which may have been responsible for the slow drug release. From Figure 1, it can be deduced that thermal treatment plays a major role in the release of theophylline. This is obvious, since the dissolution profiles of both the heat-treated F2 tablets, i.e., the HME tablets made of extruded filaments and the 3D-printed tablets, showed faster dissolution rates than the PM tablets under various dissolution con-ditions. For example, the DE% of the HME and printed tablets of F2 at 10 dpm were 53.91 and 57.12%, while this was lower for the PM tablets at 45.32%. This was also observed in our previous work where distilled water was used as the dissolution medium [16]. The effect of heat on promoting the dissolution rate can be attributed to the fast cooling of the melted/softened material, leading to an increased free volume and better water penetra-tion to amorphous polymers, as described in other studies [16,23,24]. Another possible explanation is that, during heat treatment, the theophylline particles highly disperse in the molten polymer matrix, and because theophylline is a water-soluble drug, this can improve water penetration to the polymer matrix and facilitate the drug release [3]. A similar pattern was seen for the F2 tablets in other dissolution conditions, i.e., the dipping rate of 20 dpm, and ionic strengths of 0.2 M and 0.4 M. With the F1 formulation, a similar effect of heat treatment was observed for the HME tablets, but the 3D-printed F1 tablets were anomalous with an unusually low DE%. Biomedicines 2023, 11, 375 8 of 17 treated F2 tablets, i.e., the HME tablets made of extruded filaments and the 3D-printed tablets, showed faster dissolution rates than the PM tablets under various dissolution conditions. For example, the DE% of the HME and printed tablets of F2 at 10 dpm were 53.91 and 57.12%, while this was lower for the PM tablets at 45.32%. This was also observed in our previous work where distilled water was used as the dissolution medium [16]. The effect of heat on promoting the dissolution rate can be attributed to the fast cooling of the melted/softened material, leading to an increased free volume and better water penetration to amorphous polymers, as described in other studies [16,23,24]. Another possible explanation is that, during heat treatment, the theophylline particles highly disperse in the molten polymer matrix, and because theophylline is a water-soluble drug, this can improve water penetration to the polymer matrix and facilitate the drug release [3]. A similar pattern was seen for the F2 tablets in other dissolution conditions, i.e., the dipping rate of 20 dpm, and ionic strengths of 0.2 M and 0.4 M. With the F1 formulation, a similar effect of heat treatment was observed for the HME tablets, but the 3D-printed F1 tablets were anomalous with an unusually low DE%. Table 3 shows that the dissolution efficiency values (DE%) of the F1 formulation were increased when the agitation rate was increased from 10 dpm to 20 dpm for the PM and HME tablets, but this was not the case for the 3D-printed tablets. For example, the DE% increased from 16.57% to 19.82% in the case of the physical mixtures, whereas this value remained the same in the case of the 3D-printed tablets (around 13% at both agitation rates). This indicates that the 3D-printed F1 tablets were more robust against the agitation rate; hence, they were expected to produce a more consistent release profile in vivo under various GI motility rates. On the other hand, the 3D-printed F2 formulation showed some variability, where an increase in the agitation rate increased its DE% value to some extent. Despite this, the change in the dissolution efficiency was much smaller for the 3D-printed tablets compared with the HME and PM tablets, with percentage changes of 31, 22 and 19% for the PM, HME and 3D-printed tablets, respectively, when the agitation rate increased from 10 to 20 dpm. This once again shows that the 3D-printed tablets showed more resistance against the agitation rate compared to the tablets made from PM or HME. Generally speaking, an increase in the DE% at a high agitation rate could be due to the depletion of the surface gel layer (diffusion layer) due to the agitation. The thickness of the surface diffusion layer is inversely correlated with the dissolution rate and its depletion is expected to increase the drug diffusion based on the Noyes–Whitney equation [1]. The depletion of the gel layer can also restrict the effect of gelling on controlling the drug release, making the drug release at 20 dpm faster than at a low agitation speed (10 dpm). Table 3 also gives the similarity factors calculated for the same tablet tested at two dif- ferent agitation rates. The similarity values show a significant difference in the dissolution profile of all F2 tablets compared to the insignificant change for all F1 tablets, indicating a higher sensitivity of the F2 formulations to the agitation rate compared to the F1 formula- tion. For example, the DE% of PM tablets with the F2 formulation increased from 45.32% at 10 dpm to 65.35% at 20 dpm, showing a significant difference in the dissolution profiles with a low similarity factor of f 2 = 34.27%. On the other hand, the F1 formulation showed a much lower increase in DE%, from 16.57 to 19.82%, with a high similarity of dissolution profiles at various agitation rates with f 2 = 72.27%. The high sensitivity of the F2 formula- tion could be due to the polymeric matrix containing only the hydrophilic polymer HPC, which is probably more susceptible to gel layer depletion [6] than the polymer matrix in F1, which has 35% of the non-swellable polymer, EC, in its composition. Overall, the 3D-printed F1 tablets have the highest f 2 values when the release profiles are compared between the 10 and 20 dpm agitation rates (Table 3). This means that the 3D-printing method along with the use of EC improved the formulation robustness to agitation changes. This could be attributed to both the hydrophobic composition (EC) and the hard and solid structure of the printed tablets, making them more resistant to hydrodynamic mechanical stress. Biomedicines 2023, 11, 375 9 of 17 The unexpectedly low DE% of the 3D-printed F1 tablets, which was also seen under different dipping rates (Table 3) and ionic strengths (Table 4), may be justified by the lack of disintegration of the 3D-printed tablets (Figure 2). Both the PM and printed tablets showed a higher integrity during dissolution compared to HME tablets, with slower disintegration. As proven in our previous work [16], the disintegration of HME tablets can be attributed to the high water uptake of this F1 formulation. Although the printed tablets were prepared from hot-melt extruded filaments, they showed no disintegration during the dissolution period, which also correlates with the high hardness of these tablets (>550 N), as reported in our previous work [16]. Table 4. DE% values obtained from dissolution testing performed with a dipping rate of 10 dpm at two different ionic strengths, the baseline ionic strength (0 M KCl added) and 0.2 M added KCl, along with similarity (f 2) and difference (f 1) factors for these two profiles. Formulation/Method DE% 0 M KCl 0.2 M KCl Similarity Factor Difference Factor Inference F1 PM tablets F1 HME tablets F1 printed tablets F2 PM tablets F2 HME tablets F2 printed tablets 16.57 33.76 13.12 45.32 53.91 57.12 17.05 32.48 13.03 44.27 55.76 54.01 85.19 73.62 94.55 80.25 76.46 70.97 6.91 8.16 4.98 4.10 4.20 5.83 Similar Similar Similar Similar Similar Similar Figure 2. F1 tablets after dissolution test. From left to right: printed, HME and PM tablets. The three types of tablets had the same cylindrical geometry with the same SA/V (0.8 mm−1) before dissolution. However, the residual after dissolution (shown in this figure) showed that PM and printed tablets preserved their integrity while HME tablets showed erosion, which could be responsible for the increased release % of HME tablets compared to PM and printed ones, as shown in Figure 1. It is also notable that in the case of the F2 formulation under high agitation (20 dpm), the difference between the DE% values (Table 3) was not significant, with the DE% values varying between 65 and 69%. Figure 3 shows the release rate of the various tablets at 10 dpm. The release rate was calculated using a zero-kinetics model that showed a good fit for the parts of the release profiles in each pH medium (R2 values > 0.99). The F1 tablets contained 35% of the hydrophobic polymer, EC, while F2 contained just HPC (Table 1). This difference in the polymer composition led to considerable differences in the release rates, as discussed earlier. Figure 3 shows that for F1, there was a burst release in the first hour of the test at a pH of 1.2, followed by more steadiness in the release rate for the rest of the dissolution run at increasing pH levels. For example, the HME tablets showed drug release rates of 0.27, 0.19, 0.16, 0.16, 0.13 and 0.12 mg/min at pH values of 1.2, 2.2, 5.8, 6.8, 7.2 and 7.5, respectively. However, F2 in general showed a burst release with a sounder gradual decrease in the release rate over time (Figure 3). For instance, the HME tablets started with a release rate of 0.44 mg/min as a burst release at a pH of 1.2, which then dropped to 0.36, 0.26, 0.22, 0.14 and 0.10 mg/mL as the pH increased to 2.2, 5.8, 6.8, 7.2 and 7.5, respectively. It must be noted that the solubility of theophylline, a very weak acid of pKa 8.81, was not expected to change considerably at pH variations [25]. Biomedicines 2023, 11, x FOR PEER REVIEW 9 of 17 Figure 2. F1 tablets after dissolution test. From left to right: printed, HME and PM tablets. The three types of tablets had the same cylindrical geometry with the same SA/V (0.8 mm−1) before dissolution. However, the residual after dissolution (shown in this figure) showed that PM and printed tablets preserved their integrity while HME tablets showed erosion, which could be responsible for the increased release % of HME tablets compared to PM and printed ones, as shown in Figure 1. It is also notable that in the case of the F2 formulation under high agitation (20 dpm), the difference between the DE% values (Table 3) was not significant, with the DE% values varying between 65 and 69%. Table 3. Parameters extracted from dissolution profiles, including DE% values and similarity (f2) and difference (f1) factors of the dissolution profiles of the various tablets dissolved at different dip-ping rates. Formulation/Method DE% * f2 f1 Inference 10 dpm 20 dpm F1 PM tablets 16.57 19.82 72.27 16.45 Similar F1 HME tablets 33.76 39.14 62.10 13.67 Similar F1 printed tablets 13.12 13.33 99.08 1.93 Similar F2 PM tablets 45.32 65.35 34.27 29.96 Different F2 HME tablets 53.91 68.90 40.36 21.52 Different F2 printed tablets 57.12 69.41 44.71 17.53 Different * No added KCl in dipping rate studies. Ionic strength was from the buffer composition (baseline ionic strength). Figure 3 shows the release rate of the various tablets at 10 dpm. The release rate was calculated using a zero-kinetics model that showed a good fit for the parts of the release profiles in each pH medium (R2 values > 0.99). The F1 tablets contained 35% of the hydro-phobic polymer, EC, while F2 contained just HPC (Table 1). This difference in the polymer composition led to considerable differences in the release rates, as discussed earlier. Fig-ure 3 shows that for F1, there was a burst release in the first hour of the test at a pH of 1.2, followed by more steadiness in the release rate for the rest of the dissolution run at in-creasing pH levels. For example, the HME tablets showed drug release rates of 0.27, 0.19, 0.16, 0.16, 0.13 and 0.12 mg/min at pH values of 1.2, 2.2, 5.8, 6.8, 7.2 and 7.5, respectively. However, F2 in general showed a burst release with a sounder gradual decrease in the release rate over time (Figure 3). For instance, the HME tablets started with a release rate of 0.44 mg/min as a burst release at a pH of 1.2, which then dropped to 0.36, 0.26, 0.22, 0.14 and 0.10 mg/mL as the pH increased to 2.2, 5.8, 6.8, 7.2 and 7.5, respectively. It must be noted that the solubility of theophylline, a very weak acid of pKa 8.81, was not expected to change considerably at pH variations [25]. The burst release, in the beginning, is likely to be driven by the release of the drug on the surface of the tablets. This is followed by a decrease in the release rate over time, which can be explained by two mechanisms as follows. First, in the case of water-soluble drugs such as theophylline, the release rate is related to the amount of drug available for disso-lution, which decreases over time, causing the release rate to decrease [26]. The second reason is the decrease in the tablet’s surface area as a result of dissolution during the dis-solution testing, which, according to the Noyes–Witney equation, can decrease the release rate by the end of the test [1]. Biomedicines 2023, 11, 375 10 of 17 Figure 3. Bar charts comparing release rate at each pH medium during dissolution test at 10 dpm for PM, HME and printed tablets in both formulations. The burst release, in the beginning, is likely to be driven by the release of the drug on the surface of the tablets. This is followed by a decrease in the release rate over time, which can be explained by two mechanisms as follows. First, in the case of water-soluble drugs such as theophylline, the release rate is related to the amount of drug available for dissolution, which decreases over time, causing the release rate to decrease [26]. The second reason is the decrease in the tablet’s surface area as a result of dissolution during the dissolution testing, which, according to the Noyes–Witney equation, can decrease the release rate by the end of the test [1]. 3.2. Effect of Ionic Strength The initial ionic strength of the buffer solutions was used as the reference for data comparison. A dipping rate of 10 dpm was used for the ionic strength study. Parameters extracted from the dissolution profiles recorded with modified ionic strengths are listed in Tables 4 and 5. Data from Table 4 shows that an increased ionic strength due to the addition of 0.2 M KCl had no significant effect on the dissolution profiles of any of the tablets. However, according to Table 5, when the ionic strength was further increased by adding 0.4 M KCl, different dissolution profiles were observed for several types of tablets. Table 5. Dissolution efficiency values (DE%) obtained from dissolution testing performed with a dipping rate of 10 dpm using buffers with two different ionic strengths, the baseline ionic strength (with 0 M KCl added) and with 0.4 M KCl added, along with the similarity (f 2) and difference (f 1) factors for these two dissolution profiles. Formulation/Method DE% 0 M KCL 0.4 M KCL Similarity Factor Difference Factor F1 PM tablets F1 HME tablets F1 printed tablets F2 PM tablets F2 HME tablets F2 printed tablets 16.49 33.66 13.08 45.2 53.81 57.12 14.07 23.03 9.85 31.6 39.27 36.97 75.98 45.63 72.77 37.90 40.83 32.89 17.02 32.11 24.52 30.04 26.8 53.71 Inference Similar Different Similar Different Different Different Table 5 shows that, in general, an increase in the added ionic strength from 0 to 0.4 M caused a reduction in the release rate, resulting in a low similarity between the release profiles, as seen from the low f 2 values for several tablet types. Furthermore, the DE% of formulation F2 was more sensitive to ionic strength changes compared to the F1 formulation. When the percentage change in the DE% as a result of the 0.4 M KCl addition was calculated for both the F1 and F2 formulations, the results showed that the changes in the DE% for F1 were 15, 25 and 27%, whereas these values for the F2 formulation were 32, 30 and 35% for the PM, HME and 3D-printed tablets, respectively. This indicates that the F2 tablets were not consistent when the composition and ionic strength of the media was modified. However, for F1, the high f 2 values for the PM and 3D-printed tablets Biomedicines 2023, 11, x FOR PEER REVIEW 10 of 17 Figure 3. Bar charts comparing release rate at each pH medium during dissolution test at 10 dpm for PM, HME and printed tablets in both formulations. 3.2. Effect of Ionic Strength The initial ionic strength of the buffer solutions was used as the reference for data comparison. A dipping rate of 10 dpm was used for the ionic strength study. Parameters extracted from the dissolution profiles recorded with modified ionic strengths are listed in Tables 4 and 5. Data from Table 4 shows that an increased ionic strength due to the addition of 0.2 M KCl had no significant effect on the dissolution profiles of any of the tablets. However, according to Table 5, when the ionic strength was further increased by adding 0.4 M KCl, different dissolution profiles were observed for several types of tablets. Table 5 shows that, in general, an increase in the added ionic strength from 0 to 0.4 M caused a reduction in the release rate, resulting in a low similarity between the release profiles, as seen from the low f2 values for several tablet types. Furthermore, the DE% of formulation F2 was more sensitive to ionic strength changes compared to the F1 formula-tion. When the percentage change in the DE% as a result of the 0.4 M KCl addition was calculated for both the F1 and F2 formulations, the results showed that the changes in the DE% for F1 were 15, 25 and 27%, whereas these values for the F2 formulation were 32, 30 and 35% for the PM, HME and 3D-printed tablets, respectively. This indicates that the F2 tablets were not consistent when the composition and ionic strength of the media was modified. However, for F1, the high f2 values for the PM and 3D-printed tablets show the consistency of their release profiles when varying the dissolution media and a resistance to the increased ionic strength. A more detailed comparison of the dissolution profiles can be seen in Figures 4 and 5, where the dissolution rates at various pH values of the dissolution medium are illus-trated. The release rates of the F2 tablets (Figure 4) showed an interesting pH-dependant pattern, where adding 0.2 M ionic strength led to an increase in the release rate at a pH of 1.2 and also at some higher pH values. This increase in the release rate could be justified by the effect of alkaline ions such as potassium on improving the solubility of theophylline [27]. Table 4. DE% values obtained from dissolution testing performed with a dipping rate of 10 dpm at two different ionic strengths, the baseline ionic strength (0 M KCl added) and 0.2 M added KCl, along with similarity (f2) and difference (f1) factors for these two profiles. Formulation/Method DE% Similarity Factor Difference Factor Inference 0 M KCl 0.2 M KCl F1 PM tablets 16.57 17.05 85.19 6.91 Similar F1 HME tablets 33.76 32.48 73.62 8.16 Similar F1 printed tablets 13.12 13.03 94.55 4.98 Similar F2 PM tablets 45.32 44.27 80.25 4.10 Similar F2 HME tablets 53.91 55.76 76.46 4.20 Similar F2 printed tablets 57.12 54.01 70.97 5.83 Similar Biomedicines 2023, 11, 375 11 of 17 show the consistency of their release profiles when varying the dissolution media and a resistance to the increased ionic strength. A more detailed comparison of the dissolution profiles can be seen in Figures 4 and 5, where the dissolution rates at various pH values of the dissolution medium are illustrated. The release rates of the F2 tablets (Figure 4) showed an interesting pH-dependant pattern, where adding 0.2 M ionic strength led to an increase in the release rate at a pH of 1.2 and also at some higher pH values. This increase in the release rate could be justified by the effect of alkaline ions such as potassium on improving the solubility of theophylline [27]. Figure 4 shows that the release rate at a pH of 1.2 decreased when 0.4 M KCl was added. As a result of the 0.4 M KCL addition, in the first hour of dissolution at a pH of 1.2, the theophylline release rate of the F2 formulation was decreased from 0.35, 0.44 and 0.46 mg/min for the PM, HME and 3D-printed tablets to 0.30, 0.29 and 0.33 mg/min, respectively, with PM showing the least reduction. This observation is opposite to what was expected from the chaotropic effect of K +; however, considering the complexity of the formulation and manufacturing process, it is obvious that there are several other processes that may be more dominant in controlling the drug release than the mere pure drug factors. For example, the effect of KCl ions on the gelling and salting out of HPC polymers and water penetration into various manufactured tablets may be affected by the excess salt, which will need to be further investigated. In addition, it has been suggested before that a high ionic strength can reduce the drug release from certain matrix formulations by reducing the water available for drug transportation [18,28]. Figure 4. Release rate in each pH medium during dissolution test at different added ionic strengths for F2. Biomedicines 2023, 11, x FOR PEER REVIEW 12 of 17 the free volume of the polymer and a decrease in the density following these processes reported earlier [16]. Other studies have also shown that with increased free volume, the water uptake will be higher and this will affect mostly the free water available for trans-portation rather than the bound water; hence, the gel’s ability to retard the drug release will be reduced, and higher release rates will be observed [28]. Figure 4. Release rate in each pH medium during dissolution test at different added ionic strengths for F2. Similarly, the preparation method clearly affected the release behaviour of the F1 for-mulation (Figure 5). Although all tablets showed a decrease in the release rate at the in-creased ionic strength (0.4 M KCl), this decrease was more pronounced in the HME tab-lets. At a pH of 6.8, the HME tablets had a release rate of 0.16 mg/min when no KCl was added; this dropped to 0.09 mg/min (43% reduction) with 0.4 M KCl. However, the PM and 3D-printed tablets showed a lower reduction of 37% and 8%, respectively, in the re-lease rate at this pH as a result of the 0.4 M KCl addition. Similarly, in a pH of 2.2, the release rate dropped by 28%, 48% and 35% for the PM, HME and 3D-printed tablets, re-spectively. Contrary to the PM and 3D-printed F1 tablets, the HME tablets with the F1 formulation showed disintegration during dissolution in 0 M KCl compared to 0.4 M, where the tablet integrity was preserved during dissolution. This high sensitivity of the HME tablets to ionic strength may be due to a change in the hydrophilic/hydrophobic balance between EC and HPC as a result of the hot-melt extrusion, a behaviour that was also seen in our previous work [16]. Biomedicines 2023, 11, 375 12 of 17 Figure 5. Release rate in each pH medium during dissolution test at different ionic strength additions for F1. Increasing the concentration of ions in the medium can reduce the amount of free water molecules available for the polymer hydration, since ions will compete with the polymer for binding with water molecules, leaving the polymer with less hydration/water penetration than what is required for the drug release [3,29]. Two types of water are found within the hydrated polymer, bound water and free water. The bound water is responsible for gelling behaviour while the free water is responsible for the transportation of solutes and thus, drug release and polymer dissolution [28]. When ions are added, free water can be reduced while maintaining gelling behaviour and tablet integrity, which would decrease the drug release [3,4]. It is worth mentioning that the 3D-printing method increased the sensitivity of the F2 formulation to the 0.4 M added KCl, with this tablet showing the least similarity of its release profile at f 2 = 32.89% (Table 5). This might be due to the free volume changes under the heating/fast cooling process that may change the swelling behaviour of HPC and make it more susceptible to the ionic strength salting-out effect [30]. The release rates with 0.4 M KCl can provide an understanding for how the preparation method may change the drug release behaviours of the formulation. As for F2, in the PM tablets, the release rate with 0.4 M KCl decreased over time, with 0.30, 0.19, 0.15, 0.07, Biomedicines 2023, 11, x FOR PEER REVIEW 13 of 17 Figure 5. Release rate in each pH medium during dissolution test at different ionic strength addi-tions for F1. At high pH values such as 7.2 and 7.5, the 3D-printed tablets had an increase in their release rate by 5% and 25% with 0.4 M of added ionic strength compared to 0 M, which was absent in the PM and HME tablets, which had lower release rates at a high ionic strength (Figure 5). The same pattern was seen for 0.2 M KCl, where the release rate from the 3D-printed tablets, unlike that of the PM and HME tablets, increased by 60% and 122% at pH values of 7.2 and 7.5 compared to the release rate at 0 M KCl. This increase was also seen in the F2 formulations, both for the 3D-printed and HME tablets, which could be attributed to the thermal effect leading to a decreased density and an increased specific volume after the heating/cooling process, and this could also be associated with expanded voids within the printed matrix compared to the HME and PM tablets, as discussed in the authors’ previous work [16]. This increase in specific volume, in turn, could increase the water penetration, improving the theophylline solubility and release, especially in the presence of basic ions at higher pH values. The dissolution rate-increasing effect of 0.4 KCl at high pH values, which is unique to tablets with a heat-treated manufacture, may indicate a substantial change in the microstructure of the formulation due to the melting and fast cooling processes. Biomedicines 2023, 11, 375 13 of 17 0.07 and 0.08 mg/min at pH values of 1.2, 2.2, 5.8, 6.8, 7.2 and 7.5, respectively. This is a similar pattern to its release in a baseline medium with no KCl (Figure 4). However, after extruding and 3D printing, the release pattern of the F2 formulation changed in the dissolution medium containing 0.4 M KCl (Figure 4) and the reductions in the release rate over time became much smaller. In the HME tablets, the release rate was 0.29 mg/min at a pH of 1.2, and it decreased to 0.22, 0.19, 0.22, 0.17 and 0.13 mg/min at pH values of 2.2, 5.8, 6.8, 7.2 and 7.5, respectively. Similarly, in the printed tablets, the release rate was 0.33 mg/min at a pH of 1.2, which then decreased to 0.19, 0.16, 0.15, 0.17 and 0.15 mg/min at pH values of 2.2, 5.8, 6.8, 7.2 and 7.5, respectively. This indicates some major physical changes after the thermal process and fast cooling, which was also seen by an increase in the free volume of the polymer and a decrease in the density following these processes reported earlier [16]. Other studies have also shown that with increased free volume, the water uptake will be higher and this will affect mostly the free water available for transportation rather than the bound water; hence, the gel’s ability to retard the drug release will be reduced, and higher release rates will be observed [28]. Similarly, the preparation method clearly affected the release behaviour of the F1 for- mulation (Figure 5). Although all tablets showed a decrease in the release rate at the increased ionic strength (0.4 M KCl), this decrease was more pronounced in the HME tablets. At a pH of 6.8, the HME tablets had a release rate of 0.16 mg/min when no KCl was added; this dropped to 0.09 mg/min (43% reduction) with 0.4 M KCl. However, the PM and 3D-printed tablets showed a lower reduction of 37% and 8%, respectively, in the release rate at this pH as a result of the 0.4 M KCl addition. Similarly, in a pH of 2.2, the release rate dropped by 28%, 48% and 35% for the PM, HME and 3D-printed tablets, respectively. Contrary to the PM and 3D-printed F1 tablets, the HME tablets with the F1 formulation showed disintegration during dissolution in 0 M KCl compared to 0.4 M, where the tablet integrity was preserved during dissolution. This high sensitivity of the HME tablets to ionic strength may be due to a change in the hydrophilic/hydrophobic balance between EC and HPC as a result of the hot-melt extrusion, a behaviour that was also seen in our previous work [16]. At high pH values such as 7.2 and 7.5, the 3D-printed tablets had an increase in their release rate by 5% and 25% with 0.4 M of added ionic strength compared to 0 M, which was absent in the PM and HME tablets, which had lower release rates at a high ionic strength (Figure 5). The same pattern was seen for 0.2 M KCl, where the release rate from the 3D-printed tablets, unlike that of the PM and HME tablets, increased by 60% and 122% at pH values of 7.2 and 7.5 compared to the release rate at 0 M KCl. This increase was also seen in the F2 formulations, both for the 3D-printed and HME tablets, which could be attributed to the thermal effect leading to a decreased density and an increased specific volume after the heating/cooling process, and this could also be associated with expanded voids within the printed matrix compared to the HME and PM tablets, as discussed in the authors’ previous work [16]. This increase in specific volume, in turn, could increase the water penetration, improving the theophylline solubility and release, especially in the presence of basic ions at higher pH values. The dissolution rate-increasing effect of 0.4 KCl at high pH values, which is unique to tablets with a heat-treated manufacture, may indicate a substantial change in the microstructure of the formulation due to the melting and fast cooling processes. 3.3. Release Kinetics The Korsmeyer–Peppas model was applied to the dissolution data and its coefficients, K and n, were calculated. The model fit to all the data was excellent, with R2 values over 0.99. The resulting n values for all formulations are listed in Table 6. The diffusional exponent (n) is an indication of the mechanism of drug release [31], and its values can indicate how different dipping rates and ionic strengths affected the release mechanisms. If n < 0.45 (for tablet shape), the release mechanism is through the diffusion of the drug out of the matrix according to the concentration gradient. This is the Fickian diffusion method, Biomedicines 2023, 11, 375 14 of 17 where solvent diffusion is less impactful than polymer chain mobility. If n approaches 0.85, it means that swelling/gelling behaviour governs the release of the drug and the main mechanism of drug release is erosion. In this case, polymer relaxation happens and the chain mobility is slower than water diffusion. Anomalous diffusion for 0.45 < n < 0.85 happens when the water mobility in the matrix allows diffusion based on the concentration while swelling is happening and controlling the release. Therefore, in anomalous mechanisms, both diffusion and swelling/dissolution control the drug release [32]. It follows that any change in swelling behaviour can change the n value. Since the tablets prepared in this study had similar geometries in terms of their surface area and shape, the changes in the n values would have been driven by changes in the formulation under extrusion and high temperatures or by dissolution conditions. Table 6. The Korsmeyer–Peppas n values for drug release from F1 and F2 tablets under different dissolution conditions of dipping rate (10 or 20 dpm) and added KCl concentrations. Formulation/Method 10 dpm, 0 M KCl 20 dpm, 0 M KCl 10 dpm, 0.2 M KCl 10 dpm, 0.4 M KCl F1 PM tablets F1 HME tablets F1 printed tablets F2 PM tablets F2 HME tablets F2 printed tablets 0.469 0.687 0.606 0.733 0.844 0.873 0.469 0.716 0.582 0.765 0.913 0.907 0.391 0.587 0.642 0.687 0.771 0.791 0.389 0.553 0.613 0.527 0.767 0.692 Table 6 shows that increasing the dipping rate could increase the n values to some extent, especially for F2. This is probably because of the higher depletion of the surface gel layer, making erosion of the swelled layer more dominant for the drug release. On the other hand, the n values decreased with increasing ionic strength. The F1 HME tablets had an n value of 0.687 at 0 M KCl, which decreased to 0.587 and 0.553 with 0.2 M and 0.4 M KCl, respectively. Studies have shown that the more the swelling controls the release mechanism, the higher the n value is [32], which means that the swelling/erosion role in the drug release decreases under high ionic strengths. This claim agrees with the effect of the ionic strength on the swelling behaviour by reducing the free water available for drug transport within the gel/swollen layer [3,29]. In terms of the effect of the manufacturing method, changes in the n values showed that the thermal process of formulations can make the swelling/erosion influence sounder on the release mechanism. For example, the release mechanism of F2 moved from anomalous with n = 0.733 in the PM tablets to super case II after 3D printing with n = 0.873 (near zero- order release). This could have resulted from the increase in free volume. Water sorption can be faster and higher due to an increased free volume under the heating/cooling process, while 3D printing allows swelling to be the main release mechanism [16,32]. The same pattern was observed at 20 dpm (Table 6). Another example is F1 in a high-ionic-strength medium (0.4 M KCl). Without any thermal process, the PM tablets had Fickian diffusion (with n = 0.389), meaning the swelling effect was minimal, probably due to the effect of KCl ions on polymer dehydration. However, the drug release mechanism for the 3D-printed tablets became anomalous with n = 0.613, meaning the formulation underwent physical changes under extrusion and high temperatures that made the swelling/erosion mechanism more effective due to the thermal process. It can be concluded that the manufacturing method, and also the dissolution conditions, can alter the mechanism of drug release from diffusion to erosion or vice versa. The 3D printing of tablets can introduce, as a manufacturing method, several physical changes that impact the release profile of the same formulation made into tablets by conventional tableting methods. Few comparison studies have been performed in this area, and changes in geometry, hardness, porosity and true density were not included in the dissolution data interpretation [17,32]. Thus, the authors have carried out more in-depth Biomedicines 2023, 11, 375 15 of 17 and integral comparison studies in their previous work, including all aforementioned factors [16]. The FDM printing method was compared to the HME and PM methods, and it was found that the printed tablets had the highest hardness and porosity, as discussed in this paper. In addition, the authors proved the decreased true density in printed tablets because of the fast cooling process, distinguishing their study from other pharmaceutical comparison studies. The current work gives a deeper understanding of how the FDM 3D-printing method impacts a formulation’s robustness to dissolution conditions, which is an extension of the authors’ previous work [16]. Overall, increasing the dipping rate can make the drug release faster, while increasing the ionic strength can achieve the opposite. Introducing a hydrophobic polymer, EC, seems to reduce the formulation’s sensitivity to dissolution variables. The influence of the decreased true density in printed tablets— associated with an increased free volume as discussed previously—on the release profile was obvious in alkaline buffers. As mentioned in other studies [28], the increase in free volume will allow more free water in the matrix, which can justify a higher release rate for hydrophilic formulations (containing HPC) in the printed tablets than in non-thermally processed tablets (PM ones). Changes in the release kinetics after printing can be also linked to a free volume increase (true density decrease) after thermal printing. When a water-swellable polymer is used, such as HPC in this study, water penetration can be higher due to an increased free volume induced by the heating/fast cooling process. This can cause the release mechanism to be controlled more by swelling/erosion, giving a high diffusional exponent (n value) in the Korsmeyer–Peppas model. This study shows how both the formulation composition and manufacturing method affect a tablet’s robustness to dissolution behaviour, which helps when developing robust tablets for in vivo dissolution conditions based on the used tableting method. 4. Conclusions The effect of three tablet manufacturing methods, i.e., conventional tablet compression from a physical mixture, compression of hot-melt extruded (HME) filaments, and 3D FDM printing, on the release of theophylline measured under simulated fed and fasted conditions was studied using a biodissolution tester (USP 3 apparatus). The dissolution efficiency (DE%) increased by increasing the dipping rate, regardless of the formulation composition, while raising the ionic strength beyond a certain level decreased the DE%, indicating the effect of KCl ions on salting out the swellable polymers, and hence prolonging the drug release. Despite having similar drug-release patterns, the hydrophilic formulation with just the HPC polymer was highly sensitive to changes in the dipping rate and ionic strength compared to the formulation containing both HPC and the hydrophobic polymer, EC. The hydrophilic formulation also showed a high variability in the dissolution behaviour under different agitation rates and ionic strengths, indicating that under in vivo conditions, this formulation will probably be susceptible to changes under fed and fasted states in the GI tract, while the EC-containing formulation was more robust to variable dissolution conditions. The manufacturing method also impacted the formulation’s dissolution profile and its variability under different dissolution conditions. When EC was included in the formulation, the 3D-printing and conventional tableting methods produced more consistent dissolution profiles under variable dissolution conditions compared to the HME method. However, the 3D-printing method increased the variability of the dissolution profile as a result of the KCl addition when the hydrophilic formulation was used, indicating possible variable release behaviour in vivo under different fed and fasted conditions. This observa- tion is mostly a formulation-related effect rather than a 3D-printing manufacturing effect, since the other two manufacturing methods also produced similarly high variability upon a 0.4 M KCl addition. Thus, the 3D-printing method seems to be preferable for hydrophobic formulations, and it produces a consistent release profile under different agitations and ionic strengths. The release mechanism was also dependent on the manufacturing method, where the thermal treatment in both the 3D-printing and HME methods could significantly increase the n values of the Korsmeyer–Peppas model. The results prove that the method Biomedicines 2023, 11, 375 16 of 17 of manufacturing is as critical as the composition in the design of robust formulations with consistent release profiles under various physiological conditions. Author Contributions: Conceptualization, A.N.; methodology, N.N. and S.C.; software, N.N. and A.N.; validation, N.N., M.L. and A.N.; formal analysis, N.N.; investigation, N.N., S.C. and A.N.; resources, A.N.; data curation, N.N., M.L., T.G. and A.N.; writing—original draft preparation, N.N.; writing—review and editing, M.L., T.G. and A.N.; visualization, A.N.; supervision, M.L., T.G. and A.N.; project administration, A.N. All authors have read and agreed to the published version of the manuscript. Funding: This research received no external funding. Institutional Review Board Statement: Not applicable. Informed Consent Statement: Not applicable. Data Availability Statement: Not applicable. Conflicts of Interest: The authors declare no conflict of interest. References 1. Mudie, D.M.; Amidon, G.L.; Amidon, G.E. Physiological Parameters for Oral Delivery and In vitro Testing. Mol. Pharm. 2010, 7, 2. 3. 1388–1405. [CrossRef] [PubMed] Vinarov, Z.; Abdallah, M.; Agundez, J.A.G.; Allegaert, K.; Basit, A.W.; Braeckmans, M.; Ceulemans, J.; Corsetti, M.; Griffin, B.T.; Grimm, M.; et al. Impact of gastrointestinal tract variability on oral drug absorption and pharmacokinetics: An UNGAP review. Eur. J. Pharm. Sci. 2021, 162, 105812. [CrossRef] [PubMed] Johnson, J.L.; Holinej, J.; Williams, M.D. Influence of ionic strength on matrix integrity and drug release from hydroxypropyl cellulose compacts. Int. J. Pharm. 1993, 90, 151–159. [CrossRef] 6. 5. 4. Mitchell, K.; Ford, J.L.; Armstrong, D.J.; Elliot, P.N.C.; Hogan, R.; Hogan, J.E. The influence of additives on the cloud point, disintegration and dissolution of hydroxypropylmethylcellulose gels and matrix tablets. Int. J. Pharm. 1990, 66, 233–242. [CrossRef] Gray, V.A. Power of the Dissolution Test in Distinguishing a Change in Dosage Form Critical Quality Attributes. AAPS PharmSciTech 2018, 19, 3328–3332. [CrossRef] [PubMed] Asare-Addo, K.; Levina, M.; Rajabi-Siahboomi, A.R.; Nokhodchi, A. Study of dissolution hydrodynamic conditions versus drug release from hypromellose matrices: The influence of agitation sequence. Colloids Surf. B Biointerfaces 2010, 81, 452–460. [CrossRef] [PubMed] Pillay, V.; Fassihi, R. Unconventional Dissolution Methodologies. J. Pharm. Sci. 1999, 88, 843–851. [CrossRef] Klein, S. The Use of Biorelevant Dissolution Media to Forecast the In Vivo Performance of a Drug. AAPS J. 2010, 12, 397–406. [CrossRef] [PubMed] Ramos Pezzini, B.; Gomes Ferraz, H. Bio-dis and the paddle dissolution apparatuses applied to the release characterization of ketoprofen from hypromellose matrices. AAPS PharmSciTech 2009, 10, 763–771. [CrossRef] [PubMed] 7. 8. 9. 10. Rohrs, B.R.; Burch-Clark, D.L.; Witt, M.J.; Stelzer, D.J. USP dissolution apparatus 3 (reciprocating cylinder): Instrument parameter effects on drug release from sustained release formulations. J. Pharm. Sci. 1995, 84, 922–926. [CrossRef] [PubMed] 11. Yu, L.X.; Wang, J.T.; Hussain, A.S. Evaluation of USP apparatus 3 for dissolution testing of immediate-release products. AAPS PharmSci. 2002, 4, 1–5. [CrossRef] 12. Pezzini, B.R.; Issa, M.G.; Duque, M.D.; Ferraz, H.G. Applications of USP apparatus 3 in assessing the In Vitro release of solid oral dosage forms. Braz. J. Pharm. Sci. 2015, 51, 265–272. [CrossRef] 13. Hamed, R. Physiological parameters of the gastrointestinal fluid impact the dissolution behavior of the BCS class IIa drug valsartan. Pharm. Dev. Technol. 2018, 23, 1168–1176. [CrossRef] [PubMed] 14. Hamed, R.; AlJanabi, R.; Sunoqrot, S.; Abbas, A. The effect of pH, buffer capacity and ionic strength on quetiapine fumarate release from matrix tablets prepared using two different polymeric blends. Drug Dev. Ind. Pharm. 2017, 43, 1330–1342. [CrossRef] [PubMed] 15. Klanˇcar, U.; Markun, B.; Baumgartner, S.; Legen, I. A novel beads-based dissolution method for the in vitro evaluation of extended release HPMC matrix tablets and the correlation with the in vivo data. AAPS J. 2013, 15, 267–277. [CrossRef] [PubMed] 16. Nashed, N.; Lam, M.; Ghafourian, T.; Pausas, L.; Jiri, M.; Majumder, M.; Nokhodchi, A. An Insight into the Impact of Thermal Process on Dissolution Profile and Physical Characteristics of Theophylline Tablets Made through 3D Printing Compared to Conventional Methods. Biomedicines 2022, 10, 1335. [CrossRef] [PubMed] 17. Zhang, J.; Feng, X.; Patil, H.; Tiwari, R.V.; Repka, M.A. Coupling 3D printing with hot-melt extrusion to produce controlled release tablets. Int. J. Pharm. 2017, 519, 186–197. [CrossRef] [PubMed] Biomedicines 2023, 11, 375 17 of 17 18. Asare-Addo, K.; Conway, B.R.; Larhrib, H.; Levina, M.; Rajabi-Siahboomi, A.R.; Tetteh, J.; Boateng, J.; Nokhodchi, A. The effect of pH and ionic strength of dissolution media on in-vitro release of two model drugs of different solubilities from HPMC matrices. Colloids Surf. B. Biointerfaces 2013, 111, 384–391. [CrossRef] [PubMed] 19. Asare-Addo, K.; Levina, M.; Rajabi-Siahboomi, A.R.; Nokhodchi, A. Effect of ionic strength and pH of dissolution media on theophylline release from hypromellose matrix tablets—Apparatus USP III, simulated fasted and fed conditions. Carbohydr. Polym. 2011, 86, 85–93. [CrossRef] 20. Zhang, Y.; Huo, M.; Zhou, J.; Zou, A.; Li, W.; Yao, C.; Xie, S. DDSolver: An add-in program for modeling and comparison of drug dissolution profiles. AAPS J. 2010, 12, 263–271. [CrossRef] 21. Ritger, P.L.; Peppas, N.A. A simple equation for description of solute release I. Fickian and non-fickian release from non-swellable devices in the form of slabs, spheres, cylinders or discs. J. Control. Release 1987, 5, 23–36. [CrossRef] 22. Vetchy, D.; Vetcha, M.; Rabiskova, M.; Gryczova, E.; Bartosikova, L. Comparison in vitro felodipine release rate from the original versus generic product with controlled release of the drug. Medicina 2007, 43, 326–331. [CrossRef] [PubMed] 23. Wang, J.; Hopmann, C.; Kahve, C.; Hohlweck, T.; Alms, J. Measurement of specific volume of polymers under simulated injection molding processes. Mater. Des. 2020, 196, 109136. [CrossRef] 24. Guo, J.-H.; Robertson, R.E.; Amidon, G.L. Influence of Physical Aging on Mechanical Properties of Polymer Free Films: The Prediction of Long-Term Aging Effects on the water Permeability and Dissolution Rate of Polymer Film-Coated Tablets. Pharm. Res. 1991, 8, 1500–1504. [CrossRef] 25. Available online: https://pubchem.ncbi.nlm.nih.gov/compound/Theophylline (accessed on 27 December 2022). 26. Kim, S.W.; Bae, Y.H.; Okano, T. Hydrogels: Swelling, Drug Loading, and Release. Pharm. Res. 1992, 9, 283–290. [CrossRef] 27. Cohen, J.L. Analytical Profiles of Drug Substances; Florey, K., Ed.; European University Institute: Florence, Italy; Academic Press: Cambridge, MA, USA, 1975; pp. 466–493. 28. Abdelmola, F. Influence of Voids on Water Uptake in Polymer Panel. Ph.D. Thesis, Florida Atlantic University, Boca Raton, FL, 29. 30. USA, 2018. Judah, H.L.; Liu, P.; Zarbakhsh, A.; Resmini, M. Influence of buffers, ionic strength, and pH on the volume phase transition behavior of acrylamide-based nanogels. Polymers 2020, 12, 2590. [CrossRef] Siepmann, J.; Peppas, N.A. Modeling of drug release from delivery systems based on hydroxypropyl methylcellulose (HPMC). Adv. Drug Deliv. Rev. 2001, 48, 139–157. [CrossRef] 31. de Naylor, T.V. Permeation Properties. In Comprehensive Polymer Science and Supplements; Allen, G., Bevington, J.C., Eds.; Pergamon 32. Press: Oxford, UK; New York, NY, USA, 1989; pp. 643–668. Shi, K.; Salvage, J.P.; Maniruzzaman, M.; Nokhodchi, A. Role of release modifiers to modulate drug release from fused deposition modelling (FDM) 3D printed tablets. Int. J. Pharm. 2021, 597, 120315. [CrossRef] Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
10.1177_16094069221150106
Regular Article Governance Diaries: An Approach to Researching Marginalized People’s Lived Experiences in Difficult Settings International Journal of Qualitative Methods Volume 22: 1–12 © The Author(s) 2023 DOI: 10.1177/16094069221150106 journals.sagepub.com/home/ijq Miguel Loureiro1 , Anuradha Joshi1, Katrina Barnes2, and Eg´ıdio Chaimite3 Abstract How do chronically poor and marginalized citizens interact with and make claims to the different public authorities that exist in fragile, conflict and violence-affected contexts? In other words, how does governance from below look like in difficult settings? Given the centrality of the ‘leave no one behind’ agenda, an understanding of how such populations meet their governance needs can help identify the constraints to achieving development for all in these challenging settings. We wanted to research these questions comparatively, to see if there were common features of response in different contexts, with the presence of various kinds of non-state actors, diverse histories of colonialism and authoritarianism, and widely different social norms. In this article we describe the governance diaries approach, an iterative alternative to large-n surveys and multi-sited ethnographies we developed in the process of answering these questions. Governance diaries, working as a qualitative panel data, are a suitable approach for researching complex behavior that changes over time as large-n surveys are insufficiently dynamic to trace the processes behind change (lacking sensitivity) and ethnographic studies often have limited generalizability (lacking comparability). We describe here how this approach works and the challenges and opportunities it offers for research. Keywords ethnography, methods in qualitative inquiry, mixed methods, par participatory action research, conversation analusis Introduction What does governance look like for those governed in settings affected by fragility, conflict, and violence? How can we find out? This article describes an approach that we evolved in the process of answering this question— governance diaries. We wanted to understand what empowerment and accountability meant to chronically poor and marginalized people living in these settings. How do people navigate the formal and in- formal institutions that govern their lives? What historical and socio-political influences shape their behaviors? We wanted to research these questions comparatively, to see if there were common features of response in contexts with different types and levels of conflict, the presence of various kinds of non- state actors, diverse histories of colonialism and authoritari- anism, and widely different social norms. Most of our understandings of empowerment and ac- countability come from places with relatively stable and ef- fective states. Historically, accountability gains have emerged through broad-based movements for socio-economic rights, yet lessons from these places have limited relevance in fragile contexts where fear and trauma born of experiences of re- pression and violence make social and political action rare. Fragile contexts are marked by weak and fragmented state institutions which lack legitimacy, whilst simultaneously other non-state actors control territory, provide public services and claim to represent the population. This makes claim-making 1Research Fellow, Institute of Development Studies, University of Sussex, Brighton, UK 2Evidence Uptake and Learning Lead, Oxfam GB, Oxford, UK 3Researcher, Instituto de Estudos Sociais e Económicos (IESE), Maputo, Mozambique Corresponding Author: Miguel Loureiro, Institute of Development Studies, University of Sussex, Library road, Brighton BN1 9RE, UK. Email: m.loureiro@ids.ac.uk Creative Commons Non Commercial CC BY-NC: This article is distributed under the terms of the Creative Commons Attribution-NonCommercial 4.0 License (https://creativecommons.org/licenses/by-nc/4.0/) which permits non-commercial use, reproduction and distribution of the work without further permission provided the original work is attributed as specified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/open-access-at-sage). 2 International Journal of Qualitative Methods complex, as it is unclear who is accountable for what, and on what basis. These features, combined with low levels of trust, suggest that informal networks carry greater weight. Yet, researching marginalized groups’ experiences in fragile set- tings is difficult: challenges around gaining permissions, ac- cessing populations, developing trust and ensuring safety of both respondents and researchers, mean conventional research methods are either time-intensive or infeasible. While de- veloping this research, we needed a methodology that allowed us to capture poor and marginalized households’ diversity of experiences with a medium ‘n’ that allowed us to make some mid-level generalizable observations across contexts about how governance takes place and what for accountability. it means This article focuses on the approach we developed for meeting these needs –governance diaries.1 Governance diaries involve following the same set of households/individuals over a long period with regular visits, enabling in-depth research questions, following household events as they unfold. We used governance diaries in Mozambique, Myanmar, and Pakistan from 2017–2021, in two phases with an interim analysis and reflection period in between. In the first phase we focused the diaries on 164 poor and marginalized households to enquire how they interacted with and made claims on the diversity of public authorities present in their context, and whether these interactions advanced empowerment or accountability. Phase 1’s key finding was that poor and marginalized households rarely engage directly with state or non-state authorities; instead, they do nothing, self-provide, or most commonly, go through intermediaries. Households were usually unable to choose the intermediaries to approach – these were fixed for the group, or issue, or location. The discovery of the intermediaries’ importance led us to use governance diaries with them in phase 2 to understand their role and how they navigated different public authorities. The rest of the article is devoted to the approach rather than the substantive findings.2 The next section focuses on the challenges in exploring the lives of hard-to-reach populations and concludes showing how governance diaries were able to overcome some constraints. The third section details our re- search design and operationalization of key concepts tailored to our country contexts. In the fourth section we elaborate on field implementation and the fifth section describes some of the tools we used. In the sixth and seventh sections we offer some reflections on the approach, as well as on the insights that the approach allowed us to surface. Finally, we conclude with some ideas on the relevance of governance diaries to other research, and to adaptive practice. Researching Grassroots Experiences in Difficult Contexts Gaining a deep understanding of marginalized groups’ lived experiences is a daunting task, particularly so in fragile, conflict and violence-affected contexts (Atkinson and Flint, 2001; Clark, 2006; Cohen and Arieli, 2011; Jacobsen and Landau, 2003; Khan Mohmand et al., 2017). Key among the challenges are: · high levels of insecurity leading to an atmosphere of fear and distrust, restricting the open or easy flow of information between researchers and respondents. · cultural, social, and economic constraints, from lan- guage barriers to never having engaged with research, to relative positions in local hierarchies. · ideological, religious, and political affiliation related barriers, particularly when belonging to minority or powerless groups. · technical, administrative, and legal obstacles, such as mobility limitations and state regulatory frameworks. Overcoming these challenges to get at the questions we were interested in required extensive of fieldwork, for building trust and uncovering the real meaning of processes we heard about or observed. The challenges also make it more difficult to operationalize abstract concepts such as ‘accountability’ which may mean different things in different places, and for different groups (Fox, 2022). Moreover, comparing processes and indicators across contexts with varying levels and com- binations of fragility, violence and conflict required an iter- ative process of data collection, analysis, reflection, and adaptation (Khan Mohmand et al., 2017). Conventionally, there are two main approaches to such research: ethnographies and large-n surveys. Ethnographic work is better placed to explore more sensitive aspects of everyday life, as time spent embedded in local communities allows researchers to build relationships of trust with the respondents while observing as much as possible of their everyday lives. While ethnographic studies provide rich, descriptive information and insights about target populations, their small sample sizes, and purposive sampling limits generalizability (Lyberg et al., 2014). This is particularly problematic if a research aim is to compare cases across lo- cations: the attributes of different settings, as well as differ- ences in the nature and prevalence of hard-to-reach populations, lead to variations in problems making compar- ison difficult (Smith, 2014). Thus, while ethnographies are useful in concept building, they are less valuable in theory testing. By contrast, large-n surveys provide one-off snapshots that can present generalizable, comparative pictures across a larger set of cases suitable for theory testing. Large-n com- parative studies help eliminate competing hypotheses and isolate the explanatory, most often by having a control group (Jacobsen & Landau, 2003). Yet, such surveys are insuffi- ciently dynamic to trace the processes behind change: they do not allow for trust building or permit sufficient time for re- searchers to observe ‘the field’. We developed the governance diary approach drawing on the advantages of both approaches outlined above.3 It involves researchers interviewing the same households (or other actors) Loureiro et al. 3 regularly over several months to follow specific questions/ issues while tracing events as they unfold in real time. The idea is to generate a qualitative medium-n panel data set that allows for in-depth probing of certain themes in a comparative fashion, while enabling trust-building with respondents. We find that governance diaries are suitable for researching complex behavior that changes over time. They appear par- ticularly suitable for exploring issues that people are un- comfortable talking about to outsiders. Regular follow-up breaks down respondent reticence over time, thus reducing misinterpretation. The more entries in the ‘diary,’ the more trust between researchers and respondents and a deeper emerging understanding of issues being explored. In Myan- mar for instance, discussion themes changed from less sen- sitive health-related issues in the first couple of visits, to more sensitive conflict-related stories in subsequent ones. In Pakistan, a few households that mentioned health expenses in the first visits disclosed later that these expenses were actually money spent on drugs by some household members. Besides building trust, capturing people’s experiences over time provided us with new insights into how state and non-state institutions shift in relevance, legitimacy, and trust; and why these perceptions change over time. Designing Research and Operationalizing Concepts We started by defining and operationalizing key concepts and identifying our units of analysis. Here we highlight our key choices, our rationale, and how these changed along the way. Empowerment and Accountability We used Eyben, et al.’s (2008: 6) work as a starting point for conceptualizing empowerment, ‘when individuals and orga- nized groups are able to imagine their world differently and to realize that vision by changing the relations of power that have been keeping them in poverty’. We saw empowerment as both an action – the act of gaining control over decisions and resources that affects one’s life – as well as a state of being (objective and subjective), where people have a greater voice over decision making, allowing them to expand their choices (as well as the possibility of making those choices), and eventually giving them increased control over their own lives. For accountability we used Schedler’s (1999) conceptuali- zation as a broad two-way relationship between two parties, that incorporates both answerability (obligation to inform and justify actions) and enforcement (rendering judgements with attached sanctions or rewards). These initial definitions were refined during the research as it became clear that accountability in the strict sense of answerability and en- forcement did not quite capture the processes we observed; rather attention to sources of expectations and obligations on the one hand, and processes of scrutiny and judgements on the (Anderson greater forthcoming). other were importance of Public Authority We defined public authorities as formal and informal insti- tutions which ‘can undertake core governance functions: protection from external threats and managing external rela- tions; peaceful resolution of internal conflicts; and providing or facilitating the provision of a range of collective goods and services’ (Unsworth, 2010: 9). Using this definition allowed us to focus on functions rather than on form, while being more neutral about the processes and actors involved – inside or outside the formal state This broad definition had to be re- defined when faced with examples from the field. How were we to assess the healthcare claims that a woman heading a household made to a relative who seemed to have authority over her actions, as he was both a respected member of the community and a relative? After much discussion we agreed to exclude familial sources of authority unless they were accepted as responsible for public goods more broadly and were coincidentally related to households under study. Op- erationally we defined public authority as people, organiza- tions, or institutions who households considered responsible for the provision of particular public goods or services de- livered to a wide range of people within the community – e.g., traditional local governance institutions, armed groups, local state institutions, religious leaders – who had the legitimacy and capacity to carry out their functions. We did not hold a priori assumptions about the role that any of these played, as they played different functions across the three countries in our diaries. Public Goods Poor and marginalized households engage with public au- thorities on a range of governance issues. To choose the issues to follow, we began by focusing on core state functions that people might expect authorities to provide. Our choice of functions was steered by Stewart and Brown’s (2009) oper- ational concept of state fragility involving three dimensions, namely: authority failures; service failures; and legitimacy failures. We chose a core function from each dimension: security; health provision; and revenue collection.4 Over the course of the research, we reframed security into security, justice, and conflict resolution, as we could see an intercon- nection in the households’ stories of accessing security and justice and the existing conflict resolution mechanisms availed by them. We also added themes which emerged as important, specifically social protection, employment, and poverty, ac- cessing legal documentation.5 resources services, (other) and and 4 International Journal of Qualitative Methods Household Selection Our unit of analysis in phase 1 was the household, as gov- ernance issues are usually experienced by and responded to by households, rather than individuals.6 We collectively defined household as a group of people living together and sharing meals. We selected chronically poor households using Collins et al. (2009: 190, 195)’s definition: households that display evidence of deprivation of basic human needs that had existed over a long period of time (many months and often years) … such households are poor or “at the bottom end”. In this definition, we tested for local interpretations of households considered marginalized within each research community, either because of ascriptive identities, religion, or household characteristics. To account for intersectionality and contextual specificities (Crenshaw, 1989; Lyberg et al., 2014) our selection process took three stages. We started with a typology based on our previous knowledge of poverty and marginalization in each country: household composition; social status; assets and occupation; and type of dwelling. We then discussed and validated this categorization with country field researchers and refined the list at country level. For instance, while ethnicity was a key factor in Mozambique and Myanmar, it did not play a role in household marginalization in any of the sites in Pakistan. Instead, being a religious minority did, as did belonging to lower-ranked tribes. Not owning land was an important marker in most of Pakistan while not so much in Mozambique or Myanmar, as in these countries’ land arrangements revolve around user rights rather than ownership. Finally, we validated the categorization with local communities and purposely selected 10 to 15 households in each location during our first two visits. In this selection we aimed to have a significant number of female-headed households, a particularly marginalized population group in any country. Anticipating that some households were likely to drop out, we added a few additional households in each lo- cation during the second and third visits through snowball sampling,7 ended up with a total of 164 households across the three countries of which 40% (in Pakistan) to 54% (in Myanmar) were female-headed.8 Intermediary Selection In phase 2, we followed individuals or organizations acting as governance intermediaries as identified by our respondents from phase 1.9 Intermediaries were the first link in the gov- ernance chain that households reached out to for addressing an issue outside of the family. Thus, our set of intermediaries were identified by marginalized households rather than those that might have been identified by from the top down. In other words, these were the real brokers in mediating between the the poor and public authorities. Overwhelmingly male, intermediaries varied widely across the three countries and included village elders and respected individuals, political party workers, religious leaders, and members of social movements. As in phase 1, anticipating that intermediaries would drop out, we started with about 20–30 intermediaries roughly divided across four locations in each of the three countries. In the end we followed 81 intermediaries across the three countries. We started interviewing intermediaries in person; however, Covid-19 forced the countries into lockdown starting in late March and April 2020. We consequently switched physical visits to phone and voice over IP services. This was prob- lematic not only because intermediaries were often busy with Covid-19 issues and did not have the time or patience to speak with us, but they were also uncomfortable with the perceived insecurity of phone/online conversations (IDS, 2020). Yet the pandemic gave us an opportunity to witness how intermediaries across different locations dealt with the same crisis: how they interacted with citizens and public authorities, their roles, practices and strategies, and how they made themselves essential to the functioning of local governance systems. Implementing Governance Diaries Selecting Researchers to Build Trust: Gaining access and Gatekeepers Access and trust are key challenges of doing research with the most marginalized. Distrust is heightened by fragility, vio- lence, and conflict –speaking to the wrong people can lead to violence from powerholders, whether local armed groups or powerful political interests. We took the strategic decision of choosing field researchers10 who were closely linked to the locations, to reduce distrust and facilitate access, as well as to ensure that the interviews could be conducted in local lan- guages. This was achieved differently in each country. In one country, researchers were graduate students at a national university who returned home for the interviews. In another, we recruited junior faculty from regional universities who relocated to the research sites for substantial periods. In the third country, where ongoing ethnic conflict was an issue, we trained local civil society organization staff from the same ethnic background and language of each location. Having field researchers with existing connections to the locations meant our teams could gain access more easily both through local gatekeepers but also directly with households and intermediaries – in several cases the field team acted both as field researchers and gatekeepers. The researchers who were local worked through several local gatekeepers with whom they had previous connections with or formed con- nections with during the first research stage. The researchers who relocated for the study duration took time to establish networks and relationships to support the work before be- ginning interviews. Finally, the country team working through Loureiro et al. 5 local civil society organizations chose organizations well- known locally for service provision. Building and strength- ening local networks and relationships improved access and increased our chances that gatekeepers would trust us. In turn, this trust between local researchers and gatekeepers increased respondent willingness to take part in both phases: most households and intermediaries opened-up as they came to recognize the field researchers as people who understood their lived experiences. Working with and through local field researchers was also key to avoiding public authorities considering our activities suspicious. Research in fragile, conflict and violence-affected settings (FCVAS) can pose potential risks not only to re- spondents, but also to researchers (Campbell, 2017). It is crucial to be mindful of cultural, social, and political sensi- tivities and power relations, being aware of whom to ask what, as well as how to ask it (Goodhand, 2000). Conducting this research through a partnership of research institutes and an international NGO, and with local field researchers allowed us to have in place some of the tools and approaches aid agencies have developed for staff security (Mazurana and Gale, 2013), as well as the familiarity and experience of the settings to which Goodhand (2000) refers.11 Building Conversational Communities Among Researchers From the outset, the research plan involved periods of re- flection and analysis between field visits. The period be- tween the monthly field visits was used to transcribe and clean interviews, sense check what was being heard, identify issues for further probing, discuss new tools and start the analysis process. Every three months, there was a more intense period of reflection. Finally, there was a gap between the two phases for analysis and rethinking of strategy. The reflections took place through ‘conversational communities’ (Gudeman and Rivera, 1990). Throughout our research, we created three layers of conversational communities: a local one, where the field researchers were paired to share de- scriptive and reflective notes as a unit; a national one, where country research teams would reflect on their monthly ob- servations with the principal investigators and each other, and plan subsequent visits; and the international one, where principal investigators shared their observations and chal- lenges and tried to keep a comparative element to the whole process. As well as using conversations to improve data quality, a(nother) key activity of our conversational communities centered on unpacking the words and expressions respondents used to tell their stories, and the meanings behind them. To ensure we were questioning and measuring the same concepts across locations, we used a similar iterative procedure to that we had used for household selection. Keeping true to the intended meaning of our concepts entailed seeking agreement within the different possible interpretations by validating translations across a series of in-country actors, namely the households and communities, the intermediaries, the research teams, academics, think-tanks and research institutes, activ- ists, and development actors. A Menu of Field Tools Conducting interviews in these settings, even by local re- searchers, poses several challenges. A general atmosphere of suspicion in FCVAS influences the interviewing process, shaping not only the kind of questions we can ask, but also the strategies we must adopt in asking these questions (Goldstein, 2014). Interviewing in FCVAS requires patience, subtlety, and above all flexibility (Kovats-Bernat, 2002). The multitude of unpredictable parameters which could restrain the inter- viewing process (particularly during an active conflict) forced us to adopt a reflective/adaptive approach to fieldwork. In these situations, rigid questionnaires can appear threatening and inappropriate (Barakat and Ellis, 1997), so it is best to let respondents lead the discussion, as the more comfortable they feel, the more likely they will be put at ease resulting in more detail (Goldstein, 2014). We used conversations rather than formal to simply interview different households and observe their everyday activities, but to engage in long, relaxed conversations. Conversations were particularly valuable to help build trust and informality be- tween the field researchers and households; it allowed re- searchers and respondents to be at ease. interviews. The aim was not As part of the method, we used several tools from an array of methods suitable for research on social and political action in FCVAS (Khan Mohmand et al., 2017). The aim was to create a menu which field teams could use iteratively (Chambers, 1994) and allowed us to probe deeper into re- spondents’ answers, and them to reflect on their answers, while simultaneously reducing respondent fatigue (H¨oglund, 2011). Among these tools three were key for our research: ethnographic tools, institutional mapping, and anchoring vignettes. Ethnographic Tools We largely used three ethnographic tools: participant obser- vation; in-depth interviews with individuals and focus groups (including life histories); and thick description. Participant observation allowed for a holistic awareness of events as they unfolded, giving a more comprehensive understanding of what really matters to respondents. Central to this reliance on observation was the need to maintain an ‘anthropological eye’, that is, a sensibility to local culture. Being both insiders and outsiders enabled researchers to ask questions and make observations that local inhabitants would not usually think of. Positionality though, can become an issue when re- searchers are also gatekeepers in their own research. As most of our field research teams belonged to the study locations, there was a risk they would take local everyday life for 6 International Journal of Qualitative Methods granted and not record observations. Therefore, a significant element of the training focused on getting the field teams to become detached observers. We asked them to pretend they had landed from Mars; everything is new so worth recording and questioning. Having an anthropological eye was not only useful for observing how people reacted to special events, but also to ordinary ones. For instance, in the third visit to one of the households, during a moment of silence when the household had already hinted there was nothing new to report, one of the researchers noticed a cow outside the house. When asked if the cow was theirs, the household head replied it was not, but that she oversaw it. Probing deeper, we uncovered a local informal social protection arrangement through which poorer households rear non- poor households’ livestock in exchange for a cut when the cattle is sold. The iterative nature of the diaries meant that incomplete or unexpected observations shared and discussed amongst us in a previous visit, became follow-up questions to probe in the next. These follow-up questions, along with issues discussed across the three countries became semi-structured thematic mental prompt lists for the research teams. Having a prompt list in their minds allowed the field researchers to bring respondents back to the main issue when they moved away from the theme.12 Quite a few of these unstructured and semi-structured interviews developed into focus groups, as other respondents within the households joined the conversation with their opinions and stories. Life histories are useful tools for capturing processes of change. Life history interviews and life event analysis allowed for in- dividuals to discuss not only themselves and their lives, but also the social, economic, and political spaces that they inhabit. We adapted life histories to our needs, and in particular life event analysis, by focusing not on the life of the respondent – be it the entire household, or the intermediary – but on the life of the events themselves. the household head, The third ethnographic tool we used was thick description (Geertz, 1973), where we not only explain the behavior of individuals and groups within a society, but also its context, as understanding the local context is crucial to fill in certain blanks.13 Thick description was composed not only of facts but also of commentary and interpretation: what we find out as researchers is inherently connected with how we find it out (Emerson et al., 2011). Thus, it was critical that the field teams documented their own activities, circumstances, and emotional responses to fieldwork, as these shaped the process of observing and recording others’ lives (Emerson et al., 2011). Thick description also allowed us to engage with narrative analysis as an analytical method, to interpret social meaning in respondents’ everyday lives and inter- actions with different public authorities by analyzing their stories. It allowed us a richer thematic approach when an- alyzing the data. Institutional Mapping As the research unfolded, we felt we needed to bring in other tools not only to probe deeper into the nature of the relationships between households, intermediaries, and public authorities, but also to counteract survey fatigue. We decided to introduce institutional mapping as a participatory tool that could deal with these two issues.14 Institutional mapping allowed us to see the perceived importance, accessibility, and impact of different public authorities to different households, as well as the key institutions existing in their communities, and how they related both to each other and to external agencies involved in service delivery and the administration of programs. In each location we selected which households’ stories were most noteworthy and could help us gain a deeper knowledge of people’s perceptions of their local institutions – including different kinds of formal and informal public au- thorities – and their access (or lack of) to them. The main objective was to see which actors helped households solve problems. The field teams started by helping households identify key actors and their relative power in helping them solve problems (the more powerful the actor, the bigger the circle representing it), while simultaneously noting the basis that households used for determining power. Then the field teams asked how accessible these actors were to the house- holds, using distance from the center – household – as a measure of accessibility. Part of the exercise was to probe the map, to ask how and why households chose to identify those actors, the criteria they used to determine power, and the level of relationship between different actors. Anchoring Vignettes and Hypothetical Scenarios Studies based on small samples and in-depth interviews rarely allow for conducting longitudinal or geographic comparisons unless very carefully selected (Jacobsen and Landau, 2003). We used anchoring vignettes with households in phase 1 to measure empowering and disempowering situations when holding public authorities accountable, and with intermedi- aries in phase 2 to ascertain levels of answerability.15 These, we felt, would allow for cross-location analysis. Measuring is difficult not only empowerment and disempowerment technically, but also conceptually (Masset, 2015). As such, we developed a set of hypothetical scenarios against which we could calibrate respondents’ self-assessments (King et al., 2004). Each vignette described a situation related to an empowering/disempowering moment of a fictitious character. We used an amalgamation of real stories to construct them locally, making them plausible within each context. We chose a health-related event as this was the theme brought out by most households. After each vignette the field teams probed the rationale for the respondents’ answers. Following on from this reflection, the field teams used the exercise to inquire about respondents’ claim-making abilities with existing public Loureiro et al. 7 authorities. Similarly, we prepared anchoring vignettes to measure perceptions of accountability components (infor- mation, justification, and enforcement) among intermediaries with two scenarios: one mediating a land dispute resolution for rural locations, the other mediating citizen access to public security for urban locations. In Myanmar, due to scheduling issues of the project,16 we did not manage to use anchoring vignettes; however, we used similar vignettes within 18 focus group discussions to unpack the role of trust. These hypothetical scenarios – based on real stories that had emerged from the governance diaries – al- lowed us to create a conversation around how people would act if the hypothetical situation were to happen in their lo- cation. This enabled a slightly broader triangulation of what respondents had said in the household interviews, while also allowing for a reality check of the proposed actions of others from within their community to try to counter potential re- sponse bias. These hypothetical scenario vignettes allowed us to ask questions about trust in various governance actors indirectly and understand people’s attitudes and behavior in situations in which authorities were potentially involved. Asking directly whether participants trust certain authorities would have been too sensitive, making participants uncom- fortable about speaking openly; this way, the groups were able to have these conversations whilst allowing them to save face and not be speaking about specific individuals. Reflections on the Research Process The research process encountered several challenges, from the operationalization of key concepts across different contexts and languages, to respondent fatigue, to the amount of time required for training and accompaniment of the field teams due to the iterative nature of the approach and the need to carry out analysis as one went along. There were twelve languages used across the project, each having words for key concepts that either had additional implications, or several words for the concept with slightly different nuances. In Myanmar, for example, the term ‘authority’ often relates to ‘legitimacy’, while in Mozambique people use the term to refer to the police. In Myanmar, the literal word ‘trust’ in local languages has more spiritual meanings, and so we needed to find other ways to get at the concept without using the word. The definition for who could be considered public authority was also tricky – we were asked multiple times why God was not on our list of governance actors. The iterative nature of the governance diaries allowed us to come back from the field, reflect on the stories, and think of other ways of enquiry: sometimes because we had reached a methodological impasse and wanted to probe deeper; other times to look at issues from a different angle or epistemo- logical framework; or to reformulate an abstract query into something more concrete to allow respondents to reflect on their actions. Yet this iteration also posed challenges. Every three months the country teams would bring together field researchers and principal investigators to reflect on and un- dertake a preliminary analysis of trends that were coming out. However, due to skill sets, methodological issues, and con- fusion over roles, these were often spent on troubleshooting methodologies rather than analysis: whatever time we planned for training, it was never enough. We chose local field teams based on the need for local expertise, but as the visits were monthly, we did not employ them full-time– and naturally, they had other commitments. The local researchers were keen on the research and really wanted to learn; however, it was a gradual process to develop their confidence and create ownership of the research. At the end of phase 1, there were three key process-related lessons: first, the need to collectively develop and agree definitions, concepts and meanings, units of study, and targets right from the beginning. Second, there will always be a need for more training and more accompaniment than planned, due to the unpredictability of FCVAS and the fact that often we are doing remote research dependent on gatekeepers who may not have the same level of understanding of the research process. Third, while it is important to have an initial plan/design, it is imperative to keep the whole process flexible both in terms of tools and methodologies used and critical reflection in and between countries – be ready to embrace change. Our phase 2 research with intermediaries posed even greater challenges. To start, intermediaries were more asser- tive than households, largely because their role demanded it of them. This meant they had less patience with repetitive questions as well as being more demanding of some form of compensation for their time. After the initial few visits Covid- 19 struck, which meant we could no longer interview in person. As mentioned earlier, this complicated fieldwork as intermediaries became even busier solving local problems and did not want to talk on the phone or online for fear of sur- veillance. As a result, we lost several intermediaries and had to come up with alternative ways to keep in contact during Covid-19 restrictions.17 A concern we had from the beginning was that the research should not be extractive. The households have had long ex- perience of subjugation by different configurations of the state (repressive colonial, repressive post-colonial, authoritarian, neoliberal), in which local formal and informal public au- thorities played a key role, and we wanted the research to be a positive and reflective process for them. All participating households and intermediaries knew they would not receive payment for participating in the research, that there were no benefits, and that they could withdraw at any point.18 And yet, most households – particularly those in rural areas – kept talking to us, as did quite a few intermediaries even during the pandemic. Three key factors played a role: there was someone there to listen to them – both as catharsis as well as the feeling that someone cared about them; we could potentially help with their lives;19 and the different tools employed avoided mo- notonous repetitiveness and provided a welcome distraction in their everyday routines. 8 International Journal of Qualitative Methods Trust started developing between respondents and field teams from the third visit onwards. The idea of communi- cating with people who came from provincial capitals, places not easily accessible for those living in the most remote districts, served as a stimulus for the respondents to keep speaking with us. The fact that these highly educated uni- versity and NGO researchers came to their homes to hear about their lives made many feel important.20 Just the fact that we made recurring visits showed that we were worried about their situation, they said; even those who were initially reluctant to provide details became more receptive after the third visit. Some would even initiate the conversations themselves, as they already knew the nature of the research. One of the Pakistan researcher’s notes from a third visit reflects this: We have started feeling that people are becoming more com- fortable in talking to us as time passes. They discuss their issues more openly now. We felt this in the case of Mr. X, as he talked about his father’s murder in great length. He told us all the events that happened before, during and after the murder. He even shared the responses of the womenfolk in his family after the murder. Men in this region usually don’t share the stories of their women and the events that happen within the family, except to the closest of their friends. So, in a way, we are becoming his friends. By the third visit, we felt he was waiting for our arrival and was desperate to share the story. Governance diaries, the interviews, have be- come a source of catharsis for him. The regular visits took on a cathartic tone, including for women. As one field researcher noted: In the first two visits, [women] kept their answers very short; but now they have started opening up about their issues. One of the major problems with interviewing women in this region is that generally they don’t speak to males except those within the family. Women here are often kept aside from public social and political life: since their childhood they are taught to respect and obey their male members, starting with their brothers and fa- thers. At every point in their lives, they are told to recognize their inferior status in local society as compared to their male counterparts. In phase 2 we continued fieldwork in 12 locations out of the original 20 from phase 1. Households in these 12 locations played a key role helping us identify intermediaries at the start of the second phase and many kept interacting informally with field researchers when they visited these locations. While the interaction between field researchers and most households from the other eight locations stopped once phase 1 ended, there were still a few that infrequently would contact field researchers for assistance or to mark festive occasions such as Eid or Christmas. If not friends, they at least became good acquaintances and someone to network with. Reflections on Findings We have outlined above some of the opportunities and challenges presented by governance diaries. Yet, what evi- dence do we have that the approach generates useful insights that would not be generated otherwise? While it is beyond the scope of this article to elaborate on the results of the research, here we offer a brief glimpse into the insights and their im- plications.21 Traditional understandings of change from below suggest that marginalized groups: (a) express voice on a given issue or grievance; (b) mobilize (usually publicly) to hold appropriate authorities to account; and then (c) authorities will respond in a positive or negative fashion. Our research using the governance diaries in FCVAS challenged these assump- tions on several counts and showed the nuances and ambi- guities of how the process of claim-making works for marginalized groups. First, in FCVAS, histories of violence and fear mean that it is hard to change internalized norms and deep emotions of powerlessness that affect the possibility of voice. The over- whelming majority of the households felt disempowered and had low expectations of public authorities. There were, how- ever, interesting ways of coping with this marginalization – of rationalizing their situation and relative impotence. In all three countries, when people had no access to healthcare they chose deliberate ignorance, choosing not to find out about their ill- nesses as they could not afford treatment. In one location in Mozambique, when crimes were committed and criminals were known, people would resign themselves by saying, ‘[The criminals] are being resourceful; they are also getting by’. Fate was often invoked as an explanation, and many turned to God or prayer as a positive action to solve problems. Secondly, unsurprisingly, we found rapidly closing civil society spaces and a history of authoritarian regimes making collective action rare and risky. In these settings it is difficult for marginalized people to collectivize or even find a common identity. In our 164 households across the three countries, we only found two clear instances of collective action to either substitute public authorities or to hold them accountable. Understanding these instances of positive outliers seems to offer a promising start to unlocking constraints to account- ability action from below. Thirdly, and what triggered phase 2 of our research, marginalized people rarely reach state or non-state authorities directly but work through intermediaries who mediate their claims. Intermediaries navigate these diverse sources of au- thority, working across formal and informal local governance systems, and being the deciders themselves at times as they too exercise significant authority. There are strong pressures within the local governance system involving intermediaries to resolve things locally. Higher authorities decline to take on problems till serious efforts at local levels have failed. These authorities impose hierarchies and punish ‘skipping levels’. As one intermediary from Myanmar noted, ‘I always try the to make the big crime small, and the small crime best Loureiro et al. 9 disappear’. Self-provision is common, with low expectations of higher-level authorities. This does not mean that people see self-provision as the best solution; it is often the least bad. Overall, we see local governance systems as a web of networks that take different shapes depending on the location. Standard structures rarely apply, even within one region. Across all three countries we see a diversity of actors and institutions key for local-level decision making and gover- nance needs. These public authorities are neither inherently good nor bad: just because poor and marginalized people identify them as legitimate public authorities does not mean they deliver, are just, or accountable. Conclusions The governance diaries approach uses a basket of methods and sits between a medium ‘n’ survey which gathers information in a single snapshot, and a detailed ethnography taking place in several locations simultaneously over an extended period. This hybridity offers several advantages. Repeated interac- tions with the same respondents combined with adapting to use different methods for different needs – for both data collection and analysis – allows for a relatively open-ended agenda in FCVAS, and to slowly ‘tighten the net’ by focusing on particular stories. Thus, the approach probes deeper into households’ access to different public authorities while limiting respondent fatigue. In the process, households can also visualize and reflect on their position vis-`a-vis different public authorities. Similarly, understanding how intermedi- aries deal with issues presented to them enables a triangu- lation of the household findings, as well as looking upwards to see how the intermediaries fit within broader governance networks. Analysis of the diaries gave us pause to think about how people at the margins imagine the state and public authority. In all our focus countries people thought of state and public authority in ways different from our conceptions, variously invoking the state as an absent father who is remiss in providing as in Mozambique, or as an arm of repression for people who want to be left alone. Alternative forms of public authority were perceived to have different degrees of legitimacy and credibility in terms of delivering on public goods. While we used this approach to understand governance issues in FCVAS, it can be used for other questions in different settings where researchers are interested in a ground-up view of how particular services, institutions or discourses are ex- perienced and perceived. The approach is particularly valuable in places where: (a) there is likely to be limited trust between populations of interest and outsiders, particularly around re- search; (b) where the environment is rapidly changing re- quiring shifting responses from people; or (c) where the subject matter is particularly sensitive and one-shot ap- proaches are unlikely to generate accurate data. One could imagine using the approach with other hard-to-reach populations, such as internally displaced people (IDPs) and refugees in camps, those living with stigmatized illnesses, unemployed youth in industrialized countries, or even with frontline workers or mid-level administrators in hierarchical bureaucracies. In addition to these, we can imagine the governance diaries approach used as an independent, real-time, accompaniment for development programs, to sense test whether theories of change and assumptions work, or whether they need adap- tation. Such a strategy can strengthen monitoring, evaluation and learning tools. Given the current emphasis on adaptive including ‘thinking and working politically’; approaches, ‘doing development differently’; and ‘program driven itera- tive adaptation’; governance diaries can offer an additional source of evidence on what is working, what is not, and what the reasons might be for the observed interim results. We are already seeing governance diary adaptations to explore the extreme poor’s urban livelihood strategies in Bangladesh pre- Covid 19 (Devereux and Shahan, 2020) and the new poor’s during and post-Covid 19 (Durdiner diaries, under the CLEAR project), the intersectionality of gender and health in urban informal settlements in Sierra Leone (Conteh et al., 2021), household financial governance and coping strategies in D.R.C. (Stys et al., 2021), women’s struggles against backlash in South Asia (struggle diaries), peoples’ struggles during the military coup in Myanmar (emergency diaries), and the impact of Covid-19 on young people’s lives in Nepal and Indonesia (livelihood diaries). The urgent need to understand how public policies and programmes affect the most marginalized make the search for appropriate tools and methods that can offer insights a priority for development practitioners and researchers. Governance diaries offers a powerful practical approach to meet this need – it is relatively cost-effective, offers real-time qualitative data, and can be used in comparative research to produce mid-range generalizable insights that have wide implications. Acknowledgments We would like to thank our respondents across all locations in all three countries for their openness and teachings. We would also like to thank the country field teams, without whom this research would have not been possible: in Mozambique Andissene Andissene, Abudo Gimo, Ana Paula Meque, Muaziza Omar, Domingos Sa´ıte, Gerson Selemane, and Gaspar Tocoloa; in Pakistan Affaf Ahmed, Muddabir Ali, Haider Butt, Zaibunissa Ejaz, Haider Kaleem, Danyal Khan, Noor ur Rehman, Hamid Riaz, and Rizwan Wazir. Unfortunately, we cannot name the Myanmar research team members for security reasons. Finally, we would like to thank the following people for their ideas, suggestions, and research support at different stages of the research: Colin Anderson, Stephanie de Chassy, Evelina Dagnino, Jenny Edwards, Salvador Forquilha, John Gaventa, Duncan Green, Ammar Jan, Jane Lonsdale, Aung Myo Min, L´ucio Posse, Jo Rowland, and Alex Shankland. 10 International Journal of Qualitative Methods Declaration of Conflicting Interests The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article. Funding The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This re- search was conducted as part of the Action for Empowerment and Accountability (A4EA) Research Programme (PO 7239), funded by the UK Foreign, Commonwealth and Development Office (FCDO – formerly DFID) based at the Institute of Development Studies, University of Sussex. ORCID iD Miguel Loureiro  https://orcid.org/0000-0001-6793-0649 Notes 1. The approach was inspired by the ‘Portfolios of the Poor’ in which poor households were recruited to keep diaries about their finance (Collins et al., 2009). Parallel to those financial diaries, we settled on ‘governance diaries’ to name the similar approach we were planning to understand governance, as it captured the spirit of the approach, even though not literally true. As is ex- plained further later, these are not diaries in the conventional sense kept by participants, but diaries in the sense of noting events as they unfold. 2. For a summary of the key findings see Anderson et al. (forthcoming). 3. Governance diaries are not actual diaries as most of the research participants in these settings are illiterate. Moreover, physical (written or visual) diaries in these settings can expose vulnerable people to further risk. 4. We chose revenue collection as a proxy for understanding the legitimacy and representativeness of the relevant public authority and the extent to which people experienced it not as coercion, but as an implicit fiscal social contract (see Brautigam et al., 2008). 5. Land registration titles, identity cards, etc. 6. We chose household over family as the unit of analysis because of the latter’s borders being more fluid and relating mostly to kinship (Das, 1973; Loureiro, 2013). 7. At the end of phase 1, the attrition rate was about 25 per cent. 8. We selected the household head as the main respondent for the household but allowing other members to participate. Initially women and younger members within male-headed households would engage less with the field researchers, regardless of the field researchers’ gender. Over time and with increasing trust these members started to interact more and be more vocal, with a certain variation across countries: from a less conservative Mozambique to a more conservative Pakistan. 10. Along with ten principal researchers, there was a total of 16 field researchers across all countries and phases. In the first phase, there were four field researchers visiting 79 households across seven Myanmar locations (assisted by 12 staff members of two local CSOs), six field researchers visiting 47 households across eight Mozambican locations, and six field researchers visiting 38 households across five Pakistani locations. In the second phase we reduced the number of locations to four in each country and therefore reduced the number of field researchers as well: four four visited 31 visited 33 intermediaries intermediaries in Mozambique, and three visited 17 intermedi- aries in Pakistan. in Myanmar, 11. Research ethics in FCVAS are both more difficult to negotiate and more important than in other settings. It is particularly critical to uphold informed consent and always follow the ‘do no harm’ principle. Throughout both phases we maintained informed consent by regularly reminding all households and intermediaries of their power to decline to answer any question and to withdraw at any time without negative repercussions. 12. This was particularly useful when we started tracking specific stories within each household. 13. There are excellent examples of ethnographic thick descriptions in FCVAS such as Daniel’s (1996) work in Sri Lanka; Nordstrom and Robben’s (1995) collection of essays by anthropologists who have experienced political violence first-hand; Smyth and Robinson’s (2001) edited volume on ethical and methodologi- cal issues while researching violently divided societies; and Mazurana et al.’s (2013) edited volume of lessons and reflections by a group of academics, journalists, and filmmakers on their own role within FCVAS. 14. A participatory visual method to identify and represent people’s perceptions of key institutions and individuals, their relation- ships, and importance. 15. Due to the impact of Covid-19 on our research, we only used accountability anchoring vignettes in Pakistan. 16. One of the advantages of the iterative nature of the approach is the possibility of adding methodological tools at any time. The disadvantage of this in a cross-country comparative study is that it takes time for each team to learn how to use each tool, par- ticularly a new method. As we devised anchoring vignettes at the end of phase 1, the Myanmar team did not integrate them in their final visits, as they felt the focus was primarily to gather stories from the households. 17. In one instance, the interview was conducted whilst jogging together in the park. 18. Although this had not been promised, we did pay each household a small sum at the end of the research. 19. The field teams often assisted households with reading gov- ernment documents or medical prescriptions or putting them in contact with pro bono lawyers and NGOs. 9. By governance intermediaries we mean individuals or organi- zations who for diverse reasons are approached by and play a mediating role – formally or informally – between chronically poor and marginalized households and public authorities. 20. This was not true for Myanmar, however, where ethnic tensions made local populations extremely suspicious of outsiders. 21. For more details on the findings see Anderson et al. forthcoming, and on the implications of the findings see Barnes et al., 2021. Loureiro et al. References Anderson, C. (forthcoming) ‘Unpacking accountability relationships: Insights from the Action for Empowerment and Accountability research programme’. Development Policy Review. Online ahead of print. Anderson, C., Joshi, A., BarnesAhmed, K.A., Ali, M., Chaimite, E., Forquilha, S., Khan, D., Khan, R., Loureiro, M., Posse, L., Rowlands, J., & Shankland, A. (forthcoming) Everyday gov- ernance in areas of contested power: Insights from Mozambi- que, Myanmar and Pakistan. Development Policy Review. Online ahead of print. Atkinson, R., & Flint, J. (2001). Accessing hidden and hard-to-reach populations: Snowball research strategies. In Social research up- date (pp. 1–4). University of Surrey, Department of Sociology, 33. Barakat, S., & Ellis, S. (1997). Researching under fire: Issues for consideration when collecting data and information in war circumstances, with specific reference to relief and recon- struction projects. Disasters, 20(2), 149–156. https://doi.org/10. 1111/j.1467-7717.1996.tb00525.x Barnes, K., Anderson, C., de Chassy, S., Ahmed, A., Ali, M., AungChaimite, M.E., Joshi, A., Khan, D., Loureiro, M., Posse, L., Rowlands, J., Shankland, A., & Wazir, R. (2021). Under- standing governance from the margins: What does it mean in practice?’ A4EA Policy and practice paper. Brighton: Institute of Development Studies. Brautigam, D., Fjeldstad, O.-H., & Moore, M. (2008). Taxation and state-building in developing countries: Capacity and consent. New York: Cambridge University Press. Campbell, S. (2017). Ethics of research in conflict environments. Journal of Global Security Studies, 2(1), 89–101. https://doi. org/10.1093/jogss/ogw024 Chambers, R. (1994). The origins and practice of participatory rural appraisal. World Development, 22(7), 953–969. https://doi.org/ 10.1016/0305-750x(94)90141-4 Clark, J. A. (2006). Field research methods in the Middle East. Political Science and Politics, 39(3), 417–423. https://doi.org/ 10.1017/s1049096506060707 Cohen, N., & Arieli, T. (2011). Field research in conflict environ- ments: Methodological challenges and snowball sampling. Journal of Peace Research, 48(4), 423–435. https://doi.org/10. 1177/0022343311405698 Collins, D., Morduch, J., Ratherford, S., & Ruthven, O. (2009). Portfolios of the poor: How the world’s poor live on $2 a day. Princeton University Press. Conteh, A., Wilkinson, A., & Macarthy, J. (2021). Exploring gender, health, and intersectionality in informal settlements in freetown. Gender & Development, 29(1), 111–129. https://doi.org/10. 1080/13552074.2021.1885215 Crenshaw, K. (1989) Demarginalizing the intersection of race and sex: a black feminist critique of antidiscrimination doctrine, feminist theory, and antiracist politics (140, pp. 139–167). University of Chicago Legal Forum. Daniel, E. (1996). Charred lullabies: Chapters in an anthropology of violence. Princeton University Press. 11 Das, V. (1973). The structure of marriage preferences: An account from Pakistani fiction. Man, 8(1), 30–45. https://doi.org/10. 2307/2800608 Devereux, S., & Shahan, A. (2020). Improving the lives of the urban extreme poor (ILUEP): Understanding urban livelihood tra- jectories in Bangladesh. Research report for concern world- wide: Round 2. Centre for Social Protection, Institute of Development Studies and Development Research Initiative. Emerson, R., Fretz, R., & Shaw, L. (2011). Writing ethnographic fieldnotes (2nd edition). Chicago University Press. Eyben, R., Kabeer, N., & Cornwall, A. (2008). Conceptualising em- powerment and the implications for pro-poor growth: A paper for the DAC poverty network. Brighton: Institute of Development Studies. Fox, J. (2022). ‘Accountability keywords’. Working paper. Ac- countability Research Center, American University. Geertz, C. (1973). The interpretation of cultures: Selected essays. New York: Basic Books. Goldstein, D. M. (2014). Qualitative research in dangerous places: Becoming an ‘ethnographer’ of violence and personal safety. DSD Working Papers on Research Security. 1(1), 1–18. Goodhand, J. (2000). Research in conflict zones: Ethics and ac- countability. Forced Migration Review, 8(1), 12–15. Gudeman, S., & Rivera, A. (1990). Conversations in Colombia: The domestic economy in life and text. Cambridge University Press. H¨oglund, K. (2011). Comparative field research in warn-torn soci- eties. In K. H¨oglund, & M. ¨Oberg (Eds), Understanding peace research: Methods and challenges. Routledge. IDS (2020). ‘Mediating between the state and its poor and marginalized during covid-19’. Institute of Development Studies. www.ids.ac. uk/opinions/mediating-between-the-state-and-its-poor-and- marginalised-during-covid-19/ (accessed 5 May 2022). Jacobsen, K., & Landau, L. (2003). The dual imperative in refugee research: Some methodological and ethical considerations in social science research on forced migration. Disasters, 27(3), 185–206. https://doi.org/10.1111/1467-7717.00228 Khan Mohmand, S., Justino, P., Dowd, C., Scott-Villiers, P., Befani, B., Loureiro, M., & Shaw, J. (2017). Innovative methods for research on social and political action in fragile and conflict-affected settings. IDS working paper 487. Brighton: Institute of Development Studies. King, G., Murray, C., Salomon, J., & Tandon, A. (2004). Enhancing the validity and cross-cultural comparability of measurement in survey research. American Political Science Review, 98(1), 191–207. https://doi.org/10.1017/s000305540400108x Kovats-Bernat, J. C. (2002). Negotiating dangerous fields: Pragmatic strategies for fieldwork amid violence and terror. American Anthro- pologist, 104(1), 208–222. https://doi.org/10.1525/aa.2002.104.1.208 Loureiro, M. (2013). Migrant’s love(s): The changing nature of re- lationships in post-earthquake Pakistan-administered kashmir. Asian Journal of Women’s Studies, 19(2), 36–65. https://doi.org/ 10.1080/12259276.2013.11666148 Lyberg, L., & Harknes, J., Mohler, P., & Stange, M., & Jpec, L. (2014). A review of quality issues associated with studying hard-to-survey populations. In R. Tourangeau, B. Edwards, T. P. Johnson, K. M. Wolter, & N. Bates (Eds), Hard-to-Survey populations. : Cambridge University Press. 12 International Journal of Qualitative Methods Masset, E. (2015). Measuring empowerment in rural India using vignettes. Journal of Development Effectiveness, 7(3), 346–356. https://doi.org/10.1080/19439342.2015.1068355 Mazurana, D., & Gale, L. (2013). Preparing for research in active conflict zones: Practical considerations for personal safety. In D. Mazurana, K. Jacobsen, & L. Gale (Eds), Research methods in conflict settings: A view from below (pp. 277–292). New York: Cambridge University Press. Mazurana, D., Jacobsen, K., & Gale, L. (Eds), (2013). Research methods in conflict settings: A view from below. New York: Cambridge University Press. Nordstrom, C., & Robben, A. (Eds), (1995). Fieldwork under fire: Contemporary studies of violence and survival. University of California Press. Schedler, A. (1999). Conceptualizing accountability. In A. Schedler, L. Diamond, & M. Plattner (Eds), The self-restraining state: Power and accountability in new democracies. Lynne Rienner. Smith, T. (2014). Hard-to-Survey populations in comparative per- spective. In R. Tourangeau, B. Edwards, T. Johnson, K. Wolter, & N. Bates (Eds), Hard-to-Survey populations. Cambridge University Press. Smyth, M., & Robinson, G. (Eds), (2001). Researching violently issues. Pluto divided societies: Ethical and methodological Press. Stewart, F., & Brown, G. (2009). ‘Fragile states’. CRISE working paper 51. Centre for Research on Inequality, Human Security and Ethnicity. Stys, P., Kirk, T., Muhindo, S., Balume, B., Mazuri, P., Tchumisi, I., N’simire, S., & Green, D. (2021). ‘An experimental method- ology forstudying household financial governance and coping mechanisms in Goma, DRC’. In Working Paper. London School of Economics, 1/21. Unsworth, S. (2010). An upside down view of governance. Centre for the Future State, Institute of Development Studies.
10.1111_disa.12560
doi:10.1111/disa.12560 The policy landscape and challenges of disaster risk financing: navigating risk and uncertainty Olivia G. Taylor Lecturer, Department of Geography, University of Sussex, United Kingdom A more anticipatory, pre-agreed response is a shared goal of many in the disaster manage- ment and humanitarian communities. This paper considers the emerging policy landscape of disaster risk financing (DRF), which is taken here to include mechanisms that allow agencies to act in advance of disasters occurring, as well as those that aim to respond earlier to dis- asters which have already happened. What they both have in common is no longer waiting for needs to become apparent before responding; however, this creates a challenge for practi- tioners because of the potential for acting erroneously. This paper provides a more cohesive way of understanding approaches in this policy area through the shared challenge of decision- making under the condition of uncertainty. Drawing on expert interviews and science and technology studies theory, it sets out some recommendations on how practitioners can nav- igate risk and uncertainty better within DRF and in a more nuanced way. Keywords: anticipation, anticipatory action, disaster risk finance, decision-making, risk, uncertainty Introduction Historically, disaster risk reduction and preparedness have made up a small proportion of overall expenditure on disaster response (Kellett and Caravani, 2013). In recent years, however, movement towards a more anticipatory, pre-agreed approach, which I refer to broadly in this paper as disaster risk financing (DRF), has become a key goal of many in the disaster response and humanitarian sectors. A watershed year in this respect was 2021, which saw a series of key events and signifi- cant new commitments to fund risk financing. During the G7 (Group of Seven) meeting hosted in the United Kingdom in June 2021, the Governments of Germany and the UK respectively committed GBP 120 million and GBP 125 million of new financing (approx- imately USD 160 million and USD 140 million) for pre-arranged disaster risk financing for vulnerable communities (UK Presidency of the G7, 2021). In September 2021, the United Nations Office for the Coordination of Humanitarian Affairs (UN OCHA) convened a high-level event focused on anticipatory action, in partnership with the Governments of Germany and the UK, at which a number of countries and agencies further boosted their commitments to these approaches (United Nations and the Governments of Germany and the United Kingdom, 2021). For example, Germany announced its ambition to double Disasters, 2023, 47(3): 745–765. © 2022 The Authors Disasters © 2022 ODI This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. 746 its contribution to anticipatory action by 2022, while the Government of Ireland com- mitted to directing approximately 25 per cent of its humanitarian funding straight to mechanisms that support anticipatory action (United Nations and the Governments of Germany and the United Kingdom, 2021, p. 2). Sentiment regarding the current policy shift towards more anticipatory disaster financ- ing was summed up by the current Under-Secretary-General for Humanitarian Affairs and Emergency Relief Coordinator, Martin Griffiths. Speaking at the High-Level Human- itarian Event on Anticipatory Action on 9 September 2021, he stated that: ‘The humani- tarian system must be as anticipatory as possible, and only as reactive as necessary’ (Griffiths, 2021, p. 2). The significant increase in the momentum behind anticipatory and risk financing approaches has resulted in an emergent and rapidly evolving policy area. One notice- able characteristic is the complex terminology for particular mechanisms. The two most common phrases used are ‘anticipatory action’, which usually refers to anticipatory financ- ing mechanisms implemented by humanitarian agencies, and ‘disaster risk financing’, which usually refers to mechanisms used by development financing institutions to provide rapid financing in the aftermath of a disaster. While there are differences between the mechanisms in this area, they also have a great deal in common, specifically relating to the challenge of decision-making when agencies are no longer waiting for disaster needs to become apparent before responding. This is critical, because it opens up decision-making to interpretation and raises important questions regarding how decisions are made, as well as the possibility of making decisions in error. Underlying this is the challenge of navigating risk and uncertainty in decision- making, which is the central focus of this paper. For reasons that I explain further in the second section, I adopt a broad definition of mechanisms that fall within the scope of DRF. These are based on: (i) information about or measures of disaster risk; (ii) pre-arranged finance and plans; and (iii) an instrument to enact a response. This definition is purposefully broad; it does not refer, for instance, to ‘anticipatory’ use of forecasts, but rather to information and measures of risk, whether that is a forecast, an expert advisory alert, or a proxy measurement of a hazard (such as windspeed) in order to inform and trigger a response. This allows diverse policy mech- anisms ranging from index-based insurance to forecast-based financing (FbF)1 to be understood as different tools within the same policy landscape. This paper explores the policy landscape of DRF and provides a more cohesive way of thinking about the different mechanisms in use, spanning those that permit agencies to act in advance of disasters occurring, as well as those that aim to respond earlier to disasters that have already happened. I unpack the central policy narratives supporting DRF, relating to efficiency and effectiveness. These demonstrate the tensions across the sector and differences in how agencies approach and define the central arguments. Next, I discuss some of the challenges to implementing DRF in practice. As a result of acting based on information which is inherently incomplete, acting in the face of uncertainty is a critical challenge confronting practitioners in this field. I bring to bear literature from science and technology studies (STS) to outline some of the ways in which both risk and Olivia G. Taylor 747 uncertainty could be better understood and outline some recommendations for practi- tioners in this field. Methodology This paper draws on qualitative research, including 27 expert policy interviews, combined with participant observation during key events in the DRF community and desk-based reviews of policy literature. Interviews were semi-structured in nature and organised through a mixture of purposive and snowball sampling to ensure that key agencies were represented. It is important to note that participants are categorised in a broad way as: humanitarian practitioner; donor; DRF specialist; catastrophe modeller; and researcher. This is necessary for two reasons. First, DRF is a small sector, so broad categories are important to maintain participant anonymity and to ensure that interviews are non- identifiable. Second, this sector is composed of new collaborations of expertise, spanning policy, climate science, finance, and humanitarian practice. Table 1 lists the organisations represented in empirical data by broad descriptive category and exemplar job titles from among the categories of interview participants. Interviews were supplemented with attendance at key events relevant to the sector between 2018 and 2021. These included: • the Red Cross’ Global Dialogue Platform on Forecast-based Financing, and later the Global Dialogue Platform on Anticipatory Humanitarian Action,2 in Berlin, Germany, in September 2018 and November 2019; • the Global Facility for Disaster Reduction and Recovery (GFDRR)’s Understanding Risk Conference in Mexico City, Mexico, in May 2018; • the United Nations’ (UN) Global Platform for Disaster Risk Reduction in Geneva, Switzerland, in May 2019; and • the Conference of the Parties to the United Nations Framework Convention on Climate Change (UNFCCC) in Madrid, Spain, in December 2019. I participated in two further multi-day virtual conferences during 2020 and 2021: the Red Cross’ Virtual Global Dialogue Platform for Anticipatory Humanitarian Action in December 2020 and 2021; and the Insurance Development Forum’s Virtual Summit in June 2021. Attending conferences and public sessions allowed for the triangulation of key narratives, especially conferences that spanned the climate, disaster risk reduction, and humanitarian communities. It also enabled me to follow the emergence of new approaches, vocabularies, and particular mechanisms within the DRF space. Disaster risk financing: background and policy landscape DRF is emerging at a time when global humanitarian needs are reaching their highest level in decades (UN OCHA, 2020), a situation exacerbated by the COVID-19 pandemic. The total value of unmet humanitarian appeals has been increasing over the past several The policy landscape and challenges of disaster risk financing: navigating risk and uncertainty 748 Table 1. Examples of organisations and job titles of research participants included in the sample Interview number Category Organisations included in the category Job title examples within the category 1 2 3 5 9 11 17 21 23 26 10 12 13 20 4 6 16 18 7 8 14 15 22 19 24 25 27 Humanitarian practitioner Humanitarian practitioner Humanitarian practitioner Humanitarian practitioner Food and Agriculture Organization of the United Nations (FAO) Crisis Anticipation Adviser Global Coordinator FbF Senior Officer International Federation of Red Cross and Red Crescent Societies (IFRC) Red Cross Red Crescent Climate Centre Humanitarian practitioner START Network Humanitarian practitioner World Food Programme (WFP) Humanitarian practitioner Humanitarian practitioner Humanitarian practitioner Humanitarian practitioner Catastrophe modeller Private consultants Financial Sector Specialist Oasis Loss Modelling Framework World Bank Foreign, Commonwealth and Development Office (FCDO) German Federal Foreign Office United Nations Office for the Coordination of Humanitarian Affairs (UN OCHA) Humanitarian Affairs Officer Senior Desk Officer Centre for Disaster Protection Consultant START Network World Bank Technical Lead on Crisis Anticipa- tion and Risk Financing German Red Cross Adviser for Policy and Advocacy IFRC Manager University of Reading Catastrophe modeller Catastrophe modeller Catastrophe modeller Donor Donor Donor Donor DRF expert DRF expert DRF expert DRF expert DRF expert Researcher Researcher Researcher Researcher Note: the organisations and job titles are grouped in this way to ensure that individual interviews are not identifiable. Source: author. Olivia G. Taylor 749 years, from USD 8.9 billion in 2016 to USD 13.1 billion in 2020, excluding the total value of COVID-19 relevant appeals in 2020, of which a further USD 5.7 billion was unmet (Development Initiatives, 2021, p. 33). This is set against the backdrop of a long-term trend of rising global humanitarian funding over the past decade, yet the percentage of humani- tarian appeal requirements that are met by funding has declined from 63 per cent in 2011 to 52 per cent in 2020 (Development Initiatives, 2021, p. 33). Even prior to the COVID-19 pandemic, key actors were making the case that more anticipatory financing was the only way to resolve the ongoing problem of humanitar- ian needs outstripping financing. For instance, Mark Lowcock, who served as the UN Under-Secretary-General for Humanitarian Affairs and Emergency Relief Coordinator from March 2017–June 2021, argued: We are now seeking almost US$27 billion for 2019, for the appeals from the UN, NGOs and others that I coordinate. We have raised almost $16 billion so far. That’s a record . . . But it leaves a large gap. It would be nice to think we can fill the gap just by raising more money. But we can’t. We also have to make the money we have go further. The best way to do that is to change our current system from one that reacts, to one that anticipates (Lowcock, 2019, pp. 1–2). Most recently, the confluence of COVID-19 and existing drivers of humanitarian crises, such as conflict and climate change, has served to underline calls for more anticipatory financing, as well as increasing the coherence between development and humanitarian assistance. While these recent pressures have drawn more attention to the need to change the status quo of disaster response, there is a longer trend that has catalysed calls for altering the paradigm of disaster response. Among participants interviewed for this research, the Horn of Africa crisis of 2011–12 was seen as a key example of the perceived failure to respond to disasters in a timely way, even when credible warning information had been available (Bailey, 2012). The slow response to this event was seen as a systemic failure across the humanitarian sector. One of the recommendations to follow was that agencies should do more to act despite uncertainty, no longer ‘waiting for certainty before responding’ (Hillbruner and Moloney, 2012, p. 1). Wider lessons learned included the need to ensure that scientific information is used better in decision-making, and that decision-makers act on it (Humanitarian Emergency Response Review, 2011). One research participant referred to the period of time after this crisis as representing a ‘step change’ (Interview 8, DRF specialist), which led to a focus on a more anticipatory approach and arguably laid the groundwork for DRF. Indeed, some of the findings resulted in the use of triggers for action, one of the key characteristics of DRF. As one participant explained: ‘What we have seen in the history . . . of humanitarian actions is there’s a lot of early warning systems that have absolutely no consequence because there is no obligation to take an action based on a warning. So, what we’re trying to do is to force that’ (Interview 18, donor). To deliver this, there has been a significant enlargement of the policy landscape of DRF initiatives. In 2017, the InsuResilience Global Partnership for Climate and Disaster Risk The policy landscape and challenges of disaster risk financing: navigating risk and uncertainty 750 Finance and Insurance Solutions was launched by the G7 to provide climate risk insur- ance for 400 million people in developing countries by 2020 (InsuResilience, 2018). In the same year, the Government of the UK launched the Centre for Disaster Protection, to provide a technical advisory for DRF, although this was not fully operational until 2019. The Centre’s work spans sovereign disaster financing mechanisms to working with humanitarian agencies on risk financing (DFID, 2017),3 which highlights some of the increasing interconnections between sovereign and humanitarian disaster financing through DRF. In 2018, the Global Risk Financing Facility was launched in partnership with the InsuResilience Global Partnership, to pilot further disaster risk financing tools, implemented by the World Bank and GFDRR (2018). Defining DRF: complex terminology This policy area has rapidly diversified and is now a complex array of agencies, mecha- nisms, and projects often with their own methodologies and vocabularies, leading to a complicated and sometimes confusing debate around terminology. The two most common definitions used in this wider sector are ‘anticipatory action’ and ‘disaster risk financing’. In this subsection I discuss some of the history of the different terminologies and explain why the term DRF is adopted in this paper. ‘Disaster risk financing’ as a term was initially used by the World Bank and can be traced back to the programme name for a World Bank and GFDRR stream of work on sovereign insurance, market development, and partnerships with the private sector, titled the ‘Disaster Risk Financing and Insurance Programme’ (DRFIP), which was launched in 2011. Subsequently, in their influential book Dull Disasters?, Daniel Clarke and Stefan Dercon (2016), who had both been affiliated with the DRFIP, argued for a more rules- based approach to financing disaster response, by combining what they defined as: ‘A coordinated plan for post-disaster action agreed in advance; A fast, evidence-based decision-making process; and Financing on standby to ensure that the plan can be implemented’ (Clarke and Dercon, 2016, p. 3). In so doing, they provided one of the first overarching definitions that could be used to describe emerging mechanisms across the sector. With regard to definitions, it is notable that many practitioners who work in this sector use the terms DRF and anticipatory action interchangeably. This was the perspective of one humanitarian research participant who argued that FbF should be seen as one tool within a broader landscape of DRF, on the condition that wider considerations of disaster risk management are recognised: ‘The current definition of DRF from the World Bank only focusses on the response element—a bunch of instruments to ensure liquidity for response. But this is changing, it’s really . . . looking towards holistic perspectives on disaster risk management . . . that’s what I hope DRF will become in the future, and in that principle, in that definition, I will say that FbF is a tool within DRF’ (Interview 2, humanitarian practitioner). Many others interviewed for this research insisted that the difference was really a semantic one. For example, one participant argued: ‘if you talk to a government or you Olivia G. Taylor 751 talk to people at risk, they don’t give two craps about what you call it. They care about what you’re trying to do for them’ (Interview 8, DRF specialist). Another contended: ‘we’re all for having an open definition of it [referring to anticipation]. . . . I think it’s good for all of us in the sector to have something loose’ (Interview 1, humanitarian practitioner). Nevertheless, there has been an ongoing debate about terminology in the sector, and significant resources have been invested in trying to find consensus. For example, a number of agencies, including the Centre for Disaster Protection, the Red Cross Red Crescent Climate Centre, and UN OCHA, commissioned a joint ‘thesaurus’ of anticipa- tory action to: ‘enable reflection on the similarities and differences in the way organizations use language associated with the concept of anticipatory humanitarian action’ and to enable mutual understanding (De Wit, 2019, p. 5). In September 2021, the newly formed Anticipation Hub4 hosted an event titled ‘Linking Anticipatory Action to Risk Financ- ing’, in order to assess the connections between anticipatory action and risk financing (InsuResilience Global Partnership and the Anticipation Hub, 2021). Despite conclud- ing that the sector needed ‘to stop silo-approaches across the disaster management and crisis response spectrum’ (InsuResilience Global Partnership and the Anticipation Hub, 2021, p. 1), the event was clearly premised on there being a clear distinction between antici- patory action and risk financing. As noted in the introduction, the main distinctions that are usually drawn between the two approaches are: • the temporality of the mechanism, whether it is anticipatory or ‘ex-ante’; and • by what type of agency it is implemented/funded, and whether that is an actor with a humanitarian mandate. However, I argue that this is not the most useful way to think about the mechanisms in this sector. On the first point, while a ‘more anticipatory’ (or less late) response is a key policy objective for the sector, it is not a clear defining characteristic because temporal distinctions are often difficult to apply in practice to disasters. For instance, it has long been pointed out that the phases of mitigation and preparedness ‘pre disaster’, and response and recovery ‘post disaster’, are rarely as neatly defined in practice as they seem in the dis- aster management cycle depicted in the disaster studies literature (Neal, 1997; Contreras, 2016). Moreover, it is difficult to ascertain the onset of impacts of many slow-onset haz- ards such as droughts, which feature prominently among the risks responded to through such mechanisms, which blurs the boundary between early action and early response (Wilkinson et al., 2018). Thus, while anticipatory response is a key policy objective, this research cautions against relying on ‘anticipation’ as a defining characteristic, even though the goal of a ‘less late’ response is clear and important. In relation to the second point, part of the desire to distinguish between humanitarian and development financing in this sector is linked to understandable concerns about the potential loss of humanitarian impartiality. This sentiment was summed up by a senior policymaker at the German Federal Foreign Office, during a panel session at the 2018 The policy landscape and challenges of disaster risk financing: navigating risk and uncertainty 752 Global Dialogue Platform conference. He recommended that different approaches of risk financing ‘be kept separate so that all approaches are not mixed up. . . . Ultimately, humanitarian financing is obligated to human needs and not political considerations’ (German Red Cross, 2018, p. 19). However, the hybridisation of mechanisms that has occurred in the years since suggests that it is now increasingly difficult to distinguish neatly between ‘humanitarian’ and ‘development’ finance. This is exemplified by the poten- tial use of the World Bank’s Crisis Response Window to fund humanitarian response through UN OCHA’s Anticipatory Action Frameworks, a partnership that was developed through the first UN OCHA Anticipatory Action pilot in Somalia in 2020 (Getliffe, 2021). Other examples of ‘hybrid’ mechanisms have also emerged, such as partnerships between insurance and humanitarian response through the START Network’s (2020) ARC Replica insurance policy, and the newly launched START Ready framework (START Network, 2021). Consequently, the type of funding or implementation agency is no longer a helpful way to distinguish between anticipatory action and disaster risk financing. As such, the definition I adopt here is a broad approach to DRF entailing: (i) informa- tion or measures of disaster risk; (ii) pre-arranged finance and plans; and (iii) a mechanism to enact a response. Understanding DRF policy objectives and narratives: efficiency and effectiveness The notion of a quicker response to a disaster, anticipating rather than reacting, is strongly intuitive. As the saying goes, ‘prevention is better than cure’, a theme oft-repeated by the research participants. Here, I discuss the main policy narratives in DRF: that it leads to both a more efficient and effective disaster response and is therefore the way to ‘square the circle’ of increasing humanitarian response costs. I explore these narratives and how they are understood across the sector to reveal some of the underlying contestation relat- ing to the objectives of DRF, as well as practical challenges to implementation. First, the financing gap between growing humanitarian need and available financing is frequently cited in policy and advocacy materials as the central driver of a more effi- cient and effective response via DRF. It was a key point in Mark Lowcock’s speech, referred to earlier, that we can no longer simply raise more money to meet humanitarian needs: ‘We . . . have to make the money we have go further. The best way to do that is to change our current system from one that reacts, to one that anticipates’ (Lowcock, 2019, p. 2). Certainly, DRF is an evolution from the status quo of ‘ex-post’ disaster response, which has been likened to the passing of a ‘begging bowl’ around donors to raise funds after a disaster happens (Clarke and Dercon, 2016). This contributes to a fragmented and politi- cised response that is poorly matched with post-disaster needs, which are often contingent on funding cycles in donor countries, with little relevance to needs on the ground (Talbot and Barder, 2016). Moreover, an earlier response can avert harmful coping strategies and protect livelihoods contributing to long-term development gains (Wilkinson et al., 2018). Olivia G. Taylor 753 However, the link between disasters and humanitarian financing needs is, in practice, more complex. As Swithern (2018) has written, there are many reasons why disaster impacts do not correlate directly with the scale of humanitarian funding appeals. In addition, a recent meta review by the World Meteorological Organization (2021) con- cluded that while weather-related disasters have increased over the past 50 years, they have caused more damage but fewer deaths, mostly as a result of improved forecasting and disaster risk reduction activities. The relative importance of the rationales for a more effective and efficient response is also complex and varies between actors in the DRF sector. For example, there has been significant investment in ‘cost–benefit’ research into various aspects of resilience pro- gramming, early action, and preparedness by several bilateral donors, in particular the UK and the United States. For instance, Cabot Venton et al. (2013) were commissioned to conduct a report into The Economics of Early Response and Disaster Resilience for the UK’s Department for International Development (DFID), and later a similar report, Economics of Resilience to Drought: Kenya Analysis, was commissioned by the United States Agency for International Development (Cabot Venton, 2017). And a Return on Investment for Emergency Preparedness Study was produced for the United Nations Chil- dren’s Programme and the World Food Programme by the Boston Consulting Group and funded by DFID in 2015 (UNICEF and WFP, 2015). However, some agencies other- wise very active in the policy space are notable because of their absence in funding such reports, especially the German Federal Foreign Office, which was an early funder of the Red Cross’ FbF work (German Red Cross, 2015). The robustness of such ‘cost–benefit’ evidence has been scrutinised more carefully as the sector has evolved. A recent policy paper that reviewed the evidence base for anticipa- tory action raises the point that easily reproducible and catchy numbers produced by return on investment and cost–benefit studies ‘can obscure the quality of and underlying assumptions behind these numbers’ (Weingärtner, Pforr, and Wilkinson, 2020, p. 34). Nonetheless, the findings of such reviews were cited to this author by the research par- ticipants, even if they were sceptical about them, which demonstrates how much these studies have cut through. For example, one stated: ‘I’m a bit sceptical about . . . the num- bers like the data that says you can act . . . what is it five or six times you say before it’s worse than a late response’ (Interview 6, donor). This interviewee was referring to the key statistic in the Cabot-Venton et al. (2013, p. 1) review: ‘for every early response to a correctly forecast crisis, early responses could be made 2–6 times to crises that do not materialise, before the cost of a single late response is met’. Efficiency was also a subject of debate among the study participants. Overall, humani- tarian practitioners were cautious, with one asserting that ‘the interesting thing about aid money is we want to give it away’ (Interview 4, humanitarian practitioner). Another interviewee who worked at the interface between different specialisms in DRF outlined the differences in view between humanitarian practitioners and others in the sector: ‘If we talk to humanitarian actors . . . in my experience some of them get the bang for the buck argument. . . . They get it but they’re like “no, that’s not what we’re here for, we’re The policy landscape and challenges of disaster risk financing: navigating risk and uncertainty 754 here to help people”. So, you have to frame it as you could help more people with the same . . . amount of money’ (Interview 15, DRF expert). While the notion that acting earlier can make responses more effective makes sense in principle, it is also notable that this has been harder to evidence across different mech- anisms and hazards. In particular, the usefulness of actions that can be employed in the window of opportunity between the warning of a hazard and the impacts of the ensuing disaster being felt has been questioned. A working group titled ‘Early Actions: Why do we always end up with chlorine tablets?’ discussed this issue during the sixth Global Dia- logue Platform on FbF (German Red Cross, 2018). Chlorine tablets are regularly distrib- uted prior to a flood or cyclone hazard and are, of course, indispensable for preventing water-borne disease. Participants in the session pointed out, however, that they are dis- seminated as part of agreed protocols because they are small and easy to preposition within the time available, but as a result, they are used in preference to other actions that would be more aligned with long-term risk reduction activities (German Red Cross, 2018). Of course, the potential effectiveness of early actions varies significantly between haz- ards, something that practitioners acknowledge. Commenting on this, one humanitari- an practitioner noted that while timely action will reduce human suffering, ‘it doesn’t mean that the disaster will be totally prevented. Of course, it will really depend on the hazard . . . like for drought I’m more inclined to say that we have enough lead time. . . . But for a cyclone . . . I mean Idai5, we could have had the most amazing FbF in place but still the houses will be totally destroyed’ (Interview 2, humanitarian practitioner). Thus, the use of arguments pertaining to efficiency and effectiveness of DRF varies between different mechanisms and hazard contexts, and they have evolved over time. There is an underlying recognition that acting in advance is no ‘silver bullet’ for signifi- cant efficiency savings, and does not overcome the challenge of mitigating the impacts of major hazards, while the relationships between disaster events, humanitarian financ- ing needs, and the response are complex. However, the mutually reinforcing narrative of DRF approaches being both more efficient and effective is highly intuitive and very powerful, especially in the wider context of pressures on the humanitarian financing system. As one participant put it: ‘there is a clear understanding that disaster risk financ- ing instruments are super essential in the future. It is clear we are going to have more disasters, and the money that is located at this moment for humanitarian action is not going to be enough for the type of events that we will have in 10, 20 years’ (Interview 23, humanitarian practitioner). The challenges of decision-making in DRF Acting based on information rather than existing needs is the key to taking a more antici- patory approach in disaster response, yet it also presents significant challenges in terms of decision-making. As De Wit (2019, p. 34) argues in her discussion of the language used in anticipatory approaches in this sector: ‘questions around temporality have moral impli- cations for finding a common understanding of when decisions are taken and actions Olivia G. Taylor 755 planned, how you justify those choices, and how they can be funded’. This challenge was also articulated by participants in this research, such as one who explained that: ‘early action is . . . to a certain extent open to interpretation. . . . This is why it’s fundamental to work on coherence and common approaches because that way we govern this, we manage this uncertainty, we manage the questions around the evidence, and we render it credible’ (Interview 5, humanitarian practitioner). In this section I look at how questions concerning decision-making are navigated. It is understandable that the need to justify actions taken based on DRF information is a primary concern, and this is reflected throughout the defining pillars of DRF mecha- nisms. Each of the components of the definition of DRF, despite variations in terminol- ogy, can be understood as contributing to a robust process for decision-makers to use to take action. For example, the aspect which requires that financing and plans be pre- arranged is described in one policy document as creating ‘certainty about what finance will be available’ (Montier, Harris, and Ranger, 2019, p. 4), giving disaster managers and decision-makers the confidence to act. This was further expressed during a keynote session at the 2018 Global Dialogue Platform conference, where certainty of finance was portrayed as ‘the “glue” to assure that early action is taken ahead of a disaster based on a scientific decision-making process’ (German Red Cross, 2018, p. 19). Moreover, the third pillar of DRF, a mechanism to trigger a response, is intended to overcome any potential inertia created by uncertainty. For instance, one participant explained its pur- pose to this author as follows: ‘the function of triggers is not to tell you what to do, but when to act . . . you’re changing the default from hesitating and wondering to taking action’ (Interview 18, donor). However, the variation in how the pillars of DRF are defined reveals the lack of con- sensus on the specifics of what creates a robust decision-making process. As outlined above, I adopt a definition of DRF as requiring ‘information or measures of disaster risk’. This was chosen as a broad and encompassing criterion; to choose a more specific defini- tion might have excluded some mechanisms. However, there are significant differences in how this component of DRF is defined across the policy literature. By way of example, Clarke and Dercon (2016, p. 3) adopt a loose definition in Dull Disasters?, referring to: ‘A fast, evidence-based decision making process’. Others place more emphasis on warn- ing information that provides a quantifiable output, such as a policy document written by practitioners from the START Network, whose definition of DRF underlines ‘quantify- ing risks in advance’ (Montier, Harris, and Ranger, 2019). These differentiations point to questions about what makes information sufficiently credible to use in DRF mechanisms. This is especially important when making compari- sons across different types of hazards within the remit of DRF mechanisms, which range from volcanoes to cyclones and droughts, and which require insights from diverse phys- ical sciences—not to mention other types of humanitarian crises covered by some mech- anisms, such as conflicts or migration flows. There are no clear answers to these questions, and the different definitions and methodologies adopted across the sector show that in many cases, each implementing agency is finding its own way of managing them. The policy landscape and challenges of disaster risk financing: navigating risk and uncertainty 756 From ‘acting on uncertainty’ to ‘acting based on risk’: how risk and uncertainty are understood in DRF policy narratives An irony in DRF is that one of the original objectives of this policy shift was to over- come the inability or unwillingness of agencies to act in the face of uncertainty, which is perceived as having been a principal aspect of the failure to respond in a timely way to past emergencies. This was one of the conclusions drawn from the Horn of Africa crisis in 2011–12, which as discussed earlier, was a significant turning point in encour- aging more anticipatory, pre-agreed approaches. In an influential review of that event, Hillier and Dempsey (2012, p. 15) contend that: ‘Early response requires acting on uncer- tainty’. In this policy paper, the authors discuss quantifying uncertainty and adopting a risk management approach in humanitarian response decisions, but their arguments do not shy away from acknowledging uncertainty in decision-making based on such information. They state: ‘Forecasts involve uncertainty: they are inevitably based on data which is not totally comprehensive and are tinged with judgement; the earlier the warn- ing, the less accurate it is likely to be’ (Hillier and Dempsey, 2021, p. 15). However, in the years since, it is important to note how the DRF policy space has taken shape, in many cases moving towards a policy language that focuses heavily on risk at the expense of uncertainty. Indeed, this is codified in part in the term disaster risk financing, and it is no surprise, therefore, that the idea of acting ‘based on risk’ has become one of the defining characteristics of the paradigm shift. This is even consistent in work that spotlights humanitarian mechanisms and does not use the specific terminology of DRF, such as De Wit’s (2019, p. 6) Thesaurus, in which she summarises the paradigm shift of anticipation as ‘acting based on risk’. The notion of ‘acting based on risk’ or a ‘risk-based’ approach is reflected widely in the policy literature, such as in policy documents of the START Network. Quoting one of its donors: ‘We are trying to shift to a risk-based approach instead of needs based, with more preparedness and early action’ (START Network, 2019, p. 8). And this notion is also evident in the language used in policy materials of UN OCHA in the statement released after the High-Level Humanitarian Event on Antici- patory Action in 2021: ‘The humanitarian system must shift away from a solely reactive response to crises towards an increasingly proactive, anticipatory approach – acting on risks instead of only reacting to needs’ (United Nations and the Governments of Germany and the UK, 2021, p. 2). Despite the sense of there being a clear distinction between acting on uncertainty and acting ‘based on risk’, one of the key findings of this research is that when asked how they thought about risk and uncertainty in their work, practitioners had very diverse understandings and attitudes. A key difference among them related to whether or not they viewed quantifiable uncertainty as a form of risk, or as a valid form of uncertainty. The former view was more commonly held by economists and social scientists. For instance, one research participant who had worked at an economic research institution said that: ‘uncertainty would be to describe the fact that you didn’t know what your probability risks are’ (Interview 15, DRF expert). This reflects a ‘Knightian’ definition of risk and uncertainty—named after the economist Frank Knight—whereby risk is associated Olivia G. Taylor 757 with quantifiable uncertainty and uncertainty refers to anything that cannot be numeri- cally quantified. In contrast, modellers and physical scientists often considered ‘quantifiable uncer- tainty’ to be a valid and important type of uncertainty. Indeed, this is what probabilistic modelling is designed to communicate: stating and quantifying predictive uncertainty (Gneiting, 2008). For example, one participant with a technical background explained: ‘there is [sic] two levels. . . . One is the uncertainty you absolutely cannot quantify because mathematically you just can’t do it. . . . So, there is that box of stuff we can’t quantify. And then there is stuff we can quantify because we actually do have some data and you can use mathematical approaches to quantify uncertainty around that data’ (Interview 13, catastrophe modeller). Another case was highlighted by a participant with a forecast- ing background, who described a recent situation within the FbF community where two cyclone models contradicted each other. They argued that a full understanding of uncer- tainty requires taking into account uncertainty that lies beyond the scope of a forecast model. In their words: You’ve got the uncertainties that you can quantify, a sort of stochastic one, so you can say like a 50 per cent chance of a flood . . . but you know that there’s the uncertainty that you can’t quantify or characterise . . . that the ensemble is not representing . . . you would have an ensemble forecast of tropical cyclones and you’ve got an ECMWF 6 ensemble that says one thing and a Met Office7 ensemble that says another thing, and if they were characterising uncertainty well, then the ensembles’ spread would be overlapping in both of them. But if they both say separate things, then what do you do? Because there’s uncertainty that goes beyond what that ensemble is representing (Interview 27, researcher). Highlighting these differences is not intended to diagnose a lack of understanding per se, but rather to show the complexity of these concepts. Risk, for example, is not a singular, objective metric ‘out there’ that can be measured in a uniform way across differ- ent settings or hazards. As I explore further below, risk and uncertainty are complicated, determined by degrees and forms of knowledge, and specific to particular hazards and contexts, hence they are difficult to convey across disciplinary boundaries. Lastly, this research also identified a sense among practitioners that language pertain- ing to uncertainty was sometimes unwelcome in their work on DRF, particularly among those working in government donor agencies. For instance, one participant stated that in their view, ‘risk is seen in a very clear way in government in particular . . . you have risk registers etc. There’s a very formalised . . . “how you deal with risk” manual, that we all have to comply with. . . . But uncertainty is seen as “I don’t know the answer” and that tends to paralyse people . . . even the word still seems to scare people. So, it’s actually better to talk about managing risk, you know, uncertainty being a risk’ (Interview 15, DRF expert). The view that uncertainty was difficult to talk about in a public sector context was sup- ported by another research participant. From the perspective of financial services, they pointed out that: ‘the concept of dealing with uncertainty is pretty well ingrained in the The policy landscape and challenges of disaster risk financing: navigating risk and uncertainty 758 finance sector in a way that I think it is not ingrained in the public sector. . . . [It] is very challenging for the public sector to get its head around’ (Interview 13, catastrophe modeller). Taken together, these factors result in a policy sector that often focuses on risk at the expense of uncertainty. In some cases, this leads to eliminating uncertainty from the policy discourse. This is evident in Mark Lowcock’s two important speeches delivered in 2018 and 2019 on the subject of anticipatory humanitarian financing.8 In the two speeches combined, the word ‘risk’ is used a total of 36 times, but the word ‘uncertainty’ is not used at all (Lowcock, 2018, 2019). In other cases, such as during the Global Dialogue Platform conference in 2018, session convenors of one side event asserted that the use of data and triggers ‘help us eliminate uncertainty about when and how to act’ (German Red Cross, 2019, p. 22). What do we know? What can we predict? What can we foresee?: navigating risk and uncertainty in DRF As demonstrated, DRF uses information about potential future hazards and disasters to take action, instead of waiting for such events to happen and responding in the aftermath. As a result, acting in the face of uncertainty is a critical challenge for practitioners, as acknowledged by Hillier and Dempsey (2012, p. 15), who argue that while reducing uncer- tainty is important, ‘[e]arly response requires acting on uncertainty’. This sentiment was also echoed by the Head of Country Programmes at the Centre for Disaster Protection, when speaking at a public webinar.9 She suggested that DRF fundamentally requires reflec- tion about: ‘What do we know? What can we predict? What can we foresee?’. However, practitioners interviewed in this research described many difficulties in talking about uncertainty in DRF, and as noted above, there is a tendency in formal policy spaces to focus on risk at the expense of uncertainty. In this section I draw insights from science and technology studies to make some sug- gestions as to how both risk and uncertainty can be better understood in relation to DRF. STS is a diverse discipline that seeks to comprehend better the role of knowledge, spe- cifically scientific knowledge, in policy and society. STS has explored in particular how knowledge is intimately tied up with risk and uncertainty, determining the boundaries of what we know and what we do not know (Stirling, 2007, 2009, 2010), while also highlight- ing how knowledge shapes our perceptions of risk and uncertainty (Lash, Szerszynski, and Wynne, 1998; Wynne, 1998). Both of these points are particularly relevant to develop- ing a better understanding of risk and uncertainty as it pertains to DRF, nuancing what is often implied as a binary distinction between risk and uncertainty, as well as of why practitioners with different disciplinary backgrounds and perspectives approach risk and uncertainty differently. First, it is helpful to trace back common definitions of risk and uncertainty used in both academia and practice today. One of the foundational early definitions came from the economist Frank Knight, who defined risk as anything to which we can assign numerical probabilities, whereas uncertainty is anything that cannot be constrained statistically Olivia G. Taylor 759 (Knight, 1921/2006). His original theorisation was tied up with his ‘theory of profit’, in which he argued that making profit required decision-making in the face of uncertainty, because anything which could be constrained numerically could be insured against, and thus any losses could be recuperated with insurance. Knight’s distinction between risk and uncertainty is reflected in later theories on risk and uncertainty, most notably the sociologist Ulrich Beck’s (1992) ‘risk society’ thesis. Beck (1992) argued that the shift from an industrial society to a risk society is defined by risks becoming increasingly ‘incal- culable’. According to this view, novel ‘modernity’ risks include events such as nuclear fallouts or pandemics that are not statistically predictable and cannot be constrained by risk methodologies based on calculating likelihoods, and hence cannot be insured against. Beck (1992) concluded that such non-insurable risks define the modern era as a ‘risk society’. Beck’s ideas have proved to be a major provocation with respect to risk, uncertainty, and politics, despite numerous critiques and iterations in thinking, such as his later work on the ‘world risk society’ (Beck, 2009), which responded to critiques of Eurocentrism and acknowledges the role of governments as a backstop insurer in times of catastrophe. One of the key aspects of Beck’s argument which is relevant here, however, is his use of Knight’s ideas about whether or not something is numerically predictable, and thus the use of ‘insurability’ as a key distinguishing feature between risk and uncertainty. Other work has challenged the binary distinction between what is numerically pre- dictable and what is not, focusing on the distinction between risk and uncertainty. Empirically, it has been shown that many modern insurance techniques blur the distinc- tions between calculative and non-calculative techniques because they include aspects of ‘intuition’ and non-quantitative approaches (Bougen, 2003; O’Malley, 2003, 2004). While insurance is seen as a stereotypically ‘risk-based’ practice in the thinking of both Knight and Beck, many insurance practices are, in fact, characterised by educated guess- work and hunches (Bougen, 2003), where ‘knowability’ is not a clear binary. More recent analysis of the global reinsurance industry further supports this argu- ment. For example, Jarzabkowski, Bednarek, and Spee (2015) provide an account of rein- surance as a financial market for hedging against ‘unknown unknowns’, based on col- lective practices that span both technical and contextual expertise. This complexity and nuance concerning the distinction between risk and uncertainty was highlighted during this study by participants who had come to DRF from catastrophe modelling and reinsurance. For instance, one participant commented that: ‘there is definitely a sort of, I would say intuition that builds up over time, and I think has built up with people in the industry who have been using these models for 20 years or so’ (Inter- view 13, catastrophe modeller). However, another participant with a modelling back- ground felt that humanitarian practitioners were less comfortable with this complexity in modelling, stating that for the humanitarian community, ‘things are often about next year or the next three years . . . they want a very fixed answer, often people look very much only at the single value output and say, “Oh it’s right or wrong”’ (Interview 10, catastro- phe modeller). The policy landscape and challenges of disaster risk financing: navigating risk and uncertainty 760 Going beyond the specific contours of what is calculable and what is not, it is impor- tant to concentrate instead on how knowledge is tied up with how we define risk and uncertainty, while being mindful of the complex nature of knowledge itself. Stirling’s (2009) work on the condition of ‘incertitude’ is useful here, which he sees as extending beyond risk and uncertainty to encompass the conditions of ignorance, uncertainty, ambiguity, and risk. He argues that the boundaries between these different conditions are distinguished by degrees of knowledge relating to two parameters: the extent of knowl- edge of possible outcomes; and the extent of knowledge of the likelihoods of such outcomes (Stirling, 2009). Like others have shown in terms of the ‘fuzzy’ boundary between risk and uncertainty discussed above, Stirling (2009) asserts that the difference between what is known and what is unknown is not always as clear as one might like to think, because the nature of scientific knowledge is not linear, monolithic, or additive. Instead, knowledge is diverse, and often tacit; in many cases, knowing more does not confirm previous knowledge, but rather undermines or destabilises what we thought we knew before (Stirling, 2009). Thus, what most risk and uncertainty scholars share is the view that risk is distin- guished from uncertainty by degrees of knowledge. This insight is particularly pertinent to DRF, because it brings into question the binary between risk and uncertainty that is implied in the focus on ‘acting based on risk’ as opposed to uncertainty. The informa- tion and methodologies being used to trigger more anticipatory action might produce a numerical output, but in practice this is a lot ‘fuzzier’ than it might seem, and there is still inherent uncertainty in this process. Second, focusing on the role of knowledge in our understanding of risk and uncer- tainty in DRF opens up space to consider how practitioners understand risk and uncer- tainty, and how this is influenced by different disciplinary perspectives and epistemologies, or ways of knowing. This is a particularly important consideration in DRF because the sector requires insights from a wide range of expertise, from hydrologists and climate scientists to actuaries and humanitarian practitioners. The resulting policy landscape is highly interdisciplinary, and this has contributed to different ways of thinking about risk and uncertainty, as discussed in the section on the challenges of decision-making in DRF. STS theory has again made important contributions to comprehending the role of knowledge and positionality in risk and uncertainty. For example, in an account of sci- ence and policy in the UK in the wake of the Chernobyl disaster in 1986 in the then Soviet Union, Wynne (1998) explores the advice given to Cumbrian sheep farmers and the dif- ference in approach between a farming and scientific perspective. He contrasts lay knowl- edge with expert knowledge to show how epistemology is crucial to our understanding of uncertainties, demonstrating that the sheep farmers’ tacit knowledge led them to be sceptical about assumptions of predictability, prediction, and control made by the scien- tific community (Wynne, 1998). Indeed, the fact that epistemology and cultural factors play a significant role in determining perceptions of risk has been widely demonstrated in disaster studies literature (Bankoff, 2003; Krüger et al., 2015; Binder and Baker, 2017). However, reflexive analyses of understanding of risk and uncertainty among members of the disaster studies community itself are much less common (Hewitt, 2015). Olivia G. Taylor 761 The differences identified in this research regarding how practitioners thought about risk and uncertainty in their work further supports this statement. While the different conceptions discussed earlier are all consistent with the overarching argument that the difference between risk and uncertainty is determined by degrees of knowledge, they place emphasis in multiple ways that can cause confusion in a sector such as DRF. This study does not suggest that we should try to reconcile these perspectives in a unified way of thinking about risk and uncertainty: different objectives and scopes of work are both good reasons why natural scientists and modellers think about risk and uncertainty in a different way to economists and policymakers, for example. While there has been significant interdisciplinary exchange and learning in this sector, such as the aforementioned A Thesaurus for Anticipatory Humanitarian Action (De Wit, 2019), this research recommends that practitioners engage more thoroughly with ques- tions of knowledge and explore how risk and uncertainty are perceived and understood by both different individuals and agencies. This would help to build understanding and awareness of why risk and uncertainty mean different things to different people, how this is expressed, and how it can be better managed in DRF. Conclusion In analysing the emerging policy area of DRF, this paper has made three key contribu- tions. First, it provides a more cohesive way of defining the sector, which has a number of different terminologies associated with it, most notably in relation to ‘anticipatory action’ and ‘disaster risk financing’. While considerations such as temporality are very important, mechanisms in this wider sector have a great deal in common because of the way in which they link information about disaster risk with action, to facilitate a more timely response. This presents a shared challenge around acting based on information which is inherently incomplete, and therefore acting in the face of uncertainty. Focusing on this common problem is potentially useful for cutting through some of the termi- nological complexity of this emerging policy space, and drawing shared lessons, most notably about how risk and uncertainty can be better understood and navigated. Second, the paper has discussed and unpacked the policy narratives in the sector, revealing some of the underlying contestation relating to the objectives of DRF. Although the central policy narratives of a more efficient and effective response are highly intui- tive and appear to be mutually reinforcing, these goals are difficult to achieve and vary between hazard contexts and agencies. These complexities point to some of the practical difficulties in implementing DRF, which should not be underestimated. Third, the paper has spotlighted the central challenge of DRF: acting based on infor- mation rather than on existing needs. While this is key to the potential that DRF offers in terms of a better disaster response, it also raises significant obstacles in terms of decision-making. Despite the obvious importance of managing uncertainty and questions concerning evidence in the DRF sphere, the paper highlights the tendency in recent years to replace The policy landscape and challenges of disaster risk financing: navigating risk and uncertainty 762 discussions of uncertainty—at least in formal policy documents and speeches—with a vocabulary that concentrates narrowly on ‘risk-based’ decision-making. This relates in part to the sentiment among some participants interviewed in this research, especially those from the donor community, that uncertainty is difficult to recognise explicitly in their work. In practice, however, the field of risk financing and anticipation necessarily grapples with what we know, and by corollary, what we do not know. Drawing from STS theory, the paper concludes that we should think about risk and uncertainty from the standpoint of knowledge, which bounds what is known and what is not known, and recognise that this is often much less clearly distinguishable than we would like to think. Acknowledgements I would like to thank the Natural Environment Research Council in the UK for funding the SHEAR Studentship Cohort programme. I am also very grateful to the anonymous peer reviewers and editors whose comments improved this paper, as well as to all of the interview participants from across the risk financing sector for taking the time to con- tribute to this research. The author declared no potential conflicts of interest with respect to the research, author- ship, and/or publication of this paper. This research was funded by the NERC Science for Humanitarian Emergencies and Resilience Studentship Cohort (SHEAR SSC), grant number: NE/R007799/1, and the SHEAR ForPAc project, grant number: NE/P000673/1. Data availability statement Research data are not shared.10 Correspondence Olivia G. Taylor, Department of Geography, Arts Building C, University of Sussex, Falmer, Brighton BN1 9SJ, UK. E-mail: o.g.taylor@sussex.ac.uk Endnotes 1 Detailed information on the methodology for different mechanisms within DRF are beyond the scope of this paper, but for a broad typology and explanation of different mechanisms see Willitts-King et al. (2020). 2 This event changed name during this research, to represent the shifting focus from a particular method- ology of FbF towards ‘anticipatory humanitarian action’, but the content of the conferences spanned both anticipatory action and risk financing more generally. Olivia G. Taylor 763 3 It was announced in June 2020 that the UK’s Department for International Development (DFID) would merge with the Foreign and Commonwealth Office (FCO) to form a new department, the Foreign, Com- monwealth and Development Office (FCDO), launched in September 2020. Here I refer to and reference ‘DFID’ when the issue at-hand predates this change, or when the document being cited was published by DFID prior to the merger. More information about the amalgamation is available at: https://publications. parliament.uk/pa/cm5801/cmselect/cmfaff/809/80902.htm (last accessed on 20 December 2022). 4 The Anticipation Hub aims to share knowledge and experiences to jointly enhance and scale up anticipa- tory action globally, and brings together the German Red Cross, the International Federation of Red Cross and Red Crescent Societies, and the Red Cross Red Crescent Climate Centre. For more information, see: https://www.anticipation-hub.org/ (last accessed on 13 January 2023). 5 Referring to Cyclone Idai, a major tropical cyclone that hit Mozambique, Zimbabwe, and Malawi in March 2019. 6 ECMWF refers to the European Centre for Medium-Range Weather Forecasts, an independent intergov- ernmental organisation that conducts meteorological research and operational forecasting. 7 Referring to the UK’s Meteorological Office, a national weather service that performs operational forecasting. 8 The first speech, in March 2018, was titled ‘A Casement Lecture: Towards a Better System for Humanitar- ian Financing’, part of a high-level series of lectures at Iveagh House, organised by the Department of Foreign Affairs, Ireland. A follow-up was delivered at the London School of Economics and Political Science in the UK in December 2019, titled ‘Anticipation saves lives: How data and innovative financing can help improve the world’s response to humanitarian crises’. 9 The recorded event, titled the ‘InsuResilience Sectoral Community Workshop: Linking Anticipatory Action to Risk Financing’, took place on 20 September 2021. Timestamp: 8.56. https://www.youtube.com/ watch?v=W6ZB4p4kSgo (last accessed on 20 December 2022). 10 Owing to the nature of this study, participants were assured that any data used would remain anonymous/ de-identified and would not be shared. References Bailey, R. (2012) Famine Early Warning and Early Action: The Cost of Delay. July. Chatham House (The Royal Institute of International Affairs), London. Bankoff, G. (2003) Cultures of Disaster: Society and Natural Hazard in the Philippines. Routledge, London. Beck, U. (1992) Risk Society: Towards a New Modernity. Sage Publications, London. Beck, U. (2009) World at Risk. Polity Press, Cambridge. Binder, S.B. and C.K. Baker (2017) ‘Culture, local capacity, and outside aid: a community perspective on disaster response after the 2009 tsunami in American Sāmoa’. Disasters. 41(2). pp. 282–305. Bougen, P.D. (2003) ‘Catastrophe risk’. Economy and Society. 32(2). pp. 253–274. Cabot Venton, C. (2017) Economics of Resilience to Drought: Kenya Analysis. January. United States Agency for International Development, Washington, DC. Cabot Venton, C., C. Fitzgibbon, T. Shitarek, L. Coulter, and O. Dooley (2013 The Economics of Early Response and Resilience: Summary of Findings. https://assets.publishing.service.gov.uk/media/57a08a0bed915d6 22c000521/61114_Summary_of_Findings_Final_July_22.pdf (last accessed on 22 December 2022). Clarke, D.J. and S. Dercon (2016) Dull Disasters? Oxford University Press, Oxford. Contreras, D. (2016) ‘Fuzzy boundaries between post-disaster phases: the case of L’Aquila, Italy’. International Journal of Disaster Risk Science. 7(3). pp. 277–292. De Wit, S. (2019) Getting Ahead of Crises: A Thesaurus for Anticipatory Humanitarian Action. Centre for Disaster Protection, Oxford. Development Initiatives (2021) Global Humanitarian Assistance Report 2021. Development Initiatives, Bristol. The policy landscape and challenges of disaster risk financing: navigating risk and uncertainty 764 DFID (Department for International Development) (2017) Business Case and Summary: Centre for Global Disaster Protection. 205231. Published: October 2017. https://devtracker.fcdo.gov.uk/projects/GB-1-205231/ documents (last accessed on 22 December 2022). German Red Cross (2015) Framework: Action Plan of the Federal Foreign Office for Humanitarian Adaptation to Climate Change. German Red Cross, Berlin. German Red Cross (2018) Report: 6th Global Dialogue Platform on Forecast-based Financing. German Red Cross, Berlin. German Red Cross (2019) Report: 7th Global Dialogue Platform on Forecast-based Financing. German Red Cross, Berlin. Gettliffe, E. (2021) UN OCHA Anticipatory Action: Lessons from the 2020 Somalia Pilot. Working Paper. 9 October. Centre for Disaster Protection, London. Gneiting, T. (2008) ‘Editorial: probabilistic forecasting’. Journal of the Royal Statistical Society: Statistics in Society. Series A. 171(2). pp. 319–321. Griffiths, M. (2021) Opening Remarks by the Under-Secretary-General for Humanitarian Affairs and Emer- gency Relief Coordinator at the High-Level Event on Anticipatory Action. United Nations, New York, NY, US, 9 September. https://reliefweb.int/report/world/opening-remarks-under-secretary-general- humanitarian-affairs-and-emergency-relief-0 (last accessed on 22 December 2022). Hewitt, K. (2015) ‘Framing disaster in the “global village”: cultures of rationality in risk, security and news’. In F. Krüger et al. (eds.) Cultures and Disasters: Understanding Cultural Framings in Disaster Risk Reduction. Routledge, London. pp. 35–52. Hillbruner, C. and G. Moloney (2012) ‘When early warning is not enough—lessons learned from the 2011 Somalia famine’. Global Food Security. 1(1). pp. 20–28. Hillier, D. and B. Dempsey (2012) A Dangerous Delay: The Cost of Late Response to Early Warnings in the 2011 Drought in the Horn of Africa. Joint Agency Briefing Paper. 18 January. Oxfam GB for Oxfam International and Save the Children, Oxford. Humanitarian Emergency Response Review (2011) Humanitarian Emergency Response Review. Chaired by Lord (Paddy) Ashdown, 28 March. Department for International Development, London. InsuResilience (2018) Concept Note: Shaping the InsuResilience Global Partnership. Final consultation document. May. https://www.insuresilience.org/wp-content/uploads/2022/10/20180502_Concept-Note_ InsuResilience-Global-Partnership.pdf (last accessed on 22 December 2022). InsuResilience Global Partnership and the Anticipation Hub (2021) 1st Workshop of the Sectoral Community ‘Linking Anticipatory Action to Risk Financing’. Virtual call, 20 September. https://www.anticipation- hub.org/Documents/Reports/Sectoral-Community-AA-DRF_20Sep2021-WS_REPORT.pdf (last accessed on 22 December 2022). Jarzabkowski, P., R. Bednarek, and P. Spee (2015) Making a Market for Acts of God: The Practice of Risk Trading in the Global Reinsurance Industry. Oxford University Press, Oxford. Kellett, J. and A. Caravani (2013) Financing Disaster Risk Reduction: A 20 Year Story of International Aid. September. Overseas Development Institute, London, and the Global Facility for Disaster Reduction and Recovery at the World Bank, Washington, DC. Knight, F. (1921/2006) Risk, Uncertainty and Profit. Fourth edition. Dover Publications, Mineola, NY. Krüger, F., G. Bankoff, T. Cannon, B. Orlowski, and L. Schipper (2015) Cultures and Disasters: Understanding Cultural Framings in Disaster Risk Reduction. Routledge, New York, NY. Lash, S., B. Szerszynski, and B. Wynne (1998) Risk, Environment and Modernity: Towards a New Ecology. Sage Publications, London. Lowcock, M. (2018) A Casement Lecture: Towards a Better System for Humanitarian Financing. Iveagh House, Dublin, Ireland, 23 March. https://cerf.un.org/document/casement-lecture-towards-better-system- humanitarian-financing-usgerc-mark-lowcock (last accessed on 22 December 2022). Lowcock, M. (2019) Anticipation Saves Lives: How Data and Innovative Financing Can Help Improve the World’s Response to Humanitarian Crises. London School of Economics and Political Science, London, UK, Olivia G. Taylor 765 2 December. https://reliefweb.int/report/world/mark-lowcock-under-secretary-general-humanitarian- affairs-and-emergency-relief (last accessed on 22 December 2022). Montier, E., C. Harris, and N. Ranger (2019) Disaster Risk Financing in Concert: How Co-ordinated Disaster Risk Financing Can Save More Lives. Opinion Piece. September. START Network, London. Neal, D.M. (1997) ‘Reconsidering the phases of disasters’. International Journal of Mass Emergencies and Disasters. 15(2). pp. 239–264. O’Malley, P. (2003) ‘Governable catastrophes: a comment on Bougen’. Economy and Society. 32(2). pp. 275–279. O’Malley, P. (2004) Risk, Uncertainty and Government. Cavendish Publishing, Portland, OR. START Network (2019) The START Financing Facility. Board Paper. 10 December. https://startprogrammes. app.box.com/s/hm2sq6pjqv3q0fxy0gm0j3lbxp4r7w6e (last accessed on 22 December 2022). START Network (2020) ARC Replica Pay-out: Senegal 2020. Internal Evaluation. December. https://start- network.app.box.com/v/ARC-Replica-Evaluation-2020-EN (last accessed on 22 December 2022). START Network (2021) START Ready. Briefing Document. https://www.anticipation-hub.org/Documents/ Other/START_READY_-_2_pager.pdf (last accessed on 22 December 2022). Stirling, A. (2007) ‘Risk, precaution and science: towards a more constructive policy debate. Talking point on the precautionary principle’. EMBO Reports. 8(4). pp. 309–315. Stirling, A. (2009) ‘Risk, uncertainty and power’. Seminar. 597 (May). pp. 33–39. Stirling, A. (2010) ‘Keep it complex’. Nature. 468(7327). pp. 1,029–1,031. Swithern, S. (2018) Underfunded Appeals: Understanding the Consequences, Improving the System. EBA Report 2018:09. Expert Group for Aid Studies, Stockholm. Talbot, T. and O. Barder (2016) Payouts for Perils: Why Disaster Aid is Broken, and How Catastrophe Insurance Can Help to Fix It. CGD Policy Paper 087. Center for Global Development, Washington, DC. UK Presidency of the G7 (2021) Statement by the UK Presidency of the G7: The Road to COP. G7 Summit, Carbis Bay, Cornwall, UK, 11–13 June 2021. https://www.g7uk.org/the-road-to-cop-statement-by-the-uk- presidency-of-the-g7/ (last accessed on 22 December 2022). UN OCHA (United Nations Office for the Coordination of Humanitarian Affairs) (2020) Global Humani- tarian Overview 2021. UN OCHA, Geneva. UNICEF (United Nations Children’s Fund) and WFP (World Food Programme) (2015) Return on Investment for Emergency Preparedness study. Final Report. January. Boston Consulting Group, Boston, MA. United Nations and the Governments of Germany and the United Kingdom (2021) Co-chairs’ Statement. High-Level Humanitarian Event on Anticipatory Action, United Nations, New York, NY, US, 10 September. https://www.unocha.org/sites/unocha/files/AnticipatoryActionHLE2021_CoChairsStatement.pdf (last accessed on 22 December 2022). Weingärtner, L., T. Pforr, and E. Wilkinson (2020) The Evidence Base on Anticipatory Action. World Food Programme, Rome. Wilkinson, E. et al. (2018) Forecasting Hazards, Averting Disasters: Implementing Forecast-based Early Action at Scale. Report. March. Overseas Development Institute, London. Willitts-King, B., L. Weingärtner, F. Pichon, and A. Spencer (2020) Risk-informed Approaches to Humani- tarian Funding: Using Risk Finance Tools to Strengthen Resilience. Report. May. Overseas Development Institute, London. World Bank and GFDRR (Global Facility for Disaster Reduction and Recovery) (2018) Global Risk Financing Facility: Supporting Early Action to Climate Shocks, Disasters and Crises. Information Sheet. World Bank, Washington, DC. World Meteorological Organization (2021) WMO Atlas of Mortality and Economic Losses from Weather, Climate and Water Extremes (1970–2019). WMO-No. 1267. World Meteorological Organization, Geneva. Wynne, B. (1998) ‘May the sheep safely graze? A reflexive view of the expert–lay knowledge divide’. In S. Lash, B. Szerszynski, and B. Wynne (eds.) Risk, Environment and Modernity: Towards a New Ecology. Sage Pub- lications, London. pp. 44–83. The policy landscape and challenges of disaster risk financing: navigating risk and uncertainty
10.3389_fpsyt.2022.1086038
fpsyt-13-1086038 January 13, 2023 Time: 17:35 # 1 OPEN ACCESS EDITED BY Nuno Madeira, University of Coimbra, Portugal REVIEWED BY Sandra Vieira, Institute of Psychiatry, Psychology and Neuroscience, King’s College London, United Kingdom Marco Simoes, University of Coimbra, Portugal João Valente Duarte, University of Coimbra, Portugal *CORRESPONDENCE Diana Prata diana.prata@kcl.ac.uk SPECIALTY SECTION This article was submitted to Neuroimaging, a section of the journal Frontiers in Psychiatry RECEIVED 31 October 2022 ACCEPTED 29 December 2022 PUBLISHED 19 January 2023 CITATION Tavares V, Vassos E, Marquand A, Stone J, Valli I, Barker GJ, Ferreira H and Prata D (2023) Prediction of transition to psychosis from an at-risk mental state using structural neuroimaging, genetic, and environmental data. Front. Psychiatry 13:1086038. doi: 10.3389/fpsyt.2022.1086038 COPYRIGHT © 2023 Tavares, Vassos, Marquand, Stone, Valli, Barker, Ferreira and Prata. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms. TYPE Original Research PUBLISHED 19 January 2023 DOI 10.3389/fpsyt.2022.1086038 Prediction of transition to psychosis from an at-risk mental state using structural neuroimaging, genetic, and environmental data Vânia Tavares1,2, Evangelos Vassos3,4, Andre Marquand5,6, James Stone7, Isabel Valli8,9, Gareth J. Barker10, Hugo Ferreira1 and Diana Prata1,11* 1Instituto de Biofísica e Engenharia Biomédica, Faculdade de Ciências, Universidade de Lisboa, Lisbon, Portugal, 2Faculdade de Medicina, Universidade de Lisboa, Lisbon, Portugal, 3Social, Genetic and Developmental Psychiatry Centre, Institute of Psychiatry, Psychology and Neuroscience, King’s College London, London, United Kingdom, 4National Institute for Health Research Maudsley Biomedical Research Centre, South London and Maudsley National Health System Trust, London, United Kingdom, 5Donders Centre for Cognitive Neuroimaging, Donders Institute for Brain, Cognition and Behaviour, Radboud University, Nijmegen, Netherlands, 6Department of Cognitive Neuroscience, Radboud University Medical Centre, Nijmegen, Netherlands, 7Brighton and Sussex Medical School, University of Sussex, Brighton, United Kingdom, 8Department of Psychosis Studies, Institute of Psychiatry, Psychology and Neuroscience, King’s College London, London, United Kingdom, 9Institut d’Investigacions Biomèdiques August Pi i Sunyer, University of Barcelona, Barcelona, Spain, 10Department of Neuroimaging, Institute of Psychiatry, Psychology and Neuroscience, King’s College London, London, United Kingdom, 11Department of Old Age Psychiatry, Institute of Psychiatry, Psychology and Neuroscience, King’s College London, London, United Kingdom Introduction: Psychosis is usually preceded by a prodromal phase in which patients are clinically identified as being at in an “At Risk Mental State” (ARMS). A few studies have demonstrated the feasibility of predicting psychosis transition from an ARMS using structural magnetic resonance imaging (sMRI) data and machine learning (ML) methods. However, the reliability of these findings is unclear due to possible sampling bias. Moreover, the value of genetic and environmental data in predicting transition to psychosis from an ARMS is yet to be explored. Methods: In this study we aimed to predict transition to psychosis from an ARMS using a combination of ML, sMRI, genome-wide genotypes, and environmental risk factors as predictors, in a sample drawn from a pool of 246 ARMS subjects (60 of whom later transitioned to psychosis). First, the modality-specific values in predicting transition to psychosis were evaluated using several: (a) feature types; (b) feature manipulation strategies; (c) ML algorithms; (d) cross-validation strategies, as well as sample balancing and bootstrapping. Subsequently, the modalities whose at least 60% of the classification models showed an balanced accuracy (BAC) statistically better than chance level were included in a multimodal classification model. Results and discussion: Results showed that none of the modalities alone, i.e., neuroimaging, genetic or environmental data, could predict psychosis from an ARMS statistically better than chance and, as such, no multimodal classification model was trained/tested. These results suggest that the value of structural MRI data and genome-wide genotypes in predicting psychosis from an ARMS, which has been fostered by previous evidence, should be reconsidered. KEYWORDS machine learning, biomarker, schizophrenia, ARMS, prognosis Frontiers in Psychiatry 01 frontiersin.org fpsyt-13-1086038 January 13, 2023 Time: 17:35 # 2 Tavares et al. 10.3389/fpsyt.2022.1086038 1. Introduction Psychosis is a severe condition usually within the context of a mental disorder such as a schizophrenia, some neurological disorders (e.g., Alzheimer’s disease) or other medical conditions (e.g., induced by drugs or illicit substances), characterized by a disconnection from reality (1). The onset of psychosis, when in the context of a mental disorder, is typically preceded by a prodromal phase that lasts months to years (2); and usually starts early during adolescence and precedes the onset of psychotic symptoms by 10 or more years (3). In this prodromal phase, subtle and subjectively experienced disturbances in mental processes emerge (basic symptoms). These are the first manifestations of the neurobiological processes underlying psychosis and are mainly distinguished from other symptoms (i.e., positive or negative symptoms) by their self-experience nature (4). As the course of the psychotic illness evolves, increasingly disabling behavioral symptoms start to emerge, generally called negative symptoms, in particular a reduction of motivation and/or expressiveness (5). Additionally, cognitive deficits in attention, memory, reasoning, lack of concentration and executive functioning appear (6). Lastly, positive symptoms emerge, such as hallucinations, delusions, disorganized speech, and behavior (1). A patient may be clinically identified as being at a late prodromal phase of psychosis or having an “At Risk Mental State” (hereinafter: ARMS) if they present a functional decline in association with one or more of the following commonly used criteria (2, 7): (1) attenuated psychotic symptoms (APS), such as delusions, hallucinations, or disorganized speech with a frequency of at least once per week in the past month; (2) a brief limited intermittent psychotic (BLIP) episode lasting less than 1 week which resolves without antipsychotic medication; or (3) a genetic liability to psychosis or schizotypal traits, i.e., having either a first-degree relative with psychosis or a schizotypal personality disorder. Transition to psychosis from an ARMS may be evaluated based on the severity, frequency, and total duration of the psychotic symptoms, i.e., when the subject experiences a first episode of psychosis (FEP). Subjects with an ARMS and seeking help have a transition rate to psychosis of about 9% in the first 6 months and 25% in the first 3 years (8) and, in particular, an increased risk of transition to schizophrenia of 15.7% within an average period of 2.35 years, as shown by a meta-analysis (9). Thus, most of the people with an ARMS who later develop a psychotic illness will be diagnosed with schizophrenia. Furthermore, since about 70% of subjects diagnosed with an ARMS never develop a full- blown psychotic illness (9), these people may benefit from a less intensive treatment to ameliorate symptoms or need no treatment at all. Such increase in treatment cost-effectiveness would represent a substantial decrease in healthcare costs, and treatment burden to patients, including pharmacological side effects. However, there is no method for distinguishing between individuals with an ARMS who will subsequently develop a psychotic illness from those who will not (i.e., before a FEP onset). Given the above need, an effective, precise, and quantitative tool for the prediction of transition to psychosis from an ARMS has been sought by several studies employing machine learning (ML) methods and structural magnetic resonance imaging (sMRI). Indeed, several studies have consistently showed prediction of transition to psychosis from as ARMS with accuracies ranging between 74 and 84% (10– 15). Transition to psychosis from an ARMS using only sMRI and ML was first predicted using whole-brain gray matter volume metrics with an accuracy of 82% [(15 ARMS who transitioned to psychosis (ARMS-T) and 18 who did not (ARMS-NT)] (10). This finding was later replicated: (a) in an independent sample by the same group [balanced accuracy (BAC) = 84%, 16 ARMS-T and 21 ARMS-NT] (11); (b) combining both these samples (BAC = 80%, 33 ARMS-T and 33 AMRS-NT) (12); (c) using also one of the above samples for graph- extracted network metrics from cortical gyrification (BAC = 81%, 16 ARMS-T and 63 ARMS-NT) (15), and regional gray matter metrics (BAC = 74%, 16 ARMS-T and 19 ARMS-NT) (14); and (d) using regional gray matter metrics in an independent sample (BAC = 77%, 17 ARMS-T and 17 ARMS-NT; specificity of a replication sample of individuals with an ARMS who did not develop psychosis = 68%, 40 ARMS-NT) (13). To date, only two, relatively small, ARMS samples have been used for sMRI and ML analysis: FETZ (10, 12, 15) and FePsy (11, 12, 14). Thus, the robustness and generalizability of the above findings are still unclear due to possible specific sample characteristics, i.e., small sample sizes (from 33 individuals to at most 79 individuals with ARMS), with several studies stemming from a single site (10, 11, 13–15) or a combination of previously studied sites (12), which makes them not actual replications, with one exception (13). Interestingly, to the best of our knowledge, genetic data has been explored for the prediction of the transition to psychosis from an ARMS only once (16). In this study, a schizophrenia polygenic risk score (PRS) was able to predict transition to psychosis in individuals with an European [area under the curve (AUC) = 0.65; 32 ARMS- T and 92 ARMS-NT] and with a Non-European (AUC = 0.59; 48 ARMS-T and 156 ARMS-NT) ancestry, respectively. This is despite there being several classification studies showing that genetic markers can predict schizophrenia (17–22), FEP (23) or ARMS (23), both of individual polymorphisms (18, 19, 21, 23) or, composite polygenic scores (20–22), and gene expression profiles (24). From an environmental exposure perspective, and to the best of our knowledge, environmental data have never been explored for predicting individual transition to psychosis from an ARMS. The combination of neuroimaging measures and genetics or environmental measures, using ML, has, to the best of our knowledge, been explored once to predict ARMS prognosis (i.e., transition to psychosis from an ARMS) in a study running in parallel to ours (25). Therein, a large sample from the PRONIA project (26 ARMS-T and 308 ARMS-NT from 7 sites) was used to build a sequential stacked multimodal model using clinical-neurocognitive (including environmental data), genetic (in the form of a PRS for schizophrenia) and neuroimaging (in the form of voxel-based gray matter volume maps) data and - unlike the present study–human prognostic ratings, showing a final balanced accuracy in predicting transition to psychosis of 86%. In the present longitudinal prognostic biomarker study, we aimed to explore the use of ML models trained with sMRI, genetic, and environmental baseline data to predict the individual-level transition to psychosis from an ARMS within a 2-year follow up. While providing such preliminary (given the unprecedented data combinations/features and a limited sample size) evidence at the multimodal level, we took the opportunity to attempt to replicate previous promising sMRI-ML findings of studies using similar or smaller sample size (10–15). Methods-wise, we used naturalistically diverse samples but balanced them for demographic (age and sex) and imaging (scan acquisition sMRI protocol) variables. We set out to train and test modality-specific models first and then, provided Frontiers in Psychiatry 02 frontiersin.org F r o n t i e r s TABLE 1 Socio-demographic and clinical information of the At Risk Mental State (ARMS) sample with structural MRI data. Protocol 1 Protocol 2 Protocol 3 Group comparison i n P s y c h a t r y i Age at baseline (years) Age at follow-up or transition (years) Age at scan (years) Interval between baseline and scan age (years) ARMS-T (n = 14) 23.2 ± 3.4 [15.6 26.9] 25.6 ± 4.2 [17.3 33.4] 23.0 ± 3.6 [17.5 27.8] –0.2 ± 1.4 [–2.3 1.9] ARMS-NT (n = 19) 24.5 ± 4.8 [19.2 34.5] 32.7 ± 5.2 [22.6 43.9] 23.9 ± 4.8 [18.5 34.8] –0.5 ± 1.1 [–2.3 2.1] ARMS-T (n = 3) 26.2 ± 7.0 [20.1 33.8] 29.2 ± 5.4 [20.2 38.8] 27.0 ± 8.2 [20.2 36.1] 0.9 ± 1.3 [0.1 2.4] ARMS-NT (n = 16) 24.5 ± 5.2 [17.8 35.3] 28.8 ± 5.6 [22.9 43.1] 25.1 ± 5.4 [18.6 37.4] 0.5 ± 0.5 [0.1 2.1] ARMS-T (n = 6) 23.4 ± 4.5 [17.5 29.2] 25.2 ± 4.8 [18.3 31.0] 24.1 ± 4.8 [18.3 30.8] 0.6 ± 0.5 [0.2 1.6] ARMS-NT (n = 41) 21.8 ± 4.3 [17.1 33.1] 25.6 ± 4.8 [20.3 41.2] 22.4 ± 4.6 [17.7 38.3] 0.6 ± 1.0 [0.0 5.1] Sex (male/female) 11/3 9/10 2/1 14/2 3/3 19/22 Handednessa (right/left/ambidextrous) 12/0/1 16/0/0 3/0/0 13/1/0 4/0/0 36/4/0 0 3 Self-reported ethnicity (White/Black/Asian/other) 7/5/1/1 11/5/1/2 2/1/0/0 13/1/1/1 4/1/1/0 19/19/1/2 Years of education IQ (z-standardized)b GAF at baseline GAF at follow-upc CAARMS at baselined CAARMS at follow-upe 13.4 ± 2.1 [10 18] –1.1 ± 1.1 [–2.1 1.0] 52.9 ± 16.0 [35 90] 49.3 ± 18.6 [10 69] 33.2 ± 13.0 [9 56] 19.6 ± 23.0 [0 63] 13.1 ± 1.9 [10 17] 0.0 ± 1.1 [–2.1 1.8] 57.8 ± 11.4 [35 75] 58.5 ± 17.9 [20 94] 28.4 ± 15.3 [8 58] 11.6 ± 10.9 [0 31] 11.7 ± 2.3 [9 13] 0.1 ± 0.1 [0.0 0.2] 58.7 ± 3.2 [55 61] 27.3 ± 6.8 [22 35] 29.3 ± 21.9 [12 54] 42.0 ± 43.3 [6 90] 14.1 ± 2.6 [11 20] 0.5 ± 0.9 [–1.3 1.6] 58.6 ± 9.9 [41 75] 62.3 ± 13.5 [46 93] 23.2 ± 14.3 [0 51] 14.7 ± 18.4 [0 54] 15.2 ± 2.5 [11 18] −0.1 ± 1.3 [–2.1 1.6] 50.3 ± 11.4 [35 65] 52.5 ± 20.0 [30 78] 39.7 ± 24.1 [0 69] 42.7 ± 42.1 [0 102] 13.0 ± 2.2 [9 20] 0.1 ± 1.1 [–2.1 3.5] 53.6 ± 14.8 [0 75] 66.2 ± 13.6 [33 87] 28.5 ± 16.7 [0 81] 15.5 ± 17.2 [0 60] Protocol: p = 0.271 Transition: p = 0.592 Protocol × Transition: p = 0.447 Protocol: p = 0.027* Transition: p = 0.099 Protocol × Transition: p = 0.025* Protocol: p = 0.261 Transition: p = 0.499 Protocol × Transition: p = 0.541 Protocol: p < 0.001*** Transition: p = 0.419 Protocol × Transition: p = 0.795 Protocol × Transition: Protocol 1: p = 0.070 Protocol 2: p = 0.422 Protocol 3: p = 1 Protocol × Transition: Protocol 1: p = 0.448 Protocol 2: p = 1 Protocol 3: p = 1 Protocol × Transition: Protocol 1: p = 0.933 Protocol 2: p = 0.530 Protocol 3: p = 0.212 Protocol: p = 0.298 Transition: p = 0.966 Protocol × Transition: p = 0.024* Protocol: p = 0.427 Transition: p = 0.252 Protocol × Transition: p = 0.923 Protocol: p = 0.402 Transition: p = 0.475 Protocol × Transition: p = 0.877 Protocol: p = 0.064 Transition: p < 0.001*** Protocol × Transition: p = 0.095 Protocol: p = 0.505 Transition: p = 0.153 Protocol × Transition: p = 0.824 Protocol: p = 0.082 Transition: p = 0.001*** Protocol × Transition: p = 0.262 f r o n t i e r s i n o r g . Data format: mean ± standard deviation [min max]. Information not available for a1 ARMS-T and 3 ARMS-NT (Protocol 1), 2 ARMS-NT (Protocol 2), 2 ARMS-T and 1 ARMS-NT (Protocol 3); b1 ARMS-T and 1 ARMS-NT (Protocol 2), 1 ARMS-NT (Protocol 3); c2 ARMS and 5 ARMS-NT (Protocol 1), 4 ARMS-NT (Protocol 2), 8 ARMS-NT (Protocol 3); d2 ARMS-T and 7 ARMS-NT (Protocol 1), 3 ARMS-NT (Protocol 2), 2 ARMS-NT (Protocol 3); e3 ARMS-T and 6 ARMS-NT (Protocol 1), 3 ARMS-NT (Protocol 2), 8 ARMS-NT (Protocol 3). ARMS, at-risk mental state; ARMS-T, individuals at ARMS that did not transition to psychosis; ARMS-NT, individuals at ARMS that did not transitioned to psychosis. *p < 0.05; ***p < 0.001. f p s y t - 1 3 - 1 0 8 6 0 3 8 J a n u a r y 1 3 , 2 0 2 3 T i m e : 1 7 : 3 5 # 3 T a v a r e s e t a l . . 1 0 3 3 8 9 / f p s y t . 2 0 2 2 . 1 0 8 6 0 3 8 fpsyt-13-1086038 January 13, 2023 Time: 17:35 # 4 Tavares et al. 10.3389/fpsyt.2022.1086038 these performed above chance-level, a multimodal one. For the sMRI data, we used state-of-the-art preprocessing and ML pipelines; and explored several unprecedented combinations of brain structural measures, feature manipulation and cross-validation (CV) strategies. For the genetic data, we explored several approaches: a schizophrenia individual GWA-implicated SNPs (27), and a brain- PRS (26), specific expression Quantitative Trait Loci (eQTL) score. For the environmental data, we employed a schizophrenia environmental risk score (ERS) (28), and individual risk factors. 2. Materials and methods 2.1. Sample description The total sample consisted of 246 individuals with an ARMS, recruited at first presentation from consecutive referrals to the Outreach and Support in South London (OASIS) high-risk service, South London and Maudsley NHS Foundation Trust (29). The presence of ARMS was assessed using the CAARMS, a detailed clinical assessment (30). When the subjects were first diagnosed as having an ARMS (i.e., baseline) a set of data were acquired: (a) a sMRI scan; (b) genome-wide genotypes; and (c) assessment of environmental risk exposures. Subjects were labeled as transitioned to psychosis (ARMS-T) if they later presented a FEP or as not- transitioned to psychosis (ARMS-NT) if they did not present a FEP within a period of at least 2 years. For a detailed description of the recruitment, inclusion and exclusion criteria please refer to the Supplementary material. Additional socio-demographic including: and clinical measures were also assessed at baseline, age; sex; handedness; self-reported ethnicity; full scale intelligence TABLE 2 Socio-demographic and clinical information of the At Risk Mental State (ARMS) sample with genetic data and an European ancestry. ARMS-T (n = 21) ARMS-NT (n = 54) Group comparison Age at baseline (years) Age at follow-up or transition (years) 23.8 ± 5.3 [15.6 33.8] 25.3 ± 5.9 [17.3 38.8] 22.5 ± 4.0 [14.6 34.5] 27.9 ± 5.1 [18.5 43.9] Sex (male/female) 14/7 30/24 Years of education IQ (z-standardized)a GAF at baseline GAF at follow-upb CAARMS at baselinec CAARMS at follow-upd 13.0 ± 2.2 [10.0 18.0] 0.1 ± 1.0 [–2.1 2.2] 54.0 ± 15.7 [0 80] 47.8 ± 24.3 [0 79] 37.6 ± 17.5 [6 69] 24.4 ± 27.9 [0 90] 12.0 ± 4.4 [0 18.0] 0.2 ± 1.0 [–2.1 1.8] 53.6 ± 16.0 [0 78] 59.2 ± 21.0 [0 94] 29.9 ± 16.2 [0 81] 12.4 ± 14.0 [0 60] p = 0.284 p = 0.069 p = 0.380 p = 0.292 p = 0.678 p = 0.923 p = 0.050 p = 0.097 p = 0.019* quotient measured by the National Adult Reading Test (31); years of education; and global assessment of function using the GAF instrument tool at baseline and at follow-up (32), and CAARMS (at baseline and follow-up) (30). Regarding the sMRI, genetic and environmental sub-samples: 99, 135 and all the 246 individuals with an ARMS had a baseline sMRI scan (Table 1), genome-wide genotyped data (Table 2), and environmental risk factors assessment data (Table 3), respectively (more details in the Supplementary material). Over the 2-years follow-up period, 23, 41, and 60 individuals at an ARMS from each of the previous sub-samples developed psychosis (AMRS-T) and the remaining 15, 94, and 186 did not (ARMS-NT), respectively. Moreover, part of the study’s data collection occurred under the Genetic and Psychosis (GAP) umbrella project (33). Ethics approval was obtained by the NHS South East London Research Ethics Committee (Project GAP; Ref. 047/04), consistent with the Helsinki Declaration of 1975 (as revised in 2008) and all subjects gave written informed consent. Socio-demographic and clinical variables were analyzed using a two-tailed independent t-test or a Univariate Analysis of Variance (ANOVA) for continuous data and a chi-square test or Fisher’s exact test (if there were less than 5 subjects in one group) for ordinal data (Tables 1–3). These statistical analyses were performed using the Statistical Package for the Social Sciences 26 (SPSS 26 for Windows, Chicago, IL, USA). 2.2. Structural neuroimaging data 2.2.1. Structural magnetic resonance imaging acquisition Structural magnetic resonance imaging (sMRI) scans were acquired with one of two scanners (one with a field strength of 1.5T, three 3-Dimensional enhanced fast gradient echo protocols (detailed description in Supplementary material). the other 3T) using one of 2.2.2. Image processing High spatial resolution volumetric T1-weighted images were processed with the Computational Anatomy Toolbox for Statistical Parametric Mapping (SPM) –12 (CAT12; v10921), an SPM12 add- on (v69092) using default settings and MATLAB (9.3) as we have described elsewhere (34) (detailed description in Supplementary material). In summary, gray and white matter volumes for 64 regions-of-interest is in the Supplementary Table 1) were extracted using the Hammers atlas (35). Additionally, regional-based cortical thickness and surface measures (i.e., folding measures)–gyrification index, the depth of sulci and the measurement of local surface complexity were extracted for 68 ROIs (description of each ROI is in the Supplementary Table 2) defined by the Desikan–Killiany atlas (36). (ROIs; description of each ROI 2.2.3. Image quality control The quality of each processed image was empirically assessed using the quality assurance framework of CAT12 (detailed description in the Supplementary material). We set the subject’s image inclusion threshold at D (sufficient), i.e., only subjects whose Data format: mean ± standard deviation [min max]. Information not available for a2 ARMS-T and 9 ARMS-NT; b4 ARMS-NT; c1 ARMS and 9 ARMS-NT; d1 ARMS-T and 3 ARMS-NT. ARMS, at-risk mental state; ARMS-T, individuals at ARMS that did not transition to psychosis; ARMS-NT, individuals at ARMS that did not transitioned to psychosis. *p < 0.05. 1 http://www.neuro.uni-jena.de/cat/ 2 http://www.fil.ion.ucl.ac.uk/spm/ Frontiers in Psychiatry 04 frontiersin.org fpsyt-13-1086038 January 13, 2023 Time: 17:35 # 5 Tavares et al. 10.3389/fpsyt.2022.1086038 images had an image quality rate of A (excellent) to D (sufficient) (in a scale that goes up to F–unacceptable/failed) were included in the final sample, as it has been shown that typical scientific (clinical) data get good-to-satisfactory ratings (37). All this study’s images passed the above criteria and thus were included in all analyses (see Supplementary material for more details). migrant; (4) belonging to an ethnic minority; (5) the upbringing urbanicity level; (6) the parental age at birth; (7) the presence of childhood trauma; and (8) the season of birth (detailed description of how the risk for psychosis was assessed in each factor is in Supplementary material). 2.3. Genetic data Genotyping procedures have been previously described (26, 38). In summary, samples were genotyped at two different sites with two distinct chips (Illumina HumanCore Exome BeadChip and Genome-wide Human SNP Array 6.0). A standard quality control screening (exclusion of SNPs with low minor allele frequency, high genotypic failure and not in Hardy Weinberg equilibrium) followed by imputation procedures were conducted. Then, samples from both sites were merged by keeping only the overlapped imputed SNPs followed by a second quality control screening. Finally, a population stratification analysis was conducted with principal component analysis (PCA) to select only subjects with a European ancestry (the number of subjects per self-reported ethnicity is in the Supplementary Table 3). For a detailed description see the Supplementary material. 2.4. Environmental data Each subject was assessed on at least one of eight environmental risk factors: (1) tobacco and (2) cannabis consumption; (3) being TABLE 3 Socio-demographic and clinical information of the At Risk Mental State (ARMS) sample with environmental data (with less than 20% of the environmental risk factors missing). ARMS-T (n = 37) ARMS-NT (n = 97) Group comparison Age at baseline (years) Age at follow-up or transition (years)a 23.6 ± 4.8 [15.6 33.6] 25.6 ± 5.6 [17.3 39.2] 21.9 ± 3.7 [14.6 33.1] 27.1 ± 4.7 [18.5 41.2] Sex (male/female) 22/15 50/47 Years of educationb IQ (z-standardized)c GAF at baselined GAF at follow-upe CAARMS at baselinef CAARMS at follow-upg 13.2 ± 2.7 [8 18] −0.3 ± 1.0 [–2.1 2.2] 55 ± 12.5 [35 90] 50.4 ± 19.9 [10 88] 30.9 ± 19.4 [0 78] 29.7 ± 31.2 [0 102] 13.3 ± 2.0 [9 18] 0.1 ± 1.0 [–2.1 3.5] 56.7 ± 8.6 [40 80] 63.2 ± 14.2 [20 94] 28.3 ± 16.0 [0 81] 13.3 ± 14.2 [0 60] p = 0.027* p = 0.131 p = 0.411 p = 0.686 p = 0.049* p = 0.523 p =< 0.001* p = 0.478 p =<0.001* Data format: mean ± standard deviation [min max]. Information not available for a1 ARMS-T; b5 ARMS-T and 6 ARMS-NT; c7 ARMS-T and 13 ARMS-NT; d5 ARMS-T and 4 ARMS-NT; e5 ARMS-T and 8 ARMS-NT; f6 ARMS-T and 13 ARMS-NT; g4 ARMS-T and 8 ARMS-NT; subject. ARMS, at-risk mental state; ARMS-T, individuals at ARMS that did not transition to psychosis; ARMS-NT, individuals at ARMS that did not transition to psychosis. *p < 0.05. 2.5. Machine learning approach Several ML strategies to generate prediction models for transition to psychosis from sMRI data using our ARMS sample were investigated (Figures 1, 2). These include: (a) sample balancing and bootstrapping; and testing several: (b) feature types; (c) feature manipulation approaches; and (d) CV approaches. The analyses were conducted using the neuroimaging ML tool NeuroMiner v1.0 ELESSAR3 for sMRI data, chosen given that it was used in the previous above-mentioned ARMS prognosis studies and provided therein high accuracy results (12, 39, 40), and R software 4.0.5 (41) for genetic (16) and environmental data. As detailed below, we have used SVM on the neuroimaging data since that is the approach which not only is more often employed with sMRI data but also that which has shown higher accuracies in psychiatric diagnostic classifications using sMRI data including in the ARMS population (10–14) which we herein attempt to replicate. We have used elastic-net algorithm for the genetic data (SNPs and eQTL scores) and environmental risk factors as it a well-suited method for dealing with high-dimensional data and possibly correlated data; and it performs an embedded feature selection and model fitting at once. The PRS and the environmental risk score were analyzed with logistic regression, given that only one feature was used. 2.5.1. Sample balancing and bootstrapping The final sample used in the ML analyses was defined by all the ARMS-T subjects available (23 subjects for the sMRI predictors, 19 for the PRS predictor, 21 for the SNP’s alleles predictors, 21 for eQTL scores predictors, 37 for the ERS predictor, and 17 for the individual environmental predictors), and the same number of ARMS-NT subjects randomly selected to match the ARMS-T for age and sex (for each data modality), and for scan acquisition protocol (for sMRI data). The matching criteria for age and sex were based on the non-rejection of the null hypothesis (i.e., p > 0.05) that the ARMS-T and ARMS-NT groups had the same median age (tested with a two-sided Mann–Whitney U-test) and sex proportions (tested with a two-sided chi-square statistic). The matching for the scan acquisition protocol was done in a one-to-one manner, i.e., the number of ARMS-NT subjects is the same as the number of ARMS-T. within each protocol Of note, we have considered the approach of applying a class- weighted support vector machine for our neuroimaging measures and have detected that differences in terms of accuracies between a model with weights vs. no-weights (considering the full unbalanced samples) were practically null (results not shown)–and therefore we did not pursue that approach. Then, each subsampling was repeated five times, i.e., 5 bootstrapped samples were created, and the subsequent ML analyses were conducted for each of the bootstrapped sample. 3 http://proniapredictors.eu/neurominer/index.html Frontiers in Psychiatry 05 frontiersin.org fpsyt-13-1086038 January 13, 2023 Time: 17:35 # 6 Tavares et al. 10.3389/fpsyt.2022.1086038 FIGURE 1 Overall machine learning approach taken for assessing the predictive value, i.e., the accuracy, of each type of extracted neuroimaging, genetic or environmental feature in predicting transition to psychosis from an At Risk Mental State (ARMS). ERS, environmental risk score; eQTL score, expression quantitative trait loci; PRS, polygenic risk score; ROIGM, regional-based gray matter volumes; ROISurface, surface-based regional cortical thickness, and gyrification, sulci, and complexity indexes; ROIWM, regional-based white matter volumes; SNP, single nucleotide polymorphism; VMGM, voxel-based gray matter volume maps; VMWM, voxel-based white matter volume maps. 2.5.2. Feature types 2.5.2.1. Structural magnetic resonance imaging data Individual ML models were trained and validated for each of the following brain measures: (a) voxel-based gray matter (VBGM) maps (297,811 initial features); (b) voxel-based white matter (VBWM) maps (204,706 initial features); (c) regional-based gray (ROIGM) and (d) white (ROIWM) matter volumes (each with 64 initial features) scaled to the total intracranial volume (TIV); and (e) surface-based regional cortical thickness, and gyrification, sulci, and complexity indexes (ROISurface; 272 initial features). Each feature is scaled between 0 and 1 before entering a support vector machine (SVM) classification algorithm. 2.5.2.2. Genetic data We tested whether a PRS which we have previously found to predict (R2 = 0.94) a cross-sectional diagnosis of FEP (vs. healthy controls) would be a good longitudinal predictor for ARMS prognosis. Following the same methodology (26), this PRS was computed as the sum of SNPs alleles statistically associated with schizophrenia in a GWAS meta-analysis study (42) weighted by the effect size of that association (more details in Supplementary material). In addition, two other novel prediction models using the present ARMS sample were trained and tested. One used SNPs’ alleles (79,247 SNPs) as predictors and the other used eQTL scores of genes expressed in brain tissue (141 genes across several brain tissues). Both SNPs and genes’ eQTL scores were chosen as the ones most associated with psychosis as ascertained in a recent meta-analysis (27). The eQTL score of each gene was extracted with the eGenScore which we developed and published previously (43) and it was computed as the sum of the alleles of SNPs showing a statistically significant association with the brain gene expression in a standard genomic and transcriptomic sample weighted by the size of that effect (further details available in Supplementary material). 2.5.2.3. Environmental data We tested whether an ERS for psychosis which we have previously developed (28) would be a good longitudinal predictor for ARMS prognosis. Only subjects with less than 20% of missing information (i.e., missing data for less than 2 environmental risk factors) were considered for the ERS-based ML analysis. Therefore, the final sample included 37 ARMS-T subjects and 97 ARMs-NT subjects. Then, each environmental risk factor (see Section “2.4. Environmental data”) was used as an individual feature in the model. For this ML analysis only subjects with information for all the environmental risk factors (i.e., with no missing information) were considered (i.e., 17 ARMS-T and 49 ARMS-NT subjects). Further details available in Supplementary material. 2.5.3. Feature manipulation Feature manipulation was performed only in ML analyses using sMRI data. In particular, feature dimensionality reduction was performed for VBGM and VBWM features using robust PCA (44, Frontiers in Psychiatry 06 frontiersin.org fpsyt-13-1086038 January 13, 2023 Time: 17:35 # 7 Tavares et al. 10.3389/fpsyt.2022.1086038 FIGURE 2 Scheme of the cross-validation (CV) approach taken to train, test, and validate classification models trained with (A) neuroimaging data and support vector machines (SVM); genetic (single nucleotide polymorphisms or expression quantitative trait loci) or environmental (environmental risk factors) data and elastic-net; or (B) genetic (polygenic risk score) or environmental (environmental risk score) data and logistic regression. 45). Here the robust PCA was applied during the inner CV cycle (see Section “2.5.5. Cross-validation”). The number of principal components that were retained explained up to 80% of the variance in the data and were limited by the inner CV cycle’s sample size, n, i.e., a maximum of only n/2 components could indeed be extracted. Supplementary Table 5 shows the maximum number of principal components that can be extracted for each inner CV cycle in each CV scheme that was used (see also Section “2.5.5. Cross-validation”) (for detailed description see the Supplementary material). Feature selection was performed on regional brain features (i.e., ROIGM, ROIWM, and ROISurface) using a greedy forward search feature selection algorithm. This is a stepwise algorithm that starts with an empty set of features and then tests the predictive value of every single feature, selecting the ones improving the overall accuracy across the inner CV cycle folds (see Section “2.5.5. Cross- validation”). The final set of features is, then, composed by the 10% most predictive variables. Additionally, no feature selection, i.e., using the total number regional brain features, was also tested. 2.5.4. Machine learning algorithm Binary classification of transition to psychosis from an ARMS (i.e., ARMS-T vs. ARMS-NT) was performed using linear SVM for sMRI data, and logistic regression and elastic net for both genetic and environmental data. 2.5.4.1. Support vector machine classification Binary classification of transition to psychosis from an ARMS (i.e., ARMS-T vs. ARMS-NT) using sMRI data was performed using linear SVM (46, 47). In this study we exclusively used a linear kernel SVM to reduce the risk of overfitting the data (given our final sample size being relatively small). Furthermore, the linear SVM classifier has a penalty parameter C that controls the trade- off between having zero training error and allowing misclassification. Herein, a parameter search was carried out to identify the optimal C value (i.e., 2l, l [−5 : 1 : 4]) in the inner CV cycle (see Section “2.5.5. Cross-validation”). (ERS) data was performed using simple logistic regression. A threshold of 0.5 was applied to the probability of observing i.e., an ARMS-T (see Supplementary material the outcome, for more details). 2.5.4.3. Elastic net for classification transition to psychosis Binary classification of from an ARMS (i.e., ARMS-T vs. ARMS-NT) using genetic (psychosis- associated SNPs or eQTL scores of psychosis-associated genes) or environmental (environmental risk factors) data was performed using logistic regularized regression with elastic net (48) using l1 and λ values hyperparameters search to identify the optimal (regression weights shrinkage) (i.e., l1 0 : 0.1 : 1; λ 0.01 : 0.01 : 1) in the inner CV cycle (see Section “2.5.5. Cross-validation”) (for detailed description see the Supplementary material). The elastic net was implemented using the “glmnet” v4.0 R package. 2.5.5. Cross-validation Each model (trained with sMRI, psychosis-associated SNPs or eQTL scores of psychosis-associated genes and environmental risk factors) was trained in a nested-CV scheme for hyperparameter tuning (in the inner CV cycle) and to estimate the generalizability of the trained prediction model and its performance (in the outer CV cycle) (Figure 2A). For more details see the Supplementary material. For sMRI models, we tested three different nested-CV schemes: (a) leave-one scan acquisition protocol-out (LSO); (b) leave-one per group from the same scan acquisition protocol-out (LPO); and (c) classic 5-fold CV. For the remaining sMRI, genetic (trained with psychosis-associated SNPs or eQTL scores of psychosis-associated genes data) and environmental (trained with environmental risk factors data) models, nested-CV was defined with an inner 5- fold and an outer leave-one per group-out (LPO) CV schemes. Furthermore, the logistic regression (trained with genetic–PRS–and environmental–ERS–data) was trained and tested in a simple LPO CV scheme (Figure 2B). 2.5.6. Performance measures 2.5.4.2. Logistic regression for classification Binary classification of transition to psychosis from an ARMS (i.e., ARMS-T vs. ARMS-NT) using genetic (PRS) or environmental Each model’s performance was evaluated using measures derived from the confusion matrix: sensitivity; specificity; BAC; positive likelihood ratio; negative likelihood ratio; and diagnostic odds ratio Frontiers in Psychiatry 07 frontiersin.org fpsyt-13-1086038 January 13, 2023 Time: 17:35 # 8 Tavares et al. 10.3389/fpsyt.2022.1086038 (DOR). Moreover, permutation testing was used to test if the BAC was higher than chance–50%–with a statistical significance of 5% (For a detailed description of each measure see the Supplementary material). The prediction ability of each tested combination of feature type, feature manipulation, and CV scheme was defined as significant if the BAC was higher than chance–50% in at least 3 out of 5 bootstrapped samples. evaluated by testing the statistical significance of the median BAC across bootstrapped samples using a one-tailed Wilcoxon signed rank test (i.e., to test if the median BAC across bootstrapped samples is higher than chance– 50%, with a statistical significance level of 5%). P-values were not adjusted for multiple comparisons due to non-independence of the samples used in each statistical test. 3. Results Overall, the BAC of the classification models trained and validated using each combination of feature type (i.e., ROIGM, ROIWM, ROISurface, VBGM, or VBWM–for sMRI data; PRS, psychosis-associated SNPs or psychosis-associated brain eQTL score genes scores–for genetic data; or ERS or individual environmental risk factors–for environmental data), feature manipulation (i.e., feature dimensionality reduction through PCA; no feature selection; or forward feature selection), CV scheme (i.e., LSO CV; LPO CV; or 5-fold CV), and bootstrapped sample (i.e., one of the 5 samples) ranged from 37 to 67% for the classification models trained with sMRI (Tables 4, 5 and Figures 3, 4), from 26 to 62% for the models trained with genetic data (Table 6 and Figure 5) and from 38 to 61% for models trained with environmental data (Table 6 and Figure 6). The prediction ability of each model was not significant as less than 3 bootstrapped samples per each feature type showed a BAC statistically higher than chance–50%. 4. Discussion This study aimed to predict transition to psychosis from an ARMS using ML applied to quantitative data across modalities– i.e., neuroimaging (sMRI), genetics (genome-wide genotypes), and environment–collected when subjects first sought clinical help (i.e., at baseline) and were identified with an ARMS. This is, to the best of our knowledge, the first study: (1) of longitudinal design exploring sMRI, genetic and environmental data to predict the development of a psychotic disorder from a prodromal stage; and (2) when considering each modality individually, exploring a range of approaches (for genetics and environmental data) and/or feature combinations (for sMRI data). 4.1. Prediction of transition to psychosis using structural neuroimaging data In this study we applied ML to structural neuroimaging data using a relatively larger sample and an ML approach, improved to the best of our ability, to detect transition to psychosis from an ARMS and to replicate previous positive findings of accuracies 74 to 84% of six studies, which together used 3 independent samples (10–15). For this, we decided: to: (1) use only the most recent versions of the image processing tools (i.e., CAT12) and ML tools (i.e., NeuroMiner); (2) replicate as accurately as possible the methods that were described in the abovementioned MRI papers since it was not possible to access their processing and ML pipelines; (3) add a layer of ML generalizability by bootstrapping and fitting a model to each subsample; and (4) overcome previous studies’ limitations (e.g., sample unbalancing for demographics). Furthermore, we explored, for the first time, the use of whole brain white matter volume and regional white matter volume, cortical thickness, and surface-based brain gyrification, sulci depth, and complexity indexes with ML to predict transition to psychosis. Unexpectedly, we did not replicate previous findings. After balancing the samples for binary classification of transition to psychosis accounting for age, sex, and the three different scan acquisition protocols to avoid overoptimistic results, the performance of all tested combinations (i.e., of feature type–ROIGM, ROIWM, ROISurface, VBGM, or VBWM; feature manipulation–feature dimensionality reduction through PCA, no feature selection, or forward feature selection; and CV scheme–LSO CV, LPO CV, or 5-fold CV) were not significantly better than chance level. Compared to the previous studies reporting high balanced accuracies (74 to 84%) in predicting transition to psychosis from sMRI maps (10–15), the current study has some advantages. First, this study’s sample is drawn from a more naturalistic ARMS population as it includes subjects whose sMRI images were acquired using three different scan acquisition protocols. Training a classification model with data from different centers potentially increases its generalizability. Only one of the previous transition to psychosis prediction studies used a two-site group balanced sample (12), combining the samples reported in two previous studies by the same authors (10, 11). The main differences between this report and our study are the following: (a) Their sample was larger than our balanced bootstrapped samples (i.e., 36% larger than ours, measured as the absolute value of the change in sample size, divided by the average of the size of the two samples). However, we tested our ML models on five balanced subsamples (i.e., through bootstrapping), allowing us to obtain a measure of generalizability of these models’ performance. Moreover, they do not present a measure of the statistical significance of the model’s BAC, which we do herein. (b) They controlled the effect of site on the classification using partial correlations during the training phase of the CV cycle, whereas we controlled it by keeping the same proportion of subjects at an ARMS that transitioned to psychosis and those who did not in each scan protocol during the training phase of the CV cycle (i.e., when using the LPO CV scheme as the previous study did). Additionally, we also guaranteed that the pair of subjects left out for testing/validation were from the same site. This potentially increases the generalizability of the classification model by training it with a more heterogeneous sample (and, as explained above, more naturalistic) and diminishing the effect of site on the testing/validation classification accuracy, which is not taken into account in the previous report (12). Second, we trained our classification models with samples balanced for group (subjects at an ARMS who later transitioned to psychosis and who did not), age at scan and sex. Balancing for group is important to avoid biasing the classification model to the most represented group and it was not taken into account by three out of six previous reports (10, 11, 14). Moreover, the effects of age (49) and sex (50) on brain structure, rate of transition to psychosis from ARMS (2), and prevalence of psychosis (3, 51), have been consistently reported and, therefore, should be taken into Frontiers in Psychiatry 08 frontiersin.org fpsyt-13-1086038 January 13, 2023 Time: 17:35 # 9 Tavares et al. 10.3389/fpsyt.2022.1086038 TABLE 4 Performance measures of each structural magnetic resonance imaging (sMRI) classification model based on brain regional features across bootstrapped samples. ROIGM No-FS 55.7 ± 6.4 [47.8, 65.2] 55.7 ± 10.4 [43.5, 69.6] 55.7 ± 5.5 [47.8, 63.0] 1.3 ± 0.3 [0.9, 1.9] 0.8 ± 0.2 [0.6, 1.1] 1.7 ± 0.8 [0.8, 3.0] 1 47.8 ± 6.1 [43.5, 56.5] 54.8 ± 6.6 [47.8, 60.9] 51.3 ± 4.8 [45.7, 58.7] 1.1 ± 0.2 [0.8, 1.4] 1.0 ± 0.2 [0.7, 1.2] 1.2 ± 0.5 [0.7, 2] 0 42.6 ± 3.6 [39.1, 47.8] 54.8 ± 12.5 [34.8, 65.2] 48.7 ± 6.6 [39.1, 56.5] 1.0 ± 0.3 [0.7, 1.4] 1.1 ± 0.3 [0.8, 1.6] 1.0 ± 0.5 [0.4, 1.7] 0 LSO CV scheme SE (%) SP (%) BAC (%) PLR NLR DOR Significant models LPO CV scheme SE (%) SP (%) BAC (%) PLR NLR DOR Significant models 5-fold CV scheme SE (%) SP (%) BAC (%) PLR NLR DOR Significant models FFS 59.1 ± 11.7 [47.8, 78.3] 40.9 ± 10.5 [26.1, 52.2] 50.0 ± 3.8 [43.5, 52.2] 1.0 ± 0.1 [0.8, 1.1] 1.0 ± 0.2 [0.8, 1.4] 1.0 ± 0.3 [0.6, 1.3] 0 67.0 ± 7.9 [56.5, 73.9] 44.3 ± 10.8 [34.8, 60.9] 55.7 ± 7.1 [45.7, 65.2] 1.3 ± 0.3 [0.9, 1.8] 0.8 ± 0.3 [0.5, 1.3] 1.9 ± 1.1 [0.7, 3.6] 0 40.9 ± 5.8 [34.8, 47.8] 40.9 ± 7.3 [30.4, 47.8] 40.9 ± 2.4 [37.0, 43.5] 0.7 ± 0.1 [0.6, 0.8] 1.5 ± 0.2 [1.3, 1.9] 0.5 ± 0.1 [0.3, 0.6] 0 ROIWM No-FS 57.4 ± 19.1 [30.4, 82.6] 46.1 ± 14 [21.7, 56.5] 51.7 ± 5.6 [43.5, 58.7] 1.1 ± 0.2 [0.7, 1.4] 0.9 ± 0.2 [0.7, 1.2] 1.3 ± 0.5 [0.6, 2.0] 0 49.6 ± 8.5 [39.1, 60.9] 52.2 ± 5.3 [43.5, 56.5] 50.9 ± 5.2 [43.5, 56.5] 1.0 ± 0.2 [0.8, 1.3] 1.0 ± 0.2 [0.8, 1.3] 1.1 ± 0.4 [0.6, 1.7] 0 59.1 ± 6.6 [52.2, 69.6] 40.9 ± 8.5 [30.4, 52.2] 50.0 ± 5.1 [45.7, 56.5] 1.0 ± 0.2 [0.9, 1.2] 1.0 ± 0.3 [0.7, 1.3] 1.1 ± 0.5 [0.7, 1.8] 0 ROISurface FFS No-FS FFS 62.6 ± 13.3 [52.2, 82.6] 27.8 ± 22.3 [0.0, 56.5] 45.2 ± 6.6 [37.0, 54.3] 0.9 ± 0.2 [0.7, 1.2] 1.5 ± 0.7 [0.8, 2.5] 0.6 ± 0.5 [0.0, 1.4] 0 39.1 ± 9.2 [26.1, 47.8] 50.4 ± 12.5 [34.8, 69.6] 44.8 ± 6.6 [37.0, 54.3] 0.8 ± 0.3 [0.5, 1.3] 1.3 ± 0.3 [0.9, 1.5] 0.7 ± 0.4 [0.3, 1.5] 0 45.2 ± 8.5 [34.8, 56.5] 53 ± 11.3 [34.8, 65.2] 49.1 ± 2.5 [45.7, 52.2] 1.0 ± 0.1 [0.9, 1.1] 1.1 ± 0.1 [0.9, 1.3] 0.9 ± 0.2 [0.7, 1.2] 0 41.7 ± 14.6 [26.1, 60.9] 61.7 ± 17.8 [34.8, 82.6] 51.7 ± 7.4 [43.5, 63.0] 1.2 ± 0.4 [0.8, 1.8] 1.0 ± 0.3 [0.6, 1.4] 1.4 ± 0.9 [0.6, 2.9] 1 53.9 ± 6.6 [43.5, 60.9] 54.8 ± 5.0 [47.8, 60.9] 54.3 ± 4.3 [50.0, 60.9] 1.2 ± 0.2 [1.0, 1.6] 0.8 ± 0.1 [0.6, 1.0] 1.5 ± 0.6 [1.0, 2.4] 1 53.0 ± 10.4 [39.1, 65.2] 54.8 ± 6.6 [47.8, 60.9] 53.9 ± 5.2 [47.8, 60.9] 1.2 ± 0.2 [0.9, 1.6] 0.9 ± 0.2 [0.6, 1.1] 1.5 ± 0.6 [0.8, 2.4] 0 39.1 ± 20.6 [17.4, 65.2] 63.5 ± 12.1 [43.5, 73.9] 51.3 ± 9.8 [43.5, 67.4] 1.1 ± 0.6 [0.6, 2.1] 1.0 ± 0.3 [0.5, 1.2] 1.5 ± 1.6 [0.5, 4.3] 1 52.2 ± 6.9 [43.5, 60.9] 54.8 ± 12.9 [39.1, 69.6] 53.5 ± 7.5 [43.5, 60.9] 1.2 ± 0.3 [0.8, 1.6] 0.9 ± 0.3 [0.6, 1.3] 1.5 ± 0.8 [0.6, 2.4] 0 57.4 ± 8.4 [47.8, 69.6] 53 ± 11.7 [39.1, 65.2] 55.2 ± 9.3 [47.8, 67.4] 1.3 ± 0.5 [0.9, 2.0] 0.9 ± 0.3 [0.5, 1.1] 2.0 ± 1.6 [0.8, 4.3] 1 Measures for each tested combination of brain regional feature type [i.e., regional-based gray (ROIGM) and white (ROIWM) matter volume; and surface-based regional cortical thickness, gyrification, sulci, and complexity indexes (ROISurface)], feature selection [i.e., no feature selection (No-FS); and forward feature selection (FFS)], and cross-validation (CV) scheme [i.e., leave-one scan acquisition protocol-out (LSO) CV; leave-one per group-out (LPO) CV; and 5-fold CV] are presented. Statistical significance of the balanced accuracy (BAC) for each bootstrapped sample was tested using permutation testing with a significance level of 5%. Data format: mean ± standard deviation [min max]. DOR, diagnostic odds ratio; NLR, negative likelihood ratio; PLR, positive likelihood ratio; SE, sensitivity; SP, specificity. Significant models: number of models with statistically significant BAC higher than 50%. account in these studies. All previous reports (and the current study) matched transition proportion for age and sex (10–14), except for one (15). Das and colleagues reported a statistically significant and better than chance level BAC in predicting transition to psychosis using a sample unbalanced for both group and sex. Although they used a ML algorithm with class (i.e., group) weighing–which in summary increases the influence of the minority class when training the model by assigning higher weights to rare cases, the authors Frontiers in Psychiatry 09 frontiersin.org fpsyt-13-1086038 January 13, 2023 Time: 17:35 # 10 Tavares et al. 10.3389/fpsyt.2022.1086038 TABLE 5 Performance measures of each structural magnetic resonance imaging (SMRI) classification model based on voxel-wise features across bootstrapped samples. LSO CV scheme LPO CV scheme 5-fold CV scheme SE (%) SP (%) BAC (%) PLR NLR DOR Significant models VBGM 20.9 ± 34.8 [0, 82.6] 72.2 ± 35.7 [8.7, 91.3] 46.5 ± 3.6 [43.5, 52.2] 0.6 ± 0.6 [0.0, 1.5] 1.3 ± 0.4 [1.0, 2.0] 0.7 ± 0.9 [0.0, 2.2] 0 VBWM 46.1 ± 38.2 [4.3, 78.3] 53 ± 35.4 [21.7, 95.7] 49.6 ± 2.4 [45.7, 52.2] 0.9 ± 0.3 [0.3, 1.1] 1.0 ± 0.1 [0.9, 1.1] 0.8 ± 0.4 [0.1, 1.1] 1 VBGM 47.0 ± 10.4 [34.8, 60.9] 55.7 ± 8.4 [43.5, 65.2] 51.3 ± 7.5 [45.7, 63.0] 1.1 ± 0.4 [0.8, 1.8] 1.0 ± 0.3 [0.6, 1.2] 1.3 ± 1.0 [0.7, 3.1] 0 VBWM 50.4 ± 11.3 [34.8, 60.9] 53.0 ± 7.1 [47.8, 65.2] 51.7 ± 2.8 [47.8, 54.3] 1.1 ± 0.1 [0.9, 1.2] 0.9 ± 0.1 [0.8, 1.1] 1.1 ± 0.2 [0.8, 1.4] 0 VBGM 30.4 ± 10.2 [21.7, 43.5] 51.3 ± 7.8 [43.5, 60.9] 40.9 ± 2.4 [37.0, 43.5] 0.6 ± 0.1 [0.5, 0.8] 1.4 ± 0.1 [1.3, 1.5] 0.4 ± 0.1 [0.2, 0.6] 0 VBWM 41.7 ± 8.5 [34.8, 56.5] 52.2 ± 6.1 [43.5, 60.9] 47.0 ± 6.8 [41.3, 58.7] 0.9 ± 0.3 [0.7, 1.4] 1.1 ± 0.3 [0.7, 1.4] 0.9 ± 0.7 [0.5, 2.1] 0 Measures for each tested combination of voxel-wise feature type [i.e., voxel-based gray (VBGM) and white (VBWM) matter volume maps], feature dimensionality reduction through principal component analysis and cross-validation (CV) scheme [i.e., leave-one scan acquisition protocol-out (LSO) CV; leave-one per group-out (LPO) CV; and 5-fold CV] are presented. Statistical significance of the balanced accuracy (BAC) for each bootstrapped sample was tested using permutation testing with a significance level of 5%. Data format: mean ± standard deviation [min max]. DOR, diagnostic odds ratio; NLR, negative likelihood ratio; PLR, positive likelihood ratio; SE, sensitivity; SP, specificity. Significant models: number of models with statistically significant BAC higher than 50%. FIGURE 3 Balanced accuracy across bootstrapped samples for each tested combination of regional feature type [i.e., regional-based gray and white matter volume; and surface-based regional cortical thickness, gyrification, sulci, and complexity indexes (surface-based regional measures)], feature selection [i.e., no feature selection; and forward feature selection (FFS)], and cross-validation (CV) scheme [i.e., leave-one scan acquisition protocol-out (LSO) CV; leave-one per group-out (LPO) CV; and 5-fold CV]. Dots represent the balanced accuracy value in each of the five bootstrapped samples and are red colored if the balanced accuracy is statistically significant (i.e., p < 0.05) or blue colored if it is not (i.e., p > 0.05). The statistical significance of the balanced accuracy in each bootstrapped sample was evaluated through permutation testing. performed an unspecified correction for sex effect (as well as for age and TIV effects) to the data during the training CV cycle. This approach may not be the most appropriate given the known effect of sex on brain structure (50) and the, abovementioned, empirically tested association between sex and group (i.e., transition to psychosis from an ARMS vs. no transition) (15), which makes sex a potential confounder in this analysis. Furthermore, in three of the six previous reports, the effects of age and sex were corrected before entering the ML analysis (10), and during the training CV cycle (11, 15) using partial correlations (10, 11) or an unspecified method (15)– which we did not perform. Correction for age effects in ML analysis has been previously shown to increase classification accuracy in Alzheimer’s disease, when it is estimated from healthy subjects (52). Correction for effects of no interest in ML analyses should be done with extreme caution as it can easily remove relevant subject-specific information (53). This is especially important when the correction Frontiers in Psychiatry 10 frontiersin.org fpsyt-13-1086038 January 13, 2023 Time: 17:35 # 11 Tavares et al. 10.3389/fpsyt.2022.1086038 FIGURE 4 Balanced accuracy across bootstrapped samples for each tested combination of voxel-wise feature type [i.e., voxel-based gray (VBGM) and white (VBWM) matter volume maps], feature dimensionality reduction through principal component analysis and cross-validation (CV) scheme [i.e., leave-one scan acquisition protocol-out (LSO) CV; leave-one per group-out (LPO) CV; and 5-fold CV. Dots represent the balanced accuracy value in each of the five bootstrapped samples and are red colored if the balanced accuracy is statistically significant (i.e., p < 0.05) or blue colored if it is not (i.e., p > 0.05). The statistical significance of the balanced accuracy in each bootstrapped sample was evaluated through permutation testing. TABLE 6 Performance measures of: (1) a genetic schizophrenia polygenic risk score (PRS), (2) a list of psychosis-associated single nucleotide polymorphisms (SNPs), (3) expression quantitative trait loci (eQTL) scores (43) of a list of psychosis-associated genes expressed in the brain; (4) an environmental schizophrenia risk score (ERS), and (5) a list of schizophrenia-associated environmental risk factors, classification models across bootstrapped samples. PRS SNP eQTL score ERS Environmental risk factors SE (%) SP (%) BAC (%) PLR NLR DOR 42.1 ± 20.0 [21.1, 63.2] 46.3 ± 11.4 [31.6, 57.9] 44.2 ± 15.3 [26.3, 60.5] 0.9 ± 0.5 [0.3, 1.5] 1.4 ± 0.8 [63.6, 2.5] 1.0 ± 1.0 [0.1, 2.4] Significant models 0 41.9 ± 13.6 [23.8, 61.9] 50.5 ± 16.4 [28.6, 66.7] 46.2 ± 10.7 [33.3, 61.9] 0.9 ± 0.4 [0.5, 1.6] 1.3 ± 0.6 [0.6, 2.2] 1.0 ± 1.0 [0.2, 2.6] 0 61.0 ± 17.0 [47.6, 85.7] 31.4 ± 23.2 [4.8, 57.1] 46.2 ± 4.9 [40.5, 52.4] 0.9 ± 0.1 [0.8, 1.1] 1.9 ± 1.1 [0.9, 3.0] 0.7 ± 0.4 [0.3, 1.2] 1 44.9 ± 5.1 [29.7, 56.8] 50.8 ± 8.2 [45.9, 64.9] 47.8 ± 8.8 [37.8, 60.8] 1.0 ± 0.4 [0.6, 1.6] 1.1 ± 0.3 [0.7, 1.5] 1.0 ± 0.8 [0.4, 2.4] 0 10.6 ± 4.9 [5.9, 17.6] 70.6 ± 7.2 [64.7, 82.4] 40.6 ± 2.5 [38.2, 44.1] 0.4 ± 0.1 [0.2, 0.5] 1.3 ± 0.1 [1.1, 1.4] 0.3 ± 0.1 [0.2, 0.4] 0 Statistical significance of the balanced accuracy (BAC) for each bootstrapped sample was tested using permutation testing with a significance level of 5%. Data format: mean ± standard deviation [min max]. DOR, diagnostic odds ratio; NLR, negative likelihood ratio; PLR, positive likelihood ratio; SE, sensitivity; SP, specificity. Significant models: number of models with statistically significant BAC higher than 50%. is being performed in a non-healthy (i.e., non-standard) population, because the effect of external variables such as age and sex might be modulated by the presence of the disease (e.g., being at ARMS or having schizophrenia). Third, this study’s sample is composed of subjects whose clinical diagnosis of an ARMS was based on having a schizotypal personality disorder or on the subject’s familial-high risk coupled with functioning decline and on the CAARMS (54), which mainly evaluates positive symptoms. These were not the same criteria as those used in the previous studies predicting transition to psychosis from an ARMS. These previous studies all used samples of subjects clinically assessed with tools that evaluate not only positive symptoms, but also basic and negative symptoms (10–12, 14, 15), except one (13), which included only familial-high risk subjects in its sample. This potentially increases the inclusion of subjects in the early phase of the psychosis prodrome (characterized by the presence of basic and negative symptoms), whereas our sample includes mainly subjects in the late prodromal phase of psychosis (characterized mainly by the presence of positive symptoms) (2). Therefore, our results suggest that previously reported accuracies in predicting transition to psychosis may be population-specific, poorly generalizable to differently clinically characterized populations (as ours herein). 4.2. Prediction of transition to psychosis using genetic data In this study we applied ML to genetic data and used three types of genetic features to detect transition to psychosis from an ARMS: (a) a schizophrenia PRS that we have previously shown to distinguish Frontiers in Psychiatry 11 frontiersin.org fpsyt-13-1086038 January 13, 2023 Time: 17:35 # 12 Tavares et al. 10.3389/fpsyt.2022.1086038 FIGURE 5 Balanced accuracy across bootstrapped samples for each model trained with the polygenic risk score, the list of psychosis-associated single nucleotide polymorphism (SNPs) or with the list of psychosis-associated genes for which an expression quantitative trait loci (eQTL) score was extracted. Dots represent the balanced accuracy value in each of the 5 bootstrapped samples and are red colored if the balanced accuracy is statistically significant (i.e., p < 0.05) or blue colored if it is not (i.e., p > 0.05). The statistical significance of the balanced accuracy in each bootstrapped sample was evaluated through permutation testing. FIGURE 6 Balanced accuracy across bootstrapped samples for each model trained with the environmental risk score or with each environmental risk factors as features. Dots represent the balanced accuracy value in each of the 5 bootstrapped samples and are red colored if the balanced accuracy is statistically significant (i.e., p < 0.05) or blue colored if it is not (i.e., p > 0.05). The statistical significance of the balanced accuracy in each bootstrapped sample was evaluated through permutation testing. FEP patients from healthy controls (26) and ARMS-T from ARMS- NT (16), (b) a set of psychosis-associated SNPs previously associated with schizophrenia in a recent GWAS meta-analysis (27), and (c) a brain-specific expression Quantitative Trait Loci (eQTL) score including the latter genes. Genetic data showed a poor performance in predicting transition to psychosis from an ARMS. SNPs-based classification models have been previously shown to classify schizophrenia (18, 19, 21), and FEP patients (23) (vs. healthy controls) better than chance level, but not subjects at an ARMS vs. healthy controls or FEP patients (23). Furthermore, one of these studies has selected a list of SNPs from the Psychiatric Genomics Consortium 2 (PGC2) (21, 42), which potentially overlaps with the ones selected in this study (27). Despite the (scarce) evidence of the potential of PRS for schizophrenia (20–22) to classify schizophrenia patients (vs. healthy controls) and the one report showing the schizophrenia PRS’s ability to predict transition to psychosis (16) we were not able to predict transition to psychosis from an ARMS using this type of genetic feature. Although the latter study (16) used a larger sample (i.e., 106% higher than ours, measured as the absolute value of the change in sample size, divided by the average of the size of the two samples) to train the PRS-based model, sample balancing in terms of group and age or sex were not taken into account or that was unclear, respectively. Furthermore, herein we applied a bootstrapped sample approach to estimate generalizability of the PRS-based model by assuring that each bootstrapped sample met the balancing conditions for group, age, and sex–which does not seem to be the case in that study (16). Furthermore, another possible explanation for the PRS negative results is that although the genetic architecture, conveyed through a PRS, has been shown to differ between patients with schizophrenia and healthy controls, one cannot exclude the possibility that it is specific to schizophrenia (a fully developed psychotic disorder), and might even be present in all subjects at an ARMS, i.e., those who later transition to psychosis and those who do not. The constellation of genetic variations (i.e., SNPs) that might confer susceptibility to transition to psychosis already from a prodromal stage is not necessarily the same as the one for schizophrenia (when drawn in comparison to healthy controls). This Frontiers in Psychiatry 12 frontiersin.org fpsyt-13-1086038 January 13, 2023 Time: 17:35 # 13 Tavares et al. 10.3389/fpsyt.2022.1086038 may justify the advantage of using a less hypothesis-based approach for the selection of genetic features (as we did by pre-selecting a large list of SNPs and performing an embedded feature selection using elastic net regression). Lastly, using a PRS formula made specifically for transition to psychosis from an ARMS would require a larger and independent sample to estimate SNP effect sizes, which might be better provided by multicenter projects, such as NAPLS 2 (55) and PRONIA4 over the next years. Expression Quantitative Trait Loci (eQTL) scores for psychosis associated genes expressed in the brain were also not able to predict transition to psychosis from an ARMS. Only one previous study has shown the predictive value of gene expression profiling in the frontal brain region in classifying schizophrenia patients (vs. healthy controls) (17). In the present study, instead of actual gene expression measures we used a proxy for a-genetically regulated component of the expression of genes, the eQTL scores. Although we have computed eQTL scores only for the genes having a validated eQTL score model (43), this does not guarantee that the estimated gene expression represents (or correlates perfectly with) the real levels of the expression. Furthermore, although we have selected the initial list of genes as the ones most associated with schizophrenia (vs. healthy controls), this selection did not take into account the expression profile of these genes in the brain, and we have computed an eQTL score for several brain tissues. A future improvement of this step would be to test an eQTL scores-based model with a selection of genes that: (a) are highly expressed in the brain in healthy subjects, and (b) their expression is associated to a schizophrenia diagnosis, or even better with the transition to psychosis from an ARMS. 4.3. Prediction of transition to psychosis using environmental data In this study we applied, for the first time, ML to environmental data using two types of features to detect transition to psychosis from an ARMS: (a) a schizophrenia ERS which we have previously reported (28), and (b) a set of environmental risk factors as predictors. Overall, neither environmental risk assessment, could predict transition to psychosis from an ARMS with an averaged accuracy, i.e., across bootstrapped samples, better than chance level. Although we know of no similar longitudinal ARMS transition study, the closest other report using ML and environmental data to diagnose schizophrenia (vs. healthy controls) (22) also found a BAC not statistically better than chance level, even having included features such as the presence of obstetric complications and of developmental anomalies, the parental socio-economic status; and –without feature selection– trained and tested the model in a 13 times larger, albeit age, sex, and group -unbalanced, sample (103 patients and 337 controls) than ours (22). However, due to the still poorly understood environmental risk mechanisms one cannot exclude the lack of statistical power as a potential explanation for these negative findings including ours. The ML model trained with the ERS for schizophrenia, which we have tested as an (admittedly limited) exploratory predictor of the transition to psychosis from an ARMS, showed a poor performance, i.e., a BAC similar to chance level. Indeed, ERS is a composite score of individual risk factors computed under the assumption that the risk factors are completely independent (28), which has been shown 4 http://pronia.eu not to be the case (56)–i.e., intercorrelated risk factors may inflate the ERS estimation. This crude approach may limit the ability of the ERS to capture the detailed environmental architecture underlying psychosis. Moreover, just as for a PRS, an ERS for schizophrenia may not be a good substitute of a potential ERS for transition to psychosis from an ARMS (57). Lastly, our criterion for training and testing a fully multimodal ML model with modalities that would show an ML model performance statistically better than chance (i.e., 50%) predicting transition to psychosis from an ARMS in at least 3 of 5 bootstrapped samples was not fulfilled given that none of the modality-based ML models survived that threshold. This conservative criterion was chosen given the already small sample size available for the training of the multimodal ML model, i.e., only 6 ARMS-T and 23 ARMS-NT (only this subset of subjects had data for the three data modalities, simultaneously). The decrease in sample size, remarkably impairs the prediction power of the model, i.e., its accuracy. Without previous evidence of the ability to predict transition to psychosis from an ARMS by modality supporting its integration in a multimodal ML model, negative results from this multimodal model would be highly difficult to explain, as they could theoretically be explained by the increase of noise in the model due to the inclusion of features that did show previous predictive ability or by the lack of predictive power due to the very small sample size. Moreover, the parallel-to-ours, multi-site study, albeit very group-unbalanced (only 26 ARMS-T patients vs. 308 ARMS-NT), from the PRONIA project, showed that a stacked model combining similar data to our study’s plus human prognostic ratings could predict transition to psychosis with a balanced accuracy of 86% and a good geographical generalizability (25). This multimodal approach was showed to improve biological-based unimodal models by 15% (VBGM volume maps-based model) and 20% (PRS for schizophrenia-based model). As such, the replication of this promising finding, following the same multimodal approach as that study, using in our study’s sample and data features co-existing in both samples, would be interesting as an additional method to ascertain whether our negative findings are due to lack of power or to no discriminability with our feature sets. 4.4. Limitations This study was limited by several factors. First, and foremost, the small sample size may have limited the performance of classification models, even though our sample size was informed by previous ML studies showing 74–84% accuracies in predicting transition to psychosis from an ARMS (10–15). Indeed, this is a critical limitation when dealing with high dimensional data, such as neuroimaging and genetics–which we have used herein. Although we have taken measures to avoid overfitting and an overestimation of the classification models’ performance such as artificially increasing the sampling through bootstrapping and employing CV strategies, this might not be enough to overcome this limitation. Indeed, our complementary analysis comparing the models’ training and testing performance (results in the Supplementary material) is indicative that some of the tested classification models (mainly trained with neuroimaging or with SNPs) might suffer from some degree of overfitting. Ultimately, we cannot determine whether our negative findings were due to lack of power to obtain a good performance or due to a true lack of association between the predictors and the transition to psychosis from an ARMS (and hence inflated findings Frontiers in Psychiatry 13 frontiersin.org fpsyt-13-1086038 January 13, 2023 Time: 17:35 # 14 Tavares et al. 10.3389/fpsyt.2022.1086038 from previous studies). This is one of the reasons why replication studies in independent datasets are essential in ML literature. As a final note, a power analysis for this study design would have been the most informative way to define the sample size needed to achieve an accuracy in predicting transition to psychosis from an ARMS better than chance level. However, this is not a trivial task in ML analysis and there is no established method to perform this analysis as there is for univariate analysis [for examples of studies exploring innovative ways of computing sample size for classification problems see Refs. (58, 59)] and, therefore, it was not performed. Second, in order to dilute possible confounding effects in the developed classification models we have restricted the samples used to train the models to: (a) be class-balanced, i.e., with the same number of ARMS-T and ARMS-NT subjects; (b) be matched for age, sex, scanning acquisition protocols for neuroimaging data; (c) include subjects with European ancestry only for genetic data; and (d) limit the proportion of missing data for the environment data. Although this has artificially homogenized the study sample thus avoiding the presence of statistical confounders, it has deemed the sample to be less representative of the ARMS population. Third, overall, the findings of this study are only valid to young help-seeking individuals, i.e., that are clinically screened for ARMS criteria, and whose ARMS diagnosis was based on having a schizotypal personality disorder or on the subject’s familial-high risk coupled with functioning decline and on the CAARMS (54), which mainly evaluates positive symptoms. 5. Conclusion and future directions In this study, we explored the value of using exclusively quantitative and multimodal data (i.e., as predictors) to predict transition to psychosis from an ARMS. Overall, we found that, contrary to what has been previously reported, sMRI could not predict transition to psychosis from an ARMS. We have employed several ML strategies aiming to replicate the highly promising previous positive sMRI findings (74–84%) (10–15). This is even though our sample was larger than four of the above 6 studies (10, 11, 13, 14), respectively (Conversely, our sample was smaller than two of the above studies [Das et al. (15); Koutsouleris et al. (12), respectively]. This points to the need for a cautious interpretation of small sample size studies. Also, we could not replicate the one previous evidence of the value of the schizophrenia PRS in predicting transition to psychosis. Moreover, and to the best of our knowledge, we explored for the first time the value of environment in the prediction of psychosis already from a prodromal stage. Lastly, the genetic and the environmental data used could not predict transition to psychosis from an ARMS. In summary, the present study should serve as a call for caution and skepticism regarding the currently achievable prognostic and diagnostic biomarker development goals, with the existing modeling tools and data measurement tools. Additionally, our study’s methodological approaches tailored to each data modality, may serve as suggestive proofs-of-concept for the exploration of future multimodal datasets, either for novel discovery or replication of previous promising findings, across psychiatric disorders, not exclusive to ARMS. We further suggest larger samples (in the several hundreds) should be employed for both model training and testing, given the inherent high data dimensionality (specially of neuroimaging and genetics) and the little established relevance of individual features. Although still heterogeneity in phenotypic measurements is increased in larger they bring not only statistical power but ecological samples, generalizability, and thus carry a higher potential to be clinically useful. This is best achieved with consortia multi-center studies which are increasingly common albeit not without challenges (60). Alternatively, methods for synthetic generation of data such as the Generative Adversarial Networks (GAN)-based are also a promising avenue for sample size augmentation, now starting to be applied in the clinical research field (61). Last, but not least, we recommend the use of objective and quantitative criteria-based tools for the assessment of a ML biomarker’s clinical applicability, once high effect size and accuracy estimates are achieved, such as one we have previously proposed (62). Data availability statement The datasets presented in this article are not readily available include public data sharing. should be directed to the because ethics approval did not Requests the datasets corresponding author. to access Ethics statement The studies involving human participants were reviewed and approved by NHS South East London Research Ethics Committee. The patients/participants provided their written informed consent to participate in this study. Author contributions VT ran most data preprocessing and statistical analyses and drafted the manuscript. EV coordinated genotyping and advised the genetic and environmental data analysis. AM and HF provided advise on imaging data processing and machine learning analysis. JS and IV collected imaging data. GB provided advise on imaging data processing. DP collected genetic and environmental data, co- designed the study, ran preliminary data preprocessing and machine learning analyses, and supervised the study. All authors revised the manuscript and agreed with its final version. Funding This study, VT received support from Fundação para a Ciência e a Tecnologia (FCT) Ph.D. fellowship PD/BD/114460/2016 and DSAIPA/DS/0065/2018 grants; DP received primary support from National Institute for Health Research (NIHR) PDF-2010-03-047 grant, and additionally from FCT FCT-IF/00787/2014, LISBOA- 01–0145-FEDER-030907, and DSAIPA/DS/0065/2018 grants, and a European Commission (EC) Marie Curie Career Integration Grant (FP7-PEOPLE-2013-CIG 631952). EV was part-funded by the NIHR Maudsley Biomedical Research Centre at South London and Maudsley NHS Foundation Trust and King’s College London. IV was supported by EC’s Horizon 2020 Marie Skłodowska-Curie grant (Ref. 754550, project BITRECS) and “La Caixa” Foundation (LCF/PR/GN18/50310006). Frontiers in Psychiatry 14 frontiersin.org fpsyt-13-1086038 January 13, 2023 Time: 17:35 # 15 Tavares et al. 10.3389/fpsyt.2022.1086038 Acknowledgments We thank Prof. Philip McGuire for his invaluable guidance during data design and collection, the OASIS team, and all volunteers with an ARMS who made this study possible. organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher. Conflict of interest Author disclaimer The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. The reviewer SV declared a shared affiliation with the authors EV, IV, GB, and DP to the handling editor at the time of review. Publisher’s note The views expressed were those of the authors and not necessarily those of any of the above sponsors. Supplementary material All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fpsyt.2022.1086038/ full#supplementary-material References 1. Schrimpf L, Aggarwal A, Lauriello J. Psychosis. Contin Lifelong Learn Neurol. (2018) 24:845–60. doi: 10.1212/CON.0000000000000602 2. Fusar-Poli P, Borgwardt S, Bechdolf A, Addington J, Riecher-Rössler A, Schultze- Lutter F, et al. The psychosis high-risk state. JAMA Psychiatry. (2013) 70:107. doi: 10.1001/ jamapsychiatry.2013.269 3. Kahn R, Sommer I, Murray R, Meyer-Lindenberg A, Weinberger D, Cannon T, et al. Schizophrenia. Nat Rev Dis Prim. (2015) 1:15067. doi: 10.1038/nrdp.2015.67 4. Schultze-Lutter F, Ruhrmann S, Fusar-Poli P, Bechdolf AG, Schimmelmann B, Klosterkotter J. Basic symptoms and the prediction of first-episode psychosis. Curr Pharm Des. (2012) 18:351–7. doi: 10.2174/138161212799316064 5. Correll C, Schooler N. Negative symptoms in schizophrenia: A review and clinical guide for recognition, assessment, and treatment. Neuropsychiatr Dis Treat. (2020) 16:519–34. doi: 10.2147/NDT.S225643 6. Sheffield J, Karcher N, Barch D. Cognitive deficits in psychotic disorders: A lifespan perspective. Neuropsychol Rev. (2018) 28:509–33. doi: 10.1007/s11065-018-9388-2 7. Yung A, McGorry P, McFarlane C, Jackson H, Patton G, Rakkar A. Monitoring and care of young people at incipient risk of psychosis. Schizophr Bull. (1996) 22:283–303. doi: 10.1093/schbul/22.2.283 8. Salazar De Pablo G, Radua J, Pereira J, Bonoldi I, Arienti V, Besana F, et al. Probability of transition to psychosis in individuals at clinical high risk: An updated meta-analysis. JAMA Psychiatry. (2021) 78:970–8. doi: 10.1001/jamapsychiatry.2021.0830 9. Fusar-Poli P, Bechdolf A, Taylor M, Bonoldi I, Carpenter W, Yung A, et al. At risk for schizophrenic or affective psychoses? A meta-analysis of DSM/ICD diagnostic outcomes in individuals at high clinical risk. Schizophr Bull. (2013) 39:923–32. doi: 10.1093/schbul/ sbs060 10. Koutsouleris N, Meisenzahl E, Davatzikos C, Bottlender R, Frodl T, Scheuerecker J, et al. Use of neuroanatomical pattern classification to identify subjects in at-risk mental states of psychosis and predict disease transition. Arch Gen Psychiatry. (2009) 66:700–12. doi: 10.1001/archgenpsychiatry.2009.62 11. Koutsouleris N, Borgwardt S, Meisenzahl E, Bottlender R, Möller H, Riecher-Rössler A. Disease prediction in the at-risk mental state for psychosis using neuroanatomical biomarkers: Results from the fepsy study. Schizophr Bull. (2012) 38:1234–46. doi: 10.1093/ schbul/sbr145 12. Koutsouleris N, Riecher-Rössler A, Meisenzahl E, Smieskova R, Studerus E, Kambeitz-Ilankovic L, et al. Detecting the psychosis prodrome across high-risk populations using neuroanatomical biomarkers. Schizophr Bull. (2015) 41:471–82. doi: 10.1093/schbul/sbu078 13. Zarogianni E, Storkey A, Johnstone E, Owens D, Lawrie S. Improved individualized prediction of schizophrenia in subjects at familial high risk, based on neuroanatomical data, schizotypal and neurocognitive features. Schizophr Res. (2017) 181:6–12. doi: 10. 1016/j.schres.2016.08.027 15. Das T, Borgwardt S, Hauke D, Harrisberger F, Lang U, Riecher-Rössler A, et al. Disorganized gyrification network properties during the transition to psychosis. JAMA Psychiatry. (2018) 75:613–22. doi: 10.1001/jamapsychiatry.2018.0391 16. Perkins D, Loohuis L, Barbee J, Ford J, Jeffries C, Addington J, et al. Polygenic risk score contribution to psychosis prediction in a target population of persons at clinical high risk. Am J Psychiatry. (2020) 177:155–63. doi: 10.1176/appi.ajp.2019.18060721 17. Struyf J, Dobrin S, Page D. Combining gene expression, demographic and clinical data in modeling disease: A case study of bipolar disorder and schizophrenia. BMC Genomics. (2008) 9:531. doi: 10.1186/1471-2164-9-531 18. Yang H, Liu J, Sui J, Pearlson G, Calhoun V. A hybrid machine learning method for fusing fMRI and genetic data: Combining both improves classification of schizophrenia. Front Hum Neurosci. (2010) 4:192. doi: 10.3389/fnhum.2010.00192 19. Aguiar-Pulido V, Seoane J, Rabuñal J, Dorado J, Pazos A, Munteanu C. Machine learning techniques for single nucleotide polymorphism - disease classification models in schizophrenia. Molecules. (2010) 15:4875–89. doi: 10.3390/molecules1507 4875 20. Chen J, Wu J, Mize T, Shui D, Chen X. Prediction of schizophrenia diagnosis by integration of genetically correlated conditions and traits. J Neuroimmune Pharmacol. (2018) 13:532–40. doi: 10.1007/s11481-018-9811-8 21. Vivian-Griffiths T, Baker E, Schmidt K, Bracher-Smith M, Walters J, Artemiou A, et al. Predictive modeling of schizophrenia from genomic data: Comparison of polygenic risk score with kernel support vector machines approach. Am J Med Genet Part B Neuropsychiatr Genet. (2019) 180:80–5. doi: 10.1002/ajmg.b.32705 22. Antonucci L, Pergola G, Pigoni A, Dwyer D, Kambeitz-Ilankovic L, Penzel N, et al. A pattern of cognitive deficits stratified for genetic and environmental risk reliably classifies patients with schizophrenia from healthy control subjects. Biol Psychiatry. (2020) 87:697–707. doi: 10.1016/j.biopsych.2019.11.007 23. Pettersson-Yeo W, Benetti S, Marquand A, Dell’Acqua F, Williams S, Allen P, et al. Using genetic, cognitive and multi-modal neuroimaging data to identify ultra-high-risk and first-episode psychosis at the individual level. Psychol Med. (2013) 43:2547–62. doi: 10.1017/S003329171300024X 24. Fromer M, Roussos P, Sieberts S, Johnson J, Kavanagh D, Perumal T, et al. Gene expression elucidates functional impact of polygenic risk for schizophrenia. Nat Neurosci. (2016) 19:1442–53. 25. Koutsouleris N, Dwyer D, Degenhardt F, Maj C, Urquijo-Castro M, Sanfelici R, et al. Multimodal machine learning workflows for prediction of psychosis in patients with clinical high-risk syndromes and recent-onset depression. JAMA Psychiatry. (2021) 78:195–209. 26. Vassos E, Di Forti M, Coleman J, Iyegbe C, Prata D, Euesden J, et al. An examination of polygenic score risk prediction in individuals with first-episode psychosis. Biol Psychiatry. (2017) 81:470–7. doi: 10.1016/j.biopsych.2016.06.028 14. Zarogianni E, Storkey A, Borgwardt S, Smieskova R, Studerus E, Riecher-Rössler A, et al. Individualized prediction of psychosis in subjects with an at-risk mental state. Schizophr Res. (2019) 214:18–23. doi: 10.1016/j.schres.2017.08.061 27. Pardiñas A, Holmans P, Pocklington A, Escott-Price V, Ripke S, Carrera N, et al. Common schizophrenia alleles are enriched in mutation-intolerant genes and in regions under strong background selection. Nat Genet. (2018) 50:381–9. Frontiers in Psychiatry 15 frontiersin.org fpsyt-13-1086038 January 13, 2023 Time: 17:35 # 16 Tavares et al. 10.3389/fpsyt.2022.1086038 28. Vassos E, Sham P, Kempton M, Trotta A, Stilo S, Gayer-Anderson C, et al. The Maudsley environmental risk score for psychosis. Psychol Med. (2020) 50:2213–20. doi: 10.1017/S0033291719002319 45. Hubert M, Rousseeuw P, Vanden Branden K. ROBPCA: A new approach to robust principal component analysis. Technometrics. (2005) 47:64–79. doi: 10.1198/ 004017004000000563 29. Broome M, Woolley J, Johns L, Valmaggia L, Tabraham P, Gafoor R, et al. Outreach and support in south London (OASIS): implementation of a clinical service for prodromal psychosis and the at risk mental state. Eur Psychiatry. (2005) 20:372–8. doi: 10.1016/j. eurpsy.2005.03.001 30. Phillips L, Yung A, McGorry P. Identification of young people at risk of psychosis: validation of Personal Assessment and Crisis Evaluation Clinic intake criteria. Aust N Z J Psychiatry. (2000) 34Suppl:S164–9. doi: 10.1046/j.1440-1614.2000.00798.x 31. Nelson H. The National Adult Reading Test (NART): Test Manual. Wind UK NFER-Nelson. (1982) 124:0–25. 32. American Psychiatric Association. Diagnostic and Statistical Manual of Mental Disorders, Text Revision (DSM-IV-TR). Fourth ed. (Vol. 1). Arlington, VA: American Psychiatric Association (2000). 33. Murray R, Mondelli V, Stilo S, Trotta A, Sideli L, Ajnakina O, et al. The influence of risk factors on the onset and outcome of psychosis: What we learned from the GAP study. Schizophr Res. (2020) 225:63–8. doi: 10.1016/j.schres.2020.01.011 34. Tavares V, Prata D, Ferreira H. Comparing SPM12 and CAT12 segmentation pipelines: a brain tissue volume-based age and Alzheimer’s disease study. J Neurosci Methods. (2020) 334:108565. doi: 10.1016/j.jneumeth.2019.108565 35. Hammers A, Allom R, Koepp M, Free S, Myers R, Lemieux L, et al. Three- dimensional maximum probability atlas of the human brain, with particular reference to the temporal lobe. Hum Brain Mapp. (2003) 19:224–47. doi: 10.1002/hbm.10123 36. Desikan R, Ségonne F, Fischl B, Quinn B, Dickerson B, Blacker D, et al. An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. Neuroimage. (2006) 31:968–80. doi: 10.1016/j.neuroimage. 2006.01.021 37. Dahnke, R, Ziegler G, Grosskreutz J, Gaser C. Retrospective Quality Assurance of MR Images. Seattle, WA: Human Brain Mapping (2013). Availale online at: http://dbm.neuro. uni-jena.de/HBM2013/Dahnke01.pdf 38. Bramon E, Pirinen M, Strange A, Lin K, Freeman C, Bellenguez C, et al. A genome-wide association analysis of a broad psychosis phenotype identifies three loci for further investigation. Biol Psychiatry. (2014) 75:386–97. doi: 10.1016/j.biopsych.2013.03 .033 39. Kambeitz-Ilankovic L, Meisenzahl E, Cabral C, von Saldern S, Kambeitz J, Falkai P, et al. Prediction of outcome in the psychosis prodrome using neuroanatomical pattern classification. Schizophr Res. (2016) 173:159–65. doi: 10.1016/j.schres.2015.03 .005 40. Koutsouleris N, Kambeitz-Ilankovic L, Ruhrmann S, Rosen M, Ruef A, Dwyer D, et al. Prediction models of functional outcomes for individuals in the clinical high-risk state for psychosis or with recent-onset depression: A multimodal, multisite machine learning analysis. JAMA Psychiatry. (2018) 75:1156–72. doi: 10.1001/jamapsychiatry.2018. 2165 41. R Core Team. R: A Language and Environment for Statistical Computing. Vienna: R Foundation for Statistical Computing (2014). 42. Ripke S, Neale B, Corvin A, Walters J, Farh K, Holmans P, et al. Biological insights from 108 schizophrenia-associated genetic loci. Nature. (2014) 511:421–7. doi: 10.1038/ nature13595 43. Tavares V, Monteiro J, Vassos E, Coleman J, Prata D. Evaluation of genotype-based gene expression model performance: A cross-framework and cross-dataset study. Genes (Basel). (2021) 12:1531. doi: 10.3390/genes12101531 44. Hubert M, Rousseeuw P, Verdonck T. Robust PCA for skewed data and its outlier map. Comput Stat Data Anal. (2009) 53:2264–74. doi: 10.1016/j.csda.2008.05.027 46. Cortes C, Vapnik V. Support-vector networks. Mach Learn. (1995) 20:273–97. doi: 10.1007/BF00994018 47. Burges C. A tutorial on support vector machines for pattern recognition. Data Min Knowl Discov. (1998) 2:121–67. 48. Zou H, Hastie T. Regularization and variable selection via the elastic net. J R Stat Soc Ser B Stat Methodol. (2005) 67:301–20. doi: 10.1111/j.1467-9868.2005.00503.x 49. Fjell A, Walhovd K. Structural brain changes in aging: courses, causes and cognitive consequences. Rev Neurosci. (2010) 21:187–221. doi: 10.1515/REVNEURO.2010.21.3.187 50. Ruigrok A, Salimi-Khorshidi G, Lai M, Baron-Cohen S, Lombardo MV, Tait R, et al. A meta-analysis of sex differences in human brain structure. Neurosci Biobehav Rev. (2014) 39:34–50. doi: 10.1016/j.neubiorev.2013.12.004 51. Castillejos M, Martín-Pérez C, Moreno-Küstner B. Incidence of psychotic disorders and its association with methodological issues. A systematic review and meta-analyses. Schizophr Res. (2019) 204:458–9. doi: 10.1016/j.schres.2018.07.031 52. Falahati F, Ferreira D, Soininen H, Mecocci P, Vellas B, Tsolaki M, et al. The effect of age correction on multivariate classification in Alzheimer’s Disease, with a focus on the characteristics of incorrectly and correctly classified subjects. Brain Topogr. (2016) 29:296–307. doi: 10.1007/s10548-015-0455-1 53. Wachinger C, Rieckmann A, Pölsterl S. Detect and correct bias in multi-site neuroimaging datasets. Med Image Anal. (2021) 67:101879. doi: 10.1016/j.media.2020. 101879 54. Yung A, Yung A, Pan Yuen H, Mcgorry P, Phillips L, Kelly D, et al. Mapping the onset of psychosis: The comprehensive assessment of at-risk mental states. Aust New Zeal J Psychiatry. (2005) 39:964–71. doi: 10.1080/j.1440-1614.2005.01714.x 55. Addington J, Cadenhead K, Cornblatt B, Mathalon D, McGlashan T, Perkins D, et al. North American Prodrome Longitudinal Study (NAPLS 2): Overview and recruitment. Schizophr Res. (2012) 142:77–82. doi: 10.1016/j.schres.2012.09.012 56. Guloksuz S, Rutten B, Pries L, Ten Have M, De Graaf R, Van Dorsselaer S, et al. The complexities of evaluating the exposome in psychiatry: A data-driven illustration of challenges and some propositions for amendments. Schizophr Bull. (2018) 44:1175–9. doi: 10.1093/schbul/sby118 57. Padmanabhan J, Shah J, Tandon N, Keshavan M. The “polyenviromic risk score”: Aggregating environmental risk factors predicts conversion to psychosis in familial high-risk subjects. Schizophr Res. (2017) 181:17–22. doi: 10.1016/j.schres.2016. 10.014 58. Figueroa R, Zeng-Treitler Q, Kandula S, Ngo L. Predicting sample size required for classification performance. BMC Med Inform Decis Mak. (2012) 12:8. doi: 10.1186/1472- 6947-12-8 59. Dobbin K, Simon R. Sample size planning for developing classifiers using high-dimensional DNA microarray data. Biostatistics. (2007) 8:101–17. doi: 10.1093/ biostatistics/kxj036 60. Tognin S, Van Hell H, Merritt K, Winter-Van Rossum I, Bossong M, Kempton M, et al. Towards precision medicine in psychosis: Benefits and challenges of multimodal multicenter studies - PSYSCAN: Translating neuroimaging findings from research into clinical practice. Schizophr Bull. (2020) 46:432–41. 61. Laino M, Cancian P, Politi L, Della Porta M, Saba L, Savevski V. Generative adversarial networks in brain imaging: A narrative review. J Imaging. (2022) 8:83. doi: 10.3390/jimaging8040083 62. Prata D, Mechelli A, Kapur S. Clinically meaningful biomarkers for psychosis: A systematic and quantitative review. Neurosci Biobehav Rev. (2014) 45:134–41. doi: 10. 1016/j.neubiorev.2014.05.010 Frontiers in Psychiatry 16 frontiersin.org
10.1111_jbfa.12686
Asymmetric trading responses to credit rating announcements from issuer- Asymmetric trading responses to credit rating announcements from issuer- versus investor-paid rating agencies versus investor-paid rating agencies Quan Pham Minh Nguyen, HX Do, A Molchanov, L Nguyen, NH Nguyen Publication date Publication date 01-01-2023 Licence Licence This work is made available under the CC BY 4.0 licence and should only be used in accordance with that licence. For more information on the specific terms, consult the repository record for this item. Document Version Document Version Published version Citation for this work (American Psychological Association 7th edition) Citation for this work (American Psychological Association 7th edition) Nguyen, Q. P. M., Do, H., Molchanov, A., Nguyen, L., & Nguyen, N. (2023). Asymmetric trading responses to credit rating announcements from issuer- versus investor-paid rating agencies (Version 1). University of Sussex. https://hdl.handle.net/10779/uos.23495603.v1 Published in Published in Journal of Business Finance and Accounting Link to external publisher version Link to external publisher version https://doi.org/10.1111/jbfa.12686 Copyright and reuse: Copyright and reuse: This work was downloaded from Sussex Research Open (SRO). This document is made available in line with publisher policy and may differ from the published version. Please cite the published version where possible. Copyright and all moral rights to the version of the paper presented here belong to the individual author(s) and/or other copyright owners unless otherwise stated. For more information on this work, SRO or to report an issue, you can contact the repository administrators at sro@sussex.ac.uk. Discover more of the University’s research at https://sussex.figshare.com/ Received: 16 June 2020 Revised: 6 December 2022 Accepted: 9 January 2023 DOI: 10.1111/jbfa.12686 A R T I C L E Asymmetric trading responses to credit rating announcements from issuer- versus investor-paid rating agencies Quan M. P. Nguyen1 Lily Nguyen4 Nhut H. Nguyen5 Hung Xuan Do2,3 Alexander Molchanov2 1Department of Accounting and Finance, University of Sussex, Brighton, UK 2School of Economics and Finance, Massey University, Auckland, New Zealand 3International School, Vietnam National University, Hanoi, Vietnam 4UQ Business School, University of Queensland, Brisbane, Queensland, Australia 5Department of Finance, Auckland University of Technology, Auckland, New Zealand Correspondence Alexander Molchanov, School of Economics and Finance, Massey University, Auckland, New Zealand. Email: a.e.molchanov@massey.ac.nz Abstract The credit rating industry has traditionally followed the “issuer-pays” principle. Issuer-paid credit rating agencies (CRAs) have faced criticism regarding their untimely release of negative rating adjustments, which is attributed to a conflict of interests in their business model. An alternative model based on the “investor-pays” principle is arguably less subject to the conflict of interest problem. We examine how investors respond to changes in credit ratings issued by these two types of CRAs. We find that investors react asymmetrically: They abnormally sell equity stakes around rating downgrades by investor-paid CRAs, while abnormally buying around rating upgrades by issuer-paid CRAs. Our study suggests that, through their trades, investors capital- ize on value-relevant information provided by both types of CRAs, and a dynamic trading strategy taking advantage of this information generates significant abnormal returns. K E Y W O R D S credit ratings, institutional investors, trading strategy J E L C L A S S I F I C AT I O N G11, G24 This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. © 2023 The Authors. Journal of Business Finance & Accounting published by John Wiley & Sons Ltd. J Bus Fin Acc. 2023;1–29. wileyonlinelibrary.com/journal/jbfa 1 2 1 INTRODUCTION NGUYEN ET AL. The credit rating sector has long been dominated by three major issuer-paid credit rating agencies (CRAs): Standard and Poor’s (S&P), Moody’s Investors Service (Moody’s) and Fitch Ratings. These issuer-paid CRAs extract fees directly from bond issuers, which might lead to potential conflicts of interest when they provide rating services to those issuers. Issuer-paid CRAs tend to delay the release of negative ratings (Cornaggia & Cornaggia, 2013; J. He et al., 2012; Skreta & Veldkamp, 2009) while giving favorable ratings to stocks in their owners’ portfolios (Kedia et al., 2017). Baghai and Becker (2018) find evidence that issuer-paid CRAs assign higher ratings even to those issuers who pay them for non-rating services. The lack of timeliness in negative rating adjustments in high-profile bankruptcies, such as Enron (2001), WorldCom (2002) and Lehman Brothers (2008), is often presented as evidence of such conflicts. For example, on September 10, 2008—the day Lehman Brothers announced its bankruptcy—S&P and Moody’s had them rated at A2 and A, respectively, and only adjusted the credit ratings down after the bankruptcy announcement. The entry of investor-paid CRAs (e.g., Egan-Jones Ratings [EJR] and Rapid Ratings) has changed the dynamics of the credit rating industry. These CRAs are paid by the end users of their ratings, such as institutional investors, and the con- flict of interest problem is potentially alleviated. Extant literature documents significant evidence of high rating quality of investor-paid CRAs. Cornaggia and Cornaggia (2013) show that Rapid Ratings provides more timely downgrades for defaulting bonds than Moody’s downgrades, which results in significant loss avoidance for investors. Xia (2014) con- siders the entry of EJR as a natural experiment to assess issuer-paid CRAs’ reactions to potential competition from a new player. They find that due to reputational concerns, credit ratings issued by S&P tend to become more responsive and informative following the EJR entry. Beaver et al. (2006) and Bruno et al. (2016) report that EJR’s credit ratings are more accurate and timely than Moody’s, even after its successful registration as a nationally recognized statisti- cal rating organization in December 2007. X. Hu et al. (2019) find corroborating evidence in a non-US setting. Using the introduction of China Bond Rating (CBR) in 2010, a CRA that combines a public utility model and an investor-paid model, the authors show that the CBR entry triggers a significant reduction in rating inflation and improvements in information quality of credit rating announcements by nine traditional issuer-paid CRAs in China. Given the rise of investor-paid CRAs, the competition they bring about and the information content of their credit ratings relative to issuer-paid CRAs, it is crucial to understand whether and how financial market participants uti- lize credit ratings provided by both issuer- and investor-paid CRAs for their benefit. Xia (2014) and Berwart et al. (2019) find that stocks with downgrade announcements by EJR experience significantly more negative returns than following downgrades by issuer-paid CRAs, whereas EJR upgrades apparently do not trigger a positive response from investors. Investigating the reaction of institutional investors to EJR’s rating changes, Bhattacharya et al. (2019) find that these investors are more responsive to its rating announcements than to other trading signals. They also show that institutional investors who follow EJR’s credit rating announcements outperform those who ignore these signals. We contribute to this strand of literature and examine the value relevance of credit rating changes issued by both types of CRAs. We argue that investor-paid CRAs cannot completely dominate traditional issuer-paid CRAs that have long-term positions in the credit rating sector. As argued in previous studies, issuer-paid CRAs only tend to delay negative credit rating announcements due to the potential conflict of interest (Cornaggia & Cornaggia, 2013; He et al., 2012; Skreta & Veldkamp, 2009). In contrast, issuer-paid CRAs are likely less conservative in issuing rating upgrades since it would be in their interest to cater positive ratings to their clients (e.g., Bolton et al., 2012; Griffin et al., 2013). Hence, it remains unclear whether investors show different trading patterns in responding to negative and positive credit rat- ing adjustments from issuer- and investor-paid CRAs. The answer to this question is important as it provides a better understanding of the relevance and viability of different types of CRAs. We use institutional investors and mutual funds’ changes in stock holdings around rating announcements as a proxy for market reaction. We consider EJR as a representative of investor-paid CRAs, while the “Big Three” CRAs (S&P, 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License NGUYEN ET AL. 3 Moody’s and Fitch) are representatives of issuer-paid CRAs. We find that institutional investors abnormally decrease their equity holdings surrounding investor-paid rating downgrades but do not respond to any issuer-paid rating down- grades. On the contrary, they significantly increase their equity holdings around issuer-paid rating upgrades but remain unresponsive to investor-paid rating upgrades. These results suggest that institutional investors and mutual funds consider investor-paid CRAs’ rating downgrades as being timely and informative for their trading as opposed to issuer-paid CRAs’ rating downgrades. Further, they regard issuer-paid rating upgrades as having more value-relevant information than investor-paid rating upgrades. In the main analysis, we use quarterly mutual fund (S12) holdings and quarterly institutional (13F) holdings provided by Thomson Reuters. We also use daily institutional trades provided by Abel Noser Corporation to measure institutional reactions to credit rating adjustments.1 We then examine whether investors can profit from trading decisions in response to rating changes. We con- struct and compare four trading strategies: (1) a “dynamic” strategy—selling following investor-paid negative signals and buying following issuer-paid positive signals, (2) a “naïve” strategy—selling following negative signals and buy- ing following positive signals from any rating agency, (3) an “EJR-based” strategy—selling following negative signals and buying following positive signals announced by EJR and (4) an “issuer-paid CRA-based” strategy—selling follow- ing negative signals and buying following positive signals by any of the issuer-paid CRAs. Following Jagolinzer et al. (2011), we compute returns for each trading strategy adjusting for common risk factors using the Fama–French five- factor model. The trading strategy analysis is performed in two steps. First, we construct “notional” trading strategies to acknowledge the fact that any market player with access to credit ratings can potentially benefit from these strate- gies. These results also correspond to equally weighted returns of an investor who trades on every signal consistent with a given strategy. While all four strategies outperform the buy-and-hold strategy, we find that the dynamic strat- egy produces the highest returns, offering an average difference in annualized risk-adjusted returns of up to 5.02% over the other three strategies for a 1-month holding period. Second, using aggregate credit rating changes and insti- tutional investors’ quarterly stock holding changes from S12 and 13F data, we find that all four trading strategies earn substantially higher abnormal returns than the corresponding notional strategy returns and that the dynamic strategy consistently exhibits the highest abnormal returns. Finally, since Abel Noser Corporation provides daily trading data for institutional investors, we use them as an alternative dataset to identify trading strategies based on cumulative net buy around announcement dates. We thus explicitly acknowledge that an institution can dynamically switch between strategies and potentially follow multiple strategies at a time. Our results confirm the superiority of the dynamic trad- ing strategy. More importantly, it outperforms the other three strategies by more than 10% per annum for a 1-month holding period and up to 7.26% per annum for a 9-month holding period. This outperformance is more than twice the notional strategies’ corresponding outperformance; hence, they are consistent with the argument that institutional investors have advanced trading skills and knowledge (Puckett & Yan, 2011) to exploit the informative announcements in the financial markets. Our study contributes to the literature in several important ways. First, we add to the knowledge of the relationship between the quality of credit ratings and market participants’ behavior. The related literature finds that the high qual- ity of investor-paid CRA ratings creates a reputational concern for issuer-paid CRAs, which motivates them to improve the overall quality of credit ratings (e.g., Berwart et al., 2019; Bruno et al., 2016; Ramsay, 2011; Xia, 2014). For example, Xia (2014) finds that following EJR’s appearance, S&P ratings started to reflect credit risks more accurately. Similarly, Ramsay (2011) discovers that the entry of Rapid Ratings—another investor-paid CRA—motivates major issuer-paid CRAs to improve the quality of credit ratings. X. Hu et al. (2019) provide evidence of significant improvements in credit rating informativeness in the China bond market after the introduction of CBR, a combined public utility and investor- paid CRA. However, the impact of credit rating quality on investors’ behavior remains underexamined. Our study fills this gap by examining the role of timeliness of credit rating adjustments—a proxy for credit rating quality—in driving institutional investors’ behavior. 1 We thank the editor and referee for this constructive suggestion. 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 4 NGUYEN ET AL. Second, our findings enrich the understanding of how institutional investors, as professional players, analyze and react dynamically to negative and positive rating adjustments obtained from different sources over time. Baker and Mansi (2002) report interesting results regarding the view of institutional investors toward credit ratings. They find that a majority of institutional investors value credit ratings in their investment decisions and place significant importance on rating timeliness. Although they also generally agree on the accuracy of ratings in reflecting firms’ creditworthiness, they believe that ratings could either overstate or understate credit risk. Therefore, institutional investors tend to also rely on their own internal analysis before responding to credit rating news. Cantor et al. (2007) find from their survey that investment managers in the United States and Europe share remarkably similar usage of credit ratings to conduct their investment activities. He (2021) finds that transient institutional investors tend to trade more intensively in low credit rating firms following their earnings announcements. Bhattacharya et al. (2019) find that institutional investors who follow EJR’s rating announcements significantly focus on EJR rating news rather than important equity trading signals, such as analyst recommendations, earnings announcements and earnings forecast revisions. They also find that institutional investors who persistently follow EJR’s credit rating announcements out- perform those who do not embrace these signals. Our study extends their findings by providing new evidence that investors with access to rating announcements could dynamically exploit the value-relevant information of negative and positive rating signals provided by both investor-paid and issuer-paid CRAs in making their trading decisions. Our results show that while such trading behavior is generally profitable, institutional investors evidently earn the high- est abnormal profits. Finally, the reported abnormal profits that continue to exist up to at least 6 months suggest that investors underreact to the information content of credit rating announcements, particularly to negative signals provided by investor-paid EJR and positive signals given by issuer-paid CRAs. The remainder of the paper is organized as follows. Section 2 summarizes data collection, variable measurements and summary statistics. Section 3 presents the methodology and empirical results. Robustness checks are presented in Section 4. Section 5 concludes. 2 SAMPLE SELECTION, VARIABLE MEASUREMENTS AND SUMMARY STATISTICS 2.1 Sample selection We consider two quarterly institutional holding databases to extract institutional investors’ trading activities: mutual fund (S12) holdings and institutional (13F) holdings provided by Thomson Reuters. The S12 holdings database pro- vides data on mutual fund holdings of US securities at the end of each quarter. The 13F holdings database provides a similar data structure at the institutional (i.e., investment company or fund family) level.2 Our analysis includes all US equity mutual funds and institutional investors that have at least 65% of their assets in common stocks (e.g., Amihud & Goyenko, 2013; Cremers & Petajisto, 2009).3 The final samples include 8566 mutual funds and 8656 institutional investors. As mentioned above, we focus on two types of CRAs: investor- and issuer-paid CRAs. EJR is a representative of investor-paid CRAs, while the “Big Three” represent issuer-paid CRAs. Credit rating data are sourced from Egan-Jones Rating Company4 and Bloomberg for the period from 1999 to 2017 to match with the S12 and 13F holding data. The credit rating databases include two types of rating information: rating warning announcements5 and official rating 2 Note that Form 13F is only required for institutional investment managers with more than $100 million in assets under management. 3 We also consider alternative thresholds such as 50%, 60% and 70% as robustness checks. The results are consistent and available upon request. 4 We wish to thank the Egan-Jones Rating Company for sharing its historical rating data. 5 Based on the data availability, there are two types of rating warning announcements: outlook and developing signals. These signals are normally announced before official rating adjustments. 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License NGUYEN ET AL. 5 adjustments.6 The databases also report the date of each credit rating adjustment. As we are interested in corporate credit ratings, sovereign credit and asset-backed securities ratings are excluded. 2.2 Variable definitions Since credit ratings are represented by different combinations of letters and numbers (e.g., AAA/Aaa, AA+/Aa1, AA/Aa2, AA-/Aa3), several prior studies follow Gande and Parsley (2005) to construct a unique “comprehensive credit rating” (CCR) scale to quantify alphabetic ratings (Alsakka & ap Gwilym, 2012; Chen et al., 2016; Dimitrov et al., 2015; Drago & Gallo, 2016). Based on the features of credit rating data availability, we follow Joe and Oh’s (2018) rating con- version scale. The numeric score for letter rating and warning (single) signals are shown in Appendix A.7 In addition, we also follow the literature (Chen et al., 2016; Vu et al., 2015) to measure the significance of the credit rating event for firm n at time t as the change in CCR, ΔCCRn,t: ΔCCRn,t = CCRn,t − CCRn,t−1. (1) To match the frequency of fund holding data, we aggregate changes in credit rating adjustment on a quarterly basis. For instance, in the first quarter of 2010, S&P announces two credit rating adjustments for firm n, a single downgrade (i.e., −1 notch) on February 1, 2010, and a double downgrade (i.e., −2 notches) on March 2, 2010, and the aggregate credit rating change by S&P for firm n in the first quarter of 2010 is −3 notches. We use abnormal mutual fund and institutional investors’ trading as a proxy for investors’ responses, measured by quarterly abnormal net buy, NBi,n,q. NBi,n,q = nbi,n,q − average nbi,n,q, (2) where nbi,n,q is the quarterly net buy by mutual fund or institutional investor i on stock n measured as dollar stock holding in quarter q minus quarter q − 1, normalized by the stock’s total market value at the end of the quarter q.8 Average nbi,n,q denotes the average value of nbi,n,q in the period from quarter q − 4 to q − 1 as follows: average nbi,n,q = ∑−4 k = −1 nbi,n,q+k 4 . (3) We follow Chemmanur et al. (2016) to convert NBi,n,q into basis points. 2.3 Control variables We also follow the related literature (Bernile et al., 2015; Bhattacharya et al., 2019; Henry et al., 2017) to control for a vector of firm characteristics related to institutional trading activities. The control variables include firm size, profitability, stock idiosyncratic volatility, Z-score, analyst coverage, interest coverage, firm age, leverage, high-tech 6 Official rating adjustments are basically divided into two types: positive and negative signals. These signals can also include single and multiple events. In our study, a single event is either a one-notch upgrade or downgrade, and a multiple event is either a multiple-notch upgrade (downgrade) or a combined event of a rating warning announcement and an official rating adjustment. 7 Gande and Parsley (2005) count positive and negative outlooks as one notch. In our study, to highlight the impacts of official upgrades (downgrades), positive and negative outlooks are counted as 0.5 notch and positive and negative developments as 0.25. 8 This is to follow the merit of Chemmanur et al. (2009) who estimate institutional net buy based on shares traded and shares outstanding. 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 6 NGUYEN ET AL. dummy and an S&P 500 index inclusion dummy. The descriptions of control variables and their sources are presented in Appendix B. 2.4 Summary statistics Table 1 presents summary statistics of mutual funds (S12) and institutional investors (13F). The number of mutual funds (institutional investors) has gradually increased from 3364 (1751) in 1999 to 4752 (4130) in 2017. The number of stocks held by mutual funds (institutional investors) has been relatively stable, ranging from 576 (562) in 1999 to 653 (692) in 2017. On average, each institutional investor holds 109 stocks in their portfolio in 1999. The number gradually increases to 162 in 2017. These figures are almost double those of mutual funds, which are at 52 stocks in 1999 and 99 stocks in 2017. Mutual funds’ (institutional investors’) stock holdings have sharply increased from $301 (267) billion in 1999 to $2573 (1168) billion in 2017. On average, each mutual fund holds $89 million worth of stocks in 1999, and the amount increases to $541 million in 2017. The figures for institutional investors are $153 million in 1999 and $283 million in 2017. Table 2 displays summary statistics of credit rating events. The first row of panel A shows the number of unique firms that each CRA provides credit rating announcements over the sample period of 1999–2017. Despite being a relatively new player in the credit rating industry, EJR provides credit ratings for 1502 firms, which are only slightly fewer than S&P (1432 firms) but more than double the coverage by either Moody’s (645) or Fitch (502). EJR is also the only CRA that provides developing signals, whereas the traditional issuer-paid CRAs do not provide such service during our sample period.9 We split our rating announcements into negative and positive events and present them in panel A, sections 1 and 2. There are 2628 (2504), 1172 (546), 355 (172) and 200 (64) negative (positive) combined events10 assigned by EJR, S&P, Moody’s and Fitch, respectively. In addition, the sample comprises 2013 (1896), 1187 (1342), 370 (541) and 578 (549) solo downgrades (upgrades) and 415 (264), 428 (114), 187 (48) and 163 (70) multiple downgrades (upgrades) announced by EJR, S&P, Moody’s and Fitch, respectively. Panel A also shows 1648 (1910), 1537 (730), 440 (299) and 278 (80) negative (positive) outlook signals by these CRAs, respectively. Panel B of Table 2 presents the distribution of credit rating adjustments. Regarding the total number of rating events, EJR issues about 20% more rating changes than all issuer-paid CRAs’ events combined. Within each CRA, EJR has more positive than negative rating announcements. This is opposite to the issuer-paid CRAs, which announce more negative rating adjustments than positive ones. Regarding the magnitude of rating adjustments, Fitch, on aver- age, seems to provide the boldest adjustments, compared to the other CRAs. For example, the mean absolute value of negative rating adjustments is 1.174 for Fitch, while that is 1.041, 1.050 and 1.068 for EJR, S&P and Moody’s, respec- tively. Negative rating adjustments are generally larger in absolute value than positive rating adjustments. The median column in panel B suggests that S&P is relatively more conservative in their negative rating adjustments: 50% of their negative rating events have a median value of 0.5 notch. Table 3 presents the descriptive statistics of control variables computed around credit rating changes. Observa- tions are divided into three groups: The first group includes firms rated by EJR and S&P, the second group is for firms rated by EJR and Moody’s and the third group covers firms rated by EJR and Fitch. The N column shows the number of fund-firm-quarter observations. The first group has the largest number of observations in both S12 and 13F sam- ples, followed by groups three and two. The third group includes, on average, larger and older firms. This appears to be 9 EJR derives its “watch” assignments from the difference between the current and projected ratings. No difference between the two results in a “stable” watch, a higher projected rating results in a “positive” or “POS” watch and a lower projected rating results in a “negative” or “NEG” watch. The absence of a projected rating results in a “developing” or “DEV” watch or no watch being populated. The addition of a POS or NEG is at the discretion of the analyst or Rating Committee and usually results from the direction the rate is expected to move over time. See https://www.egan-jones.com/public/download/methodologies/ 20210510/EJR_Main_Methodologies_V15a.pdf 10 A combined event is a multiple announcement when a CRA adjusts both credit rating score and outlook (or developing) signal. 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License NGUYEN ET AL. 7 ) s I I ( F 3 1 s r o t s e v n i l a n o i t u t i t s n I ) s F M ( 2 1 S s d n u f l a u t u M e u l a v k c o t s l a t o T e u l a v k c o t s l a t o T e u l a v k c o t s l a t o T I I h c a e y b d l e h e u l a v k c o t s l a t o T s k c o t s . o N F M h c a e y b d l e h s F M y b d l e h s k c o t s . o N ) s n o i l l i b ( ) s n o i l l i b ( s I I y b d l e h I I r e p s k c o t s . o N s I I . o N ) s n o i l l i b ( ) s n o i l l i b ( F M r e p s k c o t s . o N s F M o N . 3 5 1 0 . 5 8 1 0 . 6 5 1 0 . 7 4 1 0 . 0 7 1 0 . 9 9 1 0 . 6 0 2 0 . 7 2 2 0 . 5 4 2 0 . 4 8 1 0 . 7 5 1 0 . 6 9 1 0 . 7 0 2 0 . 7 2 2 0 . 9 5 2 0 . 4 8 2 0 . 2 6 2 0 . 9 5 2 0 . 3 8 2 0 . 1 1 2 0 . 7 6 2 8 4 3 4 1 3 0 0 3 7 4 3 5 3 4 4 9 4 5 7 5 1 8 6 4 4 5 9 5 4 9 6 5 4 3 6 2 2 7 3 6 8 2 2 0 1 8 0 0 1 3 3 0 1 8 6 1 1 0 2 6 9 0 1 3 2 1 8 2 1 0 3 1 9 3 1 4 4 1 1 4 1 2 4 1 2 4 1 4 3 1 5 3 1 2 4 1 8 4 1 2 5 1 9 5 1 5 6 1 1 6 1 9 5 1 2 6 1 3 4 1 2 6 5 2 1 6 1 1 6 9 2 6 7 3 6 4 5 6 9 6 6 5 6 6 5 6 6 2 4 6 2 4 6 1 5 6 1 7 6 6 7 6 1 9 6 9 9 6 6 0 7 6 9 6 2 9 6 6 5 6 1 5 7 1 5 7 8 1 2 1 0 2 3 4 0 2 6 4 0 2 7 8 1 2 1 0 4 2 9 2 5 2 0 8 7 2 0 6 9 2 3 2 9 2 0 0 9 2 0 7 0 3 2 8 1 3 0 4 3 3 3 9 5 3 3 5 8 3 9 8 9 3 0 3 1 4 9 1 8 2 9 8 0 0 . 2 0 1 0 . 1 9 0 0 . 7 8 0 0 . 8 9 0 0 . 1 3 1 0 . 2 5 1 0 . 3 8 1 0 . 4 9 1 0 . 3 5 1 0 . 6 5 1 0 . 8 0 2 0 . 4 1 2 0 . 1 4 2 0 . 9 1 3 0 . 7 8 3 0 . 7 8 3 0 . 9 1 4 0 . 1 4 5 0 . 9 1 2 0 . 1 0 3 1 8 3 1 5 3 8 4 3 8 9 3 5 2 5 6 0 6 8 1 7 5 4 8 9 1 7 8 8 6 7 5 8 7 2 9 4 4 0 1 9 5 3 1 5 2 7 1 3 9 7 1 4 2 0 2 3 7 5 2 7 5 9 2 5 1 6 9 6 5 7 9 7 2 8 4 8 6 8 7 8 1 9 8 9 8 9 6 9 9 9 0 0 1 8 9 9 9 0 0 1 9 9 7 8 6 7 5 8 0 6 4 1 6 8 1 6 7 2 6 4 4 6 7 5 6 3 5 6 2 4 6 8 2 6 8 2 6 8 3 6 9 4 6 7 5 6 3 7 6 0 8 6 3 8 6 3 7 6 3 5 6 2 4 6 4 6 3 3 6 2 7 3 3 6 8 3 4 8 9 3 2 7 0 4 3 0 0 4 2 0 0 4 3 2 9 3 5 6 3 4 4 9 6 4 3 1 4 4 3 1 1 4 3 3 3 4 8 2 3 4 4 5 2 4 4 5 4 4 0 4 6 4 6 3 8 4 2 5 7 4 7 1 2 4 r a e Y 9 9 9 1 0 0 0 2 1 0 0 2 2 0 0 2 3 0 0 2 4 0 0 2 5 0 0 2 6 0 0 2 7 0 0 2 8 0 0 2 9 0 0 2 0 1 0 2 1 1 0 2 2 1 0 2 3 1 0 2 4 1 0 2 5 1 0 2 6 1 0 2 7 1 0 2 e g a r e v A i l s c i t s i t a t s s g n d o h ) F 3 1 ( r o t s e v n i l a n o i t u t i t s n i d n a ) 2 1 S ( d n u f l a u t u M 1 E L B A T 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 8 NGUYEN ET AL. TA B L E 2 Credit rating sample statistics Panel A: Rating changes Egan-Jones Ratings (EJR) Standard and Poor (S&P) Moody’s Investors Service (Moody’s) Fitch Ratings (Fitch) Number of firms rated Section 1: Negative events Negative developing Negative outlook Negative combine event Single downgrade Multiple downgrade Section 2: Positive events Positive developing Positive outlook Positive combine event Single upgrade Multiple upgrade 1502 186 1648 2628 2013 415 741 1910 2504 1896 264 1432 – 1537 1172 1187 428 – 730 546 1342 114 645 – 440 355 370 187 – 299 172 541 48 502 – 278 200 578 163 – 80 64 549 70 Panel B: The distribution of rating changes N Mean Std. dev. P1 P25 Median P75 P99 EJR negative event EJR positive event S&P negative event S&P positive event 6885 1.041 7315 0.909 4325 1.050 2730 1.024 Moody’s negative event 1352 1.068 Moody’s positive event 1060 0.884 Fitch negative event 1219 1.174 Fitch positive event 762 1.177 0.758 0.740 0.933 0.961 0.736 0.423 1.054 0.867 0.250 0.500 0.250 0.500 0.500 0.500 0.500 0.500 0.500 0.500 0.500 0.500 0.500 0.500 0.500 1.000 1.000 0.750 0.500 1.000 1.000 1.000 1.000 1.000 1.250 4.000 1.000 3.750 1.000 5.000 1.000 5.500 1.500 3.500 1.000 2.500 1.000 5.500 1.000 5.000 Note: The table presents credit rating events announced by EJR (investor-paid credit rating agency [CRA]) and S&P, Moody’s and Fitch (issuer-paid CRAs). Panel A displays the number of firms rated and the number of rating events (negative and pos- itive separately) announced by each CRA after being merged with COMPUSTAT, CRSP and S12/13F data. Panel B presents summary statistics for credit rating changes of each CRA, where the magnitude of a rating change is calculated as the total number of notches by which a rating agency changes a firm’s credit rating. consistent with EJR’s and Fitch’s policy of rating veteran firms. For instance, the mean market capitalization in the S12 (13F) sample in the third group is $36,874 ($44,837) million, while the number is $30,333 ($37,638) million for group one and $12,836 ($16,899) million for group two. The mean Z-scores in the S12 (13F) sample are 2.12 (2.11), 1.72 (1.70) and 1.84 (1.86) for the first, second and third groups, respectively. These means are relatively close to the conventional threshold of 1.8 but above the risk level of a financially healthy firm. Leverage ratios are similar across all three groups. Finally, the median interest coverage ratio is slightly lower for firms in group two than for firms in groups one and three. 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License NGUYEN ET AL. 9 TA B L E 3 Descriptive statistics for firms covered in S12 and 13F databases Panel A: Mutual fund (S12) database Firms rated by EJR and S&P Firms rated by EJR and Moody’s Firms rated by EJR and Fitch N Mean Median N Mean Median N Mean Median Ln(MV) ROA IDIO_RISK Z-SCORE 2,808,671 10.32 9.37 1,079,231 2,808,671 0.04 0.04 1,079,231 2,808,671 0.02 0.01 1,079,231 2,808,671 2.12 1.82 1,079,231 ANALYST_COVERAGE 2,808,671 7.15 6.78 1,079,231 Ln(AGE) 2,808,671 3.21 3.33 1,079,231 9.46 0.03 0.02 1.72 6.33 2.99 8.59 1,716,964 10.51 0.04 1,716,964 0.04 0.02 1,716,964 0.01 1.54 1,716,964 1.84 5.84 1,716,964 7.38 3.00 1,716,964 3.31 INTEREST_COVERAGE 2,808,671 15.65 9.26 1,079,231 11.78 7.15 1,716,964 14.39 LEVERAGE S&P_500 2,808,671 0.33 0.27 1,079,231 2,808,671 0.65 1.00 1,079,231 HIGH_TECH 2,808,671 0.03 0.00 1,079,231 0.33 0.48 0.02 0.31 1,716,964 0.37 0.00 1,716,964 0.76 0.00 1,716,964 0.02 9.65 0.04 0.01 1.63 7.00 3.47 8.59 0.28 1.00 0.00 Panel B: Institutional investors (13F) database Firms rated by EJR and S&P Firms rated by EJR and Moody’s Firms rated by EJR and Fitch N Mean Median N Mean Median N Mean Median Ln(MV) ROA IDIO_RISK Z-SCORE 3,180,369 10.54 9.65 1,084,625 3,180,369 0.04 0.04 1,084,625 3,180,369 0.01 0.01 1,084,625 3,180,369 2.11 1.86 1,084,625 ANALYST_COVERAGE 3,180,369 7.44 7.02 1,084,625 Ln(AGE) 3,180,369 3.32 3.40 1,084,625 9.73 0.03 0.02 1.70 6.53 3.05 8.75 2,040,834 10.71 0.04 2,040,834 0.04 0.02 2,040,834 0.01 1.51 2,040,834 1.86 6.00 2,040,834 7.62 3.04 2,040,834 3.44 INTEREST_COVERAGE 3,180,369 16.01 9.80 1,084,625 12.65 7.29 2,040,834 15.12 LEVERAGE S&P_500 3,180,369 0.35 0.28 1,084,625 3,180,369 0.68 1.00 1,084,625 HIGH_TECH 3,180,369 0.03 0.00 1,084,625 0.33 0.49 0.02 0.31 2,040,834 0.38 0.00 2,040,834 0.76 0.00 2,040,834 0.02 9.87 0.04 0.01 1.66 7.30 3.56 9.26 0.28 1.00 0.00 Note: The table presents the summary statistics of control variables, which are defined in Appendix B. Statistics are computed around credit rating announcements. 3 MAIN RESULTS 3.1 Institutional responses to issuer- and investor-paid rating adjustments We now examine institutional investors’ responses to credit rating signals announced by issuer- and investor-paid CRAs. To ensure that reactions are comparable, we construct three paired samples, which include firms rated by EJR and each of the major issuer-paid CRAs: EJR and S&P, EJR and Moody’s and EJR and Fitch. We estimate the following 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 10 NGUYEN ET AL. regression for each of the paired samples: NBi, n,q = 𝛼 + 𝛽1NEGn,q + 𝛽2POSn,q + 𝛽3NEGn,q∗EJRn,q + 𝛽4POSn,q∗EJRn,q 𝛾kCONTROLS(n,q) +q 1 + 𝛽5EJR(n,q) + 𝜃qQuarterFEq ∑q 1 ∑k 1 + 𝛿iInvestorFEi + 𝜑nFirmFEn + 𝜀i,n,q i∑ 1 n∑ 1 , (4) ∑ ∑ q and denote it by where NBi,n,q, defined in equation (2), denotes mutual fund (institutional investor) i’s abnormal dollar net buy of firm n’s stock for credit rating adjustments in quarter q. We sum all ΔCCRs, as defined in equation (1), for each firm n in quarter ∑ ΔCCRn,q > 0 and ΔCCRn,q < 0. Therefore, an increase in NEGn,q (POSn,q) represents an POSn,q as absolute increase in aggregate credit rating downgrade (upgrade) for firm n in quarter q. CONTROLSn,q represents a set of firm-level control variables as described in Table 3. QuarterFEq denotes quarter-specific dummy variables to control for differences in institutional trading behavior that can be induced by various economic conditions in different quar- ΔCCRn,q.11 We then define NEGn,q as | ∑ ΔCCRn,q < 0 and zero if ΔCCRn,q > 0 and zero if ΔCCRn,q| if ΔCCRn,q if ∑ ∑ ∑ ters. InvestorFEi (FirmFEn) is used to control for investor- (firm-) characteristics that are not captured by CONTROLSn,q. In this model, NEGn,q and POSn,q are interacted with EJRn,q, a dummy variable that equals one for EJR’s credit rating announcements and zero otherwise. 𝜀i,n,q is a random error. The results of equation (4) are presented in Table 4. We find significant asymmetries in the abnormal trading of insti- tutional investors and mutual funds in relation to EJR’s and issuer-paid CRAs’ rating announcements. These results are robust to the inclusion of control variables and fixed effects. For example, for firms that are rated by EJR and S&P, columns 1 and 2 show significant increases in mutual funds’ and institutional investors’ net buy of stocks with an aggregate positive change in S&P’s rating adjustments in a given quarter. The POS coefficient is positive and signifi- cant across the regression specifications. Its magnitude is also economically meaningful. For example, the 0.0571 basis point coefficient in column 2 of panel B is equivalent to an average increase of $215,824 in abnormal institutional net buy over the respective quarter with a one-notch upgrade.12 Institutional investors, however, react significantly less to positive rating changes issued by EJR. The EJR*POS interaction coefficient is negative in almost every model. The F- test results for the overall impact of rating upgrades by EJR, that is, the sum of POS and EJR*POS coefficients, indicate that both mutual funds and institutional investors are unresponsive to EJR’s positive rating changes. Table 4 shows the opposite results for rating downgrades. Institutional investors and mutual funds apparently find EJR’s negative rating adjustments more informative than S&P’s announcements. While the NEG coefficient shows no clear pattern across specifications, the EJR*NEG is negative and statistically and economically significant across the models. For example, the −0.1368 coefficient of EJR*NEG in column 2 of panel B indicates that a firm receiving a one- notch rating downgrade by EJR experiences an average decrease of $517,077 in abnormal institutional net buy over the respective quarter, compared to a similar downgrade by S&P. The F-test results for the overall impact of rating downgrades by EJR, that is, the sum of NEG and EJR*NEG coefficients, indicate that the total effect of EJR downgrades is statistically and economically significant. We find similar asymmetric responses by institutional investors to upgrades and downgrades for firms that are rated by EJR and Moody’s. For example, a POS coefficient of 0.0931 in column 4 indicates that abnormal institutional net buy, on average, increases by $156,543 over the quarter in which a one-notch aggregate rating upgrade by Moody’s takes place. The combined effect of POS + EJR*POS shows that institutional investors do not react to EJR’s upgrades as opposed to Moody’s. However, the results for rating downgrades support the notion that institutional investors respond to EJR’s rather than Moody’s downgrades. The EJR*NEG coefficient in column 4 of Panel B is −0.2707, indi- cating that EJR downgrades are associated with, on average, a decrease of $455,170 in abnormal institutional net buy, compared to Moody’s downgrades. The significant F-test results for the sum of NEG and EJR*NEG in columns (3) 11 In our analysis, we exclude firm-quarter observations that EJR and the paired issuer-paid CRA have different credit rating signals in a quarter. 12 The increase is calculated by multiplying the POS coefficient of 0.0571 by the average market capitalization (e10.54 = $37,798 million) of firms in the EJR and S&P group in panel B of Table 3 and dividing the result by 10,000 (since the net buy is in basis points). 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License NGUYEN ET AL. 11 TA B L E 4 Abnormal trading responses to credit rating adjustments—S12 and 13F samples Panel A: Mutual funds’ abnormal holding changes EJR versus S&P EJR versus Moody’s EJR versus Fitch (1) (2) (3) (4) (5) (6) Intercept 0.1379*** −0.0856** 0.4049*** −0.3387*** 0.1770*** −0.5467*** NEG POS (0.0183) (0.0346) (0.0445) (0.0828) (0.0234) (0.0481) 0.0105* −0.0013 −0.004 0.0234 0.0211** 0.0286** (0.0062) (0.0073) (0.024) (0.0283) (0.009) (0.0122) 0.0318*** 0.0324*** 0.0241** 0.0308** 0.0441*** 0.0612*** (0.0087) (0.0108) (0.0121) (0.0123) (0.0127) (0.0179) EJR×NEG −0.0199*** −0.0103* −0.0416* −0.0512* −0.0139 −0.0187 EJR×POS −0.0257*** −0.0262** −0.0123 −0.0152 −0.0424*** −0.0587*** (0.0074) (0.0057) (0.0253) (0.0292) (0.0102) (0.0135) EJR −0.2294*** −0.2410*** −0.5808*** −0.6357*** −0.2948*** −0.2884*** (0.0097) (0.012) (0.0303) (0.0337) (0.0138) (0.0191) Control variables: No Yes No Yes No Yes (0.0084) (0.0099) (0.0267) (0.030) (0.0124) (0.0165) F-tests: NEG + EJR×NEG −0.0094** −0.0115** −0.0456*** −0.0278*** 0.0073 0.0100 POS + EJR×POS 0.0061 0.0062 0.0118 0.0156 0.0017 0.0025 (0.0042) (0.005) (0.0084) (0.0096) (0.0051) (0.0063) (0.0044) (0.0056) (0.0083) (0.0095) (0.0055) (0.0072) Fixed effects: Investor FE Firm FE Quarter FE N Adj. R2 Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes 3,582,992 2,808,671 1,273,265 1,079,231 2,291,401 1,716,964 0.002 0.003 0.003 0.004 0.002 0.002 Panel B: Institutional investors’ abnormal holding changes EJR versus S&P EJR versus Moody’s EJR versus Fitch (1) (2) (3) (4) (5) (6) Intercept 0.6663*** 2.0098*** 2.4527*** 2.7954*** 0.2020*** −1.3221*** NEG POS (0.0548) (0.1136) (0.1483) (0.2731) (0.0568) (0.1378) −0.0278 −0.0665*** 0.0124 0.0511 −0.0383 0.0785** (0.0219) (0.0257) (0.0827) (0.0959) (0.0234) (0.0346) 0.0330* 0.0571** 0.1087*** 0.0931*** 0.0893** 0.1089** (0.0181) (0.0285) (0.0325) (0.0224) (0.0352) (0.0531) EJR×NEG −0.1555*** −0.1368*** −0.2474*** −0.2707*** −0.0083 −0.0624 (0.026) (0.0305) (0.0875) (0.0992) (0.0273) (0.0387) (Continues) 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 12 TA B L E 4 (Continued) NGUYEN ET AL. Panel B: Institutional investors’ abnormal holding changes EJR versus S&P EJR versus Moody’s EJR versus Fitch (1) (2) (3) (4) (5) (6) EJR×POS 0.0071 −0.0288 −0.0956*** −0.0923*** −0.1045*** −0.1373** EJR −0.7735*** −1.0401*** −2.6222*** −2.6376*** −0.3948*** −0.3114*** (0.035) (0.0433) (0.0276) (0.027) (0.039) (0.0575) Control variables: No Yes No Yes No Yes (0.0292) (0.0342) (0.0974) (0.1075) (0.0336) (0.0466) F-tests: NEG + EJR×NEG −0.1833*** −0.2034*** −0.2350*** −0.2196*** −0.0467* 0.0162 POS + EJR×POS 0.0401 0.0283 0.0131 0.0008 −0.0151 −0.0285 (0.0145) (0.0174) (0.0305) (0.0334) (0.0247) (0.0189) (0.0265) (0.0207) (0.033) (0.0361) (0.0173) (0.0238) Fixed effects: Investor FE Firm FE Quarter FE N Adj. R2 Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes 4,088,703 3,180,369 1,281,861 1,084,625 2,729,378 2,040,834 0.001 0.001 0.002 0.002 0.001 0.001 Note: The table reports OLS regression results for mutual fund (S12) and institutional investor (13F) abnormal holding changes in response to credit rating adjustments announced by EJR and issuer-paid CRAs. The dependent variable, defined in equa- tion (2), is a mutual fund’s (institutional investor’s) abnormal net buy of a stock during a quarter. Based on a firm’s aggregate credit rating change in a quarter, we define NEG as the absolute value of a negative change and zero otherwise and POS as the value of a positive change and zero otherwise. Therefore, an increase in NEG (POS) represents an absolute increase in the firm’s aggregate downgrade (upgrade) in that quarter. EJR is a dummy variable that equals one for EJR’s credit rating announce- ments and zero otherwise. Detail descriptions of firm-level control variables are described in Appendix B. Standard errors in parentheses are adjusted for heteroskedisticity and clustering at the firm and quarter levels. ***, ** and * indicate statistical significance at the 1%, 5% and 10% levels, respectively. and (4) indicate that the total effect of EJR downgrades is statistically and economically strong. The results for firms jointly rated by EJR and Fitch in columns 5 and 6 are less clear. While the POS coefficient is significant, indicating that investors react to Fitch’s upgrades, EJR*NEG and EJR + EJR*NEG are mostly insignificant. We investigate this further in Section 3.5. Overall, the results in Table 4 suggest that mutual funds and institutional investors find that credit rating upgrades are more informative; hence, they respond accordingly when issued by S&P or Moody’s rather than by EJR. In contrast, they find that negative rating adjustments are more value-relevant when they are announced by EJR than by S&P or Moody’s. These findings are consistent with the argument that institutional investors are well-equipped to assess the informativeness of credit rating announcements. Previous studies have shown that issuer-paid CRAs tend to delay rat- ing downgrades due to conflict of interests (e.g., Cornaggia & Cornaggia, 2013) but still issue timely rating upgrades (e.g., Kedia et al., 2017). Brogaard et al. (2019) also find that upgrades issued by issuer-paid CRAs do convey new infor- mation. In contrast, investor-paid CRAs tend to be more timely in rating downgrade adjustments (e.g., Berwart et al., 2019; Johnson, 2004). 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License NGUYEN ET AL. 13 3.2 Do CRAs behave the way we assume they do? The findings in the previous section that institutional investors respond more to positive rating announcements by major issuer-paid CRAs and to negative rating announcements by the investor-paid EJR suggest a lead-lag in the timeliness of credit rating announcements between these two types of CRAs. We now empirically examine this. As before, we separately consider three pairs: EJR and S&P, EJR and Moody’s and EJR and Fitch. For each firm rated by each pair of CRAs, the credit rating score is adjusted multiple times by two paired CRAs throughout the sample period. We investigate the lead–lag relationship of each CRA pair for upgrades and downgrades separately. Based on the announcement timeline and the relative magnitude of consecutive rating adjustments, three scenarios are possible. First, when one CRA issues a rating adjustment that is relatively larger in magnitude than the subsequent adjustment announced by the other CRA, the leading CRA is classified as a “major leader.” Second, when one CRA issues a rating adjustment relatively smaller in magnitude than the subsequent adjustment announced by the other CRA, the following CRA is classified as a “major confirmer.” Third, if a rating adjustment by one CRA is followed by an adjustment of the same magnitude by the other CRA, we classify the leading CRA as an “equal magnitude leader.” We then perform a binominal test with the null hypothesis that the relative frequencies that both CRAs in a pair hold for a specific role are equal. In Table 5, section 1 reports the results for negative events, and section 2 shows the results for positive events. Panels A, B and C present the results for EJR and S&P, EJR and Moody’s and EJR and Fitch, respectively. The results generally confirm our expectations that EJR issues relatively larger rating adjustments than the issuer-paid CRAs when these adjustments are downgrades. For example, EJR’s downgrades are larger than S&P’s subsequent down- grades 56.95% ( = 422/(422 + 319)) of the time, which is statistically higher than 43.05% of the time when S&P plays the role of a major leader. The comparison is even higher for EJR than Moody’s in Panel B, at 67.14% ( = 141/(141 + 69)) versus 32.86%. When EJR follows S&P or Moody’s after their respective negative rating adjustments, EJR tends to issue larger negative adjustments more frequently than when the other two CRAs follow EJR’s downgrades with larger magnitudes. The major confirmer row for negative events confirms these differences statistically. There are no statisti- cal differences between EJR and the other CRAs in the frequency of being an equal magnitude leader. However, we find no evidence of EJR’s leading role, compared to Fitch in the issuance of negative signals. Fitch apparently issues larger negative adjustments more frequently than EJR, although these frequency differences are not statistically significant. The results for positive events in Section 2 of Table 5 indicate that all three issuer-paid CRAs tend to issue larger rating upgrades more frequently than EJR. These frequency differences are statistically significant for both cases when these traditional CRAs are major leaders or major confirmers. There are no significant frequency differences in being an equal magnitude leader, except for the EJR and Fitch pair where EJR leads Fitch more often when they issue positive rating adjustments of the same magnitude. Overall, the findings in this table support the results in Table 4 that EJR’s negative rating announcements are apparently more timely and value-relevant to institutional investors than those rating downgrades by the other CRAs. However, the issuer-paid CRAs’ positive rating announcements are valued more by institutional investors than EJR’s rating upgrades. 3.3 Profitability of asymmetric trading strategies 3.3.1 Notional trading strategies We now investigate whether a trading strategy based on credit rating signals with the highest information content can generate superior returns. Credit rating announcements are, in principle, available to all investors—not just 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 14 NGUYEN ET AL. e u l a v - p h c t i F R J E e u l a v - p s ’ y d o o M R J E e u l a v - p P & S R J E h c t i F d n a R J E : C l e n a P s ’ y d o o M d n a R J E : B l e n a P P & S d n a R J E : A l e n a P 7 9 6 2 0 . 7 1 1 1 0 . 5 2 8 7 0 . 1 0 0 0 0 < . 1 0 0 0 0 < . ) . % 8 3 3 5 ( 2 4 1 ) . % 2 6 6 4 ( 4 2 1 1 0 0 0 0 < . ) . % 6 8 2 3 ( 9 6 ) . % 4 1 7 6 ( 1 4 1 2 0 0 0 0 . ) . % 5 0 3 4 ( 9 1 3 ) . % 5 9 6 5 ( 2 2 4 ) t ( r e d a e l r o j a M ) . % 6 0 5 5 ( 6 3 1 ) . % 4 9 4 4 ( 1 1 1 2 1 0 0 0 . ) . % 4 8 8 3 ( 2 8 ) . % 4 1 1 6 ( 9 2 1 2 1 0 0 0 . ) . % 2 1 4 4 ( 4 3 3 ) . % 8 8 5 5 ( 3 2 4 ) 1 + t ( r e m r i f n o c r o j a M ) . % 5 9 0 5 ( 7 0 1 ) . % 5 0 9 4 ( 3 0 1 8 9 4 5 0 . ) . % 5 4 7 4 ( 5 6 ) . % 5 5 2 5 ( 2 7 7 9 9 2 0 . ) . % 5 8 7 4 ( 8 7 2 ) . % 5 1 2 5 ( 3 0 3 ) t ( r e d a e l e d u t i n g a m l a u q E s t n e v e e v i t a g e N : 1 n o i t c e S ) . % 3 7 2 7 ( 2 1 1 ) . % 7 2 7 2 ( 2 4 2 1 2 0 0 . ) . % 6 5 8 5 ( 6 0 1 ) . % 4 4 1 4 ( 5 7 ) . % 7 7 0 8 ( 6 2 1 ) . % 3 2 9 1 ( 0 3 3 7 9 0 0 . ) . % 5 2 6 5 ( 9 9 ) . % 5 7 3 4 ( 7 7 1 0 0 0 0 < . 1 0 0 0 0 < . ) . % 8 4 1 6 ( 5 6 2 ) . % 2 5 8 3 ( 6 6 1 ) t ( r e d a e l r o j a M ) . % 8 4 3 6 ( 9 9 2 ) . % 2 5 6 3 ( 2 7 1 ) 1 + t ( r e m r i f n o c r o j a M s t n e v e e v i t i s o P : 2 n o i t c e S 4 4 5 0 0 . ) . % 6 9 1 4 ( 0 6 ) . % 4 0 8 5 ( 3 8 6 4 4 9 0 . ) . % 4 2 0 5 ( 4 0 1 ) . % 6 7 9 4 ( 3 0 1 7 6 2 9 0 . ) . % 9 7 9 4 ( 5 3 2 ) . % 1 2 0 5 ( 7 3 2 ) t ( r e d a e l e d u t i n g a m l a u q E d e t a r s m r i f r o f s t n e m e c n u o n n a r o f e r a C d n a B , A s l e n a P . s t n e m e c n u o n n a g n i t a r t i d e r c e v i t a g e n d n a e v i t i s o p n i s A R C d i a p - r o t s e v n i d n a - r e u s s i l f o e o r e v i t a l e r e h t s w o h s e b a t l s i h T : e t o N n a h t e d u t i n g a m n i r e g r a l y l e v i t a l e r s i t a h t j t n e m t s u d a g n i t a r a s e u s s i A R C e n o n e h w , t s r i F . s o i r a n e c s e e r h t r e d i s n o c e W ’ . ) h c t i F d n a s y d o o M P, & S ( s A R C ” e e r h T g i B “ e h t f o e n o d n a R J E y b n i r e l l a m s y l e v i t a l e r s i t a h t j t n e m t s u d a g n i t a r a s e u s s i A R C e n o n e h w , d n o c e S ” . r e d a e l r o j a m “ a s a d e i f i s s a l c s i A R C g n d a e i l e h t , A R C r e h t o e h t y b d e c n u o n n a t n e m t s u d a j t n e u q e s b u s e h t n a y b d e w o l l o f s i A R C e n o y b t n e m t s u d a g n i t a r a f i j , d r i h T ” . r e m r i f n o c r o j a m “ a s a d e i f i s s a l c s i A R C g n w o i l l o f e h t , A R C r e h t o e h t y b d e c n u o n n a t n e m t s u d a t n e u q e s b u s e h t n a h t e d u t i n g a m j n i ( y c n e u q e r f e v i t a l e r e h t d n a s e m i t f o r e b m u n e h t w o h s e w , r i a p h c a e r o F ” . r e d a e l e d u t i n g a m l a u q e “ n a s a A R C g n d a e i l e h t y f i s s a l c e w , A R C r e h t o e h t y b e d u t i n g a m e m a s e h t f o t n e m t s u d a j . l e o r c i f i c e p s a n i r i a p A R C h c a e f o y c n e u q e r f e v i t a l e r e h t e r a p m o c o t t s e t i l a i m o n b a m o r f s i e u l a v - p . l e o r c i f i c e p s a s d o h A R C a t a h t l ) s t e k c a r b s l a n g i s e v i t i s o p d n a e v i t a g e n n i s A R C d i a p - r o t s e v n i d n a - r e u s s i l f o e o r e v i t a l e r e h T 5 E L B A T 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License NGUYEN ET AL. 15 institutions. Therefore, we begin by analyzing “notional” trading strategies available to a hypothetical investor with timely access to credit ratings. The first strategy we consider is the “dynamic strategy”—selling following EJR’s negative rating signals and buy- ing following issuer-paid CRAs’ positive rating signals. This trading strategy is our main interest. The second one is the ‘naïve strategy’—selling following negative signals and buying following positive signals from any rating agency. The third strategy is the “EJR-based strategy”– selling following negative signals and buying following positive signals announced by EJR. The fourth strategy is the “issuer-paid CRA-based strategy”—selling following negative signals and buying following positive signals issued by any of the “Big Three” CRAs. We also add a passive strategy as an additional benchmark—investing in the S&P 500 index. We measure the profitability for each trading strategy as follows. First, we examine various holding periods of k months (where k = 1, 3, 6, 9 and 12) starting from the rating announcement date t to day t + 5. We follow Jagolinzer et al. (2011) and estimate abnormal returns after adjusting for common risk factors. Specifically, for each day in the [0, 5] window, risk-adjusted return is the intercept (alpha) from the Fama and French (2015) five-factor model estimated over a holding period of k months: (Rn − Rf ) = 𝛼 + 𝛽1 ( Rmkt − Rf ) + 𝛽2SMB + 𝛽3HML + 𝛽4RMW + 𝛽5CMA + ei, (5) where Rn is the daily return of firm n; Rf is the daily risk-free rate; Rmkt is the CRSP (Center for Research in Secu- rity Prices) value-weighted market return; SMB, HML, RMW and CMA are size, book-to-market, operating profitability and investment factors, respectively.13 For notional strategies, we assume that investors trade in accordance with a credit rating signal, that is, selling (buying) if the signal is negative (positive). Therefore, if the announcement is a rat- ing downgrade, we multiply daily alphas in equation (5) by (−1) to represent risk-adjusted returns to investors’ sales. This adjustment does not apply for investors’ purchases following a rating upgrade. We calculate the risk-adjusted alpha for firm n’s rating announcement t as the simple average of alphas over the [0, 5] window and denote it by αn,t. We use equal weightings to calculate the firm’s mean alpha for each announcement event as with these hypothetical transactions we do not have data on investors’ buy and sell values. The event alphas, αn,t, are then grouped into appropriate trading strategies described above, and a t-test is per- formed across all rating announcements in a given strategy. We also test the mean difference in the value-weighted risk-adjusted returns (i.e., weighted by market capitalization) between strategies with a two-sample t-test and report the results in Table 6. 14 All returns are annualized. We find that all four strategies outperform the buy-and-hold of the S&P 500 index. In addition, consistent with our expectations, the dynamic strategy yields higher abnormal returns than all other strategies. Over the 1-month investment horizon, the dynamic strategy outperforms the other three strate- gies by an annualized risk-adjusted return ranging from 4.22% to 5.02%. Its outperformance is statistically significant for up to 6 months. 3.3.2 Institutional trading strategies We now examine trading strategies based on institutional transactions. The returns on notional strategies can be interpreted as equally weighted returns of an institution trading around every credit rating announcement consis- tent with a certain strategy. By explicitly considering institutional transactions, we acknowledge that institutions may 13 We thank Kenneth French for sharing data on the five risk factors in his website, http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library. html 14 We adjust firms’ market capitalization for inflation following the merit of Acharya and Pedersen (2005). Specifically, we first calculate the ratio of CRSP total market value at the end of month m – 1 (relative to the credit event month) to CRSP total market value at the end of 1998 (just before our sample starts). We then divide a firm’s market capitalization in month m by this ratio before using it as a weight in the t-test. Our (unreported) results are robust when unadjusted market capitalization is used. 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 0.0259 (0.0248) 0.0035 (0.0156) (0.0034) −0.0068 (0.0358) 0.0011 (0.0012) 0.0224 (0.0293) 0.0102 (0.025) 0.0327 (0.0435) 0.0248 (0.0248) 0.0024 (0.0157) 16 NGUYEN ET AL. TA B L E 6 Notional trading strategy profitability Holding periods (1) Dynamic 1 month 3 months 6 months 9 months 12 months 0.0824*** 0.0655*** 0.0462*** 0.0317*** (0.0104) (0.0079) (0.0061) (0.0109) (2) Naïve 0.0356*** 0.0349*** 0.0203*** 0.0141** (0.0075) (0.0046) (0.0038) (0.0068) (3) EJR-based 0.0322*** 0.0327*** 0.0344*** 0.0147*** 0.0156*** (0.0086) (0.0041) (0.003) (0.0023) (4) Issuer-paid CRA-based 0.0402*** 0.0369*** 0.019** (5) S&P 500 index (0.0134) 0.0011 (0.0012) (0.0093) 0.0011 (0.0012) (0.008) 0.0011 (0.0012) (0.0012) 0.0132 (0.016) 0.0011 0.0468*** 0.0307*** 0.0259*** 0.0176 (0.0129) (0.0092) (0.0072) (0.0129) 0.0502*** 0.0328*** 0.0118* 0.017 (0.0135) (0.0089) (0.0068) (0.0112) 0.0422** 0.0286** 0.0272*** 0.0185 (0.0169) (0.0122) (0.0101) (0.0194) 0.0813*** 0.0644*** 0.0451*** 0.0306*** (0.0105) (0.008) (0.0062) 0.0345*** 0.0337*** 0.0192*** (0.011) 0.013* (0.0076) (0.0048) (0.004) (0.0069) (1)–(2) (1)–(3) (1)–(4) (1)–(5) (2)–(5) (3)–(5) (4)–(5) 0.0311*** 0.0316*** 0.0333*** 0.0136*** 0.0145*** (0.0087) (0.0043) (0.0032) (0.0026) 0.0391*** 0.0358*** 0.0179** (0.0134) (0.0094) (0.0081) 0.0121 (0.016) (0.0036) −0.0079 (0.0358) Note: This table reports and compares annualized risk-adjusted returns on four notional trading strategies. “Dynamic” is a strategy that sells a stock when it receives an EJR’s negative rating adjustment and buys the stock when its rating is upgraded by an issuer-paid CRA. The “naïve” strategy is simply to sell (buy) a stock following a negative (positive) signal from any rating agency. For the “EJR-based” strategy, an investor sells (buys) a stock when EJR announces a downgrade (upgrade) in the stock’s credit rating. The “issuer-paid CRA-based” strategy involves selling (buying) a stock following a rating downgrade (upgrade) from an issuer-paid CRA. A buy-and-hold of the S&P 500 index is included as a benchmark strategy. We measure the prof- itability for each trading strategy as follows. For each day in the time window [0, 5] surrounding each rating announcement on a firm, risk-adjusted return is the intercept (or alpha) from the Fama–French five-factor model estimated over a holding period. The firm-event alpha is then calculated as a simple average of the estimated alphas in the assessment window. We mul- tiply the firm-event alpha by (−1) to represent risk-adjusted returns to investors’ sales following a rating downgrade. Based on the firm’s rating change, we assign its event alpha to one of the trading strategies. Finally, we use firms’ inflation-adjusted market capitalization as weights and assess the strategies’ performance using one and two sample t-tests. Standard errors of the t-test for the mean and difference in means are in parentheses. ***, ** and * indicate statistical significance at the 1%, 5% and 10% levels, respectively. follow multiple strategies at a time, switch in and out of strategies, and may not trade on every signal consistent with a given strategy. Since holding data in S12 and 13F forms are available on a quarterly basis, we assume that institutional holding adjustments (and aggregate credit rating changes) happen on the last day of each quarter.15 15 In Section 4.2, we assume that trading occurs on the first day of each quarter. Our findings are robust. 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License NGUYEN ET AL. 17 ∑ First, given firm n’s aggregate rating change in quarter q, We form trading strategies and estimate their returns in calendar time at the institutional investor level as follows. ΔCCRn,q, and institutional investor i’s dollar net buy of the firm’s stock, nbi,n,q, we classify this firm into a specific strategy. Since we assume that an institutional investor can con- currently follow multiple strategies, a firm can be assigned to more than one trading strategy. We multiply the stock’s ΔCCRn,q < 0 and nbi,n,q < 0 to reflect stock returns to sales following a rating downgrade. Next, returns by (−1) if for each institutional investor i, we calculate portfolio returns for each strategy using the absolute values of stock net buy as weights.16 These portfolio returns are computed for each day over an investment horizon of k months starting ∑ at the end of the quarter. We repeat this process for each quarter in our sample period. Finally, the risk-adjusted return for each strategy is obtained using a pooled cross-sectional time series regression of the Fama–French five-factor model. We use a two-sample t-test to examine the difference in mean risk-adjusted returns between strategies. The strategy alphas are reported in Table 7. Generally, while all four trading strategies provide positive risk-adjusted profits for up to 12 months after credit rating announcements, the dynamic strategy that mimics the typical institu- tional response to credit rating adjustments yields the highest returns. For example, its 1-month annualized return is 17.54% and 16.73% for the mutual fund and institutional samples, respectively. While the return magnitude decreases with the holding period, it significantly outperforms all other strategies for the holding periods of at least 6 months. Among the other strategies, the strategy following issuer-paid CRAs’ credit announcements, while still significantly outperforming the buy-and-hold, yields the lowest returns. Overall, the results in Table 7 are consistent with our expectations that credit rating announcements have valuable information content and that the most value-relevant announcements are downgrades by the investor-paid EJR and upgrades by the issuer-paid CRAs. Our findings illustrate that institutional investors that dynamically change their trading behavior based on the advantages and disadvantages of credit rating information are likely to make abnormal profits beyond those of naïve trading strategies. Having said that, the prolonged abnormal profits are consistent with underreaction to credit rating information, especially to the information content of EJR’s credit downgrades and the issuer-paid CRAs’ credit upgrades. 3.4 Alternative institutional trading data In the preceding analysis of institutional trading, we rely on S12 and 13F quarterly data. In this section, we conduct the analysis using daily transaction-level data provided by the Abel Noser Corporation. G. Hu et al. (2018) describe several important features of Abel Noser’s institutional trading data. The dataset covers at least 12% of the total CRSP trading volume, 233 million transactions with $37 trillion in traded volume. It also records equity transactions by a large number of institutions from January 1999 to September 2011.17 Despite its limited availability, the data on institutional investors’ daily trading activities enable us to better capture their trading responses to credit rating adjustments.18 We winsorize institutional trading data at the 1st and 99th percentiles to minimize the effect of out- liers. After matching with our credit rating samples, we find 1126, 1259, 509 and 420 firms rated by EJR, S&P, Moody’s and Fitch, respectively. 16 As a robustness check, we calculate each strategy’s returns using stock returns and their associated institutional net buy values across all institutional investors in the quarter. Hence, we have one value-weighted return per strategy per day in a holding horizon. The unreported results are robust in both statistical significance and economic magnitude. 17 While Abel Noser arguably provides “cleaner” transaction-level data, we rely on S12 and 13F data in the main analysis, as Abel Noser does not provide data for research purposes after 2011. We note that our results are robust to the choice of the dataset. 18 Due to Abel Noser’s high level of coverage, several prior studies have used these data to investigate institutional trading behavior. G. Hu et al. (2018) summarize 55 publications that use these data. 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 18 NGUYEN ET AL. TA B L E 7 Trading strategy profitability—S12 and 13F samples Panel A: Mutual funds’ trading strategy profitability Holding periods (1) Dynamic 1 month 3 months 6 months 9 months 12 months 0.1754*** 0.1421*** 0.1242*** 0.0791*** 0.0608*** (0.0003) (0.0015) (0.0009) (0.001) (0.001) (2) Naïve 0.1398*** 0.1114*** 0.0974*** 0.0766*** 0.059*** (0.0099) (0.0034) (0.001) (0.001) (0.001) (3) EJR-based 0.1416*** 0.1158*** 0.1004*** 0.0803*** 0.0633*** (0.0003) (0.001) (0.0003) (0.0003) (0.0003) (4) Issuer-paid CRA-based 0.125*** 0.0986*** 0.086*** 0.0322*** 0.0384*** (0.0114) (0.0025) (0.0018) (0.0015) (0.0023) (5) S&P 500 index 0.033 0.033 0.033 0.033 (1)–(2) (1)—(3) (1)–(4) (1)–(5) (2)–(5) (3)–(5) (4)–(5) (0.0453) (0.0453) (0.0453) (0.0453) 0.0356*** 0.0307*** 0.0268*** 0.0025** (0.0099) (0.0037) (0.0011) 0.0338*** 0.0263*** 0.0238*** (0.0004) (0.0018) (0.0007) (0.0011) −0.0011 (0.0007) 0.033 (0.0453) 0.0018 (0.0011) −0.0025 (0.0111) 0.0504*** 0.0436*** 0.0382*** 0.0469*** 0.0224*** (0.0114) (0.0025) (0.0018) (0.0016) (0.0023) 0.1424*** 0.1091*** 0.0912*** 0.0461*** 0.0278*** (0.0045) (0.0041) (0.0025) (0.0025) (0.0025) 0.1068*** 0.0784*** 0.0643*** 0.0436*** 0.026*** (0.0109) (0.0051) (0.0015) (0.0015) (0.0015) 0.1085*** 0.0828*** 0.0674*** 0.0473*** 0.0303*** (0.0045) (0.0039) (0.0016) 0.092*** 0.0656*** 0.053*** (0.0172) (0.0042) (0.0034) (0.0015) −0.0008 (0.0034) (0.0023) 0.0054 (0.0034) Panel B: Institutional investors’ trading strategy profitability Holding periods (1) Dynamic 1 month 3 months 6 months 9 months 12 months 0.1673*** 0.1455*** 0.1091*** 0.0728*** 0.0618*** (0.0139) (0.0015) (0.0007) (0.0004) (0.0004) (2) Naïve 0.1175*** 0.1143*** 0.0898*** 0.0649*** 0.0535*** (0.0178) (0.0046) (0.0014) (0.0014) (0.0014) (3) EJR-based 0.1236*** 0.1191*** 0.0953*** 0.0727*** 0.062*** (0.0093) (0.0009) (0.0003) (0.0002) (0.0003) (4) Issuer-paid CRA-based 0.1202*** 0.0804*** 0.0678*** 0.0483*** 0.0384*** (0.0198) (0.0062) (0.0045) (0.0038) (0.0057) (5) S&P 500 index 0.033 0.033 0.033 0.033 0.033 (0.0453) (0.0453) (0.0453) (0.0453) (0.0453) (1)–(2) 0.0498** 0.0312*** 0.0194*** 0.0078*** 0.0084*** (0.0226) (0.0048) (0.0014) (0.0014) (0.0014) (Continues) 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License NGUYEN ET AL. TA B L E 7 (Continued) 19 Panel B: Institutional investors’ trading strategy profitability Holding periods 1 month 3 months 6 months 9 months 12 months (1)–(3) (1)–(4) (1)–(5) (2)–(5) (3)–(5) (4)–(5) 0.0438*** 0.0264*** 0.0138*** 0.0001 (0.0168) (0.0017) (0.0005) (0.0004) −0.0001 (0.0005) 0.0471** 0.0651*** 0.0414*** 0.0245*** 0.0235*** (0.0204) (0.0063) (0.0045) (0.0038) (0.0057) 0.1343*** 0.1125*** 0.0761*** 0.0397*** 0.0288*** (0.0147) (0.004) (0.0024) (0.0023) (0.0023) 0.0845*** 0.0813*** 0.0567*** 0.0319*** 0.0205*** (0.0183) (0.0059) (0.0018) (0.0018) (0.0018) 0.0905*** 0.0861*** 0.0623*** 0.0397*** 0.0289*** (0.0103) (0.0039) (0.0015) (0.0015) 0.0872*** 0.0474*** 0.0347*** 0.0153** (0.0298) (0.0095) (0.0077) (0.0077) (0.0023) 0.0054 (0.0077) Note: This table reports and compares annualized risk-adjusted returns on four institutional trading strategies. “Dynamic” is a strategy that sells a stock when it receives an EJR’s negative rating adjustment and buys the stock when its rating is upgraded by an issuer-paid CRA. The “naïve” strategy is simply to sell (buy) a stock following a negative (positive) signal from any rating agency. For the “EJR-based” strategy, an investor sells (buys) a stock when EJR announces a downgrade (upgrade) in the stock’s credit rating. The “issuer-paid CRA-based” strategy involves selling (buying) a stock following a rating downgrade (upgrade) from an issuer-paid CRA. A buy-and-hold of the S&P 500 index is included as a benchmark strategy. We measure the prof- itability for each trading strategy as follows. First, based on a firm’s aggregate rating change in a quarter and an institutional investor’s net buy of the firm’s stock, we classify it to a specific strategy. We multiply the stock’s returns by (−1) if the quar- terly rating change is negative and the investor exhibits a net sale of the stock to reflect stock returns to sales following a rating downgrade. Next, for each institutional investor, we calculate portfolio returns for each strategy using the absolute values of stock net buy as weights. These portfolio returns are computed for each day over an investment horizon starting at the end of the quarter. We repeat this process for each quarter in our sample period. Finally, the risk-adjusted return for each strat- egy is obtained using a pooled regression of the Fama–French five-factor model. We use a two-sample t-test to examine the difference in mean risk-adjusted returns between strategies. Robust standard errors are in parentheses. ***, ** and * indicate statistical significance at the 1%, 5% and 10% levels, respectively. 3.4.1 Institutional trading response to credit rating announcements We investigate abnormal institutional trading surrounding a stock’s credit rating adjustments in the time window [0, 5].19 Day 0 is the date of a credit rating event. We consider institutions’ trading activities up to five days after the credit rating adjustment to account for investors’ potential gradual reactions while also avoiding confounding effects that can appear in longer windows. With detailed transaction data, we calculate institutional investor i’s abnormal net buy of stock n over the [0, 5] day window around a credit rating announcement, AN_NBi,n,w, as follows: +5∑ AN_NBi,n,w = k = 0 (AN_nbi,n,t+k − average ANnbi,n,t), (6) 19 We also consider two different time windows [−2, 5] and [−2, 1] for robustness. The purpose is to account for institutional investors’ pre-reactions because of potential information leakage (e.g., Bhattacharya et al., 2019). Our results are robust. 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 20 NGUYEN ET AL. where AN_nbi,n,t is the daily dollar volume bought minus daily dollar volume sold scaled by the stock’s 1-month-lagged market capitalization as shown in equation (7). average AN_nbi,n,t is the average value of AN_nbi,n,t in the period from day t − 371 to day t − 6 prior to the announcement date t as shown in equation (8). AN_nbi,n,t = BOUGHTi,n,t − SOLDi,n,t MARKET_CAPi,n,m−1 , average AN_nbi,n,t = ∑−371 k = −6 AN_nbi,n,t+k 365 . We then estimate the following model, which is similar to equation (4), for each of the paired samples: AN_NBi,n,w = 𝛼 + 𝛽1NEGn,t + 𝛽2POSn,t + 𝛽3NEGn,t∗EJRn,t + 𝛽4POSn,t∗EJRn,t + 𝛽5EJRn,t + 𝛾kCONTROLSn,t + t∑ 1 𝜃tQuarterFEt . + i∑ 1 𝛿iInvestorFEi + 𝜑nFirmFEn + 𝜀i,n,w k∑ 1 n∑ 1 (7) (8) (9) Depending on the numeric change in the CCR scale, ΔCCRn,t, for firm n on the adjustment date t, we define NEGn,t as |ΔCCRn,t| if ΔCCRn,t < 0 and zero if ΔCCRn,t > 0,and POSn,t as ΔCCRn,t if ΔCCRn,t > 0,and zero if ΔCCRn,t < 0. Therefore, an increase in NEGn,t (POSn,t) represents an absolute increase in credit rating downgrade (upgrade) for firm n at credit event t. Control variables and fixed effects are described in equation (4). We present the results in Table 8. The results are qualitatively similar to those based on S12 and 13F data— institutional investors react asymmetrically to credit rating announcements made by EJR and issuer-paid CRAs.20 Columns 1 and 2 show the results for firms jointly rated by EJR and S&P. Institutional investors’ net buy increases significantly around S&P’s positive rating adjustments. The POS coefficient is positive and significant across all speci- fications. The 0.1655 basis point coefficient in column 2 is equivalent to an average increase of $316,914 in abnormal net buy over [0, 5] days around the S&P’s one-notch rating upgrade announcements. However, the insignificant F-test results for the overall impact of rating upgrades by EJR, that is, the sum of POS and EJR*POS coefficients, indicate that institutional investors are unresponsive to EJR’s positive rating changes. We document opposite results for rating downgrades. The EJR*NEG coefficient is negative and statistically and eco- nomically significant across all models. For example, the −0.1191 coefficient of EJR*NEG in column 2 shows that a one-notch downgrade announcement by EJR is equivalent to a decrease of $228,063 in abnormal institutional net buy over the [0, 5] day window compared to a similar announcement by S&P. The F-test for the overall impact of EJR downgrades, that is, the sum of NEG and EJR*NEG coefficients, indicates that the effect is strong statistically and economically. We find similarly asymmetric responses for firms jointly rated by S&P and Moody’s in columns 3 and 4. For exam- ple, the POS coefficient of 0.4078 in column 4 indicates that abnormal institutional net buy, on average, increases by $281,586 over the [0, 5] day window surrounding a credit rating upgrade by Moody’s. The F-test results for the sum of NEG and EJR*NEG coefficients in column 4 indicate that EJR’s downgrades, on average, are associated with a signif- icant decrease of $111,378 in abnormal institutional net buy over the [0, 5] day window. The results in columns 5 and 6 do not exhibit any robust and significant difference in the response of institutional investors around credit rating changes for firms covered by both EJR and Fitch. All coefficients of interest are statistically insignificant. We further investigate investor reactions to Fitch ratings in Section 3.5. 20 In order to assess the potential impact of changes in the sample period, we perform the analysis of S12 and 13F data on two sub-periods: 1999–2011 (to match Abel Noser data coverage) and 2012–2017. The results are presented in Tables A11 and A12 in the Online Appendix and are qualitatively similar across sub-periods. 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License NGUYEN ET AL. 21 TA B L E 8 Abnormal trading responses to credit rating adjustments—Abel Noser sample EJR versus S&P EJR versus Moody’s EJR versus Fitch (1) (2) (3) (4) (5) (6) Intercept 0.3095 −0.4884 −0.3379 −0.4840 0.8414 −0.4015 NEG POS (1.8238) (2.1424) (4.9528) (4.7255) (1.7814) (2.0034) 0.0470* 0.0538 0.0966 0.0925 0.0285 0.0259 (0.0266) (0.0408) (0.1154) (0.1302) (0.0233) (0.0448) 0.2450*** 0.1655*** 0.3324** 0.4078*** 0.074** 0.0818 (0.0300) (0.0498) (0.1319) (0.1387) (0.0353) (0.0711) EJR×NEG −0.1014*** −0.1191** −0.1984* −0.2538* −0.0295 −0.0636 EJR×POS −0.2125*** −0.1232** −0.2723** −0.4453*** 0.0023 −0.0436 (0.0322) (0.0478) (0.1204) (0.1342) (0.0349) (0.0568) (0.0360) (0.0563) (0.1387) (0.1481) (0.0443) (0.0807) EJR 0.1316*** 0.1254** 0.2429** 0.4005*** 0.0507 0.0744 (0.0365) (0.0493) (0.1234) (0.1282) (0.0491) (0.0738) Control variables: No Yes No Yes No Yes F-tests: NEG + EJR×NEG −0.0544*** −0.0654** −0.1018** −0.1613*** −0.0011 −0.0377 POS + EJR×POS 0.0325 0.0423 0.0601 −0.0375 0.0763*** 0.0381 (0.0209) (0.0273) (0.0468) (0.0487) (0.0279) (0.0369) (0.0216) (0.0272) (0.0496) (0.0523) (0.0285) (0.0393) Fixed effects: Investor FE Firm FE Quarter FE N Adj. R2 Yes Yes Yes Yes No Yes Yes Yes Yes Yes No Yes Yes Yes Yes Yes No Yes 429,268 304,731 114,354 93,867 207,587 133,788 0.010 0.004 0.016 0.010 0.006 0.008 Note: The table reports OLS regression results for institutional investors’ abnormal trading around credit rating adjustments announced by EJR and issuer-paid CRAs. The dependent variable, defined in equation (6), is an institutional investor’s abnor- mal net buy of a stock over the [0, 5] day window. We define NEG as the absolute value of a rating downgrade and zero otherwise and POS as the value of a rating upgrade and zero otherwise. Therefore, an increase in NEG (POS) represents an absolute increase in a firm’s downgrade (upgrade). EJR is a dummy variable that equals one for EJR’s credit rating announce- ments and zero otherwise. Detail descriptions of firm-level control variables are described in Appendix B. Standard errors in parentheses are adjusted for heteroskadisticity and clustering at the firm and quarter levels. ***, ** and * indicate statistical significance at the 1%, 5% and 10% levels, respectively. 3.4.2 Institutional trading profits in response to credit rating announcements We now turn to assess institutional trading strategy profitability based on Abel Noser data. First, we use the Fama–French five-factor model, as shown in equation (5), to estimate risk-adjusted returns, that is, alphas, for each day in the [0, 5] window around a credit rating announcement. We multiply a daily alpha by (−1) if an institutional investor’s trades on that day represent a net sale. We calculate the average alpha over the assessment window using the institutional investor’s absolute daily net trade values as weights. We then assign this firm-institution-event alpha to different trading strategies based on the institutional net buy over the assessment window and the credit rating 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 22 NGUYEN ET AL. signal. Finally, we assess the performance of these strategies using firms’ market capitalization as weights in one and two-sample t-tests. Strategy trading profits and their significance are presented in Table 9. The results are consistent with those based on S12 and 13F data. The dynamic strategy that follows EJR’s negative signals and other CRAs’ positive signals earns the highest returns, compared to the other strategies. For example, for the 1-month investment horizon, the dynamic strategy outperforms the other three strategies by an annualized value-weighted risk-adjusted return ranging from 10.17% to 10.95%. This outperformance is approximately twice as much as the corresponding outperformance of notional strategies. Although this outperformance decreases with the investment horizon, it is still statistically signif- icant for up to 9 months. Among the other three active strategies, following EJR’s signals alone apparently generates the best returns, whereas following issuer-paid CRAs’ signals only yields the least profits. We also compare the four trading strategies to a passive strategy—a buy-and-hold annual return of the S&P 500 index. We observe that all trading strategies outperform the index for up to 6 months, except the strategy following issuer-paid CRAs’ rating announcements. 3.5 The case of Fitch ratings We note that investor reactions around credit rating changes for firms jointly rated by EJR and Fitch are different from those covered by S&P and Moody’s. In the S12 and 13F samples, we observe significant reactions to Fitch upgrades and no significant reactions to EJR downgrades. We document essentially no significant reactions in Able Noser data. This has prompted us to investigate this further.21 Fitch has traditionally held a smaller market share relative to Moody’s and S&P (Becker & Milbourn, 2011; Livingston & Zhou, 2016). This may have influenced both their rating behavior (Beatty et al., 2019; Hirth, 2014) and investor reaction. Our empirical analysis suggests that Fitch differs in rating behavior from other issuer-paid CRAs. First, as reported in Table 5, not only does Fitch lead EJR in positive events (as expected) but is also the only issuer-paid CRA to lead EJR in negative announcements (although the difference is not statistically significant). Furthermore, our unreported analysis also shows that Fitch leads S&P and Moody’s in both positive and negative announcements. We believe this is consistent with Fitch providing more timely rating announcements in order to increase their market share. Second, we look at the information content of Fitch announcements by constructing two additional trading strategies: the “Fitch-based strategy”—buying on Fitch upgrades and selling on Fitch downgrades (for the sake of completeness, we also create “S&P-based strategy” and “Moody’s-based strategy”), and the “modified dynamic strategy”—buying on credit upgrades by the “Big Three” and selling on Fitch downgrades. Our unreported results show that the Fitch-based strategy not only outperforms a simple buy-and-hold of the S&P 500 index but also produces bet- ter returns than the issuer-paid CRA-based strategy, particularly over longer time periods. This suggests that Fitch’s announcements actually have higher information content than other issuer-paid CRAs. The modified dynamic strategy is the second-best performing strategy, suggesting that Fitch’s negative announcements have substantial information content. However, the “dynamic strategy”—buying on positive issuer-paid CRA announcements and selling on EJR’s negative announcements—yields the best returns, which is consistent with our main hypothesis. Finally, we investigate institutional investors’ reactions to Fitch’s announcements in greater detail. In the main anal- ysis, institutions do not appear to react significantly to either positive or negative announcements in the sample of firms jointly rated by Fitch and EJR, despite evidence that both CRAs’ announcements have significant information content. We posit that as Fitch leads EJR in negative signals (although insignificantly), the lack of significant reaction to EJR’s negative announcements may be due to the dilution of investors’ reaction to both Fitch and EJR’s announce- ments. Investors do not react to Fitch’s announcements in a significant way (even though these announcements have significant informational content), and this still weakens investors’ reactions to subsequent announcements by EJR. 21 We thank an anonymous referee for this suggestion. 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License NGUYEN ET AL. 23 TA B L E 9 Trading strategy profitability—Abel Noser sample 1 month 3 months 6 months 9 months 12 months (3) EJR-based 0.0609*** 0.047*** 0.0476*** 0.0129 Holding periods (1) Dynamic (2) Naïve (4) Issuer-paid CRA-based (5) S&P 500 index (1)–(2) (1)–(3) (1)–(4) (1)–(5) (2)–(5) (3)–(5) (4)–(5) 0.1626*** 0.1098*** 0.0728*** 0.0714** (0.0223) 0.0534* (0.0317) (0.0188) 0.0468* (0.0267) (0.0224) 0.0188 (0.0129) (0.033) 0.0078 (0.0227) (0.0113) 0.0531* (0.0297) −0.0163 (0.0621) (0.0127) 0.0395 (0.0254) −0.0163 (0.0621) 0.1092*** 0.063* (0.0388) (0.0327) (0.0114) −0.0213 (0.026) −0.0163 (0.0621) 0.054*** (0.0171) 0.1017*** 0.0628*** 0.0253 (0.025) (0.0227) (0.0188) (0.0217) −0.0012 (0.0265) −0.0163 (0.0621) 0.0635** (0.028) 0.0584* (0.0309) 0.1095*** 0.0703*** 0.0941*** 0.0726** (0.0306) (0.0262) (0.0275) (0.0287) 0.0297 (0.0496) 0.0008 (0.0451) 0.0049 (0.0909) 0.0184 (0.0463) −0.0163 (0.0621) 0.0289 (0.0515) 0.0248 (0.1036) 0.0113 (0.0525) 0.1789*** 0.1261*** 0.0891*** 0.0877*** 0.046 (0.0232) (0.0195) (0.0226) (0.0332) 0.0697** 0.0631** 0.0351*** 0.0241 (0.0323) (0.0272) (0.013) (0.0227) 0.0772*** 0.0633*** 0.0639*** 0.0292 (0.0129) 0.0694* (0.0376) (0.0137) 0.0558 (0.0382) (0.0116) −0.005 (0.0434) (0.0218) 0.0151 (0.053) (0.0497) 0.0171 (0.0451) 0.0212 (0.091) 0.0347 (0.0617) Note: This table reports and compares annualized risk-adjusted returns on four institutional trading strategies. “Dynamic” is a strategy that sells a stock when it receives an EJR’s negative rating adjustment and buys the stock when its rating is upgraded by an issuer-paid CRA. The “naïve” strategy is simply to sell (buy) a stock following a negative (positive) signal from any rating agency. For the “EJR-based” strategy, an investor sells (buys) a stock when EJR announces a downgrade (upgrade) in the stock’s credit rating. The “issuer-paid CRA-based” strategy involves selling (buying) a stock following a rating downgrade (upgrade) from an issuer-paid CRA. A buy-and-hold of the S&P 500 index is included as a benchmark strategy. We measure the prof- itability for each trading strategy as follows. For each day in the time window [0, 5] surrounding each rating announcement on a firm, risk-adjusted return is the intercept (or alpha) from the Fama–French five-factor model estimated over a holding period. We multiply a daily alpha by (−1) if an institutional investor’s trades on that day represent a net sale. Next, we calculate the average alpha over the assessment window using the institutional investor’s absolute daily net trade values as weights. We then assign a stock’s event alpha to different trading strategies based on its institutional net buy over the assessment window and the credit rating signal. Finally, we use firms’ inflation-adjusted market capitalization as weights and assess the strate- gies’ performance using one and two sample t-tests. Standard errors of the t-test for the mean and difference in means are in parentheses. ***, ** and * indicate statistical significance at the 1%, 5% and 10% levels, respectively. 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 24 NGUYEN ET AL. To investigate this, we remove negative announcements led by Fitch. Our unreported results are consistent with our expectations, that is, investors’ reaction to EJR’s negative announcements becomes negative and significant in three out of four specifications, which is consistent with the results reported in panels A and B of Table 4 for EJR and S&P and EJR and Moody’s. 4 ROBUSTNESS TESTS 4.1 Hedge trading strategies Although our findings on trading strategy profits suggest that investors generally underreact to the information con- tent of credit rating signals, some investors may, in fact, overreact to rating events, which may result in subsequent profits for contrarian trading strategies (Ellul et al., 2011). Alternatively, CRAs could provide upward-biased credit ratings to some relationship firms (e.g., Baghai & Becker, 2018), and some institutional investors may be sophisticated enough to detect these overrated firms and respond in the opposite way around rating announcements, which sub- sequently earns them abnormal profits. We account for the effect of this potential contrarian trading strategy on the performance of our main strategies as follows. We follow the steps in Section 3.3.2 to construct portfolios that are opposite to our main strategies. For example, an institutional investor that increases its net holding of a stock with an aggregate downgrade by EJR and decreases its net holding of a stock with an aggregate upgrade by issuer-paid CRAs is considered to be a dynamic contrarian strategy. A dynamic hedge portfolio is then defined as longing the dynamic portfolio and shorting the dynamic contrarian portfolio. Finally, we estimate the four hedge portfolios’ risk-adjusted returns using the Fama–French five-factor model and report the results in Table A1. The dynamic hedge portfolio remains the best performer, and its outperformance relative to the other hedge portfolios is qualitatively similar in magnitude and statistical significance to the results in Table 7. 4.2 Alternative assumptions on the timing of trades In the main analysis of quarterly S12 and 13F data, we assume that credit rating adjustments and fund-stock holding changes happen on the final day of each quarter. In this robustness check, we make an alternative assumption that these changes occur on the first day of each quarter. Trading strategy profitability is then re-estimated, and the results are presented in Table A2 in the Online Appendix. The results are consistent with those in the main analysis that the dynamic strategy significantly outperforms all other strategies considered. We also consider alternative event windows in the analysis of daily Abel Noser data: [−2, 1] and [−2, 5] trading days. First, these time windows include the two days prior to credit rating adjustments to control for potential information leakage before official rating adjustments (Bhattacharya et al., 2019). Second, we also choose short time windows to control for any effect of clusters of rating signals (e.g., Alsakka & ap Gwilym, 2012; Gande & Parsley, 2005; Vu et al., 2015). In other words, shorter time windows enable us to avoid any information contamination problems caused by the appearance of other information in the financial market in longer time windows. The results for the two alter- native event windows are presented in Table A3 in the Online Appendix and are consistent with the main findings. Institutional investors still exhibit asymmetric trading behavior to the issuer- and investor-paid credit rating signals, abnormally buying on issuer-paid CRAs’ positive rating adjustments and abnormally selling on EJR’s negative rating adjustments in both alternative time windows. All four active trading strategies earn significant profits in similar patterns as in Table 4. They outperform the buy- and-hold return of the S&P 500 index for up to a 9-month horizon. Most importantly, the dynamic trading strategy is the best performer over all other strategies. The robust results of institutional trading strategies constructed sur- rounding alternative event windows of [−2, 1] and [−2, 5] days are reported in Table A4 in the Online Appendix. We 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License NGUYEN ET AL. 25 also report the results for notional trading strategies for these alternative windows, and the results exhibit similar patterns as shown in Table A5 in the Online Appendix. 4.3 Raw institutional trading Our next robustness check analyzes “raw” reactions (i.e., unadjusted for the average of past trading activities) of insti- tutional investors to credit rating announcements. We perform the analysis on both S12 and 13F quarterly data (Table A6) and in the [0, 5] day window on Abel Noser daily data (Table A7). The results are highly consistent with the main findings that institutional investors tend to abnormally sell stocks of firms with EJR’s negative rating announcements but ignore positive ones; however, their net buy increases substantially surrounding issuer-paid CRAs’ positive rating announcements. 4.4 Combined issuer-paid CRA In another robustness check, we treat all three issuer-paid CRAs as a combined issuer-paid CRA. We then investigate institutional investor’s trading activities surrounding negative and positive rating signals by EJR and the combined issuer-paid CRA. The results are reported for the S12, 13F and Abel Noser Sample in Table A8 in the Online Appendix. The results are consistent: Institutional investors tend to abnormally sell stocks surrounding negative signals issued by EJR and abnormally buy stocks surrounding positive signals issued by the combined issuer-paid CRA. 4.5 Excluding non-trading observations In our main analysis, abnormal net buy is set at zero if institutional investors have no trading activities surrounding credit rating adjustments. In this final robustness check, we exclude these non-trading observations. We find robust results in Table A9 for S12 and 13F data and Table A10 for Abel Noser data in the Online Appendix. The results are robust. After excluding non-trading observations, institutional investors still have asymmetric responses, abnormally increasing (decreasing) stock holdings surrounding positive (negative) rating signals by issuer- (investor-) paid CRAs. Overall, these robustness tests confirm our main findings that institutional investors who have advanced trading skills selectively react to credit rating signals from different sources based on their relative informational values. 5 CONCLUSION This study investigates institutional investors’ responses to credit rating adjustments announced by the investor-paid EJR and the “Big Three” issuer-paid CRAs. In recent years, traditional issuer-paid CRAs have faced criticism regard- ing lack of timeliness in negative signals in many infamous scandals such as Enron (2001), WorldCom (2002) and Lehman Brothers (2008). Meanwhile, investor-paid CRAs, particularly EJR, have built a good reputation regarding the timeliness of their negative rating adjustments. As a result, institutional investors with advanced trading skills and sophistication (Puckett & Yan, 2011) are likely to dynamically switch between following investor- and issuer-paid CRAs based on the timeliness of credit rating information. We document considerable asymmetries in institutional investors’ responses to issuer- and investor-paid CRA announcements. They react by abnormally selling following EJR’s negative signals and abnormally buying following issuer-paid CRAs’ positive signals. The results differentiate our paper from the existing literature. Several prior stud- ies show that institutional investors simply tend to be more sensitive to negative rather than positive signals. Our study 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 26 NGUYEN ET AL. finds that institutional investors, as professional players, have their own responses to the lack of timeliness criticism by following investor-paid CRA’s negative signals. They still maintain faith in positive issuer-paid rating announce- ments due to no evidence of their delays. The results are robust across different databases from which the institutional investors’ trading activities are extracted. We also document that a dynamic trading strategy based on selling following the investor-paid CRA’ negative sig- nals and buying following issuer-paid CRAs’ positive signals produces superior returns. While any investor can take advantage of these strategies, institutional investors evidently achieve higher returns. Although we document the highly dynamic behavior of institutions in responding to important market signals, our results imply that market partic- ipants tend to underreact to positive signals by issuer-paid CRAs and negative signals by investor-paid CRA. Therefore, the information content of these signals is not fully reflected in prices at the announcement time, thus leading to opportunities to earn abnormal returns by trading following these signals. As further information in support of CRA creditworthiness predictions is released, abnormal returns are generally dissipated. Our results are consistent with this view—abnormal returns decrease as holding periods increase. The difference between dynamic strategy and naïve strategy returns becomes substantially smaller in the 12-month holding period. Given that discrepancies in credit rating quality between issuer-paid CRAs and investor-paid CRAs are not limited only to the US bond market (e.g., X. Hu et al., 2019), we believe there are some interesting avenues for future research such as whether institutional investors also respond asymmetrically to credit rating announcements by issuer-paid and investor-paid CRAs in an international setting (e.g., China); if so, whether such asymmetric responses are conditional on some firm or market level shocks. ACKNOWLEDGMENTS We would like to thank Professor Pope, the JBFA editor and an anonymous referee for their constructive and valu- able comments. Our thanks are also to participants at the New Zealand Finance Colloquium (NZFC) for their useful comments. Open access publishing facilitated by Massey University, as part of the Wiley - Massey University agreement via the Council of Australian University Librarians. DATA AVAILABILITY STATEMENT The data that support the findings of this study are available from third parties. Restrictions apply to the availability of these data, which were used under license for this study. Data are available from the authors with the permission of third parties. ORCID Alexander Molchanov https://orcid.org/0000-0003-0133-3811 REFERENCES Acharya, V., & Pedersen, L. (2005). Asset pricing with liquidity risk. Journal of Financial Economics, 77, 375–410. Alsakka, R., & ap Gwilym, O. (2012). Foreign exchange market reactions to sovereign credit news. Journal of International Money and Finance, 31, 845–864. Amihud, Y., & Goyenko, R. (2013). Mutual fund’s R2 as predictor of performance. Review of Financial Studies, 26(3), 667–694. Baghai, R. P., & Becker, B. (2018). Non-rating revenue and conflicts of interest. Journal of Financial Economics, 127(1), 94–112. Baker, H. K., & Mansi, S. A. (2002). Assessing credit rating agencies by bond issuers and institutional investors. Journal of Business Finance and Accounting, 29(9), 1367–1398. Beatty, A., Gillette, J., Petacchi, R., & Weber, J. (2019). Do rating agencies benefit from providing higher ratings? Evidence from the consequences of municipal bond ratings recalibration. Journal of Accounting Research, 57(2), 323–354. Beaver, W. H., Shakespeare, C., & Soliman, M. T. (2006). Differential properties in the ratings of certified versus non-certified bond-rating agencies. Journal of Accounting and Economics, 43, 303–334. Becker, B., & Milbourn, T. (2011). How did increased competition affect credit ratings? Journal of Financial Economics, 101(3), 493–514. 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License NGUYEN ET AL. 27 Bernile, G., Sulaeman, J., & Wang, Q. (2015). Institutional trading during a wave of corporate scandals: “Perfect Payday”? Journal of Corporate Finance, 34(C), 191–209. Berwart, E., Guidolin, M., & Milidonis, A. (2019). An empirical analysis of changes in the relative timeliness of issuer-paid vs. investor-paid ratings. Journal of Corporate Finance, 59, 88–118. Bhattacharya, U., Wei, & Xia, H. (2019). Follow the money: Investor trading around investor-paid credit rating changes. Journal of Corporate Finance, 58, 68–91. Bolton, P., Freixas, X., & Shapiro, J. (2012). The credit ratings game. Journal of Finance, 67(1), 85–111. Brogaard, J., Koski, J. L., & Siegel, A. F. (2019). Do upgrades matter? Evidence from trading volume. Journal of Financial Markets, 43, 54–77. Bruno, V., Cornaggia, J., & Cornaggia, K. J. (2016). Does regulatory certification affect the information content of credit ratings? Management Science, 62(6), 1578–1597. Cantor, R., ap Gwilym, O., & Thomas, S. (2007). The use of credit ratings in investment management in the US and Europe. Journal of Fixed Income, 17(2), 13–26. Chemmanur, T. J., He, S., & Hu, G. (2009). The role of institutional investors in seasoned equity offerings. Journal of Financial Economics, 94(3), 384–411. Chemmanur, T. J., Li, Y., Xie, J., & Zhu, Y. (2016). Noisy signalling through open market share repurchase progrqams and infor- mation production by institutions (Working Paper. 2018 Asian Finance Association (AsianFA) Conference, 2016 Financial Intermediation Research Society (FIRS) conference, 2015 European Finance Association (EFA) conference). https://ssrn. com/abstract=2543397 Chen, S. S., Chen, H. Y., Chang, C. C., & Yang, S. L. (2016). The relation between sovereign credit rating revisions and economic growth. Journal of Banking and Finance, 64, 90–100. Cornaggia, J., & Cornaggia, K. (2013). Estimating the costs of issuer-paid credit ratings. Review of Financial Studies, 26(9), 2229– 2269. Cremers, K. J. M., & Petajisto, A. (2009). How active is your fund manager? A new measure that predicts performance. Review of Financial Studies, 22(9), 3329–3365. Dimitrov, V., Palia, D., & Tang, L. (2015). Impact of the Dodd-Frank act on credit ratings. Journal of Financial Economics, 115, 505–520. Drago, D., & Gallo, R. (2016). The impact and the spillover effect of a sovereign rating announcement on the euro area CDS market. Journal of International Money and Finance, 67, 264–286. Ellul, A., Jotikasthira, C., & Lundblad, C. T. (2011). Regulatory pressure and fire sales in the corporate bond market. Journal of Financial Economics, 101, 596–620. Fama, E. F., & French, K. R. (2015). A five-factor asset pricing model. Journal of Financial Economics, 116, 1–22. Gande, A., & Parsley, D. (2005). New spillovers in the sovereign debt market. Journal of Financial Economics, 75, 691–734. Griffin, J. M., Nickerson, J., & Tang, D. Y. (2013). Rating shopping or catering? An examination of the response to competitive pressure for CDO credit ratings. Review of Financial Studies, 26(9), 2270–2310. He, G. (2021). Credit rating, post-earnings-announcement drift, and arbitrage from transient institutions. Journal of Business Finance & Accounting, 48, 1434–1467. He, J., Qian, J., & Strahan, P. E. (2012). Are all ratings created equal? The impact of issuer size on the pricing of mortgage-backed securities. Journal of Finance, 67(6), 2097–2137. Henry, D., Nguyen, L., & Pham, V. H. (2017). Institutional trading before dividend reduction announcements. Journal of Financial Markets, 36, 40–55. Heron, R. A., & Lie, E. (2009). What fraction of stock option grants to top executives have been backdated or manipulated. Management Science, 55, 513–525. Hirth, S. (2014). Credit rating dynamics and competition. Journal of Banking and Finance, 29, 100–112. Hu, X., Huang, H., Pan, Z., & Shi, J. (2019). Information asymmetry and credit rating: A quasi-natural experiment from China. Journal of Banking and Finance, 106, 132–152. Hu, G., Jo, K. M., Wang, Y. A., & Xie, J. (2018). Institutional trading and Abel Noser data. Journal of Corporate Finance, 52, 143– 167. Jagolinzer, A. D., Larcker, D. F., & Taylor, D. J. (2011). Corporate governance and the information content of insider trades. Journal of Accounting Research, 49(5), 1249–1274. Joe, D. Y., & Oh, F. D. (2018). Spillover effects within business groups: The case of Korean chaebols. Management Science, 64(3), 1396–1412. Johnson, R. (2004). Rating agency actions around the investment-grade boundary. Journal of Fixed Income, 13, 25–38. Kedia, S., Rajgopal, S., & Zhou, X. (2017). Large shareholders and credit ratings. Journal of Financial Economics, 124, 632–653. Livingston, M., & Zhou, L. (2016). Information opacity and Fitch bond ratings. Journal of Financial Research, 39(4), 329–357. Puckett, A., & Yan, X. (2011). The interim trading skills of institutional investors. Journal of Finance, 66(2), 601–633. 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 28 NGUYEN ET AL. Ramsay, J. (2011). Testimony concerning oversight of the credit rating agencies Post-Dodd-Frank. https://www.sec.gov/news/ testimony/2011/ts072711jr.htm Skreta, V., & Veldkamp, L. (2009). Ratings shopping and asset complexity: A theory of ratings inflation. Journal of Monetary Economics, 56, 678–695. Vu, H., Alsakka, R., & ap Gwilym, O. (2015). The credit signals that matter most for sovereign bond spreads with split rating. Journal of International Money and Finance, 53, 174–191. Xia, H. (2014). Can investor-paid credit rating agencies improve the information quality of issuer-paid rating agencies? Journal of Financial Economics, 111, 450–468. SUPPORTING INFORMATION Additional supporting information can be found online in the Supporting Information section at the end of this article. How to cite this article: Nguyen, Q. M. P., Do, H. X., Molchanov, A., Nhut, L., & Nguyen, N. H. (2023). Asymmetric trading responses to credit rating announcements from issuer- versus investor-paid rating agencies. Journal of Business Finance & Accounting, 1–29. https://doi.org/10.1111/jbfa.12686 APPENDIX APPENDIX A: NUMERIC TRANSFORMATION OF ALPHANUMERICAL RATING CODES Investment grade Speculative grade Credit eventsa Rating AAA (Aaa) AA+ (Aa1) AA (Aa2) AA- (Aa3) A+ (A1) A (A2) A- (A3) BBB+ (Baa1) BBB (Baa2) BBB− (Baa3) Score 22 21 20 19 18 17 16 15 14 13 Rating BB+ (Ba1) BB (Ba2) BB− (Ba3) B+ (B1) B (B2) B− (B3) CCC+ (Caa1) CCC (Caa2) CCC− (Caa3) CC (Ca) C SD, D Score 12 11 10 9 8 7 6 5 4 3 2 1 Single upgrade Positive outlook Positive developing Stable Negative developing Negative outlook Single downgrade Score 1 0.5 0.25 0 −0.25 −0.5 −1 aSingle upgrade (downgrade) is a credit rating announcement when a rating agency adjusts the firm’s credit rating by one letter rating higher (lower; e.g., up from AA+ to AAA or down from AA+ to AA). A positive (negative) outlook is a credit rating review when a CRA adjusts its short-term expectations about the firm from being stable to positive (negative). A positive (negative) developing is a credit rating signal when a CRA adjusts its long-term expectations about the firm from being stable to positive (negative). 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License NGUYEN ET AL. 29 APPENDIX B: FIRM-LEVEL VARIABLE DEFINITIONS AND DATA SOURCES Variable Ln(MV) ROA Description Data source The natural log of total market capitalization in the quarter CRSP The ratio of operating income before depreciation to total COMPUSTAT assets in the quarter IDIO_RISK The standard deviation of residual returns from the Kenneth R. French & Fama–French three-factor model using daily stock returns from day t − 31 to day t − 1 CRSP Z-SCORE Alman’s Z-score that presents the probability that a firm will COMPUSTAT go into bankruptcy within 2 years ANALYST_COVERAGE The average number of analysts covering a firm in the quarter CRSP Ln(AGE) The natural log of number of years since a firm’s first CRSP appearance on CRSP database INTEREST_COVERAGE The ratio of earnings before interest, tax and depreciation and COMPUSTAT amortization to total interest expense in the quarter LEVERAGE The ratio of sum of long-term debt and debt in current COMPUSTAT liabilities to total assets in the quarter S&P_500 A binary variable that equals one if a firm is included in the S&P S&P 500 Index 500 list HIGH_TECH A binary variable that equals one if a firm’s Standard Industry Classification (SIC) code is between 7370 and 7379 (Heron and Lie, 2009) and zero otherwise CRSP 14685957, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jbfa.12686 by Test, Wiley Online Library on [15/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
10.1002_jcv2.12126
Received: 23 May 2022- Accepted: 28 November 2022 O R I G I N A L A R T I C L E DOI: 10.1002/jcv2.12126 Social, emotional and behavioural difficulties associated with persistent speech disorder in children: A prospective population study Yvonne Wren1,2 | Emma Pagnamenta3 | Faith Orchard4 Alan Emond5 | Kate Northstone6 | Laura Louise Miller7 | Susan Roulstone8 | Tim J. Peters2 | 1Bristol Speech and Language Therapy Research Unit, North Bristol NHS Trust, Bristol, UK 2Bristol Dental School, University of Bristol, Bristol, UK 3School of Psychology and Clinical Language Sciences, University of Reading, Reading, UK 4School of Psychology, University of Sussex, East Sussex, UK 5Centre for Academic Child Health, Bristol Medical School, Bristol, UK 6Population Health Sciences, Bristol Medical School, Oakfield House, University of Bristol, Bristol, UK 7Oakfield House, University of Bristol, Bristol, UK 8Faculty of Health and Applied Sciences, University of the West of England, Bristol, UK Correspondence Yvonne Wren, Bristol Dental School, University of Bristol, Oakfield Grove, Bristol BS8 2BN, UK. Email: yvonne.wren@bristol.ac.uk Funding information Wellcome Trust, Grant/Award Number: 217065/Z/19/Z; Medical Research Council, Grant/Award Number: G0501804 ID 76829 Abstract Purpose: Social, emotional and behavioural difficulties (SEBD) in childhood are associated with negative consequences across the life course. Children with developmental language disorder have been identified as being at risk of developing SEBD but it is unclear whether a similar risk exists for children with speech sound disorder, a condition which impacts on children's ability to make themselves un- derstood and has been shown to be associated with poor educational outcomes. Methods: Participants were children who attended the 8‐year‐old clinic in the Avon Longitudinal Study of Parents and Children (N = 7390). Children with speech sound disorder that had persisted beyond the period of typical speech acquisition (persistent speech disorder [PSD]) at age 8 were identified from recordings and transcriptions of speech samples (N = 263). Parent‐, teacher‐ and child‐reported questionnaires and interviews including the Strengths and Difficulties Question- naire, Short Moods and Feelings Questionnaire and measures for antisocial and risk‐ taking behaviour were used to provide outcome scores for SEBD at 10–14 years in a series of regression analyses. Results: Following adjustment for biological sex, socio‐economic status and Intel- ligence Quotient, children with PSD at age 8 were more likely to show peer prob- lems at age 10–11 years compared with their peers, as reported by teachers and parents. Teachers were more likely to report problems with emotionality. Children with PSD were no more likely to report symptoms of depression than their peers. No associations were observed between PSD, risk of antisocial behaviour, trying alcohol at age 10 or smoking cigarettes at age 14. Conclusions: Children with PSD may be at risk in terms of their peer relationships. This could impact on their wellbeing and, while not observed at this age, may lead to depressive symptoms in older childhood and adolescence. There is also the potential that these symptoms may impact on educational outcomes. K E Y W O R D S ALSPAC, antisocial behaviours, depression, emotional and behavioural difficulties, risk‐taking, social, speech disorder This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, pro- vided the original work is properly cited. © 2023 The Authors. JCPP Advances published by John Wiley & Sons Ltd on behalf of Association for Child and Adolescent Mental Health. JCPP Advances. 2023;3:e12126. https://doi.org/10.1002/jcv2.12126 wileyonlinelibrary.com/journal/jcv2 - 1 of 11 2 of 11 - INTRODUCTION There is a growing body of research on the individual and societal impact of childhood social, emotional and behavioural difficulties (SEBD), with evidence of immediate and long‐term consequences. This includes the development and continuity of internalising and externalising disorders (Agnew‐Blais et al., 2016; Kovacs & Dev- lin, 1998) and an increased risk of adverse outcomes across a range of domains of functioning. Emotional problems in childhood have been associated with increased risk of educational underachieve- ment, unemployment, substance abuse, teenage pregnancy, poor physical health, and future suicidal behaviour (Bridge et al., 2006; Clayborne et al., 2019; Essau et al., 2014; Keenan‐Miller et al., 2007; Woodward et al., 2001). Similarly, behavioural problems in childhood are also predictive of a wide range of outcomes in adulthood, including future crime, substance use, mental health, risky sexual behaviour and violent partner relationships (Fergusson et al., 2005). The potential impact of SEBD on children's futures raises important questions about the aetiology of these difficulties, and how to iden- tify children at risk. ‘Speech, language and communication needs’ (SLCN) is an um- brella term which encompasses difficulties with understanding and using language (such as vocabulary and grammar) as well as diffi- culties with the production and fluency of speech and the use of language in social contexts. SLCN are associated with a wide range of negative outcomes, including education, social participation and wellbeing and mental health (Botting et al., 2016; Durkin et al., 2017; Lee et al., 2020; McCormack et al., 2009; Snowling et al., 2006; Yew & O'Kearney, 2013). Specifically, an increased rate of conduct problems has been found in children with developmental language disorder (DLD) and adolescents with a history of DLD in comparison with typically developing peers (Conti‐Ramsden et al., 2013; St Clair et al., 2011) while other studies have highlighted the long‐term im- pacts of DLD on social, emotional and behavioural outcomes, friendships and bullying (Botting et al., 2016; Charman et al., 2015; Durkin et al., 2017; Lindsay & Dockrell, 2000; Norbury et al., 2016; van den Bedem, Dockrell, van Alphen, Kalicharan et al., 2018; Yew & O'Kearney, 2013). Associations have also been found between SLCN WREN ET AL. Key points � Social, emotional and behavioural difficulties (SEBD) in childhood are associated with negative consequences in older childhood and adulthood. � Children with developmental language disorder (DLD) have been identified as being at risk of developing SEBD. � It is unclear whether a similar risk exists for children with Persistent Speech Disorder (PSD)—a condition which has been shown to impact on educational outcomes. � Children with PSD are more likely to have problems with peer relationships. � Children with PSD are more likely to show emotionality at school. � While depressive symptoms are not more common in this population at age 8 years compared with peers, this may change in adolescence and poor peer relationships may contribute to this. � Children with PSD are no more likely than their peers to show antisocial and risk‐taking behaviours. � Future research should consider whether children with PSD are more or less likely to show SEBD as they get older and also which treatments work best to alleviate these if they are likely to present. � Children with PSD or a history of PSD should be iden- tified at school to ensure appropriate monitoring and support are in place. � Speech and Language Therapists, education and health staff should be made aware that children with PSD are at risk of experiencing difficulties with peer relationships and emotionality in school and that the profile of these needs may change over time. � Intervention for PSD in early childhood for speech may lead to a reduction in the negative sequelae seen in older childhood and adolescence. and mental health problems, such as anxiety and depression Kalicharan et al., 2018). A recent longitudinal study has also exam- (Beitchman et al., 2001, 2014; Botting et al., 2016; Conti‐Ramsden & ined the developmental pathways that may mediate a relationship Botting, 2008; Wadman et al., 2011). between DLD and externalising problems, suggesting a role for Theoretical frameworks for understanding the association be- emotional competence (van den Bedem, Dockrell, van Alphen, de tween SLCN and SEBD are emerging. The quality of relationships Rooij et al., 2018). may act as a risk or protective factor. For example, a qualitative study Whilst some types of SLCN such as DLD have been researched carried out with 11 children with speech and language disorders over widely, much less is known about social, emotional and behavioural the course of 6 months found that difficulties with relationships was outcomes for children with speech sound disorder, a condition that a risk factor for wellbeing, whereas positive relationships appeared to impacts on children's production of speech and ability to make be a protective factor (Lyons & Roulstone, 2018). Moreover, Forrest themselves understood. Furthermore, of the research that does et al. (2018) found that emotional difficulties experienced by ado- exist, findings have been mixed and samples often include a wide lescents with DLD could be partially accounted for by peer problems range of speech difficulties. For example, McCormack et al. (2011) at age 7. Other studies carried out with adolescents with DLD that found increased rates of SEBD at ages 7–9‐years‐old while Beitch- have found an association with an increased risk of depression have man et al. (2001) showed no difference compared to controls in suggested that a relationship between language difficulties and adolescence. However, both studies used broad eligibility criteria internalising symptoms may be moderated by bullying victimisation and included children with a range of speech, voice and fluency (Kilpatrick et al., 2019) or mediated by maladaptive emotional difficulties rather than focussing exclusively on children with speech regulation strategies (van den Bedem, Dockrell, van Alphen, sound disorder. Lewis et al. (2016) looked specifically at children SOCIAL, EMOTIONAL AND BEHAVIOURAL DIFFICULTIES ASSOCIATED WITH PERSISTENT SPEECH DISORDER - 3 of 11 with speech sound disorder and found no association with symp- 7, the entire cohort was invited to attend clinics for direct assess- toms of depression, anxiety, internalising or externalising problems ment of growth and development on an annual basis. Speech was in adolescence and adulthood. However, this study focussed on pre‐ assessed at the 8‐year clinic. schoolers, an age at which speech sound difficulties in many children The study website contains details of all the data that are resolve within the period of typical speech acquisition. It is there- available through a fully searchable data dictionary and variable fore possible that associations with SEBD may only be present for search tool at http://www.bristol.ac.uk/alspac/researchers/our‐data/. the subset of children who have speech sound disorder continuing beyond the period of typical speech acquisition (known as persistent speech disorder [PSD]). PSD has an estimated prevalence of 3.6% at Participants 8 years of age (Wren et al., 2016). Prospective longitudinal studies have identified a number of predictors of PSD, including male sex, A total of 7390 children attended the clinic at 8 years of age, where family history, atypical errors, lower socioeconomic status, hearing a speech assessment was conducted, including a recording of loss and suspected coordination difficulties (Eadie et al., 2015; continuous speech by trained assessors. Observation during the Morgan et al., 2017; Wren et al., 2016) as well as associations be- assessment identified 991 (13.4%) children as having some unusual tween PSD and poorer literacy and educational outcomes, after features in their speech which could qualify them as atypical. A total accounting for biological sex, socioeconomic status and Intelligence of 580 (7.8%) of these demonstrated only errors that can be clas- Quotient (IQ) (Wren et al., 2021). It is not known whether there is sified as common clinical distortions (Shriberg, 1993). The re- an increased risk of SEBD for this group, or what the role of po- cordings of the remainder of those identified with atypical speech tential confounders such as biological sex, socioeconomic status and (N = 411, 5.5%) were transcribed and analysed by speech and lan- IQ might be. Identifying the relationships between PSD and SEBD guage therapists to determine percentage consonants correct scores can ultimately inform education and health services to ensure that (Shriberg et al., 1997). This process was repeated for a random children at risk are identified and offered appropriate support. The sample of 50 children from the rest of the cohort to provide aim of the current study was therefore to address the current gap in normative data. The PSD group comprised those children whose the evidence to investigate the social, emotional and behavioural percentage consonants correct scores, were more than 1.2 standard outcomes of children with PSD in older childhood using data from a deviations below the mean (N = 263, 3.6%). Those whose speech large prospective, population‐based sample—the Avon Longitudinal was identified as atypical but whose percentage consonants correct Study of Parents and Children (ALSPAC)—including adjustment for score did not reach this threshold and those whose errors were biological sex, socio‐economic status, non‐verbal and verbal IQ. limited to common clinical distortions were excluded from further Three research questions drove the analyses: analysis. This was because previous research on the dataset has revealed distinct differences in the profiles of children within these � Are children with PSD at 8 years more likely to present with SEBD groups with regards to demographic, cognitive and speech motor in older childhood than those without? skills, suggesting that they should not be considered equivalent to � To what extent do children with PSD at 8 years present with the rest of the cohort (Wren et al., 2012). Moreover, we were keen symptoms related to depression in older childhood? to explore the associations for those children with the most � Are children with PSD at 8 years more likely to show antisocial impacted speech production. and risk‐taking behaviours in older childhood than those without? The total sample size for this study was therefore 6662 children, consisting of a case group of children with confirmed PSD (N = 263) and a control group comprising the rest of the cohort (N = 6399). METHODS Cohort study numbers Measures The Avon Longitudinal Study of Parents and Children (ALSPAC, Outcomes for behaviour and depression were measured using the www.bristol.ac.uk/alspac) is a transgenerational, observational. Strengths and Difficulties Questionnaire (SDQ, Goodman, 1997) at Population‐based study of health and development across the life- age 10–11 and an adapted version of the Short Moods and Feel- span. Pregnant women resident in Avon, UK with expected dates of ings Questionnaire (sMFQ, Angold et al., 1995) at age 10. Out- delivery between 1 April 1991 and 31 December 1992 were invited comes for antisocial and risk‐taking behaviours were measured to take part. The initial number of pregnancies enroled was 14,541 using a range of interview questions during clinics at ages 10, 11 and of these, there were a total of 14,676 foetuses, resulting in and 14. 14,062 live births and 13,988 children who were alive at 1 year of The SDQ is a brief behavioural screening tool which can be age. completed by parents and teachers in 5 min for children aged 4–17. It Detailed information on parents and children in the ALSPAC comprises 25 questions and includes positive and negative attributes sample is available from Fraser et al. (2013) and Boyd et al. (2013). across five subscales: emotional symptoms, conduct problems, hy- Data have been collected via self‐report questionnaires, face‐to‐face peractivity/inattention, peer relationship problems and prosocial clinical assessments, birth, medical, and educational records, and behaviour. The SDQ has been validated against existing tools—the from biological samples, all collected prospectively at multiple time- Child Behaviour Checklist (Achenbach, 1991) and the Rutter ques- points during pregnancy and throughout childhood. From the age of tionnaires (Elander & Rutter, 1996)—and shown to be able to 4 of 11 - WREN ET AL. distinguish between low‐ and high‐risk groups (Goodman, 1997; Two of the antisocial and risk‐taking behaviours (tried alcohol at Goodman & Scott, 1999). It has also been shown to be better than age 10; tried smoking at age 14) were binary yes/no responses where the Rutter questionnaires at identifying strengths and pro‐social ‘no’ was the reference group, while the antisocial score at age 10 was behaviours, and problems with inattention and peer relationships dichotomised such that a score of 0 provided the reference group. (Goodman, 1997). Follow‐up interviews with parents of children from the high‐risk groups suggested that the SDQ was better than the Child Behaviour Checklist at detecting problems with inattention and Potential confounders hyperactivity and as good at it as identifying internalising and externalising problems (Goodman & Scott, 1999). Biological sex, as recorded in birth records, and socio‐economic The sMFQ is a self‐report screening measure of childhood status, based on highest level of maternal education (specifically, depression, designed for use in epidemiological studies. It is a uni- whether mothers' education at finished at the end of compulsory factorial scale comprising 13 items. It has been validated against the schooling at age 16 or whether they had gone on to complete Children's Depression Inventory (Kovacs, 1983) and the Diagnostic optional schooling to 18 or had completed degree level education), Interview Schedule for Children depression scale (Costello were included as confounders in all analyses. Analyses were also et al., 1982), showing an ability to distinguish between children with adjusted for performance on the Wechsler Intelligence Scale for depression and controls (Angold et al., 1995). Children (WISC III—Wechsler et al., 1992). Specifically, scores for Behaviour and depression each of the verbal and performance subtests as well as total IQ scores were used to ensure that neither general intelligence nor language skill could account for the results. The following sub‐tests were administered by trained psychologists: Information, Similar- Parents completed the SDQ when their children were aged 11 and ities, Arithmetic, Vocabulary, and Comprehension (verbal subtests), teachers were asked to complete it when the children were in their Picture Completion, Coding, Picture Arrangement, Block Design final year of primary education, aged 10–11. The SDQ scores were and Object Assembly (performance subtests). Children were not normally distributed, so scores were dichotomised such that assessed on the WISC on the same day as the speech samples were those in the bottom 10% of scores (highest 10% for prosocial) are collected. considered problematic and the reference group is the top 90% (bottom 90% for prosocial). At age 10, the sMFQ statements were read aloud to the child Statistical analysis who was asked to indicate how they had been feeling or acting in the previous two weeks, and whether the statement was ‘true’, ‘some- For dichotomous outcomes, logistic regression was performed and times true’ or ‘not at all true’. Scores of 2, 1 and 0 were allocated to odds ratios (ORs) and 95% confidence intervals (CIs) are presented. each response respectively and summed together (range 0–26). As For each of the SDQ subscales we considered the ‘non‐problematic’ with the SDQ, scores were not normally distributed and so were 90% of the population as the reference group. For continuous out- dichotomised into children whose scores suggested that they were in comes, linear regression was performed with regression coefficients a depressed category (score of 12 or above, Angold et al., 1995) and and 95% CIs presented. those not. Antisocial and risk‐taking behaviours A number of adjusted models are presented (model 0 being un- adjusted). Model 1 adjusted for biological sex and maternal education. In model 2 either performance IQ (model 2a) or verbal IQ (model 2b) was added. Finally model 3 added total IQ. At each point, all available data were used. All analyses were conducted Outcomes relating to various antisocial behaviours were collected at using Stata (Version 13 Stata Corp, Texas, USA). age 10 via structured interviews where children were told they would be asked some questions about whether their friends or they had done something that could get them into trouble. They were RESULTS assured of confidentiality and also told that everybody would be asked the same questions. They were first asked if their friends had The available sample size for each of the outcomes and confounders taken part in a particular activity and then asked if they had taken are summarised in Table 1. Compared with the baseline cohort part. A composite score for antisocial behaviour was derived from (children alive at 1 year of age), those who attended the 8‐year clinic the responses to 11 questions asking about deliberately missing were more likely to be female and to have mothers with higher levels school, destroying something for fun, setting fire to something, of education. stealing, getting into fights, cruelty to animals, being in trouble with Descriptive statistics for the confounding variables are sum- the police, smoking cigarettes, drinking alcohol, being offered drugs marised in Table 2. There were significantly more boys in the case and smoking cannabis. An overall antisocial score was derived group as compared with controls (p < 0.001). There was a higher (range). proportion of mothers with lower levels of education in the case At the age of 14, children completed a questionnaire where they group (p = 0.019). Cases had lower mean performance IQ score were asked if they had ever smoked cigarettes. (p = 0.009), verbal IQ scores (p = 0.038) and total IQ scores 199 197 197 198 199 144 144 144 144 144 Cases 209 SOCIAL, EMOTIONAL AND BEHAVIOURAL DIFFICULTIES ASSOCIATED WITH PERSISTENT SPEECH DISORDER - 5 of 11 T A B L E 1 Summary of variables and sample size Age when collected Number available Controls Cases Variable name Variable type Variable source Well‐being—Parent report Prosocial Hyperactivity Emotionality Conduct Outcome Outcome Outcome Outcome Strengths and Difficulties Questionnaire Strengths and Difficulties Questionnaire Strengths and Difficulties Questionnaire Strengths and Difficulties Questionnaire Peer problems Outcome Strengths and Difficulties Questionnaire Well‐being—Teacher report Prosocial Hyperactivity Emotionality Conduct Outcome Outcome Outcome Outcome Strengths and Difficulties Questionnaire Strengths and Difficulties Questionnaire Strengths and Difficulties Questionnaire Strengths and Difficulties Questionnaire Peer problems Outcome Strengths and Difficulties Questionnaire 11 11 11 11 11 10–11 10–11 10–11 10–11 10–11 5008 5000 5001 5008 5008 3677 3617 3617 3616 3617 Depression Variable type Variable source Age when collected Controls Depression Outcome Short Moods and Feelings Questionnaire 10 5483 Number available Anti‐social/risk taking behaviours Variable type Variable source Age when collected Controls Tried alcohol Smoking Anti‐social score Outcome Outcome Outcome Item from anti‐social questionnaire Questionnaire item Anti‐social questionnaire 10 14 10 5492 3986 5660 Cases 209 162 222 Number available Confounders Biological sex Variable type Variable source Confounder Midwifery records Age when collected Antenatal Highest level of maternal education Confounder Mother's questionnaire at 32 weeks gestation Antenatal reported by the mother Performance IQ (subscale of WISC) Confounder Wechsler Intelligence Scale for Children 8 years clinic 6309 Verbal IQ (subscale of WISC) Confounder Wechsler Intelligence Scale for Children 8 years clinic 6319 Combined IQ (total of WISC) Confounder Wechsler Intelligence Scale for Children 8 years clinic 6292 Abbreviations: IQ, intelligence quotient; WISC, Wechsler Intelligence Scale for Children. Number available Controls Cases 6399 5924 263 242 258 257 256 (p = 0.003). Thus, all of these variables were controlled for in the problematic group for more subscales: prosocial (unadjusted OR: analyses. 1.78 [95% CI: 1.11–2.87]); hyperactivity (unadjusted OR: 1.78 [95% Table 3 presents the results for behaviour and depression. The CI: 1.14–2.80]); emotionality (unadjusted OR: 1.93 [95% CI: 1.25– OR gives the increase or decrease in odds of being in the ‘problem- 2.97]); and peer problems (unadjusted OR: 2.78 [95% CI: 1.85– atic’ category (top 10% for each subscale of the SDQ, except pro- 4.16]). The associations for teacher report of prosocial behaviour social [bottom 10%] or in the ‘depressed’ category for the mental and hyperactivity were no longer evident in any of the adjusted health measure) for cases compared with controls. models. However, strong associations remained, particularly for The unadjusted models for the SDQ showed that case children peer problems, with the fully adjusted OR attenuated to 2.30 [95% were more likely to be rated by parents as being in the problematic CI: 1.49–3.57]. The association with emotionality was also attenu- group for prosocial (unadjusted OR: 1.58 [95% CI: 1.32–2.23]) and ated but remained after full adjustment (OR = 1.80 [956% CI: 1.14, peer problems (unadjusted OR: 2.42 [95% CI: 1.65–3.55]). After 2.85]). taking confounders into account, the association with prosocial dis- The unadjusted model (model 0) for depression showed that case appeared; however, the association with peer problems remained children were more likely to be classified as depressed compared albeit slightly attenuated (e.g., model 3: adjusted OR 1.81 [95% CI: with controls (OR: 1.73 [95% CI: 1.18–2.54]), but the association was 1.18–2.78]). no longer evident in any of the adjusted models. The unadjusted models for the SDQ showed that case children Table 4 presents the results of the associations with antisocial were more likely to be rated by teachers as being in the and risk‐taking behaviours. No associations were evident. 6 of 11 - T A B L E 2 confounders Summary of descriptive statistics for the emotionality and hyperactivity (Conti‐Ramsden et al., 2013; Lindsay WREN ET AL. Confounder Biological sex Controls Cases Male (49.0%) 3096 (48.4%) 169 (64.3%) & Dockrell, 2000). Findings from the Millennium Cohort Study have suggested that peer problems may be a mediator between language disorder and emotional difficulties such that positive peer relationships act as a protective factor (Forrest et al., 2018), while meta‐analyses of both Female (51.0%) p < 0.001 3303 (51.6%) 94 (35.7%) cross‐sectional and longitudinal studies have demonstrated signifi- Maternal education <O level (21.7%) O levela (35.2%) 1265 (21.4%) 70 (28.9%) 2093 (35.3%) 79 (32.6%) A levelsb or higher (38.4%) 2566 (43.3%) 93 (38.4%) p = 0.019 Performance IQ cant relationships between peer victimisation and internalising problems (Hawker & Boulton, 2000; Reijntjes et al., 2010). In this study however, there was stronger evidence of an association with peer problems, some evidence for an association with emotionality in school and no evidence of an association with behavioural problems. This profile was unexpected and is in contrast with broader findings of increased rates of behavioural problems in other profiles of SLCN, Mean (sd) p = 0.009 107.2 (16.6%) 101.0 (18.9%) such as children with DLD (Conti‐Ramsden et al., 2013; St Clair Verbal IQ Mean (sd) p = 0.038 99.8 (17.0%) 93.9 (19.0%) Total IQ et al., 2011). Our findings may suggest a different and more specific link between communication difficulties as a result of PSD and peer problems, although it is also possible that emotional and behavioural difficulties may become more apparent at a later stage in adolescence, Mean (sd) p = 0.003 104.3 (16.3%) 97.6 (18.9%) not assessed in the current study. For example, Forrest et al. (2018) Abbreviation: IQ, intelligence quotient. aCompulsory examssinations taken at age 16. bOptional examinations taken at age 18. DISCUSSION found that emotional difficulties that were identified in adolescence could partly be accounted for by peer problems earlier in childhood in DLD. This large‐scale population study, with a carefully defined group of children with PSD that included adjustment for language provides new insights into the specific relationships that may exist between PSD and later psychosocial outcomes. Previous studies that have compared children with speech sound disorder alone with children This study investigated whether PSD was associated with SEBD using with speech sound disorder with comorbid language impairment or data from a large, prospective, population‐based sample. Following language impairment alone have found that associations with later adjustment for biological sex, socio‐economic status and verbal and SEBD and psychiatric disorders were related to difficulties with lan- performance IQ, children with PSD at age 8 were more likely to show guage rather than speech (Beitchman et al., 2001; Lewis et al., 2016). peer problems at age 10–11 as reported by teachers and parents. Teachers were also more likely to report problems with emotionality. Children with PSD were found to be more likely to report symptoms of depression than their peers though this was explained by the To what extent do children with PSD present with symptoms related to depression? confounders of biological sex, maternal education, performance and verbal IQ rather than their speech status. They were no more likely When all confounders were taken into account, children with PSD to become involved in antisocial and risk‐taking behaviour in later were no more likely to report symptoms of depression than their childhood and early adolescence than their peers. peers at age 10. While this is encouraging, depression typically pre- Are children with PSD more likely to present with SEBD? sents in mid‐adolescence (Orchard et al., 2017), so it is possible that there was not enough variability in symptoms at age 10, and that this relationship may emerge later in childhood. Furthermore, there is some indication in the literature that parents and children report different symptom severity (Eg et al., 2018; Orchard et al., 2019), Children with PSD at age 8 were more likely to present with certain with a lack of clarity regarding which report is more ‘accurate’. types of SEBD but not all. They were more likely to present with peer Future work would benefit from examining parental report as well as problems at age 10–11 as reported by teachers and parents even child report (De Los Reyes et al., 2013) and considering children's after controlling for biological sex, maternal education, verbal and reports in older adolescence. performance IQ. Teacher's reports of emotionality and peer problems was more likely to be associated with case status after adjustment, suggesting children might show different behaviours in school or that certain behaviours may be perceived differently at home and at Are children with PSD more likely to show antisocial and risk‐taking behaviours? school. These findings are consistent with previous studies that have No associations were observed between PSD and risk of antisocial identified an association between the broader group of children with behaviour, trying alcohol at age 10, or smoking cigarettes at age 14. SLCN and difficulties with peer relationships and bullying (Durkin This is similar to Beitchman et al. (2001) who found no evidence of et al., 2017; Forrest et al., 2018; McCormack et al., 2009, 2011) and increased rates of substance abuse and antisocial personality SOCIAL, EMOTIONAL AND BEHAVIOURAL DIFFICULTIES ASSOCIATED WITH PERSISTENT SPEECH DISORDER - 7 of 11 l a b r e v , n o i t a c u d e l a n r e t a m Q I e c n a m r o f r e p d n a n o i t a c u d e l a n r e t a m Q I l a b r e v d n a r o f d e t s u d A j : 3 l e d o M d e t s u d A j : b 2 l e d o M , x e s l a c i g o o b l i , x e s l a c i g o o b i l r o f n o i t a c u d e l a n r e t a m e c n a m r o f r e p d n a t n e i t o u q e c n e g i l l e t n i , x e s l a c i g o o b l i r o f r o f d e t s u d A j : 1 l e d o M d n a x e s l a c i g o o b i l ) Q I ( n o i t a c u d e l a n r e t a m d e t s u d a n U j : 0 l e d o M d e t s u d A j : a 2 l e d o M e r i a n n o i t s e u Q s e i t l u c fi f i D d n a s h t g n e r t S f o e l a c s ‐ b u s h c a e r o f y r o g e t a c l c i t a m e b o r p e h t n i i g n e b r o f ) ] s I C [ s l a v r e t n i e c n e d fi n o c % 5 9 ( ) s R O ( s o i t a r s d d o : s e m o c t u O h t l a e H l a t n e M d n a i g n e b l l e W 3 E L B A T s l o r t n o c s u s r e v s e s a c r o f ) Q F M s ( e r i a n n o i t s e u Q s g n i l e e F d n a s d o o M t r o h S e h t y b d e n fi e d s a d e s s e r p e d i g n e b r o f d n a ) Q D S ( e u l a v ‐ p ] I C % 5 9 [ R O e u l a v ‐ p ] I C % 5 9 [ R O e u l a v ‐ p ] I C % 5 9 [ R O e u l a v ‐ p ] I C % 5 9 [ R O e u l a v ‐ p ] I C % 5 9 [ R O e m o c t u O ) 1 1 e g a ( t r o p e r t n e r a P : Q D S 5 0 1 0 . ] 6 9 1 . , 4 9 0 [ . 6 3 1 . 6 0 1 0 . ] 6 9 1 . , 4 9 0 [ . 6 3 1 . 5 2 1 0 . ] 3 9 1 . , 2 9 0 [ . 3 3 1 . 6 6 0 0 . ] 1 0 2 . , 8 9 0 [ . 0 4 1 . 9 0 0 0 . ] 3 2 2 . , 2 3 1 [ . 8 5 1 . ] 7 5 2 5 / 5 0 8 = n [ l a i c o s o r P 8 8 6 0 . ] 7 5 1 . , 0 5 0 [ . 9 8 0 . 2 9 6 0 . ] 7 5 1 . , 1 5 0 [ . 9 8 0 . 8 5 9 0 . ] 1 7 1 . , 7 5 0 [ . 9 9 0 . 3 8 6 0 . ] 0 5 1 . , 6 6 0 [ . 2 1 1 . 9 3 3 0 . ] 3 1 2 . , 7 7 0 [ . 8 2 1 . ] 7 9 1 5 / 0 6 3 = n [ y t i v i t c a r e p y H 8 6 6 0 . ] 4 0 2 . , 3 6 0 [ . 4 1 1 . 3 7 8 0 . ] 5 0 2 . , 4 6 0 [ . 5 1 1 . 3 0 7 0 . ] 1 0 2 . , 3 6 0 [ . 2 1 1 . 3 4 0 . ] 0 2 2 . , 2 7 0 [ . 5 2 1 . 8 6 3 0 . ] 5 1 2 . , 5 7 0 [ . 7 2 1 . ] 5 9 1 5 / 1 4 3 = n [ y t i l a n o i t o m E 7 7 3 0 . ] 9 3 1 . , 2 4 0 [ . 6 7 0 . 9 9 3 0 . ] 1 4 1 . , 2 4 0 [ . 7 7 0 . 3 1 4 0 . ] 2 4 1 . , 3 4 0 [ . 8 7 0 . 8 9 4 0 . ] 8 4 1 . , 5 4 0 [ . 1 8 0 . 7 6 9 0 . ] 2 7 1 . , 7 5 0 [ . 9 9 0 . ] 6 0 2 5 / 2 7 3 = n [ t c u d n o C 7 0 0 0 . ] 8 7 2 . , 8 1 1 [ . 1 8 1 . 7 0 0 0 . ] 8 7 2 . , 8 1 1 [ . 1 8 1 . 3 0 0 0 . ] 0 9 2 . , 5 2 1 [ . 0 9 1 . 1 0 0 0 0 < . ] 1 1 3 . , 7 3 1 [ . 7 0 2 . 1 0 0 0 < . ] 5 5 3 . , 5 6 1 [ . 2 4 2 . ] 7 0 2 5 / 7 2 4 = n [ l s m e b o r p r e e P ) 1 1 – 0 1 e g a ( t r o p e r r e h c a e T : Q D S 3 4 2 0 . ] 7 2 2 . , 1 8 0 [ . 6 3 1 . 3 4 2 0 . ] 7 2 2 . , . 8 0 [ 6 3 1 . 7 5 1 0 . ] 0 4 2 . , 7 8 0 [ . 4 4 1 . 4 9 0 0 . ] 0 5 2 . , 3 9 0 [ . 3 5 1 . 7 1 0 0 . ] 7 8 2 . , 1 1 1 [ . 8 7 1 . ] 1 6 7 3 / 7 3 3 = n [ l a i c o s o r P 4 5 7 0 . ] 0 8 1 . , 6 5 0 [ . 8 0 1 . 2 5 6 0 . ] 5 8 1 . , 8 6 0 [ . 2 1 1 . 5 3 4 0 . ] 3 0 2 . , 4 7 0 [ . 2 2 1 . 5 2 1 0 . ] 5 3 2 . , 0 9 0 [ . 5 4 1 . 2 1 0 0 . ] 0 8 2 . , 4 1 1 [ . 8 7 1 . ] 1 6 7 3 / 9 8 3 = n [ y t i v i t c a r e p y H 2 1 0 0 . ] 5 8 2 . , 4 1 1 [ . 0 8 1 . 1 0 0 . ] 9 8 2 . , 6 1 1 [ . 3 8 1 . 4 0 0 0 . ] 6 0 3 , 4 2 1 [ . 4 9 1 . 2 0 0 0 . ] 0 2 3 . , 1 3 1 [ . 5 0 2 . 3 0 0 0 . ] 7 9 2 . , 5 2 1 [ . 3 9 1 . ] 1 6 7 3 / 4 1 4 = n [ y t i l a n o i t o m E 4 5 9 0 . ] 3 8 1 . , 0 6 0 [ . 5 0 1 . 5 0 8 0 . ] 7 8 1 . , 2 6 0 [ . 7 0 1 . 6 9 6 0 . ] 5 9 1 . , 4 6 0 [ . 2 1 1 . 1 6 5 0 . ] 3 0 2 . , 8 6 0 [ . 8 1 1 . 7 0 1 0 . ] 2 5 2 . , 1 9 0 [ . 2 5 1 . ] 0 6 7 3 / 9 2 3 = n [ t c u d n o C 1 0 0 0 0 < . ] 7 5 3 . , 9 4 1 [ . 0 3 2 . 1 0 0 0 0 < . ] 8 6 3 . , 4 5 1 [ . 8 3 2 . 1 0 0 0 0 < . ] 6 6 3 . , 4 5 1 [ . 7 3 2 . 1 0 0 0 0 < . ] 6 9 3 . , 1 7 1 [ . 0 6 2 . 1 0 0 0 0 < . ] 6 1 4 . , 5 8 1 [ . 8 7 2 . ] 1 6 7 3 / 3 8 3 = n [ l s m e b o r p r e e P 3 5 6 0 . ] 5 7 1 . , 0 7 0 [ . 0 1 1 . 3 8 5 0 . ] 0 8 1 . , 2 7 0 [ . 4 1 1 . 4 3 4 0 . ] 7 8 1 . , 7 7 0 [ . 0 2 1 . 7 2 1 0 . ] 4 1 2 . , 1 9 0 [ . 9 3 1 . 5 0 0 0 . ] 4 5 2 . , 8 1 1 [ . 3 7 1 . t r o p e r d l i h C : e r o c s n o i s s e r p e D ] 2 9 6 5 / 8 6 5 = n [ . 2 0 3 5 – 0 4 4 3 = 3 l e d o m ; 6 2 3 5 – 2 5 4 3 = b 2 l e d o m ; 7 1 3 5 – 2 5 4 3 = a 2 l e d o m ; 9 0 3 5 – 0 5 4 3 = 1 l e d o m ; 2 8 7 5 – 8 9 4 3 = 0 l e d o m : e m o c t u o o t i g n d r o c c a s e i r a v N : e t o N ) 0 1 e g a ( t r o p e r d l i h C : e r i a n n o i t s e u Q s g n i l e e F d n a s d o o M t r o h S 8 of 11 - l a n r e t a m , x e s , n o i t a c u d e d n a l a b r e v Q I e c n a m r o f r e p l a c i g o o b l i r o f n o i t a c u d e l a n r e t a m r o f d e t s u d A j : b 2 l e d o M , x e s l a c i g o o b i l r o f d e t s u d A j : a 2 l e d o M n o i t a c u d e l a n r e t a m e c n a m r o f r e p d n a , x e s l a c i g o o b l i d n a x e s l a c i g o o b i l r o f d e t s u d A j : 1 l e d o M Q I l a b r e v d n a ) Q I ( t n e i t o u q e c n e g i l l e t n i n o i t a c u d e l a n r e t a m d e t s u d a n U j : 0 l e d o M d e t s u d A j : 3 l e d o M s l o r t n o c s u s r e v s e s a c r o f y r o g e t a c e c n e r e f e r e h t n i i g n e b r o f ) ] s I C [ s l a v r e t n i e c n e d fi n o c % 5 9 ( ) s R O ( s o i t a r s d d o : s r u o i v a h e b i g n k a t k s i r d n a l a i c o s ‐ i t n A 4 E L B A T e u l a v ‐ p ] I C % 5 9 [ R O e u l a v ‐ p ] I C % 5 9 [ R O e u l a v ‐ p ] I C % 5 9 [ R O e u l a v ‐ p ] I C % 5 9 [ R O e u l a v ‐ p ] I C % 5 9 [ R O e m o c t u O 8 3 2 0 . ] 6 7 0 . , 0 1 0 [ . 3 4 0 . 4 5 2 0 . ] 1 8 0 . , 1 1 0 [ . 4 4 0 . 5 3 4 0 . ] 9 7 1 . , 1 1 0 [ . 4 4 0 . 9 7 2 0 . ] 8 8 1 . , 1 1 0 [ . 6 4 0 . 5 3 7 0 . ] 1 6 0 . , 6 2 0 [ . 2 8 0 . ) o n ( 1 1 e g a l o h o c l a d e i r T ] 1 0 7 5 / 2 0 6 5 = n [ 0 9 0 0 . ] 6 0 1 . , 6 4 0 [ . 0 7 0 . 3 0 1 0 . ] 7 0 1 . , 7 4 0 [ . 1 7 0 . 2 9 0 0 . ] 6 0 1 . , 6 4 0 [ . 0 7 5 . 5 0 1 0 . ] 8 0 1 . , 7 4 0 [ . 1 7 0 . 1 6 9 0 . ] 8 4 1 . , 9 6 0 [ . 1 0 1 . ) 0 ( 0 1 e g a e r o c s l a i c o s i t n A 3 9 0 0 . ] 6 0 1 . , 5 4 0 [ . 9 6 0 . 3 3 1 0 . ] 0 1 1 . , 8 4 0 [ . 3 7 1 . 3 9 0 0 . ] 6 0 1 . , 5 4 0 [ . 0 7 0 . 3 4 1 0 . ] 1 1 1 . , 8 4 0 [ . 3 7 0 . 7 0 2 0 . ] 5 1 1 . , 3 5 0 [ . 8 7 0 . ] 2 8 7 5 / 5 5 9 4 = n [ ] 8 4 1 4 / 2 0 1 3 = n [ ) o n ( 4 1 e g a i g n k o m S . 2 0 3 5 – 9 5 1 2 = 3 l e d o m ; 6 2 3 5 – 0 7 1 2 = b 2 l e d o m ; 7 1 3 5 – 4 6 1 2 = a 2 l e d o m ; 7 8 3 5 – 8 8 1 2 = 1 l e d o m ; 2 8 7 5 – 5 2 3 2 = 0 l e d o m : N : e t o N WREN ET AL. disorder in adolescents with a history of speech problems in early childhood. As with the measures for depression however, it is possible that associations might be seen at slightly older ages, particularly for trying alcohol. Limitations of this study Whilst there were some clear strengths of this study such as the large sample, prospective design and diverse range of SEBD measures, it is important to note some key limitations. Firstly, direct assessment was used to determine case status for this study but nevertheless, diag- nosis of PSD was not confirmed by clinical assessment. Moreover, we did not exclude children with PSD and other comorbidities which may have confounded our results. Secondly, although the sample size used in the analysis was substantial in comparison to other epidemiological studies (even after missing data), it lacked the diversity seen in the UK population today. The results would be strengthened by replicating with a more diverse sample. Thirdly, the findings presented could be explained by residual confounding. While we have adjusted for known factors, there may be other unknown factors which have impacted on the associations observed but which we have not yet identified. Fourthly, the measures used for behaviour and depression outcomes, the SDQ and the sMFQ are screening assessments rather than clinical diagnostic tools. Nevertheless they were designed specifically for the purposes of large‐scale epidemiological research where clin- ical assessment is not feasible. Finally, the analyses reported a num- ber of outcome measures from a range of different timepoints but none of the measures were themselves repeated, thus providing a series of cross‐sectional analyses using a longitudinal dataset. Future work could consider some of the questions arising from this work such as whether signs of depression or early use of alcohol are observed in children with PSD when they are older and also whether the problems observed in this cohort with peer relationships become more noticeable as children progress through school. Clinical implications This study highlights the importance of identifying children with PSD, including those with a history of PSD, throughout their primary and secondary school years in order to ensure appropriate monitoring and support are in place. Speech and Language Therapists, education and health staff need to be aware that children with PSD are at risk of experiencing difficulties with peer relationships and emotionality in school, in addition to poorer education outcomes (Wren et al., 2021), and that the profile of these needs may change over time. Our findings support intervention for PSD in early childhood due to the potential that improvement in speech may lead to a reduction in the negative sequelae seen in older childhood and adolescence. This study has also raised important implications for future research in investigating SEBD in older children with PSD to examine later emerging wellbeing and mental health difficulties, which have been reported in other populations of children with SLCN (Beitchman et al., 2001, 2014; Botting et al., 2016; Conti‐Ramsden & Botting, 2008; Wadman et al., 2011). This future research is also important as emotional problems in childhood have been associated SOCIAL, EMOTIONAL AND BEHAVIOURAL DIFFICULTIES ASSOCIATED WITH PERSISTENT SPEECH DISORDER - 9 of 11 with increased risk of a range of negative health and wellbeing out- DATA AVAIL ABILI TY STATEMENT comes in later life (Bridge et al., 2006; Clayborne et al., 2019; Essau The datasets generated during and/or analysed during the current et al., 2014; Keenan‐Miller et al., 2007; Woodward et al., 2001). study are not publicly available but access can be requested via a proposal to the ALSPAC study team. CONCLUSIONS ETHI CAL CONSI DERATIONS Ethical approval for the study was obtained from the ALSPAC Ethics Children with PSD at 8 years of age show an increase in SEBD be- and Law Committee and the Local Research Ethics Committees tween the ages of 10 and 14 in specific areas, particularly problems http://www.bristol.ac.uk/alspac/researchers/research‐ethics/. Infor- with peer relationships and emotionality in school, compared with med consent for the use of data collected via questionnaires and peers according to their parents and teachers. However, they appear clinics was obtained from participants following the recommenda- to be no more likely than their peers to report depressive symptoms tions of the ALSPAC Ethics and Law Committee at the time. at age 10 or to become involved in antisocial and risk‐taking behaviour at age 11–14. Whilst it is important to be aware that the findings may be affected by the respondent (for example, parent, child or teacher) or the age of measurement, the results do suggest that PSD can lead to some problems with SEBD. Professionals working with children with PSD should be aware of the potential for ORCID Yvonne Wren https://orcid.org/0000-0002-1575-453X Faith Orchard https://orcid.org/0000-0002-5324-5007 difficulties in peer relationships and be mindful of the possibility of an REFERENCES impact on wellbeing in older childhood as well as on educational outcomes. AUTH OR CON TRIBUTIONS Yvonne Wren: Investigation; Project administration; Writing – orig- inal draft. Emma Pagnamenta: Writing – original draft; Writing – review & editing. Faith Orchard: Writing – original draft; Writing – review & editing. Tim J. Peters: Supervision; Writing – review & editing. Alan Emond: Conceptualization; Funding acquisition. Kate Northstone: Formal analysis; Writing – review & editing. Laura Louise Miller: Formal analysis; Writing – review & editing. Susan Roulstone: Conceptualization; Funding acquisition; Methodology; Project administration; Supervision; Writing – review & editing. ACK NOWLEDGEMENTS The UK Medical Research Council and Wellcome Trust (Grant ref: 217065/Z/19/Z) and the University of Bristol provide core support for ALSPAC. This publication is the work of the authors who will serve as guarantors for the contents of this paper. A comprehen- sive list of grants funding is available on the ALSPAC website (http://www.bristol.ac.uk/alspac/external/documents/grant‐acknowl- edgements.pdf); This research was specifically funded by UK Medical Research Council Grant G0501804 ID 76829. We are extremely grateful to all the families who took part in this study, the midwives for their help in recruiting them, and the whole ALSPAC team, which includes interviewers, computer and labora- tory technicians, clerical workers, research scientists, volunteers, managers, receptionists and nurses. We are particularly grateful to the speech team that collected and transcribed the speech samples. We thank Lawrence D. Shriberg for his advice on case identifica- tion. We also acknowledge the contribution of Sue Loader, speech and language therapist, who had the foresight to initiate the collection of speech and language data from the children when they were aged 25 months. CONFLICT OF INTEREST The authors have declared that they have no competing or potential conflicts of interest. Achenbach, T. M. (1991). Manual for the child behavior checklist 14–18 and 1991 profile. University of Vermont Department of Psychiatry. Agnew‐Blais, J. C., Polanczyk, G. V., Danese, A., Wertz, J., Moffitt, T. E., & Arseneault, L. (2016). Evaluation of the persistence, remission, and emergence of attention‐deficit/hyperactivity disorder in young adulthood. JAMA Psychiatry, 73(7), 713–720. https://doi.org/10.10 01/jamapsychiatry.2016.0465 Angold, A., Costello, E. J., Messer, S. C., Pickles, A., Winder, F., & Silver, D. (1995). Development of a short questionnaire for use in epidemio- logical studies of depression in children and adolescents. Interna- tional Journal of Methods in Psychiatric Research, 5, 237–249. Beitchman, J. H., Brownlie, E. B., & Bao, L. (2014). Age 31 mental health outcomes of childhood language and speech disorders. Journal of the American Academy of Child & Adolescent Psychiatry, 53(10), 11 02–1110. https://doi.org/10.1016/j.jaac.2014.07.006 Beitchman, J. H., Wilson, B., Johnson, C. J., Atkinson, L., Young, A., Adlaf, E., Escobar, M., & Douglas, L. (2001). Fourteen‐year follow‐up of speech/language‐impaired and control children: Psychiatric ou- tcome. Journal of the American Academy of Child & Adolescent Psy- chiatry, 40(1), 75–82. https://doi.org/10.1097/00004583‐2001010 00‐00019 Botting, N., Toseeb, U., Pickles, A., Durkin, K., & Conti‐Ramsden, G. (2016). Depression and anxiety change from adolescence to adulthood in individuals with and without language impairment. PLoS One, 11(7), e0156678. https://doi.org/10.1371/journal.pone.0156678 Boyd, A., Golding, J., Macleod, J., Lawlor, D. A., Fraser, A., Henderson, J., Molloy, L., Ness, A., Ring, S., & Davey Smith, G. (2013). Cohort pro- file: The ‘children of the 90s’—The index offspring of the Avon Longitudinal Study of Parents and Children. International Journal of Epidemiology, 42(1), 111–127. https://doi.org/10.1093/ije/dys064 Bridge, J. A., Goldstein, T. R., & Brent, D. A. (2006). Adolescent suicide and suicidal behavior. Journal of Child Psychology and Psychiatry, 47(3‐4), 372–394. https://doi.org/10.1111/j.1469‐7610.2006.01615.x Charman, T., Ricketts, J., Dockrell, J. E., Lindsay, G., & Palikara, O. (2015). Emotional and behavioural problems in children with language im- pairments and children with autism spectrum disorders. International Journal of Language & Communication Disorders, 50(1), 84–93. https:// doi.org/10.1111/1460‐6984.12116 Clayborne, Z. M., Varin, M., & Colman, I. (2019). Systematic review and meta‐analysis: Adolescent depression and long‐term psycho- social outcomes. Journal of the American Academy of Child & Adolescent Psychiatry, 58(1), 72–79. https://doi.org/10.1016/j.jaac. 2018.07.896 Conti‐Ramsden, G., & Botting, N. (2008). Emotional health in adolescents with and without a history of specific language impairment (SLI). Journal of Child Psychology and Psychiatry, 49(5), 516–525. https:// doi.org/10.1111/j.1469‐7610.2007.01858.x 10 of 11 - Conti‐Ramsden, G., Mok, P. L., Pickles, A., & Durkin, K. (2013). Adolescents with a history of specific language impairment (SLI): Strengths and difficulties in social, emotional and behavioral functioning. Research in Developmental Disabilities, 34(11), 4161–4169. https://doi.org/10. 1016/j.ridd.2013.08.043 Costello, A., Edelbrock, C., Kalas, R., Kessler, M., & Klaric, S. A. (1982). Diagnostic interview schedule for children (DISC). National Institute of Mental Health. De Los Reyes, A., Thomas, S. A., Goodman, K. L., & Kundey, S. M. (2013). Principles underlying the use of multiple informants' reports. Annual Review of Clinical Psychology, 9(1), 123–149. https://doi.org/10.1146/ annurev‐clinpsy‐050212‐185617 Durkin, K., Toseeb, U., Botting, N., Pickles, A., & Conti‐Ramsden, G. (2017). Social confidence in early adulthood among young people with and without a history of language impairment. Journal of Speech, Lan- guage, and Hearing Research, 60(6), 1635–1647. https://doi.org/10. 1044/2017_jslhr‐l‐16‐0256 Eadie, P., Morgan, A., Ukoumunne, O. C., Ttofari Eecen, K., Wake, M., & Reilly, S. (2015). Speech sound disorder at 4 years: Prevalence, comorbidities, and predictors in a community cohort of children. Developmental Medicine and Child Neurology, 57(6), 578–584. https:// doi.org/10.1111/dmcn.12635 Eg, J., Bilenberg, N., Costello, E. J., & Wesselhoeft, R. (2018). Self‐and parent‐reported depressive symptoms rated by the mood and feel- ings questionnaire. Psychiatry Research, 268, 419–425. https://doi. org/10.1016/j.psychres.2018.07.016 Elander, J., & Rutter, M. (1996). Use and development of the Rutter parents' and teachers' scales. International Journal of Methods in Psychiatric Research, 6(2), 63–78. https://doi.org/10.1002/(sici)1234‐ 988x(199607)6:2<63::aid‐mpr151>3.3.co;2‐m Essau, C. A., Lewinsohn, P. M., Olaya, B., & Seeley, J. R. (2014). Anxiety disorders in adolescents and psychosocial outcomes at age 30. Journal of Affective Disorders, 163, 125–132. https://doi.org/10.1016/ j.jad.2013.12.033 Fergusson, D. M., Horwood, L. J., & Ridder, E. M. (2005). Show me the child at seven: The consequences of conduct problems in childhood for psychosocial functioning in adulthood. Journal of Child Psychology and Psychiatry, 46(8), 837–849. https://doi.org/10.1111/j.1469‐7610. 2004.00387.x Forrest, C. L., Gibson, J. L., Halligan, S. L., & St Clair, M. C. (2018). A lon- gitudinal analysis of early language difficulty and peer problems on later emotional difficulties in adolescence: Evidence from the Mil- lennium Cohort Study. Autism & Developmental Language Impairments, 3, 1–15. https://doi.org/10.1177/2396941518795392 Fraser, A., Macdonald‐Wallis, C., Tilling, K., Boyd, A., Golding, J., Davey Smith, G., Henderson, J., Macleod, J., Molloy, L., Ness, A., Ring, S., Nelson, S. M., & Lawlor, D. A. (2013). Cohort profile: The Avon Longitudinal Study of Parents and Children: ALSPAC mothers cohort. International Journal of Epidemiology, 42(1), 97–110. https:// doi.org/10.1093/ije/dys066 Goodman, R. (1997). The Strengths and Difficulties Questionnaire: A research note. Journal of Child Psychology and Psychiatry, 38(5), 581–586. https://doi.org/10.1111/j.1469‐7610.1997.tb01545.x Goodman, R., & Scott, S. (1999). Comparing the Strengths and Difficulties Questionnaire and the child behavior checklist: Is small beautiful? Journal of Abnormal Child Psychology, 27(1), 17–24. https://doi.org/ 10.1023/a:1022658222914 Hawker, D. S., & Boulton, M. J. (2000). Twenty years' research on peer victimization and psychosocial maladjustment: A meta‐analytic re- view of cross‐sectional studies. Journal of Child Psychology and Psy- chiatry, 41(4), 441–455. https://doi.org/10.1111/1469‐7610.00629 Keenan‐Miller, D., Hammen, C. L., & Brennan, P. A. (2007). Health out- comes related to early adolescent depression. Journal of Adolescent Health, 41(3), 256–262. https://doi.org/10.1016/j.jadohealth.2007. 03.015 Kilpatrick, T., Leitão, S., & Boyes, M. (2019). Mental health in adolescents with a history of developmental language disorder: The moderating effect of bullying victimisation. Autism & Developmental Language Impairments, 4, 2396941519893313. https://doi.org/10.1177/2396 941519893313 WREN ET AL. Kovacs, M. (1983). The Children's Depression Inventory: A self‐rated depres- sion scale for school‐aged youngsters. University of Pittsburgh School of Medicine, Department of Psychiatry, Western Psychiatric Insti- tute and Clinic. Kovacs, M., & Devlin, B. (1998). Internalizing disorders in childhood. The Journal of Child Psychology and Psychiatry and Allied Disciplines, 39(1), 47–63. https://doi.org/10.1017/s0021963097001765 Lee, Y. C., Chen, V. C., Yang, Y. H., Kuo, T. Y., Hung, T. H., Cheng, Y. F., & Huang, K. Y. (2020). Association between emotional disorders and speech and language impairments: A national population‐based study. Child Psychiatry and Human Development, 51(3), 355–365. https://doi.org/10.1007/s10578‐019‐00947‐9 Lewis, B. A., Patton, E., Freebairn, L., Tag, J., Iyengar, S. K., Stein, C. M., & Taylor, H. G. (2016). Psychosocial co‐morbidities in adolescents and adults with histories of communication disorders. Journal of Com- munication Disorders, 61, 60–70. https://doi.org/10.1016/j.jcomdis. 2016.03.004 Lindsay, G., & Dockrell, J. (2000). The behaviour and self‐esteem of chil- dren with specific speech and language difficulties. British Journal of Educational Psychology, 70(4), 583–601. https://doi.org/10.1348/ 000709900158317 Lyons, R., & Roulstone, S. (2018). Well‐being and resilience in children with speech and language disorders. Journal of Speech, Language, and Hearing Research, 61(2), 324–344. https://doi.org/10.1044/2017_ jslhr‐l‐16‐0391 McCormack, J., Harrison, L. J., McLeod, S., & McAllister, L. (2011). A na- tionally representative study of the association between communi- cation impairment at 4–5 years and children's life activities at 7–9 years. Journal of Speech, Language, and Hearing Research, 54(5), 1328–1348. https://doi.org/10.1044/1092‐4388(2011/10‐0155) McCormack, J., McLeod, S., McAllister, L., & Harrison, L. J. (2009). A systematic review of the association between childhood speech impairment and participation across the lifespan. International Jour- nal of Speech Language Pathology, 11(2), 155–170. https://doi.org/10. 1080/17549500802676859 Morgan, A., Eecen, K. T., Pezic, A., Brommeyer, K., Mei, C., Eadie, P., Reilly, S., & Dodd, B. (2017). Who to refer for speech therapy at 4 years of age versus who to “watch and wait”. The Journal of Pediatrics, 185, 200–204. https://doi.org/10.1016/j.jpeds.2017.02.059 Norbury, C. F., Gooch, D., Wray, C., Baird, G., Charman, T., Simonoff, E., Vamvakas, G., & Pickles, A. (2016). The impact of nonverbal ability on prevalence and clinical presentation of language disorder: Evi- dence from a population study. Journal of Child Psychology and Psy- chiatry, 57(11), 1247–1257. https://doi.org/10.1111/jcpp.12573 Orchard, F., Pass, L., Cocks, L., Chessell, C., & Reynolds, S. (2019). Exam- ining parent and child agreement in the diagnosis of adolescent depression. Child and Adolescent Mental Health, 24(4), 338–344. https://doi.org/10.1111/camh.12348 Orchard, F., Pass, L., Marshall, T., & Reynolds, S. (2017). Clinical charac- teristics of adolescents referred for treatment of depressive disor- ders. Child and Adolescent Mental Health, 22(2), 61–68. https://doi. org/10.1111/camh.12178 Reijntjes, A., Kamphuis, J. H., Prinzie, P., & Telch, M. J. (2010). Peer victimization and internalizing problems in children: A meta‐analysis of longitudinal studies. Child Abuse & Neglect, 34(4), 244–252. https://doi.org/10.1016/j.chiabu.2009.07.009 Shriberg, L. D. (1993). Four new speech and prosody‐voice measures for genetics research and other studies in developmental phonological disorders. Journal of Speech, Language, and Hearing Research, 36(1), 105–140. https://doi.org/10.1044/jshr.3601.105 Shriberg, L. D., Austin, D., Lewis, B. A., McSweeny, J. L., & Wilson, D. L. (1997). The percentage of consonants correct (PCC) metric: Ex- tensions and reliability data. Journal of Speech, Language, and Hearing Research, 40(4), 708–722. https://doi.org/10.1044/jslhr.40 04.708 Snowling, M. J., Bishop, D. V. M., Stothard, S. E., Chipchase, B., & Kaplan, C. (2006). Psychosocial outcomes at 15 years of children with a pre- school history of speech‐language impairment. Journal of Child Psy- chology and Psychiatry, 47(8), 759–765. https://doi.org/10.1111/j. 1469‐7610.2006.01631.x SOCIAL, EMOTIONAL AND BEHAVIOURAL DIFFICULTIES ASSOCIATED WITH PERSISTENT SPEECH DISORDER - 11 of 11 St Clair, M. C., Pickles, A., Durkin, K., & Conti‐Ramsden, G. (2011). A longitudinal study of behavioral, emotional and social difficulties in individuals with a history of specific language impairment (SLI). Journal of Communication Disorders, 44(2), 186–199. https://doi.org/ 10.1016/j.jcomdis.2010.09.004 van den Bedem, N. P., Dockrell, J. E., van Alphen, P. M., de Rooij, M., Samson, A. C., Harjunen, E. L., & Rieffe, C. (2018). Depressive symptoms and emotion regulation strategies in children with and without developmental language disorder: A longitudinal study. In- ternational Journal of Language & Communication Disorders, 53(6), 1110–1123. https://doi.org/10.1111/1460‐6984.12423 van den Bedem, N. P., Dockrell, J. E., van Alphen, P. M., Kalicharan, S. V., & Rieffe, C. (2018). Victimization, bullying, and emotional competence: Longitudinal associations in (pre) adolescents with and without developmental language disorder. Journal of Speech, Language, and Hearing Research, 61(8), 2028–2044. https://doi.org/10.1044/2018_ jslhr‐l‐17‐0429 Wadman, R., Botting, N., Durkin, K., & Conti‐Ramsden, G. (2011). Changes in emotional health symptoms in adolescents with specific language impairment. International Journal of Language & Communication Dis- orders, 46(6), 641–656. https://doi.org/10.1111/j.1460‐6984.2011. 00033.x Wechsler, D., Golombok, S., & Rust, J. (1992). WISC‐III UK Wechsler intel- ligence scale for children: UK manual. The Psychological Co-tnqh_9; rporation. Woodward, L., Fergusson, D. M., & Horwood, L. J. (2001). Risk factors and life processes associated with teenage pregnancy: Results of a pro- spective study from birth to 20 years. Journal of Marriage and Family, 63(4), 1170–1184. https://doi.org/10.1111/j.1741‐3737.2001.01 170.x Wren, Y., Miller, L. L., Peters, T. J., Emond, A., & Roulstone, S. (2016). Prevalence and predictors of persistent speech sound disorder at eight years old: Findings from a population cohort study. Journal of Speech, Language, and Hearing Research, 59(4), 647–673. https://doi. org/10.1044/2015_jslhr‐s‐14‐0282 Wren, Y., Pagnamenta, E., Peters, T. J., Emond, A., Northstone, K., Miller, L. L., & Roulstone, S. (2021). Educational outcomes associated with persistent speech disorder. International Journal of Language & Communication Disorders, 56(2), 299–312. https://doi.org/10.1111/ 1460‐6984.12599 Wren, Y. E., Roulstone, S. E., & Miller, L. L. (2012). Distinguishing groups of children with persistent speech disorder: Findings from a prospec- tive population study. Logopedics Phoniatrics Vocology, 37(1), 1–10. https://doi.org/10.3109/14015439.2011.625973 Yew, S. G. K., & O'Kearney, R. (2013). Emotional and behavioural out- comes later in childhood and adolescence for children with specific language impairments: meta‐analyses of controlled prospective studies. Journal of Child Psychology and Psychiatry, 54(5), 516–524. https://doi.org/10.1111/jcpp.12009 How to cite this article: Wren, Y., Pagnamenta, E., Orchard, F., Peters, T. J., Emond, A., Northstone, K., Miller, L. L., & Roulstone, S. (2023). Social, emotional and behavioural difficulties associated with persistent speech disorder in children: A prospective population study. JCPP Advances, 3(1), e12126. https://doi.org/10.1002/jcv2.12126
10.1167_jov.23.1.2
Journal of Vision (2023) 23(1):2, 1–13 1 The effect of a short-wave filtering contact lens on color appearance Billy R. Hammond John Buch Lisa M. Renzi-Hammond Jenny M. Bosten Derek Nankivil Vision Sciences Laboratory, Behavioral and Brain Sciences Program, Department of Psychology, University of Georgia, Athens, GA, USA Research & Development, Johnson & Johnson Vision Care, Inc., Jacksonville, FL, USA Institute of Gerontology, Department of Health Promotion and Behavior, University of Georgia, Athens, GA, USA University of Sussex, Falmer, UK Research & Development, Johnson & Johnson Vision Care, Inc., Jacksonville, FL, USA We assessed the effect of a contact lens that filters short-wavelength (SW) visible light on color appearance. These effects were modeled and measured by direct comparison to a clear contact lens. Sixty-one subjects were enrolled, and 58 completed as cohort; 31 were 18 to 39 years old (mean ± SD, 29.6 ± 5.6), 27 were 40 to 65 years old (50.1 ± 8.1). A double-masked contralateral design was used; participants randomly wore a SW-filtering contact lens on one eye and a clear control lens on the other eye. Subjects then mixed three primaries (including a short-wave primary, strongly within the absorbance of the test lens) until a perceived perfect neutral white was achieved with each eye. Color appearance was quantified using chromaticity coordinates measured with a spectral radiometer within a custom-built tricolorimeter. Color vision in natural scenes was simulated using hyperspectral images and cone fundamentals based on a standard observer. Overall, the chromaticity coordinates of matches that were set using the SW-filtering contact lens (n = 58; x = 0.345, y = 0.325, u(cid:2) = 0.222, v(cid:2) = 0.470) and clear contact lens (n = 58; x = 0.344, y = 0.325, u(cid:2) = 0.223, v(cid:2) = 0.471) were not significantly different, regardless of age group. Simulations indicated that, for natural scenes, the SW-filtering contact lens that was evaluated changes L/(L+M) and S/(L+M) chromatic contrast by no more than −1.4% to +1.1% and −36.9% to +5.0%, respectively. Tricolorimetry was used to measure color appearance in subjects wearing a SW-filtering lens in one eye and a clear lens in the other, and the results indicate that imparting a subtle tint to a contact lens, as in the SW-filtering lens that was evaluated, does not alter color appearance for younger or older subjects. A model of color vision predicted little effect of the lens on chromatic contrast for natural scenes. Introduction Tinted filters are often used in photography to alter the color appearance of images. It has also been hypothesized (e.g., Simunovic et al., 2012) that intraocular filters, such as blue-light filtering intraocular lenses (BLF IOLs) or lenticular or retinal macular pigments (MPs) have the potential to alter color perception. Bornstein (1973) argued, for example, that semantic differences in basic color terminology across cultures might be explained by group differences in average MP levels. Bornstein hypothesized that the high MP levels of some groups (e.g., individuals near the tropical equator who have a high intake of carotenoid-rich foods) would cause “a reduction in the perception of blueness.” The increasing yellowness of Claude Monet’s lens has often been used as an explanation for the slow disappearance of shorter wave colors in his impressionistic paintings (Steele & O’Leary, 2001). The retinal and lenticular pigments do, in fact, screen the foveal cones from a significant amount of light in the short-wave (SW) end of the visible spectrum (400–500 nm). Further, this amount Citation: Hammond, B. R., Buch, J., Renzi-Hammond, L. M., Bosten, J. M., & Nankivil, D. (2023). The effect of a short-wave filtering contact lens on color appearance. Journal of Vision, 23(1):2, 1–13, https://doi.org/10.1167/jov.23.1.2. https://doi.org/10.1167/jov.23.1.2 Received April 22, 2022; published January 3, 2023 ISSN 1534-7362 Copyright 2023 The Authors This work is licensed under a Creative Commons Attribution 4.0 International License. Journal of Vision (2023) 23(1):2, 1–13 Hammond et al. 2 can dramatically differ between individuals (Curran Celentano, Burke, & Hammond, 2002). At peak absorbance (460 nm, the dominant wavelength of blue sky light), MP screening (Hammond, Wooten, & Snodderly, 1997) may vary by more than 1 log unit of optical density (= 0.2 to over 1.3). When one adds SW screening of visible light by the crystalline lens (= 0.9 to over 1.6 at 407) (Wooten, Hammond, & Renzi, 2007), this creates a situation of fairly dramatic individual differences in the amount of SW light incident on the photoreceptors. Note that, although the eye therefore naturally filters light at 460 nm, the same wavelength associated with peak melatonin suppression (Rea, Bullough, & Figueiro, 2002), the test lens in this study does not. Nonetheless, empirical data show that these natural individual differences seem to have very little impact on color perception. For example, Stringham, Hammond, Wooten, and Snodderly, (2006) and Stringham and Hammond (2007) measured π 1 sensitivity and hue cancellation values in subjects with a wide range of MP density. Comparisons were made both across and within subjects (different locations on the retina, where MPs were either dense or mostly absent, were compared). No relation was found; the visual system presumably increases gain to offset filtering by even the highest levels of MP density. The visual system does not operate like a passive detector; rather, it can adapt to large variations in ambient lighting and compensate for similarly large variations in stable intraocular filters. The light-filtering characteristics of the natural lens, tinted intraocular implants, and macular pigment are all relatively stable features of the eye. When filtering changes quickly, however, the system requires time to renormalize. Delahunt, Webster, Ma, and Werner (2004) found that chromatic mechanisms following cataract surgery and implantation of a clear IOL took months to reset. Tregillus, Werner, and Webster (2016) used yellow spectacle lenses and tested renormalization each day over the span of a week. They found that some adaptive processes (such as simultaneous contrast) were quite fast (minutes), but full adaptation to the lenses (color perception similar to no lenses) often took hours (this process may be accelerated in individuals who habitually wear tinted lenses) (Engel, Wilkins, Mand, Helwig, & Allen, 2016). The results of Tregillus et al. (2016) indicate that alterations in color perception might be different with lenses that are extrinsic to the eye, such as spectacles or contact lenses, compared to intraocular filters that are stable over time. Alzahrani et al. (2020) modeled the effects of commercially available BLF lenses on the perception of color and concluded that they are “capable of reducing the perception of blue colours … by 5-36 per cent.” Thus, the literature may lead the practitioner to ask whether such filters affect color perception. Figure 1. Transmittance of the SW-filtering (orange) and control (blue) contact lenses. SW-filtering contact lenses contains chromophores that intentionally absorb specific wavelengths (380–450 nm) of visible light. Chromophores that filter in this range have a yellowish hue or “tint” imparted to them when viewed against a white background. The test lens also contains a bluish compound to normalize the cosmetic acceptance of the lens, thus resulting in a slightly teal color. This should not be confused with “tinted” contact lenses that contain pigments or dyes for the purpose of changing cosmetic appearance or making a lens easier to see for handling. See Figure 1 for the spectral transmission of the test and control lens used in this study. The test lens has the highest filtering (∼60%) within the range of 380 to 450 nm of any known soft contact lens. The control lens contains an ultraviolet-absorbing compound and a blue visibility tint and is considered relatively clear compared to the test lens. In this study, we assessed whether SW-filtering contact lenses would influence color appearance (as opposed to chromatic discrimination measured by tests such as the FM-100 hue test). Two approaches were used. The first was to measure color appearance using additive tricolorimetry where three color primaries (red, green, and blue) are mixed to achieve a perfect neutral white. Subject settings are plotted in CIE space as chromaticity coordinates. This type of geometric representation of color space allows specification of the appearance of any mixture of light, rather than its spectral composition, by specific coordinates. Disruptions from normal trichromatic color vision are represented as alterations in the relative amounts of the color primaries used to create a perfect perceived neutral white. Because each variable is on a continuous scale, even small alterations in color appearance can be assessed. Journal of Vision (2023) 23(1):2, 1–13 Hammond et al. 3 Although it is informative to characterize the effectiveness of such filters using precisely controlled and calibrated stimuli, any functional impact of tinted lenses will be realized in the natural world through their effect on light in natural scenes. To estimate the impact of a filter on color perception in natural environments, we also modeled the visual responses of individual observers to light in the real world. Thus, as a complement to the aforementioned empirical study, we developed modeling tools that permit quantitative characterization of the effect of filters on the color vision of individually specified observers. We present summary statistics of the predicted effects of the tested SW-filtering lenses on the color vision of a standard observer for a series of hyperspectral images of natural scenes. Methods Design of the empirical study and subjects A prospective, randomized, double-masked, contralateral design was used. Subjects were habitual wearers of spherical silicone hydrogel soft contact lenses with best-corrected visual acuity of 20/25 or better in each eye. Ishihara pseudoisochromatic plates were used to screen for color vision deficiency, and a basic clinical exam was used to exclude any overt ocular pathology (no subjects had to be excluded). A total of 61 subjects were enrolled from one clinical site (Georgia Center for Sight, Greensboro, GA). All enrolled subjects were randomly assigned to one of the two lens sequences (test/control or control/test) and completed the study (the lenses were worn throughout the single test session). All 61 subjects were included in the safety population, which included 31 subjects 18 to 39 years old (50.8%) and 30 subjects 40 to 65 years old (49.2%). Two age groups were selected to assess any change due to age; for example, lens yellowing with age is a possible additive influence (Artigas, Felipe, Navea, Fandino, & Artigas, 2012). Of the 61 subjects, three subjects were excluded from the per-protocol population due to lens dispensing with incorrect product codes and missing data. The per-protocol population (cohort) included 58 subjects, with 31 subjects 18 to 39 years old (29.6 ± 5.6; 53.4%) and 27 subjects 40 to 65 years old (50.1 ± 8.1; 46.6%). Of the 58 subjects from the per-protocol population, 44 were females (75.9%), and 14 were males (24.1%). The majority of subjects were white (42; 72.4%) and non-Hispanic or Latino (56; 96.6%), and their average age was 39.2 ± 12.34 years. The available contact lens powers were −1.00 through −6.00 diopters (D) in 0.25-D steps and were fit to achieve a plano spherical over-refraction OD and OS. There were no adverse events. Ethics The study was performed in accordance with “Clinical investigation of medical devices for human subjects” (ISO 14155:2011) and followed the tenets of the Declaration of Helsinki. Written informed (and verbal) consent was obtained from all subjects. The protocol was approved by the Sterling Institutional Review Board, Atlanta, GA. Apparatus and procedure A custom designed tricolorimeter was constructed to determine and specify the locus of perceptual white within the CIE chromaticity diagram (for additional details and a schematic, see Hammond, Wooten, Saint, & Renzi-Hammond, 2021). The optical system was built around two integrating spheres (Labsphere, North Sutton, NH). Each hemisphere was drilled for two apertures (1-inch diameter). The light source was a 1-inch-diameter, chip-on-board array of cool white light-emitting diodes (LEDs; 6500 color temperature). This light array was reflected within the sphere and diffused in all directions creating a Lambertian emitter where the luminance toward an observer was independent of the viewing position; the perception was one devoid of all texture and perfectly uniform. This light passed through a filter assembly composed of a red (R), a green (G), and a blue (B) filter (Wratten filters; #26, #40, and #47, respectively; Edmund Optics, Barrington, NJ). The tripartite filter assembly was mounted onto a vertical/horizontal micrometer. Adjustment of the filter along the horizontal direction resulted in either (1) an increase in red and a decrease in green, or (2) a decrease in red and an increase in green. Adjustment of the filter along the vertical direction resulted in either (1) an increase in both red and green and a decrease in blue, or (2) a decrease in both red and green and an increase in blue. With this, a wide ratio of the three colored filters could be set to sample a large subset of the CIE chromaticity diagram ranging from clearly red or green or blue to a perfect white and all colors in between. Optical baffles were used to prevent stray light between the various components. After passing through the RGB filter, light entered the second sphere, which served to additively mix the R, G, and B components thoroughly so that the emitted light was also Lambertian and color appearance was constant across the perceived target for a given RGB setting. Light from the second sphere was transmitted through a lens and then passed into a beam splitter where half the light was reflected onto a second lens that focused the light on the detector of a spectral radiometer, which calculated the chromaticity Journal of Vision (2023) 23(1):2, 1–13 Hammond et al. 4 coordinates. The other half of the collimated light was directed through an eye cup and into the eye. For an emmetrope, the image would be in sharp focus. For a myope or a hyperope, however, the image would be in front of the retina or behind the retina, respectively. This design allowed for control over the focus of the image for any observer by simply increasing or decreasing the distance between the aperture of the second sphere and the final lens. It is one example of a class of telecentric lens assemblies, and it provided an image magnification that was constant in size no matter the distance between the aperture and lens. For our application, the eye cup was fixed because it was the reference point for eye position. Therefore, we mounted the spheres on a platform that could be translated along the z-axis. The observer varied the lens position by turning a dial until perfect focus was achieved. Overview of the psychophysical technique Subjects wore the contact lenses for approximately 25 minutes before beginning an experimental session with the tristimulus colorimeter. Before the start of the session, the experimenter demonstrated to the subject how turning the knob changed the color appearance of the test stimulus and explained that the goal was to identify “pure white” (or “snow white,” no identifiable tints or colors). The tester then moved the control knobs to achieve a maximal saturation starting point that participants reliably identified as “red.” From that point, the subject verbally guided the experimenter’s adjustments until no hint of hue was perceived, with a criterion point of “pure white.” The same procedure (below) was repeated for maximal saturation starting points that participants reliably identified as red, green, and blue. Finding “white” From each of these three starting points, the experimenter systematically adjusted one axis (e.g., red–green) at a time and instructed the subject to state when the visual field was as close to white as possible along that axis, at which point the experimenter stopped turning that dial. Then, the experimenter moved to the other axis (red–green–blue in this example) and adjusted that dial until the visual field was as close to white as possible. This continued, back and forth, with small, fine-tuned adjustments, until the subject reported that the visual field was pure white without any tint of color. When an approximate white setting was obtained, subjects were asked to look away from the device for approximately 5 seconds, and then back into the eye piece to double check that the visual field did not contain any tint of color at that setting. They were asked such questions as, “If you had to give what you are seeing a color name, what would you call it?” Also, using that setting, the experimenter turned each knob to bracket the area (meaning, at the point where the subject first perceived a tint). Four trials, one from each starting point, were collected for each eye. For each measurement, the spectroradiometer provided measures of the stimulus color as in chromaticity coordinates (x, y and u(cid:3), v(cid:3)) and stimulus illuminance (lux). Design of the modeling study: Simulating color vision in natural scenes Hyperspectral images Any shifts in color chromaticity conferred by the SW-filtering lens may not equate to changes in color appearance, as chromatic adaptation may rebalance the color appearance to compensate for initial changes in receptor signals caused by the filter. However, the colored filter may still affect the range of color contrasts available to the observer. We were interested in predicting the effect of the filter on the available color gamut for natural scenes. Colors in RGB images have very different spectra from those in natural scenes, and correspondingly the effects of filters on RGB images will be very different from their effects on natural scenes. It is therefore not possible to fully simulate the effect of filters on the color perception of natural scenes solely from RGB images; instead, hyperspectral images are required, where full spectral information is available at each pixel. To this end, we used five sets of hyperspectral images. Three images sets were obtained from the literature (Chakrabarti & Zickler, 2011; Nascimento, Ferreira, & Foster, 2002; Parraga, Brelstaff, Troscianko, & Moorhead, 1998), and one image set was acquired in support of this study. The image sets obtained from the literature were chosen from those available for their high precision and lack of spectral artifacts. A calibrated IQ camera (Specim Spectral Imaging Oy, Ltd., Oulu, Finland) was used to gather a custom set of hyperspectral images of scenes created in the lab under controlled lighting and outdoor scenes. The custom image set consisted of 35 images featuring a mixture of manmade and colored objects, including objects where color discrimination is often critical (e.g., maps, colored markers, flowers, traffic lights, colored threads and cables, fruit). The spatial resolution of the images was 512 × 512, and the native spectral resolution was 204 wavebands between 400 and 1000 nm. These custom images were calibrated using a PR655 spectroradiometer (PhotoResearch, Chatsworth, CA), and colored stimuli were presented on a cathode-ray tube (CRT) monitor (Mitsubishi, Tokyo, Japan). By measuring a series of the same stimuli with both the PR655 and the Specim IQ, a vector of mean Journal of Vision (2023) 23(1):2, 1–13 Hammond et al. 5 Figure 2. Calibration of Specim IQ. Nine colored stimuli were presented on a Mitsubishi Diamond Pro CRT monitor. The spectra of the stimuli are shown measured using a PR655 spectroradiometer (red line) and using the calibrated Specim IQ hyperspectral camera (blue open circles). wavelength-specific scaling factors was calculated that allowed the camera-specific intensities of the Specim IQ to be transformed into radiance units. The wavelengths output by the Specim IQ were not calibrated, but the calibration was checked using an independent set of stimuli (Figure 2). The two sets of publicly available hyperspectral images captured by Foster, Amano, Nascimento, and Foster (2006) are mostly of outdoor scenes including foliage, flowers, countrysides, and cityscapes. The images are 820 × 820 spatially with 33 spectral wavebands between 400 and 720 nm. The Nascimento et al. (2002) dataset consists of eight images with reflectance information only (i.e., as if the scenes were imaged under equal energy white). The Parraga et al. (1998) dataset consists of 29 images of outdoor scenes, including trees, foliage, flowers, bark, and soil. The images provided radiance for 31 wavebands between 400 and 700 nm with a spatial resolution of 256 × 256. The Chakrabarti and Zickler (2011) dataset includes 25 indoor images and 38 outdoor images. Several images were eliminated from Chakrabarti and Zickler’s full set because they contained spectral artifacts (camera saturation). The spatial resolution of the images was 1040 × 1392 with intensities at 31 wavebands between 420 nm and 720 nm. The images were calibrated to account for camera sensitivity but were not radiance calibrated. Individual observer The individual observer model was based on the Stockman and Sharpe (2000) nomogram (Figure 3). The standard normal observer was defined using in vitro peak receptoral sensitivities of 558.9 nm, 530.3 nm, and 430.7 nm. Results of the nomogram, showing example normalized cone fundamentals created between 420 nm and 560 nm, are shown in Figure 3. Additional model features included macular pigment optical density (= 0.35); age-dependent lens density function using Journal of Vision (2023) 23(1):2, 1–13 Hammond et al. 6 Figure 3. (Left) Stockman and Sharpe (2000) nomogram for simulating cone fundamentals of arbitrary peak sensitivity. (Right) The same simulated cone fundamentals for a 20-year-old observer with a macular pigment density of 0.35 and optical density of 0.38. equations provided by Pokorny, Smith, and Lutze (1987) (e.g., 20 years); receptoral optical density (e.g., 0.38, 0.38, and 0.3 for L, M, and S as recommended by Stockman and Sharpe, 2000); and scaling factors for the relative peak heights of the L and M cone fundamentals so that L = 1.5M as recommended by Stockman and Sharpe (2000). The scaling factor for the S-cone fundamental is arbitrary. Quantifying the impact of a filter on color vision After defining the observer, each hyperspectral image was transformed to observer-specific LMS values (with and without the additional filter). First, the observer-specific matrices of MacLeod and Boynton (1979) chromaticity coordinates were created and translated to center the white point at (0, 0), which was calculated as the average MacLeod–Boynton chromaticity coordinates S/(L+M) and L/(L+M) over all pixels in the image. Gamut changes were then calculated relative to the white point. The impact of the filter was calculated as the difference between the absolute MacLeod–Boynton chromaticity coordinates of the scene with the filter and the absolute MacLeod–Boynton chromaticity coordinates of the scene without the filter. Using the absolute values ensures that any enhancement in chromatic contrast from white (along the bipolar chromaticity dimension) is a positive number in the difference metric, whereas any reduction in chromatic contrast from white is a negative number. The size of the change in pixel chromatic contrast caused by the filter was quantified as a percentage of 95% of the gamut of the scene without the filter, which was calculated as difference between the 2.5th and 97.5th percentiles of the MacLeod–Boynton chromaticity coordinates of all pixels in the image. Simulations were conducted for macular pigment optical density = 0.35 and lens age = 20. Results Empirical study For this within-subject comparison, the chromaticity coordinates of eyes with the SW-filtering test contact lens (n = 58; x = 0.345, y = 0.325, u(cid:3) = 0.222, v(cid:3) = 0.470) was not significantly different from eyes with the clear contact lens (n = 58; x = 0.344, y = 0.325, u(cid:3) = 0.223, v(cid:3) = 0.471) (see Figure 4, Table 1). This was also true when the subjects were separated by age into young (18–39 years, n = 31) and older (40–65 years, n = 27) groups. Within-subject differences indicate that Figure 4. Subjective white point of all subjects. The size of the marker (open circle) represents the standard deviation, and the asterisk indicates the range. Journal of Vision (2023) 23(1):2, 1–13 Hammond et al. 7 Age group: 18–39 years Age group: 40–65 years Total Test lens (n = 31) Control lens (n = 31) Test lens (n = 27) Control lens (n = 27) Test lens (n = 58) Control lens (n = 58) Color appearance (u(cid:3)) Mean (SD) Median Min–Max Color appearance (v(cid:3)) Mean (SD) Median Min–Max Color appearance (x) Mean (SD) Median Min–Max Color appearance (y) Mean (SD) Median Min–Max 0.222 (0.004) 0.222 0.212–0.233 0.223 (0.005) 0.222 0.212–0.236 0.223 (0.007) 0.223 0.210–0.240 0.223 (0.005) 0.223 0.213–0.232 0.222 (0.005) 0.222 0.210–0.240 0.223 (0.005) 0.222 0.212–0.236 0.471 (0.007) 0.472 0.459–0.486 0.471 (0.007) 0.470 0.459–0.485 0.468 (0.015) 0.470 0.400–0.486 0.470 (0.007) 0.470 0.457–0.486 0.470 (0.011) 0.471 0.400–0.486 0.471 (0.007) 0.470 0.457–0.486 0.345 (0.010) 0.343 0.332–0.368 0.344 (0.010) 0.344 0.322–0.371 0.345 (0.012) 0.344 0.317–0.366 0.345 (0.011) 0.344 0.325–0.370 0.345 (0.011) 0.343 0.317–0.368 0.344 (0.011) 0.344 0.322–0.371 0.326 (0.010) 0.326 0.308–0.349 0.325 (0.010) 0.325 0.308–0.348 0.324 (0.011) 0.324 0.308–0.353 0.325 (0.012) 0.323 0.304–0.346 0.325 (0.010) 0.325 0.308–0.353 0.325 (0.011) 0.325 0.304–0.348 Table 1. Chromaticity for the younger (left), older (middle), and combined (right) groups with SW-filtering test and clear control lenses in the per-protocol (cohort) population. Min = minimum; Max = maximum. Figure 5. Histograms of the within-subject differences in subjective white points. Color difference ((cid:3)E) and direction (θ ) (top) and change in x, (cid:3)x, y, and (cid:3)y (bottom). Here, the vector (cid:3)E is given by (cid:3)E = test − v(cid:3) vector (θ ) is given by θ = tan−1((v(cid:3) control )2, and the angle of the control )2 + (v(cid:3) test − u(cid:3) control )/(u(cid:3) test − v(cid:3) test − u(cid:3) control )). (u(cid:3) (cid:2) Journal of Vision (2023) 23(1):2, 1–13 Hammond et al. 8 Illuminance (lux) Age group: 18–39 years Age group: 40–65 years Total Test lens (n = 31) 113 (6) 115 103–125 Control lens (n = 31) 113 (6) 113 102–123 Test lens (n = 27) 116 (5) 115 104–126 Control lens (n = 27) 116 (6) 117 103–128 Test lens (n = 58) 114 (5) 115 104–126 Control lens (n = 58) 114 (6) 115 102–128 Mean (SD) Median Min–Max Table 2. Illuminance for the younger (left), older (middle), and combined (right) groups with SW-filtering test and clear control lenses in the per-protocol (cohort) population. the vast majority of subjects had a color difference ((cid:3)E) less than 0.01 with no systematic bias in the color direction (θ ). Similarly, within-subject differences in x, (cid:3)x, y, and (cid:3)y were consistently less than 0.02 (Figure 5). We also measured the energy required to make the matches (illuminance) and found no differences across lens types or age (Table 2). Although participants were not directly asked about their subjective experiences wearing the lens, no participants indicated noticing a difference that specifically related to the SW-filtering test lens. Modeling study: Results of the analysis of hyperspectral images Calculated chromatic contrast changes in L/(L+M) as a percentage of 95% of the unfiltered image gamut for each dataset show that the test lens is predicted to cause very little change in L/(L+M) chromatic contrast. Results vary by scene, and each dataset is composed of different scenes, so we present all results and report the range across datasets herein. L/(L+M) chromatic contrast is predicted to change by, at most, –1.4% and +1.1%. Between 1% and 16% of pixels (for different datasets) exhibit L/(L+M) chromatic contrast reductions of more than 0.5%, and between 1% and 7% of pixels exhibit L/(L+M) chromatic contrast enhancements of more than 0.5%. Although the change is minor, between 35% and 67% of pixels exhibit L/(L+M) chromatic contrast reductions, and the average reduction is between 0.0% and 0.1% (Figure 6, left). Conversely, calculated chromatic contrast changes in S/(L+M) as a percentage of 95% of the unfiltered image gamut show that the test lens is predicted to generally degrade S/(L+M) chromatic contrast. Results varied slightly by scene and dataset such that the S/(L+M) chromatic contrast for individual pixels Figure 6. Percentile changes in chromatic contrast for L/(L+M) (left) and S/(L+M) (right) as a percentage of the original image gamut for age = 20 years and macular pigment optical density = 0.35. The collective dataset is comprised of five sets of hyperspectral images: Four image sets were obtained from the literature: Chakrabarti and Zickler (2011); outdoor, Chakrabarti & Zickler (2011); indoor, Nascimento et al. (2002); and Parraga et al. (1998). One image set was acquired using the Specim IQ hyperspectral camera. Journal of Vision (2023) 23(1):2, 1–13 Hammond et al. 9 Figure 7. Hyperspectral image of a colored pencil set and assessments of the impact of the filter. (Top left) Original hyperspectral image recast in RGB. (Top right) Filtered hyperspectral image recast in RGB. (Middle left) Change in L/(L+M) as a percentage of 95% of the original image gamut. (Middle right) Change in S/(L+M) as a percentage of 95% of the original image gamut. (Bottom left) Change in L+M as a percentage of 95% of the original image gamut. (Bottom right) Scatterplot of the chromatic contrast of pixels in the image shown with (blue) and without (red) the filter. Journal of Vision (2023) 23(1):2, 1–13 Hammond et al. 10 was predicted to change by, at most, –36.9% and +5.0%. Between 82% and 92% of pixels exhibited S/(L+M) chromatic contrast reductions of more than 0.5%, and between 5% and 10% of pixels exhibited S/(L+M) chromatic contrast enhancements of more than 0.5%. Between 87% and 94% of pixels exhibited S/(L+M) chromatic contrast reductions, and the average reduction was between 3.6 and 8.2% (Figure 6, right). A single example of a hyperspectral image of a color pencil set (from the custom Specim IQ dataset) is provided to illustrate the impact of the filter (Figure 7). Qualitatively, comparing the image with and without the filter (Figure 7, top row), the image appears a bit darker, but the colors are generally well represented in the filtered image. Quantitatively, we see that along L/(L+M) some of the red colors are enhanced slightly by as much as 0.5% and some of the blue colors are degraded slightly by as much as 0.5%. Along S(L+M), most of the image content is degraded, and the orange and blue were most degraded by as much as 10% (Figure 7, middle row). The luminance of the image was reduced slightly by as much as 3% (Figure 7, bottom left). The gamut is generally well preserved with some contraction along S/(L+M), again driven by the orange and blue surfaces in the image (Figure 7, bottom right). Discussion The term “color” is used in many different ways across numerous fields, but, broadly, color is a perception. Unlike many perceptions, however, the perception of color can be very precisely quantified using the CIE colorimetry system. We used a tristimulus colorimeter along with a spectral radiometer and spectral computations based on color matching functions. Subjects were asked to mix three primaries to achieve a perfect white (a color sensation without hue). This was done on subjects wearing a SW-filtering test contact lens in one eye and a clear contact lens in the other. Based on this contralateral design of the clinical Figure 8. (Upper left) Filter transmittance (orange) and the normalized S-cone responsivity with no lens (blue). (Upper right) Normalized S-cone responsivity after applying the transmittance functions of either the SW-filtering test (orange) or the clear control (blue) contact lens. (Bottom left) Spectral radiance of the blue primary filtered by the SW-filtering test (orange) and the clear control (right) contact lenses. The peak amplitude and wavelength of the best-fit Gaussian is shown at the top right. (Bottom right) Relative S-cone activation to the blue primary with the SW-filtering test (orange) and the clear control (right) contact lenses. The cumulative total relative S-cone activation is shown in the top right. Journal of Vision (2023) 23(1):2, 1–13 Hammond et al. 11 phase of our study, with 58 subjects younger and older, we found that subtle tinting of a contact lens in the violet/blue region of the spectrum did not alter color appearance values in normal trichromatic subjects. Note that we tested color vision monocularly using a contralateral design. Hence, it is possible that wearing the lens in both eyes using a between-subject design may have yielded different results. This, however, seems unlikely given the fact that color perception tends to be yoked (e.g., Roth, Pelizzone, & Hermès, 1989), nor is it uncommon to find intraocular differences in natural chromophores (e.g., yellowing of the crystalline lens) (Artigas et al., 2012) that exceeds the tinting in the test lens. Tinting contact lenses has been suggested as a means of influencing chromatic discrimination (not color appearance) in patients with color vision deficiencies (Elsherif, Salih, Yetisen, & Butt, 2021). Heavy tinting (such as is sometimes done with lenses used in sports) has been shown to influence chromatic discrimination in normal subjects (Harris & Cabrera, 1976; Laxer, 1990). The effects of tinting as one might see in more common use on color appearance, however, has not been tested until this study. Our results suggest that the effects of a modest SW filter on the white point are negligible. However, the SW filter did reduce S-cone responsivity by more than 20% for wavelengths below 433 nm, and the blue primary of the colorimeter has a full-width-at-half-maximum bandwidth spanning the range of 428 to 458 nm. As a result, the SW filter we tested in this study had a very modest impact on the blue primary of the tristimulus colorimeter, shifting the peak wavelength by only 1.6 nm and decreasing the amplitude of the irradiance by only 5.4%, and reducing the relative S-cone response by 8% (Figure 8). Given the modest impact of the SW filter on the blue primary, it is not incredibly surprising that effects on color perception were negligible. Whereas this result may not be terribly surprising, it is important to note that this type of primary is generally similar to those found in common LED display technology, where the blue primary of the display does not emit much light below 428 nm. Thus, this null result suggests that the SW-filtering lens does not adversely impact color perception of displays with primaries that are similar to the colorimeter presented herein. Moreover, any functional impact of tinted lenses will be realized in the natural world through their effect on light in natural scenes. To estimate the impact of a SW-filtering lens on color perception in natural environments, we modeled the visual responses of individual observers to light in the real world. Summary statistics of the predicted effects of the tested tinted lenses on the gamut of colors available to a standard observer in a series of hyperspectral images acquired in natural scenes indicate that the SW-filtering lenses are expected to degrade L/(L+M) chromatic contrast by 0% to 0.1% on average and to degrade S/(L+M) chromatic contrast by between 3.6% and 8.2% on average. In summary, the SW-filtering lenses are generally not expected to change the gamut of color contrasts appreciably; however, substantial degradation of S/(L+M) contrast is expected in a minority of scene content. Evidence in the literature supports the notion that visual adaptation occurs over multiple time scales (Wark, Fairhall, & Rieke, 2009). Short-term changes in perceived contrast have been observed with chromatic contrast adaptation for a duration of only 1 hour (Tregillus & Webster, 2014). However, our experiments were conducted after each participant wore the contact lenses for about 25 minutes. The total experimental duration was approximately 2 hours, and the experiments was not designed to assess adaptation. It is thus plausible that adaptation was partially responsible for the consistency in white point observed comparing clear and SW-filtering contact lenses. Conclusion “I wept when I saw the color of the sea—how can a mere color make one cry?” (Ludwig Boltzmann, as quoted Greenstein, 1991). Color perception is profoundly important. Eyecare professionals are at once both willing to embrace tinted lenses to improve quality of life while being concerned that such lenses may degrade color perception. In this study, the impact of a short-wave filter on color appearance was assessed experimentally by tricolorimetry and theoretically by simulating a standard observer in a series of hyperspectral images acquired in natural scenes. Results indicate that the short-wave filtering lens that was evaluated generally does not appreciably alter color appearance as predicted by the model. Keywords: color vision, color appearance, blue-light filter, short-wave filter, contact lenses Acknowledgments The authors thank Jie Xu for providing individual subject data and summary descriptive statistics from the clinical study. The authors also recognize the support of Lauren Hacker, OD, in participant recruitment. Supported by Johnson & Johnson Vision Care, Inc. Commercial relationships: B.R. Hammond, Johnson & Johnson Vision Care, Inc. (F); J. Buch, Johnson & Johnson Vision Care, Inc. (E, I, P); L.M. Renzi- Hammond, Johnson & Johnson Vision Care, Inc. (F); Journal of Vision (2023) 23(1):2, 1–13 Hammond et al. 12 J.M. Bosten, Johnson & Johnson Vision Care, Inc. (C); D. Nankivil, Johnson & Johnson Vision Care, Inc. (E, I, P). Corresponding author: Billy R. Hammond. Email: bhammond@uga.edu. Address: Vision Sciences Laboratory, Behavioral and Brain Sciences Program, Department of Psychology, University of Georgia, Athens, GA, USA. References Alzahrani, H. S., Khuu, S. K., & Roy, M. (2020). Modelling the effect of commercially available blue-blocking lenses on visual and non-visual functions. Clinical and Experimental Optometry, 103(3), 339–346. Artigas, J. M., Felipe, A., Navea, A., Fandino, A., & Artigas, C. (2012). Spectral transmission of the human crystalline lens in adult and elderly persons: Color and total transmission of visible light. Investigative Ophthalmology & Visual Science, 53(7), 4076–4084. Bornstein, M. H. (1973). The psychophysiological component of cultural difference in color naming and illusion susceptibility. Behavior Science Notes, 8(1), 41–101. Chakrabarti, A., & Zickler, T. (2011). Statistics of real-world hyperspectral images. In: Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition. Piscataway, NJ: Institute of Electrical and Electronics Engineers. Curran Celentano, J., Burke, J. D., & Hammond, B. R. (2002). In vivo assessment of retinal carotenoids: Macular pigment detection techniques and their impact on monitoring pigment status. The Journal of Nutrition, 132(3), 535S–539S. Delahunt, P. B., Webster, M. A., Ma, L., & Werner, J. S. (2004). Long-term renormalization of chromatic mechanisms following cataract surgery. Visual Neuroscience, 21(3), 301. Elsherif, M., Salih, A. E., Yetisen, A. K., & Butt, H. (2021). Contact lenses for color vision deficiency. Advanced Materials Technologies, 6(1), 2000797. Engel, S. A., Wilkins, A. J., Mand, S., Helwig, N. E., & Allen, P. M. (2016). Habitual wearers of colored lenses adapt more rapidly to the color changes the lenses produce. Vision Research, 125, 41–48. Foster, D. H., Amano, K., Nascimento, S. M. C., & Foster, M. J. (2006). Frequency of metamerism in natural scenes. Journal of the Optical Society of America A: Optics and Image Science, and Vision, 23, 2359–2372. Greenstein, G. (1991). Science: The Bulldog: A profile of Ludwig Boltzmann. The American Scholar. 60(1):97–105. Hammond, B. R., Wooten, B. R., & Snodderly, D. M. (1997). Individual variations in the spatial profile of human macular pigment. Journal of the Optical Society of America A: Optics and Image Science, and Vision, 14(6), 1187–1196. Hammond, B. R., Wooten, B. R., Saint, S. E., & Renzi-Hammond, L. (2021). The effects of a blue-light filtering versus clear intraocular implant on color appearance. Translational Vision Science & Technology, 10(12), 25. Harris, M. G., & Cabrera, C. R. (1976). Effect of tinted contact lenses on color vision. American Journal of Optometry and Physiological Optics, 53(3), 145–148. Laxer, M. (1990). Soft tinted contact lenses and color discrimination. International Contact Lens Clinic, 17(3-4), 88–91. MacLeod, D. I., & Boynton, R. M. (1979). Chromaticity diagram showing cone excitation by stimuli of equal luminance, Journal of the Optical Society of America, 69, 1183–1186. Nascimento, S. M. C., Ferreira, F. P., & Foster, D. H. (2002). Statistics of spatial cone-excitation ratios in natural scenes. Journal of the Optical Society of America A: Optics and Image Science, and Vision, 19, 1484–1490. Parraga, C. A., Brelstaff, G., Troscianko, T., & Moorhead, I. R. (1998). Color and luminance information in natural scenes. Journal of the Optical Society of America A: Optics and Image Science, and Vision, 15(3) 563–569. Pokorny, J., Smith, V. C., & Lutze, M. (1987). Aging of the human lens, Applied Optics, 26, 1437–1440. Rea, M. S., Bullough, J. D., & Figueiro, M. G. (2002). Phototransduction for human melatonin suppression. Journal of Pineal Research, 32(4), 209–213. Roth, A., Pelizzone, M., & Hermès, D. (1989). The two equation method I. Results in normal color vision. In: Colour Vision Deficiencies IX: Proceedings of the Ninth Symposium of the International Research Group on Colour Vision Deficiencies. Dordrecht: Springer; 317–323. Simunovic, M. P. (2012). On seeing yellow: The case for, and against, short-wavelength light–absorbing intraocular lenses. Archives of Ophthalmology, 130(7), 919–926. Steele, M., & O’Leary, J. P. (2001). Monet’s cataract surgery. The American Surgeon, 67(2), 196. Stockman, A., & Sharpe, L. T. (2000). Spectral sensitivities of the middle- and long-wavelength Journal of Vision (2023) 23(1):2, 1–13 Hammond et al. 13 sensitive cones derived from measurements in observers of known genotype. Vision Research, 40, 1711–1737. Stringham, J. M., & Hammond, B. R., Jr. (2007). Compensation for light loss due to filtering by macular pigment: Relation to hue cancellation. Ophthalmic and Physiological Optics, 27(3), 232–237. Stringham, J., Hammond, B. R., Wooten, B. R., & Snodderly, D. M. (2006). Compensation for light loss due to filtering by macular pigment: Relation to the S-cone pathway. Optometry and Vision Sciences, 83(12), 887–894. Tregillus, K., & Webster, M. A. (2014). Dynamics of color contrast adaptation. Journal of the Optical Society of America A: Optics and Image Science, and Vision, 31(4), A314–A321. Tregillus, K. E., Werner, J. S., & Webster, M. A. (2016). Adjusting to a sudden “aging” of the lens. Journal of the Optical Society of America A: Optics and Image Science, and Vision, 33(3), A129– A136. Wark, B., Fairhall, A., & Rieke, F. (2009). Timescales of inference in visual adaptation. Neuron, 61(5), 750–761. Wooten, B. R., Hammond, B. R., & Renzi, L. M. (2007). Using scotopic and photopic flicker to measure lens optical density. Ophthalmic and Physiological Optics, 27(4), 321–328.
10.3390_molecules28020474
Article Structure-Based Identification of Natural-Product-Derived Compounds with Potential to Inhibit HIV-1 Entry Nneka Ugwu-Korie 1, Osbourne Quaye 1 Emmanuel Broni 4,5,6 , Yash Gupta 7 , Edward Wright 2 , Sylvester Languon 1,3 , Odame Agyapong 4,5 , , Prakasha Kempaiah 7 and Samuel K. Kwofie 1,4,* 1 West African Centre for Cell Biology of Infectious Pathogens, Department of Biochemistry, Cell and Molecular Biology, College of Basic and Applied Sciences, University of Ghana, Legon, Accra P.O. Box LG 54, Ghana School of Life Sciences, University of Sussex, Brighton BN1 9QG, UK 2 3 Cellular and Molecular Biomedical Sciences Program, University of Vermont, Burlington, VT 05405, USA 4 Department of Biomedical Engineering, School of Engineering Sciences, College of Basic & Applied Sciences, University of Ghana, Legon, Accra P.O. Box LG 77, Ghana 5 Department of Parasitology, Noguchi Memorial Institute for Medical Research (NMIMR), College of Health Sciences (CHS), University of Ghana, Legon, Accra P.O. Box LG 581, Ghana 6 Department of Medicine, Loyola University Medical Center, Maywood, IL 60153, USA 7 Infectious Diseases, Mayo Clinic, Jacksonville, FL 32224, USA * Correspondence: skkwofie@ug.edu.gh; Tel.: +233203797922 Abstract: Broadly neutralizing antibodies (bNAbs) are potent in neutralizing a wide range of HIV strains. VRC01 is a CD4-binding-site (CD4-bs) class of bNAbs that binds to the conserved CD4- binding region of HIV-1 envelope (env) protein. Natural products that mimic VRC01 bNAbs by interacting with the conserved CD4-binding regions may serve as a new generation of HIV-1 en- try inhibitors by being broadly reactive and potently neutralizing. This study aimed to identify compounds that mimic VRC01 by interacting with the CD4-bs of HIV-1 gp120 and thereby inhibit- ing viral entry into target cells. Libraries of purchasable natural products were virtually screened against clade A/E recombinant 93TH057 (PDB: 3NGB) and clade B (PDB ID: 3J70) HIV-1 env protein. Protein–ligand interaction profiling from molecular docking and dynamics simulations showed that the compounds had intermolecular hydrogen and hydrophobic interactions with conserved amino acid residues on the CD4-binding site of recombinant clade A/E and clade B HIV-1 gp120. Four potential lead compounds, NP-005114, NP-008297, NP-007422, and NP-007382, were used for cell-based antiviral infectivity inhibition assay using clade B (HXB2) env pseudotype virus (PV). The four compounds inhibited the entry of HIV HXB2 pseudotype viruses into target cells at 50% inhibitory concentrations (IC50) of 15.2 µM (9.7 µg/mL), 10.1 µM (7.5 µg/mL), 16.2 µM (12.7 µg/mL), and 21.6 µM (12.9 µg/mL), respectively. The interaction of these compounds with critical residues of the CD4-binding site of more than one clade of HIV gp120 and inhibition of HIV-1 entry into the target cell demonstrate the possibility of a new class of HIV entry inhibitors. Keywords: HIV-1 entry; inhibition; CD4-binding site; VRC01; gp120; virtual screening; small com- pound mimetics Citation: Ugwu-Korie, N.; Quaye, O.; Wright, E.; Languon, S.; Agyapong, O.; Broni, E.; Gupta, Y.; Kempaiah, P.; Kwofie, S.K. Structure-Based Identification of Natural -Product-Derived Compounds with Potential to Inhibit HIV-1 Entry. Molecules 2023, 28, 474. https://doi.org/10.3390/ molecules28020474 Academic Editor: Francisco Torrens Received: 3 September 2022 Revised: 15 December 2022 Accepted: 24 December 2022 Published: 4 January 2023 1. Introduction Copyright: © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). Human immunodeficiency virus (HIV) entry is a complex interaction of host cell CD4 receptor with chemokine co-receptors (CXCR4 and CCR5) and the HIV env protein (trimer of gp120 non-covalently bound to gp41). The process comprises attachment to the host cell, binding of CD4 receptor, co-receptor binding, cell fusion, and deposition of viral genetic material for viral replication [1]. Targeting HIV entry into the host cell presents opportunities for therapeutic intervention as a treatment and prevention strategy for viral infection [2,3]. Only two viral entry inhibitors (enfuvirtide (fusion inhibitor) and maraviroc (CCR5 antagonist)) have been approved for use in the clinical treatment of HIV by the US Molecules 2023, 28, 474. https://doi.org/10.3390/molecules28020474 https://www.mdpi.com/journal/molecules molecules Molecules 2023, 28, 474 2 of 29 Food and Drug Administration (FDA) [2]. However, there are limitations associated with the clinical use of these entry inhibitors for the treatment of HIV infection. Treatment with maraviroc is associated with the emergence of CXCR4 tropic viruses [4,5], and enfuvirtide is a large polypeptide administered subcutaneously which is associated with painful injection sites and therefore limits the clinical use of the drug [6]. Broadly neutralizing antibodies (bNAbs) of human immunodeficiency type-1 (HIV-1) are monoclonal antibodies that solely target the env protein of HIV-1. They possess excep- tional potency and breadth against variant strains of HIV-1 due to their ability to interact with conserved regions of HIV env protein [7–9]. These bNAbs have been shown to com- pletely prevent simian–human immunodeficiency virus (SHIV) infections [10], suppress viremia in HIV-1-infected individuals [11], and effectively control HIV-1 infection and suppress viremia in humanized mice after therapy was discontinued for about sixty days— owing to the longer half-life of the antibodies compared to antiretrovirals (ARVs) [12–14]. CD4-binding-site (CD4bs) bNAbs exhibit high potency and breadth to variant HIV strains, since the binding site is one of the most functionally conserved sites of the HIV-1 env protein and is required for initial CD4 binding and successful infection of a host cell by HIV [15]. VRC01 is an example of CD4bs bNAbs with a neutralization breadth of about 80–90% of HIV strains and a high potency (IC50 of 0.33 µg/mL) against HIV group M [8] and clade B HIV-1 (IC50 between 10 and 25 µg/mL in a TZM-bl neutralization assay) [16]. As a CD4-bs, VRC01 shares certain similarities with CD4. The VRC01 binding site on HIV-1 gp120 is 98% similar to that of CD4 [17], although CD4 prefers gp120 in an unbound state, while VRC01 has an affinity for gp120 in both CD4-bound and unbound states [17]. Other similarities between CD4 and VRC01 include the hydrogen bonding interaction between Arg59CD4 and Gly54VRC01 with Asp368gp120 [17]. VRC01 has demonstrated efficacy in the prevention of HIV in humanized mice, prolonged the time of HIV rebound during treatment interruption, and is currently in clinical trials for prevention of HIV transmissions in human [18–20]. Furthermore, VRC01 infusions into chronically HIV-1-infected indi- viduals showed that VRC01 temporarily prevented viral replication in study participants, implying that at higher VRC01 concentrations, long-term viral rebound regulation could be achieved [21]. Even though HIV bNAbs such as VRC01 have the potential of being broad and potent HIV entry inhibitors due to promising results as a potential therapeutic agent, the antibodies are saddled with challenges that are associated with development as therapeutics. The challenges include laborious and time-consuming production techniques, high cost of production, instability of the developed antibody, delivery systems of the antibodies, and risk of treatment failure due to variation in immunological responses of antibodies therapy [22]. Moreover, the different clades have mutations in gp120 protein affecting the neutralization potential of these antibodies [23]. To overcome these challenges, small molecule mimetics of large biological molecules can be developed as therapeutic agents against HIV infection. Natural products (NP) and their derivatives have traditionally played an important role in drug discovery [24]. Libraries of NPs and their derivatives are distinguished by the complexity and diversity of their structures in addition to their molecular rigidity when compared to the synthetic small compound libraries [25]. These characteristics confer an advantage when inhibiting protein–protein interactions [26]. Compounds derived from NPs have been shown to inhibit critical steps in HIV infections [27,28]. Arylnaphthaline lignin glycoside, “patentiflorin A”, which was derived from the Justica gendarussa plant, showed inhibitory activity against HIV isolates resistant to azidothymidine (AZT) and nevirapine (NVP). Calanolide A and B derived from the mangosteen family tree Calophyllum lanigerum have exhibited inhibition of non-nucleoside reverse transcriptase (NNRT) and are currently in development. Furthermore, Rheum palmatum L, Kuwanon-L, and Betulinic acid have demonstrated HIV integrase inhibition [27]. In silico techniques have been used in the early stages of drug discovery to identify hit compounds and elucidate the mechanism of action of natural products, thereby low- Molecules 2023, 28, 474 3 of 29 ering time and costs and increasing efficiency [29,30]. Previous studies have combined cheminformatics-based and experimental-based studies to identify inhibitors for various diseases [31,32]. Computational approaches such as molecular docking, high-throughput virtual screening, molecular dynamics simulations, and quantitative structure–activity rela- tionship (QSAR) have also been deployed in HIV research to discover small compounds with potential activity against HIV [33–36]. This study, therefore, sought to use a computational approach to identify compounds that mimic the broadly neutralizing antibody VRC01 in interaction with conserved amino acid residues of CD4-bs and assess the in vitro inhibitory activity of these mimetic com- pounds on HIV-1 infection using pseudotype technology. 2. Results 2.1. In Silico High-Throughput Screening of Natural Compound Mimetics of VRC01 2.1.1. Protein Structure Extraction The crystal structure of PDB ID: 3NGB with a resolution of 2.68 Å was obtained from the Protein Data Bank (PDB). 3NGB is a complex of the crystallized antigen-binding fragment (Fab) of the broadly neutralizing antibody VRC01 and the core gp120 of HIV-1 clade A/E recombinant 93TH057. The protein consists of twelve chains comprising G, A, D, I, H, B, E, J, L, C, F, and K. Chains G, A, D, and I are HIV-1 envelope glycoprotein, chains H, E, J, and B are the heavy chains of VRC01 Fab, while chains L, C, F, and K are the light chains of VRC01 Fab. The core gp120 trimer (chains G, A, and D) consists of outer and inner domain with truncated N- and C-termini and V1/V2 and V3 variable loops [17,37]. 2.1.2. Determination of gp120 Core Residues of Interaction with VRC01 The site of vulnerability on the gp120 site is the highly conserved contact site for the CD4 receptor on the gp120 outer domain. VRC01 contacts amino acid residues of gp120 at positions 275–283, 364–371, 421–437, and 455–478, which represent 98% of the CD4-binding site on the HIV-1 gp120 (Figure 1). Figure 1. HIV-1 gp120 site of vulnerability for VRC01. (A) gp120 trimer (3J70; HXB2) with the site of vulnerability (red), (B) gp120 monomer (chain A) with the site of vulnerability (red), and (C) gp120 amino acid sequence (chain A) with residues of interaction with VRCO1 (red boxes). Molecules 2023, 28, x FOR PEER REVIEW 4 of 30 Figure 1. HIV-1 gp120 site of vulnerability for VRC01. (A) gp120 trimer (3J70; HXB2) with the site of vulnerability (red), (B) gp120 monomer (chain A) with the site of vulnerability (red), and (C) gp120 amino acid sequence (chain A) with residues of interaction with VRCO1 (red boxes). 2.1.3. Molecular Dynamics (MD) Simulation The results obtained from the MD simulation of the receptor indicated that the root-mean-square deviation (RMSD) of the protein backbone steadily increased from 0.075 nm at 0 ns to 0.175 nm at 1 ns. There was a constant fluctuation of the RMSD between 0.15 nm (lowest point) and slightly above 0.225 nm (highest point) within the period of 1 ns and 9 ns (Figure 2). The system stabilized after 9 ns between about 0.175 nm and 0.2 nm till the end of the simulation. The low RMSD after 9 ns is indicative of fewer fluctuations and a more stable protein structure [38]. Molecules 2023, 28, 474 4 of 29 2.1.3. Molecular Dynamics (MD) Simulation The results obtained from the MD simulation of the receptor indicated that the root- mean-square deviation (RMSD) of the protein backbone steadily increased from 0.075 nm at 0 ns to 0.175 nm at 1 ns. There was a constant fluctuation of the RMSD between 0.15 nm (lowest point) and slightly above 0.225 nm (highest point) within the period of 1 ns and 9 ns (Figure 2). The system stabilized after 9 ns between about 0.175 nm and 0.2 nm till the end of the simulation. The low RMSD after 9 ns is indicative of fewer fluctuations and a more stable protein structure [38]. Figure 2. Root-mean-square deviation (RMSD) of the MD simulation of HIV-1 gp120 monomer using GROMACS. A plot of RMSD in nanometers (nm) against time in nanoseconds (ns). The RMSD increased from 0 ns to 1 ns and fluctuated till 9 ns, after which it stabilized till the end of the simulation. 2.1.4. Ligand Generation, Preparation, and Molecular Docking Six purchasable natural product-derived compound libraries were downloaded from three natural compound companies. A total of 27,824 distinct ligands were generated from these compound libraries (Table 1). The ligands were minimized and optimized using the default setting in OpenBabel via PyRx. A total of 27,824 ligands generated from six natural product libraries (Table 1) were used for molecular docking using AutoDock Vina. The AutoDock molecular docking program assumes a rigid receptor and a conformationally flexible ligand. Different conformational poses of the ligands were docked into a rigid receptor (HIV gp120) and scored according to the binding energies of the receptor–ligand complexes. The best ligand poses with the lowest binding energies were selected and visualized in PyMOL. 2.1.5. Protein–Ligand Interaction Profiling LIGPLOT was used to elucidate the amino acid residues in the binding site of the receptor (HIV gp120) that interact with the ligands [39], and 2D schematic diagrams that show the protein–ligand interactions were obtained. Ligand NP-005114 had nine hydrogen bond interactions with eight amino acid residues and hydrophobic interactions with eleven amino acid residues in the CD4-bs of HIV gp120 (Figure 3A,B). Ligand NP-008297 had ten hydrogen bond interactions with amino acid residues and hydrophobic interactions with eleven amino acid residues in the CD4-bs of HIV gp120 (Figures S3A and S4A). Ligand NP- 007382 formed hydrogen bond interactions with five and seven hydrophobic interactions Molecules 2023, 28, x FOR PEER REVIEW 5 of 30 Figure 2. Root-mean-square deviation (RMSD) of the MD simulation of HIV-1 gp120 monomer using GROMACS. A plot of RMSD in nanometers (nm) against time in nanoseconds (ns). The RMSD increased from 0 ns to 1 ns and fluctuated till 9 ns, after which it stabilized till the end of the simulation. 2.1.4. Ligand Generation, Preparation, and Molecular Docking Six purchasable natural product-derived compound libraries were downloaded from three natural compound companies. A total of 27,824 distinct ligands were generated from these compound libraries (Table 1). The ligands were minimized and optimized using the default setting in OpenBabel via PyRx. A total of 27,824 ligands generated from six natural product libraries (Table 1) were used for molecular docking using AutoDock Vina. The AutoDock molecular docking program assumes a rigid receptor and a conformationally flexible ligand. Different conformational poses of the ligands were docked into a rigid receptor (HIV gp120) and scored according to the binding energies of the receptor–ligand complexes. The best ligand poses with the lowest binding energies were selected and vis-ualized in PyMOL. Table 1. Summary of natural product compound libraries used for virtual screening. S/N Company Library Number of Distinct Compounds 1 AnalytiCon Discovery Fragment from nature (FRGx) 247 2 Macrocyclic semi-synthetic compounds (MACROx) 2040 3 Purified natural product screening compounds (MEGx) 3781 4 SPECS Pre-plated 20k diverse compound library 20,636 5 Pre-plated natural product library 373 6 InterBioScreen (IBS) IBS 2017_sep_Natual compound Libraries 747 Total 27,824 2.1.5. Protein–Ligand Interaction Profiling LIGPLOT was used to elucidate the amino acid residues in the binding site of the receptor (HIV gp120) that interact with the ligands [39], and 2D schematic diagrams that Molecules 2023, 28, 474 5 of 29 with amino acid residues in the CD4-bs (Figures S3B and S4B). The binding energies and interacting amino acid residues are summarized in Table 2. Table 1. Summary of natural product compound libraries used for virtual screening. S/N Company Library Number of Distinct Compounds 1 2 3 4 5 6 AnalytiCon Discovery Fragment from nature (FRGx) Macrocyclic semi-synthetic compounds (MACROx) Purified natural product screening compounds (MEGx) SPECS Pre-plated 20k diverse compound library Pre-plated natural product library InterBioScreen (IBS) IBS 2017_sep_Natual compound Libraries Total 247 2040 3781 20,636 373 747 27,824 Figure 3. HIV gp120 of recombinant clade A/E-NP-005114 complex interaction profile. (A) Molecular representation of HIV gp120 of recombinant clade A/E in complex with compound NP-005114 (represented as sticks, firmly docked into CD4-bs (colored hot pink) of HIV gp120). (B) Protein– ligand interaction profile evaluated using LIGPLOT. In (B), the green dotted lines represent hydrogen bond interactions, while red arcs with spikes represent hydrophobic interactions. Residues of the receptor are shown as orange lines with black and red dots, while the ligand is shown as purple lines with black and red dots. Ligand NP-005114 had 9 hydrogen bonds and 11 hydrophobic bonds with residues in the CD4-bs of HIV gp120. Molecules 2023, 28, x FOR PEER REVIEW 6 of 30 show the protein–ligand interactions were obtained. Ligand NP-005114 had nine hydro-gen bond interactions with eight amino acid residues and hydrophobic interactions with eleven amino acid residues in the CD4-bs of HIV gp120 (Figure 3A,B). Ligand NP-008297 had ten hydrogen bond interactions with amino acid residues and hydrophobic interac-tions with eleven amino acid residues in the CD4-bs of HIV gp120 (Figures S3A and S4A). Ligand NP-007382 formed hydrogen bond interactions with five and seven hydrophobic interactions with amino acid residues in the CD4-bs (Figures S3B and S4B). The binding energies and interacting amino acid residues are summarized in Table 2. (A) (B) Figure 3. HIV gp120 of recombinant clade A/E-NP-005114 complex interaction profile. (A) Molecu-lar representation of HIV gp120 of recombinant clade A/E in complex with compound NP-005114 (represented as sticks, firmly docked into CD4-bs (colored hot pink) of HIV gp120). (B) Protein–ligand interaction profile evaluated using LIGPLOT. In (B), the green dotted lines represent hy-drogen bond interactions, while red arcs with spikes represent hydrophobic interactions. Residues of the receptor are shown as orange lines with black and red dots, while the ligand is shown as purple lines with black and red dots. Ligand NP-005114 had 9 hydrogen bonds and 11 hydropho-bic bonds with residues in the CD4-bs of HIV gp120. Table 2. Binding energies and interacting amino acid residues of recombinant clade A/E and clade B HIV-1 gp120. Recombinant Clade A/E Clade B Compound Binding En-ergy (kcal/mol) Amino Acid Residues Binding En-ergy (kcal/mol) Amino Acid Residues NP-008297 −10.3 Glu370, Gln258, Pro470, Ile371, Pro363, Ser365, Gln362, Gly471, Thr455, Gly472, and Asn425. −7.9 Asp368, Ser365, Ser364, Gly366, Asn425, and Met426. NP-000088 −9.7 Glu370, His375, Gln258, Gly472, Gln363, and Ser365 −7.1 Gly473, Thr283, Ser364, and Ser365. NP-007382 −9.6 Pro470, Gly366, Ile371, Gly473, and Ile475. −7.2 Asp457, Arg469, Ser364, Ser365, and Gly472. NP-005003 −9.3 Asp457, Asn280, Pro470, and Arg469. −7.3 Thr283 and Ala281. Molecules 2023, 28, 474 6 of 29 Table 2. Binding energies and interacting amino acid residues of recombinant clade A/E and clade B HIV-1 gp120. Compound Binding Energy (kcal/mol) Amino Acid Residues Binding Energy (kcal/mol) Amino Acid Residues Recombinant Clade A/E Clade B NP-008297 −10.3 NP-000088 NP-007382 NP-005003 NP-007422 NP-001800 NP-005114 NP-004255 FRG-00075 −9.7 −9.6 −9.3 −9.3 −9.2 −9.1 −9 −7.4 Glu370, Gln258, Pro470, Ile371, Pro363, Ser365, Gln362, Gly471, Thr455, Gly472, and Asn425. Glu370, His375, Gln258, Gly472, Gln363, and Ser365 Pro470, Gly366, Ile371, Gly473, and Ile475. Asp457, Asn280, Pro470, and Arg469. Gln258, Ile371, Thr257, Met373, Glu370, Asn425, Asp280, and Asp457. Glu370, Trp427, Gly429, Gln105, and Asn474. Gln258, Ile371, Gln362, Pro363, Gly471, Asp457, Asn280, and Gly472. Trp427, Gly429, Ile371, Gln258, Pro470, and Gly472. Glu370, Ile371, Ser365, Arg469, and Gly472. −7.9 −7.1 −7.2 −7.3 −6.7 −7.8 −8.2 −7.3 −6.8 Asp368, Ser365, Ser364, Gly366, Asn425, and Met426. Gly473, Thr283, Ser364, and Ser365. Asp457, Arg469, Ser364, Ser365, and Gly472. Thr283 and Ala281. Asp474, Arg476, Trp427, Gly473, Asn425, Asp368, and Glu370. Ser365 and Gln363. Arg476, Gly472, Pro470, Gly366, Ser364, Glu370, Asn425, Trp427, and Met426. Arg476, Asp474, Thr455, and Pro470. Gly431 and Ser375. 2.1.6. High-Throughput Virtual Screening (vHTS) and Analysis The top 100 ligands from each of the six libraries (total 600) ranked according to their binding energies were analyzed for binding interactions with HIV-1 gp120. Out of a total of 470 ligands with high binding affinity analyzed from the selected 600 ligands, 386 ligands had hydrogen bond interactions with HIV gp120, while 84 ligands had no hydrogen bond interactions with HIV gp120. Of the 386 ligands with hydrogen bond interactions with gp120, 263 ligands had hydrogen bond interactions with less than four conserved amino acid residues within the site of vulnerability, while 41 had hydrogen bond interactions with four or more amino acid residues in the site of vulnerability on the HIV-1 gp120 (Figure S5). The best nine compounds were selected for in vitro evaluations based on their binding energies and the number of amino acid residues interacting with the HIV-1 gp120 CD4-binding site. The molecular structures of the nine selected compounds are shown in (Table 3). The binding energies of the best nine compounds ranged from −10.3 to −6.4 kcal/mol (Table 2). The lower value of the binding energies corresponds with a higher binding affinity to HIV-1 gp120. Table 3. Chemical structure of the nine selected compounds. Compound ID Chemical Structure NP-000088 Molecules 2023, 28, x FOR PEER REVIEW 7 of 30 NP-007422 −9.3 Gln258, Ile371, Thr257, Met373, Glu370, Asn425, Asp280, and Asp457. −6.7 Asp474, Arg476, Trp427, Gly473, Asn425, Asp368, and Glu370. NP-001800 −9.2 Glu370, Trp427, Gly429, Gln105, and Asn474. −7.8 Ser365 and Gln363. NP-005114 −9.1 Gln258, Ile371, Gln362, Pro363, Gly471, Asp457, Asn280, and Gly472. −8.2 Arg476, Gly472, Pro470, Gly366, Ser364, Glu370, Asn425, Trp427, and Met426. NP-004255 −9 Trp427, Gly429, Ile371, Gln258, Pro470, and Gly472. −7.3 Arg476, Asp474, Thr455, and Pro470. FRG-00075 −7.4 Glu370, Ile371, Ser365, Arg469, and Gly472. −6.8 Gly431 and Ser375. 2.1.6. High-Throughput Virtual Screening (vHTS) and Analysis. The top 100 ligands from each of the six libraries (total 600) ranked according to their binding energies were analyzed for binding interactions with HIV-1 gp120. Out of a total of 470 ligands with high binding affinity analyzed from the selected 600 ligands, 386 lig-ands had hydrogen bond interactions with HIV gp120, while 84 ligands had no hydrogen bond interactions with HIV gp120. Of the 386 ligands with hydrogen bond interactions with gp120, 263 ligands had hydrogen bond interactions with less than four conserved amino acid residues within the site of vulnerability, while 41 had hydrogen bond interac-tions with four or more amino acid residues in the site of vulnerability on the HIV-1 gp120 (Figure S5). The best nine compounds were selected for in vitro evaluations based on their binding energies and the number of amino acid residues interacting with the HIV-1 gp120 CD4-binding site. The molecular structures of the nine selected compounds are shown in (Table 3). The binding energies of the best nine compounds ranged from -10.3 to -6.4 kcal/mol (Table 2). The lower value of the binding energies corresponds with a higher binding affinity to HIV-1 gp120. Table 3. Chemical structure of the nine selected compounds. Compound ID Chemical Structure NP-000088 NP-004255 OOHHOHOOOOHOHHOOOOOOHOHOHHOHOHOHOOHOOOOHOHOHOOOOHOOH Molecules 2023, 28, 474 7 of 29 Table 3. Cont. Compound ID NP-004255 Chemical Structure NP-005003 NP-005114 NP-007382 NP-007422 Molecules 2023, 28, x FOR PEER REVIEW 7 of 30 NP-007422 −9.3 Gln258, Ile371, Thr257, Met373, Glu370, Asn425, Asp280, and Asp457. −6.7 Asp474, Arg476, Trp427, Gly473, Asn425, Asp368, and Glu370. NP-001800 −9.2 Glu370, Trp427, Gly429, Gln105, and Asn474. −7.8 Ser365 and Gln363. NP-005114 −9.1 Gln258, Ile371, Gln362, Pro363, Gly471, Asp457, Asn280, and Gly472. −8.2 Arg476, Gly472, Pro470, Gly366, Ser364, Glu370, Asn425, Trp427, and Met426. NP-004255 −9 Trp427, Gly429, Ile371, Gln258, Pro470, and Gly472. −7.3 Arg476, Asp474, Thr455, and Pro470. FRG-00075 −7.4 Glu370, Ile371, Ser365, Arg469, and Gly472. −6.8 Gly431 and Ser375. 2.1.6. High-Throughput Virtual Screening (vHTS) and Analysis. The top 100 ligands from each of the six libraries (total 600) ranked according to their binding energies were analyzed for binding interactions with HIV-1 gp120. Out of a total of 470 ligands with high binding affinity analyzed from the selected 600 ligands, 386 lig-ands had hydrogen bond interactions with HIV gp120, while 84 ligands had no hydrogen bond interactions with HIV gp120. Of the 386 ligands with hydrogen bond interactions with gp120, 263 ligands had hydrogen bond interactions with less than four conserved amino acid residues within the site of vulnerability, while 41 had hydrogen bond interac-tions with four or more amino acid residues in the site of vulnerability on the HIV-1 gp120 (Figure S5). The best nine compounds were selected for in vitro evaluations based on their binding energies and the number of amino acid residues interacting with the HIV-1 gp120 CD4-binding site. The molecular structures of the nine selected compounds are shown in (Table 3). The binding energies of the best nine compounds ranged from -10.3 to -6.4 kcal/mol (Table 2). The lower value of the binding energies corresponds with a higher binding affinity to HIV-1 gp120. Table 3. Chemical structure of the nine selected compounds. Compound ID Chemical Structure NP-000088 NP-004255 OOHHOHOOOOHOHHOOOOOOHOHOHHOHOHOHOOHOOOOHOHOHOOOOHOOHMolecules 2023, 28, x FOR PEER REVIEW 8 of 30 NP-005003 NP-005114 NP-007382 NP-007422 NHNONHNHOHNONHOHOOHOHNONHOOOHOHOOOOOHOHOHOHOOHOHOHOHOOHOHOOOOOHOHOHOOOOOOOHOHOOHHOHOOOHOHOOHOOHMolecules 2023, 28, x FOR PEER REVIEW 8 of 30 NP-005003 NP-005114 NP-007382 NP-007422 NHNONHNHOHNONHOHOOHOHNONHOOOHOHOOOOOHOHOHOHOOHOHOHOHOOHOHOOOOOHOHOHOOOOOOOHOHOOHHOHOOOHOHOOHOOHMolecules 2023, 28, x FOR PEER REVIEW 8 of 30 NP-005003 NP-005114 NP-007382 NP-007422 NHNONHNHOHNONHOHOOHOHNONHOOOHOHOOOOOHOHOHOHOOHOHOHOHOOHOHOOOOOHOHOHOOOOOOOHOHOOHHOHOOOHOHOOHOOHMolecules 2023, 28, x FOR PEER REVIEW 8 of 30 NP-005003 NP-005114 NP-007382 NP-007422 NHNONHNHOHNONHOHOOHOHNONHOOOHOHOOOOOHOHOHOHOOHOHOHOHOOHOHOOOOOHOHOHOOOOOOOHOHOOHHOHOOOHOHOOHOOH Molecules 2023, 28, 474 8 of 29 Chemical Structure Table 3. Cont. Compound ID NP-008297 NP-001800 FRG-00075 2.1.7. Determination of HIV HXB2 Envelope Protein Structure Human immunodeficiency virus clade B clone 2 (HIV HXB2) served as the reference HIV strain. HXB2 envelope protein was used in the production of HIV-1 PVs for use in cell- based viral infectivity inhibition assays. The cDNA sequence of the HXB2 envelope protein (gp160) was retrieved from the HIV sequence database (Figure S6A) and converted into an amino acid sequence using the ExPASy Translate tool [40]. The HXB2 gp160 amino acid sequence was used to query the ExPASy server platform. The protein sequence (UniProtKB ID: P04578) with the highest similarity score to our query protein sequence was retrieved from UniProtKB (Figure S6B). The 3D structure of the protein (PDB ID: 3J70), with 100% structural similarity to P04578, was retrieved from Protein Data Bank. The structure consists of HIV HXB2 gp160 (chains D, P, U, E, Q, and V) in complex with CD4 (chains C, O, and T) and 17b antibody (chains A, M, R, B, N, and S) (Figure S6C). The HIV HXB2 gp120 trimer (chains D, P, and U) was extracted from 3J70 using PyMOL (Figure S6D). 2.1.8. Virtual Screening and Analysis of HIV HXB2 Envelope Protein The nine selected compounds were used for virtual screening on HIV-1 clade B en- velope protein. The compounds were docked against the HXB2 gp120 monomer and Molecules 2023, 28, x FOR PEER REVIEW 9 of 30 NP-008297 NP-001800 FRG-00075 2.1.7. Determination of HIV HXB2 Envelope Protein Structure Human immunodeficiency virus clade B clone 2 (HIV HXB2) served as the reference HIV strain. HXB2 envelope protein was used in the production of HIV-1 PVs for use in cell-based viral infectivity inhibition assays. The cDNA sequence of the HXB2 envelope protein (gp160) was retrieved from the HIV sequence database (Figure S6A) and con-verted into an amino acid sequence using the ExPASy Translate tool [40]. The HXB2 gp160 amino acid sequence was used to query the ExPASy server platform. The protein sequence (UniProtKB ID: P04578) with the highest similarity score to our query protein sequence was retrieved from UniProtKB (Figure S6B). The 3D structure of the protein (PDB ID: 3J70), with 100% structural similarity to P04578, was retrieved from Protein Data Bank. The structure consists of HIV HXB2 gp160 (chains D, P, U, E, Q, and V) in complex with CD4 (chains C, O, and T) and 17b antibody (chains A, M, R, B, N, and S) (Figure S6C). The HIV HXB2 gp120 trimer (chains D, P, and U) was extracted from 3J70 using PyMOL (Fig-ure S6D). OHOOOHOOHOHOOOHOHOOOOHHOOHOHNOHNONHOOHNNHHNOONOHOHHNOHNHNNOFFFAbsMolecules 2023, 28, x FOR PEER REVIEW 9 of 30 NP-008297 NP-001800 FRG-00075 2.1.7. Determination of HIV HXB2 Envelope Protein Structure Human immunodeficiency virus clade B clone 2 (HIV HXB2) served as the reference HIV strain. HXB2 envelope protein was used in the production of HIV-1 PVs for use in cell-based viral infectivity inhibition assays. The cDNA sequence of the HXB2 envelope protein (gp160) was retrieved from the HIV sequence database (Figure S6A) and con-verted into an amino acid sequence using the ExPASy Translate tool [40]. The HXB2 gp160 amino acid sequence was used to query the ExPASy server platform. The protein sequence (UniProtKB ID: P04578) with the highest similarity score to our query protein sequence was retrieved from UniProtKB (Figure S6B). The 3D structure of the protein (PDB ID: 3J70), with 100% structural similarity to P04578, was retrieved from Protein Data Bank. The structure consists of HIV HXB2 gp160 (chains D, P, U, E, Q, and V) in complex with CD4 (chains C, O, and T) and 17b antibody (chains A, M, R, B, N, and S) (Figure S6C). The HIV HXB2 gp120 trimer (chains D, P, and U) was extracted from 3J70 using PyMOL (Fig-ure S6D). OHOOOHOOHOHOOOHOHOOOOHHOOHOHNOHNONHOOHNNHHNOONOHOHHNOHNHNNOFFFAbsMolecules 2023, 28, x FOR PEER REVIEW 9 of 30 NP-008297 NP-001800 FRG-00075 2.1.7. Determination of HIV HXB2 Envelope Protein Structure Human immunodeficiency virus clade B clone 2 (HIV HXB2) served as the reference HIV strain. HXB2 envelope protein was used in the production of HIV-1 PVs for use in cell-based viral infectivity inhibition assays. The cDNA sequence of the HXB2 envelope protein (gp160) was retrieved from the HIV sequence database (Figure S6A) and con-verted into an amino acid sequence using the ExPASy Translate tool [40]. The HXB2 gp160 amino acid sequence was used to query the ExPASy server platform. The protein sequence (UniProtKB ID: P04578) with the highest similarity score to our query protein sequence was retrieved from UniProtKB (Figure S6B). The 3D structure of the protein (PDB ID: 3J70), with 100% structural similarity to P04578, was retrieved from Protein Data Bank. The structure consists of HIV HXB2 gp160 (chains D, P, U, E, Q, and V) in complex with CD4 (chains C, O, and T) and 17b antibody (chains A, M, R, B, N, and S) (Figure S6C). The HIV HXB2 gp120 trimer (chains D, P, and U) was extracted from 3J70 using PyMOL (Fig-ure S6D). OHOOOHOOHOHOOOHOHOOOOHHOOHOHNOHNONHOOHNNHHNOONOHOHHNOHNHNNOFFFAbs Molecules 2023, 28, 474 9 of 29 interacted with the conserved CD4-bs on HXB2 gp120 (Figure 4A). NP-005114 had hydro- gen bond interactions with eleven amino acids and hydrophobic interactions with four amino acids on the CD4-bs (Figure 4B). There were four and ten hydrogen and hydrophobic interactions, respectively, between NP-008297 (Figure S7A,B) and the amino acids in the gp120 CD4-bs. Compound NP-007382 had five hydrogen bond interactions and twelve hydrophobic interactions with the amino acids of CD4-bs (Figure S7C,D). Figure 4. HIV gp120 of clade B-NP-005114 complex interaction profile. (A) Molecular representation of HIV gp120 of clade B in complex with compound NP-005114 (represented as sticks, firmly docked into CD4-bs of HIV gp120). (B) Protein–ligand interaction profile of NP-005114 and clade B gp120. NP-005114 had eleven hydrogen bond interactions and four hydrophobic interactions with amino acids in the CD4-bs. 2.1.9. RMSD of Clade A/E-Ligand Complexes RMSD was used to assess the binding stability of the complexes. For all the complexes, the C-alpha and backbone RMSDs were the same. For the clade A/E-NP-005114 complex, the RMSD of the protein backbone was observed to increase steadily till about 20 ns, where it stabilized with an average RMSD of 2.5 Å. At ~35 ns, the RMSD rose to 3.0 Å and remained stable with an average of 2.75 Å till the end of the 100 ns period. The ligand (NP-005114) was observed to remain stable throughout the 100 ns simulation period. For about 20 ns, the NP-005114 had an average RMSD of 2.2 Å and increased to an average RMSD of 3.0 Å till the end. For the clade A/E-NP-007382 complex, the RMSD values of the backbone ranged between 2.0 and 3.25 Å. The clade A/E-NP-007382 complex maintained an average RMSD of 3.0 Å throughout the 100 ns simulation period. NP-007382 had the least ligand RMSD, maintaining an average of 0.8 Å throughout the simulation. For the clade A/E-NP-007422 complex, the backbone’s RMSD rose to ~3.25 Å around 20 ns and then fell to 2.75 Å and remained stable till ~55 ns, where the RMSD fell to 2.5 Å, which was maintained till the end of the 100 ns simulation period. The RMSD of the ligand (NP-007422) was also observed to increase and then maintained stability till about 40 ns with an average ligand RMSD of 3 Å and then rose to 4.75 Å, maintaining this average till 100 ns. The clade A/E-NP-008297 complex demonstrated a very stable RMSD throughout the 100 ns simulation period with RMSD (average RMSD of 3 Å) ranging from 1.5 to 3.5 Å (Figure 5). Molecules 2023, 28, x FOR PEER REVIEW 10 of 30 2.1.8. Virtual Screening and Analysis of HIV HXB2 Envelope Protein The nine selected compounds were used for virtual screening on HIV-1 clade B en-velope protein. The compounds were docked against the HXB2 gp120 monomer and in-teracted with the conserved CD4-bs on HXB2 gp120 (Figure 4A). NP-005114 had hydro-gen bond interactions with eleven amino acids and hydrophobic interactions with four amino acids on the CD4-bs (Figure 4B). There were four and ten hydrogen and hydropho-bic interactions, respectively, between NP-008297 (Figure S7A,B) and the amino acids in the gp120 CD4-bs. Compound NP-007382 had five hydrogen bond interactions and twelve hydrophobic interactions with the amino acids of CD4-bs (Figure S7C,D). (A) (B) Figure 4. HIV gp120 of clade B-NP-005114 complex interaction profile. (A) Molecular representa-tion of HIV gp120 of clade B in complex with compound NP-005114 (represented as sticks, firmly docked into CD4-bs of HIV gp120). (B) Protein–ligand interaction profile of NP-005114 and clade B gp120. NP-005114 had eleven hydrogen bond interactions and four hydrophobic interactions with amino acids in the CD4-bs. 2.1.9. RMSD of Clade A/E-Ligand Complexes RMSD was used to assess the binding stability of the complexes. For all the com-plexes, the C-alpha and backbone RMSDs were the same. For the clade A/E-NP-005114 complex, the RMSD of the protein backbone was observed to increase steadily till about 20 ns, where it stabilized with an average RMSD of 2.5 Å. At ~35 ns, the RMSD rose to 3.0 Å and remained stable with an average of 2.75 Å till the end of the 100 ns period. The ligand (NP-005114) was observed to remain stable throughout the 100 ns simulation pe-riod. For about 20 ns, the NP-005114 had an average RMSD of 2.2 Å and increased to an average RMSD of 3.0 Å till the end. For the clade A/E-NP-007382 complex, the RMSD values of the backbone ranged between 2.0 and 3.25 Å. The clade A/E-NP-007382 complex maintained an average RMSD of 3.0 Å throughout the 100 ns simulation period. NP-007382 had the least ligand RMSD, maintaining an average of 0.8 Å throughout the simu-lation. For the clade A/E-NP-007422 complex, the backbone’s RMSD rose to ~3.25 Å around 20 ns and then fell to 2.75 Å and remained stable till ~55 ns, where the RMSD fell to 2.5 Å, which was maintained till the end of the 100 ns simulation period. The RMSD of the ligand (NP-007422) was also observed to increase and then maintained stability till Molecules 2023, 28, 474 10 of 29 Figure 5. RMSD analysis of 100 ns MD simulation trajectory. The RMSD plots for clade A/E- (A) NP-005114, (B) NP-007382, (C) NP-007422, and (D) NP-008297 complexes. 2.1.10. RMSD of HXB2-Ligand Complexes The protein RMSD was observed to stabilize after 2.5 ns with an average of 3.2 Å till about 12.5 ns, where fluctuations were observed. The protein RMSD stabilized again with an average value of 4.2 Å and remained consistent till the end of the 20 ns simulation period. For the HXB2-NP-008297 complex, stability was observed after 1.5 ns with an average RMSD of 5 Å, and this remained constant till about 7.5 ns. A reduction in RMSD was then observed with an average of 4 Å till after 12.5 ns, where fluctuations were observed. The protein achieved stability with an RMSD of 4.2 Å at 17.5ns till the end of the simulation. In the case of the HXB2-NP-007422 complex, the RMSD values ranged between 0.8 and 10.5 Å. The system exhibited fluctuations till about 7 ns and remained stable with an average RMSD of 6.1 Å till after 12.5 ns, where minimal fluctuations were observed. The system regained stability after 15 ns till the end of the simulation period with an average RMSD of 7.5 Å. Significant fluctuations were observed in the HXB2-NP-005114 complex throughout the simulation. However, the HXB2-NP-005114 complex tends to be stabilized at 15 ns till the 20 ns with an average RMSD value of 5.8 Å (Figure S8). 2.1.11. RMSF of Clade A/E-Ligand Complexes To identify the movement of the protein residues in the active site and conformational changes, RMSF was calculated. All four clade A/E-ligand complexes had fluctuations in similar regions. Leu122 to Pro206 and Pro299 to Ile326 were observed to have shown high fluctuations in all four complexes. Clade A/E-NP-007382 and clade A/E-NP-008297 also demonstrated comparatively higher fluctuations around residues Gly397 to Gly412 and Gly459 to Ser464, respectively. The higher RMSF demonstrates that these residues are undergoing conformational changes during the simulation. Notably, however, the overall Molecules 2023, 28, x FOR PEER REVIEW 11 of 30 about 40 ns with an average ligand RMSD of 3 Å and then rose to 4.75 Å, maintaining this average till 100 ns. The clade A/E-NP-008297 complex demonstrated a very stable RMSD throughout the 100 ns simulation period with RMSD (average RMSD of 3 Å) ranging from 1.5 to 3.5 Å (Figure 5). (A) (B) (C) (D) Figure 5. RMSD analysis of 100 ns MD simulation trajectory. The RMSD plots for clade A/E- (A) NP-005114, (B) NP-007382, (C) NP-007422, and (D) NP-008297 complexes. 2.1.10. RMSD of HXB2-Ligand Complexes The protein RMSD was observed to stabilize after 2.5 ns with an average of 3.2 Å till about 12.5 ns, where fluctuations were observed. The protein RMSD stabilized again with an average value of 4.2 Å and remained consistent till the end of the 20 ns simulation period. For the HXB2-NP-008297 complex, stability was observed after 1.5 ns with an av-erage RMSD of 5 Å, and this remained constant till about 7.5 ns. A reduction in RMSD was then observed with an average of 4 Å till after 12.5 ns, where fluctuations were ob-served. The protein achieved stability with an RMSD of 4.2 Å at 17.5ns till the end of the simulation. In the case of the HXB2-NP-007422 complex, the RMSD values ranged be-tween 0.8 and 10.5 Å. The system exhibited fluctuations till about 7 ns and remained stable with an average RMSD of 6.1 Å till after 12.5 ns, where minimal fluctuations were ob-served. The system regained stability after 15 ns till the end of the simulation period with an average RMSD of 7.5 Å. Significant fluctuations were observed in the HXB2-NP-005114 complex throughout the simulation. However, the HXB2-NP-005114 complex tends to be stabilized at 15 ns till the 20 ns with an average RMSD value of 5.8 Å (Figure S8). 2.1.11. RMSF of Clade A/E-Ligand Complexes To identify the movement of the protein residues in the active site and conforma-tional changes, RMSF was calculated. All four clade A/E-ligand complexes had fluctua-tions in similar regions. Leu122 to Pro206 and Pro299 to Ile326 were observed to have Molecules 2023, 28, 474 11 of 29 trend suggests greater flexibility of the residues but a stable binding of the compounds since they present uniform fluctuations in most of the residues (Figure S9). 2.1.12. RMSF of HXB2-Ligand Complexes The RMSF plots revealed that all four compounds elicited some degree of fluctuations in similar regions of the HXB2 protein. Notably, there were fluctuations between residues Lys130 to Ala204 and Thr303 to Asn325 in all the HXB2-ligand complexes, with the latter exhibiting a greater degree of fluctuation. Similarly, the higher RMSF demonstrates that these residues are changing their conformation during the simulation. The overall trend observed here again suggests greater flexibility of the residues but a stable binding of the compounds since most of the interacting residues exhibited uniform fluctuations (Figure S10). 2.1.13. Clade A/E-Ligand Complex Molecular Interaction For the clade A/E-NP-005114 complex, NP-005114 interacted with Thr258, Ser365, Ile371, Met373, Arg469, Pro470, Gly472, Asn474, and Asn478 for more than 40% of the simulation period. It was also observed to interact with residues Thr257, Gln258, Met373, Thr455, Arg456, Pro470, and Gly472, mainly via hydrogen bonds. Ile371 and Arg469 were also involved in hydrophobic contacts with NP-005114. Pro470 was observed to interact with NP-005114 with an interaction fraction more than 1.0 (Figure 6). For more than 40% of the 100 ns simulation period, NP-007382 formed interactions with Gln258, Ser365, Gly366, Ile371, Asn425, Trp427, Arg469, Pro470, Gly472, Gly473, and Asn474. Ile371, Asn425, Arg469, and Gly472 were mostly involved in hydrogen bonding. Ile371 was involved in both hydrogen and hydrophobic interactions. NP-007422 interacted with Gln258, Glu370, Ile371, Met373, Gly472, and Gly473 for more than 40% of the period. It was also observed to interact with Gln258, Gly472, and Gly473 with an interaction fraction greater than 1.0. NP-007422 was observed to interact with the clade A/E by forming hydrogen bonds with Thr257, Gln258, Asn280, Glu370, Ile371, Met373, His375, Tyr384, Asn425, Gln432, Asp457, Arg469, Gly472, and Gly473 during the simulation. For the clade A/E-NP-008297 complex, Gln258, Ala281, Thr283, Ser365, Glu370, Ile371, Thr372, Met373, His375, Tyr384, Asn425, Trp427, Arg469, Pro470, Gly472, and Gly473 formed contacts with NP-008297 for more than 40% of the 100 ns simulation period. Ile371, His375, Trp427, and Arg469 were involved in hydrophobic interactions while Thr257, Gln258, Ala281, Thr283, Ser365, Gly366, Glu370, Met373, Tyr384, Asn425, Arg469, Pro470, Gly472, and Gly373 were mainly involved in hydrogen bonds. The interaction analysis shows on average 60% of the ligands are involved in interactions and each molecule has a few unique interactions. This analysis strongly supports the scope of feature amalgamation as well as rational improvement to design more specific derivatives. 2.1.14. HXB2-Ligand Complex Molecular Interaction NP-005114 was observed to interact with Asn280, Trp427, Thr455, Arg456, and Asp457 mainly via hydrogen bonding during the simulation. NP-007382 formed bonds with Ala281, Lys282, Asn425, and Trp427 for more than 40% of the simulation time. It interacted with Ala281 and Lys282 mainly via hydrogen bonds and interacted mainly with Trp427 via hydrophobic bonds. NP-008297 also formed contacts with Thr283, Ser365, Gly366, Asp368, Ile371, Asn425, Pro470, and Asp477 for more than 40% of the 20 ns simulation period. It interacted with Thr283, Gly366, Asn425, and Asp477 mainly via hydrogen bonding. Moreover, it interacted mainly with Ile371 and Trp427 via hydrophobic contacts and Asp368 via an ionic bond. Asp368 had interaction fraction values of approximately 1.75, signifying that Asp368 interacted with HXB2 for more than 100% of the simulation period due to multiple simultaneous formations of interactions with NP-008297 [41]. NP- 007422 was observed to form bonds with Ser365, Thr455, Arg456, Asp457, Gly458, and Asp474 for more than 40% of the 20 ns simulation period. NP-007422 interacted mainly with Asn280, Thr455, Arg456, Gly458, and Asp474 via hydrogen bonding. The formation of multiple hydrogen bonds between HXB2 and NP-007422 could influence its activity [42] Molecules 2023, 28, 474 12 of 29 (Figure S11). The interactions show clade independent interactions, and differences seen are due to differences in the conformations. Figure 6. Analysis of the molecular interactions and the type of contacts with clade A/E throughout MD simulation. Normalized stacked bar chart of clade A/E residues interacting with (A) NP-005114, (B) NP-007382, (C) NP-007422, and (D) NP-008297. Hydrogen bonds, hydrophobic bond, ionic interactions, and water bridges are represented as green, grey, red, and blue, respectively. 2.1.15. Origin and Sources of Selected Compounds Of the nine selected natural product-derived compounds, six were obtained from plants, three from micro-organisms, and one was a synthetic fragment of the natural compound (Table 4). Table 4. The origin and sources of the test compounds. Purity data are as provided by the manufacturer. Compound Structure Organism Name Purity NP-000088 NP-004255 NP-005003 NP-005114 NP-007382 NP-007422 NP-008297 NP-001800 FRG-00075 Plant Plant Micro-organism Plant Micro-organism Plant Plant Micro-organism Fragment Mentha piperita Terminalia chebula Aspergillus Terminalia chebula Actinomycete Withania somnifera Ginkgo biloba Fungi N/A 79 92 97 99 98 96 94 79 70 2.1.16. Pharmacological Profiling The pharmacological and physicochemical properties of the nine selected compounds were obtained from SwissADME (Tables 5 and 6). The compounds have molecular weights between 335.72 and 895.09 Dalton. Rotatable bonds are predictors of ligand conformational flexibility and small compound bioavailability [43], with rotatable bonds less than or equal to seven being ideal in drug discovery; eight of the compounds fell within this range. Pharmacological indices such as lipophilicity, solubility, gastrointestinal (GI) absorption, brain–blood barrier (BBB) permeant, P-glycoprotein (P-gp) substrate, and Cytochrome Molecules 2023, 28, x FOR PEER REVIEW 13 of 30 (A) (B) (C) (D) Figure 6. Analysis of the molecular interactions and the type of contacts with clade A/E through-out MD simulation. Normalized stacked bar chart of clade A/E residues interacting with (A) NP-005114, (B) NP-007382, (C) NP-007422, and (D) NP-008297. Hydrogen bonds, hydrophobic bond, ionic interactions, and water bridges are represented as green, grey, red, and blue, respectively. 2.1.14. HXB2-Ligand Complex Molecular Interaction NP-005114 was observed to interact with Asn280, Trp427, Thr455, Arg456, and Asp457 mainly via hydrogen bonding during the simulation. NP-007382 formed bonds with Ala281, Lys282, Asn425, and Trp427 for more than 40% of the simulation time. It interacted with Ala281 and Lys282 mainly via hydrogen bonds and interacted mainly with Trp427 via hydrophobic bonds. NP-008297 also formed contacts with Thr283, Ser365, Gly366, Asp368, Ile371, Asn425, Pro470, and Asp477 for more than 40% of the 20 ns sim-ulation period. It interacted with Thr283, Gly366, Asn425, and Asp477 mainly via hydro-gen bonding. Moreover, it interacted mainly with Ile371 and Trp427 via hydrophobic con-tacts and Asp368 via an ionic bond. Asp368 had interaction fraction values of approxi-mately 1.75, signifying that Asp368 interacted with HXB2 for more than 100% of the sim-ulation period due to multiple simultaneous formations of interactions with NP-008297 [41]. NP-007422 was observed to form bonds with Ser365, Thr455, Arg456, Asp457, Gly458, and Asp474 for more than 40% of the 20 ns simulation period. NP-007422 inter-acted mainly with Asn280, Thr455, Arg456, Gly458, and Asp474 via hydrogen bonding. The formation of multiple hydrogen bonds between HXB2 and NP-007422 could influence its activity [42] (Figure S11). The interactions show clade independent interactions, and differences seen are due to differences in the conformations. 2.1.15. Origin and Sources of Selected Compounds Of the nine selected natural product-derived compounds, six were obtained from plants, three from micro-organisms, and one was a synthetic fragment of the natural com-pound (Table 4). Molecules 2023, 28, 474 13 of 29 450 inhibitor were used to predict the absorption, distribution, metabolism, and excretion (ADME) profiles of the compounds which describe their pharmacokinetic properties. The lipophilicity value for the selected compounds was between −0.98 and 1.91. The ideal lipophilicity value of compounds lies between 1 and 3 [44], with only four of the selected compounds falling within this range. Eight of the compounds were predicted to be soluble in water (Table 6). The summation of all the polar atoms of a molecule is defined by the topological polar surface area (TPSA) [45]. The TPSA value is a metric used to define cell permeability by a drug. The ideal TPSA usually falls within 140 to 90 ´Å, above 140 ´Åindicates poor permeability into a cell [46], and below 90 indicates penetration into the blood–brain barrier (BBB) [47]. None of the compounds fell within this range. Furthermore, none of the test compounds were BBB permeants and inhibitors of the CYP 450 enzyme (Table 6). However, nine of the compounds had low GI absorption and were substrates of the P-gp transporter (Table 6). Table 5. Physicochemical properties of the 9 selected compounds. Compound Name Molecular Weight (Dalton) Number of Rotatable Bonds Number of H-bond Acceptors Number of H-bond Donors NP-008297 NP-004255 NP-000088 NP-007422 NP-005114 NP-007382 NP-005003 NP-001800 FRG-00075 740.66 634.45 610.56 782.91 636.46 596.57 723.78 800.90 335.72 9 2 7 9 5 5 5 7 2 16 15 15 14 15 12 9 9 4 9 11 8 9 11 4 9 9 2 Table 6. Adsorption, distribution, metabolism, and excretion properties of the 9 selected compounds. Compound Name NP-008297 NP-004255 NP-000088 NP-007422 NP-005114 NP-007382 NP-005003 NP-001800 FRG-00075 TPSA * Lipophilicity Water Solubility GI Absorption ¥ BBB Permeant £ Pgp Substrate § CYP Inhibitor 275.5 310.66 234.29 245.29 310.66 186.12 251.16 251.16 71.18 1.79 −0.87 −0.43 0.07 −3.12 1.91 −0.98 0.61 1.3 Soluble Soluble Soluble Soluble Soluble Soluble Soluble Insoluble Soluble Low Low Low Low Low Low Low Low High No No No No No No No No No Yes Yes Yes Yes No Yes Yes Yes Yes No No No No No No No No No * Topological polar surface area. § P-glycoprotein substrate. ¥ gut absorption. £ cross the blood–brain barrier. CYP450 conversion. 2.1.17. Toxicity Profiling of the Selected Compounds The result of the toxicity profile of the selected compounds predicted using ProTox-II is summarized in Table S2 [48]. ProTox-II generates the toxicity profiles of small compound molecules by measuring parameters such as immunotoxicity, hepatotoxicity, cytotoxic- ity, mutagenicity, carcinogenicity, and toxicity class [48]. Toxicity class ranges from 1–6 (1 = highly toxic/fatal, 6 = least toxic) with probability range of 0–1 (0 = not likely, and 1 = very likely). Eight of the compounds belonged to the less toxic drug class associated with high LD50, except for NP-007422, which belongs to toxicity class 2 (Table S2). Moreover, eight of the compounds were inactive for carcinogenicity and cytotoxicity, whilst seven were predicted as immunotoxins. None of the compounds was hepatotoxic or mutagenic. Molecules 2023, 28, x FOR PEER REVIEW 15 of 30 Table 6. Adsorption, distribution, metabolism, and excretion properties of the 9 selected com-pounds. Compound Name TPSA * Lipophilicity Water Solu-bility GI Absorption ¥ BBB Perme-ant £ Pgp Sub-strate § CYP Inhibitor ȡ NP-008297 275.5 1.79 Soluble Low No Yes No NP-004255 310.66 -0.87 Soluble Low No Yes No NP-000088 234.29 -0.43 Soluble Low No Yes No NP-007422 245.29 0.07 Soluble Low No Yes No NP-005114 310.66 -3.12 Soluble Low No No No NP-007382 186.12 1.91 Soluble Low No Yes No NP-005003 251.16 -0.98 Soluble Low No Yes No NP-001800 251.16 0.61 Insoluble Low No Yes No FRG-00075 71.18 1.3 Soluble High No Yes No * Topological polar surface area. § P-glycoprotein substrate. ¥ gut absorption. £ cross the blood–brain barrier. ȡ CYP450 conversion. 2.1.17. Toxicity Profiling of the Selected Compounds The result of the toxicity profile of the selected compounds predicted using ProTox-II is summarized in Table S2 [48]. ProTox-II generates the toxicity profiles of small com-pound molecules by measuring parameters such as immunotoxicity, hepatotoxicity, cyto-toxicity, mutagenicity, carcinogenicity, and toxicity class [48]. Toxicity class ranges from 1-6 (1 = highly toxic/fatal, 6 = least toxic) with probability range of 0 - 1 (0 = not likely, and 1 = very likely). Eight of the compounds belonged to the less toxic drug class associated with high LD50, except for NP-007422, which belongs to toxicity class 2 (Table S2). Moreo-ver, eight of the compounds were inactive for carcinogenicity and cytotoxicity, whilst seven were predicted as immunotoxins. None of the compounds was hepatotoxic or mu-tagenic. 2.2. Cell-Based Viral Infectivity Inhibition Assay 2.2.1. In Vitro Cell Cytotoxicity Assessment To evaluate the cytotoxic effect of the test compounds on target TZM-bl cells, cell viability was determined using the Alamar blue assay. Compounds FGR-0075, NP-00603, NP-00088, NP-001800, NP-005114, NP-008297, and NP-007422 had no cytotoxic effect on TZM-bl cells at the concentration tested (Figure S12). Compounds NP-007382 and NP-004255 exhibited cytotoxic effects. The CC50 was calculated in Figure S13. Ursolic acid, a known cytotoxic compound, was used as a positive control. 2.2.2. Determination of 50% Cytotoxicity Concentration (CC50) To determine the 50% cytotoxic concentration (CC50) of compounds that had a cyto-toxic effect on the TZM-bl cells, a non-linear dose–response regression analysis was used to calculate CC50. CC50 was defined as the concentration of a compound that is required to reduce viable cells by 50%. Compound NP-007382 had CC50 of 47.6 µM (28.4 µg/mL), while NP-004255 had CC50 of 14.4 µM (9.1 µg/mL), respectively (Figure S13). 2.2.3. Pseudotype Virus (PV) Titration and Determination of 50% Tissue Culture Infec-tivity Dose (TCID50) PV titration was carried out to determine the number of viable PVs produced. A vi-able PV was defined as a PV that can infect a target cell and integrate the luciferase re-porter gene, leading to the expression of the reporter gene. The quantity of light produced by the luciferase enzyme (relative light unit RLU) is measured with a luminometer and directly correlates with the number of viable PVs in a harvest. TCID50 was defined as the number of PVs that can infect 50% of target cells (TZM-bl cells). A positive infection was Molecules 2023, 28, x FOR PEER REVIEW 15 of 30 Table 6. Adsorption, distribution, metabolism, and excretion properties of the 9 selected com-pounds. Compound Name TPSA * Lipophilicity Water Solu-bility GI Absorption ¥ BBB Perme-ant £ Pgp Sub-strate § CYP Inhibitor ȡ NP-008297 275.5 1.79 Soluble Low No Yes No NP-004255 310.66 -0.87 Soluble Low No Yes No NP-000088 234.29 -0.43 Soluble Low No Yes No NP-007422 245.29 0.07 Soluble Low No Yes No NP-005114 310.66 -3.12 Soluble Low No No No NP-007382 186.12 1.91 Soluble Low No Yes No NP-005003 251.16 -0.98 Soluble Low No Yes No NP-001800 251.16 0.61 Insoluble Low No Yes No FRG-00075 71.18 1.3 Soluble High No Yes No * Topological polar surface area. § P-glycoprotein substrate. ¥ gut absorption. £ cross the blood–brain barrier. ȡ CYP450 conversion. 2.1.17. Toxicity Profiling of the Selected Compounds The result of the toxicity profile of the selected compounds predicted using ProTox-II is summarized in Table S2 [48]. ProTox-II generates the toxicity profiles of small com-pound molecules by measuring parameters such as immunotoxicity, hepatotoxicity, cyto-toxicity, mutagenicity, carcinogenicity, and toxicity class [48]. Toxicity class ranges from 1-6 (1 = highly toxic/fatal, 6 = least toxic) with probability range of 0 - 1 (0 = not likely, and 1 = very likely). Eight of the compounds belonged to the less toxic drug class associated with high LD50, except for NP-007422, which belongs to toxicity class 2 (Table S2). Moreo-ver, eight of the compounds were inactive for carcinogenicity and cytotoxicity, whilst seven were predicted as immunotoxins. None of the compounds was hepatotoxic or mu-tagenic. 2.2. Cell-Based Viral Infectivity Inhibition Assay 2.2.1. In Vitro Cell Cytotoxicity Assessment To evaluate the cytotoxic effect of the test compounds on target TZM-bl cells, cell viability was determined using the Alamar blue assay. Compounds FGR-0075, NP-00603, NP-00088, NP-001800, NP-005114, NP-008297, and NP-007422 had no cytotoxic effect on TZM-bl cells at the concentration tested (Figure S12). Compounds NP-007382 and NP-004255 exhibited cytotoxic effects. The CC50 was calculated in Figure S13. Ursolic acid, a known cytotoxic compound, was used as a positive control. 2.2.2. Determination of 50% Cytotoxicity Concentration (CC50) To determine the 50% cytotoxic concentration (CC50) of compounds that had a cyto-toxic effect on the TZM-bl cells, a non-linear dose–response regression analysis was used to calculate CC50. CC50 was defined as the concentration of a compound that is required to reduce viable cells by 50%. Compound NP-007382 had CC50 of 47.6 µM (28.4 µg/mL), while NP-004255 had CC50 of 14.4 µM (9.1 µg/mL), respectively (Figure S13). 2.2.3. Pseudotype Virus (PV) Titration and Determination of 50% Tissue Culture Infec-tivity Dose (TCID50) PV titration was carried out to determine the number of viable PVs produced. A vi-able PV was defined as a PV that can infect a target cell and integrate the luciferase re-porter gene, leading to the expression of the reporter gene. The quantity of light produced by the luciferase enzyme (relative light unit RLU) is measured with a luminometer and directly correlates with the number of viable PVs in a harvest. TCID50 was defined as the number of PVs that can infect 50% of target cells (TZM-bl cells). A positive infection was Molecules 2023, 28, 474 14 of 29 2.2. Cell-Based Viral Infectivity Inhibition Assay 2.2.1. In Vitro Cell Cytotoxicity Assessment To evaluate the cytotoxic effect of the test compounds on target TZM-bl cells, cell viability was determined using the Alamar blue assay. Compounds FGR-0075, NP-00603, NP-00088, NP-001800, NP-005114, NP-008297, and NP-007422 had no cytotoxic effect on TZM-bl cells at the concentration tested (Figure S12). Compounds NP-007382 and NP- 004255 exhibited cytotoxic effects. The CC50 was calculated in Figure S13. Ursolic acid, a known cytotoxic compound, was used as a positive control. 2.2.2. Determination of 50% Cytotoxicity Concentration (CC50) To determine the 50% cytotoxic concentration (CC50) of compounds that had a cyto- toxic effect on the TZM-bl cells, a non-linear dose–response regression analysis was used to calculate CC50. CC50 was defined as the concentration of a compound that is required to reduce viable cells by 50%. Compound NP-007382 had CC50 of 47.6 µM (28.4 µg/mL), while NP-004255 had CC50 of 14.4 µM (9.1 µg/mL), respectively (Figure S13). 2.2.3. Pseudotype Virus (PV) Titration and Determination of 50% Tissue Culture Infectivity Dose (TCID50) PV titration was carried out to determine the number of viable PVs produced. A viable PV was defined as a PV that can infect a target cell and integrate the luciferase reporter gene, leading to the expression of the reporter gene. The quantity of light produced by the luciferase enzyme (relative light unit RLU) is measured with a luminometer and directly correlates with the number of viable PVs in a harvest. TCID50 was defined as the number of PVs that can infect 50% of target cells (TZM-bl cells). A positive infection was defined as one with 2.5 times the average RLU of the “cells only” control wells. TCID50 was calculated in the TCID50 MACRO sheet, using the modified Spearman and Karber methods [49], and was determined to be 5657 TCID50/mL. The cut-off for the luminescence was set at 10,000 RLU, and the recommended volume of PVs that will produce 10,000 RLU was calculated to be 105 µL. 2.2.4. Viral Infectivity Inhibition Assay The concentration of a compound where the RLU (response) is reduced by 50% relative to the “cells + PVs” control was defined as the 50% inhibition concentration (IC50) of that compound. Only four (NP-005114, NP-008297, NP-007422, and NP-007382) out of the nine tested compounds exhibited a dose-dependent inhibition of viral infection of TZM-bl cells. NP-005114 had IC50 of 15.2 µM (9.7 µg/mL, R2 = 0.9331, [95%CI = 7.2–12.9]), NP-007382 had IC50 of 21.6 µM (12.9 µg/mL, R2 = 0.9216, [95%CI= 10.2–15.9]), NP-008297 had IC50 of 10.1 µM (7.5 µg/mL, R2 = 0.9557, [95%CI = 6.0–9.2]), and NP-007422 had IC50 of 16.2 µM (12.7 µg/mL, R2 = 0.8925 [95%CI = 8.7–18.3]) (Figure 7A–D). Compound NP-004255, however, exhibited a dose-independent inhibition of viral entry (Figure S14). There was no observed viral inhibition by compounds FRG-00075, NP-005003, NP-000088, and NP-001800 (Figure S14). 2.2.5. Selectivity Index The selectivity index (SI) parameter is used to access a compound’s in vitro efficacy in the inhibition assay by evaluating the cytotoxic effect versus the antiviral effect. SI was defined as the ratio of the 50% cytotoxic concentration (CC50) to the 50% inhibitory concentration (IC50) and was calculated for compound NP-007382 with CC50 of 47.6 µM and IC50 of 21.6 µM. The SI was determined as 2.2. Molecules 2023, 28, 474 15 of 29 Figure 7. Cell-based viral infectivity inhibition assay. Nonlinear regression plot of percentage inhibition (response) against the Log10 of compound concentration (dose) at a 95% confidence interval. (A) NP-005114 had IC50 of 15.2 µM (9.7 µg/mL), (B) NP-007382 had IC50 of 21.6 µM (12.9 µg/mL), (C) NP-008297 had IC50 of 10.1 µM (7.5 µg/mL), and (D) NP-007422 had IC50 of 16.2 µM (12.7 µg/mL). Error bars represent the standard error of the mean (SEM) of triplicate wells. 3. Discussion This study used a pharmacoinformatics approach to identify compounds derived from natural products (plants and micro-organisms) that mimic VRC01 bNAbs by interacting with similar binding sites as VRC01 (CD4-binding site) on the HIV gp120. The CD4-binding site (CD4-bs) is a conserved region on the HIV-1 gp120 that is crucial for the CD4-gp120 interaction required for successful HIV-1 infection. Out of 27, 824 compounds analyzed, nine compounds (NP-008297, NP-004255, NP-000088, NP-007422, NP-005114, NP-007382, NP-005003, NP-001800, and FRG-00075) that interacted with the CD4-bs of gp120 in silico were confirmed in vitro using pseudotype virus technology. While the compounds’ purity range was around 70%, natural compounds are secondary metabolites that are formed by a variety of cascades. They have similar properties with different compounds upstream of the synthetic pathway as well as other metabolic branches of similar compounds de- rived from similar precursors. This makes isolating a single compound very difficult. The Molecules 2023, 28, x FOR PEER REVIEW 16 of 30 defined as one with 2.5 times the average RLU of the “cells only” control wells. TCID50 was calculated in the TCID50 MACRO sheet, using the modified Spearman and Karber methods [49], and was determined to be 5657 TCID50/mL. The cut-off for the luminescence was set at 10,000 RLU, and the recommended volume of PVs that will produce 10,000 RLU was calculated to be 105 µL. 2.2.4. Viral Infectivity Inhibition Assay The concentration of a compound where the RLU (response) is reduced by 50% rela-tive to the “cells + PVs” control was defined as the 50% inhibition concentration (IC50) of that compound. Only four (NP-005114, NP-008297, NP-007422, and NP-007382) out of the nine tested compounds exhibited a dose-dependent inhibition of viral infection of TZM-bl cells. NP-005114 had IC50 of 15.2 µM (9.7 µg/mL, R2 = 0.9331, [95%CI = 7.2-12.9]), NP-007382 had IC50 of 21.6 µM (12.9 µg/mL, R2 = 0.9216, [95%CI= 10.2-15.9]), NP-008297 had IC50 of 10.1µM (7.5µg/mL, R2 = 0.9557, [95%CI = 6.0-9.2]), and NP-007422 had IC50 of 16.2 µM (12.7 µg/mL, R2 = 0.8925 [95%CI = 8.7-18.3]) (Figure 7 A–D). Compound NP-004255, however, exhibited a dose-independent inhibition of viral entry (Figure S14). There was no observed viral inhibition by compounds FRG-00075, NP-005003, NP-000088, and NP-001800 (Figure S14). Molecules 2023, 28, 474 16 of 29 best example is curcumin; while there is a known structure for curcumin, the prepara- tions, even pharmaceutical ones, are a mixture of many curcuminoids (77% curcumin, 17% desmethoxycurcumin, and 3% bisdemethoxycurcumin) with slight variations in structure but highly similar chemical properties, making pure curcumin isolation a high-loss pro- cess [50]. Similarly, we expect these compound mixtures to have highly similar compounds, and many of them might be sharing the pharmacophore; thus, what we are seeing could be a cumulative effect. In silico analysis of the test compounds revealed that the test compounds had high binding affinities to the critical amino acid residues of CD4-bs of clade A/E recombinant and clade B HIV-1 gp120. The ligands were found to be firmly docked in the binding domain of CD4-bs of both recombinant clades A/E and clade B (HXB2) HIV-1 gp120 (Figure 3A, 4A and S3). The binding energies of the screened ligands ranged from −8.2 to −4.8 kcal/mol for the gp120 of clade B and −10.3 to −7.4 kcal/mol for recombinant clade A/E gp120 (Table 2). Analysis of the molecular interaction between the test compounds and HIV-1 gp120 showed that the compounds exhibited intermolecular hydrogen and hydrophobic bonding with key amino acid residues of the CD4-bs gp120 of both clade B (Figure 4B and S7) and recombinant clade A/E (Figure Figure 3B and S4). HIV-1 clade A/E and clade B gp120 were used as the protein receptor for the virtual screening to compare the protein–ligand interaction of our test compounds to the gp120-VRC01 [17] and CD4-gp120 [51] interactions, respectively. Protein–ligand interaction analysis showed that NP-005114, NP-008297, NP-007382, and NP007422 partially mimic CD4 and VRC01 in their interaction with the CD4-binding sites of HIV-1 gp120. The stability of the protein–ligand complexes was assessed by subjecting the complexes to molecular dynamics simulations. For all the complexes, the RMSDs of the protein backbone, side chains, and heavy atoms demonstrated similar trends. The RMSD trajectory plots of the clade A/E complexes indicated stability in all protein–ligand complexes by the end of the 100 ns simulation. The RMSD plots also revealed that the HXB2-NP-007382, HXB2-NP-008297, and HXB2- NP-007422 complexes demonstrated more stability than the HXB2-NP-005114 complex. Overall, these results highlight the binding stability of the complexes and underscore the mechanistic basis for the inhibition of HIV-1 entry. A previous study showed that the conformational equilibrium for the HXB2 backbone is reached after 5 ns, similar to the results obtained herein [52]. The lead compounds had intermolecular hydrogen and hydrophobic interactions with the critical amino acid residues of the CD4-bs, which are important for the binding of VRC01 and CD4 to gp120. Compounds NP-005114, NP-008297, NP-007382, and NP007422 had biomolecular hydrogen and hydrophobic interactions with Asp368 gp120, Glu370 gp120, Ile371 gp120, Asn425 gp120, Met426 gp120, Trp427 gp120, Gly473 gp120, Met475 gp120, and Asp457 gp120, which are critical and conserved residues in the VRC01-gp120 interaction [17] and Phe43 cavity. The Phe43 cavity is a functionally crucial and conserved pocket on HIV gp120, where CD4 binds [51,53,54] and alterations in the amino acid residues lead to a decrease in CD4 binding [55]. The molecular interaction profiles of the protein– ligand complexes, which were assessed using MD simulation to determine the nature of the bonds formed between the protein and the ligand in a dynamic environment, are consistent with what were observed during molecular docking. This is indicative of the predicted stability of the intermolecular hydrogen and hydrophobic bonds formed between the protein and ligand, which partially mimic the CD4–gp120 and VRC01-gp120 interactions. Many small compounds that have been developed to target the CD4-bs of HIV-1 gp120 [56] interact with the Phe43 cavity in CD4-bs and have demonstrated broad activity against various strains of HIV-1 entry into target cells [57]. The interactions of these compounds with gp120 of clades A/E and B indicate possible conservation of the CD4-bs [58], despite constant variation of the HIV envelope protein due to mutations caused by error-prone HIV reverse transcriptase. Although it is also indicative of the possibility of interactions with gp120 from several strains of HIV, more work needs to be carried out to determine the interaction of the lead compounds with other HIV clades. Molecules 2023, 28, 474 17 of 29 Small compounds with the desired bioactivity are assessed for drug-likeness (Lipin- ski’s rule of five) using their physicochemical properties. Lipinski’s rule allows not more than 5 and 10 hydrogen bond donors and acceptors, respectively, LogP ≤ 5, and molecular mass (MW) of ≤500 Daltons [59]. Compounds NP-005114, NP-00892, and NP-007422 violated three of Lipinski’s rules (MW>500, H-bond donors, and acceptors not more than 5 and 10, respectively), while compound NP-007382 violated two rules (MW > 500, H-bond acceptors >10) (Table 5). The implication of violating Lipinski’s rules is that the compounds are predicted to have poor pharmacokinetic properties (ADME) and oral bioavailability. However, Lipinski’s rule does not predict pharmacological activity [59]. The absorption, distribution, metabolism, and excretion (ADME) properties of the compounds describe their pharmacokinetics. Parameters including lipophilicity, solubility, gastrointestinal (GI) absorption, brain–blood barrier (BBB) permeant, P-glycoprotein (P- gp) substrate, and Cytochrome 450 inhibitor are also used to predict pharmacokinetic properties (Table 6). Lipophilicity is an indication of the solubility, permeability, selectivity, and promiscuity of a drug compound. Lipophilicity > 5 has been associated with poor receptor selectivity, high metabolic turnover, low solubility, and poor absorption. However, compounds with low lipophilicity values are also linked to poor ADME properties [44]. The ideal lipophilicity value lies between 1 and 3; only NP-008297 and NP-005114 fell within this range. All lead compounds have low GI absorption, possibly due to the complex interplay of many factors such as molecular weight, cell permeability, and solubility of the compounds; however, these properties can be optimized in the developmental stage of drug discovery [60,61]. TPSA value is a metric used to define cell permeability by a drug compound. The ideal TPSA usually falls within 140 to 90 ´Å, above 140 ´Åindicates poor permeability into a cell [44], and below 90 indicates penetration into the blood–brain barrier (BBB) [47]. None of the compounds fell within this range. The blood–brain barrier (BBB) permeability index indicated the ability of the drug compound to be deposited into the central nervous system (CNS) [62]; none of the compounds was BBB permeant (Table 6). P-glycoprotein (P-gp) is a transmembrane efflux pump that is important in the efflux and uptake of the drug compound [63]. NP-008297 and NP-007382 were substrates of the P-gp transporter. Cytochrome P450 (CYP450) enzymes play a significant role in drug metabolism, excretion, and drug–drug interactions [64]. Compounds that inhibit CYP enzymes interfere with the metabolism of other drugs that are activated by the enzyme. None of the compounds were inhibitors of the CYP 450 enzymes. There is scope for rational improvement of the leads to further derive an inhibitor with a better bioavailability profile. The toxicity profile of the selected compounds predicted using ProTox-II is sum- marized in Table S2 [48]. ProTox-II generates the toxicity profiles of small compound molecules by measuring parameters such as immunotoxicity, hepatotoxicity, cytotoxicity, mutagenicity, carcinogenicity, and toxicity class [48]. Toxicity class ranges from 1 to 6 (1 = highly toxic/fatal, 6 = least toxic) with probability range of 0–1 (0 = not likely, 1 = very likely). All the compounds except NP-007422 and NP-005003 belonged to the less toxic drug class associated with LD50 ≥ 1000 (Table S2). None of the compounds were shown to exhibit hepatotoxicity and mutagenicity. FRG-00075 was predicted as carcinogenic whilst NP-005114 and FRG-00075 were predicted to be immunotoxins. NP-007422 was also predicted to be cytotoxic and highly lethal. Further research is required to optimize the structures of the potential leads using the pharmacophores to circumvent these toxicity predictions. Worthy of note is the disparity between the in silico predicted toxicity profile and the observed in vitro cytotoxicity. Despite compounds NP-004255 and NP-007382 being predicted as non-cytotoxic in HeLa cell lines, they demonstrated some cytotoxicity in vitro. This could be because ProTox-II predicts by using molecular and pharmacophore similarity of lead compounds to known compounds in its database, hence a given limitation if the lead compound does not have high similarity to known toxic compounds in the database [65]. Further transcription profiling may reveal the mechanism/off-target behind observed cytotoxicity. Molecules 2023, 28, 474 18 of 29 Although the in silico predictions are beneficial in identifying possible bioactive compounds, there is a need for further confirmation using experimental methods. Inhibition of viral entry by the test compounds was examined using molecularly cloned HXB2 PV. Only four compounds, NP-005114, NP-007382, NP-008297, and NP-007422, exhibited a dose-dependent inhibition of clade B (HXB2) env PV entry into TZM-bl cells, with IC50 values of 15.2 µM (9.7 µg/mL), 21.6 µM (12.9 µg/mL), 10.1 µM (7.5 µg/mL), and 16.2 µM (12.7 µg/mL), respectively. Compound NP-004255, however, exhibited a dose-independent inhibition of viral entry (Figure S14). The antiviral compound NP-004255 could not be differentiated from the cytotoxic effect on the target cells. There was no observed viral inhibition by compounds FRG-00075, NP-005003, NP-000088, and NP-001800 (Figure S14). From previous literature [20,66–68], the IC50 values of VRC01 in the neutralizing panel of clade B HIV and clade B env PV ranged from 0.1 to 50 µg/mL. The IC50 of NP-005114, NP-008297, and NP-007382 obtained from this study is well within this range. Several small compound HIV-1 entry inhibitors, targeting the CD4-bs of the HIV-1 gp120 [56], have been discovered recently. However, only BMS-663068 and NBD-556, as well as their derivatives, have shown good antiretroviral activity and are in the preclinical and clinical trials of the drug development process [69–71]. BMS-663068 is a more potent derivative of BMS-626529, with inhibition to viral entry against a panel of PVs at IC50 of between 0.0001 and 9.5 µg/mL with varying sensitivity towards different clades of HIV-1 [72]. NBD-556 and its derivatives were shown to inhibit entry of HIV-1HXB2 PVs at IC50 of 20.0 µM [73]. Further testing is needed with complete viruses and different clades, especially newly emerged clade C [74]. Compound NP-005114 was extracted from the seed of Terminalia chebula (Table 4). T. chebula is a medicinal plant popularly referred to as the “king of medicine” due to a wide array of bioactivity of compounds extracted from the plant. Compounds extracted from T. chebula fruit are HIV-1 integrase and reverse transcriptase inhibitors [75,76]. So far, no compound extracted from T. chebula has been characterized as an HIV-1 entry inhibitor [77,78]. NP-005114 has exhibited an HIV-1 entry inhibitory property. NP-008297 was extracted from the leaf of Ginkgo biloba. G. biloba crude extracts have shown activity against HIV-1 reverse transcriptase RNase H and protease [77,79]. This study has shown the HIV-1 entry inhibition activity of NP-008297. Compound NP-007422 was derived from Withania somnifera, popularly known as Ashwagandha. Ashwagandha herb is popular for its anxiolytic and adaptogenic effects [80] and has been found to reverse neuron cells toxicity induced by β-Amyloid in HIV-associated neurocognitive disorders (HAND) [81]. Unlike the others, NP-007382 was extracted from actinomycetes, which are micro-organisms that have been a source of active compounds against HIV-1 [77]. 4. Materials and Methods 4.1. In Silico High-Throughput Screening of Natural Compound Mimetics of Vrc01 4.1.1. Protein Structure Extraction The 3D structure of crystallized antigen-binding fragment (Fab) of the broadly neu- tralizing antibody, VRC01, in complex with core gp120 of HIV-1 clade A/E recombinant 93TH057 (PDB ID: 3NGB), with a resolution of 2.68 Å was downloaded from the Protein Data Bank (PDB). The core gp120 trimer consisting of outer and inner domains with trun- cated N- and C-termini, as well as V1/V2 and V3 variable loops, were extracted from the complex as protein chains A, G, and D using PyMOL 1.74 (Schrödinger, Inc., NY, USA) [82]. 4.1.2. Determination of gp120 Core Residues of Interaction with VRC01 VRC01 is a CD4-binding site (CD4bs) broadly neutralizing antibody (bNAb) which makes contact with 98% of gp120 sites of vulnerability [17]. The gp120 vulnerability site is the highly conserved CD4 receptor contact site on gp120 outer domain. The gp120 amino acid residues that interact with VRC01 and CD4 were obtained elsewhere [17]. Molecules 2023, 28, 474 19 of 29 4.1.3. Receptor Molecular Dynamics (MD) Simulation MD simulation of the HIV-1 gp120 core (PDB ID: 3NGB) was carried out to determine the conformation, stability, and dynamics of the structure. A 10 ns MD simulation was performed using GROMACS 5.1.4 [83]. The CHARMM27 force field was adopted to prepare the topology input file for the protein. The protein was solvated by SPCE water molecules and immersed in a 1 nm thick cubic periodic water box. Before the simulation, a short minimization of 500 steps using the steepest descent method was carried out to remove possible distortion in the protein structure caused by the addition of water to the system. Eight Cl- ions were added to neutralize the system. The system was equilibrated at a temperature of 300 K and normal pressure for 50 ps to restrain the heavy atoms of the proteins to their starting position to allow water molecules to saturate the system. The final production simulation was performed for 10 ns under similar conditions as the equilibration steps. The root-mean-square displacement (RMSD) of the minimized protein heavy-atom concerning the resolved X-ray structure was calculated and plotted using GRACE 5.1.4 [84]. The final minimized GROMACS protein file was visualized with Visual Molecular Dynamics (VMD) 1.9.3 version [85] and saved as frames in PDB format for further analysis. 4.1.4. Receptor Preparation The minimized HIV-1 gp120 was prepared for molecular docking using AutoDock- Tools version 4.2.6 (Scripps Research, La Jolla, CA, USA). Water molecules were removed from the structure, polar hydrogen atoms were added, and non-polar hydrogens were merged with a parent carbon atom. Gasteiger partial charges of the atoms were calculated and added. The protein file was then converted to AutoDock compatible format (pdbqt). The energy grid box was set around the conserved residues of HIV-1 gp120 with a dimen- sion of 42 ´Å × 50 ´Å × 56 ´Å, and coordinates of X = 52.318 ´Å, Y = 26.081 ´Å, and Z = 46.493 ´Å for the receptor macromolecules. 4.1.5. Ligand Compound Library Generation Libraries of natural compounds were obtained from three commercially available natural product-derived compound databases consisting of AnalytiCon Discovery, Specs, and InterBioScreen (IBS). Structures of compounds retrieved from these databases were downloaded as a single structure-data file (SDF) format. 4.1.6. Ligands Preparation SDF format files of the ligand libraries were split into individual compounds using a split utility module in Open Babel 2.3.1 [86] accessible via the PyRx 0.8 [87]. The compounds were optimized and converted to AutoDock compatible (PDBQT) format. A total of 27,824 ligands were minimized for molecular docking. 4.1.7. High-Throughput Virtual Screening High-throughput virtual screening of the compound libraries was performed on a Linux Operating System High-Performance Computing System (Zuputo) hosted by the West African Centre for Cell Biology of Infectious Pathogens (WACCBIP). AutoDock Vina 1.1.1 (Scripps Research, La Jolla, CA, USA), a molecular docking tool, was used for the virtual screening of compound libraries. AutoDock Vina utilizes rigid receptor and flexible ligand docking mode and empirical scoring function to rank protein–ligand complexes [88]. The previously prepared HIV-1 gp120 protein’s PDBQT file with defined grid box dimensions (42 ´Å × 50 ´Å × 56 ´Å, coordinates of: x = 52.318 ´Å, y = 26.081 ´Å, z = 46.493 ´Å) and exhaustiveness set value 9 (amount of computational effort used during a docking experiment; default value 8), along with all optimized ligands, was used for docking. A custom bash script was prepared for the docking procedures. All virtual Molecules 2023, 28, 474 20 of 29 screening results were aggregated and written to a single log file. The log file contained binding energies for each of the docked complexes. 4.1.8. Virtual Screening Result Analysis AutoDock Vina scores the binding affinities of the ligand to the receptor using empiri- cal data obtained from the summation of energies contributed by the receptor and ligand interactions, which is calculated as the total atom pair distance-dependent interactions of the protein and ligand [89]. The binding energy values are inversely related to the binding affinity, i.e., the lower the binding energy value, the higher the binding affinity. The single log file containing the binding energies in (kcal/mol) of all processed ligands was extracted into a comma-separated value (CSV) file format and was analyzed using Microsoft Excel 2016 (Microsoft Corporation, One Microsoft Way, Redmond, WA, USA). The docked poses of each ligand were visualized using PyMOL. The poses with the best fit in the binding site with the lowest binding energies were selected for further analysis. 4.1.9. Protein–Ligand Interaction Profiling The protein–ligand complex interaction profiles were determined using LIGPLOT version 1.4.5 [39]. LIGPLOT analyzes the molecular interaction between proteins and ligands and generates 2D schematic representations of the interactions in terms of bond types. The bond types include hydrogen and hydrophobic interactions, as well as bond lengths and interacting residues. Hydrogen bond interactions are represented by broken green lines of distinct length, while arcs with spikes pointing towards the ligands represent the hydrophobic interactions. 4.1.10. HIV-1 HXB2 Envelope Protein Human immunodeficiency virus clade B clone 2 (HIV HXB2) was used as the reference HIV strain. HXB2 envelope protein was used in the production of HIV-1 PVs utilized in the cell-based viral infectivity inhibition assay. Virus-like particles are not full-length genomes and lack nonstructural proteins required for viral replication and jumping elements for recombination and thus are categorized in RG1 of biosafety handling. 4.1.11. Determination of HIV-1 HXB2 Envelope Protein Structure The nucleic acid sequence of HXB2 envelope protein (gp160) that was cloned into the pCAGGGS plasmid was retrieved from the HIV sequence database. The nucleic acid sequence was converted to an amino acid sequence using the ExPASy Translate tool, which translates nucleic acid sequences into an amino acid sequence. The HXB2 gp160 amino acid sequence was retrieved from the ExPASy Translate tool server as a FASTA file and used as a query for BLASTP via the ExPASy platform. BLASTP is used to query the protein sequence database to identify similar sequences. The protein sequence with the highest similarity score to our query protein sequence was retrieved from UniProt Knowledgebase (UniProtKB). UniProtKB is a comprehensive protein database that contains annotations and functional information of the proteins. A structure similarity search was performed using the UniProt ID of the protein sequence to retrieve the 3D protein structure (PDB ID: 3J70) from Protein Data Bank. 4.1.12. Protein Structure Extraction and Energy Minimization of HXB2 gp120 The 3D structure of HXB2 gp160 with the entire variable regions in complex with CD4 and antibody 17b was retrieved from Protein Data Bank (PDB ID: 3J70) [90]. The gp120 trimer was extracted from the protein complex using PyMOL 1.74 [82] as protein chains D, P, and U. The energy of the extracted gp120 protein was then minimized using Swiss-Pdb Viewer 4.2 [91] with default settings. Swiss-PDB Viewer [91] utilizes the Gronin- gen Molecular Simulation computer program package (GROMOS 43B1) [83] force field to minimize the energy of the protein and repair distorted geometries of the protein structure. The minimized protein was extracted and saved as a PDB file. Molecules 2023, 28, 474 21 of 29 4.1.13. Receptor Preparation In preparation for molecular docking, the minimized HIV-1 HXB2 gp120 was opti- mized using AutoDockTools version 4.2.6 (Scripps Research, La Jolla, CA, USA). Water molecules were removed from the structure to prevent interference with the docking pro- cess. Polar hydrogens were added, and non-polar hydrogens were merged with a parent carbon atom. Gasteiger partial charges of the atoms were calculated and added. The protein file was then converted to AutoDock compatible format (pdbqt). The energy grid box was set around the conserved residues of HIV-1 gp120 at dimensions (54 ´Å × 40 ´Å × 44 ´Å) with coordinates of X = 249.640 ´Å, Y = 230.870 ´Å, and Z = 198.249 ´Å. After optimization, the protein was used as a receptor for molecular docking. The resultant complex was analyzed, and the protein–ligand interaction was profiled using LIGPLOT [39]. 4.1.14. The Root-Mean-Square Deviation (RMSD) of the Complexes The stability of the protein–ligand complexes was assessed using molecular dynam- ics simulations performed with the Desmond module in Schrödinger Suite (Schrödinger Release 2021-2, Schrödinger, LLC, New York, NY, USA). A 20 ns MD simulation was per- formed for the HXB2-ligand complexes since previous study has shown that the backbone reaches conformational equilibrium after 5 ns [52]. However, for clade A/E-ligand com- plexes, extended 100 ns MD simulations were performed. The MD simulations were aimed at understanding the stability, dynamic fluctuations, and molecular interactions in the protein–ligand complexes. The docked conformers of NP-005114, NP-007382, NP-007422, and NP-008297 with each of the receptors were used for the MD simulations. 4.1.15. The Root-Mean-Square Fluctuation (RMSF) of the Complexes The RMSF trajectories of the complexes were also investigated. The RMSF plot pro- vides information on the flexibility of the various regions of a protein, which is related to the crystallographic B-factors [92]. Protein regions with significantly high fluctuations represent the areas involved in ligand binding and catalysis [93]. 4.1.16. Molecular Interactions under Dynamic Simulation Molecular interactions and the type of bonds were investigated to understand the contacts between the proteins and stability of the complex in presence of water (TIP3p) and ions similar to physiological condition (0.15 M NaCl) and each ligand throughout the simu- lation period. Various types of interactions including hydrogen bonds, hydrophobic bonds, ionic bonds, and water bridges were formed during simulation, which are represented in the stacked bar charts as green, grey, red, and blue, respectively. The stacked bar charts represent the bonds over the simulation trajectory. All the ligands kept the water from entering the binding site, indicating stable binding. Few water molecules participated in the complex, indicating more stability in physiological conditions. Moreover, the complexes transitioned to lower energy states without ligands flying off, validating the interaction and establishing ligand strain is far less than interaction energies. 4.1.17. Physicochemical, Pharmacokinetics, and Drug-Likeness Properties Prediction SwissADME (Swiss Institute of Bioinformatics, Amphipôle, Quartier UNIL-Sorge, 1015 Lausanne, Switzerland) was used to predict the relevant physicochemical, pharmacokinet- ics, and drug-likeness properties of the shortlisted ligands [94]. The Simplified Molecular Input Line Entry Specification (SMILES) formats of the query ligand files were used to generate parameters such as physicochemical properties, lipophilicity (trans-membrane movement), absorption, distribution, metabolism, and excretion (ADME) profiles. The others include gastrointestinal absorption, blood–brain barrier permeability, P-glycoprotein substrate, cytochrome enzyme inhibition, water solubility, and drug-likeness (Lipinski’s rule of 5). Molecules 2023, 28, 474 22 of 29 4.1.18. Toxicity Profile Prediction The toxicity profiles of selected compounds were predicted using ProTox-II [48]. ProTox-II utilizes compound pharmacophore fingerprint, structural similarity, and machine learning models designed from both in vitro and in vivo assay data to predict immunotoxi- city, hepatotoxicity, cytotoxicity, mutagenicity, and carcinogenicity. The toxicity profiles of the compounds have a probability range of 0–1 (0 = not likely, 1 = very likely), predicted lethal dose (LD50), and toxicity class which ranges from 1 to 6 (1 = highly toxic/fatal, 6 = least toxic). 4.2. Cell-Based Viral Infectivity Inhibition Assay 4.2.1. Reagents Complete cell culture medium was prepared using Dulbecco’s Modified Eagle Medium (DMEM) supplemented with L-glutamine, streptomycin, penicillin, and heat-inactivated fetal bovine serum (FBS), all from Thermo Fisher (Thermo Fisher, Waltham, MA, USA). 10X Trypsin-EDTA (0.5%), Phosphate Buffered Saline (PBS) and Dimethyl sulfoxide (DMSO) were purchased from Sigma Aldrich. 4.2.2. Plasmids Plasmids p8.91, pCSFLW, and pCAGGS-HXB2 were donated by Dr. Edward Wright (University of Sussex, Brighton BN1 9QG, UK). The p8.91 plasmid described in PNAS by Naldidni et al., 1996 [95], expresses HIV gag-pol, tat, and rev. pCSFLW is a lentiviral vector expressing firefly luciferase, and pCAGGS-HXB2-env is a mammalian expression vector expressing HIV-1 envelope protein (gp160) clade B clone 2. 4.2.3. Cell Lines TZM-bl cells are HeLa cells with HIV-1-tat-directed luciferase reporter gene, engi- neered to express CD4+ receptors and co-receptors. HEK 293T/17 is an adherent human kidney cell line that is highly transfectable and capable of producing retroviruses in high titer. Both TZM-bl and HEK 293T/17 cell lines were donated by Dr. Edward Wright (University of Sussex, Brighton BN1 9QG, UK). 4.2.4. Cell Culture TZM-bl and HEK 293T/17 cells were cultured in Dulbecco’s Modified Eagle’s Medium (DMEM) containing 10% fetal calf serum + 50 unit/mL penicillin + 50 µg/mL streptomycin in a 10 cm culture dish containing 10 mL of media and incubated at 37 ◦C in 5% CO2. The cell monolayer was washed twice with PBS and treated with 1.5 ml of 1X trypsin- ethylenediaminetetraacetic acid (EDTA) to dissociate adherent cells. The re-suspended cells were seeded into 10 mL of new culture plates for transfection assay. 4.2.5. Production of HIV Pseudotype Viruses (PV) Clade B (HXB2) HIV-1 PV was used to test the ability of the compounds to inhibit HIV-1 entry into target cells. HIV PVs were produced by transfecting HEK293T-17 cells, using the protocol described by Longo and colleagues [96]. Briefly, plasmid DNA mix (pCAGGSHXB2, p8.91, and pCSFLW) and transfection reagent polyethyleneimine (PEI) mix from Polysciences, Inc. (Polysciences Inc., Warrington, PA, USA), were prepared in separate Eppendorf tubes containing serum-free media (OptiMEM). The contents of the two tubes were mixed and incubated for about 20 min and afterward added to HEK293T-17 cells. The transfection plate was incubated overnight at 37 ◦C (5% CO2). The media were changed after 14 h, with culture supernatant containing HIV-1 PVs harvested at 48- and 72-h post-transfection. Molecules 2023, 28, 474 23 of 29 4.2.6. Determination of Tissue Culture Infectivity Dose (TCID50) of Harvested Pseudotype Virus (PV) In a 96-well plate, 25 µL of harvested PVs was serially diluted (1/5 dilution) in quadruple across the plate (in complete media) and incubated at 37 ◦C (5% CO2) for 30 min. Approximately 2 × 104 (100 µL) of TZM-bl cells was added, and the plate was incubated for 48 h at 37 ◦C (5% CO2). The control wells contain only TZM-bl cells. After 48 h, spent media were discarded, and 50 µL of a 50:50 mix of media and Bright Glo substrate was added to the well and incubated for 3 min. The luminescence of each well was quantified using the Promega GloMax®Explorer Multimode Microplate Reader (Promega Corporation, Madison, WI, USA) at a 0.3 shake rate. TCID50 was calculated in the TCID50 MACRO sheet using the modified Spearman and Karber method [49]. TCID50 was defined as the number of PVs that can infect 50% of target cells (TZM-bl cells). The cut-off for the luminescence was set at 10,000 RLU. 4.2.7. Preparation of Test Compounds The nine test compounds were purchased from the natural compounds screening li- braries of AnalytiCon Discovery (https://ac-discovery.com/screening-libraries/ (accessed 18 February 2019), with purity ranging from 70 to 90 (Table 4). The compounds were prepared and diluted to 1 mg/mL in 10% DMSO. The final concentration of DMSO in the downstream analysis was 0.1% (v/v). The maximum concentrations of test compounds used for downstream assay have been summarized in Table S1. 4.2.8. Evaluation of Compounds Cytotoxicity Alamar blue assay was conducted to evaluate the effect of the test compounds on TZM-bl cells. A volume of 100 µL (approximately 2 × 104 cells) of TZM-bl cells was seeded overnight in a 96-well plate, treated with serially diluted test compounds, and incubated for 48 h at 37 ◦C and 5% CO2. Positive controls were cells treated with ursolic acid, a known cytotoxic compound, negative controls were cells with no compound treatment, with compounds and media only serving as color control, and media only as blank. Absorbance was read at 570 nm after 48 h. The assay was set up in triplicates, and mean values were calculated for analysis. Cell viability was calculated relative to the absorbance of the negative control wells (TZM-bl cells only), which were set as 100% of absorbance, using the equation below Cell viability (%) = (cid:34) Abssample – Absblank AbsNeg_control – Absblank (cid:35) × 100% (1) 4.2.9. Viral Infectivity Inhibition Assay The ability of test compounds to inhibit HIV entry into target TZM-bl cells was tested in a viral infectivity inhibition assay previously described by [97]. Briefly, test compounds were serially diluted and incubated with HIV-1 PVs for 1 h at 37 ◦C, after which TZM-bl cells were added and incubated for 48 h at 37 ◦C. The controls included TZM-bl cells only (assay control), HIV PVs only (negative control), and HIV PVs with TZM-bl cells (positive control). The relative light unit (RLU) from the assay was read using the Promega GloMax®Explorer Multimode Microplate Reader (Promega Corporation, Madison, WI, USA). The IC50 was calculated using non-linear regression dose–response analysis with GraphPad Prism version 8 as described by [97], with slight modification. The modification was by normalization of the Y-axis (response) to 0% = PVs only control and 100% = cells + PVs only. A cells-only control was set up as background RLU. 5. Conclusions To our knowledge, this is the first study on natural product-derived compounds that mimic HIV-1 broadly neutralizing antibody VRC01 by interacting with the CD4-bs of HIV-1 gp120 of recombinant clades A/E and B. The study combined varieties of in silico Molecules 2023, 28, 474 24 of 29 techniques including molecular docking and in vitro screening. We have identified four compounds with the potential to inhibit HIV-1 clade B (HXB2) entry into target TZM-bl cells. The study further identified novel insights into the binding mechanisms including potential residues critical for interactions using molecular dynamics simulations. The molecules were shown to have reasonably good pharmacological profiles and were predicted as drug-like. The molecules could serve as templates for the design of next-generation HIV-1 inhibitors of therapeutic importance. The study warrants the experimental evaluation of these compounds with known entry inhibitors. Supplementary Materials: The following supporting information can be downloaded at: https:// www.mdpi.com/article/10.3390/molecules28020474/s1, Figure S1: A typical 96-well plate layout for Alamar blue cell viability assay; Figure S2: Viral infectivity inhibition assay plate layout; Figure S3: Compounds A) NP-008297 and B) NP-007382 docked into CD4-bs (hot pink shaded) of recombinant clade A/E HIV gp120 (blue molecular surface); Figure S4: Protein–ligand interaction profiles of recombinant clade A/E HIV gp120- A) NP-008297 and B) NP-007382 complexes; Figure S5: Summary of virtual high-throughput screening. The best 600 ligands were selected from the ligand library based on binding energies, and 470 ligands of the selected ligands were analyzed. A total of 304 ligands had hydrogen bond interactions with the CD4 binding site (CD4-bs) amino acid residues. Forty-one had hydrogen bond interactions with 4 or more amino acid residues in the CD4-bs. The best 10 compounds were selected from the 41 for in vitro analysis; Figure S6: Determination of the 3D structure of HIV HXB2. A) The amino acid sequence of the HXB2 envelope protein retrieved from the HIV database, B) BlastP similarity search result of HXB2 env amino acid sequence (sequence alignment), C) 3D structure of 3J70, which consists of HXB2 gp160 in complex with CD4 and 17b antibody, and D) HXB2 gp120 extracted from 3J70 using PyMOL; Figure S7: Molecular interactions between the CD4-bs of clade B HIV gp120 (blue molecular surface) and compounds NP-008297 (A and B) and NP-007382 (C and D); Figure S8: RMSD analysis of MD simulation trajectory. The plots for A) HXB2-NP-005114, B) HXB2-NP-007382, C) HXB2-NP-007422, and D) HXB2-NP-008297 complexes; Figure S9: RMSF analysis of MD simulation trajectory. The plots for A) clade A/E-NP-005114, B) clade A/E-NP-007382, C) clade A/E-NP-007422, and D) clade A/E-NP-008297 complexes. For the RMSF plots, the residue index 1 maps to Val44; Figure S10: RMSF analysis of MD simulation trajectory. The plots for A) HXB2-NP- 005114, B) HXB2-NP-007382, C) HXB2-NP-007422, and D) HXB2-NP-008297 complexes; Figure S11: Analysis of the molecular interactions and the type of contacts with HXB2 throughout MD simulation. Normalized stacked bar chart of HXB2 residues interacting with A) NP-005114, B) NP-007382, C) NP-007422, and D) NP-008297. Hydrogen bond, hydrophobic bond, ionic interactions, and water bridges are represented as green, grey, red, and blue, respectively; Figure S12: Alamar blue cell viability assay shows the tested compounds had no cytotoxic effect on the TZM-bl cells. The positive control, ursolic acid, is a known cytotoxic compound. Error bars represent the standard error of the mean (SEM) of triplicate wells; Figure S13: Determination of 50% cytotoxicity concentration (CC50.). Alamar blue cell viability assay shows NP-004255 and NP-007382 had a cytotoxic effect on the TZM-bl cells (A and B). The positive control, ursolic acid, is a known cytotoxic compound. CC50 was determined using dose–response non-linear regression analysis. (C) NP-004255 with CC50 of 14.4 µM (9.1 µg/mL) and (D) NP-007382 with CC50 of 47.6 µM (28.4 µg/mL). Error bars represent the standard error of the mean (SEM) of triplicate wells; Figure S14: Viral infectivity inhibition assay for A) FRG-00075, B) NP-000088, C) NP-001800, D) NP-004255, and E) NP-005003. A dose-independent inhibitory activity was observed for NP-004255. No inhibition was observed for compounds FRG-00075, NP-0005003, NP-000088, and NP-001800. Error bars represent standard error of the mean (SEM); Table S1: Maximum concentration of test compound used for cell viability and viral inhibition; Table S2: Toxicity profiles of the 9 selected compounds predicted using ProTox-II. Author Contributions: N.U.-K., O.Q., E.W. and S.K.K. conceptualized the project. N.U.-K., S.K.K., and O.A. performed the cheminformatics analysis. N.U.-K. and S.L. performed the biological assays. E.B., S.K.K., P.K. and Y.G. contributed to the molecular dynamics simulations. N.U.-K. wrote the first draft of the manuscript. All authors have read and agreed to the published version of the manuscript. Funding: This work was supported by funds from a World Bank African Centres of Excellence grant (ACE02-WACCBIP: Awandare). Institutional Review Board Statement: Not applicable. Molecules 2023, 28, 474 25 of 29 Informed Consent Statement: Not applicable. Data Availability Statement: Not applicable. Acknowledgments: The authors are grateful to Edwards Wright from the University of Sussex for providing the plasmids and cell lines used for the biological assays. We would like to thank the University of Ghana’s West African Centre for Cell Biology of Infectious Pathogens for providing us with free high-performance computing time. Conflicts of Interest: The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results. Sample Availability: Samples of the compounds are available from the authors. References 1. Wilen, C.B.; Tilton, J.C.; Doms, R.W. HIV: Cell binding and entry. Cold Spring Harb. Perspect. Med. 2012, 2, a006866. [CrossRef] 2. Henrich, T.J.; Kuritzkes, D.R. HIV-1 entry inhibitors: Recent development and clinical use. Curr. Opin. Virol. 2013, 3, 51–57. [CrossRef] 3. Haqqani, A.A.; Tilton, J.C. Entry inhibitors and their use in the treatment of HIV-1 infection. Antivir. Res. 2013, 98, 158–170. 4. 5. 6. [CrossRef] Gulick, R.M.; Lalezari, J.; Goodrich, J.; Clumeck, N.; DeJesus, E.; Horban, A.; Nadler, J.; Clotet, B.; Karlsson, A.; Wohlfeiler, M.; et al. Maraviroc for Previously Treated Patients with R5 HIV-1 Infection. N. Engl. J. Med. 2008, 359, 1429–1441. [CrossRef] Fätkenheuer, G.; Nelson, M.; Lazzarin, A.; Konourina, I.; Hoepelman, A.I.M.; Lampiris, H.; Hirschel, B.; Tebas, P.; Raffi, F.; Trottier, B.; et al. Subgroup analyses of maraviroc in previously treated R5 HIV-1 infection. N. Engl. J. Med. 2008, 359, 1442–1455. [CrossRef] Lazzarin, A.; Clotet, B.; Cooper, D.; Reynes, J.; Arastéh, K.; Nelson, M.; Katlama, C.; Stellbrink, H.-J.; Delfraissy, J.-F.; Lange, J.; et al. Efficacy of Enfuvirtide in Patients Infected with Drug-Resistant HIV-1 in Europe and Australia. N. Engl. J. Med. 2003, 348, 2186–2195. [CrossRef] 7. Wibmer, C.K.; Moore, P.L.; Morris, L. HIV broadly neutralizing antibody targets. Curr. Opin. HIV AIDS 2015, 10, 135–143. [CrossRef] 8. Wu, X.; Yang, Z.-Y.; Li, Y.; Hogerkorp, C.-M.; Schief, W.R.; Seaman, M.S.; Zhou, T.; Schmidt, S.D.; Wu, L.; Xu, L.; et al. Rational Design of Envelope Identifies Broadly Neutralizing Human Monoclonal Antibodies to HIV-1. Science 2010, 329, 856–861. [CrossRef] [PubMed] Binley, J.M.; Wrin, T.; Korber, B.; Zwick, M.B.; Wang, M.; Chappey, C.; Stiegler, G.; Kunert, R.; Zolla-Pazner, S.; Katinger, H.; et al. Comprehensive cross-clade neutralization analysis of a panel of anti-human immunodeficiency virus type 1 monoclonal antibodies. J. Virol. 2004, 78, 13232–13252. [CrossRef] 9. 10. Hessell, A.J.; Rakasz, E.G.; Poignard, P.; Hangartner, L.; Landucci, G.; Forthal, D.N.; Koff, W.C.; Watkins, D.I.; Burton, D.R. Broadly neutralizing human anti-HIV antibody 2G12 is effective in protection against mucosal SHIV challenge even at low serum neutralizing titers. PLoS Pathog. 2009, 5, e1000433. [CrossRef] 11. Caskey, M.; Klein, F.; Lorenzi, J.C.C.; Seaman, M.S.; West, A.P.; Buckley, N.; Kremer, G.; Nogueira, L.; Braunschweig, M.; Scheid, J.F.; et al. Corrigendum: Viraemia suppressed in HIV-1-infected humans by broadly neutralizing antibody 3BNC117. Nature 2016, 535, 580. [CrossRef] 12. Horwitz, J.A.; Halper-Stromberg, A.; Mouquet, H.; Gitlin, A.D.; Tretiakova, A.; Eisenreich, T.R.; Malbec, M.; Gravemann, S.; Billerbeck, E.; Dorner, M.; et al. HIV-1 suppression and durable control by combining single broadly neutralizing antibodies and antiretroviral drugs in humanized mice. Proc. Natl. Acad. Sci. USA 2013, 110, 16538–16543. [CrossRef] 13. Klein, F.; Halper-Stromberg, A.; Horwitz, J.A.; Gruell, H.; Scheid, J.F.; Bournazos, S.; Mouquet, H.; Spatz, L.A.; Diskin, R.; Abadir, A.; et al. HIV therapy by a combination of broadly neutralizing antibodies in humanized mice. Nature 2012, 492, 118–122. [CrossRef] 14. Halper-Stromberg, A.; Lu, C.L.; Klein, F.; Horwitz, J.A.; Bournazos, S.; Nogueira, L.; Eisenreich, T.R.; Liu, C.; Gazumyan, A.; Schaefer, U.; et al. Broadly neutralizing antibodies and viral inducers decrease rebound from HIV-1 latent reservoirs in humanized mice. Cell 2014, 158, 989–999. [CrossRef] Scheid, J.F.; Mouquet, H.; Ueberheide, B.; Diskin, R.; Klein, F.; Oliveira, T.Y.K.; Pietzsch, J.; Fenyo, D.; Abadir, A.; Velinzon, K.; et al. Sequence and Structural Convergence of Broad and Potent HIV Antibodies That Mimic CD4 Binding. Science 2011, 333, 1633–1637. [CrossRef] 15. 16. Yu, W.H.; Su, D.; Torabi, J.; Fennessey, C.M.; Shiakolas, A.; Lynch, R.; Chun, T.W.; Doria-Rose, N.; Alter, G.; Seaman, M.S.; et al. Predicting the broadly neutralizing antibody susceptibility of the HIV reservoir. JCI Insight 2019, 4, e130153. [CrossRef] 17. Zhou, T.; Georgiev, I.; Wu, X.; Yang, Z.Y.; Dai, K.; Finzi, A.; Do Kwon, Y.; Scheid, J.F.; Shi, W.; Xu, L.; et al. Structural basis for broad and potent neutralization of HIV-1 by antibody VRC01. Science 2010, 329, 811–817. [CrossRef] Molecules 2023, 28, 474 26 of 29 18. Bar, K.J.; Sneller, M.C.; Harrison, L.J.; Justement, J.S.; Overton, E.T.; Petrone, M.E.; Salantes, D.B.; Seamon, C.A.; Scheinfeld, B.; Kwan, R.W.; et al. Effect of HIV Antibody VRC01 on Viral Rebound after Treatment Interruption. N. Engl. J. Med. 2016, 375, 2037–2050. [CrossRef] 19. Gautam, R.; Nishimura, Y.; Pegu, A.; Nason, M.C.; Klein, F.; Gazumyan, A.; Golijanin, J.; Buckler-White, A.; Sadjadpour, R.; Wang, K.; et al. A single injection of anti-HIV-1 antibodies protects against repeated SHIV challenges. Nature 2016, 533, 105–109. [CrossRef] 20. Mayer, K.H.; Seaton, K.E.; Huang, Y.; Grunenberg, N.; Isaacs, A.; Allen, M.; Ledgerwood, J.E.; Frank, I.; Sobieszczyk, M.E.; Baden, L.R.; et al. Safety, pharmacokinetics, and immunological activities of multiple intravenous or subcutaneous doses of an anti-HIV monoclonal antibody, VRC01, administered to HIV-uninfected adults: Results of a phase 1 randomized trial. PLoS Med. 2017, 14, e1002435. [CrossRef] 21. Cale, E.M.; Bai, H.; Bose, M.; Messina, M.A.; Colby, D.J.; Sanders-Buell, E.; Dearlove, B.; Li, Y.; Engeman, E.; Silas, D.; et al. Neutralizing antibody VRC01 failed to select for HIV-1 mutations upon viral rebound. J. Clin. Investig. 2020, 130, 3299–3304. [CrossRef] [PubMed] 22. Awi, N.J.; Teow, S.-Y. Antibody-Mediated Therapy against HIV/AIDS: Where Are We Standing Now? J. Pathog. 2018, 2018, 8724549. [CrossRef] [PubMed] 23. Arantes, I.; Ribeiro-Alves, M.; de Azevedo, S.S.D.; Delatorre, E.; Bello, G. Few amino acid signatures distinguish HIV-1 subtype B pandemic and non-pandemic strains. PLoS ONE 2020, 15, e0238995. [CrossRef] [PubMed] 24. Newman, D.J.; Cragg, G.M. Natural Products as Sources of New Drugs from 1981 to 2014. J. Nat. Prod. 2016, 79, 629–661. [CrossRef] [PubMed] 25. Barnes, E.C.; Kumar, R.; Davis, R.A. The use of isolated natural products as scaffolds for the generation of chemically diverse screening libraries for drug discovery. Nat. Prod. Rep. 2016, 33, 372–381. [CrossRef] [PubMed] 26. Lawson, A.D.G.; Maccoss, M.; Heer, J.P. Importance of Rigidity in Designing Small Molecule Drugs to Tackle Protein-Protein Interactions (PPIs) through Stabilization of Desired Conformers. J. Med. Chem. 2018, 61, 4383–4389. [CrossRef] [PubMed] 27. Cary, D.C.; Peterlin, B.M. Natural Products and HIV/AIDS. AIDS Res. Hum. Retroviruses 2018, 34, 31–38. [CrossRef] 28. Chinsembu, K.C. Chemical diversity and activity profiles of HIV-1 reverse transcriptase inhibitors from plants. Rev. Bras. 29. Farmacogn. 2019, 29, 504–528. [CrossRef] Fang, J.; Liu, C.; Wang, Q.; Lin, P.; Cheng, F. In silico polypharmacology of natural products. Brief. Bioinform. 2017, 19, 1153–1171. [CrossRef] 30. Yi, F.; Li, L.; Xu, L.J.; Meng, H.; Dong, Y.M.; Liu, H.B.; Xiao, P.G. In silico approach in reveal traditional medicine plants pharmacological material basis. Chin. Med. 2018, 13, 33. [CrossRef] 31. Crentsil, J.A.; Yamthe, L.R.T.; Anibea, B.Z.; Broni, E.; Kwofie, S.K.; Tetteh, J.K.A.; Osei-Safo, D. Leishmanicidal Potential of 32. Hardwickiic Acid Isolated From Croton sylvaticus. Front. Pharmacol. 2020, 11, 753. [CrossRef] [PubMed] Stevanovic, S.; Sencanski, M.; Danel, M.; Menendez, C.; Belguedj, R.; Bouraiou, A.; Nikolic, K.; Cojean, S.; Loiseau, P.M.; Glisic, S.; et al. Synthesis, in silico, and in vitro evaluation of anti-leishmanial activity of oxadiazoles and indolizine containing compounds flagged against anti-targets. Molecules 2019, 24, 1282. [CrossRef] [PubMed] 33. Panigrahi, D.; Mishra, A.; Sahu, S.K. Rational in silico drug design of HIV-RT inhibitors through G-QSAR and molecular docking study of 4-arylthio and 4-aryloxy-3-iodopyridine-2(1-H)-one derivative. Beni-Suef Univ. J. Basic Appl. Sci. 2020, 9, 48. [CrossRef] Sinha, C.; Nischal, A.; Bandaru, S.; Kasera, P.; Rajput, A.; Nayarisseri, A.; Khattri, S. An In silico Approach for Identification of Novel Inhibitors as a Potential Therapeutics Targeting HIV-1 Viral Infectivity Factor. Curr. Top. Med. Chem. 2015, 15, 65–72. [CrossRef] 34. 35. Chen, Y.; Tian, Y.; Gao, Y.; Wu, F.; Luo, X.; Ju, X.; Liu, G. In silico Design of Novel HIV-1 NNRTIs Based on Combined Modeling Studies of Dihydrofuro[3,4-d]pyrimidines. Front. Chem. 2020, 8, 164. [CrossRef] 36. Kirchmair, J.; Distinto, S.; Roman Liedl, K.; Markt, P.; Maria Rollinger, J.; Schuster, D.; Maria Spitzer, G.; Wolber, G. Development of Anti-Viral Agents Using Molecular Modeling and Virtual Screening Techniques. Infect. Disord.—Drug Targets 2012, 11, 64–93. [CrossRef] 37. Bonsignori, M.; Scott, E.; Wiehe, K.; Easterhoff, D.; Munir Alam, S.; Hwang, K.-K.; Cooper, M.; Xia, S.-M.; Zhang, R.; Montefiori, D.C.; et al. Inference of the HIV-1 VRC01 Antibody Lineage Unmutated Common Ancestor Reveals Alternative Pathways to Overcome a Key Glycan Barrier. Immunity 2018, 49, 1162–1174. [CrossRef] 38. Kufareva, I.; Abagyan, R. Methods of Protein Structure Comparison. In Methods in Molecular Biology; NIH Public Access: Clifton, NJ, USA, 2011; Volume 857, pp. 231–257. 39. Laskowski, R.A.; Swindells, M.B. LigPlot+: Multiple Ligand–Protein Interaction Diagrams for Drug Discovery. J. Chem. Inf. Model. 2011, 51, 2778–2786. [CrossRef] 40. Artimo, P.; Jonnalagedda, M.; Arnold, K.; Baratin, D.; Csardi, G.; de Castro, E.; Duvaud, S.; Flegel, V.; Fortier, A.; Gasteiger, E.; et al. ExPASy: SIB bioinformatics resource portal. Nucleic Acids Res. 2012, 40, W597–W603. [CrossRef] 41. Choudhary, M.I.; Shaikh, M.; tul-Wahab, A.; ur-Rahman, A. In silico identification of potential inhibitors of key SARS-CoV-2 3CL hydrolase (Mpro) via molecular docking, MMGBSA predictive binding energy calculations, and molecular dynamics simulation. PLoS ONE 2020, 15, e0235030. [CrossRef] 42. Wen, C.C.; Kuo, Y.H.; Jan, J.T.; Liang, P.H.; Wang, S.Y.; Liu, H.G.; Lee, C.K.; Chang, S.T.; Kuo, C.J.; Lee, S.S.; et al. Specific plant terpenoids and lignoids possess potent antiviral activities against severe acute respiratory syndrome coronavirus. J. Med. Chem. 2007, 50, 4087–4095. [CrossRef] [PubMed] Molecules 2023, 28, 474 27 of 29 43. Veber, D.F.; Johnson, S.R.; Cheng, H.; Smith, B.R.; Ward, K.W.; Kopple, K.D. Molecular Properties That Influence the Oral Bioavailability of Drug Candidates. J. Med. Chem. 2002, 45, 2615–2623. [CrossRef] [PubMed] 44. Arnott, J.A.; Kumar, R.; Planey, S.L. Lipophilicity Indices for Drug Development. J. Appl. Biopharm. Pharmacokinet. 2013, 1, 31–36. 45. [CrossRef] Fernandes, J.; Gattass, C.R. Topological polar surface area defines substrate transport by multidrug resistance associated protein 1 (MRP1/ABCC1). J. Med. Chem. 2009, 52, 1214–1218. [CrossRef] 46. Pajouhesh, H.; Lenz, G.R. Medicinal chemical properties of successful central nervous system drugs. NeuroRx 2005, 2, 541–553. [CrossRef] 47. Hitchcock, S.A.; Pennington, L.D. Structure−Brain Exposure Relationships. J. Med. Chem. 2006, 49, 7559–7583. [CrossRef] 48. Drwal, M.N.; Banerjee, P.; Dunkel, M.; Wettig, M.R.; Preissner, R. ProTox: A web server for the in silico prediction of rodent oral toxicity. Nucleic Acids Res. 2014, 42, W53–W58. [CrossRef] 49. Lei, C.; Yang, J.; Hu, J.; Sun, X. On the Calculation of TCID50 for Quantitation of Virus Infectivity. Virol. Sin. 2021, 36, 141–144. [CrossRef] 50. Péret-Almeida, L.; Cherubino, A.P.F.; Alves, R.J.; Dufossé, L.; Glória, M.B.A. Separation and determination of the physico-chemical characteristics of curcumin, demethoxycurcumin and bisdemethoxycurcumin. Food Res. Int. 2005, 38, 1039–1044. [CrossRef] 51. Kwong, P.D.; Wyatt, R.; Robinson, J.; Sweet, R.W.; Sodroski, J.; Hendrickson, W.A. Structure of an HIV gp120 envelope glycoprotein in complex with the CD4 receptor and a neutralizing human antibody. Nature 1998, 393, 648–659. [CrossRef] 52. Wood, N.T.; Fadda, E.; Davis, R.; Grant, O.C.; Martin, J.C.; Woods, R.J.; Travers, S.A. The influence of N-linked glycans on the molecular dynamics of the HIV-1 gp120 V3 loop. PLoS ONE 2013, 8, e80301. [CrossRef] [PubMed] 53. Moebius, U.; Clayton, L.K.; Abraham, S.; Harrison, S.C.; Reinherz, E.L. The human inmaunodeficiency virus gpl20 binding site on CD4: Delineation by quantitative equilibrium and kinetic binding studies of mutants in conjunction with a high-resohtion cd4 atomic structure. J. Exp. Med. 1992, 176, 507–517. [CrossRef] [PubMed] 54. Kowalski, M.; Potz, J.; Basiripour, L.; Dorfman, T.; Goh, W.C.; Terwilliger, E.; Dayton, A.; Rosen, C.; Haseltine, W.; Sodroski, J. Functional regions of the envelope glycoprotein of human immunodeficiency virus type 1. Science 1987, 237, 1351–1355. [CrossRef] [PubMed] 55. Olshevsky, U.; Helseth, E.; Furman, C.; Li, J.; Haseltine, W.; Sodroski, J. Identification of individual human immunodeficiency virus type 1 gp120 amino acids important for CD4 receptor binding. J. Virol. 1990, 64, 5701–5707. [CrossRef] 56. Lu, L.; Yu, F.; Cai, L.; Debnath, A.K.; Jiang, S. Development of Small-molecule HIV Entry Inhibitors Specifically Targeting gp120 or gp41. Curr. Top. Med. Chem. 2016, 16, 1074–1090. [CrossRef] 57. Dey, B.; Berger, E.A. Blocking HIV-1 gp120 at the Phe43 cavity: If the extension fits. Structure 2013, 21, 871–872. [CrossRef] 58. Zhou, T.; Xu, L.; Dey, B.; Hessell, A.J.; Van Ryk, D.; Xiang, S.H.; Yang, X.; Zhang, M.Y.; Zwick, M.B.; Arthos, J.; et al. Structural definition of a conserved neutralization epitope on HIV-1 gp120. Nature 2007, 445, 732–737. [CrossRef] 59. Barret, R. Lipinski’s Rule of Five. In Therapeutical Chemistry; Elsevier Ltd.: Oxford, UK, 2018; pp. 97–100. [CrossRef] 60. Dahan, A.; Miller, J.M. The solubility-permeability interplay and its implications in formulation design and development for poorly soluble drugs. AAPS J. 2012, 14, 244–251. [CrossRef] 61. Amidon, G.L.; Lennernäs, H.; Shah, V.P.; Crison, J.R. A theoretical basis for a biopharmaceutic drug classification: The correlation of in vitro drug product dissolution and in vivo bioavailability. Pharm. Res. 1995, 12, 413–420. [CrossRef] 62. Pardridge, W.M. The blood-brain barrier: Bottleneck in brain drug development. NeuroRX 2005, 2, 3–14. [CrossRef] 63. Heller, F.; Voelker, S.; Wacharamanotham, C.; Borchers, J. Transporters. In Proceedings of the Proceedings of the 33rd Annual ACM Conference Extended Abstracts on Human Factors in Computing Systems—CHI EA ’15, Seoul, Republic of Korea, 18–23 April 2015. 64. McDonnell, A.M.; Dang, C.H. Basic review of the cytochrome p450 system. J. Adv. Pract. Oncol. 2013, 4, 263–268. [PubMed] 65. Banerjee, P.; Eckert, A.O.; Schrey, A.K.; Preissner, R. ProTox-II: A webserver for the prediction of toxicity of chemicals. Nucleic Acids Res. 2018, 46, W257–W263. [CrossRef] [PubMed] 66. Magaret, C.A.; Benkeser, D.C.; Williamson, B.D.; Borate, B.R.; Carpp, L.N.; Georgiev, I.S.; Setliff, I.; Dingens, A.S.; Simon, N.; Carone, M.; et al. Prediction of VRC01 neutralization sensitivity by HIV-1 gp160 sequence features. PLoS Comput. Biol. 2019, 15, e1006952. [CrossRef] [PubMed] 67. Cohen, Y.Z.; Lorenzi, J.C.C.; Krassnig, L.; Barton, J.P.; Burke, L.; Pai, J.; Lu, C.L.; Mendoza, P.; Oliveira, T.Y.; Sleckman, C.; et al. Relationship between latent and rebound viruses in a clinical trial of anti-HIV-1 antibody 3BNC117. J. Exp. Med. 2018, 215, 2311–2324. [CrossRef] [PubMed] 68. Walker, L.M.; Huber, M.; Doores, K.J.; Falkowska, E.; Pejchal, R.; Julien, J.-P.; Wang, S.-K.; Ramos, A.; Chan-Hui, P.-Y.; Moyle, M.; et al. Broad neutralization coverage of HIV by multiple highly potent antibodies. Nature 2011, 477, 466–470. [CrossRef] [PubMed] 69. Ray, N.; Hwang, C.; Healy, M.D.; Whitcomb, J.; Lataillade, M.; Wind-Rotolo, M.; Krystal, M.; Hanna, G.J. Prediction of Virological Response and Assessment of Resistance Emergence to the HIV-1 Attachment Inhibitor BMS-626529 During 8-Day Monotherapy With Its Prodrug BMS-663068. JAIDS J. Acquir. Immune Defic. Syndr. 2013, 64, 7–15. [CrossRef] [PubMed] Molecules 2023, 28, 474 28 of 29 70. Lalezari, J.P.; Latiff, G.H.; Brinson, C.; Echevarría, J.; Treviño-Pérez, S.; Bogner, J.R.; Thompson, M.; Fourie, J.; Sussmann Pena, O.A.; Mendo Urbina, F.C.; et al. Safety and efficacy of the HIV-1 attachment inhibitor prodrug BMS-663068 in treatment-experienced individuals: 24 week results of AI438011, a phase 2b, randomised controlled trial. Lancet HIV 2015, 2, e427–e437. [CrossRef] [PubMed] 71. Meuser, M.E.; Murphy, M.B.; Rashad, A.A.; Cocklin, S. Kinetic Characterization of Novel HIV-1 Entry Inhibitors: Discovery of a Relationship between Off-Rate and Potency. Molecules 2018, 23, 1940. [CrossRef] 72. Pancera, M.; Lai, Y.-T.; Bylund, T.; Druz, A.; Narpala, S.; O’Dell, S.; Schön, A.; Bailer, R.T.; Chuang, G.-Y.; Geng, H.; et al. Crystal structures of trimeric HIV envelope with entry inhibitors BMS-378806 and BMS-626529. Nat. Chem. Biol. 2017, 13, 1115–1122. [CrossRef] 73. Zhao, Q.; Ma, L.; Jiang, S.; Lu, H.; Liu, S.; He, Y.; Strick, N.; Neamati, N.; Debnath, A.K. Identification of N-phenyl-N(cid:48)-(2,2,6,6- tetramethyl-piperidin-4-yl)-oxalamides as a new class of HIV-1 entry inhibitors that prevent gp120 binding to CD4. Virology 2005, 339, 213–225. [CrossRef] Singh, S.; Kumar, A.; Brijwal, M.; Choudhary, A.; Singh, K.; Singh, R.; Husain, M.; Dar, L. Intra-Clade C signature polymorphisms in HIV-1 LTR region: The Indian and African lookout. Virus Res. 2021, 297, 198370. [CrossRef] [PubMed] 74. 75. Ahn, M.J.; Chul, Y.K.; Ji, S.L.; Tae, G.K.; Seung, H.K.; Lee, C.K.; Lee, B.B.; Shin, C.G.; Huh, H.; Kim, J. Inhibition of HIV-1 integrase by galloyl glucoses from Terminalia chebula and flavonol glycoside gallates from Euphorbia pekinensis. Planta Med. 2002, 68, 457–459. [CrossRef] [PubMed] 76. Gambari, R.; Lampronti, I. Inhibition of immunodeficiency type-1 virus (HIV-1) life cycle by medicinal plant extracts and 77. plant-derived compounds. Adv. Phytomed. 2006, 2, 299–311. Salehi, B.; Kumar, N.; ¸Sener, B.; Sharifi-Rad, M.; Kılıç, M.; Mahady, G.; Vlaisavljevic, S.; Iriti, M.; Kobarfard, F.; Setzer, W.; et al. Medicinal Plants Used in the Treatment of Human Immunodeficiency Virus. Int. J. Mol. Sci. 2018, 19, 1459. [CrossRef] [PubMed] Singh, I.P.; Bharate, S.B.; Bhutani, K.K. Anti-HIV natural products. Curr. Sci. 2005, 89, 269–290. 78. 79. Lee, J.; Hattori, M.; Kim, J. Inhibition of HIV-1 Protease and RNase H of HIV-1 Reverse Transcriptase Activities by Long Chain 80. Phenols from the Sarcotestas of Ginkgo biloba. Planta Med. 2008, 74, 532–534. [CrossRef] Salve, J.; Pate, S.; Debnath, K.; Langade, D. Adaptogenic and Anxiolytic Effects of Ashwagandha Root Extract in Healthy Adults: A Double-blind, Randomized, Placebo-controlled Clinical Study. Cureus 2019, 11, e6466. [CrossRef] 81. Kurapati, K.R.V.; Atluri, V.S.R.; Samikkannu, T.; Nair, M.P.N. Ashwagandha (Withania somnifera) Reverses β-Amyloid1-42 Induced Toxicity in Human Neuronal Cells: Implications in HIV-Associated Neurocognitive Disorders (HAND). PLoS ONE 2013, 8, e77624. [CrossRef] 82. Yuan, S.; Chan, H.C.S.; Hu, Z. Using PyMOL as a platform for computational drug design. Wiley Interdiscip. Rev. Comput. Mol. Sci. 2017, 7, e1298. [CrossRef] 83. Abraham, M.J.; Murtola, T.; Schulz, R.; Páll, S.; Smith, J.C.; Hess, B.; Lindahl, E. GROMACS: High performance molecular simulations through multi-level parallelism from laptops to supercomputers. SoftwareX 2015, 1–2, 19–25. [CrossRef] 84. Turner, P. XMGRACE, Version 5.1. 19; Center for Coastal and Land-Margin Research, Oregon Graduate Institute of Science and Technology: Beaverton, OR, USA, 2005. 85. Humphrey, W.; Dalke, A.; Schulten, K. VMD: Visual molecular dynamics. J. Mol. Graph. 1996, 14, 33–38. [CrossRef] [PubMed] 86. O’Boyle, N.M.; Banck, M.; James, C.A.; Morley, C.; Vandermeersch, T.; Hutchison, G.R. Open Babel: An open chemical toolbox. J. Cheminform. 2011, 3, 33. [CrossRef] 87. Dallakyan, S.; Olson, A.J. Small-Molecule Library Screening by Docking with PyRx. In Methods in Molecular Biology; NIH Public Access: Clifton, NJ, USA, 2015; Volume 1263, pp. 243–250. ISBN 9781118435762. [CrossRef] 88. Huang, S.-Y. Comprehensive assessment of flexible-ligand docking algorithms: Current effectiveness and challenges. Brief. Bioinform. 2017, 19, 982–994. [CrossRef] [PubMed] 89. Trott, O.; Olson, A.J. Software news and update AutoDock Vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. J. Comput. Chem. 2010, 31, 455–461. [CrossRef] [PubMed] 90. Rasheed, M.; Bettadapura, R.; Bajaj, C. Computational Refinement and Validation Protocol for Proteins with Large Variable Regions Applied to Model HIV Env Spike in CD4 and 17b Bound State. Structure 2015, 23, 1138–1149. [CrossRef] 91. Kaplan, W.; Littlejohn, T.G. Swiss-PDB Viewer (Deep View). Brief. Bioinform. 2001, 2, 195–197. [CrossRef] 92. Cheng, X.; Ivanov, I. Molecular Dynamics. In Methods in Molecular Biology; NIH Public Access: Clifton, NJ, USA, 2012; Volume 929, pp. 243–285. ISBN 9781627030496. 93. Dong, Y.W.; Liao, M.L.; Meng, X.L.; Somero, G.N. Structural flexibility and protein adaptation to temperature: Molecular dynamics analysis of malate dehydrogenases of marine molluscs. Proc. Natl. Acad. Sci. USA 2018, 115, 1274–1279. [CrossRef] 94. Daina, A.; Michielin, O.; Zoete, V. SwissADME: A free web tool to evaluate pharmacokinetics, drug-likeness and medicinal chemistry friendliness of small molecules. Sci. Rep. 2017, 7, 42717. [CrossRef] 95. Naldini, L.; Blömer, U.; Gallay, P.; Ory, D.; Mulligan, R.; Gage, F.H.; Verma, I.M.; Trono, D. In vivo gene delivery and stable transduction of nondividing cells by a lentiviral vector. Science 1996, 272, 263–267. [CrossRef] Molecules 2023, 28, 474 29 of 29 96. Longo, P.A.; Kavran, J.M.; Kim, M.S.; Leahy, D.J. Transient mammalian cell transfection with polyethylenimine (PEI). In Methods 97. in Enzymology; Elsevier: Amsterdam, The Netherlands, 2013; Volume 529, pp. 227–240. [CrossRef] Ferrara, F.; Temperton, N.; Ferrara, F.; Temperton, N. Pseudotype Neutralization Assays: From Laboratory Bench to Data Analysis. Methods Protoc. 2018, 1, 8. [CrossRef] [PubMed] Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
10.1080_2162402x.2020.1790716
OncoImmunology ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/koni20 IL15 synergizes with radiotherapy to reprogram the tumor immune contexture through a dendritic cell connection Karsten A. Pilones, Maud Charpentier, Elena Garcia-Martinez & Sandra Demaria To cite this article: Karsten A. Pilones, Maud Charpentier, Elena Garcia-Martinez & Sandra Demaria (2020) IL15 synergizes with radiotherapy to reprogram the tumor immune contexture through a dendritic cell connection, OncoImmunology, 9:1, 1790716, DOI: 10.1080/2162402X.2020.1790716 To link to this article: https://doi.org/10.1080/2162402X.2020.1790716 © 2020 The Author(s). Published with license by Taylor & Francis Group, LLC. Published online: 09 Jul 2020. Submit your article to this journal Article views: 1266 View related articles View Crossmark data Citing articles: 3 View citing articles Full Terms & Conditions of access and use can be found at https://www.tandfonline.com/action/journalInformation?journalCode=koni20 ONCOIMMUNOLOGY 2020, VOL. 9, NO. 1, 1–2 https://doi.org/10.1080/2162402X.2020.1790716 AUTHOR’S VIEW IL15 synergizes with radiotherapy to reprogram the tumor immune contexture through a dendritic cell connection Karsten A. Pilones a, Maud Charpentier a, Elena Garcia-Martinezb, and Sandra Demaria a,c aDepartment of Radiation Oncology, Weill Cornell Medicine, New York, NY, USA; bHematology and Oncology Department, Hospital Universitario Morales Meseguer, Murcia, Spain; cDepartment of Pathology and Laboratory Medicine, Weill Cornell Medicine, New York, NY, USA ABSTRACT IL15 is a key cytokine for the activation and survival of anti-tumor effectors CD8+ T and NK cells. Recently published preclinical studies demonstrate that the therapeutic activity of IL15 requires conventional dendritic cells type 1 (cDC1). Radiotherapy cooperates with IL15 by enhancing cDC1 tumor infiltration via interferon type 1 activation. ARTICLE HISTORY Received 28 June 2020 Accepted 29 June 2020 KEYWORDS IL-15; radiotherapy; dendritic cells; interferon type 1; CD8 T-cells Cytokines are key regulators of the immune response and among the first biologics to be tested for the ability to elicit anti-tumor immune responses.1 Interleukin-15 (IL15) is a common gamma- chain cytokine that promotes the proliferation, survival and activa- tion of natural killer (NK) and CD8+ T cells, and was ranked by Martin Cheever at the top of a list of immunotherapy agents with the potential to treat cancer.2 Early phase clinical trials have shown that recombinant human IL15 (rhIL15) administered subcuta- neously (s.c.) has good tolerability and activity, measured as signifi- cant NK and CD8+ T-cell expansion, but limited anti-cancer effects.3 IL15 is produced as a heterodimer with the alpha chain of the IL15 receptor (IL15Rα), and it is this heterodimer that has biolo- gical activity,4 prompting the development of similar constructs to increase the stability and activity of IL15 administered therapeuti- cally. One such construct, het-IL15, was recently tested by Bergamaschi and colleagues in two different mouse models of carcinoma, colon MCA38 and lung TC1.5 They observed an increased tumor infiltration, proliferation and survival of CD8+ T and NK cells in mice treated with intraperitoneal het-IL15 that was associated with a significantly slower tumor growth as com- pared to untreated mice. Further analyses confirmed a gene sig- nature associated with activated, proliferative, cytotoxic lymphocytes in het-IL15-treated tumors with granzyme A and B among the most upregulated transcripts. In addition, NK and CD8+ T cells infiltrating the tumor of het-IL15-treated mice pro- duced XCL1, a chemokine that drives conventional dendritic cells type 1 (cDC1) to the tumor.6 Consistently, they observed an increased accumulation of cDC1, which are specialized in cross- presenting tumor antigens to CD8+ T cells.6 The cDC1 recruited to the tumor, in turn, secreted CXCL9 and CXCL10 in response to het-IL15 and IFN-γ released by T and NK cells, attracting more effector T cells. Overall, this study suggests that het-IL15 boosts anti-tumor immune responses by amplification of a cycle of recruitment and activation of CD8+ T cells that hinges upon cDC1. Our group tested rhIL15 administered subcutaneously in three mouse models of carcinoma, the TSA breast cancer, MCA38 colon cancer and LLC1 lung cancer and found that it did not have any anti-tumor effect.7 The discrepancy with the results of Bergamaschi et al.,5 may reflect the improved biological activity of IL15 administered as a heterodimer with IL15Rα chain. Additionally, it may be due to the delayed administration in our study (day 12 versus day 5 post-tumor inoculation in experiments performed by Bergamaschi et al), when the tumor microenviron- ment is well-established. The lack of clinical activity of single-agent IL15, even when given as a complex with the IL15Rα chain to patients with advanced cancer,8 suggests that IL15 is unable by itself to overcome the immune suppression associated with tumor progression. However, we found that IL15 synergized with focal tumor radiotherapy leading to improved control and often com- plete regression of the irradiated tumor and protective memory responses in cured mice. The combination of IL15 and radio- therapy to one tumor also led to inhibition of a synchronous non- irradiated tumor (abscopal effect).7 These responses were abro- gated by CD8+ T cell depletion, and in the absence of cDC1, in Batf3-deficient mice. Similar to Bergamaschi et al.,5 we observed an increase in cDC1 in the tumor of wild-type mice treated with IL15, but only when given together with radiation. Radiation itself increased cDC1 presence in the tumor, but the addition of IL15 significantly enhanced this effect, and increased the expression of costimulatory molecules CD80, CD86 and CD40 on intra-tumoral cDC1, and the priming of tumor-antigen specific CD8+ T cells in draining lymph nodes. Further investigation showed that cancer cell-intrinsic interferon type I (IFN-I) upregulation by radiation was required for the synergy of radiotherapy with IL15. Mice bearing the lung tumor LLC1, which expresses very low levels of the cytosolic DNA sensor cGAS and is thus unable to respond to radiation with increased IFNβ production, did not show any improvement in control of the irradiated tumor or survival when given IL15. Additionally, in TSA tumor-bearing mice the thera- peutic synergy between IL15 and radiation was observed when radiation was given as 8 Gy doses on three consecutive days, a regimen that optimally activates IFN-I, but not as a single dose CONTACT Sandra Demaria © 2020 The Author(s). Published with license by Taylor & Francis Group, LLC. This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial License (http://creativecommons.org/licenses/by-nc/4.0/), which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited. Department of Radiation Oncology, Weill Cornell Medicine, New York, NY 10065, USA szd3005@med.cornell.edu 2 K. PILONES ET AL. Figure 1. Proposed model of the interactions between radiation and IL15 in the tumor microenvironment promoting anti-tumor immunity. (Left) In advanced tumors optimized radiation therapy regimens elicit cancer-intrinsic activation of the type I IFN pathway and tumor antigen exposure promoting the recruitment of cDC1 that cross-present tumor antigens and transpresent the subcutaneously administered recombinant IL15, resulting in cytotoxic CD8+ T cell infiltration, tumor control and establishment of anti-tumor effector and memory responses. (Right) In more immunogenic/less advanced tumors systemic treatment with bio- logically active IL15-IL15Rα dimers (het-IL15) expands and activates preexisting anti- tumor CD8+ T and NK cell, which infiltrate the tumor and secrete XCL1 to recruit cDC1. cDC1, in turn, promote further tumor infiltration by CD8+ T and NK cells. of 20 Gy, which is a poor inducer of IFN-I pathway in this model.9 The increase in intra-tumoral cDC1 was abrogated by antibody- mediated blockade of the IFN-I receptor, consistent with the role of IFN-I in cDC1 recruitment. Taken together, data from these studies support a model whereby IL15 activates and expands a preexisting effector popula- tion of CD8+ and NK cells which produce chemokines to attract cDC1 to the tumor. The latter, in turn, produce CXCL10 and amplify the cancer-immunity cycle10 by attracting and activating more effector cells. In advanced tumors, when this effector response is largely suppressed, radiation is required to jumpstart the cancer-immunity cycle by inducing cancer cell-intrinsic IFN-I that attracts cDC1, and CXCL10 that attracts effector T cells (Figure 1). Recruited cDC1, which express IL15Rα, can then cross-present tumor antigens released by radiation and trans- present IL15 resulting in activation of CD8+ T cells within the tumor as well as the draining lymph nodes. Thus, the ability of IL15 to foster the cross-talk of CD8+ T and NK cells with cDC1 appears to be a critical component of IL15-induced reprogram- ming of the tumor immune contexture toward tumor rejection. as agent single It remains to be investigated if the key role of cDC1 identified in the preclinical studies will also be seen in patients treated with IL15. Several early trials are ongoing testing het-IL15 and another construct, ALT-803, (NCT03054909, NCT02452268, NCT02099539), whereas rh-IL15 is being tested in combination with other agents but not with radiotherapy (NCT03388632, NCT02689453, NCT03759184, NCT03905135, NCT04150562). Our results suggest is a promising candidate for testing with IL15, but caution that the intrinsic tumor expression of cytosolic DNA sensors and the radiation dose and fractionation should be considered in the design of such studies. that radiotherapy Disclosure of potential conflicts of interest The authors declared that no conflict of interest exists related to this work, but S.D. has received honorarium from Lytix Biopharma, EMD Serono, and Mersana Therapeutics for advisory service, and research grants from Lytix Biopharma and Nanobiotix. EGM has served as consultant or speaker for Roche, AstraZeneca, Clovis and Pharmamar, research funding from Roche and financial support from AstraZeneca, Roche, Pharmamar, Pfizer, Bristol Meyer Squibb and MSD. Funding This work was supported by grants from the NIH (R01 CA198533 and R01CA201246), The Chemotherapy Foundation and the Breast Cancer Research Foundation (to S.D.). Breast Cancer Research Foundation (to S. D.). Elena García-Martínez was supported by a grant from the Spanish Breast Cancer Group (GEICAM) ORCID Karsten A. Pilones Maud Charpentier Sandra Demaria http://orcid.org/0000-0002-0982-3260 http://orcid.org/0000-0002-7343-4163 http://orcid.org/0000-0003-4426-0499 References 1. Garcia-Martinez E, Smith M, Buque A, Aranda F, de la Pena FA, Ivars A, Canovas MS, Conesa MAV, Fucikova J, Spisek R, et al. Trial Watch: Immunostimulation with recombinant cytokines for cancer therapy. Oncoimmunology. 2018;7:e1433982. doi:10.1080/2162402X.2018.1433982. 2. Cheever MA. Twelve immunotherapy drugs that could cure cancers. Immunol Rev. 2008;222:357–368. doi:10.1111/j.1600-065X.2008.00604.x. 3. Miller JS, Morishima C, McNeel DG, Patel MR, Kohrt HEK, Thompson JA, Sondel PM, Wakelee HA, Disis ML, Kaiser JC, et al. A First-in-Human Phase I Study of Subcutaneous Outpatient Recombinant Human IL15 (rhIL15) in Adults with Advanced Solid Tumors. Clin Cancer Res. 2018;24:1525–1535. doi:10.1158/1078-0432.CCR-17-2451. 4. Steel JC, Waldmann TA, Morris JC. Interleukin-15 biology and its in cancer. Trends Pharmacol Sci. therapeutic 2012;33:35–41. doi:10.1016/j.tips.2011.09.004. implications 5. Bergamaschi C, Pandit H, Nagy BA, Stellas D, Jensen SM, Bear J, Cam M, Valentin A, Fox BA, Felber BK, et al. Heterodimeric IL-15 delays tumor growth and promotes intratumoral CTL and dendri- tic cell accumulation by a cytokine network involving XCL1, IFN-gamma, CXCL9 and CXCL10. J Immunother Cancer. 2020;8:e000599. doi:10.1136/jitc-2020-000599. 6. Bottcher JP, Reis ESC. The Role of Type 1 Conventional Dendritic Cells in Cancer Immunity. Trends Cancer. 2018;4:784–792. doi:10.1016/j.trecan.2018.09.001. 7. Pilones KA, Charpentier M, Garcia-Martinez E, Daviaud C, Kraynak J, Aryankalayil J, Formenti SC, Demaria S. Radiotherapy cooperates with IL15 to induce antitumor immune responses. Cancer Immunol Res OnlineFirst June. 2020;12:2020. doi:10.1158/ 2326-6066.CIR-1119-0338. 8. Margolin K, Morishima C, Velcheti V, Miller JS, Lee SM, Silk AW, Holtan SG, Lacroix AM, Fling SP, Kaiser JC, et al. Phase I Trial of ALT-803, A Novel Recombinant IL15 Complex, in Patients with Advanced Solid Tumors. Clin Cancer Res. 2018;24:5552–5561. doi:10.1158/1078-0432.CCR-18-0945. 9. Vanpouille-Box C, Alard A, Aryankalayil MJ, Sarfraz Y, Diamond JM, Schneider RJ, Inghirami G, Coleman CN, Formenti SC, Demaria S. DNA exonuclease Trex1 regulates radiotherapy-induced tumour immunogenicity. Nat Commun. 2017;8:15618. doi:10.1038/ncomms15618. 10. Chen DS, Mellman cancer-immunity cycle. immuni.2013.07.012. I. Oncology meets the Immunity. 2013;39:1–10. doi:10.1016/j. immunology:
10.3390_ijms24010865
Article The Inhibitory Properties of a Novel, Selective LMTK3 Kinase Inhibitor , Andrea Lauer Betrán 1,†, Athanasios Papakyriakou 2 Alessandro Agnarelli 1,† Mark Samuels 1, Panagiotis Papanastasopoulos 1, Christina Giamas 1, Erika J. Mancini 1, Justin Stebbing 3, John Spencer 4 and Georgios Giamas 1,* , Angeliki Ditsiou 1 , Chiara Cilibrasi 1 , Viviana Vella 1 , 1 Department of Biochemistry and Biomedicine, School of Life Sciences, University of Sussex, 2 Brighton BN1 9QG, UK Institute of Biosciences and Applications, National Centre for Scientific Research “Demokritos”, 15341 Athens, Greece 3 Department of Surgery and Cancer, Imperial College, London SW7 2BX, UK 4 Sussex Drug Discovery Centre, School of Life Sciences, University of Sussex, Brighton BN1 9QG, UK * Correspondence: g.giamas@sussex.ac.uk † These authors contributed equally to this work. Abstract: Recently, the oncogenic role of lemur tyrosine kinase 3 (LMTK3) has been well established in different tumor types, highlighting it as a viable therapeutic target. In the present study, using in vitro and cell-based assays coupled with biophysical analyses, we identify a highly selective small molecule LMTK3 inhibitor, namely C36. Biochemical/biophysical and cellular studies revealed that C36 displays a high in vitro selectivity profile and provides notable therapeutic effect when tested in the National Cancer Institute (NCI)-60 cancer cell line panel. We also report the binding affinity between LMTK3 and C36 as demonstrated via microscale thermophoresis (MST). In addition, C36 exhibits a mixed-type inhibition against LMTK3, consistent with the inhibitor overlapping with both the adenosine 5(cid:48)-triphosphate (ATP)- and substrate-binding sites. Treatment of different breast cancer cell lines with C36 led to decreased proliferation and increased apoptosis, further reinforcing the prospective value of LMTK3 inhibitors for cancer therapy. Keywords: LMTK3; kinase inhibitor; breast cancer 1. Introduction Protein kinases are a large family of enzymes responsible for catalyzing protein phos- phorylation. They are involved in critical mechanisms regulating different cellular func- tions, including proliferation, cell cycle, apoptosis, motility, growth, and differentiation [1]. The deregulation of protein kinase activity contributes to various human diseases and disorders, including cancer [2]. Therefore, it is not surprising that the kinome is considered an attractive target for the treatment of several tumors, leading to a shift in the clinical management of cancer and improved patient outcome [3]. However, despite promising results, the inevitable development of drug resistance, largely due to the activation of com- plementary and/or compensatory pathways, remains a major limitation for this therapeutic approach [4,5]. Lemur tyrosine kinase 3 (LMTK3) is a dual specificity serine/threonine kinase com- posed of a transmembrane helical segment, a kinase domain, and a C-terminal intrinsically disordered region [6]. Studies have put forward a physiological role for LMTK3 in neu- ron trafficking where LMTK3 knockout can cause behavioral abnormalities in mice [7]. Although information regarding the function of LMTK3 in normal physiology is limited, its oncogenic role has been well established so far in various tumor types, including blad- der, lung, and colorectal cancer, among others, highlighting it as a potential therapeutic target [8–22]. LMTK3 was originally identified as an important regulator of estrogen re- ceptor alpha (ERα) activity in breast cancer (BC) following a whole human kinome siRNA Citation: Agnarelli, A.; Lauer Betrán, A.; Papakyriakou, A.; Vella, V.; Samuels, M.; Papanastasopoulos, P.; Giamas, C.; Mancini, E.J.; Stebbing, J.; Spencer, J.; et al. The Inhibitory Properties of a Novel, Selective LMTK3 Kinase Inhibitor. Int. J. Mol. Sci. 2023, 24, 865. https://doi.org/ 10.3390/ijms24010865 Academic Editor: Jan Korabecny Received: 22 September 2022 Revised: 23 November 2022 Accepted: 2 December 2022 Published: 3 January 2023 Copyright: © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). Int. J. Mol. Sci. 2023, 24, 865. https://doi.org/10.3390/ijms24010865 https://www.mdpi.com/journal/ijms International Journal of Molecular Sciences Int. J. Mol. Sci. 2023, 24, 865 2 of 15 screen [8]. Specifically, LMTK3 was shown to directly protect ERα from ubiquitin-mediated proteasomal degradation and indirectly promote ERα transcription through the PKC/AKT signaling pathway [8]. Follow-up studies have further supported that elevated levels of LMTK3 in BC are associated with poorer overall survival (OS) and disease-free survival (DFS) [12]. Moreover, LMTK3 has also been implicated in endocrine [13] and chemotherapy resistance in BC [14], while us and others have described an involvement of LMTK3 in different signaling pathways [13,23]. Recently, using robust in vitro and cell-based screening and selectivity assays com- bined with biophysical analyses, we identified and characterized a highly selective small- molecule adenosine 5(cid:48)-triphosphate (ATP)-competitive LMTK3 inhibitor, namely C28, that acts by degrading LMTK3 via the ubiquitin-proteasome pathway [2]. Overall, C28 exhib- ited effective anticancer effects in several cancer cell lines, as well as in vivo BC mouse models (xenograft and transgenic) [2]. Here, we report the inhibitory properties of another compound (C36) against LMTK3, further supporting the rationale that the development and optimization of LMTK3 inhibitors can have prospective value to cancer patients. 2. Results 2.1. Selectivity Profile of C36 Inhibitor Considering the oncogenic role of LMTK3, a library encompassing 28,716 compounds (Charles River Discovery Research Services, Chesterford Research Park, UK Ltd.; formerly known as BioFocus DPI Ltd.) was screened using robust in vitro and cell-based assays identifying a potent small-molecule ATP-competitive LMTK3 inhibitor (C28), as previously described [2]. Among the hit compounds that were identified, C36 also emerged as a potential selective LMTK3 inhibitor (Figure 1A). To obtain a more detailed analysis of the selectivity profile of C36, we performed a radioactive filter binding assay, screening this inhibitor against a series of 140 kinases [24]. Our results identified 16 kinases whose activity was reduced by >50% in the presence of 1 µM C36 (Figure 1B) compared to 18 kinases when using C28 as previously described [2]. Figure 1. Cont. Int. J. Mol. Sci. 2023, 24, x FOR PEER REVIEW 3 of 18 Int. J. Mol. Sci. 2023, 24, 865 3 of 15 Figure 1. Selectivity of C36 toward LMTK3. (A) Chemical structure of C36. (B) Selectivity profile of C36 (1 µM) against 140 kinases using radioactive filter binding assay (MRC International Centre for Kinase Profiling unit). The data are displayed as the percentage of activity remaining of assay duplicates with a SD. Only kinases with >50% decrease in their activity are shown. The relative IC50 values are also presented. (C) TREEspot interaction map depicting the kinome phylogenetic grouping, with kinases interacting with C36 (5 µM) represented as red circles (DiscoverX KINOMEscan [25]). Kinases whose binding affinity was inhibited by C36 to less than 10% of the control (DMSO) are shown in the table. Lower numbers indicate the most probable hits to bind to C36. The larger the diameter of the circle, the higher the C36 binding affinity to the respective kinase active site. (D) The IC50 value for C36 against LMTK3cat (LMTK3 kinase domain) was determined by in vitro kinase assay. The intensities of the bands on the autoradiogram have been quantified using ImageJ software and normalized to total protein levels based on Coomassie Blue stained membranes. DMSO has been used as a control (E) MST binding curves for C36 (Kd = 1.87 ± 0.2 µM, red curve) and C28 (Kd = 2.50 ± 0.4 µM, green curve) with LMTK3, showing fraction bound on the Y axis and drug concentration (M) on the X axis. More specifically, fraction bound is calculated as the ratio between the emitted fluorescence of LMTK3-C36/C28 complex and the curve amplitude [26]. The error bars represent the SD of each data point calculated from three independent experiments. Binding check analysis reveals no interaction between DMSO (control) and LMTK3 kinase domain (signal to noise ratio: 1.2) (Figure S2). (F) IC50 values for C36 in FDCP1 and FDCP1-LMTK3 cell lines. Error bars represent the means ± SD from three independent experiments. To further examine the specificity of C36, we used an active site-directed competition binding assay (DiscoverX KINOMEscan, San Diego, CA, USA [25]) and quantitatively mea- sured the interactions between C36 and 403 purified human kinases. Figure 1C represents a TREEspot interaction map of our compound against 403 kinases. C36 was tested at a 5 µM final concentration and the red circles indicate kinases to which C36 binds at their active site at this concentration. In addition, the size of the circle is also proportional to the binding affinity of C36 to the respective kinase (i.e., the larger the diameter of the circle, the higher the binding affinity of C36). The data shown in the table (Figure 1C) display the most significant hits from the TREEspot interaction map. The percentage of DMSO (control) is also indicated, with 10% being the highest amount used. In particular, the lower the numbers in the “% control” column, the more probable C36 binds the kinase active site. Int. J. Mol. Sci. 2023, 24, x FOR PEER REVIEW 3 of 18 Int. J. Mol. Sci. 2023, 24, 865 4 of 15 The S(35) selectivity index of C36 was 0.114, as measured by the percentage of the kinome inhibited below 35% of the control at this concentration using the following equation: S(35) = number of kinase with % Ctrl < 35 number of kinases tested (1) Interestingly, the selectivity score for C36 (0.114) was lower when compared to C28 (0.186) [2], indicating a higher in vitro selectivity of C36 versus C28. More specifically, C36 inhibited the activity of 16 out of a total of 403 kinases by >90% (Figure 1C) compared to 33 out of 403 for C28 [2], with 5 of them overlapping (MYLK4, FLT3, GSG2, TRKA, HIPK4). The selectivity profile of C36 was determined using assays with different underlying principles (radioactive filter binding assay and active site-directed competition binding assay). However, it is noteworthy that there was an overlap of kinases targeted by C36 that have been identified via both assays (namely PIM1, TRKA, PDGFRA, MKNK1, MKNK2). This further validates the reliability of the obtained results. Dose-dependent in vitro 32P γ-ATP radiolabeled kinase assays revealed high efficiency of C36 to inhibit LMTK3 at low concentrations (<1 µM), as measured by the phosphorylation of substrate heat shock protein 27 (HSP27) by LMTK3 (Figure 1D). The half maximal inhibitory concentration (IC50) of C36 for LMTK3 was approximately 100 nM, as shown by the quantification of the in vitro kinase assay (Figure 1D) [2]. Moreover, as demonstrated by microscale thermophoresis (MST), C36 and C28 displayed comparable affinities to LMTK3 (1.87 ± 0.2 µM and 2.50 ± 0.4 µM, respectively) (Figure 1E). More specifically, this assay measures the movement of fluorescently tagged biomolecules in solution (NHS-647 red dye) through a temperature gradient produced by an infrared (IR) laser [26]. This physical phenomenon is also referred to as “thermophoresis” [27,28]. Since the thermophoretic behavior of a biomolecule depends on its hydration shell, charge, and size [27], the binding of a ligand/drug (in our case C36 and C28) to a molecule of interest (in our case LMTK3) will change the thermophoresis of the molecule of interest [27]. This change in thermophoretic behavior can then be used to analyze the dissociation constant (Kd) [27]. Following this, we used the interleukin-3 (IL-3)-dependent murine bone marrow- derived cell line FDCP-1 and engineered an LMTK3-transformed clone (FDCP-1/BCR- LMTK3) that relies on the constitutive expression of catalytically active LMTK3 for its survival and proliferation, as described previously [2]. Using this cell-based approach, we assessed the potency of compound C36 and determined the IC50 by tracking the cellular viability of FDCP-1 parental and FDCP-1/BRC-LMTK3. As shown in Figures 1F and S1, C36 displayed a higher inhibition of cell viability with FDCP-1/BCR-LMTK3 than the FDCP-1 parental cell line, indicating a C36 inhibition dependent on LMTK3. Taken together, we report the identification of a novel LMTK3 inhibitor (C36), displaying a high in vitro selectivity profile. 2.2. Biochemical/Mechanistic Investigation of C36 Binding to LMTK3 To investigate the mechanism of action of C36, we examined the effect of increasing HSP27 substrate concentrations on the inhibitory activity of the compound in the presence of constant ATP concentration. Data from the steady-state analysis were fitted to the Michaelis–Menten equation (Figure 2A). Our results from a single technical replicate revealed that the presence of C36 resulted in an increase of Km (0.486 µM from 0.364 µM in the absence of C36) with a significantly lower Vmax (26.1 µmol/min from 59.0 µmol/min in the absence of C36). Next, we investigated the effect of increasing concentrations of ATP at a fixed substrate (HSP27) concentration of 0.6 µM. Similarly, the presence of C36 resulted in a significant increase of the apparent Km (0.048 µM from 0.023 µM in the absence of C36) accompanied by a substantial decrease in Vmax (18.7 µmol/min from 87.6 µmol/min in the absence of C36, Figure 2B). It is noteworthy to emphasize that these results come from a single replicate and further analysis is required to confirm these data. So far, these results indicate a mixed-type inhibition of LMTK3 by C36, where the inhibitor may overlap Int. J. Mol. Sci. 2023, 24, 865 5 of 15 with both the ATP- and the substrate HSP27-binding sites without exclusively being a competitive inhibitor of ATP, or the substrate alone. Figure 2. Identification of C36 as a potent inhibitor against LMTK3. (A) Kinetic analysis of C36 inhibition with respect to HSP27 concentration (fixed ATP concentration). Kinetic parameters (Km and Vmax) were determined from nonlinear regression fit of the initial reaction rates as a function of HSP27 concentration to the Michaelis–Menten equation using GraphPad Prism 8.01 software (GraphPad Software, Inc., San Diego, CA, USA). (B) Kinetic analysis of C36 inhibition as a function of ATP concentration (fixed HSP27 concentration of 0.6 µM). Kinetic parameters (Km and Vmax) were determined from nonlinear regression fit of the initial reaction rates as a function of ATP concen- tration to the Michaelis–Menten equation using GraphPad Prism 8.01 software. (C) Characteristic melting plots obtained from CD spectroscopy for LMTK3 in the absence (DMSO) and presence of inhibitor (C36). (D) Characteristic melting curves obtained from thermal shift assay measurements. (E) Molecular model of LMTK3 in the active state with bound ATP and a peptide fragment of insulin receptor substrate 2 (IRS2). The kinase domain of LMTK3 is shown with green color, the bound ATP is color-coded with yellow C atoms, and the substrate with grey C atoms; blue is for N, red is for O, yellow is for S, and orange is for P. (F) Docked pose of C36 in the active state of ligand-free LMTK3. Inset is a close-up view illustrating residue-specific interactions. C36 is shown with purple C atoms and LMTK3 residues with cyan C atoms. Int. J. Mol. Sci. 2023, 24, x FOR PEER REVIEW 6 of 18 Figure 2. Identification of C36 as a potent inhibitor against LMTK3. (A) Kinetic analysis of C36 in-hibition with respect to HSP27 concentration (fixed ATP concentration). Kinetic parameters (Km and Vmax) were determined from nonlinear regression fit of the initial reaction rates as a function of HSP27 concentration to the Michaelis–Menten equation using GraphPad Prism 8.01 software (GraphPad Software, Inc., San Diego, CA, USA). (B) Kinetic analysis of C36 inhibition as a function of ATP concentration (fixed HSP27 concentration of 0.6 µM). Kinetic parameters (Km and Vmax) were determined from nonlinear regression fit of the initial reaction rates as a function of ATP concentra-tion to the Michaelis–Menten equation using GraphPad Prism 8.01 software. (C) Characteristic melt-ing plots obtained from CD spectroscopy for LMTK3 in the absence (DMSO) and presence of Int. J. Mol. Sci. 2023, 24, 865 6 of 15 In addition, we assessed the ability of C36 to bind LMTK3 in solution by monitoring the thermal denaturation of the enzyme in the presence and absence of C36 using a thermal shift assay and circular dichroism (CD) spectroscopy. Both methods displayed a single transition in the thermal melting curves, while the thermal unfolding of LMTK3 was irreversible due to protein aggregation. Our results revealed a minimal influence of the thermal stability of LMTK3 in the presence of C36, with the thermal shift assay indicating a decrease in the Tm of LMTK3 by −0.3 ± 0.1 ◦C (from 52.5 to 52.2 ◦C in the presence of C36, Figure 2C), and CD spectroscopy showing a small increase of Tm by 0.4 ± 0.1 ◦C in the presence of C36 (from 53.0 to 53.4 ◦C in the presence of C36, Figure 2D). However, it is important to emphasize that these experiments are not a direct measure of the binding affinity of C36 to LMTK3 due to the intrinsic limitations of these methods [29,30]. Therefore, a change in Tm value, whether significant or not, cannot be used to infer binding of our compound to LMTK3. Similarly, one cannot infer that our compound does not bind LMTK3 either. Overall, we conclude that C36 does not have any effect on the thermostability of LMTK3. Considering that LMTK3 in the absence of ATP and substrate is mainly in the inactive state, the results of the thermal shift assay and CD spectroscopy experiments suggest that C36 has a poor affinity for the inactive state of LMTK3 in solution. Taken together, these results indicate that C36 has no effect in the thermodynamic stability of inactive LMTK3, which contrasts with C28, which displayed a statistically significant stabilization of LMTK3 in the absence of ATP and/or substrate [2]. With the aim to present a putative model of LMTK3 with bound C36 that is in accor- dance with the above-mentioned results, we prepared a homology model of LMTK3 in the active state and carried out docking of C36. The inactive state of LMTK3 remains the only available X-ray structure where the ATP-binding site is occluded by the DYG-motif Tyr314 [2]. Considering the potentially low affinity of C36 for the LMTK3 inactive state and the relatively high sequence identity between the kinase domain of insulin receptor (IRK) and LMTK3 (37%) (Figure S3), we thus employed the X-ray structure of IRK in complex with ATP and a peptidic substrate (PDB ID: 3bu5) [31] as a template for modelling of LMTK3 in the active state. Our docking results suggest that C36 could bind adjacent to the ATP-binding site of LMTK3 and interact with the substrate as well (Figure 2E,F). This binding mode is also in accordance with the mixed-type inhibition profile of C36, as observed in the kinetic analysis. 2.3. C36 Exhibits Potent Anticancer Activity in Different Human Cancer Cell Lines We then investigated the potential use of C36 as an anticancer strategy by examining the viability of various BC cell lines in the presence of increasing concentrations of C36. As shown in Figure 3A, C36 was able to inhibit the growth of MCF7, T47D, and MDA-MB-231 BC cells, with IC50 values ranging from 16.19 µM to 18.38 µM. Following this, we submitted C36 to the Developmental Therapeutics Program (DTP) of the National Cancer Institute (NCI) and screened it against a panel of 60 human cancer cell lines [32]. Interestingly, our results showed that at a 10 µM dose, C36 inhibited all cancer cell lines by >40% (Figure 3B). Finally, we investigated the apoptotic properties of C36 in the aforementioned BC cell lines using annexin V and 7-AAD (7-amino-actinomycin D) staining. As shown in Figure 4A,B, following treatment for 96 h C36 exhibited an apoptotic effect at 20 µM in MCF7 and T47D BC cell lines, respectively. For MCF7 cells, apoptotic effects of C36 were also detected at 10 µM (Figure 4A). Lastly, no apoptosis was detected in MDA-MB-231 BC cell line, even when treated with 20 µM C36 (Figure 4C). Specifically, MCF7 and T47D cell lines displayed late apoptotic effects after 20 µM C36 treatment (Figure S4). Taken together, these results show that different BC cell lines display sensitivity to C36 treatment in terms of proliferation and apoptosis. Int. J. Mol. Sci. 2023, 24, 865 7 of 15 Figure 3. C36 impairs the viability of various human cancer cell lines. (A) Viability of BC cell lines treated with increasing concentrations of C36 for 72 h. The IC50 values are means from three independent experiments. (B) One-dose screening of C36 (10 µM; 24 h) on the NCI-60 panel of tumor cell lines. The percent growth of C36-treated cells is shown. Negative values represent lethality. Int. J. Mol. Sci. 2023, 24, x FOR PEER REVIEW 8 of 18 Int. J. Mol. Sci. 2023, 24, 865 8 of 15 Figure 4. Apoptotic effect of C36 on different human breast cancer cell lines. MCF7 (A), T47D (B), and MDA-MB-231 (C) were treated with increasing concentrations of C36 for 96 h. The percentages of apoptotic and dead cells were analyzed by Annexin V and 7-AAD staining. Results are expressed as means ± SEM; * p < 0.05, ** p < 0.01. Int. J. Mol. Sci. 2023, 24, x FOR PEER REVIEW 10 of 18 Figure 4. Apoptotic effect of C36 on different human breast cancer cell lines. MCF7 (A), T47D (B), and MDA-MB-231 (C) were treated with increasing concentrations of C36 for 96 h. The percentages Int. J. Mol. Sci. 2023, 24, 865 9 of 15 2.4. Pharmacological Properties of C36 The metabolic stability of C36 was also analyzed by incubating this drug with mouse hepatic microsomes (1 µM initial concentration, 0.25 mg protein/mL). Our results show that C36 was metabolized, relatively quickly, with a half-life value of 22 min and a high intrinsic clearance (Clint) value of 132 µL/min/mg (Figure S5A). This microsomal stability assay also produced two detectable putative metabolites, the most abundant of which is shown in Figure S5B. In addition, C36 showed low passive permeability in Caco-2 monolayer experiments (A > B; Papp of 5.1 × ≤10−6 cm/s), which could limit its absorption in in vivo studies, although its efflux ratio of 2.1 is not excessive (Figure S5A). 3. Discussion Despite the increasingly established role of LMTK3 in several cancer types and its central role in a number of well-described signaling pathways [8,10,15,20], currently there are no drugs in clinical trials targeting this oncogenic kinase. Here, we report a new tool compound, namely C36, which exhibits anticancer activity against a variety of cancer cell lines that is at least partly mediated by LMTK3 [2]. Based on our data, we propose that C36 not only competes with the LMTK3 ATP-binding site but also with the substrate-binding site in the kinase active state. Our molecular model suggests that C36 can interact with the ATP-binding site of LMTK3 and with the substrate as well, confirming the mixed-type inhibitory profile of C36. It is well established that most protein kinase inhibitors in clinical development mainly target the highly conserved ATP-binding site and thus are likely to have many off-target effects against kinases unrelated to diseases. Therefore, inhibitors like C36 that also possess competitive properties towards the kinase’s substrates are considered more selective and are expected to be promising therapeutic agents [33]. Recently, we reported the first tool compound (C28) against LMTK3 that displays anticancer activity in a variety of cancer cell lines and in vivo BC mouse models [2]. C36 and C28 displayed comparable affinities to LMTK3, as shown by MST. Importantly, C36 has a higher selectivity to purified human kinases when compared to C28 highlighting it as a promising candidate for drug development against LMTK3. Moreover, C36 demonstrates a strong antiproliferative effect against different cancer cell lines. Based on our results, C36 also exhibited apoptotic effects against BC cell lines (MCF7 and T47D) following a longer treatment exposure (96 h) and higher drug concentration (20 µM) than required to induce apoptosis following C28 treatment (72 h and 10 µM) [2]. The analysis of C36 metabolic stability was performed by incubating this drug with mouse hepatic microsomes. The results indicate C36 was metabolized relatively quickly. This cell-based system is not the ideal indicator for mouse in vivo studies; however, it is a commonly used steppingstone that can correlate well with liver microsomal stability in human and in vivo activity in mice [34–36]. Moreover, given the short half-life of C36, in vivo oral treatment at 0.25 mg protein/ml would be attainable only for a short period of time. In addition, given the short half-life of the drug (22 min) and its low passive permeability (A > B; Papp of 5.1 × ≤ 10−6 cm/s), work is currently underway, focusing on the design of C36 analogues and testing their effects in xenograft models of BC. Additional testing in other types of non-cancerous cell lines will provide further validation regarding C36 selectivity against cancer cell lines. Future work will focus on studying the specific molecular interactions between C36 and the kinase domain of LMTK3 by performing co-crystallization experiments. Investi- gating which amino acid residues are involved will likely shed light onto the mechanism of action of C36, furthering our knowledge. From our preliminary data presented in this manuscript, C36 decreases the rate of HSP27 phosphorylation by LMTK3 kinase do- main. Therefore, additional experiments might include co-crystallizing the entire complex (LMTK3-HSP27-C36) in order to understand, in more detail, the specific molecular inter- actions involved. Additional C36 analogues are also being synthesized to improve the binding affinity of the drug to LMTK3 and to increase its inhibitory properties against LMTK3 activity both in vitro (kinetic analysis) and in breast cancer cell lines. Int. J. Mol. Sci. 2023, 24, 865 10 of 15 Since LMTK3 has been shown to have a fundamental role in breast cancer progression and since there are no current drugs available targeting this oncogenic kinase, LMTK3 inhibitors could represent a valid alternative treatment to breast cancer patients. More specifically, LMTK3 inhibitors could be combined alongside aromatase inhibitors (AIs) as an alternative to treatment with CDK4/6 inhibitors to improve patient outcome in estrogen receptor positive (ER+) BC [37,38]. Likewise, given the aberrant expression of LMTK3 in triple negative breast cancer (TNBC) and studies showing that LMTK3 inhibition results in inhibition of TNBC cell proliferation, migration and invasion [14,16,17], the use of LMTK3 inhibitors could have beneficial effects for this clinically unmet category of BC patients. In addition, since the mechanism of emergence of endocrine and chemotherapy re- sistance in BC remains largely unclear [2], there is a need to treat these patients in a more focused way. Based on our previous studies, inhibition of LMTK3 appears to be implicated in re-sensitization of cells to tamoxifen and doxorubicin treatment [8,12–14]. Consequently, an LMTK3 drug could be used alongside established therapies to increase the sensitivity of tumors to treatment and/or potentially overcome resistance. Ultimately, this paper provides a steppingstone for the development and optimization of oral LMTK3 inhibitors, including C36, for use in clinical applications, either as a monotherapy or as a combination therapy in breast cancer. Finally, immunotherapy has become an established mainstay in cancer treatment and new drugs are being promptly developed for use in clinical settings [39]. Monoclonal antibodies (mAb), Ab-drug conjugates (ADCs), and cancer vaccines all represent different types of immunotherapies used in the treatment of BC [39] and other cancers. Currently, there are no immunotherapy programs specifically targeting LMTK3 in BC. However, the combination of immuno-therapeutic drugs (immune checkpoint inhibitor atezolizumab (Tecentriq®, Genentech, San Francisco, CA, USA)) and chemotherapeutic agents (nabPTX (Abraxane®, Celgene, Summit, NJ, USA)) has already been applied for the treatment of TNBC [34]. Based on this, novel LMTK3 inhibitors may be used in combination with immunotherapy and chemotherapy drugs [40] to improve the treatment of BC. 4. Materials and Methods 4.1. Cell Lines MCF7, T47D, and MDA-MB-231 cell lines were purchased from ATCC. MCF7 and MDA-MB-231 were maintained in low glucose DMEM (Sigma Aldrich, St. Louis, MO, USA, #D6046-500ML) supplemented with 10% FBS (Sigma Aldrich, #F7524-500ML) and 1% Penicillin/Streptomycin (Sigma Aldrich, #P0781-100ML). T47D cell line was maintained in RPMI-1640 medium (Sigma Aldrich, #R5886-500ML) supplemented with 10% FBS (Sigma Aldrich, #F7524-500ML) and 1% L-glutamine/Penicillin/Streptomycin solution (Sigma Aldrich, #G1146-100ML). 4.2. Cell Death and Apoptosis Cells were treated with increasing concentrations of C36 for 96 h. After collection, cells were stained with the Muse Annexin V Dead Cell Kit according to the manufacturer’s protocol (Millipore, Burlington, MA, USA, #MCH100105). Cells were then analyzed using the Muse Cell Analyzer (Millipore). Statistical analysis was performed with GraphPad Prism 8.0.1 software. In particular, one-way ANOVA analysis of variance with Dun- net post-hoc test for multiple comparison was performed. Statistical significance refers to the sample compared to the control (DMSO). p values < 0.05 were considered to be statistically significant. 4.3. Cell Viability Assay Cell viability assay was performed as previously described [41]. Mammalian cells were cultured at 3000 cells/well in 96-well plates (Corning, Corning, NY, USA, #3603). FDCP1 and LMTK3-transformed FDCP1 cells were plated at 5000 cells/well in 384-well plates (Aurora Biotechnologies, Poway, CA, USA, cat. no. 2030-10200). Cell viability was Int. J. Mol. Sci. 2023, 24, 865 11 of 15 assessed using the CellTiter-Glo luminescent cell viability assay (Promega, Madison, WI, USA, #G7572), as previously described [14]. Data analysis was performed with GraphPad Prism 8.01 software. In particular, we performed a nonlinear regression (curve fit) analysis by using the “dose–response–inhibition” model (log(inhibitor) vs. response–variable slope (four parameters)) to calculate the IC50 values as previously described [2]. 4.4. In Vitro Kinase Assay 32P γ-ATP in vitro kinase assays were performed in-house, as we have previously described [40]. The intensities of the bands on the autoradiograms have been quantified using ImageJ 1.53t software (Wayne Rasband and contributors, National Institutes of Health, Madison, WI, USA) and normalized to total protein levels based on Coomassie Blue stained membranes. DMSO has been used as a control. 4.5. Kinase Inhibitor Competition Binding Assay The selectivity profiling of C36 kinase inhibitor at 5 µM was analyzed using Discov- erX KINOMEscan competition binding assay against a panel of 403 kinases [25]. The KINOMEscan screening platform uses a novel active site-directed competition binding assay to measure interactions between a specific compound and approximately 400 kinases in a quantitative manner. The KINOMEscan assay does not require ATP and therefore reports true thermodynamic interaction affinities, instead of IC50 values, which usually depend on the ATP concentration. In particular, “hits” are detected by measuring the amount of kinase captured in test versus control samples by using qPCR, which is a method that detects the associated DNA label [25]. 4.6. Microscale Thermophoresis (MST) Purified LMTK3 protein was labelled with an NHS-647 red dye (NanoTemper Tech- nologies, München, Germany), following the manufacturer’s protocol. Serial dilutions of C36 (200 µM–0.61 nM) and C28 (200 µM–3.05 nM) in MST buffer (50 mM Tris pH 7.4, 150 mM NaCl, 10 mM MgCl2, 0.05% Tween 20, 2% DMSO) were mixed with 50 nM NHS- 647-labeled LMTK3 and loaded into standard glass capillaries (Monolith NT.115 Capillaries, NanoTemper Technologies). The final DMSO concentration was kept below 5%, as indi- cated by Ref. [26]. Thermophoresis analysis was performed over 20 sec on a Monolith NT.115 instrument (80% LED, 60% MST power) at 24 ◦C. The MST curves were fitted using NT Analysis software (NanoTemper Technologies) to obtain Kd values for binding. 4.7. Thermal Shift Assay A thermal shift assay was performed using Roche LightCycler 96 real-time polymerase chain reaction (RT-PCR) instrument, with excitation and emission wavelengths set to 533 and 572 nm, respectively. Solutions comprising 16 µL of 5.4 µM LMTK3 in 200 mM tris buffer (pH 8.0), 200 mM NaCl, and 4 µL of 50× SYPRO orange (Sigma-Aldrich, St. Louis, MO, USA) and 0.2 µL of either dimethyl sulfoxide (DMSO) or C36 in DMSO (final concentration of 10 µM C36, 1% (v/v) DMSO, 4.3 µM LMTK3, and 10× SYPRO orange). The temperature range spanned from 25 ◦C to 80 ◦C at a scan rate of 1 ◦C/min. Data analysis was performed in LightCycler 96 (v1.1, Roche, Mannheim, Germany) software using the melting curve analysis, and Tm values were determined as the first negative derivative of the fluorescence with respect to the temperature. 4.8. CD Spectroscopy CD spectroscopy was performed using a Jasco J-715 instrument (Jasco, Tokyo, Japan) equipped with a PTC-348 temperature control unit. Temperature increased from 20 ◦C to 90 ◦C at an increment of 1 ◦C/min, and data points were acquired every 0.2 ◦C by monitoring a wavelength of 230 nm. For thermal stability experiments, LMTK3 samples of 5.4 µM in 200 mM tris buffer (pH 8.0) and 200 mM NaCl were treated with either DMSO 0.4% (v/v) or 8.3 µM C36 in DMSO (0.4%) to a total volume of 120 µL in 0.1 cm cuvettes. Data analysis Int. J. Mol. Sci. 2023, 24, 865 12 of 15 was performed in GraphPad Prism 8.01 software by fitting data in the transition region to a Boltzmann sigmoidal. Apparent Tm values were determined as the point at which the transition was 50% complete. 4.9. Molecular Modelling of LMTK3 with Bound C36 The homology model of LMTK3 in the active state and the X-ray structure of the kinase domain of human insulin receptor (IRK), in complex with ATP and a peptidic-substrate (PDB ID: 3bu5) [31], were prepared using Modeller v9.24 (University of California San Francisco, San Francisco, CA, USA) [42]. The alignment is shown in the Supplementary Figure S3. The model with the lowest DOPE score was employed for docking of C36 using AutoDock v4.2 (The Scripps Research Institute, La Jolla, CA, USA) [43] with de- fault parameters, except for the number of docking rounds set to 100, and number of energy evaluations set to 10 million. Results were clustered with a rmsd tolerance of 2.0 Å (Supplementary Figure S6), and the top-ranked pose was selected as the putative bound conformation of C36 in the active state of LMTK3 (Figure 2F). The model of LMTK3 in complex with ATP and substrate (Figure 2E) was generated by superimposing the bound ATP and peptide substrate from the insulin receptor X-ray structure onto the model of active LMTK3, and after energy minimization with positional restraints on all Cα atoms (10 kcal × mol−1 × Å−2) using AMBER v16 (UCSF, San Francisco, CA, USA) [43]. 4.10. Caco-2 Permeability Assay The bi-directional Caco-2 cell permeability assay was performed as described in the BioFocus DPI Ltd. Standard Operating Procedure, ADME-SOP-49. Caco-2 cells (ECACC) were seeded onto 24-well Transwell plates at 2 × 105 cells per well and used in confluent monolayers after a 21-day culture at 37 ◦C under 5% CO2. Test and control compounds (propranolol, vinblastine), prepared in DMSO, were added (10 µM, 0.1% DMSO final, n = 2) to donor compartments of the Transwell plate assembly in assay buffer (Hanks balanced salt solution supplemented with 25 mM HEPES, adjusted to pH 7.4) for both apical to basolateral (A > B) and basolateral to apical (B > A) measurements. Incubations were performed at 37 ◦C, with samples removed from both donor and acceptor chambers at T = 0 and 1 h and compound analyzed by mass spectrometry (LC-MS/MS) including an analytical internal standard. Apparent permeability (Papp) values were determined from the relationship: Papp = [CompoundAcceptor T=end] × VAcceptor/([CompoundDonor T=0] × VDonor)/incubation time × VDonor/Area × 60 × 10−6 cm/s. V is the volume of each Transwell compartment (apical 125 µL, basolateral 600 µL), and concentrations are the relative MS responses for compound (normalized to internal standard) in the donor chamber before incubation and acceptor chamber at the end of the incubation. Area = area of cells exposed for drug transfer (0.33 cm2). Efflux ratios (Papp B > A/Papp A > B) were calculated for each compound from the mean Papp values in each direction. A finding of good permeability B > A, but poor permeability A > B, suggests that a compound is a substrate for an efflux transporter, such as P-glycoprotein. Lucifer Yellow (LY) was added to the apical buffer in all wells to assess viability of the cell layer. As LY cannot freely permeate lipophilic barriers, a high degree of LY transport indicates poor integrity of the cell layer and wells with a LY Papp > 10 × 10−6 cm/s were rejected. Note that an integrity failure in one well does not affect the validity of other wells on the plate. Compound recovery from the wells was determined from MS responses (normalized to internal standard) in donor and acceptor chambers at the end of incubation compared to response in the donor chamber pre-incubation. Recoveries < 50% suggest compound Int. J. Mol. Sci. 2023, 24, 865 13 of 15 solubility, stability, or binding issues in the assay, which may reduce the reliability of a result. 4.11. Compound Stability in Mouse Hepatic Microsomes Microsomal stability assays were performed as described in the BioFocus DPI Ltd. Standard Operating Procedure, ADME-SOP-84, using pooled hepatic microsomes from mouse (Xenotech/1210302, Kansas City, KS, USA). Test and control compounds (dex- tromethorphan and midazolam), prepared in DMSO, were incubated at an initial con- centration of 1 µM (0.25% DMSO final, n = 2) with microsomes (0.25 mg protein/ml) at 37 ◦C in the presence and absence of the cofactor, NADPH (1 mM). Aliquots were re- moved at 0, 5, 10, 20, and 40 min for termination of reactions and compound extraction with acetonitrile containing an analytical internal standard. Samples were centrifuged and the supernatant fractions were analyzed for parent compound by mass spectrometry (LC-MS/MS). The amount of compound remaining (expressed as %) was determined from the MS re- sponse in each sample relative to that in the T = 0 samples (normalized for internal standard). Ln plots of the % remaining were used to determine the half-life for compound disappearance using the relationship: t1/2 (min) = −0.693/λ, where λ is the slope of the Ln % remaining vs. time curve. The in vitro intrinsic clearance (CLint) (µL/min/mg microsomal protein) was calculated 2 (min) × (1/mg of microsomal protein/ml) × 1000. using the formula: CLint = 0.693 × 1/t 1 4.12. NCI-60 Human Tumor Cell Line Screen The NCI-60 panel of tumor cell lines utilizes a variety of different cancerous cell lines to identify and characterize novel compounds that inhibit the growth or exert a lethal effect on these tumor cells. This screen encompasses 60 cell lines from leukemia, melanoma, and cancers of the colon, brain, ovary, lung, prostate, breast, and kidney. In our case, 60 different cell lines were treated with 10 µM C36 for 24 h. Following this, growth inhibition and lethality were measured [32]. Supplementary Materials: The following supporting information can be downloaded at: https:// www.mdpi.com/article/10.3390/ijms24010865/s1. Author Contributions: G.G. conceived the project, planned and oversaw the execution of all the work. A.A., A.L.B., V.V., M.S., C.C. and A.D. performed the biochemical and cell-based experiments. A.P. performed the biophysical experiments, molecular modelling and contributed to the interpretation of the results and writing the respective parts. A.A., A.L.B., A.D., P.P., E.J.M., J.S. (Justin Stebbing), J.S. (John Spencer) and C.G. helped with the preparation of the figures. All authors contributed to the writing and editing of this manuscript. All authors have read and agreed to the published version of the manuscript. Funding: This work was supported by Action Against Cancer (Grant number: G1868). Institutional Review Board Statement: Not applicable. Informed Consent Statement: Not applicable. Data Availability Statement: All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Materials. Additional data related to this paper may be requested from the authors. Acknowledgments: We thank the members of our lab and other collaborators for the useful and constructive discussions. Conflicts of Interest: Georgios Giamas is editor of Cancer Gene Therapy and founder/chief scientific officer of Stingray Bio. JS’s COI can be found at: https://www.nature.com/onc/editors (accessed on 10 October 2022); none are relevant here. No other conflicts are declared. Int. J. Mol. Sci. 2023, 24, 865 14 of 15 References 1. 2. 3. 4. 5. 6. Cicenas, J.; Zalyte, E.; Bairoch, A.; Gaudet, P. Kinases and Cancer. Cancers 2018, 10, 63. [CrossRef] [PubMed] Ditsiou, A.; Cilibrasi, C.; Simigdala, N.; Papakyriakou, A.; Milton-Harris, L.; Vella, V.; Nettleship, J.E.; Lo, J.H.; Soni, S.; Smbatyan, G.; et al. The structure-function relationship of oncogenic LMTK3. Sci. Adv. 2020, 6, eabc3099. [CrossRef] [PubMed] Bhullar, K.S.; Lagarón, N.O.; McGowan, E.M.; Parmar, I.; Jha, A.; Hubbard, B.P.; Rupasinghe, H.P.V. Kinase-targeted cancer therapies: Progress, challenges and future directions. Mol. Cancer 2018, 17, 48. [CrossRef] Blume-Jensen, P.; Hunter, T. Oncogenic kinase signalling. Nature 2001, 411, 355–365. [CrossRef] [PubMed] Krause, D.S.; Van Etten, R.A. Tyrosine kinases as targets for cancer therapy. N. Engl. J. Med. 2005, 353, 172–187. [CrossRef] Ditsiou, A.; Gagliano, T.; Samuels, M.; Vella, V.; Tolias, C.; Giamas, G. The multifaceted role of lemur tyrosine kinase 3 in health and disease. Open Biol. 2021, 11, 210218. [CrossRef] 7. Montrose, K.; Kobayashi, S.; Manabe, T.; Yamamoto, T. Lmtk3-KO Mice Display a Range of Behavioral Abnormalities and Have 8. 9. an Impairment in GluA1 Trafficking. Neuroscience 2019, 414, 154–167. [CrossRef] Giamas, G.; Filipovi´c, A.; Jacob, J.; Messier, W.; Zhang, H.; Yang, D.; Zhang, W.; Shifa, B.A.; Photiou, A.; Tralau-Stewart, C.; et al. Kinome screening for regulators of the estrogen receptor identifies LMTK3 as a new therapeutic target in breast cancer. Nat. Med. 2011, 17, 715–719. [CrossRef] Jacob, J.; Favicchio, R.; Karimian, N.; Mehrabi, M.; Harding, V.; Castellano, L.; Stebbing, J.; Giamas, G. LMTK3 escapes tumour suppressor miRNAs via sequestration of DDX5. Cancer Lett. 2016, 372, 137–146. [CrossRef] 10. Li, Z.; Wu, J.; Ji, M.; Shi, L.; Xu, B.; Jiang, J.; Wu, C. Prognostic role of lemur tyrosine kinase 3 in postoperative gastric cancer. Mol. 11. 12. 13. 14. Clin. Oncol. 2014, 2, 756–760. [CrossRef] Shi, H.; Wu, J.; Ji, M.; Zhou, Q.; Li, Z.; Zheng, X.; Xu, B.; Deng, H.; Zhao, W.; Wu, C.; et al. Serum lemur tyrosine kinase 3 expression in colorectal cancer patients predicts cancer progression and prognosis. Med. Oncol. 2013, 30, 754. [CrossRef] [PubMed] Stebbing, J.; Filipovic, A.; Ellis, I.O.; Green, A.R.; D’Silva, T.R.; Lenz, H.J.; Coombes, R.C.; Wang, T.; Lee, S.C.; Giamas, G. LMTK3 expression in breast cancer: Association with tumor phenotype and clinical outcome. Breast Cancer Res. Treat. 2012, 132, 537–544. [CrossRef] Stebbing, J.; Filipovic, A.; Lit, L.C.; Blighe, K.; Grothey, A.; Xu, Y.; Miki, Y.; Chow, L.W.; Coombes, R.C.; Sasano, H.; et al. LMTK3 is implicated in endocrine resistance via multiple signaling pathways. Oncogene 2013, 32, 3371–3380. [CrossRef] [PubMed] Stebbing, J.; Shah, K.; Lit, L.C.; Gagliano, T.; Ditsiou, A.; Wang, T.; Wendler, F.; Simon, T.; Szabó, K.S.; O’Hanlon, T.; et al. LMTK3 confers chemo-resistance in breast cancer. Oncogene 2018, 37, 3113–3130. [CrossRef] 15. Wakatsuki, T.; LaBonte, M.J.; Bohanes, P.O.; Zhang, W.; Yang, D.; Azuma, M.; Barzi, A.; Ning, Y.; Loupakis, F.; Saadat, S.; et al. Prognostic role of lemur tyrosine kinase-3 germline polymorphisms in adjuvant gastric cancer in Japan and the United States. Mol. Cancer Ther. 2013, 12, 2261–2272. [CrossRef] [PubMed] 16. Xu, Y.; Zhang, H.; Lit, L.C.; Grothey, A.; Athanasiadou, M.; Kiritsi, M.; Lombardo, Y.; Frampton, A.E.; Green, A.R.; Ellis, I.O.; et al. The kinase LMTK3 promotes invasion in breast cancer through GRB2-mediated induction of integrin β1. Sci. Signal. 2014, 7, ra58. [CrossRef] 17. Xu, Y.; Zhang, H.; Nguyen, V.T.; Angelopoulos, N.; Nunes, J.; Reid, A.; Buluwela, L.; Magnani, L.; Stebbing, J.; Giamas, G. LMTK3 Represses Tumor Suppressor-like Genes through Chromatin Remodeling in Breast Cancer. Cell Rep. 2015, 12, 837–849. [CrossRef] [PubMed] 18. Zhang, K.; Chen, L.; Deng, H.; Zou, Y.; Liu, J.; Shi, H.; Xu, B.; Lu, M.; Li, C.; Jiang, J.; et al. Serum lemur tyrosine kinase-3: A novel biomarker for screening primary non-small cell lung cancer and predicting cancer progression. Int. J. Clin. Exp. Pathol. 2015, 8, 629–635. 19. Lu, L.; Yuan, X.; Zhang, Q.; Zhang, H.; Shen, B. LMTK3 knockdown retards cell growth and invasion and promotes apoptosis in thyroid cancer. Mol. Med. Rep. 2017, 15, 2015–2022. [CrossRef] 20. Klug, L.R.; Bannon, A.E.; Javidi-Sharifi, N.; Town, A.; Fleming, W.H.; VanSlyke, J.K.; Musil, L.S.; Fletcher, J.A.; Tyner, J.W.; Heinrich, M.C. LMTK3 is essential for oncogenic KIT expression in KIT-mutant GIST and melanoma. Oncogene 2019, 38, 1200–1210. [CrossRef] Jiang, T.; Lu, X.; Yang, F.; Wang, M.; Yang, H.; Xing, N. LMTK3 promotes tumorigenesis in bladder cancer via the ERK/MAPK pathway. FEBS Open Bio 2020, 10, 2107–2121. [CrossRef] [PubMed] 21. 22. Xu, Z.; Qi, X.; Zhang, X.; Yu, L. Preoperative serum LMTK3 as a novel biomarker in non-small cell lung cancer. Tumour Biol. 2014, 35, 5007–5011. [CrossRef] [PubMed] 23. Wang, C.; Yang, M.; Gu, X.; Gu, Y. Lemur tyrosine kinase-3 (LMTK3) induces chemoresistance to cetuximab in colorectal cancer via the ERK/MAPK pathway. Bioengineered 2021, 12, 6594–6605. [CrossRef] [PubMed] 24. Kinase Screen. Available online: https://www.kinase-screen.mrc.ac.uk/ (accessed on 1 September 2022). 25. Eurofins DiscoverX. Available online: https://www.discoverx.com/home (accessed on 1 September 2022). 26. Rainard, J.M.; Pandarakalam, G.C.; McElroy, S.P. Using Microscale Thermophoresis to Characterize Hits from High-Throughput Screening: A European Lead Factory Perspective. SLAS Discov. 2018, 23, 225–241. [CrossRef] 27. Entzian, C.; Schubert, T. Studying small molecule-aptamer interactions using MicroScale Thermophoresis (MST). Methods 2016, 97, 27–34. [CrossRef] Int. J. Mol. Sci. 2023, 24, 865 15 of 15 28. Magnez, R.; Bailly, C.; Thuru, X. Microscale Thermophoresis as a Tool to Study Protein Interactions and Their Implication in Human Diseases. Int. J. Mol. Sci. 2022, 23, 7672. [CrossRef] 29. Gao, K.; Oerlemans, R.; Groves, M.R. Theory and applications of differential scanning fluorimetry in early-stage drug discovery. Biophys. Rev. 2020, 12, 85–104. [CrossRef] 30. Khrapunov, S. Circular dichroism spectroscopy has intrinsic limitations for protein secondary structure analysis. Anal. Biochem. 2009, 389, 174–176. [CrossRef] 31. Wu, J.; Tseng, Y.D.; Xu, C.F.; Neubert, T.A.; White, M.F.; Hubbard, S.R. Structural and biochemical characterization of the KRLB region in insulin receptor substrate-2. Nat. Struct. Mol. Biol. 2008, 15, 251–258. [CrossRef] 32. DTP Developmental Therapeutics Program. Available online: https://dtp.cancer.gov/discovery_development/nci-60/ (accessed on 1 September 2022). 33. Han, K.C.; Kim, S.Y.; Yang, E.G. Recent advances in designing substrate-competitive protein kinase inhibitors. Curr. Pharm. Des. 2012, 18, 2875–2882. [CrossRef] 34. Perryman, A.L.; Stratton, T.P.; Ekins, S.; Freundlich, J.S. Predicting Mouse Liver Microsomal Stability with “Pruned” Machine Learning Models and Public Data. Pharm. Res. 2016, 33, 433–449. [CrossRef] [PubMed] 35. Quintieri, L.; Fantin, M.; Palatini, P.; De Martin, S.; Rosato, A.; Caruso, M.; Geroni, C.; Floreani, M. In vitro hepatic conversion of the anticancer agent nemorubicin to its active metabolite PNU-159682 in mice, rats and dogs: A comparison with human liver microsomes. Biochem. Pharmacol. 2008, 76, 784–795. [CrossRef] [PubMed] 37. 36. Palmer, B.D.; Thompson, A.M.; Sutherland, H.S.; Blaser, A.; Kmentova, I.; Franzblau, S.G.; Wan, B.; Wang, Y.; Ma, Z.; Denny, W.A. Synthesis and Structure−Activity Studies of Biphenyl Analogues of the Tuberculosis Drug (6S)-2-Nitro-6-{[4- (trifluoromethoxy)benzyl]oxy}-6,7-dihydro-5H-imidazo[2,1-b][1,3]oxazine (PA-824). J. Med. Chem. 2010, 53, 282–294. [CrossRef] Slamon, D.J.; Neven, P.; Chia, S.; Fasching, P.A.; De Laurentiis, M.; Im, S.A.; Petrakova, K.; Bianchi, G.V.; Esteva, F.J.; Martín, M.; et al. Phase III Randomized Study of Ribociclib and Fulvestrant in Hormone Receptor-Positive, Human Epidermal Growth Factor Receptor 2-Negative Advanced Breast Cancer: MONALEESA-3. J. Clin. Oncol. Off. J. Am. Soc. Clin. Oncol. 2018, 36, 2465–2472. [CrossRef] [PubMed] Sledge, G.W., Jr.; Toi, M.; Neven, P.; Sohn, J.; Inoue, K.; Pivot, X.; Burdaeva, O.; Okera, M.; Masuda, N.; Kaufman, P.A.; et al. MONARCH 2: Abemaciclib in Combination With Fulvestrant in Women With HR+/HER2- Advanced Breast Cancer Who Had Progressed While Receiving Endocrine Therapy. J. Clin. Oncol. Off. J. Am. Soc. Clin. Oncol. 2017, 35, 2875–2884. [CrossRef] 39. Hashemzadeh, N.; Dolatkhah, M.; Adibkia, K.; Aghanejad, A.; Barzegar-Jalali, M.; Omidi, Y.; Barar, J. Recent advances in breast 38. cancer immunotherapy: The promising impact of nanomedicines. Life Sci. 2021, 271, 119110. [CrossRef] 40. Giamas, G.; Hirner, H.; Shoshiashvili, L.; Grothey, A.; Gessert, S.; Kühl, M.; Henne-Bruns, D.; Vorgias, C.E.; Knippschild, U. Phosphorylation of CK1delta: Identification of Ser370 as the major phosphorylation site targeted by PKA in vitro and in vivo. Biochem. J. 2007, 406, 389–398. [CrossRef] Sali, A.; Blundell, T.L. Comparative protein modelling by satisfaction of spatial restraints. J. Mol. Biol. 1993, 234, 779–815. [CrossRef] 41. 42. Morris, G.M.; Huey, R.; Lindstrom, W.; Sanner, M.F.; Belew, R.K.; Goodsell, D.S.; Olson, A.J. AutoDock4 and AutoDockTools4: Automated docking with selective receptor flexibility. J. Comput. Chem. 2009, 30, 2785–2791. [CrossRef] 43. Case, D.A.; Cheatham, T.E., 3rd; Darden, T.; Gohlke, H.; Luo, R.; Merz, K.M., Jr.; Onufriev, A.; Simmerling, C.; Wang, B.; Woods, R.J. The Amber biomolecular simulation programs. J. Comput. Chem. 2005, 26, 1668–1688. [CrossRef] Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
10.1098_rsos.220808
royalsocietypublishing.org/journal/rsos Review Cite this article: Dienes Z. 2023 The credibility crisis and democratic governance: how to reform university governance to be compatible with the nature of science. R. Soc. Open Sci. 10: 220808. https://doi.org/10.1098/rsos.220808 Received: 17 June 2022 Accepted: 6 January 2023 Subject Category: Science, society and policy Subject Areas: psychology Keywords: university governance, democracy, open democracy, open science Author for correspondence: Zoltan Dienes e-mail: z.dienes@sussex.ac.uk The credibility crisis and democratic governance: how to reform university governance to be compatible with the nature of science Zoltan Dienes School of Psychology, University of Sussex, Brighton, UK ZD, 0000-0001-7454-3161 To address the credibility crisis facing many disciplines, change is needed at the institutional level. Science will only function optimally if the culture by which it is governed becomes aligned with the way of thinking required in science itself. The paper suggests a series of graduated reforms to university governance, to radically reform the operation of universities. The reforms are based on existing established open democratic practices. The aim is for universities to become consistent with the flourishing of science and research more generally. 1. Introduction Many areas of science have been facing difficulties in credibility with a sense that the scientific process is not as healthy as it could be. There is low replicability of studies ([1], chapter 2), possibly associated with a failure of a field to self-correct [2]; and at the same time, there is a hyper-competitive culture aligned with perverse incentives that may reward substandard science [3]. The solutions to this credibility crisis will surely involve multiple levels of reform [4]. Specifically, cultural change is needed at the institutional level, which is the level that this paper addresses. Initially, a simple account of how knowledge grows is presented. Then a brief historical overview is provided of the growth of knowledge and its relation to how decisions are made in the broader social context. Classical Athens is taken as an example of matching between the governance of the society as a whole and the growth of knowledge. Next, the state of governance of UK universities is considered. Finally, open democratic practices that are Athenian-like and that have been tested in politics are considered for how a university might be governed in an open democratic way in order for the university to align itself with the way knowledge grows. © 2023 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. 2 r o y a l s o c i e t y p u b i . l i s h n g o r g / j o u r n a l / r s o s R . S o c . O p e n S c i . 1 0 : 2 2 0 8 0 8 2. How does knowledge grow? Popper [5] asked ‘How does knowledge grow?’ His answer in general, is by trial and error; try ideas out and see what works; reject those that do not work. How could that process be enhanced? According to Popper, traditionally schools of thought were passed down from master to disciple with an aim to impart a doctrine pure and unchanged. There is in such a tradition a hierarchy with roles filled by people by virtue of their characteristics: master and disciple. This is a good way of conserving knowledge as it is, but not for promoting the growth of knowledge. But consider an alternative, which Popper calls the critical tradition: the master says ‘Here is my idea to solve a problem; can you improve on it?’ It can be difficult to encourage others to do better than oneself. Thus, the critical tradition as a second-order tradition that is passed on from mentor to pupil has to be constantly fought for: businesses, religious leaders, politicians and even academics will regularly try to stamp out criticism of their ideas. The critical tradition occurs when there is a culture of considering arguments for their own sake, with small regard for the authority of the person stating them. That is, in a critical tradition roles are fluid, and what is important is the quality of ideas. Taking part in a critical tradition may be psychologically easier when people see ideas, theories and data as having an objective existence apart from themselves, with properties that must be discovered; this is what it means to consider arguments for their own sake. Then people can refute a theory without thinking they themselves have been harmed [6]; cf. also [7]. Let an open society be a society in which such a critical tradition is encouraged (cf. [8]). Let democracy be an open society in which there are institutions that encourage a critical tradition independent of any individual. Thus an autocratic ruler may promote an open society if that was the sort of thing that ruler liked; but the society would not as such be a democracy, because the existence of the open society would depend on the whims of a particular ruler. 3. Lessons from history There is intriguing evidence of a historical relationship between the growth of knowledge and the existence of an open society, especially democracy. Popper [5] suggests how Thales, around 600 BCE, proposed a natural principle for how the world works, only for his apparent student, Anaximander to come up with something logically better, starting a critical tradition. For the next several hundred years in Athens, there was an astonishing flourishing of knowledge, in mathematics, astronomy, history, psychology and medicine. Knowledge and open critical discussion continued into the extended Greek and Roman world for some time AD, but the critical tradition gradually withered. For example, the Epidemics of the Hippocratic corpus (fifth century BCE) mainly indicated how their treatments failed ([9], chapter 5) in contrast to case histories from later centuries (consider the numerous triumphs Galen, fl. second century AD, described in outwitting other doctors ([10], e.g. chapter 7); after Galen, there were no students of his who tried to produce better solutions, at least not the for many centuries). Almost exactly contemporaneously with the rise and decline of flourishing of knowledge, there was a rise and decline in Athenian-style democracy. In Athens itself, the initial reforms of Solon (600 BCE) were strengthened by Cleisthenes (coming up to 500 BCE), then after further gradual refinements, Athenian democracy was formally ended in 332 BC by the conquest of Alexander the Great. However, as seen as part of an ecology of about 1000 Greek states, democracy robustly continued well into the second century BCE, with the number of democracies actually increasing for some time [11]. Over several hundred years from before until after the classic period, these Greek states showed shifting mixtures of democracies and elite control. Based on archaeological and historical evidence, Ober argued that cultural flourishing (and wealth) followed not the rise of conservative institutions but tracked the development of democratic rule-egalitarian institutions. There was both an explosion of knowledge and the implementation of a robust long-lasting democracy during the time of classic Athens, followed after a delay by a slow stagnation in knowledge growth. At other times and places, when there was an open society, knowledge also flourished—even without democracy. Saliently, from 800–1400 Arabic science drew from Greeks and Indians, and made major progress [12]. There was not democracy, but absolute rule by Abbasid caliphs. Any given caliph might support an open society (e.g. initially especially Al-Ma’mun), in the sense of encouraging critical discussion of ideas (e.g. Al-Khalili [12], pp. 67–68). When a caliph, or others in positions of authority, supported the free exploration of ideas, knowledge grew; when a subsequent caliph was more conservative, learning shrank (e.g. [12], p. 230, cf. also p. 194). The relation of general openness during this time and scientific progress bears further investigation. 3 r o y a l s o c i e t y p u b i . l i s h n g o r g / j o u r n a l / r s o s R . S o c . O p e n S c i . 1 0 : 2 2 0 8 0 8 The critical tradition sustained in the Arab world eventually found its way to the medieval Italian states. From about 1100, these states were already exploring democratic governance, using mixtures of lot and election (e.g. [13,14]). The assembly politics often characterizing decision making in those states entailed considering arguments for their own sake even if the person voicing them may be of low rank. Thus, the ground was laid for exploring new ideas; and later there is indeed the outpouring of new forms in art, invention and later science. Ferris [15] picks up the story of the intertwining of democratic values and the growth in science in the 1600s onwards, noting how science developed in the most liberal countries, and how conversely scientists were key people pushing for democratic change. A further explosive growth in knowledge occurred alongside the progressive rejection of authoritarian political values and the development of liberal values, from the enlightenment onwards. Yet the growth of knowledge has also occurred when society as a whole was not especially open, and conversely, science often did not occur where there were even democratic institutions. An example of the first point is the steady growth of knowledge throughout Chinese history. Up until about 1400, China was hundreds, and sometimes thousands of years, ahead of Europe in technological development. Needham [16] spent decades documenting how many innovations came from China, for example, China pioneered inoculations ( possibly tenth century and certainly by 16th); China developed mechanical clocks six centuries before Europe. True, China throughout its history has valued scholars very highly and had a well-organized civil service based on exams rather than (explicitly) on pedigree. But this time the emperor in principle had the final say on any matter (including an edict that held from 653 forbidding the private possession of astronomical instruments ([17], p. 228)), and there is little evidence of democratic processes (except briefly for a period in the Zhou dynasty, 1050–221 BCE ([14], p. 150)). There was the constant development of technology—but not the explosive development of science. Needham asked why did modern science not develop in China and only in Europe? Why did knowledge of the physical world grow steadily in China, and yet not explode like it did in Europe? Needham suggested that ‘There was no modern science in China because there was no democracy’ [16, p. 152]. Needham pointed out that science is indifferent to who makes the argument; thus ‘these civilizations which have developed an … exaggerated respect for teachers, will have to modify it’ [16, p. 140]. In sum, ‘there is a real kinship between the scientific ‘ namely, skepticism, anti-authoritarianism, not letting others mind and the democratic mentality, decide on aims or assessment of evidence, ‘a give and take, a live and let live attitude’ [16, p. 143]. (For a review of other hypotheses by Needham and others to address this question, see [17].) throughout Conversely, Stasavage [14] and Graeber & Wengrow [18] present historical, archaeological and anthropological evidence for democracy, in the sense of decision by assemblies, being a common solution to the problem of political governance, and often in large-scale societies. If that is true, why did science not emerge multiple times? Debates in assemblies, to the extent criticism of any individual’s views are welcome and not just tolerated, promote a critical tradition. And a critical tradition allows knowledge to grow, but it need not be specifically scientific knowledge. Graeber & Wengrow argue for the political sophistication of the indigenous Americans, whose skills were honed in assembly politics, and who could hold their own if not best European intellectuals of the 1600s in political and social arguments. Indeed, according to Graeber and Wengrow, native American arguments may have, unacknowledged, transformed the course of European intellectual history. Consider also the democratic politics in India at the time of the start of Buddhism (Mahaparinibbana sutta in [19]), which occurred contemporaneous with the development of ideas about the mind which continue to influence modern thinking (e.g. [20]). Clearly there is no deterministic relation between science and democracy; but there is synergy. The lesson to draw from history is that when science and democracy occur together, science is facilitated. Science might happen without democracy, and democracy without science—but put them together and see what happens. 4. Athenian-style democracy Athenian democracy emphasized the selection of arguments that could in principle be provided by anyone, rather than primarily the selection of people according to their traits as such [21]. Let us consider some details, because they will be useful for considering modern reform (for a readable overview with the specific aim of implementing the principles in modern corporate institutional governance, see [22]; see also [23,24]. Athenians were divided into 10 tribes pooling different types of citizens (urban, rural, coastal) into in-groups to unite pre-existing cultural divisions (consider the use of houses or colleges in some universities). Day-to-day business was organized by the Council of 500 classical Athens council of 500 (Boule) agenda sets agenda and implements decisions general administration and supervision of the system decisions law courts LOT Tribe 1 Tribe 2 popular assembly (Ecclesia) – all citizens (30k) discuss and approve motions nomothetai – checks new laws for consistency with old LOT Tribe 10 . . . (the citizens) 4 r o y a l s o c i e t y p u b i . l i s h n g o r g / j o u r n a l / r s o s R . S o c . O p e n S c i . 1 0 : 2 2 0 8 0 8 Figure 1. Sketch of classic Athenian democracy. In Athens circa fourth century BC, eligible citizens were assigned by lot, i.e. randomly, to the main governing bodies and the law courts. Of about 1000 posts, 90% were determined by lot and the rest by election. The majority of citizens would have spent at least a year at some point in their life in the main governing body, the boule. Some form of such governance lasted for hundreds of years in Athens and other Greek states. (the boule), consisting of 50 eligible citizens selected by random lot from each tribe for a period of a year. Each month the 50 from one tribe would set the agenda for the business of the day, and prescribe the meetings of the Assembly (which we will come to). Other duties include overseeing the work of all officials and financial oversight. Each day officials from the other tribes in the Council make decisions concerning the agenda. The leader of the Council was selected daily from the tribe in charge of agenda-setting that month: one person was the nominal head of state for one day alone! In sum, central decision-making was deliberately integrated over many people, chosen by lot from the citizenship. Roughly weekly, the citizens, or the subset who turned up (maybe a fifth of all citizens on any one occasion), formed the Assembly (ecclesia). The agenda was set by Council, and in principle any citizen could speak on any motion, before a vote to decide each proposal. One further institution is worth mentioning: the nomothetai. This was a panel formed by lot from eligible citizens to reflectively consider arguments for and against proposed general laws before a final decision was actually made to accept them. We will draw on this important institution later. Figure 1 for a sketch of the overall structure of Athenian democracy. 5. The state of UK universities Coming into this millennium, many UK universities were democracies at the school (i.e. faculty) level. School meetings were decision-making bodies, with decisions made usually by majority vote. Decisions fed up to the central university level. Senior management at the central level then had to do their best to render coherent at the institutional level the way these parts fitted together. Subsequently, in about the first decade of the millennium, many UK universities moved to more or less complete non-democratic top-down control, where senior management made decisions, and the Deans of schools were to work out how to implement those decisions to management’s satisfaction. The Dean rather than faculty had final say at school level. The Vice-Chancellor (VC) was appointed by Council with no involvement by faculty at large (with the members of Council being appointed by Council). These reforms occurred in the tradition of ‘New Public Management’, a philosophy of public sector management which started to be implemented in the Thatcher years, and has become dominant in the UK in its higher education policy [25–27]; see [28], for a history of UK education sympathetic to this approach). In many universities in the UK, top-down control is exerted by management, who are perceived to be a class separate from academics with no real accountability [29]. These managers may apply strong pressure on academics to achieve key performance indicators. The result is managerialism: the worth of an academic (for getting jobs, promotion, respect) is determined by set performance goals, typically including getting grants and publishing in high-impact-factor journals. The first duty of a researcher is not high-quality research by their own judgement, but to fulfil the agenda of an increasingly large class of managers (whose agenda is to boost e.g. world rankings and other metrics). It is not obvious 5 r o y a l s o c i e t y p u b i . l i s h n g o r g / j o u r n a l / r s o s R . S o c . O p e n S c i . 1 0 : 2 2 0 8 0 8 in project management the failure of complete top-down control that top-down control by senior managers is the best way of dealing with a rapidly changing and unpredictable environment that is necessarily what the interface of knowledge and ignorance consists in. For in unpredictable environments for aid agencies, see Honig [30]; and its failure in the secondary education sector, Honig [31]. Honig [31] presents evidence that top-down management with close monitoring and control backfires when those managed are already there because they want to be: such management produces both selection effects, the loss of good people, and motivational changes in those that remain (for the latter see also [32], chapter 5). Martin, reviewing the history of UK university governance with an eye on the management literature, asks, ‘Why, when the management literature of the last two decades has stressed the benefits of flatter organizational structures, of decentralization and local initiative … have many universities been intent on moving in precisely the opposite direction of greater centralization with a more hierarchical, organizational structure, top-down management … and ever more cumbersome and intrusive procedures?’ [27 p. 7].1 And of specific relevance to science, Xu et al. [38] found that scientific teams with a flat rather than hierarchical structure produced more novel ideas and a higher long-term citation impact. farms [40]. When an expert needs to exercise judgement Attempting to incentivize ‘performance’ when what really counts as performance cannot be easily measured—as is the case in attempting to understand the unknown—will generally backfire [30,39]. Simple-minded performance targets famously backfire even for simple problems. Consider the rat tails of Hanoi. At the beginning of the twentieth century the French colonial rulers of Hanoi wanted the city rid of rats. To bring on board the local population, money was offered for each rat tail delivered as proof of a killed rat. Yet the rats only increased. It turned out the locals, being resourceful, set up in an unpredictable environment, rat sustained incentives to maximize something just because it is measurable, typically distort best practice. The best person to judge strategy and tactics for dealing with a research problem is the researcher themselves. Incentivizing them to, for example, apply for grants, will distort how time is best spent. If a researcher needs a grant to further their research, they have no need of a manager to tell them to get a grant. Conversely, if their time would be better spent writing up that file drawer of papers, incentivizing them to apply for grants only promotes inefficiency. But the problem may be far worse than this. Key performance indicators filtering down from senior management as pressure on individual researchers may not just waste time—it may damage the integrity of science. It may produce, in effect, rat farms. In simulation studies, Smaldino & McElreath [41] consider a population of different laboratories who differ in the degree of p-hacking they engage in while competing for grants. Given reasonable assumptions, those laboratories that p-hack the most are most successful in obtaining grant money— and thereby produce more progeny laboratories (via the PhD students and post docs they train) which carry on the same culture as the parent laboratory. The current environment of intense competition for grant money plausibly promotes low-integrity science (while of course being consistent with some successful laboratories that do value rigour, as shown, for example, by their commitment to open science). Smaldino et al. [42] consider solutions. In their simulations, the way to break the effect of the selection of p-hackers, was to borrow an idea from classical Athens, selection by lot: award grants by random lottery for those submissions that passed a minimal standard of methodological rigour. (Such a procedure not only can restore integrity, it also ensures a lack of discrimination based on gender, race, or institution in grant allocation.2) While the concept of p-hacking is not relevant to all disciplines, the argument plausibly generalizes for any similar process where cutting corners to the detriment of the quality of the research nonetheless allows outcomes that look convincing. Satisfying key performance indicators typically involves publishing in journals with high impact factors. Often high-impact-factor journals are run by for-profit companies charging high publishing fees. A principle of how science should function is that anyone should be able to contribute according to only the quality of their contribution. High publishing fees mean that only those able to join a rich 1Leaf [33] describes widely varying levels of democratic governance across different US Universities, yet indicates the direction of change is toward less democracy. For the rise of managerialism in US Universities, see Ginsberg [34]. For the variability in extent of democratic governance across Europe, but the same direction of change as the UK, see de Boer & File [35]. And for the headlong embrace of managerialism by Australian Universities, see Biggs [36] and Hil [37]. 2For the greater efficiencies that would be achieved by grant lotteries, see Gross & Bergstrom [43]. For a different solution, see Brette [44]. 6 r o y a l s o c i e t y p u b i . l i s h n g o r g / j o u r n a l / r s o s R . S o c . O p e n S c i . 1 0 : 2 2 0 8 0 8 person’s club, at an intuitional or individual functioning of science [45]. And it may be worse than that. level, can contribute. This undermines the proper Does pressure to publish in high-impact journals produce p-hacked and less reproducible research? Direct evidence for pressure to publish producing poor quality publications as a general relationship is not yet in (consider the attempt by [46]). But there are some specific connections. Cash bonuses for publishing in high-impact-factor journals, as has been practised in Australia and China, is associated with retractions of papers [46]. Further, Fang & Casadevall [47] found a proportional relation between retractions and journal impact factor; most of these retractions were due to fraud [48]. While the retraction–impact factor relation may in part be because articles in high- rather than low-impact-factor journals are subject in his investigation of research quality across a range of scientific disciplines. He found a negative relation between journal impact factor and various measures of methodological rigour3 (also recently found by [51], in management science; and [52] found weakly negative relations in behavioural science and neuroscience). So if the papers are less good methodologically on average, why do they get published in higher impact-factor journals? Presumably because the authors were good at selling their results. In sum, managerialism at university level, situated in a dysfunctional ecosystem, selects for p-hacking/ corner-cutting salespeople. this explanation is ruled out by Brembs [49] to greater scrutiny, How is the rise of managerialism experienced by university staff? Shattock & Horvath [53] conducted extensive onsite interviews with staff at all levels of each university, at UK universities that spanned a ‘The wide range of rankings, to explore staff experiences. They found widespread dissatisfaction. sense that conditions for the pursuit of high quality academic work have worsened and are continuing to worsen is widespread, even in institutions that are most obviously successful. Criticisms that universities have become too top-down in their governance, and are insufficiently bottom-up, that good academic work is stifled by over-regulation and bureaucracy, and that too much academic business is handled by non-academic professionals, are commonplace’ [53, p. 104]. Similarly, Erickson et al. [29] in a survey of 5888 academic staff in the UK higher education sector, found only 10% of university staff were satisfied with senior management. The question is, in terms of university governance, could we be doing better? In the next section we consider democratic solutions. 6. Open democracy Hierarchical university governance may contribute to damage to scientific integrity. So what mode of governance might actually promote the growth of knowledge? That is, what way of governing universities would be consistent with a culture of the critical tradition, the tradition of carefully considering and selecting arguments rather than people? Arguments concerning the running of an institution can only be considered as such, by anyone with a stake in them, if there is transparency of information. Relevant information must be readily accessible ( just as is required for science to function). What organizational structure produces transparency in a way that is useful? Consider the following assumptions: (1) Each academic has information about how the university is working. (2) Different people have different relevant information. (3) Information will be expressed when a person has to use it to make decisions. (4) Those decisions will make best use of the information if people making decisions have to live by them. (5) To select the best ideas (that integrate information well), ideas should be selected for, not people. Ways of governing that satisfy these assumptions include those of Athenian-like democracy that we considered earlier. Athenian-like democracy has inspired a number of open democratic practices that have been explored in the political context in the last few decades [54]. We will review some of these practices now. 3Again, this finding is consistent with some papers in high-impact-factor journal being rigorous. For example, plausibly if the papers are Registered Reports, then they may have good methodology (e.g. [50]). But then consider the quality of Registered Reports in PCI RR, which is free to authors and readers, and whose rigour the reader can assess for themselves because the whole process is open: https://rr.peercommunityin.org/. 7 r o y a l s o c i e t y p u b i . l i s h n g o r g / j o u r n a l / r s o s R . S o c . O p e n S c i . 1 0 : 2 2 0 8 0 8 Before describing concrete ideas, consider an objection. Given the complex environment in which universities now operate, don’t we need decisions made not democratically by the uninformed but rather by experts (cf. [55]), with a top-down governance structure that thus allows nimble and agile decisions? As against this, if the above five assumptions are accepted, some form of democracy may foster decision-making where the most information is maximally integrated in the time available4: structures embedding such practices should produce networks between people closer to small-world networks than a strict top-down hierarchy could; and such networks allow more global integrated information [57]. Open democracy can ensure decisions are made by the well informed, as we discuss. And it is rare to hear modern universities described as nimble and agile [29,53]. Open democracy also presents a smorgasbord of ways of being democratic with different time scales of operation. Indeed, in Athens, important decisions were sometimes made or reversed very quickly, even in the space of a day ([22], e.g. pp. 133–134). Sometimes decisions should be fast, sometimes slow and reflective. With this in mind, let us see what is on offer. 6.1. The deliberative poll immigration or, The deliberative poll was developed by Fishkin [58,59], and will be used to illustrate the more general class of mini-publics, procedures by which people are selected from the population of citizens to deliberate an issue—such as citizen’s assemblies, citizen’s juries or citizens’ initiative reviews (which we will consider below) (see [60], for a review). Consider a difficult issue that concerns a community, and for which a reflective and informed decision is needed that takes into account diverse viewpoints—such as Brexit, in a university setting, the principles for allocating resources to schools in a context of shifting student demand. Randomly select 200–300 people from the total population (one may decide to over-represent certain groups of people particularly affected by the issue). Given the selection is approximately random, everyone has a chance to contribute and no one is selected simply because they have a vested interest. The total sample is allocated into groups of 15 to discuss the chosen issue. The discussants are given information packs, each pack prepared by experts or protagonists of different views. The discussions are moderated to encourage everyone to contribute more or less equally, and for debate to focus on arguments per se. There is an opportunity for discussants to ask a panel of experts any questions that remain unresolved. After several meetings, discussants anonymously vote: as Fishkin puts it, the vote is then what the people would think if they had reflected. The deliberative poll has been used in many countries over the last few decades, for example in the UK, Europe, Australia, China, the USA, Canada and Mongolia (for critical discussion see [61], chapter 3; [60], chapter 3). To take an example, in Ze Guo Township in China, the issue was how to spend the council’s money [58]. Thirty projects were listed by the council. Two hundred and seventy-five citizens were randomly selected; of these 235 completed the poll, so the final sample was close to random. Fishkin provides evidence that participants became more informed as a result of the discussions, that there was little to no domination by privilege, that there was little to no group polarization, and that priorities shifted toward projects that would benefit the whole town. Of course, the devil would be in the details to get such good procedural outcomes (e.g. well-trained moderators). People chose a sewage treatment plant, park and a main road; not, for example, a fancy town square. These choices surprised officials but they acted on the results of the poll. A minipublic can be used to either set the agenda of a committee, or to make decisions on issues defined by a committee, and both roles may be useful in a university. To adapt the minipublic to a university, one might adjust the number of people selected, or the length and number of meetings, according to the issue considered. Note the similarity of the minipublic in finalizing decisions to the nomethetai in the Athenian model. 6.2. Participatory budgeting Engaging the community as a whole in planning budgets was pioneered in Porto Alegre in Brazil in 1989 onwards in a procedure called participatory budgeting (see [60], for an introduction). In this case, neighbourhood assemblies voted locals to represent the neighbourhood for a year in a local committee 4Information integration in different organizational structures could be tested by injecting information into an organization at different places and then at a later time measure whether that information has been used to make decisions at different points in the organizational structure; compare measuring information integration in the human brain [56]. 8 r o y a l s o c i e t y p u b i . l i s h n g o r g / j o u r n a l / r s o s R . S o c . O p e n S c i . 1 0 : 2 2 0 8 0 8 to plan projects and their priorities. People were also elected to one of several thematic committees (education, transport, etc.) to plan projects and their priorities for each theme. Locals from each neighbourhood and thematic assembly were also voted to an overall coordinating central committee. The central committee decided projects and allocated a substantial proportion of the council’s budget. The committees also involved people with relevant technical and financial expertise. A person could be elected for no more than two terms. The choice of local people, the shortness of the term and the limited number of terms per person is what roughly corresponds to random sampling in deliberative polls in the sense of being the mechanism that limits entrenchment of certain people in decision making. Participatory budgeting was regarded as so successful it has been taken up in different forms in more than 2700 governments (though in 2017, participatory budgeting was suspended in Porto Alegre itself [62]. For a university following the Porto Alegre model, schools could, as local neighbourhoods, select faculty on a rotating basis to a school budgeting committee for a year to decide school projects. Staff from relevant groups could be selected on a rotating basis to committees devoted to certain themes at institutional level, such as IT, catering, grounds, etc. Similarly, each school could select on a rotating basis two representatives for a central budgeting committee to determine institution-wide spending and to finalize decisions of the other committees. In the participatory budget model, selection is by local election; but people could be selected (semi) randomly. 6.3. The citizens’ initiative review The citizens’ initiative review [63] is a one-group minipublic where the group jointly summarizes the best arguments pro and contra a proposal, and also summarizes how they voted. The aim is to provide an information leaflet for a referendum on the proposal that reflects the range of views ordinary people would have if they reflected on the issues and made themselves informed. Crucially, therefore the information leaflet is not provided by vested interests. The citizens review initiative has been used extensively in the state of Oregon, USA, where studies indicate that voters appreciated the information and were better informed as a result of it. 6.4. Allocation of citizens to roles Democracy is often associated with voting, as that is how our current representative democracies work. But when academics vote for people in the university to be on senate, council or some other position, often the vote is based on limited information, such as what school they work in, or what other committees they have been on. The facts given may be of marginal relevance to how that person would contribute to that role. And if the information is enough to be seen as relevant, voting is a mechanism for selecting people who stand out from other people, in other words, for selecting elites [13]. Having the resources or motivation to promote oneself in an election is not the same thing as having the qualities that would make one good at the job the election is for. The medieval Italian city states realized this. But they also thought selection strictly by lot may not select the best people for the job, even if it ensures the top jobs are not all held by people of a certain class. So these city states devised various combinations of election and lot to try to obtain the benefits of both procedures (see [64], chapter 7, for a systematic analysis of the possibilities). For example, one could select by lot a group of candidates, who are then chosen by election. An institution could decide what combination of processes will assign people to posts. One counter-defence of selection by lot alone, i.e. without election, is that the repeated use of such a procedure educates the citizenship; further if each person is merely a constraint in a global process of selecting ideas (as in science) the notion of selecting the right person may lose some of its relevance. Note that once people have been selected for the executive committee by an open democratic process, such a committee can in principle act in real time as fast as any committee now, depending on details. 7. Reforming the university Figure 2 shows a schematic summary of the flow of control from the top down in a hypothetical modern university. The simplest way of seeing how radical reform could be made is keeping the same structure but assigning people to roles by an open democratic process (figure 3a). Just as in Athens, assignment by a schematic possible structure of a UK university now top-down control senior management council senate Dean 1 Dean 2 Dean 3 committees e.g. teaching and learning School 1 School 2 School 3 Figure 2. Schematic diagram of top-down governance as might exist in a current UK university. Control runs from top to bottom. Senate allows some pushback on academic matters but senate may only in practice have the power for suggesting that senior management or council reconsider. (a) new democratic governance model? (b) new democratic governance model? council executive committee; constantly selected from each of the schools minipublics: agenda decisions selection (sortition?) council proposals decisions executive committee; constantly selected from each of the schools selection (sortition?) selection (sortition?) committees e.g. teaching and learning decisions selection (sortition ?) committees e.g. teaching and learning decisions School 1 School 2 School 3 School 1 School 2 School 3 9 r o y a l s o c i e t y p u b i . l i s h n g o r g / j o u r n a l / r s o s R . S o c . O p e n S c i . 1 0 : 2 2 0 8 0 8 (c) new democratic governance model? (d) new democratic governance model? council proposals decisions executive committee; constantly selected from each of the schools minipublics: agenda decisions selection (sortition?) enough signatures triggers proposal selection (sortition ?) committees e.g. teaching and learning decisions council proposals decisions executive committee; constantly selected from each of the schools minipublics: agenda decisions selection (sortition?) enough signatures triggers proposal proposals decisions assembly of citizens selection (sortition?) committees e.g. teaching and learning decisions citizen initiative review? School 1 School 2 School 3 School 1 School 2 School 3 Figure 3. A series of democratic changes to the governance structure shown in figure 2. (a) The same command flow through committees could be kept as in figure 2, but people assigned to committees by sortition (lot). (b) The executive committee could use minipublics to set certain agendas or make decisions. (c) To this could be added a standing right for decisions to be reviewed by the executive committee if a petition with enough signatures is submitted. (d) There could also be a general assembly of citizens to which the executive committee could submit some decisions. So that the assembly can make an informed choice, such referenda could be supported by citizen review initiatives. lot does not mean there are no restrictions. Some jobs may be open only to senior lecturers or above for example. Or some committees may require experience of other committees. While this one change is structurally simple, it is of course a radical first move. It involves the abolishment of senior management. Some people, notably senior management, might think this a possibly catastrophic first move. This possible first move is presented to illustrate how one can keep other things the same, yet radically change the democratic nature of governance to be similar to the style of governance that once thrived in a complex nation for over a hundred years. In practice, open democracy should be explored in small steps. One could first of all set up a minipublic, with commitment from senior management to abide by its decisions, on an issue of importance that needs 10 r o y a l s o c i e t y p u b i . l i s h n g o r g / j o u r n a l / r s o s R . S o c . O p e n S c i . 1 0 : 2 2 0 8 0 8 time to consider, for example, the university’s response to proposed pension reforms. The issue needs reflection and requires becoming immersed in relevant information, and should not be decided solely by people with one type of axe to grind. Figure 3b illustrates the addition of minipublics. Once their use has been explored and finessed, more changes could be explored. Figure 3c shows a further step that could be taken as an initial small step. To increase recurrent information flow through the system, so that it functions as close as possible to a self-organizing dynamic system that maximizes global constraint satisfaction, if enough signatures are obtained for a petition, the executive committee could be obliged to reconsider a decision, then provide information for why they kept it or changed the decision. Finally figure 3d shows the addition of an assembly of citizens. On some occasions the executive committee may wish decisions be decided by referenda. In this case, citizens’ initiative reviews should be used to provide unbiased information about what is at stake. Whether an institution makes any one of the changes suggested in this sequence should be a slow process of exploration. 8. Conclusion This paper outlined some broad principles for making decision making in a university more closely match decision making in science, arguing both that open democracy allows good decisions (else why do we use it in science, one of our most successful endeavours?); and that good science will be promoted when embedded in a broader culture that operates in the same way as itself. But much was left unaddressed. Who counts as a citizen? That will need to be addressed by an institution (bearing in mind we should be accountable to students, in a way we are not in the current system). What about professional services? Just as politicians need a civil service that has expertise and will offer alternative proposals, so academics need professional services to allow the university to run. The organization of professional services has not been addressed, but presumably some of the same principles could apply to their governance. What about Council, the ultimate governing body of a university? Shattock and Horvath ([53], p. 100) are scathing about how Councils are currently formed in terms of what is expected of them. An institution will need to decide how best to make Council better informed and more accountable through open democratic practices. The argument is not at all that the credibility crisis in science, and the current dissatisfaction with university governance, is the fault of any particular individuals, most of whom are simply people trying to make the best decisions. The problem is the structure in which senior management operate. No amount of focus groups to determine key catch phrases to repeat as slogans will change that. The current governance structure is almost designed to be divisive and demoralizing. Just like an Abbasid caliph, a current senior manager or VC of a modern university may somehow benignly run a happy and smooth operation that promotes the flourishing of knowledge despite the system. Until the next VC comes along. Let’s make the system itself work. One concern is whether open democracy will increase the admin load on people. If committees maintained the same numbers as currently, the average committee load ceteris paribus remains the same. There is an incentive difference, however, that may reduce average committee burden: while professional bureaucrats have an incentive to maximize number of committee meetings (are there not more data on key performance indicators and ‘academic drivers’ to be drilled down into?), people who are committee- averse have an incentive to reduce them. On the other hand, deliberative polls formed to consider an issue will increase admin time. The trade-off is that this increase in time is time spent being informed about how things work, being a part of its workings, and making a difference to how things work. A related concern is random lot means people may be selected who consider they are already being overworked. One could be allowed to refuse such an assignment a certain number of times in a certain period. But in the end the system will only function if people are committed to being citizens. The current pay of senior management could be split among citizens, and people could opt out of being citizens. People who opt out would not be taken seriously when they complained bitterly about decisions that were made. But is democracy weak under pressure, inherently irrational, maybe susceptible to mob rule? Tangian ([65], p. 294) argued that Athenian democracy would make unstable decisions for ‘situations close to controversy [i.e. 50/50 for and against in the population] because a negligible disturbance results in a significant change. At the controversial point, the decisive majority opinion, whether positive or negative, depends on the positions of just a few individuals.’ Similarly, there are long-standing theorems, such as Arrow’s theorem, that seem to show a group of people making decisions by voting are necessarily irrational in one way or another. Deutsch ([66], chapter 13) points out that these 11 r o y a l s o c i e t y p u b i . l i s h n g o r g / j o u r n a l / r s o s R . S o c . O p e n S c i . 1 0 : 2 2 0 8 0 8 arguments, that might be used against democracy, including Tangian’s just mentioned, presume there are fixed options regarded by people with fixed preferences. But the point of critical discussion is to change one’s preferences and allow new options to unfold. That is the real substance of decision making. Open democracy allows preferences and lists of options to change in a rational way. And to the extent that having people form small minipublics with moderated discussion promotes sensitivity to reason, democracy need not amount to mob rule. How could these reforms be implemented in practice? Popper [67] argued for piecemeal social engineering, that is taking small steps at any one time, as the politician is aware that the perfect state, if attainable, is far distant, and each change along the way will have unintended consequences, which are best dealt with one by one. Could a university VC be persuaded there is at least one decision that is important for people, yet one they could be willing to give up to a minipublic? It is sometimes in a leader’s interest to have handed over difficult decisions to a public body, such the Citizens’ Assemblies to consider abortion or marriage equality in Ireland [68]. Or if a VC genuinely considers a pension deal is the best one for the university, and they trust fully informed rational discussion will lead to the same conclusion, why not, by having a minipublic decide, save themselves from losing goodwill for the rest of their term? Or why not start with a decision about which they have no axe to grind, but staff do care about—might that not help staff morale? Might not a university who started along this path become a beacon for other universities, be the one that stood out from the other sheep by not copying their nearest neighbour sheep? Start with a single decision. Then finesse the procedure to make it better. Gradually finessing the procedure will take effort; but as indicated there are reasons to think it is worth the effort. In fact, open democracy may uniquely solve a key problem. Fukuyama [69] argued that all political systems face the problem of elite entrenchment bringing about inevitable decay of the system. Analogous arguments may apply to universities, not just nation states, when there is a limited pool of managers who, schooled by the same system, approach problems in similar ways. Despite what Fukuyama claimed, there is a way of hindering elite entrenchment: by constantly randomly selecting citizens to act as decision makers, there is a constant input of new viewpoints. Injections of randomness are necessary for creativity. Of course, people with many good ideas may still be especially influential. The goal is to make sure that in taking on board ideas, what matters is the quality of the ideas. Thus, good ideas should be selected, whatever their source. The benefits of open democracy may also be motivated in the light of McGregor’s [70] influential distinction between two theories that management might have about the psychology of people. According to Theory X, ‘the average human being has an inherent dislike of work … [and so] must be coerced, controlled, directed, threatened with punishment to get them to put forth adequate effort … ’ (p. 43). By contrast, according to Theory Y, ‘The expenditure of physical and mental effort in work is as natural as play or rest … [a person] will exercise self-direction and self-control to the service of objectives to which [they are] committed … the capacity to exercise a high degree of imagination, ingenuity and creativity in the solution of organizational problems is widely, not narrowly, distributed … ’ (pp. 59–60). A relevant principle may be that when management treats people in ways that express certain expectations of them (e.g. Theory X or Y), management tends to get what it expects (for a business case study, see [71]). Likewise, requiring academics to fulfil narrow key performance indicators, may tend to produce academics who do as they are directed (and no more). Managers who subscribe to Theory X may regard this as proof of their position—even as science and organizational creativity suffers. Yet trusting faculty to solve important organizational and research problems, as open democracy requires, may yield superior outcomes, by promoting those qualities presumed by trust (see [31], and [39] for a review of relevant evidence). Current problems with the university environment that were identified earlier included the overuse of key performance indicators and the use of a top-down governance structure. Yet democracy as such is orthogonal to both specific decisions concerning working conditions (e.g. whether staff are incentivized by particular metrics), and also by whether the flow of control runs from top down or bottom up in terms of committee structure. A democratic university may (or may not) decide to incentivize, for example, grant income in promotion criteria. If they do, this will be a decision whose details will be finessed by those who daily confront what the real trade-offs are. And given general dissatisfaction with the way senior management attempts to incentivize academics now, a democratic university may in at least some institutions come to different decisions in detail. Importantly, the decision will be made by faculty knowing they are trusted to make decisions, because open democracy embodies Theory Y thinking. When metrics are decided locally and with light touch by the experts, and used with judgement as a means and not an end, they can be useful and need not be demotivating [39]. Similarly, the flow of control can still be democratically top down as in figure 3, allowing greater coherence across the university (democratic decisions can apply to the whole university), or bottom 12 r o y a l s o c i e t y p u b i . l i s h n g o r g / j o u r n a l / r s o s R . S o c . O p e n S c i . 1 0 : 2 2 0 8 0 8 up, allowing different schools more independence. The two directions of flow can both be democratic because the people at the bottom and the top are, over time, the same when there is assortment by lot, or the use of deliberative polls. What the direction of the flow of control is for a given institution can be worked out according to the needs of a particular institution (say, by a deliberative poll). And in whichever direction control goes, once again a crucial difference remains compared with the status quo: An open democratic system embodies Theory Y thinking. Once a democratic organizational structure is in place, the decisions that result can be jointly owned. It will be us who made the decisions. We can say: this is our house, we built it. Our state. We as citizens may make a mess of it, as we invariably must, as any decision-making procedure will. But it will be our mess, our problems to solve—together. Data accessibility. This article has no additional data. Authors’ contributions. Z.D.: conceptualization, writing—original draft. Conflict of interest declaration. I declare I have no competing interests. Funding. I received no funding for this study. Acknowledgements. Thanks to Jason Chin and David Mellor for valuable comments. Thanks also to Josh Ober who gave talks at Sussex in 2015 which helped inspire me, in conversation with Jorg Huber. References 1. 6. 4. 3. 2. 5. Ritchie S. 2020 Science fictions: exposing fraud, bias. Negligence and hype in science. London, UK: The Bodley Head. Ioannidis JPA. 2012 Why science is not necessarily self-correcting. Pers. Psychol. Sci. 7, 645–654. (doi:10.1177/1745691612464056) Edwards MA, Roy S. 2017 Academic research in the 21st century: maintaining scientific integrity in a climate of perverse incentives and hypercompetition. Environ. Eng. Sci. 34, 51–61. (doi:10.1089/ees.2016.0223) Stewart S, Pennington CR, Silva G, Ballou N, Butler J, Dienes Z, Jay C, Rossit S, Samara A. 2022 Reforms to improve reproducibility and quality must be coordinated across the research ecosystem: the view from the UKRN Local Network Leads. BMC Research Notes. See: https://psyarxiv.com/xzfa2 Popper KR. 1963 Conjectures and refutations. London, UK: Routledge. Popper KR. 1979 Objective knowledge: an evolutionary approach. Oxford, UK: Oxford University Press. Flourishing Science Think Tank. 2022 Beyond kindness: a proposal for the flourishing of science and scientists. (doi:10.31231/osf.io/ 4zrmd) Nottourno MA. 2000 Science and the open society: the future of Karl Popper’s philosophy. Budapest, Hungary: Central University Press. Lloyd CER. 1970 Early Greek science: Thales to Aristotle. London, UK: Ghatto & Windus. 10. Mattern SP. 2013 The prince of medicine: Galen in the Roman empire. Oxford, UK: Oxford University Press. Ober J. 2015 The rise and fall of classical Greece (The Princeton history of the ancient world). Princeton, NJ: Princeton University Press. Al-Khalili J. 2012 The house of wisdom: how Arabic science saved ancient knowledge and gave us the Renaissance. London, UK: Penguin. 13. Manin B. 1997 The principles of representative government. Cambridge, UK: Cambridge University Press. 12. 11. 9. 8. 7. 14. 15. 16. 17. 18. Stasavage D. 2020 The decline and rise of democracy: a global history from antiquity to today. Princeton, NJ: Princeton University Press. Ferris T. 2011 The science of liberty: democracy, reason, and the laws of nature. New York, NY: Harper Collins. Needham J. 1969 The grand titration: science and society in east and west (China: history, philosopy, economic). London, UK: George Allen & Unwin. Lu J. 2011 Reassessing the Needham question: what forces impeded the development of modern science in China after the 15th century? Concord Rev. 21, 209–255. Graeber D, Wengrow D. 2021 The Dawn of everything: a new history of humanity. Dublin, Ireland: Allen Lane. 20. 19. Walshe M. 1995 The long discourses of the Buddha: a translation of the Digha Nikaya. Boston, MA: Wisdom Publications. Lovell M, Dienes Z. 2022 Minimal mindfulness of the world as an active control for a full mindfulness of mental states intervention: a registered report and pilot study, in principle acceptance of version 4 by Peer Community in Registered Reports. https://osf.io/tx54k Recommendation See: https://rr. peercommunityin.org/articles/rec?id=45 Saxonhouse AH. 2006 Free speech and democracy in ancient Athens. Cambridge, UK: Cambridge University Press. 21. 23. 22. Manville B, Ober J. 2003 A company of citizens: what the world’s first democracy teaches leaders about creating great organizations. Harvard, MA: Harvard Business Review Press. Cartledge P. 2016 Democracy: a life. Oxford, UK: Oxford University Press. Rhodes PJ. (translator). 2002 The Athenian constitution. London, UK: Penguin. Collini S. 2017 Speaking of universities. London, UK: Verso. Ferlie E, Musselin C, Andresani G. 2008 The steering of higher education systems: a 24. 26. 25. public management perspective. High. Educ. 56, 325. (doi:10.1007/s10734- 008-9125-5) 27. Martin BR. 2017 What’s happening to our universities? Prometheus 34, 7–24. 28. Willets D. 2017 A university education. Oxford, 29. 30. 31. 32. 33. 34. 35. 36. 37. 38. UK: Oxford University Press. Erickson M. 2021 The UK higher education senior management survey: a statactivist response to managerialist governance. Stud. High. Educ. 46, 2134–2151. (doi:10.1080/ 03075079.2020.1712693) Honig D. 2020 Navigation by judgment: why and when top down management of foreign aid doesn’t work. Oxford, UK: Oxford University Press. Honig D. 2022 Managing for Motivation as Public Performance Improvement Strategy in Education & Far Beyond. See: https://bsc.cid. harvard.edu/publications/managing-motivation- public-performance-improvement-strategy- education-far-beyond. Brown R, Pehrson S. 2019 Group processes: dynamics within and between groups, 3rd edn. New Jersey, NJ: Wiley. Leaf MJ. 2018 An anthropology of academic governance and institutional democracy: the community of scholars in America. Cham, Switzerland: Palgrave Macmillan. Ginsberg B. 2011 The fall of the faculty: the rise of the all-administrative university and why It matters. New York, NY: Oxford University Press. De Boer H, File J. 2009 Higher Education Governance Reforms across Europe. See: https:// ris.utwente.nl/ws/portalfiles/portal/5147233/ c9hdb101+modern+project+report.pdf. Biggs JB. 2013 Changing universities: a memoir about academe in different places and times. Scarborough, Australia: Strictly Literary. Hil R. 2012 Whackademia: an insider’s account of the troubled university. Sydney, Australia: UNSW Press. Xu F, Wu L, Evans J. 2022 Flat teams drive scientific innovation. Proc. Natl Acad. Sci. USA 119, e2200927119. (doi:10.1073/pnas. 2200927119) 42. 41. 43. 40. 44. 39. Muller JZ. 2018 The tyranny of metrics. Princeton, NJ: Princeton University Press. Vann MG. 2003 Of rats, rice, and race: the great Hanoi rat massacre, an episode in French colonial history. Fr. Colon. Hist. 4, 191–203. (doi:10.1353/fch.2003.0027). S2CID 143028274. Smaldino PE, McElreath R. 2016 The natural selection of bad science. R. Soc. Open Sci. 3, 3160384160384. (doi:10.1098/rsos.160384) Smaldino PE, Turner MA, Contreras Kallens PA. 2019 Correction to ‘Open science and modified funding lotteries can impede the natural selection of bad science’. R. Soc. Open Sci. 6, 6191249191249. (doi:10.1098/rsos.191249) Gross K, Bergstrom CT. 2019 Contest models highlight inherent inefficiencies of scientific funding competitions. PLoS Biol. 17, e3000065. (doi:10.1371/journal.pbio.3000065) Brette R. 2022 Le modèle managérial de la recherche - Critique et alternative [A critique of the managerial model of research]. Med. Sci. (Paris) 2022, 84–88. French. (doi:10.1051/medsci/ 2021247). English: See: http://romainbrette.fr/ WordPress3/wp-content/uploads/2022/01/A- critique-of-the-managerial-model-of-research.pdf Racimo F, Galtier N, De Herde V, Aubert Bonn N, Phillips B, Guillemaud T, Bourguet D. 2022 Ethical publishing: how do we get there? Zenodo 14, 15. (doi:10.5281/zenodo.6289488) Fanelli D. 2020 Pressure to publish: what effects do we see? In Gaming the metrics: misconduct and manipulation in academic research (eds M Biagioli, A Lippman). Cambridge, MA: MIT Press. Fang FC, Casadevall A. 2011 Retracted science and the retraction index. Infect. Immun. 79, 3855–3859. (doi:10.1128/IAI.05661-11) Fang FC, Steen RG, Casadevall A. 2012 Misconduct accounts for the majority of retracted scientific publications. Proc. Natl Acad. 47. 48. 46. 45. 49. 50. 51. 52. 53. 54. 55. Sci. USA 109, 17– 028–17 033. (doi:10.1073/ pnas.1212247109) Brembs B. 2018 Prestigious science journals struggle to reach even average reliability. Front. Hum. Neurosci. 12, 37. (doi:10.3389/fnhum. 2018.00037) Dienes Z. 2020 The inner workings of Registered Reports. PsyArXiv preprint. (doi:10.31234/osf.io/ yhp2a) Kepes S, Keener SK, McDaniel MA, Hartman NS. 2022 Questionable research practices among researchers in the most research-productive management programs. J. Org. Behav. 43, 1190–1208. (doi:10.1002/job.2623) Dougherty MR, Horne Z. 2022 Citation counts and journal impact factors do not capture some indicators of research quality in the behavioural and brain sciences. R. Soc. Open Sci. 9, 9220334220334. (doi:10.1098/rsos.220334) Shattock M, Horvath A. 2021 The governance of British higher education: the impact of governmental, financial and market pressures. Bloomsbury Higher Education Research. London, UK: Bloomsbury Academic. Landemore H. 2020 Open democracy: reinventing popular rule for the 21st century. Princeton, NJ: Princeton University Press. Jones G. 2020 10% less democracy: why you should trust elites a little more and the masses a little less. Stanford, CA: Stanford University Press. 56. Massimini M, Ferrarelli F, Huber R, Esser SK, Singh H, Tononi G. 2005 Breakdown of cortical effective connectivity during sleep. Science 309, 2228–2232. (doi:10.1126/science.1117256) Latora V, Marchiori M. 2001 Efficient behavior of small-world networks. Phys. Rev. Lett. 87, 198701. (doi:10.1103/PhysRevLett.87.198701) Fishkin JS. 2011 When the people speak: deliberative democracy and public consultation. Oxford, UK: Oxford University Press. 57. 58. 59. 63. 62. 61. 60. Fishkin JS. 2018 Democracy when the people are thinking: revitalizing our politics through public deliberation. Oxford, UK: Oxford University Press. Smith G. 2009 Democratic innovations: designing institutions for citizen participation. Cambridge, UK: Cambridge University Press. O’Flynn I. 2022 Deliberative democracy. Cambridge, UK: Polity Press. Abers R, King R, Votto D, Brandao I. 2018 Porto Alegre: participatory budgeting and the challenge of sustaining transformative change. Case study, World Resources Institute. See: https://www.wri.org/research/porto-alegre- participatory-budgeting-and-challenge- sustaining-transformative-change. Gastil J, Knobloch KR. 2020 Hope for democracy: how citizens can bring reason back into politics. Oxford, UK: Oxford University Press. 64. McCormick JP. 2011 Machiavellian democracy. Cambridge, UK: Cambridge University Press. Tangian A. 2020 Analytic theory of democracy: history, mathematics, and applications. Cham, Switzerland: Springer. Deutsch D. 2011 The beginning of infinity: explanations that transform the world. London, UK: Allen Lane. Popper KR. 1945/1974 The open society and its enemies: volume 1, Plato. London, UK: Routledge. Farrell DM, Suiter J. 2019 Reimagining democracy: lessons in deliberative democracy from the Irish frontline. Ithaca, NY: Cornell University Press. Fukuyama F. 2014 Political order and political decay: from the industrial revolution to the globalization of democracy. London, UK: Profile Books. 66. 68. 67. 69. 65. 70. McGregor D. 1960 The human side of enterprise, 71. annotated edn. New York, NY: McGraw Hill. Semler R. 1993 Maverick! The success story behind the world’s most unusual workplace. London, UK: Random House Business Books. 13 r o y a l s o c i e t y p u b i . l i s h n g o r g / j o u r n a l / r s o s R . S o c . O p e n S c i . 1 0 : 2 2 0 8 0 8
10.1038_s41467-023-35891-9
Article https://doi.org/10.1038/s41467-023-35891-9 Universality of light thermalization in multimoded nonlinear optical systems Received: 28 July 2022 Accepted: 5 January 2023 Check for updates ; , : ) ( 0 9 8 7 6 5 4 3 2 1 ; , : ) ( 0 9 8 7 6 5 4 3 2 1 Qi Zhong1,5, Fan O. Wu 1,5, Absar U. Hassan1, Ramy El-Ganainy 2,3 & Demetrios N. Christodoulides 1,4 Recent experimental studies in heavily multimoded nonlinear optical systems have demonstrated that the optical power evolves towards a Rayleigh–Jeans (RJ) equilibrium state. To interpret these results, the notion of wave turbulence founded on four-wave mixing models has been invoked. Quite recently, a different paradigm for dealing with this class of problems has emerged based on thermodynamic principles. In this formalism, the RJ distribution arises solely because of ergodicity. This suggests that the RJ distribution has a more general origin than was earlier thought. Here, we verify this universality hypothesis by investigating various nonlinear light-matter coupling effects in physically accessible multimode platforms. In all cases, we find that the system evolves towards a RJ equilibrium—even when the wave-mixing paradigm completely fails. These observations, not only support a thermodynamic/ probabilistic interpretation of these results, but also provide the foundations to expand this thermodynamic formalism along other major disciplines in physics. Nonlinear optics plays a crucial role in a wide range of modern sci- ence and technologies. These include optical cavity microcombs1,2, high-power light sources3, cavity optomechanics4,5, nonlinear topo- logical and non-Hermitian photonics6–10, bioimaging11,12, as well as classic/quantum networks and signal processing13–16. While nonlinear interactions widely vary in strength and differ from one material system to another, their vast majority can still be described using an relies on perturbative framework that underlying theoretical analysis17. Particularly, by expressing the electric polarization vector as a Taylor series expansion in terms of the driving electric field, one can classify nonlinear optical effects into several, largely indepen- dent processes such as those associated with second harmonic and sum/difference frequency generation and multi-wave mixing interactions17. A few decades ago, this same paradigm was adopted by Zakharov and colleagues to study optical nonlinear propagation effects when an infinite number of Fourier components is involved—a field of research that is nowadays known as wave turbulence18. In this seminal work, it was shown that such a system can be described by a Boltzmann-like kinetic model that admits a steady-state solution in the form of a Rayleigh–Jeans (RJ) distribution. In this regard, it was conjectured that the RJ law results as a mere byproduct of the non- linear attractor dynamics taking place during multi-wave mixing19. In developing this model, several assumptions were made. Firstly, it was implicitly assumed that four-wave mixing dominates the interaction process. Secondly, the so-called random phase approximation20 was employed to omit off-resonant interaction terms. Meanwhile, recent progress in the general area of multimode fiber optics21–29 has enabled a new generation of nonlinear experimental setups where the RJ distribution (power allocation among modes) was successfully observed for the first time30–33. The clear demonstration of RJ ther- malization in such settings has been touted as evidence in support of the wave turbulence theory. While reaching such a conclusion does not seem to pose a problem from a practical point of view, it is unsettling at a more fundamental level. In essence, adopting the wave 1CREOL, College of Optics and Photonics, University of Central Florida, Orlando, FL 32816, USA. 2Department of Physics, Michigan Technological University, Houghton, MI 49931, USA. 3Henes Center for Quantum Phenomena, Michigan Technological University, Houghton, MI 49931, USA. 4Ming Hsieh Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, CA 90089, USA. 5These authors contributed equally: Qi Zhong, Fan O. Wu. e-mail: ganainy@mtu.edu; demetri@creol.ucf.edu Nature Communications | (2023) 14:370 1 Article https://doi.org/10.1038/s41467-023-35891-9 turbulence hypothesis is to a great extent analogous to attempting to infer, for example, the nature of the interactions between gas molecules solely from the Maxwell–Boltzmann distribution. Even more importantly, while the laws of simple thermodynamic systems like gases can be developed from either classical (Newtonian) kinetic theories or quantum mechanical perspectives, this is by no means necessary, given that the corresponding equations of state can be derived from purely entropic principles—in total disregard to the underlying collisional mechanisms. So, the question naturally arises: is the RJ distribution an actual byproduct of multi-wave mixing pro- cesses or does it represent a much more general result that has little to do with the specifics of the inherent nonlinearity involved? Quite recently, a different approach for studying light thermali- zation was put forward on the basis of statistical mechanics and thermodynamics34–38. While this latter theoretical framework reaches similar conclusions to those derived from the aforementioned kinetic theories18,19 as far as the RJ distribution is concerned, its perspective of optical thermalization is fundamentally different. Being founded on notions from statistical mechanics, this paradigm34,35 allows one to predict and interpret the RJ law emerging in a microcanonical system from purely entropic considerations. In this regard, the RJ equilibrium state macroscopically manifests itself because it is ergodically asso- ciated with a largest number of microstates (in phase space) and thus it can be considered a byproduct of probability theory—an aspect that has little to do with the nature of the underlying nonlinearity involved. If this is indeed the case, then in analogy with statistical mechanics of gases, the RJ thermalization should occur in systems with more generic nonlinearities beyond the wave mixing paradigm as illustrated in Fig. 1. The situation is however more complex. Nonlinear optical systems often exhibit two constants of motion, i.e., the power and the Hamil- tonian. The first, which describes the conservation of optical power, is analogous to the number of particles in a gas system. The second, however, when expressed in the linear eigenbasis, involves both a linear and a nonlinear component. Thus, strictly speaking, such a system is not necessarily expected to relax to a RJ distribution. Only under the condition that the linear part is constant, the RJ distribution can be anticipated. In reality, however, even under weak nonlinear conditions, the linear part of the Hamiltonian is only quasi-conserved. In other words, the analogy between multimoded nonlinear optical arrangements and idealized thermodynamic systems involving two constants of motion is not formal, which further complicates the question about thermalization in nonlinear optical systems and the physical mechanism responsible for observing the RJ distribution. Non-equilibrium state Equilibrium state FWM SHG OM MWM w/o WM y c n a p u c c O Maximize the entropy y c n a p u c c O RJ Energy level Energy level Fig. 1 | Conceptual illustration of thermalization in a nonlinear multimode optical system. Similar to thermalization in matter, the nature of the interaction forces (like forces between gas molecules) is irrelevant. Here, we show that light thermalization into a Rayleigh–Jeans (RJ) distribution can take place under a wide range of nonlinear conditions beyond the traditional four-wave mixing (FWM) paradigm. These include second harmonic generation (SHG), multi-wave mixing (MWM), optomechanical (OM) cascaded interactions between optical and mechanical modes, and even scenarios where the system cannot be described by any wave mixing expansion (w/o WM). In this work, we critically examine the manner in which optical thermalization processes unfold in nonlinear environments with dif- ferent types of nonlinearities such as those arising from optomecha- interactions (where wave mixing interpretations are rather nical cumbersome) and those associated with photorefractive crystals (where above certain power thresholds, standard perturbative wave mixing expansions are not possible). In addition, we consider also artificial nonlinear systems with nonanalytic and discontinuous non- linear functions that cannot be described by any convergent poly- nomial and demonstrate that such set-ups can also reach the RJ equilibrium distribution. Our work thus establishes the universality of the thermalization towards the RJ state in nonlinear optical systems, and, in doing so, presents compelling evidences in favor of the more general entropic view of optical thermalization as opposed to the more restrictive four-wave mixing paradigm. Results Before we proceed, perhaps it would be useful to highlight some of the basic notions upon which the optical thermodynamic approach relies on. As in the case of standard statistical mechanics39, the entropy of the optical multimode arrangement can be built within a microcanonical ensemble formalism by accounting all possible microstates, each containing information as to the energy/power and phase distribution among all modes in the system. In defining the macrostates, the energy/power distribution is retained while the phase information is omitted40 (being superfluous given that it is uniformly distributed within the range 0 to 2π). In this respect, the nonlinear interaction acts merely as an agent that enables a chaotic reshuffling of optical energy among modes and therefore facilitates thermalization. On the other hand, the specifics of nonlinearity are inconsequential. Optical ther- modynamic equilibrium is then reached when entropy is maximized over all possible microstates under the constraints dictated by the two constants of motion35. Kerr nonlinearity We begin our analysis by first considering a Kerr nonlinear multimode tight-binding model—a one-dimensional photonic array comprised of M evanescently coupled single-mode waveguides with nearest neigh- bor coupling41,42 (a situation most relevant to experimental imple- mentations), as shown in Fig. 2a. Under these conditions, light propagation along z in such a lattice can be described by the following normalized discrete nonlinear Schrödinger equation43: i dam dz + am(cid:1)1 + am + 1 + ∣am∣2am = 0, ð1Þ (cid:2) m = 1 PM ∣am∣2 = PM j = 1 where am is the field amplitude at site m, and the last term denotes Kerr nonlinear effects. Equation (1) exhibits two constants of motion. The first invariant (denoting power conservation) is given by ∣cj∣2, where cj is the field amplitude component P = associated with supermode ∣ψji of the linear array (i.e., the normal modes obtained by diagonalizing Eq. (1) in the absence of the nonlinear term). The complex amplitudes cj at any distance z are obtained by projecting the state ∣ψ of the system on the linear supermodes as expressed in the local representation (i.e., in terms of am). The second invariant is associated with the optical Hamiltonian comprised of a linear HL and a nonlinear HNL component, i.e., H = HL + HNL where ∣am∣4, where aM+1 = 0 H because of the truncated boundary condition. Under weak nonlinear conditions, the contribution from the linear term HL dominates, and as a result one can define a quasi-invariant internal energy by εj∣cj∣2, where εj = 2 cosð jπ Þ are the eigenvalues U (cid:3) (cid:1)H M + 1 associated with the linear supermodes ∣ψji. As indicated above, by using purely entropic principles, one can show that light propagating in such a system evolves towards a thermal state obeying the RJ m + 1 + a* PM j = 1 Þ and H mam + 1 L = (cid:1) ðama* PM PM NL = L = m = 1 m = 1 1 2 Nature Communications | (2023) 14:370 2 Article https://doi.org/10.1038/s41467-023-35891-9 Fig. 2 | Thermalization of light in nonlinear waveguide arrays with different nonlinearities. Linear and nonlinear couplings in three optical lattices when acted upon by three different nonlinearities: a a Kerr nonlinearity, c cascade χ(2) process, and e optomechanical nonlinearities, as described by Eqs. (1), (3), and (4), respectively. Numerical simulations provide the modal occupancies after therma- lization in all these three scenarios, in good agreement with the predicted Rayleigh–Jeans (RJ) distributions (black lines), as shown in b, d, and f. The insets display a monotonic increase in entropy S as well as the invariants of the motion U and P. Note that in all cases, numerical simulations are performed over ensemble averages. The thermal fluctuations of quasi-invariants (when applicable) are indi- cated by gray lines, depicting the instantaneous values of U and P around their mean values. In all cases, the nonlinear array has M = 100 sites and the dashed lines represent the initial occupancies for the linear optical supermodes. distribution34,35: ∣cj∣2 = (cid:1) T μ + εj , ð2Þ where T and μ represents the optical temperature and chemical potential, respectively. In general, the equilibrium values of T, μ can be predicted from the initial conditions, i.e., from the invariants P and U34,36,38. For instance, for a lattice with M = 100 elements, an input excitation ∣cj∣2 = 0.05(εj + 2) (dashed line in Fig. 2b) leads to P = 10 and U = − 9.9, which in turn predicts T = 0.15 and μ = − 2.5 (see Supplemen- tary Note 1). The size of the systems considered in this study is large enough so as to guarantee the extensivity of the entropy and the self- consistency of the thermodynamic formulation used44. By numerically integrating Eq. (1), we find that the equilibrium modal occupancies ∣cj∣2 are consistent with the theoretically predicted RJ distribution (Fig. 2b). The inset panel in Fig. 2b shows that during propagation, the optical j = 1 lnð∣cj∣2Þ monotonically increases until it reaches a entropy S = maximum (as expected by the second law of thermodynamics) while the optical energy U remains quasi-invariant. PM Cascade second order χ(2) nonlinearity In order to demonstrate the universality of RJ thermalization, we now investigate a variety of scenarios. In this respect, we consider cascade second order χ(2) nonlinear processes unfolding in waveguide arrays, governed by the following normalized coupled evolution equations45–47: ða2 m = 1 mb* PM PM mbmÞ(cid:4) m = 1 (cid:1) 1 2 mam + 1 PM j = 1 m + 1 + a* where am and bm are the local site field amplitudes associated with the fundamental and the second-harmonic frequency, and Δ is the phase mismatch. Here the linear coupling among bm is neglected46, as illustrated in Fig. 2c. This system exhibits two constants of motion: the ð∣am∣2 + ∣bm∣2Þ and the Hamiltonian total optical power P = m + a*2 Δ∣bm∣2 + 1 ½ama* H = (see Sup- 2 plementary Note 2). Under weak nonlinear conditions, the field in the fundamental frequency am dominates, and therefore its power and ∣cj∣2, and energy can be regarded as quasi-invariants, i.e., Pa = Ua = (cid:1) εj∣cj∣2, where cj is the field amplitude of the corresponding supermode . If indeed this system can thermalize through the χ(2) process under these two invariants, one should then anticipate a RJ distribution once equilibrium is reached. To confirm this hypothesis, we numerically simulated Eq. (3) with Δ = 1, M = 100 when the first 30 modes in the fundamental frequency were evenly excited (dashed line in Fig. 2d). As shown in Fig. 2d, after a non-equilibrium prethermaliza- tion stage, the quantities Pa and Ua eventually settle to Pa = 8:3 and Ua = −15.1, i.e., they remain invariants. For this set of values, once thermal equilibrium is attained, our theory predicts T = 0.016 and μ = −2.007, in excellent agreement with our numerical simula- tions (Fig. 2d). PM j = 1 Optomechanical nonlinearity Next, we consider a lossless nonlinear optomechanical cavity array where the intracavity optical fields and the vibrational motions are described by the following evolution equations48: dam dz + am(cid:1)1 + am + 1 + a* i dbm i dz (cid:1) Δbm + a2 m = 0, mbm = 0, Nature Communications | (2023) 14:370 ð3Þ i i dam dt dbm dt (cid:1) ðam(cid:1)1 + am + 1 Þ + amðbm + b* mÞ = 0, (cid:1) Ωbm + ∣am∣2 = 0: ð4Þ 3 Article https://doi.org/10.1038/s41467-023-35891-9 m = 1 m = 1 PM ∣am∣2 = PM Þ + ∣am∣2ðbm + b* PM j = 1 ½(cid:1)ðama* Here am and bm stands for the optical field and the mechanical oscil- lation amplitude in cavity m, respectively (Fig. 2e), while the parameter Ω represents a normalized angular frequency of the mechanical resonance. Synchronization between driven optomechanical oscilla- tors have been investigated in earlier studies and it was shown that the synchronization dynamics follow the generic features of the Kur- amoto model49. Here, instead, we are interested in the nonlinear dynamics of coupled optomechanical oscillators in the absence of the driving force. We proceed by first noting that the above system exhibits two invariants: the number of “photons” in the cavities ∣cj∣2, and the overall Hamiltonian of the sys- Pa = tem H = mam + 1 m + 1 + a* (see Supplementary Note 3), where cj denotes the field amplitude of the jth optical supermode. As before, under weakly nonlinear conditions and when the normalized Ω is large, such as Ω = 8 in our numerical simulations, one finds that the Hamiltonian associated with the optical field is a quasi-invariant, εj∣cj∣2. Even in this more Ua = complex scenario, the RJ distribution emerges at thermal equilibrium as a result of ergodicity as can be seen in Fig. 2f. In all cases, a good agreement was found to exist between numerical simulations and the theoretically anticipated RJ distribution once Pa, Ua were specified by initial conditions. Note that in this case, it is impossible to associate a multi-wave mixing process to the optical nonlinearity—an aspect that dispels the wave turbulence paradigm. Interestingly, unlike their photon counterparts, the mechanical vibrations themselves do not display a pair of (quasi-)invariants P and U (see Supplementary Note 3), and therefore cannot thermalize to a RJ equilibrium state in the same manner. the linear part of mÞ (cid:1) Ω∣bm∣2(cid:4) m + 1 + a* PM j = 1 ½(cid:1)ðama* mam + 1 PM Þ(cid:4) = m = 1 Nonlinearity described by a smooth but nowhere analytic function So far, we have analyzed thermalization effects in multimode systems where the nonlinearities conform to standard Taylor series expan- sions. Naturally, one may ask whether the RJ thermalization process can indeed manifest itself in more general nonlinear settings. To address this question, we now consider optical lattices involving generalized intensity-dependent nonlinearities F(x) as described by50: i dam dz + am(cid:1)1 + am + 1 + Fð∣am∣2Þam = 0: ð5Þ PM j = 1 m = 1 PM ½ama* m + 1 + a* ∣cj∣2 as well as the Hamiltonian Here the optical power P = H = mam + 1 + Gð∣am∣2Þ(cid:4) of the system are still con- served, where G(x) is the antiderivative of F(x) (i.e., dG(x)/dx = F(x), and G(0) = 0). As before, in the weak nonlinear regime, i.e., F(x) ≪ 1, the linear part of the Hamiltonian U = (cid:1) εj∣cj∣2 is a quasi-invariant. PM j = 1 1 PN ðxÞ = First, we consider the case where F(x) is chosen to be a smooth (infinitely differentiable) function everywhere, yet nowhere analytic (i.e., it does not have a convergent Taylor series representation). This function, which we will henceforth denote as F1(x). For example, here we construct such a nonanalytic function via Fourier series n = (cid:1)N hn expði2πnxÞ, where the Fourier coefficients hn are F random variables chosen such that their amplitudes drop with n faster than the reciprocal of any polynomial but slower than exponential51–53 (see Supplementary Note 4). This condition guarantees that in the limit N → ∞, the function F1(x) is infinitely differentiable but nowhere ana- lytic. In other words, this function has a Taylor series but its radius of convergence tends to 0 as N → ∞. From a practical point of view, one can choose N to be large enough so as the function F1(x) does not have a proper Taylor series within the range of interest of the intensities involved in our simulations. Figure 3a shows one such possible func- tion F1(x) used in our computations. In this case, numerical simulations carried out on Eq. (5) clearly indicate that the RJ distribution still emerges upon thermalization, as shown in Fig. 3b. While these results clearly support the universality hypothesis for RJ thermalization, they still do not provide compelling evidence, mainly because the function F1(x) is continuous. In this case, the Stone–Weierstrass theorem54 guarantees that it can be still represented by a polynomial expansion, even though it does not correspond to its Taylor series. Thus, in this Fig. 3 | Thermalization of light in nonlinear lattices involving generalized intensity-dependent nonlinearities F(x). a An example of non-analytic function used in our simulations. b Corresponding Rayleigh–Jeans (RJ) distribution (T = 0.15, μ = −2.5) occurring after thermalization. c A discontinuous multi-step function used in our simulations. d Again this nonlinearity leads to a RJ distribution. e A saturable nonlinearity described by F 1 + x, and (f) its corresponding RJ distribution. In ðxÞ = x 3 (b) and (d), the initial excitation conditions are exactly the same and as a result they attain the same RJ allocation, an aspect indicating universality in thermaliza- tion.The insets have been plotted in a manner similar to Fig. 2. As before, here we used M = 100 and the initial mode occupancies are represented by the dashed lines. Nature Communications | (2023) 14:370 4 Article https://doi.org/10.1038/s41467-023-35891-9 scenario one could still argue that the underlying nonlinear interac- tions do arise from a series of higher-order wave mixing terms. A discontinuous nonlinearity function In order to assert the universality of RJ thermalization, i.e., being of a purely entropic (ergodic) origin that goes beyond the wave mixing picture, we next consider a nonlinearity that is described by a dis- continuous multi-step function55–57 such as that depicted in Fig. 3c, denoted as F2(x). Due its discontinuous nature, the function F2(x) cannot be analytically represented by a polynomial expansion across its entire domain. In other words, the wave mixing paradigm com- pletely fails in this case. Interestingly, even in this case, the system thermalizes and reaches a RJ equilibrium state as shown in Fig. 3d, in full accord with theoretically anticipated results. This latter example demonstrates once and for all that optical thermalization in multi- mode systems has a more fundamental origin—rooted in the system’s ergodicity rather than in the intricate nature of the nonlinear interac- tions involved. In other words, the onset of a RJ distribution does not necessarily require the presence of any multi-wave mixing mechan- isms. Instead, it is simply the outcome of the maximizing the entropy itself. Note that the simulations depicted in Fig. 3b, d were carried out for the same parameters and initial conditions (M = 100, P = 10, U = −9.9). Interestingly, despite the profound differences in their nonlinearity, they all settle exactly at the same RJ distribution with T = 0.15 and μ = −2.5. This further supports our hypothesis. In other words, as indicated before, one cannot infer the nature of the inter- molecular collision processes from the Maxwell–Boltzmann distribu- tion as manifested in actual gases. 3 ðxÞ = x Saturable nonlinearity We finally extend this discussion to more realistic material systems. For instance, consider photorefractive crystals where the nonlinearity is saturable57,58 F 1 + x, as shown in Fig. 3e. In the domain where x > 1, F3(x) does not have a Taylor representation but instead has a Laurent series expansion59: F3(x) = 1 − x−1 + x−2 − x−3 + . . . . Obviously, in this regime, the nonlinear interaction cannot be described by a simple wave mixing approach. Yet, assuming that ergodicity holds, and given that two invariants P and U still exist, as per our previous arguments, this should lead to RJ thermalization. This is verified using numerical simulations as shown in Fig. 3f. To ensure the validity of our conclu- sions, the values of the local intensities ∣am∣2 have been monitored during our simulations so as the F3(x) function was predominantly within the Laurent series expansion (see Supplementary Note 5). Discussion In conclusion, we have critically examined the manner in which optical thermalization processes unfold in nonlinear multimode environments and showed that the RJ distribution law is universal: it can manifest itself even in systems where the multi-wave mixing picture fails. These results extend the notion of wave thermalization beyond the original wave turbulence hypothesis that is founded on the premise of wave mixing interactions. In other words, through the use of counterexamples we demonstrated that nonlinear wave mix- ing may be sufficient but by no means necessary. Importantly, it would seem that, in some cases, these processes may not be in fact responsible for thermalization. Instead, our results suggest that RJ equilibrium is obtained because of ergodicity and entropy max- imization as expected by the second law of thermodynamics. These observations, not only support a thermodynamic/probabilistic interpretation of these results, but also provide appropriate foun- dations to expand the thermodynamic formalism in other physical settings governed by classical bosonic interactions. Finally, of inter- est would be to investigate the prospects of devising a formal proof that would dictate the universality of thermalization processes under general nonlinear conditions. Methods Numerical simulation All the simulation results in this work are obtained by numerically integrating the nonlinear equations of motion described by Eqs. (1), and (3)–(5). Due to the finite size of the system, the modal occupancies ∣cj∣2 fluctuate around their equilibrium values. Thus, the final equili- brium state can be evaluated either by calculating the time (distance) average or ensembles average. In this work, we adopted the latter strategy. In particular, for each simulation in Figs. 2 and 3, we have employed 400 ensembles, each of which corresponds to a random initial condition that have the same intensity profile (i.e., same values for ∣cj∣2) but a different phase distribution. Data availability The data that support the findings of this study are available from the corresponding authors upon request. References 1. 2. Bao, H. et al. Laser cavity-soliton microcombs. Nat. Photonics 13, 384–389 (2019). Kippenberg, T. J., Holzwarth, R. & Diddams, S. A. Microresonator-based optical frequency combs. Science 332, 555–559 (2011). 3. Wright, L. G., Christodoulides, D. N. & Wise, F. W. Spatio- temporal mode-locking in multimode fiber lasers. Science 358, 94–97 (2017). Kippenberg, T. J. & Vahala, K. J. Cavity optomechanics: back-action at the mesoscale. Science 321, 1172–1176 (2008). 4. 5. Aspelmeyer, M., Kippenberg, T. J. & Marquardt, F. Cavity opto- mechanics. Rev. Mod. Phys. 86, 1391–1452 (2014). 6. Xia, S. et al. Nontrivial coupling of light into a defect: the interplay of 7. nonlinearity and topology. Light Sci. Appl. 9, 147 (2020). Kirsch, M. S. et al. Nonlinear second-order photonic topological insulators. Nat. Phys. 17, 995–1000 (2021). 8. Wimmer, M. et al. Observation of optical solitons in PT-symmetric lattices. Nat. Commun. 6, 7782 (2015). 9. Xia, S. et al. Nonlinear tuning of PT symmetry and non-Hermitian topological states. Science 372, 72–76 (2021). 10. Bongiovanni, D. et al. Dynamically emerging topological phase transitions in nonlinear interacting soliton lattices. Phys. Rev. Lett. 127, 184101 (2021). 11. Bezryadina, A. et al. Nonlinear self-action of light through biological suspensions. Phys. Rev. Lett. 119, 058101 (2017). 12. Zhang, S. et al. Recent advances in nonlinear optics for bio-imaging applications. Opto-Electron. Adv. 3, 200003 (2020). 13. Tan, M. et al. RF and microwave photonic temporal signal proces- sing with Kerr micro-combs. Adv. Phys. X 6, 1838946 (2021). 14. Reimer, C. et al. High-dimensional one-way quantum processing implemented on d-level cluster states. Nat. Phys. 15, 148–153 (2019). 15. Ramos, A., Fernández-Alcázar, L., Kottos, T. & Shapiro, B. Optical phase transitions in photonic networks: a spin-system formulation. Phys. Rev. X 10, 031024 (2020). 16. Shi, C., Kottos, T. & Shapiro, B. Controlling optical beam thermali- zation via band-gap engineering. Phys. Rev. Res. 3, 033219 (2021). 17. Boyd, R. W. Nonlinear Optics (Academic Press, 2008). 18. Zakharov, V. E., L’vov, V. S. & Falkovich, G. Kolmogorov Spectra of Turbulence I: Wave Turbulence (Springer, 2012). 19. Picozzi, A. et al. Optical wave turbulence: towards a unified none- quilibrium thermodynamic formulation of statistical nonlinear optics. Phys. Rep. 542, 1–132 (2014). 20. Abarbanel, H. D., Rabinovich, M. I. & Sushchik, M. M. Introduction to Nonlinear Dynamics for Physicists (World Scientific, 1993). 21. Kibler, B. & Béjot, P. Discretized conical waves in multimode optical fibers. Phys. Rev. Lett. 126, 023902 (2021). Nature Communications | (2023) 14:370 5 Article https://doi.org/10.1038/s41467-023-35891-9 22. Mafi, A. Pulse propagation in a short nonlinear graded-index 43. Christodoulides, D. N. & Joseph, R. I. Discrete self-focusing in multimode optical fiber. J. Lightwave Technol. 30, 2803–2811 (2012). 23. Nazemosadat, E. & Mafi, A. Nonlinear multimodal interference and saturable absorption using a short graded-index multimode optical fiber. J. Opt. Soc. Am. B 30, 1357–1367 (2013). 24. Pourbeyram, H., Agrawal, G. P. & Mafi, A. Stimulated Raman scat- tering cascade spanning the wavelength range of 523 to 1750 nm using a graded-index multimode optical fiber. Appl. Phys. Lett. 102, 201107 (2013). 25. Teğin, U. et al. Controlling spatiotemporal nonlinearities in multi- mode fibers with deep neural networks. APL Photonics 5, 030804 (2020). 26. Rahmani, B., Loterie, D., Konstantinou, G., Psaltis, D. & Moser, C. Multimode optical fiber transmission with a deep learning network. Light Sci. Appl. 7, 69 (2018). 27. Mumtaz, S., Essiambre, R.-J. & Agrawal, G. P. Nonlinear propagation in multimode and multicore fibers: generalization of the Manakov equations. J. Lightwave Technol. 31, 398–406 (2013). 28. Chen, T. et al. All-fiber passively mode-locked laser using nonlinear multimode interference of step-index multimode fiber. Photonics Res. 6, 1033–1039 (2018). nonlinear arrays of coupled waveguides. Opt. Lett. 13, 794–796 (1988). 44. Selim, M. A., Wu, F. O., Ren, H., Khajavikhan, M. & Christodoulides, D. Thermodynamic description of the near- and far-field intensity patterns emerging from multimode nonlinear waveguide arrays. Phys. Rev. A 105, 013514 (2022). 45. Trillo, S. & Torruellas, W. Spatial Solitons (Springer, 2001). 46. Iwanow, R. et al. Observation of discrete quadratic solitons. Phys. Rev. Lett. 93, 113902 (2004). 47. Siviloglou, G. A. et al. Observation of discrete quadratic surface solitons. Opt. Express 14, 5508–5516 (2006). 48. Xiong, H., Gan, J. & Wu, Y. Kuznetsov-Ma soliton dynamics based on the mechanical effect of light. Phys. Rev. Lett. 119, 153901 (2017). 49. Heinrich, G., Ludwig, M., Qian, J., Kubala, B. & Marquardt, F. Col- lective dynamics in optomechanical arrays. Phys. Rev. Lett. 107, 043603 (2011). 50. Yang, J. Nonlinear Waves in Integrable and Nonintegrable Systems (SIAM, 2010). 51. Trefethen, N. Smoothies: nowhere analytic functions. https://www. chebfun.org/examples/stats/Smoothies.html (2020). 29. Poletti, F. & Horak, P. Dynamics of femtosecond super- 52. Trefethen, N. Notes of a numerical analyst: random smoothies. LMS continuum generation in multimode fibers. Opt. Express 17, 6134–6147 (2009). 30. Pourbeyram, H. et al. Direct observations of thermalization to a Rayleigh–Jeans distribution in multimode optical fibres. Nat. Phys. 18, 685–690 (2022). Newsletter 43 (2021). 53. Bilodeau, G. G. The origin and early development of non-analytic infinitely differentiable functions. Arch. Hist. Exact Sci. 27, 115–135 (1982). 54. Stone, M. H. The generalized weierstrass approximation theorem. 31. Mangini, F. et al. Statistical mechanics of beam self-cleaning in Math. Mag. 21, 237–254 (1948). GRIN multimode optical fibers. Opt. Express 30, 10850–10865 (2022). 32. Podivilov, E. V. et al. Thermalization of orbital angular momentum beams in multimode optical fibers. Phy. Rev. Lett. 128, 243901 (2022). 33. Baudin, K. et al. Classical Rayleigh-Jeans condensation of light waves: observation and thermodynamic characterization. Phys. Rev. Lett. 125, 244101 (2020). 34. Wu, F. O., Hassan, A. U. & Christodoulides, D. N. Thermodynamic theory of highly multimoded nonlinear optical systems. Nat. Pho- tonics 13, 776–782 (2019). 35. Makris, K. G., Wu, F. O., Jung, P. S. & Christodoulides, D. N. Statis- tical mechanics of weakly nonlinear optical multimode gases. Opt. Lett. 45, 1651–1654 (2020). 36. Parto, M., Wu, F. O., Jung, P. S., Makris, K. & Christodoulides, D. N. Thermodynamic conditions governing the optical temperature and chemical potential in nonlinear highly multimoded photonic sys- tems. Opt. Lett. 44, 3936–3939 (2019). 37. Wu, F. O., Jung, P. S., Parto, M., Khajavikhan, M. & Christodoulides, D. N. Entropic thermodynamics of nonlinear photonic chain net- works. Commun. Phys. 3, 216 (2020). 38. Efremidis, N. K. & Christodoulides, D. N. Fundamental entropic processes in the theory of optical thermodynamics. Phys. Rev. A 103, 043517 (2021). 39. Pathria, R. K. Statistical Mechanics (Elsevier, 2016). 40. Wu, F. O. et al. Thermalization of light’s orbital angular momentum in nonlinear multimode waveguide systems. Phys. Rev. Lett. 128, 123901 (2022). 41. Leblond, H., Kremer, D. & Mihalache, D. Few-cycle spatiotemporal optical solitons in waveguide arrays. Phys. Rev. A 95, 043839 (2017). 55. Musslimani, Z. H., Segev, M., Christodoulides, D. N. & Soljačić, M. Composite multihump vector solitons carrying topological charge. Phys. Rev. Lett. 84, 1164–1167 (2000). 56. Musslimani, Z. H., Segev, M. & Christodoulides, D. N. Multi- component two-dimensional solitons carrying topological charges. Opt. Lett. 25, 61–63 (2000). 57. Musslimani, Z. H., Soljačić, M., Segev, M. & Christodoulides, D. N. Interactions between two-dimensional composite vector solitons carrying topological charges. Phys. Rev. E 63, 066608 (2001). 58. Jia, P., Li, Z., Hu, Y., Chen, Z. & Xu, J. Visualizing a nonlinear response in a Schrödinger wave. Phys. Rev. Lett. 123, 234101 (2019). 59. Arfken, G., Weber, H. & Harris, F. Mathematical Methods for Physi- cists: A Comprehensive Guide (Academic Press, 2012). Acknowledgements This work was partially supported by ONR MURI (Grant No. N00014-20-1- 2789), AFOSR MURI (Grant Nos. FA9550-20-1-0322 and FA9550-21-1- 0202), National Science Foundation (Grant Nos. DMR-1420620 and EECS-1711230), MPS Simons Collaboration (Simons Grant No. 733682), W. M. Keck Foundation, U.S.-Israel Binational Science Foundation (Grant No. 2016381), and US Air Force Research Laboratory (Grant No. FA86511820019). Author contributions R.E.-G. and D.N.C. conceived the idea. Q.Z. and F.O.W. developed the theory and conducted the simulations with feedback from A.U.H. All the authors contributed in preparing the manuscript. 42. Pelinovsky, D. E., Sukhorukov, A. A. & Kivshar, Y. S. Bifurcations and stability of gap solitons in periodic potentials. Phys. Rev. E 70, 036618 (2004). Competing interests The authors declare no competing interests. Nature Communications | (2023) 14:370 6 Article https://doi.org/10.1038/s41467-023-35891-9 Additional information Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41467-023-35891-9. Correspondence and requests for materials should be addressed to Ramy El-Ganainy or Demetrios N. Christodoulides. Peer review information Nature Communications thanks Mario Ferraro, Mikko Huttunen, and the other, anonymous, reviewer(s) for their con- tribution to the peer review of this work. Reprints and permissions information is available at http://www.nature.com/reprints Publisher’s note Springer Nature remains neutral with regard to jur- isdictional claims in published maps and institutional affiliations. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/ licenses/by/4.0/. © The Author(s) 2023 Nature Communications | (2023) 14:370 7
10.7554_elife.82232
ReSeaRCh aRTICLe Loss of aquaporin- 4 results in glymphatic system dysfunction via brain- wide interstitial fluid stagnation Ryszard Stefan Gomolka1, Lauren M Hablitz2, Humberto Mestre2,3, Michael Giannetto2, Ting Du2,4, Natalie Linea Hauglund1, Lulu Xie2, Weiguo Peng1,2, Paula Melero Martinez1, Maiken Nedergaard1,2*, Yuki Mori1* 1Center for Translational Neuromedicine, University of Copenhagen, Copenhagen, Denmark; 2Center for Translational Neuromedicine, University of Rochester Medical Center, Rochester, United States; 3Department of Neurology, University of Pennsylvania, Philadelphia, United States; 4School of Pharmacy, China Medical University, Shenyang, China Abstract The glymphatic system is a fluid transport network of cerebrospinal fluid (CSF) entering the brain along arterial perivascular spaces, exchanging with interstitial fluid (ISF), ulti- mately establishing directional clearance of interstitial solutes. CSF transport is facilitated by the expression of aquaporin- 4 (AQP4) water channels on the perivascular endfeet of astrocytes. Mice with genetic deletion of AQP4 (AQP4 KO) exhibit abnormalities in the brain structure and molec- ular water transport. Yet, no studies have systematically examined how these abnormalities in structure and water transport correlate with glymphatic function. Here, we used high- resolution 3D magnetic resonance (MR) non- contrast cisternography, diffusion- weighted MR imaging (MR- DWI) along with intravoxel- incoherent motion (IVIM) DWI, while evaluating glymphatic function using a standard dynamic contrast- enhanced MR imaging to better understand how water trans- port and glymphatic function is disrupted after genetic deletion of AQP4. AQP4 KO mice had larger interstitial spaces and total brain volumes resulting in higher water content and reduced CSF space volumes, despite similar CSF production rates and vascular density compared to wild- type mice. The larger interstitial fluid volume likely resulted in increased slow but not fast MR diffusion measures and coincided with reduced glymphatic influx. This markedly altered brain fluid transport in AQP4 KO mice may result from a reduction in glymphatic clearance, leading to enlargement and stagnation of fluid in the interstitial space. Overall, diffusion MR is a useful tool to evaluate glymphatic function and may serve as valuable translational biomarker to study glym- phatics in human disease. Editor's evaluation This important investigation is of interest to neuroimaging scientists and neurophysiologists studying the glymphatic system. Using a multi- modal approach including magnetic resonance and histolog- ical methods, this work provides substantial data interrogating the effect of removing aquaporin- 4 (AQP4) from the mouse brain parenchyma on the structural morphology and interstitial fluid dynamics stagnation. In particular, the authors provide convincing evidence that deletion of AQP4 in mice results in increased interstitial volume, likely due to increased resistance to parenchymal CSF efflux. *For correspondence: nedergaard@sund.ku.dk (MN); yuki.mori@sund.ku.dk (YM) Competing interest: The authors declare that no competing interests exist. Funding: See page 27 Received: 28 July 2022 Preprinted: 29 July 2022 Accepted: 08 February 2023 Published: 09 February 2023 Reviewing Editor: Saad Jbabdi, University of Oxford, United Kingdom Copyright Gomolka et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 1 of 36 Research article Introduction Aquaporin (AQP) channels facillitate passive water transport across cell membranes (Preston et al., 1992; Li and Wang, 2017). Aquaporin- 4 (AQP4) water channels are highly enriched in astrocytic endfeet that ensheath the cerebral vasculature (Nagelhus and Ottersen, 2013), covering up to 20–60% of the perivascular endfeet membrane facing the vessel wall (Nielsen et al., 1997; Wolburg et al., 2011; Rash et al., 1998; Rasmussen et al., 2022). This high density of water channels in astro- cytic vascular endfeet is remarkable because brain endothelial cells are essentially devoid of the AQP1 water channels expressed by endothelial cells in peripheral tissues (Bonomini and Rezzani, 2010). AQP4 has primarily been studied in the context of pathophysiology, such as ischemia, traumatic brain injury, or hydrocephalus (Iacovetta et  al., 2012; Trillo- Contreras et  al., 2018; Urushihata et  al., 2021; Trillo- Contreras et  al., 2021; Katada et  al., 2014; Bloch et  al., 2006). Genetic deletion of AQP4 (AQP4 KO) in mice results in an increased brain water content (Katada et al., 2014; Li et al., 2009), an expanded interstitial fluid (ISF) volume fraction (Katada et  al., 2014; Yao et  al., 2008), decreased ventricular volume (Trillo- Contreras et al., 2018; Li et al., 2009), and decreased capacity to buffer interstitial potassium (K+) ions (Amiry- Moghaddam et al., 2003b; Strohschein et al., 2011). Yet, why AQP4 was enriched at the endfeet of the astrocytes was unknown until the discovery of the glymphatic system (Iliff et al., 2012). The glymphatic system is comprised of a network of annular perivascular spaces formed by astro- cytic endfeet ensheathing the vascular walls. Perivascular spaces form a low resistance pathway enabling cerebrospinal fluid (CSF) and ISF exchange, promoting the clearance of interstitial solutes from the brain (Iliff et al., 2012). AQP4 KO mice exhibit a 25–60% decrease in glymphatic CSF tracer influx (Mestre et  al., 2018a; Hablitz et  al., 2020; Zhang et  al., 2019), and acute pharmacolog- ical blockade of AQP4 inhibits glymphatic transport using TGN- 020 inhibitor (Huber et  al., 2009; Harrison et  al., 2020; Takano and Yamada, 2020), reducing severity of brain edema and lesion volume after ischemic injury (Igarashi et al., 2011; Sun et al., 2022). However, in cell- based assays TGN- 020 inhibitor failed to inhibit AQP4, bringing into question the true molecular mechanisms of its action (Verkman et al., 2017), discussed in Choi et al., 2021. Deletion of AQP4 also accelerates buildup of neurotoxic protein waste in neurodegenerative models of Alzheimer’s (Xu et  al., 2015; Ishida et al., 2022) and Parkinson’s disease (Cui et al., 2021). In humans, a common single nucleotide AQP4 polymorphism is correlated to changes in slow- wave non- REM sleep and cognition (Ulv Larsen et al., 2020), consistent with increased glymphatic function during sleep (Xie et al., 2013; Eide et al., 2021). Recent reports highlight potential roles of AQP4, especially in edema formation after hypoxia due to the spinal cord injury, and in early and acute phases of stroke (Salman et al., 2022; Kitchen et al., 2020; Sylvain et al., 2021). Thus, the evidence suggests that it is the vascular polarized AQP4 expression in the astrocytic vascular endfeet functionally crucial for fluid transport, and not neces- sarily total AQP4 levels in the tissue (Mestre et al., 2018a; Hablitz et al., 2020; Sylvain et al., 2021; Amiry- Moghaddam et al., 2003a; Eide and Hansson, 2018). Thus, AQP4 plays a strategic role in facilitating CSF influx across the vascular endfeet of astrocytes, waste clearance, and proper sleep architecture. However, it is not clear how AQP4 facilitates glym- phatic fluid transport in part because of the lack of non- invasive whole- brain in vivo measurement of fluid dynamics. The purpose of this study was to characterize the impact of genetic AQP4 deletion in mice on brain- water morphometry and transport by employing state- of- the- art multi- modal in vivo magnetic resonance imaging (MRI) alongside more traditional physiological and histological approaches to measure vascular density, distribution of AQP4 across the brain, brain- water content, ISF volume, and CSF production. Using fully non- invasive high- resolution 3D MR cisternography, we assessed struc- tural differences between the intracranial and CSF space volumes of AQP4 KO and wildtype (WT) mice, as well as the water molecular diffusion and pseudodiffusion using standard diffusion- weighted imaging (DWI) and intravoxel- incoherent motion (IVIM) DWI. Finally, we superimposed these methods with standard dynamic contrast- enhanced MRI to generate the first comprehensive evaluation of brain- fluid movement in the AQP4 KO mouse model. Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 2 of 36 Neuroscience Research article Results AQP4 KO mice having larger brain volumes and smaller CSF spaces First, we hypothesized that AQP4 deletion would alter macroscopic features of the adult murine brain as previously reported (Trillo- Contreras et al., 2018; Katada et al., 2014; Li et al., 2009; Yao et al., 2008; Haj- Yasein et  al., 2011). We employed 3D constructive interference in steady- state (CISS)- based CSF space volumetry and cisternography to delineate highly T2- weighted water signal allowing high- resolution mapping of the brain fluid compartments. We found the brain volume 5–10% larger in AQP4 KO than in WT mice (p<0.01, mean ± SD KO = 521 ± 13 vs. WT = 477 ± 24 mm3; Figure 1A) with no significant differences between animal age (p=0.697), body weight (p=0.7662) or signal- to- noise ratio (p=0.1385) of the 3D- CISS images from two genotypes (Table 1– Methods). This difference in brain volume coincided with ~6% increase in the brain water content measured ex vivo (p<0.05, min- max of 2–11%, KO = 3.66 ± 0.09 vs. WT = 3.46 ± 0.05 ml/g dry brain weight; Figure 1D; Li et al., 2009; Haj- Yasein et al., 2011). Total CSF volume was estimated as 2.4–4.4% of the parenchymal volume among all animals and no difference in delineated CSF space volumes was found between KO and WT (p=0.1255, KO = 15.1 ± 1.9 vs. WT = 17.7 ± 2.4 mm3). Yet the whole segmented CSF space to brain volume ratio was 23–29% smaller in KO compared to WT (p<0.05, KO = 2.99 ± 0.43 % vs. WT = 3.86 ± 0.41 %; Figure 1B). This difference was mainly noted within the ventricular system comprising of the lateral, third and fourth ventricles (p<0.05; Figure  1C), consistent with previous ex vivo (Li et  al., 2009) or lower resolution 2D assessment with low- field strength MRI (Trillo- Contreras et al., 2018; Trillo- Contreras et al., 2021; Li et al., 2009). The most prominent difference was in the lateral (p<0.05, mean ± SEM WT- KO = 0.57 ± 0.15 %) and third ventricles (p<0.01, WT- KO = 0.13 ± 0.03 %), but not the fourth ventricle (p=0.6623; Appendix  1—figure 1A). These changes in the ventricular volume were not driven by CSF production, since we found similar CSF volume production in AQP4 KO and WT mice (Figure  1E) using a newly developed in vivo approach (Liu et  al., 2020). No difference in the CSF spaces of the parietal cisterns was noted (p=0.4589; Appendix 1—figure 1A). Perhaps most surprisingly, no differences were noted in the segmented perivascular CSF space between KO and WT (p=0.1623; Figure 1C) or its individual components (skull base/ Circle of Willis, p=0.9307; basilar artery, p=0.4286; anterior/posterior perivascular spaces, p=0.2486; Appendix 1—figure 1A). Quantification of interstitial space volume using real- time iontophoresis with tetramethylammonium (TMA) (Odackal et  al., 2017), showed that both awake and ketamine/xylazine anesthetized AQP4 KO mice exhibited a larger interstitial space (p<0.05 for both, Figure 1F). The relative enlargement in the interstitial space volume fraction, α, that occured in response to ketamine/xylazine administration did not differ between the two genotypes (p=0.9186, Δαawake- K/X KO = 0.090 ± 0.047 vs. WT = 0.093 ± 0.059; Mann- Whitney U- test). We found no difference in tortuosity between genotypes (p=0.1412, Figure 1F). Thus, deletion of AQP4 is linked to an expansion of the interstitial space volume fraction as well as in total brain volume, with no clear abnormalities in the glymphatic influx paths such as the size of the larger periarterial spaces. Genetic loss of AQP4 alters water diffusivity independent of the microvascular density We next asked how deletion of AQP4 affected the brain’s water mobility within the brain parenchyma and CSF compartments. First, we used a standard DWI model which, by assuming monoexponen- tial signal decay using apparent diffusion coefficient (ADC), provides very sensitive but non- specific scoring for cellularity, the integrity of the cell membranes, and difference in intracellular and ISF volumes (Le Bihan et al., 1988) or their composition and viscosity (Le Bihan et al., 1986; Le Bihan and Iima, 2015). Second, we applied a biexponential intravoxel- incoherent motion (IVIM) DWI model (Le Bihan et al., 1988; Le Bihan et al., 1986) to measure passive molecular water diffusion (D) sepa- rately from the water motion affected by tissue perfusion (Fournet et al., 2017; Federau, 2017; Vieni et al., 2020). ADC and D (IVIM) provided similar results (Pearson’s linear correlation, r=0.94, P<0.0001, Appendix 1—figure 1B), and no differences were found within all 5 large CSF space regions (Table 2, Appendix  1—figure 1C). Both revealed increased slow diffusion measures within the brain paren- chyma, with KO animals exhibiting ADC and D 5.7 ± 1.5 % higher than in WT (Figure  2A and C), consistent with previous ADC estimates in KO animals using lower resolution at 7 Tesla MR and Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 3 of 36 Neuroscience Research article Figure 1. 3D- CISS MRI, CSF and interstitial space volumetry in vivo. Overlaid 3D surface images of the co- registered and averaged 3D- CISS brain volumes, and whiskers- box plots comparing (A) the brain volumes and (B) segmented CSF space volumes from 6 WT and 5 AQP4 KO mice (see Figure 1—source data 1). (C) Overlaid 3D surface reconstruction of the co- registered whole CSF spaces segmented from 3D- CISS from all WT (gray) and KO (blue) animals, along with whiskers- box plot comparisons of main CSF compartments volumes segmented: whole ventricular (left top), lateral (middle top) and third ventricular (right top), and whole perivascular space at the skull base (PVS; left bottom). (D) Whiskers- box plots comparing brain water content from 5 WT and 3 KO ex vivo (Figure 1—source data 2). (E) Whiskers- box plots comparing CSF production rates measured in 6 WT and 5 KO in vivo (Figure 1—source data 3). (F) Whiskers- box plots for the extracellular space volume and tortuosity measured using real- time iontophoresis with tetramethylammonium (TMA) in awake (17 WT, 6 KO) and K/X anesthetized (16 WT, 7 KO) animals (Figure 1—source data 4). Legend: ns- not significant, *-p<0.05, **-p<0.01; Mann- Whitney U- test (A–E), one- way ANOVA with Bonferroni’s post- hoc correction (F). The online version of this article includes the following source data for figure 1: Source data 1. 3D- CISS CSF space volumetry in vivo - source data. Source data 2. Brain water content ex vivo - source data. Source data 3. CSF production in vivo - source data. Source data 4. ISF space volume estimation with TMA - source data. Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 4 of 36 Neuroscience Research article r o f s r e t e m a r a p n o i t i s i u q c a d n a s l o c o t o r p I R M f o s l i a t e d ) B ( d n a , d e m r o f r e p s t n e m i r e p x e 8 n i d e s u i s l a m n a e h t f o s c i t s i r e t c a r a h c i c h p a r g o m e d f o y r a m m u S . 1 e b a T l ; l e g n a p fl i - A F ; o h c e o t e m i t - E T ; n o i t i t e p e r o t e m i t - R T i ; e v e c e r / t i m s n a r t - x R / x T ; n o i t i s i u q c a f o e m i t - A T ; l i o c R M d e o o c - y l l l i a c n e g o y r c – e b o r P o y r C : d n e g e L . y d u t s t n e r r u c e h t n i l d e y o p m e , s e c a p s F S I d n a i F S C n a r b e h t f o n o i t a u a v e l I - P S F e u r T ; e c n e u q e s n o i s s e c e r p e e r f e t a t s - y d a e t s I - P S F ; s n o i t i s i u q c a d e t a e p e r y l t n e d n e p e d n i r o n o i t a m r o f e g a m i r o f s e g a r e v a – . p e R / . v A ; w e v i f o d e fi l - V O F - p ; a n g a m a n r e t s i c – M C ; l e g n a p fl i l e b a i r a v – A F V ; l R T e b a i r a v – R T V ; e c n e u q e s g n g a m i i l r a n a p - o h c e - I P E ; e c n e u q e s n o i s s e c e r p e e r f e t a t s - y d a e t s d e c n a a b e u r t l . ) 5 0 . 0 > > p ( i g n d n fi t n a c fi n g i s i t o n y l l a c i t s i t a t s - S N l ; e b a l i a v a t o n a t a d – A / N ; t s e t - U y e n t i h W - n n a M m o r f l e u a v - p l a c i t s i t a t s p S N S N S N - - - - - 0 3 ± 6 7 1 1 1 ± 5 7 1 4 1 ± 2 8 1 ) D S ± - - - - - ] m p b [ n o i t a r i p s e R n a e m ( l l a r e v O p ) D S ± n a e m ( l l a r e v O ] g [ i t h g e w y d o B S N S N S N A / N S N S N A / N A / N 1 . 2 ± 3 . 7 2 0 . 3 ± 4 . 2 2 2 . 2 ± 3 . 7 2 6 . 4 ± 3 . 4 2 2 . 4 ± 7 . 8 2 0 3 – 5 2 0 3 – 5 2 0 3 – 5 2 p S N S N S N S N S N S N S N S N ) D S ± ] % [ l e a M T W O K ] s k e e w [ e g A n a e m ( l l a r e v O s l a m n a i . f o o N A 8 . 1 ± 8 . 3 1 7 . 0 ± 4 . 0 1 6 . 1 ± 4 . 3 1 0 . 0 ± 0 . 0 1 0 . 1 ± 6 . 4 1 5 . 0 ± 4 . 5 1 0 . 0 ± 0 . 0 1 0 . 1 ± 0 . 2 1 7 . 2 7 7 . 1 4 5 . 5 4 0 5 0 5 . 5 4 5 . 2 6 5 5 6 6 6 4 6 6 0 2 5 5 6 5 - 6 5 8 3 l y r t e m u o v S S C D 3 - I n o i s s e r p x e 4 P Q A y t i s n e d r a u c s a V l n o i t c u d o r p F S C l e m u o v e c a p s F S I g n i s u n o i t a m i t s e A M T t n e t n o c r e t a w n a r B i I W D - R M I R M - E C D ) d e t a g - y r o t a r i p s e r ( i n m 0 3 – 0 2 2 . 1 1 × 4 . 4 1 × 2 . 6 1 7 0 3 3 ) s e c i l s 6 1 , p a g m m 2 . 0 ( 5 . 0 × 5 1 . 0 × 5 1 . 0 3 0 9 0 3 0 0 6 3 ) l a x a i ; c i r t e m u o v ( l I P E - D 2 i n m 7 2 8 . 2 1 × 8 . 2 1 × 2 . 9 1 0 6 2 ) 0 . 1 × 6 . 1 × 0 . 2 ( 3 3 0 . 0 × 3 3 0 . 0 × 3 3 0 . 0 2 0 5 6 . 2 2 . 5 ) l a t t i g a s ; e b o r P o y r C ( m m / 2 s ) 2 ( 1 8 0 3 , ) 2 ( 1 7 0 2 , ) 2 ( 4 6 5 1 , ) 2 ( 7 5 0 1 , 4 5 8 , 9 4 6 , 5 4 4 , 2 4 3 , 8 3 2 , 7 9 1 , 5 6 1 , 3 1 1 , 2 9 , 0 7 , 9 5 , 0 5 , 0 4 : ) ) 1 > . v A ( l s e u a v - b ( * ) s e m i t n o i t a r a p e s d n a n o i t a r u d t n e d a r g r o i f s m 0 1 = Δ d n a 3 = δ ( ) o v i v x e d n a o v i v n i ( I W D - R M i n m 0 9 8 . 2 1 × 8 . 2 1 × 2 . 9 1 1 8 7 1 . 0 × 1 . 0 × 1 . 0 1 5 1 3 6 . 1 6 2 . 3 ) l a t t i g a s n o i t c e n i - j M C a v i I R M - e C D ; e b o r P o y r C ( I P S F - D 3 y r t e m u o v l e c a p s F S C R M I P S F e u r T - D 3 A T e z i s x i r t a M i ] x p / z H [ i h t d w d n a B ) l n o i t a o p r e t n i ( ] 3 m m [ e z i s l e x o V . p e R / . v A ] g e d [ A F ] s m [ E T ] s m [ R T ) n o i t a t n e i r o e c i l s s r e t e m a r a p n o i t i s i u q c a d n a s e c n e u q e s R M B ; l i o c x R / x T ( e c n e u q e S i n m 0 2 h 3 1 × 2 . 6 1 × 2 . 6 1 1 7 6 i n m 0 4 r h 1 1 × 2 . 6 1 × 2 . 6 1 1 7 6 0 . 3 × 1 . 0 × 1 . 0 2 0 . 3 × 1 . 0 × 1 . 0 3 0 9 A F V 1 . 3 1 . 3 R T V 0 0 0 2 1 ; c i r t e m u o v ( l E R A R - D 2 ) l a x a i ; c i r t e m u o v ( l E R A R - D 2 ) l a x a i ) ° 0 9 d n a ° 5 4 : A F V ( ) s m 5 1 , 0 5 , 0 8 , 0 0 1 , 0 0 3 , 0 0 5 , 0 0 8 , 0 0 0 1 , 0 0 0 2 , 0 0 0 4 , 0 0 5 6 , 0 0 0 9 , 0 0 0 2 1 : R T V ( . e m i t n o i t a r a p e r p t n e d a r g o t e u d i , p u - t e s e h t n a h t i r e h g h y l t h g i l s e r a i s n o i t c e r i d g n d o c n e n o i s u f f i d 3 m o r f l s e u a v d e r u s a e m f o s e g a r e v a d e t n e s e r P * i g n p p a m 1 T - o v v i x e m o t n a h p s d a e b o r c M i Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 5 of 36 Neuroscience Research article Table 2. Summary of findings for (A) average ADC and D- IVIM, (B) direction- wise MR diffusion and pseudodiffusion among 21 ROI assessed, along with statistical scoring. Asterisks reflect ‘p’ significance values from the nonparametric Mann- Whitney test comparing diffusion measures between KO vs. WT animals ROI- wise (total n=21, balanced groups), along with the magnitude of the difference expressed by the inequality sign. Legend: OLF- olfactory, CA / RSP- cingulate / retrosplenial, VIS- visual (V1), SS- somatosensory (S1), AUD- auditory, HIP- hippocampus, PERI- perirhinal, TH- thalamus, HAB- habenula, HY- hypothalamus, MB- midbrain, PAG- periaqueductal gray, HB- hindbrain; CP- caudate putamen, WM- white matter; 3V- third ventricle, LV- lateral ventricle, 4V- fourth ventricle, PCS- pericisternal space, CoW- Circle of Willis, CB- cerebellum. NS- no significant difference, *-p<0.05, **-p<0.01, by means of Mann- Whitney U- test. A Average ADC Average D- IVIM ROI OLF CA / RSP VIS SS AUD HIP PERI TH HAB HY MB PAG HB CP WM 3V LV 4V PCS CoW CB Finding KO>WT - - KO>WT KO>WT KO>WT - KO>WT - - KO>WT KO>WT KO>WT KO>WT KO>WT - - - - - - Cerebral cortex Brain stem Cerebral nuclei and tracts CSF space Cerebellum Table 2 continued on next page Significance Finding Significance * NS NS * * ** NS ** NS NS * * ** ** * NS NS NS NS NS NS - - - KO>WT KO>WT KO>WT - KO>WT - - KO>WT - KO>WT KO>WT KO>WT - - - - - - NS NS NS * * * NS ** NS NS * NS ** * * NS NS NS NS NS NS IVIM D*, Fp, Fp x D* - - - - - - - - - - - - - - - - - - - - KO>WT Fp = 0.0649 Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 6 of 36 Neuroscience Research article Table 2 continued B IVIM ADC / D Direction Z (cranio- caudal) Direction X (bilateral) Direction Y (ventral- dorsal) Direction Z (cranio‐ caudal) Direction X (bilateral) Direction Y (ventral‐ dorsal) ROI Finding Signif. Finding Signif. Finding Signif. Finding Signif. Finding Signif. Finding Signif. OLF - CA / RSP D* KO>WT Cerebral cortex Brain stem Cerebral nuclei and tracts CSF space Cerebellum VIS SS AUD HIP PERI TH HAB HY MB PAG HB CP WM 3V LV 4V PCS CoW CB - - - - D* KO>WT - - - - - - - D* KO>WT - - - - - - NS * NS NS NS NS ** NS NS NS NS NS NS NS ** NS NS NS NS NS NS Fp x D* KO>WT - - - - - - - - - - - - - - Fp / Fp x D* KO>WT - - - - - * NS NS NS NS NS NS NS NS NS NS NS NS NS NS ** / * NS NS NS NS NS - - - - - - - - - - - - - - - - - - - - - NS NS NS NS NS NS NS NS NS NS NS NS NS NS NS NS NS NS NS NS NS KO>WT **/* - - KO>WT KO>WT KO>WT NS NS -/* */* */* - - KO>WT KO>WT KO>WT KO>WT - NS KO>WT KO>WT **/** KO>WT - - KO>WT KO>WT KO>WT NS NS -/* */** */* - - KO>WT KO>WT KO>WT KO>WT **/** KO>WT NS NS */- **/ P=0.056 **/* */- */- */- NS NS */- **/- **/- */- - - KO>WT KO>WT KO>WT NS NS */- */- */* KO>WT **/* - KO>WT - - NS */- NS NS KO>WT **/- KO>WT KO >WT - */- */* NS - - - - - - - NS NS NS NS NS NS NS - - - - - - - NS KO>WT */- NS NS NS NS NS NS - - - - - - NS NS NS NS NS NS time- dependent diffusion (Pavlin et al., 2017). This was evident in 10 out of 15 parenchymal regions assessed (Figure  2A), with the largest differences visible in 4 brainstem areas (thalamic, midbrain, periaquaductal gray, and hindbrain) as well as 4 cortical regions (olfactory, somatosensory, auditory, and hippocampal), and the caudate region (min. p<0.05 for all). Overall, these results support the hypothesis of increased ISF space volume, with no difference in slow water mobility within the large CSF spaces in KO animals. To evaluate whether a fast bulk displacement of intravascular water protons due to capillary perfu- sion may contribute to our findings, we performed an additional scoring for differences associated with intra- voxel pseudodiffusive fluid regimes (Le Bihan and Iima, 2015), using IVIM diffusion model. IVIM may reveal pathophysiological impairment in the microcirculation by estimating perfusion fraction (Fp) and pseudodiffusion coefficient (fast diffusion, D*; Le Bihan and Turner, 1992; Henkelman, 1990; Henkelman et al., 1994), but can also detect general fluid dynamics associated with macromolecules such as proteins or biological polymers (Le Bihan, 2019). Transferring IVIM measures to standard perfusion measurements, D* can be associated with mean transit time, Fp with a flowing blood fraction that is correlated with vessel density or cerebral blood volume, and the product of Fp × D* with rela- tive cerebral blood flow in each voxel. We found no differences in average D*, Fp and the product of Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 7 of 36 Neuroscience Research article Figure 2. Diffusion- weighted MRI in vivo and ex vivo, vascularity, and AQP4 cellular surface expression ex vivo. Radar plots showing statistical significances for the differences between the slow diffusion measures among 15 parenchymal ROI assessed, for average ADC (A) and D, and in (B) 3 diffusion- encoding direction separately; (C) Whiskers- box plots for the mean of the average ADC values among 6 AQP4 KO and 6 WT animals analyzed, including: the regions showing the most significant differences (top); exemplary regions showing significant differences (middle); exemplary regions showing no differences (bottom), by means of the Mann- Whitney U- test (Figure 2—source data 1); (D) mean and 95% confidence intervals of ADC and D calculated in a water phantom (+0.001 mM/ml gadobutrol) and 3 water phantoms filled with Sephadex- G25 microbeads of fine, moderate and coarse sizes; (E) Correlation plots of the calculated mean ± SD of ADC and D to the T1 relaxation values and free fluid volumes obtained from the phantoms using MRI (left), and micro- computed tomography (µCT; middle) (Figure 2—source data 2). Single- slices of turbo- spin echo (RARE) MR images from respective phantoms (right, upper) along with µCT images of the central portion of the respective phantoms (right, lower) filled with 1:1 solution of Ominpaque 350 contrast agent and 0.9% NaCl. Semi- transparent red area marks the free fluid space considering only voxels above 75th percentile Figure 2 continued on next page Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 8 of 36 Neuroscience Research article Figure 2 continued of Hounsfield units (HU) intensity distribution in each µCT image; (F) Exemplary immunohistochemistry images (Olympus UplanXApo 10 x/numerical aperture 0.40, ∞/compatible cover glass thickness 0.17 mm/field number 26.5 mm, no immersion liquid) (left) from the hippocampal area from WT and KO animal (magenta vascular labeling) along with whiskers- box plot for the comparison of the mean vascular density among 17 regions analyzed (right) (Figure 2—source data 3); (G) Region- wise correlation plot of calculated ADC to vascular density among 12 regions analyzed with both methods, in WT and KO animals; (H) Exemplary image for vasculature (green, AlexaFluor 488) and AQP4 (magenta) immunohistochemistry staining (Olympus UplanXApo 60 x/numerical aperture 1.42, ∞/compatible cover glass thickness 0.17 mm/field number 26.5 mm, oil immersion) (left), bar- plot comparison of the mean AQP4 channel expression among 11 ROI assessed from 4 WT mice (right) (Figure 2—source data 4); (I) ROI- wise correlation plot comparing average ADC and D with the mean AQP4 channel expression among 10 regions assessed with both methods. Legend: OLF- olfactory, CA / RSP- cingulate / retrosplenial, VIS (V1)- visual, SS (S1)- somatosensory, M1- motorcortex, AUD- auditory, HIP- hippocampus, PERI- perirhinal, INS- insular, TH- thalamus, HAB- habenula, HY- hypothalamus, MB- midbrain, PAG- periaqueductal gray, HB- hindbrain; CP- caudate putamen, WM- white matter; PCS- pericisternal space, CoW- Circle of Willis, EPD- ependymal layer around lateral ventricles; SD- standard deviation; ns- not significant, *-p<0.05, **-p<0.01, by means of Mann- Whitney U- test (A–C, F), Kruskal- Wallis one- way ANOVA with Dunn’s correction (D, H). All correlation plots show respective regression lines along with semi- transparent areas marking 95% confidence intervals of the fitting. The highest obtained Pearson’s linear or Spearman’s range correlation scores are reported and considered significant if correlation value >0.5 with p<0.05, and non- zero regression slope. The online version of this article includes the following source data for figure 2: Source data 1. DWI and IVIM- DWI estimates in vivo - source data. Source data 2. DWI, IVIM- DWI, and T1 estimates in the phantom - source data. Source data 3. Immunohistochemistry: vascular labeling - source data. Source data 4. Immunohistochemistry: AQP4 channel labeling - source data. Fp × D* within all parenchymal regions from KO and WT mice (Table 2A). This suggests similar blood perfusion, and is supported with our histological analysis showing no significant differences in the vascular density between the genotypes (Figure 2F), with a trend towards a small increase within the thalamus and olfactory bulb (11 and 15 %, respectively; both p≈0.1). The lack of change in both IVIM measures and vessel density supports the conclusion that no tangible differences in microcirculation 17O washout between KO and exist between KO and WT and is supported by no difference in the H2 WT at 9.4 Tesla MR (Zhang et al., 2019). Directional water diffusion as a measure of anatomical differences in AQP4 KO mice Application of both DWI and IVIM models confirmed increased slow MR diffusion in the brain paren- chyma, with no difference in the fast MR diffusion (psuedodiffusion) in AQP4 KO mice (Table  2A, Appendix  1—figure 1B–C). We next investigated tissue orientation- specific water mobility restric- tions, by assessing the diffusion separately for cranio- caudal (slice, Z), in- plane bilateral (X), and ventral- dorsal (Y) encoding directions. Overall, the largest differences in both ADC and D were found in the direction parallel to the main orientation of neuronal tracts (i.e. bilateral for the auditory cortex, bilateral and ventral- dorsal for hindbrain, ventral- dorsal for midbrain, cranio- caudal for caudate and thalamus). ADC revealed a similar degree of increased water diffusion in KO compared with WT, among 3 brain stem regions (thalamus, periaqueductal gray, hindbrain; min. p=0.04; Figure 2B) and 4 cortical regions (visual, somatosensory, auditory, hippocampal; min. p<0.03) in both bilateral and ventral- dorsal directions. D highlighted the main differences present only in the cortical somatosensory, audi- tory, and hippocampal regions (Figure 2A vs. Figure 2B). In the cranio- caudal direction, ADC model highlighted main differences between the genotypes in the same 3 brain stem regions (Figure 2B) and 3 cortical regions (olfactory, auditory, hippocampal), and the largest difference was visible in the olfac- tory, thalamic, and the caudate regions (p<0.01). In this direction, D had higher sensitivity (Figure 2B) and the largest difference in D appeared in the parenchymal brain stem areas neighboring the ventric- ular spaces (i.e. thalamus and periaqueductal gray; p<0.01), which possibly reflects the disrupted ependymal cell layer around ventricles and cerebral aqueduct in AQP4 KO mice (Li et al., 2009). Fast water diffusion was altered between KO and WT solely in the cranio- caudal (Z) and bilateral (X) directions (Appendix 1—figure 1C), and were usually associated with a difference in slow diffu- sion markers in at least one orthogonal direction (Table 2). D* was different between the genotypes in the cranio- caudal direction within cingulate/retrosplenial and perirhinal cortex, and white matter (p<0.01, Table 2B). This might reflect possible differences in the rate of water passage orthogonally Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 9 of 36 Neuroscience Research article to the differences in associated fluid perfusion markers. Differences in Fp and Fp  × D* were found only in the olfactory area and within the 3rd ventricle (min. p<0.05; Table 2B), suggesting respectively possible alterations in interstitial fluid efflux pathways and in CSF passage within the smaller volume of 3rd ventricle in KO. Overall, slightly increased average ADC and D along with no difference in average IVIM measures suggest existence of larger interstitial space volume in AQP4 KO, perhaps as a result of increased water exchange time (Urushihata et al., 2021), and without tangible alterations in parenchymal blood perfusion. We tested this hypothesis by mimicking increasing interstitial space volume using three water phantoms filled with Sephadex G- 25 microbeads of coarse, moderate and fine particle size (Figure 2D and E). The coarse beads will have greater spaces between the particles than the fine (see MRI and µCT images in Figure 2E), yet all microbeads possess the same porosity so a similar exchange rate between stored and free water pools is expected. We found both ADC and D increased along with the free fluid space volume surrounding the microbeads, as expected (Lee et al., 2016), with no difference in IVIM measures between the phantoms with microbeads (Appendix 1—figure 1E). This was reflected with high linear correlation between ADC and D, and T1 relaxation times (r=0.977, p<0.05) and free fluid volume fractions estimated (r=0.97, p<0.01) using MRI and µCT in all phantoms (Figure 2D- E). We conclude that these increased ADC and D with increased free water pools support larger interstitial space volume in AQP4 KO compared to WT mice. Reduced gadolinium-based MRI tracer influx into AQP4 KO mouse brain Our noninvasive MR measurements showed increased ADC values along with increased brain volume and reduced CSF space in AQP4 KO mice. Next, we tested whether these measures were associated with reduction in the gadolinium CSF tracer influx from the cisterna magna into the brain parenchyma by means of standard DCE- MRI (Figure  3A). As of particular importance for studying AQP4, it is worth noting that tracer transport (here gadobutrol) does not directly reflect the movement of water. The water can move into the tissue not only through the paracellular gap between astrocytic endfeet but also via diffusive transcellular exchange. The transport of membrane- impermeable CSF tracers, however, is limited to paracellular transport between the gaps of astrocytic endfeet (Salman et al., 2022; Salman et al., 2021). In contrast to prior report (Li et al., 2009), but consistent with the glymphatic model where ventric- ular fluid dynamics are upstream of cisterna magna injections (Iliff et al., 2012), no differences in the tracer distribution were found in the ventricular systems. From the cisterna magna (Figure 3A), the CSF tracer dispersed via the subarachnoid space cisterns to the Circle of Willis, and then dorsally along the middle cerebral artery into the brain parenchyma and anteriorly toward the olfactory bulb (Figure 3B), consistent with the previous reports using fluorescent tracers (Iliff et al., 2012; Mestre et al., 2018a). Importantly, there were no differences in the tracer distribution within the perivascular space at the basal cistern (Figure 3B – ‘Circle of Willis’), consistent with the perivascular space volume not differing between the two genotypes (Figure  1C, Appendix  1—figure 1A). Yet, the peak and overall magnitude of the parenchymal signal enhancement were significantly lower in KO than WT brains (lowest p<0.01; Figure 3B–D). Differences were especially visible in the striatum (p<0.01), thal- amus (p<0.05), hippocampus (p<0.01), and visual and cingulate/retrosplenial cortex (p<0.01 for both; Figure 3C). Within parenchyma, the difference increased with time from infusion and was largest in the cortex and hippocampus (p<0.001 at T=80 min; Figure 3C and E, and Appendix 1—figure 2A). Finally, while the tracer in WT accumulated around the venous sinuses as previously reported (Iliff et al., 2013), the AQP4 KO mice exhibited substantially less accumulation (p=0.0148; Figure 3D). The reduced parenchymal influx of contrast agent after cisterna magna injection in KO vs. WT mice are consistent with current models of glymphatic function (Iliff et al., 2012; Mestre et al., 2018a; Kress et al., 2014; Mestre et al., 2018b). Next, to identify possible differences in the tracer dynamics between KO and WT mice, we calculated area under curve, arrival time, time- to- peak, peak intensity, and duration of significant from baseline parenchymal tracer accumulation (Table  3). Overall, the arrival time was similar, if slightly longer in KO than in WT mice. The largest delay in the tracer arrival time was visible in the lateral ventricles and caudate nucleus in KO (Table  3). The duration of time- to- peak was  ~30% longer in KO among all regions, with the largest difference visible in the hippocampus, midbrain, Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 10 of 36 Neuroscience Research article Figure 3. Dynamic contrast- enhanced MRI in vivo (Figure 3—source data 1). (A) 3D multiplanar reconstructions of dynamic- contract enhanced (DCE) MRI – sagittal slices from mean images of 6 WT (top) and 5 AQP4 KO (bottom) using 3D- FISP, after applying gadobutrol injection via cisterna magna (CM); mean ± SD DCE time- curves from WT (gray) and AQP4 KO (blue): (B) main CSF compartments ventrally and caudo- cranially from CM, (C) parenchymal regions where significant differences between WT and KO were found, (D) ventricular and CSF efflux regions. (E) 3D multiplanar reconstruction of DCE- MRI from mean 3D FISP images of 6 WT animals, along with Allen Brain Atlas- based segmentation maps color- coded according to p- significance value from nonparametric Two- way ANOVA with post- hoc, showing WT vs. KO differences in the CSF tracer dynamics at 60th minute after gadobutrol injection start. Legend: ns- not significant, *-p<0.05, **-p<0.01, by means of nonparametric Two- way Anova with Bonferroni’s post- hoc. The online version of this article includes the following source data for figure 3: Source data 1. DCE- MRI in vivo - source data. thalamus, cerebellum and superior sagittal sinus, though this trend was not significant due to high variability. In regions neighboring or related to the CSF spaces and dorsal cortex (i.e. olfactory and somatosensory cortex, hypothalamus, periaqueductal gray, hindbrain or perivascular) time- to- peak was longer in WT, which might be an overall effect of lower tracer penetration in KO. The relative WT- KO differences of peak intensities and parenchymal accumulation durations were always positive and moderately correlated (Pearson’s linear correlation of relative peak intensity to relative duration difference of r=0.7212, p=0.0003; Appendix  1—figure 2B), which suggests that duration was shorter in KO due to smaller tracer penetration into the brain. This would be confirmed with the area under the DCE curve that consistently was smaller in KO (Table 3, ‘Mean AUC’). Together, delayed tracer arrival time and time- to- peak, lower peak intensity, and duration of tracer accumulation, as well as smaller area under the DCE curve in KO mice indicate a reduced tracer parenchymal influx and higher parenchymal resistance associated with lack of AQP4 chan- nels (Mestre et al., 2018a). Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 11 of 36 Neuroscience Research article ; e s i w - I O R t s e t - U y e n t i h W – n n a M c i r t e m a r a p n o n m o r f ) t a t s - p ( s e u a v l l a c i t s i t a t s - p d e t a c o s s a h t i i w g n o a l s e v r u c E C D e h t r e d n u s a e r a m o r f n a e m – C U A n a e M : d n e g e L . e s i w - n a r t s i s e u a v l ) D S ( i n o i t a v e d d r a d n a t s ± n a d e m d n a i t s e t k n a r - d e n g i s n o x o c W d e l i l i a t - o w t a g n i s u s c i t s i t a t s e s i w - r i a p c i r t e m a r a p n o n i d e t a c o s s a h t i w g n o a l , i e c m e t a m r e t t i l T W 6 d n a O K 5 n i ) I W D - R M l f o s n a e m y b d e z y a n a e s o h t g n h c t a m i ( I O R 1 2 m o r f s r e t e m a r a p d e v i r e d - E C D . 3 e b a T l ; l o r t n o c e p y t d l i w – ) ( + / + 4 p q A ; l n o i t a u m u c c a r e c a r t l a m y h c n e r a p e h t g n k c m m i i i , t n e m e c n a h n e l a n g i s e n i l e s a b e h t m o r f t n e r e f f i d y l t n a c fi n g i s i f o n o i t a r u d – n o i t a r u D r e c a r t l a i t i t s r e t n i f o n o i t a r u d r e t r o h s e h t s n a e m n g i s ‘ - ‘ e v i t a g e n ( e c n e r e f f i d e v i t a e r l % 0 0 1 × T W ) / ) O K - T W ( ( – Δ . l e r ; e c n e r e f f i d O K - T W – Δ i ; e c m O K 4 P Q A – ) - / - ( 4 p q A d r a d n a t s f o n a e m – ¶ . c o h - t s o p s ' i n o r r e f n o B h t i w a v o n A y a w - o w T – # . l 2 e b a T r o , 3 d n a 2 i s e r u g F n o i t p a c e e s i , s n o i t a v e r b b a I O R r o F i ; ) e c m O K n i l n o i t a u m u c c a . 1 0 . 0 < p – * * , 5 0 . 0 < p – * . e s i w I O R s n o i t a v e d i 8 . 5 1 – 9 . 1 2 – 4 . 1 5 – 6 . 8 2 – 2 3 5 2 8 1 5 1 0 0 1 - < A / N 0 . 3 6 – 5 . 0 1 – 2 . 3 1 – 8 . 0 1 – 1 . 8 – 6 . 1 1 – 4 . 4 2 – 3 . 0 1 – 8 . 0 1 – 5 . 3 1 – 1 . 1 1 – 2 . 6 1 – 6 . 2 3 – 8 . 0 1 – 8 . 9 6 – 3 . 1 2 – 7 1 4 3 3 3 3 3 4 3 8 3 1 3 5 3 3 3 2 3 2 3 1 3 9 2 3 3 6 1 7 3 8 3 2 3 7 3 1 2 4 2 6 4 8 3 8 3 7 3 7 3 3 4 1 4 9 3 7 3 7 3 6 3 7 3 3 4 7 3 3 5 7 4 0 7 0 9 0 9 5 2 9 3 6 6 3 7 1 1 A / N A / N 0 9 0 9 0 9 0 9 4 5 0 9 1 6 1 6 0 9 9 6 7 4 0 5 1 6 0 9 0 9 7 6 4 7 7 5 8 5 8 5 1 2 3 5 1 3 7 2 8 5 8 3 6 1 0 2 3 3 8 5 5 7 1 3 0 8 0 9 0 9 2 3 0 9 0 9 0 9 0 9 0 9 9 5 4 8 5 7 6 6 0 9 3 7 0 5 9 5 6 7 0 9 0 9 5 7 3 4 9 5 4 5 2 1 7 6 5 4 3 5 3 5 4 5 3 2 2 4 5 3 8 2 4 5 7 3 5 1 3 2 4 3 4 5 8 3 9 2 5 . 5 2 6 . 5 5 7 . 2 7 5 . 2 1 0 . 0 5 2 . 2 5 4 . 5 2 2 . 6 2 0 . 0 2 6 . 4 1 6 . 5 3 1 . 1 1 2 . 4 1 9 . 5 3 8 . 3 2 6 . 4 1 5 . 6 1 1 . 6 1 7 . 1 2 4 . 1 7 5 . 7 1 7 ± 2 3 7 ± 7 3 9 1 ± 4 2 9 7 2 1 9 7 5 2 2 4 4 5 4 6 3 6 0 1 7 2 3 3 8 0 1 6 4 9 5 1 6 5 4 5 0 1 3 2 1 6 5 1 1 3 4 1 1 5 2 0 8 9 3 3 7 8 1 1 5 1 3 2 1 2 2 0 8 7 8 7 8 0 5 3 3 3 1 9 3 5 0 1 7 9 3 4 2 2 0 8 1 7 5 1 7 7 7 9 6 5 9 1 - 1 - 4 0 8 1 – 7 - 8 - 1 - 5 7 1 – 5 4 1 - 2 - 0 0 1 1 - 0 4 2 – 2 1 – 2 5 0 9 0 9 6 1 9 8 1 8 0 9 0 9 0 9 4 3 3 8 8 3 8 3 0 9 3 5 4 2 1 3 6 4 0 9 8 8 9 5 1 6 9 8 9 8 0 2 9 8 3 6 3 8 2 8 9 8 9 3 6 6 3 4 2 4 9 8 1 5 4 2 1 4 5 4 0 9 4 6 7 4 0 0 1 - 1 - 0 2 - 1 - 0 2 - 0 1 - 2 - 0 4 - 2 - 0 0 1 - 5 - 2 - 1 - 6 7 8 7 7 5 . 7 5 3 – 2 1 1 4 4 0 1 3 – 2 8 5 . 6 1 3 – 2 6 1 1 3 7 6 7 7 6 7 5 . 5 4 3 – 2 9 4 3 8 3 – 2 4 5 . 4 1 3 – 2 5 6 1 6 S N S N * S N S N S N * * S N S N S N S N * * S N S N * * * * S N S N * S N 9 1 6 5 6 1 6 9 2 5 8 3 4 0 4 3 2 2 6 1 9 5 6 2 2 5 0 . 0 7 5 5 2 2 9 9 1 6 6 7 6 3 1 0 4 7 0 2 7 2 6 6 1 9 9 6 1 7 3 5 4 5 4 8 5 2 3 3 1 5 3 1 7 4 3 9 7 3 2 8 0 9 7 5 2 1 6 4 7 4 2 2 0 4 7 5 2 4 2 1 8 7 6 1 1 5 4 4 5 1 9 2 9 6 3 7 1 3 5 9 6 3 0 9 9 1 6 7 2 3 2 8 9 4 2 1 3 6 1 5 1 1 4 4 0 1 8 2 2 2 1 5 5 5 6 0 3 4 1 1 5 1 6 8 8 2 8 2 3 8 5 9 6 3 7 Δ . l e r ] % [ 4 p q A ) - / - ( 4 p q A ) + / + ( ) - / - ( 4 p q A ) + / + ( 4 p q A d n E t r a t S d n E t r a t S ] % [ Δ . l e r 4 p q A ) - / - ( 4 p q A ) + / + ( Δ 4 p q A ) - / - ( 4 p q A ) + / + ( Δ 4 p q A ) - / - ( 4 p q A ) + / + ( t a t s - p 4 p q A ) - / - ( 4 p q A ) + / + ( e m i t l n o i t a u m u c c a r e c a r t l a i t i t s r e t n I ] n m i [ n o i t a r u D ] n m i [ ] . u . a [ y t i s n e t n i k a e P ] n m i [ k a e p o t - - e m T i ] n m i [ e m i t l a v i r r A ] . u . a [ C U A n a e M O K . s v T W # s e i r e s - e m i t ? t n e r e f f i d S N * * * * S N S N * * 2 5 0 . 0 = P * S N S N * S N S N S N * * S N S N S N S N * S N F L O P S R / A C S I V S S D U A I P H I R E P H T B A H Y H B M G A P B H P C M W M C S C P V 3 V L S S S B C I O R l a c i t r o C m e t s i n a r B / i l e c u N s t c a r t e c a p s F S C l a d u a C 1 0 0 0 . 0 < P , O K > T W 1 0 0 0 . 0 < p , O K > T W , O K ~ T W 1 7 9 2 . 0 = p 1 0 . 0 < p , O K < T W 1 0 0 0 . 0 < p , O K > T W ) t s e t k n a r - d e n g i s n o x o c W l i ( e c n e r e f f i d T W . s v O K 7 2 ± 1 8 3 2 ± 3 6 8 . 2 ± 5 . 6 2 . 2 ± 0 . 5 ¶ 6 0 3 1 ± 3 1 0 4 ¶ 3 5 6 1 ± 5 2 1 6 D S ± N a D e M I Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 12 of 36 Neuroscience Research article AQP4 distribution determines regional differences in the parenchymal fluid flow Increased AQP4 KO brain volume is associated with decreased ventricular volumes, increased slow MR diffusion markers, reduced glymphatic transport, and small region- specific, non- significant differences in the vascular density (Figure 2F). Using immunohistochemistry for AQP4, we found heterogenous expression in WT animals, with the highest expression in the hippocampus, thalamus, and habenula (Figure  2H; Yao et  al., 2008; Hubbard et  al., 2015). Next, we asked which factor reflected the changes in MR- DWI or DCE- MRI derived markers more: vascular density or local AQP4 expression? The vascular density did not correlate with ADC or IVIM measures in both KO and WT animals (Figure 2G, Appendix 1—figure 1D). However, when investigating tracer accumulation dynamics, we found vascular density having a low to moderate correlation with arrival time (Spearman’s rho=0.6715, p=0.0201) and moderate to high correlation with time- to- peak in WT animals (rho=0.8411, p=0.0013; Appendix 1—figure 2C). In KO animals, only a low range correlation with time- to- peak was identified (rho=0.6084, p=0.0398). Interestingly, in WT mice AQP4 expression did not correlate with changes in the MR tracer dynamics (Appendix  1—figure 2D), but had a moderate linear correlation with both ADC (r=0.77, p=0.0092) and D (r=0.74, p=0.0137; Figure 2I). Finally, there was no correlation found between the vascular density and AQP4 expression (both Pearson’s and Spearman’s correlation absolute value <0.52 and p>0.16), or between MR diffusion and tracer dynamics parameters (correla- tion value <0.5 and p>0.05 for any comparison) except for low correlation between the area under the DCE curve and ADC for both genotypes (WT: r=0.63, p=0.0272; KO: r=0.64, p=0.0239). Thus, high vascular density predicts fast CSF tracer influx consistent with the notion that CSF is pumped in along the perivascular spaces surrounding arteries (Iliff et al., 2012; Iliff et al., 2013). Also, the correlation between slow MR diffusion, ADC and D, with AQP4 density across 10 regions in WT mice (Figure 2I) suggests that AQP4 expression is higher in regions with relatively larger interstitial fluid volume, possibly reflecting that AQP4 is upregulated in response to stagnation of interstitial fluid in wildtype mice. Discussion Using non- invasive high- resolution MR CSF space volumetry and cisternography in vivo, we found increased brain volume and decreased CSF pool volume, mainly in the ventricular compartment, in mice genetically lacking the water channel AQP4, alongside increased brain- water content (Table 4). Changes in brain water content and CSF pool size were not explained by changes in CSF production or the volume of the larger perivascular CSF spaces. Next, we investigated the brain water mobility in AQP4 KO animals using standard MR- DWI and IVIM- DWI. Measures of fast MR diffusion and vascular density were also unchanged between KO and WT mice, although KO animals exhibited a higher variability in vascular density. Slow diffusion (ADC and D) estimates were increased within the paren- chyma of KO animals and so was the cortical interstitial space volume measured using the real- time ionophoresis TMA technique. Finally, we asked whether AQP4 expression or local vascular density correlated to slow diffusion, fast diffusion (IVIM), or with measures of decreased CSF MR tracer influx into the AQP4 KO brains. In WT animals, slow diffusion measures were correlated with AQP4 expres- sion and differential vascular density was nonlinearly correlated to measures of tracer accumulation. AQP4 KO animals had a very low correlation of vascular density to time- to- peak tracer accumulation. These correlations suggest that the vascular network provides a highway for perivascular CSF inflow and thereby drives the initial tracer distribution within the parenchyma. Increased AQP4 expression in regions manifesting high ADC or D in WTs, possible due enlarged interstitial volume, may reflect a compensatory upregulation of AQP4 due to fluid stagnation consistent with the notion that AQPs reduce parenchymal resistance and facilitate the water and solute movement. Consistent with this hypothesis, recent studies report dynamic AQP4 relocalization leading to changes in signaling path- ways (Salman et al., 2022). Thus, our data overall suggest that the markedly altered brain fluid trans- port in AQP4 KO mice may result from a reduction in glymphatic fluid export, leading to stagnation of ISF and enlargement of the interstitial space. The interstitial fluid stagnation will in turn reduce CSF influx and give rise to an overall reduction in glymphatic transport. Why would AQP4 play a role in export of brain interstitial fluid? In general, AQPs increase membrane water permeability and are present in kidney and exocrine organs where fluid transport is Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 13 of 36 Neuroscience n o i s u f f i d l l a n i d n a e g a r e v a r o f t n e s e r p e c n e r e f f i d ( l e c i r t n e v d r 3 , e c a p s l r a u c i r t n e v l i o i t a r e m u o v n a r b / e c a p s F S C r e l l a m s % 9 2 – 2 2 l s e m u o v n a r b r e g r a i l % 0 1 – 5 - - y h p a r g o n r e t s i c D 3 Research article l l i a c g o o t s i h d n a l i l a c g o o i s y h p 5 d n a I R M 3 f o s n a e m y b i , s l a m n a T W d n a O K n e e w t e b d n u o f s e c n e r e f f i d t s e g r a l e h t f o s n o g e r i e h t s t h g i l h g h i t n o f c i l a t i l d o B . y d u t s t n e r r u c e h t n i d e t n e s e r p s g n d n fi i f o y r a m m u s e v i t p i r c s e D . 4 e b a T l e c n e r e f f i d t s e g r a l f o n o g e R i T W o t d e r a p m o c O K n i i s g n d n fi l a r e n e G . d e i l p p a s d o h t e m t n e m s s e s s a o v v i n i i g n g a m i e c n a n o s e r c i t e n g a M t n e m e r u s a e M , s n o g e r i y a r g l a t c u d e u q a i r e p , i n a r b d n h i ) s n o i t c e r i d , s u m a a h t l - , s u p m a c o p p h i l , s u m a a h t e h t n i n o i s s e r p x e t s e g r a l e h t n i n o i s s e r p x e 4 P Q A s u o n e g o r e t e h ) T W y l n o ( l a u n e b a h i n a r b n o i s s e r p x e 4 P Q A d n a l s u m a a h t e h t n i y t i s n e d r a u c s a v l r e g r a l r o f d n e r t a e r a y r o t c a f l o T W o t y t i s n e d r a u c s a v l r a l i m i s T W o t n o i t c u d o r p F S C r a l i m i s y t i s n e d r a u c s a V l n o i t c u d o r p F S C t n e t n o c i r e t a w n a r b r e g r a l % 6 ~ t n e t n o c r e t a w n a r B i l y g o o t s i h d n a o v v i x e e c n e r e f f i d t s e g r a l f o n o g e R i T W o t d e r a p m o c O K n i i s g n d n fi l a r e n e G t n e m e r u s a e M r e g r a l l e m u o v e c a p s F S I A M T s i s e r o h p o n o i e m i t - l a e R o v v i n i s t n e m e r u s a e m l i l a c g o o t s i h d n a l l i a c g o o i s y h P s u p m a c o p p h d n a i x e t r o c y r o t i d u a D d n a C D A i r e h g h % 6 – 5 s u p m a c o p p h i , I O R l a c i t r o c : x u fl n i - s u n i s l a t t i g a s r o i r e p u s : x u fl f e - n o i t a u c a v e d n a x u fl n i r e c a r t l a m y h c n e r a p d e c u d e r ` a n g a m a n r e t s i c a i v g n g a m i i r e c a r t F S C c m a n y D i l e c i r t n e v d r 3 e h t n i y l n o * D × p F d n a p F r e h g H i i g n g a m i i d e t h g e w - n o i s u f f i d D 2 e v i s a v n i - n o n e v i s a v n i Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 14 of 36 Neuroscience Research article driven by relatively small osmotic gradients produced by plasma membrane ion transporters (Salman et  al., 2022; Salman et  al., 2021). AQPs facilitate near- isomolar transepithelial fluid transport and AQP deletions have previously been shown to be associated with reduced fluid secretion (Verkman, 2009). A recent study showed that AQP4 sharply reduces outflow of interstitial fluid (Plá et al., 2022). It has been shown that brain interstitial fluid leaves by multiple pathways including perivenous efflux, along cranial and spinal nerves and also along the ventricular and pial surfaces (Rasmussen et  al., 2022). The fluid is then absorbed by meningeal and cervical lymphatic vessels for export to the venous system. AQP4 is intensely expressed in glia limitans facing the ependymal layers of the ventricles and also the pial membrane surrounding the brain surfaces. Recent studies showed dynamic and revers- ible AQP4 channel membrane relocalization for regulation of local water homeostasis (Salman et al., 2022) in response to hypothermia and hypotonic treatment in cultured rat (Kitchen et al., 2015) and human primary cortical astrocytes (Salman et al., 2017) without change in AQP4 mRNA levels, and nonuniformity of these responses among AQP4 subtypes (Ciappelloni et  al., 2019). Similarly, the vascular endfeet of astrocytes plastered around both arterioles, capillaries and veins in both rodent and human brain express high levels of AQP4 in the membrane facing the vessel wall (Nielsen et al., 1997; Rash et  al., 1998; Oberheim et  al., 2009; Zeppenfeld et  al., 2017). In fact, the intensity of AQP4 immunofluorescence signal in endfeet plastered around veins is almost double of those surrounding arteries (Iliff et al., 2012). Fibrous astrocytes in white matter tracts are also recognized for their high AQP4 expression (Lundgaard et al., 2014). Thus, AQP4 is present at high density at most if not all glymphatic efflux paths, and therefore also in a prime position to mediate outflow of interstitial fluid. Some of the most critical findings presented here were that the changes in the brain water content, CSF pool size, and interstitial spaces were not due to changes in CSF production in AQP4 KO mice or in perivascular space volume. This is consistent with growing evidence that CSF distribution is depen- dent on arousal state and circadian timing rather than being dictated by the rate of CSF production (Hablitz et al., 2020; Xie et al., 2013; Liu et al., 2020). Furthermore, vascular density or fast diffusion estimates were not altered by AQP4 deletion, suggesting that the vasculature and blood perfusion remains the same even after genetic deletion of AQP4. Instead, we show that slow MR diffusion measures are increased mostly due to an enlarged inter- stitial space. Only slight 5–6% increase in the mean ADC and D might result from superposition of opposing effects of reduced transmembrane permeability (reducing ADC) and increased extracellular space (increasing ADC) as concluded previously using time- dependent diffusion MRI and Latour’s model of long- time diffusion behavior (Pavlin et al., 2017). Similarly, evaluation of ADC using multi- b- value- multi- diffusion- time DWI provided higher ADC’s in healthy hemispheres of mice subjected to contralateral ischemic stroke, reflecting larger interstitial space in KO (Urushihata et  al., 2021). The enlarged interstitial space in both awake and anesthetized AQP4 KO mice is also consistent with previous reports under anesthesia (Yao et  al., 2008; Amiry- Moghaddam et  al., 2003b; Papado- poulos et al., 2004). Our findings in the water phantom filled with Sephadex microbeads of similar porosity but different sizes also confirmed increase in both ADC and D resulting from increased free fluid volume (Figure 2D). Our findings support a glymphatic model whereby cerebrospinal fluid is driven by vascular move- ment into the brain alongside the perivascular space, and AQP4 at the vascular astrocytic endfoot enables ISF and solute movement from the parenchyma. These changes in the micro- and macro- scopic CNS fluid compartments could be due, specifically, an increased resistance towards glymphatic fluid efflux caused by lack of AQP4 channels along the perivenous space (Xie et al., 2013; Plá et al., 2022). Fluid accumulation in the interstitial space would, in turn, increase resistance toward periarte- rial CSF influx explaining the overall suppression of glymphatic transport. One group has previously published evidence against the importance of AQP4 in glymphatic fluid transport (Smith et  al., 2017; Smith and Verkman, 2018). This group’s finding is contradicted by multiple independently generated datasets using different transgenic lines of mice with deletion of the AQP4 or α-syntrophin (Snta1) genes, different fluorescent and radioisotope- labeled tracers of 17O at 9.4T MRI (Zhang et al., 2019) 4.5–70 kDa size (Mestre et al., 2018a; Hablitz et al., 2020) or H2 across a wide age range of 6–24 weeks. A meta- analysis of all published studies demonstrated a signif- icant decrease in CSF tracer transport in AQP4 KO mice compared to wildtype and meta- regression suggested that differences in anesthesia, age, and tracer delivery explained the opposing results. Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 15 of 36 Neuroscience Research article Since we have discussed the controversy in details (Mestre et  al., 2020), the only additional note is that observations reported here add additional support to a key role of AQP4 in glymphatic flow. Glymphatic disruption has been observed in preclinical models of Alzheimer’s disease (AD; Peng et al., 2016; Harrison et al., 2020). In humans, alterations in MR- DWI have been seen in mild cogni- tive impairment and symptomatic AD (Kantarci et al., 2001; Kantarci et al., 2002; Kulkarni et al., 2020). Here, a linear correlation between vascular density and time- to- peak tracer accumulation across multiple brain regions was found in both KO and WT mice. We also found overall increased slow water diffusion in AQP4 KO mice, which was most pronounced in regions that normally exhibit higher AQP4 expression. These observations suggest that vascular density determines the speed of tracer distribution whereas AQP4 expression determines interstitial fluid exchange. Interestingly, increased slow MR diffusion is also found in AD patients, compared to those with mild cognitive impairment and healthy controls (Bergamino et al., 2020). Perhaps known alterations in AQP4 distri- bution and abundance in AD (Zeppenfeld et al., 2017; Boespflug et al., 2018; Simon et al., 2018) drive increased interstitial fluid stagnation, decreasing glymphatic function past what is expected in natural aging (Kress et  al., 2014). Extending this speculation, perhaps regional loss of AQP4 may explain subregion- dependent susceptibility to neurodegeneration by driving local interstitial fluid and protein stagnation increasing the risk of aggregation prone proteins. The complex approach we used here, based on cutting- edge non- invasive MR techniques including high- resolution 3D non- contrast cisternography with sophisticated automatic CSF volume estimation, DWI along with IVIM- DWI, stan- dard DCE- MRI, and traditional ex vivo histology and physiological measures, may have potential to answer some of these fundamental questions on how cellular pathology and glymphatic dysfunction contributes to proteinopathies. Materials and methods Animals and experimental setups Animal approval was received from both the University of Copenhagen Animal Experiment Inspec- torate and the University of Rochester Medical Center Committee on Animal Resources. The same AQP4 knockout (Aqp4(-/-), KO) mouse line (Thrane et al., 2011), regularly cross- breed with wildtype (WT) mice was bred in both Copenhagen and Rochester, and in total n=97 10–16 weeks old AQP4 KO and WT (Aqp4(+/+)) littermates mice on a C57BL/6 background (Mestre et al., 2018a) was used (Table 1A). All animals were group- housed (up to 5 mice/cage) with ad- libitum access to food and water, temperature (22 ± 2°C), and humidity- controlled (55 ± 10%) environment with a 12/12 hr light/ dark cycle. The animals were subdivided randomly into groups, and each group underwent one of the experimental paradigms including three in vivo magnetic resonance imaging (MRI) experiments: 3D CSF space volumetry and cisternography, 2D diffusion- weighted imaging (DWI), dynamic contrast- enhanced (DCE) MRI via cisterna magna; or in vivo CSF production, and three ex vivo measurements: vascular density, AQP4 expression, brain water content. At the end of each in vivo experiment, mice were sacrificed via K/X overdose and cervical dislocation. MRI All MRI scanning was performed at 9.4T MRI device (BioSpec 94/30USR, Bruker BioSpin, Ettlingen, Germany) in the head- first prone position with animal’s body temperature maintained at 37 ± 1°C with a thermostatically- controlled waterbed and monitored, along with the respiratory rate, by an MR compatible remote monitoring system (SA Instruments, NY, USA). Non-contrast MRI volumetry and cisternography To achieve high- spatial resolution of MRI CSF space volumetry and cisternography a 3D constructive interference in steady- state (CISS) sequence along with a cryogenically cooled Tx/Rx quadrature- resonator (CryoProbe, Bruker BioSpin) and 240 mT/m gradient coil (BGA- 12S) were used. During MRI, the animals (5 KO and 6 WT) were anesthetized under Ketamine/Xylazine (i.p. K/X: 100/10 mg/kg) and underwent acquisition of two 3D- TrueFISP volumes of opposite phase encoding direction (i.e.: 0° and 180°) (Table  1B), for further 3D- CISS image calculation. The complete MRI protocol lasted over an hour so every animal was implanted with a permanent intraperitoneal PE- 10 catheter in the abdominal area, connected to a 1 mL syringe filled with K/X solution. The syringe was kept outside Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 16 of 36 Neuroscience Research article MR during whole imaging, and animals received a single supplementary dose of K/X after the first TrueFISP volume acquired. No difference in age (p=0.697 for KO vs. WT using Mann- Whitney U- test), body weight (p=0.7662), respiration rate during MRI (p>0.99) as well as signal- to- noise ratio (SNR) of computed 3D- CISS images (3.5 ± 0.4 vs 3.7 ± 0.2; p=0.1385) was found between KO and WT animals (Table 1A) so no animals were excluded from further analysis. DWI and IVIM-DWI To assess differences in the brain water mobility between the genotypes, echo- planar- imaging (EPI)- based diffusion- weighted imaging (DWI) was performed using a room- temperature volumetric Tx/Rx resonator (in. ø40 mm) and 1500mT/m gradient coil (BFG6S, Bruker). The animals (6 KO and 6 WT) were anesthetized with K/X (i.p. 100/10 mg/kg) and underwent DWI with 17 b- values measured in 3 orthogonal directions of diffusion encoding gradients (Table 1B). To reduce the effect of respiratory motion (Federau et al., 2013), all DWI images were acquired with respiratory- gating in exhale, assisted by the remote monitoring system (see above). To minimize the influence of deep anesthesia and long scanning time on the measurements, the imaging protocol encompassed solely DWI lasting <40 min so no supplementary K/X was required. No difference in age (p>0.99 for KO vs. WT; Mann- Whitney U- test), body weight (p=0.935), respiration rate during MRI (p=0.632) was found between KO and WT so no animals were excluded from further analysis (Table 1A). It is worth mentioning, that there is an ongoing debate on efficacy of IVIM modelling in reflecting phenomena in microvascular network or related to tissue microarchitecture (Fournet et  al., 2017; Meeus et al., 2017; Paschoal et al., 2018; Schneider et al., 2019; Niendorf et al., 1996). We have aimed to provide an optimal setup for DWI (Liao et al., 2021; Lemke et al., 2011) by measuring MR diffusion signal using 30ms echo time (TE) and 17 b- values with increased averaging for b- values≥1000 s2/mm (see Table 1B). Based on or preliminary assessment (unpublished) application of higher than minimal available (here minimal  ~22ms) TE would reduce the influence of ghosting and perfusion- related artefact, and higher averging would reduce the influence of possible Rician noise at high b- values. Although IVIM estimates were reported to depend on TE (Führes et  al., 2022; Bisdas and Klose, 2015), mostly for perfusion fraction, our evaluation focused predominantly on the slow diffusion component. Furthermore, by sampling signal up to 3000 s2/mm b- value, which may lead to presented slight lower ADC values, we aimed for depicting dominant signal from extracellular space at high b- values (Le Bihan, 2019; Cihangiroglu et al., 2009; Clark and Le Bihan, 2000; Niendorf et al., 1996). However indicatively useful, higher order models and the models focused on separating hindered MR diffusion signal according to assumption on microarchitecture (Latour et  al., 1994; Palombo et  al., 2020; Burcaw et  al., 2015; Kaden et  al., 2016; Wu and Zhang, 2019; Olesen et al., 2022; Pfeuffer et al., 1998) or assessment of diffusion signal distribution (Roth et al., 2008; Benjamini and Basser, 2019; Slator et al., 2021; Ronen et al., 2006) are beyond presented general evaluation. CM cannulation and DCE-MRI For DCE- MRI, cisterna magna (CM) cannulation based on the previous studies (Xavier et al., 2018) was performed in mice (5 KO and 6 WT; Table 1A) anesthetized with Ketamine/Dexmedetomidine (i.p. K/Dex: 75/1 mg/kg). After exposing CM, a 30 G copper cannula (out. ø0.32 mm; Nippon Tokushukan, Mfg, Tokyo, Japan), attached to a PE- 10 tube filled with aCSF, was inserted into CM and fixed in position with a drop of cyanoacrylate glue followed by a drop of glue accelerator. The incision site and the skull were then covered by a mixture of cyanoacrylate glue and dental cement. Subsequently, the exposed PE- 10 tubing was attached to a cannula filled with contrast agent (gadobutrol, 20 mM; Gadovist, Bayer Pharma AG, Leverkusen, Germany) and the animals were moved to MR scanner. The filled cannula was then connected to a syringe in a microinfusion pump. Head movements during the scanning were minimized by fixing the animals head in an MR- compatible stereotactic holder with ear bars, and the animals were put into MR with the head centered under the CryoProbe. The scanning protocol based on the previously described (Stanton et al., 2021) and whole mouse brain pre- and post- contrast T1- weighted DCE- MRI was acquired with 1 min temporal and 100 µm isotropic spatial resolutions using 3D- FISP sequence (Table 1B). DCE- MRI continued over 90 measurements (90 min), and T1- enhancing contrast agent was infused into the CM (gadobutrol, 1 µL/min for 10 min) after the first three baseline scans (i.e., after 3 min). Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 17 of 36 Neuroscience Research article Microbeads phantom ex vivo To verify whether ADC and D may reflect the volume of interstitial fluid (ISF) space in the brain paren- chyma, we performed an additional DWI measurements by mimicking increasing ISF space volume in 3 water phantoms filled with Sephadex G- 25 microbeads (Sephadex G- 25; Sigma- Aldrich, St. Louis, MO, USA) of coarse (100–300 μm), medium (50–150 μm), fine particle size (20–80 μm of wet particle size). All microbeads possess the same porosity <5 kD so a similar exchange rate between stored and free water pools is expected. Each phantom was of the same in- house design (in. volume ~0.5 ml.), formed from of thick plexiglas tube (~25 mm in. long, out./in. ø10/5 mm) with a thread plug at both sides. The thread plug was made from the same tube, with a standard bonded polyester micro- filter placed inside (Appendix  1—figure 1E) to prevent evacuation of the microbeads. For DWI, microbeads for each phantom were placed initially in a distilled water for 24  hr, to achieve their maximal size. Afterwards, a 2 ml syringe filled with distilled water solution of 0.001 mM/ml gadobutrol was attached to the thread plug on one side of the phantom and the phantom was filled with the microbeads. Gadobutrol was used to obtain optimal MR signal shortening for the echo time used in the same EPI sequence as employed in vivo (Table 1B). Any possible air bubbles remaining between the microbeads were carefully removed using a 1 µl inoculating loop. Subsequently, the other side of the phantom was closed with the thread plug, the phantom was flushed with ~1.8 ml gadobutrol solution in the attached syringe, and the solitary plug was sealed with a rubber syringe cap and para- film. The syringe with a residual of gadobutrol solution at the other end was left to support pressure equalization inside the phantom. DWI was performed 6 times in each phantom, with a central slice imaged in 1/3 phantom’s distal portion from the syringe. To verify the diffusion values in a free water environment, DWI was performed 4 times in the same phantom filled solely with gadobutrol solution. Free fluid volume estimation in the phantom To confirm that the free fluid volume surrounding the microbeads increases with increasing particles size, additional measurements were performed employing MR T1 mapping and contrast- enhanced micro computed- tomography (µCT). For T1 mapping, all phantoms were prepared de novo and scanned jointly using spin- echo sequence (2D- RARE; Table 1B) with variable repetition times. For each phantom, dry microbeads were put for 24  hr directly into the phantom lumen filled with 0.001  mM/ml gadobutrol solution. Before MRI, any possible air bubbles were removed, and the phantom was sealed (as above). Microbeads of different size were expected to differently infiltrate surrounding water, and partially gadobutrol. Thus, the relative concentration of gadobutrol was expected to be altered compared to original solution, and shorter T1 relaxations were expected to be observed for phantoms with decreasing microbeads size. For reference, T1 mapping was performed in the phantom filled solely with gadobutrol solution. To correct T1 maps for B1- filed inhomogeneities using double- angle method (Insko and Bolinger, 1993; Cunningham et al., 2006), all phantoms were scanned with the same spin- echo sequence with a maximal repetition time used for T1 mapping, and using two different flip angles (Table 1B). For free fluid space estimation using µCT, all phantoms previously scanned for DWI were unsealed and flushed with a distilled water solution via 5  ml syringe to remove gadobutrol. Subsequently, distilled water was replaced with a 1:1 dilute of non- ionic iodine Omnipaque 350 contrast agent (Iohexol, 350 mg iodine/ml; GE Healthcare AS) in normal saline via 3 ml syringe. Each phantom was scanned using Vector4uCT system (MILabs, Utrecht, Netherlands) using the scan parameters of 15 μm isomteric resolution, 50 kVp, 0.24 mA (75ms exposure), 360 degrees rotation, 0.2 degree rotation step, 2 frames for averaging, 0.5 mm thick beam aluminum filter, Hann filter with cone- beam reconstruction. To verify the relation between MR diffusion values and the T1 relaxation times as well as free fluid estimates using µCT, DWI was performed 4 times in each phantom (as above). AQP4 channel staining ex vivo C57Bl/6 brain sections collected previously (Hablitz et  al., 2020, ) were newly reimaged and an entirely new analysis done for current experiment. Originally, mice (4  WT; Table  1A) were cardiac perfused with AlexaFluor 488 conjugated wheat germ agglutinin (Thermofisher Scientific) at 15 µg/ mL in 20 mL of phosphate buffered saline (PBS), and then perfused with 20 mL PBS with 4% parafor- maldehyde (PFA). Subsequently, the brains were extracted and were immersed in 4% PFA overnight. Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 18 of 36 Neuroscience Research article This step labels the vasculature and fixes the brain tissue. Brains were sectioned coronally at 100 µm using a vibratome (Leica VT1200s) and equivalent sections from +1.2 to –1.8 bregma were stained for AQP4. Floating sections were permeabilized with 0.1% Triton- X- 100 in PBS, blocked with 7% normal donkey serum (Jackson Immunoresearch) in PBS with 0.03% Triton- X- 100, and then incubated with Anti- AQP4 primary antibody (AB3594, Millipore; 1:1000 dilution) overnight. The sections were after- wards washed three times with PBS and incubated with Alexa 594 fluorophore- linked donkey anti- rabbit secondary antibody (A21207, Invitrogen; 1:500 dilution) and DAPI (D1306, Invitrogen; 1:2000 dilution). Stained sections were mounted with Fluoromount G (Thermofisher Scientific). Vascular density measurements ex vivo Vascular density was measured in mice brains (6 AQP4 KO and 6 C57Bl6; Table  1A) prepared in few steps. All mice were cardiac perfused with AlexaFluor 488 conjugated wheat germ agglutinin at 15 ug/mL in 20 mL of PBS, and then perfused with 20 mL 4% PFA. Afterwards, the brains were extracted and immersed in 4% PFA overnight. 100 µm coronal sections were taken using a vibratome (Leica 1200  S). Anterior brain sections at  +1.2  mm bregma, and posterior brain sections –1.8  mm bregma were mounted in Prolong Gold media with DAPI (Invitrogen). Equivalent sections were used for all biological replicates. CSF production measurements in vivo The measurements were performed in mice (5 KO and 6 WT; Table 1A) as described previously (Liu et al., 2020). Mice were anesthetized with K/X (100 mg/kg / 20 mg/mL i.p.) and 2% isoflurane, and placed in a stereotactic frame. Their scalp was shaved, and the skin was cleaned with a chlorhexidine swab followed by an alcohol wipe to remove the chlorhexidine. An iodine solution was applied and left to dry, and the scalp was opened and the skin retracted. The exposed skull was irrigated with sterile saline and cleaned by applying sterile cotton swabs, and a sterilized stainless- steel light- weight head plate (0.9×19×12 mm dimension), equipped with a round hole of at the center (in. ø9.0 mm), was attached to the mouse skull using a mixture of dental cement with cyano- acrylate glue (Sweeney et  al., 2019). Pre- operatively as well as for three days post- surgery the mice received Banamine (1.1 mg/kg) subcutaneously as an analgesic. The mice were trained to tolerate positioning in the head plate stand (Cat# MAG- 1, Narishige International USA Inc), as well as a restraint tube in three daily training sessions, each lasting 30 min for 3 days post- surgery. For measuring CSF production during wakefulness, the mice were anesthetized with 2% isoflurane during cannula implantation. A cannula (30 G needle) attached to artificial CSF filled PE- 10 tubing was implanted into the right lateral ventricle through a small burr hole (AP = −0.1 mm, ML = 0.85 mm, DV = –2.00 mm from the postion of bregma). The cannula was fixed to the skull with dental cement and the opposite end of the PE tubing was sealed by high- temperature cautery. Once the cannula was in place, the neck was flexed 90  degrees and the headplate was attached to the head stand. Then a separate cannula was inserted into the CM and advanced into the 4th ventricle. One micro- liter of mineral oil was infused into the 4th ventricle to block the exit of CSF into the subarachnoid space. All incisions were infiltrated with 0.25% bupivacaine topical anesthetic to prevent the animal from experiencing post- surgical pain. The measurements were collected while the animal rested in an open cylinder restraint tube (9 cm length, in. ø3.5 cm), to which the animals were accustomed during training. Anesthesia was discontinued and CSF production was measured in head- fixed mice in 10 min intervals for 65 min, after a 30- min recovery period. Tetramethylammonium microiontophoresis for interstitial fluid space volume estimation Real- time iontophoresis with tetramethylammonium (TMA) was performed in mice (8 KO and 20 WT, body weight not recorded) as adapted from the previous studies (Nicholson, 1993; Xie et al., 2013). The single barrel iontophoresis microelectrode (tip out. ø2–3 µm) contained 150 mM tetramethylam- monium (TMA)- chloride and 10 µM Alexa 488. A series of currents of 20 nA, 40 nA and 80 nA were applied by a dual channel microelectrode pre- amplifier. For measurements of TMA, microelectrodes (out. ø2–3  µm) were fabricated from double- barreled theta- glass using a tetraphenylborate- based ion exchanger. The TMA barrel was backfilled with 150 mM TMA chloride, while the reference barrel contained 150 mM NaCl and 10 µM Alexa 568. All recordings were obtained by inserting the two Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 19 of 36 Neuroscience Research article electrodes to a depth of 150  µm below the cortical surface. Recording electrodes were inserted 2.5 mm lateral and 2 mm posterior to bregma. The electrode tips were imaged after insertion using 2- photon excitation to determine the exact distance between the electrodes (typically ~150 µm). The TMA signal was calculated by subtracting the voltage measured by the reference barrel from the voltage measured by the ion- detecting barrel using a dual- channel microelectrode pre- amplifier. The Nikolsky equation was used for calibration of the TMA electrodes based on measurements obtained in electrodes containing 0.5, 1, 2, 4, and 8 mM TMA- chloride in 150 mM NaCl. The TMA measure- ments were acquired relative to similar recordings obtained in 0.3% agarose prepared from a solution containing 0.5 mM TMA and 150 mM NaCl. A custom- made software in Matlab (v. R2019a, The Math- works, Inc, Natick, MA.), ‘Walter’, developed by C. Nicholson was used to calculate α and λ values (Nicholson, 1993; Xie et al., 2013). Brain water content ex vivo Anesthetized animals (3 KO and 5 WT, gender and body weight not recorded; Table 1A) were decap- itated, the whole brains were taken out and weighed immediately (Wwet). Brain tissue was dried at 65 °C for 48 h until it reached a constant weight, and brain were re- weighed (Wdry). The ratio between the difference of Wwet - Wdry and Wdry was considered reflecting the brain water content (ml H2O/g dry weight). Data processing and statistical analysis MRI All acquired images were visually checked and no presence of significant artefacts influencing morphological and functional assessment was found. Further processing pipelines were applied, and included motion- correction, spatial co- registration, and automatic or semi- automatic pre- and final postprocessing. All described statistical analyses were performed in GraphPad Prism 8 (GraphPad Software) and Matlab. The results coming from statistical comparisons were considered significant for p<0.05 after post- hoc correction, when applicable. 3D-CISS volumetry and cisternography All 3D- CISS volumes were calculated in few steps using in- house pre- processing pipeline (Appendix 1—figure 3A). For each animal, every 3D- TrueFISP volume acquired with two repetitions was motion- corrected (10 times or until no further improvement) and averaged. Subsequently, the second averaged 3D- TrueFISP volume (180° encoding direction) was subjected to rigid- body regis- tration (6 df.) to the first volume (0° encoding direction). Both motion- correction and registration were performed in AFNI (Oakes et al., 2005), and aimed to reduce the influence of random motion on the computed 3D- CISS image. Finally, 3D- CISS image was computed as a maximum intensity projection from 2 co- registered 3D- TrueFISP volumes, resulting in an image of almost completely removed banding artifacts. Afterwards, every computed 3D- CISS volume underwent semi- automatic brain parenchyma image extraction using the ‘Segment 3D’ tool in ITK- SNAP (Yushkevich et  al., 2006), to remove the regions outside the brain parenchyma image from further analysis. Brain paren- chyma was considered as the brain tissue volume surrounded with dark regions of skull image and including intracerebral vessels. To correct for intensity inhomogeneities coming from the B0 field and the surface profile of the CryoProbe (B1), the extracted brain parenchyma image underwent bias field correction using FSL (Zhang et al., 2001) (0.5 sigma, 20 mm FWHM, 4 iterations). At each step of pre- processing, the results were visually checked and confirmed for correctness. Automatic CSF space segmentation For a single, bias- corrected 3D- CISS volume the ventricular and perivascular CSF spaces were sepa- rated from the brain parenchyma image in 3 dimensions using an in- house fully automatic adaptive algorithm in Matlab (Gomolka et al., 2021), in four steps (Appendix 1—figure 3B). First, high- intensity regions, as branches of the optic nerve’s residual after the semi- automatic brain extraction and not adjacent to the parenchyma in all image slices, were excluded using a bounding box enclosing the brain (Appendix  1—figure 3B ‘Bounding box and adaptive thresholding’). The bounding box was Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 20 of 36 Neuroscience Research article automatically computed based on a maximization of the voxels intensity variance slice- wise, sepa- rately in three orthogonal planes. The parenchyma volume surrounded by the bounding box was enclosed and the solitary regions were removed based on their geometrical properties calculated slice- wise in the sagittal plane: eccentricity≥0.5, roudness≥0.5, perimeter<0.005% of the brain paren- chyma voxels count. The resulting brain image mask was geometrically dilated with a disk kernel of 11 pixels diameter, to enclose potentially removed or non- continuous parenchymal regions. The non- continuity appeared in case of residuals from banding artifacts at the borders of the skull and the ethmoidal bone. Subsequently, brain parenchyma volume was updated according to the resulting mask, for further automatic segmentation of the CSF space. Second, initial CSF space separation was performed by means of an adaptive intensity threshold and calculation of cumulative distribution of voxels intensities>0 from the separated and bias- corrected brain volume. As the overall distribution of the brain intensities differed slightly between the 3D- CISS images due to their SNR, a threshold- correcting factor (denoted fc) was calculated by means of the formula: fc = σd µd+σd (1) where µd is a mean and σd is a standard deviation (SD) of aggregated brain parenchyma voxels intensity distribution. Separation of the CSF space was performed assuming that CSF intensities reflect those ≥95th percentile of the aggregated intensities distribution. Hence, the intensity of each voxel was rescaled according to the formula: Ir = ( I − 1.33 σd × σd , ) (2) where I and Ir are the original and the rescaled voxels intensities. The rescaled voxels intensities were encoded using a floating point precision in range between −1.33 × σd and a distribution peak close to the image SNR calculated as µd/σd (i.e. maximum rescaled intensity ~10 with the mean value varying between 3 and 5 among all analyzed images). Third, all the rescaled voxels possessing nega- tive intensity (i.e. brain parenchyma) were assigned to 0, and a new aggregated distribution of the rescaled voxels of >0 intensity was computed. Subsequently, all voxels intensities ≤95.5th percentile of the new distribution were assigned to 0 to keep only the high intensity CSF seed regions. For the images of SNR  >4 (i.e. lower contribution of Rician noise), the new distribution threshold was set of ≤95.5 +fc. The correction factor fc accounted for subtle intensity changes and did not result in the threshold exceeding the 97th percentile. A mask image of the initially separated CSF space was computed by assigning all remaining nonzero voxels to 1, for further application of a region- growing algorithm. The final segmentation was performed using a 2D region- growing algorithm applied slice- wise, consecutively in sagittal, axial and coronal planes (Appendix 1—figure 3B- ‘in Sagittal plane’, ‘in Axial plane’, ‘in Coronal plane’). In each slice, the algorithm reconsidered the voxels at the boundary of the CSF space in horizontal, vertical and diagonal directions separately. The algorithm based on exten- sion of the method for contrast calculation applied to study properties of hemorrhagic and ischemic regions in clinical CT images (Nowinski et al., 2014; Gomolka et al., 2017). The calculations were performed considering the initially separated CSF space mask and the original 3D- CISS image, in two steps: (1) CSF mask dilation; (2) CSF mask erosion. CSF mask dilation To assure that only voxels belonging to the CSF space and not affected by the partial volume from the surrounding parenchyma were included into calculation, the contrast was computed for the initially separated CSF boundary voxels reflecting the intensities ≥97.5th - fc percentile (i.e., µd + 2 × σd) of the aggregated intensities distribution from not intensity- rescaled brain parenchyma image. The boundary contrast was obtained as a ratio of an absolute difference between considered boundary voxel intensity and the mean intensity of the n consecutive voxels to the left/top/left- diagonal to the sum of the boundary voxel intensity and the mean intensity of the n consecutive voxels to the right/ bottom/right- diagonal in the original CISS image (absolute relative CSF/brain parenchyma contrast in horizontal/vertical/diagonal directions, respectively). The contrast was calculated for n from 1 to 4 Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 21 of 36 Neuroscience Research article in the sagittal and from 1 to 3 voxels in the axial and coronal planes. The voxels at nth distance from the boundary were included into the updated CSF mask if their absolute relative contrast was <2%. CSF space erosion The boundary of the updated CSF mask was recalculated, and the voxels were reconsidered using the same method for the contrast calculation (see above). Herein, however, the contrasts were calculated for n from 1 to 2 in the sagittal and from 1 to 3 in the axial and coronal planes, to avoid removing small regions belonging to the perivascular space around the main cerebral arteries. The voxels at nth distance were removed from the updated CSF mask if their intensity in the original CISS image was ≤95.5th - fc percentile (for sagittal, and <95.5th - fc for axial and coronal planes) of the aggregated intensity distribution, and the absolute relative contrast to the respective boundary voxel was >2.5% (for sagittal, and ≥2.5% for the other planes). To further remove remaining false- positively segmented single voxels and to enclose wrongly opened larger regions the final CSF mask was subjected to 3D median filtration with a 3×3 voxels kernel (Appendix 1—figure 3B ‘Filtration and labeling’). Final CSF volume was visually assessed for correctness by RSG and YM, by overlaying with its parental 3D- CISS volume in ITK- SNAP. CSF compartments labeling and statistical comparison The delineated CSF space was separated manually from the final CSF mask in ITK- SNAP into seven compartments (Appendix 1—figure 3B – ‘Filtration and labeling’), for further statistical comparison: lateral ventricles; third ventricle; fourth ventricle; basilar artery; basal perivascular space at the skull base surrounding the Circle of Willis; parietal perivascular spaces and cisterns (ventrally from the position of posterior cerebral artery, via space neighboring the transverse sinuses and dorsally to the junction of the superior sagittal sinus and transverse sinuses); remaining perivascular space within the olfactory area, surrounding anterior cerebral and frontopolaris arteries, middle cerebral arteries branches, and posterior cisterns including pontine and cisterna magna. For supplementary compar- ison, the segmented lateral, third and fourth ventricular spaces were considered jointly as the ventric- ular space, and the basilar, basal and the remaining anterior/posterior CSF spaces were considered jointly as the whole perivascular space. Number of voxels was counted, and the volume of each segment was calculated by multiplying the voxels count by the voxel dimension from the original 3D- CISS image, for subsequent statistical comparison. To compensate for the brain capsule volume differences and provide a reliable measure of the brain’s CSF space volume between animals, a ratio of the CSF to the brain volume (intracranial volume) was calculated for each delineated CSF segment as: RatioCSF space = CSFcompartment volume CSFwhole segmented volume . Brainvolume− (3) The ratios obtained for each of the CSF compartments, as well as the segmented brain volumes were compared between KO and WT animals using nonparametric Mann- Whitney U- test. DWI and IVIM-DWI Preprocessing For every animal, all DWI volumes acquired were subjected to motion- correction in AFNI to avoid influence of random and subtle frame- to- frame image displacements. The motion- correction was performed (4 times or until no further improvement) in reference to the first image acquired using the first b- value (b0), and the results were visually confirmed for correctness by RSG. Images acquired with different number of repetitions were averaged according to b- value and diffusion encoding direction, for subsequent calculation of diffusion measures in all 3 directions. Separately, an averaging took place considering only b- values, to calculate an average diffusion images. To normalize the image intensities and reduce influence of nonstationary noise, volume- wise normalization of voxels intensi- ties was performed in all images from different b- values using a mean value of a background signal defined in a half- ball VOI including 704 voxels outside the visible tissue image. The VOI was manually set in an averaged b0 image and applied to every averaged volume from each b- value for the back- ground signal depiction. The intensity normalization was performed according to formula: Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 22 of 36 Neuroscience Research article DWIbval norm = DWIbval − µbval noise − 6 × σbval noise , ) (4) ( where DWIbval and DWIbval norm are the original and normalized voxels intensities, and µbval noise and σbval noise are respectively the mean and SD of the background signal at b- value. To reduce the influence of distortion artefacts, all DWI images were subjected to slice- wise spatial smoothing using a 0.5 sigma [3×3] Gaussian filter. To automatically exclude the voxels outside the brain regions in subsequent diffusion curve fitting, a brain mask image was calculated from b0 image and applied to all images from different b- values. The mask image excluded all the voxels below the mean + 0.5 × SD of the b0 image intensity. As initial threshold excluded voxels ventrally from the brain image, the mask was subjected to morphological dilation with a diamond shape kernel of 2 pixels diameter, and subse- quent image filling to include sole obsolete voxels inside the mask. Diffusion-curves fitting All calculations were performed in Matlab using in- house computational pipelines and curve- fitting implementations based on the least squares method. For a standard DWI model, monoexponential curve- fitting was performed voxel- wise using all 17 b- values. The ADC and estimated S0 images were calculated voxel- wise based on the formula: Sbval = S0 × bval e − × ADC . ) (5) ( For IVIM model, a two- step algorithm was used as considered providing more robust and reliable results compared to standard biexponential curve- fitting (Lee et al., 2016). In the first step, a perfu- sion fraction (Fp) was estimated as a voxel- wise ratio of the difference between b0 and approximated S0 to b0 image. Voxels intensities in the S0 image were estimated from the linear regression approx- imation considering the logarithm intensities of the volumes from the IVIM threshold to the highest b- value. The threshold b- value used was closest to 250 s2/mm (here ~238 s2/mm). The slow diffusion D was estimated as the slope coefficient of the linear regression function. Subsequently, for the voxels where a positive and nonzero Fp value was calculated (presence of fast diffusion) a biexponential curve- fitting was performed using the calculated Fp and D: Sbval = S0 × Fp × − e ( ( bval D∗ × ) + 1 Fp − ( ) bval × − e ( × D , )) (6) where D is the slow (‘pure’) molecular water diffusion, D* is the fast diffusion (pseudodiffusion). Regions of interest definition and statistical comparison For further statistical comparison, the mean and SD of ADC, Fp, D and D* signal intensities were calculated within 21 regions of interest (ROI), for every animal brain image separately. The regions were chosen based on reported high Aqp4 expression (Hsu et al., 2011; Hubbard et al., 2015) and were drawn manually using ITK- SNAP in estimated S0 image, and overlaid on the calculated diffusion parameters maps. The defined ROIs included from 48 to 250 voxels depending on the anatomical structure and excluding neighboring distortion artefact, and covered the position of cortical (olfactory area - OLF; cingulate/retrosplenial cortex - CA/RSP; visual - VIS; somatosensory - SS; auditory - AUD; hippocampus - HIP; perirhinal - PERI), brain stem (thalamus - TH; habenula - HAB; hypothalamus - HY; midbrain - MB; periaqueductal gray - PAG; hindbrain - HB), cerebral nuclei and tracts (caudate putamen - CP; white matter including striatal regions - WM); CSF space (lateral ventricles - LV; third ventricle - 3 V; fourth ventricle - 4 V; pericisternal space - PCS; perivascular space within the Circle of Willis - CoW) and the cerebellar ROI (cerebellum - CB). Application of automatic segmentation template was avoided due to variable influence of distortion and ghosting artefacts at large span of 17 b- values measured, resulting in erroneous template registration and delineation of signal from specific anatomical structures. The regions were set manually by RSG in reference to the Allen Brain Atlas, and verified by YM and MN. Due to presence of distortion and ghosting artefacts resulting in possible inhomogeneous noise properties of EPI images acquired, voxel intensities within delineated ROIs were considered coming from independent and non- uniform signal distributions (i.e., comparing ventral vs. dorsal ROI). There- fore, mean ADC, D, Fp, D* and Fp × D* were calculated excluding <0  and>99 percentiles of their intensity distribution within each ROI (see calculated values in the supplementary Excel file), and were Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 23 of 36 Neuroscience Research article compared ROI- wise between KO and WT animals using a nonparametric Mann- Whitney U- test. To confirm reliability of the slow diffusion measures, Pearson’s linear correlation was calculated between the mean ADC and D from all animals. DCE-MRI Preprocessing Time- series of FISP volumes acquired were motion- corrected and spatially normalized with ANTs (Avants et  al., 2014) and co- registered to the respective baseline image, subject- wise. For every animal, the co- registration process was repeated twice using rigid- body (6 df.) and twice using an affine transformation (12 df.) to assure accuracy. Subsequently, the intensities from the first volume acquired after the dummy scan (i.e., of unstable steady- state MR signal) were subtracted voxel- wise from succeeding volumes to reveal only the regions enhanced with the contrast agent. To normalize the CSF signal in each time- series their voxel intensities were subjected to Gaussian normalization using the first 3D- FISP volume. The resulting images were smoothed with a 3×3×3 voxels kernel of [0.2, 1, 0.2] weights along each axis, to reduce the influence of possible artifacts after the automatic registration and subtracting the baseline volume. Volumes of interest definition and statistical comparison For further statistical comparison, DCE 3D- FISP signal was derived from 21 ROI reflecting those from DWI analysis (see above), in each animal individually. All ROI were defined in a 3D template volume, based on Reference Space Model from Allen Brain Atlas, of the same spatial resolution as 3D- FISP volumes acquired. The reference volume was automatically registered in AFNI using affine registra- tion (6 df.) to the baseline DCE volume. Mean DCE signal was depicted from every 3D- FISP volume in every animal ROI- wise (90 signal intensity values / ROI), for further statistical comparison. The first volume following the dummy scan was rejected from further consideration as it had previously been used for the signal normalization. To compare ROI- wise DCE signals between KO and WT animals, a repeated measurements Two- way ANOVA along with post- hoc Bonferroni’s correction was used in search for differences between the signal changes at specific time points and compared to the base- line signal before the gadolinium infusion. Further calculation of DCE- derived scores of CSF tracer arrival time, relative time- to- peak, peak intensity, and duration of significant from baseline intersti- tial tracer accumulation was performed based on the rank- sum scores provided by Dunn’s multiple comparison from nonparametric Friedman’s one- way ANOVA, performed separately in KO and WT groups. Arrival time was considered as a difference in time between the baseline time point, and time point where the negative rank- sum score was at least three times that from the baseline. Time- to- peak was calculated as a time difference between the start of the gadobutrol infusion and the subsequent time point at which the maximum negative rank- sum score was calculated (i.e. corresponding to the peak CSF signal intensity within the group). The peak intensity was considered as a relative signal increase occurring at the time- to- peak. Duration of significant from baseline interstitial tracer accumu- lation was obtained as a time difference between the last and the first time points, where DCE signal increase was significantly different from baseline. Area under the DCE curve was calculated for each subject as a sum of DCE signal amplitudes over all the measured time points. DCE scores derived for KO and WT groups were compared ROI- wise using nonparametric two- tailed Wilcoxon signed- rank test. Microbeads MR phantom ex vivo DWI and IVIM measures were computed using the same processing pipeline as for in vivo imaging. To avoid influence of artefacts due to field inhomogeneity or at the border of the phantom lumen, all diffusion values were depicted in a 1.5 mm diameter circular region around the center of the phantom image, in all slices. Mean diffusion values from all phantoms were compared using nonparametric Kruskal- Wallis one- way ANOVA with Dunn’s post- hoc. T1 mapping Calculation of T1 maps was performed using an in- house algorithm in Matlab. The voxel intensities from spin- echo images of the phantoms were approximated over the increasing repetition times, by inversely solving the equation with the least squares method: Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 24 of 36 Neuroscience Research article STR = S0 × TR T1 1 − e ( ( )) + ε , (7) where TR is a repetition time, S0 is the estimate of the proton density, STR is the intensity of the considered voxel from the corresponding slice, and is an estimator of a random error. To reduce influence of transmit inhomogeneity on T1 estimations, B1- correction was applied using dual- angle method based on relationship between an effective and nominal flip angles (Cunningham et  al., 2006) and by applying voxel- wise a correction factor dependent on the intensities from two spin- echo volumes acquired with 45° and 90° flip angles. ε Free water volume fraction Evaluation of the free water volume fraction inside the phantoms with microbeads was performed based on percentage of the contrast- filled volume, quantified in 8 bits gray scale axial images. First, the central portion of the phantom volume, reflecting that from DWI analysis, was manually separated in ImageJ (v. 1.53 j, NIH, USA). The free fluid volume regions (filled with Omnipaque 350 solution) were of clearly higher Hounsfield Units (HU) intensities compared to those from the microbeads. Thus, the voxels belonging to the free fluid were empirically verified occupying >75th percentile of the cumulative HU distribution in all the separated volume images, from all phantoms. Therefore, the voxels belonging to the free fluid were counted three times, considering only voxels above 75th, 80th, and 85th percentile of HU distribution. For each phantom, the free fluid volume fraction was estimated as a mean voxel count from all thresholds, divided by the voxels count in the separated volume image. As objective separation of the fluid space was impossible in the phantom with the fine microbeads and lower HU thresholds would consider voxels affected by partial volume from the microbeads, proposed free fluid volume fraction estimation was found sufficient to correct for the changes in shape of voxels HU distribution from different phantoms. AQP4 staining and vascular density image analysis Brain sections were imaged using a conventional fluorescence macroscope (Stereo Investigator with objective UplanXApo 10x /numerical aperture 0.40, ∞/compatible cover glass thickness 0.17  mm/ field number 26.5 mm, no immersion liquid (air); Olympus) and subsequently analyzed using ImageJ. Multiple FOVs were acquired in a 1360 × 1024 px / 1392.44 ×1048.43 µm frames (1.048 µm2/pixel), and subsequently aligned and stitched together to provide image of entire brain section. Both AQP4 and vascular staining were assessed in a complimentary anterior and posterior entire brain sections (cf. AQP4 channel staining ex vivo). To minimize the reader- associated bias, all image analyses were performed by the investigator (MG) blinded to the image content and animal genotype, and in randomized images. AQP4 staining For each brain hemisphere in each section, all ROIs were manually drawn according to the Allen Brain Atlas and using auto fluorescence from the green channel following visible anatomical landmarks and confirming AQP4 expression visibility. Each subregion measured from 0.03 to 1.69 mm2. Therefore, a universal threshold was applied manually to all images as a preferred method for reduction of the background signal influence. Subsequently, a mean area fraction covered by AQP4 in both hemi- spheres/sections was measured for each ROI. In total, mean AQP4 expression was calculated for 11 ROIs (retrosplenial cortex - RSP, visual - VIS; somatosensory - SS; auditory - AUD; hippocampus - HIP; perirhinal - PERI; thalamus - TH; habenula - HAB; hypothalamus - HY; pericisternal - PCS; white matter - WM) in 4  WT mice. A nonparametric Kruskal- Wallis one- way ANOVA with Dunn’s post- hoc was employed to compare the mean AQP4 channel expressions between ROIs. Vascular density ROIs were manually outlined around each anatomical subregion according to the Allen Brain Atlas, and using visible anatomical landmarks. Due to differences in fluorescent labeling intensity, each region was thresholded individually to isolate labeled blood vessels from the background, so each region measured from 0.023 to 4.67 mm2 within each hemisphere/section. Area fraction of blood vessels above the threshold was measured for each ROI. In total, mean vascular density was calculated from multiple subregions for 17 ROI (olfactory area - OLF; cingulate cortex - CA; retrosplenial cortex Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 25 of 36 Neuroscience Research article - RSP; primary visual - V1; primary somatosensory - S1; primary motor area - M1; auditory - AUD; hippocampus - HIP; perirhinal - PERI; insular - INS; thalamus - TH; habenula - HAB; hypothalamus - HY; caudate putamen - CP; white matter - WM; pericisternal - PCS; ependymal around lateral ventricles - EPD) in 6 KO and 6 WT animals. Further statistical comparison was performed assuming inhomoge- neous signal distribution properties between different ROI (similarly as for DWI). Hence, considering independent measurements of vascular densities among ROI analyzed and due to small group size, nonparametric Mann- Whitney U- test was employed to compare the vessel densities from KO and WT animals ROI- wise. For ROI- wise correlation analysis, a mean value of AQP4 expressions as well as vascular densities at ROI was calculated from all respective animals strain- wise. TMA measurements in vivo and remaining statistics For ISF space volume estimation with TMA: to keep uniformity of obtained α and λ values distribu- tion within KO and WT groups altered by uneven group size, only the animals expressing α within mean ± 1.5 × SD of α distribution within the group (within ~90% of cumulative distribution) were kept. Therefore, two animals from KO and three animals from WT group were removed from among awake batch, and one KO and four WTs were removed from the K/X anesthetized batch of animals (6–7 KO and 16–17 WT remaining). The obtained α and λ values form TMA were compared using one- way ANOVA with Bonferroni’s post- hoc. Differences in α and λ values between awake and anesthetized animals, total excreted CSF volumes, brain water contents, as well as demographic characteristics parameters (Table 1A) between KO and WT mice were compared using Mann- Whiteny U- test. Visualizations Whisker- box and correlation plots were generated using Graphpad and radar plots comparing DWI and IVIM values ROI- wise were generated using OriginPro (v. 2020, OriginLab Corporation, Northampton, MA). For the purpose of 3D- CISS volumetry and cisternography figures plotting or to visually depict the CSF signal changes within the DCE- MRI time- series, all 3D maximum intensity projection and multiplanar reconstruction images were generated using Amira version 6.2 (Thermo Fisher Scientific, Waltham, MA, USA). 3D- CISS surface reconstructions were performed using Mango Image Analysis software (v.4.1, Research Imaging institute, UTHSCA). To representatively visualize both AQP4 channel and vascular immunohistochemistry staining, a confocal microscopy (objective UplanXApo 60x / numerical aperture 1.42, ∞ / compatible cover glass thickness 0.17 mm / field number 26.5 mm, oil immersion; Olympus) was performed in a representa- tive section from a single WT animal. Single FOVs were acquired in a 401×401 px / 166.14×166.13 µm frames (0.172 µm2/pixel) and overlaid in ImageJ to visualize both the position of AQP4 channels as well as the vasculature in a representative image. Acknowledgements This study was supported by Lundbeck Foundation (R359- 2021- 165), Novo Nordisk Foundation (NNF20OC0066419), and the Dr. Miriam and Sheldon G Adelson Medical Research Foundation. National Institutes of Health grants (R01AT011439 and U19NS128613), The U.S. Army Research Office MURI (W911NF1910280), Human Frontier Science Program (RGP0036) and The Simons Foundation (811237). The views and conclusions contained in this article are solely those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the National Institutes of Health, the Army Research Office, or the U.S. Government. The U.S. Govern- ment is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein. The funding agencies have taken no part on the design of the study, data collection, analysis, interpretation, or in writing of the manuscript. We thank Dan Xue for assistance with the illustrations, and Palle Koch for help with manufacturing MRI phantoms. Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 26 of 36 Neuroscience Research article Additional information Funding Funder Lundbeckfonden Novo Nordisk Fonden National Institutes of Health Army Research Office Human Frontier Science Program Simons Foundation Adelson Family Foundation Grant reference number Author Maiken Nedergaard Maiken Nedergaard Maiken Nedergaard Maiken Nedergaard Maiken Nedergaard Maiken Nedergaard Maiken Nedergaard The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Author contributions Ryszard Stefan Gomolka, Conceptualization, Software, Formal analysis, Validation, Investigation, Visu- alization, Methodology, Writing - original draft, Writing – review and editing; Lauren M Hablitz, Formal analysis, Supervision, Validation, Investigation, Methodology, Writing – review and editing; Humberto Mestre, Formal analysis, Validation, Investigation, Methodology, Writing – review and editing; Michael Giannetto, Formal analysis, Validation, Investigation, Visualization; Ting Du, Formal analysis, Valida- tion, Investigation; Natalie Linea Hauglund, Resources, Validation, Investigation; Lulu Xie, Weiguo Peng, Paula Melero Martinez, Validation, Investigation; Maiken Nedergaard, Conceptualization, Supervision, Funding acquisition, Methodology, Project administration, Writing – review and editing; Yuki Mori, Conceptualization, Supervision, Investigation, Visualization, Methodology, Project adminis- tration, Writing – review and editing Author ORCIDs Ryszard Stefan Gomolka Lauren M Hablitz Humberto Mestre Michael Giannetto Natalie Linea Hauglund Maiken Nedergaard Yuki Mori http://orcid.org/0000-0001-9797-1062 http://orcid.org/0000-0001-6159-7742 http://orcid.org/0000-0001-5876-5397 http://orcid.org/0000-0002-4338-8709 http://orcid.org/0000-0002-2198-6329 http://orcid.org/0000-0001-6502-6031 http://orcid.org/0000-0003-4208-0005 Ethics All experiments were performed based on approval received from both the Danish Animal Experi- ments Inspectorate (License number: 2020- 15- 0201- 00581) and the University of Rochester Medical Center Committee on Animal Resources (UCAR, Protocol 2011- 023). Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.82232.sa1 Author response https://doi.org/10.7554/eLife.82232.sa2 Additional files Supplementary files •  MDAR checklist •  Source data 1. Aggregated data set, including all source data. Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 27 of 36 Neuroscience Research article Data availability Entire data from the paper is available in the .xls data file attached. The attached data file is subdi- vided into separate sheets, each for a single experiment and accompanied with respective heading and descriptions, and provides the possibility of replicating all figures and statistics. A summary of data is presented in the tables and figures within the paper. A detailed description of an author algo- rithm for CSF space segmentation from 3D- CISS images, as well as DWI analysis, is provided in the Materials and Methods section (page 14 onward). References Amiry- Moghaddam M, Otsuka T, Hurn PD, Traystman RJ, Haug FM, Froehner SC, Adams ME, Neely JD, Agre P, Ottersen OP, Bhardwaj A. 2003a. An alpha- syntrophin- dependent pool of AQP4 in astroglial end- feet confers bidirectional water flow between blood and brain. PNAS 100:2106–2111. DOI: https://doi.org/10.1073/pnas. 0437946100, PMID: 12578959 Amiry- Moghaddam M, Williamson A, Palomba M, Eid T, de Lanerolle NC, Nagelhus EA, Adams ME, Froehner SC, Agre P, Ottersen OP. 2003b. Delayed K+ clearance associated with aquaporin- 4 mislocalization: phenotypic defects in brains of alpha- syntrophin- null mice. PNAS 100:13615–13620. DOI: https://doi.org/10. 1073/pnas.2336064100, PMID: 14597704 Avants BB, Tustison NJ, Stauffer M, Song G, Wu B, Gee JC. 2014. The insight toolkit image registration framework. Frontiers in Neuroinformatics 8:44. DOI: https://doi.org/10.3389/fninf.2014.00044, PMID: 24817849 Benjamini D, Basser PJ. 2019. Water mobility spectral imaging of the spinal cord: parametrization of model- free Laplace MRI. Magnetic Resonance Imaging 56:187–193. DOI: https://doi.org/10.1016/j.mri.2018.12.001, PMID: 30584915 Bergamino M, Nespodzany A, Baxter LC, Burke A, Caselli RJ, Sabbagh MN, Walsh RR, Stokes AM. 2020. Preliminary assessment of intravoxel incoherent motion diffusion- weighted MRI (IVIM- DWI) metrics in Alzheimer’s disease. Journal of Magnetic Resonance Imaging 52:1811–1826. DOI: https://doi.org/10.1002/jmri. 27272, PMID: 32621405 Bisdas S, Klose U. 2015. IVIM analysis of brain tumors: an investigation of the relaxation effects of CSF, blood, and tumor tissue on the estimated perfusion fraction. MAGMA 28:377–383. DOI: https://doi.org/10.1007/ s10334-014-0474-z, PMID: 25475914 Bloch O, Auguste KI, Manley GT, Verkman AS. 2006. Accelerated progression of kaolin- induced hydrocephalus in aquaporin- 4- deficient mice. Journal of Cerebral Blood Flow and Metabolism 26:1527–1537. DOI: https://doi. org/10.1038/sj.jcbfm.9600306, PMID: 16552421 Boespflug EL, Simon MJ, Leonard E, Grafe M, Woltjer R, Silbert LC, Kaye JA, Iliff JJ. 2018. Targeted assessment of enlargement of the perivascular space in Alzheimer’s disease and vascular dementia subtypes implicates astroglial involvement specific to Alzheimer’s disease. Journal of Alzheimer’s Disease 66:1587–1597. DOI: https://doi.org/10.3233/JAD-180367, PMID: 30475760 Bonomini F, Rezzani R. 2010. Aquaporin and blood brain barrier. Current Neuropharmacology 8:92–96. DOI: https://doi.org/10.2174/157015910791233132, PMID: 21119879 Burcaw LM, Fieremans E, Novikov DS. 2015. Mesoscopic structure of neuronal tracts from time- dependent diffusion. NeuroImage 114:18–37. DOI: https://doi.org/10.1016/j.neuroimage.2015.03.061, PMID: 25837598 Choi YH, Laaker C, Hsu M, Cismaru P, Sandor M, Fabry Z. 2021. Molecular mechanisms of neuroimmune crosstalk in the pathogenesis of stroke. International Journal of Molecular Sciences 22:9486. DOI: https://doi. org/10.3390/ijms22179486, PMID: 34502395 Ciappelloni S, Bouchet D, Dubourdieu N, Boué-Grabot E, Kellermayer B, Manso C, Marignier R, Oliet SHR, Tourdias T, Groc L. 2019. Aquaporin- 4 surface trafficking regulates astrocytic process motility and synaptic activity in health and autoimmune disease. Cell Reports 27:3860–3872.. DOI: https://doi.org/10.1016/j.celrep. 2019.05.097, PMID: 31242419 Cihangiroglu M, Uluğ AM, Firat Z, Bayram A, Kovanlikaya A, Kovanlikaya I. 2009. High b- value diffusion- weighted MR imaging of normal brain at 3T. European Journal of Radiology 69:454–458. DOI: https://doi.org/ 10.1016/j.ejrad.2007.11.023, PMID: 18162352 Clark CA, Le Bihan D. 2000. Water diffusion compartmentation and anisotropy at high B values in the human brain. Magnetic Resonance in Medicine 44:852–859. DOI: https://doi.org/10.1002/1522-2594(200012) 44:6<852::aid-mrm5>3.0.co;2-a, PMID: 11108621 Cui H, Wang W, Zheng X, Xia D, Liu H, Qin C, Tian H, Teng J. 2021. Decreased AQP4 expression aggravates ɑ-synuclein pathology in Parkinson’s disease mice, possibly via impaired glymphatic clearance. Journal of Molecular Neuroscience 71:2500–2513. DOI: https://doi.org/10.1007/s12031-021-01836-4, PMID: 33772424 Cunningham CH, Pauly JM, Nayak KS. 2006. Saturated double- angle method for rapid B1+ mapping. Magnetic Resonance in Medicine 55:1326–1333. DOI: https://doi.org/10.1002/mrm.20896, PMID: 16683260 Eide PK, Hansson HA. 2018. Astrogliosis and impaired aquaporin- 4 and dystrophin systems in idiopathic normal pressure hydrocephalus. Neuropathology and Applied Neurobiology 44:474–490. DOI: https://doi.org/10. 1111/nan.12420, PMID: 28627088 Eide PK, Vinje V, Pripp AH, Mardal KA, Ringstad G. 2021. Sleep deprivation impairs molecular clearance from the human brain. Brain 144:863–874. DOI: https://doi.org/10.1093/brain/awaa443, PMID: 33829232 Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 28 of 36 Neuroscience Research article Federau C, Hagmann P, Maeder P, Müller M, Meuli R, Stuber M, O’Brien K. 2013. Dependence of brain intravoxel incoherent motion perfusion parameters on the cardiac cycle. PLOS ONE 8:e72856. DOI: https://doi. org/10.1371/journal.pone.0072856, PMID: 24023649 Federau C. 2017. Intravoxel incoherent motion MRI as a means to measure in vivo perfusion: a review of the evidence. NMR in Biomedicine 30:e80. DOI: https://doi.org/10.1002/nbm.3780, PMID: 28885745 Fournet G, Li J- R, Cerjanic AM, Sutton BP, Ciobanu L, Le Bihan D. 2017. A two- pool model to describe the IVIM cerebral perfusion. Journal of Cerebral Blood Flow and Metabolism 37:2987–3000. DOI: https://doi.org/10. 1177/0271678X16681310, PMID: 27903921 Führes T, Riexinger AJ, Loh M, Martin J, Wetscherek A, Kuder TA, Uder M, Hensel B, Laun FB. 2022. Echo time dependence of biexponential and triexponential intravoxel incoherent motion parameters in the liver. Magnetic Resonance in Medicine 87:859–871. DOI: https://doi.org/10.1002/mrm.28996, PMID: 34453445 Gomolka RS, Chrzan RM, Urbanik A, Kazmierski R, Grzanka AD, Nowinski WL. 2017. Quantification of image contrast of infarcts on computed tomography scans. The Neuroradiology Journal 30:15–22. DOI: https://doi. org/10.1177/1971400916678226, PMID: 28059673 Gomolka R, Nedergaard M, Mori Y, ESMRMB. 2021. 38th Annual Scientific Meeting. MAGMA. p. 95–96. Hablitz LM, Plá V, Giannetto M, Vinitsky HS, Stæger FF, Metcalfe T, Nguyen R, Benrais A, Nedergaard M. 2020. Circadian control of brain glymphatic and lymphatic fluid flow. Nature Communications 11:4411. DOI: https:// doi.org/10.1038/s41467-020-18115-2, PMID: 32879313 Haj- Yasein NN, Vindedal GF, Eilert- Olsen M, Gundersen GA, Skare Ø, Laake P, Klungland A, Thorén AE, Burkhardt JM, Ottersen OP, Nagelhus EA. 2011. Glial- conditional deletion of aquaporin- 4 (AQP4) reduces blood- brain water uptake and confers barrier function on perivascular astrocyte endfeet. PNAS 108:17815– 17820. DOI: https://doi.org/10.1073/pnas.1110655108, PMID: 21990350 Harrison IF, Ismail O, Machhada A, Colgan N, Ohene Y, Nahavandi P, Ahmed Z, Fisher A, Meftah S, Murray TK, Ottersen OP, Nagelhus EA, O’Neill MJ, Wells JA, Lythgoe MF. 2020. Impaired glymphatic function and clearance of tau in an Alzheimer’s disease model. Brain 143:2576–2593. DOI: https://doi.org/10.1093/brain/ awaa179, PMID: 32705145 Henkelman RM. 1990. Does IVIM measure classical perfusion? Magn Reson Med 16:470–475. DOI: https://doi. org/10.1002/mrm.1910160313 Henkelman RM, Neil JJ, Xiang QS. 1994. A quantitative interpretation of IVIM measurements of vascular perfusion in the rat brain. Magn Reson Med 32:464–469. DOI: https://doi.org/10.1002/mrm.1910320407 Hsu MS, Seldin M, Lee DJ, Seifert G, Steinhäuser C, Binder DK. 2011. Laminar- specific and developmental expression of aquaporin- 4 in the mouse hippocampus. Neuroscience 178:21–32. DOI: https://doi.org/10.1016/ j.neuroscience.2011.01.020, PMID: 21256195 Hubbard JA, Hsu MS, Seldin MM, Binder DK. 2015. Expression of the astrocyte water channel aquaporin- 4 in the mouse brain. ASN Neuro 7:1759091415605486. DOI: https://doi.org/10.1177/1759091415605486, PMID: 26489685 Huber VJ, Tsujita M, Nakada T. 2009. Identification of aquaporin 4 inhibitors using in vitro and in silico methods. Bioorganic & Medicinal Chemistry 17:411–417. DOI: https://doi.org/10.1016/j.bmc.2007.12.040, PMID: 18182301 Iacovetta C, Rudloff E, Kirby R. 2012. The role of aquaporin 4 in the brain. Veterinary Clinical Pathology 41:32–44. DOI: https://doi.org/10.1111/j.1939-165X.2011.00390.x, PMID: 22250904 Igarashi H, Huber VJ, Tsujita M, Nakada T. 2011. Pretreatment with a novel aquaporin 4 inhibitor, TGN- 020, significantly reduces ischemic cerebral edema. Neurological Sciences 32:113–116. DOI: https://doi.org/10. 1007/s10072-010-0431-1, PMID: 20924629 Iliff JJ, Wang M, Liao Y, Plogg BA, Peng W, Gundersen GA, Benveniste H, Vates GE, Deane R, Goldman SA, Nagelhus EA, Nedergaard M. 2012. A paravascular pathway facilitates CSF flow through the brain parenchyma and the clearance of interstitial solutes, including amyloid β. Science Translational Medicine 4:147ra111. DOI: https://doi.org/10.1126/scitranslmed.3003748, PMID: 22896675 Iliff JJ, Lee H, Yu M, Feng T, Logan J, Nedergaard M, Benveniste H. 2013. Brain- Wide pathway for waste clearance captured by contrast- enhanced MRI. The Journal of Clinical Investigation 123:1299–1309. DOI: https://doi.org/10.1172/JCI67677, PMID: 23434588 Insko EK, Bolinger L. 1993. Mapping of the radiofrequency field. Journal of Magnetic Resonance, Series A 103:82–85. DOI: https://doi.org/10.1006/jmra.1993.1133 Ishida K, Yamada K, Nishiyama R, Hashimoto T, Nishida I, Abe Y, Yasui M, Iwatsubo T. 2022. Glymphatic system clears extracellular tau and protects from tau aggregation and neurodegeneration. The Journal of Experimental Medicine 219:e20211275. DOI: https://doi.org/10.1084/jem.20211275, PMID: 35212707 Kaden E, Kelm ND, Carson RP, Does MD, Alexander DC. 2016. Multi- Compartment microscopic diffusion imaging. NeuroImage 139:346–359. DOI: https://doi.org/10.1016/j.neuroimage.2016.06.002, PMID: 27282476 Kantarci K, Jack CR Jr, Xu YC, Campeau NG, O’Brien PC, Smith GE, Ivnik RJ, Boeve BF, Kokmen E, Tangalos EG, Petersen RC. 2001. Mild cognitive impairment and Alzheimer disease: regional diffusivity of water. Radiology 219:101–107. DOI: https://doi.org/10.1148/radiology.219.1.r01ap14101, PMID: 11274543 Kantarci K, Xu Y, Shiung MM, O’Brien PC, Cha RH, Smith GE, Ivnik RJ, Boeve BF, Edland SD, Kokmen E, Tangalos EG, Petersen RC, Jack CR Jr. 2002. Comparative diagnostic utility of different Mr modalities in mild cognitive impairment and Alzheimer’s disease. Dementia and Geriatric Cognitive Disorders 14:198–207. DOI: https://doi.org/10.1159/000066021, PMID: 12411762 Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 29 of 36 Neuroscience Research article Katada R, Akdemir G, Asavapanumas N, Ratelade J, Zhang H, Verkman AS. 2014. Greatly improved survival and neuroprotection in aquaporin- 4- knockout mice following global cerebral ischemia. FASEB Journal 28:705–714. DOI: https://doi.org/10.1096/fj.13-231274, PMID: 24186965 Kitchen P, Day RE, Taylor LHJ, Salman MM, Bill RM, Conner MT, Conner AC. 2015. Identification and molecular mechanisms of the rapid tonicity- induced relocalization of the aquaporin 4 channel. The Journal of Biological Chemistry 290:16873–16881. DOI: https://doi.org/10.1074/jbc.M115.646034, PMID: 26013827 Kitchen P, Salman MM, Halsey AM, Clarke- Bland C, MacDonald JA, Ishida H, Vogel HJ, Almutiri S, Logan A, Kreida S, Al- Jubair T, Winkel Missel J, Gourdon P, Törnroth- Horsefield S, Conner MT, Ahmed Z, Conner AC, Bill RM. 2020. Targeting aquaporin- 4 subcellular localization to treat central nervous system edema. Cell 181:784–799.. DOI: https://doi.org/10.1016/j.cell.2020.03.037, PMID: 32413299 Kress BT, Iliff JJ, Xia M, Wang M, Wei HS, Zeppenfeld D, Xie L, Kang H, Xu Q, Liew JA, Plog BA, Ding F, Deane R, Nedergaard M. 2014. Impairment of paravascular clearance pathways in the aging brain. Annals of Neurology 76:845–861. DOI: https://doi.org/10.1002/ana.24271, PMID: 25204284 Kulkarni P, Grant S, Morrison TR, Cai X, Iriah S, Kristal BS, Honeycutt J, Brenhouse H, Hartner JC, Madularu D, Ferris CF. 2020. Characterizing the human APOE epsilon 4 knock- in transgene in female and male rats with multimodal magnetic resonance imaging. Brain Research 1747:147030. DOI: https://doi.org/10.1016/j.brainres. 2020.147030, PMID: 32745658 Latour LL, Svoboda K, Mitra PP, Sotak CH. 1994. Time- dependent diffusion of water in a biological model system. PNAS 91:1229–1233. DOI: https://doi.org/10.1073/pnas.91.4.1229, PMID: 8108392 Le Bihan D, Breton E, Lallemand D, Grenier P, Cabanis E, Laval- Jeantet M. 1986. Mr imaging of intravoxel incoherent motions: application to diffusion and perfusion in neurologic disorders. Radiology 161:401–407. DOI: https://doi.org/10.1148/radiology.161.2.3763909, PMID: 3763909 Le Bihan D, Breton E, Lallemand D, Aubin ML, Vignaud J, Laval- Jeantet M. 1988. Separation of diffusion and perfusion in intravoxel incoherent motion MR imaging. Radiology 168:497–505. DOI: https://doi.org/10.1148/ radiology.168.2.3393671, PMID: 3393671 Le Bihan D, Turner R. 1992. The capillary network: a link between IVIM and classical perfusion. Magnetic Resonance in Medicine 27:171–178. DOI: https://doi.org/10.1002/mrm.1910270116, PMID: 1435202 Le Bihan D, Iima M. 2015. Diffusion magnetic resonance imaging: what water tells us about biological tissues. PLOS Biology 13:e1002203. DOI: https://doi.org/10.1371/journal.pbio.1002203, PMID: 26204162 Le Bihan D. 2019. What can we see with IVIM MRI? NeuroImage 187:56–67. DOI: https://doi.org/10.1016/j. neuroimage.2017.12.062, PMID: 29277647 Lee JH, Cheong H, Lee SS, Lee CK, Sung YS, Huh J- W, Song J- A, Choe H. 2016. Perfusion assessment using intravoxel incoherent motion- based analysis of diffusion- weighted magnetic resonance imaging: validation through phantom experiments. Investigative Radiology 51:520–528. DOI: https://doi.org/10.1097/RLI. 0000000000000262, PMID: 26895196 Lemke A, Stieltjes B, Schad LR, Laun FB. 2011. Toward an optimal distribution of B values for intravoxel incoherent motion imaging. Magnetic Resonance Imaging 29:766–776. DOI: https://doi.org/10.1016/j.mri. 2011.03.004, PMID: 21549538 Li X, Kong H, Wu W, Xiao M, Sun X, Hu G. 2009. Aquaporin- 4 maintains ependymal integrity in adult mice. Neuroscience 162:67–77. DOI: https://doi.org/10.1016/j.neuroscience.2009.04.044, PMID: 19393298 Li C, Wang W. 2017. Molecular biology of aquaporins. Advances in Experimental Medicine and Biology 969:1– 34. DOI: https://doi.org/10.1007/978-94-024-1057-0_1, PMID: 28258563 Liao YP, Urayama SI, Isa T, Fukuyama H. 2021. Optimal model mapping for intravoxel incoherent motion MRI. Frontiers in Human Neuroscience 15:617152. DOI: https://doi.org/10.3389/fnhum.2021.617152, PMID: 33692677 Liu G, Mestre H, Sweeney AM, Sun Q, Weikop P, Du T, Nedergaard M. 2020. Direct measurement of cerebrospinal fluid production in mice. Cell Reports 33:108524. DOI: https://doi.org/10.1016/j.celrep.2020. 108524, PMID: 33357428 Lundgaard I, Osório MJ, Kress BT, Sanggaard S, Nedergaard M. 2014. White matter astrocytes in health and disease. Neuroscience 276:161–173. DOI: https://doi.org/10.1016/j.neuroscience.2013.10.050, PMID: 24231735 Meeus EM, Novak J, Withey SB, Zarinabad N, Dehghani H, Peet AC. 2017. Evaluation of intravoxel incoherent motion fitting methods in low- perfused tissue. Journal of Magnetic Resonance Imaging 45:1325–1334. DOI: https://doi.org/10.1002/jmri.25411, PMID: 27545824 Mestre H, Hablitz LM, Xavier AL, Feng W, Zou W, Pu T, Monai H, Murlidharan G, Castellanos Rivera RM, Simon MJ, Pike MM, Plá V, Du T, Kress BT, Wang X, Plog BA, Thrane AS, Lundgaard I, Abe Y, Yasui M, et al. 2018a. Aquaporin- 4- dependent glymphatic solute transport in the rodent brain. eLife 7:e40070. DOI: https:// doi.org/10.7554/eLife.40070, PMID: 30561329 Mestre H, Tithof J, Du T, Song W, Peng W, Sweeney AM, Olveda G, Thomas JH, Nedergaard M, Kelley DH. 2018b. Flow of cerebrospinal fluid is driven by arterial pulsations and is reduced in hypertension. Nature Communications 9:4878. DOI: https://doi.org/10.1038/s41467-018-07318-3, PMID: 30451853 Mestre H, Mori Y, Nedergaard M. 2020. The brain’s glymphatic system: current controversies. Trends in Neurosciences 43:458–466. DOI: https://doi.org/10.1016/j.tins.2020.04.003, PMID: 32423764 Nagelhus EA, Ottersen OP. 2013. Physiological roles of aquaporin- 4 in brain. Physiological Reviews 93:1543– 1562. DOI: https://doi.org/10.1152/physrev.00011.2013, PMID: 24137016 Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 30 of 36 Neuroscience Research article Nicholson C. 1993. Ion- Selective microelectrodes and diffusion measurements as tools to explore the brain cell microenvironment. Journal of Neuroscience Methods 48:199–213. DOI: https://doi.org/10.1016/0165-0270(93) 90092-6, PMID: 8412303 Nielsen S, Nagelhus EA, Amiry- Moghaddam M, Bourque C, Agre P, Ottersen OP. 1997. Specialized membrane domains for water transport in glial cells: high- resolution immunogold cytochemistry of aquaporin- 4 in rat brain. The Journal of Neuroscience 17:171–180. DOI: https://doi.org/10.1523/JNEUROSCI.17-01-00171.1997, PMID: 8987746 Niendorf T, Dijkhuizen RM, Norris DG, van Lookeren Campagne M, Nicolay K. 1996. Biexponential diffusion attenuation in various states of brain tissue: implications for diffusion- weighted imaging. Magnetic Resonance in Medicine 36:847–857. DOI: https://doi.org/10.1002/mrm.1910360607, PMID: 8946350 Nowinski WL, Gomolka RS, Qian G, Gupta V, Ullman NL, Hanley DF. 2014. Characterization of intraventricular and intracerebral hematomas in non- contrast CT. The Neuroradiology Journal 27:299–315. DOI: https://doi. org/10.15274/NRJ-2014-10042, PMID: 24976197 Oakes TR, Johnstone T, Ores Walsh KS, Greischar LL, Alexander AL, Fox AS, Davidson RJ. 2005. Comparison of fMRI motion correction software tools. NeuroImage 28:529–543. DOI: https://doi.org/10.1016/j.neuroimage. 2005.05.058, PMID: 16099178 Oberheim NA, Takano T, Han X, He W, Lin JHC, Wang F, Xu Q, Wyatt JD, Pilcher W, Ojemann JG, Ransom BR, Goldman SA, Nedergaard M. 2009. Uniquely hominid features of adult human astrocytes. The Journal of Neuroscience 29:3276–3287. DOI: https://doi.org/10.1523/JNEUROSCI.4707-08.2009, PMID: 19279265 Odackal J, Colbourn R, Odackal NJ, Tao L, Nicholson C, Hrabetova S. 2017. Real- time iontophoresis with tetramethylammonium to quantify volume fraction and tortuosity of brain extracellular space. Journal of Visualized Experiments 24:55755. DOI: https://doi.org/10.3791/55755, PMID: 28784968 Olesen JL, Østergaard L, Shemesh N, Jespersen SN. 2022. Diffusion time dependence, power- law scaling, and exchange in gray matter. NeuroImage 251:118976. DOI: https://doi.org/10.1016/j.neuroimage.2022.118976, PMID: 35168088 Palombo M, Ianus A, Guerreri M, Nunes D, Alexander DC, Shemesh N, Zhang H. 2020. SANDI: a compartment- based model for non- invasive apparent soma and neurite imaging by diffusion MRI. NeuroImage 215:116835. DOI: https://doi.org/10.1016/j.neuroimage.2020.116835, PMID: 32289460 Papadopoulos MC, Manley GT, Krishna S, Verkman AS. 2004. Aquaporin- 4 facilitates reabsorption of excess fluid in vasogenic brain edema. FASEB Journal 18:1291–1293. DOI: https://doi.org/10.1096/fj.04-1723fje, PMID: 15208268 Paschoal AM, Leoni RF, Dos Santos AC, Paiva FF. 2018. Intravoxel incoherent motion MRI in neurological and cerebrovascular diseases. NeuroImage. Clinical 20:705–714. DOI: https://doi.org/10.1016/j.nicl.2018.08.030, PMID: 30221622 Pavlin T, Nagelhus EA, Brekken C, Eyjolfsson EM, Thoren A, Haraldseth O, Sonnewald U, Ottersen OP, Håberg AK. 2017. Loss or mislocalization of aquaporin- 4 affects diffusion properties and intermediary metabolism in gray matter of mice. Neurochemical Research 42:77–91. DOI: https://doi.org/10.1007/s11064- 016-2139-y, PMID: 28039592 Peng W, Achariyar TM, Li B, Liao Y, Mestre H, Hitomi E, Regan S, Kasper T, Peng S, Ding F, Benveniste H, Nedergaard M, Deane R. 2016. Suppression of glymphatic fluid transport in a mouse model of Alzheimer’s disease. Neurobiology of Disease 93:215–225. DOI: https://doi.org/10.1016/j.nbd.2016.05.015, PMID: 27234656 Pfeuffer J, Flögel U, Dreher W, Leibfritz D. 1998. Restricted diffusion and exchange of intracellular water: theoretical modelling and diffusion time dependence of 1H NMR measurements on perfused glial cells. NMR in Biomedicine 11:19–31. DOI: https://doi.org/10.1002/(sici)1099-1492(199802)11:1<19::aid-nbm499>3.0.co;2-o, PMID: 9608585 Plá V, Bork P, Harnpramukkul A, Olveda G, Ladrón- de- Guevara A, Giannetto MJ, Hussain R, Wang W, Kelley DH, Hablitz LM, Nedergaard M. 2022. A real- time in vivo clearance assay for quantification of glymphatic efflux. Cell Reports 40:111320. DOI: https://doi.org/10.1016/j.celrep.2022.111320, PMID: 36103828 Preston GM, Carroll TP, Guggino WB, Agre P. 1992. Appearance of water channels in Xenopus oocytes expressing red cell CHIP28 protein. Science 256:385–387. DOI: https://doi.org/10.1126/science.256.5055.385, PMID: 1373524 Rash JE, Yasumura T, Hudson CS, Agre P, Nielsen S. 1998. Direct immunogold labeling of aquaporin- 4 in square arrays of astrocyte and ependymocyte plasma membranes in rat brain and spinal cord. PNAS 95:11981–11986. DOI: https://doi.org/10.1073/pnas.95.20.11981, PMID: 9751776 Rasmussen MK, Mestre H, Nedergaard M. 2022. Fluid transport in the brain. Physiological Reviews 102:1025– 1151. DOI: https://doi.org/10.1152/physrev.00031.2020, PMID: 33949874 Ronen I, Moeller S, Ugurbil K, Kim DS. 2006. Analysis of the distribution of diffusion coefficients in cat brain at 9.4 T using the inverse Laplace transformation. Magnetic Resonance Imaging 24:61–68. DOI: https://doi.org/ 10.1016/j.mri.2005.10.023, PMID: 16410179 Roth Y, Ocherashvilli A, Daniels D, Ruiz- Cabello J, Maier SE, Orenstein A, Mardor Y. 2008. Quantification of water compartmentation in cell suspensions by diffusion- weighted and T (2) -weighted MRI. Magnetic Resonance Imaging 26:88–102. DOI: https://doi.org/10.1016/j.mri.2007.04.013, PMID: 17574364 Salman MM, Kitchen P, Woodroofe MN, Brown JE, Bill RM, Conner AC, Conner MT. 2017. Hypothermia increases aquaporin 4 (AQP4) plasma membrane abundance in human primary cortical astrocytes via a calcium/ transient receptor potential vanilloid 4 (TRPV4)- and calmodulin- mediated mechanism. The European Journal of Neuroscience 46:2542–2547. DOI: https://doi.org/10.1111/ejn.13723, PMID: 28925524 Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 31 of 36 Neuroscience Research article Salman MM, Kitchen P, Iliff JJ, Bill RM. 2021. Aquaporin 4 and glymphatic flow have central roles in brain fluid homeostasis. Nature Reviews. Neuroscience 22:650–651. DOI: https://doi.org/10.1038/s41583-021-00514-z, PMID: 34408336 Salman MM, Kitchen P, Halsey A, Wang MX, Törnroth- Horsefield S, Conner AC, Badaut J, Iliff JJ, Bill RM. 2022. Emerging roles for dynamic aquaporin- 4 subcellular relocalization in CNS water homeostasis. Brain 145:64–75. DOI: https://doi.org/10.1093/brain/awab311, PMID: 34499128 Schneider MJ, Gaass T, Ricke J, Dinkel J, Dietrich O. 2019. Assessment of intravoxel incoherent motion MRI with an artificial capillary network: analysis of biexponential and phase- distribution models. Magnetic Resonance in Medicine 82:1373–1384. DOI: https://doi.org/10.1002/mrm.27816, PMID: 31131482 Simon MJ, Wang MX, Murchison CF, Roese NE, Boespflug EL, Woltjer RL, Iliff JJ. 2018. Transcriptional network analysis of human astrocytic endfoot genes reveals region- specific associations with dementia status and tau pathology. Scientific Reports 8:12389. DOI: https://doi.org/10.1038/s41598-018-30779-x, PMID: 30120299 Slator PJ, Hutter J, Marinescu RV, Palombo M, Jackson LH, Ho A, Chappell LC, Rutherford M, Hajnal JV, Alexander DC. 2021. Data- Driven multi- contrast spectral microstructure imaging with inspect: integrated spectral component estimation and mapping. Medical Image Analysis 71:102045. DOI: https://doi.org/10. 1016/j.media.2021.102045, PMID: 33934005 Smith AJ, Yao X, Dix JA, Jin BJ, Verkman AS. 2017. Test of the “ glymphatic ” hypothesis demonstrates diffusive and aquaporin- 4- independent solute transport in rodent brain parenchyma. eLife 6:e27679. DOI: https://doi. org/10.7554/eLife.27679, PMID: 28826498 Smith AJ, Verkman AS. 2018. The “glymphatic” mechanism for solute clearance in alzheimer’s disease: game changer or unproven speculation? FASEB Journal: Official Publication of the Federation of American Societies for Experimental Biology 32:543–551. DOI: https://doi.org/10.1096/fj.201700999 Stanton EH, Persson NDÅ, Gomolka RS, Lilius T, Sigurðsson B, Lee H, Xavier ALR, Benveniste H, Nedergaard M, Mori Y. 2021. Mapping of CSF transport using high spatiotemporal resolution dynamic contrast- enhanced MRI in mice: effect of anesthesia. Magnetic Resonance in Medicine 85:3326–3342. DOI: https://doi.org/10.1002/ mrm.28645, PMID: 33426699 Strohschein S, Hüttmann K, Gabriel S, Binder DK, Heinemann U, Steinhäuser C. 2011. Impact of aquaporin- 4 channels on K+ buffering and gap junction coupling in the hippocampus. Glia 59:973–980. DOI: https://doi. org/10.1002/glia.21169, PMID: 21446052 Sun C, Lin L, Yin L, Hao X, Tian J, Zhang X, Ren Y, Li C, Yang Y. 2022. Acutely inhibiting AQP4 with TGN- 020 improves functional outcome by attenuating edema and peri- infarct astrogliosis after cerebral ischemia. Frontiers in Immunology 13:870029. DOI: https://doi.org/10.3389/fimmu.2022.870029, PMID: 35592320 Sweeney AM, Plá V, Du T, Liu G, Sun Q, Peng S, Plog BA, Kress BT, Wang X, Mestre H, Nedergaard M. 2019. In vivo imaging of cerebrospinal fluid transport through the intact mouse skull using fluorescence macroscopy. Journal of Visualized Experiments 1:e74. DOI: https://doi.org/10.3791/59774, PMID: 31403617 Sylvain NJ, Salman MM, Pushie MJ, Hou H, Meher V, Herlo R, Peeling L, Kelly ME. 2021. The effects of trifluoperazine on brain edema, aquaporin- 4 expression and metabolic markers during the acute phase of stroke using photothrombotic mouse model. Biochimica et Biophysica Acta. Biomembranes 1863:183573. DOI: https://doi.org/10.1016/j.bbamem.2021.183573, PMID: 33561476 Takano K, Yamada M. 2020. Contrast- Enhanced magnetic resonance imaging evidence for the role of astrocytic aquaporin- 4 water channels in glymphatic influx and interstitial solute transport. Magnetic Resonance Imaging 71:11–16. DOI: https://doi.org/10.1016/j.mri.2020.05.001, PMID: 32446035 Thrane AS, Rappold PM, Fujita T, Torres A, Bekar LK, Takano T, Peng W, Wang F, Rangroo Thrane V, Enger R, Haj- Yasein NN, Skare Ø, Holen T, Klungland A, Ottersen OP, Nedergaard M, Nagelhus EA. 2011. Critical role of aquaporin- 4 (AQP4) in astrocytic ca2+ signaling events elicited by cerebral edema. PNAS 108:846–851. DOI: https://doi.org/10.1073/pnas.1015217108, PMID: 21187412 Trillo- Contreras JL, Ramírez- Lorca R, Hiraldo- González L, Sánchez- Gomar I, Galán- Cobo A, Suárez- Luna N, Sánchez de Rojas- de Pedro E, Toledo- Aral JJ, Villadiego J, Echevarría M. 2018. Combined effects of aquaporin- 4 and hypoxia produce age- related hydrocephalus. Biochimica et Biophysica Acta. Molecular Basis of Disease 1864:3515–3526. DOI: https://doi.org/10.1016/j.bbadis.2018.08.006, PMID: 30293570 Trillo- Contreras JL, Toledo- Aral JJ, Villadiego J, Echevarría M. 2021. Aquaporin- 4 mediates permanent brain alterations in a mouse model of hypoxia- aged hydrocephalus. International Journal of Molecular Sciences 22:9745. DOI: https://doi.org/10.3390/ijms22189745, PMID: 34575909 Ulv Larsen SM, Landolt H- P, Berger W, Nedergaard M, Knudsen GM, Holst SC. 2020. Haplotype of the astrocytic water channel AQP4 is associated with slow wave energy regulation in human NREM sleep. PLOS Biology 18:e3000623. DOI: https://doi.org/10.1371/journal.pbio.3000623, PMID: 32369477 Urushihata T, Takuwa H, Takahashi M, Kershaw J, Tachibana Y, Nitta N, Shibata S, Yasui M, Higuchi M, Obata T. 2021. Exploring cell membrane water exchange in aquaporin- 4- deficient ischemic mouse brain using diffusion- weighted MRI. European Radiology Experimental 5:44. DOI: https://doi.org/10.1186/s41747-021-00244-y, PMID: 34617156 Verkman AS. 2009. Knock- out models reveal new aquaporin functions. Handb Exp Pharmacol 1:359–381. DOI: https://doi.org/10.1007/978-3-540-79885-9_18, PMID: 19096787 Verkman AS, Smith AJ, Phuan PW, Tradtrantip L, Anderson MO. 2017. The aquaporin- 4 water channel as a potential drug target in neurological disorders. Expert Opinion on Therapeutic Targets 21:1161–1170. DOI: https://doi.org/10.1080/14728222.2017.1398236, PMID: 29072508 Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 32 of 36 Neuroscience Research article Vieni C, Ades- Aron B, Conti B, Sigmund EE, Riviello P, Shepherd TM, Lui YW, Novikov DS, Fieremans E. 2020. Effect of intravoxel incoherent motion on diffusion parameters in normal brain. NeuroImage 204:116228. DOI: https://doi.org/10.1016/j.neuroimage.2019.116228, PMID: 31580945 Wolburg H, Wolburg- Buchholz K, Fallier- Becker P, Noell S, Mack AF. 2011. Structure and functions of aquaporin- 4- based orthogonal arrays of particles. International Review of Cell and Molecular Biology 287:1–41. DOI: https://doi.org/10.1016/B978-0-12-386043-9.00001-3, PMID: 21414585 Wu D, Zhang J. 2019. Evidence of the diffusion time dependence of intravoxel incoherent motion in the brain. Magnetic Resonance in Medicine 82:2225–2235. DOI: https://doi.org/10.1002/mrm.27879, PMID: 31267578 Xavier ALR, Hauglund NL, von Holstein- Rathlou S, Li Q, Sanggaard S, Lou N, Lundgaard I, Nedergaard M. 2018. Cannula implantation into the cisterna magna of rodents. Journal of Visualized Experiments 1:57378. DOI: https://doi.org/10.3791/57378, PMID: 29889209 Xie L, Kang H, Xu Q, Chen MJ, Liao Y, Thiyagarajan M, O’Donnell J, Christensen DJ, Nicholson C, Iliff JJ, Takano T, Deane R, Nedergaard M. 2013. Sleep drives metabolite clearance from the adult brain. Science 342:373–377. DOI: https://doi.org/10.1126/science.1241224, PMID: 24136970 Xu Z, Xiao N, Chen Y, Huang H, Marshall C, Gao J, Cai Z, Wu T, Hu G, Xiao M. 2015. Deletion of aquaporin- 4 in APP/PS1 mice exacerbates brain Aβ accumulation and memory deficits. Molecular Neurodegeneration 10:58. DOI: https://doi.org/10.1186/s13024-015-0056-1, PMID: 26526066 Yao X, Hrabetová S, Nicholson C, Manley GT. 2008. Aquaporin- 4- deficient mice have increased extracellular space without tortuosity change. The Journal of Neuroscience 28:5460–5464. DOI: https://doi.org/10.1523/ JNEUROSCI.0257-08.2008, PMID: 18495879 Yushkevich PA, Piven J, Hazlett HC, Smith RG, Ho S, Gee JC, Gerig G. 2006. User- guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability. NeuroImage 31:1116– 1128. DOI: https://doi.org/10.1016/j.neuroimage.2006.01.015, PMID: 16545965 Zeppenfeld DM, Simon M, Haswell JD, D’Abreo D, Murchison C, Quinn JF, Grafe MR, Woltjer RL, Kaye J, Iliff JJ. 2017. Association of perivascular localization of aquaporin- 4 with cognition and alzheimer disease in aging brains. JAMA Neurology 74:91–99. DOI: https://doi.org/10.1001/jamaneurol.2016.4370, PMID: 27893874 Zhang Y, Brady M, Smith S. 2001. Segmentation of brain Mr images through a hidden Markov random field model and the expectation- maximization algorithm. IEEE Transactions on Medical Imaging 20:45–57. DOI: https://doi.org/10.1109/42.906424, PMID: 11293691 Zhang Y, Xu K, Liu Y, Erokwu BO, Zhao P, Flask CA, Ramos- Estebanez C, Farr GW, LaManna JC, Boron WF, Yu X. 2019. Increased cerebral vascularization and decreased water exchange across the blood- brain barrier in aquaporin- 4 knockout mice. PLOS ONE 14:e0218415. DOI: https://doi.org/10.1371/journal.pone.0218415, PMID: 31220136 Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 33 of 36 Neuroscience Research article Appendix 1 Supplementary correlational analysis To associate the findings from MR- DWI, a correlation between the mean AQP4 expression and mean ADC and D from the measured parenchymal ROI was computed for WTs. Additionally, to confirm whether ADC and D may reflect the free fluid volume, and thus the ISF space volume in the brain, Pearson’s linear correlation was calculated between the mean DWI and IVIM, and T1 values and free water volumes fractions from the phantoms. Similarly, the estimated mean vascular densities were correlated to the mean ADC and IVIM diffusion values and to the MR tracer dynamics scores ROI- wise, for both KO and WT separately. In search for potential relationships, correlation was computed further between AQP4 expression and the vascular density as well as between MR diffusion and MR tracer dynamics parameters ROI- wise. The correlations were calculated only for the set of parenchymal ROI assessed with both compared methods, and considered significant for the highest value of linear Pearson’s r- or range Spearman’s rho- correlation>50%, with pP<0.05 and non- zero slope of the regression line. Appendix 1—figure 1. Supplementary results for MR CSF space volumetry and MR diffusion evaluation. (A) Whiskers- box plot comparison of segmented CSF regions, where no statistical difference between KO and WT was found, along with overlaid 3D surface images of separate compartments (colored) from exemplary WT animal: CSF space at the skull base - Circle of Willis; perivascular space (PVS) around the basilar artery and its branches; PVS anteriorly and posteriorly, remaining after separating the ventricular and the space at the skull base; CSF space in the fourth ventricle; parietal PVS and cisterns (Figure 1—source data 1). (B) Correlation plot for ADC and D slow diffusion obtained from 6 AQP4 KO and 6 WT animals analyzed. (C) Radar plot showing statistical significances for the differences between IVIM scores, found among 5 parenchymal and 1 CSF space ROI assessed, for average and in 3 diffusion- encoding directions jointly (Figure 2—source data 1). (D) Correlation plots for ROI- wise comparison between IVIM measures and vascular densities from 12 parenchymal ROI analyzed with both methods in KO and WT animals, along with calculated linear and range correlation scores. (E) Fast diffusion IVIM measures calculated in a distilled water phantom (+0.001 mM/ml gadobutrol) and water phantoms filled with Sephadex- G25 microbeads of fine, moderate, and coarse sizes (Figure 2—source data 2). Legend: OLF- olfactory, CA / RSP- cingulate / retrosplenial, VIS (V1)- visual, SS (S1)- somatosensory, AUD- auditory, HIP- hippocampus, PERI- perirhinal, TH- thalamus, HAB- habenula, HY- hypothalamus, MB- midbrain, PAG- periaqueductal gray, HB- hindbrain; CP- Appendix 1—figure 1 continued on next page Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 34 of 36 Neuroscience Research article Appendix 1—figure 1 continued caudate putamen, WM- white matter; 3V- third ventricle, LV- lateral ventricle, 4V- fourth ventricle, PCS- pericisternal space, CB- cerebellum, D*-pseudodiffusion (fast diffusion), Fp- perfusion fraction; ns- not significant, *-p<0.05, **-p<0.01, by means of Mann- Whitney U- test (A, C), Kruskal- Wallis one- way ANOVA with Dunn’s correction (E). All correlation plots show respective regression lines along with semi- transparent areas marking 95% confidence intervals of the fitting. The highest obtained Pearson’s linear or Spearman’s range correlation scores reported and considered significant if correlation value >0.5 with P<0.05, and non- zero regression slope. Appendix 1—figure 2. Supplementary results for dynamic contrast- enhanced MRI in vivo and correlational analysis. (A) 3D multiplanar reconstructions of dynamic- contract enhanced (DCE) MRI – midsagittal, parasagittal, and lateral along with orthogonal axial and coronal slices from mean 3D FISP images from 5 AQP4 KO and 6 WT along with segmentation maps color- coded according to p- significance value from nonparametric Two- way ANOVA with post- hoc, for 3 time points: 30 minutes (top), 60 minutes (middle) and 90 minutes (full time, bottom) after applying gadobutrol injection via cisterna magna (CM) (Figure 3—source data 1). (B) Correlation plot for the relative duration and peak intensity differences between AQP4 KO and WT animals analyzed. (C) Correlation plots between DCE- derived scores and the vascular densities from 12 parenchymal ROI analyzed with both methods in AQP4 KO and WT animals. (D) Correlation plots between DCE- derived scores and AQP4 expression from 10 parenchymal ROI analyzed with both methods in AQP4 KO and WT animals. Legend: OLF- olfactory, CA / RSP- cingulate / retrosplenial, VIS (V1)- visual, SS (S1)- somatosensory, AUD- auditory, HIP- hippocampus, PERI- perirhinal, TH- thalamus, HAB- habenula, HY- hypothalamus, MB- midbrain, PAG- periaqueductal gray, HB- hindbrain; CP- caudate putamen, WM- white matter; 3V- third ventricle, LV- lateral ventricle, CoW- Circle of Willis, CM- cisterna magna, SSS- superior sagittal sinus; NS- not significant. All correlation plots show respective regression lines along with semi- transparent areas marking 95% confidence intervals of the fitting. The highest obtained Pearson’s linear or Spearman’s range correlation scores reported and considered significant if correlation value >0.5 with P<0.05, and non- zero regression slope. Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 35 of 36 Neuroscience Research article Appendix 1—figure 3. Schematic flow chart of the algorithm for the 3D- CISS- based CSF space volumetry and cisternography in Aqp4(-/-) (KO) and Aqp4(+/+) (WT) mice, including initial preprocessing (bias field correction and brain extraction) of the computed 3D- CISS volume (A) and subsequent automatic CSF space segmentation (B), based on representative 3D- CISS volume from a single animal. Gomolka et al. eLife 2023;12:e82232. DOI: https://doi.org/10.7554/eLife.82232 36 of 36 Neuroscience
10.1021_acs.inorgchem.2c03737
pubs.acs.org/IC Article Chiral Co3Y Propeller-Shaped Chemosensory Platforms Based on 19F‑NMR Gabrielle Audsley, Harry Carpenter, Nsikak B. Essien, James Lai-Morrice, Youssra Al-Hilaly, Louise C. Serpell, Geoffrey R. Akien, Graham J. Tizzard, Simon J. Coles, Cristina Pubill Ulldemolins,* and George E. Kostakis* Cite This: Inorg. Chem. 2023, 62, 2680−2693 Read Online ACCESS Metrics & More Article Recommendations *sı Supporting Information ABSTRACT: Two propeller-shaped chiral CoIII 3YIII complexes built from fluorinated ligands are synthesized and characterized by IR, UV−vis, circular single-crystal X-ray diffraction (SXRD), dichroism (CD), elemental analysis, thermogravimetric analysis (TGA), electron spray ionization mass spectroscopy (ESI-MS), and NMR (1H, 13C, and 19F). This work explores the sensing and discrimination abilities of these complexes, thus providing an innovative sensing method using a 19F NMR chemosensory system and opening new directions in 3d/4f chemistry. Control experiments and theoretical studies shed light on the sensing mechanism, while the scope and limitations of this method are discussed and presented. ■ INTRODUCTION The detection and differentiation of chiral enantiomers are important in synthetic and pharmaceutical chemistry, but they can prove challenging.1−6 Two enantiomers may possess different chemical and physical properties; thus, misidentifying their configuration can jeopardize biological and pharmaco- logical activities. For example, the stereochemistry of drugs can significantly affect their activity due to the inherent chirality found in the environments of biological systems. New biologically active chiral compounds are ever-increasing, where approximately 60% of all pharmaceutical drugs are chiral. Developments into asymmetric synthesis routes have stemmed from this increase in the production of chiral compounds; however, a complete stereochemical analysis including the absolute configuration, enantiomeric excess (ee), and total concentration has limited the discovery process. Conventional high-performance liquid chromatography (HPLC) separates the enantiomers, which are subsequently stereochemically analyzed. The HPLC technique involves a chiral column packed with a chiral stationary phase where enantiomers can be separated. Chiral additives can also be added to the mobile phase to separate enantiomers or form diastereomers beforehand; however, the extensive cost of these chiral columns is a limiting factor for scientific developments. Therefore, the discovery of other ease methods is in need. The stereochemical discrimination can occur using spectro- scopic methods, including circular dichroism (CD) and fluorescence,7−12 by monitoring absorbance intensity change and NMR, presenting chemically shifted signals for different chiral molecules or complexes.13−16 For the latter, a host− guest complex consisting of a chiral substrate sample interacts with a chiral detector molecule, transferring chiral information and inducing a change in the chiral environment, observed as split signals of precise chemical shifts in the corresponding NMR spectrum. NMR chemosensors for chiral discrimination are typically limited to aromatic-based compounds, facilitating signal splitting by inducing a significant shielding effect. 1H NMR dominates the differentiation of organic or biologically relevant molecules, but an increase in the size of these molecules and structural complexity leads to overlapping resonances on an already narrow spectral range, thus making discrimination impossible. Known 1H NMR analytical methods for chiral determination include the addition of chiral solvating agents (CSAs),17−19 reagents (CLSRs),20 and chiral derivatizing agents (CDAs).21 For example, Pirkle’s alcohol22 determines the absolute config- uration and enantiomeric purity of chiral molecules. Although Pirkle’s alcohol regularly used to analyze chiral issues have arisen related to the inaccurate molecules, determination of ee due to resonance overlap present in the 1H NMR spectra�a result of a narrow spectral width lanthanide shift is still chiral Received: October 23, 2022 Published: January 30, 2023 © 2023 The Authors. Published by American Chemical Society 2680 https://doi.org/10.1021/acs.inorgchem.2c03737 Inorg. Chem. 2023, 62, 2680−2693 Inorganic Chemistry pubs.acs.org/IC Article accompanied by the complicated spectra that arise with large organic compounds. Scheme 1. Previous and Current Complexes Used for Amine Sensing with 19F NMR Hence, incorporating molecules the development of heteronuclear-based sensors represents a reasonable alternative to overcome these hindrances. Methods that bear phosphorous atoms are developed to monitor changes with 31P{1H} NMR due to observable splitting of signals and a broad spectral range compared to 1H NMR spectra.15,23 Moreover, the 19F nucleus of spin quantum number 1/2 has 83% the sensitivity of the 1H nucleus and is found in 100% natural abundance; with great receptivity, it yields strong signals on an NMR spectrum.24−31 19F can also be easily incorporated into organic compounds as it mimics the 1H nucleus in many environments and benefits from a lack of background interference due to its low natural occurrence;32,33 thus, it can be used to probe the structure and dynamics of many large and complex biomolecules such as proteins.34 The broad detection window of 19F NMR is conducive to easily distinguishable split signals and narrow overlapping resonan- ces, increasing spectral resolution and providing the simpler deconvolution of complicated spectra. In 2015, Zhao and Swager35 presented a 19F chemosensory method to detect multiple chiral amines. The sensing strategy consisted of a palladium complex built from a chiral pincer ligand, of which the scaffold is facile in preparation alongside having well- known coordination chemistry (Scheme 1, upper).36,37 When varying the substituent R groups of the ligand, the authors envisioned an optimized design by moving the fluorine atoms closer to the analyte, confining the chiral pocket. This change successfully demonstrated more pronounced differences in the chemical shifts of the analytes, allowing for more time-efficient In addition to the detection with improved resolution. integrating these 19F assignment of absolute configuration, NMR spectra gave values in agreement with the actual enantiopurities of the chosen analytes. Following Zhao and Swager’s work, Song et al. designed a chiral sensor based on octahedral rhodium complexes containing fluorine (Scheme 1, middle).38 The advantage of two- coordinating site, which permits the sensing of monoamines, diamines, and amino acids. The scaffold was previously used as a successful chiral catalyst in asymmetric catalytic reac- tions.39,40 Screening tests identified DMSO-d6 as the optimum solvent, presenting a broader chemical shift difference of 0.21 ppm.. this complex is its Complexes formed from organic ligands, transition metal (3d) ions, and/or lanthanide (4f) ions constitute a large class of materials esteemed for their wide range of properties, including luminescence,41−43 magnetism,44−47 and cataly- sis.48−55 Synthetic-wise, in 3d−4f chemistry, the 3d or 4f metal ions can be targeted and selectively substituted by 3d or 4f ions with similar coordination properties, often without altering the topology of the core. This synthetic alteration has ion and 3d−3d ion been found in numerous 4f−4f substitutions56−60 but is proven difficult for the 4f−4f ion exchange due to the lanthanide contraction. The simultaneous presence of 3d and 4f elements in a single molecule represents an elegant advantage for studying mechanisms and ration- alizing experimental data. For example, YIII can be included in this category because it has a size and Lewis acidity similar to HoIII and its diamagnetic character permits monitoring with NMR (1H, 13C, 15N, 19F, or 89Y). It may also be possible to exchange 3d ions in the same oxidation state without distorting the core topology and permitting monitoring with UV−vis, EPR, or NMR. Similarly, the 4f entity may be replaced with GdIII (8S7/2 ground state) so that electron paramagnetic resonance can identify changes in the coordination environ- ment within the complex; this is a process that we have successfully demonstrated in the past.48,51 Recently, we identified the Lewis acidic catalytic efficacy of two chiral propeller-shaped [CoIII 3YIIIL6] complexes built from a known Schiff base ligand; other groups investigated the magnetic properties of these components.61−67 These two compounds retain their structure and chirality in solvent media for a prolonged period, over 6 months. The Y metal sits in the center with the CoIII centers and associated pairs of organic ligands forming the propeller wings. Each CoIII center sits in the center of an almost perfect octahedron, while the central YIII center bonds to six O atoms in a trigonal antiprismatic geometry. Given (a) the high ionic radii of 4f ions, thus being difficult ligand exchange with solvent substrates, (b) that amines are known to coordinate to lanthanide metal centers in organic media,68,69 (c) that the coordination geometry of the Co centers is fulfilled, and (d) that the 4f ion is captured in a trigonal antiprism, which limits access to solvent-substrate molecules only via axial sites, thus replicating the coordination sphere of the non-labile Pd and Rh examples,35,38 we envisioned these propeller-shaped molecules to control their 2681 https://doi.org/10.1021/acs.inorgchem.2c03737 Inorg. Chem. 2023, 62, 2680−2693 Inorganic Chemistry pubs.acs.org/IC Article as an ideal platform for limited ligand exchange with the solvent-lattice systems. Moreover, the development of 19F- NMR chemosensors relies on modifying already well-under- stood species; thus, we embarked on a project modifying the ligand used in our previous studies67 by adding a fluorine antenna on the organic scaffold to assemble a 3d−4f entity that can serve as a chiral 19F-NMR chemosensory platform. 19F NMR can then be used to monitor the sensing ability of the system, boasting a broad spectral range and avoidance of complicated and overlapping resonances, the key to efficiency, simplicity, and rapidity with a chiral detection method. The scope and limitations of our approach are discussed herein. ■ RESULTS AND DISCUSSION triethylamine would harvest Ligand and Complex Synthesis. The ligands (H2LR, H2LS and its racemic version H2L, see Table S1) can be synthesized in one high-yielding condensation reaction from commercial fluorinated salicylaldehyde and the corresponding amino alcohol (see the ESI). These ligands have never been used in coordination chemistry to our knowledge. Following our synthetic protocol, the assembly of the ligands with nitrate or chloride CoII/YIII salts in CH3CN or EtOH solvent media and the presence of the tetranuclear propeller-shaped entity; however, screening tests (Table S1) were carried out to optimize the synthetic procedure and obtain the targeted complexes in high purity. Both protic and aprotic, polar, and non-polar solvents were screened as the ligand can form a hydrogen bond, and it was unclear whether or not this would benefit or disadvantage the crystallization process. X-ray-quality crystals were only grown in EtOH, MeOH, and MeCN. Preliminary crystallographic characterization studies (see the ESI) identified that EtOH was not to be the solvent of choice; an additional non-coordinating ligand can be found in the lattice along with the expected 3YIIIL6][H2L]·x(EtOH)}. tetranuclear MeOH and MeCN produced single crystals of [CoIII 3YIIIL6]· x(MeOH) and [CoIII respectively, but comparatively, MeCN produced a higher yield and higher- purity crystals and additionally presented the same crystal product for both S and R enantiomers when using the chloride salts of both yttrium and cobalt sources. Then, the synthesis was repeated with nitrate salts of yttrium and cobalt instead of the chloride, and these reactions gave the highest yields. Last but not least, reactions at higher concentrations proved to be less efficient. targeted species{[CoIII 3YIIIL6]·x(MeCN), formula [CoIII Single X-ray Diffraction Characterization. Both com- pounds CR and CS crystallized from acetonitrile were characterized at the NCS UK facility. Both crystals diffract poorly at higher angles; therefore, an exact allocation of the lattice solvent molecules proves challenging, but a co- crystallized ligand, as was the case in the EtOH samples, cannot be identified. Both compounds are isostructural and 3YIIIL6]·x(MeCN), but the possess a general lattice acetonitrile molecules slightly deviate (x is 5.5 for R (CR) and 5.25 for S (CS), respectively). The relevant crystallographic tables, including crystal data, structure refine- ment, and bond lengths, are found in the Supplementary Material. Both compounds provide the targeted propeller- shaped structure in which the YIII center sits in the middle of a trigonal antiprism (Figure 1). The propeller itself is chiral, and this motif has been seen in other tetranuclear complexes and molecular compounds.66,70−78 Notably, in both compounds, the Flack parameter, an indication of chirality retention,79,80 Figure 1. X-ray structures of CR (upper) and CS (lower). Lattice (CH3CN) molecules and H atoms are omitted for clarity. Color code: Y (light blue), Co (pink), C (gray), N (blue), O (red), F (green). deviates from the typical values. This notion indicates that the racemization process may occur during crystallization (13 and 21% for CR and CS, respectively), as the crystals are formed and collected after 1 week. This racemization process may be an outcome of the ligands or the propeller shape,73 reversing their chirality. We collected better-quality data for the CR derivative using a Cu source to validate this notion further (Table S3). The analysis provides a Flack parameter of 7%, which is very close to 5% of the enantiomerically pure Fe4 compound,75 indicating that the enantiomeric purity of both CR and CS is of good levels. Given that in both cases (CR and CS), racemization is unavoidable under the current synthetic circumstances and the results presented in the following sections, we did not attempt to validate the Flack parameter of CS. NMR Characterization. As CS and CR are diamagnetic, retrieving 1H NMR without complications was possible. Comparing the 1H NMR spectra of both S- and R-derivatives of H2L and C, there is a clear upfield shift of the imine protons on the transformation of ligand to complex (Figure S2), consistent with a lesser deshielding effect present from the nitrogen now coordinating to cobalt with electronegativity drawn further away from the imine carbon. The C−H proton on the adjacent C to the N has shifted downfield, indicating an increase in shielding due to cobalt coordination. Additionally, the OH peak is lost due to ligand-to-metal coordination.7 The 19F NMR spectra (Figure S3) of both S- and R- derivatives of H2L and C also demonstrated a change in the chemical shift of the resultant resonance, confirming the formation of a new species. Notably, both the 1H and 19F thus NMR spectra of each pair of derivatives overlay, demonstrating an identical chemical shift pattern. 2682 https://doi.org/10.1021/acs.inorgchem.2c03737 Inorg. Chem. 2023, 62, 2680−2693 Inorganic Chemistry pubs.acs.org/IC Article CD. CD is a light absorption spectroscopy that “quantifies” chirality by measuring the differential absorption of optically active molecules in left- and right-handed circularly polarized light; thus, it is possible to investigate the structural features of optically active chiral molecules. CD measurements of enantiomers should be inequivalent but opposite. The spectra of both pairs of H2L and C in solution, acetonitrile 1 mM, confirmed active optical species, demonstrating opposite rotation (Figure 2). However, minor differences in the spectra Figure 2. Circular dichroism spectra of H2LS versus H2LR (upper); CS versus CR (lower). (i.e., minima and maxima at 285, 310, 410 nm for CS) may be attributed to the racemization process of the propeller or the ligands, evidenced by the Flack parameter (see above). Thermogravimetric Analysis. We conducted a thermog- ravimetric analysis (under a N2 atmosphere) of both compounds. Both complexes are thermally stable up to 220 °C (Figure S4). The initial weight loss (∼8%, 100 °C) corresponds to lattice CH3CN and possibly absorbed H2O molecules. Above 220 °C, a three-step decomposition is taking place. Taking into account the molecular formula of the crystal structure {CS 5.33 (CH3CN)} and that the anticipated 3YIIIO6] is 18.72%, we envisage that theoretical value of [CoIII decomposition completes possibly beyond 1000 °C. Titration Studies. The exploration of new 19F NMR chemosensory systems relies on repurposing already known systems;35,38 thus, our initial aim was to introduce a broad scope of chiral analytes used in previously known sensing procedures35,38 and assess the sensing ability of CS and CR. To reiterate how the 19F NMR chemosensor works, introducing an analyte to the sensing system induces a change in the environment of the complex, forming a new complex {CS + A}, which will be subsequently recognized as a new, second signal in the 19F NMR spectra. The type of sensor−analyte interaction, covalent bond, hydrogen bond, or aromatic interactions is unclear until further evidence is obtained. From previous works, amines are known to coordinate to lanthanide metal centers68 as well as the recent chiral sensing studies.15,35,38 It is suggestive that the amines will coordinate with the YIII center; however, due to the nature of the ligand with O and NH moieties, hydrogen bonding interactions between the complex and the analyte cannot be excluded. We used a range of chiral and non-chiral analytes. Preliminary titration tests were performed to standardize the process. Stock solutions of both enantiomers of the analytes were prepared in the following concentrations: 1, 2, 10, and 20 mM, following known protocols.35,38 Each concentration was trialed to decipher optimum reacting conditions�looking for intensity of signals and the chemical shift difference between enantiomeric pairs in the corresponding 19F NMR spectra. 1 mM stock solutions of both CS and CR were prepared in 10 Table 1. 19F NMR Spectra Using a Different Sensor (1 mM, CDCl3): Amine Ratios (1:1, 1:2, 1:10, 1:20, CDCl3) (T = 303 K) 2683 https://doi.org/10.1021/acs.inorgchem.2c03737 Inorg. Chem. 2023, 62, 2680−2693 Inorganic Chemistry pubs.acs.org/IC Article mL CDCl3, and 350 μL of A1 with each complex derivative (in all possible combinations) was then added directly to an NMR tube and recorded instantaneously. Preliminary testing began with investigating the stability of CS and CR in both CDCl3 and CD3OD, looking for any degradation or change in the spectrum over some time (Tables S5 and S6). Both solvents had been favored in previous works35,38 with their use dependent on the dissolving analyte�whether they required a protic or an aprotic environment. As demonstrated in Tables S5 and S6, there was no change in the 19F NMR spectra of either complex; however, a slight shift of the principal peak −130.06 ppm (CDCl3) over −131.22 ppm (CD3OD) is observed. Notably, 19F NMR spectra recorded after 1 week or month proved complex stability, representing a significant advantage to this sensing method; thus, stock solutions can be prepared and stored. In CD3OD, a coordinating solvent, an additional peak was found at approximately −139 ppm, which may correspond to {C-(CD3OD)x}, where x = 1 or 2 species. We then used CS to detect A1 in 64 scans, which requires approximately 2 min of running time in CDCl3 and at 30 °C (Table 1). Initial efforts of 1:1 and 1:2 ratios, as seen to be most successful in previous studies,35,38 did not present a split signal; therefore, increasing the amine loading was necessary. In fact, this was found to be an advantage as an analyte signal was achievable using less complex. Both the 1:10 and 1:20 ratios demonstrated amine peaks, but the chemical shift difference between enantiomeric pairs for 1:20 was broader, allowing for more direct discrimination and boasting the advantage of requiring less complex. Therefore, we chose to use the ratio 1:20 for further experimental studies. Then, we examined if additional scans (128 vs 64) shall improve the sensing performance and incorporate compound CR (Table S6). A 20 mM concentration of A1 (in all possible combinations) in CDCl3 at 30 °C with CS and CR was run for 128 scans (Table S7). CS allowed for better chiral discrimination with a broader enantiomeric difference in chemical shift compared to CR. However, as the chemical shift differences were not too dissimilar, 64 scans were deemed more efficient at half the running time. Repeats were then undertaken using A1, but changing the solvent system from CDCl3 to CD3OD; however, no analyte response can be detected at T = 0 (Table S8). An analyte peak appeared at T = 96 h, suggesting that CD3OD can be used for the detection studies but not with an immediate effect. After discovering that the sensor can work at large amine-to- complex ratios, the ratio was increased from 1:20 to 1:50 to explore how far this factor may be extended (Table S9). The resulting 19F NMR spectra confirmed the sensing of the A1 analyte. However, the signals were slightly less prominent than with the 1:20 ratio; the chemical shift differences improved by approximately 0.1 ppm in comparison. These enantiomeric differences were better than or comparable to previous works of chiral sensing using 19F sensors,35,38 thus making this complex a competitor against current well-defined metal complex sensors. This notion may suggest that a sensor:amine ratio of 1:50 allows for more rapid discrimination of further analytes. Analyte Scope and Limitations. After achieving successful results for the sensing of A1, further analytes were introduced (Scheme 2). No additional analyte peak was present for compounds A2, A3 (monoamines), and A4 (diol). Then, we incorporated other amino alcohols (A5−A8) and Scheme 2. The Type of Analytes Tested in This Work amino acids (A9−A12) with the same 1,2-amino alcohol backbone compared to A1. For the amino acids, we used CD3OD instead of CDCl3 and NaOMe for dissolving purposes, as demonstrated in previous studies.38 Amino alcohols A5−A7 were successfully sensed but required different ratios (1:100, 1:50, and 1:20, respectively, Table 2), as in some instances, the amine signal was not intense enough to account for a resolved signal. All amino acids (A9−A12, Table 3) were sensed successfully using a 1:50 ratio. Both complexes at 0 h are stable when an excess of amino acid is added. The sensed amino alcohol and amino acid analytes (19F NMR spectra found in Tables S10 and S11) were both chiral and non-chiral species. While the chiral species demonstrated how C can successfully sense and discriminate between the S- and R-configurations of each analyte by a chemical shift difference, the non-chiral species (A5 and A12) were introduced to assess the sensing ability of this complex concerning molecular recognition. A split signal correspondent of a non-chiral species introduces a further use for these complexes and chiral discrimination, which is the sensing of species that do not require chiral discrimination by their configurational derivative or induce changes in the chiral environment. For example, this can be developed to detect and recognize specific organic or biological molecules. A8 was not sensed, implying that the five-membered ring holding the 1,2- amino alcohol motif sterically hinders the ability of the motif to approach, and thus interact with, the sensor. To overcome this obstacle, we attempted to obtain the corresponding propeller- shaped structure (Scheme 3), replacing the bulky phenyl ring with a methyl group. Large red block-shaped crystals were instantaneously formed, almost quantitively; however, prelimi- nary and sole single X-ray diffraction characterization (Figure S6) identified the formation of the neutral CoIII species, as shown in Scheme 3. Future efforts will optimize the reaction conditions to obtain the targeted propeller-based structure. the mechanism for Plausible Sensing Mechanisms. Recognition of chiral and non-chiral analytes makes this complex−analyte interaction interesting. The 1,2-amino alcohol motif must be present on the analyte for sensing to is is unclear whether occur. Further, representative of an induced change in the chiral environment 3YIIIL6] by affecting the inherent chirality of coordination to the central YIII or instead supports the notion that a ligand exchange interaction facilitates this additional peak or a weak H-bonding interaction occurs, as shown in the split signal the [CoIII it 2684 https://doi.org/10.1021/acs.inorgchem.2c03737 Inorg. Chem. 2023, 62, 2680−2693 Inorganic Chemistry pubs.acs.org/IC Article Table 2. 19F NMR Spectra of Complex (1 mM, CDCl3) and different Amino Alcohols (xx mM, CDCl3, T = 303 K)a complex CS CR CS CS CR CR CS CR CR CS analyte A5 A5 A6S A6R A6S A6R A7 A7 A8 A8 ratio 1:100 1:100 1:50 1:50 1:50 1:50 1:20 1:20 1:20 1:20 time (h) complex peak complex + analyte peak difference 0 0 0 0 0 0 0 0 0 0 −129.98 −129.98 −129.98 −129.98 −129.98 −129.98 −129.98 −129.98 −129.98 −129.98 −132.22 −132.06 −132.72 −132.62 −132.65 −132.56 −132.70 −132.72 no peak no peak 0.10 0.09 aThe corresponding 19F NMR spectra are given in the order of the table. Data for A8 are omitted for clarity. Scheme 4. In the first case (A, Scheme 4), the resultant mono- adduct holds a capped trigonal antiprism geometry and would in the 19F NMR spectrum, correspond to a split signal differentiating from the original peak. As YIII tends to favor a coordination number of 7 or 8, coordination of a further analyte to the YIII metal center from the remaining plane cannot be excluded and would correspond to a bis-adduct. Here, the geometry becomes bicapped trigonal antiprismatic and would correspond to an additional signal in the 19F NMR spectrum. It is relevant to reiterate the additional signal that accompanied the C peak in CD3OD (see Table 3 and S6) at approximately −139 ppm. Coordination of CD3OD to the YIII 3YIIIL6(CD3OD)x} (x = 1 or 2) metal center can form {CoIII and, with the further introduction of an analyte, would observe three individual peaks corresponding to {CoIII 3YIIIL6}, {CoIII 3YIIIL6 (CD3OD)x}, similar to Table 4. In the second and third scenarios (B and C, Scheme 4), given that all analytes bear the same 1,2-amino alcohol motif, the additional signal found in the 19F NMR spectrum can be the result of a partial (B) or full (C) ligand exchange process that yields the formation of a new compound 3YIIIL5L′}, as this has been proposed in a previous {CoIII work.38 Last but not least, numerous weak hydrogen bonding interactions between the complex and the analyte can occur in the solutions (a representative example, D, is drawn), which may cause the appearance of a second peak in the 19F NMR spectrum. 3YIIIL6(analyte)}, and {CoIII Control Experiments and Theoretical Studies. We performed computational calculations; recorded 1H, 89Y, and additional 19F NMR data; and used other analytes to probe the mechanistic path. Initially, we performed 1H-decoupled 19F NMR experiments to validate the shape and nature of the observed peaks (Figure 3A). The data were recorded with the same number of scans, and the samples had the same concentration. In the decoupled spectrum, the multiplet peak of the complex appears as single with the same intensity; however, the “analyte+complex” peak retains its broad character, indicative of a chemical exchange and several types of interactions. Using the chiral analytes containing the N−N motif such as SS or RR diphenylethy- lenediamine in the 1:20 ratio (Figure 3B) provided a second peak at −139.08 ppm. These data were recorded with the inevitable use of coordinating d6-DMSO solvent for solubility purposes. Comparing the data in Table 4, this peak can be attributed to the {CoIII 3YIIIL6(DMSO)x} (x = 1 or 2) species; therefore, sensing of analytes with the NN pocket proves challenging. Then, we recorded a sample in CDCl3 (Figure 3C) containing CR, (R)-(−)-2-phenylglycinol, and (S)-(+)-2- amino-1-propanol in a 1:100:100 ratio at 1 mM concentration. The samples of CR with (R)-(−)-2-phenylglycinol and (S)- (+)-2-amino-1-propanol in a 1:100 ratio were also recorded and shown for convenience. These data indicate that CR cannot be used to discriminate between two analytes with similar pocket sizes since only one, possibly average, peak at 2685 https://doi.org/10.1021/acs.inorgchem.2c03737 Inorg. Chem. 2023, 62, 2680−2693 Inorganic Chemistry pubs.acs.org/IC Article Table 3. 19F NMR Spectra of Complex (1 mM, CD3OD) and Different Amino Acids (50 mM, CD3OD)a complex CS CR CS CS CS CR CR CR CS CS CS CR CR CR CS CS CS CR CR CR CS amine ratio A9D A9L A9D/L A9D A9L A9D/L A10D A10L A10D/L A10D A10L A10D/L A11D A11L A11D/L A11D A11L A11D/L A12 1:50 1:50 1:50 1:50 1:50 1:50 1:50 1:50 1:50 1:50 1:50 1:50 1:50 1:50 1:50 1:50 1:50 1:50 1:50 solvent CD3OD CD3OD CD3OD CD3OD CD3OD CD3OD CD3OD CD3OD CD3OD CD3OD CD3OD CD3OD CD3OD CD3OD CD3OD CD3OD CD3OD CD3OD CD3OD CD3OD CD3OD complex peak complex + analyte peak complex + solvent upfield peak enantiomer difference −131.23 −131.23 −131.23 −131.23 −131.23 −131.23 −131.23 −131.23 −131.23 −131.23 −131.23 −131.23 −131.23 −131.23 −131.23 −131.23 −131.23 −131.23 −131.23 −131.23 −131.23 −134.69 −134.66 −134.62 −135.13 −135.16 −135.07 −135.09 −135.31 −135.09 −135.29 −135.29 −135.24 −135.29 −135.29 −135.27 −135.31 −138.99 −138.99 −139.28 −139.28 −139.28 −139.28 −139.28 −139.28 −138.99 −138.99 −138.99 −138.99 −138.99 −138.99 −138.99 −138.99 −138.99 −138.99 −138.99 −138.99 −138.99 0.03 0.03 0.22 0.00 0.00 aThe corresponding 19F NMR spectra are given in the order of the table (T = 303 K). Scheme 3. The Anticipated and Obtained Species Following the Synthetic Recipe Described in This Work −132.15 ppm is observed. Next, our studies concentrated on increasing the temperature to facilitate ligand exchange, if any, or alter the H-bonding interactions, thus providing better spectral resolution and improved chemical shift differences (Figure 3D). These data indicate that a full ligand (mechanism C, Scheme 4) exchange process takes place since two new peaks appear; one corresponds to the free ligand (−126 ppm) and the second (∼163 ppm) to an unknown species. Then, we recorded 1H−89Y data for the complex and the complex with an excess of phenylglycinol (Figure S7). The data of the the YIII center with the complex show an interaction of methylenic protons of the ligand (Figure S7 up), which is retained when an excess of phenylglycinol is present (Figure S7 down); however, no other peak that would suggest the presence of a {CoIII 3YIIIL6(analyte)x}, since the analyte is in excess, is observed. Last, we recorded the data of the same sample (complex + analyte ratio 1:20) after 4 and 10 days, and three peaks can be identified (Figure 3E). These three peaks 2686 https://doi.org/10.1021/acs.inorgchem.2c03737 Inorg. Chem. 2023, 62, 2680−2693 Inorganic Chemistry pubs.acs.org/IC Article Scheme 4. Plausible Sensing Mechanisms can be attributed to the free ligand (H2L), the complex (C), and the complex + analyte. After 10 days, the intensity of the complex peak significantly drops, whereas the intensity of the peak that corresponds to the free ligand significantly increases, signifying that a dynamic ligand substitution is responsible for the sensing process. We then performed theoretical calcu- lations, trialed B3LYP/SDD and M06/Def2-TZVP levels in the recently reported Rh system,38 for the first time, to evaluate their efficacy (Figure S8, Tables S10 and S11, respectively), and identified optimum performance with the M06/Def2- TZVP level (Table S10). Encouraged by the excellent agreement of the computed 19F NMR chemical shifts, we 2687 https://doi.org/10.1021/acs.inorgchem.2c03737 Inorg. Chem. 2023, 62, 2680−2693 Inorganic Chemistry pubs.acs.org/IC Article Table 4. Computed 19F NMR Chemical Shifts for the Resulting {CoIII Mechanisms in Scheme 4 and Comparison to the Reported Experimental NMR Values in This Work 3YIIIL6(Analyte)x} Complexes According to the agreement (%) complex with analyte mechanism solvent experimental 19F NMR shift (ppm) complex only (Co3YL6) complex onlyc (Co3YL6) A1R A1S A1Sc A6R A6S A6Sc A9R A9S A1R A1S A6R A6S A1R A1S A6R A6S A9R A9S 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 aCalculations were performed at the B3LYP/SDD level of theory with the polarizable continuum model (PCM) as the implicit solvent model. bFluorobenzene (C6H5F) was used as the 19F NMR reference calculated at the same level of theory. cThis calculation was performed at the M06/ Def2-TZVP level. dSaturated system with six ligands partially exchanged. none none partial ligand exchange partial ligand exchange partial ligand exchange partial ligand exchange partial ligand exchange partial ligand exchange partial ligand exchange partial ligand exchange 6× partial ligand exchanged 6× partial ligand exchange 6× partial ligand exchange 6× partial ligand exchange ligand substitution ligand substitution ligand substitution ligand substitution ligand substitution ligand substitution computed 19F NMR shift (ppm)a,b −130.30 −133.32 −129.99 −130.43 −127.81 −130.45 −130.64 −127.92 −129.29 −129.64 −129.11 −130.43 −130.64 −131.64 −131.25 −131.11 −131.38 −131.30 −129.21 −129.36 −131.23 −131.23 −132.05 −132.12 −132.12 −132.72 −132.62 −132.62 −134.66 −134.69 −132.05 −132.12 −132.72 −132.62 −132.05 −132.12 −132.72 −132.62 −134.66 −134.69 99.3% 98.4% >99.9% 98.7% 96.7% 98.4% 98.4% 97.6% 96% 96.2% 99.3% 98.7% 98.5% 99.2% 99.1% 99.2% 99.1% 98.9% 96% 92/9% MeOD MeOD CDCl3 CDCl3 CDCl3 CDCl3 CDCl3 CDCl3 MeOD MeOD CDCl3 CDCl3 CDCl3 CDCl3 CDCl3 CDCl3 CDCl3 CDCl3 MeOD MeOD computed the 19F NMR chemical shifts for the four plausible mechanisms shown in Scheme 4. As depicted in Table 4, there is an excellent agreement between the computed 19F NMR chemical shifts of the {CoIII 3YIIIL6} complex alone in MeOD for both levels of theory employed (entries 1 and 2) but slightly better for the B3LYP/SDD system (entry 1). Our attempts to compute the complex with the mono- and bis-adduct, which means {CoIII 3YIIIL6 + (analyte)x, where x = 1 or 2, and analyte (A1, A6, and A9) (Scheme 4, mechanism A) suggest that complex {CoIII 3YIIIL6} is very stable in all cases (Figure S9, Table S12). The computed 19F NMR chemical shifts for the resulting complex analytes from the partial ligand exchange mechanism (Scheme 4, mechanism B) are in very good agreement with the experimental values (Table 4, entries 3−10). A slightly better agreement was observed for the computed NMR values for the {CoIII 3YIIIL6} complexes when a complete (six times) ligand exchange takes place (Table 4, entries 11−14). Under the experimental evidence gathered in this work, we can argue that the best agreement can be met 3YIIIL5A} species from the ligand with the resulting {CoIII substitution mechanism (C, Scheme 4, Table 4, entries 15 to 20). Bearing all these in mind, and considering that complex C is incapable of sensing diols and diamines, the latter due to solubility issues, our chemical intuition inclines mechanism C, Scheme 4, as the most probable. Comparison with Previous Methods. Referring to the previous work of Zhao and Swager,35 utilizing a Pd-pincer complex and 19F NMR for the sensing of chiral amines, it is vital to explore the differences in their chiral chemosensory system as opposed to our system. First, the Pd sensor is unsuitable for amino acids; it allows discrimination of amino alcohols using 1:1 and 1:2 ratios of sensor:analyte with chemical shift differences of enantiomeric pairs ranging from 0.1 to 1 ppm. Comparatively, the discrimination of amino the sensor alcohol enantiomeric pairs using these CoIII 3YIII complexes presented a range of 0.02−0.18 ppm, with the higher results of a sensor:amine ratio 1:50. In addition to the advantages�the need to use less amount of to detect and discriminate between the analytes and boasting broader chemical shift differences�the window for detection, which is the shift difference between the original sensor peak and analyte peak, is broader. For the Pd sensor, the maximum window is 2.5 ppm, whereas using the CS or CR, this difference increases to 8 ppm. This increased difference benefits from a more straightforward interpretation of the resultant spectra and differentiation between existing complex and new analyte signals. Limitations arise with the complex where it cannot sense monoamines, as they lack the significant 1,2-amino alcohol motif present in both amino alcohol and amino acid analogues�the discrimination of monoamines is possible using the Pd complex; however, it again requires a more extensive complex loading with a ratio of 1:1 or 1:2. Moreover, ee determination is possible using the palladium sensor but not doable for CS or CR. Song et al.38 used fluorinated Rh complex and discriminated chiral enantiomers of monoamines, diamines, and amino acids, thus outperforming CS or CR with enantiomeric differences of up to 0.21 ppm. The ratio of the sensor to the analyte is 1:1. They also found that changing the solvent from CDCl3 to DMSO-d6 increases the chemical shift difference between the two signals of an enantiomeric pair reproduced in other literature using 19F chemosensing systems.16 However, examining the subsequent 19F NMR spectra for both amino alcohols and amino acids in this Rh system, a pair of fluorine resonances can be found per analyte, as on coordination of the analyte to the Rh sensor, the sensor becomes asymmetric with each fluorine inequivalent. Consequently, it is unclear which signal of each pair of signals should be used to compare the 2688 https://doi.org/10.1021/acs.inorgchem.2c03737 Inorg. Chem. 2023, 62, 2680−2693 Inorganic Chemistry pubs.acs.org/IC Article Figure 3. (A) A comparison of the 1H-decoupled 19F NMR of the CR complex with phenylglycinol. (B) The 19F NMR spectra of the CR complex with SS or RR diphenylethylenediamine in a 1:20 ratio in the CDCl3:DMSO 9:1 ratio. (C) The 19F NMR of the CR complex with different amino alcohols in the 1:100 ratio at 1 mM concentration. (D) Variable-temperature 19F NMR data indicating that the increase in the temperature facilitates the ligand exchange. (E) 19F NMR data of complex (1 mM, CDCl3) + analyte (1 mM, CDCl3) samples left undisturbed after 4 and 10 days, and other compounds for comparison at T = 303 K. enantiomer difference in chemical shift. We only observe an additional fluorine resonance representing the analyte along- side the original peak and providing more facile discrimination using uncomplicated spectra for any amino alcohol and the amino acid analyte in question. To follow, taking even the farthest signal distance of each enantiomeric pair, a maximum of 1 ppm is found for both amino acids and amino alcohols, which again is less than the most significant values for our complexes�0.22 ppm for amino acids and 0.18 ppm for amino alcohols (both in the 1:50 ratio). Interestingly, when optimizing sensing conditions, Song et al. found that increasing the equivalence of the sensor and extending the reaction time to 2 h with a reading taken at 50 °C showed more precise fluorine signals, which is therefore indicative of equilibrium favoring the coordinated analyte complex product. A more general comparison of Zhao and Swager and Song et al.’s complexes35,38 versus ours highlights the incorporation of low-cost, and non-toxic metals and the use of abundant, commercially available ligands for the synthesis of CS and CR. 2689 https://doi.org/10.1021/acs.inorgchem.2c03737 Inorg. Chem. 2023, 62, 2680−2693 Inorganic Chemistry pubs.acs.org/IC Article ■ CONCLUSIONS We present the first example of a 3d/4f 19F-NMR chemo- sensory system and identify the scope and limitations of this method. These air-stable and easy-to-make complexes are built from non-toxic and inexpensive metals and retain their structure in the solution for a prolonged period; however, the crystallization solvent may impact a needless racemization process or improvise unnecessary impurities. Complexes CS and CR are applicable to sense a specific type of analytes bearing an NH2CX-CHX-OH pocket via a dynamic ligand exchange mechanism (C, Scheme 4). Despite the limited analyte library, our method imposes an extreme sensor:analyte ratio (1:20 or 1:50) and a broad sensing window (8 ppm over 0.8 ppm, as seen in other studies), which are advantageous over other techniques and thus can be used to detect and recognize organic or biological molecules bearing this specific motif at millimolar concentrations. Future work will focus on modifying the existing propeller-shaped motif by (a) replacing the central YIII unit of the propeller-shaped structure with GdIII or EuIII/TbIII ions to allow sensing investigations of the related species with EPR or fluorescence, (b) modifying the organic ligand to enhance the F signal or remove aromatic interactions, and (c) overcoming the racemization effect during crystal- lization, which will become prone to investigate the chemo- sensing amine abilities of system with CD spectroscopy. this 3d/4f ■ ASSOCIATED CONTENT Data Availability Statement CCDC deposition numbers 2195856−2195858 contain the supplementary crystallographic data for this paper. *sı Supporting Information The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.inorgchem.2c03737. Copies of ESI-MS data, 1H and 19F NMR, TGA, UV− vis, and coordinates for theoretical calculations (PDF) Accession Codes CCDC 2195856−2195858 contain the supplementary crys- tallographic data for this paper. These data can be obtained free of charge via www.ccdc.cam.ac.uk/data_request/cif, or by emailing data_request@ccdc.cam.ac.uk, or by contacting The Cambridge Crystallographic Data Centre, 12 Union Road, Cambridge CB2 1EZ, UK; fax: +44 1223 336033. ■ AUTHOR INFORMATION Corresponding Authors Cristina Pubill Ulldemolins − Department of Chemistry, School of Life Sciences, University of Sussex, Brighton BN1 9QJ, UK; Present Address: Department of Nutrition, Food Sciences and Gastronomy, Faculty of Pharmacy and Nutrition, INSA-University of Barcelona, Barcelona; Email: C.Pubill-Ulldemolins@sussex.ac.uk George E. Kostakis − Department of Chemistry, School of Life Sciences, University of Sussex, Brighton BN1 9QJ, UK; orcid.org/0000-0002-4316-4369; Email: G.Kostakis@ sussex.ac.uk Authors Gabrielle Audsley − Department of Chemistry, School of Life Sciences, University of Sussex, Brighton BN1 9QJ, UK Harry Carpenter − Department of Chemistry, School of Life Sciences, University of Sussex, Brighton BN1 9QJ, UK Nsikak B. Essien − Department of Chemistry, School of Life Sciences, University of Sussex, Brighton BN1 9QJ, UK James Lai-Morrice − Department of Chemistry, School of Life Sciences, University of Sussex, Brighton BN1 9QJ, UK; orcid.org/0000-0001-6319-0698 Youssra Al-Hilaly − Sussex Neuroscience, School of Life Sciences, University of Sussex, Brighton BN1 9QG, UK; Chemistry Department, College of Science, Mustansiriyah University, Baghdad 10001, Iraq; 2289-4597 orcid.org/0000-0003- Louise C. Serpell − Sussex Neuroscience, School of Life Sciences, University of Sussex, Brighton BN1 9QG, UK; orcid.org/0000-0001-9335-7751 Geoffrey R. Akien − Department of Chemistry, Lancaster University, Lancaster LA1 4YB, UK Graham J. Tizzard − UK National Crystallography Service, Chemistry, University of Southampton, Southampton SO1 71BJ, UK Simon J. Coles − UK National Crystallography Service, Chemistry, University of Southampton, Southampton SO1 71BJ, UK; orcid.org/0000-0001-8414-9272 Complete contact information is available at: https://pubs.acs.org/10.1021/acs.inorgchem.2c03737 Author Contributions G.A. performed the synthesis and characterization of CS and CR and performed sensing studies. H.C. and N.B.E. expanded the synthetic and sensing aspects. J.L.-M. and C.P.U. performed theoretical calculations. Y.A-.H. and L.C.S. provided the equipment and performed CD studies. G.R.A. performed VT 19F NMR, and 89Y NMR studies. G.E.K., G.J.T., and S.J.C. performed crystallographic analysis. G.E.K. conceptualization. All authors contributed to writing this article. Notes The authors declare no competing financial interest. ■ ACKNOWLEDGMENTS G.E.K. thanks the EPSRC UK National Crystallography Service at the University of Southampton for collecting the crystallographic data for CS and CR,81 as well as Prof. Annie K. Powell for CHN analysis data. N.B.E. thanks the TETFUND for financial support. We thank Dr. Ramon Gonzalez-Mendez, University of Sussex, for the ESI-MS data. ■ REFERENCES Ibnusaud, (1) Willia, K.; Lee, E. Importance of Drug Enantiomers in Clinical Pharmacology. Drugs. Springer October 1985, 30, 333−354. (2) Nguyen, L. A.; He, H.; Pham-Huy, C. Chiral Drugs: An Overview. Int. J. Biomed. Sci. 2006, 2, 85−100. (3) Brooks, W. H.; Guida, W. C.; Daniel, K. G. The Significance of Chirality in Drug Design and Development. Curr. Top. Med. Chem. 2011, 11, 760−770. (4) Polavarapu, P. L.; Scalmani, G.; Hawkins, E. K.; Rizzo, C.; Jeirath, N.; I.; Habel, D.; Nair, D. S.; Haleema, S. Importance of Solvation in Understanding the Chiroptical Spectra of Natural Products in Solution Phase: Garcinia Acid Dimethyl Ester. J. Nat. Prod. 2011, 74, 321−328. (5) Allred, T. K.; Manoni, F.; Harran, P. G. Exploring the Boundaries of “Practical:” De Novo Syntheses of Complex Natural Product-Based Drug Candidates. Chem. Rev. 2017, 117, 11994− 12051. (6) Li, X.; Duan, M.; Yu, P.; Houk, K. N.; Sun, J. Organocatalytic Enantioselective Dearomatization of Thiophenes by 1,10-Conjugate Addition of Indole Imine Methides. Nat. Commun. 2021, 12, 4881. 2690 https://doi.org/10.1021/acs.inorgchem.2c03737 Inorg. Chem. 2023, 62, 2680−2693 Inorganic Chemistry pubs.acs.org/IC Article (7) Bentley, K. W.; Wolf, C. Stereodynamic Chemosensor with Selective Circular Dichroism and Fluorescence Readout for in Situ Determination of Absolute Configuration, Enantiomeric Excess, and Concentration of Chiral Compounds. J. Am. Chem. Soc. 2013, 135, 12200−12203. (8) Shcherbakova, E. G.; Brega, V.; Minami, T.; Sheykhi, S.; James, T. D.; Anzenbacher, P. Toward Fluorescence-Based High-Through- put Screening for Enantiomeric Excess in Amines and Amino Acid Derivatives. Chem. − Eur. J. 2016, 22, 10074−10080. (9) De los Santos, Z. A.; MacAvaney, S.; Russell, K.; Wolf, C. Tandem Use of Optical Sensing and Machine Learning for the Determination of Absolute Configuration, Enantiomeric and Diaster- eomeric Ratios, and Concentration of Chiral Samples. Angew. Chem., Int. Ed. 2020, 59, 2440−2448. (10) De los Santos, Z. A.; Lynch, C. C.; Wolf, C. Optical Chirality Sensing with an Auxiliary-Free Earth-Abundant Cobalt Probe. Angew. Chem., Int. Ed. 2019, 58, 1198−1202. (11) Bentley, K. W.; Proano, D.; Wolf, C. Chirality Imprinting and Direct Asymmetric Reaction Screening Using a Stereodynamic Brønsted/Lewis Acid Receptor. Nat. Commun. 2016, 7, 12539. (12) Dragna, J. M.; Pescitelli, G.; Tran, L.; Lynch, V. M.; Anslyn, E. V.; Di Bari, L. In Situ Assembly of Octahedral Fe(II) Complexes for the Enantiomeric Excess Determination of Chiral Amines Using Circular Dichroism Spectroscopy. J. Am. Chem. Soc. 2012, 134, 4398− 4407. (13) Chaudhari, S. R.; Suryaprakash, N. Simple and Efficient Methods for Discrimination of Chiral Diacids and Chiral Alpha- Methyl Amines. Org. Biomol. Chem. 2012, 10, 6410−6419. (14) Xu, Z.; Liu, C.; Zhao, S.; Chen, S.; Zhao, Y. Molecular Sensors for NMR-Based Detection. Chem. Rev. 2019, 119, 195−230. (15) Chen, Z.; Yang, M.; Sun, Z.; Zhang, X.; Xu, J.; Bian, G.; Song, L. Chiral Discrimination by a Binuclear Pd Complex Sensor Using 31P{1H} NMR. Anal. Chem. 2019, 91, 14591−14596. (16) Dong, C.; Xu, Z.; Wen, L.; He, S.; Wu, J.; Deng, Q. H.; Zhao, Y. Tailoring Sensors and Solvents for Optimal Analysis of Complex Mixtures Via Discriminative 19F NMR Chemosensing. Anal. Chem. 2021, 93, 2968−2973. (17) Seo, M. S.; Kim, H. 1H NMR Chiral Analysis of Charged Molecules via Ion Pairing with Aluminum Complexes. J. Am. Chem. Soc. 2015, 137, 14190−14195. (18) Puentes, C. M.; Wenzel, T. J. Phosphated Cyclodextrins as Water-Soluble Chiral NMR Solvating Agents for Cationic Com- pounds. Beilstein J. Org. Chem. 2017, 13, 43−53. (19) Sun, Z.; Chen, Z.; Wang, Y.; Zhang, X.; Xu, J.; Bian, G.; Song, L. Chiral Discrimination of Varied Ammonium Compounds through 1h Nmr Using a Binuclear Ti Complex Sensor. Org. Lett. 2020, 22, 589−593. (20) Hinckley, C. C. Paramagnetic Shifts in Solutions of Cholesterol and the Dipyridine Adduct of Trisdipivalomethanatoeuropium(III). A Shift Reagent. J. Am. Chem. Soc. 1969, 91, 5160−5162. J. A.; Dull, D. L.; Mosher, H. S. α-Methoxy-α- (21) Dale, the Trifluoromethylphenylacetic Acid, a Versatile Reagent Determination of Enantiomeric Composition of Alcohols and Amines. J. Org. Chem. 1969, 34, 2543−2549. (22) Pirkle, W. H.; Sikkenga, D. L.; Pavlin, M. S. Nuclear Magnetic Resonance Determination of Enantiomeric Composition and Absolute Configuration of γ-Lactones Using Chiral 2,2,2-Trifluoro- 1-(9-Anthryl)Ethanol. J. Org. Chem. 1977, 42, 384−387. (23) Bravo, J.; Cativiela, C.; Chaves, J. E.; Navarro, R.; Urriolabeitia, E. P. 31P NMR Spectroscopy as a Powerful Tool the Determination of Enantiomeric Excess and Absolute Configurations of α-Amino Acids. Inorg. Chem. 2003, 42, 1006−1013. (24) Gimenez, D.; Phelan, A.; Murphy, C. D.; Cobb, S. L. 19F NMR as a Tool in Chemical Biology. Beilstein J. Org. Chem. 2021, 17, 293− 318. (25) Chen, H.; Viel, S.; Ziarelli, F.; Peng, L. 19F NMR: A Valuable Tool for Studying Biological Events. Chem. Soc. Rev. 2013, 42, 7971− 7982. for for (26) Rosenau, C. P.; Jelier, B. J.; Gossert, A. D.; Togni, A. Exposing the Origins of Irreproducibility in Fluorine NMR Spectroscopy. Angew. Chem., Int. Ed. 2018, 57, 9528−9533. (27) Norton, R. S.; Leung, E. W. W.; Chandrashekaran, I. R.; MacRaild, C. A. Applications of 19F-NMR in Fragment-Based Drug Discovery. Molecules 2016, 21, 860. (28) Buchholz, C. R.; Pomerantz, W. C. K. 19F NMR Viewed through Two Different Lenses: Ligand-Observed and Protein- for Fragment-Based Drug Observed 19F NMR Applications Discovery. RSC Chem. Biol. 2021, 2, 1312−1330. (29) Cobb, S. L.; Murphy, C. D. 19F NMR Applications in Chemical Biology. J. Fluorine Chem. 2009, 130, 132−143. (30) Boeszoermenyi, A.; Ogórek, B.; Jain, A.; Arthanari, H.; Wagner, G. The Precious Fluorine on the Ring: Fluorine NMR for Biological Systems. J. Biomol. NMR 2020, 74, 365−379. (31) Dolbier, W. R. Guide to Fluorine NMR for Organic Chemists; John Wiley & Sons, Inc., 2016, DOI: 10.1002/9781118831106. (32) Yu, J. X.; Hallac, R. R.; Chiguru, S.; Mason, R. P. New Frontiers and Developing Applications in 19F NMR. Prog. Nucl. Magn. Reson. Spectrosc. 2013, 70, 25−49. (33) Yu, J.; Kodibagkar, V.; Cui, W.; Mason, R. 19F: A Versatile Reporter for Non-Invasive Physiology and Pharmacology Using Magnetic Resonance. Curr. Med. Chem. 2005, 12, 819−848. (34) Danielson, M. A.; Falke, J. J. Use of 19F NMR to Probe Protein Structure and Conformational Changes. Annu. Rev. Biophys. Biomol. Struct. 1996, 25, 163−195. (35) Zhao, Y.; Swager, T. M. Simultaneous Chirality Sensing of Multiple Amines by 19F NMR. J. Am. Chem. Soc. 2015, 137, 3221− 3224. (36) Reed, J. E.; White, A. J. P.; Neidle, S.; Vilar, R. Effect of Metal Coordination on the Interaction of Substituted Phenanthroline and Pyridine Ligands with Quadruplex DNA. Dalton Trans. 2009, 14, 2558−2568. (37) Yamnitz, C. R.; Negin, S.; Carasel, I. A.; Winter, R. K.; Gokel, G. W. Dianilides of Dipicolinic Acid Function as Synthetic Chloride Channels. Chem. Commun. 2010, 46, 2838−2840. (38) Wang, W.; Xia, X.; Bian, G.; Song, L. A Chiral Sensor for Recognition of Varied Amines Based on 19 F NMR Signals of Newly Designed Rhodium Complexes. Chem. Commun. 2019, 55, 6098− 6101. (39) Wang, C.; Chen, L. A.; Huo, H.; Shen, X.; Harms, K.; Gong, L.; Meggers, E. Asymmetric Lewis Acid Catalysis Directed by Octahedral Rhodium Centrochirality. Chem. Sci. 2015, 6, 1094−1100. (40) Huo, H.; Shen, X.; Wang, C.; Zhang, L.; Röse, P.; Chen, L. A.; Harms, K.; Marsch, M.; Hilt, G.; Meggers, E. Asymmetric Photoredox Transition-Metal Catalysis Activated by Visible Light. Nature 2014, 515, 100−103. (41) Jankolovits, J.; Kampf, J. W.; Pecoraro, V. L. Insight into the Structural Versatility of the Ln(III)[15-Metallacrown-5] Platform by Comparing Analogs with Ni(II), Cu(II), and Zn(II) Ring Ions. Polyhedron 2013, 52, 491−499. (42) Bo, L.; Wang, S.; Schipper, D.; Yang, X.; Zhu, T.; Tao, J. Self- Assembly of Luminescent Zn-Ln (Ln = Sm and Nd) Nanoclusters with a Long-Chain Schiff Base Ligand. New J. Chem. 2018, 42, 7241− 7246. (43) Zhang, F.; Li, Z.; Ge, T.; Yao, H.; Li, G.; Lu, H.; Zhu, Y. Four Novel Frameworks Built by Imidazole-Based Dicarboxylate Ligands: Hydro(Solvo)Thermal Synthesis, Crystal Structures, and Properties. Inorg. Chem. 2010, 49, 3776−3788. (44) Liu, J. L.; Wu, J. Y.; Chen, Y. C.; Mereacre, V.; Powell, A. K.; Ungur, L.; Chibotaru, L. F.; Chen, X. M.; Tong, M. L. A Heterometallic FeII-DyIII Single-Molecule Magnet with a Record Anisotropy Barrier. Angew. Chemie - Int. Ed. 2014, 53, 12966−12970. (45) Schmitz, S.; Kovalchuk, A.; Martín-Rodríguez, A.; van Leusen, J.; Izarova, N. V.; Bourone, S. D. M.; Ai, Y.; Ruiz, E.; Chiechi, R. C.; Kögerler, P.; Monakhov, K. Y. Element-Selective Molecular Charge Transport Characteristics of Binuclear Copper(II)-Lanthanide(III) Complexes. Inorg. Chem. 2018, 57, 9274−9285. 2691 https://doi.org/10.1021/acs.inorgchem.2c03737 Inorg. Chem. 2023, 62, 2680−2693 Inorganic Chemistry pubs.acs.org/IC Article (46) Dey, A.; Acharya, J.; Chandrasekhar, V. Heterometallic 3d−4f Complexes as Single-Molecule Magnets. Chem. - Asian J. 2019, 14, 4433−4453. (47) Wilson, L. R. B.; Coletta, M.; Evangelisiti, M.; Piligkos, S.; Dalgarno, S. J.; Brechin, E. K. The Coordination Chemistry of P-Tert- Butylcalix[4]Arene with Paramagnetic Transition and Lanthanide Metal Ions: An Edinburgh Perspective. Dalton Trans. 2022, 51, 4213−4226. (48) Griffiths, K.; Kumar, P.; Akien, G. R.; Chilton, N. F.; Abdul- Sada, A.; Tizzard, G. J.; Coles, S. J.; Kostakis, G. E. Tetranuclear Zn/ 4f Coordination Clusters as Highly Efficient Catalysts for Friedel Crafts Alkylation. Chem. Commun. 2016, 52, 7866−7869. (49) Griffiths, K.; Kostakis, G. E. Transformative 3d-4f Coordination Cluster Carriers. Dalton Trans. 2018, 47, 12011−12034. (50) Griffiths, K.; Gallop, C. W. D.; Abdul-Sada, A.; Vargas, A.; Navarro, O.; Kostakis, G. E. Heteronuclear 3 d/DyIII Coordination Clusters as Catalysts in a Domino Reaction. Chem. - A Eur. J. 2015, 21, 6358−6361. (51) Griffiths, K.; Tsipis, A. C.; Kumar, P.; Townrow, O. P. E.; Abdul-Sada, A.; Akien, G. R.; Baldansuren, A.; Spivey, A. C.; Kostakis, G. E. 3d/4f Coordination Clusters as Cooperative Catalysts for Highly Diastereoselective Michael Addition Reactions. Inorg. Chem. 2017, 56, 9563−9573. (52) Sampani, S. I.; McGown, A.; Vargas, A.; Abdul-Sada, A.; Tizzard, G. J.; Coles, S. J.; Spencer, J.; Kostakis, G. E. Solvent-Free Synthesis and Key Intermediate Isolation in Ni 2 Dy 2 Catalyst Development in the Domino Ring-Opening Electrocyclization Reaction of Furfural and Amines. J. Org. Chem. 2019, 84, 6858−6867. (53) Wang, L.; Xu, C.; Han, Q.; Tang, X.; Zhou, P.; Zhang, R.; Gao, G.; Xu, B.; Qin, W.; Liu, W. Ambient Chemical Fixation of CO2 Using a Highly Efficient Heterometallic Helicate Catalyst System. Chem. Commun. 2018, 54, 2212−2215. (54) Xu, R.; Hua, L.; Li, X.; Yao, Y.; Leng, X.; Chen, Y. Rare-Earth/ Zinc Heterometallic Complexes Containing Both Alkoxy-Amino- Bis(Phenolato) and Chiral Salen Ligands: Synthesis and Catalytic Application for Copolymerization of CO2 with Cyclohexene Oxide. Dalt. Trans. 2019, 48, 10565−10573. (55) Ramirez, B. L.; Lu, C. C. Rare-Earth Supported Nickel Catalysts for Alkyne Semihydrogenation: Chemo- And Regioselectiv- ity Impacted by the Lewis Acidity and Size of the Support. J. Am. Chem. Soc. 2020, 142, 5396−5407. (56) Chen, S.; Mereacre, V.; Zhao, Z.; Zhang, W.; Zhang, M.; He, Z. Targeted Replacement: Systematic Studies of Dodecanuclear {MIII6LnIII6} Coordination Clusters (M = Cr, Co; Ln = Dy, Y). Dalton Trans. 2018, 47, 7456−7462. (57) Goura, J.; Guillaume, R.; Rivière, E.; Chandrasekhar, V.; Rivie, E. Hexanuclear, Heterometallic, Ni3n3 Complexes Possessing O- Capped Homo- and Heterometallic Structural Subunits: SMM Behavior of Inorg. Chem. 2014, 53, 7815−7823. (58) Chandrasekhar, V.; Dey, A.; Das, S.; Rouzières, M.; Clérac, R. Syntheses, Structures, and Magnetic Properties of a Family of Heterometallic Heptanuclear [Cu5Ln2] (Ln = Y(III), Lu(III), Dy(III), Ho(III), Er(III), and Yb(III)) Complexes: Observation of SMM Behavior for the Dy(III) and Ho(III) Analogues. Inorg. Chem. 2013, 52, 2588−2598. (59) Moreno Pineda, E.; Chilton, N. F.; Tuna, F.; Winpenny, R. E. P.; McInnes, E. J. L. Systematic Study of a Family of Butterfly-Like {M2Ln2} Molecular Magnets (M = MgII, MnIII, CoII, NiII, and CuII; Ln = YIII, GdIII, TbIII, DyIII, HoIII, and ErIII). Inorg. Chem. 2015, 54, 5930−5941. (60) Hooper, T. N.; Schnack, J.; Piligkos, S.; Evangelisti, M.; Brechin, E. K. The Importance of Being Exchanged: [Gd(III)4M- (II)8(OH)8(L)8(O2CR)8]4+ Clusters for Magnetic Refrigeration. Angew. Chem., Int. Ed. 2012, 51, 4633−4636. (61) Zhu, Y.-Y.; Zhang, Y.-Q.; Yin, T.-T.; Gao, C.; Wang, B.-W.; Gao, S. A Family of Co(II)Co(III)3 Single-Ion Magnets with Zero- Field Slow Magnetic Relaxation: Fine Tuning of Energy Barrier by the Dysprosium Analogue. Remote Substituent and Counter Cation. Inorg. Chem. 2015, 54, 5475−5486. (62) Mayans, J.; Font-Bardia, M.; Di Bari, L.; Górecki, M.; Escuer, A. Chiral [MnIIMnIII3M′] (M′=NaI, CaII, MnII) and [MnIIMnIII6- NaI2] Clusters Built from an Enantiomerically Pure Schiff Base: Synthetic, Chiroptical, and Magnetic Properties. Chem. - A Eur. J. 2018, 24, 18705−18717. (63) Pilichos, E.; Escuer, A.; Font-Bardia, M.; Mayans, J. Chiral Versus Non-Chiral [MnIII6MnIINaI], [MnIII6MnII2NaI2] and [MnIII3MnIINaI] Clusters Derived from Schiff Bases or the Fight for Symmetry. Chem. - A Eur. J. 2020, 26, 13053−13062. (64) Escuer, A.; Mayans, J.; Font-Bardia, M.; Górecki, M.; Bari, L. D. Syntheses, Structures, and Chiroptical and Magnetic Properties of Chiral Clusters Built from Schiff Bases: A Novel [MnIIMnIII6NaI2] Core. Dalton Trans. 2017, 46, 6514−6517. (65) Hu, P.; Wang, X. N.; Jiang, C. G.; Yu, F.; Li, B.; Zhuang, G. L.; Zhang, T. Nanosized Chiral [Mn6Ln2] Clusters Modeled by Enantiomeric Schiff Base Derivatives: Synthesis, Crystal Structures, and Magnetic Properties. Inorg. Chem. 2018, 57, 8639−8645. (66) Mayans, J.; Font-Bardia, M.; Escuer, A. Chiroptical and Magnetic Properties of Star-Shaped Fe III 4 Complexes from Chiral Schiff Bases. Structural and Magnetic Correlations Based on Continuous Shape Measures. Dalton Trans. 2018, 47, 8392−8401. (67) Spivey, A. C.; Kostakis, G. E. et al., to be submitted. (68) Veits, G. K.; Read de Alaniz, J. Dysprosium(III) Catalysis in Organic Synthesis. Tetrahedron 2012, 68, 2015−2026. (69) Yu, D.; Thai, V. T.; Palmer, L. I.; Veits, G. K.; Cook, J. E.; Read De Alaniz, J.; Hein, J. E. Importance of Off-Cycle Species in the Acid- Catalyzed Aza-Piancatelli Rearrangement. J. Org. Chem. 2013, 78, 12784−12789. (70) Singh, R.; Banerjee, A.; Colacio, E.; Rajak, K. K. Enantiopure Tetranuclear Iron(III) Complexes Using Chiral Reduced Schiff Base Ligands: Synthesis, Structure, Spectroscopy, Magnetic Properties, and DFT Studies. Inorg. Chem. 2009, 48, 4753−4762. (71) Clemens, J. B.; Kibar, O.; Chachisvilis, M. A Molecular Propeller Effect for Chiral Separation and Analysis. Nat. Commun. 2015, 6, 1−10. (72) Zhang, Y.; Calupitan, J. P.; Rojas, T.; Tumbleson, R.; Erbland, G.; Kammerer, C.; Ajayi, T. M.; Wang, S.; Curtiss, L. A.; Ngo, A. T.; Ulloa, S. E.; Rapenne, G.; Hla, S. W. A Chiral Molecular Propeller Designed for Unidirectional Rotations on a Surface. Nat. Commun. 2019, 10, 1−9. (73) Shimizu, Y.; Shoji, Y.; Hashizume, D.; Nagata, Y.; Fukushima, T. Sensing the Chirality of Various Organic Solvents by Helically Arranged π-Blades. Chem. Commun. 2018, 54, 12314−12317. (74) Cornia, A.; Mannini, M.; Sessoli, R.; Gatteschi, D. Propeller- Shaped Fe4 and Fe3M Molecular Nanomagnets: A Journey from Crystals to Addressable Single Molecules. Eur. J. Inorg. Chem. 2019, 552−568. (75) Zhu, Y.-Y.; Guo, X.; Cui, C.; Wang, B.-W.; Wang, Z.-M.; Gao, S. An Enantiopure FeIII4 Single-Molecule Magnet. Chem. Commun. 2011, 47, 8049−8051. (76) Ninova, S.; Lanzilotto, V.; Malavolti, L.; Rigamonti, L.; Cortigiani, B.; Mannini, M.; Totti, F.; Sessoli, R. Valence Electronic Structure of Sublimated Fe4 Single-Molecule Magnets: An Exper- imental and Theoretical Characterization. J. Mater. Chem. C 2014, 2, 9599−9608. (77) Katoono, R.; Kawai, H.; Fujiwara, K.; Suzuki, T. Dynamic Molecular Propeller: Supramolecular Chirality Sensing by Enhanced Chiroptical Response through the Transmission of Point Chirality to Mobile Helicity. J. Am. Chem. Soc. 2009, 131, 16896−16904. (78) Martinez, A.; Guy, L.; Dutasta, J. P. Reversible, Solvent-Induced Chirality Switch in Atrane Structure: Control of the Unidirectional Motion of the Molecular Propeller. J. Am. Chem. Soc. 2010, 132, 16733−16734. (79) Flack, H. D. On Enantiomorph-polarity Estimation. Acta Crystallogr. Sect. A 1983, 39, 876−881. (80) Watkin, D. J.; Cooper, R. I. Howard Flack and the Flack Parameter. Chemistry 2020, 2, 796−804. 2692 https://doi.org/10.1021/acs.inorgchem.2c03737 Inorg. Chem. 2023, 62, 2680−2693 Inorganic Chemistry pubs.acs.org/IC Article (81) Coles, S. J.; Gale, P. A. Changing and Challenging Times for Service Crystallography. Chem. Sci. 2012, 3, 683−689. 2693 https://doi.org/10.1021/acs.inorgchem.2c03737 Inorg. Chem. 2023, 62, 2680−2693
10.7554_elife.58744
RESEARCH ARTICLE Seizures are a druggable mechanistic link between TBI and subsequent tauopathy Hadeel Alyenbaawi1,2,3, Richard Kanyo1,4, Laszlo F Locskai1,4, Razieh Kamali-Jamil1,5, Miche` le G DuVal4, Qing Bai6, Holger Wille1,5, Edward A Burton6,7, W Ted Allison1,2,4* 1Centre for Prions & Protein Folding Disease, University of Alberta, Edmonton, Canada; 2Department of Medical Genetics, University of Alberta, Edmonton, Canada; 3Majmaah University, Majmaah, Saudi Arabia; 4Department of Biological Sciences, University of Alberta, Edmonton, Canada; 5Department of Biochemistry, University of Alberta, Edmonton, Canada; 6Department of Neurology, University of Pittsburgh, Pittsburgh, United States; 7Geriatric Research, Education and Clinical Center, Pittsburgh VA Healthcare System, Pittsburgh, United States Abstract Traumatic brain injury (TBI) is a prominent risk factor for dementias including tauopathies like chronic traumatic encephalopathy (CTE). The mechanisms that promote prion-like spreading of Tau aggregates after TBI are not fully understood, in part due to lack of tractable animal models. Here, we test the putative role of seizures in promoting the spread of tauopathy. We introduce ‘tauopathy reporter’ zebrafish expressing a genetically encoded fluorescent Tau biosensor that reliably reports accumulation of human Tau species when seeded via intraventricular brain injections. Subjecting zebrafish larvae to a novel TBI paradigm produced various TBI features including cell death, post–traumatic seizures, and Tau inclusions. Bath application of dynamin inhibitors or anticonvulsant drugs rescued TBI-induced tauopathy and cell death. These data suggest a role for seizure activity in the prion-like seeding and spreading of tauopathy following TBI. Further work is warranted regarding anti-convulsants that dampen post-traumatic seizures as a route to moderating subsequent tauopathy. Introduction Traumatic brain injury (TBI) is a leading cause of mortality and disability worldwide (Hay et al., 2016; Nguyen et al., 2016; Rimel et al., 1981). It is also a prominent risk factor for neurodegeneration and dementia, such as chronic traumatic encephalopathy (CTE) (Chauhan, 2014; Gardner and Yaffe, 2015; Uryu et al., 2007). TBI can result from direct physical insults, from rapid acceleration and deceleration of the brain, or from shock wave impacts such as pressure waves emanating from explosive blasts (Cruz-Haces et al., 2017). Regardless, the primary mechanisms have much in com- mon and the neuropathology in TBI and CTE patients includes the wide distribution of hyperphos- phorylated Tau pathology, axonal degeneration, and neuronal (Hay et al., 2016; Johnson et al., 2013; McKee et al., 2015; Ojo et al., 2016). The mechanisms whereby physical injury is translated into progressive Tau pathology remain unresolved and represent prospective therapeutic targets. Progress on this front is hampered by lack of access to suitable models: apply- ing physical injury to a cell culture is difficult and poorly represents the complex biopathology that intertwines many multifaceted aspects of brain physiology. loss The progressive deposition of hyperphosphorylated Tau protein in filamentous forms is a defining hallmark of tauopathies, which includes Alzheimer’s disease (AD), CTE, and several other dementias. Each of the tauopathies affects distinct brain regions and has a unique clinical presentation (Kovacs, 2017; Orr et al., 2017). Early in CTE, hyperphosphorylated Tau is accumulated in a cluster *For correspondence: ted.allison@ualberta.ca Competing interests: The authors declare that no competing interests exist. Funding: See page 28 Received: 09 May 2020 Accepted: 07 December 2020 Published: 02 February 2021 Reviewing editor: Stephen C Ekker, Mayo Clinic, United States This is an open-access article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose. The work is made available under the Creative Commons CC0 public domain dedication. Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 1 of 33 Research article Neuroscience eLife digest Traumatic brain injury can result from direct head concussions, rapid head movements, or a blast wave generated by an explosion. Traumatic brain injury often causes seizures in the short term and is a risk factor for certain dementias, including Alzheimer’s disease and chronic traumatic encephalopathy in the long term. A protein called Tau undergoes a series of chemical changes in these dementias that makes it accumulate, form toxic filaments and kill neurons. The toxic abnormal Tau proteins are initially found only in certain regions of the brain, but they spread as the disease progresses. Previous studies in Alzheimer’s disease and other diseases where Tau proteins are abnormal suggest that Tau can spread between neighboring neurons and this can be promoted by neuron activity. However, scientists do not know whether similar mechanisms are at work following traumatic brain injury. Given that seizures are very common following traumatic brain injury, could they be partly responsible for promoting dementia? To investigate this, researchers need animal models in which they can measure neural activity associated with traumatic brain injury and observe the spread of abnormal Tau proteins. Alyenbaawi et al. engineered zebrafish so that their Tau proteins would be fluorescent, making it possible to track the accumulation of aggregated Tau protein in the brain. Next, they invented a simple way to perform traumatic brain injury on zebrafish larvae by using a syringe to produce a pressure wave. After this procedure, many of the fish exhibited features consistent with progression towards dementia, and seizure-like behaviors. The results showed that post-traumatic seizures are linked to the spread of aggregates of abnormal Tau following traumatic brain injury. Alyenbaawi et al. also found that anticonvulsant drugs can lower the levels of abnormal Tau proteins in neurons, preventing cell death, and could potentially ameliorate dementias associated with traumatic brain injury. These drugs are already being used to prevent post-traumatic epilepsy, but more research is needed to confirm whether they reduce the risk or severity of Tau-related neurodegeneration. of perivascular neurons and glia in the depths of cortical sulci. Later in CTE, Tau pathology is wide- (Hay et al., 2016; spread and incorporates cortical and subcortical gray-matter areas Johnson et al., 2012; McKee et al., 2015). This broad spreading of Tau pathology in CTE can also be observed following TBI ascribed to single trauma events (Johnson et al., 2012). This spreading of Tauopathy is consistent with a prion-like mechanism; indeed, brain homogenates from mice sub- jected to TBI can initiate p-Tau pathology when injected into healthy wild-type mice (Zanier et al., 2018). The recipient mice develop a p-Tau pathology similar to single severe TBI patients, which then spreads from injection sites to distant regions, behaving similarly to bona fide prions (Zanier et al., 2018). Beyond TBI, the self-propagation and prion-like spread of Tau aggregates is thought to play a key role in the progression of other tauopathies such as AD (Iba et al., 2013; Iba et al., 2015; Mudher et al., 2017; Narasimhan et al., 2017; Sanders et al., 2014). Mechanisms of Tau spread- ing, and the therapeutic targets they offer, have principally been defined in vitro and include tunnel- ling nanotubes and extracellular vesicles (EVs such as exosomes and synaptic vesicles) and their uptake via endocytosis (Colin et al., 2020; Demaegd et al., 2018; Evans et al., 2018). In AD and other tauopathies, observations from patients and mice have highlighted the capacity of Tau seeds to spread trans-synaptically (Goedert et al., 1989; Pickett et al., 2017). Moreover, it has been shown that neuronal activity serves an important role in the spread of Tau pathology and general proteostasis (Pickett et al., 2017; Wu et al., 2016; Yamada et al., 2014). Stimulation of neuronal activity increased the extracellular release of Tau to the media in vitro and enhanced Tau pathology in a mouse model of familial frontotemporal dementia (Pickett et al., 2017; Wu et al., 2016; Yamada et al., 2014). Whether similar mechanisms of Tau release and spread occur following TBI remains unknown. In this light, an intriguing aspect of TBI is the prominence of post-traumatic seizures that might be predicted to initiate the aggregation and/or exacerbate the spread of Tau pathology. Seizures are one of the key consequences of all types of TBI, and they have been more commonly reported in patients who suffered from blast injuries (Asikainen et al., 1999; Salinsky et al., 2015). Although Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 2 of 33 Research article Neuroscience the exact prevalence remains undetermined (Lucke-Wold et al., 2015), it is anticipated that over 50% of TBI patients with severe injuries develop seizures or post-traumatic epilepsy (Kovacs et al., 2014). A link between seizures and Tau pathology is suggested by increased prevalence of seizures in AD patients and animal models of AD (Sa´nchez et al., 2018; Yan et al., 2012). Whether reducing post-traumatic seizures can delay or minimize the progression of tauopathy has yet to be fully explored. This knowledge gap is due in part to a lack of accessible in vivo models that can report the pro- gression and spread of tauopathy, or that allow neural activity associated with TBI to be measured and manipulated. To address these issues, we engineered a tauopathy biosensor transgenic zebra- fish that develops GFP+ puncta when Tau aggregates within the brain or spinal cord. Additionally, we introduce a simple medium-throughput method to induce TBI in zebrafish larvae. Combining these novel approaches, we found that post-traumatic seizures correlate strongly with spreading tau pathology following TBI. Manipulating this seizure activity mitigated Tau aggregation and revealed a critical role for endocytosis in the prion-like spread of Tau seeds in vivo following TBI. The results implicate seizures and dynamin-dependent endocytosis in the from our novel spread of Tau seeds, thereby offering potential therapeutic targets. in vivo TBI model Results transgenic Engineering and validation of tauopathy reporter lines Previous reports describe the assessment and quantification of Tau inclusions in living cells (typically Human Embryonic Kidney cells), via measuring aggregation of fluorescent proteins fused to Tau pro- tein, providing sensitive detection of pathological Tau species and strain variants (Kaufman et al., 2016; Sanders et al., 2014; Woerman et al., 2016). Tau is predominantly expressed in the neurons of the CNS, and we reasoned that fluorescent biosensor tools would have good potential to reveal additional phenotypes when expressed in these cells and moreover, that prion-like mechanisms of tauopathy spread are best modeled in an intact brain (e.g. vectored by blood and glymphatic circu- lation, ventricles, axonal projections, and immune systems). Therefore, we engineered a tauopathy biosensor zebrafish that expresses a fluorescent Tau reporter protein. Our genetically encoded fluorescent reporter protein was composed of the sequence of the human Tau core-repeat domain fused to green fluorescent protein (GFP) with a linker sequence and is referred to here as Tau4R-GFP (Figure 1A and its Figure 1—figure supplement 1A). Contrasting previous in vitro models, our biosensor did not feature any pro-aggregation mutations in the human Tau repeats; this design was intended to minimize spontaneous aggregation events. The expression of the biosensor protein in zebrafish was under the control of the pan-neuronal promoter neuronal eno- lase 2 (eno2, see Bai et al., 2007), which drives expression throughout the CNS (Figure 1B and its Figure 1—figure supplement 1C). We deployed the transgene in a transparent zebrafish line (the ‘Casper’ background [White et al., 2008]) to facilitate analysis beyond the early larval development stages (when pigmentation would otherwise begin to obscure microscopy). We isolated a stable transgenic (Tg) line that expresses the Tau4R-GFP biosensor reporter robustly and clearly in the CNS (Figure 1B), Tg(eno2:Hsa.MAPT_Q244-E372(cid:0)EGFP)ua3171, and assigned it allele number ua3171. Simultaneously, we expressed the same biosensor in vitro to validate the construct we deployed in vivo (Figure 1—figure supplement 1A and B). Both in HEK293T cells and Tg zebrafish, immuno- blotting using anti-GFP antibody detected our Tau-4R-GFP reporter protein at the expected size of ~45 Kd, similar to a SOD1:GFP biosensor protein of similar predicted size, and an appropriately larger size relative to GFP protein alone (Figure 1C and its Figure 1—figure supplement 1B’). We assessed the capacity of our Tau4R-GFP biosensor to report the presence of Tau pathology via transducing brain homogenates into cells. Brain homogenates burdened with tauopathy, from transgenic mice expressing mutant human Tau (Tg TauP301L), were compared to normal non-Tg mouse homogenates as a negative control. Congruent with findings obtained in past similar cell assays (Sanders et al., 2014), GFP-positive (GFP+) inclusions were detected only when cells were transduced with brain homogenate containing pathogenic human Tau fibrils (from Tg TauP310L mice) (Figure 1—figure supplement 1D). The in vitro assay detection rate was approximately 5% of cells having GFP+ inclusions in total, with 2% of cells forming multiple nuclear puncta and ~3% forming one cytoplasmic inclusion, whereas various negative controls consistently displayed 0% of cells with Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 3 of 33 Research article Neuroscience Figure 1. Validating tauopathy fluorescent biosensor in vitro and in zebrafish. The biosensor Tau4R-GFP was validated for its ability to detect tauopathy seeds in vitro and in zebrafish. (A) Schematic of Tau4R-GFP ‘Tau biosensor’ that contains the four binding repeats (4R) region of wild-type human Tau linked to green fluorescent protein (GFP; see also Figure 1—figure supplement 1A). (B) Transgenic zebrafish engineered to express Tau4R- Figure 1 continued on next page Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 4 of 33 Research article Neuroscience Figure 1 continued GFP biosensor throughout neurons of the CNS. Wild-type GFP is also abundant in the heart, which serves as a marker of the transgene being present but is otherwise irrelevant to our analyses. Scale bar ffi1 mm. (C) Western blot on zebrafish brain confirmed production of Tau4R-GFP at the expected size, similar to a SOD1-GFP biosensor and coordinately larger than GFP alone. (D) Human Tau fibril precipitated from transgenic (Tg TauP301L) mouse brain homogenates using PTA and assessed by EM. (E) Application of PTA-purified brain homogenate induced the formation of Tau inclusions similar to clarified brain homogenate (scale bar 50 mm; compare to Figure 1— figure supplement 1D), but application of equivalent preparations from non-Tg mice produced no GFP+ inclusions. (F–I) Tau biosensor zebrafish detects disease-associated human Tau fibrils following intraventricular injection of brain homogenate. Crude brain homogenates were microinjected into the hindbrain ventricle of Tau4R-GFP zebrafish larvae at 2 days post-fertilization, and Tau inclusions were analyzed at several time points. (F) Tau biosensor zebrafish larvae developed readily apparent GFP+ inclusions in the brain and spinal cord (Figure 1—figure supplement 2) when injected with brain homogenate burdened with Tau pathology (from Tg mice) but not from healthy brain homogenate (F’, from non-Tg mice). F’ inset shows many adjacent cells exhibiting GFP+ Tau aggregates. (G) Tau biosensor zebrafish injected with human Tau fibrils (within Tg mouse brain homogenate) developed significantly more aggregates on the spinal cord compared to uninjected control and other control groups, including compared to wildtype mouse brain homogenate (**p=0.0033 ***p=0.0006, ordinary two-way ANOVA and Tukey’s multiple comparison test). (H) Same data as in G, expressed as the percentage of larval fish showing Tau aggregates in the spinal cord, and (I) those same fish also showed Tau aggregates in the brain, over time. n = number of individual larvae. Images in E and F are 5 days post-application or post-injection, respectively. The online version of this article includes the following figure supplement(s) for figure 1: Figure supplement 1. Quantification GFP+ inclusions in HEK cells expressing Tau4R-GFP biosensor. Figure supplement 2. Quantification GFP+ inclusions in CNS of zebrafish expressing Tau4R-GFP biosensor, following injection of human Tau into the zebrafish hindbrain ventricle. Figure supplement 3. Movement of some Tau puncta over time following injection of zebrafish larvae with brain homogenate burdened with human tauopathy. inclusions (Figure 1—figure supplement 1E). To verify that Tau aggregates in the clarified brain homogenate caused the GFP+ puncta, we purified Tau aggregates from the tissue samples using PTA precipitations (Woerman et al., 2016). Tau fibrils purified from these preparations were charac- terized via EM analysis (Figure 1D). Transducing these preparations (in contrast to control prepara- tions derived from non-Tg mice) produced fluorescent puncta in the Tau4R-GFP reporter cells (Figure 1E), confirming the ability of our Tau4R-GFP chimeric protein to report Tau aggregation. Validation of in vivo Tau biosensor via intraventricular brain injections of Tau fibrils To test if the Tau4R-GFP biosensor can report the in vivo progression of tauopathy, we emulated intracerebral injection methods that induce (prion-like) Tau pathology in mice (Clavaguera et al., 2013; Guo et al., 2016; Peeraer et al., 2015). We injected clarified brain homogenate laden with human Tau fibrils, prepared as above from Tg mice, into the hindbrain ventricle of 2 days post-fertili- zation (dpf) Tau biosensor zebrafish (Figure 1—figure supplement 2A). The injected larvae and con- trol groups were monitored daily for up to 4 days post-injection (dpi). Biosensor larvae injected with human Tau fibrils (from Tg mouse brain) developed GFP+ puncta, reflective of Tau aggregation in the brain (Figure 1F,F”). These Tau inclusions were prominent near the ventricle wall as well as in sensory neurons along the spinal cord, when injected with brain homogenate from human-tau trans- genic mouse (Figure 1—figure supplement 2B). These puncta sometimes appeared to have either a lone dot-like shape or were similar to the multiple nuclear puncta detected in vitro, in which three to four small puncta are clustered together, or in other instances were more diffuse and concen- trated outside the nucleus (Figure 1F,F”). Repeated assessment of the location of Tau aggregates on the spinal cord of the same individuals over multiple days, using somite numbers as landmarks, suggested a movement of some of these puncta over time (Figure 1—figure supplement 3). The abundance of GFP+ spinal cord inclusions was progressive and significantly higher in larvae injected with pathogenic Tau brain homogenate compared with various controls (Figure 1G. p=0.0006 and p=0.0033 at 3dpi and 4dpi, respectively compared to injection of healthy wild-type control brain, ordinary two-way ANOVA and Tukey’s multiple comparison test). Few larvae in the Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 5 of 33 Research article Neuroscience control groups developed spontaneous inclusions but the number of the larvae and the abundance of those inclusions were minimal (Figure 1G). A total of 80% and 35% of the larvae injected with human Tau fibrils developed puncta in the brain and spinal cord, respectively (Figure 2H and I). On the other hand, a lower proportion of TAu biosensor larvae developed ‘spontaneous’ inclusions post-injection with the control brain homogenate from wild-type mice (Figure 2H and I). Tau aggre- gates were detected on the spinal cord region as early as 2 dpi. Intriguingly, a small percentage of larvae developed sporadic GFP+ Tau aggregates regardless of treatment. Visualizing the data as dis- tributions of larvae with particular abundances of GFP+ inclusions (Figure 1—figure supplement 2C) highlights a trend where most larvae did not develop aggregates unless they were injected with brain homogenate containing fibrillar, pathogenic human Tau species. In those cases, the biosensor larvae developed an abundant number of aggregates. Overall, these data confirm the ability of our biosensor model to detect pathogenic Tau species in vivo. Like other protein misfolding diseases, tauopathies reflect a proteostatic imbalance wherein the clearance of pathological Tau species is insufficient relative to accumulation (Chiti and Dobson, 2006; Lim and Yue, 2015). We reasoned that if the Tau biosensor larvae are faithfully reflecting Tau proteostasis concepts in vivo, then this could be revealed via inhibition of the proteasome with MG- 132. Larvae treated with MG-132 had GFP+ puncta in their brains (Figure 2A) at a rate approxi- mately double to the occurrence of spontaneous GFP+ inclusions (Figure 2B). Following injection of mouse brain homogenate containing human Tau fibrils, applying the proteasome inhibitor MG-132 substantially enhanced the percentage of larvae bearing Tau4R-GFP+ inclusions in the brain (to ~70%, Figure 2B), relative to equivalent larvae without MG-132 (~36%, Figure 2B). Figure 2. Protein-only induction of Tau puncta in vivo detected in biosensor zebrafish. Injections of synthetic Tau fibrils into Tau4R-GFP zebrafish induced GFP+ puncta in brains and spinal cord. (A,B) Inhibiting the proteosome with MG-132 enhanced the percentage of larvae bearing GFP+ inclusions in the brain following injection of tau-laden brain homogenate. (C) Synthetic human Tau proteins were fibrillized as confirmed via EM analysis. Human Tau fibrils were microinjected into the larval hindbrain at 2 days post-fertilization, and Tau inclusions were analyzed at 3 days post injections. (D) Tau aggregates were only observed after injection of Tau fibrils, not monomers (*p=0.0104, Kruskal Wallis test). (E) Tau aggregates (same data as D, presented as distribution of larvae that displayed various amounts of GFP+ puncta) appear only after injection of Tau fibrils, not monomers. Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 6 of 33 Research article Neuroscience It was striking that the zebrafish Tau biosensor was robustly able to discriminate brain homoge- nates that were burdened with human Tau aggregates versus those that were not. However, we con- sidered an alternative explanation for the data: the difference may not depend directly on human Tau in the brain homogenate but could instead reflect other bioactive components of the degener- ating Tg mouse brain. To verify that the formation of GFP+ puncta in zebrafish can be seeded by a protein-only injection, we delivered synthetic human Tau protein (2N4R). After confirming the recombinant Tau proteins were appropriately fibrillized via EM (Figure 2C), we delivered them by intraventricular injections as described above. Similar to previous data with brain homogenate, the larvae that were injected with synthetic Tau fibrils developed inclusions proximal to the brain ven- tricles as well as along the spinal cord at 3–6 dpi. The abundance of Tau aggregates along the spinal cord was significantly higher in larvae injected with the synthetic Tau fibrils compared to larvae injected with Tau monomers or to the non-injected group (p=0.0104) (Figure 2D). The distribution of larvae based on the number of Tau aggregates they accumulated also supported these findings (Figure 2E). In sum, the Tau4R biosensor deployed in the CNS of larval zebrafish was able to report Tau species, and further revealed the prion-like induction of tauopathy via protein-only seeding in vivo. Introduction of the first TBI model for larval zebrafish We next sought to deploy our Tau biosensor in a tauopathy model that enables higher throughput than can be achieved with intraventricular injection methods. We considered TBI) as an inducer of the tauopathy in CTE; further, we were encouraged that innovations in this realm could fill an unmet need for a high-throughput, genetically tractable in vivo model of these devastating concussive inju- ries. Although a few methods have been reported to induce TBI in adult zebrafish that are compara- ble to mammalian TBI methods (Maheras et al., 2018; McCutcheon et al., 2017), no such methods were available for zebrafish larvae (although McCutcheon et al., 2016 argue their application of exogenous glutamate may help address aspects of excitotoxicity associated with such insults). Here, we introduce and validate a simple and inexpensive method to induce TBI in zebrafish larvae. Investi- gating TBI in larvae offers substantial benefits regarding experimental throughput, economy, acces- sibility of drug and genetic interventions, and bioethics. We devised a traumatic injury paradigm by loading zebrafish larvae (~12 individuals in their typical E3 liquid growth media) into a syringe with a closed valve stopper, and applying a hit on the plunger to produce a pressure wave through the fish body akin to pressure or shock waves experienced dur- ing human blast injury (Nakagawa et al., 2011; Figure 3A). To challenge the method’s reproducibil- ity, and to permit manipulation of injury intensity, a series of defined masses were dropped on the syringe plunger. Technical variability, anticipated from larvae being in different orientations and positions within the syringe, was reduced by applying the injury three times to each group of larvae (except where noted otherwise) while repositioning the syringe between each injury. To assess if our method faithfully induced TBI similar to injury from pressure waves, we examined multiple markers known to be associated with blast-induced TBI, including cell death, hemorrhage, blood flow abnor- malities, and tauopathy (Bir et al., 2012; Kovacs et al., 2014; Nakagawa et al., 2011). Additionally, we evaluated the occurrence of post-traumatic seizure activity and increases in neuronal activity acutely associated with the trauma. We established the TBI method via empirical testing of various parameters, restricting ourselves to materials and methods that can be adopted inexpensively, with a goal of consistently inducing a robust injury (see phenotypes below) vs. a tradeoff with maximizing survival of the larvae. Subse- quent to this optimization, we were able to characterize the pressure induced within the syringe dur- ing each injury (Figure 3B–D). The maximum pressure induced was near 170 kPa (Figure 3B). The dynamics of the pressure change events during TBI (Figure 3B) imply that dropping the heavier weights led to the weight bouncing and producing a secondary increase in pressure (e.g. at ~175 or ~275 ms in Figure 3B). The maximal pressure induced varied from ~130 to ~175 kPa in an approx- imately linear fashion depending on the mass of the weight dropped (Figure 3D). The mean pres- sure change over the first 300 ms of the TBI also increased in a nearly linear fashion, and increased by nearly an order of magnitude when dropping weights of 30 g compared to 300 g (Figure 3C). We evaluated TBI-induced hemorrhage via the use of Tg(gata1a:DsRed) larvae that have red fluo- rescence in their blood cells (Traver et al., 2003). Hemorrhage was observed variably in larvae when a heavy weight (300 g) was used to induce the traumatic injury (Figure 3E). Further, approximately Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 7 of 33 Research article Neuroscience Figure 3. Zebrafish larvae subjected to traumatic brain injury (TBI) exhibited various biomarkers of TBI. (A) A novel TBI model for larval zebrafish: to induce blast injury, zebrafish larvae were loaded into a syringe with a stopper. A defined weight was dropped on the syringe plunger from a defined height, producing a pressure wave through the fish body akin to pressure waves experienced during human blast injury. (B) Dynamics of the pressure increase after dropping weights of varying masses in our TBI model. (C,D) The mean and maximum pressures generated, respectively, by various Figure 3 continued on next page Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 8 of 33 Research article Figure 3 continued Neuroscience weights applied in the TBI model. Dots represent individual trials. (E) Hemorrhage after TBI was observed in some of the larvae fish using Tg[gata1a: DsRed] transgenic zebrafish that express DsRed in erythrocytes, as indicated by white arrows. Lateral view of larval heads with anterior at the left. Scale bar ffi250 mm. (F) Increased cell death in the brain of 4 dpf larvae subjected to TBI as indicated by immunostaining of activated Caspase-3 (magenta). Positive and negative controls for immunostaining are in supplement. Nuclei were stained with DAPI in gray for reference. These are dorsal views of larval zebrafish brains with anterior at the left. Scale bar is 100 mm. (G) Seizure-like clonic shaking is observed in a subset of larvae after TBI. Movie frames are displayed from Video 2. These frames (left and right panels) are separated by ~400 ms in time, and are lateral views of the larval zebrafish trunk (akin to red box in G’). Control fish without TBI show little movement except obvious blood flow. Following TBI, larvae show bouts of calm (bottom left) interspersed (~400 ms later) with bouts of intense seizure-like convulsions (Stage III seizures; bottom right). (H) Larvae subjected to TBI also displayed Stage II seizures, that is weaker seizures that manifest as hypermotility and are detected using a previously optimized behavioral tracking software system – seizures are significantly more intense following TBI compared to the control group (***p=0.0013, paired t-test; dots are raw data for each larva, mean is plotted ± SE). The online version of this article includes the following figure supplement(s) for figure 3: Figure supplement 1. Traumatic brain injury (TBI)-induced cell death. half of the TBI larvae showed abnormalities in blood flow including a temporary reduction or com- plete absence of blood circulation (Video 1), consistent with abnormalities detected in rodent TBI models (Bir et al., 2012). Subsequently, we assessed apoptosis in the TBI larvae, observing that our TBI method induced cell death in larvae as detected by staining for active Caspase-3 (Figure 3F and its Figure 3—figure supplement 1). The number of active-Caspase-3-positive cells was negligible in the control groups compared to a mean of 62 apoptotic cells in TBI larvae (SEM ±9.17, n = 3) and 75 (SEM ±4, n = 2) in positive-control-larvae (cell death induced with camptothecin, CPT; Figure 3—fig- ure supplement 1). These data all align well with existing animal models of TBI with respect to mim- icking characteristic features of human TBI, and support the effectiveness of our method in inducing TBI in larval zebrafish. TBI-treated larvae exhibited post-traumatic seizure-like behavior and increased neuronal activity during trauma Post-traumatic seizures are one of the most frequent conditions associated with traumatic brain inju- ries and, despite being prevalent, remain poorly understood in TBI patients (Kovacs et al., 2014). Post-traumatic seizures were overtly apparent in a subset (approximately 40%) of zebrafish larvae after they were subjected to TBI. In some instances, the activity was highly reminiscent of Stage III seizures (defined previously in larval zebrafish as the most intense seizures; Liu and Baraban, 2019) with bouts of intense clonic convulsions and arrhythmic shaking (Video 2; exemplar frames from the movie are in Figure 3G). Other individuals exhibited hypermotility that is exactly consistent with past definitions of less intense Stage I or Stage II seizures. We quantified the latter seizure activity via behavioral tracking software (which we had previously optimized and validated for quantifying seizures in larval zebrafish [Kanyo et al., 2020a; Leighton et al., 2018]) and determined that larvae subjected to TBI exhibited seizure-like activity that was significantly higher than the control group (p=0.0013) (Figure 3H). Video 1. TBI-induced blood flow abnormalties. https://elifesciences.org/articles/58744#video1 Video 2. TBI-induced seizures and blood flow abnormalities. https://elifesciences.org/articles/58744#video2 Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 9 of 33 Research article Neuroscience Seizures are caused by abnormal and excessive neuronal excitability (Stafstrom and Carmant, 2015). To document bursts of neuronal activity during the brain trauma, if any, we utilized a geneti- cally encoded calcium imaging CaMPARI reporter (calcium modulated photoactivatable ratiometric integrator) expressed throughout the CNS. CaMPARI fluoresces green in baseline conditions, and permanently converts to red fluorescent emission if high intracellular calcium levels (i.e. neural activ- ity) occur coincident with application of ‘photoconverting’ intense 405 nm blue light. We subjected our allele of Tg(elavl3:CaMPARI)ua3144 larvae (Kanyo et al., 2020b) to TBI, coincident with brief application of photoconverting light (405 nm light provided by an LED array directed at the syringe, as described in Figure 4A). A sharp increase in neuronal activity during TBI was evident, especially in the hindbrain region as indicated by enhanced red emission (Figure 4B). CaMPARI allows robust quantification of neural activity expressed as a ratio of red:green fluorescent emission, which con- firmed that neuronal excitability increases significantly in response to brain trauma (Figure 4C and D). Notably, this combination of newly introduced methods of TBI being integrated with CaMPARI optogenetic methods (where the stable/irreversible changes from green to red fluorescent report- age allows a ratiometric quantification in a subsequent microscopy session) offers the rare ability to assess neural activity on un-restrained (free-swimming) subjects during TBI. In sum, our data reveal a substantial burst of neural activity occurs during TBI, and that zebrafish larvae exposed to TBI subse- quently exhibit a significantly higher propensity for spontaneous seizures. TBI on tau-biosensor-zebrafish-larvae induced GFP+ puncta After validating that our method was able to induce TBI upon zebrafish larvae, we next asked whether TBI induces Tau aggregates in our Tau biosensor model. Initially, we evaluated if our TBI method would induce aggregation of fluorescent proteins in models expressing GFP alone or other biosensor proteins such as SOD1-GFP (that is also designed to report prion-like protein aggrega- tion). Following TBI, and regardless of injury intensity, no GFP+ aggregates were detected in these controls (Figure 5—figure supplement 1). Similar results were obtained with other transgenic zebra- fish that express GFP in motor neurons (data not shown). Further, our Tau4R-GFP fish additionally express an unmodified GFP variant in the active heart muscle, and this robust GFP showed no sign of aggregation following TBI. Remarkably, in these same individual Tau4R-GFP larvae we detected Tau4R-GFP biosensor GFP+ puncta in both brains and spinal cords following TBI (Figure 5A–B). The abundance of GFP+ puncta increased with time following the injury (Figure 5C–D). To determine if the severity of tauopathy varies coordinately with severity of the traumatic injury, we assessed the impact of different masses. Although some variability is evident, a dose-response relationship is apparent such that the 65 g, 100 g, and 300 g weights induced more Tau aggregates compared to the control and 30 g weight (Figure 5—figure supplement 4). The heaviest weight (300 g) induced significantly more Tau aggregates versus the control group or the group with the 30 g weight (p<0.01 and p<0.001, respectively). Therefore, we decided to use both the 300 and 65 g weights for subsequent experiments. We evaluated whether dropping the light weight once or multiple times would affect the number of Tau aggregates on the spinal cord as well as dropping the weight once on 3 consecutive days, perhaps reminiscent of repetitive sports injury. We observed an increase in the abundance of Tau aggregates when the weight was dropped multiple times during 1 day, or over 3 consecutive days, but this increase was not statistically significant (Figure 5—figure supple- ment 4). The GFP+ Tau aggregates formed in the brain region following TBI tend to form fused shapes (Figure 5—figure supplement 4A) reminiscent of the spontaneous aggregates described above. The aggregates on the spinal cord, however, had similar shapes to aggregates detected post brain- injections, but with qualitatively less brightness in some instances. Multi-day monitoring of individual larvae (Figure 5—figure supplements 2 and 3) revealed variation in formation of Tau aggregates amongst individual TBI larvae. We monitored the abundance of Tau aggregates within individual fish over time following TBI and found that the average tauopathy significantly increased compared to the control group (p=0.0224 at 3dpti and *p=0.0312 at 4dpti (days post-traumatic injury), Figure 5C,D). Analysis of distribution of larvae binned into the number of Tau4R-GFP+ puncta at 3 dpti showed that more larvae developed Tau4R-GFP+ puncta compared to the control group (inset in Figure 5D). Considering that many of the larvae subjected to TBI formed Tau4R-GFP+ puncta in the brain that had a fused pattern (Figure 5—figure supplement 4A), we focused on Tau Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 10 of 33 Research article Neuroscience Figure 4. Neural activity increases during traumatic brain injury (TBI) as measured in CaMPARI zebrafish larva. (A) Schematic of TBI using CaMPARI (Calcium Modulated Photoactivatable Ratiometric Integrator) to optogenetically quantify neuronal excitability. Three dpf CaMPARI larvae were freely swimming while subjected to TBI, coincident with exposure to 405 nm photoconversion light. CaMPARI fluorescence permanently photoconverts from green to red emission only if the photoconversion light is applied while neurons are active (high intracellular [Ca2+]). The ratio of red:green emission is Figure 4 continued on next page Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 11 of 33 Research article Figure 4 continued Neuroscience stable such that it is quantifiable via subsequent microscopy. (B) Increased neural activity during TBI is represented by increased red:green emission (red pseudocolored to magenta) in the hindbrain of larvae (B”), compared to larvae not receiving TBI (B’) or fish not exposed to photoconverting light (‘no PC’ in panel B). These representative maximum intensity projection images show dorsal view of zebrafish brain (anterior at top), including merged, or red or green channels alone. (C) Heatmaps encode the CaMPARI signal (higher neural activity = higher red:green = hotter colors), highlighting location of increased neural activity during TBI relative to control larvae not receiving TBI. (D) Quantification of CaMPARI output in the hindbrain area reveals a significant increase in the neuronal excitability during TBI compared to control group not receiving TBI (**p=0.0087, Mann-Whitney test. Each data point is an individual larva). Scale bars = 100 mm. aggregates that formed on the spinal cord as their abundance could be most efficiently quantified compared to aggregates that formed in the brain. Post-traumatic seizure intensity influences tauopathy progression Considering the clinical prominence of post-traumatic seizures following TBI, and the suggested role of cell stress and increased neural activity in promoting protein misfolding diseases (Kovacs et al., 2014; Sa´nchez et al., 2018), we speculated that post-traumatic seizures might form a causal link between TBI and subsequent tauopathy. We first asked if a correlation exists between seizure inten- sity and extent of tauopathy. Following TBI, some larvae exhibited seizure-like movements, while some did not seem to move abnormally relative to untreated fish (Figure 3H). We sorted the larvae subjected to TBI into groups exhibiting the seizure-like behavior and those that displayed no overtly abnormal movement. Larvae exhibiting seizure-like behavior after TBI went on to develop abundant spinal cord aggregates (fivefold increase, p<0.001) in comparison to larvae that showed no seizure- like response to TBI (Figure 5E). To assess the hypothesis that seizure activity has a causal role in increasing the abundance of Tau aggregates in our TBI model, we employed convulsant and anti-convulsant drugs to modulate the seizure intensity and in vivo neural activity. We selected drugs that are well-established to behave similarly in zebrafish as in mammals, although it is perhaps notable that the multi-day drug applica- tion used here is longer than the acute applications typically considered in zebrafish (Ellis et al., 2012). Our hypothesis predicted that decreasing seizure-like activity following TBI would reduce tau- opathy. Indeed, applying the anti-convulsant drug Retigabine (RTG), that opens voltage-gated potassium channels (KCNQ, Kv7), resulted in a significant decrease in the abundance of GFP+ puncta (p=0.0107) with many TBI larvae not developing any Tau4R-GFP aggregates (Figure 5F). Similarly, intensifying post-traumatic seizures via application of the convulsant kainate increased the abundance on tauopathy fourfold (p<0.0001. Figure 5G) in a dose-dependent manner (Fig- ure 5—figure supplement 5). Kainate did not increase Tau4R-GFP+ puncta in the absence of TBI. Surprisingly, the convulsant 4-aminopyridine did not increase tauopathy (explored below). To assess if the impacts of kainate and retigabine on tauopathy were directly due to their modula- tion of post-traumatic seizures, we applied effective doses of each in concert. Co-application of kai- nate and retigabine following TBI produced an abundance of Tau4R-GFP+ puncta that was indistinguishable from larvae receiving TBI without pharmacology (Figure 5G). TBI-induced cell death was likewise correlated with the intensity of post-traumatic seizures. Co- application of kainate and retigabine following TBI increased or decreased, respectively, the abun- dance of cell death in a manner coordinate with the tauopathy (Figure 5H). Overall, convulsant and anti-convulsant drugs acted to increase and decrease TBI-induced tauop- athy, respectively. The drugs appear to be specific – their individual impacts on tauopathy and cell death are largely attributable to their epileptic and anti-epileptic modulation of post-traumatic seiz- ures, because when kainate and retigabine were applied concurrently they negated each others’ effects. Appearance of tauopathy following TBI requires endocytosis To further examine increased seizure activity after TBI, we applied 4-aminopyridine (4-AP), a Kv chan- nel blocker and convulsant drug. We predicted that raising the level of seizure activity would elevate tauopathy abundance in our TBI model, aligning with our observations following application of kai- nate (above). Surprisingly, higher doses of 4-AP consistently abrogated the appearance of Tau larvae with 200 or 800 mM of 4-AP for a prolonged period (38 hr, aggregates. Treating TBI Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 12 of 33 Research article Neuroscience Figure 5. Traumatic brain injury (TBI) induces tauopathy in larval zebrafish. (A) GFP+ Tau puncta are detected in the brain of Tau4R-GFP biosensor zebrafish at 5 days post-traumatic brain injury (dpti). A 300 g weight was used to induce TBI throughout this figure. (B) Tau aggregates formed on the spinal cord as a result of the TBI as shown by arrows. (C) Tauopathy significantly increases over time following TBI compared to control group (No TBI) Figure 5 continued on next page Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 13 of 33 Research article Neuroscience Figure 5 continued *p=0.0264, **p=0.007, two-way ANOVA with Tukey’s multiple comparison test. (D) The number of Tau aggregates in spinal cord significantly increases over time following TBI compared to control group (*p=0.0224 at 3dpti and *p=0.0312 at 4dpti). Inset: Tau4R-GFP zebrafish larvae subjected to TBI develop more GFP+ puncta compared to the control group by 3dpti (inset plot is similar to Figure 2E). Figure 5—figure supplements 2 and 3 plot this in individual fish. (E–H) Post-traumatic seizures link TBI to tauopathy. (E) Following TBI, larvae displaying post- traumatic seizures developed many more Tau aggregates relative to those not displaying post-traumatic seizures (**p=0.0011, Kruskal-Wallis ANOVA with Dunn’s multiple comparison test). (F) Inhibiting post-traumatic seizures with the anti-convulsant retigabine (RTG, 10 mM) significantly decreased the abundance of GFP+ puncta in the spinal cord (p=0.0107, Mann-Whitney test). (G) Increasing post-traumatic seizure using the convulsant kainate (KA, 100 mM; see dose-response in Figure 5—figure supplement 5) significantly increased the formation of Tau aggregation following TBI; this effect was prevented by co-treatment with anti-convulsant RTG. ****p<0.0001, ordinary one-way ANOVA with Tukey’s multiple comparison test. (H) Blunting post-traumatic seizures with RTG reduced TBI-related cell death. The main impact of RTG was specific to its anticonvulsant modulation of seizures because its effects were reversed by convulsant KA. Color scheme in panel C applies to other panels. n = number of zebrafish larvae. ****p<0.0001, ordinary one-way ANOVA with Tukey’s multiple comparison test. Scale bars = 500 mm. The online version of this article includes the following figure supplement(s) for figure 5: Figure supplement 1. Traumatic brain injury (TBI) did not induce GFP+ puncta in transgenic zebrafish larvae expressing SOD1-GFP. Figure supplement 2. Longitudinal analysis of individual fish after traumatic brain injury (TBI) shows various patterns of TAu inclusion formation and clearance in their brains. Figure supplement 3. Longitudinal analysis of individual fish after traumatic brain injury (TBI) shows various patterns of Tau inclusion formation and clearance in their spinal cords. Figure supplement 4. Increasing intensity of traumatic brain injury (TBI) significantly increased Tau4R-GFP puncta abundance, but only modest insignificant increases in Tau puncta were observed with an increasing number of successive brain injuries. Figure supplement 5. Intensifying seizures following traumatic brain injury (TBI) increased abundance of GFP+ Tau puncta in a dose-dependent manner. beginning 24 hr post-traumatic injury) significantly inhibited the abundance of Tau4R-GFP+ puncta in the TBI group (Figure 6A–B and its Figure 6—figure supplement 1A-B). Analysis of the distribu- tion of larvae linked to the number of Tau aggregates supported this finding with no zebrafish larvae developing aggregates in groups treated with 4-AP (Figure 6—figure supplement 1C). It is worth noting that 4-AP is commonly used in zebrafish models of epilepsy, but rarely used for prolonged treatment. To evaluate if the time at which treatments are administered plays a role in this unex- pected result, we treated larvae with 200 mM 4-AP at earlier time points, specifically during TBI and 1.5 hr later. We kept the duration of 4-AP treatment the same as previous experiments (38 hr). We found that administering 4-AP during different time windows relative to the TBI did not measurably alter the inhibitory action of 4-AP on the abundance of Tau aggregates (Figure 6—figure supple- ment 1D). A similar observation was made when the duration of the 4-AP treatment was reduced to 24 hr (Figure 6—figure supplement 1E). Next, we considered if this unexpected inhibition of tauopathy by high-dose 4-AP convulsant is a direct consequence of increased neural activity (e.g. perhaps via neural exhaustion). We found that larvae receiving TBI and 4-AP continued to exhibit a lack of Tau aggregates when co-treated with anti-convulsant retigabine (p<0.0001) (Figure 6B). This suggested that high doses of 4-AP block the formation of Tau aggregates via a mechanism independent of its convulsant activity. To resolve a mechanism whereby high doses of 4-AP reduced Tau pathology, contrary to our pre- dictions above regarding neural hyperactivity, we considered previous in vitro work that demon- strated high concentrations of 4-AP cause reduced endocytosis of synaptic vesicles (Cousin and Robinson, 2000). To examine if the inhibitory actions of 4-AP on the abundance of Tau aggregates in our TBI model is consistent with an endocytosis inhibition mechanism, we treated our tau biosen- sor larvae post-traumatic injury with Pyrimidyn-7 (P7), a potent dynamin inhibitor that is known to block endocytosis (McGeachie et al., 2013), and analyzed the propagation of Tau pathology by quantifying the number of Tau inclusions. Owing to the potency of P7 and its impact on the survival of larvae, we treated the larvae with it for 24 hr at 3 mM. Similar to the findings with 4-AP, P7 Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 14 of 33 Research article Neuroscience Figure 6. Tauopathy induced by traumatic brain injury (TBI) was attenuated by 4-aminopyridine (4-AP) via mechanisms independent of seizures. (A) Tau4R-GFP biosensor zebrafish larvae subjected to TBI and treated with the convulsant 4-AP show no brain puncta. Scale bar = 200 mm. (B) 4-AP significantly reduced (apparently eliminated) the abundance of GFP+ puncta in the brain and spinal cord compared to untreated TBI control. Results from alternative doses and timings of 4-AP are reported in Figure 6—figure supplement 1. The impact of 4-AP on tauopathy appears to be independent of its actions on post-traumatic seizures because reducing the latter with anti-convulsant retigabine (RTG) had no measurable effect. (C-F) Pharmacological inhibition of endocytosis reduced tauopathy following TBI. (C) Blocking endocytosis with Pyrimidyn-7 (P7) treatments significantly inhibited the formation of Tau4R-GFP+ puncta following TBI in zebrafish larvae (***p=0.001). (D) Dyngo 4a treatment significantly reduced Tau aggregates in the spinal cord (**p=0.0025) in a manner similar to P7. Figure 6 continued on next page Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 15 of 33 Research article Neuroscience Figure 6 continued (E) 4-AP treatment significantly inhibited the formation of Tau4R-GFP+ puncta in the spinal cord (***p=0.0003) of Tau biosensor line that also express human Tau (0N4R) after traumatic brain injury compared to untreated TBI control group. (F) A notable reduction in Tau aggregates was observed in the same line after treatment with P7 drug. Statistical analysis shows no significance difference (ns, p=0.2223) between groups. n = number of larvae. ****p<0.0001, Kruskal-Wallis ANOVA with Dunn’s multiple comparison test used throughout this Figure. The online version of this article includes the following figure supplement(s) for figure 6: Figure supplement 1. 4-Aminopuridine (4-AP) abrogates TBI-induced Tau aggregates when applied at various doses or for various times. treatments significantly inhibited the formation of Tau4R-GFP+ puncta in TBI larvae (p=0.001) (Figure 6C). We assessed further the role of endocytosis by employing another dynamin- inhibitor drug, Dyngo 4a, that is less potent than P7 (McCluskey et al., 2013). We obtained similar results in which Dyngo 4a treatments significantly reduced tauopathy in our TBI model (Figure 6D). To deter- mine if these results are applicable to human tau, we induced traumatic brain injury on double-trans- genic larvae expressing both human Tau (the 0N4R human Tau isoform, see Bai et al., 2007) and our Tau biosensor reporter, followed by treatment with either 4-AP or P7. Apart from the untreated control, both groups treated with 4-AP or P7 exhibited a noticeable reduction in abundance of Tau aggregates. While the decrease in the case of P7 was not statistically significant, statistical analysis showed significance after 4-AP treatments (p=0.0003) (Figure 6E and F). These findings confirmed the ability of 4-AP and dynamin inhibitors of reducing human Tau aggregates in our TBI larvae. Discussion The consequences of concussive blasts and TBI extend beyond the proximate injury – they are prom- inent risk factors for devastating dementias including AD, CTE and other tauopathies. Identifying the causal links that entwine TBI with subsequent tauopathies would inspire improved diagnostics and therapeutics. Investigating these mechanisms has been hampered by lack of tractable models, since cell culture platforms cannot faithfully represent the injury or the response-to-injury or the treatment thereof. Indeed, TBI and tauopathies are complex tissue and systems-level events with pathobiology progressing on a backdrop of dynamic vigorous neural function, prion-like vectoring of misfolded proteins via glymphatic and blood vasculature, immune and support cells, sleep physiology, homeo- static regulation and complex drug metabolism. Rats have been the favored animal model for TBI, and mice can complement this as insightful models of tauopathy, yet both are challenged by expense, ethical considerations, and CNS tissues that are relatively inaccessible to (longitudinal) visu- alization of cellular events in living individuals (Bodnar et al., 2019; Marklund, 2016; Meconi et al., 2018; Pham et al., 2019). Here, our zebrafish models replicate human TBI and tauopathy; the model is imperfect (e.g. see Limitations below) but by addressing many of these challenges in an accessible and vibrantly active vertebrate brain, we offer an innovative approach for the study of prion-like events, tauopathy and/or TBI. Here, we introduce a simple method for delivering TBI to larval zebrafish that can be scaled to high-throughput and adopted at low expense. The tractability and transparency of zebrafish larvae allowed us to deploy genetically encoded fluorescent reporters that were validated to (i) uniquely quantify neural activity on freely behaving animals during TBI and (ii) effectively document prion-like tauopathy in individual subjects over multiple days. The accessibility of this platform to pharmacol- ogy allowed us to query cell biology events in vivo and support a role for endocytosis in prion-like progression and TBI-induced tauopathy. Further, anti-convulsant drugs were potent mitigators of the tauopathy and cell death that emerged subsequent to TBI (schematized in Figure 7); these effects were attributable to suppression of post-traumatic seizures (as proven by the anticonvulsant’s therapeutic effects being reversed by co-application of convulsant drugs). It remains to be seen if these data have any bearing on the long-term clinical management of TBI patients, and we newly speculate that prophylactic application of anti-convulsants (already common for blunting of patient’s post-traumatic seizures) might hinder progression of tauopathies including CTE and AD. If true, then debates regarding the optimal regimen of anti-epileptics for TBI patients should consider their potential for providing long-term benefits on dementias. Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 16 of 33 Research article Neuroscience Figure 7. Graphical summary: anticonvulsants reverse the tauopathy and cell death exhibited by zebrafish larvae following traumatic brain injury (TBI). Tauopathy was reported via aggregation of a genetically encoded chimeric protein, Tau4R-GFP, that was expressed throughout the central nervous system. TBI led to seizures, and subsequently to Tau aggregation and cell death (akin to chronic traumatic encephalopathy, CTE). The tauopathy and cell death were ameliorated, producing healthy larvae, by blocking seizures with anti-convulsants; these effects were specific insomuch that they could be reversed by co-application of convulsants. Prion-like tauopathy induced by TBI It is well established that TBI induces tauopathy, and over the past decade, numerous in vitro and in vivo studies have supported that Tau proteins possess prion-like properties. Indeed seeding, tem- plated misfolding (conversion) and spread to synaptically connected regions has been documented in tauopathies such as AD and FTD (Ayers et al., 2018; de Calignon et al., 2012; Goedert et al., 2017a; Goedert et al., 2017b; Iba et al., 2015; Woerman et al., 2016). Various mechanisms have been proposed for the transcellular transfer of Tau seeds including release mechanisms via exo- somes, or cellular uptake mechanisms via endocytosis (Demaegd et al., 2018; Evans et al., 2018; Wang et al., 2017; Wu et al., 2013). Nonetheless, these suggested mechanisms were postulated based on in vitro evidence as there is a lack of appropriate models that can visualize and manipulate the prion-like spread of Tau pathology between tissues in a vibrant brain. It is unclear if these mecha- nisms are universal in progression of all tauopathies or if there are factors and mechanisms that are unique to each disease. Our zebrafish models allow us to study the progression and spread of TBI- induced tauopathy longitudinally in living animals, an experimental advantage that is unmatched among TBI animal models. Regarding TBI, a large knowledge gap exists regarding how Tau seeds are released and/or inter- nalized by adjacent (or far-flung) cells - indeed the prion-like properties of Tau species following TBI had not been assessed until very recently (Woerman et al., 2016; Zanier et al., 2018). Moreover, the focus in the literature has mostly been directed toward repetitive mild trauma as it is most asso- ciated with CTE, yet the various forms of TBI all are considered risk factors for neurodegeneration. The recent revelation that most TBI patients, whether they suffered from single or repetitive brain Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 17 of 33 Research article Neuroscience trauma, all exhibited Tau pathology similar to CTE (Washington et al., 2016; Zanier et al., 2018) suggests all forms of TBI might incorporate tauopathies. Our mode of TBI on larval zebrafish entails a pressure wave that most closely mimics a blast injury (e.g. as experienced by military personnel or civilians near an explosion, and presumably producing injury throughout the body), but the etiology leading to tauopathy probably has many similarities regardless of the mode of the initiating TBI. Post-traumatic seizures accelerate tauopathy following TBI We inspected the role of seizure activity and/or neuronal excitability, as well as the role of dynamin- dependent endocytosis, during the progression of tauopathy after TBI. We focussed our attention on seizure activity in part because seizures frequently occur in TBI patients following blast traumatic injury (Englander et al., 2014; Kovacs et al., 2014). We hypothesized that neuronal excitability and seizure activity after TBI can play a role in accelerating the wide dissemination of Tau pathology. As such, we introduced two new approaches to test this hypothesis. The first approach was to engineer a novel in vivo Tau biosensor model in zebrafish that can visualize pathological Tau spreading and accumulation within the intact and vibrant CNS. The Tau biosensor zebrafish express human tau4R- GFP reporter protein, and we confirmed its ability to detect Tau seeds from various sources both in vivo and in vitro, similar to previously engineered in vitro models (Kaufman et al., 2016; Sanders et al., 2014). Our second approach was to introduce and optimize an elegantly simple technique to cause pressure-wave induced TBI, similar to human blast TBI, in zebrafish larvae. We endeavoured to inflict injury on larval zebrafish rather than adults because of the synergistic advan- tages that larval zebrafish provide: these include economical access to large numbers of individuals and associated statistical power, and the tractability of larvae for high-throughput in vivo screening of therapeutic agents (Saleem and Kannan, 2018). Larval zebrafish provide a large economic advan- tage compared to adults, with respect to time, cost per individual and space consumed in animal housing. Moreover, injuring animals in larval stages is viewed as an ethically favorable Replacement [sensu ‘the three Rs’ of Russell and Burch, 1959] compared to injuring adult subjects. Thus, regard- less of any bioethical considerations based on taxonomy, larval fish (that are accessible early in their development via external fertilization of eggs) are ethically advantageous to rodents (that are acces- sible for TBI only at postnatal stages) when considering highly invasive procedures like TBI. The latter conclusion relies on the assumption that the knowledge gained is of value, that is relevant to appre- ciating disease etiology. Our data argue that our TBI methods are germane to clinical aetiology, because (akin to existing animal models of TBI) we were able to confirm the presence of various markers associated with brain injury, such as cell death, abnormalities in blood flow, hemorrhage and the occurrence of post-trau- matic seizures. Recently, post-traumatic seizures were also observed in adult zebrafish when TBI was delivered via focused ultrasound (Cho et al., 2020) offering an additional opportunity for zebrafish to be an important tool and bridge organism in the field of neurotrauma. The post-traumatic seizures apparent in our TBI model led us to consider the neural events occur- ring during the TBI, and their potential bearing on the correlation between neural activity and tauop- athy. Few studies examine how TBI impacts neuronal circuits, especially in vivo, and these typically consider events several hours or days after brain trauma (Bugay et al., 2020). This may be of impor- tance when considering evaluating the reasons behind the developments of post-traumatic seizures and epilepsy. In a controlled cortical impact model of TBI, an initial decrease or loss in neuronal activity is recorded after injury before a rise in neuronal activity is noted (Ping and Jin, 2016). Whether this occurs in different types of TBI, like blast TBI, was unexamined. To address this, we performed TBI on larval zebrafish expressing CaMPARI, a genetically encoded optogenetic reporter of neural activity. CaMPARI is particularly ideal for this question, as its reportage of neural activity (a stable and quantifiable shift from green to red fluorescence) occurs only during user-defined times and that reportage is relatively permanent. This allowed us to quantify the CNS activity that had occurred during TBI, by characterizing the ratio of red:green fluorescent emission using confocal microscopy after the TBI injury was completed. This approach therefor allows relatively easy access to quantifying neural activity during injury in an unencumbered freely swimming animal. Here, we revealed for the first time a snapshot of neurons becoming active at the moment of TBI. Our results demonstrated an increase in neuronal excitability upon TBI, which may contribute to the frequency of post-traumatic seizures observed in our model, other blast TBI models and TBI patients (Bugay et al., 2020). The increases in neuronal activity were especially prominent in the hindbrain; Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 18 of 33 Research article Neuroscience this may be due to this region being susceptible to injury in our paradigm, and/or represent an out- put that is related to increased movement (e.g. this region is home to robust motoneurons that would be hyperactive during seizures). It will be of interest to resolve what types of neurons and neu- rochemistry are most impacted by the injury. Notably, the subsequent Tau4R-GFP aggregates were also most prominent in this hindbrain region, and while this could be coincidental, or perhaps an artefact of a hidden differential Tau4R-GFP abundance, it is also possibly a result of increased neural activity. Regarding the etiology of tauopathy subsequent to TBI, the CaMPARI quantification pro- vided us important validation that neural activity was substantively impacted by TBI, complementing the evidence of increased seizure-like movements. This supported our rationale that convulsant and anti-convulsant drug treatments might modulate neural activity and thereby accelerate or decelerate tauopathy accumulation. Indeed, we had noted the occurrence of post-traumatic seizures in most of our TBI samples, which is in agreement with the prevalence of seizures in blast TBI patients and TBI rodent models (Bugay et al., 2020; Kovacs et al., 2014). However, whether post-traumatic seizures contribute to prion-like spreading of Tau pathology (observed after TBI or not) was unknown. Beyond TBI, several investigations have supported an association between Tau pathology and seizures (Sa´nchez et al., 2018; Tai et al., 2016). Studies on epileptic human temporal structure revealed accumulation of Tau aggregates (Sa´nchez et al., 2018). In 3XTg AD mice, induced chronic epilepsy was associated with changes of inter-neuronal p-Tau expression (Yan et al., 2012). Additionally, data obtained from postmortem analysis of patient tissues with AD and drug resistant epilepsy uncovered a correlation between symptomatic seizures, increased Braak staging and accelerated Tau accumulation (Thom et al., 2011). Interestingly, the presence of tau deposits in epileptic patients and the similarity of its pathology to CTE suggest a conceivable role for seizures influencing the progression of Tau pathology in a similar manner to TBI (Puvenna et al., 2016). Indeed, our data from the application of the convulsant kainate here support the role of post-traumatic seizure in enhancing Tau abun- dance and cell death in our TBI model (Figure 5G,H). This finding is in line with observations in a patient with epilepsy and a history of head injury, in which progressive Tau pathology was noted (Geddes et al., 1999; Thom et al., 2011). Intriguingly, reducing seizure activity after TBI via anti- convulsant drugs was able to significantly reduce tauopathy and cell death, providing further evi- dence of the relationship between seizures and tauopathy in TBI (Figure 5F,G and H). The mecha- nism of drug action appears to be dominated by its anticonvulsant properties, because its effects were reversed by co-application of convulsants. Thus, anti-convulsants are intriguing as a route to slowing progression of tauopathy following TBI, and it is encouraging that they are already com- monly deployed to prevent post-traumatic seizures. Endocytosis mediates prion-like spread of tauopathy One particular convulsant drug, 4-AP, inhibited tauopathy in our TBI model (Figure 6A and B), con- trary to our hypothesis that seizure intensity is positively correlated with tauopathy following TBI. 4- AP is a voltage-gated potassium channel blocker that enhances neuronal firing activity and has been used often in zebrafish seizure studies (Kanyo et al., 2020b; Kasatkina, 2016; Liu and Baraban, 2019; Lundh, 1978; Winter et al., 2017). Yet, 4-AP is rarely administered for prolonged treatments such as those we deployed here, for example past studies rarely exceed one hour of 4-AP (Winter et al., 2017). Thus, we considered that our high dose and prolonged stimulation with 4-AP may have led to off-target effects; we confirmed this insomuch that the 4-AP’s inhibition of tauop- athy was not related to its convulsant properties (as determined by 4-AP’s effects being unaltered by potent anti-convulsants (Figure 6B)). Indeed, previous in vitro work revealed high concentrations or prolonged stimulation with 4-AP has off-target effects via inhibiting dynamin, which is important for the endocytosis of synaptic vesicles at the nerve terminals (Cousin and Robinson, 2000). The inhibition of endocytosis observed in that study was independent of 4-AP-dependent seizure activity. We further queried the potential role of dynamin-dependent endocytosis in the prion-like pro- gression of Tau pathology after TBI by applying endocytosis inhibitors that target dynamin. Dynamin is a GTPase involved in two mechanisms of endocytosis that are important for synaptic vesicle trans- port (Singh et al., 2017). Empirical work on human stem-cell-derived neurons has indicated that Tau aggregates are internalized via dynamin-dependent endocytosis and that blocking other endocytosis pathways independent of dynamin, such as bulk endocytosis and macropinocytosis, did not disrupt Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 19 of 33 Research article Neuroscience Tau uptake (Evans et al., 2018). On the contrary, inhibiting dynamin significantly decreased the internalization of Tau aggregates. Our results are in line with the previously mentioned findings that show Tau progression in TBI models depends on dynamin-dependent endocytic pathways - blocking them with two different inhibitors (and with 4-AP) dramatically lessened the abundance of Tau seeds (Figure 6C–F). Hence, our findings not only provide in vivo validation of past in vitro works, but also suggest mechanisms underlying prion-like spreading of Tau seeds in TBI and CTE that could aid in developing therapeutic strategies. Limitations of our approach Our tauopathy biosensor, human Tau4R-GFP, was deployed in vivo and uniquely able to detect sig- nificant increases (and decreases) in the abundance of Tau aggregates following various insults and treatments, typically in a dose-dependent manner and in harmony with expected trends. Consider- ing this success, it remains perplexing and intriguing that a subset of larvae exhibit GFP+ Tau puncta despite receiving no known tauopathy-inducing insults. This suggests that the larvae express the transgene at a level near to a threshold for producing spontaneous aggregates. We performed selective breeding to minimize these occurrences and tentatively believe, after too few generations, that genetics of the fish is a factor – substantial genetic variation exists in zebrafish in-bred lines (Balik-Meisner et al., 2018; Guryev et al., 2006). However, we acknowledge the variation could be a minor technical artefact rather than biological. It remains to be determined if our biosensor can self-aggregate in some situations, or if these aggregates represent transient moments of imbalance in the proteostatic cycle of Tau aggregation and clearance. Regardless of this variation amongst indi- vidual larvae, the effects of neurotrauma and drug interventions were rational and robust when con- sidering the mean of several animals; the large sample sizes available with the larval zebrafish model can overcome some of this variability. More optimistically, this inter-individual variation and stochas- tic appearance of tauopathy, in a high-throughput model, could be leveraged to newly appreciate aspects of spontaneous AD or other non-familial tauopathies. Regardless, future work will also need to characterize the biochemistry and biophysics of the human Tau inclusions in zebrafish compared to patients or rodent models. Regarding our TBI methods, further refinements may yet be able to improve consistency of the injury and reduce the apparent variability between individuals. This variability is real, but somewhat offset by the large sample sizes attainable: our TBI methods offer the potent advantages of zebrafish larvae with respect to genetic and drug accessibility in high-throughput formats, while also retaining the critical in vivo complexity required to investigate disease etiology and treatments. Further, it remains to be established if the mechanisms we reveal are ubiquitous across the various forms of TBI: our model fills a gap by supplying a rare ‘closed head’ TBI model (as opposed to the majority of animal models that access the brain by removing skull elements prior to brain injury, see exceptions by Meconi et al., 2018; Mychasiuk et al., 2014). Our model might be most relevant to brain trauma experienced by the human fetus (e.g. during car collisions or domestic abuse), considering the devel- opmental stage and aqueous media. Moreover, the injury induced herein is presumably not limited to the brain, and probably injures various tissues including the spinal cord; this should be considered when interpreting the data. Further work is also needed to appreciate how the physics of our blast injury is altered by occurring at a small scale (e.g. larval brain is <500 mm). At this point, we are left to assume that the cellular and physiological aspects of TBI we consider here are sufficiently similar across all classes of TBI, and thus the knowledge gleaned may be variably applicable. Finally, we have chosen to restrict our analysis to study of larval fish. While this offers many logisti- cal and ethical advantages detailed above, it limits our study to acute effects occurring over the course of several days. Conclusions from such work, once refined and validated using the power of the in vivo zebrafish larva model, should be tested in rodent models where it is equally time-consum- ing to assess the long-term efficacy of treatments on these progressive late-onset dementias. Conclusion Currently, no available treatments are applicable to all tauopathies, which remain as devastating and inevitably fatal dementias. Zebrafish larvae, fostered by appropriate innovations, now offer a potent complement both to rodent models of TBI and to cellular models of tauopathy. Our engineered fish allowed us to reveal post-traumatic seizures as a druggable mechanistic link between TBI and the Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 20 of 33 Research article Neuroscience prion-like progression of tauopathy. Intriguingly, our conclusions have potential for translation to TBI clinics where anti-convulsants are already in use as prophylactics for post-traumatic epilepsy, though further work remains to address if they mitigate (the risk or severity of) later progression of CTE, AD or other tauopathies. Key resources table Materials and methods Reagent type (species) or resource Strain (zebrafish) Strain (zebrafish) Strain (zebrafish) Strain (Zebrafish) Genetic reagents (zebrafish) Designation Tg(eno2:hsa. MAPT-ires-egfp)Pt406 Tg(eno2: SOD1-GFP) ua3181 Tg[elavl3: CaMPARI (W391F+V398L)]ua1344 Tg(eno2:Hsa. MAPT_Q244- E372(cid:0)EGFP)ua3171 Multisite Gateway technology (BP Clonase II Enzyme mix and LR Clonase II Plus enzyme) Cell line (Homo-sapiens) HEK293T Sequence- based reagent Biological sample (mouse) Antibody Antibody Antibody Antibody GFP_R Whole brains Anti GFP (rabbit monoclonal) Anti-b-actin (rabbit polyclonal) Anti-Active- Caspase-3 (rabbit polycolonal) Alexa Fluor 647 (chicken anti- rat IgG) Continued on next page Source or reference Burton’s Lab (Bai et al., 2007) This paper Identifiers ZFIN ID: ZDB-ALT- 080122–6 N/A In house allele (Kanyo et al., 2020b) established using vector provided by Eric Schreiter’s lab This paper N/A Additional information Zebrafish that express human four repeat TAU zebrafish biosensor engineered to detect human SOD1 aggregation zebrafish expressing the calcium sensor CaMPARI Zebrafish biosensor engineered to detect human Tau aggregation Guo and Lee, 2011; Kwan et al., 2007 Thermo Fisher ATCC Provided by Dr. David Westaway’s laboratory This paper Tissues were provided by Dr. David Westaway (Eskandari-Sedighi et al., 2017; Murakami et al., 2006) Abcam Sigma-Aldrich BD Pharmingen Cat# 11789020 Cat# 12538120 ZFIN ID: ZDB-PUB-170809–10 Cat# CRL- 3216, RRID:CVCL_0063 PCR primer TCTCGTTGG GGTCTTTGCTC Isolated from wild -type mice with 129/SvEvTac genetic background, and TgTauP301L mice Cat# ab183734, RRID:AB_2732027 WB(1:3000) Cat# A2066, RRID:AB_476693 Cat# 559565, RRID:AB_397274 WB (1:10000) IHC (1:500) Invitrogen Cat# A-21472, RRID:AB_2535875 WB (1:500) Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 21 of 33 Research article Neuroscience Continued Reagent type (species) or resource Peptide, recombinant protein Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Chemical compound, drug Software, algorithm Software, algorithm Other Other Other Other Other Designation Human MAPT (2N4R) Kanic acid monohydrate Retigabine 4-Aminopyridine (4AP) K+ channel blocker Pyrimidyn-7 (P7) Dynamin inhibitor Dyngo 4a Lab Chart 7 (software) Geneious Prime (bioinformatics software) Human MAPT (Gene block) Power lab Data acquisition device (equipment) Piezoresistive pressure transducer (equipment) Lipofectamine. 2000 (transfection reagents) DAPI stain Source or reference rPeptide Identifiers Cat# T-1001–2 Additional information Resuspended to 2 mg/ml before use Sigma Aldrich K0250 Toronto research chemicals Sigma Abcam Abcam R189050 Cat# 275875–1G Cat# ab144501 Cat# ab120689 AD Instruments geneious.com Version 8 Ordered from IDT AD Instruments 2/26 AD Instruments Cat# MLT844 Invitrogen Cat# 11668–019 50 mM concentration supplied in DMSO This is aa 244–372 of the full-length TAU 2N4R with seven- amino acid C-terminal linker (RSIAGPA) Thermo Fisher D1306 (1 mg/mL) Animal ethics and zebrafish husbandry Zebrafish were raised and maintained following protocol AUP00000077 approved by the Animal Care and Use Committee: Biosciences at the University of Alberta, operating under the guidelines of the Canadian Council of Animal Care. The fish were raised and maintained within the University of Alberta fish facility under a 14/10 light/dark cycle at 28˚C as previously described (Westerfield, 2000). Generating transgenic tauopathy reporter zebrafish To engineer the transgenic Tau4R-GFP reporter zebrafish, the human wild-type MAPT sequence of the four-microtubule binding repeat domain (aa 244–372 of the full-length TAU 2N4R, NCBI linker (RSIAGPA) was NC_000017.11, protein id NP_005901) with a seven-amino acid C-terminal ordered as a gene block from IDT. The gene block was subcloned into a middle entry cloning vector (Multisite Gateway technology, ThermoFisher). This was recombined with the p5E-enolase2 and Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 22 of 33 Research article Neuroscience p3E-GFP components into destination vector pDestTol2CG2 of the Tol2kit (Guo and Lee, 2011; Kwan et al., 2007). The destination vector contained a reporter construct [encompassing EGFP driven by the cardiac myosin light chain (clmc) promoter that helps identify stable transgenic zebra- fish]. The resulting plasmid pDestTol2CG2.eno2:Tau4R-GFP was delivered in a 10 ml injection solu- tion including 750 ng/ml of the construct mixed with 250 ng/ml Tol2 transposase mRNA, 1 ml of 0.1M KCL, and 20% phenol red. The solution was injected into the single-cell embryos of Casper zebrafish line (transparent zebrafish line)(White et al., 2008). Injected embryos were screened for mosaic expression of the Tau4R-GFP transgene at 2 days post-fertilization (dpf) using a Leica M165 FC dis- secting microscope. F0 mosaic fish were raised to adulthood and outcrossed. Successful F1 embryos were identified by their abundant expression of Tau4R-GFP in the CNS and the green heart marker. The stable transgenic line Tg(eno2:Hsa.MAPT_Q244-E372(cid:0)EGFP)ua3171 was assigned the allele num- ber ua3171. An equivalent transgenic zebrafish biosensor was engineered to detect human SOD1 aggrega- tion. Subcloning from existing vectors (Pokrishevsky et al., 2018) produced pDestTol2CG2.eno2: SOD1-GFP and similar transgenesis methods engineered the Tg[eno2:SOD1-GFP] zebrafish line that was assigned allele number ua3181. Cell culture and generation of tauopathy reporter stable cell line To move the Tau4R-GFP reporter above into a vector appropriate for cell culture, BamHI and Xhol nuclease restriction enzymes were employed to remove the Tau4R-GFP fragment from pDest tol2CG2.eno2.Tau4R-GFP.pA. The Tau4R-GFP fragment was subcloned into the pCDNA3.1 vector using a T4 DNA ligase enzyme. Sequencing of the cloned vector with the following reverse primer for GFP (TCTCGTTGGGGTCTTTGCTC) confirmed the proper orientation. Purification of the plasmid was conducted with the Qiagen purification kit. HEK293T cells were grown in Dulbecco’s modified Eagle’s medium (GibcoTM, ThermoFisher) supplemented with 10% fetal bovine serum and 1% peni- cillin/streptomycin. All cells were maintained at 37˚C in a humidified 5% CO2 incubator. For passag- ing cells, cells were washed with phosphate-buffered saline (PBS) before trypsinization with 0.05 Trypsin-EDTA (Sigma Aldrich, T4174). HEK293T cells were plated at 1 (cid:2) 106 cells/well in six-well plates. Cells were transfected with pcDNA3.1.Tau4R-GFP plasmid 24 hr after plating using lipofectamine 2000 reagents according to the manufacturer’s guidelines. Briefly, 4 mg of pcDNA3.1.Tau4R-GFP was diluted in 250 ml of Opti- MEM media (GibcoTM, ThermoFisher). The expression of the fluorescent reporter was confirmed the next day through microscopic analysis. A stable cell line was established by replating the trans- fected cells at a 1:10 dilution and selecting in DMEM media containing 1200 mg/ml geneticin (Gib- coTM, ThermoFisher). Expression of the fused fluorescent proteins in the stable cell lines was confirmed using fluorescent microscopy. Polyclonal cells and monoclonal cells were grown to conflu- ency in 10 cm dishes, then stored in liquid nitrogen until use. Immunoblotting of cell lysate and zebrafish brain lysate For cell lysate preparation, cells were washed with cold PBS, then collected and incubated with cold lysis buffer (150 mM NaCl, 50 mM Tris-HCl (pH 8), 1 mM EDTA and 1% Nonidet P-40) supplemented with protease inhibitor (Cocktail Set III; Millipore) for 10 min on ice. Cells were lysed using a bio-vor- texer homogenizer for 20 s for two rounds. The lysate was centrifuged at 13,000 rpm for 10 min at 4˚C. The supernatant was collected, and the protein concentration was determined using the Qubit Protein Assay Kit (Invitrogen). For zebrafish brain lysate preparation, the brains of adult zebrafish were dissected. Brains were homogenized in cell lysis buffer (20 mM HEPES, 0.2 mM EDTA, 10 mM NaCl, 1.5 mM MgCl2, 20% glycerol, 0.1% Triton-X) with protease inhibitor and phospSTOP (Sigma-Aldrich) in the case of pt406 Tg. Brains were lysed using a bio-vortexer homogenizer and sonicated for 3 s for one round. Sam- ples were centrifuged as above and concentration of the samples was assessed in a Qubit fluorome- ter (Invitrogen). For immunoblotting, 30–40 mg of the total protein was combined with 2X sample buffer (Sigma- Aldrich) and boiled for 10 min before loading in 11% SDS-PAGE. Electrophoresis was performed using the Bio-Rad Power PAC system in running buffer (25 mM Tris base, 192 mM glycine and 0.1% SDS). The gel was transferred to a PVDF membrane using a wet transfer system. All membranes Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 23 of 33 Research article Neuroscience were blocked for one hour in protein-free blocking buffer PBS (ThermoFisher) or TBST with 5% milk and then incubated with primary antibody overnight at 4˚C with gentle agitation. The primary anti- bodies used in this study include rabbit monoclonal GFP (abcam, EPR14104) at 1:3000 dilution, rab- bit anti-b-actin (Sigma-Aldrich, A2066) at 1:10,000. All membranes were washed three times with 1X TBST before incubation with secondary antibody (goat-anti-mouse) HRP or HRP-conjugated anti-rab- bit at 1:5000 dilution (Jackson ImmunoResearch) for 1 hr at room temperature. The membranes were washed for the final time before visualization using Pierce ECL Western Blotting Substrate (ThermoFisher) on a ChemiDoc (Biorad). For stripping and re-probing, the membranes were stripped using mild stripping buffer (199.8 mM Glycine, 0.1% SDS, and 1% Tween 20 with a pH of 2.2) before blocking them and repeating the methods described before. Immunohistochemistry Larvae were fixed overnight in 4% paraformaldehyde, either 1 day after being subjected to TBI or following the subsequent application of drugs as indicated. Immunostaining of Activated-Caspase3 on whole-mount larvae was carried out as previously described (Duval et al., 2014). Larvae were washed with 0.1 M PO4 with 5% sucrose three times before washing with 1% Tween in H2O (pH 7.4), and then (cid:0)20˚C acetone. Larvae were incubated in PBS3+ containing 10% normal goat serum for 1 hr and then incubated with primary antibody with 2% normal goat serum in PBS3+. The primary anti- body used was polyclonal Anti-Active-Caspase-3 (BD Pharmingen, 559565) at 1:500 dilution. The secondary antibody applied was Alexafluor 647 anti-rabbit at 1:200 dilution (Invitrogen). Larvae were counterstained with 4’,6-diamidino-2-phenylindole (DAPI) (ThermoFisher) for 30 min. Preparations of mouse brain homogenate (crude and PTA precipitated) Brains from TgTauP301L mice and non-Tg littermate controls (129/SvEvTac genetic background) were provided by Dr. David Westaway and Dr. Nathalie Daude (Eskandari-Sedighi et al., 2017; Murakami et al., 2006). Crude brain homogenate was prepared by homogenizing the brains to 10% (wt/vol) in calcium- and magnesium-free DPBS that included a protease inhibitor and phosSTOP, using a glass homogenizer and power gen homogenizer (Fisher Scientific). Samples were then centri- fuged at 13,000 rpm for 15 min at 4˚C. The clear supernatant was collected, aliquoted and stored in (cid:0)80˚C until use for experiments. The phosphotungstate anion (PTA)-precipitated brain homogenate was prepared as described (Woerman et al., 2016). Briefly, 10% (wt/vol) brain homogenate was prepared as reported above and mixed with a final concentration of 2% sarkosyl (Sigma Aldrich) and 0.5% benzonase (Sigma Aldrich, E1014), and then incubated at 37˚C for two hours with constant agitation in an orbital shaker. Sodium PTA (Sigma Aldrich) was dissolved in ddH2O, and the pH was adjusted to 7.0 before it was added to the samples at a final concentration of 2% (vol/vol). The samples were then incu- bated overnight under the previous conditions. The next day, the samples were centrifuged at 16,000 g for 30 mins at room temperature. The supernatant was discarded, while the resulting pellet was resuspended in 2% (vol/vol) PTA in ddH2O (pH 7.0) and 2% sarkosyl in DPBS. The samples were next incubated for one hour before the second centrifugation. The supernatant was removed and the pellet was re-suspended in DPBS. An aliquot of 5 ml of PTA purified brain homogenate was employed for electron microscopy (EM) analysis to confirm the presence of fibrils in each sample. Tau fibrillization and EM analysis Synthetic human Tau protein (wildtype full-length monomers) was purchased as a lyophilized powder (rPeptide, T-1001–2) and resuspended in ddH2O at a concentration of 2 mg/ml. The recombinant protein was fibrillized as described previously (Guo and Lee, 2011). Recombinant Tau was incubated with 40 mM low-molecular-weight heparin and 2 mM DTT in 100 mM sodium acetate buffer (pH 7.0) at 37˚C, thereafter being agitated for seven days. The fibrillization mixture was centrifuged at 50,000 g for 30 mins, and the resulted pellet was resuspended in 100 mM sodium acetate buffer (pH 7.0) without heparin or DTT. Successful fibrillization was verified by EM. Negative staining for EM analysis of fibrils was conducted as described elsewhere (Eskandari- Sedighi et al., 2017). Briefly, 400 mesh carbon-coated copper grids (Electron Microscopy Sciences) were glow-discharged for 40 s before adding the sample aliquots. PTA-purified brain homogenates or synthetic Tau fibrils (5 mL) were applied on the top of the grid for 1 min. These grids were washed Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 24 of 33 Research article Neuroscience using 50 ml each of 0.1M and 0.01M ammonium acetate and negatively stained with 2 (cid:2) 50 ml of fil- tered 2% uranyl acetate. After removing excess stain and drying, the grids were examined with a Tecnai G20 transmission electron microscope (FEI Company) with an acceleration voltage of 200 kV. Electron micrographs were recorded with an Eagle 4k (cid:2) 4 k CCD camera (FEI Company). Liposome-mediated transduction of brain homogenate into tauopathy reporter cells Polyclonal Tau4R-GFP cells were plated at 2 (cid:2) 105 per well in 24-well plates. Cells were transduced the next day, using 40 ml of 10% clarified brain homogenate combined with Opti-MEM to a final vol- ume of 50 ml. A further 48 ml of Opti-MEM and 2 ml of Lipofectamine-2000 (Invitrogen) was added to the previous Opti-MEM mixture to a total volume of 100 ml and incubated for 20 min. The liposome mixture was applied to the cells for 18 hr, and cells were then washed with PBS, trypsinized, and re- plated on coated coverslips (ThermoFisher) for imaging and analysis. For PTA-precipitated brain homogenate, 1:10 dilution of precipitated fibrils was used for the transfection. 5 ml of PTA-purified fibrils was diluted in 45 ml Opti-MEM to a final volume of 50 ml. The previous Opti-MEM mixture was added to 47 ml of Opti-MEM and 3 ml of Lipofectamine-2000 and incubated in room temperature for 2 hr as described in Safar et al., 1998; Woerman et al., 2016. The mixture was added to cells, washed after 18 hr and re-plated before analysis exactly as men- tioned previously. Quantification of the percentage of cells with positive inclusion Prior to imaging, transfected cells were fixed 2% PFA in PBS for 15 mins. Samples were then washed twice with PBS then stained with DAPI (1:3000 from 1 mg/ml stock) for six mins. Cells were imaged using a Zeiss LSM 700 scanning confocal microscope featuring Zen 2010 software (Carl Zeiss, Ober- kochen, Germany). Due to increased brightness of the GFP+ puncta formed after introduction of brain homogenate, GFP exposure was minimized for those cells only. To quantify the GFP+ puncta, a total of nine images were collected and analyzed for each condition, each with ~100 cells. DAPI- positive nuclei were utilized to determine the number of cells per image. The number of cells with inclusions (multiple nuclear inclusions or one cytoplasmic puncta) were counted and the percentage was calculated. Brain ventricle injections into tauopathy reporter larvae Injections into the larval zebrafish brain (intraventricular space) were performed as described previ- ously with few modifications (Gutzman and Sive, 2009). Embryos at 2 dpf (days post-fertilization) were removed from their chorions and anesthetized with 4% tricaine (MS-222, Sigma Aldrich). The embryos were placed in a 1% agarose-coated dish with small holes. Under a stereomicroscope, the immobilized embryos were oriented so that the brain ventricles were accessible for injections. The injection was carried out via pulled capillary tubes mounted in a micromanipulator. The injection vol- ume was calibrated to 5 nL by injection into mineral oil and measurement with an ocular micrometer. Thereafter, the needle containing the injection solutions was placed through the roof plate of the hindbrain and 5–10 nL of either 10% clarified brain homogenate (TgTauP301L mice or wildtype litter- mate control), or synthetic tau, were mixed with 20% dextran Texas Red fluorescent dye (Invitrogen) and injected into the ventricles. For all the brain injection experiments, an uninjected control group and control group injected only with 20% red dextran fluorescent dye in PBS were included. After the injections, embryos were screened using a Leica M165 FC dissecting microscope and appropri- ately injected larvae were gathered for further analysis. The injections were considered appropriate if they had sharp edges and non-diffuse dye in the ventricle (Figure 1—figure supplement 2). Lar- vae receiving improper injections, in which the needle was inserted too deep in the brain ventricles resulting in the dye being visible outside the ventricle space and/or in the yolk, were excluded from analysis. Microscopy analysis of GFP-positive puncta in Tau reporter larvae For the microscopic analysis of GFP-positive inclusions, larvae that were either injected or treated with traumatic injury, along with the control groups, were anesthetized via tricaine at the indicated time point (two, three, four, or five days post-injection (dpi) or post-traumatic injury (dpti) depending Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 25 of 33 Research article Neuroscience on the experiment). Images for GFP-positive puncta on the brain area or lateral line above the spinal cord were taken using a Leica M165 FC dissecting microscope and the number of GFP-positive puncta were manually counted. TBI paradigm for zebrafish larvae To induce TBI, 10–12 unanesthetized larvae (3 dpf) were loaded into a 10-ml syringe with 1 ml of E3 media. The syringe was blocked using a stopper valve to ensure no larvae or media left the syringe upon compression of the plunger. The syringe was held vertical using a metal tube holder at the bot- tom end of a 48’ tube apparatus. A defined weight (between 30 and 300 g) was dropped manually from the top of the tube. The tube diameter was matched to (slightly greater than) the weight’s diameter to enhance repeatability. This was either done once or repeated three times, with either 65 or 300 g weights. Once larvae were subjected to the TBI, they were moved back to a petri dish with fresh media and maintained for further analysis. Quantifying the pressure induced during TBI To characterize the dynamic changes in pressure that occurred within the syringe during the TBI events, the stopper valve attached to the syringe (described immediately above) was replaced with a piezoresistive pressure transducer (#MLT844 AD Instruments, Colorado Springs, CO). Events were monitored via a PowerLab 2/26 data acquisition device and LabChart 7 software (AD Instruments). The pressure transducer was zeroed to report gauge pressure (pressure changes relative to atmo- spheric pressure) and was calibrated against a manometer (Fisherbrand Traceable from Thermo- scientific, Ottawa ON). After each weight drop, the syringe apparatus was reset to remove any air bubbles and the pressure transducer was zeroed. Time courses of induced pressure were reported over a 350 msec time frame with 50 msec of base line recording, while mean and maximum pressure values were calculated from the initial 300 msec following the impact of the weight. Recording blood flow following TBI Abnormalities of blood flow and circulation resulted from TBI was detected 5 to 10 mins after larvae were subjected to TBI. The blood flow in the tail area of zebrafish larvae, either those subjected to TBI or uninjured controls, was recorded using Leica DM2500 LED optical microscope. Measuring the Seizure-like phenotype in TBI larvae The seizure-like behavior and activity of zebrafish larvae post-traumatic injury experiment was quanti- fied via behavioral tracking software as described in our recent publications (Kanyo et al., 2020a; Leighton et al., 2018). Briefly, control larvae or larvae subjected to TBI using 65 g weight, were placed individually in wells of 96-well plates. The locomotor and seizure activity were assessed 40 min after the TBI through EthoVision XT-11.5 software (Noldus, Wageningen, Netherlands). The hypermotility of larvae is a manifestation of Stage I and Stage II seizures (previously defined via application of epileptic drugs), whereas more intense Stage III seizures are arrhythmic convulsions that manifest as reduced macroscopic movement in this assay (Kanyo et al., 2020a; Leighton et al., 2018; Liu and Baraban, 2019). Measuring neuronal activity during TBI using CaMPARI We used a recently described (Kanyo et al., 2020b) in-house allele of transgenic zebrafish express- ing the calcium sensor CaMPARI, line Tg[elavl3:CaMPARI (W391F+V398L)]ua3144. Due to a federal moratorium on importing zebrafish into Canada (Hanwell et al., 2016), we remade these fish using the Tol2 transgenesis system (Fisher et al., 2006) and a vector gifted by Eric Schreiter’s lab and published in Fosque et al., 2015. The transgene was bred onto the transparent Casper background. Larvae with robust CNS expression of CaMPARI were loaded into a 20 ml syringe containing 1 ml E3 media (prepared as per recipe in Westerfield, 2000, but without ethylene blue) and were exposed to a 405 nm LED array (Loctite), which illuminated the syringe entirely. Larvae were exposed for 10 s, with the LED array at a distance of 7.5 cm from the syringe, while being subjected to TBI using the 300 g weight as described above. Following this photoconversion of CaMPARI dur- ing TBI, larvae were anesthetized in 0.24 mg/mL tricaine methanesulfonate (MS-222, Sigma Aldrich) Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 26 of 33 Research article Neuroscience and embedded in 2% low-gelling agarose (A4018, Sigma Aldrich) for analysis under confocal microscopy. CaMPARI imaging began with acquisition of Z-stacks (8 mm steps) using a laser point-scanning confocal microscope (Zeiss 700, 20x/0.8 Objective), and visualized as maximum intensity projections. The hindbrain area was analyzed, as it was the brain region most responsive to TBI. To specifically isolate the brain regions and obtain data points, a 3D area was isolated by creating a surface with Imaris 7.6 (Bitman, Zuerich) and the mean fluorescence intensities of the green and red channel intensities were calculated. Data points were presented as a red/green ratio for each individual larva and interpreted as relative neural activity, which is defined as red photoconverted CaMPARI in ratio to green CaMPARI (Fosque et al., 2015; Kanyo et al., 2020b). Bath application of drugs Tau biosensor larvae were treated with 20 mM of the proteasome inhibitor MG-132 at 2dpf, follow- ing injections with brain homogenate from Tg human Tau mice. The treatment was left for 48 hr before changing the media and evaluating the percentage of larvae developing GFP+ puncta in the brain region. For Kainic acid or kainate treatment (KA), the doses (5, 50, 100, 150, and 200 mM) were selected based on previous use of KA in zebrafish larvae (Kim et al., 2010; Menezes et al., 2014), and added within 6 hr after TBI. For 4-aminopyridine (4-AP), one of two doses of 4-AP (200 or 800 mM) were added either six or 24 hr after TBI, as indicated. For Retigabine (RTG) treatment, 10 mM was used to treat TBI larvae beginning 6 hr after TBI. Doses of 4-AP and RTG were selected based on our previ- ous experience using them to affect seizures (Kanyo et al., 2020b). Unless otherwise stated, KA, 4- AP and/or RTG were applied to larvae for 38 hr, then a fresh drug-free E3 media was added. The formation of GFP-positive puncta was analyzed at four to five days post injury. Pyrimidyn-7 (P7), the dynamin inhibitor, was purchased at a 50 mM concentration supplied in DMSO (Abcam). Larvae that were subjected to TBI were treated within six hours following the injury with 3 mM of P7. The dose was chosen based on the previous use of the P7 drug on zebrafish larvae (Verweij et al., 2019). The larvae were incubated with the drug for 20 hr, after which they were transferred to a fresh plate with drug-free media. Dyngo 4a, another dynamin inhibitor (McCluskey et al., 2013), was purchased from (Abcam) and 4 mM of Dyngo 4a was used to treat lar- vae as previously explained with P7. The formation and abundance of GFP-positive puncta was eval- uated as previously described at 4 days post-traumatic injury (dpti). For some experiments, the ‘Tau biosensor’ transgenic zebrafish were bred to a separate Tg line that express human four repeat TAU Tg(eno2:hsa.MAPT-ires-egfp)Pt406 throughout the zebrafish CNS (Bai et al., 2007). Statistics All statistical analyses were performed using GraphPad Prism Software (Version 7, GraphPad, San Diego, CA). Sample sizes appropriate for our conclusions were estimated iteratively as the variance in each of our new methods became apparent; dose-response curves and significant differences amongst these dose were used to judge that any detected impacts of subsequent interventions would be valid. All experiments were independently replicated at least twice, individual larvae were the sampling unit (reported on Figures), and no outliers or other data were excluded. The experi- menters were blinded to the treatments prior to quantifying outcomes. Paired t-tests were used to compare between two groups, except for when sample sizes were too small to assess normality wherein Mann-Whitney tests were used. For comparison between three or more groups at various time points or the same time point, two-way and ordinary one-way ANOVA were used followed by post-hoc Mann-Whitney U tests and Kruskal-Wallace multiple comparison tests, respectively. Acknowledgements We acknowledge Nathalie Daude and David Westaway provided mouse brain samples, advice, and access to cell culture infrastructure. Gavin Neil and Jenna Bratvold contributed to SOD1-GFP cloning and zebrafish transgenesis via modifying a vector provided by Neil Cashman and Edward Pokrishev- sky. Mark Loewen provided advice on hydraulic measures of pressure. Sue-Ann Mok, Satya Kar, Oksana Suchowersky, David Westaway and Brian Christie provided comments on an earlier version of the manuscript. Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 27 of 33 Research article Neuroscience Funding to HA was from the Saudi Arabia Cultural Bureau and Majmaah University. RK was sup- ported by SynAD postdoctoral fellowship funded via Alzheimer Society of Alberta and Northwest Territories through their Hope for Tomorrow program and the University Hospital Foundation. LFL received Studentships from Alberta Innovates and NSERC. MGD received Studentships from Alberta Innovates and CIHR. Operating funds to EB were from CurePSP 655-2018-06 and 468–08, U.S. Department of Veterans Affairs BX003168, and NIH NS080881; The contents of this article do not represent the views of the United States government. Operating funds to HW were from Alberta Innovates and the Alzheimer Society of Alberta and Northwest Territories through the joint Alberta Alzheimer’s Research Program (AARP 201700005). Operating funds to WTA were also from the joint AARP (201700018), and from anonymous donors. Funders played no role in study design, prioritiza- tion, data collection or interpretation, or decision to submit the work for publication. Additional information Grant reference number Author Funding Funder Alberta Innovates - Health So- lutions Natural Sciences and Engi- neering Research Council of Canada Canadian Institutes of Health Research Alberta Innovates Bio Solutions 201700005 Alberta Innovates Bio Solutions 201700018 Anonymous Donors Laszlo F Locskai Miche` le G DuVal Laszlo F Locskai Miche` le G DuVal Holger Wille W Ted Allison W Ted Allison National Institutes of Health NS080881 Edward A Burton CurePSP 655-2018-06 and 468-08 Edward A Burton U.S. Department of Veterans Affairs BX003168 Majmaah University Saudi Arabia Cultural Bureau in Ottawa Edward A Burton Hadeel Alyenbaawi Hadeel Alyenbaawi The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Author contributions Hadeel Alyenbaawi, Conceptualization, Resources, Formal analysis, Supervision, Funding acquisition, Investigation, Visualization, Methodology, Writing - original draft, Project administration, Writing - review and editing; Richard Kanyo, Formal analysis, Supervision, Investigation, Visualization, Method- ology, Writing - review and editing; Laszlo F Locskai, Resources, Formal analysis, Supervision, Investi- gation, Visualization, Methodology, Writing - review and editing; Razieh Kamali-Jamil, Investigation, Writing - review and editing; Miche` le G DuVal, Qing Bai, Resources, Investigation, Writing - review and editing; Holger Wille, Edward A Burton, Resources, Supervision, Writing - review and editing; W Ted Allison, Conceptualization, Resources, Formal analysis, Supervision, Funding acquisition, Investi- gation, Visualization, Methodology, Project administration, Writing - review and editing Author ORCIDs Miche` le G DuVal Holger Wille Edward A Burton W Ted Allison http://orcid.org/0000-0001-6975-8117 http://orcid.org/0000-0001-5102-8706 http://orcid.org/0000-0002-8072-4636 https://orcid.org/0000-0002-8461-4864 Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 28 of 33 Research article Neuroscience Ethics Animal experimentation: Our Discussion invokes animal research ethics as an advantageous aspect of our technology development. Our Methods section begins with the following: Animal Ethics and Zebrafish Husbandry Zebrafish were raised and maintained following protocol AUP00000077 approved by the Animal Care and Use Committee: Biosciences at the University of Alberta, operat- ing under the guidelines of the Canadian Council of Animal Care. Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.58744.sa1 Author response https://doi.org/10.7554/eLife.58744.sa2 Additional files Supplementary files . Source data 1. Source data and statistics. . Transparent reporting form Data availability All data generated or analysed during this study are included in the manuscript and supporting files. Source data files have been provided in Source Data 1. References Asikainen I, Kaste M, Sarna S. 1999. Early and late posttraumatic seizures in traumatic brain injury rehabilitation patients: brain injury factors causing late seizures and influence of seizures on long-term outcome. Epilepsia 40: 584–589. DOI: https://doi.org/10.1111/j.1528-1157.1999.tb05560.x, PMID: 10386527 Ayers JI, Giasson BI, Borchelt DR. 2018. Prion-like spreading in tauopathies. Biological Psychiatry 83:337–346. DOI: https://doi.org/10.1016/j.biopsych.2017.04.003, PMID: 28506438 Bai Q, Garver JA, Hukriede NA, Burton EA. 2007. Generation of a transgenic zebrafish model of tauopathy using a novel promoter element derived from the zebrafish eno2 gene. Nucleic Acids Research 35:6501–6516. DOI: https://doi.org/10.1093/nar/gkm608, PMID: 17897967 Balik-Meisner M, Truong L, Scholl EH, Tanguay RL, Reif DM. 2018. Population genetic diversity in zebrafish lines. Mammalian Genome 29:90–100. DOI: https://doi.org/10.1007/s00335-018-9735-x, PMID: 29368091 Bir C, Vandevord P, Shen Y, Raza W, Haacke EM. 2012. Effects of variable blast pressures on blood flow and oxygen saturation in rat brain as evidenced using MRI. Magnetic Resonance Imaging 30:527–534. DOI: https:// doi.org/10.1016/j.mri.2011.12.003, PMID: 22285875 Bodnar CN, Roberts KN, Higgins EK, Bachstetter AD. 2019. A systematic review of closed head injury models of mild traumatic brain injury in mice and rats. Journal of Neurotrauma 36:1683–1706. DOI: https://doi.org/10. 1089/neu.2018.6127, PMID: 30661454 Bugay V, Bozdemir E, Vigil FA, Chun SH, Holstein DM, Elliott WR, Sprague CJ, Cavazos JE, Zamora DO, Rule G, Shapiro MS, Lechleiter JD, Brenner R. 2020. A mouse model of repetitive blast traumatic brain injury reveals Post-Trauma seizures and increased neuronal excitability. Journal of Neurotrauma 37:248–261. DOI: https:// doi.org/10.1089/neu.2018.6333, PMID: 31025597 Chauhan NB. 2014. Chronic neurodegenerative consequences of traumatic brain injury. Restorative Neurology and Neuroscience 32:337–365. DOI: https://doi.org/10.3233/RNN-130354, PMID: 24398724 Chiti F, Dobson CM. 2006. Protein misfolding, functional amyloid, and human disease. Annual Review of Biochemistry 75:333–366. DOI: https://doi.org/10.1146/annurev.biochem.75.101304.123901, PMID: 16756495 Cho SJ, Park E, Telliyan T, Baker A, Reid AY. 2020. Zebrafish model of posttraumatic epilepsy. Epilepsia 61: 1774–1785. DOI: https://doi.org/10.1111/epi.16589, PMID: 32592416 Clavaguera F, Akatsu H, Fraser G, Crowther RA, Frank S, Hench J, Probst A, Winkler DT, Reichwald J, Staufenbiel M, Ghetti B, Goedert M, Tolnay M. 2013. Brain homogenates from human tauopathies induce Tau inclusions in mouse brain. PNAS 110:9535–9540. DOI: https://doi.org/10.1073/pnas.1301175110, PMID: 236 90619 Colin M, Dujardin S, Schraen-Maschke S, Meno-Tetang G, Duyckaerts C, Courade JP, Bue´ e L. 2020. From the prion-like propagation hypothesis to therapeutic strategies of anti-tau immunotherapy. Acta Neuropathologica 139:3–25. DOI: https://doi.org/10.1007/s00401-019-02087-9, PMID: 31686182 Cousin MA, Robinson PJ. 2000. Ca(2+) influx inhibits dynamin and arrests synaptic vesicle endocytosis at the active zone. The Journal of Neuroscience 20:949–957. PMID: 10648699 Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 29 of 33 Research article Neuroscience Cruz-Haces M, Tang J, Acosta G, Fernandez J, Shi R. 2017. Pathological correlations between traumatic brain injury and chronic neurodegenerative diseases. Translational Neurodegeneration 6:20. DOI: https://doi.org/10. 1186/s40035-017-0088-2, PMID: 28702179 de Calignon A, Polydoro M, Sua´ rez-Calvet M, William C, Adamowicz DH, Kopeikina KJ, Pitstick R, Sahara N, Ashe KH, Carlson GA, Spires-Jones TL, Hyman BT. 2012. Propagation of Tau pathology in a model of early alzheimer’s disease. Neuron 73:685–697. DOI: https://doi.org/10.1016/j.neuron.2011.11.033, PMID: 22365544 Demaegd K, Schymkowitz J, Rousseau F. 2018. Transcellular spreading of Tau in tauopathies. ChemBioChem 19: 2424–2432. DOI: https://doi.org/10.1002/cbic.201800288, PMID: 30133080 Duval MG, Oel AP, Allison WT. 2014. gdf6a is required for cone photoreceptor subtype differentiation and for the actions of tbx2b in determining rod versus cone photoreceptor fate. PLOS ONE 9:e92991. DOI: https://doi. org/10.1371/journal.pone.0092991, PMID: 24681822 Ellis LD, Seibert J, Soanes KH. 2012. Distinct models of induced hyperactivity in zebrafish larvae. Brain Research 1449:46–59. DOI: https://doi.org/10.1016/j.brainres.2012.02.022, PMID: 22386495 Englander J, Cifu DX, Diaz-Arrastia R. 2014. Information/education page seizures and traumatic brain injury. Archives of Physical Medicine and Rehabilitation 95:1223–1224. DOI: https://doi.org/10.1016/j.apmr.2013.06. 002, PMID: 24862307 Eskandari-Sedighi G, Daude N, Gapeshina H, Sanders DW, Kamali-Jamil R, Yang J, Shi B, Wille H, Ghetti B, Diamond MI, Janus C, Westaway D. 2017. The CNS in inbred transgenic models of 4-repeat tauopathy develops consistent Tau seeding capacity yet focal and diverse patterns of protein deposition. Molecular Neurodegeneration 12:72. DOI: https://doi.org/10.1186/s13024-017-0215-7, PMID: 28978354 Evans LD, Wassmer T, Fraser G, Smith J, Perkinton M, Billinton A, Livesey FJ. 2018. Extracellular monomeric and aggregated Tau efficiently enter human neurons through overlapping but distinct pathways. Cell Reports 22: 3612–3624. DOI: https://doi.org/10.1016/j.celrep.2018.03.021, PMID: 29590627 Fisher S, Grice EA, Vinton RM, Bessling SL, Urasaki A, Kawakami K, McCallion AS. 2006. Evaluating the biological relevance of putative enhancers using Tol2 transposon-mediated transgenesis in zebrafish. Nature Protocols 1: 1297–1305. DOI: https://doi.org/10.1038/nprot.2006.230, PMID: 17406414 Fosque BF, Sun Y, Dana H, Yang CT, Ohyama T, Tadross MR, Patel R, Zlatic M, Kim DS, Ahrens MB, Jayaraman V, Looger LL, Schreiter ER. 2015. Neural circuits labeling of active neural circuits in vivo with designed calcium integrators. Science 347:755–760. DOI: https://doi.org/10.1126/science.1260922, PMID: 25678659 Gardner RC, Yaffe K. 2015. Epidemiology of mild traumatic brain injury and neurodegenerative disease. Molecular and Cellular Neuroscience 66:75–80. DOI: https://doi.org/10.1016/j.mcn.2015.03.001, PMID: 2574 8121 Geddes JF, Vowles GH, Nicoll JA, Re´ ve´ sz T. 1999. Neuronal cytoskeletal changes are an early consequence of repetitive head injury. Acta Neuropathologica 98:171–178. DOI: https://doi.org/10.1007/s004010051066, PMID: 10442557 Goedert M, Spillantini MG, Jakes R, Rutherford D, Crowther RA. 1989. Multiple isoforms of human microtubule- associated protein tau: sequences and localization in neurofibrillary tangles of Alzheimer’s disease. Neuron 3: 519–526. DOI: https://doi.org/10.1016/0896-6273(89)90210-9, PMID: 2484340 Goedert M, Eisenberg DS, Crowther RA. 2017a. Propagation of Tau aggregates and neurodegeneration. Annual Review of Neuroscience 40:189–210. DOI: https://doi.org/10.1146/annurev-neuro-072116-031153, PMID: 2 8772101 Goedert M, Masuda-Suzukake M, Falcon B. 2017b. Like prions: the propagation of aggregated Tau and a- synuclein in neurodegeneration. Brain 140:266–278. DOI: https://doi.org/10.1093/brain/aww230, PMID: 2765 8420 Guo JL, Narasimhan S, Changolkar L, He Z, Stieber A, Zhang B, Gathagan RJ, Iba M, McBride JD, Trojanowski JQ, Lee VM. 2016. Unique pathological Tau conformers from Alzheimer’s brains transmit Tau pathology in nontransgenic mice. Journal of Experimental Medicine 213:2635–2654. DOI: https://doi.org/10.1084/jem. 20160833, PMID: 27810929 Guo JL, Lee VM-Y. 2011. Seeding of normal Tau by pathological Tau conformers drives pathogenesis of Alzheimer-like tangles. Journal of Biological Chemistry 286:15317–15331. DOI: https://doi.org/10.1074/jbc. M110.209296 Guryev V, Koudijs MJ, Berezikov E, Johnson SL, Plasterk RH, van Eeden FJ, Cuppen E. 2006. Genetic variation in the zebrafish. Genome Research 16:491–497. DOI: https://doi.org/10.1101/gr.4791006, PMID: 16533913 Gutzman JH, Sive H. 2009. Zebrafish brain ventricle injection. Journal of Visualized Experiments 26:1218. DOI: https://doi.org/10.3791/1218 Hanwell D, Hutchinson SA, Collymore C, Bruce AE, Louis R, Ghalami A, Allison WT, Ekker M, Eames BF, Childs S, Kurrasch DM, Gerlai R, Thiele T, Scott I, Ciruna B, Dowling JJ, McFarlane S, Huang P, Wen X-Y, Akimenko M-A, et al. 2016. Restrictions on the importation of zebrafish into Canada associated with spring viremia of carp virus. Zebrafish 13:S-153–-150. DOI: https://doi.org/10.1089/zeb.2016.1286 Hay J, Johnson VE, Smith DH, Stewart W. 2016. Chronic traumatic encephalopathy: the neuropathological legacy of traumatic brain injury. Annual Review of Pathology: Mechanisms of Disease 11:21–45. DOI: https://doi.org/ 10.1146/annurev-pathol-012615-044116, PMID: 26772317 Iba M, Guo JL, McBride JD, Zhang B, Trojanowski JQ, Lee VM. 2013. Synthetic tau fibrils mediate transmission of neurofibrillary tangles in a transgenic mouse model of alzheimer’s-like tauopathy. Journal of Neuroscience 33: 1024–1037. DOI: https://doi.org/10.1523/JNEUROSCI.2642-12.2013, PMID: 23325240 Iba M, McBride JD, Guo JL, Zhang B, Trojanowski JQ, Lee VM. 2015. Tau pathology spread in PS19 Tau transgenic mice following locus coeruleus (LC) injections of synthetic Tau fibrils is determined by the LC’s Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 30 of 33 Research article Neuroscience afferent and efferent connections. Acta Neuropathologica 130:349–362. DOI: https://doi.org/10.1007/s00401- 015-1458-4, PMID: 26150341 Johnson VE, Stewart W, Smith DH. 2012. Widespread t and amyloid-b pathology many years after a single traumatic brain injury in humans. Brain Pathology 22:142–149. DOI: https://doi.org/10.1111/j.1750-3639.2011. 00513.x, PMID: 21714827 Johnson VE, Stewart JE, Begbie FD, Trojanowski JQ, Smith DH, Stewart W. 2013. Inflammation and white matter degeneration persist for years after a single traumatic brain injury. Brain 136:28–42. DOI: https://doi.org/10. 1093/brain/aws322, PMID: 23365092 Kanyo R, Leighton PLA, Neil GJ, Locskai LF, Allison WT. 2020a. Amyloid-b precursor protein mutant zebrafish exhibit seizure susceptibility that depends on prion protein. Experimental Neurology 328:113283. DOI: https:// doi.org/10.1016/j.expneurol.2020.113283, PMID: 32165257 Kanyo R, Wang CK, Locskai LF, Li J, Allison WT, Kurata HT. 2020b. Functional and behavioral signatures of Kv7 activator drug subtypes. Epilepsia 61:1678–1690. DOI: https://doi.org/10.1111/epi.16592, PMID: 32652600 Kasatkina LA. 2016. 4-Aminopyridine sequesters intracellular Ca2+ which triggers exocytosis in excitable and non-excitable cells. Scientific Reports 6:34749. DOI: https://doi.org/10.1038/srep34749, PMID: 27703262 Kaufman SK, Sanders DW, Thomas TL, Ruchinskas AJ, Vaquer-Alicea J, Sharma AM, Miller TM, Diamond MI. 2016.In Tau prion strains dictate patterns of cell pathology, progression rate, and regional vulnerability in Vivo. Neuron 92:796–812. DOI: https://doi.org/10.1016/j.neuron.2016.09.055, PMID: 27974162 Kim YH, Lee Y, Lee K, Lee T, Kim YJ, Lee CJ. 2010. Reduced neuronal proliferation by proconvulsant drugs in the developing zebrafish brain. Neurotoxicology and Teratology 32:551–557. DOI: https://doi.org/10.1016/j.ntt. 2010.04.054, PMID: 20420900 Kovacs SK, Leonessa F, Ling GS. 2014. Blast TBI models, neuropathology, and implications for seizure risk. Frontiers in Neurology 5:47. DOI: https://doi.org/10.3389/fneur.2014.00047, PMID: 24782820 Kovacs GG. 2017. Tauopathies. In: Alafuzoff I. rina (Ed). Handbook of Clinical Neurology. Elsevier. p. 355–368. Kwan KM, Fujimoto E, Grabher C, Mangum BD, Hardy ME, Campbell DS, Parant JM, Yost HJ, Kanki JP, Chien CB. 2007. The Tol2kit: a multisite gateway-based construction kit for Tol2 transposon transgenesis constructs. Developmental Dynamics 236:3088–3099. DOI: https://doi.org/10.1002/dvdy.21343, PMID: 17937395 Leighton PLA, Kanyo R, Neil GJ, Pollock NM, Allison WT. 2018. Prion gene paralogs are dispensable for early zebrafish development and have nonadditive roles in seizure susceptibility. Journal of Biological Chemistry 293: 12576–12592. DOI: https://doi.org/10.1074/jbc.RA117.001171 Lim J, Yue Z. 2015. Neuronal aggregates: formation, clearance, and spreading. Developmental Cell 32:491–501. DOI: https://doi.org/10.1016/j.devcel.2015.02.002, PMID: 25710535 Liu J, Baraban SC. 2019. Network properties revealed during Multi-Scale calcium imaging of seizure activity in zebrafish. Eneuro 6:ENEURO.0041-19.2019. DOI: https://doi.org/10.1523/ENEURO.0041-19.2019, PMID: 30 895220 Lucke-Wold BP, Nguyen L, Turner RC, Logsdon AF, Chen YW, Smith KE, Huber JD, Matsumoto R, Rosen CL, Tucker ES, Richter E. 2015. Traumatic brain injury and epilepsy: underlying mechanisms leading to seizure. Seizure 33:13–23. DOI: https://doi.org/10.1016/j.seizure.2015.10.002, PMID: 26519659 Lundh H. 1978. Effects of 4-aminopyridine on neuromuscular transmission. Brain Research 153:307–318. DOI: https://doi.org/10.1016/0006-8993(78)90409-2, PMID: 210882 Maheras AL, Dix B, Carmo OMS, Young AE, Gill VN, Sun JL, Booker AR, Thomason HA, Ibrahim AE, Stanislaw L, Dallego JC, Ngo CN, Chen A, Fortini BK, Spence RD. 2018. Genetic pathways of neuroregeneration in a novel mild traumatic brain injury model in adult zebrafish. Eneuro 5:ENEURO.0208-17.2017. DOI: https://doi.org/10. 1523/ENEURO.0208-17.2017, PMID: 29302617 Marklund N. 2016. Rodent models of traumatic brain injury: methods and challenges. Methods in Molecular Biology 1462:29–46. DOI: https://doi.org/10.1007/978-1-4939-3816-2_3, PMID: 27604711 McCluskey A, Daniel JA, Hadzic G, Chau N, Clayton EL, Mariana A, Whiting A, Gorgani NN, Lloyd J, Quan A, Moshkanbaryans L, Krishnan S, Perera S, Chircop M, von Kleist L, McGeachie AB, Howes MT, Parton RG, Campbell M, Sakoff JA, et al. 2013. Building a better dynasore: the dyngo compounds potently inhibit dynamin and endocytosis. Traffic 14:1272–1289. DOI: https://doi.org/10.1111/tra.12119, PMID: 24025110 McCutcheon V, Park E, Liu E, Wang Y, Wen XY, Baker AJ. 2016. A model of excitotoxic brain injury in larval zebrafish: potential application for High-Throughput drug evaluation to treat traumatic brain injury. Zebrafish 13:161–169. DOI: https://doi.org/10.1089/zeb.2015.1188, PMID: 27028704 McCutcheon V, Park E, Liu E, Sobhebidari P, Tavakkoli J, Wen XY, Baker AJ. 2017. A novel model of traumatic brain injury in adult zebrafish demonstrates response to injury and treatment comparable with mammalian models. Journal of Neurotrauma 34:1382–1393. DOI: https://doi.org/10.1089/neu.2016.4497, PMID: 27650063 McGeachie AB, Odell LR, Quan A, Daniel JA, Chau N, Hill TA, Gorgani NN, Keating DJ, Cousin MA, van Dam EM, Mariana A, Whiting A, Perera S, Novelle A, Young KA, Deane FM, Gilbert J, Sakoff JA, Chircop M, McCluskey A, et al. 2013. Pyrimidyn compounds: dual-action small molecule pyrimidine-based dynamin inhibitors. ACS Chemical Biology 8:1507–1518. DOI: https://doi.org/10.1021/cb400137p, PMID: 23642287 McKee AC, Stein TD, Kiernan PT, Alvarez VE. 2015. The neuropathology of chronic traumatic encephalopathy. Brain Pathology 25:350–364. DOI: https://doi.org/10.1111/bpa.12248, PMID: 25904048 Meconi A, Wortman RC, Wright DK, Neale KJ, Clarkson M, Shultz SR, Christie BR. 2018. Repeated mild traumatic brain injury can cause acute neurologic impairment without overt structural damage in juvenile rats. PLOS ONE 13:e0197187. DOI: https://doi.org/10.1371/journal.pone.0197187, PMID: 29738554 Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 31 of 33 Research article Neuroscience Menezes FP, Rico EP, Da Silva RS. 2014. Tolerance to seizure induced by kainic acid is produced in a specific period of zebrafish development. Progress in Neuro-Psychopharmacology and Biological Psychiatry 55:109– 112. DOI: https://doi.org/10.1016/j.pnpbp.2014.04.004, PMID: 24743104 Mudher A, Colin M, Dujardin S, Medina M, Dewachter I, Alavi Naini SM, Mandelkow EM, Mandelkow E, Bue´ e L, Goedert M, Brion JP. 2017. What is the evidence that tau pathology spreads through prion-like propagation? Acta Neuropathologica Communications 5:99. DOI: https://doi.org/10.1186/s40478-017-0488-7, PMID: 2925 8615 Murakami T, Paitel E, Kawarabayashi T, Ikeda M, Chishti MA, Janus C, Matsubara E, Sasaki A, Kawarai T, Phinney AL, Harigaya Y, Horne P, Egashira N, Mishima K, Hanna A, Yang J, Iwasaki K, Takahashi M, Fujiwara M, Ishiguro K, et al. 2006. Cortical neuronal and glial pathology in TgTauP301L transgenic mice: neuronal degeneration, memory disturbance, and phenotypic variation. The American Journal of Pathology 169:1365–1375. DOI: https://doi.org/10.2353/ajpath.2006.051250, PMID: 17003492 Mychasiuk R, Farran A, Angoa-Perez M, Briggs D, Kuhn D, Esser MJ. 2014. A novel model of mild traumatic brain injury for juvenile rats. Journal of Visualized Experiments 8:51820. DOI: https://doi.org/10.3791/51820 Nakagawa A, Manley GT, Gean AD, Ohtani K, Armonda R, Tsukamoto A, Yamamoto H, Takayama K, Tominaga T. 2011. Mechanisms of primary blast-induced traumatic brain injury: insights from shock-wave research. Journal of Neurotrauma 28:1101–1119. DOI: https://doi.org/10.1089/neu.2010.1442, PMID: 21332411 Narasimhan S, Guo JL, Changolkar L, Stieber A, McBride JD, Silva LV, He Z, Zhang B, Gathagan RJ, Trojanowski JQ, Lee VMY. 2017. Pathological tau strains from human brains recapitulate the diversity of tauopathies in nontransgenic mouse brain. The Journal of Neuroscience 37:11406–11423. DOI: https://doi.org/10.1523/ JNEUROSCI.1230-17.2017, PMID: 29054878 Nguyen R, Fiest KM, McChesney J, Kwon C-S, Jette N, Frolkis AD, Atta C, Mah S, Dhaliwal H, Reid A, Pringsheim T, Dykeman J, Gallagher C. 2016. The international incidence of traumatic brain injury: a systematic review and Meta-Analysis. Canadian Journal of Neurological Sciences / Journal Canadien Des Sciences Neurologiques 43:774–785. DOI: https://doi.org/10.1017/cjn.2016.290 Ojo JO, Mouzon B, Algamal M, Leary P, Lynch C, Abdullah L, Evans J, Mullan M, Bachmeier C, Stewart W, Crawford F. 2016. Chronic repetitive mild traumatic brain injury results in reduced cerebral blood flow, axonal injury, Gliosis, and increased T-Tau and tau oligomers. Journal of Neuropathology & Experimental Neurology 75:636–655. DOI: https://doi.org/10.1093/jnen/nlw035, PMID: 27251042 Orr ME, Sullivan AC, Frost B. 2017. A brief overview of tauopathy: causes, consequences, and therapeutic strategies. Trends in Pharmacological Sciences 38:637–648. DOI: https://doi.org/10.1016/j.tips.2017.03.011, PMID: 28455089 Peeraer E, Bottelbergs A, Van Kolen K, Stancu IC, Vasconcelos B, Mahieu M, Duytschaever H, Ver Donck L, Torremans A, Sluydts E, Van Acker N, Kemp JA, Mercken M, Brunden KR, Trojanowski JQ, Dewachter I, Lee VM, Moechars D. 2015. Intracerebral injection of preformed synthetic tau fibrils initiates widespread tauopathy and neuronal loss in the brains of tau transgenic mice. Neurobiology of Disease 73:83–95. DOI: https://doi.org/ 10.1016/j.nbd.2014.08.032, PMID: 25220759 Pham L, Shultz SR, Kim HA, Brady RD, Wortman RC, Genders SG, Hale MW, O’Shea RD, Djouma E, van den Buuse M, Church JE, Christie BR, Drummond GR, Sobey CG, McDonald SJ. 2019. Mild Closed-Head injury in conscious rats causes transient neurobehavioral and glial disturbances: a novel experimental model of concussion. Journal of Neurotrauma 36:2260–2271. DOI: https://doi.org/10.1089/neu.2018.6169, PMID: 30 843474 Pickett EK, Henstridge CM, Allison E, Pitstick R, Pooler A, Wegmann S, Carlson G, Hyman BT, Spires-Jones TL. 2017. Spread of tau down neural circuits precedes synapse and neuronal loss in the rTgTauEC mouse model of early Alzheimer’s disease. Synapse 71:e21965. DOI: https://doi.org/10.1002/syn.21965 Ping X, Jin X. 2016. Transition from initial hypoactivity to hyperactivity in cortical layer V pyramidal neurons after traumatic brain injury In Vivo. Journal of Neurotrauma 33:354–361. DOI: https://doi.org/10.1089/neu.2015. 3913, PMID: 26095991 Pokrishevsky E, McAlary L, Farrawell NE, Zhao B, Sher M, Yerbury JJ, Cashman NR. 2018. Tryptophan 32- mediated SOD1 aggregation is attenuated by pyrimidine-like compounds in living cells. Scientific Reports 8: 15590. DOI: https://doi.org/10.1038/s41598-018-32835-y, PMID: 30349065 Puvenna V, Engeler M, Banjara M, Brennan C, Schreiber P, Dadas A, Bahrami A, Solanki J, Bandyopadhyay A, Morris JK, Bernick C, Ghosh C, Rapp E, Bazarian JJ, Janigro D. 2016. Is phosphorylated tau unique to chronic traumatic encephalopathy? phosphorylated tau in epileptic brain and chronic traumatic encephalopathy. Brain Research 1630:225–240. DOI: https://doi.org/10.1016/j.brainres.2015.11.007, PMID: 26556772 Rimel RW, Giordani B, Barth JT, Boll TJ, Jane JA. 1981. Disability caused by minor head injury. Neurosurgery 9: 221–228. PMID: 7301062 Russell WMS, Burch RL. 1959. The Principles of Humane Experimental Technique. Methuen: Johns Hopkins. Safar J, Wille H, Itri V, Groth D, Serban H, Torchia M, Cohen FE, Prusiner SB. 1998. Eight prion strains have PrP (Sc) molecules with different conformations. Nature Medicine 4:1157–1165. DOI: https://doi.org/10.1038/2654, PMID: 9771749 Saleem S, Kannan RR. 2018. Zebrafish: an emerging real-time model system to study Alzheimer’s disease and neurospecific drug discovery. Cell Death Discovery 4:45. DOI: https://doi.org/10.1038/s41420-018-0109-7, PMID: 30302279 Salinsky M, Storzbach D, Goy E, Evrard C. 2015. Traumatic brain injury and psychogenic seizures in veterans. Journal of Head Trauma Rehabilitation 30:E65–E70. DOI: https://doi.org/10.1097/HTR.0000000000000057, PMID: 24901325 Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 32 of 33 Research article Neuroscience Sa´ nchez M, Garcı´a-Cabrero A, Sa´ nchez-Elexpuru G, Burgos D, Serratosa J. 2018. Tau-Induced pathology in epilepsy and dementia: notions from patients and animal models. International Journal of Molecular Sciences 19:1092. DOI: https://doi.org/10.3390/ijms19041092 Sanders DW, Kaufman SK, DeVos SL, Sharma AM, Mirbaha H, Li A, Barker SJ, Foley AC, Thorpe JR, Serpell LC, Miller TM, Grinberg LT, Seeley WW, Diamond MI. 2014. Distinct tau prion strains propagate in cells and mice and define different tauopathies. Neuron 82:1271–1288. DOI: https://doi.org/10.1016/j.neuron.2014.04.047, PMID: 24857020 Singh M, Jadhav HR, Bhatt T. 2017. Dynamin functions and ligands: classical mechanisms behind. Molecular Pharmacology 91:123–134. DOI: https://doi.org/10.1124/mol.116.105064, PMID: 27879341 Stafstrom CE, Carmant L. 2015. Seizures and epilepsy: an overview for neuroscientists. Cold Spring Harbor Perspectives in Medicine 5:a022426. DOI: https://doi.org/10.1101/cshperspect.a022426, PMID: 26033084 Tai XY, Koepp M, Duncan JS, Fox N, Thompson P, Baxendale S, Liu JY, Reeves C, Michalak Z, Thom M. 2016. Hyperphosphorylated tau in patients with refractory epilepsy correlates with cognitive decline: a study of temporal lobe resections. Brain 139:2441–2455. DOI: https://doi.org/10.1093/brain/aww187, PMID: 27497924 Thom M, Liu JY, Thompson P, Phadke R, Narkiewicz M, Martinian L, Marsdon D, Koepp M, Caboclo L, Catarino CB, Sisodiya SM. 2011. Neurofibrillary tangle pathology and Braak staging in chronic epilepsy in relation to traumatic brain injury and hippocampal sclerosis: a post-mortem study. Brain 134:2969–2981. DOI: https://doi. org/10.1093/brain/awr209, PMID: 21903728 Traver D, Paw BH, Poss KD, Penberthy WT, Lin S, Zon LI. 2003. Transplantation and in vivo imaging of multilineage engraftment in zebrafish bloodless mutants. Nature Immunology 4:1238–1246. DOI: https://doi. org/10.1038/ni1007, PMID: 14608381 Uryu K, Chen XH, Martinez D, Browne KD, Johnson VE, Graham DI, Lee VM, Trojanowski JQ, Smith DH. 2007. Multiple proteins implicated in neurodegenerative diseases accumulate in axons after brain trauma in humans. Experimental Neurology 208:185–192. DOI: https://doi.org/10.1016/j.expneurol.2007.06.018, PMID: 17826768 Verweij FJ, Revenu C, Arras G, Dingli F, Loew D, Pegtel DM, Follain G, Allio G, Goetz JG, Zimmermann P, Herbomel P, Del Bene F, Raposo G, van Niel G. 2019. Live tracking of Inter-organ communication by endogenous exosomes in Vivo. Developmental CellIn 48:573–589. DOI: https://doi.org/10.1016/j.devcel.2019. 01.004, PMID: 30745143 Wang Y, Balaji V, Kaniyappan S, Kru¨ ger L, Irsen S, Tepper K, Chandupatla R, Maetzler W, Schneider A, Mandelkow E, Mandelkow EM. 2017. The release and trans-synaptic transmission of tau via exosomes. Molecular Neurodegeneration 12:5. DOI: https://doi.org/10.1186/s13024-016-0143-y, PMID: 28086931 Washington PM, Villapol S, Burns MP. 2016. Polypathology and dementia after brain trauma: does brain injury trigger distinct neurodegenerative diseases, or should they be classified together as traumatic encephalopathy? Experimental Neurology 275:381–388. DOI: https://doi.org/10.1016/j.expneurol.2015.06.015, PMID: 26091850 Westerfield M. 2000. The Zebrafish Book: A Guide for the Laboratory Use of Zebrafish (Danio rerio. Fourth Edition. Eugene: University of Oregon press. White RM, Sessa A, Burke C, Bowman T, LeBlanc J, Ceol C, Bourque C, Dovey M, Goessling W, Burns CE, Zon LI. 2008. Transparent adult zebrafish as a tool for in vivo transplantation analysis. Cell Stem Cell 2:183–189. DOI: https://doi.org/10.1016/j.stem.2007.11.002, PMID: 18371439 Winter MJ, Windell D, Metz J, Matthews P, Pinion J, Brown JT, Hetheridge MJ, Ball JS, Owen SF, Redfern WS, Moger J, Randall AD, Tyler CR. 2017. 4-dimensional functional profiling in the convulsant-treated larval zebrafish brain. Scientific Reports 7:6581. DOI: https://doi.org/10.1038/s41598-017-06646-6, PMID: 28747660 Woerman AL, Aoyagi A, Patel S, Kazmi SA, Lobach I, Grinberg LT, McKee AC, Seeley WW, Olson SH, Prusiner SB. 2016. Tau prions from Alzheimer’s disease and chronic traumatic encephalopathy patients propagate in cultured cells. PNAS 113:E8187–E8196. DOI: https://doi.org/10.1073/pnas.1616344113, PMID: 27911827 Wu JW, Herman M, Liu L, Simoes S, Acker CM, Figueroa H, Steinberg JI, Margittai M, Kayed R, Zurzolo C, Di Paolo G, Duff KE. 2013. Small misfolded tau species are internalized via bulk endocytosis and anterogradely and retrogradely transported in neurons. Journal of Biological Chemistry 288:1856–1870. DOI: https://doi.org/ 10.1074/jbc.M112.394528 Wu JW, Hussaini SA, Bastille IM, Rodriguez GA, Mrejeru A, Rilett K, Sanders DW, Cook C, Fu H, Boonen RA, Herman M, Nahmani E, Emrani S, Figueroa YH, Diamond MI, Clelland CL, Wray S, Duff KE. 2016. Neuronal activity enhances tau propagation and tau pathology in vivo. Nature Neuroscience 19:1085–1092. DOI: https:// doi.org/10.1038/nn.4328, PMID: 27322420 Yamada K, Holth JK, Liao F, Stewart FR, Mahan TE, Jiang H, Cirrito JR, Patel TK, Hochgra¨ fe K, Mandelkow EM, Holtzman DM. 2014. Neuronal activity regulates extracellular tau in vivo. Journal of Experimental Medicine 211:387–393. DOI: https://doi.org/10.1084/jem.20131685, PMID: 24534188 Yan XX, Cai Y, Shelton J, Deng SH, Luo XG, Oddo S, Laferla FM, Cai H, Rose GM, Patrylo PR. 2012. Chronic temporal lobe epilepsy is associated with enhanced Alzheimer-like neuropathology in 3(cid:2)Tg-AD mice. PLOS ONE 7:e48782. DOI: https://doi.org/10.1371/journal.pone.0048782, PMID: 23155407 Zanier ER, Bertani I, Sammali E, Pischiutta F, Chiaravalloti MA, Vegliante G, Masone A, Corbelli A, Smith DH, Menon DK, Stocchetti N, Fiordaliso F, De Simoni M-G, Stewart W, Chiesa R. 2018. Induction of a transmissible tau pathology by traumatic brain injury. Brain 127:2685–2699. DOI: https://doi.org/10.1093/brain/awy193 Alyenbaawi et al. eLife 2021;10:e58744. DOI: https://doi.org/10.7554/eLife.58744 33 of 33
10.1016_j.ejca.2022.10.026
European Journal of Cancer 178 (2023) 128e138 Available online at www.sciencedirect.com ScienceDirect j o u r n a l h o m e p a g e : w w w . e j c a n c e r . c o m Original Research Content validity of the EORTC quality of life questionnaire QLQ-C30 for use in cancer Kim Cocks a,*, Jane R. Wells b, Colin Johnson c, Heike Schmidt d, Michael Koller e, Simone Oerlemans f, Galina Velikova g, Monica Pinto h, Krzysztof A. Tomaszewski i, Neil K. Aaronson j, Elizabeth Exall b, Chelsea Finbow b, Deborah Fitzsimmons k, Laura Grant a, Mogens Groenvold l,m, Chloe Tolley a, Sally Wheelwright n, Andrew Bottomley o On behalf of the European Organisation for Research and Treatment of Cancer (EORTC) Quality of Life Group a Adelphi Values, Patient-Centered Outcomes, Bollington, Cheshire, UK b Formerly of Adelphi Values, Patient-Centered Outcomes, Bollington, Cheshire, UK c Surgical Unit, Faculty of Medicine, University of Southampton, Southampton, UK d Institute of Health and Nursing Science and University Clinic and Outpatient Clinic for Radiotherapy, Medical Faculty, Martin Luther University Halle Wittenberg, Germany e Center for Clinical Studies, University Hospital Regensburg, Regensburg, Germany f Netherlands Comprehensive Cancer Organisation (IKNL), Utrecht, the Netherlands g Leeds Institute of Medical Research at St James’s, University of Leeds, St James’s University Hospital, Leeds Cancer Centre, Leeds Teaching Hospitals NHS Trust, St James’s University Hospital, Leeds, UK h Strategic Health Services Department, Istituto Nazionale Tumori e IRCCS- Fondazione G Pascale, Napoli, Italy i Faculty of Medicine and Health Sciences, Andrzej Frycz Modrzewski Krakow University, Krakow, Poland j Division of Psychosocial Research & Epidemiology, The Netherlands Cancer Institute, Amsterdam, the Netherlands k University of Swansea, Wales, UK l The Palliative Care Research Unit, Department of Geriatrics and Palliative Medicine, Bispebjerg and Frederiksberg Hospital, Copenhagen, Denmark m Department of Public Health, University of Copenhagen, Copenhagen, Denmark n Brighton and Sussex Medical School, University of Sussex, Brighton, UK o Quality of Life Department, European Organisation for Research and Treatment of Cancer, Brussels, Belgium Received 4 July 2022; received in revised form 24 October 2022; accepted 25 October 2022 Available online 1 November 2022 * Corresponding author. E-mail address: kim.cocks@adelphivalues.com (K. Cocks). @drkimcocks (K. Cocks) https://doi.org/10.1016/j.ejca.2022.10.026 0959-8049/ª 2022 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http:// creativecommons.org/licenses/by-nc-nd/4.0/). K. Cocks et al. / European Journal of Cancer 178 (2023) 128e138 129 Abstract Aim: The European Organisation for Research and Treatment of Cancer Quality of Life Questionnaire Core 30 (QLQ-C30) is among the most widely used patient-reported outcome measures in cancer research and practice. It was developed prior to guidance that content should be established directly from patients to confirm it measures concepts of interest and is appropriate and comprehensive for the intended population. This study evaluated the content validity of the QLQ-C30 for use with cancer patients. Methods: Adults undergoing cancer treatment in Europe and the USA participated in open- ended concept elicitation interviews regarding their functional health, symptoms, side- effects and impacts on health-related quality of life. Thematic analysis was conducted, and similarities across cancer types, disease stages and countries or languages were explored. Results: Interviews with 113 patients with cancer (85 European, 28 USA) including breast, lung, prostate, colorectal and other cancers were conducted between 2016 and 2020. Concep- tual saturation was achieved. The most frequently reported concepts were included in the QLQ-C30 conceptual framework. QLQ-C30 items were widely understood across language versions and were relevant to patients across cancer types and disease stages. While several new concepts were elicited such as difficulty climbing steps or stairs, weight loss, skin problems and numbness, many were not widely experienced and/or could be considered sub-concepts of existing concepts. Conclusions: The QLQ-C30 demonstrates good evidence of content validity for the assessment of functional health, symptom burden and health-related quality of life in patients with localised-to-advanced cancer. ª 2022 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). KEYWORDS Quality of life; Neoplasms; Interviews 1. Introduction The European Organisation for Research and Treat- ment of Cancer (EORTC) Quality of Life Questionnaire Core 30 (QLQ-C30) is a patient-reported outcome measure (PROM) designed to assess the functional health, symptom burden and health-related quality of life (HRQoL) of patients with localised-to-advanced cancer. The original version was published in 1993 [1]. The current version (QLQ-C30 v3.0) contains 30 items and has been used since 1997 [2], translated and vali- dated in over 120 languages and used in more than 5000 studies worldwide. The QLQ-C30 (https://qol.eortc.org/) scales assessing includes 15 scales: five functional emotional physical functioning, social functioning; nine multi- and single-item scales assess- ing fatigue, nausea and vomiting, pain, dyspnoea, insomnia, appetite loss, constipation, diarrhoea, finan- cial difficulties and a global health status/QoL scale. The EORTC measurement model supplements the QLQ-C30 with disease- or treatment-specific modules (e.g. for breast cancer) [3,4]. functioning, functioning, functioning cognitive role and Despite wide acceptance of the QLQ-C30, evidence of content validity specifically for use in patients with cancer has not been demonstrated (though all recently developed disease-specific modules debriefed in patient interviews have included the QLQ-C30 concepts). Demonstrating a PROM’s content validity involves generating qualitative evidence from patients that the questionnaire framework and items are appropriate and the QLQ-C30’s conceptual comprehensive relative to its intended measurement concept, population and context of use [5]. During framework development, and items were based on oncologist and researcher consensus [4,6], before undergoing validity and reli- ability testing with patients [7]. However, since initial development, more detailed guidelines and standards have been established for demonstrating PROMs’ con- tent validity, to inform measurement concepts and acceptability of item and instruction wording [8e11]. focussing on patient input of the health, functional To provide evidence of the QLQ-C30’s suitability for use with contemporary cancer patients, qualitative exploration symptom burden and HRQoL impact of cancer is required. The extent to which this experience is consistent between patients from populations that differ in terms of de- mographic and clinical characteristics, and geographic locations must also be understood. Further, evidence of patients’ understanding and relevance of QLQ-C30 items in these populations is needed. This research aimed to explore the extent to which the QLQ-C30 is an appropriate and comprehensive measure of functional health, symptom burden and HRQoL for use with cancer patients across disease stages. Due to the exten- sive use of the QLQ-C30 Version 3, the objective was not only to provide current evidence for the content validity in order to support its ongoing use but also to highlight any gaps that need to be considered when including the QLQ-C30 in a study due to the expanded portfolio of novel cancer treatments. If the study had 130 K. Cocks et al. / European Journal of Cancer 178 (2023) 128e138 found significant problems with the understanding or relevance of the items in the QLQ-C30, recommenda- tions would have been made to identify alternative wording to improve reliability. The qualitative analysis was conducted by a team outside the EORTC Quality of Life Group (QLG), who specialise in PROM develop- ment and validation. 2. Material and methods This qualitative interview study was conducted with patients undergoing treatment across a range of cancer types and stages in Europe and the USA. The objectives of this research were: 1. to identify the functional health, symptom burden, and HRQoL concepts that impacted by cancer and treatment; 2. to use these findings to determine how appropriate the QLQ-C30 is for the target population; and 3. to evaluate from the patient perspective, the extent to which the domains and items of the QLQ-C30 are understood and perceived as relevant. National or local ethical approval was sought as required prior to the conduct of any study activities with patients. All patients provided written informed consent before participating. The study collaborators met bian- nually at the EORTC Quality of Life Group meetings during the project to discuss progress and emerging re- sults. The Adelphi Values team responsible for conduct of some of the interviews and the full qualitative analysis were not, and are not, members of the EORTC Quality of Life Group (QLG). 2.1. Patient sample Eligible patients were at least 18 years of age with a clinician-confirmed diagnosis of cancer and were receiving cancer treatment for cancer (including surgical treatment within the previous three months). Patients had to be literate and fluent in the local language and willing to participate in an interview. Complete eligi- bility criteria are provided in Supplementary File 1. A purposive sampling approach with recruitment quotas ensured a diverse sample in terms of demographic and clinical characteristics and geographic location. A minimum of 20 patients with each of the four most common types of cancer (breast, lung, prostate and colorectal) and 40 patients with other cancers were tar- geted for recruitment. Patients were also sampled ac- cording to age, disease severity and sex. Patients were recruited from multiple locations in the US, Germany (two sites) and England (two sites), and one site each in Italy, the Netherlands and Poland. Given the relatively small sample size per site, the local investigators aimed to recruit a range of patients across these groups, rather than aiming to meet specific quotas per site. In Europe, patients were identified by the local investigator at each recruiting site and approached by a research team member to determine whether they were interested in participating. In the US, patients were identified by a third-party recruitment agency working with selected clinicians/oncologists. The clinicians/on- cologists approached potentially relevant patients to determine whether they were interested in participating. No record of refusal to participate was recorded, and no patients withdrew from the study after agreeing to participate. Adequacy of sample size was assessed through concept saturation [12]. Once all interviews had been conducted, the interviews were grouped into six sets of 18 or 19 transcripts. The concepts elicited in each group were compared with the next group and saturation was considered achieved if no new concepts relevant to the research question were elicited in the final set of tran- scripts. All concepts were elicited in the first or second set of interviews. 2.2. Patient interviews A semi-structured interview guide was developed in English, with expert clinical input, and translated into the local language where appropriate. The guide was developed in a way that reduced the potential for confirmation bias (the tendency to elicit or interpret information consistent with existing hypotheses). The interview guide is provided in Supplementary File 2. In Europe, interviews were conducted face-to-face by investigator or research team member. In- the local terviewers underwent a comprehensive study briefing and were advised to cover as much of the interview guide as possible. In the US, interviews were conducted via telephone [13] by research team members (JW, EE) who are experienced qualitative researchers based in the UK. All interviews were conducted in the local language and were intended to last approximately 60 min. The interview started with open-ended concept elicitation to promote spontaneous discussion about disease- and treatment-related symptoms and how those symptoms impacted patients’ lives. Subsequently, there was targeted (probed) discussion about the 15 QLQ-C30 domains to ensure all concepts included in the instrument were explored in detail. Finally, a structured cognitive debriefing of the paper version of the QLQ-C30 explored patients’ understanding of each item, the relevance of the item to their illness experience and appropriateness of the recall period, instructions and response options. All interviews were audio- recorded. 2.3. Qualitative analysis Content validity of the QLQ-C30 was assessed through qualitative analysis of the interview data. Audio files K. Cocks et al. / European Journal of Cancer 178 (2023) 128e138 131 were transcribed verbatim, de-identified and translated to English where relevant. Transcript analysis was first conducted by the local investigator at each study site using a thematic framework approach, which is used to classify and organise data according to key themes, concepts and emergent categories [14]. The framework consisted of an excel spreadsheet listing the existing QLQ-C30 item concepts and items. Analysis was based on the existing QLQ-C30 conceptual framework, which grouped items into concepts and was used as a basis for structuring the interview guide. To map issues identified during the interviews onto the concepts and items in the framework, interview quotes relating to each concept or item were placed in the relevant section. Symptoms were grouped at concept level, which related to the existing QLQ-C30 conceptual framework and then sub-concepts were added based on information elicited from the in- terviews. New emerging concepts not included in the QLQ-C30 were added to the framework. Research team members (JW, EE, CF) collated the locally completed frameworks. To ensure the framework was consistently and appropriately interpreted, central quality control was conducted by JW, EE and CF whereby an initial three transcripts were cross-checked with the framework for the accuracy of completion. Further transcripts were checked if the initial check identified inaccuracies or inconsistencies. Updates were made as required in consultation with the local site. Finally, the number of patients who reported each issue was counted. When analysing the concept elicitation data, a concept was considered elicited if the patient stated they personally experienced it. In the conceptual mapping exercise, concepts elicited by at least 10 patients were reviewed to determine if they are assessed in the QLQ- C30 or could be considered a sub-concept of a QLQ-C30 concept. An item or instruction was considered understood if the patient supplied explicit verbal evidence of under- standing, including rewording the item or instruction in their own words, supplying an appropriate example or coherently answering a follow-up probe. An item was considered relevant if the patients indi- cated they would select ‘A little’, ‘Quite a bit’ or ‘Very much’ from the response scale, or explicitly stated they had experienced the concept assessed by the item since their cancer diagnosis. Sub-group analyses were con- ducted by cancer site and disease stage to determine if each item concept was relevant across these sub-groups. An item was deemed relevant to a sub-group if one or more patients from that group reported the item was relevant to their experience. Recall period was considered sufficiently understood if the patient reported they were thinking back to the appropriate timeframe. For most items, this was during the past week, but for the physical functioning domain items, current ability is reported (e.g. ‘Do you have any trouble taking a long walk?’). Adequacy of the response scale was assessed by asking patients how easy or diffi- cult it was to select a response using the response scale. Patients were also asked for their opinion on the length of the QLQ-C30. 3. Results 3.1. Patient sample 113 interviews were conducted, and interview length ranged from 20 to 90 min. The sample had approxi- mately equal proportions of patients with a primary cancer diagnosis of breast (n Z 19, 17%), lung (n Z 19, 17%), prostate (n Z 19, 17%) and colorectal (n Z 15, 13%) cancer. The sample included a proportion of pa- tients with haematological cancer (n Z 12, 11%), skin cancer (n Z 8, 7%) and other cancer types/sites (n Z 21, 19%) such as head and neck, and ovarian. Patients’ ages ranged from 23 to 89 years (mean 63.5 years), and there were a similar number of males (n Z 62, 55%) and females (n Z 51, 45%) in the sample. Most patients had been diagnosed with cancer for one year or less (n Z 79, 70%) and were approximately equally spread across cancer stage: metastatic (n Z 43, 38%), locally advanced (n Z 37, 33%) and localised (n Z 28, 25%). Most patients had an Eastern Cooper- ative Oncology Group status of 0 or 1 (n Z 92, 81%) and over half of the patients had a chronic comorbid condition in addition to cancer (n Z 71, 63%). Infor- mation relating to treatments was collected from clini- cians, with a focus on clinical treatment types rather than supportive treatments. Further clinical character- istics of the patient sample can be found in Table 1. Patients’ current active treatment was mainly systemic therapy (single agent or combination therapy; n Z 71, 63%). Details of the sample’s current and past treat- ments can be found in Table 2. 3.2. Patient interview findings All concepts reported by 10 or more patients were reported within the first two sets of transcripts. Data are supportive of concept saturation having been achieved across the total sample. 3.2.1. Core concepts associated with HRQoL of cancer patients and QLQ-C30 concept mapping Findings from 112 interviews (the concept elicitation recording for one patient became corrupted) are pre- sented. Patients discussed their experience of cancer, including symptoms, treatment side-effects and impacts on HRQoL. Concepts discussed by at three patients are illustrated in Fig. 1. A further 26 concepts were spontaneously reported by one or two patients each. least Concept mapping to the QLQ-C30 showed that the 13 concepts most frequently elicited are covered by the 132 K. Cocks et al. / European Journal of Cancer 178 (2023) 128e138 Table 1 Patient demographic and clinical characteristics (N Z 113). Description Poland (n Z 5) 63.2 (53e72) Italy (n Z 10) 62.9 (23e80) Age, average (range) Sex, n (%) Male Female Primary cancer diagnosis, n (%) Breast Lung Prostate Colorectal Haematological Multiple myeloma (Non) Hodgkin lymphoma Hodgkin lymphoma Skin Melanoma Other Skin Head and Neck Ovarian Oesophageal Cervical Pancreatic Bone Other cancer Salivary gland (adenoid cystic carcinoma) Cancer of unknown primary (CUP) Leiomyosarcoma of the scrotum Adrenal carcinoma Cardia (stomach) carcinoma Time since cancer diagnosis, n (%) 0e6 months 7 months to 1 year 2e5 years 6þ years Current disease stage, n (%) Metastatic Locally advanced Localised Remission Missing Data/Not applicable ECOG status, n (%) 0 1 2 3 4 4 (80) 1 (20) 1 (20) 0 3 (60) 1 (20) 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 4 (80) 1 (20) 0 0 1 (20) 3 (60) 1 (20) 0 0 1 (20) 2 (40) 1 (20) 1 (20) 0 6 (60) 4 (40) 2 (20) 0 1 (10) 1 (10) 0 1 (10) 0 2 (20) 0 1 (10) 0 0 0 0 1 (10) 1 (10) 0 0 0 0 2 (20) 4 (40) 2 (20) 2 (20) 4 (40) 5 (50) 1 (10) 0 0 1 (10) 5 (50) 2 (20) 2 (20) 0 Netherlands (n Z 16) 60.9 (30e80) Germany (n Z 18) 61.4 (24e79) US (n Z 28) 62.1 (42e89) UK (n Z 36) 66.8 (45e84) Total (n Z 113) 63.5 (23e89) 9 (56) 7 (44) 3 (19) 3 (19) 3 (19) 2 (13) 0 1 (6) 1 (6) 0 0 0 1 (6) 0 1 (6) 0 0 0 0 0 0 1 (6) 3 (19) 8 (50) 2 (13) 3 (19) 11 (69) 3 (19) 2 (13) 0 0 7 (44) 3 (19) 0 6 (38) 0 11 (61) 7 (39) 1 (6) 5 (28) 0 0 4 (22) 0 1 (6) 0 0 3 (17) 0 0 0 0 1 (6) 0 1 (6) 1 (6) 1 (6) 0 6 (33) 1 (6) 8 (44) 3 (17) 9 (50) 2 (11) 3 (17) 0 4 (22) 0 16 (89) 2 (11) 0 0 7 (25) 21 (75) 4 (14) 5 (18) 5 (18) 5 (18) 2 (7) 2 (7) 0 0 2 (7) 0 2 (7) 0 1 (4) 0 0 0 0 0 0 0 15 (54) 4 (14) 4 (14) 5 (18) 7 (25) 10 (36) 10 (36) 1 (4) 0 10 (36) 14 (50) 3 (11) 1 (4) 0 25 (69) 11 (31) 8 (22) 6 (17) 7 (19) 6 (17) 0 0 0 4 (11) 0 0 0 3 (8) 0 2 (6) 0 0 0 0 0 0 18 (50) 13 (36) 2 (6) 3 (8) 11 (31) 14 (39) 11 (31) 0 0 20 (56) 13 (36) 3 (8) 0 0 62 (55) 51 (45) 19 (17) 19 (17) 19 (17) 15 (13) 6 (5) 4 (4) 2 (2) 6 (5) 2 (2) 4 (4) 3 (3) 3 (3) 2 (2) 2 (2) 2 (2) 1 (1) 1 (1) 1 (1) 1 (1) 1 (1) 48 (43) 31 (27) 18 (16) 16 (14) 43 (38) 37 (33) 28 (25) 1 (1) 4 (4) 39 (35) 53 (47) 11 (10) 10 (9) 0 QLQ-C30: feeling tired (n Z 99), pain (n Z 94), impact on sleep (n Z 86), need for rest (n Z 73), impact on hobbies and leisure activities (n Z 78), impact on ac- tivities around the house (n Z 56), nausea (n Z 62), ability to walk long distances (n Z 79), impact on appetite (n Z 79), worry (n Z 75), weakness (n Z 65), diarrhoea (n Z 60) and impact on work (n Z 51). New concepts reported by 10 or more patients that were considered distinct from those included in the QLQ-C30 included climbing steps or stairs (n Z 37), weight loss (n Z 20), skin problems (n Z 19), numbness (n Z 15), issues with urination (n Z 13), hair loss or thinning (n Z 12), cough or phlegm (n Z 12), swelling (n Z 12), dizziness or faintness (n Z 10), weight gain (n Z 10) and feeling more emotional (n Z 10). Impact on sexual functioning was reported by seven patients. Concepts reported by fewer than three patients and conceptual mapping are provided in Supplementary File 3. Patients were asked if there were any concepts or items missing from the questionnaire. Only seven con- cepts were identified as missing by a maximum of four K. Cocks et al. / European Journal of Cancer 178 (2023) 128e138 133 Table 2 Patient current and past treatments (N Z 113). Description Poland (n Z 5) Italy (n Z 10) Netherlands (n Z 16) Germany (n Z 18) US (n Z 28) UK (n Z 36) Total (n Z 113) Current active treatment, n (%) Surgery Radiation Systemic therapyesingle agent Systemic therapyecombination Other treatment 3 (60) 2 (40) 0 1 (20) 4 (40) 0 3 (30) 3 (30) Immunotherapy (unspecified) Immunotherapy (Pembrolizumab) Palbociclib and Fulvestrant Targeted therapy (Dabrafenib/Trametinib) 0 0 0 0 Treatment received in the past, n (%) Surgery Systemic therapyecombination Radiation Systemic therapyesingle agent Hormonal Other treatment Ibandrovate and Vexemetasane Stem cell transplant 4 (80) 0 1 (20) 0 0 0 0 0 0 0 0 7 (70) 5 (50) 4 (40) 1 (10) 0 0 0 0 0 4 (25) 7 (44) 1 (6) 0 0 0 7 (44) 9 (56) 3 (19) 2 (13) 0 0 0 0 3 (17) 5 (28) 5 (28) 0 0 0 0 9 (50) 10 (56) 5 (28) 5 (28) 0 0 0 3 (11) 3 (11) 18 (64) 6 (21) 0 0 1 (4) 0 14 (50) 4 (14) 11 (39) 5 (18) 1 (4) 0 2 (7) 0 4 (11) 19 (53) 13 (36) 0 3 (8) 0 1 (3) 17 (47) 5 (14) 5 (14) 6 (17) 7 (19) 1 (3) 0 10 (9) 12 (11) 49 (43) 35 (31) 1 (1) 3 (3) 1 (1) 1 (1) 58 (51) 33 (29) 29 (26) 19 (17) 8 (7) 1 (1) 2 (2) patients each (impact on sexual functioning, issues with relationship with urination, healthcare professionals, dizziness and swelling). side effects, treatment 3.2.2. Cognitive debriefing Patients (N Z 112; debriefing data not available for one patient who chose to end the interview early) shared their feedback on the QLQ-C30. Due to various factors (e.g. time constraints), not all patients were asked all debriefing questions. Findings are reported out of the number of patients who were asked each debriefing question. 3.2.3. Item relevance Item relevance ranged from 45% (vomiting) to 96% (tired), with an average of 78% and median of 80%. The proportion of patients reporting each item as relevant to their experience is detailed in Fig. 2. Sub-group analysis to examine the relevance of QLQ- C30 items across the largest cancer-site groups (breast, lung, prostate, colorectal, haematological and skin) demonstrated that all but three items were relevant across all sub-groups, with at least one patient from each sub-group indicating the item was relevant. Items assessing lack of appetite, vomiting and diarrhoea were found to be relevant in all groups except skin cancer. However, while no patients with skin cancer reported that these QLQ-C30 items were relevant in the cognitive debriefing section, a small number of these patients re- ported during the initial concept elicitation discussion that they had experienced a lack of appetite (n Z 2) and diarrhoea (n Z 1), confirming the relevance of all but the concept of vomiting in this population. Sub-group analysis by disease stage (localised, locally advanced and metastatic) demonstrated that all QLQ-C30 items were relevant to patients across all sub-groups (relevant to between 23% and 92% of patients in each group). 3.2.4. Item understanding For each QLQ-C30 item, at least 90% of patients asked demonstrated understanding. A similar proportion of patients demonstrated understanding across the lan- guage versions tested. 3.2.5. General findings Almost all patients asked (n Z 54/57, 95%) demon- strated understanding of the QLQ-C30 instructions, providing evidence that the instructions are sufficiently clearly worded and appropriate. No patients suggested rewording the instructions. Regarding the recall periods, items 1 to 5 (the phys- ical functioning domain) have no specified recall period. Items 6 to 30 specify a seven-day recall period. Over half of the patients asked (n Z 60/104, 58%) indicated that they adhered to the recall period, while 37% (n Z 38/ 104) indicated they did not adhere to the recall period. For six patients asked, it was unclear whether they were adherent to the recall period due to the lack of infor- mation available. Most patients who were asked (n Z 94/107, 88%) re- ported that it was easy to select responses to the QLQ-C30 items. A small number of patients (n Z 12/107, 11%) re- ported that it was somewhat difficult to select responses. The reasons provided were in relation to item content, item wording or finding it difficult to choose a response. Most patients asked (n Z 89/102, 87%) indicated that the length of the questionnaire is appropriate. A small number of patients (n Z 8/102, 8%) responded with respect to the duration of the interview rather than the questionnaire. Two patients (n Z 2/102, 2%) felt the 134 K. Cocks et al. / European Journal of Cancer 178 (2023) 128e138 w e i v r e t n i n i d e t r o p e r s t p e c n o C Tired Pain Need for rest Hobbies and Leisure Around the house Nausea Walking long distances Appe(cid:2)te Diarrhoea Work Weak Worried Sleep Climbing steps/stairs Shortness of breath Family life Outside of the home Strenuous ac(cid:2)vi(cid:2)es Walking short distances Cons(cid:2)pa(cid:2)on Depressed Lack of energy Taste Social rela(cid:2)onships Weight loss Anxiety Skin problems Vomi(cid:2)ng Numbness Sadness Concentra(cid:2)on Excericise and fitness Issues with urina(cid:2)on Angry Swelling 40 38 37 37 37 36 35 33 32 32 31 29 29 74 70 58 51 5 4 12 24 4 2 22 34 27 26 25 25 25 24 23 20 20 19 19 1 15 15 14 14 13 13 12 0 20 40 16 24 23 30 15 42 42 42 54 40 36 39 31 30 39 42 2 15 27 11 24 11 37 25 37 45 16 25 8 5 5 13 23 13 17 17 8 35 71 37 32 37 11 13 6 37 87 84 92 90 93 58 46 18 24 19 36 22 36 40 40 38 15 16 26 4 5 38 9 13 18 97 97 97 99 99 100 60 w e i v r e t n i n i d e t r o p e r s t p e c n o C Cough/phlegm Scared/fearful Frustrated Hair loss/thinning Irritable Memory Weight gain Stomach bloa(cid:2)ng Dizziness/faitness More emo(cid:2)onal Shock Self care Acid reflux/indeges(cid:2)on Sexual func(cid:2)oning Self-image Muscle weakness Financial impact Tense Bowel irregulari(cid:2)es Mouth blisters Vision problems Ea(cid:2)ng/Swallowing Stress Reduced immune response Embarrassment Lack of mo(cid:2)va(cid:2)on Dry mouth/lips General psychological impacts Bending down Annoyed Loss of autonomy Change in body temperature Voice loss Muscle cramp/spasm Standing Guilt/shame 3 12 12 12 12 11 10 10 10 10 10 9 8 7 7 7 6 6 6 6 6 6 6 6 5 5 5 4 4 4 4 4 3 3 3 3 3 50 43 21 53 100 100 100 100 102 102 102 102 103 105 105 105 103 80 100 27 44 24 15 30 27 5 50 40 39 51 106 106 106 106 106 107 107 107 108 108 108 108 108 109 109 109 109 109 0 20 40 60 Number of pa(cid:2)ents (N=112) 80 100 Spontaneous - included in QLQ-C30 Probed - not included in QLQ-C30 Probed - included in QLQ-C30 Not experienced Spontaneous - not included in QLQ-C30 Not asked/Not clear Note: Some concept names have been shortened (e.g. ‘hobbies and leisure’ refers to concept ‘Impact on ability to do hobbies and leisure ac(cid:2)vi(cid:2)es’) Fig. 1. Concepts reported by (cid:2) 3 patients in concept elicitation interviews (N Z 112). K. Cocks et al. / European Journal of Cancer 178 (2023) 128e138 135 Fig. 2. Proportion of patients reporting QLQ-C30 items as relevant to their experience of cancer (N Z 112). questionnaire was too long, saying they would have preferred it to consist of 15e20 items. One patient felt some items were repetitive and another felt some ques- tions could be merged. Finally, one patient felt the questionnaire length was ‘just about right’ but reported struggling to think of answers towards the end. 4. Discussion Qualitative interview results from 113 patients with cancer from Europe and the US showed that concepts included in the QLQ-C30 are widely understood across language versions, and that existing items are relevant to patients across cancer types and disease stages. In this study sample, the 13 most frequently, spontaneously elicited concepts were already covered by the QLQ-C30 conceptual framework. The use of PROMs such as the QLQ-C30 in oncology clinical trials is well-established and there is growing evidence for their utility in enhancing clinical practice [15,16]. Using PROMs for routine monitoring has been shown to improve communication, clinician awareness of symptoms and symptom management, alongside improving quality of life and even survival [15,17e21]. To remain credible and support medical product eval- uation by regulators, PROMs need to show evidence that they comprehensively cover concepts that are important and relevant to the target patient population, known as content validity [10]. Guidelines for establishing content validity are focussed on eliciting concepts from patients via in- terviews and assessing patient understanding of the developed PROM using cognitive debriefing interviews. Although the QLQ-C30 has been extensively used in patients with cancer for almost 30 years, it is pertinent to evaluate content validity using current recommended methods, ensuring the concepts assessed are important and relevant to patients with cancer today and supplying evidence documenting this. Due to the widespread use of the QLQ-C30, recommending changes could have been challenging yet would have been necessary to confirm content validity. It is important to recall that the EORTC measure- ment model recommends the QLQ-C30 be used along- side disease- or symptom-specific modules or standalone questionnaires to cover additional concepts important 136 K. Cocks et al. / European Journal of Cancer 178 (2023) 128e138 and relevant to specific populations. Items from the EORTC Item Library can be added if required concepts are not covered by the selected questionnaires [22]. A short list of concepts that were considered distinct from those included in the QLQ-C30 was reported by a small proportion of patients in the open-ended concept elicitation discussion. These concepts generally fit within existing QLQ-C30 conceptual domains (e.g. additional symptoms or impacts on side-effects and further emotional functioning), except for impact on sexual functioning that is not currently covered within the QLQ-C30. If required, the EORTC suite of resources can be used alongside the QLQ-C30 to support the assessment of these concepts. Since the QLQ-C30 needs to strike a balance between patient burden and ensuring relevance across a range of patients with cancer, no new generic items were recommended for inclusion but sup- plementing the core questionnaire with additional items or scales may be an appropriate strategy depending on likely importance and relevance in the specific target population. In addition to the findings from this study, the QLQ- C30 encompasses 11 of the 12 concepts recommended by Reeve et al. as the core set of symptoms to measure in adult oncology clinical trials [23]. The remaining concept, neuropathy (referred to as numbness in this study), was reported by a small proportion of the sample (n Z 15, 13%) and was not considered prevalent enough to warrant adding the concept to the QLQ-C30. For specific pop- ulations where this concept may be more relevant, the EORTC measurement model allows for the inclusion of a chemotherapy-induced peripheral neuropathy module, the QLQ-CIPN20 alongside the QLQ-C30 [24], or for the selection of one or more items from the Item Library. To assess the relevance of items across cancer types, whether a concept was elicited within that patient subgroup was evaluated. All QLQ-C30 concepts were relevant in all but skin cancer, in which vomiting was not part of the patients’ disease experience (though nausea was relevant). Though we had a small threshold for the relevance of at least one patient within the group indicating an item/concept as relevant, given the research question asks for a binary response (is the item relevant in this group or not?), this threshold was deemed suitable. No QLQ-C30 concepts were deemed irrelevant. The item assessing vomiting had the lowest proportion of patients reporting relevance, yet nearly 50% of patients across the sample endorsed this item. This may in part reflect the sample characteristics and their current treatments, rather than indicating it as a symptom that does not warrant inclusion. Reported adherence to recall periods was acceptable, though adding the recall period stated alongside each item could be considered to improve clarity. This aligns with how the questionnaire may be delivered electroni- cally. Recent research highlighted that for the PROMIS physical function scores, the recall period (no recall period versus 24 h versus 7 days) did not impact scores [25]; therefore, the variable adherence to recall period in our study was not considered a concern for content validity. 5. Limitations Although the study was multi-national and multi- lingual, there was a high proportion of English- speaking patients in the sample. Further, US patients were recruited and interviewed differently to the Euro- pean patients, which may have led to differences between the US and European samples. The US patients were interviewed via telephone, whereas patients from all other countries were interviewed face-to-face. How- ever, there is little evidence of difference in data quality when conducting interviews in these different ways [13]. As this is a qualitative study, it reflects the specific views of the patients interviewed and may not be fully representative of, or transferable to, the target patient population. A typical cancer population would consist of many different cancer sites and stages, but to capture this in a study, sample would require a much larger sample size and a more targeted recruitment approach than is typically feasible for a qualitative study. Any sampling strategy would struggle to achieve complete representation across all possible patient and disease characteristics, though in this study, many groups were represented and the consistency in findings between groups is supportive of the EORTC model to administer the QLQ-C30 with a disease/treatment module. As is common for studies in patients with cancer, younger patients and those with more limitations (Eastern Cooperative Oncology Group scores of 2 or more) may be under-represented. Additionally, few participants relative to the total sample were reported to have been previously or currently undergoing immunotherapy or targeted therapies. However, these newer therapies were available to patients at study recruitment sites and, in some instances, were standard of care. In our study, patients on these therapies were most often categorised on the case report form as being on systemic therapy; as such, the treatment name or type was not captured. Despite these limitations, the study comprised a large sample of international patients, including the most common cancer sites and widely used treatments. 6. Conclusions Overall, this study demonstrates that the existing QLQ- C30 conceptual framework is appropriate in a large, international patient sample with different cancer sites and stages. The questionnaire items and subscales are relevant and important to patients across a range of cancers, disease stages and treatments. The EORTC measurement model recommends using the QLQ-C30 K. Cocks et al. / European Journal of Cancer 178 (2023) 128e138 137 alongside supplementary disease- or symptom-specific modules, standalone questionnaires or items from the Item Library. The QLQ-C30 demonstrates good evi- dence of content validity for use with patients with localised-to-advanced cancer. All authors meet all four criteria for authorship ac- cording to the ICMJE recommendations, and confirm they have full access to the study data and accept responsibility to submit for publication. Funding This work was supported by the EORTC Quality of Life Group [Grant number 006-2015]. Data sharing The study protocol and aggregated and anonymised data that underlie the results reported in this article will be available immediately following publication with no end date to researchers who provide a methodologically- sound proposal to achieve aims related to the approved proposal. Proposals should be submitted to www.eortc. be/services/forms/erp/request.aspx. To gain access, data requestors will need to sign a data access agreement. Author contributions Author Kim Cocks contributed through study con- ceptualisation, investigation, methodology, analysis, and writing the original draft manuscript. acquisition, funding Author Jane R Wells contributed through data curation, formal analysis, methodology, project admin- istration, validation, and writing the original draft manuscript. Author Colin Johnson contributed through study investigation, funding acquisition, conceptualisation, methodology, and review and editing the manuscript. Authors Michael Koller, Heike Schmidt, Simone Oer- lemans, Galina Velikova, Monica Pinto, and Krzysztof Tomaszewski contributed through investigation, meth- odology, and review and editing the manuscript. Authors Elizabeth Exall and Chelsea Finbow contributed through data curation, formal analysis, project administration, validation, and writing the original draft manuscript. Author Laura Grant contributed through method- ology, project administration, and review and editing the manuscript. Authors Neil Aaronson, Deborah Fitzsimmons, Mogens Groenvold, and Sally Wheelwright contributed through study conceptualisation, methodology, and review and editing the manuscript. Author Chloe Tolley contributed through study investigation, funding acquisition, conceptualisation, methodology, and review and editing the manuscript. Author Andrew Bottomley contributed through study conceptualisation, methodology, funding acquisi- tion, and review and editing the manuscript. Conflict of interest statement The authors declare the following financial interests/ personal relationships which may be considered as po- tential competing interests: The following authors declare no competing interests: Andrew Bottomley, Neil K Aaronson, Deborah Fitz- simmons, Mogens Groenvold, and Sally Wheelwright. Colin Johnson, Krzysztof Tomaszewski, Michael Koller, Monica Pinto, Simone Oerlemans, and Heike Schmidt’s respective institutions were provided funding from the EORTC Quality of Life Group to conduct the research described in this manuscript. Jane R Wells, Chelsea Finbow, Elizabeth Exall, Chloe Tolley, and Laura Grant are or were at the time of the study employees of Adelphi Values Ltd. Adelphi Values Ltd was provided a grant from the EORTC Quality of Life Group to conduct the research described in this manuscript and to develop this manuscript. Adelphi Values Ltd provides consultancy for a variety of pharmaceutical companies. Kim Cocks is an employee of Adelphi Values Ltd. However, the role of Principal Investigator on this study was not funded through the EORTC Quality of Life Group grant. Kim Cocks has received consulting fees from Endomag Ltd. Galina Velikova’s institution was provided funding from the EORTC Quality of Life Group to conduct the research described in this manuscript and other research. Galina Velikova also declares that her insti- tution has grants/contracts with Breast Cancer Now, Pfizer, and IQVIA. Galina Velikova discloses re- lationships with Novartis, Eisai, Seattle Genetics, Sanofi Advisory Board, and Roche Ester Steering Committee. Appendix A. Supplementary data Supplementary data to this article can be found online at https://doi.org/10.1016/j.ejca.2022.10.026. References [1] Aaronson NK, Ahmedzai S, Bergman B, et al. The European Organisation for Research and Treatment of Cancer QLQ-C30: a quality-of-life instrument for use in international clinical trials in oncology. J Natl Cancer Inst 1993;85(5):365e76. [2] Bjordal K, De Graeff A, Fayers P, et al. A 12 country field study of the EORTC QLQ-C30 (version 3.0) and the head and neck cancer specific module (EORTC QLQ-H&N35) in head and neck patients. Eur J Cancer 2000;36(14):1796e807. [3] Bjelic-Radisic V, Cardoso F, Cameron D, et al. An international update of the EORTC questionnaire for assessing quality of life in 138 K. Cocks et al. / European Journal of Cancer 178 (2023) 128e138 breast cancer patients: EORTC QLQ-BR45. Ann Oncol 2020; 31(2):283e8. [4] Aaronson N, Bullinger M, Ahmedzai S. A modular approach to quality-of-life assessment in cancer clinical trials. In: Cancer clinical trials. Springer; 1988. p. 231e49. [5] Smith AB, Cocks K. Content validity and legacy patient-reported outcome measures in cancer. Qual Life Res 2015;24(7):1585e6. [6] Aaronson N, Ahmedzai S, Bullinger M, et al. The EORTC core quality of life questionnaire: interim results of an international field study. Eff Cancer Qual Life 1991:185e203. [7] Osoba D, Aaronson N, Zee B, Sprangers M, Te Velde A. Modification of the EORTC QLQ-C30 (version 2.0) based on content validity and reliability testing in large samples of patients with cancer. Qual Life Res 1997;6(2). 0-0. [8] Haynes SN, Richard D, Kubany ES. Content validity in psy- chological assessment: a functional approach to concepts and methods. Psychol Assess 1995;7(3):238. [9] US Food and Drug Administration. Guidance for Industry: patient-reported outcome measures: use in medical product development to support labeling claims. 2009. [10] Patrick DL, Burke LB, Gwaltney CJ, et al. Content val- iditydestablishing and reporting the evidence in newly developed patient-reported outcomes (PRO) instruments for medical prod- uct evaluation: ISPOR PRO Good Research Practices Task Force report: part 2dassessing respondent understanding. Value Health 2011;14(8):978e88. [11] Wheelwright Sally, Bjordal Kristen, Bottomley Andrew, et al., EORTC Quality of Life Group. EORTC Quality of Life Group guidelines for developing questionnaire modules. 5th. EORTC; 2021. [12] Guest G, Bunce A, Johnson L. How many interviews are enough? An experiment with data saturation and variability. Field Methods 2006;18(1):59e82. [13] Mazar I, Lamoureux R, Ojo O, et al. Telephone versus face-to- face interviews for patient-reported outcome instrument devel- opment. Value Health 2015;18(7):A718. [14] Gale NK, Heath G, Cameron E, Rashid S, Redwood S. Using the framework method for the analysis of qualitative data in multi- disciplinary health research. BMC Med Res Methodol 2013; 13(1):1e8. [15] Basch E, Barbera L, Kerrigan CL, Velikova G. Implementation of patient-reported outcomes in routine medical care. Am Soc Clin Oncol Educ Book 2018;vol. 38:122e34. [16] Campbell R, Ju A, King MT, Rutherford C. Perceived benefits and limitations of using patient-reported outcome measures in clinical practice with individual patients: a systematic review of qualitative studies. Qual Life Res 2021:1e24. [17] Basch E, Deal AM, Dueck AC, et al. Overall survival results of a symptom trial monitoring during routine cancer treatment. JAMA 2017;318(2): 197e8. patient-reported outcomes assessing for [18] Basch E, Deal AM, Kris MG, et al. Symptom monitoring with patient-reported outcomes during routine cancer treatment: a randomized controlled trial. J Clin Oncol 2016;34(6):557. [19] Denis F, Basch E, Septans A-L, et al. Two-year survival comparing web-based symptom monitoring vs routine surveillance following treatment for lung cancer. JAMA 2019;321(3):306e7. [20] Gilbert A, Sebag-Montefiore D, Davidson S, Velikova G. Use of patient-reported outcomes to measure symptoms and health related quality of life in the clinic. Gynecol Oncol 2015;136(3): 429e39. [21] Kotronoulas G, Kearney N, Maguire R, et al. What is the value of the routine use of patient-reported outcome measures toward improvement of patient outcomes, processes of care, and health service outcomes in cancer care? A systematic review of controlled trials. J Clin Oncol 2014;32(14):1480e510. [22] Kulis D, Bottomley A, Whittaker C, et al. The use of the EORTC item library to supplement EORTC quality of life instruments. Value Health 2017;20(9):A775. [23] Reeve BB, Mitchell SA, Dueck AC, et al. Recommended patient- reported core set of symptoms to measure in adult cancer treat- ment trials. JNCI: J Natl Cancer Inst 2014;106(7):dju129. [24] Kieffer JM, Postma TJ, van de Poll-Franse L, et al. Evaluation of the EORTC chemotherapy- the psychometric properties of induced peripheral neuropathy questionnaire (QLQ-CIPN20). Qual Life Res 2017;26(11):2999e3010. [25] Condon DM, Chapman R, Shaunfield S, et al. Does recall period matter? Comparing PROMIS(cid:2) physical function with no recall, 24-hr recall, and 7-day recall. Qual Life Res 2020;29(3): 745e53.
10.1038_s41467-022-29322-4
ARTICLE https://doi.org/10.1038/s41467-022-29322-4 OPEN Structural basis for the mechanisms of human presequence protease conformational switch and substrate recognition Wenguang G. Liang1, Juwina Wijaya2, Hui Wei David Lee Clinton S. Potter3, Bridget Carragher 6, Chang Liu6, Carla M. Koehler2, Minglei Zhao 4, John V. Lin King5, Man Pan 4 & Wei-Jen Tang 3, Jordan M. Mancl1, Swansea Mo1, 3, Alex J. Noble 3, Sheng Li 1✉ 6, ; , : ) ( 0 9 8 7 6 5 4 3 2 1 for Presequence protease (PreP), a 117 kDa mitochondrial M16C metalloprotease vital mitochondrial proteostasis, degrades presequence peptides cleaved off from nuclear- encoded proteins and other aggregation-prone peptides, such as amyloid β (Aβ). PreP structures have only been determined in a closed conformation; thus, the mechanisms of substrate binding and selectivity remain elusive. Here, we leverage advanced vitrification techniques to overcome the preferential denaturation of one of two ~55 kDa homologous domains of PreP caused by air-water interface adsorption. Thereby, we elucidate cryoEM structures of three apo-PreP open states along with Aβ- and citrate synthase presequence- bound PreP at 3.3–4.6 Å resolution. Together with integrative biophysical and pharmacolo- gical approaches, these structures reveal the key stages of the PreP catalytic cycle and how the binding of substrates or PreP inhibitor drives a rigid body motion of the protein for substrate binding and catalysis. Together, our studies provide key mechanistic insights into M16C metalloproteases for future therapeutic innovations. 1 Ben-May Department for Cancer Research, The University of Chicago, Chicago, IL, USA. 2 Department of Chemistry and Biochemistry, University of California Los Angeles, Los Angeles, CA, USA. 3 National Resource for Automated Molecular Microscopy, Simons Electron Microscopy Center, New York Structural Biology Center, New York, NY, USA. 4 Department of Medicine, University of California San Diego, La Jolla, CA, USA. 5 Department of Physiology, University of California, San Francisco, San Francisco, CA, USA. 6 Department of Biochemistry and Molecular Biology, The University of Chicago, Chicago, IL, USA. email: wtang@bsd.uchicago.edu ✉ NATURE COMMUNICATIONS | (2022) 13:1833 | https://doi.org/10.1038/s41467-022-29322-4 | www.nature.com/naturecommunications 1 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29322-4 targeting sequence) at Mitochondria are vital to cellular metabolism, home- ostasis, and stress responses1,2; their defects are linked to a plethora of neurodegenerative diseases3. Assembly of mitochondria requires the coordinated action of protein import into mitochondria, coupled with processing and proteo- lysis pathways4,5. Most imported proteins contain a presequence the (also known as a mitochondrial N-terminus4. Upon entrance into mitochondria, presequences are cleaved off by mitochondrial processing peptidase and in some cases further cleaved by mitochondrial intermediate proteases4,5. Presequences are rich in hydrophobic and positively charged residues and highly toxic to mitochondria if left to accumulate4,5. Presequence protease (PreP) is a ubiquitously expressed, 117 kDa M16C clan zinc metalloprotease that localizes to the mitochon- dria matrix and cleaves presequence peptides into non-toxic pieces4,6,7. PreP also degrades amyloid β (Aβ), the cleavage pro- duct of the amyloid precursor protein (APP) that is linked to the progression of Alzheimer’s disease and is imported into mito- chondria, particularly synaptic mitochondria6–9. A homozygous Pitrm1 (which encodes PreP) knockout in mice displays embryonic lethality10. Furthermore, genetic defects in PreP are linked with human neurological disorders, e.g., cognitive dis- ability/impairment and cerebellar atrophy10–12. disease, non-neuropathic Proteomes are maintained in a healthy state by four main processes, autophagy, chaperones, ubiquitination/proteasomal degradation, and a cohort of proteases that degrade potentially cytotoxic peptides13,14. Aggregates of such peptides include amyloidogenic species that are highly cytotoxic and are associated e.g., Alzheimer’s with human neurodegenerative diseases, and Parkinson’s systemic and amyloidoses15–17. Aβ is a key initiating factor in Alzheimer’s disease and its accumulation is caused by the imbalance between Aβ production and clearance17. Aβ is the proteolytic product of the amyloid precursor protein (APP), which belongs to a small gene family including APP-like proteins that have shared func- tions in CNS development, synapse formation, brain injury, and neuroprotection18. Only through the processing of APP, a recently evolved gene within the APP family can generate Aβ and no dedicated protease has evolved for Aβ clearance18. A handful of proteases out of 569 human proteases can effectively degrade monomeric Aβ6,19,20. Aβ-degrading proteases have a broad sub- cellular distribution, e.g., extracellular milieu (insulin-degrading enzyme (IDE), matrix metalloprotease 2 (MMP2), MMP9); mitochondria (PreP), lysosomes (cathepsin B), the plasma membrane (neprilysin (NEP) and endothelin converting enzyme 1 and 2 (ECE1/2), plasmin, IDE), and the cytosol (IDE, acyl-peptide hydrolase (APEH)). Together these enzymes enable better control of Aβ levels at all intra- and extra-cellular locations where Aβ has been detected. intracellular vesicles (IDE), The formation of aggregates of amyloid peptides such as Aβ is at least a two-step process; the first is a slow, stochastic, and reversible nucleation to form small amyloid peptide seeds fol- lowed by the elongation of seeds into large amyloid fibrils, which is faster and mostly irreversible21,22. Monomeric amyloid peptide fuels the forward progression of both steps; thus, Aβ-degrading proteases that recognize and degrade monomeric amyloid pep- tides prevent the formation of amyloid fibrils19. Of the Aβ- degrading proteases, PreP belongs to the chamber-containing protease (crypt-containing peptidase, or cryptidase) family that uses a sizable catalytic chamber to engulf, unfold, and degrade IDE (M16A their clan) and M13 metalloproteases, e.g., NEP and ECE1/223. Aβ-degrading proteases also can selectively degrade other amy- loid peptides which are highly diverse in sequence and structure. This raises a major question: how do such proteases selectively degrade amyloid peptides over non-amyloidogenic peptides? substrates23. Other cryptidases include toxic aggregates6. However, Crystallographic analyses reveal that PreP and related M16C metalloproteases are composed of ~55 kDa homologous N- and C-terminal domains (PreP-N and PreP-C, respectively), which are connected by an extended helical hairpin6,7. PreP-N and PreP-C, in the closed state of PreP, form an enclosed catalytic chamber to entrap and degrade monomeric amyloidogenic peptides, thereby the preventing the formation of structure of the catalytic chamber in the PreP closed state pre- cludes the capture of its substrates such as Aβ, or the release of its reaction products, which are key steps in the PreP catalytic cycle (Supplementary Fig. 1). To date, no structure of an open state of the M16C clan of metalloprotease has been reported. Solution scattering studies indicate that human PreP in solution is mostly in a closed or partially closed state6. However, the structural basis for the closed-open transition that allows for substrate capture and release of proteolysis products, in addition to substrate-induced transition from open to the closed state, remains elusive (Sup- plementary Fig. 1). Here, we integrate cryoEM, crystallography, size-exclusion chromatography (SEC) coupled small-angle X-ray scattering (SAXS), hydrogen–deuterium exchange (HDX)–mass spectrometry (MS), site-specific mutagenesis, and chemical biol- ogy approaches to elucidate the structural basis of the open-closed transition and the mechanism of substrate recognition for human PreP. Results Solution of multiple open and substrate-bound closed struc- tures by cryo-EM. Advances in cryo-electron microscopy (cryoEM) allow the structural determination of conformational states recalcitrant to crystallography24–27, thus we used cryoEM to examine the conformational states of PreP in the absence of substrate (apo-PreP). We first used differential scanning fluori- metry (DSF) to optimize the unfolding and dissociation enthalpy of human PreP (Supplementary Fig. 2)28. Apo-PreP grids were then prepared using a Vitrobot and a dataset of 2,626 micro- graphs was collected at 300 kV on a Titan Krios at various ice thicknesses (Table 1 and Supplementary Figs. 3 and 4). During processing, we observed that the predominant classes contained particles only half the expected size of PreP (Fig. 1a and Sup- plementary Fig. 3). Following the 3D classification of 411,000 particles, four classes were obtained (Supplementary Fig. 3). The two major classes, comprising ~208,000 and ~118,000 particles, displayed an intact PreP-N domain and a denatured PreP-C domain, which both were refined to 4.2 Å (Fig. 1b and Supple- mentary Figs. 4 and 5). The third class, comprised of 50,000 particles and refined to 4.5 Å, was found to contain full-length PreP adopting a partially open conformation that we designate pO (Fig. 1b and Supplementary Fig. 3). The final class, of ~34,000 particles and refined to 5.3 Å, comprised of the intact PreP-N and partially denatured PreP-C that was in pO state (Supplementary Fig. 3). interface (AWI) the extensive denaturation of It has previously been demonstrated that more than 90% of particles derived from diverse proteins or protein complexes are adsorbed to the air–water in cryoEM29. Furthermore, extensive studies have shown that many proteins are denatured rapidly upon exposure to the AWI30. Therefore, we hypothesized that the PreP-C domain described above resulted from denaturation at the AWI. To explore this possibility, we used fiducial-less cryo-electron tomography (cryoET) to examine the distribution of PreP particles within the vitrified ice of the same cryoEM grids. This analysis of Vitrobot-prepared grids revealed that nearly all PreP particles were adsorbed to the AWI (Fig. 1c, Supplementary Fig. 6, and Movie 1). Approximately, ~88% of particles had half of the anticipated size in our cryoET analysis, consistent with 2 NATURE COMMUNICATIONS | (2022) 13:1833 | https://doi.org/10.1038/s41467-022-29322-4 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29322-4 ARTICLE Table 1 CryoEM data collection, refinement, and validation statistics. Data collection and processing Apo-PreP Aβ-bound PreP CS27-bound PreP Conformations Magnification Voltage (kV) Electron exposure (e−/Å2) Defocus range (µm) Pixel size (Å) Symmetry imposed Initial particle images (no.) Final particle images (no.) 3D-FSC sphericity Map resolution (Å) FSC threshold EMDB Refinement Map sharpening B factor (Å2) Model composition Protein residues Total atoms Substrate B factors (Å2) Protein Substrate R.m.s deviations Bond length (Å) Bond angle (°) Ramachandran (%) Favored Allowed Outliers Validation MolProbity score Poor rotamers (%) Clash score PDB Partial Open 1 (pO1) Partial Open 2 (pO2) Open (O) Partial Closed (pC) Partial Closed (pC) 130,000 300 62.36 −2.5 to −1.5 0.856 C1 356,754 130,572 0.86 3.7 0.143 EMD-22278 130,000 300 66.39 −2.0 to −1.2 0.855 C1 213,544 174,537 0.837 3.3 0.143 EMD-22281 130,000 300 66.14 −2.5 to −1.5 0.855 C1 1,799,857 330,536 0.878 4.6 0.143 EMD-22282 356,754 139,127 0.871 3.9 0.143 EMD-22279 334,473 93,593 0.861 4.0 0.143 EMD-22280 −56 966 7704 – 111 – 0.009 1.080 97.61 2.39 0.00 1.17 0.00 2.88 6XOS −88 966 7704 – 132 – 0.006 1.086 97.19 2.81 0.00 1.41 0.23 4.98 6XOT −92 965 7697 – 102 – 0.005 1.081 97.19 2.81 0.00 1.32 0.35 3.80 6XOU −40 972 7769 51 90 88 0.004 0.668 94.91 5.09 0.00 1.98 0.00 13.06 6XOV −292 966 7711 – 136 – 0.006 0.976 95.84 4.16 0.00 2.15 0.23 23.48 6XOW predominant 3D classes (80%) having a denatured PreP-C domain (Fig. 1c and Supplementary Fig. 6A). Interestingly, while PreP-N and PreP-C share a highly similar structure, PreP-N has an additional β-hairpin (Supplementary Fig. 6B). This β-hairpin extends from an α-helix that binds the catalytic zinc ion and interacts with the α-helical hairpin that links PreP-N and structure likely makes PreP-N more stable PreP-C. This than PreP-C (Supplementary Fig. 6B). Together, our data indicate that PreP-C is preferentially denatured during the vitrification process, either by the repetitive exposure to AWI and/or shear force caused by grid blotting (Supplementary Fig. 6C)31. We hypothesized that if the amount of time the sample spent on the grid prior to vitrification (dwell time) could be significantly reduced, PreP denaturation would likewise be reduced. Spotiton, a novel method of vitrifying samples using a piezoelectric dispensing head to deliver sample droplets onto a self-blotting nanowire grid, has been shown to significantly reduce the dwell time of particles at the AWI prior to vitrification31–34. We employed this technique to prepare grids using a 133 ms dwell time (compared to 1–2 s for Vitrobot) by chameleon35, a commercial version of Spotiton developed by SPT Labtech. An apo-PreP dataset of 3012 micrographs was processed to yield about 363,000 particles from these grids, which adopted well- defined 3D classes. Following 2D and 3D classification, all classes were found to contain full-length PreP particles, and no denaturation was observed (Fig. 1d, e, Table 1, and Supplemen- tary Figs. 7 and 8). 3D classification of PreP particles revealed three distinct open state structures of PreP. They were refined to an open state (O) and two partially open states (pO1 and pO2) with resolutions of 4, 3.7, 3.9, respectively. Structural models of these three states were then built and refined (Table 1, Supplementary Figs. 5B–D, 8, and Movie 2). Two substrate- bound PreP cryoEM structures were also determined. We also optimized the conditions for determining the structure of PreP in the presence of a five-fold molar excess of Aβ 1–40 by DSF. Grids were prepared by the chameleon, and a dataset of 3483 micrographs was processed to yield about 175,000 particles from a well-defined 3D class that was refined to a partially closed (pC) state of PreP at 3.3 Å resolution (Fig. 1f, Table 1; Supplementary Figs. 5E, 9, and 11). A similar approach was used to obtain a map for PreP in complex with a model presequence peptide derived from human citrate synthase (27 aa long, CS27)36, resulting in pC state PreP at 4.6 Å resolution (Fig. 1g, Table 1; Supplementary Figs. 10 and 11). Structural models of substrate-bound PreP were then built and refined (Table 1, Supplementary Figs. 5E, 11, and Movie 3). We define the substrate-bound PreP cryoEM structures as pC state that is slightly more open than the closed (C) state crystal structures of PreP solved in the presence or absence of Aβ6. Structural analysis of apo- and substrate-bound PreP reveal key states in PreP catalytic cycle and the molecular basis for substrate recognition. Comparison of the five cryoEM structures NATURE COMMUNICATIONS | (2022) 13:1833 | https://doi.org/10.1038/s41467-022-29322-4 | www.nature.com/naturecommunications 3 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29322-4 a Vitrobot b PreP-N PreP-N PreP-C PreP-N PreP-C 10 nm d e Chameleon PreP-NPreP-C PreP-N PreP-C PreP-N PreP-C 10 nm 4.2 Å, 327K PreP-N only Missing PreP-C c 5.3 Å , 34K Open (O) Missing some PreP-C 4.5 Å, 50K partial Open (pO) i c e 7 5 n m Apo-PreP 4.0 Å, 94K Open (O) Apo-PreP 3.7 Å, 131K partial Open 1 (pO1) Apo-PreP 3.9 Å, 139K pO2 f PreP-N PreP-C g PreP-N PreP-C 100 nm 100 nm PreP - Aβ 3.3 Å, 175K partial Closed (pC) PreP - CS presequence 4.6 Å, 331K parital Closed (pC) Fig. 1 CryoEM analysis of PreP. a 2D classification, b 3D classification, and c cryoET analysis of PreP alone using grid prepared by Vitrobot. Top, Schematic diagrams of the average ice thickness (solid blue lines), and particle distribution in the ice. Almost all particles are on the AWI (770 particles), and only one full particle is not absorbed into the AWI. Bottom, Comparison of an enlarged slice of tomograms with particles’ labels (left) and without particles labels (right). PreP-N and PreP-C are colored in cyan and magenta, respectively. d 2D and e 3D classification of PreP alone using grid prepared by chameleon. f, g 3D classification of PreP in complex with Aβ (f) and citrate synthase (CS) presequence (g). derived from the apo- and substrate-bound PreP reveals three key conformational states based on the degree of opening of the catalytic chamber: open (O), partially open (either pO1 or pO2), and partially closed (pC). Apo-PreP has three states, O, pO1, and pO2 that are distinct from each other with the two pO states being slightly more open than the pC or C states (Fig. 2a, Sup- plementary Fig. 12, and Table 1). Aβ- and CS27-bound PreP structures in the pC state are nearly identical to each other and are slightly more open than crystal structures of apo or Aβ-bound PreP6 (Fig. 2a, Supplementary Fig. 12 and Table 1). Similar to the Aβ-bound PreP crystal structure6, extra-densities at the catalytic cleft and hydrophobic sites away from catalytic zinc ion within the catalytic chamber of PreP were found when Aβ was present, confirming key substrate-binding sites of PreP. Of these struc- tures, the pO, pC, and C states of PreP have a small variation in the degree of opening, making their chamber inaccessible to substrate binding. Thus, only the PreP open state has a large enough opening to capture its peptide substrates and release the proteolytic products. The PreP open state differs from the rest of PreP states in two major ways. The first is mediated by the rigid body displacement between PreP-N and PreP-C domains, whereby the two halves of this chamber-containing protease open up, similar to a clamshell (Fig. 2a–c, Supplementary Fig. 12, Movie 4, and Table 1). The displacement results in a difference in the distance, angle, and contacts between these domains. Most noticeably, both the distance and angle between PreP-N and PreP-C in the PreP open state is substantially larger than the rest of PreP states while the buried surface between PreP-N and PreP-C is much reduced compared to the others. The displacement between PreP-N and likely driven by the entropically favorable PreP-C is most rigid body motion. We also observed major conformational rearrangements in two additional regions, which we term switch (Fig. 2a–c; A (aa 174–225) and switch C (aa 506–550) Supplementary Fig. 12 and Movie 4). Switch A contains two α- helices that have residues for the binding of substrate and catalytic zinc ion (Fig. 2a). The helix-turn-helix motif of the switch C region joins PreP-N and PreP-C and makes extensive contacts with an extended β-hairpin within the long α-helix of the switch A region, allowing the switch A and C region move jointly with the displacement between PreP-N and PreP-C (Fig. 2a). 3D classification revealed that most PreP particles (~74%) were in the partially open (pO) states, indicating that PreP in the absence of substrate prefers to be in a state inaccessible to substrate binding. We then used size exclusion chromatography (SEC) in line with SAXS to further assess how the distribution of PreP conformational states in solution is influenced by the presence of Aβ and presequence from citrate synthase (CS27) under physiological buffer conditions (Fig. 2d, Supplementary Table 2 and Fig. 13). SAXS is a highly effective technique to eliminate structural models that do not produce calculated scattering patterns that fit the experimental scattering profile37. The direct coupling of SEC just prior to SAXS analysis reduces large aggregates that contaminated our previously reported SAXS profile of PreP6. Consistent with the cryoEM data, the SEC-SAXS data confirms that apo-PreP in solution also prefers to adopt the pO state (72% based on OLIGOMER and 90% based on ensemble optimization modeling (EOM)) (Fig. 2, Supplementary Table 2 and Fig. 13)38–40. The fact that PreP in solution prefers the pO state rather than the open state is logical because the transition from the pO states to O state loses substantial buried surface (650–940 Å2) and a network of hydrogen bonds and salt bridges, and thus is energetically unfavorable (Supplementary Table 1). The presence of CS27 significantly reduced the Rg value from 31.6 4 NATURE COMMUNICATIONS | (2022) 13:1833 | https://doi.org/10.1038/s41467-022-29322-4 | www.nature.com/naturecommunications pC (PreP+Aβ) pO1 (Apo-PreP) 2.0Å 4.0Å Sample PreP Apo PreP+A β pC O pO Mix apo EOM apo Mix Aβ EOM Aβ Rg (Å) 31.6 Dmax (Å) 91 31.4 30.4 32.2 30.4 31.4 31.5 31.1 31.4 87 89 94 91 96 97 93 98 NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29322-4 ARTICLE Apo-PreP pO1 Apo-PreP pO2 PreP+A pC ββ b a Apo-PreP O state PreP-C PreP-N D2 D3 Switch A D1 Switch C COM N-C: 37.4Å D1-L-D4: 83.3o D4 COM N-C: 30.4Å D1-L-D4: 63.8o d c COM N-C: 30.7Å D1-L-D4: 64.1o COM N-C: 29.3Å D1-L-D4: 61.3o PreP PreP + Aβ PreP-N aligned e pO2 pO1/O pC 0.5 Å 2.0 Å 3.0 Å Color by RMSD f PreP-C aligned O P pO +S Catalysis pC or C Aβ-C Aβ-N 90º 90º Aβ 90º 90º g E205 H104 E107 H108 E205 H104 E107 H108 PreP-N PreP-N PreP-C PreP-C pO2 pO1 pC Fig. 2 Comparison of PreP structures. a Four key cryoEM structures of PreP. Conformations of PreP open (O), partial open state 1 and 2 (pO1 and pO2), and Aβ-bound partial closed states (pC) are shown in the ribbon. The distance between the center of mass (COM) of PreP-N and PreP-C and the angle for COM of D1, aa 562-564 at switch C, and D4 domain are listed below each conformation. PreP-N, PreP-C, switch A, and switch C domains are colored in cyan, magenta, yellow, and green, respectively and the color scheme is used throughout the figure. b Alignment of pO1 and pC state of PreP colored by RMSD as indicated. c Alignment of PreP-N and PreP-C in four distinct conformational states of PreP colored by RMSD as indicated. d SEC-SAXS analysis of PreP in the presence and absence of Aβ. Mix denotes the scattering profile for a mixed population of known PreP structures as calculated in OLIGOMER. EOM denotes the scattering profile resulting from ensemble optimization modeling. While the major classes resulting from EOM showed high similarity to the states observed in our cryoEM analysis, a minor class that opened to a degree beyond that which has been previously observed experimentally were used in both PreP alone and PreP + Aβ conditions. As such minor classes also inflated the Dmax compared to experimental conditions, the OLIGOMER analysis was found to produce a better fit to the experimental data than the EOM analysis. See Supplementary Fig. 13 for more details. e Model of PreP catalytic cycle. S is a substrate and P is proteolytic products. f Structural basis of PreP open state primed to capture Aβ by size and charge complementarity. The charge distribution is calculated using PYMOL APBS plugin. The negative and positive charged surfaces are shown in red and blue, respectively. g Structural comparison of the catalytic site of pO and pC states colored based on RMSD. to 30.8 Å. Given the fact that the predicted SAXS profiles and Rg values of pO and pC are nearly identical, our analysis reveals that the presence of CS27 leads PreP to be almost entirely in the closed state (100% and 91% based on OLIGOMER and EOM, respectively) (Supplementary Fig. 13). However, the presence of Aβ only slightly reduced the Rg and Dmax values, which slightly increases the percentage of closed states (from 72% to 80% using OLIGOMER and from 50 to 54% using EOM). This is consistent with only a subtle conformational switch occurring between pO and pC states (Fig. 2d, Supplementary Fig. 13, and Table 1), whereby substrate-binding promotes domain closure. It is worth noting that EOM consistently generated a minor class of PreP that is much more open than the observed cryoEM open state of PreP (Supplementary Fig. 13). Thus, the analysis using OLIGO- MER likely represents the more realistic estimation. Together, our data lead to the hypothesis that PreP undergoes the following conformational switch during the catalytic cycle (Fig. 2e): PreP is predominantly in the partially open state at the resting condition. The transition from the pO states to O states allows the capture of the substrate, and thus is a key state in the catalytic cycle. Upon opening, peptides that are rich in positively charged residues are attracted to the negatively charged, catalytic chamber of PreP-N which can further select for its substrates based on their size and conformational compatibility within the chamber (Fig. 2f)6. The catalytic site of PreP undergoes a minimal conformational change, and thus is poised to carry out proteolysis (Fig. 2g). After proteolysis, the closed to open transition allows the release of proteolytic products to initiate the next catalytic cycle. The mechanism for the conformational switches between PreP open and closed states in the presence and absence of sub- strates. The comparison of PreP cryoEM structures reveals the molecular basis for the equilibrium between the partially open and open states in the absence of substrate. PreP has three regions that undergo substantial conformational switches, defined as switch A–C (Fig. 3a). As discussed above, switch A and C move together with rigid body displacement between PreP-N and PreP- C (Figs. 2a and 3a). The switch A and C regions in the pO states have lower resolution and higher thermal B factors than the rest of PreP structures (Fig. 3b). In comparison, the resolution and thermal B factors of these regions in the open and substrate- bound states are not profoundly different from the rest of the protein (Fig. 3b). Together, this is indicative of high conforma- tional dynamics within these regions. As switch A contains NATURE COMMUNICATIONS | (2022) 13:1833 | https://doi.org/10.1038/s41467-022-29322-4 | www.nature.com/naturecommunications 5 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29322-4 Fig. 3 Conformational switch of PreP. a Three switch regions of PreP. PreP-N, PreP-C, switch A, switch B, and switch C domains are colored in cyan, magenta, yellow, red, and green, respectively and the color scheme is used throughout figures. b Four cryoEM PreP structures colored by B-factors as indicated. c Comparison of PreP open and partial closed state to show the rotation of switch B helix that governs the open-closed transition. d The detailed interactions of switch B region in the open (top) and partial closed (bottom) states of PreP. e Model depicting how the rotation of switch B region governs the preferred parital closed state in the absence of substrate and the interaction of hydrophobic residues in the substrate, e.g., Aβ induces the open to closed transition of PreP. f The interaction of Aβ with the catalytic site (box 1) and exosite (box 2) of PreP. The key residues for Aβ binding and those that form exosite are shown in stick and labeled. The density found in Aβ bound PreP structure is shown in the mesh. The 19FFAE22 in Aβ is colored in cyan. g Differential HDX between PreP in the absence and presence of Aβ mapped on the PreP-N structure. Differences in the average HDX are represented as percent change and colored with blue being slower exchange with Aβ and with red being faster. The differences in HDX from n = 2 technical replicates are shown. h Relative catalytic activities of PreP mutants that have point mutations at residues residing at the interface between PreP-N and PreP-C. The average relative catalytic activities were shown in scale bar ±SD, and the original data n = 8 was overlaid on the scale bar. Source data are provided as a Source Data file. residues involved in substrate peptide binding and catalytic zinc ion coordination, such high dynamics would render the pO state catalytically incompetent. The presence of substrate stabilizes switch A, and thus, the residues for substrate binding and cata- lysis, enabling the catalytic reaction. This suggests that PreP uses substrate-assisted stabilization as a mechanism for substrate cat- alysis. This is because amyloid peptides have a high propensity to unfold and form a β-strand, which can then form the cross-β- sheet with other amyloid peptides. The catalytic cleft of PreP selectively binds the β-strand of substrate peptide after such peptide is unfolded inside the catalytic chamber. Only peptides that tend to unfold and form the β-strand can stabilize PreP’s catalytic cleft, leading to catalysis. PreP switch B (aa 421–436, within PreP-N) represents a rotation of an α-helix when the PreP open state is compared with the rest of conformational states (pO, pC, or C) (Fig. 3c, d; Supplementary Movie 5). The rotation of switch B is particularly noticeable at residue M436, which rotates ~60° to transit through the hydrophobic pocket formed by Y450 and L467 (Fig. 3c, d). This allows a rigid body rotation of PreP-C in relationship to PreP-N (Figs. 2e and 3e). Such a rotation maintains the interactions between switch B and PreP-C, e.g., the contacts of E432 with R675 and that of H430 with N676 and the main chains of aa residues 642 and 643. Thus, there should be a minimum energy barrier for such rotation to allow the rapid shift between PreP partially open and open states driven by the entropically favorable rigid body motion between PreP-N and PreP-C. The rotation of switch B also offers the molecular basis for PreP’s substrate-induced conformational switch and substrate selectivity. Switch B has hydrophobic residues L428 and I432 that, together with the surrounding residues, form a substrate-binding exosite distal to the catalytic zinc ion (~28–33 Å away) (Fig. 3f and Supplementary Movie 5). This exosite is highly hydrophobic and has been observed to bind Aβ residues that are away from the preferred cleavage sites of Aβ (aa 19/20 or aa 20/21) (Fig. 3f)6. The interaction of the exosite with the hydrophobic residues of substrate would favor the closed state and thus promote the open to closed transition (Fig. 3e). Furthermore, this can in part 6 NATURE COMMUNICATIONS | (2022) 13:1833 | https://doi.org/10.1038/s41467-022-29322-4 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29322-4 ARTICLE explain why PreP prefers to degrade peptide substrates that are rich in hydrophobic residues, e.g., presequences and Aβ. Peptide amide HDX-MS is a powerful tool to probe protein conformational dynamics because it allows evaluation of compara- tive solvent accessibility throughout the protein37,41–45. We used HDX-MS to test the hypothesis that the conformationally dynamic switch regions are stabilized by the binding of substrates (Aβ and CS27) (Supplementary Table 3 and Figs. 14–17, and Supplemen- tary Data 1–2). Both 2.7 Å resolution crystal6 and 3.3 Å resolution cryoEM structures of Aβ-bound PreP showed that several discrete regions within the catalytic chamber of PreP are involved in the recognition of presequence and Aβ. Residues from both PreP-N (aa F123, F124, L127, 135–139, and M206) and PreP-C (aa R900 and Y906) form a catalytic cleft to bind aa18–22 of Aβ. PreP prefers to degrade peptides that are rich in hydrophobic and positively charged residues. The recognition of hydrophobic and positively charged residues is mediated by a hydrophobic pocket formed by L428, I432, F344, L465, L60, F443, L447, and Y450 and a negatively charged pocket formed by D212, E213, D377, and D716, respectively. As expected, reduced HDX was observed in segments in both PreP-N and PreP-C that are involved in the substrate binding (e.g., aa 115–140, aa 166–175, and aa 893–921) (Supplementary Table 3 and Figs. 14–17, and Supplementary Data 1–2). Consistent with our SEC-SAXS data that the binding of substrates (both Aβ and CS27) promotes PreP to transition from the open to a closed state, we also observed the reduced HDX at the interface between PreP-N and PreP-C (e.g., aa 634–641, aa 705–722, and aa 922–933) in the presence of substrate. In addition to the expected changes in substrate binding and PreP-N/PreP-C interface, we observed reduced HDX of switch A–C regions when two different PreP substrates, Aβ and CS27 were present, which confirmed our prediction (Fig. 3g and Supplementary Figs. 14–17). To probe how the interface between PreP-N and PreP-C controls the equilibrium between the pO and O states, we carried out structure-guided mutagenesis in this region at positions predicted to weaken the interaction between PreP-N and PreP-C. We found that two point mutations, D367A and Q637A modestly increased the catalytic activity of PreP (Fig. 3h). Thus, destabilization of this interface enhances, rather than diminishes, PreP’s enzymatic activity, presumably through increasing the ease with which PreP can transition through the key conformational states of its catalytic cycle. Mechanism of PreP inhibition by MitoBlocker-60. At the PreP- N and -C interface, we observed an intriguing overlap between a key conformational switch (B) and a key substrate-binding site, the exosite, whose functional role in substrate-binding and cata- lysis is largely unexplored. To further define the conformational this interface, we exploited the findings of an dynamics at in vitro high-throughput screen that identified MitoBlocker-60 (MB60, 1-(diphenylmethyl)-4-(3-methyl-4-nitrobenzoyl)piper- azine) as a potent inhibitor of PreP (Fig. 4a). MB60 inhibited the = 200 nM and triggered degradation of Aβ by PreP with an IC50 mitophagy under mitochondrial stress46. A previous CRISPRi screen showed that PreP was essential for the robust cell pro- liferation of human K562 leukemia cells47. Consistent with this notion, MB60 potently blocked cell proliferation in a dose- dependent manner without inducing cell death (Fig. 4a). To inhibition, we co-crystallized understand the mechanism of MB60 with human PreP. The structure of MB60-bound PreP at 2.3 Å resolution reveals an unexpected binding mechanism of MB60 (Fig. 4b, Table 2, and Supplementary Fig. 18). In the presence of MB60, PreP exists in a closed conformation that is nearly identical to structures of substrate-free and Aβ-bound PreP (RMSD = 0.15 Å and 0.31 Å, respectively)6. Within the catalytic chamber, two MB60 molecules wrap around each other to make intimate interactions to bury 313 Å2 and bind the distinct pockets at the PreP exosite via various contacts to bury 808 and 658 Å2 surfaces of MB60-a (pink) and MB60-b (yellow), respectively (Fig. 4b, c). For MB60-a, two phenyl groups bind a hydrophobic pocket in close contact with L60, F344, I432, M446, L447, and L467. The carbonyl group of MB60-a forms hydrogen bonds with waters coordinated by the carbonyl group of M446 and the hydroxyl group of Y383. The piperazine group of MB60-a forms a hydrogen bond with water coordinated with the side chain of Q435. The nitro group of MB60-a forms a hydrogen bond with the main chain of G382 and a cation-π interaction with Y383. For MB60-b, two phenyl groups interact with the hydrophobic pocket formed by I337, A343, F344, I451, and L464 while the nitro group of MB60-b forms a salt bridge with K431, thereby favoring the pO state. The finding that MB60 targets PreP’s exosite provide strong evidence that this site plays a critical role in PreP catalysis. To explore how MB60 affects the substrate-binding and conformational dynamics of PreP, we first used SEC-SAXS and showed that MB60 did not induce obvious changes in the SAXS profile (Fig. 4e and Supplementary Figs. 13 and 14). The effect of MB60 on the conformational dynamics of PreP was then explored by the differences between amide H/D exchange profiles of PreP alone and PreP in the presence of MB60. Most exchanges were unchanged (Fig. 4f and Supplementary Figs. 14 and 15). Of a few regions that showed a noticeable reduction in H/D exchange, the peptides around MB60 binding pockets were most prevalent. These included residues 355–391 and 418–477 (Fig. 4f and Supplemen- tary Fig. 19). This supports the notion that MB60 in solution binds the hydrophobic pockets revealed by our MB60-bound PreP structure. While reduced exchange in the switch B region is expected as it is a part of the MB60 binding site, segments in switch A and switch C regions (aa 207–217 and aa 506–525, respectively) that are away from the MB60 binding site also displayed reduced H/D exchange (Fig. 4f and Supplementary Fig. 19). This is consistent with our model that in the absence of MB60, the switch A and C regions of PreP undergo dynamic motion. Moreover, that binding of MB60 to the exosite stabilizes such motion suggests that conformational changes in the exosite are functionally coupled to those in switches A and C. It is worth noting that, similar to the binding of substrates, Aβ and CS27, segments in PreP-C that are near the catalytic site also had the reduced exchange (e.g., aa 703–723, and aa 898–923, Fig. 4f and Supplementary Fig. 19). This is consistent with our model that the binding of MB60 to the switch B region can further stabilize the pO states, leading to a better interaction between PreP-N and PreP-C. Together, our structural and HDX-MS analysis explains how MB60 prevents degradation of Aβ and other substrates. MB60 is too large to enter the catalytic chamber of PreP when PreP is in the pO state, the dominant state in the solution. Upon entrance into the catalytic chamber via the PreP open state, MB60 promotes the rotation of switch B, which in turn promotes the open to closed transition. Such interaction should disfavor the pO to O transition, preventing Aβ from accessing the catalytic chamber. Furthermore, MB60 also blocks Aβ from binding to PreP exosite even after Aβ enters the PreP catalytic chamber. The close interactions of PreP with both phenyl groups of MB60 also explain the structure-activity relationship of MB6046: loss of one phenyl group increased the IC50 value five-fold while the relocation of the nitro group within the tolyl group of MB60 increased the IC50 value 25-fold. Discussion Major advances in instruments, techniques, and methods have fueled a “resolution revolution”, making single-particle cryoEM a NATURE COMMUNICATIONS | (2022) 13:1833 | https://doi.org/10.1038/s41467-022-29322-4 | www.nature.com/naturecommunications 7 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29322-4 a d f MB60 N NO2 N O 60 50 40 30 0 0.8 1.6 3 6 MB60 ((cid:2)M) 12 25 b PreP-C PreP-C MB6060 MB60 c PreP-N PPPPPPPPPrePPPP P-N Zn Zn e Rg (Å) Dmax (Å) PreP+Mb60 31.4±0.1 100±5 PreP 31.4±0.4 93±5 Å Q435 L467 M446 PreP-C 90o PreP-N 90o aa418-449 aa467-477 aa454-466 K431 aa355-372 aa703-723 aa898-923 aa371-383 aa931-940 aa506-525 PreP-C aa207-217 PreP-N Y383 Y380 L464 aa450-453 G382 Y450 Fig. 4 Structural analysis of PreP in complex with small molecule inhibitor, MB60. a The effect of MB60 on the doubling time of K562 Leukemia cells. The chemical formula of MB60 is shown. Error bar is SD from n = 0.5 × 104 cells (initial cell numbers) over 4 independent experiments. Source data are provided as Source Data files. b Overall structure of MB60-bound PreP (PDB code: 4RPU) PreP is depicted in a ribbon representation. PreP-N, PreP-C, switch A, switch B, and switch C domains are colored in cyan, magenta, yellow, red, and green, respectively and the color scheme is used throughout figures. The carbons of two MB60 molecules are colored in yellow and pink while N and O atoms are in blue and red, respectively. The catalytic zinc ion is in gray. c 2mFo–DFc omit map of MB60 to depict the closed contact between two MB60 molecules at the exosite. The map was contoured to 1σ. d Detailed interactions of two MB60 with PreP side chains within the exosite. e SEC-SAXS scattering profile of PreP in the presence or absence of MB60 (dotted lines). Theoretical scattering profiles of open and closed PreP (solid lines) were modeled and calculated by CRYSOL. f Differential HDX between PreP in the absence and presence of MB60 mapped on the MB60 bound PreP crystal structure. Differences in the average HDX are represented as percent change and colored with blue being slower exchange with MB60 while with red being faster. powerful structural determination technique that rivals macro- molecular crystallography24–27. However, rapid protein dena- turation in the thin film generated during the grid-making process, due to the high surface area to volume ratio at the AWI, represents a major obstacle in identifying the suitable condition to vitrify protein sample for cryoEM analysis30. Our cryoET analysis reveals that PreP, a 117 kDa, monomeric enzyme with homo- logous 55 kDa N- and C-domains, is preferentially absorbed to the AWI. Such exposure likely caused preferential denaturation of the PreP-C domain. The ability of the PreP-N domain to withstand exposure to the AWI, despite its high degree of structural homology to the PreP-C domain, is likely due to enhanced sta- bility granted by the lock formed by switch A and switch C regions (Supplementary Fig. 6). The combination of self-blotting nanowire the grids and piezo dispensing utilized by the chameleon, commercial version of Spotiton, eliminated the paper blotting step and significantly reduced the time for PreP to interact with the AWI31–35. This led to the successful determination of apo- and substrate-bound PreP cryoEM structures. Thus, PreP is a com- pelling case study that demonstrates how reduced vitrification times can be used to alter the kinetics of protein denaturation at the AWI that has led to the near-atomic resolution (3.3–4.6 Å) cryoEM structures of a relatively small protein (117 kDa human PreP) that exhibits high conformational heterogeneity. Further- more, PreP provides an intriguing model protein to assess the efficacy of newly emerging theories and practices of grid chemistry and vitrification process, e.g., VitroJet or Shake-it-off, aimed at preventing protein denaturation during vitrification48,49. Our integrative structural approaches lead to the formulation of the following model for the catalytic cycle of PreP (Fig. 5A). 8 NATURE COMMUNICATIONS | (2022) 13:1833 | https://doi.org/10.1038/s41467-022-29322-4 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29322-4 ARTICLE Table 2 Data collection and structure refinement statistics of MB60-bound PreP. Data collection Beamline Wavelength (Å) Space group Cell dimension(Å) a, b, c α, β, γ Resolution (Å) Rmeas (%)a Rp.i.m (%)b c CC1/2 CC*d I/sigma Redundancyf Completeness (%) Unique reflections h Refinement Rwork g Rfree No. of atoms Protein Water B-factors Protein Substrate Water R.m.s. deviations Bond lengths (Å) Bond angles (°) Ramachandran plot (%) Favorable region Allowed region Disallowed region PDB code APS-19ID 0.9792 C2 245.6, 85.5,158.2 90.0, 127.5, 90 44.85–2.27 (2.31–2.27) 18.6 (80.1)e 9.7 (42.8)e (0.639)e (0.883)e 18.2 (2.1)e 3.5 (3.3)e 99.9 (98.0)e 119317 0.176 0.208 15,861 811 36.1 33.5 40.5 0.006 0.972 92.6 7.4 0 4RPU aRmeas = Σhkl [n/(n − 1)]1/2Σi│Ihkl,I − 〈Ihkl〉 │ ∕ Σhkl 〈Ihkl〉 bRp.i.m. = Σhkl [1/(n − 1)]1/2Σi│Ihkl,I − 〈Ihkl〉 │ ∕ Σhkl 〈Ihkl〉 cCC1/2—Pearson correlation coefficient between random half-datasets—ρx,y = cov[(x,y)/(σxσy)] dCC* = [2CC1/2/(1 + CC1/2)]1/2 eThe outer resolution shell. Values in parentheses indicate the highest resolution shell. fNobs/Nunique. gRwork = Σhkl | |Fobs | −k | Fcalc | |/Σhkl | Fobs|. hRfree, calculated the same as for Rwork but on the 5% data excluded from the refinement calculation. Both cryoEM and SAXS data indicate that apo-PreP prefers to be in the partially open state, which cannot capture substrates. The C-terminal end of the switch B helix (e.g., Met 436) is located at the region where PreP-C pivots away from PreP-N during the closed to open transition (Figs. 3d and 5A). The rigid body motion of PreP-N and PreP-C entropically drives the separation between PreP-N and PreP-C domains. Governed by the rotation of the switch B region, the rigid body motion between PreP-N and PreP-C can trigger the conformational switch of the extended loop in the switch C region, allowing PreP to transition into the open state (Figs. 3c–e and 5A). Presequences rich in positively charged residues can then be attracted to the nega- tively charged surface of the PreP-N catalytic chamber (Figs. 2f and 5A). Furthermore, the high dipole moment of Aβ permits the charge complementation of this peptide with the catalytic chamber formed by PreP-N and PreP-C, which are negatively and positively charged, respectively (Figs. 2f and 5A). Plentiful hydrophobic residues in presequences and Aβ, or the small molecule inhibitor MB60, then interacts with the hydrophobic exosite to promote the rotation of switch B helix that induces the open to closed transition (Figs. 3f and 5A). The hydrophobic residues of PreP substrates also interact with the hydrophobic pocket at the catalytic site formed by PreP-N and PreP-C, fur- ther promoting the favorable interaction between PreP-N and PreP-C (Figs. 3f and 5A). The motion between PreP open and closed states, in conjunction with the selective interaction between the PreP catalytic chamber and the peptide substrate, provide the requisite force to unfold presequences and Aβ. This leads to the exposure of a β-strand within the presequence peptide or Aβ that can complement and stabilize the catalytic cleft formed in part by the switch A region, which in turn facilitates proteolysis. PreP then transitions from the closed state to the open state to release the reaction products. The transition between PreP open and partially open/closed state occurs quite frequently because its rate needs to be faster than the rate of Aβ degradation, which is 50–200 per second6. The model described above provides the molecular basis for the key conformational changes during the PreP catalytic cycle that facilitate amyloi- dogenic peptide capture and degradation. As the loss of function mutations in human PreP are associated with neurological dis- orders, e.g., cognitive impairments/disability and cerebellar atrophy10,11, our model should provide guidance for future investigation into how to boost PreP activity for better control of mitochondrial proteostasis. NATURE COMMUNICATIONS | (2022) 13:1833 | https://doi.org/10.1038/s41467-022-29322-4 | www.nature.com/naturecommunications 9 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29322-4 Fig. 5 Comparison of key structural features of PreP and IDE for catalysis. A A model of the catalytic cycle of PreP to depict the structural basis for the conformational switch and substrate recognition. PreP-N, PreP-C, switch A, switch B, and switch C domains are colored in cyan, magenta, yellow, red, and green, respectively and the color scheme is used throughout figures. B A model of dimeric IDE to depict the allosteric regulation. The domains and switch regions are depicted as the cartoon and colored the same as the figures above. A detailed description is in the discussion. that and degrade IDE, and PreP, Our structural analysis of two members of the cryptidase family, they have distinct indicates mechanisms of substrate suggests a common framework for amyloidogenic peptide recognition and regulation (this work) with distinct specializations that support efficient cytotoxic pep- tide clearance in their distinct cellular niches6,23,37,50,51. Both enzymes belong to the M16 clan of metalloproteases and have homologous ~55 kDa N- and C-terminal domains (Fig. 5)23,51. Both enzymes also undergo the open-closed transition during their catalytic cycle and only the open state can capture substrate and release proteolytic products (Fig. 5)6,23,51. Furthermore, their catalytic cleft of both enzymes is formed between their N- and C-terminal domains, which is unstable in the absence of sub- strate. They use a substrate-assisted catalysis mechanism to selectively recognize (this work)37,50. However, there are noticeable differences between IDE and PreP. Firstly, how these two enzymes open up is quite different due to the connecting region between their N- and C-terminal domains (Fig. 5). PreP is connected by the relatively long, dual α-helical hairpin switch C region that is connected to the zinc-containing D1 domain via the switch A region. Conse- quently, PreP undergoes a “book-opening” motion along a rather large surface between the PreP-N and PreP-C domains that is guided by the rotation of the switch B region. However, the short ~40 amino acid long loop of IDE allows IDE-N and IDE-C domains to pivot along a much smaller surface between D2 and D3 domains for a “packman-like” open-close motion, leading to the maximal separation between D1 and D4 domains37. The second difference is their oligomerization state (Fig. 5). The monomeric PreP ‘rests’ in a partially open state and the transition between partial open and open state allows PreP to capture its substrates. Contrary to monomeric PreP, IDE exists as a dimer to allow for the allosteric regulation of catalysis37,50. Specifically, at amyloid peptides least one of the two subunits within the IDE dimer is typically in the open state and primed for substrate capture most of the time. The binding of substrate allosterically facilitates the opening of the other IDE subunit within the IDE dimer, leading to enhanced IDE catalysis37,50. Furthermore, these two enzymes use the exo- site located at their D2 domain to recognize different structural features of their substrates (Fig. 5). The PreP exosite recognizes hydrophobic residues of peptide substrates while the IDE exosite binds the N-terminal mainchains6,23. Despite these profound differences, we have shown that the structural analyses can be used to rationally design mutations to enhance the enzymatic activities of both enzymes (this work)50. Furthermore, small molecules that can either boost IDE activity or selectively inhibit the degradation of insulin by IDE has been discovered52–54. Future work will realize the preventative or therapeutic potential these Aβ-degrading in controlling the proteolytic activity of proteases for neurodegenerative diseases caused by amyloid peptide-mediated toxicity. Methods Expression and purification of PreP. The expression vectors for wild-type human PreP and E107Q mutant were made previously as described6. Vectors for other PreP mutants were made using QuikChange site-directed mutagenesis kit with the following primers: PreP N358A-5′cccGCttctcccttttacaaagccttg3′ & 5′ggga- gaaGCgggcccagaagtcaaga3′, PreP E367A-5′ttgattgCatctggccttggcacagacttttc3′ & 5′ gccagatGcaatcaaggctttgtaaaagggag3′, PreP P558G-5′tgtctgGcagcgttgaaagtttccga- tattg3′ & 5′cgctgCcagacaagaggcatcttgagg3′, PreP N593D-5′accGatggcatggtg- tatttccggg3′ & 5′gccatCggtgggctgggcgcagtactg3′, PreP Q637A-5′ caggctGCgcagatagaattgaagaccggagg3′ & 5′tatctgcGCagcctgctcccggtagtcaag3′, PreP E901V-5′attcgagTaaaaggcggtgcttatggtgg3′ and 5′gccttttActcgaatttctgtatgcaa- gaatttgg3′. Wild-type human PreP and various mutants were expressed and pur- ified as described6. Briefly, E. coli Rosetta(DE3) containing the expression plasmid for the desired PreP construct was grown in T7 medium at 25 °C with 300 µM IPTG induction for 20 h. Cells were harvested via centrifugation at 10,000×g for 20 min (4 °C) and the resulting pellet was resuspended in a solution containing 20 mM Tris pH 8, 500 mM NaCl, 0.5 mM EDTA (omitted for purification of active PreP) 0.3 mM phenylmethylsulfonyl fluoride (PMSF), and 1 mM benzamidine- 10 NATURE COMMUNICATIONS | (2022) 13:1833 | https://doi.org/10.1038/s41467-022-29322-4 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29322-4 ARTICLE HCl. Proteins were then purified over a Ni2+-NTA affinity column equilibrated with a solution containing 20 mM Tris pH 7.7, 100 mM NaCl, and 0.1 mM PMSF. Bound protein was washed to baseline on-column first with a solution containing 20 mM Tris pH 7.7, 0.1 mM PMSF, and 500 mM NaCl followed by that containing 20 mM Tris pH 7.7 0.1 mM PMSF 50 mM NaCl, and 5 mM imidazole, to remove weakly bound contaminants. Protein was then eluted into a solution containing 20 mM Tris pH 7.7, 0.1 mM PMSF, 50 mM NaCl, and 150 mM imidazole. The protein sample was then diluted to a NaCl concentration <25 mM and loaded on a Source Q anion exchange column equilibrated with a solution containing 20 mM Tris pH 8.0 and 0.1 mM PMSF. After loading, the column was washed to baseline with the same equilibration buffer and bound protein was eluted with a solution containing 20 mM Tris pH 8.0, 0.1 mM PMSF, and a linear gradient of NaCl from 0 to 1 M over 25 column volumes. The resulting protein peak containing PreP was then applied to a Superdex 200 column equilibrated with a solution containing 20 mM Tris pH 8.0 and 50 mM NaCl for SEC. Protein purity was assessed via SDS- PAGE, aliquots were flash-frozen in liquid nitrogen and stored at −80 °C. Differential scanning fluorimetry. To optimize conditions for cryo-EM data collection, DSF was applied to screen 40 buffers and 98 additive conditions. The DSF was carried on with about 1 mg/ml PreP and 10× Sypro Orange in 20 µl buffers using Thermo Fisher Step ONE RT-PCR. Using the melting temperature and slope as selection criteria, the following condition was identified as best for the grid making with 20 mM Tris (pH 7.7), 150 mM NaCl, 10 mM EDTA, 0.5 mM β- mercaptoethanol. CryoEM data collection and analysis. Purified PreP was further purified by Superdex 200 chromatography using a buffer containing 20 mM Tris, pH 7.7, 150 mM NaCl, 10 mM KCl, 20 mM EDTA, and 1 mM β-mercaptoethanol. Grids were prepared using either Vitrobot or chameleon35. For Vitrobot grids, Quantifoil holey carbon-coated 200 mesh copper R1.2/1.3 grids were plasma cleaned in the air for 30 s using a Solarus plasma cleaner (Gatan). Totally, 2.5 μl of the sample was applied to the grid, waited for 15 s, and blotted with one layer of standard vitrobot filter paper (grade 595 with the outer/inner diameters of 55/20 mm, respectively; Ted Pella, inc. #47000-100), force 1, 100% humidity, for 2.5–3.5 s from both sides of the grids followed immediately by plunging into liquid ethane. For chameleon prepared grids, 300 mesh carbon or gold lacey nanowire grids were plasma cleaned with O2 and H2 for 10 s using a Solarus plasma cleaner (Gatan). The grids were plunged at 133 ms. All images were acquired using a Titan Krios microscope (FEI) operated at 300 keV with a Gatan K2 direct electron detector (Gatan) in counting mode. Images were automatically acquired using Leginon55 using collection parameters as shown in Table 1. Images were processed using software integrated into RELION3.056. Frames were aligned using MotionCor2 software with electron exposure weighting57, CTF was estimated using Gctf58, particles were picked and extracted automatically using RELION3.056. Particle stacks were processed through several rounds of 2D and 3D classification. Example images and 2D class averages are shown in Supplementary Figs. 3, 7, 9, and 10. Selected classes with good sphericity values (ranging from 0.837 to 0.878) based on 3D-FSC59 were then processed for high-resolution 3D refinement (Table 1 and Supplementary Figs. 3, 7, 9, and 10). Finally, the overall map was improved by particle polishing in RELION3.056. The final resolution was estimated using gold-standard Fourier Shell Correlation (FSC = 0.143) (Supplementary Figs. 4, 8, 9, and 10). CryoEM data collection and processing statistics are listed in Table 1. Structural models were built using Aβ-bound PreP crystal structure (PDB = 4NGE) as a template6. Density fitting and structure refinement were performed using UCSF CHIMERA60, COOT61, and PHENIX62. The refinement statistics are listed in Table 1. CryoET analysis. Tilt-series were collected with Leginon55 on a Titan Krios with a Gatan K2 counting camera using the same sample preparation as for cryoEM. Tilt- series were aligned with Appion-Protomo63, electron exposure-weighted using equation 3 in Grant and Grigorieff64, and reconstructed with Tomo3D65. Sub- tomogram processing, including particle picking, alignment, and classification, was performed with Dynamo66 and reconstructed with Tomo3D65. Protein crystallization, data collection, and structure determination. PreP E107Q was modified by reductive lysine methylation prior to Superdex 200 chromatography as reported previously6. Specifically, 1–10 mg/ml PreP in buffer containing 50 mM HEPES pH 8.0, 500 mM NaCl, 5% Glycerol (v/v), and 10 mM β- mercaptoethanol was incubated with 40 mM formaldehyde and 20 mM dimethylamine-borane complex for 2 h at 4 °C, followed by an overnight incuba- tion with an additional 20 mM dimethalymine-borane complex. The reaction was then quenched for 2 h by adding glycine to 13.3 mM and DTT to 5 mM. Totally, 5–7 mg/ml lysine-methylated PreP-E107Q in buffer containing 20 mM HEPES pH 7.5, 250 mM NaCl, 2 mM DTT, and 200 μM MB60 (MolPort #001-620-747) was combined with mother liquor containing 15.0% (w/v) PEG 8000, 15 mM TCEP, 80 mM sodium cacodylate pH 6.7, 160 mM calcium acetate, and 20% (v/v) glycerol in a 1:1 (v/v) ratio for the crystallization of PreP-MB60 complex by hanging-drop vapor diffusion at 18 °C. Crystals grew for one week prior to data collection. Crystals were cryoprotected in mother liquor containing 30% (v/v) glycerol, then flash frozen in liquid nitrogen. Diffraction data were collected at beamline 19ID at Argonne National Laboratory and processed using HKL300067. The structure of PreP in complex with MB60 was determined by molecular replacement using Phaser and PreP structure (4L3T) as the search model. Model building—including the addition of missing 317–323 residues in chain A—were performed using COOT61 and refinement was done using PHENIX62. The final 2.27 Å resolution model (pdb = 4RPU) has Rwork = 20.6%. Data collection and = 18.6% and Rfree structure refinement statistics are listed in Table 2. Presumably, due to the shorter time of crystallization, only cysteine 112 was modified with the dimethylarsenic moiety while cysteine 556 was not. Cysteine 556 is in close proximity with cysteine 119 to form a disulfide bond. The absence of a disulfide bond between these residues might be due to the presence of a reducing agent during the purification and/or crystallization. SAXS data collection and analysis. SAXS data were collected at the BioCAT/ 18ID beamline at Advanced Photon Source, Argonne National Laboratory (Chi- cago, USA) beamline 12ID-B, at 23 °C using 1.1 mg/ml protein and an incident X-ray wavelength of 0.886 Å, and protein concentration of 0.5 mg/ml. For MB60- binding experiments, PreP was preincubated with 200 μM MB60 on ice prior to SAXS data collection. Data were reduced and analyzed using ATSAS68 using the photon counting PILATUS 3 1 M at room temperature (23 °C) and an incident X-ray wavelength of 1.03 Å. The 3.5 m sample-to-detector distance yielded a range of 0.005–0.33 Å−1 for the momentum transfer (q = 4π sinθ/λ where 2θ is the scattered angle between the incident and scattered beam and λ the X-ray wave- length). The PreP sample was loaded onto an SEC system (ÄKTA pure, GE Healthcare Life Sciences, Piscataway, NJ) with a GE Superdex 200 10/300 G. Totally, 2–3 mg protein was injected to Superdex 200 in the buffer containing 20 mM Tris, pH 7.7, 100 mM NaCl with/without EDTA. To remove the zinc ion from PreP, the protein was dialyzed against 500 ml 20 mM Tris pH7.7, 100 mM NaCl, 20 mM EDTA. A 5-fold molar excess of Aβ, CS27, or MB60 was mixed with PreP immediately prior to loading on the Superdex 200 in the buffer containing 20 mM Tris-HCl (pH 8.0), 100 mM NaCl, with/without 20 mM EDTA. The data were reduced and analyzed using ATSAS68. PRIMUS and GNOM in the ATSAS suite were used to determine the Rg value in reciprocal and real space, respectively68. Dmax and P(r) distribution were calculated by GNOM. Disordered regions were modeled into the structures based on the Alphafold structure of PreP (AF-Q5JRX3-F1)68 and theoretical scattering curves for different models were generated and fit to the experimental data using CRYSOL in the ATSAS suite68. The addition of missing segments into the experimental structures has been shown to substantially improve the fitting of SAXS data69. EOM was performed using EOM 2.038–40. With the N- and C-domains of PreP as input, 10,000 native-like chain models with 200 points were generated and assessed over 100 cycles of a genetic algorithm using default parameters. OLIGOMER in the ATSAS suite was used to determine the percent composition by parsimonious conformational states that best fit the observed data68. Key parameters in SAXS data acquisition, sample details, data analysis, and modeling fitting, as well as the software used, are listed in Supplementary Table 2 as recommended by publication guideline70. Hydrogen deuterium exchange–mass spectrometry. Prior to performing com- parative H/D exchange experiments, enzymatic and quench conditions that pro- duced an optimal fragmentation pattern of PreP were established as previously described71.71. Briefly, for PreP and MB60 study, 3 μl 4.7 mg/ml PreP in buffer containing 20 mM Tris-HCl (pH 7.5), and 50 mM NaCl was diluted with 9 μl of buffer A (8.3 mM Tris-HCl (pH 7.5), 50 mM NaCl in H2O) at 0 °C. For PreP and substrates study, 3 μl 10.5 mg/ml PreP in buffer containing 20 mM Tris-HCl (pH 7.7), 150 mM NaCl and 10 mM EDTA was diluted with 9 μl of buffer B (8.3 mM Tris-HCl (pH 7.7), 150 mM NaCl, 10 mM EDTA in H2O) at 0 °C. The sample was then mixed with 18 μl of ice-cold quench buffers containing 0.8% formic acid, 16.6% glycerol, and various concentrations of GuHCl (0.08, 0.8, and 1.6 M). Quenched samples were then subjected to HDX-MS apparatus for proteolysis and LC/MS analysis. The use of 0.8 M GuHCl resulted in the best sequence coverage of PreP. For HDX-MS analysis, 13 μM PreP in the presence or absence of 130 μM MB60 with buffer 8.3 mM Tris-HCl pH7.2, 50 mM NaCl, and 2.1% DMSO in H2O, 260 μM Aβ, or 260 μM CS27 in buffer B and 2.1%DMSO in H2O was incubated at room temperature for 30 min prior to chilling to 0 °C for deuteration studies. Functional hydrogen-deuterium exchange reactions were initiated by adding a 3 μl sample into 9 μl of buffer A or buffer B in D2O (pDREAD = 7.2) and incubated at 0 °C for 10, 100, 1000, 10,000, and 100,000 s72. The exchange reaction was ter- minated by adding 18 μl of ice-cold 0.8% formic acid, 0.8 M GuHCl, 16.6% glycerol for a final pH of 2.5. Quenched samples were then immediately frozen on dry ice and stored at −80 °C prior to LC/MS analysis. Un-deuterated and equilibrium- deuterated control samples are also prepared as previously described73. Frozen samples were later loaded onto a cryogenic autosampler74, thawed at 4 °C, then passed over an immobilized pepsin column (16 μl bed volume) for 30–40 s digestion. Proteolytic fragments were collected on a trap column and separated using Optimize Technologies C18 reverse-phase analytical column (Halo EC-C18 0.2 × 50 mm, 2.7 μm) with an acetonitrile linear-gradient (6.4–38.4% over 30 min). The effluent was directed into an OrbiTrap Elite Mass Spectrometer (Thermo- Fisher Scientific, San Jose, CA). Instrument settings were optimized to minimize the back-exchange75. The data was acquired in either MS1 profile mode or data- dependent MS/MS mode. Peptide identification was done with the aid of Proteome NATURE COMMUNICATIONS | (2022) 13:1833 | https://doi.org/10.1038/s41467-022-29322-4 | www.nature.com/naturecommunications 11 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29322-4 Discoverer software (ThermoFisher). Mass envelope centroids of deuterated pep- tides were calculated with HDEXaminer v2.5.1 (Sierra Analytics Inc, Modesto, CA) then converted to corresponding deuterium incorporation with corrections for back-exchange76. Deuterium uptake plots, Heat Map and Difference Maps were generated with Excel macro and MatLab scripts. Key parameters and data are included in Supplementary Table 3 and Data 1–2 according to the recommendation77. 8. Alikhani, N. et al. Decreased proteolytic activity of the mitochondrial amyloid- beta degrading enzyme, PreP peptidasome, in Alzheimer’s disease brain mitochondria. J. Alzheimers Dis. 27, 75–87 (2011). Fang, D. et al. Increased neuronal PreP activity reduces Abeta accumulation, attenuates neuroinflammation and improves mitochondrial and synaptic function in Alzheimer disease’s mouse model. Hum. Mol. Genet. 24, 5198–5210 (2015). 9. Enzymatic assay. The fluorogenic peptide substrate Mca-Y-V-A-D-A-P-K(Dnp)- OH (R&D Systems, Catalog # ES007) was used to measure the activity of PreP. The reaction was monitored on a Synergy Neo microplate reader using an excitation wavelength of 320 nm and an emission wavelength of 405 nm. Reactions were carried out at 37 °C, using 5 nM PreP with various concentrations of substrate V (5, 10, 20, or 40 μM) in 200 μL of buffer containing 20 mM Tris, pH7.7, 150 mM NaCl, 1 mM β-mercaptoethanol. Degradation of substrate V was assessed by monitoring the fluorescence increase for 10 min at 30-s intervals. To calculate enzymatic activity, background subtraction and linear regression fitting were used to find the initial velocity, whereas specific activity was determined by comparing the maximal fluorescence converted from the known quantity of substrate V by PreP. Cellular proliferation assay of human K562 leukemia cells. MB60 was pur- chased from MolPort. Leukemia K562 cells that express GFP (gifted from Luke Gilbert; engineered using K562 cells from ATCC CCL-243) were grown in RPMI- 1640 medium with 10% FBS, 0.05 mM 2-Mercaptoethanol, and penicillin/strep- tomycin. Cells at 1 × 105 viable cells/mL were added to a 96-well plate with 100 µl/ well. MB60 was added to the indicated concentrations, 0.8 to 25 µM. Cell growth was monitored continually every 4 h up to 60 h using IncuCyte S3s (Essen BioScience). The cell count from 12 to 40 h was used to calculate their cell doubling time. Reporting summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article. Data availability The data supporting the findings of this study are available from the corresponding author upon reasonable request. Structure factor amplitudes and coordinates for the crystal structures of MB-60 bound PreP and Aβ-bound PreP are deposited in the Protein Data Bank under accession number 4RPU and 4NGE, respectively. The 3D cryoEM density maps generated in this study is deposited in the Electron Microscopy Data Bank under accession codes EMD-22278 (Apo-PreP pC1 state), EMD-22279 (Apo-PreP pC2 state), EMD-22280 (Apo-PreP open state), EMD-22281 (Aβ-bound PreP), and EMD-22282 (CS27-bound PreP). The corresponding atomic coordinates are deposited in the Protein Data Bank under accession numbers 6XOS (Apo-PreP pC1 state), 6XOT (Apo-PreP pC2 state), 6XOU (Apo-PreP open state), 6XOV (Aβ-bound PreP), and 6XOW (CS27-bound PreP). EM data in the form of unprocessed micrographs is deposited in the Electron Microscopy Public Image Archive (EMPIAR) under accession number EMPIAR-10937. The tomogram shown in the figures is deposited to the EMD with the accession number EMD-25921. Raw tomography data is deposited to EMPIAR with the accession number EMPIAR-10929. The HDX-MS data are deposited in ProteomeXchange under the accession number PXD029542. SAXS data is deposited in the Small Angle Scattering Biological Data Bank (SASBDB) under accession codes SASDKK3 (Apo-PreP), SASDKL3 (MB60-PreP), SASDKM3 (CS27-PreP), and SASDKN3 (Aβ-PreP). Source data are provided with this paper. Received: 21 August 2020; Accepted: 4 March 2022; References 1. Eisner, V., Picard, M. & Hajnoczky, G. Mitochondrial dynamics in adaptive and maladaptive cellular stress responses. Nat. Cell Biol. 20, 755–765 (2018). Spinelli, J. B. & Haigis, M. C. The multifaceted contributions of mitochondria to cellular metabolism. Nat. Cell Biol. 20, 745–754 (2018). 3. DiMauro, S. & Schon, E. A. Mitochondrial disorders in the nervous system. Annu. Rev. Neurosci. 31, 91–123 (2008). Poveda-Huertes, D., Mulica, P. & Vogtle, F. N. The versatility of the mitochondrial presequence processing machinery: cleavage, quality control and turnover. Cell Tissue Res. 367, 73–81 (2017). Pfanner, N., Warscheid, B. & Wiedemann, N. Mitochondrial proteins: from biogenesis to functional networks. Nat. Rev. Mol. Cell Biol. 20, 267–284 (2019). 6. King, J. V. et al. Molecular basis of substrate recognition and degradation by human presequence protease. Structure 22, 996–1007 (2014). Johnson, K. A. et al. The closed structure of presequence protease PreP forms a unique 10,000 Å3 chamber for proteolysis. EMBO J. 25, 1977–1986 (2006). 2. 4. 5. 7. 12 10. Brunetti, D. et al. Defective PITRM1 mitochondrial peptidase is associated with A amyloidotic neurodegeneration. Embo Mol. Med. 8, 176–190 (2016). 11. Langer, Y. et al. Mitochondrial PITRM1 peptidase loss-of- function in childhood cerebellar atrophy. J. Med. Genet. 55, 599–606 (2018). 12. Brunetti, D. et al. Role of PITRM1 in mitochondrial dysfunction and neurodegeneration. Biomedicines 9, 833 (2021). 13. Hipp, M. S., Kasturi, P. & Hartl, F. U. The proteostasis network and its decline in ageing. Nat. Rev. Mol. Cell Biol. 20, 421–435 (2019). 14. Lopez-Otin, C., Blasco, M. A., Partridge, L., Serrano, M. & Kroemer, G. The Hallmarks of aging. Cell 153, 1194–1217 (2013). 15. Chiti, F. & Dobson, C. M. Protein misfolding, amyloid formation, and human disease: a summary of progress over the last decade. Annu. Rev. Biochem. 86, 27–68 (2017). 16. Greenwald, J. & Riek, R. Biology of amyloid: structure, function, and regulation. Structure 18, 1244–1260 (2010). 17. Selkoe, D. J. & Hardy, J. The amyloid hypothesis of Alzheimer’s disease at 25 years. Embo Mol. Med. 8, 595–608 (2016). 18. Muller, U. C., Deller, T. & Korte, M. Not just amyloid: physiological functions of the amyloid precursor protein family. Nat. Rev. Neurosci. 18, 281–298 (2017). 19. Sikanyika, N. L., Parkington, H. C., Smith, A. I. & Kuruppu, S. Powering amyloid beta degrading enzymes: a possible therapy for Alzheimer’s disease. Neurochem. Res. 44, 1289–1296 (2019). 20. Lopez-Otin, C. & Bond, J. S. Proteases: multifunctional enzymes in life and disease. J. Biol. Chem. 283, 30433–30437 (2008). 21. Gillam, J. E. & MacPhee, C. E. Modelling amyloid fibril formation kinetics: mechanisms of nucleation and growth. J. Phys. Condens Matter 25, 373101 (2013). 22. Morris, A. M., Watzky, M. A. & Finke, R. G. Protein aggregation kinetics, mechanism, and curve-fitting: a review of the literature. Biochim. Biophys. Acta 1794, 375–397 (2009). 23. Malito, E., Hulse, R. E. & Tang, W. J. Amyloid beta-degrading cryptidases: insulin degrading enzyme, presequence peptidase, and neprilysin. Cell Mol. Life Sci. 65, 2574–2585 (2008). 24. Danev, R., Yanagisawa, H. & Kikkawa, M. Cryo-electron microscopy methodology: current aspects and future directions. Trends Biochem. Sci. 44, 837–848 (2019). 25. Lyumkis, D. Challenges and opportunities in cryo-EM single-particle analysis. J. Biol. Chem. 294, 5181–5197 (2019). 26. Murata, K. & Wolf, M. Cryo-electron microscopy for structural analysis of dynamic biological macromolecules. Biochem. Biophys. Acta 1862, 324–334 (2018). 27. Ognjenovic, J., Grisshammer, R. & Subramaniam, S. Frontiers in cryo electron microscopy of complex macromolecular assemblies. Annu. Rev. Biomed. Eng. 21, 395–415 (2019). 28. Chari, A. et al. ProteoPlex: stability optimization of macromolecular complexes by sparse-matrix screening of chemical space. Nat. Methods 12, 859–965 (2015). 29. Noble, A. J. et al. Routine single particle cryoEM sample and grid characterization by tomography. Elife 7, e34257 (2018). 30. Glaeser, R. M. Proteins, interfaces, and cryo-EM grids. Curr. Opin. Colloid 34, 1–8 (2018). 31. Noble, A. J. et al. Reducing effects of particle adsorption to the air-water interface in cryo-EM. Nat. Methods 15, 793–795 (2018). 32. Dandey, V. P. et al. Spotiton: a new method for vitrifying samples for cryoEM. Acta Crystallogr. A 73, C1295–C1295 (2017). 33. Dandey, V. P. et al. Spotiton: new features and applications. J. Struct. Biol. 202, 161–169 (2018). Jain, T., Sheehan, P., Crum, J., Carragher, B. & Potter, C. S. Spotiton: a prototype for an integrated inkjet dispense and vitrification system for cryo- TEM. J. Struct. Biol. 179, 68–75 (2012). 34. 35. Darrow, M. C., Moore, J. P., Walker, R. J., Doering, K. & King, R. S. Chameleon: next generation sample preparation for CryoEM based on spotiton. Microsc. Microanalysis 25, 994–995 (2019). 36. Chow, K. M. et al. Mammalian pitrilysin: substrate specificity and mitochondrial targeting. Biochemistry 48, 2868–2877 (2009). 37. Zhang, Z. N. et al. Ensemble cryoEM elucidates the mechanism of insulin capture and degradation by human insulin degrading enzyme. Elife 7, e33572 (2018). NATURE COMMUNICATIONS | (2022) 13:1833 | https://doi.org/10.1038/s41467-022-29322-4 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29322-4 ARTICLE 38. Konarev, P. V., Volkov, V. V., Sokolova, A. V., Koch, M. H. J. & Svergun, D. I. PRIMUS: a Windows PC-based system for small-angle scattering data analysis. J. Appl. Crystallogr. 36, 1277–1282 (2003). 68. Franke, D. et al. ATSAS 2.8: a comprehensive data analysis suite for small- angle scattering from macromolecular solutions. J. Appl. Crystallogr. 50, 1212–1225 (2017). 39. Tria, G., Mertens, H. D. T., Kachala, M. & Svergun, D. I. Advanced ensemble modelling of flexible macromolecules using X-ray solution scattering. IUCRJ 2, 207–217 (2015). 40. Bernado, P., Mylonas, E., Petoukhov, M. V., Blackledge, M. & Svergun, D. I. Structural characterization of flexible proteins using small-angle X-ray scattering. J. Am. Chem. Soc. 129, 5656–5664 (2007). 41. Chung, K. Y. et al. Conformational changes in the G protein Gs induced by the β2 adrenergic receptor. Nature 477, 611–615 (2011). 42. Li, S. S. et al. Mechanism of intracellular cAMP sensor Epac2 activation: cAMP-induced conformational changes identified by amide hydrogen/ deuterium exchange mass spectrometry (DXMS). J. Biol. Chem. 286, 17889–17897 (2011). 43. Marciano, D. P., Dharmarajan, V. & Griffin, P. R. HDX-MS guided drug discovery: small molecules and biopharmaceuticals. Curr. Opin. Struct. Biol. 28, 105–111 (2014). 44. Maianti, J. P. et al. Anti-diabetic activity of insulin-degrading enzyme inhibitors mediated by multiple hormones. Nature 511, 94–98 (2014). 45. Deprez-poulain, R. et al. Catalytic site inhibition of insulin degrading enzyme by a small molecule induces glucose intolerance in mice. Nat. Commun. 6, 8250 (2015). 46. Wijaya, J. Utilizing Small Molecules to Study Mitochondrial Presequence- Degrading Protease. PhD thesis, (UCLA, 2015). 47. Horlbeck, M. A. et al. Mapping the genetic landscape of human. Cells Cell 174, 953–967 (2018). 48. Ravelli, R. B. G. et al. Cryo-EM structures from sub-nl volumes using pin- printing and jet vitrification. Nat. Commun. 11, 2563 (2020). 49. Rubinstein, J. L. et al. Shake-it-off: a simple ultrasonic cryo-EM specimen-preparation device. Acta Crystallogr. D Struct. Biol. 75, 1063–1070 (2019). 50. McCord, L. A. et al. Conformational states and recognition of amyloidogenic peptides of human insulin-degrading enzyme. Proc. Natl Acad. Sci. USA 110, 13827–13832 (2013). 51. Tang, W. J. Targeting insulin-degrading enzyme to treat type 2 diabetes mellitus. Trends Endocrinol. Metab. 27, 24–34 (2016). 52. Cabrol, C. et al. Small-molecule activators of insulin-degrading enzyme discovered through high-throughput compound screening. PLoS ONE 4, e5274 (2009). 53. Maianti, J. P. et al. Substrate-selective inhibitors that reprogram the activity of insulin-degrading enzyme. Nat. Chem. Biol. 15, 565–574 (2019). 54. Kukday, S. S. et al. Cell-permeable, small-molecule activators of the insulin- degrading enzyme. J. Biomol. Screen 17, 1348–1361 (2012). 55. Suloway, C. et al. Automated molecular microscopy: the new Leginon system. J. Struct. Biol. 151, 41–60 (2005). 56. Zivanov, J. et al. New tools for automated high-resolution cryo-EM structure determination in RELION-3. Elife 7, e42166 (2018). 57. Zheng, S., Polovcak, E., Armache, J. P., Cheng, Y. & Agard, D. Anisotropic correction of beam-induced motion for improved single-particle electron cryomicroscopy. Nat. Methods 14, 331–332 (2017). 58. Zhang, K. Gctf: real-time CTF determination and correction. J. Struct. Biol. 193, 1–12 (2016). 59. Tan, Y. Z. et al. Addressing preferred specimen orientation in single-particle cryo-EM through tilting. Nat. Methods 14, 793–796 (2017). 60. Pettersen, E. F. et al. UCSF chimera—a visualization system for exploratory research and analysis. J. Comput. Chem. 25, 1605–1612 (2004). 61. Emsley, P., Lohkamp, B., Scott, W. G. & Cowtan, K. Features and development of Coot. Acta Crystallogr. D Biol. Crystallogr. 66, 486–501 (2010). 62. Liebschner, D. et al. Macromolecular structure determination using X-rays, neutrons and electrons: recent developments in Phenix. Acta Crystallogr. D Struct. Biol. 75, 861–877 (2019). 63. Noble, A. J. & Stagg, S. M. Automated batch fiducial-less tilt-series alignment in Appion using Protomo. J. Struct. Biol. 192, 270–278 (2015). 64. Grant, T. & Grigorieff, N. Measuring the optimal exposure for single particle cryo-EM using a 2.6 A reconstruction of rotavirus VP6. Elife 4, e06980 (2015). 65. Agulleiro, J. I. & Fernandez, J. J. Tomo3D 2.0-exploitation of advanced vector extensions (AVX) for 3D reconstruction. J. Struct. Biol. 189, 147–152 (2015). 66. Castano-Diez, D., Kudryashev, M., Arheit, M. & Stahlberg, H. Dynamo: a flexible, user-friendly development tool for subtomogram averaging of cryo- EM data in high-performance computing environments. J. Struct. Biol. 178, 139–151 (2012). 67. Minor, W., Cymborowski, M., Otwinowski, Z. & Chruszcz, M. HKL-3000: the integration of data reduction and structure solution-from diffraction images to an initial model in minutes. Acta Crystallogr. D Biol. Crystallogr. 62, 859–866 (2006). 69. Liang, W. G., Mancl, J. M., Zhao, M. & Tang, W. J. Structural analysis of Mycobacterium tuberculosis M13 metalloprotease Zmp1 open states. Structure 29, 709–720 (2021). 70. Trewhella, J. et al. 2017 publication guidelines for structural modelling of small-angle scattering data from biomolecules in solution: an update. Acta Crystallogr. Sect. D. Struct. Biol. 73, 710–728 (2017). 71. Marsh, J. J. et al. Structural insights into fibrinogen dynamics using amide hydrogen/deuterium exchange mass spectrometry. Biochemistry 52, 5491–5502 (2013). 72. Coales, S. J. et al. Expansion of time window for mass spectrometric measurement of amide hydrogen/deuterium exchange reactions. Rapid Commun. Mass Spectrom. 24, 3585–3592 (2010). 73. Xiao, K. et al. Revealing the architecture of protein complexes by an orthogonal approach combining HDXMS, CXMS, and disulfide trapping. Nat. Protoc. 13, 1403–1428 (2018). 74. Woods, V. L., Jr. & Hamuro, Y. High resolution, high-throughput amide deuterium exchange-mass spectrometry (DXMS) determination of protein binding site structure and dynamics: utility in pharmaceutical design. J. Cell. Biochem. Supplement Suppl 37, 89–98 (2001). 75. Walters, B. T., Ricciuti, A., Mayne, L. & Englander, S. W. Minimizing back exchange in the hydrogen exchange-mass spectrometry experiment. J. Am. Soc. Mass Spectrom. 23, 2132–2139 (2012). 76. Zhang, Z. & Smith, D. L. Determination of amide hydrogen exchange by mass spectrometry: a new tool for protein structure elucidation. Protein Sci. 2, 522–531 (1993). 77. Masson, G. R. et al. Recommendations for performing, interpreting and reporting hydrogen deuterium exchange mass spectrometry (HDX-MS) experiments. Nat. Methods 16, 595–602 (2019). Acknowledgements We are grateful to Luke Gilbert for the leukemia K562 cell line labeled with GFP and Srinivas Chakravarthy at BioCAT, APS for assisting with SAXS data collection and analysis. This work was supported by the NIH grant GM121964 to W.-J. Tang, GM103622 to Tom Irving at BioCAT, APS. Some of this work was performed at the Simons Electron Microscopy Center and National Resource for Automated Molecular Microscopy and National Center for CryoEM Access and Technology located at the New York Structural Biology Center, supported by grants from the Simons Foundation (SF349247) and the NIH National Institute of General Medical Sciences (GM103310, U24 GM129539). Use of the Advanced Photon Source was supported by the U.S. Department of Energy, Office of Basic Energy Sciences, under contract No. DE-AC02- 06CH11357. Author contributions W.-J.T., W.G.L., S.L., B.C., C.S.P., and M.Z. designed the project. W.G.L. performed cryoEM grid preparation, data acquisition, and processing assisted by H.W., M.P., and C.L. and overseen by W.-J.T., M.Z., B.C., and C.S.P. A.J.N. performed cryoET data acquisition, and A.J.N., W.G.L., and J.M.M. performed the analysis. W.G.L., W.-J.T., and M.Z. built and refined cryoEM structural models. W.G.L. and S.M. performed protein purification and crystallographic data collection, built, and refined structural models. W.G.L. performed protein purification for HDX-MS and D.L. and S.L. performed HDX- MS and analysis. W.G.L. purified proteins and W.G.L. and J.M.M. performed SAXS studies. J.W. performed the inhibitor screen and characterization of MB60 and J.W. and C.M.K. provided critical reagents for the crystallographic analysis. W.G.L., W.-J.T., J.V.L.K., J.M.M., and S.L. wrote the paper. S.M., J.W., M.Z., and B.C. contributed to the paper finalization. Competing interests The authors declare no competing interests. Additional information Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41467-022-29322-4. Correspondence and requests for materials should be addressed to Wei-Jen Tang. Peer review information Nature Communications thanks Claudio Ciferri, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available. Reprints and permission information is available at http://www.nature.com/reprints Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. NATURE COMMUNICATIONS | (2022) 13:1833 | https://doi.org/10.1038/s41467-022-29322-4 | www.nature.com/naturecommunications 13 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29322-4 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/ licenses/by/4.0/. © The Author(s) 2022 14 NATURE COMMUNICATIONS | (2022) 13:1833 | https://doi.org/10.1038/s41467-022-29322-4 | www.nature.com/naturecommunications
10.1038_s41467-023-35820-w
Article https://doi.org/10.1038/s41467-023-35820-w Reconciling discrepant minor sulfur isotope records of the Great Oxidation Event Received: 19 May 2022 Accepted: 3 January 2023 Benjamin T. Uveges Roger E. Summons 1 1 , Gareth Izon 1, Shuhei Ono 1, Nicolas J. Beukes2,3 & Check for updates ; , : ) ( 0 9 8 7 6 5 4 3 2 1 ; , : ) ( 0 9 8 7 6 5 4 3 2 1 Understanding the timing and trajectory of atmospheric oxygenation remains fundamental to deciphering its causes and consequences. Given its origin in oxygen-free photochemistry, mass-independent sulfur isotope fractionation (S-MIF) is widely accepted as a geochemical fingerprint of an anoxic atmo- sphere. Nevertheless, S-MIF recycling through oxidative sulfide weathering— commonly termed the crustal memory effect (CME)—potentially decouples the multiple sulfur isotope (MSI) record from coeval atmospheric chemistry. Herein, however, after accounting for unrecognised temporal and spatial biases within the Archaean–early-Palaeoproterozoic MSI record, we demon- strate that the global expression of the CME is barely resolvable; thereby validating S-MIF as a tracer of contemporaneous atmospheric chemistry dur- ing Earth’s incipient oxygenation. Next, utilising statistical approaches, sup- ported by new MSI data, we show that the reconciliation of adjacent, yet seemingly discrepant, South African MSI records requires that the rare instances of post-2.3-billion-year-old S-MIF are stratigraphically restricted. Accepting others’ primary photochemical interpretation, our approach demands that these implied atmospheric dynamics were ephemeral, operating on sub-hundred-thousand-year timescales. Importantly, these apparent atmospheric relapses were fundamentally different from older putative oxy- genation episodes, implicating an intermediate, and potentially uniquely feedback-sensitive, Earth system state in the wake of the Great Oxidation Event. Atmospheric oxygenation represents perhaps the most profound chemical change experienced by the Earth system, revolutionising biogeochemical cycles1–4 while priming the planet for the rise of complex life5–9. Although a wealth of geological observations and geochemical data have been combined to articulate the story of Earth’s oxygenation1,5,10,11 (Fig. 1a, b), the recognition of mass-independent sulfur isotope fractionation (S-MIF) within the geological multiple sulfur isotope (MSI) record remains the most widely accepted means to directly trace the oxygen content of Earth’s early atmosphere12,13. Manifest as non-zero Δ3X-values (“Methods”; Eqs. 1 and 2)12,14,15, the generation and geological preservation of S-MIF is linked to low atmospheric oxygen concentrations in several ways: Starting with its production, anoxic SO2 photochemistry remains the only experimen- tally demonstrated means of creating S-MIF that remotely resembles the geological record13 (Fig. 1b). Here, in the absence of photon- shielding O2/O3, photolysis and/or photo(de)excitation reactions are understood to yield reduced (i.e., S0 or S8) and oxidised (i.e., H2SO4) that carry positive and negative Δ33S values, sulfur phases 1Department of Earth, Atmospheric and Planetary Sciences, Massachusetts Institute of Technology, 77 Massachusetts Avenue, Cambridge, MA 02139, USA. 2DSI-NRF Centre of Excellence for Integrated Mineral and Energy Resource Analysis, Department of Geology, University of Johannesburg, P.O. Box 524, Auckland Park 2006, South Africa. 3Deceased: Nicolas J. Beukes. e-mail: buveges@mit.edu Nature Communications | (2023) 14:279 1 Article https://doi.org/10.1038/s41467-023-35820-w a b ] ) M T A ( 2 O p [ g o l -1 -3 -5 ) ‰ ( S 3 3 Δ 15 10 5 0 −5 Riet. . m F l l i H l l a b e m T i . u . m F l l i H l l a b e m T i . l Ro. 400 200 0 c ) m ( y r a d n u o B H B T − . i o o R o t d e a c S h t p e D l Oxygenic Photosynthesis? Stromatolites Whiffs? Oxygen-Sensitive Detrital Grains 4000 3000 d p O 2 ( P A L ) 100 10-2 10-4 Steranes Animal Body Fossils Sulfate Deposits Spot Bulk Fe in Palaeosols & Red Beds Manganese Deposits Age (Ma) 2000 1000 0 Poulton et al., 2021 EBA-2 Luo et al., 2016 Izon et al., 2021 KEA-4 2,256 ± 6 Ma 2,310 ± 9 Ma 2,316 ± 7 Ma < 2,353 ± 18 Ma Gatsrand Mbr. Rooi.−TBH boundary 0 2 4 6 8 0 2 4 6 8 Δ33S (‰) Δ33S (‰) Chert Iron Formation Volcanics, Dolerite, Diabase Gray Shale, Siltstone Diamictite Carbonate Black Shale Sandstone, Quartzite ) m ( h t p e D e CB BQ m F . z a K KF 2 1 B M n o i t a m r o F a r r a g n u K F B R W 100 150 200 300 100 140 Killingsworth et al., 2018 (Barite) Phillipot et al., 2018 (Pyrite) > 2,200 Ma T 3 2,312.7 ± 5.6 Ma < 2,340 ± 22 Ma T 2 T 1 < 2,454 ± 23 Ma 0 2 4 Δ33S (‰) 6 8 Fig. 1 | Synthesised interpretations of Earth’s oxygenation. a A schematic representation of competing ideas surrounding the oxygenation of Earth’s atmosphere1 contextualised within a framework of biological innovations. b The secular evolution of Δ33S values compiled herein (Supplementary Data File 1). The grey box locates the interval over which atmospheric oxygen is thought to have first accumulated, while the horizontal bars illustrate more traditional redox indicators, whose red and blue colouration discriminates between those disclosing oxic and anoxic conditions, respectively4,5. Triangular and circular points represent spot and bulk Δ33S analyses respectively. Stratigraphic distributions of Δ33S data from the South African Kaapvaal (c core EBA-2; d core KEA-4) and Western Aus- tralian Pilbara (e cores T1–T3) cratons, whose discrepant interpretations have led to the emergence of conflicting models of atmospheric oxygenation. Each of the chronologically constrained lithostratigraphic columns follow those presented with the Δ33S data22,25,30,32,33, with the superimposed vertical grey bars illustrating the ±0.3‰ threshold for identifying S-MIF by Δ33S alone25. The lowermost hor- izontal blue band in c, d corresponds to the apparently sustained presence of mass- independent sulfur isotope fractionation (S-MIF) within the Rooihoogte Formation (Ro(oi)) seen throughout the Carletonville area22,25,30,37. The subsequent blue hor- izontal bars in panel c mark isolated instances of S-MIF within the Timeball Hill Formation (TBH) that have been interpreted to represent returns to an anoxic atmospheric state (Supplementary Information). Contrastingly, the persistence of low-magnitude non-zero Δ33S values throughout cores T2–T3 has been ascribed to the crustal memory effect, with atmospheric oxygenation occurring much earlier (blue horizontal bar, e, Supplementary Information). Lithostratigraphic abbrevia- tions: Reit. Reitfontain Member, WR Woongarra Rhyolite, BF Boolgeeda Iron For- mation, MB1, 2 the diamictites within the Meteorite Bore Member, KF Koolbye Formation, Kaz. Fm. Kazput Formation, BQ Beasley River Quartzite, CB Cheela Springs Basalt. respectively12,13,16–18. Importantly, one-dimensional photochemical models indicate that atmospheric oxygen concentrations must remain below 0.001% of modern levels to retain the photochemical exit channels necessary to communicate S-MIF to the Earth’s exogenic in terms of preservation, oxygen-deficient sulfur cycle19. Further, environments are less likely to homogenise these photochemical derivatives, thereby increasing their preservation potential and, thus, their overall likelihood of passing into the sedimentary record20. Understandably, it follows that the geological presence of S-MIF is now almost unanimously accepted as a robust geochemical fingerprint of a long-lived oxygen-free atmospheric state, persisting for more than half of Earth’s 4.5-billion-year (Ga) history. Despite the clear and direct links between oxygen and S-MIF, constraining the exact timing and nature of atmospheric oxygenation Nature Communications | (2023) 14:279 2 Article https://doi.org/10.1038/s41467-023-35820-w has proven difficult. Although it is generally agreed that oxygen began to accumulate within the atmosphere between 2.5–2.3 Ga, different readings of the S-MIF record have been used to portray dramatically different oxygenation trajectories—with some workers arguing for a 21–23, while others envisage unidirectional geologically rapid rise in O2 an oscillatory trajectory whereby pO2 repeatedly crossed the threshold necessary to resume S-MIF genesis24–26 (Fig. 1a–d). While inadequate age constraints and ambiguous stratigraphic correlations22–24,26–30 currently prevent discrimination between these models (Supplemen- tary Information), both interpretations are reliant on the assumption that sedimentary S-MIF chronicles contemporaneous atmospheric conditions. Again, this premise is not universally accepted and some contend that the operation of a so-called crustal memory effect (CME) decouples the MSI record from coeval atmospheric chemistry16,31–33. Specifically, it is argued that the late-Neoarchaean–Palaeoproterozoic MSI record is dominated by S-MIF inherited from an adulterated sea- water sulfate reservoir supplied by the oxidative weathering of older crustal pyrite, rather than being a syndepositional atmospheric signal (Fig. 1e). Hence, before we can begin to reconcile the conflicting nar- ratives of atmospheric oxygenation, the feasibility and potential magnitude of a CME must first be addressed. Results and discussion The crustal memory effect Noting a post-2.45 Ga shift toward muted positive Δ33S values, Farqu- har and Wing were the first to suggest that weathering-induced crustal recycling of S-MIF may hinder our ability to precisely reconstruct atmospheric oxygenation from the MSI record16. Subsequently, pre- dicated on a crustal 33S excess16,18,31,34,35 manifest as a cumulative data- base mean Δ33S value of ~1.5‰31, Reinhard et al. presented a series of box-model simulations designed to quantify the CME31. Their proof-of- concept approach demonstrated that once mass-independent weath- ering fluxes permeated an entirely mass-dependant marine sulfur cycle, the weathering-derived non-zero Δ33S values exceeded those imparted by mass-dependent processes (e.g., >|0.3‰|)36 for upwards of 10–100 million years31. Importantly, such longevity has the potential to induce a temporal lag between the cessation of atmospheric S-MIF production and its loss from the rock record, thereby masking the operation of atmospheric chemistry during this transformative inter- val of Earth history31. While alarming, the paucity of demonstrably contemporaneous MSI datasets has prevented empirical scrutiny of these computational predictions, leaving the significance of the CME untested. Even now, despite a recent explosion of data21,22,25,30,32,37, opinions surrounding the CME remain divided21,22,25,26,30,32,33. For instance, while emerging data from the Western Australian Turee Creek Basin support the notion of a long-lived CME32,33, purportedly time equivalent (Supplementary Information) observations from the Carletonville area of the South African Transvaal Basin do not (Fig. 1c–e)22,25,37—thereby questioning the operation of a CME as a global phenomenon. Furthermore, these spatial differences imply unrecognised biases within the MSI, with the important corollary that an unweighted cumulative mean may inef- fectively approximate the isotopic composition of the weatherable sulfur pool, thus confounding previous estimates of the persistence and magnitude of the CME31,32. To explore the source of this data–model mismatch, we start by compiling a comprehensive MSI database (“Methods”; Supplementary Data 1), which we then examine in its raw form and, in a reduced state wherein we compensate for potential sample density biases via aver- aging replicate and intra-sample analyses, termed the Spot Sample Averaged(SSA) database (“Methods”; Supplementary Information). Here, while a cursory evaluation of this new compilation confirms the plateau used by Reinhard et al.31 to infer the capture of a representative global mean Δ33S value, augmentation with the data published over the ensuing decade reveals a subtle downward trajectory, resulting in a pre-2.3 Ga mean Δ33S value of 0.74 ± 1.97‰ (1σ) for the entire database, increasing to 1.12 ± 2.06‰ (1σ) the SSA database is con- sidered (Fig. 2a). if A more detailed interrogation of the new MSI database reveals lithological, spatial, and temporal biases (Fig. 2; Supplementary Information). Though the lithological bias (e.g., carbonates vs. silici- clastics) was not found to significantly influence our analyses (“Meth- ods”; Supplementary Information; Supplementary Figs. 4 and 5), the impacts of the temporal and spatial biases were substantial. For example, even after compensating for sample replication via the SSA database, more than 70% of the pre-2.3 Ga data were found to originate from either the Kaapvaal or Pilbara cratons, with roughly two-thirds (35.5%) and Palaeoproterozoic-aged derived from Neoarchaean- (33.3%) successions (Fig. 2b, c). To estimate how these biases impact the database mean, we subjected the SSA database to bias-specific bootstrap sampling routines, arriving at two synthetic datasets, which we term the Temporally Adjusted (TA) and Craton Adjusted (CA) datasets (“Methods”). Compared to the mean Δ33S value obtained from the unweighted SSA database (1.12‰), the mean Δ33S values obtained from the TA and CA datasets were found to be lower, averaging 0.62 and 0.36‰, respectively (Fig. 2a). Consequently, this simple exercise confirms our data-based inferences, demonstrating that the magni- tude of the positive Δ33S skew apparent within the rock record is, at least partially, a function of bias. To determine how these biases propagate into our understanding of the size and longevity of the CME, we then incorporated the TA and CA datasets into an updated geochemical model22 (“Methods”). Aug- menting the model with a bootstrap sampling subroutine, the Δ33S of the seawater sulfate pool was computed utilising the size and isotopic composition of the major marine sulfate fluxes and sinks (“Methods”; Supplementary Information). Here, using the TA dataset, the esti- mated maximum Δ33S value of the seawater sulfate reservoir (SWSR) was found to be 0.35 ± 0.2‰ (1σ), which was further reduced to 0.21 ± 0.17‰ (1σ) when utilising the CA dataset (Fig. 3a–b). These bias- corrected estimates of the CME are much smaller than those reported by Reinhard et al.31 (centred on 1.5‰ and up to 2.5‰) and, indeed, those derived from our updated unadjusted database (0.65 ± 0.6‰; Fig. 3c; “Methods”), thus reaffirming the importance of our bias- correction routine. That said, while small, the fact that the globally integrated CME remains slightly positive despite our bias corrections is a pertinent observation. Inspection of the MSI record shows that the final appreciably negative Δ33S value occurs at least 60 million years before its positive Δ33S counterpart (Fig. 1b; Supplementary Informa- tion). This asymmetry implicates a muted positive CME persisting beneath an atmosphere capable of S-MIF genesis. Indeed, support for the early onset of a subdued CME is found within a temporal synthesis of molybdenum abundance and isotope data that has similarly been interpreted to signal the onset of oxidative pyrite weathering before pO2 rose sufficiently to curtail S-MIF-yielding photochemistry38. These observations, therefore, begin to reconcile conflicting interpretations derived from various atmospheric and marine-based geochemical proxies30,38, yielding a complementary picture of Earth’s oxygenation. Reminiscent of Reinhard and colleagues’ predictions31, weathering-derived Δ33S values have long residence times within the SWSR, failing to return to zero within the timespan of our model. Nevertheless, our approach significantly reduces the magnitude of the CME-derived Δ33S values, arriving at a conservative maximum of approximately 0.5‰, with <0.3‰ being more likely (Fig. 3). This estimate is more consistent with Δ33S values from Palaeoproterozoic- aged sulfate minerals (|0.1–0.2|‰) that should, in principle, record the coeval SWSR composition39. Ergo, beyond validating the MSI record as an archive of Earth’s incipient oxygenation, our results can be considered as a conservative discriminatory threshold (i.e., 0.5‰), separating weathering-induced non-zero Δ33S from those derived from syndepositional oxygen-free photochemistry. values Nature Communications | (2023) 14:279 3 Article https://doi.org/10.1038/s41467-023-35820-w a 1.5 1.0 0.5 0.0 c Mesoprot.−Present TA CA Palaeoproterozoic Neoarchaean Craton All Kaapvaal + Pilbara Other Mesoarchaean 2000 2005 2010 2015 2020 100 b 75 50 25 0 Year of Publication Palaeoarchaean Eoarchaean 0 1000 2000 3000 Number of Analyses ) ‰ ( S 3 3 Δ n a e M e v i t a u m u C l ) % ( e g a t n e c r e P e v i t a u m u C l Fig. 2 | Spatial and temporal biases within the Δ33S record. a A cumulative plot depicting the evolving mean Δ33S value of the pre-2.3 Ga data in the com- piled multiple sulfur isotope (MSI) database, with solid and dashed lines repre- senting those derived from the full database and its spot sample averaged (SSA) counterpart, respectively (“Methods”). Line colour distinguishes the craton(s) being analysed, with green signalling the full database, blue illustrating the com- bined Kaapvaal–Pilbara datasets, and orange showing the remainder. For refer- ence, the respective mean Δ33S values of the Time Adjusted (TA) and Craton Adjusted (CA) synthetic datasets are plotted as a blue-coloured circle and triangle. A vertical dashed line separates the data considered by Reinhard et al.31 (i.e., ≤2012) from those published subsequently, while the horizontal dashed line denotes the 0.3‰ Δ33S-threshold for identifying mass-independent sulfur isotope fractionation (S-MIF)25. b The cumulative percentage of the Kaapvaal–Pilbara-derived data (blue) relative to those sourced from elsewhere (orange). The horizontal dashed line denotes 50% of the dataset, while the x-axis scaling follows (a). c A time-binned distribution of Δ33S data within the unfiltered database. Mesoprot. abbreviates Mesoproterozoic. a − Temporally Adjusted (TA) Dataset b − Craton Adjusted (CA) Dataset c − Unadjusted SSA Database 3 2 1 0 −1 ) ‰ ( S 3 3 Reinhard et al. CME Weathered Sulfur SWSR 3200 2800 2400 Age (Ma) 2000 3200 2800 2400 Age (Ma) 2000 3200 2800 2400 Age (Ma) 2000 Fig. 3 | The simulated evolution of the crustal memory effect (CME). Here, the presented simulations are derived from exploration of a the Temporally Adjusted (TA), b the Craton Adjusted (CA) and c the unadjusted spot sample averaged (SSA) datasets. Colours are common throughout, with blue depicting the Δ33S value carried by the weathered sulfur flux and orange showing the Δ33S composition of the seawater sulfate reservoir (SWSR). Colour-coded lines depict the LOESS smoothed mean of the bootstrap outputs, while their envelopes indicate their 1σ uncertainty. The horizontal grey band depicts the existing CME estimate derived using the most feasible boundary conditions31, while the dashed lines bracket the 0.3‰ Δ33S-based S-MIF discriminatory threshold. Nature Communications | (2023) 14:279 4 Δ Article https://doi.org/10.1038/s41467-023-35820-w a 10 Turee Creek Group Pilbara Craton ) ‰ ( S 3 3 2 1 0 2500 2300 2100 3200 b 2800 2400 2000 Age (Ma) ) ‰ ( S 3 3 ) ‰ ( S 3 3 5 0 5 2 1 0 T1 T2 T3 Fig. 4 | Simulations of the crustal memory effect (CME) within a hypothetical partially hydrographically restricted basin adjacent to the Pilbara Craton. a Here, the blue line indicates the simulated Δ33S evolution of the Pilbara-specific weathering flux, with its associated envelope capturing the 1σ uncertainty. Super- imposed on these, the available Pilbara-derived Δ33S data are plotted in blue, while those from the Turee Creek Group are distinguished as red open symbols32. The vertical dashed line marks the Great Oxidation Event(GOE) sensu Luo et al.22, while the horizontal dashed line represents the globally expressed CME modelled using the CA dataset (Fig. 3b). The 2500–2000 Ma expansion (insert) follows (a). b Core- specific density distributions of the Δ33S data from the Boolgeeda Iron Formation and Turee Creek Group32, younging from left to right (i.e., cores T1→T3). The hor- izontal blue line and envelope follow (a). Returning to the MSI record, the small magnitude of the modelled CME, coupled with the fact that less than 5% of Δ33S values within the entire pre-2.3 Ga database exceed 5‰ (Supplementary Information; Supplementary Fig. 3), strongly supports the interpretation that large magnitude Δ33S values must originate from oxygen-free pho- tochemistry. Examples of sustained large values are last seen within the Rooihoogte Formation, Transvaal Basin (≤8.76‰), therefore signalling the presence of an oxygen-impoverished atmosphere well after 2.45 Ga22,25,30(Supplementary Information). the instance, Spatial heterogeneities in the expression of the CME Independent support for a diminished CME is found within the MSI systematics of contemporary riverine sulfates derived from Archaean- circum-zero Δ33S values aged terrains40. For (−0.01 ± 0.10‰) found within rivers draining the S-MIF-bearing Superior Craton empirically demonstrate our prediction that catchment-level integration modulates the magnitude of Δ33S values communicated to the SWSR via weathering. Nevertheless, measure- ments of sulfates from other riverine systems41 and, indeed, a South African groundwater sample40 do, in fact, feature moderate Δ33S values, demonstrating at least some capacity for the communication of 33S anomalies to weathered sulfur fluxes. Accordingly, given the clearly uneven distribution of large Δ33S values (Fig. 2; Supplementary Table 3, Supplementary Information), a hinterland featuring large Δ33S values draining into a (semi-)isolated, sulfate-impoverished, basin could hypothetically combine to promote a more pronounced and spatially heterogeneous CME. Recalling that the Pilbara Craton features some of the most pro- nounced Δ33S values seen within the MSI record (Fig. 4; Supplementary Table 3), not to mention the pronounced dichotomy seen between the Australian and South African MSI records (Fig. 1)25,30,32, we modified our bootstrap sampling subroutine to test whether the apparent persis- tence of S-MIF within the foreland Turee Creek Basin26,32 could reflect regional amplification of the CME (“Methods”). Here, drawing from a TA dataset comprising data derived solely from the Pilbara Craton, the resultant weathering flux was found to carry a much larger Δ33S value (0.68 ± 0.5‰, 1σ; Fig. 4a) when compared to our estimates of a globally felt CME. Furthermore, the good 1σ agreement between our model output and the Δ33S data obtained from the two youngest cores (T2 & T3) retrieved during the Turee Creek Drilling Project32 supports pre- vious calls for a weathering-derived component within the Turee Creek MSI record (Fig. 4a, b)32,33. Taking this further, the apparent lithological control seen in the δ34S and Δ33S systematics of the Kazput Formation32,33 implies a higher order depositional control (T3, Fig. 1e). Here, muted Δ33S values reminiscent of those we estimate for the globally integrated CME are captured within the more distal fine- grained facies, while the near-shore carbonates seem to record more pronounced Δ33S values akin to those we compute for our Pilbara- derived weathering flux (Figs. 1e, 4). These observations demand that the CME should not be considered in a vacuum and, in reality, the expression of any CME is likely dependent on the interplay between the catchment-specific sulfate weathering flux and sulfate derived from a globally mixed SWSR. This premise is further supported by the fact that barites from the Kazput Formation are distinctly different from SWSR estimates derived from globally-distributed sedimentary sulfates of similar age39 and, instead, mirror cooccurring sedimentary sulfide Δ33S values33(Fig. 1e). The persistence of S-MIF within Western Australia beyond its demise in South Africa, therefore, need not question the available radiometric age constraints on the Meteorite Bore Member that are used to link the Turee Creek and Transvaal MSI records26,32,42–44 (Fig. 1e; Supplementary Information). Contrasting with the monotonous and muted Δ33S values observed within the Turee Creek Group, several discrete instances of elevated Δ33S values of up to 2.9‰ are observed to perturb a near-zero baseline within the Timeball Hill Formation, South Africa25. While the Kaapvaal Craton is also known to harbour elevated Δ33S values, appli- cation of an analogous craton-specific CME simulation returns a broadly similar weathering flux (0.84 ± 0.6‰), whose theoretical maximum (c.1.5‰) is barely half of the most extreme Δ33S value reported from the Timeball Hill (Figs. 1 c, 5; Supplementary Table 3; Supplementary Fig. 7). Consequently, the pronounced, yet highly variable, nature of these post-2.3 Ga 33S-enrichements strongly argue against derivation from a regionally amplified CME and, instead, require an alternate explanation. Constraining the duration of post-2.3 Ga S-MIF occurrences The CME as previously defined31,32, and adopted herein, does not account for the physical delivery of erosion-derived sulfides. While such detritus has been widely publicised as a potential source of atmospherically estranged sulfur with Δ33S values of effectively any magnitude29,45,46, it is hard to envisage a scenario where detrital signals can repeatedly dominate the MSI record without detection—a stance strengthened given the apparent predominance of authigenic pyrites throughout the Carletonville area22,25,30,37. Consequently, while we await a detailed and dedicated grain-scale isotopic appraisal designed to unequivocally dismiss a detrital explanation of the elevated Time- ball Hill Δ33S values, it seems rationalisations reliant on recycling can be dismissed. Indeed, Poulton and colleagues25 argued that the elevated Timeball Hill Δ33S values recorded primary atmospheric dynamics (Fig. 1c). Enigmatically, however, as testified by a more recent quad- ruple sulfur isotope appraisal reporting only sub-0.3‰ Δ33S values within a proximal (<5 km) and indisputably equivalent core (KEA-4), these supposed atmospheric oscillations are not universally resolved30 (Fig. 1c, d). Consequently, providing the inferred atmospheric sig- nificance of the elevated Δ33S values seen within the Timeball Hill Formation holds true, then the recognised kilometre-scale dichotomy apparent within the Carletonville area requires that the non-zero Δ33S values must persist over sufficiently stratigraphically restricted inter- vals to escape detection via conventional sampling campaigns. To constrain the duration of the Timeball-Hill-housed S-MIF- bearing intervals and, therefore, elucidate their mechanistic driver(s), Nature Communications | (2023) 14:279 5 Δ Δ Δ Article a ) m ( i w o d n W F M S - I f o i s s e n k c h T e g a r e v A 10 5 2 1.42 1 0.5 0.2 0.02 Centered on S-MIF data of Poulton et al. (2021) https://doi.org/10.1038/s41467-023-35820-w Data Source Poulton et al.(2021) This Study b 800 900 1000 ) m ( h t p e D 1100 1200 1300 c 1.5 m 2.5 m 804 807 2 m 1012 ~40 cm 1275.5 1330 3 cm 1.4 m 1335 1 2 33S (‰) 3 0 20 40 60 Percentage (%) of S-MIF-Bearing Samples 0 2 4 6 0 33S (‰) Fig. 5 | Statistical- and data-based constraints on the longevity of post- Rooihoogte-aged S-MIF-bearing intervals. a Colour-coded frequency distribu- tions resulting from 1000 replicate 60-sample bootstrap sampling routines, tar- geting synthetic cores with variably thick mass-independent sulfur isotope fractionation (S-MIF)-bearing windows (y-axis) positioned using the pre-existing EBA-2 Δ33S dataset25, as exemplified using the sample from 1224.0 m core depth (insert; Supplementary Fig. 10). The median and 25th and 75th quartiles are shown as vertical white lines. b Pre-existing Δ33S data from core EBA-225 (blue circles), aug- mented with our new quadruple sulfur isotope(QSI) measurements (red triangles), are superimposed on the 0.3‰ Δ33S-based S-MIF identification threshold (vertical grey band). The green shading in (b) locates the intervals shown in (c), providing an expanded view of four stratigraphic intervals where Δ33S values exceeding 0.3‰ were reported25. Given the bounding mass-dependant samples, the annotated vertical orange lines denote the maximum thickness of S-MIF-bearing intervals, which, in reality, are likely much smaller; their constraint, however, awaits detailed high-resolution chemical and petrographic conformation. The colour coding fol- lows (b). we started with a simplified Bayesian approach (“Methods”; Supple- mentary Fig. 9). Here, after converting from the relative abundance of S-MIF in each core, we found the most likely average stratigraphic thickness of each S-MIF re-emergence was 1.42 m, equating to some 0.3–3 million years when combined with average compacted sedi- mentation rates (“Methods”). Significantly, however, while this Bayesian-derived estimate best explains the available data, Izon and colleagues’ inability to detect any S-MIF within their higher-resolution search of the Timeball Hill Formation30 (Supplementary Table 5) remains salient, emphasising the inherently conservative nature of this estimate (“Methods”). Accordingly, we designed a series of simulations to constrain the likelihood of avoiding S-MIF within a 60-sample sampling campaign akin to that reported by Izon et al.30 (Fig. 5a, Supplementary Fig. 9; “Methods”; Supplementary Table 5). Strikingly, however, from our 1000 bootstrap replicates, only twice (i.e., 0.2% of the replicates) were we able to avoid detecting S-MIF after prescription of the Bayesian-derived 1.42-metre-thick S-MIF window. Rather, given these boundary conditions, on average, we detected S-MIF 6 ± 2 (1σ) times per replicate, equating to 7–13% of the population size (Fig. 5a; Supplementary Table 6). Extending our statistical approach to increasingly narrower S-MIF windows reveals that their thickness must recede to well below a metre before the likelihood of avoiding S-MIF becomes statistically feasible, implying, in turn, that any re-emergence of S-MIF was likely short lived (Fig. 5a; Supplementary Table 6). To empirically test the statistically inferred brevity of these S-MIF- bearing intervals, we analysed the MSI systematics of additional sam- ples from core EBA-2, focusing on those bracketing the most pro- nounced Δ33S value reported from the Timeball Hill Formation (+2.9‰; “Methods”; Fig. 5; Supplementary Data 2)25. Here, as predicted, several samples within approximately 30 cm of the S-MIF-bearing sample found at 1275.5 m core-depth were found to possess mass-dependent MSI systematics, demonstrating that S-MIF was restricted to, at most, 40 centimetres of core (Fig. 5b, c). The presence of several other stratigraphically isolated S-MIF-bearing samples within a metre of those possessing mass-dependant MSI systematics25 (Fig. 5b, c) sup- ports our inferences and justifies their extension to the wider Timeball Hill Formation. Combining these analytical observations with the preceding statistical analyses reduces the temporal duration of the post-2.33 Ga instances of S-MIF by an order of magnitude, thus sig- nalling the return of S-MIF for perhaps tens-of-thousands of years. Implications for the operation of the atmosphere after 2.33 Ga Having established that any occurrences of S-MIF in the Timeball Hill Formation must have been short lived, we can now begin to assess the implications for the operation of the atmosphere after 2.3 Ga. First, acknowledging that c.40% of the pre-2.3 Ga Δ33S record falls within 0.3‰ of zero (Supplementary Fig. 8), it could be argued that the ele- vated Δ33S values seen within the Timeball Hill Formation signal deposition beneath an oxygen-free backdrop. Here, rather than a repeated return to an S-MIF yielding anoxic atmospheric state, this stance requires that the Timeball-Hill-housed S-MIF recurrences record some unknown localised preservation bias, leading to only sporadic expression of elevated Δ33S values. Such an interpretation, however, beyond contradicting the available Δ36S/Δ33S data, is also inconsistent with the wider Δ33S database. For instance, if we exclude the data from the crustally influenced Turee Creek Group, the over- whelming majority (>90%) of the 2.3–2.1 Ga Δ33S record comprises values between –0.3 and 0.3‰ (Supplementary Fig. 8). Thus, rather than mirroring the older MSI record derived from an almost unan- imously accepted anoxic background state, the Timeball-Hill-derived Nature Communications | (2023) 14:279 6 Δ Δ Article https://doi.org/10.1038/s41467-023-35820-w Δ33S distribution is much more reminiscent of the post-GOE record. Within this framework, the loss of S-MIF within the Rooihoogte For- mation appears to mark a fundamental transition within the Earth system, with the global predominance of S-MIF seen prior to the Timeball Hill Formation (Supplementary Fig. 8) signalling a dominantly anoxic background state that yields to a more oxidising equivalent at around 2.3 Ga22,30. Accepting the premise that the post-2.3 Ga atmosphere was pre- dominantly oxygenated enough to inhibit S-MIF production/pre- servation, we are left with describing a mechanism for the short-lived re-emergences of S-MIF in the Timeball Hill Formation. Given that stratospheric volcanic eruptions are known to communicate Δ33S values as large as |2.0|‰ to contemporary ice-core records47–49, such events could hypothetically act as a geologically instantaneous source of S-MIF in more ancient settings. Indeed, the sulfate-impoverished oceans of the Palaeoproterozoic22,45,50,51, in principle, would offer little buffering capacity against such a point-source injection of S-MIF. That said, while S-MIF genesis and preservation against a modern oxyge- nated backdrop is a good starting point, its application to the Timeball Hill Formation and, indeed, the wider sedimentary record is proble- matic. Aside from the distinct differences between the ice-core (−4.3)49 and rock-housed Δ36S/Δ33S records (–0.9)14, their mode of preservation is also fundamentally different. In essence, the survival of ice-core- housed Δ33S anomalies is not only a function of evading dilution with mass-dependently fractionated sulfur, but it specifically requires time- resolved differential rainout. Here, rather than spatially separating two 2-), the ice-core isotopically distinct sulfur phases (i.e., S8 and SO4 housed S-MIF bearing sulfates are separated in time over the course of several years by the relatively rapid and pristine accumulation of the snow/ice pack47–49. In the absence of this time separation, mass-balance requires that 33S anomalies would be eradicated when the two pools were re-mixed47. Given the marine setting of the Timeball Hill Forma- tion, it is realistic to assume that any volcanogenic sulfates would be homogenised, and therefore any 33S anomaly would be erased by mixing within the SWSR and the subsequent diagenetic realm. Theo- retically, these issues could be circumnavigated by the model pro- posed by Gallagher et al., who envisage the volcanic injection of subduction-derived recycled sedimentary sulfur carrying positive Δ33S values52. While we concede that such a model offers an intriguing, not to mention potentially substantial, point-source of S-MIF, recalling the logic presented previously for the CME, we dispute its relevance in this case. Rather, we anticipate that, similar to the mixing that occurs under surficial weathering, homogenisation within the subduction system would minimise the Δ33S value of the re-entrained sulfur pool to values smaller than those seen in the Timeball Hill Formation (Supplementary Information, Supplementary Table 4.1–4.3). All told, we are unable to explain the transient post-2.3 Ga instances of S-MIF without invoking high-order atmospheric dynamics with episodic relapses toward an anoxic atmospheric state. While we stress that the pre-Rooihoogte-aged oxidation episodes await defini- tive demonstration (Supplementary Information), it is clear that the S-MIF variability seen within the Timeball Hill Formation implicates atmospheric dynamics operating on fundamentally different time- scales than their alleged precursors24–26. Such a system not only requires a background state susceptible to transient atmospheric perturbations, but also one that is conducive to S-MIF preservation. Preservation of significantly positive S-MIF signatures is thought to require a small SWSR53, which implies slow rates of oxidative weath- ering of continental sulfides and, by extension, relatively low pO2. It follows, therefore, that the periodic reappearance of S-MIF in the Timeball Hill Formation may signal a protracted, intermediate, and extremely sensitive atmospheric state that was uniquely susceptible to perturbation as oxygen contents vacillated around the threshold for S-MIF genesis and preservation. Speculatively, such a state could have been established and maintained through the interplay of biological feedbacks54,55 encountered as organisms gradually evolved the bio- chemical machinery to thrive in increasingly more oxidising regimes5. Superimposed on this intermediary atmospheric state, large injections of reducing gasses, perhaps sourced via the emplacement of large igneous provinces or, more likely, through climate-paced methane fluxes, could have episodically outpaced the biological O2 flux to ephemerally reinstate S-MIF production56. Interestingly, support for these claims can be found in recent photochemical modelling efforts that demonstrate the feasibility of 10–100 thousand-year returns to oxygen-free S-MIF yielding atmospheres, providing the predominant boundary conditions remain near to those necessary for S-MIF genesis56. The prediction of a highly dynamic and intermediate atmo- spheric state represents a paradigm shift, departing from the con- ventional view that atmospheric oxygenation occurred as a step function between bistable endmember compositions57,58. As such, there is now a pressing need for more detailed high-resolution MSI studies to better constrain the nature and duration of this transitionary period, and thereby refine the framework of Earth’s most significant redox revolution. Methods Isotopic nomenclature Following convention, sulfur isotope data are expressed in delta (δ) notation, reflecting permille (‰) deviations of the less abundant iso- tope, X (33,34,36S), normalised to 32S, relative to the same ratio in the international reference standard, Vienna Canyon Diablo Troilite, (VCDT): δ3X S = ½ð 3X S= 32SÞ =ð 3X S= 32SÞ VCDT (cid:2)1(cid:3) × 1000 sample ð1Þ Most processes fractionate S-isotopes mass-dependently, whereby δ3XS ≈ λ3X × δ34S, with respective λ3X values of 0.515 and 1.91 for δ33S and δ36S. Departure from mass-dependent behaviour, termed mass-independent fractionation (MIF), however, is expressed in capital-delta (Δ) notation, where: Δ3X S = 1000 × ½lnðδ3X S=1000 + 1Þ(cid:2)λ 3X × lnðδ34S=1000 + 1Þ(cid:3) ð2Þ Database description The sulfur isotope data used herein (n = 10,765 with at least Δ33S values) were compiled from 84 studies, forming a new and com- prehensive database complete with metadata (Supplementary Infor- mation; Table 1; Supplementary Data 1). Given that many studies fail to report δ33S and δ36S values, to assess the effects of non-linear mixing in Δ–Δ space14,36 we calculated the missing δ-values via algebraic rear- rangement of Eqs. 1 and 2. Overall, however, our presented solutions were found to be broadly equivalent irrespective of whether the mean Δ33S values were used directly or whether the Δ33S values were calcu- lated retrospectively. To remove complications arising from the myr- iad of published sample lithological descriptions, our assessment of lithological bias uses a simplified parameter (simp.lith), allowing the compiled data to be allotted to one of ten generalised lithotypes (i.e., high energy siliciclastic, low energy siliciclastic, carbonates, sulfates, chemical, microbial, altered/late addition, glacial/diamictite/detrital, igneous/volcanic sed., and none provided). The individual lithologies assigned to each of our generalised lithotypes are listed in Supple- mentary Table 6. Finally, to ensure that inclusion of replicate sample analyses did not impact our analyses and resultant conclusions, we compared the full database to a Spot Sample Averaged database (SSA) where all stratigraphic replicates and intra-sample measurements (e.g., SIMS, KrF spot fluorination, in-situ laser, SHRIMP, CO2 spot laser fluorination, LA MC-ICP-MS) were averaged. All database interroga- tion, statistical analysis and figure generation was performed in R via Nature Communications | (2023) 14:279 7 Article https://doi.org/10.1038/s41467-023-35820-w Table 1 | Description of data and metadata included in multi- ple sulfur isotope (MSI) database Database header Description Craton The craton from which the sample is derived Supergroup The supergroup from which sample is derived Group Subgroup Formation Member Core age.max age.min age.mean age.source Sample.ID Replicate Lithology Mineralogy Phase core.rim Depth d34s sig.d34s D33S sig.D33S D36S sig.D36S The group from which the sample is derived The subgroup from which the sample is derived The formation from which the sample is derived The member from which the sample is derived The core from which the sample is derived. If sampled from outcrop, listed as ‘Outcrop’ Maximum published nearest age for a given sample Minimum published nearest age for a given sample Mean of age.min and age.max Source(s) for age.min and age.max Sample identifier. If not provided, the ID was constructed in the form of “Core-Depth” Column with a unique identifier for replicate analyses of the same sample. This includes individual spots from the same bulk sample Reported lithology Reported mineralogy. In general, for bulk samples, unless explicitly stated within the source publication, this was left blank Phase of sulfur measured (i.e., sulfide, sulfate, sulfide + organic S, organic S, total S) For spot measurements, denotes whether the spot sam- pled the core or rim of the mineral analysed. Not always specified Metres core-depth where the sample was taken. If sample was from an outcrop, a stratigraphic height is given, where applicable δ34S (‰) Reported uncertainty in δ34S (‰) Δ33S (‰) Reported uncertainty in Δ33S (‰) Δ36S (‰) Reported uncertainty in Δ36S (‰) Analysis.Type Type of analysis (e.g., bulk, SF6; SIMS; EA-CF-IRMS etc.) Chemistry Source DOI Date Preparative chemistry performed prior to analysis Source of the data Digital Object Identifier of published data source Year of publication the Tidyverse family of packages59, as well as the patchwork60, ggridges61 and MetBrewer62 packages. Besides forming the backbone of our statistical analyses, the compiled database has been used to create an interactive HTML interface that allows the construction of user-defined time vs. Δ33S plots (Supplementary Information). The HTML interface was gener- ated using R and the Plotly package63. Modelling the crustal memory effect The base geochemical model was run in MATLAB. The model follows ref. 22., however, we alter the oxidative pyrite weathering constant to emulate empirically demonstrated half-order kinetics64 (Eq. 3; Table 2). The sulfur isotope model was constructed using the flux outputs of the base geochemical model, including oxygen concentrations (Ox) and Eqs. 4–10 (Table 2). Here, the isotopic imbalance between the various sulfur reservoirs was achieved via the direct sequestration of a given proportion of S8 (FS8py) within the sedimentary pyrite pool, commu- nicating a positive Δ33S value to the continental reservoir and its Table 2 | Description of variables used in Eqs. 4–9 Variable S Si FVSulf FVS8 Fws Fbpy Fbgyp FS8py DSulf DVSulf DVS8 Dws Dbpy(cid:2)s Dbpy Dbgyp Description Seawater sulfate reservoir (SWSR) at time = tstep SWSR at time = tstep Flux of volcanic sulfur into the SWSR in the form of sulfate (cid:2) dt Flux of volcanic sulfur into the SWSR the form of S8 Flux of weathered sulfur into the SWSR Flux of buried pyrite out of the SWSR Flux of buried gypsum out of the SWSR Flux of S8 directly into the buried pyrite phase Δ33S of the SWSR Δ33S of volcanic sulfur into the SWSR in the form of sulfate Δ33S of volcanic sulfur into the SWSR the form of S8 Δ33S of weathered sulfur into the SWSR Δ33S of buried pyrite out of the SWSR Δ33S of buried pyrite including the direct incorporation of S8 Δ33S of buried gypsum out of the SWSR negative counterpart to the seawater sulfate reservoir. Ox = O2 0:5 O2 0:5 + K 0:5 wpy bpy(cid:2)s = D D Sulf bgyp = D D Sulf S = S i + F VSulf + F VS8 + F ws (cid:2) F bpy (cid:2) F bgyp ð3Þ ð4Þ ð5Þ ð6Þ D Sulf × S i + D VSulf × F VSulf + D VS8 × F D Sulf = VS8 + D S ws × F ws (cid:2) D bpy(cid:2)s × F bpy (cid:2) D bgyp × F bgyp D bpy = D Sulf × F F bpy + D bpy + F S8py VS8 × F S8py D ws = Output from bootstrap sampling subroutine Time Weight = e(cid:2)0:001 × ðt (cid:2)t(cid:2)1Þ step ð7Þ ð8Þ ð9Þ ð10Þ The Δ33S value of the weathered sulfur flux at each time step (tstep) was calculated directly from the database using a bootstrap sampling subroutine. Here, two different versions of the database were analysed, with each optimised to minimise biases: First, we generated a Tem- porally Adjusted (TA) synthetic dataset whose overall Δ33S distribution reflected that of the precursor SSA database, yet featured a consistent sample density within each 50-million-year time-bin. Next, we derived a second Craton Adjusted (CA) synthetic dataset, using the same parameters as before, however, now we stipulated that once featured within a given time-bin each craton was equally represented. For the model run using the unadjusted database, the raw form of the SSA database was utilised using the same model parameters as above. Naturally, to protect the integrity of our findings, samples that were initially described as altered and/or late addition were excluded from our analyses. Nature Communications | (2023) 14:279 8 Article https://doi.org/10.1038/s41467-023-35820-w The Δ33S value of the weathered sulfur flux at each time step (tstep), was calculated directly from the TA and CA synthetic datasets using a MATLAB-based bootstrap sampling subroutine. Here, each calculation was restricted to older samples (i.e., t > tstep) whose likelihood of selection was reduced as t and tstep diverged, via an age weighting equation (Eq. 10) modelled after Ref. 65. Importantly, an oxygen scalar (Ox; Eq. 3) designed to curb sulfide weathering at low pO2 was also employed, allowing sulfide weathering to become more pronounced as atmospheric oxygen rose. Using these weighting protocols to quasi- regulate the age and sulfur phase of selection, the bootstrap sampling subroutine then selected n samples and calculated their mean. Repeating this process m times then allowed us to assess the variability of the output. For instance, considering Fig. 3, the lighter coloured envelopes depict the standard deviation (±1σ) of the bootstrap means, while the central line gives the mean of the replicate bootstrap outputs (i.e., the grand mean). After exploring the sensitivity of the model to these parameters, respective n and m values of 10 and 20 were selected for the final model (Supplementary Information; Supplementary Fig. 6). Given its targeting of the same sulfide pool as oxidative weath- ering, photochemically catalysed pyrite oxidation, as recently descri- bed by Hao et al.66, would not significantly change our isotopic estimates of the post-2.3 Ga CME and, therefore, is not considered in the model. This process, we concede, could provide an under- appreciated source of sulfate to the oceans that has the potential to drive the Δ33S composition of the SWSR closer to 0‰ prior to the intensification of oxidative sulfide weathering. Statistical analyses of post-Rooihoogte-aged S-MIF To better characterise the prevalence of S-MIF after 2.33 Ga within the Transvaal Basin, and thus ascertain its regional–global significance, we employed a simple Bayesian approach coupled with a data- parameterised25,30 bootstrap-driven synthetic sampling experiment (Fig. 5; Supplementary Figs. 9 and 10). Here, to avoid potential sampling density biases introduced by the presence of variably sized and dis- tributed diabase intrusions, we focused our attention on the intrusion- free lower Timeball Hill Formation prior to the Gatsrand Member (Fig. 2; Supplementary Fig. 10). In each of the explored cores, the expression of the lower Timeball Hill Formation is roughly equivalent, spanning 168.5, 151.6, and 158.3 m in cores EBA-1, EBA-2, and KEA-4, respectively22,25,30,37. Comparison between these different expressions of the lower Timeball Hill Formation is presented in Supplementary Table 5. Using the parameters outlined in Supplementary Table 5, we started with a simplistic Bayesian approach to estimate the likelihood that a given stratigraphic interval is S-MIF bearing. In essence, we can liken the mass-dependency preserved within the lower Timeball Hill MSI record to a biased coin flip, where S-MIF-bearing samples (i.e., Δ33S ≥ 0.3) are assigned heads and their mass-dependant counterparts (i.e., Δ33S ≤ 0.3) assigned tails. Here, our hypothetical coin is biased, meaning that the odds of returning heads or tails is unfairly weighted (i.e., ≠50:50). Consequently, informed by the outcomes of a given number of coin flips, we are effectively determining the bias (θ) of this hypothetical coin. In this context, each coin flip equates to an analy- tical Δ33S measurement and θ to the likelihood of randomly selecting a S-MIF-bearing sample from a core. Bayesian analysis requires the definition of a prior distribution and a likelihood distribution. Accordingly, applying Poulton and col- leagues’ data25 to Eqs. 11 and 12, a prior was generated using a beta distribution: pðθ∣a,bÞ = θa(cid:2)1 × ð1 (cid:2) θÞb(cid:2)1 Bða,bÞ Bða,bÞ = Z 1 0 dθ θa(cid:2)1 × ð1 (cid:2) θÞb(cid:2)1 ð11Þ ð12Þ This prior describes the probability of a particular θ, given the parameters a and b, or p(θ|a,b), which represents the respective number of ‘heads’ (S-MIF) and ‘tails’ (S-MDF) in cores EBA-1 and EBA-2. Here, the mean θ of the resultant distribution was 0.203 and its mode was 0.190 (Supplementary Fig. 9a), empirically consistent with the 20.3% (=13/64) S-MIF detection frequency reported by Poulton et al.25 Next, a Bernoulli likelihood distribution was generated using the data presented by Izon et al.30 (Eq. 13): pðD∣θÞ = θz × ð1 (cid:2) θÞN(cid:2)z ð13Þ with N representing the number of samples in the lower Timeball Hill Formation, and z being the number of S-MIF-bearing samples (60 and 0, respectively; Supplementary Fig. 9b). This distribution describes the likelihood of reproducing the data (D) at a given θ, or p(D|θ). Returning to our numismatic analogy, Eq. 13 gives the likelihood of returning ‘tails’ 60 times given a specific weighting of a coin. Naturally, given Izon and colleagues’ 0% S-MIF detection frequency in the lower Timeball Hill Formation within core KEA-430, the mode of the resultant Bernoulli likelihood distribution was 0. Finally, by combining the previously described prior (i.e., p(θ|a,b)) and likelihood (i.e., p(D|θ)) distributions a posterior distribution was constructed using Bayes theorem (Eqs. 14 and 15; Supplementary Fig. 9c): pðθ∣z,NÞ = θz + a(cid:2)1 × ð1 (cid:2) θÞN(cid:2)z + b(cid:2)1 Bðz + a,N (cid:2) z + bÞ ð14Þ BðZ + a,N (cid:2) z + bÞ = Z 1 0 dθ θz + a(cid:2)1 × ð1 (cid:2) θÞN(cid:2)z + b(cid:2)1 ð15Þ This final distribution describes the updated probability of θ given the new data parameters z and N, or, p(θ|z,N). The resultant mean of the posterior distribution was 0.105 and its mode was 0.100. Given the combined probabilities in the posterior distribution, the most likely θ, or probability of randomly selecting an S-MIF-bearing sample (i.e., the coin returning ‘heads’) is 10.0–10.5%. This likelihood can then be translated into the proportion of a given core possessing mass- independent versus mass-dependant S-isotope systematics. Returning to the proximal (<5 km) Carletonville cores, given that the lower Timeball Hill Formation is roughly of equivalent thickness, we assume that the relative occurrence of S-MIF and the stratigraphic interval over which it’s expressed would be broadly analogous within each core. Maintaining these assumptions, we predict that 10.25% of the lower Timeball Hill Formation within core EBA-2 (151.6 m), equating to 15.54 m, should be S-MIF-bearing, which, if accommodated by the 11 S- MIF-bearing samples found in EBA-225, results in an average S-MIF- bearing interval of 1.42 m. Pursuing this tactic further, we devised a bootstrap sampling method that would allow us to estimate the likelihood of failing to find S-MIF within a 60-sample sample set taken at random from synthetic cores constructed using different stratigraphic distributions of S-MIF. Designed to emulate the typical sample thickness used by Izon et al.30 and, indeed, that typically administered by core repositories, we con- structed a 150-m-long synthetic core comprising 3000 5-cm-thick samples. Then, by centring a hypothetical S-MIF window of a given thickness (i.e., Δ33S > 0.3‰) reported by Poulton et al.25, we prescribed the mass- dependency of the surrounding samples (Supplementary Fig. 10). Selecting 60 samples from the resultant synthetic core, and repeating the operation 1000 times, then yielded a frequency distribution illus- trating how often S-MIF would be expected to be detected within a hypothetical 60-sample analytical campaign. Repeated analysis with variably thick S-MIF windows (0.02–10.0 m), demonstrated that non- (0.02–10.0 m) on the S-MIF-bearing samples Nature Communications | (2023) 14:279 9 Article https://doi.org/10.1038/s41467-023-35820-w detection of S-MIF within a 60-sample population only became likely as the size of the S-MIF window diminished (Fig. 5a; Supplementary Table 6). Recalling our Bayesian analysis, the likelihood of randomly selecting 0 or 11 S-MIF-bearing samples is roughly equivalent given a 1.42-m-thick S-MIF window. That said, it is important to stress that the likelihood of achieving either is very low, falling outside of 2σ of the mean (Fig. 5a). It follows, therefore, that while our Bayesian approach yields a S-MIF window that most adequately explains the current data, it is unlikely to accurately describe the true distribution of S-MIF. Indeed, it is only when the thickness of the prescribed S-MIF window drops to well below a metre do the chances of avoiding S-MIF within a 60-sample analytical campaign become probable (Fig. 5a; Supple- mentary Table 6). Sulfur isotope analysis Augmenting previously published EBA-2 data22, herein we report an additional 32 quadruple sulfur isotope (QSI) measurements. First, to ensure consistency, we targeted two previously reported samples from the Rooihoogte Formation22 before measuring a new sample suite spanning the Timeball Hill Formation (Supplementary Data 2). While previous work over-looked acid volatile sulfur (AVS)22, herein we follow the sequential two-step approach detailed in Izon et al.30 that liberates AVS upon reaction with ethanoic 6 M HCl followed by the release of chromium reducible sulfur (CRS; principally pyrite) via reduction using acidified 2 M CrCl2. In both AVS- and CRS-yielding refluxes, the resultant H2S was swept into zinc acetate traps where it was captured as zinc sulfide before being converted to silver sulfide (Ag2S) with silver nitrate. Given the relative paucity of QSI data within core EBA-2, we universally converted the Ag2S to sulfur hexafluoride (SF6) via overnight reaction with excess of fluorine gas (F2) at 300 °C. The resultant SF6 was then purified cryogenically at liquid-nitrogen- temperatures before final isolation by preparative gas chromato- graphy within the Geobiology laboratory at MIT. The QSI isotope composition of the pure SF6 was measured by dual-inlet gas-source isotope ratio mass spectrometry using a Thermo-Finnigan MAT 253 equipped with four collectors arranged to measure the intensity of SF5+ ion beams at mass/charge ratios (m/z) of 127, 128, 129 and 131 (32SF5 +). Over the course of this study, several fluorinations of the IAEA-distributed reference materials S1, -S2 and -S3 were performed, returning 1σ δ34S, Δ33S and Δ36S uncertainties of better than 0.4, 0.022 and 0.16, respectively (Supplementary Data File Table SX2). Here, the resultant VCDT-normalised values are inseparable from their certified values and, indeed, those reported initially by Ono et al.67 (Table SX2). Importantly, the reproducibility of these pure Ag2S standards approximates those calculated from repeated sample processing (i.e., extraction–fluorination; Supple- mentary Data 2). +, and 36SF5 +, 34SF5 +, 33SF5 Data availability The compiled database and the MSI data generated in this study are provided in Supplementary Data 1 and 2 respectively with an inter- active HTML interface plot of the geological Δ33S record through time provided as Supplementary Data 3. Code availability The code used in this study can be found online in a GitHub repository https://github.com/buveges/Sulfur-MSI-Database an updating version of the MSI database. The MATLAB model was ori- ginally created by S.O. for Luo et al.22, and subsequently modified and augmented by S.O. and B.T.U. for this study. The R code was writ- ten by B.T.U. along with References 1. Lyons, T. W., Reinhard, C. T. & Planavsky, N. J. The rise of oxygen in Earth’s early ocean and atmosphere. Nature 506, 307–315 (2014). 2. 3. Stüeken, E. E., Kipp, M. A., Koehler, M. C. & Buick, R. The evolution of Earth’s biogeochemical nitrogen cycle. Earth Sci. Rev. 160, 220–239 (2016). Reinhard, C. T. et al. Evolution of the global phosphorus cycle. Nature 541, 386–389 (2017). 4. Holland, H. Volcanic gases, black smokers, and the Great Oxidation 5. Event. Geochim. Cosmochim. Acta 66, 3811–3826 (2002). Fischer, W. W., Hemp, J. & Valentine, J. S. How did life survive Earth’s great oxygenation? Curr. Opin. Chem. Biol. 31, 166–178 (2016). 6. Gold, D. A., Caron, A., Fournier, G. P. & Summons, R. E. Paleopro- 7. terozoic sterol biosynthesis and the rise of oxygen. Nature 543, 420–423 (2017). Schopf, J. W. Geological evidence of oxygenic photosynthesis and the biotic response to the 2400-2200 Ma “Great Oxidation Event”. Biochem 79, 165–177 (2014). 8. Hodgskiss, M. S. W., Crockford, P. W., Peng, Y., Wing, B. A. & Horner, T. J. A productivity collapse to end Earth’s Great Oxidation. Proc. Natl Acad. Sci. USA 116, 17207–17212 (2019). 9. Sánchez-Baracaldo, P., Bianchini, G., Wilson, J. D. & Knoll, A. H. Cyanobacteria and biogeochemical cycles through Earth history. Trends Microbiol. 1–15. https://doi.org/10.1016/j.tim.2021.05. 008 (2021) 10. Sessions, A. L., Doughty, D. M., Welander, P. V., Summons, R. E. & Newman, D. K. The continuing puzzle of the great oxidation event. Curr. Biol. 19, R567–R574 (2009). 11. Reinhard, C. T. & Planavsky, N. J. The history of ocean oxygenation. Ann. Rev. Mar. Sci. 14, 331–353 (2022). 12. Farquhar, J., Bao, H. & Thiemens, M. Atmospheric influence of Earth’s earliest sulfur cycle. Science 289, 756–758 (2000). 13. Ono, S. Photochemistry of sulfur dioxide and the origin of mass- independent isotope fractionation in Earth’s atmosphere. Annu. Rev. Earth Planet. Sci. 45, 301–329 (2017). 14. Ono, S., Wing, B., Johnston, D., Farquhar, J. & Rumble, D. Mass- dependent fractionation of quadruple stable sulfur isotope system as a new tracer of sulfur biogeochemical cycles. Geochim. Cos- mochim. Acta 70, 2238–2252 (2006). 15. Hulston, J. R. & Thode, H. G. Variations in the S33, S34, and S36 con- tents of meteorites and their relation to chemical and nuclear effects. J. Geophys. Res. 70, 3475–3484 (1965). 16. Farquhar, J. & Wing, B. A. Multiple sulfur isotopes and the evolution of the atmosphere. Earth Planet. Sci. Lett. 213, 1–13 (2003). 17. Chen, M., Huang, Z., Campbell, I. H., Sambridge, M. & Ávila, J. N. Simultaneous fractionation of sulfur dioxide explains mass inde- pendent fractionation of sulfur isotopes in Archean sedimentary pyrites. Chem. Geol. 587, 120640 (2022). 18. Ono, S. et al. New insights into Archean sulfur cycle from mass- independent sulfur isotope records from the Hamersley Basin, Australia. Earth Planet. Sci. Lett. 213, 15–30 (2003). 19. Pavlov, A. A. & Kasting, J. F. Mass-independent fractionation of sulfur isotopes in Archean sediments: strong evidence for an anoxic Archean atmosphere. Astrobiology 2, 27–41 (2002). 20. Ono, S., Beukes, N. J., Rumble, D. & Fogel, M. L. Early evolution of atmospheric oxygen from multiple-sulfur and carbon isotope records of the 2.9 Ga Mozaan Group of the Pongola Supergroup, Southern Africa. S. Afr. J. Geol. 109, 97–108 (2006). 21. Guo, Q. et al. Reconstructing Earth’s surface oxidation across the Archean-Proterozoic transition. Geology 37, 399–402 (2009). 22. Luo, G. et al. Rapid oxygenation of Earth’s atmosphere 2.33 billion years ago. Sci. Adv. 2, e1600134 (2016). 23. Warke, M. R. et al. The Great Oxidation Event preceded a Paleo- proterozoic “snowball Earth”. Proc. Natl Acad. Sci. 117, 13314–13320 (2020). 24. Gumsley, A. P. et al. Timing and tempo of the Great Oxidation Event. Proc. Natl Acad. Sci. USA 114, 1811–1816 (2017). Nature Communications | (2023) 14:279 10 Article https://doi.org/10.1038/s41467-023-35820-w 25. Poulton, S. W. et al. A 200-million-year delay in permanent atmo- spheric oxygenation. Nature 592, 232–236 (2021). 26. Bekker, A., Krapež, B. & Karhu, J. A. Correlation of the stratigraphic cover of the Pilbara and Kaapvaal cratons recording the lead up to Paleoproterozoic Icehouse and the GOE. Earth Sci. Rev. 211, 103389 (2020). 27. Coetzee, L. L. Genetic Stratigraphy of the Paleoproterozoic Pretoria Group in the Western Transvaal. Dissertation (Rand Afrikaans Uni- versity, 2001). 28. Warke, M. R. & Schröder, S. Synsedimentary fault control on the deposition of the Duitschland Formation (South Africa): Implica- tions for depositional settings, Paleoproterozoic stratigraphic cor- relations, and the GOE. Precambrian Res. 310, 348–364 (2018). 29. Johnson, J. E. et al. Manganese-oxidizing photosynthesis before the 30. rise of cyanobacteria. Proc. Natl Acad. Sci. USA 110, 11238–11243 (2013). Izon, G. et al. Bulk and grain-scale minor sulfur isotope data reveal complexities in the dynamics of Earth’s oxygenation. Proc. Natl. Acad. Sci. USA 119, e2025606119 (2022). Earth-Science Reviews, 211, 103,389” by Pascal Phili. Earth Sci. Rev. 218, 103607 (2021). 45. Williford, K. H., Van Kranendonk, M. J., Ushikubo, T., Kozdon, R. & Valley, J. W. Constraining atmospheric oxygen and seawater sulfate concentrations during Paleoproterozoic glaciation: In situ sulfur three-isotope microanalysis of pyrite from the Turee Creek Group, Western Australia. Geochim. Cosmochim. Acta 75, 5686–5705 (2011). 46. Guy, B. M. et al. A multiple sulfur and organic carbon isotope record from non-conglomeratic sedimentary rocks of the Mesoarchean Witwatersrand Supergroup, South Africa. Precambrian Res. 216–219, 208–231 (2012). 47. Baroni, M., Thiemens, M. H., Delmas, R. J. & Savarino, J. Mass- independent sulfur isotopic compositions in stratospheric volcanic eruptions. Science 315, 84–87 (2007). 48. Burke, A. et al. Stratospheric eruptions from tropical and extra-tropical volcanoes constrained using high-resolution sulfur isotopes in ice cores. Earth Planet. Sci. Lett. 521, 113–119 (2019). 31. Reinhard, C. T., Planavsky, N. J. & Lyons, T. W. Long-term sedi- 49. Savarino, J. UV induced mass-independent sulfur isotope fractio- mentary recycling of rare sulphur isotope anomalies. Nature 497, 100–103 (2013). nation in stratospheric volcanic sulfate. Geophys. Res. Lett. 30, 2131 (2003). 32. Philippot, P. et al. Globally asynchronous sulphur isotope signals require re-definition of the Great Oxidation Event. Nat. Commun. 9, 2245 (2018). 33. Killingsworth, B. A. et al. Constraining the rise of oxygen with oxy- gen isotopes. Nat. Commun. 10, 1–10 (2019). 34. Farquhar, J. et al. Pathways for Neoarchean pyrite formation con- strained by mass-independent sulfur isotopes. Proc. Natl Acad. Sci. USA 110, 17638–17643 (2013). 35. Jamieson, J. W., Wing, B. A., Hannington, M. D. & Farquhar, J. Eval- uating isotopic equilibrium among sulfide mineral pairs in Archean ore deposits: case study from the Kidd Creek VMS deposit, Ontario, Canada. Econ. Geol. 101, 1055–1061 (2006). 36. Johnston, D. T. et al. Multiple sulfur isotope fractionations in bio- logical systems: a case study with sulfate reducers and sulfur dis- proportionators. Am. J. Sci. 305, 645–660 (2005). 50. Canfield, D. E. & Farquhar, J. Animal evolution, bioturbation, and the sulfate concentration of the oceans. Proc. Natl Acad. Sci. USA 106, 8123–8127 (2009). 51. Planavsky, N. J., Bekker, A., Hofmann, A., Owens, J. D. & Lyons, T. W. Sulfur record of rising and falling marine oxygen and sulfate levels during the Lomagundi event. Proc. Natl Acad. Sci. USA 109, 18300–18305 (2012). 52. Gallagher, M., Whitehouse, M. J. & Kamber, B. S. The Neoarchaean surficial sulphur cycle: An alternative hypothesis based on analo- gies with 20th-century atmospheric lead. Geobiology 15, 385–400 (2017). 53. Halevy, I., Johnston, D. T. & Schrag, D. P. Explaining the structure of the archean mass-independent sulfur isotope record. Science 329, 204–207 (2010). 54. Liu, C. & Zhang, J. Methane might be made by all living organisms. 37. Bekker, A. et al. Dating the rise of atmospheric oxygen. Nature 427, Nature 603, 396–397 (2022). 117–120 (2004). 38. Johnson, A. C. et al. Reconciling evidence of oxidative weathering and atmospheric anoxia on Archean Earth. Sci. Adv. 7, 1–10 (2021). 39. Crockford, P. W. et al. Claypool continued: extending the isotopic record of sedimentary sulfate. Chem. Geol. 513, 200–225 (2019). 40. Torres, M. A., Paris, G., Adkins, J. F. & Fischer, W. W. Riverine evi- dence for isotopic mass balance in the Earth’s early sulfur cycle. Nat. Geosci. 11, 661–664 (2018). 55. Shang, H., Rothman, D. H. & Fournier, G. P. Oxidative metabolisms catalyzed Earth’s oxygenation. Nat. Commun. 13, 1328 (2022). 56. Wogan, N. F., Catling, D. C., Zahnle, K. J. & Claire, M. W. Rapid timescale for an oxic transition during the Great Oxidation Event and the instability of low atmospheric O2. Proc. Natl Acad. Sci. USA 119, 2–9 (2022). 57. Goldblatt, C., Lenton, T. M. & Watson, A. J. Bistability of atmo- spheric oxygen and the Great Oxidation. Nature 443, 683–686 (2006). 41. Asael, D., Planavsky, N. J., Bellerfroid, E., Hofmann, A. & Reinhard, C. T. Sulfur non mass dependent anomalies in modern river water of Archean catchment. in Goldschmidt2017 Abstract (Gold- schmidt2017, 2017). 58. Alcott, L. J., Mills, B. J. W. & Poulton, S. W. Stepwise Earth oxyge- nation is an inherent property of global biogeochemical cycling. Science 366, 1333–1337 (2019). 59. Wickham, H. et al. Welcome to the Tidyverse. J. Open Source Softw. 42. Caquineau, T., Paquette, J.-L. & Philippot, P. U-Pb detrital zircon 4, 1686 (2019). geochronology of the Turee Creek Group, Hamersley Basin, Wes- tern Australia: timing and correlation of the Paleoproterozoic gla- ciations. Precambrian Res. 307, 34–50 (2018). 60. Pedersen, T. L. patchwork: The Composer of Plots R package https:// cran.r-project.org/ (CRAN, 2020). 61. Wilke, C. O. ggridges: Ridgeline Plots in ggplot2 R package https:// 43. Philippot, P. et al. Comment on ‘Correlation of the stratigraphic cran.r-project.org/ (CRAN, 2021). cover of the Pilbara and Kaapvaal cratons recording the lead up to Paleoproterozoic Icehouse and the GOE’ by Andrey Bekker, Bryan Krapež, and Juha A. Karhu, 2020, Earth Science Reviews. Earth Sci. Rev. 218, 103594 (2021). 62. Mills, B. R. MetBrewer: Color Palettes Inspired by Works at the Metropolitan Museum of Art R package https://cran.r-project.org/ (CRAN, 2022). 63. Sievert, C. Interactive Web-Based Data Visualization with R, plotly, 44. Bekker, A., Krapež, B., Karhu, J. A. & Chamberlain, K. Reply to and shiny (Chapman and Hall/CRC, 2020). comment on “Bekker, A., Krapež, B., Karhu, J.A., 2020. Correlation of the stratigraphic cover of the Pilbara and Kaapvaal cratons recording the lead up to Paleoproterozoic Icehouse and the GOE. 64. Johnson, A. C. et al. Experimental determination of pyrite and molybdenite oxidation kinetics at nanomolar oxygen concentra- tions. Geochim. Cosmochim. Acta 249, 160–172 (2019). Nature Communications | (2023) 14:279 11 Article https://doi.org/10.1038/s41467-023-35820-w 65. Peters, S. E. & Husson, J. M. Sediment cycling on continental and oceanic crust. Geology 45, 323–326 (2017). 66. Hao, J. et al. Anoxic photochemical weathering of pyrite on Archean continents. Sci. Adv. 8, 1–8 (2022). 67. Ono, S., Keller, N. S., Rouxel, O. & Alt, J. C. Sulfur-33 constraints on Additional information Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41467-023-35820-w. the origin of secondary pyrite in altered oceanic basement. Geo- chim. Cosmochim. Acta 87, 323–340 (2012). Correspondence and requests for materials should be addressed to Benjamin T. Uveges. Acknowledgements We dedicate this paper to our co-author Prof. Nicolas J. Beukes who sadly passed away after the acceptance of this contribution. As a col- league, mentor, and friend “Prof. Nic” worked tirelessly to bring South African geology to the world. Beyond his many contributions, Nic’s work has, and continues to, revolutionise our understanding of Earth’s evo- lution. This study was supported by the Simons Collaboration on the Origins of Life (#290361FY18 to R.E.S. & #874698 to B.T.U. and R.E.S.) and the National Science Foundation (#EAR-1338810 to R.E.S. and S.O.). G.I. acknowledges receipt of a MISTI Global Seed Award. N.J.B. acknowl- edges financial support from DSI-NRF CIMERA in South Africa. We recognise formative discussions with various members of the Sum- mons Lab. Author contributions B.T.U. and G.I. designed the study. B.T.U. compiled the chemical data, constructed the database and performed its statistical interrogation. B.T.U. implemented and augmented the geochemical model with gui- dance from S.O. N.J.B. and S.O. collected the samples from which G.I. conducted the isotopic analyses. B.T.U. and G.I. synthesised the data–model interpretations, writing the manuscript with input from all authors. R.E.S., S.O., G.I., N.J.B., and B.T.U. acquired the funding to support various aspects of this study. Peer review information Nature Communications thanks Bryan Kill- ingsworth, Guillaume Paris and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available. Reprints and permissions information is available at http://www.nature.com/reprints Publisher’s note Springer Nature remains neutral with regard to jur- isdictional claims in published maps and institutional affiliations. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/ licenses/by/4.0/. Competing interests The authors declare no competing interests. © The Author(s) 2023 Nature Communications | (2023) 14:279 12
10.1038_s41467-021-27769-5
ARTICLE https://doi.org/10.1038/s41467-021-27769-5 OPEN Phylogenetically and functionally diverse microorganisms reside under the Ross Ice Shelf 3, Zihao Zhao 3,4, Rachael J. Lappan 1,2,17, Chris Greening 3,4, Sean K. Bay 1, Clara Martínez-Pérez Daniele De Corte5, Christina Hulbe Ramunas Stepanauskas 16✉ Sergio E. Morales 11, José M. González & Federico Baltar 1,10✉ 6, Christian Ohneiser 7, Craig Stevens 8,9, Blair Thomson10, 12, Ramiro Logares 13, Gerhard J. Herndl 1,14,15, ; , : ) ( 0 9 8 7 6 5 4 3 2 1 Throughout coastal Antarctica, ice shelves separate oceanic waters from sunlight by hun- dreds of meters of ice. Historical studies have detected activity of nitrifying microorganisms in oceanic cavities below permanent ice shelves. However, little is known about the microbial In this study, we profiled the composition and pathways that mediate these activities. microbial communities beneath the Ross Ice Shelf using a multi-omics approach. Overall, beneath-shelf microorganisms are of comparable abundance and diversity, though distinct composition, relative to those in the open meso- and bathypelagic ocean. Production of new organic carbon is likely driven by aerobic lithoautotrophic archaea and bacteria that can use ammonium, nitrite, and sulfur compounds as electron donors. Also enriched were aerobic organoheterotrophic bacteria capable of degrading complex organic carbon substrates, likely derived from in situ fixed carbon and potentially refractory organic matter laterally advected by the below-shelf waters. Altogether, these findings uncover a taxonomically distinct microbial community potentially adapted to a highly oligotrophic marine environment and suggest that ocean cavity waters are primarily chemosynthetically-driven systems. 1 Department of Functional and Evolutionary Ecology, University of Vienna, Djerassiplatz 1, 1030 Vienna, Austria. 2 Centre for Microbiology and Environmental Systems Science, Division of Microbial Ecology, University of Vienna, Djerassiplatz 1, 1030 Vienna, Austria. 3 Department of Microbiology, Biomedicine Discovery Institute, Monash University, Clayton, VIC 3800, Australia. 4 Securing Antarctica’s Environmental Future, Monash University, Clayton, VIC 3800, Australia. 5 Institute for Chemistry and Biology of the Marine Environment, Carl von Ossietzky University of Oldenburg, Oldenburg, Germany. 6 School of Surveying, University of Otago, Dunedin, New Zealand. 7 Department of Geology, University of Otago, Dunedin, New Zealand. 8 National Institute of Water and Atmospheric Research, Greta Point, Wellington 6021, New Zealand. 9 Department of Physics, University of Auckland, Auckland, New Zealand. 10 Department of Marine Sciences, University of Otago, Dunedin, New Zealand. 11 Bigelow Laboratory for Ocean Sciences, East Boothbay, ME, USA. 12 Department of Microbiology, University of La Laguna, ES-38200 La Laguna, Spain. 13 Department of Marine Biology and Oceanography, Institut de Ciències del Mar (CSIC), Barcelona, Spain. 14 NIOZ, Department of Marine Microbiology and Biogeochemistry, Royal Netherlands Institute for Sea Research, Utrecht University, PO Box 59, 1790 AB Den Burg, The Netherlands. 15 Vienna Metabolomics Center, University of Vienna, Djerassiplatz 1, A-1030 Vienna, Austria. 16 Department of Microbiology and Immunology, University of Otago, Dunedin, New Zealand. 17Present address: Institute for Environmental Engineering, Department of Civil, Environmental and Geomatic Engineering, Eidgenössische Technische Hochschule (ETH) Zürich, 8093 Zurich, Switzerland. ✉ email: sergio.morales@otago.ac.nz; federico.baltar@univie.ac.at NATURE COMMUNICATIONS | (2022) 13:117 | https://doi.org/10.1038/s41467-021-27769-5 | www.nature.com/naturecommunications 1 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-27769-5 Results The water column under the Ross Ice Shelf is characterized by a steep vertical ammonium gradient. During the Ross Ice Shelf Program in December 2017, an access borehole was created by hot longitude water drilling at site HWD-2 (latitude 80.6577 S, ice shelf Ice shelves are permanent floating extensions of grounded sheets of ice that connect to a landmass. The Ross Ice Shelf, by in the world, floats atop an area the largest ~54,000 km3 ocean cavity that covers about half of the Ross Sea and hugs the coast of Antarctica (Fig. 1a). Generally over 300 m thick1, the ice shelf creates a “lid” that isolates the underlying ocean from the atmosphere and from sunlight, and exerts a direct effect on the chemical composition of the water column beneath it (in general ~700 m deep2). Waters under the permanent ice shelves are influenced by continental ice-sheet melting and are thus an important intermediary between subglacial outflow from the Antarctic continent and the open Ross Sea, and ultimately the Southern Ocean. Despite their oceanographic significance, sub-ice shelf habitats are among the least-studied ecosystems in the world’s oceans. Oceanographic and biogeochemical observations of the water cavity beneath the Ross Ice Shelf have been largely concentrated on the shelf margins, in particular at the McMurdo Ice Shelf (northwestern portion of the Ross Ice Shelf). Here, nutrient- and biomass-rich water advected from eastern McMurdo Sound likely plays an important role in sub-ice biogeochemistry of the dark ecosystem beneath the shelf front3,4. Direct observations in the grounding area have also confirmed a stratified and quiescent ocean setting5. As a result, water below the Ross Ice Shelf is reported to be exchanged with the Ross Sea with an estimated residence time of 0.9–5.4 years6,7. This allows transport of nutrients and organisms from the sea into the cavity. However, unlike other well-ventilated shelves (e.g., Amery shelf8), the proximity to open water is likely a major factor controlling bio- geochemical process in the central basin of the Ross Ice Shelf cavity. Opportunities to directly access the central sub-ice shelf cavity have been greatly limited by logistical constraints and only one expedition to date has sampled the seawater beneath the center of the Ross Ice Shelf. Sampling of the sub-ice water column took place through borehole J9, during the Ross Ice Shelf Project of 19779. The environment beneath the Ross Ice Shelf was described as “similar to the abyssal ocean in being cold and aphotic”. Within these waters, “sparse” populations of bacteria, microbial eukaryotes, and animals were observed10,11. The microbial populations were proven to be heterotrophically active and incorporated radiolabeled organic carbon molecules at very low rates comparable to the abyssal ocean10. Autotrophic activity of subsequently these microbial communities was reported and attributed to “nitrifying bacteria”12. In these aphotic ecosystems lacking photosynthetic primary production, dark carbon fixation by nitrifying microorganisms may be suf- ficient and macrofaunal populations12. Lateral inputs of organic carbon from the Ross Sea may also support these populations. However, given these studies preceded the advent of molecular techniques, the com- position of the microbial communities, their relatedness to open ocean communities, and their possible links to ecosystem function remained unexplored. to sustain observed microbial techniques In this study we accessed the waters beneath the Ross Ice Shelf to uncover the phylogenetic and functional diversity of the microbial communities under the Antarctic ice shelf. We com- bined multi-omics (metagenomics, metatran- scriptomics, single-cell genomics) with supporting biogeochemical measurements (nutrient measurements and heterotrophic bac- terial production). We show that the waters below the shelf harbor a diverse microbial community with a taxonomic composition distinct from other open ocean environments. In addition, we observed the transcription of various genes associated with lithoautotrophic and organoheterotrophic growth, uncovering the basis for previous activities reported in below-shelf waters. Fig. 1 Sampling location. a Map showing the sampling location of this study (HWD-2) and the borehole study site J9 drilled in 197710. Bathymetry and ice thickness are based on the Bedmap-2 data set1. The transparent ice surface image was sourced from the MOA2009 image map119. b (left) Thermohaline structure of the water column at station HWD-2 and defined regions. IBL, Ice basal boundary layer. V-IL, variable intermediate layer, likely modulated by tides and resulting in patches of water with variable temperature and salinity. S-IL, stratified intermediate layer. BBL, benthic boundary layer. (right) Schematic of HWD-2 drilling site depicts the sampling location of seawater samples (red circles) at 30, 180, and 330 m below the ice shelf base. 2 NATURE COMMUNICATIONS | (2022) 13:117 | https://doi.org/10.1038/s41467-021-27769-5 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-27769-5 ARTICLE ) d ( e m i t r e v o n r u T ) 1 − d 3 − m C l o m µ ( P H P ) % ( A N H ) % ( A N L l ) 1 − m s l l e c 5 0 1 × ( e c n a d n u b a l l e C ) M µ ( 2 O S i ) M µ ( – 3 4 O P ) M µ ( x O N ) M µ ( 3 H N y t i n i l a s l a c i t c a r P ) C ° ( p m e T ) m ( h t p e D 1 6 4 9 3 3 6 8 3 ) 2 0 0 ( . 0 3 0 . ) 3 0 0 ( . 0 6 0 . ) 1 . 0 ( . 4 0 . ) 4 0 ( . 0 8 1 . ) 3 0 ( 1 . 5 1 . ) 7 0 ( . 4 3 2 . ) 2 0 ( . 8 2 8 . ) 3 0 ( . 5 5 8 . ) 9 0 ( . 4 7 7 ) 1 . 0 ( . 9 0 ) 7 0 0 ( . 0 2 . 1 ) 7 0 0 ( . 0 8 0 . ) 1 . 2 ( . 0 5 6 1 . ) 7 0 ( . 0 6 6 1 . ) 3 0 ( . 0 5 6 1 ) 3 0 0 0 ( . 0 2 7 0 . ) 4 0 0 0 ( . 0 2 7 0 . ) 1 0 0 0 ( . 0 1 7 0 . ) 3 2 0 ( . 5 3 7 . ) 4 0 0 ( . 2 3 7 . ) 5 0 0 ( . 7 3 7 . ) 2 0 0 ( . 4 4 0 . ) 2 0 0 ( . 5 0 0 . ) 1 0 0 ( . 4 0 0 . ) u s p ( . 7 5 4 3 9 6 4 3 . . 6 7 4 3 6 9 . 1 − 1 9 . 1 − 3 1 . 2 − 0 8 1 0 3 3 0 3 174.4626 W), approximately 300 km from the Ross Sea and 330 km northwest of borehole J9 (Fig. 1a). The shelf ice was 370 m thick, and the underlying waters extended to 750 m below the shelf sur- face (Fig. 1b). Triplicate samples were collected at three depths: 30, 180, and 330 m below the bottom of the shelf (i.e., the ice-water interface). These depths correspond to three regions based on the thermohaline structure of the water column: a basal boundary layer just beneath the ice (IBL), the upper part of an intermediate layer characterized by highly variable temperature and salinity (V-IL), and the lower part of the intermediate layer characterized by linear stratification (S-IL). A homogeneous benthic layer was observed (BBL) but not sampled (Fig. 1b; see13 for a detailed physical oceanographic description of the study site). This structure con- firmed that the cavity is filled southward by thermohaline convec- tion in which dense, high salinity shelf water (HSSW) evolves into very cold (~−2 °C) but relatively fresh Ice Shelf Water (ISW). The temperature and salinity conditions suggest that, other than the boundary layer regions, water properties conform to Deep Ice Shelf Water, a mixture of high and low salinity shelf water and Antarctic Surface Water (AASW)13. Contrary to what previous studies detected at the shelf front3,4, other regional water masses were not present at borehole HWD2. The flow of waters beneath the drilling site was 2 cm s−1 towards the open ocean, suggesting a residence time for these waters of ca. 4 years13. This estimate is within the range of 1–6 years from previous ocean measurements6 and modeling studies2,14. the ice shelf3,4 and in deep waters of Nutrient concentrations beneath the center of the Ross Ice Shelf were generally lower than those measured at the edge of the the Ross Sea15. of Concentrations of SiO2 (165–166 µM), NOx (7.32–7.37 µM) and 3− (0.71–0.72 µM) were relatively constant across the water PO4 column (Table 1) and two- to fourfold lower than in the oceanic cavity of the McMurdo Ice Shelf at the edge of the Ross Ice Shelf3,4. In contrast, we observed a steep gradient of ammonium, with concentrations tenfold higher at the basal layer (440 nM) than in deeper waters (40–50 nM). Such high ammonium concentrations, while lower than those in open waters of the Ross Sea (which peak in summer with values >2 µM;15), were in the same range as deep (400 m) high-salinity shelf waters (HSSW) entering the front of the cavity (~500 nm;4). A similar nutrient profile was reported beneath borehole J912, where ammonium concentrations were higher beneath the ice shelf base and − − and NO2 decreased with depth, whereas values of NO3 remained constant the water column. However, concentrations of ammonium and NOx were 10- and 4-times higher at the J9 borehole than we reported for the HWD-2 3− and SiO2 were not reported)13,16,17. borehole (PO4 Microbial cell abundance ranged from 0.9 to 1.2 × 105 cells mL−1 (Table 1), which is typical for mesopelagic and upper bathypelagic open ocean environments18 and comparable to deep waters at the the McMurdo Ice Shelf4. In contrast, prokaryotic margin of heterotrophic production (PHP, a proxy for growth of hetero- trophic organisms) ranged from 0.3 to 0.6 µmol C m−3 d−1 (Table 1), which is one to two orders of magnitude lower than at the margins of the Ross Ice Shelf (~40 µmol C m−3 d−1;4) and the average global PHP rates in the mesopelagic (24 µmol C m−3 d−1) and bathypelagic (4 µmol C m−3 d−1) open ocean18. Based on these PHP rates, the turnover time of the microbial community in our study ranged between 339 and 461 days, within the same order of magnitude as the approximately 400 days reported previously at borehole J910. throughout Below-shelf microbial communities are distinct from open ocean communities. Microbial community composition beneath the Ross Ice Shelf was determined using a combination of 16S . n o i t c u d o r p c i h p o r t o r e t e h c i t o y r a k o r p P H P , s l l e c t n e t n o c d i c a c i e l c u n h g h i A N H , s l l e c t n e t n o c d i c a l c i e c u n w o l A N L , e r u t a r e p m e t p m e T . ) 3 = n ( n w o h s e r a s e u a v l ) s e s e h t n e r a p n i n o i t a v e d i d r a d n a t s ( n a e M . e l o h e r o b 2 - D W H e h t t a f l e h S e c I s s o R e h t w o l e b s n m u l o c r e t a w e h t m o r f a t a d l a c i m e h c o e g o i B 1 e l b a T NATURE COMMUNICATIONS | (2022) 13:117 | https://doi.org/10.1038/s41467-021-27769-5 | www.nature.com/naturecommunications 3 ARTICLE a e d u t i t a L 50 0 -50 b ) m 30 180 ( h t p e 330D -100 Ocean zone Bathypelagic 0 Longitude 001 002 Epipelagic Mesopelagic Ocean cavity, RIS 0 02 04 06 08 100 Relative abundance (%) Phylum Acidobacteriota Actinobacteriota Bacteroidota Chloroflexota Crenarchaeota Gemmatimonadota Latescibacterota Marinisomatota Myxococcota Nitrospinota Planctomycetota Proteobacteria SAR324 Thermoplasmatota Verrucomicrobiota Unclassified bacteria Other phyla c y t i r a l i m i s s i D 6 4 2 0 Latitude Polar Non polar Ocean zone Bathypelagic Epipelagic Mesopelagic Ocean cavity, RIS NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-27769-5 d Latitude Polar Non polar Ocean zone Bathypelagic Epipelagic Mesopelagic Ocean cavity, RIS Row Z-score 4 2 0 -2 -4 * ****** Phylum Crenarchaeota Thermoplasmatota Gemmatimonadota Myxococcota Chloroflexota Marinisomatota SAR324 clade PAUC34f AncK6 * Halobacterota Verrucomicrobiota Planctomycetota * Nanoarchaeota Hydrothermarchaeota Euryarchaeota Calditrichota Latescibacterota Aerophobota CK-2C2-2 Eremiobacterota Actinobacteriota Sva0485 Desulfobacterota Proteobacteria Poribacteria Cyanobacteria NB1-j RCP2-54 Caldatribacteriota Bacteroidota Entotheonellaeota Dadabacteria Bdellovibrionota Firmicutes Hydrogenedentes Spirochaetota Margulisbacteria Nitrospinota Fusobacteriota Nitrospirota Patescibacteria Deinococcota Acidobacteriota MBNT15 Cloacimonadota Schekmanbacteria Elusimicrobiota Dependentiae Fig. 2 Comparison of bacterial and archaeal communities in the cavity beneath the Ross Ice Shelf with open ocean environments worldwide. a Global map depicting the locations of metagenomic surveys utilized in the analysis and this study. Overlapping of symbols represent locations where multiple depths were sampled. b Phylum-level composition of microbial communities under the Ross Ice Shelf based on 16S rRNA amplicon sequencing (this study). The results for each sequencing triplicate are averaged; results for individual replicates and controls are shown in Supplementary Fig. 2a, b. Comparisons with metagenomic 16S ribosomal RNA genes (miTags) are shown in Supplementary Fig. 2c. c Cluster dendrogram depicting the average linkage hierarchical clustering based on a Bray-Curtis dissimilarity matrix of community compositions, based on the relative abundance of miTags from this study, global ocean expeditions, and Antarctic and Arctic surveys20–23. The dashed box highlights the clustering of communities in the ocean cavity under the Ross Ice Shelf with global deep-sea environments (in detail in 2d). d Heatmap visualization of calculated Z-scores from below-shelf and global deep-sea environments, based on the relative abundance of miTags grouped at phylum level. Column dendrogram shows clustering of samples according to Bray- Curtis dissimilarity index (detailed from 2b). Rows are clustered based on euclidean distance, grouping phyla that are most likely to co-occur in an environment. Asterisks mark phyla that are significantly more abundant under the Ross Ice Shelf (Kruskal-Wallis test, p < 0.05, Supplementary Data 3). Taxonomic assignment is based on the Genome Taxonomy Database (GTDB107). rRNA gene amplicon sequencing and shotgun metagenomic sequencing. The microbial community was dominated by six phyla: Proteobacteria, SAR324, Crenarchaeota (mostly Nitroso- sphaerales), Marinisomatota (formerly Marinimicrobia, SAR406 clade), Chloroflexota (mostly SAR202), and Planctomycetota (Fig. 2b). Consistent with a dark oligotrophic environment, the eukaryotic community was largely comprised of taxa typically found in the meso- and bathypelagic open ocean, including Alveolata, Dinoflagellata, and Rhizaria lineages (Supplementary Fig. 1a, Supplementary Data 1). With respect to viruses, most bacteriophages detected in the metagenomic assemblies (~50%) belonged to uncultured or unclassified taxa (Supplementary Fig. 1b, Supplementary Data 2), with the most abundant classified viruses affiliating with the family Myoviridae (~30%). We used 16S rRNA gene sequences extracted from metage- nomic reads (miTags;19) to profile the relatedness of microbial communities beneath the Ross Ice Shelf to those of marine ecosystems globally (Fig. 2a, c20–23,). This approach enabled comparison of microbial communities from available marine metagenomic datasets, while circumventing potential biases from inter-study community composition comparisons based on amplicon analyses24. In agreement with previous global metage- nomic analyses20, beta diversity analysis (Bray-Curtis dissim- ilarity) showed oceanic microbial communities cluster by depth, in especially though this was less pronounced in polar regions (Fig. 2c, d). In this global context, the communities beneath the Ross Ice Shelf form a cluster that is related to, but distinct from, those of (Fig. 2c, d). When mesopelagic polar open ocean waters compared to deep (>200 m) open ocean communities worldwide, compositional differences between open-ocean and below-shelf microbial communities are evident even at the phylum level (Fig. 2d). For example, the relative abundances of Chloroflexota, Gemmatimonadota, Marinisomatota, Myxococcota, Planctomy- cetota, and SAR324 were significantly higher under the Ross Ice Shelf, test, p = 9.4 × 10−7 − 1.9 × 10−5, full p values shown in Supplemen- tary Data 3). The phyla Halobacterota, Anck6 and PAUC34f, while typically rare in the open dark oceans, showed a tenfold increase in relative abundance in the cavity beneath the Ross Ice Shelf. Analyses restricted to polar environments using MGLM- ANOVA confirmed significant compositional differences between the ocean cavity and deep (>200 m) open-water polar environ- ments (LRT = 17333, p = 0.001, Supplementary Data 3). In addition, Indicator Species Analysis (Indval) congruently identi- fied ‘signature species’ of the ocean cavity (with respect to deep open-water polar communities) belonging to the phyla PAUC34f, Planctomycetota, and SAR324, as well as the classes Lenti- sphaeria, and SAR202 (p = 0.001–0.002, full p values shown in (Kruskal-Wallis deeper layers 4 NATURE COMMUNICATIONS | (2022) 13:117 | https://doi.org/10.1038/s41467-021-27769-5 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-27769-5 ARTICLE Fig. 3 Phylogeny of reconstructed genomes under the Ross Ice Shelf. Phylogenetic genome tree of the 235 metagenome-assembled genomes (MAGs) and single-amplified genomes (SAGs) retrieved from this study. The genomes are labeled by order, shaded by phylum, and numbered as per Supplementary Data 4. Genome characteristics (inner-to-outer circular heatmap): average genome completeness (%) at phylum level, relative abundance expressed as counts per million (CPM) and relative transcriptional activity as transcripts per million (TPM, Log10 + 1 transformed), and presence of marker genes for key metabolic pathways discussed in the main text. Supplementary Data 3). These ‘signature species’ (with IndVal p < 0.05, test statistic >0.5, Supplementary Data 3) represented on average ~10% of the community beneath the Ross Ice Shelf, reaching up to 17% in the mid water column, in comparison to an average abundance of 0.75% in deep polar open waters. Amplicon sequencing analysis provided additional taxonomic resolution of the communities under the ice shelf and confirmed the depth differentiation anticipated from oceanographic and chemical data. Significant differences in community alpha and beta diversity below the Ross Ice Shelf were observed between the basal boundary layer below the ice (30 m) and the deep water column (330 m) (p = 0.028, Supplementary Data 3, Supplemen- tary Figs. 2 and 3). The species driving these differences are described in the Supplementary Notes. Nitrifying archaea and bacteria dominate transcription under the shelf. We used a multi-omics approach to uncover the functional capacity of the microbial community beneath the Ross Ice Shelf, integrating genome-resolved metagenomics, single-cell genomics, and metatranscriptomics. We assembled 235 derepli- cated partial genomes (Fig. 3, Supplementary Figs. 4 and 5; Supplementary Data 4). These comprised 67 SAGs (single- amplified genomes) and 168 manually curated MAGs (meta- genome-assembled genomes), all with completeness >50% and contamination <5%25 (Fig. 3; Supplementary Data 4). These represent on average 50–60% of each sample’s metagenomic and including all phyla with relative metatranscriptomic reads, abundance above 0.5% (Fig. 2) and the top four most abundant genera (Supplementary Fig. 2b). Their phylogenetic diversity, metabolic traits, and relative abundances are depicted in Fig. 3. The presence and transcription of key metabolic genes in assembled and unassembled reads was used to identify prevailing metabolic pathways in the cavity under the Ross Ice Shelf. By far the most highly transcribed genes involved in autotrophic energy conservation pathways were those for oxidation of ammonium (ammonia monoxygenase, amoA) and nitrite (nitrite oxidoreductase, nxrA) (Fig. 4b). Accordingly, ammonium transporters and amoA NATURE COMMUNICATIONS | (2022) 13:117 | https://doi.org/10.1038/s41467-021-27769-5 | www.nature.com/naturecommunications 5 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-27769-5 Fig. 4 Energy conservation and carbon fixation strategies of communities beneath the Ross Ice Shelf. a Dot plot showing the metabolic potential of the 235 metagenome-assembled genomes (MAGs) and single-amplified genomes (SAGs). The size class of each point represents the number of genomes in each class that encode the gene of interest and the shading represents the average genome completeness. b Heatmaps showing the relative abundance of these genes in the three metagenomic and metatranscriptomic unassembled short reads datasets. For metagenome reads, the heatmap shows the abundance of each pathway, expressed as average gene copies per organism (across all genes listed in the pathway) calculated relative to the abundance of 14 universal single-copy ribosomal genes, with scales capped at 1. For metatranscriptome reads, the heatmap shows log10-transformed reads per kilobase million (RPKM). Where genes within the same pathway are collapsed together, the values (community percentage or RPKM) are summed. c Phylogenetic tree of protein sequences of the highly transcribed ammonia monooxygenase subunit A (amoA) gene from archaeal single-amplified genomes and unbinned metagenomic contigs shown in bold compared to reference sequences. See Supplementary Fig. 7 for a detailed version of this tree. were the most transcribed genes overall (Supplementary Fig. 6). Transcription patterns correlated with ammonium concentrations (Table 1) and relative abundance of the archaeal order Nitrosophaer- ales (Supplementary Figs. 2b, 4 and 5). Phylogenetic analysis corroborated that the most numerous amoA genes and transcripts were affiliated with Nitrosopumilus spp. (Fig. 4c, Supplementary Fig. 7), the most abundant and active archaeal lineage beneath the ice shelf (Supplementary Figs. 2b, 4 and 5), with some gammaproteo- bacterial amoA reads also detected (Fig. 4a, Supplementary Fig. 7). The metagenomic and metatranscriptomic reads of the marker gene for nitrite oxidation, nxrA, affiliated with the phyla Nitrospinota and, to a lesser extent Nitrospirota (Supplementary Data 5, Supplementary Fig. 8). In line with an autotrophic lifestyle, we identified the determinants of ammonium- or nitrite-dependent carbon fixation via the archaeal 4-hydroxybutyrate cycle (hbsC, hbsT genes) and Nitrospina reductive tricarboxylic acid cycle (aclB gene) (Fig. 4, Supplementary Figs. 9, 10 and 11; Supplementary Data 3). Consistent with these results, reconstructed genomes from the genera Nitrosopumilus and Nitrospina were among those with highest relative transcriptional activity in our dataset (S4, S5). These groups express a small fraction of their genomes (i.e., ~25% of total genes at 30 m) compared to other community members (Supplementary Fig. 4d–f), devoting most of their transcriptional effort to the key processes of carbon fixation and ammonia and nitrite oxidation, respectively. Despite being well-represented in the metatranscriptomic dataset, the relative abundance of the genus Nitrospina was low in the metagenomic dataset. For instance, the Nitrospina lineage represented by SAG_5 was among the least abundant genomes, but was highly active on the level (RNA/DNA ~270; Supplementary Fig. 5) transcriptional (Supplementary Data 4). These discrepant findings are in line with recent single-cell analyses showing Nitrospinota have high activity despite low abundance;26 it is proposed that the large cell these nitrite oxidizers are size or high mortality rates of responsible for low abundance in metagenomes and amplicon datasets compared to ammonium oxidizers26,27. their Various inorganic and organic energy sources likely support below-shelf bacteria. Many members of the microbial community are capable of supporting or surviving beneath the shelf through a 6 NATURE COMMUNICATIONS | (2022) 13:117 | https://doi.org/10.1038/s41467-021-27769-5 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-27769-5 ARTICLE Genome name Genome Class (Phylum) CAZyme diversity Transcript counts (TPM) MAG_106 SAG_36 MAG_15 MAG_22 MAG_133 MAG_104 MAG_116 MAG_137 MAG_84 MAG_79 MAG_64 MAG_92 MAG_129 MAG_155 MAG_66 MAG_77 MAG_143 MAG_160 MAG_65 MAG_99 MAG_100 MAG_105 MAG_120 MAG_122 MAG_148 MAG_58 MAG_67 MAG_74 MAG_83 MAG_86 MAG_126 MAG_130 MAG_70 MAG_111 MAG_131 MAG_37 MAG_45 MAG_47 MAG_78 MAG_94 MAG_139 MAG_128 MAG_156 MAG_60 MAG_72 MAG_75 MAG_81 MAG_88 MAG_95 SAG_30 Bacteroidia (Bacteroidota) Bacteroidia (Bacteroidota) Dehalococcoidia(Chloroflexota) Dehalococcoidia(Chloroflexota) Hydrogenedentia(Hydrogenedentota) UBA2968 (Latescibacterota) UBA2968 (Latescibacterota) UBA2968 (Latescibacterota) UBA2968 (Latescibacterota) UBA8240 (Latescibacterota) Marinisomatia (Marinisomatota) UBA4248 (Myxococcota) UBA796 (Myxococcota) UBA796 (Myxococcota) UBA796 (Myxococcota) UBA796 (Myxococcota) UBA9160 (Myxococcota) Physciphaerae (Planctomycetota) Physciphaerae (Planctomycetota) Physciphaerae (Planctomycetota) Planctomycetes (Planctomycetota) Planctomycetes (Planctomycetota) Planctomycetes (Planctomycetota) Planctomycetes (Planctomycetota) Planctomycetes (Planctomycetota) Planctomycetes (Planctomycetota) Planctomycetes (Planctomycetota) Planctomycetes (Planctomycetota) Planctomycetes (Planctomycetota) Planctomycetes (Planctomycetota) UBA1135 (Planctomycetota) UBA1135 (Planctomycetota) UBA1135 (Planctomycetota) Gammaproteobacteria (Proteobacteria) Gammaproteobacteria (Proteobacteria) Gammaproteobacteria (Proteobacteria) Gammaproteobacteria (Proteobacteria) Gammaproteobacteria (Proteobacteria) Gammaproteobacteria (Proteobacteria) Gammaproteobacteria (Proteobacteria) Lentisphaeria (Verrucomicrobiota) Verrucomicrobiae (Verrucomicrobiota) Verrucomicrobiae (Verrucomicrobiota) Verrucomicrobiae (Verrucomicrobiota) Verrucomicrobiae (Verrucomicrobiota) Verrucomicrobiae (Verrucomicrobiota) Verrucomicrobiae (Verrucomicrobiota) Verrucomicrobiae (Verrucomicrobiota) Verrucomicrobiae (Verrucomicrobiota) Verrucomicrobiae (Verrucomicrobiota) Nr. of CAZyme genes per genome 20 40 60 80 GH transcription (TPM) 100 50 0 GT transcription (TPM) 200 150 100 50 0 CBM trancription (TPM) 50 40 30 20 10 0 GH GT CBM CAZyme class 30 m 180 m 330 m 30 m 180 m 330 m 30 m 180 m 330 m Fig. 5 Relative abundance and transcription of selected carbohydrate active enzyme (CAZYme) classes. Data is displayed for reconstructed genomes (MAGs and SAGs) where CAZyme diversity was highest (top 50 genomes). Bubble plots represent the number of different genes from each CAZyme class per genome (GH, glycosyl hydrolases; GT, glycosyl transferases; CBD, genes containing carbohydrate binding domains). Heatmaps represent the total gene transcription for each CAZyme class, normalized to total transcripts per sample (transcripts per million, TPM). The data used to construct these plots is provided in Supplementary Data 7. chemoautotrophic or mixotrophic lifestyle. These include gamma- lineages, such as the Thioglobaceae (SUP05 and proteobacterial ARCTIC96BD-19) and UBA10353, which co-encode genes for the Calvin-Benson-Bassham cycle and heterotrophic metabolism. Con- sistently, RuBisCO genes (rbcL) affiliated to sulfur-oxidizing taxa (Supplementary Fig. 10) were transcribed at high levels throughout the water column (Supplementary Fig. 6). The potential of these lineages to fuel chemoautotrophy using reduced sulfur compounds as electron donors is supported by the presence and transcription of marker genes for sulfide oxidation (sqr, r-dsrA) and thiosulfate oxi- dation (soxB) (Fig. 4a, Supplementary Figs. 12, 13 and 14); (Sup- plementary Data 5 and 6). Abundant heterotrophic lineages, such as Marinisomatota and SAR324 (Fig. 2a, Supplementary Fig. 4), also encoded carbon monoxide dehydrogenases (Fig. 4a, Supplementary Fig. 15, Supplementary Data 6); carbon monoxide may serve as an energy source supporting persistence of this community, as we have recently described for other aerobic heterotrophic bacteria28,29. Genes for formate oxidation were also widespread and highly transcribed (Fig. 4b, Supplementary Fig. 6, Supplementary Data 6), whereas few community members are predicted to use H2 (Supplementary Fig. 16, Supplementary Data 6). Metabolic annotations of the derived genomes suggests that many identified taxa in this ecosystem adopt an organohetero- trophic lifestyle. Highly transcribed genes include a wide range of carbohydrate-active enzymes (CAZymes, Fig. 5,30), as well as the substrate-binding protein of the oligopeptide transporter (OppA; Supplementary Fig. 6). The highest enrichment (genes/Mbp), diversity (number of different families), and transcripts of CAZymes were detected in reconstructed genomes of the phyla Hydrogenedentota, Latescibacterota, Myxococcota, Planctomyce- tota, and Verrucomicrobia. The CAZyme-rich genomes were among the most abundant (i.e., with highest coverage) in our study (Supplementary Fig. 4) and belong to the phyla enriched under the Ross Ice Shelf with respect to deep ocean environments (Fig. 2d). These genomes contained glycoside hydrolases, polysaccharide lyases, and glycosyltransferase families required for the utilization of heterogeneous polysaccharide chains, such as alginate, rhamnose, and xylan (Supplementary Data 7). These genomic features are consistent with previous studies describing the capability of these phyla to metabolize recalcitrant organic polymers31–33. Thus, the proportion of the community differen- tially enriched in this ecosystem could be adapted to degrade refractory organic compounds persisting in the advected waters beneath the Ross Ice Shelf. In contrast to their autotrophic counterparts, these heterotrophic populations transcribed a large percentage of their genome (~80%), especially in deeper waters (Supplementary Fig. 4d–f), with transcriptional effort spreading across a variety of substrate-utilization processes. The metatranscriptome also revealed various other processes supporting life beneath the shelf. The heterotrophic majority in this system transcribed genes involved in the acquisition of inorganic and organic nitrogen and phosphorus compounds (e.g., urea, isocyanates, phosphonates, polyphosphonates; Supplemen- tary Fig. 6). Genes encoding for cold adaptation processes (e.g., NATURE COMMUNICATIONS | (2022) 13:117 | https://doi.org/10.1038/s41467-021-27769-5 | www.nature.com/naturecommunications 7 ARTICLE Ross Ice Shelf HWD-2 melting/refreezing ice NH3 ? NH3 NH3 ice basalboundary layer Nitrosopumilus spp. Nitrospina spp. S-oxidizing lithoautotrophs (e.g. Thioglobus spp.) organoheterotrophs (e.g. Latescibacterota ) advected Corg in situ produced Corg NH3 IBL NH3 CO2 HbsC GT AmoA CO2 AclB - NO2 NxrA - NO3 Sred CO2 RbcL 2- SO4 SoxB GH V-IL S-IL BBL GH GH GT Deep cavity circulation Continental Shelf NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-27769-5 estimate that the waters sampled at the borehole location have been in the cavity for as much as four years prior to sampling; this is up to 10-20-fold longer than the time predicted for marine snow from the ocean surface to reach the abyss (~6000 m38,). Likewise, the heterotrophic production rates measured in this study and at borehole J910 were among the lowest measured in environments with similar marine temperatures39. It has been suggested that production rates are highly influenced by the supply and concentration of labile dis- solved organic material39, and thus the water column beneath the ice shelf is predicted to be highly oligotrophic with respect to labile organic matter. ecosystems, including Based on these heterotrophic rates and assuming a heterotrophic prokaryotic growth efficiency of ~5% (typical of deep oceanic waters, e.g.,40.), we estimate a total organic carbon demand (i.e., the combined carbon incorporation into biomass and respiration) of ~6–12 µmol C m−3 d−1. This total carbon demand is in the same range as the carbon fixation rates reported from the environment beneath the J9 borehole (8.3 µmol C m−3 d−112). While the con- tribution of exogenous organic matter remains to be quantified, the close coupling between in situ dark carbon fixation and organic carbon demand suggests that the ecosystem beneath the Ross Ice Shelf is largely sustained by dark carbon fixation. This would differ from deep open ocean environments, where heterotrophic carbon demand significantly relies on the vertical fluxes of particulate organic carbon generated in the euphotic layer41,42. Fig. 6 Schematic illustration of the dominant bacterial and archaeal groups in the water column under the Ross Ice Shelf. Dotted lines represent the three depths sampled below the sea ice in this study (not to scale; for a scaled representation, see Fig. 1). At the lower fringe of the ice basal boundary layer (IBL), high concentrations of ammonium (from a yet unknown source) are likely to drive high relative abundance and transcriptional activity of ammonium oxidizing archaea (Nitrosopumilus ssp.) and nitrite oxidizing bacteria (Nitrospina ssp.). These, together with sulfur-oxidizing chemolithoautrotrophs (belonging to e.g., the genus Thioglobus), are likely the main source of new organic matter to this ecosystem. The representative enzymes for the metabolic pathways are displayed only once for simplicity but were detected at all depths. The heterotrophic majority is characterized by metabolically versatile bacterial lineages (e.g., belonging to the phylum Latescibacterota), encoding and transcribing multiple copies of carbohydrate-active enzymes (CAZymes, such as glycosyl transferases GT, or glycosyl hydrolases, GH). These likely feed on in-situ generated or laterally advected complex organic matter. cold-shock proteins), osmoregulation (e.g., glycine betaine transporters), and motility (i.e., flagellar apparatus) were highly transcribed (Supplementary Fig. 6). The constitutive expression of cold-shock chaperones can protect against cold-induced protein misfolding34 and is likely an adaptive response to maintain protein homeostasis at the very low water temperatures below the shelf. Furthermore, transport of compatible solutes protects the cell against freezing, hyper-osmolality, and desiccation35. Glycine betaine transporters may provide an additional advantage given these transporters were recently shown to be multifunctional, as in addition to the key they transport multiple substrates osmoregulatory compound glycine betaine36. Discussion Collectively, our results provide a detailed insight on the ecolo- gical strategies adopted by communities living in the world’s most extensive sub-ice shelf system. Oceanic cavities below ice shelf systems are uniquely different from open ocean environments in their dependence on in situ chemosynthesis and on lateral advection of food sources from open-water areas, rather than on vertical fluxes of phytoplankton-derived detrital matter37. We Our multi-omic results support this hypothesis, while unco- vering the mediators and pathways responsible for the auto- trophic and heterotrophic activities under the Ross Ice Shelf (Fig. 6). Among the lineages represented by MAGs and SAGs with the highest transcriptional activity are those originating from the chemolithoautotrophic genera Nitrosopumilus and Nitros- pina. Overall, this agrees with previous reports that aerobic ammonium-oxidizing microorganisms are widespread in Ant- arctic marine environments (e.g.,43) and that ammonium oxida- tion occurs beneath Antarctic shelves and sea ice12,44. These and other inferred facultative chemolithoautotrophs (such as facul- tative sulfur-oxidizing bacteria) are likely to be responsible for dark carbon fixation rates previously observed beneath borehole J912 and thus provide a supply of organic carbon to an ecosystem shielded from sunlight. (e.g., Nitrospina, Nitrosopumilus, The importance of dark carbon fixation has been recognized in various oceanic regions during the polar winter. Microbial lineages and Marinisomatota45–47) and enzymes (such as those mediating ammonium, nitrite, and sulfur oxidation48) that mediate che- molithoautotrophy have been observed to increase in Antarctic waters during the transition to the winter season. Likewise, comparable lineages and genes capable of sulfur compound oxi- dation have been detected in winter open waters and the central basin under the Ross Ice Shelf. Together with mounting evidence that sulfur compound oxidizers sustain carbon fixation in the wide dark open ocean (e.g.,49) and the diverse sources of reduced sulfur compounds in marine oxic environments (e.g.,50), it is plausible that these clades can also contribute to chemoauto- trophy in the oceanic cavity beneath the Ross Ice Shelf. SAR324, It is likely that ammonium is a primary energy source sustaining primary production in aphotic Antarctic waters. Consistent with this idea, ammonium oxidation rates have been reported to be higher in Antarctic coastal waters during the austral winter and to significantly support the heterotrophic demand43. In the absence of direct rate measurements in this study, we estimated the ammo- nium oxidation rates potentially supported by the standing ammonium concentrations in the water column. Our estimates for + d−1) are in accordance to rates the basal layer (~90 nM NH4 + d−1) with measured in the Southern Ocean (62 nM NH4 8 NATURE COMMUNICATIONS | (2022) 13:117 | https://doi.org/10.1038/s41467-021-27769-5 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-27769-5 ARTICLE +;43), and comparable ammonium concentrations (0.7 µM NH4 could support the heterotrophic demand in the oceanic cavity under the shelf (Supplementary Notes). These estimates suggests that the microbial communities beneath the Ross Ice Shelf can sustain ammonium oxidation at similar rates to those in the winter Antarctic Ocean and have the potential to be significant primary producers. The ammonium profile beneath the Ross Ice Shelf is intriguing. Contrary to other nutrient concentrations measured (which do not vary significantly through the water column), ammonium concentrations are significantly higher in the ice basal boundary layer compared to the deeper water samples, but comparable to those in the periphery of the shelf4. This profile (exclusive for ammonium with respect to other nitrogen species) is consistent with the reports beneath borehole J912. The proposed circulation model beneath the shelf13, by which the cavity is filled southward by dense water masses that reach its interior via deep cavity circulation, renders it unlikely that the high ammonium con- centrations detected in the fresh, northward flowing waters beneath borehole HW2D or J9 originate from the open Ross Sea. If externally sourced, nutrient concentrations would be expected to be highest in deeper waters, or else be homogenized in the water column as water masses evolve and mix in the cavity. The latter appears to be the case for the other nutrients measured in this and the J9 expedition. The exception observed in the ammonium profile suggests that this compound could be sourced beneath the ice shelf. In particular, terrestrial-origin sediments in the basal ice layer may be a significant source of ammonium to the seawater circulating beneath. Deployment of cameras at HWD2 revealed sedimentary englacial debris in the lower 20 meters of the ice shelf13. While ice melting and freezing can plausibly result in the rainout of the pellets in a sub-ice-shelf cavity, we did not witness this effect; no sediments were retrieved from the pumping samples and the microbial communities sequenced from the englacial debris and the water column were unrelated (Supplementary Fig. 2). However, temperature and salinity data from our study site (Fig. 1b,13) clearly showed ice- shelf basal melting and a supply of freshwater to the upper region of the water column, a phenomenon that could result in the observed replenishment of ammonium concentrations in this system. In free-floating sea ice, as well as in subglacial lakes, ammonium enrichments have been traditionally attributed to wet and dry atmospheric deposition, as well as in situ organic matter regeneration in brine channels, especially within older and thicker ice51–53. The latter may be also a mechanism for ammonium accumulation in deep layers of the ice shelf54, subject to solubi- lization and transport by fresh melt water. If such is the case, the ammonium transported by the ice basal boundary layer could be sourced locally (at borehole HWD2) or elsewhere upstream. Dissolved nutrients in the ice sheet or englacial debris are even- tually diluted as they circulate the interior of the shelf54, which could explain the observed higher concentrations in the water column from borehole J912, 330 km upstream from our study site. While the driving factors of the nutrient profile in the water the tenfold decrease in ammonium column remain unclear, concentrations correlate with changes in relative transcriptional activity of the ammonium-oxidizing genus Nitrosopumilus (Supplementary Fig. 4). As described in Supplementary Notes, we observed depth-related differences in microbial community composition, metabolic capabilities, and gene expression, though additional depth profiles would be required to confirm this. The community members with highest relative abundance and transcriptional activity throughout the water column included nitrifying autotrophic taxa and organoheterotrophic bacteria (Supplementary Figs. 4, 5 and 6). It is likely that the genomes with highest relative transcriptional activity represent two opposite adaptative strategies to the conditions beneath the Ross Ice Shelf. Based on the proportion of their genome expressed, nitrifiers are the surrounding environment by likely to effectively exploit expressing a reduced set of genes encoding a few metabolic pathways. The opposite is observed in the highly expressed het- erotrophic clades (Supplementary Fig. 4). By expressing up to 95% of their genome (e.g., in members of Latescibacterota and Verrucomicrobiota), the transcriptional effort of the latter is spread across a variety of process and in particular, to the exploitation of multiple substrates. These observations are con- sistent with previous studies combining expression and genomic datasets, which suggest that activity levels, substrate utilization and transcriptome diversity may be linked in defining ecological niches of microbial communities55,56. In particular, our results suggest that the most active hetero- trophic organisms are adapted to degrade complex organic com- pounds, including most of the enriched phyla in this environment, such as Myxococcota and Planctomycetota. Their capacity to degrade complex organic material from a range of sources, including potentially of both autochthonous and allochthonous origin, likely confers a major selective advantage in this highly oligotrophic ecosystem. Heterotrophy based on the consumption of recalcitrant dissolved organic carbon has been considered as one possibility for sustaining the oceanic Antarctic winter food web57, and could also be an additional support for life under the Ross Ice Shelf. Unlike organic carbon in Antarctic winter waters, which may have accumulated during the highly productive summer season, organic substrates beneath the Ross Ice Shelf potentially consist of vertically transported exudates and necromass derived from lithoautotrophic primary producers, but also recalcitrant complex organic compounds laterally transported from the Ross Sea into the shelf cavity. Decomposition of phytoplankton entering the shelf cavity is estimated at a scale of ~10 years4. Together with previous reports of diatoms in below-shelf waters9, this indicates that some photoautotrophically-derived organic matter can reach the center of the oceanic cavity. However, the metagenomes suggest that pho- tosynthetic eukaryotes (i.e., class Bacillariophyceae) make a small fraction of the eukaryotic community (0.05 %); this finding is also consistent with undetectable concentrations of chlorophyll a beneath borehole J912. Despite potentially serving as a substrate for organoheterotrophs beneath the ice shelf, phytoplankton are therefore unlikely contributors to the dissolved organic matter pool, whereas detrital sources of bacterial substrates may be more important. Further work is now needed to discriminate organic matter sources and nutrient exchange processes within the shelf. Overall, microorganisms under Antarctica’s ice shelves can thrive in some of the coldest and possibly carbon-limited marine waters, while playing a crucial role in the remineralization of nutrients to the Southern Ocean. Our results not only suggest that the waters below the Ross Ice Shelf are driven by chemo- lithoautotrophic processes, but also uncover the mechanisms responsible for sustaining that activity58. Alongside other recent reports of oceanic dark carbon fixation,27,49,59, this study also emphasizes the importance of inorganic energy sources in driving marine communities in the absence of photosynthesis. Finally, our results suggest that ammonium associated with fresh melt waters at the base of the ice is an important supply of inorganic electron donors supporting chemolithoautotrophy, and thus has a significant the microbial community. Ocean-driven basal melting, a source of freshwater and thus potentially of ammonium in the sub-ice cavity, may increase in a warming climate scenario60. Assuming that our observations are representative of the central region of the cavity under the Ross Ice Shelf, increased basal ice melting could result in an increased vulnerability of communities sup- ported by sub-ice shelf processes61, potentially leading to shifts in influence in the composition and activity of NATURE COMMUNICATIONS | (2022) 13:117 | https://doi.org/10.1038/s41467-021-27769-5 | www.nature.com/naturecommunications 9 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-27769-5 the relative biogeochemical importance of chemolithoautotrophic processes in this extensive ecosystem. These insights emphasize the importance of baseline data from existing sub-ice shelf eco- systems, such as the Ross Ice Shelf, to inform the prediction of biogeochemical impacts of climate change in the Southern Ocean. Methods Site selection and description. Sampling took place in December 2017 and was conducted by members of the Aotearoa New Zealand Ross Ice Shelf Program. Samples were collected from the sub-shelf water column at a site in the central region of the ice shelf, borehole HWD-2 (Latitude -80.6577 N, Longitude 174.4626 W), ~300 km from the shelf front and 330 km northwest of borehole J9 (Fig. 1a). The sampling site is near the glaciological boundary between ice origi- nating from the West Antarctic Ice Sheet and ice flowing from East Antarctica through Transantarctic Mountain glaciers (Fig. 1a). Sediment of terrestrial origin was observed in the lowermost ~60 m of the ice. Hot water drilling and sampling. A hot water drilling system built and operated by the Victoria University of Wellington Drilling Office was used to bore through the ice shelf, creating an access borehole with a maximum diameter of 30 cm. The borehole was used for direct sampling of water and sea floor sediments, and to conduct in situ measurements in the water column. These activities were con- ducted inside a custom-built tent that facilitated 24-h operations in any weather conditions. Seawater samples were obtained from three depths (400 m, 550 m, and 700 m from the top of the shelf, which correspond to 30 m, 180 m, and 330 m deep from the bottom of the ice shelf, respectively). These were chosen to characterize the water column under the Ross Ice Shelf while keeping the sampler ca. 40–50 m away from the seafloor and from ice crystals and sediment in the ice-shelf basal layer. The drilling water was fresh (<15 psu) and relatively warm (between −1 and +1 °C), so it remained stably floating in the borehole and did not sink into deeper layers. This, together with the advection of seawater below the ice shelf, precluded any contamination of collected seawater with the drilling water (Supplementary Fig. 2a, b). The lack of intrusion of the freshwater used for the drilling was rou- tinely checked by salinity and temperature-depth profiles. Samples were collected by in situ filtration using a McLane WTS-LV-Bore Hole filter pump fitted with a 142 mm diameter, 0.22 μm pore-size filter (Supor membrane filters, Pall Corporation). Before and after deployment, the filter holder was thoroughly cleaned to avoid sample cross-contamination. The pump head interior was also flushed after every deployment with fresh water to prevent salt crystal formation and sample contamination. This sampling approach was aimed at obtaining the most realistic representation of the microbial community’s composition and activity with the minimum possible sampling biases. Approximately 200 L of water were filtered at each depth within ca. 2 h. Thereafter, filters were placed in sterile Petri dishes and divided into seven sections using sterile scalpels and transferred to cryovials. The filtered, frozen samples were directly stored in zip lock bags in a 3 m deep borehole drilled into the cold surface snow layer until transported to Scott Base (and further airplane transport to New Zealand). The temperature of the samples deposited in the storage borehole remained stable ranging mostly between –27 °C and –28 °C (Supplementary Fig. 17). These samples were used for 16S rRNA amplicon sequencing, metagenomics, and metatranscriptomics. Water samples (150–300 mL) were also collected at the same three depths using the McLane WTS-LV-Bore Hole pump without a filter-holder in order to further minimize contamination. Once the pump was brought up, it was run in reverse to collect the water, but excluding the first 30–60 mL of water (used for rinsing). Water samples for inorganic nutrient analyses were filtered through combusted Whatman GF/F filters, collected in acid-cleaned HDPE bottles, and stored frozen until analysis in the home laboratory, following procedures recommended by the Joint Global Ocean Flux Study (JGOFS62). The liquid samples for the determination of microbial cell abundance, prokaryotic heterotrophic production, and the generation of single-cell amplified genomes (SAGs) were collected in acid- cleaned Nalgene™ opaque amber HDPE bottles, stored at 2 °C, and transported within 48 h to Scott Base to perform further laboratory analyses. The samples were imported to New Zealand under Ministry for Primary Industry permit number 2017063583 (Permit to import Restricted Biological Products of Animal Origin) issued to the University of Otago Department of Marine Science. To check for potential contamination, samples were also collected from the following sites: freshly melted snow nearby the camp area, drilling water from a reservoir tank, and sediments dislodged from the ice shelf (identified as englacial debris) and collected with the reaming tool. Water samples were filtered onto 0.22 µm polycarbonate filters (47 mm filter diameter, Millipore), and all samples were stored in cryovials and frozen. Physicochemical measurements. A SBE 19plusV2 SeaCAT Profiler CTD (Seabird Electronics, Inc.) was used to measure temperature, salinity and depth within the borehole and in the water under the Ross Ice Shelf for a detailed characterization of the water column. Furthermore, a self-contained single channel logger (RBR Solo) was attached to the frame of the WTS-LV-Bore Hole pump (at the opposite side of the water intake) for an accurate determination of the temperature and depth of the sampling casts. Samples for determining the concentrations of nitrate, dissolved 62 were colorimetrically reactive phosphorus (phosphate), ammonium and SiO2 analyzed using flow-injection analysis on a Lachat Auto-analyzer according to methods described elsewhere63. Measurements of nutrient concentrations were routinely corrected with reference blank solutions in each sample run. No anomalies were detected in the blanks, indicating no source of detectable con- tamination during the measurements. Prokaryotic abundances and heterotrophic production. Prokaryotic abundance was determined by flow cytometry. Samples (1.6 mL) were preserved with glutar- aldehyde (2% final concentration), left at 4 °C in the dark for 15 min, flash-frozen in liquid nitrogen, and stored at −80 °C until analysis. Prior to analysis, the fixed samples were thawed, stained in the dark with a DMS-diluted SYTO-13 dye (Molecular Probes Inc., 2.5 µM final concentration) for 5 min, and run on a BD AccuriTM flow cytometer with a laser emitting at 488 nm wavelength. Samples were run at low or medium speed until 10,000 events were captured. A suspension of yellow–green 1 µm latex beads (105–106 beads mL−1) was added as an internal standard (Polysciences, Inc.). Prokaryotic heterotrophic activity was estimated via the incorporation of 3H-leucine using the centrifugation method64. 3H-leucine (Perkin-Elmer, specific activity 169 Ci mmol−1) was added at saturating concentration (40 nmol L−1) to triplicate 1.2 mL subsamples. Controls were established by adding 120 µL of 50% trichloroacetic acid (TCA) to triplicate control tubes 10 min prior to radioisotope addition. The microcentrifuge tubes were incubated in the dark at 4 °C for 48 h. Incorporation of leucine in the quadruplicate tubes per sample was terminated by adding 120 µL ice-cold 50% TCA. Subsequently, the samples and the controls were kept at –20 °C until centrifugation (at ca. 12,000 × g) for 20 min followed by aspiration of the water. Finally, 1 mL of scintillation cocktail was added to the microcentrifuge tubes before determining the incorporated radioactivity after 24–48 h on a Tri-Carb 2000® Liquid Scintillation Counters scintillation counter (Perkin-Elmer) with quenching correction. The blank-corrected leucine incorporation rates were converted into prokaryotic heterotrophic production (PHP) using the theoretical conversion of 1.55 kg mol−1 leucine incorporated65–67. The rates of leucine incorporation obtained at the incubation temperature (4 °C) were converted to the in situ temperature of -2 °C using an activation energy of 72 kJ mol−1[ 67. Single cell genomics. Sample collection and analyses were performed as described previously27, see Supplementary Methods for full description. Briefly, triplicate seawater samples (1 mL) were transferred to a sterile cryovial containing 100 µL of glyTE (20 mL of 100 × TE buffer pH 8.0, 60 mL Milli-Q water and 100 mL of molecular-grade glycerol), and samples were stored at –80 °C until analysis. SAG generation was performed at the Single Cell Genomic Center at Bigelow Laboratory for Ocean Sciences (SCGC) using fluorescence-activated cell sorting and WGA-X genomic DNA amplification. Paired-end Illumina libraries were created with Nextera XT (Illumina), sequenced with NextSeq 500 (Illumina) and de novo assembled using a workflow based on SPAdes68 as previously described69. The quality of the sequencing reads was assessed using FastQC v0.11.7 (https:// www.bioinformatics.babraham.ac.uk/projects/fastqc/) and the quality of the assembled genomes was determined using CheckM v.1.0.770 and tetramer fre- quency analysis71. This workflow was evaluated for assembly errors using three bacterial benchmark cultures with diverse genome complexity and %GC, indicating no non-target and undefined bases in the assemblies and average frequencies of mis-assemblies, indels and mismatches per 100 kbp: 1.5, 3.0 and 5.069. Functional annotation was first performed using Prokka72 with default Swiss-Prot databases supplied by the software. Prokka was run a second time with a custom protein annotation database built from compiling Swiss-Prot73 entries for Archaea and Bacteria. DNA extraction, 16S rRNA gene amplicon and metagenomic sequencing. DNA was extracted using a PowerSoil® DNA Isolation Kit (MoBio, Carlsbad, CA, USA). The manufacturer’s protocol was modified to use a Geno/Grinder for 2 × 15 s instead of vortexing for 10 min and a final elution of 50 µL solution C6 (sterile elution buffer, 10 mM Tris) was used. DNA concentration was measured using a Nanodrop spectrophotometer (Thermo Fisher). The median 260/280 nm wave- length ratio was 1.5 with a lower quartile of 1.4 and an upper quartile of 1.7. Extractions were performed in triplicate for each depth under the Ross Ice Shelf (total of 9 samples) for subsequent amplicon and metagenomic sequencing. 16S rRNA gene amplicon sequencing was carried out using the Earth Microbiome Project74 protocols and standards (http://earthmicrobiome.org/ protocols-and-standards/16s/), which include the following modifications to the original 515F–806 R primer pair75 (the updated sequences, 5′- 3′, are as follows: 515 F: GTGYCAGCMGCCGCGGTAA; 806 R: GGACTACNVGGGTWTCTAAT). In brief, degeneracy was added to both the forward and reverse primers to remove known biases against Crenarachaeota/Thaumarchaeota (515 F, also called 515F- Y76) and the marine and freshwater Alphaproteobacterial clade SAR11 (806 R77,). All amplicons (independent replicates) were run on an Illumina (Foster City, CA, USA) MiSeq 250 bp × 2 run. For metagenomic sequencing, Thruplex DNA libraries 10 NATURE COMMUNICATIONS | (2022) 13:117 | https://doi.org/10.1038/s41467-021-27769-5 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-27769-5 ARTICLE (~300 bp inserts) were created from each individual DNA extraction and sequenced in an Illumina HiSeq 2500 platform (2 × 125 bp). RNA extraction and metatranscriptomic sequencing. RNA was extracted fol- lowing the RNeasy mini kit (Qiagen, Hilden, Germany) procedure and the ethanol precipitation protocol. The remaining DNA was removed with TurboDNase (Invitrogen, Carlsbad, CA, USA) and the efficiency of removal was tested with PCR. Enrichment of RNA was performed with 20 μL of sample RNA following the procedures of the MICROBEnrich (Ambion, Austin, TX, USA) and MICROBEx- press (Ambion, Austin, TX, USA) kits. Thereafter, the MessageAmp II-Bacteria kit (Invitrogen) was used to improve the subsequent amplification and purification: enriched RNA was reverse transcribed to cDNA, which was in vitro transcribed back to amplified RNA (aRNA) using the mentioned kit. Quantifications were simultaneously run with a Nanodrop spectrophotometer (Thermo Fisher) and a Qubit fluorometer (Invitrogen, Carlsbad, CA, USA) using the RNA HS Assay kit and an RNA profile generated with a Bioanalyzer 2100 (Agilent Technologies, Böblingen, Germany). aRNA was shotgun sequenced directly in an Illumina HiSeq4000 platform (CNAG, Barcelona, Spain), generating between 28–35 Gb of 2 × 101 bp reads per sample. 16S rRNA gene amplicon profiling. Paired-end 16S rRNA gene amplicon sequences were processed on the QIIME2 platform using the DADA2 pipeline to resolve exact amplicon sequence variants78,79. Raw reads were demultiplexed, yielding 302,585 reads across 16 samples. Quality plots were generated and sequences failing to pass an average base call accuracy of 99% (Phred score 20) were excluded. Low quality regions of each sequence were removed by trimming the first 13 bases of the forward and reverse reads and truncating at 150 base pairs before de-noising with DADA2 using the function qiime dada2 denoise-paired with default parameters. The final dataset contained 1228 amplicon sequence variants (ASVs) with a total frequency of 271,736. Taxonomic assignment was performed by using a Genome Taxonomy Database classifier built for the QIIME2 platform, using the SSU sequence files from GTDB ssu_r86.1_20180911 (https://osf.io/25djp/ wiki/home/). The classifier was first spliced to the 515 F/806 R primer pair using the qiime feature-classifier extract-reads, and trained using the qiime feature- classifier fit-classifier-naive-bayes command in QIIME279. The trained classifier was then used to assign the taxonomy to the ASV features using our representative reads via the function feature-classifier classify-sklearn. No sequence overlap was observed between below-shelf waters with those of control samples (e.g., drilling fluid, sediment recovered from basal ice on the shelf, snow at the camp site) (Supplementary Fig. 2), confirming absence of contamination in the water column samples. Metagenomic community profiling. Raw metagenomic and metatranscriptomic paired-end reads were quality-assessed with FastQC v0.11.7 and MultiQC v1.080. BBDuk v38.51 from the BBTools suite (https://sourceforge.net/projects/bbmap/) was used to trim adapter sequences, remove reads corresponding to Illumina’s PhiX sequencing control, trim low-quality bases (minimum quality score 20), and discard short sequences (minimum length 50 bp). The metatranscriptome reads were further processed with SortMeRNA v2.1b81 to remove reads corresponding to prokaryotic and eukaryotic ribosomal RNA, followed by BBDuk to filter low- complexity reads (entropy threshold 0.05). In addition, taxonomic profiling of bacteria, archaeal, and eukaryotic communities was performed with 16S rRNA gene sequences extracted from metagenomic reads (miTags) using a previously described protocol19. miTags were also extracted from bathypelagic samples from the Malaspina Circumnavigation expedition23, metagenomic surveys in the Arctic and Southern Ocean21, as well as metagenomic datasets from polar regions obtained from the TARA Ocean Expedition22. This allowed comparing these datasets to available miTags from epipelagic and mesopelagic samples from the TARA Ocean Expedition20. Extracted 16S and 18S rRNA gene reads were mapped to the SILVA non-redundant SSU Ref database (v.138)82 and assigned to an approximate taxonomic affiliation (nearest taxonomic unit, NTU) using PhyloFlash v3.083 (http://github.com/HRGV/phyloFlash). Bacteriophage prediction was based on identifying viral signals in the metagenomic-assembled contigs (described below) using VirSorter84. In brief, viral-like genes were identified against a curated virome database84 and a set of single-amplified viral genomes85. Abundance of viral contigs was estimated by recruitment of metagenomic reads to viral contigs and calculation of contig coverage. Open reading frames (ORFs) were detected and translated with Prodigal v.2.6.386. Taxonomic classification of the translated sequences was based on sequence homology search87 against the Uniref 100 viral database (http:// virome.dbi.udel.edu; e-value < 10−5) and used to obtain taxonomy classification of viral contigs with the anvi-import-taxonomy function from Anvi’o v.5.288. The metagenomic reads were mapped to the obtained viral contigs using Bowtie 289 (local alignment, sensitive setting). Coverage of viral contigs was calculated by metagenomic read recruitment using Anvi’o. Alpha- and beta-diversity analyses of 16S rRNA amplicons and extracted miTAGs. All statistical analyses were carried out in R v3.5.3. Data manipulation was performed using the R package tidyverse and all visualizations were made using ggplot2. Community richness and beta-diversity was calculated using the R packages Phyloseq90 and Vegan v2.5-691. In total, nine samples representing a triplicate of depth profiles were used for downstream diversity analysis of ASVs (Supplementary Fig. 3, Supplementary Data 3). Rarefaction curves were con- structed to confirm that sequencing depth adequately captured richness in each sample and rarefied using the Phyloseq rarefy_even_depth function with a sample size of 15,400, which represented the minimum sequencing depth to retain 100% of samples used for downstream analysis. Observed richness (counts) and estimated richness (Chao1) was calculated using the estimate richness function in Phyloseq. Normality of the distribution of alpha-diversity estimates was confirmed using a Shapiro-Wilk test and a one-way analysis of variance (ANOVA) to test for sig- nificant differences in richness across depth profiles. As a post-hoc, a Tukey multiple comparison of means was used to confirm which pairs of sites showed significant differences. For beta-diversity analysis on amplicon and miTag data, Bray Curtis distance matrices were calculated in Vegan and visualized using a principal coordinate analysis (PcoA). Independent permutational analysis of var- iance (PERMANOVA) based on the Bray-Curtis dissimilarities values were cal- culated with the adonis function in Vegan (999 random permutations), to test for significant differences in community structure between depth profiles. Finally, a beta-dispersion test (PERMDISP) was applied to confirm that observed differences were not influenced due to dispersion. As a post-hoc evaluation of taxa responsible for differences in microbial community structure, we performed an indicator species analysis. We used the indicator value method92 to calculate indicator values using the R package indicspecies. An individual ASV was considered a valid indicator species if the p value was < 0.05 and the Test statistic (the indicator value) was 0.5 or greater, based on 1000 random permutations93. IndVals were compared between two groups, basal layer (30 m) and mid-column samples (180 m and 330 m), with the multipatt function in the R Indicspecies package (with the option control = how(nperm = 999)). This function uses an extension of the original Indicator Value method: it looks for indicator species of both individual site groups and combinations of site groups94. Counts per NTU (at species-level resolution) of extracted miTAGs were used for comparative analyses between communities under the Ross Ice Shelf and other oceanic samples. Only bacterial and archaeal species with >4 reads per sample were included in the analyses. Samples were divided into four groups, according to sampling depth or location: below-shelf ocean cavity (depth 30–330 m, n = 9), epipelagic (depth <200 m, n = 169), mesopelagic (depth ~200–1000 m, n = 60), and bathypelagic (depth 1000–4000 m, n = 54). The Vegan function vegdist was used to calculate a Bray-Curtis dissimilarity matrix between all samples, which was visualized by hierarchical cluster analysis (average linkage method, function hclust in Vegan). Significant differences (p < 0.05) between relative abundances of taxa from deep (>200 m) open ocean communities worldwide and below-shelf communities were confirmed using a non-parametric one-way analysis of variance (Kruskal-Wallis test, function kruskal.test() in R base). The following comparisons were restricted to two groups from deep, polar environments: samples from mesopelagic and bathypelagic polar environments (n = 42) and samples from the below-shelf cavity (n = 9). As distance-based multivariate methods can confound the within- and between-group effect size and fail to account for the mean variance relationship95, a generalized linear model (GLM) approach was used via the R package mvabund96. A multivariate model was fitted using the manyglm function and negative binominal distribution. To test the multivariate hypothesis of whether species composition varied across sub-ice and open water, the anova function was used which performed an analysis of deviance using likelihood ratio tests (LRT) and PIT-trap resampling of p values using 1000 iterations. To further examine which taxa contribute to compositional changes, a series of univariate tests were performed on each taxon using the p.uni = “adjusted” argument in the anova function. IndVal values were also calculated, using the same parameters described above, to identify which species contributed most to the differences between sub-ice environments and deep open ocean waters, Further, an additional post hoc test for between-group differences was performed with analysis of similarity percentages (simper97,) on a Bray-Curtis dissimilarity matrix calculated as described above. Metagenomic assembly and binning. For assembly, metagenome paired-end reads were error corrected using Bayes Hammer implemented in SPAdes v.3.0.068, merged with BBmerge v.36.3298 and normalized to a kmer depth of 42 with BBnorm v.36.32, from the BBtools program suite. Co-assembly of metagenomes was performed with MEGAHIT v.1.1.199 with merged and unmerged reads. Metagenomic reads were mapped back to the co-assembly (min. length 1 kb) using BBmap v.36.32100 to calculate differential coverage across all samples. Contigs were binned with MetaWatt v.3.5.3101, MaxBin v.2.2.7102 and MetaBAT v.2.12.1103. Bins were automatically de-replicated and aggregated with DasTool104, then manually inspected and refined with Anvi’o v.5.288. Bins classified as Archaea, Gammaproteobacteria, Deltaproteobacteria, Gemmatimonadota, Actinobacteriota, and Chloroflexota were selected from the bulk co-assembly and used for read recruitment with a minimum identity of 70% using BBmap v.36.32. This led to less complex subsets of reads for subsequent re- assembly with a more thorough assembler (SPAdes). For each taxonomic group a separate re-assembly with SPAdes v.3.0.0 was performed followed by a new round of binning as described above and manual refinement in Anvi’o. This procedure NATURE COMMUNICATIONS | (2022) 13:117 | https://doi.org/10.1038/s41467-021-27769-5 | www.nature.com/naturecommunications 11 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-27769-5 improves assembly (i.e., number of scaffolds reduced) and consequently bin metrics such as contig length and purity of bins105. Completeness and quality of final assemblies were assessed by CheckM v.1.0.770, with bins with >50% completeness and <5% contamination (i.e., high and medium quality bins) retained for further analysis25. Genome de-replication, classification, and phylogenetic analysis. Metagenomic bins and single-cell-assembled genomes with >50% completeness were defined as MAGs and SAGs, respectively, and collectively as ‘genomes’ for simplicity. Com- parison and de-replication of genomes were performed with dRep pipeline106. In brief, genomes were grouped at an average nucleotide identity (ANI) of 99%. Representative genomes from each cluster were selected based on the highest ‘genome score’106. This analysis provided a de-replicated genomic database of population genomes. BBmap and samtools were used to recruit reads from the metagenomes (97% identity), and Anvi’o was used to calculate the interquartile (Q2Q3) mean coverage of the de-replicated genomes across samples. On average, 50–60% of each sample’s metagenomic reads mapped to the metagenomic and SAG contigs. MAGs and SAGs were taxonomically assigned using the tool GTDBTk v.0.0.6 (release 80, www.github.com/Ecogenomics/GtdbTk) in accordance to the Genome Taxonomy Database107 (Supplementary Data 4). Phylogenetic tree construction for all 235 MAGS/SAGS was performed using ribosomal protein sequences retrieved from CheckM v.1.0.770 (Fig. 3). The concatenated marker sequence for each genome was aligned using MAFFT108 and an approximate maximum-likelihood phylogenetic tree was generated using FastTree 2109 with default parameters. The tree was then visualized and annotated using the web-based tool iTOL v.6 (https://itol.embl.de). Metabolic profiling of MAGs, SAGs, and assembled unbinned reads. ORFs in binned and unbinned contigs were predicted using Prodigal v.2.6.3.86, with default noise-cut-offs followed by manual filtering using HMM cut-off scores previously described110. The predicted ORFs were automatically annotated with the standard RAST annotation pipeline111, and against the Pfam (release 32.0)112 and TIGRfam (release 15.0)113 HMM models using Interproscan 5114. Phylogenetic trees were constructed to validate findings and to determine which protein classes / lineages were present in the Ross Ice Shelf (Supplementary Figs. 7–16). Trees were constructed for AmoA, NxrA, HbsT, RbcL, AclB, DsrA, Sqr, SoxB, CoxL, and the group 1 h [NiFe]-hydrogenase (HhyL). In all cases, protein sequences retrieved from the MAGs, SAGs, and metagenomic assembled reads by homology-based searches were aligned against a subset of reference sequences from a custom database containing 51 proteins (available at https://doi.org/10.26180/ c.5230745) using ClustalW in MEGA7115. Evolutionary relationships were visualized by constructing maximum-likelihood phylogenetic trees. Specifically, initial trees for the heuristic search were obtained automatically by applying Neighbor-Join and BioNJ algorithms to a matrix of pairwise distances estimated using a JTT model, and then selecting the topology with superior log likelihood value. All residues were used, and trees were bootstrapped with 50 replicates. Annotation of carbohydrate active enzymes (CAZymes) was performed by protein search against the CAZyme HMM database (dbCAN HMMdb release 8.0) following the dbCAN2 CAZyme annotation pipeline116, with stringent parameters for all CAZyme classes (E-value <1e−15 and coverage >0.35). We quantified the number of genes in each genome encoding for different glycosyl hydrolases (GH), glycosyl transferases (GT) and containing carbohydrate binding domains (CBD) (Supplementary Data 7). Heatmaps for the 50 genomes with highest GH diversity were generated in R with ggplot2 (Fig. 6), representing their abundance in the metagenome and the metatranscriptome (as described in the section below). Comparison of abundance and expression of assembled reads. To analyze the expression of annotated ORFs, pre-processed metatranscriptomic paired reads were merged with BBmerge98. Merged and unmerged non-rRNA sequences were mapped to the metagenomic and SAG contigs (99% id) with BBmap (on average, 60% of each sample’s reads were successfully assigned). Quantification of mapped reads per identified gene was performed with the function featureCounts of the R Subread package117. The transcript abundance of each ORF was converted to transcript per million (TPM) (Eq. (1)) for each sampled depth. TPM ¼ A (cid:2) 1=ΣA (cid:2) 106 where A = reads mapped to gene/gene length (kbp). ð1Þ To minimize systematic variability of individual gene abundance, the genome interquartile (Q2Q3) mean coverage (or, for unbinned contigs, the contig’s coverage) was used to define gene abundance in the metagenome. Gene coverage was then converted to counts per million (CPM), to allow for direct comparison with TPM. CPM ¼ B (cid:2) 1=ΣB (cid:2) 106 ð2Þ where B = gene coverage. Data from sample replicates were combined for the above calculations. metatranscriptomic reads were aligned using DIAMOND v0.9.24 to the 1 manually curated protein databases described above and to the predicted ORFs that matched the additional 10 HMMs described above (Supplementary Data 6). DIAMOND mapping was performed with a query coverage threshold of >80% and a gene specific threshold of 40% (RHO), 60% (AtpA, AmoA, MmoA, CoxL, NxrA, NuoF and RbcL), 75% (HbsT), 70% (PsbA, YgfK, ARO, IsoA), (80%) PsaA, or 50% (all other databases), with data further parsed to retain only group 1 and 2 [NiFe]- hydrogenase hits. For the metagenomic data, forward reads with at least 124 bp in length were used. For the metatranscriptomic data, paired-end reads were merged with BBMerge v38.51 and merged reads of at least 124 bp in length were used. Data from sample replicates were combined for this analysis. The abundance of each gene was converted to reads per kilobase million (RPKM). RPKM ¼ X=total sample reads (cid:2) 106 ð3Þ where X = reads aligned to a gene/ gene length (kbp). The gene abundances in RPKM from the metagenomic data were further used to estimate the proportion of the community encoding these functions. The processed metagenomic reads were aligned to each of the 14 universal single-copy ribosomal marker genes available in SingleM (https://github.com/wwood/singlem) with DIAMOND using a query coverage threshold of 80%. Alignments with a bitscore below 40 were removed; the alignment counts were converted to RPKM as described above and averaged across the 14 genes to represent the abundance of a universal single-copy gene. Metabolic gene RPKM values were divided by this value to obtain the average gene copies per organism in each sample (abundance relative to a single-copy gene). Heatmaps representing the community percentage (metagenomic data) and RPKM abundance (metatranscriptomic data) were generated in R with ggplot2 (Fig. 4b). Where genes within the same pathway are collapsed together, the values (community percentage or RPKM) are summed. Reporting summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article. Data availability The data and code underlying Fig. 2a, c, d are provided in the github repository https:// github.com/ClaMtnez/Ocean_tags. The data underlying Figs. 3, 4 & 5 and Supplementary Figs. 1, 4 & 5 are provided as a Supplementary Data Files. The sequence data generated in this study have been deposited in the EMBL Nucleotide Sequence Database (ENA) database under Bioproject PRJEB35712 (metagenomic and metatranscriptomic raw reads, metagenomic and metatranscriptomic assemblies, metagenomic assembled genomes, and single-cell amplified genomes) and in the NCBI Sequence Read Archive (SRA) under Bioproject PRJNA593264 (16S rRNA gene amplicon reads). The following public databases were used in this study: Swiss-Prot database, https://www.uniprot.org/, release-2018_10; Genome Taxonomy Database, https://gtdb.ecogenomic.org/, release 80; SILVA non-redundant SSU Ref database, https://www.arb-silva.de/, v.138; UniRef 100 VIROME database, http:// virome.dbi.udel.edu; Greening lab metabolic marker gene database, https://doi.org/ 10.26180/c.5230745; CAZyme HMM database, https://bcb.unl.edu/dbCAN2/, v.8.0; Pfam HMM database, http://pfam.xfam.org/, release 32.0; and TIGRFAM HMM database, https://www.ncbi.nlm.nih.gov/genome/annotation_prok/tigrfams/, release 15.0 Received: 13 January 2021; Accepted: 30 November 2021; References 1. Fretwell, P. et al. Bedmap2: improved ice bed, surface and thickness datasets for Antarctica. Cryosph 7, 375–393 (2013). 2. Holland, D. M., Jacobs, S. S. & Jenkins, A. Modelling the ocean circulation beneath the Ross Ice Shelf. Antarct. Sci. 15, 13–23 (2003). 3. Robinson, N. J., Williams, M. J. M., Barrett, P. J. & Pyne, A. R. Observations of flow and ice-ocean interaction beneath the McMurdo Ice Shelf, Antarctica. J. Geophys. Res. Ocean. 115, 1–10 (2010). 5. 4. Vick-Majors, T. J. et al. Biogeochemistry and microbial diversity in the marine cavity beneath the McMurdo Ice Shelf, Antarctica. Limnol. Oceanogr. 61, 572–586 (2016). Begeman, C. B. et al. Ocean Stratification and Low Melt Rates at the Ross Ice Shelf Grounding Zone. J. Geophys. Res. Ocean. 123, 7438–7452 (2018). Smethie, W. M. Jr & Jacobs, S. S. Circulation and melting under the Ross Ice Shelf: estimates from evolving CFC, salinity and temperature fields in the Ross Sea. Deep Sea Res. Part I Oceanogr. Res. Pap. 52, 959–978 (2005). 7. Michel, R. L., Linick, T. W. & Williams, P. M. Tritium and carbon-14 6. Metabolic profiling of unassembled metagenome and metatranscriptome reads. The abundance of particular metabolic functions independent of assembly was calculated as previously described118. Briefly, pre-processed metagenomic and 8. distributions in seawater from under the Ross Ice Shelf Project ice hole. Sci. (80-.) 203, 445–446 (1979). Post, A. L. et al. Modern sedimentation, circulation and life beneath the Amery Ice Shelf, East Antarctica. Cont. Shelf Res. 74, 77–87 (2014). 12 NATURE COMMUNICATIONS | (2022) 13:117 | https://doi.org/10.1038/s41467-021-27769-5 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-27769-5 ARTICLE 9. Clough, J. W. & Hansen, B. L. The Ross Ice Shelf Project. Sci. (80-.) 203, 433–434 (1979). 10. Azam, F. et al. Occurrence and metabolic activity of organisms under the Ross Ice Shelf, Antarctica, at station J9. Sci. (80-.) 203, 451–453 (1979). 11. Bruchhausen, P. M. et al. Fish, crustaceans, and the sea floor under the Ross Ice Shelf. Sci. (80-.) 203, 449–451 (1979). 12. Horrigan, S. G. Primary production under the Ross Ice Shelf, Antarctica1. Limnol. Oceanogr. 26, 378–382 (1981). 13. Stevens, C. et al. Ocean mixing and heat transport processes observed under the Ross Ice Shelf control its basal melting. Proc. Natl. Acad. Sci. USA 117, 16799–16804 (2020). 14. Reddy, T. E., Holland, D. M. & Arrigo, K. R. Ross ice shelf cavity circulation, residence time, and melting: Results from a model of oceanic chlorofluorocarbons. Cont. Shelf Res. 30, 733–742 (2010). 15. Gordon, L. I. et al. Seasonal evolution of hydrographic properties in the Ross Sea, Antarctica, 1996–1997. Deep Sea Res. Part II Top. Stud. Oceanogr. 47, 3095–3117 (2000). 16. Tovar-Sánchez, A. et al., Impacts of metals and nutrients released from melting multiyear Arctic sea ice. J. Geophys. Res. C Ocean. 115 (2010). 17. Biggs, D. C., Amos, A. F., Holm-Hansen, O., Oceanographic studies of epi-pelagic ammonium distributions: the Ross Sea NH4+ flux experiment. In Antarctic Nutrient Cycles and Food Webs. 93–103 (Springer Berlin Heidelberg, 1985). 18. Arístegui, J., Gasol, J. M., Duarte, C. M. & Herndld, G. J. Microbial oceanography of the dark ocean’s pelagic realm. Limnol. Oceanogr. 54, 1501–1529 (2009). 19. Logares, R. et al. Metagenomic 16S rDNA Illumina tags are a powerful alternative to amplicon sequencing to explore diversity and structure of microbial communities. Environ. Microbiol. 16, 2659–2671 (2014). 20. Sunagawa, S. et al. Structure and function of the global ocean microbiome. Sci. (80-.) 348, 1261359 (2015). 21. Zhang, W. et al. Structure and function of the Arctic and Antarctic marine microbiota as revealed by metagenomics. Microbiome 8, 1–12 (2020). 22. Salazar, G. et al. Gene expression changes and community turnover differentially shape the global ocean metatranscriptome. Cell 179, 1068–1083.e21 (2019). 23. Duarte, C. M. Seafaring in the 21st century: the malaspina 2010 circumnavigation. Exped. Limnol. Oceanogr. Bull. 24, 11–14 (2015). 40. Baltar, F., Arístegui, J., Gasol, J. M. & Herndl, G. J. Prokaryotic carbon utilization in the dark ocean: growth efficiency, leucine-to-carbon conversion factors, and their relation. Aquat. Microb. Ecol. 60, 227–232 (2010). 41. Baltar, F. et al, Significance of non-sinking particulate organic carbon and dark CO2 fixation to heterotrophic carbon demand in the mesopelagic northeast Atlantic. Geophys. Res. Lett. 37, L09602, https://doi.org/10.1029/ 2010GL043105. 42. Reinthaler, T., van Aken, H. M. & Herndl, G. J. Major contribution of autotrophy to microbial carbon cycling in the deep North. Atlantic’s Inter. Deep Sea Res. Part II Top. Stud. Oceanogr. 57, 1572–1580 (2010). 43. Tolar, B. B. et al. Contribution of ammonia oxidation to chemoautotrophy in Antarctic coastal waters. ISME J. 10, 2605–2619 (2016). 44. Priscu, J., Downes, M., Priscu, L., Palmisano, A. & Sullivan, C. Dynamics of ammonium oxidizer activity and nitrous oxide (N20) within and beneath Antarctic sea ice. Mar. Ecol. Prog. Ser. 62, 37–46 (1990). 45. Luria, C. M., Amaral-Zettler, L. A., Ducklow, H. W. & Rich, J. J. Seasonal succession of free-living bacterial communities in coastal waters of the Western Antarctic Peninsula. Front. Microbiol. 7, 1731 (2016). 46. Grzymski, J. J. et al. A metagenomic assessment of winter and summer bacterioplankton from Antarctica Peninsula coastal surface waters. ISME J. 6, 1901–1915 (2012). 2012 610. 47. Signori, C. N., Pellizari, V. H., Enrich-Prast, A. & Sievert, S. M. Spatiotemporal dynamics of marine bacterial and archaeal communities in surface waters off the northern Antarctic Peninsula. Deep Sea Res. Part II Top. Stud. Oceanogr. 149, 150–160 (2018). 48. Williams, T. J. et al. A metaproteomic assessment of winter and summer bacterioplankton from Antarctic Peninsula coastal surface waters. ISME J. 6, 1883–1900 (2012). 2012 610. 49. Swan, B. K. et al. Potential for chemolithoautotrophy among ubiquitous bacteria lineages in the dark ocean. Sci. (80-.). 333, 1296–1300 (2011). 50. Moran, M. A. & Durham, B. P. Sulfur metabolites in the pelagic ocean. Nat. Rev. Microbiol. 17, 665–678 (2019). 2019 1711. 51. Thomas, D. N. & Dieckmann, G. S. Antarctic sea ice-a habitat for extremophiles. Sci. (80-.) 295, 641–644 (2002). 52. Wolff, E. W. Ice sheets and nitrogen. Philos. Trans. R. Soc. B Biol. Sci. 368, 20130127 (2013). 53. Christner, B. C. et al. A microbial ecosystem beneath the West Antarctic ice sheet. Nature 512, 310–313 (2014) . 24. Tremblay, J. et al. Primer and platform effects on 16S rRNA tag sequencing. 54. Smith, J. A. et al, The marine geological imprint of Antarctic ice shelves. Nat. Front. Microbiol. 6, 771 (2015). 25. Bowers, R. M. et al. Minimum information about a single amplified genome (MISAG) and a metagenome-assembled genome (MIMAG) of bacteria and archaea. Nat. Biotechnol. 35, 725–731 (2017). T. G. S. Consortium. 26. Kitzinger, K. et al. Single cell analyses reveal contrasting life strategies of the two main nitrifiers in the ocean. Nat. Commun. 11, 767 (2020). 27. Pachiadaki, M. G. et al. Major role of nitrite-oxidizing bacteria in dark ocean carbon fixation. Sci. (80-.) 358, 1046–1051 (2017). 28. Cordero, P. R. F. et al. Atmospheric carbon monoxide oxidation is a widespread mechanism supporting microbial survival. ISME J. 13, 2868–2881 (2019). Islam, Z. F. et al. Two Chloroflexi classes independently evolved the ability to persist on atmospheric hydrogen and carbon monoxide. ISME J. 13, 1801–1813 (2019). 29. 30. Cantarel, B. L. et al. The Carbohydrate-Active EnZymes database (CAZy): an expert resource for glycogenomics. Nucleic Acids Res 37, D233–D238 (2009). 31. Yilmaz, P., Yarza, P., Rapp, J. Z. & Glöckner, F. O. Expanding the world of marine bacterial and archaeal clades. Front. Microbiol. 6, 1–29 (2016). 32. Cabello-Yeves, P. J. et al, Reconstruction of diverse verrucomicrobial genomes from metagenome datasets of freshwater reservoirs. Front. Microbiol. 8, 2131 (2017). 33. Youssef, N. H. et al, In silico analysis of the metabolic potential and niche specialization of candidate phylum “Latescibacteria” (WS3). PLoS One. 10, e0127499 (2015). 34. De Maayer, P., Anderson, D., Cary, C. & Cowan, D. A. Some like it cold: understanding the survival strategies of psychrophiles. EMBO Rep. 15, 508–517 (2014). 35. Barria, C., Malecki, M. & Arraiano, C. M. Bacterial adaptation to cold. Microbiology 159, 2437–2443 (2013). 36. Noell, S. E. & Giovannoni, S. J. SAR11 Bacteria have a high affinity and multifunctional glycine betaine transporter. Environ. Microbiol. 21, 2559–2575 (2019). Commun. 10, 5635 (2019). 55. Gifford, S. M., Sharma, S., Booth, M. & Moran, M. A. Expression patterns reveal niche diversification in a marine microbial assemblage. ISME J. 7, 281–298 (2013). 56. Polz, M. F., Hunt, D. E., Preheim, S. P. & Weinreich, D. M. Patterns and mechanisms of genetic and phenotypic differentiation in marine microbes. Philos. Trans. R. Soc. B Biol. Sci. 361, 2009–2021 (2006). 57. Azam, F., Smith, D. C. & Hollibaugh, J. T. The role of the microbial loop in Antarctic pelagic ecosystems. Polar Res 10, 239–244 (1991). Ji, M. et al. Atmospheric trace gases support primary production in Antarctic desert surface soil. Nature 552, 400–403 (2017). 58. 59. Shah, V., Chang, B. X. & Morris, R. M. Cultivation of a chemoautotroph from the SUP05 clade of marine bacteria that produces nitrite and consumes ammonium. ISME J. 11, 263–271 (2017). 60. Timmermann, R. & Hellmer, H. H. Southern Ocean warming and increased ice shelf basal melting in the twenty-first and twenty-second centuries based on coupled ice-ocean finite-element modelling. Ocean Dyn. 63, 1011–1026 (2013). Ingels, J. et al, Antarctic ecosystem responses following ice-shelf collapse and iceberg calving: Science review and future research. Wiley Interdiscip. Rev. Clim. Chang. 12, 12:e682 (2021). 61. 62. Knap, A. H., Michaels, A., Close, A. R., Ducklow, H. and Dickson, A. G. (eds) (1996): Protocols for the Joint Global Ocean Flux Study (JGOFS) Core Measurements. JGOFS Report Nr. 19, vi+170 pp. Reprint of the IOC Manuals and Guides No. 29, (UNESCO, 1994). 63. Parsons, T. R., Maita, Y., Lalli, C. M., A manual of chemical and biological methods for seawater analysis (Pergamon Press, Oxford, UK, 1984). 64. Smith, D. C. & Azam, F. A simple, economical method for measuring bacterial protein synthesis rates in seawater using 3H-leucine 1. Mar. Microb. Food Webs. 6, 107–114 (1992). 65. Kirchman, D. L. In Handbook of methods in aquatic microbial ecology (Lewis Publishers Boca Raton, FL, 1993), 58, 509–512. 37. Gutt, J. et al. Biodiversity change after climate-induced ice-shelf collapse in the Antarctic. Deep Sea Res. Part II Top. Stud. Oceanogr. 58, 74–83 (2011). 38. Nowald, N. et al, in Oceans 2009-Europe (IEEE, 2009). 39. Kirchman, D. L., Morán, X. A. G. & Ducklow, H. Microbial growth in the polar oceans - role of temperature and potential impact of climate change. Nat. Rev. Microbiol. 7, 451–459 (2009). 66. Simon, M. & Azam, F. Protein content and protein synthesis rates of planktonic marine bacteria. Mar. Ecol. Prog. Ser. 51, 201–213 (1989). 67. Lønborg, C. et al. Depth Dependent Relationships between Temperature and Ocean Heterotrophic Prokaryotic Production. Front. Mar. Sci. 3, 90 (2016). 68. Bankevich, A. et al. SPAdes: a new genome assembly algorithm and its applications to single-cell sequencing. J. Comput. Biol. 19, 455–477 (2012). NATURE COMMUNICATIONS | (2022) 13:117 | https://doi.org/10.1038/s41467-021-27769-5 | www.nature.com/naturecommunications 13 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-27769-5 69. Stepanauskas, R. et al.Improved genome recovery and integrated cell-size analyses of individual uncultured microbial cells and viral particles. Nat. Commun.8, 84 (2017). 101. Strous, M., Kraft, B., Bisdorf, R. & Tegetmeyer, H. E. The binning of metagenomic contigs for microbial physiology of mixed cultures. Front. Microbiol. 3, 410 (2012). 70. Parks, D. H., Imelfort, M., Skennerton, C. T., Hugenholtz, P. & Tyson, G. W. CheckM: assessing the quality of microbial genomes recovered from isolates, single cells, and metagenomes. Genome Res. 25, 1043–1055 (2015). 71. Woyke, T. et al. Assembling the marine metagenome, one cell at a time. PLoS 102. Wu, Y.-W., Simmons, B. A. & Singer, S. W. MaxBin 2.0: an automated binning algorithm to recover genomes from multiple metagenomic datasets. Bioinformatics 32, 605–607 (2016). 103. Kang, D. D., Froula, J., Egan, R. & Wang, Z. MetaBAT, an efficient tool for One 4, e5299 (2009). 72. Seemann, T. Prokka: rapid prokaryotic genome annotation. Bioinformatics 30, 2068–2069 (2014). 73. Poux, S. et al. On expert curation and scalability: UniProtKB/Swiss-Prot as a case study. Bioinformatics 33, 3454–3460 (2017). 74. Thompson, L. R. et al. A communal catalogue reveals Earth’s multiscale microbial diversity. Nat 551, 457–463 (2017). 2017 5517681. 75. Caporaso, J. G. et al. Ultra-high-throughput microbial community analysis on the Illumina HiSeq and MiSeq platforms. ISME J. 6, 1621–1624 (2012). 76. Parada, A. E., Needham, D. M. & Fuhrman, J. A. Every base matters: assessing small subunit rRNA primers for marine microbiomes with mock communities, time series and global field samples. Environ. Microbiol. 18, 1403–1414 (2016). 77. Apprill, A., McNally, S., Parsons, R. & Weber, L. Minor revision to V4 region SSU rRNA 806R gene primer greatly increases detection of SAR11 bacterioplankton. Aquat. Microb. Ecol. 75, 129–137 (2015). 78. Callahan, B. J. et al. DADA2: High-resolution sample inference from Illumina amplicon data. Nat. Methods 13, 581–583 (2016). 79. Bolyen, E. et al. Reproducible, interactive, scalable and extensible microbiome data science using QIIME 2. Nat. Biotechnol. 37, 852–857 (2019). 80. Ewels, P., Magnusson, M., Lundin, S. & Käller, M. MultiQC: summarize analysis results for multiple tools and samples in a single report. Bioinformatics 32, 3047–3048 (2016). 81. Kopylova, E., Noé, L. & Touzet, H. SortMeRNA: fast and accurate filtering of ribosomal RNAs in metatranscriptomic data. Bioinformatics 28, 3211–3217 (2012). 82. Quast, C. et al. The SILVA ribosomal RNA gene database project: improved data processing and web-based tools. Nucleic Acids Res 41, D590–D596 (2012). 83. Gruber-Vodicka, H. R., Seah, B. K. B. & Pruesse, E. phyloFlash: Rapid small- subunit rRNA profiling and targeted assembly from metagenomes. Msystems 5, e00920–e00920 (2020). 84. Roux, S., Enault, F., Hurwitz, B. L. & Sullivan, M. B. VirSorter: mining viral signal from microbial genomic data. PeerJ 3, e985 (2015). 85. Martinez-Hernandez, F. et al. Single-virus genomics reveals hidden cosmopolitan and abundant viruses. Nat. Commun. 8, 15892 (2017). 86. Hyatt, D. et al. Prodigal: prokaryotic gene recognition and translation initiation site identification. BMC Bioinforma. 11, 119 (2010). 87. Camacho, C. et al. BLAST+: architecture and applications. BMC Bioinforma. 10, 421 (2009). 88. Eren, A. M. et al. Anvi’o: an advanced analysis and visualization platform for ‘omics data. PeerJ 3, e1319 (2015). 89. Langmead, B. & Salzberg, S. L. Fast gapped-read alignment with Bowtie 2. Nat. Methods 9, 357–359 (2012). 90. McMurdie, P. J. & Holmes, S. phyloseq: an R package for reproducible interactive analysis and graphics of microbiome census data. PLoS One 8, e61217 (2013). 91. Oksanen, J. et al, vegan: Community Ecology Package. 2019. R package version 2.5-6 (2019). 92. Dufrêne, M. & Legendre, P. Species assemblages and indicator species: the need for a flexible asymmetrycal approach. Ecol. Monogr. 67, 345–366 (1997). 93. De Cáceres, M., Legendre, P., Wiser, S. K. & Brotons, L. Using species combinations in indicator value analyses. Methods Ecol. Evol. 3, 973–982 (2012). 94. De Cáceres, M., Legendre, P. & Moretti, M. Improving indicator species analysis by combining groups of sites. Oikos 119, 1674–1684 (2010). 95. Warton, D. I., Wright, S. T. & Wang, Y. Distance-based multivariate analyses confound location and dispersion effects. Methods Ecol. Evol. 3, 89–101 (2012). 96. Wang, Y., Naumann, U., Wright, S. T. & Warton, D. I. mvabund– an R package for model-based analysis of multivariate abundance data. Methods Ecol. Evol. 3, 471–474 (2012). 97. Clarke, K. R. Non‐parametric multivariate analyses of changes in community structure. Aust. J. Ecol. 18, 117–143 (1993). 98. Bushnell, B., Rood, J. & Singer, E. BBMerge–accurate paired shotgun read accurately reconstructing single genomes from complex microbial communities. PeerJ 3, e1165 (2015). 104. Sieber, C. M. K. et al. Recovery of genomes from metagenomes via a dereplication, aggregation and scoring strategy. Nat. Microbiol. 3, 836–843 (2018). 105. Albertsen, M. et al. Genome sequences of rare, uncultured bacteria obtained by differential coverage binning of multiple metagenomes. Nat. Biotechnol. 31, 533–538 (2013). 106. Olm, M. R., Brown, C. T., Brooks, B. & Banfield, J. F. dRep: a tool for fast and accurate genomic comparisons that enables improved genome recovery from metagenomes through de-replication. ISME J. 11, 2864–2868 (2017). 107. Parks, D. H. et al. A standardized bacterial taxonomy based on genome phylogeny substantially revises the tree of life. Nat. Biotechnol. 36, 996 (2018). 108. Katoh, K. & Standley, D. M. MAFFT multiple sequence alignment software version 7: improvements in performance and usability. Mol. Biol. Evol. 30, 772–780 (2013). 109. Price, M. N., Dehal, P. S. & Arkin, A. P. FastTree 2–approximately maximum- likelihood trees for large alignments. PLoS One 5, e9490 (2010). 110. Anantharaman, K. et al. Thousands of microbial genomes shed light on interconnected biogeochemical processes in an aquifer system. Nat. Commun. 7, 13219 (2016). 111. Aziz, R. K. et al. The RAST Server: Rapid Annotations using Subsystems Technology. BMC Genomics 9, 75 (2008). 112. El-Gebali, S. et al. The Pfam protein families database in 2019. Nucleic Acids Res 47, D427–D432 (2019). 113. Haft, D. H., Selengut, J. D. & White, O. The TIGRFAMs database of protein families. Nucleic Acids Res 31, 371–373 (2003). 114. Jones, P. et al. InterProScan 5: genome-scale protein function classification. Bioinformatics 30, 1236–1240 (2014). 115. Kumar, S., Stecher, G. & Tamura, K. MEGA7: molecular evolutionary genetics analysis version 7.0 for bigger datasets. Mol. Biol. Evol. 33, 1870–1874 (2016). 116. Zhang, H. et al. dbCAN2: a meta server for automated carbohydrate-active enzyme annotation. Nucleic Acids Res 46, 95–101 (2018). 117. Liao, Y., Smyth, G. K. & Shi, W. featureCounts: an efficient general purpose program for assigning sequence reads to genomic features. Bioinformatics 30, 923–930 (2013). 118. Ortiz, M. et al, bioRxiv, in press. 119. Scambos, T. A., Haran, T. M., Fahnestock, M. A., Painter, T. H. & Bohlander, J. MODIS-based Mosaic of Antarctica (MOA) data sets: Continent-wide surface morphology and snow grain size. Remote Sens. Environ. 111, 242–257 (2007). Acknowledgements We are grateful to D. V. Meier and P. M. Leung for insightful discussions. We thank the Victoria University of Wellington Hot Water Drilling Team led by A. Pyne and D. Mendeno for fieldwork support. Staff of the Bigelow Laboratory Single Cell Genomics Center are acknowledged for generating SAG data. We also thank L. Montiel and V. Balagué from the Institut de Ciències del Mar (ICM, CSIC) for extracting RNA, and the CNAG staff for RNAseq library preparation and sequencing. Bioinformatics analyses were performed at the LiSC Cluster (University of Vienna), MARBITS platform (ICM, CSIC) and the MonARCH HPC Cluster (Monash University). This research was facilitated by the New Zealand Antarctic Research Institute (NZARI) funded Aotearoa New Zealand Ross Ice Shelf Programme, the New Zealand Antarctic Science Platform ANTA1801, the Austrian science fond (FWF) project AP3430411/21 (FB) and a Rutherford Discovery Fellowship from the Royal Society of New Zealand (FB), the US National Science Foundation grants DEB-1441717 (RS) and OCE 1335810 (RS), the Simons Foundation Grant 827839 (RS), the Austrian Science Fund project P28781-B21 (GJH), the Spanish Ministry of Science and Innovation (Spanish State Research Agency, https://doi.org/10.13039/501100011033) fellowship RYC-2013-12554 (RL) and projects CTM2015-69936-P (RL) and PID2019-110011RB-C32 (JMG), the NHMRC EL2 Fel- lowship APP1178715 (CG) and Discovery Project grant DP180101762 (CG), the ARC SRIEAS Grant SR200100005 Securing Antarctica’s Environmental Future (SKB), and the H2020 MSCA Individual Fellowship 886198 (CMP). merging via overlap. PLoS One 12, e0185056 (2017). 99. Li, D. et al. MEGAHIT v1.0: A fast and scalable metagenome assembler driven by advanced methodologies and community practices. Methods 102, 3–11 (2016). 100. Li, H. et al. The Sequence Alignment/Map format and SAMtools. Bioinformatics 25, 2078–2079 (2009). Author contributions F.B., C.H., S.E.M., and C.O. designed field experiments. F.B., S.E.M., C.H., C.O., C.S., and B.T. performed field sampling and measurements. S.E.M. and R.L. performed nucleic acid extraction and library preparation for metagenomics and metatranscriptomics, respectively. R.S. provided single-cell amplified genome sequencing. C.M.P., Z.Z., R.J.L., 14 NATURE COMMUNICATIONS | (2022) 13:117 | https://doi.org/10.1038/s41467-021-27769-5 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-27769-5 ARTICLE S.K.B,. D.D.C., B.T., J.M.G., F.B., and C.G. analyzed the data. C.M.P., C.G., and F.B. wrote the manuscript with assistance from all coauthors. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Competing interests The authors declare no competing interests. Additional information Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41467-021-27769-5. Correspondence and requests for materials should be addressed to Sergio E. Morales or Federico Baltar. Peer review information Nature Communications thanks Jeff Bowman and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/ licenses/by/4.0/. Reprints and permission information is available at http://www.nature.com/reprints © The Author(s) 2022 NATURE COMMUNICATIONS | (2022) 13:117 | https://doi.org/10.1038/s41467-021-27769-5 | www.nature.com/naturecommunications 15
10.15252_embj.2022112118
Article A critical period of prehearing spontaneous Ca2+ spiking is required for hair-bundle maintenance in inner hair cells Adam J Carlton1 , Jing-Yi Jeng1 Lara De Tomasi1, Anna Underhill1 Guy P Richardson4 , Fiorella C Grandi2 , Stuart L Johnson1,3, Kevin P Legan4 , Mirna Mustapha1,3 & Walter Marcotti1,3,* , Francesca De Faveri1 , Corn(cid:1)e J Kros4 , , Federico Ceriani1 , Abstract Sensory-independent Ca2+ spiking regulates the development of mammalian sensory systems. In the immature cochlea, inner hair cells (IHCs) fire spontaneous Ca2+ action potentials (APs) that are generated either intrinsically or by intercellular Ca2+ waves in the nonsensory cells. The extent to which either or both of these Ca2+ signalling mechansims are required for IHC maturation is unknown. We find that intrinsic Ca2+ APs in IHCs, but not those elicited by Ca2+ waves, regulate the maturation and maintenance of the stereociliary hair bundles. Using a mouse model in which the potassium channel Kir2.1 is reversibly overexpressed in IHCs (Kir2.1-OE), we find that IHC membrane hyperpolarization prevents IHCs from generating intrinsic Ca2+ APs but not APs induced by Ca2+ waves. Absence of intrinsic Ca2+ APs leads to the loss of mechanoelectrical transduction in IHCs prior to hearing onset due to progressive loss or fusion of stereocilia. RNA-sequencing data show that pathways involved in morphogenesis, actin filament- based processes, and Rho-GTPase signaling are upregulated in Kir2.1-OE mice. By manipulating in vivo expression of Kir2.1 chan- nels, we identify a “critical time period” during which intrinsic Ca2+ APs in IHCs regulate hair-bundle function. Keywords calcium waves; development; hair cell; mechanoelectrical transduction; spontaneous action potentials Subject Category Neuroscience DOI 10.15252/embj.2022112118 | Received 16 July 2022 | Revised 22 November 2022 | Accepted 28 November 2022 | Published online 3 January 2023 The EMBO Journal (2023) 42: e112118 Introduction Inner hair cells (IHCs) are the primary sensory receptors of the adult mammalian cochlea and relay acoustic information onto type I 1 School of Biosciences, University of Sheffield, Sheffield, UK 2 Gladstone Institute of Neurological Disease, San Francisco, CA, USA 3 Neuroscience Institute, University of Sheffield, Sheffield, UK 4 School of Life Sciences, University of Sussex, Falmer, Brighton, UK *Corresponding author. Tel: +44 114 2221098; E-mail: w.marcotti@sheffield.ac.uk spiral ganglion afferent neurons via the graded release of glutamate from their specialized ribbon synapses (Fuchs, 2005; Moser et al, 2020). Before hearing onset, however, which in most altricial rodents occurs at around postnatal day 12 (P12) (Mikaelian & Ruben, 1964; Ehret, 1983; Romand, 1983), IHCs exhibit patterned action potential activity that is elicited spontaneously in the absence of sound-induced stimulation by the activation of CaV1.3 Ca2+ chan- nels (Marcotti et al, 2003a; Tritsch et al, 2010; Johnson et al, 2011). This activity has been shown to drive the bursting-like firing pattern along the neural pathway of the immature auditory system (Lippe, 1994; Jones et al, 2007; Sonntag et al, 2009; Tritsch et al, 2010). As with other sensory systems (Katz & Shatz, 1996; Stellwagen & Shatz, 2002; Moody & Bosma, 2005; Blankenship & Feller, 2010), patterned peripheral firing activity was identified as being critical for the refinement of neural circuits in the brain (Clause et al, 2014, 2017; M€uller et al, 2019; Maul et al, 2022). Additionally, Ca2+ -dependent APs in IHCs have been shown to instruct the normal functional dif- ferentiation of the IHCs themselves (Johnson et al, 2007, 2013), most likely via regulating gene expression (Dolmetsch et al, 1997). However, due to the complex extracellular modulation of the intrin- sic Ca2+ action potentials in developing IHCs, the exact role of this activity is largely unknown. Spontaneous intrinsic Ca2+ action potentials first appear in the IHCs of the mouse cochlea at late embryonic stages (Marcotti et al, 2003a), and their frequency and pattern are controlled by the transiently expressed small-conductance Ca2+ current ISK2 (Marcotti + et al, 2004) and the inward rectifier K current IK1 (Marcotti et al, 1999). The frequency and pattern of the electrical activity in IHCs are also extrinsically evoked and modulated by the spontaneous release of ATP from the neighboring nonsensory cells (Tritsch et al, 2010; Wang et al, 2015; Johnson et al, 2017). This complex regulation makes it difficult to identify and separate the specific functional roles of the intrinsic and externally driven Ca2+ -dependent AP activity in IHCs. In this study, we used a mouse model in which the inward recti- channel Kir2.1 (Yu et al, 2004) was selectively overexpressed + -activated K + fier K (cid:1) 2023 The Authors. Published under the terms of the CC BY 4.0 license. The EMBO Journal 42: e112118 | 2023 1 of 20 The EMBO Journal Adam J Carlton et al in vivo in the IHCs under the control of doxycycline (DOX), lowering their membrane potential and preventing them from firing the intrinsic spontaneous Ca2+ action potentials. These “silent” IHCs, however, retained their ability to respond with AP activity to extrin- sic modulation by the ATP-induced signaling from the nonsensory cochlear cells. Our results show that prehearing IHCs require spon- taneous intrinsic Ca2+ firing to maintain the normal morphological and biophysical characteristics of the mechanoelectrical transducer apparatus for a period of time in the second postnatal week, before the onset of hearing. We also found several key genes that are upregulated in the absence of the intrinsic Ca2+ action potential activity in IHCs, several of which are involved in pathways related to maintaining cytoskeletal homeostasis. Results Overexpression of Kir2.1 (Kir2.1-OE) in cochlear IHCs in vivo prevents spontaneous firing activity The role of spontaneous Ca2+ action potential activity in IHCs, which are spikes generated intrinsically as opposed to those induced by Ca2+ waves originating in the nonsensory cells, was investigated + by conditionally overexpressing the inwardly rectifying K channel Kir2.1 (Kcnj2) in the IHCs, thereby hyperpolarizing their resting membrane potential. ; Kir2.1+/(cid:1) DOX-induced overexpression of Kir2.1 channels in the IHCs was evident from the presence of Kir2.1 immunofluorescence in the baso- lateral membrane of P6 (Fig EV1B) and P11 Kir2.1-OE mice (OtofrtTA+/(cid:1) : Fig 1B), but not in age-matched littermate con- trol mice that were also exposed to DOX (OtofrtTA+/(cid:1); Kir2.1+/+ : P6, Fig EV1A; P9-P11, Fig 1A). OHCs and nonsensory cells surrounding the hair cells showed no or very little overexpression of Kir2.1 (Appendix Fig S1), indicating specificity of the Otof promoter for tar- geting the IHCs. Prehearing IHCs overexpressing Kir2.1 showed a sig- + nificantly larger inward K current compared with control cells but + currents (Fig 1C–G, P9-P11). The larger inward normal outward K + current in the IHCs from Kir2.1-OE led to a hyperpolarized shift of K the resting membrane potential (Vm) of the IHCs of about 10 mV compared with control cells (Fig 1H). The slope conductance around the respective resting Vm values was also significantly increased in IHCs from Kir2.1-OE mice compared with control littermates (Fig 1I). The overexpression of Kir2.1 in neonatal P4 mice had a similar effect the IHC basolateral membrane on the biophysical properties of (Fig EV1D) as that described in P9-P11 IHCs (Fig 1D). We also found that the number of presynaptic ribbons, postsynaptic glutamate receptors and their co-localization in prehearing IHCs was not affected by the overexpression of Kir2.1 channels (Fig EV2A–D). In agreement with the normal morphological profile of the synapses, exocytosis in IHCs was not significantly different between the two genotypes (P = 0.4709, 2-way ANOVA, Fig EV2E and F). These data indicate that the overexpression of Kir2.1 channels is not affecting the expression of the ion channels that are normally present in develop- ing IHCs or their ribbon synapses. We then investigated the ability of IHCs to fire intrinsic and induced Ca2+ action potentials at near body temperature (34–37°C) with an in vivo endolymph-like solution surrounding the IHC hair bundles. IHC Ca2+ action potentials are elicited by the opening of Ca2+ channels that activate at around (cid:1)60 mV (Marcotti et al, 2003a). During the first postnatal week, the ionic composition of the endolymph is comparable to that of the perilymph, which contains 1.3 mM Ca2+ (Wangemann & Schacht, 1996). Under these recording conditions, spontaneous Ca2+ spiking activity was recorded from P4 control IHCs (Fig 2A). The mean spike frequency of IHCs was 2.19 (cid:3) 1.09 Hz (n = 6), and the coefficient of variation (CV) was 1.30 (cid:3) 0.63 (n = 7, duration of the recordings 45–101 s), which being greater than one, is indicative of a bursting pattern of activity as previously demonstrated (Johnson et al, 2011). In P4 Kir2.1-OE mice, due to the more hyperpolarized resting Vm, IHCs do not fire action potentials spontaneously, although they retain the ability to do so during large depolarizing current injections (Fig 2B). During the second postnatal week, spontaneous action potentials in IHCs disappear when using ex vivo cochlear preparations, which is due to a progressive hyperpolarization of the IHC resting Vm (Marcotti et al, 2003b) but could still be elicited by depolarizing current injec- tions (Fig 2C, top panel). This membrane hyperpolarization is likely to be compensated in vivo by the resting open probability of the mechanoelectrical transducer (MET) channel (Johnson et al, 2012). This is because in vivo the endolymphatic Ca2+ concentration during the second postnatal week has been estimated to be near 0.3 mM (Johnson et al, 2012), which will increase the open probability of the MET channels and thus cause the IHCs to depolarize to around the action potential threshold (Fig 2C, bottom panel; for spike fre- quency and CV see Fig 7E). We found that the IHCs from Kir2.1-OE mice failed to elicit spontaneous action potentials even in the esti- mated 0.3 mM endolymphatic Ca2+ concentration, causing the IHCs to remain silent at rest (Fig 2D). IHCs are surrounded by nonsensory cells in the greater epithelial ridge (GER, also known as Ko¨lliker’s organ: Fig 2E). The release of ATP from nonsensory cells of the GER leads to spatially and tempo- rally coordinated Ca2+ waves that propagate across the epithelium and cause IHCs to depolarize as much as 28 mV (Tritsch et al, 2010). This depolarization has been shown to produce periodic bursts of Ca2+ action potentials in IHCs (Tritsch et al, 2007, 2010; Wang et al, 2015; Johnson et al, 2017). The frequency and duration of the Ca2+ waves in the nonsensory cells were not affected by the overexpression of the Kir2.1 channels (Fig EV3A and B). Therefore, we investigated whether the more hyperpolarized IHCs (by about 10 mV) from Kir2.1-OE mice retained the ability to respond to spon- taneous Ca2+ waves originating in the GER. We found that in the presence of the estimated in vivo endolymph-like Ca2+ (0.3 mM), signals caused by the opening of the Ca2+ the Ca2+ channels in IHCs followed very closely the time course of the Ca2+ wave originating in the GER in both control (Fig 2F) and Kir2.1-OE (Fig 2G) P7-P9 mice. Moreover, the correlation between IHC Ca2+ activity and Ca2+ waves in the nonsensory cells was unaffected in Kir2.1 mice (Fig EV3C). This indicates that the large depolarization caused by the extracellular input of the nonsensory cells was necessary and sufficient to depolarize the IHCs in Kir2.1-OE mice and cause the opening of voltage-gated calcium channels. Progressive loss of mechanoelectrical transduction in IHCs lacking intrinsic Ca2+ action potentials MET currents were recorded from apical-coil IHCs by displacing their hair bundles using a 50 Hz sinusoidal force stimulus from a 2 of 20 The EMBO Journal 42: e112118 | 2023 (cid:1) 2023 The Authors Adam J Carlton et al The EMBO Journal Figure 1. Basolateral membrane properties of IHCs overexpressing Kir2.1 channels. A, B Maximum intensity projections of confocal z-stacks taken from the apical cochlear region of control (A) and littermate Kir2.1 overexpressing (B, Kir2.1-OE) mice at postnatal day 11. Inner hair cells (IHCs) were stained with antibodies against Kir2.1 (green) and the hair cell marker Myo7a (blue). At least 3 mice for each genotype were used. Scale bars: 10 lm. C, D Currents from IHCs of control (C, P9) and Kir2.1-OE (D, P10) prehearing mice. Currents were elicited by using depolarizing and hyperpolarizing voltage steps, with a nominal increment of 10 mV, from a holding potential of (cid:1)84 mV. Test potentials are shown next to some of the traces. Note that the large inward rectifier Kir2.1 current is only present in the IHC of the Kir2.1-OE mouse (D). The outward current is primarily carried by a delayed rectifier current IK. IK1 identifies the small inwardly rectifying K+ current normally expressed in IHCs. Steady-state current–voltage curves obtained from IHCs of control (P9-P11) and Kir2.1-OE (P9-P11) mice. E F, G Size of the total steady-state outward (F, IK: Control 2.88 (cid:3) 1.07 nA, n = 8; Kir2.1-OE 2.25 (cid:3) 0.97 nA, n = 6) and inward (G, Control, IK1: 0.30 (cid:3) 0.05 nA, n = 8; Kir2.1-OE, IKir2.1: 3.13 (cid:3) 1.37 nA, n = 6) K+ currents measured at 0 mV and (cid:1) 124 mV, respectively. n.s: P = 0.2836. Resting membrane potential (Vm) measured in IHCs from Control ((cid:1)62.6 (cid:3) 3.8 mV, n = 7) and Kir2.1-OE ((cid:1)73.5 (cid:3) 3.5 mV, n = 5). Slope conductance of the current measured at around the respective resting Vm (Control 1.8 (cid:3) 0.4 nS, n = 8; Kir2.1-OE 25.5 (cid:3) 9.6 nA, n = 6). H I Data information: In panels F–I, data are shown as means (cid:3) SD, and the single cell value recordings (open symbols) are plotted with the average data. All statistical tests were performed using the Student’s t-test. The number of IHCs investigated is shown above the average data points (6 control and 3 Kir2.1-OE mice). Source data are available online for this figure. piezo-driven fluid jet (Corns et al, 2018; Carlton et al, 2021). A large MET current was elicited in all IHCs tested from control (Fig 3A) and Kir2.1-OE (Fig 3B) mice at P6-P7 when their stereociliary bundles were moved towards the taller stereocilia (i.e., in the excita- tory direction) at negative membrane potentials. By stepping the membrane potential from (cid:1)124 mV to more depolarized values in 20 mV increments, the transducer current decreased in size at first and then reversed near 0 mV in IHCs from both genotypes (Fig 3A– C), consistent with the nonselective permeability of MET channels to cations. The maximal MET current at both (cid:1)124 mV and +96 mV was not significantly different between control and littermate Kir2.1- OE mice (P = 0.2269 and P = 0.3620, respectively, t-test, Fig 3D). The resting open probability of the MET channel, which is derived from the current flowing through open transducer channels in the absence of mechanical stimulation (arrows: Fig 3A and B), was also not significantly different between the IHCs from the two genotypes ((cid:1)124 mV: P = 0.3766; +96 mV: P = 0.2846, Fig 3E). At P8-P9, the size of the MET currents became more variable in the IHCs overex- pressing the Kir2.1 channel, with some cells showing a third of the current recorded from control mice (Fig 4A–C). Overall, the size of (cid:1) 2023 The Authors The EMBO Journal 42: e112118 | 2023 3 of 20 The EMBO Journal Adam J Carlton et al Figure 2. Kir2.1 overexpression prevents spontaneous, but not induced, Ca2+ action potentials in IHCs. A, B Whole-cell recordings of Ca2+ action potential activity in apical-coil IHCs from P4 control (A) and Kir2.1-OE (B) mice in the presence of 1.3 mM Ca2+ in the extracel- lular solution and at body temperature. Note that IHCs from control mice (A) fire spontaneous action potentials (40 s out of 142 s recording time), while those from overexpressing Kir2.1 IHCs (B) require a substantial current injection to elicit any spikes. For voltage-clamp data see also Fig EV1C–I. Data in Panel A,B, and Fig EV1C–I were obtained from 8 control IHCs (7 mice) and 11 Kir2.1-OE IHCs (6 mice). C, D Calcium action potentials in IHCs from control (C) and Kir2.1-OE (D) mice during the second postnatal week. IHC voltage responses were recorded during the appli- cation of a solution containing 1.3 mM Ca2+ (top panels) or 0.3 mM Ca2+ (bottom panels). The latter Ca2+ concentration (0.3 mM), which was used to mimic the estimated in vivo Ca2+ concentration in the endolymphatic compartment (Johnson et al, 2012), caused control IHCs, but not those from Kir2.1-OE mice, to elicit spontaneous action potentials (40 s out of 56 s recording time). Diagram showing a cross-section of an immature organ of Corti. IHCs: inner hair cells; GER: greater epithelial ridge, which includes nonsensory cells surrounding the IHCs. Red arrows indicate the propagation of ATP-induced Ca2+ waves from the GER towards the IHCs, which leads to their depolarization (Tritsch et al, 2007; Wang et al, 2015; Johnson et al, 2017). E F, G Representative DF/F0 traces from the IHCs and GER of P7-P9 control (F) and Kir2.1-OE (G) mice in the presence of 0.3 mM Ca2+. Spontaneous ATP-dependent Ca2+ waves from the GER (green traces) were eliciting coordinated Ca2+ signals in the IHCs from both controls and Kir2.1-OE mice. For each genotype, two separate sets of recordings from 2 mice are shown (top and bottom right), with the top traces being linked to the images on the left: before [1], during [2] and after [3]) the gen- eration of a large Ca2+ wave from the GER. For details about the frequency and duration of the Ca2+ waves, and the number of mice and recordings see Fig EV3. All recordings were obtained at body temperature. Traces are computed as pixel averages of regions of interest centred on IHCs. Source data are available online for this figure. 4 of 20 The EMBO Journal 42: e112118 | 2023 (cid:1) 2023 The Authors Adam J Carlton et al The EMBO Journal Figure 3. Mechanoelectrical transduction in Kir2.1 overexpressing IHCs is normal during the first postnatal week. A, B Saturating MET currents recorded from apical IHCs of P6 control (A) and Kir2.1-OE (B) mice in response to 50 Hz sinusoidal force stimuli to the hair bundles at C D E membrane potentials of (cid:1)124 and +96 mV. Driver voltage (DV) stimuli to the fluid jet are shown above the traces, with positive deflections of the DV being excita- tory. The arrows indicate the closure of the transducer channel in response to inhibitory bundle stimuli at (cid:1)124 and +96 mV. Peak-to-peak MET current–voltage curves from P6-P7 apical-coil IHCs of 7 control (12 IHCs) and 3 littermate Kir2.1-OE mice (7 IHCs). Recordings were obtained by mechanically stimulating the hair bundles of IHCs at the same time as stepping their membrane potential from (cid:1)124 mV to +96 mV in 20 mV increments. The two sets of data are not significantly different: P = 0.6320, 2-way ANOVA. Maximum size of the MET current recorded at (cid:1)124 mV (left panel) and +96 mV (right panel) in IHCs from both genotypes. Resting open probability (Popen) of the MET current in IHCs from the two genotypes measured at (cid:1)124 mV (left) and +96 mV (right). The resting open probability was calculated by dividing the resting MET current (the difference between the current level before the stimulus, indicated by the dashed line, and the current level at the negative phase of the stimulus when all channels are closed) by the maximum peak-to-peak MET current. Data information: All comparisons in panels D and E are not significantly different between the two genotypes (D: (cid:1)124 mV: P = 0.2269; +96 mV: P = 0.3620; E: (cid:1)124 mV: P = 0.3766; +96 mV: P = 0.2846, t-test). In panels C-E, data are shown as means (cid:3) SD, and the single cell value recordings (open symbols) are plotted behind the average data. The number of IHCs investigated is shown above the averaged data points from 7 control and 3 Kir2.1-OE mice. Source data are available online for this figure. the MET current was significantly reduced ((cid:1)124 mV: P < 0.0001; +96 mV: P = 0.0003, Fig 4D) and the resting open probability increased ((cid:1)124 mV: P = 0.0080; +96 mV: P = 0.0130, Fig 4E) in IHCs from Kir2.1-OE mice compared with controls. Since an increased resting open probability of the MET channel could be associated with changes, specifically a reduction, in the free Ca2+ inside the stereocilia, we tested this possibility by changing the intracellular Ca2+ buffering capacity by using different concentra- tions of the fast Ca2+ chelator BAPTA. Increasing the intracellular BAPTA from 0.1 to 5 mM significantly augmented the resting open probability of in IHCs from both genotypes, although at both BAPTA concentrations it was significantly higher in the IHCs of Kir2.1-OE mice (Fig EV4). This indicates that in the absence of spontaneous intrinsic firing activity in the IHCs of Kir2.1- OE mice, the MET channels are likely to have a reduced Ca2+ sensi- tivity during the second postnatal week. By P10-P11, we found that the MET current in the IHCs of Kir2.1-OE mice was very small or the MET channel absent (Fig 4F–I). At this stage, IHCs from P10-P11 Kir2.1-OE mice also failed to load with the styryl dye FM1-43 (Fig 4J), which is a permeant blocker of the hair cell MET channel and functions as an optical readout for the presence of the resting MET current (Gale et al, 2001). IHCs from Kir2.1-OE mice undergo progressive loss and fusion of the stereocilia We investigated whether the rapid reduction in the MET current was caused by defects in the growth and/or maintenance of the stereociliary bundles in IHCs. Using scanning electron microscopy we found that the hair bundles of the IHCs from Kir2.1-OE mice were able to develop a staircase structure composed of rows of stereocilia that were indistinguishable from those present in control cells (arrows: Fig 5A and B). This is consistent with the presence of a normal MET current at least up to the end of the first postnatal (cid:1) 2023 The Authors The EMBO Journal 42: e112118 | 2023 5 of 20 The EMBO Journal Adam J Carlton et al week (Fig 3). However, from about P9 onwards, IHCs from Kir2.1- OE mice started to lose the shorter third row of stereocilia (Fig 5B). A few IHCs also started to exhibit stereocilia fusion, which became more pronounced at older ages. By P26, none of the IHCs in the Kir2.1-OE mice showed normal-looking bundles, which instead exhibited profound stereocilia fusion (Fig 5C and D). Kir2.1+/(cid:1) mice that were not crossed with OtofrtTA+/(cid:1) mice showed normal hair- bundle development when treated with DOX, highlighting the speci- ficity of the Kir2.1-OE strategy (Fig EV5). These data indicate that spontaneous Ca2+ actions potential activity during the second Figure 4. 6 of 20 The EMBO Journal 42: e112118 | 2023 (cid:1) 2023 The Authors Adam J Carlton et al The EMBO Journal ◀ Figure 4. Rapid disappearance of the MET current in Kir2.1 overexpressing IHCs during the second postnatal week. A, B Saturating MET currents recorded from apical IHCs of P8 control (A) and Kir2.1-OE (B) mice. IHC hair bundles were stimulated as described in Fig 3. C Peak-to-peak MET current–voltage curves from P8-P9 apical-coil IHCs of 12 control (17 IHCs) and 13 littermates Kir2.1-OE mice (24 IHCs). The two sets of data are significantly different: P < 0.0001, 2-way ANOVA. The maximum size of the MET current measured in IHCs at (cid:1)124 mV (left panel) and +96 mV (right panel) from Kir2.1-OE mice was significantly reduced compared to that of control cells. The resting open probability (Popen) of the MET current in IHCs was significantly increased in Kir2.1-OE compared with control cells at both (cid:1)124 mV (left) and +96 mV (right). D E F, G Saturating MET currents recorded from apical IHCs of control (F, P10) and Kir2.1-OE (G, P11) mice. IHC hair bundles were stimulated as described in Fig 3. H Peak-to-peak MET current–voltage curves from P10-P11 apical-coil IHCs of 13 control (16 IHCs) and 4 littermate Kir2.1-OE mice (9 IHCs). The two sets of data are significantly different: P < 0.0001, 2-way ANOVA. The maximum size of the MET current measured in IHCs at (cid:1)124 mV (left panel) and +96 mV (right panel) from Kir2.1-OE mice is significantly reduced compared to that of control cells. Example of FM1-43 uptake by IHCs from P11 control (top) and Kir2.1-OE (bottom) mice, showing the lack of fluorescence labeling in the latter, which is an indica- tion of the lack of MET channels open at rest at this stage in the IHCs overexpressing the Kir2.1 channels. At least 3 mice for each genotype were used. I J Data information: In panels D, E, I, data are shown as means (cid:3) SD, and the single cell value recordings (open symbols) are plotted with the average data. The number of IHCs investigated is shown above the average data points from 12 control and 13 littermates Kir2.1-OE mice (panels D and E) and 13 control and 4 littermate Kir2.1-OE mice (panel I). Statistical tests in panels D, E, I was done using the t-test. The * defines the presence of statistical significance, with the P-value shown above the data. Source data are available online for this figure. postnatal week is required for the maintenance of the stereociliary bundles in the mature IHCs. Localization of bundle proteins is not affected in Kir2.1-OE mice To establish whether the progressive loss and fusion of the stere- ocilia were linked to the mislocalization of some of the key proteins expressed in the hair bundles, we performed immunostaining exper- iments on both genotypes. Stereocilia fusion has previously been documented in hair cells from mice lacking Myo6, the gene encoding for the (F-actin) minus end-directed unconventional myosin 6 (Self et al, 1999). We found that MYOSIN VI was expressed in the stere- ocilia of the IHCs from both control and littermate Kir2.1-OE mice (Fig 6A and C). The disorganized hair bundle of the IHCs from Kir2.1-OE mice also showed a normal distribution at the tip of the taller rows of stereocilia of EPS8, MYOSIN XV-isoform 1 and WHIRLIN (Fig 6B and D); key proteins required for growth and maintenance of stereocilia (Belyantseva et al, 2005; Delprat et al, 2005; Manor et al, 2011; Zampini et al, 2011). IHC action potentials exert their developmental role in stereocilia maintenance during a critical period Next, we tested whether Ca2+ action potential activity in IHCs was regulating hair-bundle maintenance during a specific time window or “critical period” of prehearing development. This was achieved by downregulating Kir2.1-OE in vivo by removing DOX from the drinking water at a specific developmental time point. Considering that the hair bundles of the IHCs in Kir2.1-OE mice were able to acquire a staircase structure and have normal MET current at the end of the first postnatal week, we sought to test whether the role of the Ca2+ -action potentials was restricted to the second week, just before hearing onset at ~P12. As for the above investigation, DOX was continuously supplied to the females from the time of conception, but for this set of experi- ments, it was then removed from the drinking water when the pups were P5. We found that 2–3 days without DOX was sufficient to strongly downregulate Kir2.1 from the membrane of the IHCs of Kir2.1-OE mice the (Fig 7A and B). This indicates that overexpression of Kir2.1 in IHCs was primarily occurring during the first postnatal week under these conditions. We then investigated whether the downregulation of Kir2.1 channels following DOX removal (Appendix Fig S2) re-established the ability of IHCs to fire intrinsic spontaneous action potentials. In 1.3 mM extracellular Ca2+ , action potentials only occurred during depolarizing current injections in the IHCs from both control and Kir2.1-OE mice in the second postnatal week (Fig 7C and D; see also Fig 2C and D). In the the in vivo endolymph-like Ca2+ presence of concentration (0.3 mM), spontaneous intrinsic firing was present not only in the IHCs of control mice (Fig 7E; see also Fig 2C) but also in Kir2.1-OE mice (Fig 7F). For long-lasting current clamp recordings, spikes occurred in a bursting pattern in both controls and Kir2.1-OE mice. The mean spike frequency (1.17 (cid:3) 0.47 Hz, n = 4) and CV (1.12 (cid:3) 0.17, n = 4 IHCs, duration of the recordings 62–125 s) in control mice were not significantly different from those measured in Kir2.1-OE mice (frequency: 1.29 (cid:3) 0.83 Hz, n = 5 IHCs, P = 0.7766; CV: 1.15 (cid:3) 0.11, n = 5, duration of the recordings 32–101 s, P = 0.7954). The IHC resting membrane potential was not signifi- cantly different between control and Kir2.1-OE mice in the presence of both 1.3 mM and 0.3 mM Ca2+ (Fig 7G). These data indicate that the removal of DOX was effective in downregulating Kir2.1 channels and restoring the normal physiology of the IHCs. We also found that when DOX was removed at P5, IHCs were able to maintain their hair-bundle structure after the onset of hearing (Fig 7H–K). These findings indicate that Ca2+ regulation via action potentials in IHCs is required for the final maturation and maintenance of the hair bundles after a critical point just before hearing onset. Identification of genes regulated by the intrinsic Ca2+ action potentials using RNA-sequencing To understand the molecular pathways underpinning the changes in the hair-bundle structure observed in the absence of the intrinsic Ca2+ action potential activity in IHCs, we performed RNA-seq on P9 controls and littermate Kir2.1-OE mice. At this age, most of the hair bundles still showed a normal-looking structure, but with some IHCs having lost the 3rd row of stereocilia and some showing stere- ocilia fusion (Fig 5B). This was associated with the onset of MET (cid:1) 2023 The Authors The EMBO Journal 42: e112118 | 2023 7 of 20 The EMBO Journal Adam J Carlton et al IHC bundle morphology progressively deteriorates in Kir2.1 overexpressing mice. Figure 5. A, B Scanning electron microscope (SEM) images showing the IHC hair-bundle structure in the apical coil of the cochlea of P11 control (A) and P8-P11 Kir2.1-OE (B) mice. Control IHCs and the large majority of P8 IHCs from Kir2.1-OE mice show a normal hair-bundle structure composed of three rows of stereocilia: tall, interme- diate and short (arrows). From about P9 in Kir2.1-OE mice, some IHCs start to lose the third row of stereocilia (arrowheads) and some already exhibit some fusion of the stereocilia (asterisk). These changes in hair-bundle structure became more prominent at P11. At least 3 mice for each genotype were used. In these panels and those below, asterisks are used to define some of the abnormal hair bundles. C, D SEM images of both IHCs and OHCs from the cochlea of P26 control (C, upper panel) and P26 Kir2.1-OE (D, upper panel) mice. Lower panels show a higher- magnification view of the hair bundle of IHCs from both genotypes, highlighting the profound disruption of the stereocilia in IHCs overexpressing Kir2.1 channels. At least 3 mice for each genotype were used. Source data are available online for this figure. 8 of 20 The EMBO Journal 42: e112118 | 2023 (cid:1) 2023 The Authors Adam J Carlton et al The EMBO Journal Figure 6. Hair-bundle proteins involved in stereociliary elongation are not affected in IHCs from Kir2.1-OE mice. A, B Maximum intensity projections of confocal z-stacks showing images of the hair bundles from apical-coil IHCs of P6 and P11 control (A) and Kir2.1-OE (B) mice immunostained with antibodies against MYOSIN VI (blue) and EPS8 (magenta). At least 3 mice for each genotype were used. C, D Confocal images of the hair bundles of P11 IHCs from control (C) and Kir2.1-OE (D) mice immunostained with antibodies against MYOSIN XV-isoform 1 (blue) and WHIRLIN (magenta). In all panels (A–D), stereocilia are labeled with phalloidin (green). Note that despite the disrupted hair-bundle structure in the IHCs overex- pressing Kir2.1 channels; the stereocilia retained a normal distribution of these bundle proteins. At least 3 mice for each genotype were used. Source data are available online for this figure. (cid:1) 2023 The Authors The EMBO Journal 42: e112118 | 2023 9 of 20 The EMBO Journal Adam J Carlton et al current reduction in at least some of the IHCs (Fig 4A–E). We rea- soned that by profiling animals at this age we could understand the early molecular response that leads to abnormal hair-bundle mor- phology. RNA-sequencing was performed on three replicates, each with eight pooled organs of Corti from four mice. Total RNA was extracted and sent for library preparation and sequencing. Sequence data were mapped to the mouse genome (mm10) using the Figure 7. 10 of 20 The EMBO Journal 42: e112118 | 2023 (cid:1) 2023 The Authors Adam J Carlton et al The EMBO Journal ◀ Figure 7. Ca2+ spikes in IHCs regulate bundle morphology over a critical period during the second postnatal week of development. A, B Maximum intensity projections of confocal z-stacks showing the IHCs of the apical cochlear region from control (A) and Kir2.1-OE (B) pups with the females being in the continuous presence of DOX in the drinking water from conception up to when the pups were P5 (upper panels). Middle and bottom panels show IHCs at P7 and P14 following the removal of DOX at P5 for both control (A) and Kir2.1-OE mice (B). IHCs were stained with antibodies against the K+ channel Kir2.1 (green) and Myosin 7a (Myo7a, blue: cell marker). Note that after 2 days following the removal of DOX, Kir2.1 overexpression was already largely downregulated. At least 3 mice for each genotype were used. E, F C, D Calcium action potentials in IHCs from control (C) and Kir2.1-OE (D) mice during the second postnatal week (P8–P9). IHC voltage responses were recorded during the application of 1.3 mM Ca2+ extracellular solution. The voltage-clamp data recorded from the same two IHCs displayed in panels C and D are shown in Appendix Fig S2; the IHCs from Kir2.1-OE mice show a strongly reduced Kir2.1 current. DOX was removed from the drinking water at P5. Spontaneous Ca2+ action potentials in IHCs from control (E, 60 s out of 92 s recording time) and Kir2.1-OE (F, 60 s out of 103 s recording time) mice during the sec- ond postnatal week in the presence of the in vivo endolymph-like 0.3 mM Ca2+. Note that in contrast to when DOX was present throughout development (Fig 2A– D), the removal of DOX at P5 restored the ability of IHCs from Kir2.1-OE mice to generate spontaneous intrinsic Ca2+ action potentials. IHC resting membrane potentials from 2 control (4 IHCs) and 3 Kir2.1-OE (7 IHCs) mice in the presence of 1.3 mM Ca2+ (left) or 0.3 mM Ca2+ (right) in the extracel- lular solution. One-way ANOVA followed by the Bonferroni’s post-test: ns, P > 0.9990 (1.3 mM Ca2+); ns, P = 0.8864 (0.3 mM Ca2+); all other comparisons were *P < 0.0001. G H, I Maximum intensity projections of confocal z-stacks showing images of the hair bundles from apical-coil IHCs of P14 control (H) and Kir2.1-OE (I) mice stained with J, K phalloidin. DOX was removed from the mother’s drinking water when the pups were P5. At least 3 mice for each genotype were used. SEM images showing the normal structure of the hair bundles of the IHCs in the apical coil of the cochlea of P14 control (J) and P14 Kir2.1-OE (K) mice. DOX was removed from the mother’s drinking water when the pups were P5. Note that the morphological profile of the hair bundles in IHCs is comparable between control and Kir2.1-OE mice, indicating that the removal of the intrinsic Ca2+ action potentials prior to the second postnatal week has no effect on the mechanoelectrical transduction apparatus. At least 3 mice for each genotype were used. Source data are available online for this figure. NextFlow RNA pipeline and gene counts were performed using Sal- mon (see Materials and Methods). These raw counts were then used as the input for differential gene expression analysis using DeSEQ2 (Love et al, 2014). After performing principal component analysis (PCA) on the top 1,000 expressed genes in the samples, we observed a clear separation between the different genotypes with PC1, which separated Kir2.1-OE and control mice, explaining 85% of the observed variance. Conversely, PC2, which mostly separated the dif- ferent biological replicates, explained 8% of variance between sam- ples (Fig 8A). As expected, we observed a 13-fold increase in Kcnj2 (Fig 8B), validating the overexpression of the Kir2.1. We next performed differential gene expression analysis (Padj < 0.05 and fold-changes >1.5), yielding 589 upregulated genes and 30 downregulated genes (Dataset EV1; Appendix Fig S3). Pathway analysis showed an enrichment of GO terms related to cell morpho- genesis, actin filament-based processes and Rho-GTPase signaling (Fig 8C). Among the differentially expressed genes were 118 upregu- lated genes with annotations related to actin filament or microtubule regulation (Fig 8D). We also noted several genes related to the Golgi body and the trans-Golgi network (TGN), for example, Golga3, Gol- ga4, and Trip11, which are all hypothesized to play a role in main- taining Golgi structure. In line with the stereocilia phenotype, we observed the upregulation of some components of the stereocilia, Myo7a and Pcdh15 (2.25 and 2.15-fold, respectively) (Fig 8E). We also performed network analysis on known protein–protein interactions on the differentially expressed genes (Fig 8G, Dataset EV2). Chromatin remodeling genes were also overrepre- sented among the upregulated genes, including DNA methylation (Dnmt1, Dnmt3a) and demethylation (Tet1, Tet2, Tet3) and histone modifying enzymes (Hdac4, Setd2, Setd5). Several components of the LINC complex that connects the nuclear lamina to the cytoskele- tal network, including the subunits of the laminin complex (Lama 1, 2, 4, 5, Lamb1,2, Lamc1 and 2) and the Nesperin family (Syne1, Syne2, Syne3) that connect the cytoskeletal network to laminin, were upregulated (Fig 8G, Dataset EV2). Mechanical signals are directly transduced from extracellular stimulus to the nuclear inte- rior through the interaction of the nesperin proteins (Khilan et al, structure and 2021). Moreover, the maintenance of nuclear Figure 8. RNA-sequencing reveals upregulation of microtubule and cytoskeletal genes in Kir2.1-OE mice. A PCA plot of each RNA library. Each point represents one pool of 4 mice (8 cochleae) for both control and littermates Kir2.1-OE mice. Principal components were calcu- lated using the top 1,000 genes after rlog transformation, using DESeq2. The percent variance for each principal component is noted on the axis. B Transcripts per million (TPM) counts of Kcnj2 for each RNA-seq library. Counts were performed using Salmon, and length normalized. Each point represents a single library. Bars represent the mean (cid:3) SD. C Top GO terms associated with the significantly (Padj. < 0.05) upregulated genes in the Kir2.1-OE mouse. D Heatmap of the counts for each RNA library for the 118 genes associated with microtubule or cytoskeletal processes. Each row is z-scored. 118 genes were all differ- entially expressed as per the previous analysis. E TPM counts of Pcdh15, Myo7a, Macf1, Ank2, Myh9, Map1a, Map1b, Cep290 and Sptbn1 for each RNA-seq library. Counts were performed using Salmon and normalized to the length of the gene sequence. Each point represents a single library. Bars represent the mean (cid:3) SD. The numbers above the data represent the multiple hypotheses corrected adjusted P-values derived from DESeq2 using the Wald test. F Results of the top 15 motifs that were overrepresented in the TSS of the upregulated genes in Kir2.1. Enrichment analysis was performed using HOMER, in the (cid:3) 2,000 bp region around the transcriptional start site of upregulated genes. G A network rendering of the top GO processes associated with the upregulated gene set. Networks were seeded with the upregulated genes and their nearest interaction and visualized using Cytoscape. Source data are available online for this figure. ▸ (cid:1) 2023 The Authors The EMBO Journal 42: e112118 | 2023 11 of 20 The EMBO Journal Adam J Carlton et al Figure 8. 12 of 20 The EMBO Journal 42: e112118 | 2023 (cid:1) 2023 The Authors Adam J Carlton et al The EMBO Journal organization is regulated by laminins, which also help to transduce mechanical strain forces into a transcriptional response. We next sought to determine which transcription factors (TFs) might be mediating the upregulated genes in Kir2.1 overexpression. Using the list of 589 upregulated genes, we used HOMER (Heinz et al, 2010) to scan the region (cid:3) 2,000 bp from the transcriptional start site (TSS) of each gene for TF binding motifs. Within the top fifteen enriched motifs were several hair cell-enriched TFs, such as Isl1, Sox9, and Gfi1 (Fig 8F). Several classic targets of SOX9, includ- ing Acan, Col2a1, Col4a1, Col5a1, Col11a1/2, Col23a1, and several ankyrin family proteins (genes: Ank1, Ank2, Ankrd11, Ankrd12) and Myo9b, were found to be SOX9 target genes in chromatin immunoprecipitation with sequencing (ChIP-seq) studies conducted in rib chondrocytes (Ohba et al, 2015). Of the 602 upregulated genes, 65% overlapped with SOX9 target genes in chondrocytes. Similarly, SOX9 ChIP-seq in the developing testis found SOX9 bound on Myo7a (Li et al, 2014). We also observed enrichment for the RFX family, which plays a conserved role in ciliogenesis in many differ- ent organisms (Lemeille et al, 2020). Discussion Here we show that spontaneous intrinsic Ca2+ action potential activ- ity present in the developing IHCs, and thus Ca2+ regulation, is cru- cial for the final stages of maturation and maintenance of the stereociliary hair bundles. The absence of the intrinsic action poten- tials during the second postnatal week led to a progressive re- absorption of stereocilia in the short 3rd row and a fusion of the tallest rows, generating “giant” stereocilia. The functional conse- quence of this hair-bundle disruption was a complete loss of mecha- noelectrical transduction prior to the onset of hearing at P12. Furthermore, we show that this intrinsic regulation of IHC develop- ment occurs during a critical time window that spans the second postnatal week of development, just before hearing onset. The RNA- sequencing analysis highlighted that absence of intrinsic APs caused the upregulation of genes involved in cytoskeleton and Rho-GTPase- related pathways, several of which have not been previously associ- ated with cochlear development. Calcium-dependent activity in the developing cochlea The initial morphological and functional differentiation of cochlear sensory hair cells depends on intrinsic genetic programs that are coordinated by a combination of transcription factors, including Ato- h1 (Bermingham et al, 1999), Helios (Chessum et al, 2018) and Tbx2 (Garc(cid:1)ıa-A~noveros et al, 2022), and microRNAs such as miR-96 (Kuhn et al, 2011). However, evidence from other sensory systems, especially from the visual system (e.g., Grubb & Thompson, 2004; Blankenship & Feller, 2010), shows that the final maturation of sen- sory pathways is driven by experience-independent Ca2+ -dependent activity, which occurs during a critical period of development. This early electrical activity has been shown to regulate several cellular responses (Berridge et al, 2000), including the remodeling of synap- tic connections (Zhang & Poo, 2001) and ion-channel expression (Moody & Bosma, 2005). In the mammalian cochlea, spontaneous Ca2+ throughout recorded have been potentials -dependent action postnatal the (cid:1) + channels and the efflux of K development of the IHCs (Kros et al, 1998; Glowatzki & Fuchs, 2000; Beutner & Moser, 2001; Marcotti et al, 2003a; Brandt et al, 2007). The firing activity of neighboring IHCs is normally synchro- nized by spontaneous intercellular Ca2+ signaling originating in the nonsensory cells via the release of ATP (Tritsch et al, 2007; Johnson et al, 2011, 2017; Wang et al, 2015; Eckrich et al, 2018). ATP acts on purinergic autoreceptors expressed in the nonsensory cells sur- rounding the IHCs, which leads to the opening of TMEM16A Ca2+ - activated Cl in the intercellular space, causing IHC depolarization (Wang et al, 2015). Although Ca2+ action potential activity in developing IHCs has been linked to the refine- ment of the tonotopic organization in the brainstem (Clause et al, 2014, 2017; M€uller et al, 2019; Maul et al, 2022) and auditory neu- ron survival (Zhang-Hooks et al, 2016), its direct role in regulating and/or maintaining IHC development is still largely unknown. A previous study has shown that increasing the IHC firing activity pre- vented linearization of their exocytotic Ca2+ dependence in the adult cochlea (Johnson et al, 2013), although both the intrinsic and ATP- dependent mechanisms could have contributed. The mouse model used in this study (Kir2.1-OE) has allowed us to specifically silence the intrinsic Ca2+ action potentials in develop- ing IHCs in vivo while retaining the ability of nonsensory cells to depolarize the IHCs via ATP-dependent Ca2+ signaling. We found that the absence of spontaneous Ca2+ action potentials that are intrinsically generated by the IHCs prevented the full maturation and maintenance of the hair bundles in IHCs, thus abolishing the mechanoelectrical transducer current that is required for the conver- sion of acoustic stimuli into electrical signals. We found that such crucial control over the hair-bundle structure and function is only established after a critical time point in the second postnatal week, just before hearing onset. Role of Ca2+-dependent action potentials in the maturation of hair cells Calcium-dependent electrical activity regulates several cellular responses (Berridge et al, 2000). Changes in intracellular Ca2+ sig- nals mediated by L-type Ca2+ channels have been implicated in regu- lating gene expression in many intracellular pathways including those associated with remodeling and development (Bading et al, 1993; Dolmetsch et al, 1997; Fields et al, 2005; Hagenston & Bading, 2011). Here we show using RNA-seq analysis that the dysregulation of Ca2+ in prehearing IHCs, which only retain the extrinsic modula- tion of the Ca2+ signals from the nonsensory cells, led to 589 upregu- lated and 30 downregulated genes. One of the characteristic phenotypes of the Kir2.1-OE mouse cochlea was the formation of giant or fused stereocilia, which was previously reported in knockout mice for the protein TRIOBP (Kita- jiri et al, 2010), which is a component of the stereocilia rootles (Pacentine et al, 2020), for the unconventional MYO6 (Self et al, 1999) that localizes at the base and all along the length of the stere- ocilia (Hertzano et al, 2008) and for a protein associated with the shaft connectors located between stereocilia (PTPRQ, Goodyear et al, 2003). RNA-seq analysis did not show any significant changes in the genes encoding the above three proteins in the cochlea of Kir2.1-OE mice but did identify 118 upregulated genes with annota- tions related to actin filament or microtubule regulation. This included cytoskeletal genes Mapt, Sptb, Plec, and Nefh, Kinesin (cid:1) 2023 The Authors The EMBO Journal 42: e112118 | 2023 13 of 20 The EMBO Journal Adam J Carlton et al superfamily proteins, which are microtubule-dependent molecular motors (Kif1a, Kif5a, Kif5c, Kif21a, Kif21b), and several components of the Rho-GTPase pathway (Rock1, Iqgap1, Iqgap2, Itpr1, Itpr2, Itpr3, Arhgap13, Argef11, Argef17, Trio, and Kalrn). Although most of the identified genes are possible novel candidates involved in hair cell development, we found some that have previously been associ- ated with hair-bundle morphology. For example, the actin-binding protein spectrin isoform SPTBN1 (Sptbn1), which is expressed in the rootlets actin filaments of the stereocilia (Furness et al, 2008), and together with TRIOBP contributes to strengthen their insertion point into the apical membrane of the hair cells (Pacentine et al, 2020), is required for the correct hair-bundle morphology. In the absence of SPTBN1 mice are deaf (Liu et al, 2019). Furthermore, the actin crosslinking family protein 7 (ACF7) and the microtubule- associated protein 1A (MAP-1A), which are encoded by the genes Macf1 and Macp1a, respectively, are also involved in the organiza- tion of the cuticular plate of the hair cells (Jaeger et al, 1994; Antonellis et al, 2014), which is the point of stereocilia insertion of the hair bundles. Finally, the non-muscle myosin Type IIA (MYH9) has been linked to both syndromic and nonsyndromic hearing loss due to the disruption of the hair cell stereociliary bundles (Mhatre et al, 2006). Among the identified transcription factors, we found enrichment for regulatory factor X (RFX), which plays a role in Materials and Methods Reagents and Tools table Reagent/Resource Experimental Models OtofrtTA (M. musculus) TetO-Kir2.1-IRES-tauLacZ Recombinant DNA Example: pCMV-BE3 Antibodies Mouse-IgG1 anti-Eps8 Rabbit-IgG anti-Whirlin Rabbit-IgG anti-Myo6 Rabbit-IgG anti-Myo15 isoform 1 Mouse-IgG1 anti-Kir2.1 Mouse-IgG1 anti-CtBP2 Mouse-IgG2a anti-GluR2 Chemicals, enzymes, and other reagents Doxycycline Vitamins Amino acids DMEM/F12 Fluo-4 AM FM1-43 Texas Red-X phalloidin ciliogenesis in many different organisms (Lemeille et al, 2020). In mammals, Rfx3 is involved in ciliary assembly and motility, and Rfx4 is known to modulate Shh signaling and regional control of cili- ogenesis (Ashique et al, 2009). Moreover, recent work has shown that the RFX family is essential for hearing in mice, with mice at 3 months of age showing a loss of stereocilia structure (Elkon et al, 2015). Altogether, indicate that these results the intrinsic Ca2+ - dependent action potential activity in IHCs during the second post- natal week is necessary to drive their full morphological and func- tional maturation into auditory sensory receptors. The absence of such activity led to the upregulation of the genetic pathways involved in the maintenance of cytoskeletal homeostasis, possibly as an attempt to repair or compensate for the progressive deteriora- tion of the actin-based hair bundles. Moreover, we found that the MET channel of the IHCs from Kir2.1-OE mice acquire a reduced Ca2+ sensitivity, which could be a potential compensatory mecha- nism for maintaining resting MET current size as the MET current is rapidly declining. Although genetic compensation responses follow- ing the mutation of genes have been described in many organisms including zebrafish and mice (e.g., El-Brolosy & Stainier, 2017; Buglo et al, 2020), their underlying mechanisms remain largely unknown. Reference or Source Identifier or Catalog Number Ozgene Pty Ltd The Jackson Lab Addgene BD Biosciences gift from Dr. Thomas Friedman Proteus Biosciences gift from Dr. Thomas Friedman Alomone Lab BD Biosciences Millipore Karidox 100 mg/ml Thermo Fisher Thermo Fisher Sigma Thermo Fisher Molecular Probes ThermoFisher n/a 009136 Cat #73021 610143 n/a 25-6791 n/a APC026 612044 MAB397 11120-037 11130-036 D8062 F14201 T3163 T7471 14 of 20 The EMBO Journal 42: e112118 | 2023 (cid:1) 2023 The Authors Adam J Carlton et al The EMBO Journal Reagents and Tools table (continued) Reagent/Resource Reference or Source Identifier or Catalog Number Software pClamp 10 Origin Python 2.7 ImageJ DeSeq2 Metascape Reactome HOMER Other Digidata 1440A Two-photon laser-scanning microscope (Bergamo II system B232) Vega3 LMU scanning electron microscope LSM 880 AiryScan microscope RNeasy Plus Micro Kit NovaSeq sequencer RRID:SCR_011323 RRID:SCR_014212 RRID:SCR_008394 RRID:SCR_003070 RRID:SCR_021038 Molecular Devices OriginLab Python Software Foundation NIH Love et al, 2014 Zhou et al, 2019 Gillespie et al, 2022 Heinz et al, 2010 Molecular Devices Thorlabs Inc Tescan Zeiss Qiagen Illumina Methods and Protocols Ethics statement Animal work was licensed by the Home Office under the Animals (Scientific Procedures) Act 1986 (PPL_PCC8E5E93) and was approved by the University of Sheffield Ethical Review Committee (180626_Mar). For ex vivo experiments mice were killed by cervical dislocation followed by decapitation. Mice had free access to food and water and a 12 h light/dark cycle. Transgenic mice The transgenic mouse line OtofrtTA expressing rtTA driven by the Otof promoter was constructed by Ozgene Pty Ltd (Bentley WA, Australia). In these mice, the expression of a target gene is controlled by a reverse tetracycline-controlled transactivator rtTA (Tet-On system: Baron & Bujard, 2000). Otof encodes for the ribbon synaptic Ca2+ sensor otofer- lin, which in the cochlea is expressed exclusively in hair cells but pri- marily in IHCs from around birth (Roux et al, 2006). Homozygous OtofrtTA mice were paired with heterozygous teto-Kir2.1-IRES-tau-lacZ mice (Kir2.1: Jackson laboratories, 009136, Yu et al, 2004). Both mouse lines (OtofrtTA and Kir2.1) were maintained on the C57BL/6N background. The resultant compound heterozygous mice, which we named Kir2.1-OE (Kir2.1-OverExpression) mice for simplicity, allowed + cell-specific overexpression of the inward rectifier K channel Kir2.1 in the IHCs when mice were treated with doxycycline (DOX). Litter- mate heterozygous OtofrtTA mice treated with DOX were used as con- trols. Pregnant, breast-feeding females and weaned pups (controls and Kir2.1-OE) were given 0.5 mg/ml of DOX daily in their drinking water, a dose that was previously optimized for the mouse cochlea (Johnson et al, 2013). Tissue preparation Cochleae were dissected out from the inner ear of the mouse using an extracellular solution composed of (in mM): 135 NaCl, 5.8 KCl, 1.3 CaCl2, 0.9 MgCl2, 0.7 NaH2PO4, 5.6 D-glucose, 10 HEPES. Sodium pyruvate (2 mM), amino acids, and vitamins were added from concentrates (Thermo Fisher Scientific, UK). The pH was adjusted to 7.48 with 1 M NaOH (osmolality ~308 mOsm/kg). The dissected cochleae were transferred to a microscope chamber and immobilized via a nylon mesh attached to a stainless-steel ring as previously described (Marcotti et al, 2003b). The chamber (vol- ume ~ 2 ml) was perfused from a peristaltic pump and mounted on the stage of an upright microscope (Olympus BX51, Japan; Leica DMLFS, Germany) equipped with Nomarski Differential Interference Contrast (DIC) optics (60× or 64× water immersion objective) and 15× eyepieces. The microscope chamber was continuously perfused with the extracellular solution by a peristaltic pump (Cole-Palmer, UK). Whole-cell electrophysiology Patch clamp experiments were performed from hair cells positioned at the 9–12 kHz region of the cochlear apical coil (M€uller et al, 2005). Recordings were performed at room temperature (20–24°C) using an Optopatch amplifier (Cairn Research Ltd, UK) as previously described (Jeng et al, 2020; Carlton et al, 2021). Patch pipettes were pulled from soda glass capillaries, which had a typical resistance in the extracellular solution of 2–3 MΩ. The intracellular solution used for the patch pipette contained (in mM): 131 KCl, 3 MgCl2, 1 EGTA- KOH, 5 Na2ATP, 5 HEPES, 10 Na-phosphocreatine (pH was adjusted with 1 M KOH to 7.28; 294 mOsm/kg). Data acquisition was con- trolled by pClamp software using Digidata 1440A (Molecular Devices, USA). In order to reduce the electrode capacitance, patch electrodes were coated with surf wax (Mr Zoggs SexWax, USA). Recordings were low-pass filtered at 2.5 kHz (8-pole Bessel), sam- pled at 5 kHz, and stored on a computer for offline analysis (Clamp- fit, Molecular Devices; Origin 2021: OriginLab, USA). Membrane potentials under voltage-clamp conditions were corrected offline for the residual series resistance Rs after compensation (usually 80%) (cid:1) 2023 The Authors The EMBO Journal 42: e112118 | 2023 15 of 20 The EMBO Journal Adam J Carlton et al and the liquid junction potential (LJP) of (cid:1)4 mV, which was mea- sured between electrode and bath solutions. Voltage-clamp proto- cols are referred to a holding potential of (cid:1)84 mV unless otherwise stated. Real-time changes in membrane capacitance (DCm) were tracked at body temperature as previously described (Johnson et al, 2005, 2017). Briefly, a 4 kHz sine wave of 13 mV RMS was applied to IHCs from the holding potential of (cid:1)81 mV and was interrupted for the duration of the voltage step. The capacitance signal from the Optopatch was filtered at 250 Hz and sampled at 5 kHz. DCm was measured by averaging the Cm trace over a 200 ms period following the voltage step and subtracting the pre- pulse baseline. Data were acquired using pClamp software and a Digidata 1440A (Molecular Devices). DCm experiments were per- formed during the local perfusion of the IHCs with 30 mM TEA, + 15 mM 4-AP (Fluka) to block the outward K currents (Johnson + et al, 2005), and 5 mM CsCl to block the inward rectifier K cur- rent (Marcotti et al, 1999). For mechanoelectrical transducer (MET) current recordings, the hair bundles of hair cells were displaced using a fluid-jet system from a pipette driven by a 25 mm diameter piezoelectric disc (Corns et al, 2014, 2018; Carlton et al, 2021). For these experiments, the intracellular solution contained (in mM): 131 CsCl, 3 MgCl2, 1 EGTA-KOH, 5 Na2ATP, 5 HEPES, 10 Na-phosphocreatine (pH was adjusted with 1 M CsOH to 7.28; 290 mOsm/kg). The extracellular solution was as described above, although for most of the record- ings we included 5 mM CsCl, which was used to block the inward + current (Marcotti et al, 1999). In order to maintain the rectifier K osmolality of the extracellular solution constant, NaCl was reduced to 130 mM in this case. The fluid-jet pipette tip had a diameter of 8–10 lm and was posi- tioned near the hair bundles to elicit a maximal MET current (typi- cally 10 lm). Mechanical stimuli were applied as 50 Hz sinusoids (filtered at 1 kHz, 8-pole Bessel). Prior to the positioning of the fluid jet by the hair bundles, any steady-state pressure was removed. The use of the fluid jet allows for the efficient displacement of the hair bundles in both the excitatory and inhibitory directions, which is essential to perform reliable measurements of the resting open prob- ability of the MET channels. Two-photon confocal Ca2+ imaging Acutely dissected cochleae were incubated for 40 min at 37°C in DMEM/F12, supplemented with fluo-4 AM at a final concentration of 10 lM (Thermo Fisher Scientific) as recently described (Ceriani et al, 2019). The incubation medium contained also pluronic F-127 (0.1%, w/v) and sulfinpyrazone (250 lM) to prevent dye sequestra- tion and secretion. Calcium signals were recorded using a two- photon laser-scanning microscope (Bergamo II System B232, Thor- labs Inc., USA) based on a mode-locked laser system operating at 925 nm, 80-MHz pulse repetition rate, < 100-fs pulse width (Mai Tai HP DeepSee, Spectra-Physics, USA). Images were captured with a 60× objective (LUMFLN60XW, Olympus, Japan) using a GaAsp PMT (Hamamatsu) coupled with a 525/40 bandpass filter (FF02- 525/40-25, Semrock). Images were analyzed offline using custom- built software routines written in Python (Python 2.7, Python Soft- ware Foundation) and ImageJ (NIH). Calcium signals were measured as relative changes in fluorescence emission intensity (DF/F0). Scanning electron microscopy (SEM) The isolated inner ear was very gently perfused with fixative for 1– 2 min through the round window. A small hole in the apical portion of the cochlear bone was made prior to perfusion to allow the fixa- tive to flow out from the cochlea. The fixative contained 2.5% vol/ vol glutaraldehyde in 0.1 M sodium cacodylate buffer plus 2 mM CaCl2 (pH 7.4). The inner ears were then immersed in the above fix- ative and placed on a rotating shaker for 2 h at room temperature. After fixation, the organ of Corti was exposed by removing the bone from the apical coil of the cochlea and then immersed in 1% osmium tetroxide in 0.1 M cacodylate buffer for 1 h. For osmium impregnation, which avoids gold coating, cochleae were incubated in solutions of saturated aqueous thiocarbohydrazide (20 min) alter- nating with 1% osmium tetroxide in buffer (2 h) twice (the OTOTO technique: Furness & Hackney, 1986). The cochleae were then dehy- drated through an ethanol series and critical point dried using CO2 as the transitional fluid (Leica EM CPD300) and mounted on speci- men stubs using conductive silver paint (Agar Scientific, Stansted, UK). The apical coil of the organ of Corti was examined at 10 kV using a Tescan Vega3 LMU scanning electron microscope in the electron microscopy unit at the University of Sheffield. Immunofluorescence microscopy As for SEM, the isolated inner ear was initially gently perfused with 4% paraformaldehyde in phosphate-buffered saline (PBS, pH 7.4) through the round window. Following this initial short fixation, the inner ear was fixed for 20 min at room temperature and then washed three times in PBS for 10 min. The apical coil of the organ of Corti was then washed in PBS, removed by fine dissection, and incubated in PBS supplemented with 5% normal goat or horse serum and 0.5% Triton X-100 for 1 h at room temperature. The samples were immunolabeled with primary antibodies overnight at 37°C, washed three times with PBS, and incubated with the secondary antibodies for 1 h at 37°C. Antibodies were prepared in 1% serum and 0.5% Tri- ton X-100 in PBS. Primary antibodies were mouse-IgG1 anti-Eps8 (1:1,000, BD Biosciences, 610,143), rabbit-IgG anti-WHIRLIN (1:200, gift from Dr. Thomas Friedman, NIH, USA); rabbit-IgG anti-MYO6 (1:150, Proteus Biosciences, 25–6,791); rabbit-IgG anti-MYO15- isoform 1 (1:1,000, gift from Dr. Thomas Friedman, NIH, USA) Israel, mouse-IgG1 anti-Kir2.1 channel APC026); mouse IgG1 anti-CtBP2 (1:200, Biosciences, #612044) and mouse IgG2a anti-GluR2 (1:200, Millipore, MAB397). F-actin was stained with Texas Red-X phalloidin (1:400, ThermoFisher, T7471) in the secondary antibody solution. Secondary antibodies were species-appropriate Alexa Fluor or Northern Lights secondary anti- bodies. Samples were mounted in VECTASHIELD (H-1000). The images from the apical cochlear region (8–12 kHz) were captured with Nikon A1 confocal microscope equipped with a Nikon CFI Plan Apo 60× Oil objective or a Zeiss LSM 880 AiryScan equipped with Plan-Apochromat 63× Oil DIC M27 objective for super-resolution images of hair bundles. Both microscopes are part of the Wolfson Light Microscope Facility at the University of Sheffield. Image stacks were processed with Fiji ImageJ software. (1:100, Alomone Lab, FM1-43 staining A 3 mM stock solution of the dye FM1-43 (T3163, Molecular Probes) was prepared in water. The dissected organs of Corti (aged P11–P12) were transferred to the bottom of a chamber filled with 16 of 20 The EMBO Journal 42: e112118 | 2023 (cid:1) 2023 The Authors Adam J Carlton et al The EMBO Journal extracellular solution, and held in position using a nylon mesh, as described above (see above: Tissue preparation). All experiments were performed at room temperature (20–24°C), as previously described (Gale et al, 2001). Briefly, the solution bathing the cochleae was very rapidly exchanged with that containing 3 lM FM1-43 for 10 s and immediately washed several times with normal extracellular solution. The cochleae were then viewed with an upright microscope equipped with epifluorescence optics and FITC filters (excitation 488 nm, emission 520 nm) using a 63× water immersion objective and a CCD camera. RNA isolation and library preparation The sensory epithelium from four control and four littermates Kir2.1-OE mice under DOX were microdissected in DNase-free ice- cold PBS 1× and immediately snap frozen in liquid nitrogen. RNA was extracted using RNeasy Plus Micro Kit (Qiagen) according to the manufacturer’s instructions. RNA quantity was established using a Nanodrop spectrophotometer and RNA integrity number (RIN) was calculated using a BioAnalyzer. All samples had RIN score greater than 9.1. Preparation of the mRNA library was per- formed using poly A enrichment and sequenced on the Illumina NovaSeq sequencer using paired-end 150 bp reads. RNA-sequencing analysis and differential gene expression The sequencing libraries were processed using the nf-core RNA pipeline (Ewels et al, 2020, https://nf-co.re/rnaseq/usage) using the standard parameters. Reads were mapped to the mouse genome (mm10). The resulting gene counts were determined using Salmon (Patro et al, 2017) and used for downstream analysis with DESeq2 (Love et al, 2014). Metascape (Zhou et al, 2019) and Reactome (Gillespie et al, 2022) were used to query for enriched GO terms and pathways in the list of differentially expressed genes. HOMER (Heinz et al, 2010) was used to find known and de novo motifs among the upregulated genes in a 2000 bp window up and down- stream of the transcriptional start site (TSS). Statistical analysis Statistical comparisons of means were made by the Student’s two- tailed t-test or, for multiple comparisons, the analysis of variance (one-way or two-way ANOVA followed by a suitable post-test) and Mann–Whitney U test (when normal distribution could not be assumed) were used. P < 0.05 was selected as the criterion for statis- tical significance. Only mean values with a similar variance between groups were compared. Average values are quoted in text and figures as means (cid:3) S.D. Animals of either sex were randomly assigned to the different experimental groups. No statistical methods were used to define sample size, which was determined based on previously published similar work from our laboratory. Animals were taken from several cages and breeding pairs over a period of several months. Most of the electrophysiological and morphological (but not imaging) experiments were performed blind to animal genotyping and in most cases, experiments were replicated at least 3 times. Data availability The data that support the findings of this study are available from corresponding author. RNA-sequencing data have been the deposited in GEO under accession number (GSE215951; http:// www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE215951). Expanded View for this article is available online. Acknowledgements The authors thank Michelle Bird (University of Sheffield) for her assistance with the mouse husbandry, and Catherine Gennery and Laila Moushtaq- Kheradmandi for their genotyping work. This work was supported by the BBSRC (BB/S006257/1 and BB/T004991/1) and Wellcome Trust (224326/Z/21/Z) to WM, the MRC (MR/S002510/1) to MM. AU was supported by a PhD studentship from the MRC DiMeN Doctoral Training Partnership to WM. For the purpose of Open Access, the author has applied a CC BY public copyright license to any author accepted manuscript version arising from this submission. Author contributions Adam J Carlton: Conceptualization; data curation; formal analysis; validation; investigation; methodology; writing – original draft; writing – review and edit- ing. Jing-Yi Jeng: Data curation; formal analysis; validation; investigation; methodology; writing – review and editing. Fiorella C Grandi: Data curation; formal analysis; validation; investigation; methodology; writing – review and editing. Francesca De Faveri: Data curation; formal analysis; validation; investigation; methodology; writing – review and editing. Federico Ceriani: Data curation; formal analysis; validation; investigation; methodology; writing – review and editing. Lara De Tomasi: Investigation; methodology. Anna Underhill: Data curation; formal analysis; investigation; methodology. Stuart L Johnson: Data curation; formal analysis; validation; investigation; methodology. Kevin P Legan: Investigation. Corn(cid:1)e J Kros: Methodology; writing – review and editing. Guy P Richardson: Investigation; methodology; writing – review and editing. Mirna Mustapha: Resources; funding acquisition; methodology; writing – review and editing. Walter Marcotti: Conceptualization; resources; data curation; formal analysis; supervision; funding acquisition; validation; investigation; methodology; writing – original draft. Disclosure and competing interests statement The authors declare that they have no conflict of interest. References Antonellis PJ, Pollock LM, Chou SW, Hassan A, Geng R, Chen X, Fuchs E, Alagramam KN, Auer M, McDermott BM Jr (2014) ACF7 Is a hair-bundle antecedent, positioned to integrate cuticular plate Actin and somatic tubulin. J Neurosci 34: 305 – 312 Ashique AM, Choe Y, Karlen M, May SR, Phamluong K, Solloway MJ, Ericson J, Peterson AS (2009) The Rfx4 transcription factor modulates Shh signaling by regional control of ciliogenesis. Sci Signal 2: ra70 Bading H, Ginty DD, Greenberg ME (1993) Regulation of gene expression in hippocampal neurons by distinct calcium signaling pathways. Science 260: 181 – 186 Baron U, Bujard H (2000) Tet repressor-based system for regulated gene expression in eukaryotic cells: principles and advances. Methods Enzymol 327: 401 – 421 Belyantseva IA, Boger ET, Naz S, Frolenkov GI, Sellers JR, Ahmed ZM, Griffith AJ, Friedman TB (2005) Myosin-XVa is required for tip localization of whirlin and differential elongation of hair-cell stereocilia. Nat Cell Biol 7: 148 – 156 (cid:1) 2023 The Authors The EMBO Journal 42: e112118 | 2023 17 of 20 The EMBO Journal Adam J Carlton et al Bermingham NA, Hassan BA, Price SD, Vollrath MA, Ben-Arie N, Eatock RA, Bellen HJ, Lysakowski A, Zoghbi HY (1999) Math1: an essential gene for the generation of inner ear hair cells. Science 284: 1837 – 1841 Berridge MJ, Lipp P, Bootman MD (2000) The versatility and universality of calcium signalling. Nat Rev Mol Cell Biol 1: 11 – 21 Beutner D, Moser T (2001) The presynaptic function of mouse cochlear inner hair cells during development of hearing. J Neurosci 21: 4593 – 4599 Blankenship AG, Feller MB (2010) Mechanisms underlying spontaneous patterned activity in developing neural circuits. Nat Rev Neurosci 11: 18 – 29 Brandt N, Kuhn S, M€unkner S, Braig C, Winter H, Blin N, Vonthein R, Knipper M, Engel J (2007) Thyroid hormone deficiency affects postnatal spiking activity and expression of Ca2+ and K+ channels in rodent inner hair cells. J Neurosci 27: 3174 – 3186 Buglo E, Sarmiento E, Martuscelli NB, Sant DW, Danzi MC, Abrams AJ, Dallman JE, Z€uchner S (2020) Genetic compensation in a stable slc25a46 mutant zebrafish: a case for using F0 CRISPR mutagenesis to study phenotypes caused by inherited disease. PLoS One 15: e0230566 Carlton AJ, Halford J, Underhill A, Jeng JY, Avenarius MR, Gilbert ML, Ceriani F, Ebisine K, Brown SDM, Bowl MR et al (2021) Loss of Baiap2l2 destabilizes the transducing stereocilia of cochlear hair cells and leads to deafness. J Physiol 599: 1173 – 1198 Ceriani F, Hendry A, Jeng JY, Johnson SL, Stephani F, Olt J, Holley MC, Mammano F, Engel J, Kros CJ et al (2019) Coordinated calcium signalling in cochlear sensory and non-sensory cells refines afferent innervation of outer hair cells. EMBO J 38: e99839 Chessum L, Matern MS, Kelly MC, Johnson SL, Ogawa Y, Milon B, McMurray M, Driver EC, Parker A, Song Y et al (2018) Helios is a key transcriptional regulator of outer hair cell maturation. Nature 563: 696 – 700 Clause A, Kim G, Sonntag M, Weisz CJ, Vetter DE, R}ubsamen R, Kandler K (2014) The precise temporal pattern of prehearing spontaneous activity is necessary for tonotopic map refinement. Neuron 82: 822 – 835 Clause A, Lauer AM, Kandler K (2017) Mice lacking the Alpha9 subunit of the nicotinic acetylcholine receptor exhibit deficits in frequency difference limens and sound localization. Front Cell Neurosci 11: 167 Corns LF, Johnson SL, Kros CJ, Marcotti W (2014) Calcium entry into stereocilia drives adaptation of the mechanoelectrical transducer current of mammalian cochlear hair cells. Proc Natl Acad Sci U S A 111: 14918 – 14923 Elkon R, Milon B, Morrison L, Shah M, Vijayakumar S, Racherla M, Leitch CC, Silipino L, Hadi S, Weiss-Gayet M et al (2015) RFX transcription factors are essential for hearing in mice. Nat Commun 6: 8549 Ewels PA, Peltzer A, Fillinger S, Patel H, Alneberg J, Wilm A, Garcia MU, Di Tommaso P, Nahnsen S (2020) The nf-core framework for community- curated bioinformatics pipelines. Nat Biotechnol 38: 276 – 278 Fields RD, Lee PR, Cohen JE (2005) Temporal integration of intracellular Ca2+ signaling networks in regulating gene expression by action potentials. Cell Calcium 37: 433 – 442 Fuchs PA (2005) Time and intensity coding at the hair cell’s ribbon synapse. J Physiol 566: 7 – 12 Furness DN, Hackney CM (1986) High-resolution scanning-electron microscopy of stereocilia using the osmium-thiocarbohydrazide coating technique. Hear Res 21: 243 – 249 Furness DN, Mahendrasingam S, Ohashi M, Fettiplace R, Hackney CM (2008) The dimensions and composition of stereociliary rootlets in mammalian cochlear hair cells: comparison between high and low-frequency cells and evidence for a connection to the lateral membrane. J Neurosci 28: 6342 – 6353 Gale JE, Marcotti W, Kennedy HJ, Kros CJ, Richardson GP (2001) FM1-43 dye behaves as a permeant blocker of the hair-cell mechanotransducer channel. J Neurosci 21: 7013 – 7025 Garc(cid:1)ıa-A~noveros J, Clancy JC, Foo CZ, Garc(cid:1)ıa-G(cid:1)omez I, Zhou Y, Homma K, Cheatham MA, Duggan A (2022) Tbx2 is a master regulator of inner versus outer hair cell differentiation. Nature 605: 298 – 303 Gillespie M, Jassal B, Stephan R, Milacic M, Rothfels K, Senff-Ribeiro A, Griss J, Sevilla C, Matthews L, Gong C et al (2022) The reactome pathway knowledgebase 2022. Nucleic Acids Res 50: D687 – D692 Glowatzki E, Fuchs PA (2000) Cholinergic synaptic inhibition of inner hair cells in the neonatal mammalian cochlea. Science 288: 2366 – 2368 Goodyear RJ, Legan PK, Wright MB, MarcottiW OA, Coats SA, Booth CJ, Kros CJ, Seifert RA, Bowen-Pope DF, Richardson GP (2003) A receptor-like inositol lipid phosphatase is required for the maturation of developing cochlear hair bundles. J Neurosci 23: 9208 – 9219 Grubb MS, Thompson ID (2004) The influence of early experience on the development of sensory systems. Curr Opin Neurobiol 14: 503 – 512 Hagenston AM, Bading H (2011) Calcium signaling in synapse-to-nucleus communication. Cold Spring Harb Perspect Biol 3: a004564 Corns LF, Johnson SL, Roberts T, Ranatunga KM, Hendry A, Ceriani F, Safieddine S, Steel KP, Forge A, Petit C et al (2018) Mechanotransduction is required for establishing and maintaining mature inner hair cells and regulating efferent innervation. Nat Commun 9: 4015 Heinz S, Benner C, Spann N, Bertolino E, Lin YC, Laslo P, Cheng JX, Murre C, Singh H, Glass CK (2010) Simple combinations of lineage-determining transcription factors prime cis-regulatory elements required for macrophage and B cell identities. Mol Cell 38: 576 – 589 Delprat B, Michel V, Goodyear R, Yamasaki Y, Michalski N, El-Amraoui A, Hertzano R, Shalit E, Rzadzinska AK, Dror AA, Song L, Ron U, Tan JT, Shitrit Perfettini I, Legrain P, Richardson G, Hardelin JP et al (2005) Myosin XVa and whirlin, two deafness gene products required for hair bundle growth, are located at the stereocilia tips and interact directly. Hum Mol Genet 14: 401 – 410 Dolmetsch RE, Lewis RS, Goodnow CC, Healy JI (1997) Differential activation of transcription factors induced by Ca2+ response amplitude and duration. Nature 386: 855 – 858 Eckrich T, Blum K, Milenkovic I, Engel J (2018) Fast Ca2+ transients of inner hair cells Arise coupled and uncoupled to Ca2+ waves of inner supporting cells in the developing mouse cochlea. Front Mol Neurosci 11: 264 Ehret G (1983) Development of hearing and response behavior to sound stimuli, behavioral studies. In Development of the Auditory and Vestibular Systems, Romand R (ed), pp 211 – 237. New York, NY: Academic Press El-Brolosy MA, Stainier DYR (2017) Genetic compensation: a phenomenon in search of mechanisms. PLoS Genet 13: e1006780 AS, Fuchs H, Hasson T et al (2008) A Myo6 mutation destroys coordination between the myosin heads, revealing new functions of myosin VI in the stereocilia of mammalian inner ear hair cells. PLoS Genet 4: e1000207 Jaeger RG, Fex J, Kachar B (1994) Structural basis for mechanical transduction in the frog vestibular sensory apparatus: II. The role of microtubules in the organization of the cuticular plate. Hear Res 77: 207 – 215 Jeng JY, Ceriani F, Olt J, Brown SDM, HolleyMC BMR, Johnson SL, Marcotti W (2020) Pathophysiological changes in inner hair cell ribbon synapses in the ageing mammalian cochlea. J Physiol 598: 4339 – 4355 Johnson SL, Marcotti W, Kros CJ (2005) Increase in efficiency and reduction in Ca2+ dependence of exocytosis during development of mouse inner hair cells. J Physiol 563: 177 – 191 Johnson SL, Adelman JP, Marcotti W (2007) Genetic deletion of SK2 channels in mouse inner hair cells prevents the developmental linearization in the Ca2+ dependence of exocytosis. J Physiol 583: 631 – 646 18 of 20 The EMBO Journal 42: e112118 | 2023 (cid:1) 2023 The Authors Adam J Carlton et al The EMBO Journal Johnson SL, Eckrich T, Kuhn S, Zampini V, Franz C, Ranatunga KM, Roberts TP, Masetto S, Knipper M, Kros CJ et al (2011) Position-dependent patterning of spontaneous action potentials in immature cochlear inner hair cells. Nat Neurosci 14: 711 – 717 Johnson SL, Kennedy HJ, Holley MC, Fettiplace R, Marcotti W (2012) The resting transducer current drives spontaneous activity in prehearing mammalian cochlear inner hair cells. J Neurosci 32: 10479 – 10483 Johnson SL, Kuhn S, Franz C, Ingham N, Furness DN, Knipper M, Steel KP, Adelman JP, Holley MC, Marcotti W (2013) Presynaptic maturation in auditory hair cells requires a critical period of sensory-independent spiking activity. Proc Natl Acad Sci U S A 110: 8720 – 8725 Johnson SL, Ceriani F, Houston O, Polishchuk R, Polishchuk E, Crispino G, Zorzi V, Mammano F, Marcotti W (2017) Connexin-mediated signaling in nonsensory cells is crucial for the development of sensory inner hair cells in the mouse cochlea. J Neurosci 37: 258 – 268 Jones TA, Leake PA, Snyder RL, Stakhovskaya O, Bonham B (2007) Spontaneous discharge patterns in cochlear spiral ganglion cells before the onset of hearing in cats. J Neurophysiol 98: 1898 – 1908 Katz LC, Shatz CJ (1996) Synaptic activity and the construction of cortical circuits. Science 274: 1133 – 1138 Khilan AA, Al-Maslamani NA, Horn HF (2021) Cell stretchers and the LINC complex in mechanotransduction. Arch Biochem Biophys 702: 108829 Kitajiri S, Sakamoto T, Belyantseva IA, Goodyear RJ, Stepanyan R, Fujiwara I, Bird JE, Riazuddin S, Riazuddin S, Ahmed ZM et al (2010) Actin-bundling protein TRIOBP forms resilient rootlets of hair cell stereocilia essential for hearing. Cell 141: 786 – 798 Kros CJ, Ruppersberg JP, R€usch A (1998) Expression of a potassium current in inner hair cells during development of hearing in mice. Nature 394: 281 – 284 Kuhn S, Johnson SL, Furness DN, Chen J, Ingham N, Hilton JM, Steffes G, Lewis MA, Zampini V, Hackney CM et al (2011) miR-96 regulates the progression of differentiation in mammalian cochlear inner and outer hair cells. Proc Natl Acad Sci U S A 108: 2355 – 2360 Lemeille S, Paschaki M, Baas D, Morl(cid:1)e L, Duteyrat JL, Ait-Lounis A, Barras E, Soulavie F, Jerber J, Thomas J et al (2020) Interplay of RFX transcription factors 1, 2 and 3 in motile ciliogenesis. Nucleic Acids Res 48: 9019 – 9036 Li Y, Zheng M, Lau YF (2014) The sex-determining factors SRY and SOX9 regulate similar target genes and promote testis cord formation during testicular differentiation. Cell Rep 8: 723 – 733 Lippe WR (1994) Rhythmic spontaneous activity in the developing avian auditory system. J Neurosci 14: 1486 – 1495 Liu Y, Qi J, Chen X, Tang M, Chu C, Zhu W, Li H, Tian C, Yang G, Zhong C et al (2019) Critical role of spectrin in hearing development and deafness. Sci Adv 5: eaav7803 Love MI, Huber W, Anders S (2014) Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol 15: 550 Manor U, Disanza A, Grati M, Andrade L, Lin H, Di Fiore PP, Scita G, Kachar B (2011) Regulation of stereocilia length by myosin XVa and whirlin depends on the Actin-regulatory protein Eps8. Curr Biol 21: 167 – 172 Marcotti W, G(cid:1)el(cid:1)eoc GS, Lennan GW, Kros CJ (1999) Transient expression of an inwardly rectifying potassium conductance in developing inner and outer hair cells along the mouse cochlea. Pflugers Arch 439: 113 – 122 Marcotti W, Johnson SL, R€usch A, Kros CJ (2003a) Sodium and calcium currents shape action potentials in immature mouse inner hair cells. J Physiol 552: 743 – 761 Marcotti W, Johnson SL, Holley MC, Kros CJ (2003b) Developmental changes in the expression of potassium currents of embryonic, neonatal and mature mouse inner hair cells. J Physiol 548: 383 – 400 Marcotti W, Johnson SL, Kros CJ (2004) A transiently expressed SK current sustains and modulates action potential activity in immature mouse inner hair cells. J Physiol 557: 613 – 633 Maul A, Huebner AK, Strenzke N, Moser T, R€ubsamen R, Jovanovic S, H€ubner CA (2022) The Cl(cid:1)-channel TMEM16A is involved in the generation of cochlear Ca2+ waves and promotes the refinement of auditory brainstem networks in mice. Elife 11: e72251 Mhatre AN, Li Y, Atkin G, Maghnouj A, Lalwani AK (2006) Expression of Myh9 in the mammalian cochlea: localization within the stereocilia. J Neurosci Res 84: 809 – 818 Mikaelian D, Ruben RJ (1964) Development of hearing in the normal CBA-J mouse: correlation of physiological observations with behavioural responses and with cochlear anatomy. Arch Otolaryngol 59: 451 – 461 Moody WJ, Bosma MM (2005) Ion channel development, spontaneous activity, and activity-dependent development in nerve and muscle cells. Physiol Rev 85: 883 – 941 Moser T, Grabner CP, Schmitz F (2020) Sensory processing at ribbon synapses in the retina and the cochlea. Physiol Rev 100: 103 – 144 M€uller M, von H€unerbein K, Hoidis S, Smolders JW (2005) A physiological place-frequency map of the cochlea in the CBA/J mouse. Hear Res 202: 63 – 73 M€uller NIC, Sonntag M, Maraslioglu A, Hirtz JJ, Friauf E (2019) Topographic map refinement and synaptic strengthening of a sound localization circuit require spontaneous peripheral activity. J Physiol 597: 5469 – 5493 Ohba S, He X, Hojo H, McMahon AP (2015) Distinct transcriptional programs underlie Sox9 regulation of the mammalian chondrocyte. Cell Rep 12: 229 – 243 Pacentine I, Chatterjee P, Barr-Gillespie PG (2020) Stereocilia rootlets: actin- based structures that are essential for structural stability of the hair bundle. Int J Mol Sci 21: 324 Patro R, Duggal G, Love MI, Irizarry RA, Kingsford C (2017) Salmon provides fast and bias-aware quantification of transcript expression. Nat Methods 14: 417 – 419 Romand R (1983) Development of the cochlea. In Development of the auditory and vestibular systems, Romand R (ed), pp 47 – 88. New York, NY: Academic Press Roux I, Safieddine S, Nouvian R, Grati M, Simmler MC, Bahloul A, Perfettini I, Le Gall M, Rostaing P, Hamard G et al (2006) Otoferlin, defective in a human deafness form, is essential for exocytosis at the auditory ribbon synapse. Cell 127: 277 – 289 Self T, Sobe T, Copeland NG, Jenkins NA, Avraham KB, Steel KP (1999) Role of myosin VI in the differentiation of cochlear hair cells. Dev Biol 214: 331 – 341 Sonntag M, Englitz B, Kopp-Scheinpflug C, R€ubsamen R (2009) Early postnatal development of spontaneous and acoustically evoked discharge activity of principal cells of the medial nucleus of the trapezoid body: an in vivo study in mice. J Neurosci 29: 9510 – 9520 Stellwagen D, Shatz CJ (2002) An instructive role for retinal waves in the development of retinogeniculate connectivity. Neuron 33: 357 – 367 Tritsch NX, Yi E, Gale JE, Glowatzki E, Bergles DE (2007) 2007 the origin of spontaneous activity in the developing auditory system. Nature 450: 50 – 55 Tritsch NX, Rodr(cid:1)ıguez-Contreras A, Crins TT, Wang HC, Borst JG, Bergles DE (2010) Calcium action potentials in hair cells pattern auditory neuron activity before hearing onset. Nat Neurosci 13: 1050 – 1052 Wang HC, Lin CC, Cheung R, Zhang-Hooks Y, Agarwal A, Ellis-Davies G, Rock J, Bergles DE (2015) Spontaneous activity of Cochlear hair cells triggered by fluid secretion mechanism in adjacent support cells. Cell 163: 1348 – 1359 (cid:1) 2023 The Authors The EMBO Journal 42: e112118 | 2023 19 of 20 The EMBO Journal Adam J Carlton et al Wangemann P, Schacht J (1996) Homeostatic mechanisms in the cochlea. In The cochlea, Dallos P, Popper A, Fay R (eds), pp 130 – 185. New York, NY: Springer Yu CR, Power J, Barnea G, O’Donnell S, Brown HE, Osborne J, Axel R, Gogos JA (2004) Spontaneous neural activity is required for the establishment and maintenance of the olfactory sensory map. Neuron 42: 553 – 566 Zampini V, R€uttiger L, Johnson SL, Franz C, Furness DN, Waldhaus J, Xiong H, Hackney CM, Offenhauser N, Di Fiore PP et al (2011) Eps8 regulates hair bundle length and functional maturation of mammalian auditory hair cells. PLoS Biol 9: e1001048 Zhang LI, Poo MM (2001) Electrical activity and development of neural circuits. Nat Neurosci 4: 1207 – 1214 Zhang-Hooks Y, Agarwal A, Mishina M, Bergles DE (2016) NMDA receptors enhance spontaneous activity and promote neuronal survival in the developing cochlea. Neuron 89: 337 – 350 Zhou Y, Zhou B, Pache L, Chang M, Khodabakhshi AH, Tanaseichuk O, Benner C, Chanda SK (2019) Metascape provides a biologist-oriented resource for the analysis of systems-level datasets. Nat Commun 10: 1523 License: This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. 20 of 20 The EMBO Journal 42: e112118 | 2023 (cid:1) 2023 The Authors
10.3847_1538-4357_acc250
The Astrophysical Journal, 946:114 (23pp), 2023 April 1 © 2023. The Author(s). Published by the American Astronomical Society. https://doi.org/10.3847/1538-4357/acc250 Effects of Magnetic Fields on Gas Dynamics and Star Formation in Nuclear Rings Sanghyuk Moon1,2 , Woong-Tae Kim1,3 , Chang-Goo Kim2 , and Eve C. Ostriker2 1 Department of Physics & Astronomy, Seoul National University, Seoul 08826, Republic of Korea; sanghyuk.moon@princeton.edu 2 Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08544, USA; cgkim@astro.princeton.edu, eco@astro.princeton.edu 3 SNU Astronomy Research Center, Seoul National University, Seoul 08826, Republic of Korea; unitree@snu.ac.kr Received 2022 November 22; revised 2023 February 6; accepted 2023 March 6; published 2023 April 7 Abstract Nuclear rings at the centers of barred galaxies are known to be strongly magnetized. To explore the effects of magnetic fields on star formation in these rings and nuclear gas flows, we run magnetohydrodynamic simulations in which there is a temporally constant magnetized inflow to the ring, representing a bar-driven inflow. The mass inflow rate is 1Me yr−1, and we explore models with a range of field strength in the inflow. We adopt the TIGRESS framework developed by Kim & Ostriker to handle radiative heating and cooling, star formation, and resulting supernova (SN) feedback. We find that magnetic fields are efficiently amplified in the ring due to rotational shear and SN feedback. Within a few 100 Myr, the turbulent component Btrb in the ring saturates at ∼35 μG (in rough equipartition with the turbulent kinetic energy density), while the regular component Breg exceeds 50 μG. Expanding superbubbles created by clustered SN explosions vertically drag predominantly toroidal fields from near the midplane to produce poloidal fields in high-altitude regions. The growth of magnetic fields greatly suppresses star formation at late times. Simultaneously, strong magnetic tension in the ring drives radially inward accretion flows from the ring to form a circumnuclear disk in the central region; this feature is absent in the unmagnetized model. Unified Astronomy Thesaurus concepts: Star formation (1569); Galaxy circumnuclear disk (581); Barred spiral galaxies (136); Stellar feedback (1602); Interstellar medium (847); Magnetohydrodynamics (1964); Magnetohy- drodynamical simulations (1966); Galaxy nuclei (609) 1. Introduction 0.1 A characteristic result of dynamical interactions between a bar and gas in disk galaxies is the formation of a pair of large- scale shocks running along the leading sides of the bar inside of loses angular corotation. Gas entering the shock front momentum and is deflected inward. In optical images, the compressed inflowing gas is seen as narrow dust lanes along which gas is funnelled toward the central regions. The observed in ~ –10 Me yr−1 and is mass inflow rate is of the order of M believed to be time-variable (Benedict et al. 1996; Regan et al. 1997; Meier et al. 2008; Elmegreen et al. 2009; Shimizu et al. 2019; Sormani & Barnes 2019). Bar-driven inflowing gas has residual angular momentum and thus forms a circumnuclear ring, which is often observed to be active in star formation. Star-forming nuclear rings are found in about ∼20% of disk galaxies in the local universe, 80% among which are barred (Comerón et al. 2010), and have star formation rates (SFRs) of ∼0.1–10 Me yr−1 (Mazzuca et al. 2008; Ma et al. 2018). While nuclear rings are sometimes found in unbarred galaxies, a majority of such galaxies have oval distortions, strong spiral arms, or close companions, all of which are thought to provide nonaxisymmetric gravitational torques similar to bars that could effectively drive gas inward (Comerón et al. 2010). Spectroscopic observations have revealed that nuclear rings are long-lived, composed of not only young star clusters formed recently but also old stellar populations with ages ranging from ∼100 Myr to a few formation histories are gigayears. The reconstructed star Original content from this work may be used under the terms of the Creative Commons Attribution 4.0 licence. Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. 1 characterized by a large time variability, involving multiple timescales ranging from a few tens of megayears to a few gigayears (e.g., Allard et al. 2006; Sarzi et al. 2007; Gadotti et al. 2019; Prieto et al. 2019; Nogueras-Lara et al. 2020). Over time, ring star formation may lead to the development of nuclear disks (Launhardt et al. 2002; Bittner et al. 2020; Gadotti et al. 2020; de Sá-Freitas et al. 2023; Sormani et al. 2022), which are also known as “disk-like bulges” as distinct from classical and box/peanut bulges (Athanassoula 2005). Recently, a number of authors have studied gas dynamics and star formation in and around nuclear rings, using numerical simulations with realistic treatment of star formation and (2019) conducted feedback. For example, Armillotta et al. hydrodynamic simulations of the interstellar medium (ISM) to study gas flows and star formation in the Central Molecular Zone (CMZ), which is believed to represent a nuclear ring in our own Milky Way. Their simulations (with a mass resolution of 2 × 103 Me) showed that the SFR of the CMZ goes through several burst-quench cycles with a mixture of a short period (∼50 Myr) and long period (∼200 Myr), although the gas mass remains relatively constant over time. Tress et al. (2020) and Sormani et al. (2020) used a higher mass resolution of <100 Me (with adaptive mass refinement depending on local density and temperature) resolving Sedov-Taylor the CMZ, blastwaves in their simulations. Contrary to Armillotta et al. (2019), these authors found that the SFR in the ring steadily increases in time in proportion to the gas mass, with the gas depletion time almost constant within a factor of ∼2. Seo et al. ran hydrodynamic simulations coupled with N-body stellar dynamics to study how a nuclear ring forms and evolves in a situation where a stellar bar forms and grows self-consistently, rather than being treated as a fixed potential. They found that star formation supernova (SN) explosions to model for most (2019) The Astrophysical Journal, 946:114 (23pp), 2023 April 1 Moon et al. in a nuclear ring is sustained for a long (>1 Gyr) period of time, and that the ring SFR correlates well with the mass inflow rate to the ring. The diversity of findings from the above studies motivated us to undertake simulations with higher resolution in the ring region than is possible to achieve with global models. A key goal was to test whether a constant mass inflow rate results in steady ring star formation, or if instead the gas mass builds up and then produces intermittent bursts of star formation. To explore whether steady versus bursty behavior in ring star formation may depend on the inflow rate, in Moon et al. (2021, I), we developed a semiglobal numerical hereafter Paper framework that provides explicit control of the mass inflow rate via boundary conditions. Paper I found that, (1) when the mass inflow rate is fixed in time, a quasi-steady equilibrium state for a wide range of Min is (0.125–8 Me yr−1), in which the SFR and depletion time are almost constant within a factor of ∼2; (2) vertical dynamical equilibrium is established within the ring gas, in which the thermal and turbulent pressures due to stellar feedback balance the gravitational field arising from both gas and stars; (3) the pressure-regulated, feedback-modulated (PRFM) star formation theory is satisfied as previously shown for featureless disks in Kim & Ostriker (2017) and Ostriker & Kim (2022), and for disks with a spiral arm potential in Kim et al. (2020), but in to disk regions where the SFR adapts to the contrast equilibrium value set by gas mass, in nuclear rings the gas mass instead adapts to the SFR set by the mass inflow rate. reached at 0.8 in M SFR ~ To understand what might have produced temporal varia- tions as well as spatial asymmetry of observed ring star formation, and how the delay between star formation and feedback affects interpretation of self-regulated equilibrium, in Moon et al. (2022, hereafter Paper II) we allowed the mass inflow rate to vary with time in a prescribed way and/or to be spatially asymmetric. Paper II found that (1) time-varying mass inflows with a sufficient oscillation amplitudecause episodic star formation, provided that the timescale of the inflow rate variations is sufficiently long (50 Myr); (2) a sudden increase of the mass inflow rate through one of the two dust lanes causes lopsided star formation in the ring, lasting no longer than a few orbital periods; and (3) the PRFM theory is still satisfied even for a nonsteady system in which the SFR and gas mass vary with time, provided that the time delay between star formation and feedback is properly taken into account. While the studies above have improved our understanding of star-forming physics in nuclear rings, they were all limited to unmagnetized models. Observations show that nuclear rings in real galaxies are quite strongly magnetized. Assuming energy equipartition between magnetic fields and cosmic rays (CRs), the average magnetic field strengths in nuclear rings inferred from radio synchrotron observations are estimated to be ∼55 μG for NGC 1097 (Beck et al. 2005), ∼63 μG for NGC 1365 (Beck et al. 2005), and ∼84 μG for NGC 5792 (Yang et al. 2022), much stronger than in spiral arms of normal disk galaxies (Beck 2015). Beck et al. (1999, 2005) mapped radio continuum emission in barred galaxies and found that the magnetic fields are predominantly parallel to the dust lanes, while penetrating the nuclear rings with a large pitch angle (∼40°). Strong magnetic fields would provide additional support for gas against gravity, potentially reducing the SFR (Pillai et al. 2015; Tabatabaei et al. 2018). Indeed, Tabatabaei et al. (2018) found a strong positive correlation between the gas depletion time and the magnetic field strength for individual giant clumps distributed along the nuclear ring of NGC 1097. No correlation was found between the depletion time and the turbulent velocity dispersion, suggesting that it may be magnetic fields rather than SN feedback that suppress ring star formation. In this paper, we present results from magnetohydrodynamic (MHD) simulations of star-forming, magnetized nuclear rings. This work extends Paper I by considering magnetized gas inflows at the domain boundaries. To focus on the effects of magnetic fields on ring star formation, we fix the mass inflow rate and characteristic ring radius (based on the imposed inflowing gas), while varying the angular momentum of magnetic field strength of the inflowing gas. By comparing the results from models with different field strengths, we quantify how magnetic fields affect dynamical evolution of nuclear rings and star formation therein. In addition to allowing us to study effects of magnetization to explore how on star formation, our models are useful magnetic fields affect accretion in the central region of galaxies. In particular, our magnetized simulations show that gas accretes inward from the star-forming nuclear ring, which could potentially lead to formation of a circumnuclear disk (CND) near the galactic center. Based on the measured strength and pitch angle of magnetic fields in the nuclear ring of NGC 1097, Beck et al. (2005) suggested that magnetic stress can drive gas accretion from a ring to fuel an active galactic nucleus. Other proposed mechanisms for gas inflows near a galaxy center include bars-within-bars (Shlosman et al. 1989), nuclear spirals (Maciejewski 2004; Kim & Elmegreen 2017), and SN feedback (Wada 2004; Tress et al. 2020). Here we use direct numerical simulations to study gas accretion and its outcomes in the presence of magnetic fields and star formation feedback. The remainder of this paper is organized as follows. In Section 2, we outline the equations that we solve, summarize the TIGRESS4 numerical framework for the ISM and star formation physics, and describe our treatment of the boundary conditions for magnetized gas inflows. In Section 3, we present the overall time evolution of our models with a focus on star formation histories, and examine gas accretion toward the center driven by magnetic stresses. In Section 4, we present the temporal evolution of the magnetic field strength in the ring and explore the effects of magnetization on the ring star formation. Finally, we summarize and discuss our results in Section 5. 2. Numerical Methods To numerically model the central kiloparsec region of a barred galaxy with high resolution, we adopt the semiglobal numerical model introduced by Paper I. In this approach, our computational domain covers only the nuclear ring and its immediate vicinity, and bar-driven mass inflows are treated by boundary conditions. Nonlinear interactions between the bar and gas leading to the gas inflows are assumed to occur outside of the computational domain, and are not explicitly modeled. inflow rate and angular Instead, we control momentum of In this section, we present the basic equations we solve (Section 2.1), the inflows using free parameters. the mass 4 TIGRESS is an acronym for “Three-phase Interstellar medium in Galaxies Resolving Evolution with Star formation and Supernova feedback.” 2 The Astrophysical Journal, 946:114 (23pp), 2023 April 1 summarize the TIGRESS framework for star formation and feedback (Section 2.2), and describe the inflow boundary conditions for magnetized gas (Section 2.3). 2.1. Governing Equations Our computational domain is a Cartesian cube with side L = 2048 pc located at the galaxy center. The domain rotates at - , corresponding 1ˆ z an angular frequency to the adopted bar pattern speed.5 We include radiative heating and cooling of the ISM, gaseous self-gravity, and a fixed external gravitational potential responsible for the background rotation curve. The governing equations we solve are 36 km s W = p kpc - 1 ¶ r ¶ t + r v · ( ) = 0, ¶ ) r v ( ¶ t =- +  · ( r vv + P  +  ) 2 r W p ´ - v r  F , tot ¶ E ¶ t +  · [ v E + ( P  +  ) · v ] =- r v ·  F -  r tot , ¶ B ¶ t =  ´ ´ v ( B ) , 2  F = self p G4 ( r + r sp ) . ( ) 1 ( ) 2 ( ) 3 ( ) 4 5 ( ) ) )  p p B - = ( 4 BB 2 ( 8 stress Here, ρ and ρsp are, respectively, the volume density of gas and young star cluster particles that form, v is the gas velocity in the rotating frame, P is the gas pressure,  is the identity matrix,  tensor, is E = ρv2/2 + P/(γ − 1) + B2/(8π) is the total energy density with adiabatic index γ = 5/3, Φtot = Φself + Φext + Φcen is the total gravitational potential, consisting of the self-gravitational potential Φself, the external gravitational potential Φext, and the 1 + , and r is the net centrifugal potential 2 cooling rate per unit volume. the Maxwell F = - W In our models, the external gravity Φext = ΦBH + Φb arises from a central supermassive black hole and a stellar bulge. We note that we do not include a nonaxisymmtric bar potential. The black hole is modeled by a Plummer potential x ( cen 2 p y ) 2 2 F = - BH GM BH 2 r + 2 r BH 6 ( ) with mass MBH = 1.4 × 108 Me and rBH = 20 pc. For the stellar bulge, we take softening length 4 r r 3 p r G r b b 0 r b 1 + + ln F = - ⎛ ⎜ ⎝ with central density ρb0 = 50 Me pc−3 and scale radius rb = 250 pc. The resulting rotation curve and circular velocity the ring position are similar to those of NGC 1097 as at reported in Onishi et al. (2015). ⎞ ⎟ ⎠ ( ) 7 . b 2 r 2 r b Moon et al. The net cooling rate of gas per unit volume in Equation (3) is given by  r = n n ( H H L - G - G PE CR ) , ( ) 8 where nH = ρ/(μHmH) is the hydrogen number density with mean molecular weight per hydrogen μH = 1.4271 assuming solar abundances. For the cooling function Λ(T), we take the fitting formula of Koyama & Inutsuka (2002; see Kim et al. 2008 for a typo-corrected version) for T < 104.2K, and the tabulated collisional ionization equilibrium cooling curve at solar metallicity of Sutherland & Dopita (1993) for T > 104.2K. The gas temperature T is related to density and pressure via the ideal equation of state P = ρkBT/(μmH), with the mean molecular weight μ(T) varying with T from μato = 1.295 for neutral gas to μion = 0.618 for fully ionized gas (Kim & Ostriker 2017). In Equation (8), ΓPE represents an idealized model of the photoelectric heating rate by far-ultraviolet (FUV) radiation impinging on dust grains, and is given by G = G PE PE,0 ⎛ ⎜ ⎝ m ( m T ) ato - - m m ion ion J FUV J FUV,0 ⎞ ⎛ ⎟ ⎜ ⎝ ⎠ + ⎞ 0.0024 , ⎠ ⎟ 9 ( ) where ΓPE,0 = 2 × 1026 erg s−1 (Koyama & Inutsuka 2002) and JFUV,0 = 2.1 × 104 erg s−1cm−2sr−1 (Draine 1978) are normal- ization factors based on the solar neighborhood conditions. The term in the first parentheses in Equation (9) reduces the photoelectric heating at high T (since realistically dust grains would sublimate), shutting this heating off completely in the fully ionized gas. The small factor in the last parentheses represents a minor contribution from the metagalactic FUV background. We use the same approximate method as in Paper I to calculate the mean FUV intensity JFUV from young star cluster particles in the simulations. For this, we first calculate the luminosity surface density ΣFUV of all star cluster particles younger than 40 Myr in the simulation domain,6 and then apply an approximate model for dust attenuation, setting J FUV = 1 - S FUV p 4 ⎜ ⎛ ⎝ ( E 2 t ^ t ^ 2 ) ⎟ ⎞ ⎠ - e n H n 0 . ( 10 ) Here, E2 is the second exponential integral, τ⊥ = κdΣ with κd = 103 cm−2 g−1 is the vertical optical depth for the mean gas surface density Σ averaged over the entire domain, and the factor in parentheses represents the average attenuation factor for a uniform-density slab with a uniform source distribution. The exponential factor represents local shielding, with n0 the density above which this shielding becomes significant. For the models presented in this paper, we take n0 = 50 cm−3, which yields a dependence of JFUV on density comparable to that obtained by applying the adaptive ray-tracing method of Kim et al. (2017; for details of this comparison, see Paper I). Inside dense regions where FUV radiation is heavily shielded, the heating is dominated by the CR ionization. The 5 Even though we do not include a bar potential explicitly, it is advantageous to work in the rotating frame, since then the nozzles for inflow streams (see Section 2.3) may be kept fixed in both space and time. 6 The FUV luminosity-to-mass ratio of the star cluster particles depends on their age, based on STARBURST99 model calculations. 3 The Astrophysical Journal, 946:114 (23pp), 2023 April 1 adopted heating rate ΓCR in Equation (8) is given by G = CR x q CR CR ⎛ ⎜ ⎝ m ( m T ) ato - - m m ion ion , ⎞ ⎟ ⎠ ( 11 ) the energy yield per where qCR = 10 eV is ionization (Glassgold et al. 2012; see also Gong et al. 2017), and ξCR denotes the CR ionization rate. The term inside the parentheses is again to shut off the CR heating in fully ionized gas. Assuming that ξCR is proportional to the SFR surface density ΣSFR and is attenuated by a factor of Σ0/Σ above a critical gas surface density Σ0 = 10.7 Me pc−2 (Neufeld & Wolfire 2017), we set x CR x= CR,0 S S SFR SFR,0 min { 1, S 0 S } , ( 12 ) where ξCR,0 = 2 × 10−16s−1 is the CR ionization rate in the solar neighborhood (Indriolo et al. 2007; Neufeld & Wolfire 2017). Equations (1)–(4) are discretized on a uniform mesh with 5123 cells: the corresponding grid spacing is Δx = 4 pc. We update the physical quantities using a version of the Athena MHD code (Stone et al. 2008), which employs the MUSCL-Hancock scheme with the constrained transport algorithm to preserve ∇ · B = 0 within machine precision (Stone & Gardiner 2009), and applies a first-order flux correction when needed (Lemaster & Stone 2009). We apply the Green’s function convolution method aided by a fast Fourier transform (e.g., Skinner & Ostriker 2015) to solve the Poisson equation (Equation (5)) with the vacuum boundary condition, i.e., Φself → 0 at infinity. 2.2. Star Formation and Feedback We handle star formation and feedback using the TIGRESS framework (Kim & Ostriker 2017; see also Paper I), which we briefly summarize here. We refer the reader to Kim & Ostriker (2017) for a more complete description. r LP simultaneously: conditions 2 p D G x ) ( (1) are met with a local sound speed cs, We create a star cluster particle whenever the following = r> three 2 c 8.86 s the threshold density based on the Larson-Penston collapse solution, (2) Φself is a local minimum,7 and (3) the velocity is converging in all directions. A portion of the gas mass in the surrounding 27 cells is converted to the initial mass of a newly created star cluster particle. Star cluster particles are allowed to accrete mass and momentum from their surroundings and merge with nearby particles within 3Δx until the onset of the first SN explosion (∼4 Myr). For orbits of star cluster particles, we solve their equations of motion x = - F -  tot 2 W p ´  x , ( 13 ) using the Boris algorithm that preserves the Jacobi integral very accurately (Boris 1970; see also the Appendix of Paper I). in the form of Star cluster particles with ages less than 40 Myr exert feedback heating (Equation (9)), CR heating (Equation (11)), and type II SN explosions. The method of energy and momentum injection for a given SN event depends on the density of the ambient medium. If the ambient density is low enough that the shell- photoelectric the Moon et al. SN = 51 10 erg formation radius is expected to be resolved, we regard the SN remnants as being in the Sedov-Taylor phase and inject 72% of the SN energy E in the form of thermal energy and the remaining 28% in the form of kinetic energy. If the ambient density is too high for the adiabatic stage of evolution to be resolved, we assume that the SN remnant has already cooled to enter radial momentum = p n as calibrated from ( H * higher-resolution simulations (Kim & Ostriker 2015a). In both cases, each SN event returns the ejecta mass Mej = 10 Me from a star cluster particle back to the ISM. the snowplow phase and inject 0.17 km s 2.8 - - 3 ) M  cm 10 ´ - 1 5 2.3. Magnetized Inflow Streams Paper I introduced the semiglobal framework that treats bar- driven mass inflows via imposed boundary conditions for hydrodynamic simulations. Here, we modify the boundary conditions slightly to handle magnetized inflows. We inject gas streams into the computational domain through two circular nozzles with radius ζin = 112 pc placed at the y-boundaries: the coordinates of the nozzle centers are (x, y, z) = ( m bin, ± L/2, 0), where bin = 512 pc is the impact parameter of the inflows (see Figure 3 of Paper I, for a schematic diagram). Here and hereafter, the upper and lower signs correspond to the upper and lower nozzles, respectively. We set the streaming velocity at the nozzles to v in =  v ( sin q ˆ x in in + cos q ˆ) y , in ( 14 ) where θin = 10° is the inclination angle of the streams relative to the y-axis. The condition of the angular momentum conservation implies that the inflow speed vin determines the location where ring forms. By setting the specific angular the nuclear momentum (in the inertial frame) of the inflows equal to Rringvrot(Rring) with the circular velocity v 1 2 dR , ) we obtain Rd ( º F ext rot v in ( x ,  L 2 ) = R R v ring rot cos ( q in ) ring L ( x ∣ 2 - W R p ) 2 sin q ∣ in . ( 15 ) We fix the ring radius to Rring = 500 pc and use Equations (14) and (15) to find the corresponding inflow velocity inside the nozzles, which varies from 72 km s−1 to 115 km s−1. The density ρin of the streams sets the mass inflow rate as  M in = ∬ r v in in cos q in dxdz , ( 16 ) 2 ) in in z b  + x [( where the integrations are performed over the two nozzles, i.e., y = ±L/2 and < . We fix the mass 2 1 2 z ] inflow rate to 1 Me yr−1 by taking ρin = 0.138 Me pc−3, corresponding to nH = 3.9 cm−3. The mean inflow speed in the  - . 1 M 2 nozzles amounts to v ( ¯ in in Radio polarization observations indicate that the magnetic fields are roughly parallel to dust lanes and point to the galactic (Beck et al. 2005; Lopez-Rodriguez et al. 2021). center Motivated by this, we take the magnetic fields inside the nozzles parallel to the inflow velocity as 91 km s r pz in cos q in 2 in º = ) 7 Φext and Φcen vary very slowly in space and thus have a negligible contribution to the gradients of Φtot. B in = 4 in v in , ( 17 ) B v in ̈ The Astrophysical Journal, 946:114 (23pp), 2023 April 1 Moon et al. Table 1 Model Parameters rates inside the ring and compare them with theoretical predictions. Model (1) Binf B100 B30 B10 Rring (2) (pc) 500 500 500 500 Min (3) (Me yr−1) 1.0 1.0 1.0 1.0 βin (4) ∞ 100 30 10 Bin,c (5) (μG) 0 1.6 3.0 5.2 with the amplitude B in = ⎜ ⎛ ⎝ in p P8 b in 1 2 ⎞ ⎠ ⎟ cos pz z 2 in ⎛ ⎜ ⎝ . ⎞ ⎟ ⎠ Bin,avg (6) (μG) 0 0.76 1.4 2.4 ( 18 ) 2 ) in z b  + x [( z = 2 1 2 ] Here, Pin = ρinkBTin/(μHmH) is the thermal pressure of the inflowing gas with temperature Tin = 2 × 104 K, βin is a plasma parameter measuring the ratio of thermal to magnetic pressure, and is the distance from the nozzle center. The cosine term ensures that the fields vanish at the nozzle boundaries, preventing the gas just outside the nozzles from accidentally acquiring too large Alfvén speeds. In Athena, the velocity and magnetic fields are cell-centered and face-centered, respectively. Despite Equation (17), the mismatch in the evaluation points of Bin and vin yields nonvanishing vin × Bin at the innermost ghost zones at early time, as explained in Appendix A. This allows seed magnetic fields to leak into our computational domain through Equation (4), which are subsequently stretched by the inflows to become parallel to the streams, smoothly matching the boundary conditions (see Section 3.1). We allow gas to freely escape from the simulation domain, but forbid inflows except through the nozzles. We accomplish this by setting the hydrodynamic variables in the ghost zones by extrapolating from the two adjacent active zones, while keeping the normal velocity at zero if the velocity is directed inward. The magnetic fields in the ghost zones are simply copied from the innermost active zones. 2.4. Models We consider four models with βin = ∞ , 100, 30, and 10 for the plasma beta parameter in the inflow. Table 1 summarizes the model parameters for all models. Column (1) lists the model names. Columns (2) and (3) give the ring radius and the mass inflow rate, respectively, which are the same for all models. Column (4) gives βin. Column (5) and (6) give the magnetic field strength at the nozzle centers Bin,c = Bin(ζ = 0) and the mean field strength inside the nozzles Bin,avg, respectively. Model B100 is our fiducial model, which has βin = 100, Bin,c = 1.6 μG, and Bin,avg = 0.76 μG. The simula- tion domain is initially filled with rarefied gas with density n cm temperature H T = 2 × 104 K, and subsequent evolution is governed entirely by the inflowing streams. 50 pc z ∣ ∣ ( exp and 10 - = )] - 5 - 3 [ 3. Evolution In this section, we describe the overall evolution of our fiducial model B100 in terms of the gas and magnetic field distribution and star formation. We also measure the accretion 5 3.1. Overall Evolution Figure 1 plots snapshots of gas surface density together with young star cluster particles as well as the projected magnetic field lines overlaid over the total field strength map in our fiducial model B100 at a few selected epochs. Figure 2 shows similar plots for models B100, B30, and B10 at a selected epoch for each model. Figure 3 plots the evolution of the radial profiles of the azimuthally averaged surface density for all models. Early evolution of model B100 is qualitatively similar to that of the unmagnetized models presented in Paper I. There is an initial transient phase during which the inflows follow nearly ballistic orbits (Figure 1(a)), but within half an orbital time (∼8 Myr), the streams from the opposite boundaries collide with each other, which drives strong shocks with a Mach number ∼16. The streams lose their orbital kinetic energy as they pass through the shocks multiple times, and form a nuclear ring with radius R ∼ Rring at t ∼ 50 Myr (corresponding to ∼3 orbital times; see Figure 1(c)). Still, the ring is elongated with the major axis precessing takes another under the external gravitational potential. It ∼50 Myr for the ring to fully circularize (Figures 1(e), (f)). The ring soon reaches a quasi-steady equilibrium where the FUV and CR heating balances the radiative cooling, the SN feedback balances the turbulent dissipation, and the thermal and turbulent pressures remain approximately constant. The result- ing total midplane pressure matches the overlying weight in the ring. The gas mass in the ring also stays roughly constant as the net mass inflow rate balances the SFR. Star formation proceeds randomly throughout the entirety of the ring. Although the resulting SN feedback disperses the gas and drives turbulence locally and temporarily, it never destroys the ring entirely nor quenches star formation completely (Section 3.2; see also Paper I). As mentioned in Section 2.3 (and Appendix A), our boundary conditions introduce weak seed magnetic fields in the active domain, which are stretched along the streams by the inflowing gas. Except for the initial ∼10 Myr, magnetic fields in the streams remain well aligned with the inflow velocity and do not exhibit systematic growth in time, although they are perturbed intermittently by strong SN feedback from the ring. form a ring, magnetic fields become As predominantly toroidal in the ring, with large fluctuations due to SN feedback. The magnetic fields in the ring become stronger and more regular with time (see Section 4.1), presumably due to both a small- and large-scale dynamo driven by SN feedback and rotational shear inside the ring, the discussion of which we defer to Section 4.4. the streams Strong magnetic fields in the ring cause the evolution of model B100 to deviate significantly from that of model Binf after t ∼ 200 Myr in two ways. First, in contrast to model Binf where the surface density profile does not change much with time (Figure 3), an accretion flow develops in model B100 from the ring toward the center, gradually filling the region inside the ring. The accreting gas piles up at the center, forming a CND with radius of ∼50 pc. Second, strong magnetic fields and associated pressure make the SFR decrease with time in model B100 (see Section 4.2). The Astrophysical Journal, 946:114 (23pp), 2023 April 1 Moon et al. Figure 1. Face-on views of model B100 at t = 10, 50, 100, 220, 250, and 285 Myr (the figure continues on the next page). The left column displays the gas surface density (color scale) and newly formed star cluster particles with age <1 Myr (circles). The right column plots streamlines of the projected magnetic fields (color scale). The black solid lines in panel (a) are the ballistic = = x trajectories that a test particle injected with vin would follow. , overlaid on maps of 2 1 2 ) y B dz y = y B dz x ) ( ò ) ( ò and ( ò ( ò  ( dz dz +   r r r r 2 x ) ) Figure 1(k) shows that much of the ring gas is concentrated in dense, trailing spiral segments with a pitch angle of 45° and azimuthal spacing of ∼100–150 pc. These spiral segments start to appear roughly at t ∼ 260–270 Myr and keep being destroyed and regenerated thereafter. The quasi-regular spacing of these spiral segments suggests that they result from the 6 The Astrophysical Journal, 946:114 (23pp), 2023 April 1 Moon et al. Figure 1. (Continued.) magneto-Jeans instability (MJI) in which magnetic tension forces from bent field lines suppress the stabilizing effect of epicyclic motions (Elmegreen 1987; Kim & Ostriker 2001; Kim et al. 2002). Indeed, the corresponding dispersion relation (Equation (21) of Kim et al. 2002) for the parameters adopted from model B100 yields the most unstable wavelength of ∼120 pc, entirely consistent with the numerical results. Due to the MJI, some spiral segments attain sufficient density to form 7 The Astrophysical Journal, 946:114 (23pp), 2023 April 1 Moon et al. Figure 2. Similar to Figure 1, but for models B100, B30, and B10 at t = 250, 130, and 130 Myr, respectively (from top to bottom). stars. We note, however, that strong shear in the ring makes the MJI operate only temporarily, preventing runaway growth of spiral segments (see Section 4.2). Figure 4 plots the spatial distribution of gas and magnetic fields in x–z slices through three consecutive positions of a moving star cluster with mass 2 × 106 Me, marked by the star symbol in each panel. Repeated SN explosions create a superbubble around the cluster. The overpressurized bubble easily expands in the vertical direction where the gas density decreases, eventually breaking out and rapidly rising up with 8 The Astrophysical Journal, 946:114 (23pp), 2023 April 1 Moon et al. Figure 3. Evolution of the radial distribution of the azimuthally averaged gas surface density for all models. Colors indicates the time interval for a temporal average. The vertical dashed lines mark the ring location Rring = 500 pc. velocities exceeding 103 km s−1. Magnetic field lines are lifted to high-|z| regions and stretched by the flows of hot gas to generate a polodial component. Figure 5 plots a volumetric rendering of the three-dimensional magnetic field geometry and gas density, showing that the magnetic fields are predominantly toroidal in the ring and poloidal in the regions away from the midplane. reach the Evolution of models B10 and B30 is qualitatively similar to that of model B100 in the sense that the inflow streams collide ring, and accretion flows to form a star-forming nuclear develop from the ring toward the center when the magnetic stress becomes strong enough (see Section 3.3). However, the models with smaller βin evolutionary stage characterized by decreased SFR and an accretion flow from the ring toward the center at an earlier time, compared to higher βin models (see Figure 2). Figure 3 shows that unlike in model the rings in models B100, B30, and B10 expand Binf, inward with time. The surface density interior to the magnetized rings increases with time due to the radial accretion flows, forming a CND characterized by the central upturn of the radial surface density profile at R  50 pc. We note that we are unable to evolve the magnetized models for arbitrarily long times, because the Alfvén speed becomes too large in the low- density region above and below the magnetized CND, severely limiting the Courant-Friedrichs-Lewy time step. 3.2. Star Formation History We define the SFR, MSF , as the total mass of star cluster particles with age less than 10 Myr, divided by 10 Myr. Figure 6 plots the temporal histories of MSF , the total gas mass 9 Mgas, the gas depletion time t dep = M  M gas SF , ( 19 ) 0.8 SF ~ and the total magnetic energy Emag inside the computational domain. For all models, there is initial transient behavior as the ring forms and star formation develops. After t ∼ 50 Myr, the in model Binf, reaching a SFR becomes almost constant –0.9 Me yr−1, with a factor of ∼2 steady-state value M stochastic fluctuations due to turbulence driven by SN feedback, similar to the models presented in Paper I. The star formation history of model B100 is very similar to that of model Binf until t ∼ 200 Myr. After t ∼ 200 Myr, however, strong magnetic fields in the ring of model B100 result in a reduced SFR. The evolution of the SFR in models B30 and B10 is qualitatively similar: it reaches a quasi-steady value at t ∼ 50–100 Myr, which is lower by a factor of a few in the models with stronger magnetic fields (smaller βin), and then starts to decline after trend of increasing Mgas evident in Figure 6(b) as well as increasing gas surface density in the ring (Figure 3) indicate that the decline of the SFR in magnetized models is not caused by the reduction in the gas mass or surface density. It is rather because a larger for star formation, as reflected in Figure 6(c), which shows that tdep increases at late times. As the SFR drops below Min , the excess gas piles up in the ring and moves toward the center to form a CND. t ∼ 120 Myr. The secular the ring gas becomes inert fraction of Figure 6(d) the magnetic energy in the computational domain exponentially increases with time, shows that The Astrophysical Journal, 946:114 (23pp), 2023 April 1 Moon et al. Figure 4. Superbubble breakout in the vertical direction. Each panel from left to right shows the density slice of model B100 at y = − 438, −433, and −419 pc, the y- position of a star cluster particle with mass 2 × 106 Me (denoted by the yellow star) at t = 196.7, 197.2, and 197.5 Myr, respectively. The streamlines in gray . The red arrows are in-plane velocity vectors, (vx, vz), with their lengths represent the magnetic fields lines with the strength B . The expanding superbubble surrounding the star cluster particle lifts the toroidal magnetic fields near the midplane to high- 2 proportional to the speed v ( x altitude regions to produce poloidal fields. 2 1 2 )+ v z 2 1 2 B ) z m 0.1 G 2 B ( x 2 B y > + + = indicative of dynamo action. We note that the magnetic energy advected with the inflow streams is very small because v∥B is near the nozzles for most of the time, and therefore cannot account for the increase of Emag (see Appendix B). We will present a more detailed analysis on the growth of magnetic fields and their effects on star formation, and discuss possible causes of the magnetic field amplification in Section 4. 3.3. Magnetically Driven Accretion Flow Figures 1 and 3 show that all magnetized models develop an accretion flow from the ring toward the center. For rotating the total mass magnetized disks, Appendix C shows that accretion rate in a quasi-steady state can be written as  M acc »  M M +  M , R ( 20 ) where MM Maxwell and Reynolds stresses, respectively, defined as are the mass accretion rates due to the and MR  M M p= 2 ¶ ( ) Rv circ ¶ R ⎡ ⎣ - ⎤ ⎦ 1 ¶á 2 R T R f ñ ¶ R , ( 21a ) 10  M R = 2 p ¶ ( ) Rv circ ¶ R ⎡ ⎣ - ⎤ ⎦ 1 ¶á 2 r R u u R ¶ R ñ f , ( 21b ) 4 acc yr M  where vcirc = (Ω − Ωp)R is the background circular velocity in the rotating frame, and (uR, uf) are the perturbation in the radial and azimuthal velocities. Figure 7 plots the temporal histories  measured at R = 100 pc for all models, in comparison of Macc with the predictions due to the Reynolds and Maxwell stresses.  In model Binf, M - on average, show- ~ ´ - 1 4 10 ing that the mass accretion rate without magnetic fields remains small for all time. While the values of Macc in the magnetized models are also small at early times, they increase rapidly to –0.03 Me yr−1 toward the end of the runs. reach M are in good agreement with MM The temporal changes of Macc , indicating that magnetic tension is the major driver of mass accretion in our models.  ~ acc Figure 8 plots the radial profiles of Macc averaged for a few selected time intervals for model B100, together with MM . The accretion rate is not constant in radius but decreases toward the center, with a slope decreasing with time. This implies that mass is being deposited at all radii <Rring, consistent with the 0.02    The Astrophysical Journal, 946:114 (23pp), 2023 April 1 Moon et al. Figure 5. Perspective visualization of the three-dimensional magnetic field structure in model B100 at t = 250 Myr. The magnetic field lines are represented by red tubes, while the gas density is volume rendered in blue-green. Note that the magnetic fields are predominantly toroidal inside the nuclear ring because of the differential rotation, and poloidal in high-|z| regions due to SN-driven outflows. the density distributions shown accretion rate near radial 3. For t = 270–300 Myr, ring is ∼0.1 Me yr−1, i.e., one-tenth of the bar-driven inflow rate, while it decreases to 0.02–0.03 Me yr−1 near the center. The radial dependence of the measured accretion rates is overall in good agreement with MM , indicating that the accretion flows are mediated mostly by the magnetic tension forces. in Figure the 4. Magnetic Fields in the Ring In this section, we analyze evolution of the regular and turbulent magnetic fields in nuclear rings and explore their effects on the ring star formation. We also discuss vertical dynamical equilibrium in the presence of magnetic fields. Finally, we discuss our results in the context of dynamo theory. 4.1. Growth of Magnetic Fields The magnetic fields inside the ring and its interior are close to axisymmetric (Figure 1), which motivates us to decompose the fields into a regular component B and an irregular, turbulent component δB as B ( R , f , z ) = B ( R z , ) + d B ( R , f , z ) , ( 22 ) where the overbar denotes an azimuthal average X R z ( , ) º 2 p 1 òp 2 0 X d f , ( 23 ) for any physical quantity X. Note that B d = by definition. 0 ∣ º B ( 2 R d ∣ 2 +f d B 2 +f º Figure 9 plots the spatial distributions of the azimuthally averaged hydrogen number density nH, the strength of the 2 1 2 + regular component B B B ) ∣ , and the strength z 2 1 2 2 2 1 2 + d d of the turbulent component B B B ) ( ∣ R z in the R–z plane, for model B100 at t = 200 and 300 Myr. At t = 200 Myr, gas and magnetic fields are concentrated mostly in the nuclear ring delineated by the black circles centered at (R, z) = (500, 0) pc with radius 200 pc, while the region outside the ring is filled with diffuse gas. At this time, magnetic fields are dominated by the turbulent component, especially outside the rings: the density-weighted (see below) mean strength of the regular and turbulent components are 8 and 22 μG, respectively. The CND that begins to form near the center at t ∼ 250 Myr due to magnetically driven accretion from the ring is visible in the t = 300 Myr panels in the right column of Figure 9, as marked by the rectangles. The CND in our models is denser and more strongly magnetized than the ring. At t = 300 Myr, the bottom two panels show that the regular magnetic field is stronger than the turbulent field in the ring. To quantify the magnetic fields within the ring, we define the density-weighted average of the regular, turbulent, and total magnetic fields as B reg, j º ∬ B trb, j º 11 r ∬ ∬ rd B B dRdz j , dRdz r 2 1 2 j dRdz ∬ r dRdz ( 24 ) ( 25 ) , The Astrophysical Journal, 946:114 (23pp), 2023 April 1 Moon et al.  at R = 100 pc for Figure 7. Temporal histories of the accretion rate Macc models Binf (black), B100 (blue), B30 (green), and B10 (red). Thin and thick solid lines correspond to the instantaneous and time-averaged (using a 10 Myr window) values, respectively, directly measured from the simulations. Dashed and dotted lines are the predicted accretion rates due to the Maxwell and Reynolds stresses, using Equation (21a) (averaged over a 10 Myr window), respectively. The increasing trend of Macc in magnetized models is well explained by the Maxwell stress.  Figure 6. Temporal histories of (a) the SFR MSF , (b) the total gas mass Mgas, (c) the gas depletion time tdep, and (d) the total magnetic energy inside the computational domain. The black, blue, green, and red lines correspond to models Binf, B100, B30, and B10, respectively. B tot, j º ∬ r B 2 1 2 j dRdz ∬ r dRdz , ( 26 ) 2 + = B j 2d B j where the integration is performed over the circular regions by definition. 2 shown in Figure 9. Note that B j Figure 10 plots the time evolution of Breg = |Breg|, Btrb = |Btrb|, and Btot = |Btot| for all models. We note that the ring is quite eccentric and undergoes damped oscillations of eccentricity before it enters a quasi-steady state at t ∼ 100 Myr, in which case B and δB do not properly represent the regular and turbulent components.8 The regular fields grow superlinearly in time (neglecting temporal fluctuations), reaching Breg ∼ 50–70 μG at the end of the simulations. In contrast, the turbulent fields grow initially but saturate at Btrb ∼ 30–40 μG. The total magnetic fields are initially dominated by the turbulent component, but are overtaken by the regular component at later times. The growth rate of the regular magnetic field at late times is largely insensitive to βin, suggesting again that the field amplification is not due to the advection of magnetic energy through the 8 During its eccentricity oscillations, the ring becomes almost circular at t ∼ 60 Myr temporarily, producing peaks of Breg, Btrb, and Btot at that time. 12 Figure 8. Radial profiles of the mass accretion rate at different epochs for model B100. Solid and dashed lines correspond to the measured accretion rate and predicted accretion rate due to the Maxwell stress, respectively. nozzles (see Appendix B). The magnetic fields grow earlier in models with smaller βin because of the stronger seed fields. The growth of magnetic fields is most likely driven by SN feedback and rotational shear, which we will discuss in Section 4.4. Figure 11 plots the temporal changes of the radial and azimuthal components of Breg for model B100 (the vertical component of Breg is negligible), showing that both compo- nents grow in time. The sign of the radial component is the opposite of the sign of the azimuthal field, implying a trailing spiral geometry (see Figure 1), consistent with observed large- scale magnetic fields in the nuclear ring of NGC 1097 (Beck et al. 1999, 2005; Lopez-Rodriguez et al. 2021). However, the pitch angle of the regular fields - 1 B ( is than θp ∼ 40° inferred from the ∼6°–12°, much smaller the field observations; indeed, we would not expect geometry probed by synchrotron emission would be directly comparable to the mass-weighted magnetic field we show here. q º - p that tan reg, reg, B ) R f The Astrophysical Journal, 946:114 (23pp), 2023 April 1 Moon et al. Figure 9. Spatial distributions from model B100 of the azimuthally averaged hydrogen number density (top), and the strength of the regular (middle) and turbulent (bottom) components of the magnetic fields at t = 200 (left column) and 300 Myr (right column). The black circles centered at (R, z) = (500, 0) pc with radius 200 pc outline the ring, while the rectangles near R = 0 in the right column mark a CND. 4.2. Effects of Magnetic Fields on Star Formation It is well known that magnetic fields inhibit star formation by providing additional pressure and tension to resist gravitational collapse (e.g., Mestel & Spitzer 1956; McKee & Zweibel 1995; Hennebelle & Inutsuka 2019; Kim et al. 2021). To assess the dynamical importance of magnetic fields relative to thermal and turbulent pressures, we measure the sound speed cs, vertical velocity dispersion σz, Alfvén speed associated with regular vA,reg and turbulent vA,trb magnetic fields of the cold–warm medium with T < 2 × 104 K at the midplane as c s = ⎛ ⎜ ⎜ ⎝ ∭ z ∭ z z =D z =-D z =D z z =-D z P dxdydz Q r Q dxdydz 1 2 ⎞ ⎟ ⎟ ⎠ , z =D z =-D z =D z z ∭ z ∭ z =-D z 2 r v z Q dxdydz r Q dxdydz 1 2 ⎞ ⎟ ⎟ ⎠ , s z = ⎛ ⎜ ⎜ ⎝ 13 ( 27 ) ( 28 ) The Astrophysical Journal, 946:114 (23pp), 2023 April 1 Moon et al. Figure 11. Temporal changes of the (a) radial and (b) azimuthal components of Breg for model B100. Note that Breg,R < 0, i.e., it points toward the center. The is roughly constant at θp ∼ 12° for - 1 B ) ( pitch angle f t = 100–240 Myr and ∼6° for t = 270–300 Myr. q º - p tan reg, reg, B R and min = 700 pc simulations performed over the annular regions between R R 300 pc , i.e., the ring regions defined in Figure 9. max = Figure 12 plots temporal histories of cs, σz, vA,reg, and vA,trb from the magnetized models, showing that cs ∼ 3 km s−1 and σz ∼ 10 km s−1 with modest variations with time. The turbulent Alfvén speed saturates at a level where the turbulent magnetic energy is roughly comparable to the kinetic energy, with vA,trb ∼ (1.5–1.8)σz. The ratio of turbulent magnetic to turbulent (Kim & kinetic energy found in previous Ostriker 2015b, 2017; Pakmor et al. 2017; Ostriker & Kim 2022) is in the range 0.2–0.5, similar or perhaps slightly below what we find here. In contrast, the regular Alfvén speed is initially constant, but exhibits secular growth toward the end of each run, eventually overtaking vA,trb and reaching vA,reg > 20 km s−1. The value of vA,reg begins to rise at ∼200 Myr in model B100, and at ∼120 Myr in models B30 and B10. Intriguingly, the period of rising vA,reg coincides with the period of rising tdep for each model (see Figure 6). Figure 13(a) plots the gas depletion time as a function of Btot, showing that tdep has a positive correlation with Btot  30 μG, while it is almost independent of Btot  30 μG. When there are strong toroidal magnetic fields, a portion of a ring cannot collapse perpendicularly to the magnetic fields unless it becomes massive enough by gathering mass along the field line. The minimum length that has to collapse along the field to become magnetically supercritical is given by L crit º 2 p B 1 2 G r = R G ring 1 2 F M M ⎛ ⎝ , ⎞ ⎠ ( 31 ) where ΦM/M is the flux-to-mass ratio of the ring (Mestel & Spitzer 1956; Chen & Ostriker 2014). We calculate ΦM/M 14 Figure 10. Temporal histories of the strength of the (a) regular, (b) turbulent, and (c) total magnetic field in the ring for models B100 (blue), B30 (green), and B10 (red), respectively. Magnetic fields are dominated by the turbulent component at early times, but become predominantly regular at late times. In the shaded regions at t < 100 Myr, the rings are not circular so that the decomposition of the magnetic fields into the regular and turbulent components is not meaningful. v A ,reg = ⎛ ⎜ ⎜ ⎝ z =D z ∭ =-D z z ∭ z p 4 B ∣ z =D z =-D z 2 ∣ Q dxdydz r Q dxdydz v A ,trb = ∭ z 4 p ⎛ ⎜ ⎜ ⎝ z =D z =-D z ∭ z d ∣ B z =D z 2 Q ∣ dxdydz r Q dxdydz =-D z 1 2 ⎞ ⎟ ⎟ ⎠ , 1 2 ⎞ ⎟ ⎟ ⎠ , ( 29 ) ( 30 ) where Θ = 1 for T < 2 × 104 K and 0 otherwise Equation (22) and related text for definitions of B∣ B2 d ∣ (see ∣ and ). Note that the integration in the horizontal directions is ∣ The Astrophysical Journal, 946:114 (23pp), 2023 April 1 Moon et al. ultimately collapse because they are magnetically subcritical. We note that at later times, the MJI operates to gather material along the ring circumference. However, when Lcrit exceeds the typical spacing ∼100 pc between the spiral segments formed by the MJI, even the mass gathered by the MJI is not enough to overcome magnetic support. Overall, Figure 13 suggests that large-scale magnetic fields tend to suppress star formation in nuclear rings. 4.3. Vertical Dynamical Equilibrium A framework developed to understand the co-regulation of galactic SFRs and ISM properties is the PRFM theory (Ostriker & Kim 2022; see also Ostriker et al. 2010; Ostriker & Shetty 2011). The theory assumes that the ISM in disk galaxies satisfies vertical dynamical equilibrium between the total midplane pressure and the weight of the overlying gas, and that it is star formation feedback that heats the gas and drives turbulence to maintain the required level of the midplane pressure. the SFR is determined by the requirement for the feedback to yield the pressure needed for vertical dynamical equilibrium, which not only depends on the gas surface density but also on the local stellar density and the velocity dispersion (or gas scale height). Here we check if the vertical dynamical equilibrium holds in the magnetized nuclear the magnetic ring, and assess the relative importance of pressure to the other pressures. In this picture, We measure the thermal, turbulent, and magnetic pressures of the cold–warm medium at the midplane as P thm = ∭ z ∭ z z =D z =-D z =D z z =-D z P dzdxdy Q , Q dzdxdy P trb = P = mag z =D z =-D z z =D z ∭ z ∭ z =-D z z =D z =-D z z =D z =-D z ∭ z ∭ z 2 r v z Q dzdxdy Q dzdxdy T zz Q dzdxdy Q dzdxdy , , ( 32 ) ( 33 ) ( 34 ) Figure 12. Temporal histories of the sound speed cs (pink), the vertical velocity dispersion σz (gray), and the Alfvén speed associated with the regular vA,reg (cyan) and turbulent vA,trb (orange) magnetic fields for models (a) B100, (b) B30, and (c) B10. The shaded region represents the epoch when the ring is not fully circularized. defined f 9 region in Figure ring the B dRdz using inside . Figure 13(b) shows that tdep has an overall M ∬ F = positive correlation with Lcrit, although different models have different tdep at a given Lcrit. Figure 13(c) plots tdep against a dimensionless ratio Lcrit/LJ, where L is the J Jeans length of the cold–warm medium in the ring,9 showing that tdep is approximately constant for Lcrit/LJ  1 but increases with Lcrit/LJ  1. This is because as Lcrit exceeds LJ, more and more Jeans-unstable clumps (smallest ones first) to 2 c G ( s fail p [ 1 2 º )] r 9 For typical ambient cold–warm medium density, we take the volume- averaged density of the cold–warm medium between R = 300 and 700 pc at the midplane, with the density cut nH > 10 cm−3 to exclude the warm, tenuous gas in the inflowing streams. For model B100 before t ∼ 200 Myr, this is ∼100 cm−3, a factor of ∼4 smaller than the mass-weighted mean. ring min = regions between R the 700 pc = where the integration in the horizontal directions is performed over and , and Θ = 1 for T < 2 × 104 K and 0 otherwise. max = R so that Πmag represents the Note that T B ( 4 zz total vertical magnetic stress, including both magnetic pressure (Boulares & Cox 1990; Piontek & and tension terms Ostriker 2007; Kim & Ostriker 2015b). The weight of the ISM is given by 300 pc 2 B z ( 8 - p p ) ) 2  = 1 A ring ∭ z = z L 2 = 0 r ¶F tot ¶ z dzdxdy , ( 35 ) where the horizontal integration is performed over the ring 2 R region as before, and A R ) ( . It follows from max equilibrium, quasi-steady that Equation + P »  if the pressures at the hor- º P P thm mid izontal and the upper boundaries of the cylindrical annulus are small compared to the midplane value. (2) + P trb pº under 2 min mag ring - Figure 14 plots Pmid and  as well as the contributions of each pressure component for magnetized models, showing 15 The Astrophysical Journal, 946:114 (23pp), 2023 April 1 Moon et al. Figure 13. Gas depletion time as a function of (a) the total magnetic field strength in the ring, (b) the magnetic critical length Lcrit = G−1/2Rring(ΦM/M) (see text), and (c) a dimensionless ratio Lcrit/LJ, where L is the Jeans length of the average cold–warm medium in the ring. Blue, green, and red symbols J correspond to models B100, B30, and B10, respectively. The yellow star in panel (a) marks the observed values for the nuclear ring of NGC 1097 (Tabatabaei et al. 2018; Prieto et al. 2019). The vertical dotted line in panel (c) marks Lcrit = LJ. The shaded region represents the range of the depletion time in model Binf for t = 100–200 Myr. 2 c G ( s p [ 1 2 = )] r Pmid »  indeed holds well once the ring enters the quasi- steady state at t ∼ 100 Myr.10 While the midplane pressure is dominated by the turbulent component at early times, the magnetic pressure dominates after t ∼ 240 Myr, ∼140 Myr, and ∼100 Myr for models B100, B30, and B10, respectively. As the ring becomes magnetically supported against the vertical gravity, the demand for the stellar feedback to replenish the thermal and turbulent pressures diminishes, causing the SFR to decline (Figure 6(a)), consistent with the PRFM theory. 4.4. Interpretation of the Field Growth In our simulations, both regular and turbulent fields grow in strength with time, although the latter saturates at ∼35 μG. The rapid growth and saturation of the turbulent magnetic fields are likely due to the SN-driven turbulence, which not only randomly stretches, twists, and folds the field lines to amplify them at small scales (Vaĭnshteĭn & Zel’dovich 1972; Childress & Gilbert 1995), but also tangles the large-scale field lines to create fluctuating components. A number of simulations of the ISM where the turbulence is naturally driven by the SN feedback have demonstrated that the turbulent magnetic fields can be amplified out of very weak seed fields (e.g., Kim & Ostriker 2015b; Rieder & Teyssier 2016, 2017; Butsky et al. 2017; Pakmor et al. 2017; Gent et al. 2021). These studies have found that the growth rate of the turbulent dynamo is sensitive to the numerical resolution because the fastest growth occurs at the smallest resolvable scale, although the saturation amplitude is almost independent of the numerical resolution. Compared to the small-scale dynamo, the large-scale dynamo responsible for the growth of ordered magnetic fields is still poorly understood. In part, the growth of Breg in our simulations is presumably due to the strong differential rotation in the ring, which stretches radial fields into the azimuthal direction. However, the most naive version of the stretching effect is not evident in our simulations. For pure differential from rotation v = vrot = RΩ(R)ef, shown can be it 10 The average midplane pressure including hot (T > 2 × 104 K) gas very well matches the weight for all times, even before t ∼ 100 Myr. For t < 100 Myr, the average midplane pressure of the cold–warm medium is somewhat higher than the weight, indicating the hot gas pressure at those times is slightly smaller than that of the cold–warm medium. 16 Equations (4)and (23), and ∇ · B = 0 that ¶ B R ¶ t = 0, ¶ B f ¶ t = - W q B , R ( 36a ) ( 36b ) d W ln = - m 1 G d R ln are exactly satisfied. Here, q º - is the rate of shear. In our simulations, q = 0.87 and Ω = 0.45 Myr−1 at R = Rring. Taking B as is true for Breg,R at t ∼ 100 R Myr in model B100, Equation (36b) would imply a growth of Bf from zero to 40 μG in 100 Myr, vastly overestimating the true growth of Breg,f shown in Figure 11. In addition, while Equation (36a) would predict Breg,R to be constant in time, Figure 11 shows that the magnitude of Breg,R in fact grows in time. The discrepancies with respect to the prediction of the simple shear model (Equation (36a)) indicate that velocity components other than vrot play an important role in governing the growth of the regular magnetic fields in our simulations. Although a quantitative analysis of is responsible for growth of Breg is beyond the scope of this work, we provide a qualitative account of the mean field growth to motivate future work. One may write a general velocity field by the large-scale dynamo that v = v rot + v blk + d v, ( 37 ) blk v rot º - v where v roughly corresponds to bulk motions of the fluid that deviate from circular rotation (e.g., bubble expansion and radial accretion), and δv is random turbulent motion. Here, v and δv are similarly defined as Equation (22) d = by definition. Substituting Equation (37) into such that v Equation (4) and applying the averaging of Equation (23), one obtains 0 ¶ B ¶ t = - W q B R e f +  ´ ( v blk ´ B ) +  ´ ( d v ´ d B ) . 38 ) In the mean-field dynamo theory for rotating systems, it is thought that the last term in Equation (38) captures the so- called α effect in which the Coriolis force yields systematic ( The Astrophysical Journal, 946:114 (23pp), 2023 April 1 Moon et al. ) · ( d ´ v midplane in combination with the Coriolis force will tend to d produce a kinetic helicity v that changes sign across the midplane (Ruzmaikin et al. 1988), and indeed we find this sign change when we measure kinetic helicity in our simulations. Thus, we can qualitatively understand the growth of Breg,R as due to the combined effect of the midplane change in sign of α with the  ´ midplane change in sign of , to make the last term in ( Equation (38) keep the same sign across the midplane. B R ) 2 As discussed above, the presence of nonzero BR will tend to produce growth of Bf, as expressed by the first term proportional to Ω on the right-hand side of Equation (38). The combination of the two effects discussed above on mean field growth is often referred to as an “α-Ω” dynamo. The h  will, however, tend to suppress B turbulent diffusion term t growth of the mean field. Also, the second term on the right- hand side of Equation (38) could in the case of our simulations represent advection of mean magnetic fields out of the ring due to the radial accretion flow, and/or dilution due to the expansion of SN remnants. Together with the turbulent diffusion, these bulk flows may account for the reduction in field growth compared to a pure α–Ω dynamo. We note that the regular magnetic fields grow faster when σz starts to decline (see Figures 11 and 12(a)), which is presumably due to the reduction of ηt with decreasing velocity dispersion (Gressel et al. 2008). One important aspect of the α effect is that while it creates large-scale magnetic helicity associated with the poloidal loops, it does so at the expense of a small-scale magnetic helicity of the opposite sign, which is associated with the internal twist of the poloidal loops (see Blackman & Brandenburg 2003, for a visual illustration), in a way that satisfies magnetic helicity conservation. As small-scale magnetic helicity (or “twists”) accumulates over time, the magnetic tension resists bending and twisting of the field lines to quench the α effect, potentially limiting the growth of regular magnetic fields (see Section 9 of Brandenburg & Subramanian 2005). Shukurov et al. (2006) showed that galactic fountain flows can transport the small- scale helicity out of the disk in the vertical direction, maintaining the dynamo against the back-reaction from the Lorentz force. In our simulations, clustered SN explosions are powerful enough to drive outflows that drag magnetic fields away from the midplane and to leave the computational domain (see Figure 4; see also Figure 8 of Paper I). Even when vertical outflows become weak as the growing magnetic fields suppress the SFR, the radial accretion flows may still be able to transport helicity out of the ring. 4.5. Resolution Dependence To see how our results depend on numerical resolution, we rerun our fiducial model B100 at lower resolution, using 2563 cells corresponding to Δx = 8 pc. Figure 15(a) compares the temporal evolution of the field strength between the 5123 and 2563 runs. Evidently, growth of the turbulent field Btrb is initially higher the higher-resolution model, and the superlinear growth stage for the regular field Breg also occurs earlier in time for the higher-resolution model. The initially faster growth of Btrb is presumably because the growth rate of the small-scale turbulent dynamo is inversely proportional to the eddy turnover time at the grid scale, as noted by Rieder & Teyssier (2016). The stronger turbulent fields at earlier times in the 5123 run also presumably lead to earlier superlinear growth for 17 Figure 14. Temporal histories of the total midplane pressure (orange) and gas weight (red), as well as the thermal (pink), turbulent (gray), and magnetic (cyan) components of the midplane pressure, for models (a) B100, (b) B30, and (c) B10. After the ring enters the steady state at t ∼ 100 Myr, Pmid »  , indicating that the vertical dynamical equilibrium holds very well. The shaded region represents the epoch when the ring is not fully circularized. twists in the field lines to produce large-scale magnetic fields, as envisaged by Parker (1955), as well as turbulent diffusion of the mean magnetic fields (Brandenburg & Subramanian 2005). For example, by assuming weak Lorentz force and isotropic the approximation turbulence, smoothing 2 ´  d ´ ´ +  h B v B ( yields , with the trans- t coefficients th » ´ d  d v v and port · ( ) 3 , where τcor is the correlation time of turbulence t )d v3 ( cor (Brandenburg & Subramanian 2005). If Breg,f has a single sign and decreases in magnitude away from the midplane, as is true in B ´ will change sign our simulations, the radial component of across the midplane. Expansion of bubbles centered on the first-order a  = )d B » - t a ( cor ) 2 The Astrophysical Journal, 946:114 (23pp), 2023 April 1 Moon et al. Figure 16. Evolution of the regular (brown) and turbulent (cyan) magnetic field strengths in the βin = 100 model with an Alfvén speed ceiling applied in order to extend the simulation run time beyond that of model B100. Thick and thin lines correspond to the runs with 5123 and 2563 cells, respectively. The 5123 run is restarted from model B100 (light-colored) at t = 266 Myr marked by the vertical dashed line, while the 2563 run has an Alfvén ceiling applied throughout its evolution. 5. Summary and Discussion 5.1. Summary Nuclear rings at the centers of barred galaxies are active in star formation (Mazzuca et al. 2008; Ma et al. 2018) and threaded by magnetic fields with a mean strength of ∼50–100 μG (Beck et al. 2005; Yang et al. 2022). To study how magnetic fields affect star formation in nuclear rings, we run MHD simulations of galactic centers. We employ the semiglobal models of Paper I, in which magnetized gas streams from two nozzles at the boundaries supply gas and magnetic fields, mimicking bar-driven gas inflows along dust lanes. We the modified TIGRESS framework (Kim & Ostriker adopt 2017) to model star formation and related FUV and SN feedback, as well as the shielding of FUV radiation and CR heating in dense environments. We fix the mass inflow rate to  - and the ring size to Rring = 500 pc, while 1 M in varying the plasma parameter βin = 10, 30, and 100 to adjust the to Bin,avg = 2.4 μG, 1.4 μG, and 0.76 μG. We also run a hydrodynamic model with unmagnetized streams for compar- ison. The magnetic fields in the streams are set parallel to the inflow velocity, motivated by observations (Beck et al. 2005; Lopez-Rodriguez et al. 2021). average field strength within the gas streams M1  yr = The main results of this work can be summarized as follows: 1. Overall Evolution: The two gas streams injected from the domain boundaries at the opposite sides collide with each other after about half an orbital time, dissipating their orbital kinetic energy via shocks. As the gas orbits gradually circularize, a well-defined nuclear ring forms at the radius where the specific angular momentum of the inflowing gas matches that of the circular orbit. At about t ∼ 100 Myr, the nuclear ring reaches a quasi-steady state in which the shape, SFR, and gas mass become approximately constant with time. Stars form randomly across the whole circumference of the ring, and the associated feedback renders the ring turbulent. When the magnetic fields in the ring become strong enough, they Figure 15. Resolution study. We compare the 5123 (Δx = 4 pc; thick) and 2563 (Δx = 8 pc; thin) runs with βin = 100 for evolution of (a) the regular (brown) and turbulent (cyan) magnetic field strengths, (b) the measured (M ; acc blue solid lines) and predicted (M ;M gold dashed lines) mass accretion rates at R = 100 pc, and (c) the depletion time tdep.  of Breg. We note that the growth rate of Breg at the time when Btrb saturates is similar in both models. The saturated field strength of Btrb is similar in both models (see also Figure 16 and related discussion of saturation). Because magnetic field growth is delayed in the low- resolution model, other characteristic evolutionary effects affected by magnetic fields (see Section 3.1) also occur later in time. For example, Figures 15(b) and (c) show that the mass accretion rate and depletion time start to increase at t ∼ 200 Myr and ∼400 Myr in the 5123 and 2563 runs, respectively, corresponding to the start of the superlinear growth of Breg. We note that even though the mass accretion rate at a given time depends on numerical resolution, it is entirely consistent with the predicted accretion rate from the instantaneous Maxwell stress. 18 The Astrophysical Journal, 946:114 (23pp), 2023 April 1 Moon et al. reduce the ring SFR. At the same time, strong magnetic torques lead to accretion flows from the ring to the galaxy center (Figures 1 and 3), where a CND grows. Due to the action of MJI combined with strong shear, at late stages, the ring forms transient trailing spiral segments, some of which undergo star formation. 2. Magnetic Fields and Their Growth: Magnetic fields in the ring can be separated into a regular and a turbulent component, where the former is defined via azimuthal averaging, and the latter is the azimuthally fluctuating residual. The turbulent component grows in strength over time and saturates at ∼30–40 μG independent of βin, likely due to the SN-driven turbulent dynamo. In contrast, the regular component does not saturate but keeps growing with time, reaching 50–70 μG at the end of the runs (Figure 10). While the turbulent fields are approxi- mately isotropic, the regular fields are dominated by the azimuthal component with a pitch angle of θp ∼ 6°–12°. The overall field direction is mostly toroidal near the midplane, but expansion of superbubbles created by clustered SNe drag the toroidal fields to produce poloidal fields in high-altitude regions (Figures 4 and 5). 3. Magnetically Driven Accretion: All of our magnetized models develop accretion flows that slowly fill the region interior to the ring and eventually form a CND with radius 50 pc at the center. This is in stark contrast to the unmagnetized model where the region interior to the ring is always filled with hot, rarefied gas. The gas accretion rates measured in the simulations are consistent with the theoretical quasi-steady rates due to the Maxwell stress, indicating that the radial accretion is driven by magnetic tension. The measured accretion rate depends on the galactocentric radius and reaches ∼(0.02–0.1)Me yr−1 at late times (Figures 7 and 8). 4. Effects of Magnetic Fields on Star Formation: When strong magnetic fields develop in the ring, they suppress star formation therein. Consequently, the gas depletion time tdep in the rings increases with the total field strength. In particular, strong regular azimuthal magnetic fields in the ring limit the radial and vertical compression that lead to collapse, unless the ring undergoes sufficient azimuthal contraction that could gather material along the field lines (Figure 13). The ring maintains vertical dynamical equilibrium instantaneously, meaning that the weight of the ISM is balanced by the midplane total pressure (Figure 14). While the magnetic pressure is negligible in the vertical force balance at early times, it becomes dominant at late times (t  240 Myr in model B100). This late-time strong magnetic support, which is mostly from the regular (nonturbulent) component of the magnetic fields, reduces the demand for SN feedback to replenish the thermal and turbulent pressures, thereby indirectly lowering the SFR, consistent with the PRFM theory of Ostriker & Kim (2022). 5.2. Discussion Infrared polarization observations indicate that Galactic magnetic fields are preferentially torodial near the CMZ and in the regions with Galactic latitudes |b| > 0.4° poloidal (Nishiyama et al. 2010). This toroidal-to-poloidal transition of the Galactic magnetic fields is consistent with our numerical 19 results that expanding superbubbles drag the toroidal fields in the rings to produce poloidal magnetic fields in high-altitude regions (Figures 4 and 5). The poloidal magnetic walls of venting superbubbles are likely illuminated by relativistic particles accelerated in situ at SN shocks, potentially creating some nonthermal radio filaments, such as Radio Arc and Sgr C filaments, found near radio bubbles (e.g., Heywood et al. 2022). For filaments without any evident source, Barkov & Lyutikov (2019) proposed that transiting pulsar wind nebulae may inject relativistic particles to make the background poloidal magnetic fields visible locally. Alternatively, Sofue (2023) proposed that represent projected wavefronts of fast MHD waves launched from SNe exploding in the ring, which locally compress the existing magnetic fields. The results of our simulations suggest that the background volume-filling magnetic fields necessary in both scenarios can be produced by superbubbles breaking out of the ring. the filaments 0.02  ~ acc The results of our simulations show that the magnetic torque produces significant inflows of gas from a nuclear ring toward the center, forming a CND. The mass accretion rate depends for model B100 at both on time and radius such that –0.03 Me yr−1 near R ∼ 50 pc while t = 300 Myr, M ∼0.1 Me yr−1 near R ∼ 500 pc. Figure 8 suggests that the radial profile of the mass accretion rate becomes flatter over time, in a way that the accretion rate near the CND increases with time. At t = 300 Myr, the CND formed in model B100 has gas mass 4 × 105 Me, which is an order of magnitude smaller than the observed CND masses ∼107 Me in nearby active galaxies (Combes et al. 2019). We note, however, that the continued mass accretion with a rate of ∼0.01–0.1 Me yr−1 is capable of producing 107 Me CNDs within 0.1–1 Gyr. We also note that, because the magnetically driven mass accretion rate is + proportional to B B , the accre- f tion rate would be higher if the magnetic fields are more loosely wrapped (i.e., larger pitch angle). 2 1 2 B ) f Tabatabaei et al. (2018) measured the magnetic field strength in 11 giant clumps in the nuclear ring of NGC 1097 and found a negative correlation between the star formation efficiency and magnetic field strength, suggesting that magnetic fields are inhibiting star formation in the nuclear ring. Figure 13(a) compares the observed average magnetic field strength of 62 μG (Tabatabaei et al. 2018) and depletion time of 5 × 108 yr (Prieto et al. 2019) with the results of our simulations. Based on our numerical results, the nuclear ring in NGC 1097 is in the regime where magnetic fields are dynamically important to suppression of star formation. ( sin 2 2 B ( R )q p = - 2 R Our simulations demonstrate the intriguing possibility of star rings together with mass formation quenching in nuclear accretion to the center, both resulting from the growth of large- scale regular magnetic fields. However, it is still uncertain how the field strength and pitch angles depend on the underlying rotation curve and feedback physics as well as the magnetiza- tion of the inflowing gas, which models a bar-driven stream. These issues may be resolved by running more realistic simulations with improved physics. First of all, while the present work assumes that magnetic fields in the gas streams are kept constant, in real galaxies, both the field strength and direction will vary with time, resulting in significant changes in the field strength and structure in nuclear rings and inward. In particular, the magnetic fields in the inflowing stream may change their polarity over a dynamical time at the bar end region and then be amplified in the ring. Reconnection with The Astrophysical Journal, 946:114 (23pp), 2023 April 1 existing fields of the opposite polarity in the ring would then prevent excessive growth of the magnetic fields. To address this issue, it will be necessary to run global simulations of barred galaxies in which gas streams along dust lanes are modeled self- consistently by nonlinear interactions with a bar potential. As Figure 10 shows, the regular magnetic fields Breg in our simulations are still growing at the end of the runs, even though the turbulent component of the magnetic field has saturated. It is, however, of great interest to determine how Breg would evolve over longer timescales. For our standard simulations, following this late- time evolution is precluded by an extremely short time step, as limited by the large Alfvén speed in locations where the density becomes low.11 However, by applying an artificial density floor to put a ceiling on the Alfvén speed of 103 km s−1, we are able to restart our fiducial model B100 from t = 266 Myr and run up to t = 600 Myr, as shown in Figure 16. This figure shows the magnetic fields saturate at Breg ∼ 100 μG around that t ∼ 520 Myr. Additionally, we find that for a corresponding lower-resolution run, the saturation level of Breg is the same, although saturation occurs at a later time (see Section 4.5). It is difficult to pinpoint what physical effect terminates the growth of Breg, but radial spreading of the ring, nonlinear quenching due to small-scale helicity, and turbulent diffusion may all play a role (e.g., Brandenburg & Ntormousi 2022). Our semiglobal framework allows us to afford a relatively high resolution of Δx = 4 pc uniformly across the entire domain such that the expansion of individual SN remnants is captured self- consistently. We note, however, that since the freefall time of star- forming clumps in the ring is shorter than the SN delay time of ∼4 Myr, early feedback in the form of radiation and stellar winds formation efficiency of may affect gas dynamics and star Moon et al. individual star-forming clumps significantly in extreme environ- ments like nuclear rings. As a final caveat, we note that considering the low ionization fraction in the densest star-forming clumps in nuclear rings, nonideal terms in the induction equation that we have neglected in this study might potentially affect the growth of magnetic fields in the ring and therefore star formation and mass accretion flows. funded by the Korean government We are grateful to the referee for an insightful report. The work of S.M. was supported by an NRF (National Research Foundation of Korea) grant (NRF- 2017H1A2A1043558-Fostering Core Leaders of the Future Basic Science Program/Global Ph. D. Fellowship Program). The work of W.-T.K. was supported by the grant of National Research Foundation of Korea (2022R1A2C1004810). The work of C.-G.K. was supported in part by NASA ATP grant No. 80NSSC22K0717. The work of ECO is partly supported by the Simons Foundation under grant 510940. Computational resources for this project were provided by Princeton Research Computing, a consortium including PICSciE and OIT at Princeton University, and by the Supercomputing Center/Korea Institute of Science and Technol- ogy Information with supercomputing resources including techni- cal support (KSC-2021-CRE-0025). Software: Athena (Stone et al. 2008), VisIt (Department Of Energy (DOE) Advanced Simulation & Computing Initiative (ASCI), 2011). Appendix A Seed Magnetic Fields In the Athena code, magnetic fields are face-centered and are updated by the edge-centered electromotive force (∝ v × B) Figure 17. Illustration of the magnetic fields and velocity vectors in the ghost cells (shaded area) that belong to the nozzle at the positive y-boundary; the white area corresponds to the adjacent active cells. The blue arrows represent the velocity vectors, defined at cell centers. The solid and dashed arrows in red indicate the face- and cell-centered magnetic field vectors, respectively, with the latter computed from the former. Filled and open circles mark the ghost and active zones, respectively. See Appendix A for details. 20 The Astrophysical Journal, 946:114 (23pp), 2023 April 1 using the constrained transport algorithm. Because our initial conditions have zero magnetic field at every active face, the only way to inject magnetic fields into the domain is by having nonzero electromotive forces at the boundaries. Even if Bin is set parallel to vin for gas streams at the boundaries in our simulations, the fact that Athena defines the magnetic fields and velocity at the face centers and cell centers, respectively, enables vin × Bin ≠ 0 at the domain boundaries, making the electromotive force nonzero and inducing seed magnetic fields in the active zones adjacent to the nozzles for an initial brief period of time. To illustrate this, Figure 17 displays a part of the the positive y-boundary. The computational domain near yellow shaded area indicates the ghost cells belonging to the upper nozzle, and the white area marks the adjacent active cells. The blue arrows at the cell centers represent (vx, vy) of the stream at t = 0, which are nonzero only in the ghost cells and zero in the active cells. The red solid arrows at the cell faces indicate (Bx, By) of the stream at t = 0, which are also nonzero only in the ghost faces and zero in the active faces, while the red dashed arrows represent the cell-centered magnetic fields computed by averaging the neighboring face-centered fields. We note that the cell faces corresponding to the domain boundary are active (i.e., By at the border between the yellow and white areas is updated via Equations (1)–(5)). Because By = 0 initially at those boundary faces, the cell-centered By at the first ghost cells (adjacent to the active domain) is reduced by half, resulting in Bin inclined to vin, the latter of which is set to Equation (14). As a result, nonvanishing electromotive forces are assigned to the edges of the outermost active cells, which subsequently induce magnetic fields into the computational domain. We stress that this process occurs only for a few megayears in the very beginning: By at the outermost active faces soon attains the same values as in the ghost faces to satisfy Equation (17). We also note that our initial conditions for the gas streams obey ∇ · B = 0 in the active domain, which is preserved by the constrained transport algorithm in Athena. Appendix B Magnetic Energy Conservation Here we consider the role of advection into the domain in the growth of magnetic energy in our simulations. We start with the equation for the rate of change of the total magnetic energy in the computational domain dE mag dt = = ò 1 p 4 ¶ ¶ t 2 B p 8 ⎜ ⎛ ⎝ ⎟ ⎞ ⎠ dV B · [  ´ ´ v ( B )] dV . ( B1 ) ò Integrating Equation (B1) by parts and applying the divergence theorem, one obtains dE mag dt = 1 p 4 ∮ [( v ´ B ) ´ B ] · d A - 1 p 4 ò v · [(  ´ B ) ´ B ] dV , ( B2 ) 11 Large Alfvén speeds in our simulations occur in cones surrounding the z- axis above and below the CND, which have very low density but moderate magnetic field strength (see the right column of Figure 9). Moon et al. where dA denotes the area element. The first term on the right- hand side of Equation (B2) represents the Poynting flux integrated over the domain boundaries, while the second term is the amount of work done by the fluid against the Lorentz force per unit time. It is evident that the first term vanishes when v∥B: there is no magnetic energy flux through the boundaries as long as the magnetic fields in the streams are parallel to the streaming velocity. One can further expand the cross products in the first term to write dE mag dt = - ∮ v · · d A - ∮ 2 B p 8 v · d A - 1 p 4 ò v · [(  ´ B ) ´ B ] dV . ( B3 ) In this form, the first and second terms correspond to the work done by the Maxwell stress  º at the boundaries and the advection of magnetic energy by the inflowing gas, respectively. Again, the two terms exactly cancel each other when v∥B. 2 ( 8 BB ( 4 - B p p  ) ) As explained in Appendix A, v is not parallel to B at the domain boundaries for the initial ∼10 Myr, in which case the advection term is not offset by the Maxwell stress term, resulting in the growth of Emag. One can estimate the maximum rate of the magnetic energy growth due to advection alone as dE mag,adv dt = - ∮ nozzles 2 B in p 8 v in · d A = - z 2 b m k T B in ò m in H H 0 in r v in in · ˆ y 2 cos pz z 2 in ⎛ ⎜ ⎝ ⎞ ⎟ ⎠ ´ 2 pz z d » 0.6  k T M in B in b m m in H H , ( B4 ) than the actual magnetic In contrast, where Equation (18) is used. For the parameters of model B100, dEmag,adv/dt ∼ 1.4 × 1049 erg Myr−1. This suggests that the total magnetic energy that would be advected (barring the work done by the Maxwell stress) into the computational domain is 1.4 × 1050 erg for the initial 10 Myr, which is a energy factor of 3 smaller Emag = 4.3 × 1050 erg at t = 10 Myr. the total magnetic energy advected into the computational domain would be 4.2 × 1051 erg at the end of the run (t = 300 Myr), which is than Emag = 7.2 × 1053 erg at the same epoch. Considering the work done by the Maxwell stress, which tends to offset the the above result magnetic energy growth by advection, suggests that the actual magnetic energy advected through the nozzles should be negligible compared to what is generated by the last term in Equation (B3) via an MHD dynamo. We conclude that while the inflow nozzles provide seed magnetic fields, it is growth via dynamo activity rather than advection into the domain that is responsible for the level of the magnetic energy at late times. about 2 orders of magnitude smaller Appendix C Mass Accretion Rates due to Maxwell and Reynolds Stresses For gas to move radially inward while moving on an approximately circular orbit, it must lose angular momentum 21 The Astrophysical Journal, 946:114 (23pp), 2023 April 1 slowly. Here we derive the theoretical accretion rates due to the Maxwell and Reynolds stresses. Multiplying the azimuthal component of Equation (2) by R yields ¶ ( r Rv f ) ¶ t +  · ( r Rv f v + e RP f ) =- e R f · (  ·  ) - W R 2 p r v R - r ¶F tot f ¶ , ( C1 ) where ef is the unit vector in the azimuthal direction. The second and third terms on the left-hand side of Equation (C1) are the angular momentum flux due to macroscopic bulk fluid motion and the microscopic thermal motion of its constituent particles, respectively. The three source terms on the right-hand side are the torque density due to the Lorentz force, the Coriolis force, and gravity, respectively. To focus on the radial mass inflow, we azimuthally average Equation (C1). Integrating the resulting equation in the vertical direction assuming the flux through the vertical boundaries is negligible, one obtains 2 r R v v f R ñ ¶ R ñ f - W á R p 2 r v R ñ , ( C2 ) ¶á 1 R 2 R T R ¶ R ¶á r Rv f ñ ¶ t + =- ¶á 1 R - 1 ∬ )p tot r f self ( 2 á ñ º f Xd dz where X for any physical quantity X, and TRf = − BRBf/(4π) is the R–f component of the Maxwell ¶ ñ » unless á ¶F ¶ ñ = á ¶F f r stress tensor. Note 0 there is a systematic azimuthal offset between the density and self-gravitational potential. The magnetic torque term is due to the magnetic tension alone. We decompose the velocity field into ordered and random components: v = vcircef + u, where vcirc ≡ vrot − RΩp is the circular velocity in the rotating reference frame and u is the random velocity. Substituting u for v and using the continuity Equation (1), Equation (C2) becomes ¶á r Ru ñ f ¶ t =- 1 R - ¶á  M acc p R 2 2 R T R ¶ R ¶ ( ) Rv circ ¶ R + 1 R ¶á 2 r R u u R ñ f ¶ R f ñ - W á R p 2 r u R ñ , ( C3 ) 2 acc º - á p r R u R  ñ is the mass accretion rate at radius where M R. Assuming a quasi-steady state (which turns out to be the case in our simulations) and neglecting the Coriolis term, which is unimportant for small R, Equation (C3) is simplified to  M acc »  M M +  M , R ( C4 ) and MR where MM are the mass accretion rates due to the Maxwell and Reynolds stress, defined in Equation (21a). Figures 7 and 8 show that the mass accretion in our simulations is dominated by MM , that is, magnetic tension. ORCID iDs https://orcid.org/0000-0002-6302-0485 Sanghyuk Moon Woong-Tae Kim https://orcid.org/0000-0003-4625-229X Chang-Goo Kim https://orcid.org/0000-0003-2896-3725 https://orcid.org/0000-0002-0509-9113 Eve C. Ostriker 22 References Moon et al. Allard, E. L., Knapen, J. H., Peletier, R. F., & Sarzi, M. 2006, MNRAS, 371, 1087 E. M., & Armillotta, L., Krumholz, M. R., Di Teodoro, McClure-Griffiths, N. M. 2019, MNRAS, 490, 4401 Athanassoula, E. 2005, MNRAS, 358, 1477 Barkov, M. V., & Lyutikov, M. 2019, MNRAS, 489, L28 Beck, R. 2015, A&ARv, 24, 4 Beck, R., Ehle, M., Shoutenkov, V., Shukurov, A., & Sokoloff, D. 1999, Natur, 397, 324 Beck, R., Fletcher, A., Shukurov, A., et al. 2005, A&A, 444, 739 Benedict, G. F., Smith, B. J., & Kenney, J. D. P. 1996, AJ, 112, 1318 Bittner, A., Sánchez-Blázquez, P., Gadotti, D. A., et al. 2020, A&A, 643, A65 Blackman, E. G., & Brandenburg, A. 2003, ApJL, 584, L99 Boris, J. P. 1970, in Proc. 4th Conf. on Numerical Simulation of Plasmas, ed. J. P. Boris & R. A. Shanny (Washington, DC: Naval Research Laboratory), 3 Boulares, A., & Cox, D. P. 1990, ApJ, 365, 544 Brandenburg, A., & Ntormousi, E. 2022, arXiv:2211.03476 Brandenburg, A., & Subramanian, K. 2005, PhR, 417, 1 Butsky, I., Zrake, J., Kim, J.-h., Yang, H.-I., & Abel, T. 2017, ApJ, 843, 113 Chen, C.-Y., & Ostriker, E. C. 2014, ApJ, 785, 69 Childress, S., & Gilbert, A. D. 1995, Stretch, Twist, Fold: The Fast Dynamo (Berlin: Springer) Combes, F., García-Burillo, S., Audibert, A., et al. 2019, A&A, 623, A79 Comerón, S., Knapen, J. H., Beckman, J. E., et al. 2010, MNRAS, 402, 2462 de Sá-Freitas, C., Fragkoudi, F., Gadotti, D. A., et al. 2022, arXiv:2211.07670 Department Of Energy (DOE), Advanced Simulation and Computing Initiative Interactive Parallel Visualization and Graphical (ASCI) 2011, VisIt: Analysis Tool, Astrophysics Source Code Library, record ascl:1103.007 Draine, B. T. 1978, ApJS, 36, 595 Elmegreen, B. G. 1987, ApJ, 312, 626 Elmegreen, B. G., Galliano, E., & Alloin, D. 2009, ApJ, 703, 1297 Gadotti, D. A., Bittner, A., Falcón-Barroso, J., et al. 2020, A&A, 643, A14 Gadotti, D. A., Sánchez-Blázquez, P., Falcón-Barroso, J., et al. 2019, MNRAS, 482, 506 Gent, F. A., Mac Low, M.-M., Käpylä, M. J., & Singh, N. K. 2021, ApJL, 910, L15 Glassgold, A. E., Galli, D., & Padovani, M. 2012, ApJ, 756, 157 Gong, M., Ostriker, E. C., & Wolfire, M. G. 2017, ApJ, 843, 38 Gressel, O., Elstner, D., Ziegler, U., & Rüdiger, G. 2008, A&A, 486, L35 Hennebelle, P., & Inutsuka, S.-i. 2019, FrASS, 6, 5 Heywood, I., Rammala, I., Camilo, F., et al. 2022, ApJ, 925, 165 Indriolo, N., Geballe, T. R., Oka, T., & McCall, B. J. 2007, ApJ, 671, 1736 Kim, C.-G., Kim, W.-T., & Ostriker, E. C. 2008, ApJ, 681, 1148 Kim, C.-G., & Ostriker, E. C. 2015a, ApJ, 802, 99 Kim, C.-G., & Ostriker, E. C. 2015b, ApJ, 815, 67 Kim, C.-G., & Ostriker, E. C. 2017, ApJ, 846, 133 Kim, J.-G., Kim, W.-T., Ostriker, E. C., & Skinner, M. A. 2017, ApJ, 851, 93 Kim, J.-G., Ostriker, E. C., & Filippova, N. 2021, ApJ, 911, 128 Kim, W.-T., & Elmegreen, B. G. 2017, ApJL, 841, L4 Kim, W.-T., Kim, C.-G., & Ostriker, E. C. 2020, ApJ, 898, 35 Kim, W.-T., & Ostriker, E. C. 2001, ApJ, 559, 70 Kim, W.-T., Ostriker, E. C., & Stone, J. M. 2002, ApJ, 581, 1080 Koyama, H., & Inutsuka, S.-i. 2002, ApJL, 564, L97 Launhardt, R., Zylka, R., & Mezger, P. G. 2002, A&A, 384, 112 Lemaster, M. N., & Stone, J. M. 2009, ApJ, 691, 1092 Lopez-Rodriguez, E., Beck, R., Clark, S. E., et al. 2021, ApJ, 923, 150 Ma, C., de Grijs, R., & Ho, L. C. 2018, ApJ, 857, 116 Maciejewski, W. 2004, MNRAS, 354, 892 Mazzuca, L. M., Knapen, J. H., Veilleux, S., & Regan, M. W. 2008, ApJS, 174, 337 McKee, C. F., & Zweibel, E. G. 1995, ApJ, 440, 686 Meier, D. S., Turner, J. L., & Hurt, R. L. 2008, ApJ, 675, 281 Mestel, L., & Spitzer, L. J. 1956, MNRAS, 116, 503 Moon, S., Kim, W.-T., Kim, C.-G., & Ostriker, E. C. 2021, ApJ, 914, 9 Moon, S., Kim, W.-T., Kim, C.-G., & Ostriker, E. C. 2022, ApJ, 925, 99 Neufeld, D. A., & Wolfire, M. G. 2017, ApJ, 845, 163 Nishiyama, S., Hatano, H., Tamura, M., et al. 2010, ApJL, 722, L23 Nogueras-Lara, F., Schödel, R., Gallego-Calvente, A. T., et al. 2020, NatAs, 4, 377 Onishi, K., Iguchi, S., Sheth, K., & Kohno, K. 2015, ApJ, 806, 39 Ostriker, E. C., & Kim, C.-G. 2022, ApJ, 936, 137 Ostriker, E. C., McKee, C. F., & Leroy, A. K. 2010, ApJ, 721, 975 The Astrophysical Journal, 946:114 (23pp), 2023 April 1 Moon et al. Skinner, M. A., & Ostriker, E. C. 2015, ApJ, 809, 187 Sofue, Y. 2023, MNRAS, 518, 6273 Sormani, M. C., & Barnes, A. T. 2019, MNRAS, 484, 1213 Sormani, M. C., Sanders, J. L., Fritz, T. K., et al. 2022, MNRAS, 512, 1857 Sormani, M. C., Tress, R. G., Glover, S. C. O., et al. 2020, MNRAS, 497, 5024 Stone, J. M., & Gardiner, T. 2009, NewA, 14, 139 Stone, J. M., Gardiner, T. A., Teuben, P., Hawley, J. F., & Simon, J. B. 2008, ApJS, 178, 137 Sutherland, R. S., & Dopita, M. A. 1993, ApJS, 88, 253 Tabatabaei, F. S., Minguez, P., Prieto, M. A., & Fernández-Ontiveros, J. A. 2018, NatAs, 2, 83 Tress, R. G., Sormani, M. C., Glover, S. C. O., et al. 2020, MNRAS, 499, 4455 Vaĭnshteĭn, S. I., & Zel’dovich, Y. B. 1972, SvPhU, 15, 159 Wada, K. 2004, in Coevolution of Black Holes and Galaxies, ed. L. C. Ho (Cambridge: Cambridge Univ. Press), 186 Yang, Y., Irwin, J., Li, J., et al. 2022, ApJ, 927, 4 Ostriker, E. C., & Shetty, R. 2011, ApJ, 731, 41 Pakmor, R., Gómez, F. A., Grand, R. J. J., et al. 2017, MNRAS, 469, 3185 Parker, E. N. 1955, ApJ, 122, 293 Pillai, T., Kauffmann, J., Tan, J. C., et al. 2015, ApJ, 799, 74 Piontek, R. A., & Ostriker, E. C. 2007, ApJ, 663, 183 Prieto, M. A., Fernandez-Ontiveros, J. A., Bruzual, G., et al. 2019, MNRAS, 485, 3264 Regan, M. W., Vogel, S. N., & Teuben, P. J. 1997, ApJL, 482, L143 Rieder, M., & Teyssier, R. 2016, MNRAS, 457, 1722 Rieder, M., & Teyssier, R. 2017, MNRAS, 471, 2674 Ruzmaikin, A., Sokolov, D., & Shukurov, A. 1988, Natur, 336, 341 Sarzi, M., Allard, E. L., Knapen, J. H., & Mazzuca, L. M. 2007, MNRAS, 380, 949 Seo, W.-Y., Kim, W.-T., Kwak, S., et al. 2019, ApJ, 872, 5 Shimizu, T. T., Davies, R. I., Lutz, D., et al. 2019, MNRAS, 490, 5860 Shlosman, I., Frank, J., & Begelman, M. C. 1989, Natur, 338, 45 Shukurov, A., Sokoloff, D., Subramanian, K., & Brandenburg, A. 2006, A&A, 448, L33 23
10.1016_j.yjsbx.2023.100085
Contents lists available at ScienceDirect Journal of Structural Biology: X journal homepage: www.sciencedirect.com/journal/journal-of-structural-biology-x Measuring the effects of ice thickness on resolution in single particle cryo-EM Kasahun Neselu a, Bing Wang b, William J. Rice b, c, Clinton S. Potter a, Bridget Carragher a,*, Eugene Y.D. Chua a, * a Simons Electron Microscopy Center, New York Structural Biology Center, New York, NY, USA b Cryo-Electron Microscopy Core, New York University Grossman School of Medicine, New York, NY, USA c Department of Cell Biology, New York University Grossman School of Medicine, New York, NY, USA A R T I C L E I N F O A B S T R A C T Keywords: Cryo-EM Ice thickness Single particle analysis Energy filter High tension Resolution Ice thickness is a critical parameter in single particle cryo-EM – too thin ice can break during imaging or exclude the sample of interest, while ice that is too thick contributes to more inelastic scattering that precludes obtaining high resolution reconstructions. Here we present the practical effects of ice thickness on resolution, and the influence of energy filters, accelerating voltage, or detector mode. We collected apoferritin data with a wide range of ice thicknesses on three microscopes with different instrumentation and settings. We show that on a 300 kV microscope, using a 20 eV energy filter slit has a greater effect on improving resolution in thicker ice; that operating at 300 kV instead of 200 kV accelerating voltage provides significant resolution improvements at an ice thickness above 150 nm; and that on a 200 kV microscope using a detector operating in super resolution mode enables good reconstructions for up to 200 nm ice thickness, while collecting in counting instead of linear mode leads to improvements in resolution for ice of 50–150 nm thickness. Our findings can serve as a guide for users seeking to optimize data collection or sample preparation routines for both single particle and in situ cryo-EM. We note that most in situ data collection is done on samples in a range of ice thickness above 150 nm so these results may be especially relevant to that community. Introduction The goal of sample preparation for single particle cryo-electron mi- croscopy (cryo-EM) is to capture the sample in optimal conditions on a cryo-EM grid. “Optimal conditions” means the biological sample is embedded in vitreous ice suspended over holes in the grid foil, has enough well-distributed particles in different orientations, and that the sample is found in ice that is as thin as possible, typically 10–100 nm (Noble et al., 2018). While the thinnest possible ice might be expected to yield the highest resolution reconstructions, there is usually a “Goldi- locks” zone for ice thickness for each sample (Olek et al., 2022). If the ice is too thin, the sample can be excluded from the holes, adopt a preferred orientation, or break during imaging. On the other hand if the ice is too thick, increased inelastic scattering from the additional ice may nega- tively affect reconstruction resolutions (Wu et al., 2016). In most cases, the thinnest possible ice that yields good particles is desirable for data collection. This ideal ice thickness depends on the sample, and can range from 15 nm for apoferritin (12 nm in diameter) (Brown & Hanssen, 2022) to 750 nm for the Giant Mimivirus (500 nm in diameter) (Xiao et al., 2005). Quite often, however, ice much thicker than the diameter of the particle is required to avoid particles adopting a preferred orientation (e.g. Huntington et al., 2022). Although ice thickness is an important parameter both for the sample integrity and optimal data collection, it is not currently possible to finely control ice thicknesses during cryo-EM sample preparation. With commonly-used plunge freezers, or even with modern automated sam- ple preparation devices such as the chameleon (Darrow et al., 2019, 2021), ice thicknesses often vary both within a grid square and across the grid. Some areas of a grid may have good particle distribution and ideal ice thickness while others may have too thin ice which excludes particles, or too thick ice that has reduced contrast. Problems of variations in ice thickness on a grid can be solved in several ways. First, by setting automated data collection parameters to only collect on the desired ice thicknesses (Brown & Hanssen, 2022; Cheng et al., 2021; Rheinberger et al., 2021). Collecting good quality data by skipping over targets with too thin or too thick ice is important * Corresponding authors. E-mail addresses: bcarr@nysbc.org (B. Carragher), echua@nysbc.org (E.Y.D. Chua). https://doi.org/10.1016/j.yjsbx.2023.100085 Received 29 November 2022; Received in revised form 10 January 2023; Accepted 23 January 2023 JournalofStructuralBiology:X7(2023)100085Availableonline24January20232590-1524/©2023TheAuthors.PublishedbyElsevierInc.ThisisanopenaccessarticleundertheCCBY-NC-NDlicense(http://creativecommons.org/licenses/by-nc-nd/4.0/). K. Neselu et al. for optimizing data collection and storage efficiency, and for achieving highest resolution reconstructions. Second, post-specimen energy filters can be used (Schr¨oder et al., 1990; Yonekura et al., 2006), which remove inelastically scattered electrons to reduce background noise, especially in regions with thicker ice. Using an energy filter should increase the upper range of ice thicknesses useful for achieving a desired resolution. Third, increasing the accelerating voltage of a microscope reduces the inelastic mean free path of scattering (Dickerson et al., 2022; Henderson, 1995; Martynowycz et al., 2021; Peet et al., 2019). This means that given the same sample thickness, electrons that have higher energy are less likely to undergo inelastic scattering than those with lower energy, and so will contribute less noise in those micrographs. While the theoretical effects of ice thickness on single particle analysis and available strategies to optimize data collection are known, the practical effects of ice thickness on single particle analysis recon- struction resolutions have to our knowledge not been experimentally quantified. To this end we collected large apoferritin datasets over a wide range of ice thickness (15–500 nm) using a variety of instrumen- tation. This included both 200 kV and 300 kV microscopes (Glacios, Arctica, and Krios); direct electron detectors operating in integrating (Glacios with Falcon3), counting (Arctica with K3 and Krios with K3), and super resolution mode (Arctica with K3); and with a 20 eV energy filter slit inserted or retracted (Krios with K3). The data were sorted into groups based on ice thickness and each batch was independently pro- cessed to measure the impact of ice thickness and imaging technique on reconstruction resolution. We show that using a 20 eV energy filter slit has a greater effect in thicker ice; that operating at 300 kV instead of 200 kV accelerating voltage provides significant resolution improve- ments at an ice thickness above 150 nm; that collecting data in super resolution mode provides the most improvement in 150–200 nm thickness; and finally that using a detector operating in counting instead of linear mode has the greatest positive effect in < 150 nm ice thickness. Our findings can serve as a guide for users seeking to optimize data collection or sample preparation routines for both single particle and in situ cryo-EM. We also note that most in situ data collection is done on samples in a range of ice thickness above 150 nm so these results may be especially relevant to that community. Methods Sample preparation Mouse apoferritin in a pET24a vector (Danev et al., 2019) was expressed in BL21(DE3) pLys cells. Cells were lysed, and apoferritin precipitated with 60% ammonium sulfate. After resuspension in 30 mM HEPES pH 7.5, 1 M NaCl, and 1 mM DTT, apoferritin was injected onto a HiTrap Q column and eluted with a 0–0.5 M NaCl gradient over 4 col- umn volumes. The elution peak was pooled and concentrated for puri- fication on a Superdex 200 16/60 column in 30 mM HEPES pH 7.5, 150 mM NaCl, and 1 mM DTT. UltrAuFoil R1.2/1.3 300 mesh grids (Quantifoil) were plasma cleaned using a Solarus II (Gatan) with Ar:O₂ (26.3:8.7) at 15 W for 10 s. 3 μL mouse 8 mg/ml apoferritin was applied onto the plasma cleaned grids. After a 30 s incubation at 100% relative humidity and 22 C the grids were blotted for 4–5 s then plunge frozen into liquid ethane using a Vitrobot Mark IV (Thermo Fisher Scientific). ◦ Data collection Cryo-EM data was collected on three different microscopes. (1) A Titan Krios (Thermo Fisher Scientific) microscope operating at 300 kV and equipped with a BioQuantum energy filter (Gatan) and K3 camera (Gatan) in counting mode. Krios data was collected either with a 20 eV energy filter slit, or with the slit open, on the same grid during the same data collection session. (2) A Talos Arctica microscope operating at 200 kV and equipped with a K3 detector operating in counting or super resolution mode. Data was collected on a different apoferritin grid. (3) A Glacios microscope operating at 200 kV and equipped with a Falcon3 camera (ThermoFisher Scientific) operating in integrating mode. Data was collected on a third apoferritin grid. Data collection parameters are found in Table 1. Leginon (Cheng et al., 2021; Suloway et al., 2005) was used for automated data collection for all sessions. Ice thickness on the Arctica and Glacios was measured by using aperture limited scattering (ALS) method, and on the Krios by using the zero loss peak (ZLP) method (Rice et al., 2018). During data collection, the incoming images were motion corrected and dose weighted with motioncor2 (Zheng et al., 2017) in Appion (Lander et al., 2010). Image processing Frame-aligned and dose-weighted images were sorted into 5 different ice thickness groups (0–50 nm, 50–100 nm, 100–150 nm, 150–200 nm, and 200–500 nm) using a Python script. The micrographs were then imported into different workspaces and processed using cry- oSPARC (Punjani et al., 2017). After importing the micrographs from each ice thickness group, the CTF was estimated. Next, the micrographs were manually curated to exclude bad micrographs, using the same exclusion criteria for all ice thickness groups. 200 micrographs were then randomly selected for further image processing. Particles were manually picked from some of these micrographs to generate good picking templates. Next, template picking was done on all 200 micro- graphs. The picks were then inspected, and obvious bad picks were excluded. The good picks were then extracted in a 256-pixel box and connected to a 2D class averaging job. The resulting 2D classifications were evaluated and only good class averages with good signal to noise ratio were kept. From the set of good particles, 2 to 4 mutually exclusive sets of 14,000 particles were created for further processing, depending on the number of particles available. Homogeneous refinement with defocus and CTF refinement was done on each set of particles, and the best and average reconstruction statistics are reported here. For the Glacios dataset, there was overfitting in the 3D reconstructions for ice thicknesses above 100 nm resulting in an overestimation of the resolu- tion. To overcome this, the same soft mask around the apoferritin den- sity was applied to all reconstructions from all Glacios ice thickness groups. Analysis Once a 3D reconstruction was obtained, the density was evaluated using UCSF Chimera (Pettersen et al., 2004). Reconstructions from the different microscopes and ice thickness groups were compared against one another to evaluate which ice thickness and microscope setup gave the best results. Linear regressions were done in Microsoft Excel using the midpoint of each ice thickness group as the value on the x-axis. Map- to-map Fourier shell correlations (FSCs) were calculated on the EMDB FSC server https://www.ebi.ac.uk/emdb/validation/fsc/. Table 1 Cryo-EM data collection parameters. Dataset “Krios (Filtered and Unfiltered)” “Arctica (Counting and Super Resolution)” “Glacios” Titan Krios 300 Talos Arctica 200 Microscope Accelerating voltage (kV) Energy filter slit width (eV) Pixel size (Å/pix) Exposure time (ms) Frame time (ms) Number of frames Total dose (e/Å2) Session name 20 1.083 2000 40 50 51.22 22may20b N/A 1.096 2400 Glacios 200 N/A 1.204 2000 50 48 50.34 22sep21a, 22sep22a 40 50 50.53 22feb15b JournalofStructuralBiology:X7(2023)1000852 K. Neselu et al. Results The thinner the ice, the better the resolution To study the effects of ice thickness on resolution, we collected apoferritin data with a wide range of ice thicknesses (15–500 nm) on the Krios, Arctica, and Glacios microscopes. We observed the expected trend that as ice thickness increases, resolution decreases (Fig. 1 & Supple- mentary Fig. 1). With data collected on differently configured micro- scopes, we can quantify the contributions from the energy filter, accelerating voltage, and detector mode, to reconstruction resolutions at varying ice thicknesses. It is important to bear in mind that the numbers presented here are for a very specific data collection scenario, and do not represent the performance limit of these microscope setups. What a reconstruction can achieve practically will also depend on the number of particles, sample size, and homogeneity. Here, we report both the best (Fig. 1 and Tables 2a and 2b) and average (Supplementary Fig. 1 and Supplementary Table 1) recon- struction statistics from mutually exclusive sets of 14,000 particles processed with the same data processing parameters and settings, so as to have a holistic view on our processing, and to report on the variability we encountered in the process. The energy filter reduces the rate of resolution decay Comparing the Krios datasets with and without the 20 eV energy filter slit shows that the main advantage of using the slit is to reduce the rate at which the resolution decays with increasing ice thickness. Fitting linear regressions into the 0–150 nm range of the resolution plot Table 2a Accompaniment table to Fig. 1A. Table of highest apoferritin reconstruction resolutions obtained from micrographs of various ice thicknesses, and with microscopes of different configurations (see Table 1 for microscope configura- tion details). 0–50 nm 3.40 2.76 50–100 nm 100–150 nm 150–200 nm 200–500 nm 4.64 2.91 10.18 3.19 9.63 6.83 2.61 2.78 3.04 4.11 2.41 2.53 2.84 3.21 2.36 2.46 2.58 2.76 2.92 9.84 8.13 8.82 6.67 Glacios Arctica (Counting) Arctica (Super Resolution) Krios (Unfiltered) Krios (Filtered) (Table 2a) reveals that both data have very similar intercepts (2.27 Å for unfiltered, and 2.30 Å for filtered), but the slope of the unfiltered data, at (cid:0) 1, is ~ 2-fold higher than that of filtered data at 0.0022 Å 0.0043 Å nm (cid:0) 1 (Table 3a). This indicates that for apoferritin at the thinnest nm possible ice, the energy filter has minimal effect; however, with every nm of increasing ice thickness, the resolution of these reconstructions from data collected without an energy filter suffer 2-fold more than with an energy filter, up to 150 nm. Similarly, the rate of B-factor decay (Table 3b) on unfiltered data is 1.6-fold worse than that of filtered data, up to 150 nm ice thickness. Above 150 nm thickness, however, the resolution of unfiltered reconstructions starts to decay more rapidly, reaching a best of only 6.67 Å in 200–500 nm ice thickness, compared to 2.92 Å for energy filtered data. Practically speaking, most single particle data collected at the Simons Electron Microscopy Center is in ice Fig. 1. (A) Plot of the best apoferritin resolutions obtained from micrographs of various ice thicknesses, and with microscopes of different configurations (see Table 1 for microscope configuration details). (B) Guinier plot B-factors from the best reconstructions versus ice thickness group. The data from each ice thickness group are plotted on the midpoint ice thickness value on the x-axis, i.e. 25, 75, 125, 175, and 350 nm. The numbers giving rise to these plots can be found in Table 2a and 2b. (C) Fig. 1(A) with a rescaled y-axis from 2.3 to 4 Å. (D) Fig. 1(B) with a rescaled y-axis from 50 to 300 Å2. JournalofStructuralBiology:X7(2023)1000853 K. Neselu et al. Table 2b Accompaniment table to Fig. 1B. Table of Guinier plot B-factors from Table 1(A) for each ice thickness group. DNE = Did not estimate; that is, the 3D refinement job did not return a B-factor. 0–50 nm 220.1 117 50–100 nm 377.2 126.6 100–150 nm 150–200 nm 200–500 nm DNE 129.6 DNE 756.8 DNE 1470.7 106 113.8 118.5 217.4 1287.1 87.5 91 108.5 128.4 766.4 85.5 90.7 98.3 104.8 116.7 Glacios Arctica (Counting) Arctica (Super Resolution) Krios (Unfiltered) Krios (Filtered) Table 3a Linear regression fits into resolution vs ice thickness plots. Fits were done into the linear portions of the graph to allow for the best comparisons between plots. DNF = did not fit. Dataset Fit range Linear regression Glacios Arctica (Counting) Arctica (Super Resolution) Krios (Unfiltered) Krios (Filtered) DNF 0–150 nm 0–150 nm 0–150 nm 0–150 nm DNF y = 0.0043x + 2.6308 y = 0.0043x + 2.4875 y = 0.0043x + 2.2708 y = 0.0022x + 2.3017 Table 3b Linear regression fits into Guinier plot B-factor vs ice thickness plots. Dataset Fit range Linear regression Glacios Arctica (Counting) Arctica (Super Resolution) Krios (Unfiltered) Krios (Filtered) DNF 0–150 nm 0–150 nm 0–150 nm 0–150 nm DNF y = 0.126x + 114.95 y = 0.125x + 103.39 y = 0.21x + 79.917 y = 0.128x + 81.9 R2 DNF 0.9704 0.9856 0.9389 0.9973 R2 DNF 0.9162 0.9799 0.871 0.9884 thickness < 100 nm, for which the improvement in resolution by inserting the energy filter slit is small. This is expected since this ice thickness range is well below the inelastic mean free path of 350 ~ 440 nm at 300 kV (Yonekura et al., 2006). Since the 20 eV slit provided the greatest resolution improvement in the thickest 200–500 nm ice thick- ness group (Fig. 1A and Table 1A), this may be of particular interest for in situ data collection from FIB-milled lamella where thickness is more likely to be in the range 150–250 nm. Increasing high tension from 200 to 300 kV has the greatest effect in thicker ice Next, we compared the Arctica counting data with the unfiltered Krios data. Since both microscopes were operated with a K3 detector in counting mode, we could concentrate on the effects of 200 vs 300 kV accelerating voltages. In ice of 0–150 nm, 200 kV data performed slightly worse than 300 kV data: linear regression fits reached intercepts of 2.63 Å (for 200 kV) vs 2.27 Å (for 300 kV), although the rates of (cid:0) 1 (Table 3a). The resolution decay were the same, at 0.0043 Å nm biggest differences were observed at > 150 nm ice thickness, where the 200 kV Arctica counting data achieved only 6.83–8.13 Å re- constructions, compared to 3.21–6.67 Å for 300 kV Krios counting data (Table 2a). The data shows that increasing the accelerating voltage from 200 to 300 kV provides the greatest improvement at the 150–200 nm thickness range. The corresponding ~ 6-fold increase in B-factors (128.4 Å2 for 300 kV vs 756.8 Å2 for 200 kV) indicates that for this ice thickness, a much larger amount of 200 kV data would need to be collected to compensate for the loss of information due to inelastic scattering. Super resolution > counting > integrating mode In integrating mode, a direct electron detector integrates the total charge imparted by an electron, distributed by the microscope’s point spread function, across several pixels. Operating in counting mode al- lows for the localization of single electron events on the camera with pixel accuracy, reducing Landau and readout noise, and improving the DQE of a detector compared to integrating mode (Gatan, 2022; Li et al., 2013). A further improvement in DQE can be gained by collecting data in super resolution mode which makes use of high-speed detector elec- tronics to determine the sub-pixel location of each electron event, digitally increasing the number of pixels by 4x (Booth, 2012; Li et al., 2013). The poorer performance of 200 kV Arctica counting data compared to 300 kV Krios counting data in 150–200 nm ice can be somewhat rescued by collecting data in super resolution mode. This improved the reconstruction resolution from 6.83 to 4.11 Å, and the B-factors from 756.8 to 217.4 Å2 (Table 2a and 2b) which are more comparable to 300 kV Krios counting data. By comparing Arctica data with Glacios data, we could compare the performance of a K3 detector operating in counting mode with a Falcon3 in integrating mode respectively. This is not an ideal comparison of counting vs integrating collection modes, since the Falcon3 and K3 have slightly different DQEs (Booth, 2019; Morado, 2020). Nevertheless, we include this data in the interest of completeness. We observed the most significant improvements from using counting mode below 150 nm ice thickness. Above 150 nm ice thickness, counting and integrating modes achieved similar resolutions and B-factors, suggesting that the noise from increased inelastic scattering and the subsequent reduction in image contrast dominates the gain in signal-to-noise from counting. We observed that our Glacios data performs poorly at all ice thick- nesses above 50 nm. While the data may appear to indicate that the resolution remains stable in 100–500 nm ice, in contrast to the other datasets where the resolutions and B-factors continued to worsen in the same ice thickness range (Fig. 1 and Supplementary Fig. 1), we believe this is just an artifact of generally poor reconstructions. Visual exami- nation of the maps for reconstructions above 100 nm thickness revealed no real structural features that might be expected for a map 9 ~ 10 Å in resolution, and instead showed that the ~ 9.5 Å reported resolutions were due to misalignments to noise (Fig. 2). Map-to-map FSCs of the maps from thicker ice calculated against the map from 0 to 50 nm ice show that the Glacios maps from > 100 nm ice thickness have, at best, a 20 Å correlation to the map from 0 to 50 nm thickness (Supplementary Fig. 2). We conclude that Glacios data collected in integrating mode in ice thicker than 100 nm produces unreliable reconstructions that are, for apoferritin, significantly worse than 7 Å. As ice thickness increases, we might expect a smooth decrease in reconstruction resolution. Instead, across all our datasets, we observed that resolution would decrease up to ~ 4 Å, after which there is a “jump” to ~ 7 Å without an intermediate 5–6 Å reconstruction (Fig. 1A and Table 1A). We hypothesize that this is because at better than ~ 4 Å there are side chain densities that reconstruction programs can align to; however, for apoferritin, which contains only alpha helices and no beta sheets or any other significant structural features, the next feature that can be aligned are alpha helices at ~ 7 Å, which results in the observed “jump” in resolution. Discussion Thicker ice produces more inelastic scattering events, which de- creases single-to-noise ratios and worsens reconstruction alignment ac- curacy, resulting in poorer reconstruction resolutions. Here, we observe that using a 20 eV energy filter slit slows down the rate of resolution decay with increasing ice thickness by ~ 2-fold on a 300 kV microscope. Using 300 kV accelerating voltage provides the greatest benefit over 200 kV at > 150 nm ice thickness, improving our apoferritin JournalofStructuralBiology:X7(2023)1000854 K. Neselu et al. Fig. 2. Apoferritin reconstructions from Glacios data collected in integrating mode, by ice thickness group. reconstruction from 6.83 to 3.21 Å in 150–200 nm ice. Using super resolution mode provides the most improvement in < 200 nm ice, and collecting data in counting instead of integrating mode improves re- constructions most noticeably in ice thinner than 150 nm. Combining these effects, we obtained the highest resolution reconstructions across all ice thickness groups from energy filtered 300 kV Krios data, followed by unfiltered Krios, 200 kV Arctica with a K3 in super resolution then counting mode, and lastly with a 200 kV Glacios with a Falcon3 in integrating mode. For 200 kV instruments, the best imaging setup of using a K3 in super resolution mode enabled high resolution re- constructions < 200 nm ice. In situations where thick (> 200 nm) ice cannot be avoided, for example with a large virus, large macromolecular complex, or in situ sample, it is most critical to use a microscope with high kV and an energy filter to obtain the highest resolution data. In thin ice (0–50 nm), the best reconstructions from our comparable 200 and 300 kV data (Arctica with K3 in counting mode and Krios unfiltered) perform similarly, at 2.76 and 2.41 Å respectively. A visual examination of the maps showed little difference between the two (Supplementary Fig. 4). The advantages of using a lower accelerating voltage for single particle cryo-EM experiments have recently been more thoroughly described (Naydenova et al., 2019; Peet et al., 2019), and the data show that 200 keV electrons are better for single particle cryo-EM than 300 keV when specimen thickness is not considered (Peet et al., 2019). Specifically, the ratio of elastic scattering at 200 keV to 300 keV is 1.24, whereas the ratio of inelastic scattering at 200 keV to 300 keV is 1.14. For specimens thinner than ~ 100 nm, electron energies lower than 300 keV were shown to contain more useful information for single particle cryo-EM (Peet et al., 2019). This improvement is likely some- what offset by the detective quantum efficiency (DQE) of existing counting direct detectors being slightly worse at 200 keV than at 300 keV. During processing, we observed that in thicker ice, reconstructions from mutually exclusive sets of 14,000 good particles randomly selected from each ice thickness group could achieve very different resolutions (Supplementary Fig. 1A). In the 200–500 nm ice thickness group, while the best reconstruction we obtained with filtered Krios data was 2.92 Å, across 4 independent reconstructions from mutually exclusive sets of 14,000 particles from the same dataset, we obtained reconstructions that ranged up to 10.74 Å, with an average of 7.39 Å (Supplementary Fig. 1A and Supplementary Table 1). This does not appear to be because some groups of 14,000 particles were from thinner ice than others. Analysis of the per-particle distribution of ice thicknesses for each of the four Krios filtered reconstructions from the 200–500 nm ice thickness group showed that particles that gave rise to the 2.92 Å reconstruction did not have significantly better ice thicknesses than the other 7–10 Å reconstructions (Supplementary Fig. 3). One hypothesis to explain this is that variability can arise during the random initialization process of 3D reconstruction: if a subset of higher quality particles happens to be selected to initialize the reconstruction, this could lead to better align- ment and resolution for that reconstruction. However, since these high- quality particles are not ubiquitous in the thick ice data, obtaining these reconstructions can be hit-or-miss. Another interesting observation was that turning off both defocus and CTF optimization during 3D recon- struction in thick ice could sometimes give higher resolution re- constructions than if we turned on both options. For example, the best unfiltered Krios 200–500 nm reconstruction achieved 6.67 Å with defocus and CTF refinement, but 3.65 Å with both options deactivated. We think this could be because in thick ice the particles have very little high-resolution signal, so the defocus and CTF optimizing algorithms are fitting to noise, and turning them off can potentially yield a better reconstruction. There are several additional considerations for improving on the existing imaging setups tested in this work. Firstly, considering that the inelastic mean free path is shorter for slower electrons, there will be more inelastic scattering in a 200 kV microscope, which means that installing a post-specimen energy filter will make a bigger impact on JournalofStructuralBiology:X7(2023)1000855 K. Neselu et al. such a setup than on a 300 kV microscope. Secondly, on a microscope with a post-specimen energy filter, when collecting data in thick ice, it would also be beneficial to reduce the width of the energy filter slit to ~ 10 eV to optimally eliminate inelastically scattered electrons and improve the reconstruction (Nakane et al., 2020; ThermoFisher Scien- tific, 2022). Thirdly, reducing inelastic scattering from ice by energy filtration will have the most benefit for small particles, since they have the lowest signal-to-noise ratios, and energy filtration also improves amplitude contrast allowing for better alignments during reconstruction (Danev et al., 2021). Since apoferritin has a molecular weight of ~ 450 kDa, this same experiment should be done with a smaller protein of < 200 kDa to better evaluate the benefits for small macromolecules. We observed that a Falcon3 detector in integrating mode (Glacios dataset) was only useful in the thinnest ice, < 50 nm. While using integrating mode on a Falcon3 reduces exposure times from 60 s to 1–2 s, making it much faster for a quick survey of the grid, the data from integrating mode is not likely to provide useful reconstructions except in very thin ice. For more challenging samples that prefer thicker ice or when working with suboptimal grids with thicker ice (as is commonly the case on a screening microscope), reconstructions obtained from the Glacios in integrating mode may not be an accurate reflection of what can be obtained from a better imaging setup. Here at the Simons Elec- tron Microscopy Center, preliminary data are commonly collected on our Glacios in integrating mode before a full data collection on a Krios instrument. The data in this paper provide a useful benchmark for how reconstructions from a Glacios dataset can be extrapolated to re- constructions from a Krios dataset, given an ice thickness range. The case can be made here for either collecting data in counting mode on the Falcon3 on our Glacios microscope, or else for upgrading the camera, say to a K3, for faster speeds and better reconstructions. Preparing samples in the thinnest ice possible remains the best global solution to obtaining high resolution. Where thick ice is necessary, for example with large macromolecules or in situ samples, using the best available imaging setup is essential for reaching high resolution with the greatest possible speed. Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Data availability Micrographs are available at EMPIAR-11397. Reconstructions are available on EMDB with the following accession codes: for Glacios 0-50 nm, EMD-29566; 50-100 nm, EMD-29567; 100-150 nm, EMD-29568; 150-200 nm, EMD-29569; 200-500 nm, EMD-29570. For Arctica in counting mode 0-50 nm, EMD-29573; 50-100 nm, EMD-29574; 100-150 nm, EMD-29575; 150-200 nm, EMD-29576; 200-500 nm, EMD-29577. For Arctica in super resolution mode 0-50 nm, EMD-29589; 50-100 nm, EMD-29591; 100-150 nm, EMD-29592; 150-200 nm, EMD-29593; 200-500 nm, EMD-29594. For Krios unfiltered 0-50 nm, EMD-29554; 50-100 nm EMD-29555; 100-150 nm, EMD-29556; 150-200 nm, EMD- 29557; 200-500 nm, EMD-29558. For Krios filtered 0-50 nm, EMD- 29536; 50-100 nm, EMD-29535; 100-150 nm, EMD-29559; 150-200 nm, EMD-29513; 200-500 nm, EMD-29393. Acknowledgements We thank Dr. Masahide Kikkawa (University of Tokyo) for the apo- ferritin plasmid, and Dr. Brian Kloss (NYSBC) for expressing and pur- ifying the apoferritin sample. This work was supported by the Simons Electron Microscopy Center and National Resource for Automated Molecular Microscopy located at the New York Structural Biology Center, supported by grants from the Simons Foundation (SF349247) and the NIH National Institute of Gen- eral Medical Sciences (GM103310). All Arctica data in this work was collected at the NYU Langone Health Cryo-EM core facility (RRID: SCR_019202). Appendix A. Supplementary data Supplementary data to this article can be found online at https://doi. org/10.1016/j.yjsbx.2023.100085. References Booth, C., 2012. K2: a super-resolution electron counting direct detection camera for Cryo-EM. Microsc. Microanal. 18 (S2), 78–79. https://doi.org/10.1017/ S1431927612002243. Booth, C. (2019). Detection Technologies for Cryo-Electron Microscopy. https://cryoem.slac. stanford.edu/s2c2/sites/s2c2.cryoem.slac.stanford.edu/files/Cameras%20for%20Cr yo-EM%20Stanford2019-Booth.pdf. Brown, H.G., Hanssen, E., 2022. MeasureIce: accessible on-the-fly measurement of ice thickness in cryo-electron microscopy. Commun. Biol. 5 (1), 817. https://doi.org/ 10.1038/s42003-022-03698-x. Cheng, A., Negro, C., Bruhn, J.F., Rice, W.J., Dallakyan, S., Eng, E.T., Waterman, D.G., Potter, C.S., Carragher, B., 2021. Leginon: new features and applications. Protein Sci. 30 (1), 136–150. https://doi.org/10.1002/pro.3967. Danev, R., Yanagisawa, H., Kikkawa, M., 2019. Cryo-electron microscopy methodology: current aspects and future directions. Trends Biochem. Sci 44 (10), 837–848. https://doi.org/10.1016/j.tibs.2019.04.008. Danev, R., Belousoff, M., Liang, Y.-L., Zhang, X., Eisenstein, F., Wootten, D., Sexton, P.M., 2021. Routine sub-2.5 Å cryo-EM structure determination of GPCRs. Nat. Commun. 12 (1), 4333. https://doi.org/10.1038/s41467-021-24650-3. Darrow, M.C., Moore, J.P., Walker, R.J., Doering, K., King, R.S., 2019. Chameleon: next generation sample preparation for CryoEM based on Spotiton. Microsc. Microanal. 25 (S2), 994–995. https://doi.org/10.1017/S1431927619005701. Darrow, M.C., Booth, T., Moore, J.P., Doering, K., Thaw, P., King, R.S., 2021. Enabling a Paradigm Shift in CryoEM Sample Preparation with chameleon. Microsc. Microanal. 27 (S1), 524–525. https://doi.org/10.1017/S1431927621002336. Dickerson, J.L., Lu, P.-H., Hristov, D., Dunin-Borkowski, R.E., Russo, C.J., 2022. Imaging biological macromolecules in thick specimens: The role of inelastic scattering in cryoEM. Ultramicroscopy 237, 113510. https://doi.org/10.1016/j. ultramic.2022.113510. Gatan. (2022). Improving DQE with Counting and Super-Resolution. https://www.gatan. com/improving-dqe-counting-and-super-resolution. Henderson, R., 1995. The potential and limitations of neutrons, electrons and X-rays for atomic resolution microscopy of unstained biological molecules. Q. Rev. Biophys. 28 (2), 171–193. https://doi.org/10.1017/S003358350000305X. Huntington, B., Zhao, L., Bron, P., Shahul Hameed, U.F., Arold, S.T., Qureshi, B.M., 2022. Thicker Ice Improves the Integrity and Angular Distribution of CDC48A Hexamers on Cryo-EM Grids. Front. Mol. Biosci. 9, 890390 https://doi.org/10.3389/ fmolb.2022.890390. Lander, G. C., Stagg, S. M., Voss, N. R., Cheng, A., Fellmann, D., Yoshioka, C., Irving, C., Mulder, A., Lau, P.-W., Potter, C. S., & Carragher, B. (2010). Appion: An integrated, database-driven pipeline to facilitate EM image processing. 16. Li, X., Mooney, P., Zheng, S., Booth, C.R., Braunfeld, M.B., Gubbens, S., Agard, D.A., Cheng, Y., 2013. Electron counting and beam-induced motion correction enable near-atomic-resolution single-particle cryo-EM. Nat. Methods 10 (6), 584–590. https://doi.org/10.1038/nmeth.2472. Martynowycz, M. W., Clabbers, M. T. B., Unge, J., Hattne, J., & Gonen, T. (2021). Benchmarking the ideal sample thickness in cryo-EM. Proceedings of the National Academy of Sciences, 118(49), e2108884118. https://doi.org/10.1073/ pnas.2108884118. Morado, D. (2020). Some timing tests of the Gatan K3. https://forum.scilifelab.se/t/some-t iming-tests-of-the-gatan-k3/139. Nakane, T., Kotecha, A., Sente, A., McMullan, G., Masiulis, S., Brown, P.M.G.E., Grigoras, I.T., Malinauskaite, L., Malinauskas, T., Miehling, J., Ucha´nski, T., Yu, L., Karia, D., Pechnikova, E.V., de Jong, E., Keizer, J., Bischoff, M., McCormack, J., Tiemeijer, P., Hardwick, S.W., Chirgadze, D.Y., Murshudov, G., Aricescu, A.R., Scheres, S.H.W., 2020. Single-particle cryo-EM at atomic resolution. Nature 587 (7832), 152–156. Naydenova, K., McMullan, G., Peet, M.J., Lee, Y., Edwards, P.C., Chen, S., Leahy, E., Scotcher, S., Henderson, R., Russo, C.J., 2019. CryoEM at 100 keV: A demonstration and prospects. IUCrJ 6 (6), 1086–1098. https://doi.org/10.1107/ S2052252519012612. Noble, A.J., Dandey, V.P., Wei, H., Brasch, J., Chase, J., Acharya, P., Tan, Y.Z., Zhang, Z., Kim, L.Y., Scapin, G., Rapp, M., Eng, E.T., Rice, W.J., Cheng, A., Negro, C.J., Shapiro, L., Kwong, P.D., Jeruzalmi, D., des Georges, A., Carragher, B., 2018. Routine single particle CryoEM sample and grid characterization by tomography. Elife 7, e34257. Peet, M.J., Henderson, R., Russo, C.J., 2019. The energy dependence of contrast and damage in electron cryomicroscopy of biological molecules. Ultramicroscopy 203, 125–131. https://doi.org/10.1016/j.ultramic.2019.02.007. Pettersen, E.F., Goddard, T.D., Huang, C.C., Couch, G.S., Greenblatt, D.M., Meng, E.C., Ferrin, T.E., 2004. UCSF Chimera—A visualization system for exploratory research JournalofStructuralBiology:X7(2023)1000856 K. Neselu et al. and analysis. J. Comput. Chem. 25 (13), 1605–1612. https://doi.org/10.1002/ jcc.20084. ThermoFisher Scientific. (2022). Selectris Applications Notes. https://cryoem.jhmi. edu/wp-content/uploads/2022/07/Selectris_apps_note_v2_1.pdf. Punjani, A., Rubinstein, J.L., Fleet, D.J., Brubaker, M.A., 2017. cryoSPARC: algorithms for rapid unsupervised cryo-EM structure determination. Nat. Methods 14 (3), 290–296. Wu, S., Armache, J.-P., Cheng, Y., 2016. Single-particle cryo-EM data acquisition by using direct electron detection camera. Microscopy 65 (1), 35–41. https://doi.org/ 10.1093/jmicro/dfv355. Rheinberger, J., Oostergetel, G., Resch, G.P., Paulino, C., 2021. Optimized cryo-EM data- acquisition workflow by sample-thickness determination. Acta Crystallogr. Sect. D Struct. Biol. 77 (5), 565–571. https://doi.org/10.1107/S205979832100334X. Rice, W.J., Cheng, A., Noble, A.J., Eng, E.T., Kim, L.Y., Carragher, B., Potter, C.S., 2018. Routine determination of ice thickness for cryo-EM grids. J. Struct. Biol. 204 (1), 38–44. https://doi.org/10.1016/j.jsb.2018.06.007. Schr¨oder, R.R., Hofmann, W., M´en´etret, J.-F., 1990. Zero-loss energy filtering as improved imaging mode in cryoelectronmicroscopy of frozen-hydrated specimens. J. Struct. Biol. 105 (1–3), 28–34. https://doi.org/10.1016/1047-8477(90)90095-T. Suloway, C., Pulokas, J., Fellmann, D., Cheng, A., Guerra, F., Quispe, J., Stagg, S., Potter, C.S., Carragher, B., 2005. Automated molecular microscopy: The new Leginon system. J. Struct. Biol. 151 (1), 41–60. https://doi.org/10.1016/j. jsb.2005.03.010. Xiao, C., Chipman, P.R., Battisti, A.J., Bowman, V.D., Renesto, P., Raoult, D., Rossmann, M.G., 2005. Cryo-electron microscopy of the giant mimivirus. J. Mol. Biol. 353 (3), 493–496. https://doi.org/10.1016/j.jmb.2005.08.060. Yonekura, K., Braunfeld, M.B., Maki-Yonekura, S., Agard, D.A., 2006. Electron energy filtering significantly improves amplitude contrast of frozen-hydrated protein at 300kV. J. Struct. Biol. 156 (3), 524–536. https://doi.org/10.1016/j. jsb.2006.07.016. Zheng, S.Q., Palovcak, E., Armache, J.-P., Verba, K.A., Cheng, Y., Agard, D.A., 2017. MotionCor2: Anisotropic correction of beam-induced motion for improved cryo- electron microscopy. Nat. Methods 14 (4), 331–332. https://doi.org/10.1038/ nmeth.4193. JournalofStructuralBiology:X7(2023)1000857
10.3389_fnbeh.2022.835753
ORIGINAL RESEARCH published: 06 April 2022 doi: 10.3389/fnbeh.2022.835753 Bottlenecks, Modularity, and the Neural Control of Behavior Anjalika Nande 1,2†, Veronika Dubinkina 3,4†, Riccardo Ravasio 5,6†, Grace H. Zhang 1† and Gordon J. Berman 7* 1 Department of Physics, Harvard University, Cambridge, MA, United States, 2 Institute for Computational Medicine, Johns Hopkins University, Baltimore, MD, United States, 3 Department of Bioengineering, University of Illinois at Urbana-Champaign, Urbana, IL, United States, 4 Carl R. Woese Institute for Genomic Biology, University of Illinois at Urbana-Champaign, Urbana, IL, United States, 5 Institute of Physics, École Polytechnique Fédérale de Lausanne, Lausanne, Switzerland, 6 The James Franck Institute, University of Chicago, Chicago, IL, United States, 7 Departments of Biology and Physics, Emory University, Atlanta, GA, United States In almost all animals, the transfer of information from the brain to the motor circuitry is facilitated by a relatively small number of neurons, leading to a constraint on the amount of information that can be transmitted. Our knowledge of how animals encode information through this pathway, and the consequences of this encoding, however, is limited. In this study, we use a simple feed-forward neural network to investigate the consequences of having such a bottleneck and identify aspects of the network architecture that enable robust information transfer. We are able to explain some recently observed properties of descending neurons—that they exhibit a modular pattern of connectivity and that their excitation leads to consistent alterations in behavior that are often dependent upon the desired behavioral state of the animal. Our model predicts that in the presence of an information bottleneck, such a modular structure is needed to increase the efficiency of the network and to make it more robust to perturbations. However, it does so at the cost of an increase in state-dependent effects. Despite its simplicity, our model is able to provide intuition for the trade-offs faced by the nervous system in the presence of an information processing constraint and makes predictions for future experiments. Keywords: neural control, modularity, bottlenecks, neural networks, robustness 1. INTRODUCTION When presented with dynamical external stimuli, an animal selects a behavior to perform—or a lack thereof—according to its internal drives and its model of the world. Its survival depends on its ability to quickly and accurately select an appropriate action, as well as to transmit information from the brain to its motor circuitry in order to physically perform the behavior. In almost all animals, however, there exists a bottleneck between the number of neurons in the brain that make cognitive decisions and the motor units that are responsible for actuating movements, thus constraining the amount of information that can be transmitted from the brain to the body (Smarandache-Wellmann, 2016; Kandel et al., 2021). Edited by: Ilona C. Grunwald Kadow, Technical University of Munich, Germany Reviewed by: Moshe Parnas, Tel Aviv University, Israel Carlotta Martelli, Johannes Gutenberg University Mainz, Germany *Correspondence: Gordon J. Berman gordon.berman@emory.edu †These authors have contributed equally to this work Specialty section: This article was submitted to Individual and Social Behaviors, a section of the journal Frontiers in Behavioral Neuroscience Received: 14 December 2021 Accepted: 14 March 2022 Published: 06 April 2022 Citation: Nande A, Dubinkina V, Ravasio R, Zhang GH and Berman GJ (2022) Bottlenecks, Modularity, and the Neural Control of Behavior. Front. Behav. Neurosci. 16:835753. doi: 10.3389/fnbeh.2022.835753 Frontiers in Behavioral Neuroscience | www.frontiersin.org 1 April 2022 | Volume 16 | Article 835753 Nande et al. Bottlenecks, Modularity, and the Neural Control of Behavior In the fruit fly Drosophila melanogaster, descending commands from the brain to the ventral nerve cord (VNC) are transmitted through approximately 300 bilaterally symmetric pairs of neurons that have their cell bodies in the brain and have axons project into the VNC (Gronenberg and Strausfeld, 1990; Hsu and Bhandawat, 2016). Recent anatomical studies have shown that these neurons exhibit a modular pattern of connectivity, with the descending neurons clustering into groups that each innervate different parts of the motor system (Namiki et al., 2018; Phelps et al., 2021). In addition to these anatomical properties, in the fruit fly, manipulating these descending neurons via optogenetics has shown that exciting individual neurons or subsets of neurons often result in dramatic and robust behavioral alterations— for example, exciting the DNg07 and DNg08 neurons reliably elicits head grooming, and exciting DNg25 elicits a fast running response (Cande et al., 2018). In many cases, however, it has been shown that exciting the same neuron in different contexts (e.g., walking and flying) often have state-dependent effects (Cande et al., 2018; Zacarias et al., 2018; Ache et al., 2019). In other words, the behavioral effect of stimulating the neuron often depends on the actions that the fly is attempting to perform. In this study, we use a simplified model of behavioral control to explore how modularity may help increase the efficiency and robustness of behavioral control given an information bottleneck. Specifically, our model predicts that modularity of behavior increases the efficiency of the network and its robustness to perturbations, but also that this modularity increases the amount of state-dependent variability in how behavioral commands are transmitted through the bottleneck. While our feed-forward model is a vast oversimplification of the complicated recurrent circuitry that lives within a fly’s ventral nerve cord, we show that it provides intuition into the trade-offs the nervous system is faced with, and makes qualitative predictions as to how the system might respond to inhibition or double-activation experiments. 2. RESULTS AND DISCUSSION Inspired by the fly ventral nerve cord, we have developed an abstracted model that aims to generate insight into the general problem of behavior control through an information bottleneck. Specifically, we assume that there is a set of N behaviors that are in an animal’s behavioral repertoire and that to perform one of these behaviors, the animal must excite a subset of M total binary “motor” neurons (e.g., task 14 requires units 1, 3, and 99 to turn- on, and all the rest to be turned off—see Figures 1A,B). However, to model the effect of having limited information transmission from the brain to the motor systems, any commands from the brain must travel through an hidden layer of R < M, N descending neurons (Namiki et al., 2018). We implemented this model using a feed-forward neural network, with the task being encoded in the top layer, the descending neurons being the hidden layer, and the motor units constituting the bottom layer (see Figure 1A). For simplicity, we assume that the brain’s intended behavioral output is represented in a one-hot encoded manner, where only one “decision” neuron is turned on at once [i.e., behavior 2 is represented by a first layer of (0, 1, 0, · · · ) ∈ {0, 1}N]. We start with the case where each behavior is randomly assigned a set of k motor neurons that must be activated. Figure 1B shows an example of this desired mapping, which we call our behavioral matrix. To perform a behavior, one of the decision neurons has to be activated and pass its signal through the network. The parameters of the network, weights {W(1) β }, are trained to perform the mapping between the top and bottom layers as accurately as possible (see details in section 4). α,β } and biases {B(1) α,β , W(2) β , B(2) Given this model, we would like to study how the network performs as a function of the bottleneck size and the sparsity of the behavioral matrix. The absolute maximum number of sequences that the network could encode is 2R as each hidden neuron can either be activated or not. However, this simple limit. In neural network is incapable of reaching the ideal Figure 1C, the bottleneck size required for accurate encoding is ∼ 20 − 60 for N = M = 100, depending on the sparsity of the behavioral matrix. These values are much larger than the minimal possible bottleneck size, R = log2 100 ≈ 7. While we will explore the potential reasons for this discrepancy shortly, we empirically define the critical bottleneck size, Rc, as the minimal number of neurons in the hidden layer sufficient to reproduce 98% of the behaviors correctly, averaged across multiple random instantiations of the behavioral matrix. See Supplementary Figure 1 for example learning and loss curves, and Supplementary Figure 2 for example values of the hidden layer and the weights of the trained network. The values of the hidden layer get more binarized (Supplementary Figures 2a,b) as its size decreases, implying that the system is getting pushed out of its dynamic range. 2.1. Characterization of the Model To explore how the statistics of the behavioral matrix affect the critical bottleneck size, we altered the sparsity of the outputs by manipulating the number of motor neurons activated per behavior (k) while keeping M = N = 100 (Figure 1C). Note that since our output size is 100 and its encoding is binary, a neural network with k and 100 − k activated motor neurons have the same statistical behavior. Thus, sparsity increases as k deviates from 50 in either direction. As evident from Figure 1C and the inset therein, as k decreases below 25, the network requires fewer neurons in the hidden layer (a lower Rc) to learn all of the behaviors perfectly, with the decrease starting around k = 25. Ultimately, for the sparsest output encoding we tested (k = 5), the network requires half the number of neurons compared to the densest (k = 50) case (Rc ≈ 24.4 ± 0.8 vs. Rc ≈ 57 ± 2), indicating that it is more difficult for our model to learn the more complicated patterns that are associated with a denser output. This effect can be more explicitly seen by plotting Rc as a function of the entropy of the behavioral matrix (Figure 1D, Equation 4). Furthermore, we note that the shape of the curve, as a function of hidden layer size, R, approaches that of a sigmoid function in the limit of dense output signal (as k approaches 50). Equivalently, sparsity can be varied by fixing k and varying the size of the output layer M (here, keeping N = 100 fixed) (Supplementary Figure 3). We again find that as the output signal becomes more sparse, that is, as M increases, it is easier to learn the mapping from behavior to motor commands. Moreover, we also notice that the learning curves split into two Frontiers in Behavioral Neuroscience | www.frontiersin.org 2 April 2022 | Volume 16 | Article 835753 Nande et al. Bottlenecks, Modularity, and the Neural Control of Behavior FIGURE 1 | Model construction and parameters. (A) The structure of the ventral nerve chord is modeled by a neural network that takes as input a task assignment represented by a binary sequence Ex of length N. The signal travels through a hidden layer (size R) to an output layer (size M), which corresponds to descending neurons and motor neurons, respectively. Each neuron in one layer communicates with all the neurons in the following layer through the weight matrices W(·) α,β , detailed in section 4. (B) An example of a behavioral matrix that indicates the motor units activated for each task. Row i corresponds to the i-th behavioral command (i.e., the i-th neuron activated in the input layer of the network). k is the number of motor neurons needed to execute a given behavior. Columns correspond to different motor neurons [i.e., the jth column indicates whether a particular motor neuron was active (gray) or not (white) in the behaviors]. (C) Fraction of behaviors learned as a function of hidden layer size R and fixed input layer size N = 100 for varying k and fixed output layer size M = 100. The inset shows the critical bottleneck size Rc as a function of k. Each point is averaged over 30 random input-output combinations. Dashed line indicates critical bottleneck threshold. (D) Values of the critical bottleneck size Rc for different values of sparsity (k = 5, 10, 20, 30, 40, 50) as a function of the behaviorial matrix entropy. Black line is the line of best fit and is provided for visual aid only. regimes (Supplementary Figures 3a,b) corresponding to when M is smaller or larger than N. When M > N, the network finds it much easier to learn with the learning ability saturating when the bottleneck size is a certain fraction of the output layer. 2.2. Modularity of Behaviors While the analyses presented in the previous section involved random mappings between behaviors and motor outputs, we now ask if imposing biologically inspired constraints on this mapping might affect the network. the behavioral matrix is Specifically, we will assume that modular, with similar behaviors (e.g., different locomotion gaits or different types of anterior grooming motions) more likely to require similar motor output patterns. This constraint is motivated from previous anatomical studies in Drosophila (Namiki et al., 2018). the efficiency of To explore the effect of modular structure on our model, we performed a set of simulations with various degrees of behavioral matrix modularity. Specifically, we fixed k = 10 and split the behavioral matrix into 5 regions (see inset in Figure 2A). If there is no active motor neuron in common between the different clusters, then we have perfect modularity [µ = 0.8, where µ is the fraction of the edges that fall within the modules minus the expected fraction within the modules for an equivalent random network (Newman, 2018), see section 4]. We then allowed for some overlap between regions to generate matrices with a spectrum of modularities (some examples given in Figure 3C) between the perfect modular limit and random mixing. We observed that the modular behavioral matrices can be learned more efficiently than random matrices, requiring far smaller critical bottleneck sizes to achieve the correct mapping of behavioral commands (Figure 2A). The perfectly modular output matrix (inset Figure 2A) was learned Frontiers in Behavioral Neuroscience | www.frontiersin.org 3 April 2022 | Volume 16 | Article 835753 Nande et al. Bottlenecks, Modularity, and the Neural Control of Behavior FIGURE 2 | Modularity in the behavioral commands reduces critical bottleneck size and affects other network properties. (A) Relationship between the size of the hidden layer R and the modularity of the behavior matrix. Each point corresponds to a set of numerical experiments with 10 different matrices around a given modularity value (see section 4 for details of data generation) for k = 10, M = N = 100. Rc is defined as the minimal hidden layer size that was able to achieve 98% accuracy in 105 epochs of training. Numbers indicate specific cases that are shown in panels (B–D) in more detail. Inset shows an example of behavioral command matrix for µ = 0.8 case (point 4). (B) Fraction of behaviors learned as a function of the hidden layer size, R for different system sizes with N = M for two levels of modularity (µ = 0.8 and µ = 0.46). Error bars correspond to the standard deviation. Results are averaged over 5 different runs with error bars corresponding to the standard deviation. (C) Values of the critical bottleneck size Rc for different values of modularity [µ = 0.8 (fully modular), 0.68, 0.57, 0.46, 0.36, 0.18] as a function of the behavioral matrix entropy. Black line is the line of best fit and is provided for visual aid only. (D) Structure of the weight matrices W1 and W2 for different modularity values. The dimensionality reduction is performed via UMAP (McInnes et al., 2018), a non-linear method that preserves local structure in the data. The point colors correspond to the colors in (A) inset: 1 (µ = 0.18, random matrix); 2 (µ = 0.46); 3 (µ = 0.56); 4 (µ = 0.8, perfectly modular matrix with 5 clusters). with only Rc = 13 neurons, which is less than half the number required for the random matrix (Rc ≃ 35) with the same amount of sparsity (Supplementary Table 1). Note that the dependence of the critical bottleneck layer size on matrix modularity is not linear, just 2 neurons overlapping between clusters makes learning much harder (Rc = 30, point #3 in Figure 2A). In addition to making the mapping easier to learn, modularity in the behavioral matrix also helps learning scale with the system size. In Figure 2B, we plot the fraction of behaviors learned as a function of the relative size of the bottleneck layer R as compared to the output layer M, for different values of the system size (we assume N = M) and for different values of the modularity. Modularity values were chosen to highlight the differences between a perfectly modular matrix (µ = 0.8) and a matrix that has a low amount of modularity (µ = 0.46) while not being completely random. For highly modular behavioral matrices (blue curves in Figure 2B), we find that the size of the output doesn’t affect the learning ability of the network, as the bottleneck occurs when the size of the hidden layer is a similar fraction of the output sizes. On the other hand, when the behavioral commands aren’t very modular, smaller system sizes learn better for a relatively smaller bottleneck size (green curves in Figure 2B). This is again a reflection of our model finding it easier to learn the simpler patterns (less entropy) of a more modular behavioral matrix (Figure 2C). The similarities between Figures 1D, 2C indicate that the entropy of the behavioral matrix is an important parameter that determines Rc, even while keeping other parameters constant. found that Finally, we imposing a modular output structure also imposes a modular structure on the weights of the learned network (Figure 2D). The modularity in the weights becomes more pronounced as the modularity of the behavioral matrix increases, found in the study of more generalized artificial neural networks (Zavatone-Veth et al., 2021). Together, these results show the that modularity in the behavioral matrix increases the network through efficiency and scaling properties of creating a representation within concomitantly modular the model. to results similar Frontiers in Behavioral Neuroscience | www.frontiersin.org 4 April 2022 | Volume 16 | Article 835753 Nande et al. Bottlenecks, Modularity, and the Neural Control of Behavior FIGURE 3 | Robustness of the network to perturbations increases with the size of the hidden layer and sparsity. (A) Schematic of the perturbation experiment. One of the hidden neurons of the trained network is artificially forced to be on, keeping all other network parameters unchanged. The network is re-run to generate new outputs for each behavioral command. (B) Robustness (fraction of outputs that are unaffected by the perturbation) averaged over the effects of activating each hidden neuron as function of the hidden layer size R, with N = M = 100, with k varied from k = 5 to k = 50. The error bars are obtained by considering 10 different behavioral matrices. The inset shows the size of the hidden layer for which such a perturbation leaves 98% of the behaviors unaffected, Rrobust, as a function of changing sparsity (varying k). (C) Example of a hidden layer perturbation on the trained networks’ behavior matrices with different modularities (all show with R = Rc). In each case, one of the hidden neurons is kept constantly activated, while the rest of the network operates according to the trained weights. White and gray colors correspond to unperturbed motor neurons, non-active and active correspondingly. Blue indicates motor units that have been turned off, and red shows motor units that have been activated. (D) Distribution of the number of behavioral commands affected by the hidden layer perturbation. Colors correspond to different degrees of modularity µ. Each distribution was calculated based on 10 different behavioral matrices, all with R = Rc. 2.3. Robustness to Perturbations of the Bottleneck Layer Although the network is capable of reproducing behavioral commands nearly perfectly when it is near the critical bottleneck, it might be prone to errors due to minor perturbations, including noise in the firing of the descending layer. Inspired by previous studies in flies where descending neurons were artificially activated (Cande et al., 2018; Ache et al., 2019), we investigate the robustness of our trained neural networks by manually activating one hidden neuron at a time. We then observe the changes in the output (see Figure 3A) to see how these activations affect the mapping between command and behavior. An example of possible outcomes on a set of behaviors under these perturbation is shown in Figure 3C (for more examples, see Supplementary Figure 4). For each behavioral command, the motor neurons can either remain unaffected—their original “active” or “non-active” state is maintained (gray and white pixels in Figure 3C) or their state gets flipped—an “active” neuron gets inactivated or vice-versa (red and blue pixels in Figure 3C). The robustness of the network with respect to the activated neuron is calculated as the number of behaviors that are conserved, that is, behavioral commands where all activated motor neurons remain unaffected. Figure 3B shows the robustness of the network to these perturbations as a function of the hidden layer size R and varying sparsity (N = M = 100 is fixed and k is varied), averaged over the effects of activating each hidden neuron and each behavioral command for a randomly generated behavioral matrix (no enforced modularity). For fixed sparsity, the fraction of behaviors that are unaffected increases as the size of the hidden layer increases. At the critical bottleneck size, for example, Rc = 35 for k = 10, 80% of behaviors were unaffected by the perturbation, indicating that the neural network has some margin of robustness. Robustness increases as we increase the hidden layer size R—the behavioral commands become less sensitive to changes in each individual hidden neuron. As long as the bottleneck layer size is less than the output layer (R < M), networks with output signals of high sparsity (lower k) are more Frontiers in Behavioral Neuroscience | www.frontiersin.org 5 April 2022 | Volume 16 | Article 835753 Nande et al. Bottlenecks, Modularity, and the Neural Control of Behavior When applying these perturbations robust on average. The robustness is bounded below by the curve corresponding to maximum output signal density k = 50 = M/2. For sufficiently dense output signals 50 ≥ k > 5, the robustness decreases monotonically with decreasing hidden layer size for the entire range of 1 ≤ R ≤ M. In contrast, the robustness of high sparsity outputs (k = 5) decreases initially with decreasing hidden layer size, but exhibits an increase in both its mean and variance at very small hidden layer sizes (R < 5). This behavior is likely caused by an all-or-nothing switching relationship between the hidden neurons and the output neurons. to more modular behavioral matrices (Figure 3C), we find that the effects of the activations to the hidden neurons lead to more correlated changes in motor outputs. For these cases at the bottleneck size Rc (which varies depending upon the modularity, see Figure 2A), when some of the hidden neurons are activated, they not only affect a certain number of behaviors, but all of these commands tend to belong to the same cluster, which is what we would expect, given the modular structure of the weights in Figure 2D. Moreover, activation of a neuron can lead to the complete switch from one type of behavior to the another. An example of this effect is shown in Figure 3C. The first matrix in this panel corresponds to a random matrix of behavioral commands (also point #1 in Figure 2A). In this case, a particular hidden neuron may be attributed to at most some set of motor neurons as its activation leads to activation of two of them and deactivation of other three. However, in the perfectly modular case, there are some neurons that are responsible for the encoding of the whole cluster (rightmost panel in Figure 3C). When a hidden neuron is activated, it causes nearly an entire module of behaviors to be altered. This is in keeping with the previous studies showing that stimulating individual descending neurons in flies can result in dramatic behavioral effects (Bidaye et al., 2014; Cande et al., 2018; Ding et al., 2019; McKellar et al., 2019). Averaging over several behavioral matrices and perturbations (Figure 3D), we observe that this pattern holds true in general, with more modular behavioral matrices affected more by perturbations at Rc. This effect is likely due to the different sizes of the hidden layer where the critical bottleneck size Rc (the minimum number of hidden layer neurons needed to ably represent all behavioral commands) occurs, for varying levels of modularity. As the size of the hidden layer controls the susceptibility toward perturbations (Figure 3B), highly modular behavioral matrices that have a much smaller Rc (Figure 2A), are affected to a larger extent by the perturbations. For example, a fully modular behavioral matrix has Rc = 13, but at this size of the hidden layer, it is only approximately 40% robust to such perturbations (Figure 4A). This example highlights a trade-off between efficient information compression in the bottleneck layer and robustness in case of failure. In general however, if the constraint is that the size of the hidden layer is fixed, modularity increases robustness to perturbations (Figure 4A). Thus, when constrained by a fixed size of the hidden layer, increasing the modularity and sparsity of the behavioral the network commands helps increase the robustness of to suffers the smallest if perturbations. However, is to operate the network at artificial the goal robustness possible critical bottleneck size for a given number of behavioral commands. 2.4. State-Dependency of Behaviors Previous experimental studies in fruit flies observed that optogenetically activated behaviors in flies often depend on their behavioral state prior to activation (Cande et al., 2018; Ache et al., 2019). This effect can be quantified by calculating the mutual information between the distribution of a fly’s behaviors before and after artificial neural activation. We refer to this effect as state-dependency. In essence, state-dependency implies that stimulating a neuron in the bottleneck layer will have varying— but predictable based on the input—behavioral results. In order to understand this experimentally observed effect within the framework of our model, we calculated the mutual information between the input and output distributions in the presence of an activated hidden neuron, while varying the size of the hidden layer and modularity (Figure 4B and Supplementary Material 5, see section 4 for details). This calculation provides a measure of how much information about the input distribution is contained in the output distribution in the presence of artificial activation. With the input distribution corresponding to the fly’s intended layer from behavioral output (the one-hot encoded initial Figure 1A) and the modified output corresponding to the set of behaviors that the artificial activation triggers, we see that increasing the bottleneck constraint (reducing R) lowers the overall mutual information—thus, it becomes harder to predict what the triggered behavior will be. On the other hand, a higher amount of modular structure in the output behavioral commands increases the mutual information for a fixed size of the hidden layer, with a maximum increase of around 0.8 corresponding to about a 30% increase between the two extreme values of modularity (µ = 0.18 and µ = 0.8) considered here. Thus, our model predicts that increased modular structure in the behavioral matrix not only increased robustness to perturbations (for a given N, M, and R), but also results in increased state-dependency. These results are consistent with the finding of state-dependency and modularity in the Drosophila VNC. In our model, this effect likely results from the fact that the model’s weights are segregated at higher modularities (Figure 2D), meaning that the effect of stimulating a given bottleneck-layer neuron will be limited to a relatively small number of output behaviors. It is worth mentioning that we find that to the robustness the mutual (Figure 4A information is proportional and Supplementary Figure 5) with a proportionality constant 1 M log2(N) (see section 4). This is a consequence of an absence of stereotypy in our simplified model, that is, multiple inputs don’t give the same output on forced activation. Given these results, we explored what predictions our model makes for two additional types of perturbation experiments that have not, to our knowledge, been systematically performed. First, we asked what the effects would be for deactivating, rather than activating, individual hidden layer neurons (Figure 5A). the As one might expect effect of deactivating individual neurons on the robustness of the network is qualitatively similar to that for activation. The network is more robust to the perturbation as the size for a binary encoded network, Frontiers in Behavioral Neuroscience | www.frontiersin.org 6 April 2022 | Volume 16 | Article 835753 Nande et al. Bottlenecks, Modularity, and the Neural Control of Behavior FIGURE 4 | Modularity improves robustness to perturbation and increases state-dependency for a fixed size of the hidden layer. (A) Robustness of the network averaged over the effects of activating each hidden neuron as a function of the hidden layer size, R and varying levels of modularity, µ. Here, robustness A(µ) is defined as the numbers of behaviors that are not affected upon forcefully activating a neuron in the network. (B) Average mutual information (defined as in Equation 9) between the input and output distributions after forced activation of each hidden neuron as a function of the size of the hidden layer R, and varying levels of modularity, µ. To highlight the effects of increasing modularity, we show the results relative to the lowest modularity. The figure for the absolute values is reported in Supplementary Figure 5. The mutual information turns out to be A(µ) × 1 5 iterations with the error bars corresponding to the standard deviation. Stars correspond to the Rc value for each value of modularity. M log2(N) due to the absence of stereotypy. (A,B) N = M = 100 and results are means over FIGURE 5 | Future excitation and inhibition experiments predict modularity is always associated with improved robustness. To highlight the effects of increasing modularity, we show the results relative to the lowest modularity, µ = 0.18 as A(µ) − A(µ = 0.18), where A is the robustness of the network upon de-activating each hidden neuron (A) and the robustness upon activating pairs of hidden neurons one at a time (B) defined as follows. The figure for the absolute values is reported in Supplementary Figure 6. The value of Rc for each modularity value is shown as stars. (A) Robustness of the network averaged over the effects of de-activating each hidden neuron as a function of the hidden layer size, R and varying levels of modularity, µ. (B) Robustness of the network averaged over the effects of activating a pair of hidden neurons as a function of the hidden layer size, R and varying levels of modularity, µ. (A,B) N = M = 100 and results are means over 5 iterations with the error bars corresponding to the standard deviation. of the the hidden layer increases. For any given size of hidden layer, modularity increases the network’s robustness to deactivating perturbations. to (rather Similarly, we also explored whether activating pairs of hidden layer neurons than individual neurons) state-dependency with modularity leads as well (Figure 5B). We find similar results in this case (averaging over all possible pairs of hidden layer units across many networks). increased 3. CONCLUSION Understanding how animals use their nervous system to control behavior is one of the key questions in neuroscience. A key component of most animal’s nervous system is an information bottleneck between cognitive decision-making in the brain and the neurons that are responsible for the performance of behaviors. In this work, we use a simple feed-forward neural network, similar to an autoencoder architecture that is commonly Frontiers in Behavioral Neuroscience | www.frontiersin.org 7 April 2022 | Volume 16 | Article 835753 Nande et al. Bottlenecks, Modularity, and the Neural Control of Behavior used in deep neural networks (Goodfellow et al., 2016), to understand the consequences of having such a bottleneck and identify different aspects of the network architecture that can still enable robust learning despite having such a constraint. For each set of network parameters, we identify the smallest size of the hidden layer (bottleneck size) that still allows near perfect learning. We find that increasing the sparsity of the output behavioral commands reduces this bottleneck size and increases the robustness of the network. In addition to sparsity, we find that an increased modularity in the behavioral commands helps to reduce the bottleneck size and increases robustness. This observation could provide an explanation for why such a modular structure has evolved in the behavioral commands in animals, so far observed in flies. Our simple model is also able to predict the experimentally observed state-dependency between behavioral states before and after the forced activation of hidden neurons. We find that lowering the size of the hidden layer reduces state-dependent variability, but state-dependency increases with increasing modularity for a fixed hidden layer size. Overall, the modular nature of the output makes it easier for the network to learn in the presence of a bottleneck, increases its robustness but also leads to a higher amount of state-dependency. it This model described here is obviously simplistic in architecture and dynamics (in that lacks them) and is highly unlikely to accurately describe the dynamical activity of ventral nerve cord function, where recurrent connections and temporal structure are important features of the system’s functioning (Reyn et al., 2014; Phelps et al., 2021). Future work would incorporate the effects of temporal dynamics, as well as using more biophysically realistic neurons. In addition, our model only includes discrete inputs, and understanding how graded controls over more continuous variables (e.g., walking or flight speed) would be interesting for future study. In addition, our interpretation of the results implicitly assumes that the information bottleneck is the fundamental constraint that evolution has to contend with, rather than modularity itself being the constraint and an information bottleneck being the answer that maximizes efficiency. While the ubiquity of information bottlenecks in most nervous systems provides indirect evidence toward our interpretation, future comparative studies will be needed to assess which of the two hypotheses is more likely. However, despite its simplicity, our model recapitulates several non-trivial features that are observed in experiment, and makes predictions as to the effects of artificially inhibiting neurons or of simultaneously stimulating multiple neurons, allowing for general principles of information-limited motor control to be elucidated, and new hypotheses to be tested. 4. MATERIALS AND METHODS 4.1. Network Architecture and Training To mimic the structure of the neural chord, we built a feed- forward fully-connected neural network with one hidden layer (see Figure 1). The network is constructed with the Python framework PyTorch. The input layer represents decision neurons of number N: they send the signal from the brain down the network leading to a certain behavioral output. The hidden layer of size R represents descending neurons of the neural chord: it transmits the signal down to the motor neurons, which are the output layer of the network of size M. We used the sigmoid as our activation function, serving as an approximation of the transmission of the neural signal. The functioning of the neural network can be understood explicitly from its mathematical definition. The first layer applies a linear transformation on the input sequence Ex via the weight matrix, W(1) α,β connecting neuron α in the first layer with neuron β in the following equation, a(1) β = Xα W(1) α,β xα − B(1) β , (1) while the second and last layer applies the activation function ρ(a) on a(1) as, a(2) β = Xα W(2) α,β ρ(a(1) α ) − B(2) β , (2) with ρ(a) given by the sigmoid ρ(x) = 1/(1+e−x) and B(1) (B(2)) is the bias, an additive constant. The output of the network is defined as f (x, W) ≡ a(2), where W contains all the parameters, comprising the biases. We fixed the size of the input layer (N = 100) throughout our experiments, while varying the sizes R, M of the hidden and output layers. We trained the network in the following fashion: we fixed the input and output matrices, i.e., decision and behavior matrices, respectively; we trained the network in a feed-forward manner using stochastic gradient descent with momentum and used the mean-squared error (MSE) loss function to assess learning performance; we stopped training after 105 epochs, which corresponds to when the loss curve flattens and the network is no longer learning. The output y = f (x, W) of the trained network is then binarized by rounding each entry (using a Heavyside step function centered around 0.5) and the trained weights and biases defining the network are saved for further analysis. Along with these parameters, the number of behaviors learnt, obtained by comparing each entry of the output y with the imposed behavior, is also stored. 4.2. Modularity We use the NetworkX 2.5 Python package to calculate modularity using the function ‘networkx.algorithms.community.modularity’ by treating the output matrix of behavioral commands as an adjacency matrix of a graph. Here modularity is defined as Newman (2018), µ = 1 2m Xij (cid:18) Aij − kikj 2m (cid:19) δ(ci, cj) (3) where m is the number of edges, Aij is the adjacency matrix, ki is the degree (number of connections of a node in a graph) of i and δ(ci, cj) is 1 if i and j are in the same community and 0 otherwise. 4.3. Entropy of the Behavioral Matrix The entropy of the behavioral matrix depends upon the number of behaviors N, size of the output layer M, sparsity k, number Frontiers in Behavioral Neuroscience | www.frontiersin.org 8 April 2022 | Volume 16 | Article 835753 Nande et al. Bottlenecks, Modularity, and the Neural Control of Behavior of modules m, and the noise σ associated with the modules (# of units active outside a module, for e.g., σ = 0 for perfect modularity). For a random behavioral matrix where for any output k random units are turned “on” the total entropy (in bits) is, S = Nlog2(cid:18) For a modular behavioral matrix with equal sized square modules (msize × msize, msize = M/m) the entropy (in bits) is given by, (4) M k (cid:19) S = Nlog2 (cid:20)(cid:18) msize k − σ (cid:19) × M − msize σ (cid:18) (cid:19)(cid:21) (5) 4.4. Data Generation The input data for all of our numerical experiments is always a 100 × 100 identity matrix. Each row of this matrix corresponds to the signal of performing one behavior from the output matrix. We generated several sets of output behavior matrices. In Figure 1, we varied the sparsity of the output matrix by changing the number of randomly activated units in a given row, i.e., the number of 1s. In Figure 3, we generated modular behavior matrices by introducing dense and sparse clusters into the output matrix. We start with 5 perfect clusters, i.e., no activated units are in common between 2 different clusters. Then, we generate matrices with different degree of modularity by deactivating some of the units within the cluster and activating the same number of units outside of the cluster so that the sparsity is preserved. In each case we generated 10 different behavior matrices for statistical purposes. 4.5. Checking the Robustness of the Network We checked the robustness of the network by forcefully activating one of the hidden layer neurons. This is achieved by setting its corresponding weight in the first weight matrix W(1) to an arbitrarily high value. We propagate the input matrix through the resulting perturbed network to get an output behavior matrix to be compared to the original output. In this way we can monitor how many of the original output behaviors were changed by the forceful activation. These steps are repeated for each individual hidden neuron and the results are averaged over the number of hidden neurons. 4.6. Mutual Information Calculation Mutual information (MI) between two distributions is the measure of the amount of information one distribution has about the other. For two discrete binary random variables X and Y embedded in RN with joint distribution P(X, Y) it is given by Cover and Thomas (2006), to one mapping between the input and output distributions and hence the MI is log2N. This perfect mapping gets perturbed on forced activation which can lead to one of the three different scenarios: (i) the input-output mapping is still unaffected, (ii) the input gets mapped to another output (stereotypy), and (iii) the input gets mapped to a completely different output that is not part of the original output distribution. This last case suggests that the input possess no information about the output. Suppose we have N inputs x and M outputs y where we assume that they follow a uniform distribution, that is, P(x) = 1/N and P(y) = 1/M. After forced activation, let ni be the number of inputs associated with each output yi where ni ≥ 0. This gives us P(x|yi) = 1 when ni > 0 and P(x|yi) = 0 when ni = 0. The ni mutual information then reads I(X, Y) = P(y) Xy∈Y Xx∈X P(x|y) log2 P(x|y) P(x) = = = P(yi) Xyi∈Y ′ Xx∈X ′ 1 ni log2 (cid:18) N ni (cid:19) P(yi) log2 (cid:18) N ni (cid:19) Xyi∈Y ′ 1 M Xyi∈Y ′ log2 (cid:18) N ni (cid:19) (7) (8) (9) (10) where X ′ is the set of ni inputs associated with each output yi, Y ′ is the set of m outputs with ni > 0. Note that in the absence of stereotypy that is, when ni is either 1 or 0, the mutual information becomes I(X, Y) = m M log2 (N) , (11) where m is the number of original outputs that were unaffected by perturbation and hence, the mutual information becomes proportional to our definition of network robustness. 4.7. Statistical Analysis Error bars in the figures are standard deviations that were calculated by averaging simulation results for 10 different output matrices unless specified otherwise. We used the UMAP (McInnes et al., 2018) method to visualize the structure in weight matrices. 4.8. Code Availability The code for both our simulations and statistical analysis, can be downloaded from: https://github.com/drahcir7/bottleneck- behaviors. I(X; Y) = Xx∈X Xy∈Y P(x, y)log2 P(x, y) P(x)P(y) (6) DATA AVAILABILITY STATEMENT where P(X) and P(Y) are the marginal distributions. In the absence of forced activation, the perfect learning case has a one The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author/s. Frontiers in Behavioral Neuroscience | www.frontiersin.org 9 April 2022 | Volume 16 | Article 835753 Nande et al. Bottlenecks, Modularity, and the Neural Control of Behavior AUTHOR CONTRIBUTIONS AN, VD, RR, and GZ performed all the analyses. GB conceived the project and advised on all aspects of the modeling and analysis. All authors wrote the manuscript. All authors contributed to the article and approved the submitted version. FUNDING GB was supported by the Simons Foundation and a Cottrell Scholar Award, a program of the Research Corporation for Science Advancement (25999). AN was supported by a grant from the US National Institutes of Health (DP5OD019851). GZ acknowledges support from the Paul and Daisy Soros Fellowship and the National Science Foundation Graduate Research Fellowship under Grant No. DGE1745303. RR was supported by the Swiss National Science Foundation under grant No. 200021-165509/1. ACKNOWLEDGMENTS The authors thank the organizers of the 2019 Boulder Summer School for Condensed Matter and Materials Physics for the opportunity to meet and start a collaboration on this project. SUPPLEMENTARY MATERIAL for this article can be found at: https://www.frontiersin.org/articles/10.3389/fnbeh. The Supplementary Material online 2022.835753/full#supplementary-material REFERENCES Ache, J. M., Namiki, S., Lee, A., Branson, K., and Card, G. M. (2019). circuits underlies 1132–1139. in Drosophila. Nat. Neurosci. sensory and motor State-dependent decoupling of behavioral flexibility doi: 10.1038/s41593-019-0413-4 22, Bidaye, S. S., Machacek, C., Wu, Y., and Dickson, B. J. (2014). Neuronal 97–101. of Drosophila walking direction. Science 344, control doi: 10.1126/science.1249964 Cande, J., Namiki, S., Qiu, J., Korff, W., Card, G. M., Shaevitz, J. W., et al. (2018). Optogenetic dissection of descending behavioral control in Drosophila. eLife 7, e34275. doi: 10.7554/eLife.34275 Cover, T. M., and Thomas, J. A. (2006). Elements of Information Theory, 2nd Edn. New York, NY: Wiley. Ding, Y., Lillvis, J. L., Cande, J., Berman, G. J., Arthur, B. J., Long, X., et al. (2019). Neural evolution of context-dependent fly song. Curr. Biol. 29, 1089.e7–1099.e7. doi: 10.1016/j.cub.2019.02.019 Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep Learning. Cambridge, MA: MIT Press. Gronenberg, W., and Strausfeld, N. J. (1990). Descending neurons supplying the neck and flight motor of diptera: physiological and anatomical characteristics. J. Comp. Neurol. 302, 973–991. doi: 10.1002/cne.903020420 Hsu, C. T., and Bhandawat, V. (2016). Organization of descending neurons in Drosophila melanogaster. Sci. Rep. 6, 20259. doi: 10.1038/srep20259 Kandel, E. R., Koester, J. D., Mack, S. H., and Siegelbaum, S. A. (2021). Principles of Neural Science, 6th Edn. New York, NY: McGraw-Hill. McInnes, L., Healy, J., Saul, N., and Groberger, L. (2018). UMAP: uniform J. Open Sour. Softw. 3, 861. manifold approximation and projection. doi: 10.21105/joss.00861 McKellar, C. E., Lillvis, J. L., Bath, D. E., Fitzgerald, J. E., Cannon, J. G., (2019). Threshold-based ordering of sequential courtship. Curr. Biol. 29, 426.e6–434.e6. J. H., et al. Simpson, actions during drosophila doi: 10.1016/j.cub.2018.12.019 Namiki, S., Dickinson, M. H., Wong, A. M., Korff, W., and Card, G. M. (2018). The functional organization of descending sensory-motor pathways in Drosophila. eLife 7, e34272. doi: 10.7554/eLife.34272 Newman, M. (2018). Networks, 2 Edn. Oxford, UK: Oxford University Press. doi: 10.1093/oso/9780198805090.001.0001 Phelps, J. S., Hildebrand, D. G. C., Graham, B. J., Kuan, A. T., Thomas, L. A., Nguyen, T. M., et al. (2021). Reconstruction of motor control circuits in adult Drosophila using automated transmission electron microscopy. Cell 184, 759.e18–774.e18. doi: 10.1016/j.cell.2020.12.013 Reyn, C. R., Breads, P., Peek, M. Y., Zheng, G. Z., Williamson, W. R., et al. (2014). A spike-timing mechanism for action selection. Nat. Neurosci. 17, 962–970. doi: 10.1038/nn.3741 Smarandache-Wellmann, C. R. (2016). Arthropod neurons and nervous system. Curr. Biol. 26, R960-R965. doi: 10.1016/j.cub.2016.07.063 Zacarias, R., Namiki, S., Card, G. M., Vasconcelos, M. L., and Moita, M. A. (2018). Speed dependent descending control of freezing behavior in Drosophila melanogaster. Nat. Commun. 9, 3697. doi: 10.1038/s41467-018-05875-1 Zavatone-Veth, J. A., Canatar, A., and Pehlevan, C. (2021). Asymptotics networks. of neural arXiv[preprint].arXiv:2106.00651. doi: 10.48550/arXiv.2106.00651 representation Bayesian learning finite in Conflict of Interest: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. Publisher’s Note: All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher. Copyright © 2022 Nande, Dubinkina, Ravasio, Zhang and Berman. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms. Frontiers in Behavioral Neuroscience | www.frontiersin.org 10 April 2022 | Volume 16 | Article 835753
10.1038_s41398-022-02206-3
Translational Psychiatry www.nature.com/tp OPEN ARTICLE Multidimensional analysis of behavior predicts genotype with high accuracy in a mouse model of Angelman syndrome 1,4 ✉ Joseph K. Tanas1, Devante D. Kerr1,2, Li Wang1, Anika Rai1, Ilse Wallaard3, Ype Elgersma3 and Michael S. Sidorov © The Author(s) 2022 Angelman syndrome (AS) is a neurodevelopmental disorder caused by loss of expression of the maternal copy of the UBE3A gene. Individuals with AS have a multifaceted behavioral phenotype consisting of deficits in motor function, epilepsy, cognitive impairment, sleep abnormalities, as well as other comorbidities. Effectively modeling this behavioral profile and measuring behavioral improvement will be crucial for the success of ongoing and future clinical trials. Foundational studies have defined an array of behavioral phenotypes in the AS mouse model. However, no single behavioral test is able to fully capture the complex nature of AS—in mice, or in children. We performed multidimensional analysis (principal component analysis + k-means clustering) to quantify the performance of AS model mice (n = 148) and wild-type littermates (n = 138) across eight behavioral domains. This approach correctly predicted the genotype of mice based on their behavioral profile with ~95% accuracy, and remained effective with reasonable sample sizes (n = ~12–15). Multidimensional analysis was effective using different combinations of behavioral inputs and was able to detect behavioral improvement as a function of treatment in AS model mice. Overall, multidimensional behavioral analysis provides a tool for evaluating the effectiveness of preclinical treatments for AS. Multidimensional analysis of behavior may also be applied to rodent models of related neurodevelopmental disorders, and may be particularly valuable for disorders where individual behavioral tests are less reliable than in AS. : , ; ) ( 0 9 8 7 6 5 4 3 2 1 Translational Psychiatry (2022) 12:426 ; https://doi.org/10.1038/s41398-022-02206-3 INTRODUCTION Rodent models have enabled mechanistic insights into the genetic causes and circuit-level manifestations of single-gene neurodevelopmental disorders (NDDs) [1–7]. As mechanism-based treatments are developed for NDDs, the effectiveness of such treatments is often first tested preclinically by assessing improve- ments in mouse behavior. Behavioral phenotypes span multiple domains in individuals with NDDs (e.g., cognitive, motor, seizures, sleep), and a wide range of corresponding behavioral assessments have been developed and deployed in mouse models [8–11]. Accurately measuring phenotypic severity across multiple beha- vioral domains is critical to properly assess the effectiveness of treatments in rodent models of NDDs. We hypothesized that multidimensional analysis of mouse behavior would enable quantification of overall behavioral severity, aggregated across multiple domains. Here we define multidimensional analysis as the multi-step process of: (a) reducing the dimensionality of large behavioral datasets using principal component analysis (PCA) [12, 13], (b) clustering data in principal component space using k-means clustering, and (c) assessing whether behaviorally defined clusters align with animal genotype. Dimensionality reduction and clustering have been validated in various mouse behavioral contexts [14–22]. Here, we tested the hypothesis that multidimensional analysis of mouse behavioral data could accurately distinguish the genotype of Ube3a mutants (a model of Angelman syndrome (AS)) from wild- type littermates. AS is an ideal disorder for testing the effective- ness of multidimensional analysis because of recent progress in developing mechanism-based treatments [23] and because behavioral testing can be performed reliably in Angelman model mice [24]. AS is a NDD caused by lack of expression of the maternal allele of UBE3A, an E3 ubiquitin ligase located on chromosome 15 [25–27]. Individuals with AS have a multifaceted behavioral phenotype that typically includes cognitive impairment, motor impairment, lack of speech, seizures, and disrupted sleep [28–30]. While mutations in maternal UBE3A are sufficient to cause AS, the majority of cases (~70%) are caused by deletions of a region of maternal chromosome 15q11–13 spanning UBE3A and neighbor- ing genes [29]. The paternal UBE3A allele is epigenetically silenced in neurons by expression of a UBE3A antisense transcript (UBE3A- ATS) [31, 32], and neuronal paternal Ube3a imprinting is conserved [33] [34]. Multiple approaches have successfully unsilenced the dormant paternal Ube3a allele in mice [35–42], and one such approach (antisense oligonucleotides targeted to UBE3A-ATS) is currently in early-stage clinical trials (NCT04428281, NCT04259281, NCT05127226). Other AS treatments such as gene replacement therapy and targeting downstream processes are also in preclinical development [43]. Developing a pipeline to test new AS treatments in mice will be in the mouse model of AS (Ube3am-/p+) 1Center for Neuroscience Research, Children’s National Hospital, Washington, DC, USA. 2Howard University, Washington, DC, USA. 3Department of Clinical Genetics and the ENCORE Expertise Center for Neurodevelopmental Disorders, Erasmus MC, Rotterdam, Netherlands. 4Departments of Pediatrics and Pharmacology & Physiology, The George ✉ Washington University School of Medicine and Health Sciences, Washington, DC, USA. email: msidorov@childrensnational.org Received: 5 May 2022 Revised: 20 September 2022 Accepted: 23 September 2022 2 J.K. Tanas et al. critical to evaluate their success, regardless of treatment mechan- ism. Recent work established a “gold standard” behavioral battery consisting of five tests (rotarod, open field, marble burying, nest building, forced swim) that are reliably impaired in Ube3am-/p+ mice [24] and are sensitive to treatment [40, 44–46]. Using this battery, we hypothesized that multidimensional analysis: (a) would enable quantification of behavior across multiple domains as a single “severity score,” (b) that this severity score would be a reliable indicator of Ube3a genotype, and (c) that this severity score is sensitive to treatment. METHODS AND MATERIALS Animals We performed multidimensional analysis using three mouse behavioral datasets. All datasets assessed behavior in male and female AS model mice (Ube3am-/p+) [34] and wild-type littermate controls (WT; Ube3am+/p+) with experimenters blind to genotype. For all datasets, experimental WT and Ube3am-/p+ littermates were generated by crossing female Ube3am+/p- mice and male WT mice. Dataset 1 used experimental mice on an F1 hybrid 129S2-C57BL6/J background, and Datasets 2 and 3 used experimental mice on a congenic C57BL6/J background. All experimental protocols were conducted in accordance with the European Commission Council Directive 2010/63/EU (CCD approval AVD101002016791; Dataset 1), or were approved by the Institutional Animal Care and Use Committee (IACUC) of Children’s National Medical Center (Dataset 2). Dataset 3 contained only previously published data and no additional mouse experiments. Dataset 1 was used for the majority of analyses (Figs. 1–3, 4a, 5, S1–S8, and S10) to assess the effectiveness of multidimensional analysis at predicting Ube3a genotype based on behavior. Dataset 1 included 286 total mice (WT: n = 148, Ube3am-/p+: n = 138) run across 10 cohorts at Erasmus Medical Center. Behavioral data from 8 of these 10 cohorts (n = 231/286 mice) were previously published [24] and 2 additional cohorts (n = 55 mice) were also included. A detailed table showing the genotypes, sex, and behavioral tests performed in each of the ten cohorts is shown in Fig. S1. Behavioral testing was performed in P60–P90 mice. Dataset 2 was used to test the hypothesis that multidimensional analysis of behavior can effectively predict Ube3a genotype using small sample sizes (Fig. 4b and S9). Dataset 2 included 24 mice (WT: n = 12, Ube3am-/p+: n = 12) tested at Children’s National Research Institute. Behavioral testing was performed in P60-P90 mice. Dataset 3 was previously published (Wolter et al.; Supplementary Fig. 4) [41] and was used to assess whether multidimensional analysis could detect behavioral improvement in Ube3am-/p+ mice treated with CRISPR- Cas9-based targeted treatment to unsilence the dormant paternal Ube3a allele (Fig. 6 and S11). Mice were treated with either viral expression of a SaCas9 gRNA targeting the Ube3a-ATS locus (Sajw33), or a negative control gRNA that did not unsilence paternal Ube3a. Bilateral i.c.v. AAV delivery of Sajw33 or control occurred at both E15.5 and P1 within the same animals, and behavioral testing began at 4 weeks and continued through 40 weeks [41]. Behavioral data included WT + control (n = 34), Ube3am-/p+ + control (n = 25), and Ube3am-/p+ + treatment (Sajw33; n = 32) groups. Behavioral testing Mice in Datasets 1 and 2 were weighed and then underwent a series of behavioral tests in the same order: rotarod, open field, marble burying, nest building, and forced swim (Fig. 1a). A subset of mice in Dataset 1 lacked data from three of these tests (weight, open field, nest building; Fig. S1a). Methods for Dataset 1 were previously published [24] and methods for Dataset 2 (below) were based upon Sonzogni et al. [24], with minor modifications. Rotarod. Mice were placed on a rotating bar that accelerated from 4 to 40 rpm across 5 min at an acceleration rate of 7.2 rpm2 (Ugo Basile model #47600). Trials were complete once the mouse fell off, if three consecutive Fig. 1 Ube3am-/p+ mice have robust behavioral impairments across multiple domains. a Experimental timeline, as described and performed by Sonzogni et al. (eight cohorts), plus two additional unpublished cohorts. b Weight (WT: n = 118, AS: n = 110). c Rotarod performance (WT: n = 148, AS: n = 138). d Distance traveled and e time spent in the center of an open field (WT: n = 118, AS: n = 110). f Marble burying performance (WT: n = 148, AS: n = 138). g Nest building performance (WT: n = 109, AS: n = 100). h Forced swim performance (WT: n = 148, AS: n = 138). See Fig. S1 for full breakdown of tests performed in each of ten cohorts. Data represent mean ± SEM; ****p < 0.0001; black: WT, red: Ube3am-/p+ (AS). i Methods for multidimensional behavioral analysis and genotype validation. First, eight behavioral measures were included in multidimensional analysis. Second, principal component analysis (PCA) reduces the dimensionality of the behavioral dataset. Each point represents one animal; these points are schematized and are not real data. Third, mice are clustered into two groups using k-means clustering by their proximity in PC space. Finally, validation reflects a comparison of clusters with the known genotype of animals. Translational Psychiatry (2022) 12:426 wrapping rotations were made, or if 5 min elapsed. Each day, the results of two trials with an inter-trial interval of one hour were averaged together. Experiments were run over the span of five consecutive days, with a two day interval before open field testing. Open field test. Mice were placed into a 42 cm square open field arena (AccuScan Instruments, Inc., Columbus, OH) and were allowed to freely move for a single 30 min trial. The center square was defined as 21 cm × 21 cm. The data was collected using the open field activity monitoring system (Omnitech Electronics, Inc. SuperFlex Open Field System) which uses photocell emitters and receptors forming an x–y grid of infrared beams. Total distance moved and time spent in the center square were recorded using infrared beam break information. Marble burying. Mice were placed individually in a 16 × 8 in cage with ~4 cm of bedding (Bed-o’Cobs 1/4” bedding) and 20 black glass marbles arranged in a 5 × 4 array for a single 30 min trial. A marble was considered buried if it was >50% covered with bedding at the end of the trial. Immediately after marble burying, mice were habituated to Nest building. single housing as well as new nesting material (Bio-Rad 7.5 × 10 cm extra thick block filter paper; 11 ± 1 g) for 5–7 days prior to testing. During testing, new nesting material was introduced on day 1 and unused material was weighed daily across five days. Forced swim test. Forced swim testing was performed on the same day immediately following the final day of nest building. Mice were placed in a 9 in cylindrical × 9.25 in tall tank filled with 23 ± 1 °C water to a height of ~60% of the container’s height. Trials lasted 6 min: the first 2 min were an acclimation period and the last 4 min were to record immobility (lack of movement or only necessary movements to keep head above water) as a percentage of total recording time. Immobility was recorded manually using a stopwatch. Mice in Dataset 3 underwent a different series of behaviors in consistent order: hindlimb clasping, fear conditioning, rotarod (again) (Fig. S11a). Methods for Dataset 3 were previously published [41]. rotarod, open field, marble burying, Multidimensional analysis of mouse behavior Multidimensional analysis of behavior consisted of a series of steps: data selection, standardization, principal component analysis (PCA), k-means clustering, and validation (Fig. 1i) [22]. Data selection: Eight total measures (weight, rotarod day 1, rotarod day 5, open field distance traveled, open field center time, marbles buried, nest building, forced swim floating time) from a series of six tests (weight, rotarod, open field, marble burying, nest building, forced swim) were included in multidimensional analysis. Tests with multiple measures include rotarod (day 1 and day 5 performance) and open field (total distance and time spent in the center of the arena). Measures considered redundant intermediate, non-independent (e.g., timepoints for rotarod and nest building) were excluded a priori from multidimensional analysis. For the majority of analyses (Figs. 2, 3a–d, 4a, 5, S3–S6, S8c, and S10), we included the subset of animals that underwent all six tests (WT: n = 88, Ube3am-/p+: n = 82). One outlier was excluded from analysis based on its position in 2PC space (PC1: −2.49, PC2: 8.62) using Grubbs’ test for outliers (alpha = 0.05), leaving a total n = 81 Ube3am-/p+ mice. Standardization: All measures were standardized using a z-score (z = (data point − group mean)/standard deviation) for different units across tests. Prior to standardization, we tested whether each behavioral measure showed significant sex differences using a two- way ANOVA with sex and genotype as factors (Fig. S2). All behaviors that showed either a significant main effect of sex or sex × genotype interaction were standardized separately in male and female animals (Fig. 1i), except where noted (Fig. S6; a separate analysis to assess the effectiveness of multidimensional analysis if sex is not accounted for). Measures with no sex differences were standardized using the entire group. Principal component analysis: We performed PCA using the pca() function in MATLAB and calculated the amount of variance explained by each PC using a Scree plot (Figs. S3, S9h, and S11b) and the loading distribution of principal components using the coefficient outputs from PCA (Figs. 5a, b and S11c). k-means clustering: k-means clustering was performed in principal component space using the kmeans() function in MATLAB with k = 2 clusters. Except where noted (Figs. 5, 6c, S4, and S11d, g), clustering was performed in 2PC space. In Figs. 5c, 6c, and S11d, g, one PC was used to account Translational Psychiatry (2022) 12:426 J.K. Tanas et al. 3 for clustering. Figure 5d, e evaluated using three and four PCs for clustering. In Fig. S4, clustering was performed on raw data (not in PC space). Validation: We compared the actual genotypes of animals to their assigned cluster and calculated the percentage correct (accuracy of clustering, Figs. 2c, 3b, c, f–h, 4b, 5c–e, S5a, S6, S10, and S11d). In order to assess the generalizability of multidimensional analysis, we varied the input parameters of analysis in three ways (Fig. 3). First, we performed multidimensional analysis on Dataset 1 (n = 169; standardized by in 2 PC space under 11 conditions (Fig. 3a–c). Each condition sex) represented the removal of a single behavioral measure or a complete behavioral test. Next, we performed multidimensional analysis using every possible combination of 3–8 total measures (Fig. 3d). Finally, we performed multidimensional analysis on different combinations of data where more animals could be included with partial behavioral profiles (Fig. 3e–h and S1). This approach resulted in four conditions with the number of behavioral measures ranging from 4 to 8 and the number of animals ranging from 169 to 286 per condition. To assess the minimal sample size needed for multidimensional analysis to effectively classify Ube3a genotype, we performed a bootstrap analysis of Dataset 1 (n = 169) using 2 PCs for clustering (Fig. 4a). For n = 3 to n = 30 with a step of n = 1, we randomly selected n animals per genotype from the overall dataset (n = 169), with replacement, and performed multidimensional analysis. We repeated 10,000 trials for each n and report the average clustering accuracy. To determine whether multidimensional analysis results in false positive effects, we performed multidimensional analysis in a homogenous sample (all 36 wild-type females). We randomly assigned mice to two groups (to model two genotypes with no true behavioral difference; Fig. S7). Behavioral rescue with Ube3a reinstatement by Cas9 gene therapy Behavioral data was available for a total of n = 91 mice in Dataset 3. However, every behavioral test was not performed in every animal. Thus, we performed multidimensional analysis in two subsets (“conditions”) of the total sample (Fig. S11e). Condition 1 (Fig. 6) included n = 33 total mice where behavioral data was available for each of thirteen total measures. Condition 2 (Fig. S11f, g) included n = 86 total mice where behavioral data was available for six of thirteen total measures. Data were standardized by sex for measures where statistically significant sex differences were observed between WT + control and AS + control groups in the total sample (P90 weight, open field center time). Statistics Statistical analysis was performed using GraphPad Prism 9 and MATLAB R2019a, R2021a, and R2022a (Mathworks). Group comparisons on individual behavioral tests (Figs. 1 and S9) were made using Student’s t-tests (weight, open field distance, open field center time, marble burying, forced swim) and two-way RM ANOVA (rotarod, nest building). Assessment of sex differences (Figs. 2, S5c, and data not shown to accompany Fig. 6/S11) were made using two-way ANOVA. Effect of treatment was assessed using one- way ANOVA and post hoc Tukey’s multiple comparison test in 1 PC space (Figs. 6c and S11g). Figure 3d used one-way ANOVA and post hoc Tukey’s multiple comparison test. Figure S7d used Student’s t test. Fisher’s exact test was used to determine if clustering accuracy was statistically different between test conditions (Figs. 3a–c, S4, and S6). For all figures, *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001. RESULTS Ube3am-/p+ mice have robust behavioral impairments Ten independent cohorts totaling 286 mice (Fig. S1; 26–30 mice per cohort) performed a series of behavioral tests in order: weight, rotarod, open field, marble burying, nest building, and forced swim (Fig. 1a). Behavioral data from eight of these cohorts (n = 231) were previously reported by Sonzogni and colleagues [24]. In the complete dataset, Ube3am-/p+ mice showed increased t(226) = 4.428, p < 0.0001), weight impaired rotarod performance (Fig. 1c; main effect of genotype: F(1,284) = 103.8, p < 0.0001), and impaired rotarod learning (Fig. 1c; genotype × time interaction: F(4,1136) = 6.792, p < 0.0001) relative to WT controls. Ube3am-/p+ mice were hypoactive in an open field (Fig. 1b; 4 J.K. Tanas et al. Fig. 2 Multidimensional behavioral analysis predicts Ube3a genotype with high accuracy. a PCA plus k-means clustering categorized two clusters, with each dot representing one mouse’s behavioral profile in 2PC space (black: cluster 1, red: cluster 2). b Actual genotype of animals (black: WT, red: AS). c An overlay of a, b with animals clustered incorrectly labeled in blue. Clustering accuracy was 94.7%. (Fig. 1d; t(226) = 8.874, p < 0.0001) despite normal time spent in the center of the arena (Fig. 1e; t(226) = 1.348, p = 0.1789). Ube3am-/p+ mice also showed impairments on marble burying (Fig. 1f; t(284) = 16.79, p < 0.0001), nest building (Fig. 1g; main effect of genotype: F(1,207) = 176.1, p < 0.0001), and forced swim tests (Fig. 1h; t(284) = 15.73, p < 0.0001) relative to WT controls. Prior work found significant sex differences in a subset of these behaviors [24], and we confirmed with the complete dataset that six of eight behavioral measures showed statistically meaningful sex differ- ences (Fig. S2). four Multidimensional analysis of behavior predicts Ube3a genotype with high accuracy From the group of six behavioral tests, we selected eight non- redundant measurements to include in multidimensional analysis (Fig. 1i). We performed multidimensional analysis using data from mice where all behavioral tests were performed within individual animals (n = 169/286; 6/10 cohorts; Fig. S1). Multidimensional analysis consisted of standardization, principal steps: component analysis (PCA), k-means clustering, and validation (Fig. 1i). First, all measures were standardized using a z-score and measures with sex differences were standardized separately by sex. PCA revealed that two principal components (PCs) are likely sufficient to capture a majority of variance in the dataset (Fig. S3); thus, we performed k-means clustering in 2PC space (Fig. 2a). Validation comparing clusters (Fig. 2a) to the actual genotype of animals (Fig. 2b) demonstrated that multidimensional analysis was 94.7% accurate in predicting Ube3a genotype from behavior (Fig. 2c). Multidimensional analysis performed better than k-means clustering of performance on individual assessments, which ranged from ~55–85% accuracy (Fig. S4). Male and female mice were equally distributed in PC space (Fig. S5), confirming that sex differences in behavior were accounted for by standardization, and likely do not represent an additional source of meaningful variance. Not accounting for the sex dependence of behaviors slightly reduced the accuracy of multidimensional analysis (from 94.7% to 91.1%), though this difference was not statistically meaningful (Fig. S6). To confirm the validity of multidimensional analysis, we demonstrated that this approach does not detect “false positive” differences in behavior between two randomly assigned groups within a homogenous group of animals (all wild- type females) (Fig. S7). Multidimensional analysis remains effective with different combinations of behavioral input For multidimensional analysis to be a valuable tool to quantify behavior in mouse models of AS and related disorders, it should generalize across multiple combinations of behavioral input. To address this question, we first performed multidimensional analysis on the same dataset (n = 169) under conditions where each individual measurement (e.g., rotarod day 5) or overall test (e.g., rotarod) were excluded from analysis. Removal of individual measures or tests resulted in a range of 92.3–96.4% accuracy, which was not statistically different from the baseline of 94.7% (Fisher’s exact tests; lowest p = 0.5092; Figs. 3a–c and S8a). Next, we assessed the clustering accuracy in each of 219 possible combinations of including between three and eight behavioral measures in analysis. Clustering accuracy decreased on average as the number of measures decreased (Fig. 3d; one-way ANOVA, F(4,261) = 21.31, p < 0.0001). We also expanded the dataset to include all animals (n = 286), and defined four conditions where different combinations of behavior were available for different subsets of animals (Fig. S1). As the number of measures decreased, the accuracy of multidimensional analysis decreased despite the sample size increasing (Figs. 3e–h and S8b). Together, these data suggest that multidimensional analysis generalizes well across different combinations of behavioral data in the Ube3am-/p+ mouse model, and gains effectiveness as more behavioral tests are included in analysis. Multidimensional analysis accurately predicts Ube3am-/p+ genotype with reasonable sample sizes for mouse behavior In practice, a behavioral study requiring >80 mice per genotype to detect group differences would likely be time and cost prohibitive. Here, we asked whether multidimensional analysis would retain high accuracy in groups with smaller sample sizes. To address this question, we performed a bootstrap analysis to predict the clustering accuracy that could be achieved across sample sizes Translational Psychiatry (2022) 12:426 J.K. Tanas et al. 5 Fig. 3 Multidimensional analysis remains effective with different combinations of behavioral data. a Removal of individual measures and individual behavioral tests resulted in clustering accuracies ranging from 92–97%. Each bar represents a condition where one measure or test was excluded; black bar represents the full dataset (n = 169) with nothing excluded. b The lowest accuracy achieved was 92.9% with weight excluded from analysis. c The highest accuracy achieved with removal of a single test was 96.5% with rotarod excluded. d Average accuracy across all combinations of measures used ranged from 82.5% (3 measures) to 93.1% (7 measures). Data represent mean ± SEM; ****p < 0.0001, one-way ANOVA. 8 measures: n = 1, 7 measures: n = 8, 6 measures: n = 28, 5 measures: n = 56, 4 measures: n = 70, 3 measures: n = 56. e Multidimensional analysis in larger sample sizes was achieved by including fewer behavioral tests. f–h Accuracy of different conditions shown in e ranges from 89.9% to 92.2%. ranging from n = 3–30 per group (Figs. 4a and S8c). We then performed multidimensional analysis on a behavioral cohort tested prospectively in a different laboratory with 12 animals per group (Dataset 2, Fig. S9). The bootstrap analysis predicted an accuracy of 90% using n = 12, and multidimensional analysis achieved 100% accuracy in this new cohort (Figs. 4b and S9). We hypothesized that accuracy was higher in a new cohort because the bootstrap analysis pulls animals randomly across multiple cohorts, introducing inter-cohort variability. We tested this hypothesis by performing multidimensional analysis separately in each of six individual cohorts from Dataset 1 (Fig. S9). Accuracy in each of these cohorts also outperformed the expectations of the bootstrap analysis, ranging from 92–100% with sample sizes of ~13–15 animals per genotype (Fig. 4a; open circles). The cohort sizes tested here are comparable to the sample sizes determined by Sonzogni et al. to be required to detect group level differences on individual behavioral tests (n = 7–21, depending on test) [24]. Overall, this work confirms that multidimensional analysis of behavior is highly accurate in Ube3a mice using typical sample sizes for mouse behavior. Fig. 4 Multidimensional analysis retains high accuracy with reasonable sample sizes. a Bootstrap analysis was performed using Dataset 1, and 10,000 trials per n. Open circles indicate the clustering accuracy within six individual cohorts. The bold circle two cohorts with the same (x,y) represents the overlap of coordinates (14, 100%). b Clustering accuracy using Dataset 2 (n = 12 per genotype) was 100%. Translational Psychiatry (2022) 12:426 6 J.K. Tanas et al. Fig. 5 Principal component 1 (PC1) is sufficient for predicting Ube3a genotype with high accuracy. a Loadings for PC1–PC4 for each behavioral measure. b Loadings for PC1 and PC2 for each measure, plotted in 2PC space. All behavioral measures except open field correlate more strongly with PC1, and both open field measures correlate more strongly with PC2. c Clustering accuracy using 1 PC is 94.7%. d Clustering accuracy using 3 PCs is 94.7%. e Increasing the number of PCs used for clustering does not have a major impact on clustering accuracy. improvement in Ube3am-/p+ mice following treatment. Fig. 6 Multidimensional analysis has the sensitivity to detect behavioral a Multidimensional analysis used 12 behavioral measures from Wolter et al. (Dataset 3). *Indicates sex difference and standardization by sex. b Animals with paternal Ube3a unsilencing (purple open circles) have a qualitatively intermediate behavioral profile between Ube3am-/p+ mutants (AS; red) and wild-type controls (WT; black). c In 1PC space, treated AS mice show overall behavioral improvement relative to control AS mice, but not full rescue. Full timeline of behavioral tests and analysis using different combinations of behavioral tests is shown in Fig. S11. Data represent mean ± SEM; ****p < 0.0001. Translational Psychiatry (2022) 12:426 Multidimensional analysis remains effective using a single PC for clustering A potentially valuable use of multidimensional behavioral analysis is the possibility of summarizing an animal’s overall phenotypic severity as a single number (PC1). Such an approach will only be valuable if (a) PC1 represents a substantial amount of the total variability in the dataset and (b) PC1 alone is sufficient to accurately predict Ube3a genotype. PC1 accounts for 37.7% of variance in the dataset (Figs. S3 and 5a) and correlates strongly with each behavioral measure tested (Fig. 5a, b). Clustering accuracy remained high (94.7%) using a single principal component for k-means clustering (Fig. 5c). Increasing the number of PCs to 3–4 (accounting for up to 75% of total variation) provided little additional benefit in predicting Ube3a genotype (Fig. 5d, e). Overall, these results suggest that PC1 alone is sufficient to accurately predict Ube3a genotype. Multidimensional analysis can detect behavioral improvement following paternal Ube3a unsilencing Behavioral testing in mouse models of neurodevelopmental disorders is often used to test the preclinical effectiveness of treatments [10]. We performed multidimensional analysis of behavior in a group of mice where CRISPR/Cas9-based targeting of Ube3a-ATS enabled unsilencing of the paternal Ube3a allele and re-expression of UBE3A protein [41]. In this cohort (Dataset 3), the series of behavioral tests performed was similar but not identical to the series of tests performed by Sonzogni and colleagues [24] (Figs. 6a and S11a). In this dataset, PC1 represented 31.0% of total variance and clustering accuracy was 100% between WT and Ube3am-/p+ non-drug control groups when using 1 PC for clustering (Fig. S11b–d). CRISPR/Cas9 treatment resulted in an amelioration, but not full correction, of overall behavioral severity in Ube3am-/p+ mice as measured by PC1 (Fig. 6b, c; main effect of group: F(2,30) = 93.66, p < 0.0001; post hoc WT/control vs. AS/ control: p < 0.0001; post hoc AS/control vs. AS/treatment: p < 0.0001). We observed a significant effect of treatment using PC1 under multiple conditions where different combinations of behavioral measures were included in the analysis (Fig. S11e–g). PUMBAA: A graphical user interface for multidimensional analysis of behavior Multidimensional analysis may be valuable for other behavioral datasets in Ube3am-/p+ mice and mouse models of related disorders. To enable widespread use of multidimensional beha- vioral analysis, we developed a graphical user interface for phenotyping using a multidimensional behavioral analysis algo- rithm (PUMBAA; Fig. S12; https://github.com/sidorovlab/PUMBAA). PUMBAA runs in a MATLAB environment but does not require users to have prior MATLAB coding knowledge. PUMBAA enables user control of analysis parameters for all steps including data selection, data standardization, principal component analysis, clustering, and validation. DISCUSSION Multidimensional analysis of behavior (Fig. 1) correctly predicted Ube3a genotype in a mouse model of AS with high accuracy (Fig. 2). This approach retained high accuracy with multiple combina- tions of behavioral data (Fig. 3) and in behavioral cohorts of a manageable size (n = ~12–15; Fig. 4). Principal component analysis enabled the simplification of an animal’s overall behavioral profile to a single severity score (PC1, Fig. 5) that (Fig. 6). We demonstrated improvement propose that multidimensional behavioral analysis provides a generalizable approach to quantify behavioral impairment and screen treatments preclinically in rodent models of neurodevelop- mental disorders. following treatment Translational Psychiatry (2022) 12:426 J.K. Tanas et al. 7 In the context of AS, the primary value of multidimensional behavioral analysis is to assess the efficacy of treatments in rodent models. Multiple promising treatments for AS are currently under development at various stages of clinical and preclinical testing [23, 43]. Current and future treatments will span multiple mechanisms of action, including directly targeting Ube3a expres- sion, targeting downstream Ube3a protein targets, and more generalized symptom-based approaches. Simplifying mouse behavior to an overall severity score will be valuable for measuring overall improvement following treatment, and for assessing the effect of treatment across development. Our analysis found that each behavioral measure in the well-established Sonzogni battery [24] contributes roughly equally to PC1 (loadings range from 0.22–0.42; removal of single measures resulted in 92.3–96.4% accuracy), and that different combinations of behavioral inputs can be used to achieve high accuracy (Figs. 3 and 5a, b). These results suggest that multidimensional behavioral analysis can generalize well across different behavioral batteries and across different laboratories. In addition, multidimensional analysis is a tool that can generalize across mouse strains, mouse lines, and across species to be applied to the new Ube3am-/p+ rat model [47]. Our study assessed behavior using the Ube3am-/p+ mouse developed by Jiang and colleagues [34], which has been the most commonly used preclinical model for AS research. However, a limit to this mouse line is that it mimics the loss of UBE3A but not other nearby genes that are also deleted in the majority of individuals with AS. In humans, AS clinical severity is typically in individuals with a deletion genotype [48]. Multi- greater dimensional analysis can be used in the future to test whether AS mouse models with larger deletions [49] have a more severe behavioral phenotype than the Ube3am-/p+ model. Multidimensional analysis revealed that overall behavioral severity was improved but not fully corrected by paternal Ube3a unsilencing at E15.5 + P1. Incomplete behavioral improvement is consistent with the results of Wolter et al. on certain individual tests (e.g., rotarod, brain weight), though they did report full correction of impairments in hindlimb clasping [41]. We hypothe- size that overall behavioral rescue was incomplete due to the amount of UBE3A reinstatement achieved using CRISPR/Cas9- based unsilencing: Wolter et al. achieved reinstatement of UBE3A protein to ~40% of WT levels in a subset of animals where Western blotting was performed [41]. An advantage of multidimensional analysis is that for future treatment studies, PC1 (as a readout of overall behavioral severity) can be correlated with the degree of UBE3A reinstatement achieved within individual animals. Beyond AS, this study provides proof of concept that multi- dimensional behavioral analysis can be applied to rodent models of related disorders. Prior studies have generally applied principal component analysis to rodent behavioral data in two contexts: (a) to assess which subset of behavioral measures are most relevant or valuable [14, 17, 18], and (b) to attempt to categorize two or more groups based on their behavior [15, 16, 22]. Here, we applied both approaches to quantify behavior in Ube3a mutants. Multi- dimensional analysis was especially effective in Ube3am-/p+ mice because behavioral impairments on individual tests are so reliable and widespread in this line [5]. However, behavioral phenotypes in other lines are often less robust. We hypothesize that multi- dimensional analysis will be particularly valuable for detecting subtle behavioral differences in established mouse models and for screening behavior in new mouse models of rare disorders. Multidimensional analysis used in this context is unlikely to result in false positive effects of genotype (Fig. S7). To enable widespread use of multidimensional behavioral analysis, we developed a graphical user to simplify and generalize analysis methods. PUMBAA enables users to perform customized multidimensional analysis in a MATLAB environment without any prior coding knowledge. interface (PUMBAA; Fig. S12) 8 J.K. Tanas et al. Our work identified a number of practical considerations for future studies using multidimensional analysis to quantify in rodent models of NDDs. First, multidimensional behavior analysis can only be performed in datasets where longitudinal behavioral testing is performed within animals. This requirement places certain limits on experimental design, such as the inclusion of tests that may be terminal (e.g., audiogenic seizures). Long- itudinal testing also presents challenges when assessing the efficacy of treatment. For example, a four-week longitudinal behavioral battery would not be appropriate for a treatment expected to last two weeks. In addition, accounting for sex differences is an important consideration for multidimensional analysis, as sex differences in behavior have been reported in Ube3am-/p+ mice and in rodent models of related disorders (Fig. S2) [24, 45, 50–53]. Accounting for sex differences in behavior resulted in a slight but not statistically meaningful improvement in the accuracy of multidimensional analysis from ~91% to ~95% (Figs. 2 and S6). Finally, our results suggest that cross-cohort behavioral variability decreases the accuracy of multidimensional analysis performed across multiple behavioral cohorts (Fig. 4). The inclusion of behavioral assessments with no group differences would not “dilute” the effectiveness of PCA; thus, multidimen- sional analysis is well-suited for analysis of broad behavioral phenotyping regimens [12, 13]. Future studies using multidimen- sional analysis of behavior may also consider additional methods (beyond percent accuracy) to quantify the degree to which two genotypes can be distinguished in PC space. Other potentially include the distance between cluster valuable parameters centroids and the spread of data within individual clusters. Overall, multidimensional behavioral analysis enables represen- tation of behavior in Ube3am-/p+ mice as a single severity score that is reliably different from wild-type controls and is sensitive to treatment. Multidimensional behavioral analysis represents a tool that may be used to evaluate the effectiveness of preclinical treatments for AS and related neurodevelopmental disorders. REFERENCES 1. Lombardi LM, Baker SA, Zoghbi HY. MECP2 disorders: from the clinic to mice and back. J Clin Investig. 2015;125:2914–23. 2. Braat S, Kooy RF. The GABAA receptor as a therapeutic target for neurodeve- lopmental disorders. Neuron. 2015;86:1119–30. 3. Krueger DD, Bear MF. Toward fulfilling the promise of molecular medicine in fragile X syndrome. Annu Rev Med. 2011;62:411–29. 4. Delling JP, Boeckers TM. Comparison of SHANK3 deficiency in animal models: implications. J Neurodev phenotypes, treatment strategies, and translational Disord. 2021;13:55. 5. Rotaru DC, Mientjes EJ, Elgersma Y. Angelman syndrome: from mouse models to therapy. Neuroscience. 2020;445:172–89. 6. Winden KD, Ebrahimi-Fakhari D, Sahin M. Abnormal mTOR activation in autism. Annu Rev Neurosci. 2018;41:1–23. 7. Golden CE, Buxbaum JD, De Rubeis S. Disrupted circuits in mouse models of autism spectrum disorder and intellectual disability. Curr Opin Neurobiol. 2018;48:106–12. 8. Silverman JL, Ellegood J. Behavioral and neuroanatomical approaches in models of neurodevelopmental disorders: opportunities for translation. Curr Opin Neurol. 2018;31:126–33. 9. Sukoff Rizzo SJ, Crawley JN. Behavioral phenotyping assays for genetic mouse models of neurodevelopmental, neurodegenerative, and psychiatric disorders. Annu Rev Anim Biosci. 2017;5:371–89. 10. Silverman JL, Yang M, Lord C, Crawley JN. Behavioural phenotyping assays for mouse models of autism. Nat Rev Neurosci. 2010;11:490–502. 11. Gulinello M, Mitchell HA, Chang Q, Timothy O’Brien W, Zhou Z, Abel T, et al. Rigor and reproducibility in rodent behavioral research. Neurobiol Learn Mem. 2019;165:106780. 12. Lever J, Krzywinski M, Altman N. Principal component analysis. Nat Methods. 2017;14:641–2. 13. Jolliffe IT, Cadima J. Principal component analysis: a review and recent devel- opments. Philos Trans A Math Phys Eng Sci. 2016;374:20150202. 14. Zimprich A, Ostereicher MA, Becker L, Dirscherl P, Ernst L, Fuchs H, et al. Analysis in the German Mouse Clinic. J Neurosci Methods. locomotor behavior of 2018;300:77–91. 15. Vannoni E, Voikar V, Colacicco G, Sanchez MA, Lipp HP, Wolfer DP. Spontaneous behavior in the social homecage discriminates strains, lesions and mutations in mice. J Neurosci Methods. 2014;234:26–37. 16. Colomb J, Winter Y. Creating detailed metadata for an R Shiny analysis of rodent behavior sequence data detected along one light-dark cycle. Front Neurosci. 2021;15:742652. 17. Sierra C, De Toma I, Cascio LL, Vegas E, Dierssen M. Social factors influence behavior in the novel object recognition task in a mouse model of Down syn- drome. Front Behav Neurosci. 2021;15:772734. 18. Heinz DE, Schottle VA, Nemcova P, Binder FP, Ebert T, Domschke K, et al. Exploratory drive, fear, and anxiety are dissociable and independent components in foraging mice. Transl Psychiatry. 2021;11:318. 19. Muehlmann AM, Bliznyuk N, Duerr I, Lewis MH. Repetitive motor behavior: further characterization of development and temporal dynamics. Dev Psychobiol. 2015;57:201–11. 20. Karamihalev S, Brivio E, Flachskamm C, Stoffel R, Schmidt MV, Chen A. Social dominance mediates behavioral adaptation to chronic stress in a sex-specific manner. Elife. 2020;9:e58723. Inayat S, Singh S, Ghasroddashti A, Qandeel, Egodage P, Whishaw IQ, et al. A Matlab-based toolbox for characterizing behavior of rodents engaged in string- pulling. Elife. 2020;9:e54540. 21. 22. Popovitz J, Mysore SP, Adwanikar H. Neural markers of vulnerability to anxiety outcomes after traumatic brain injury. J Neurotrauma. 2021;38:1006–22. 23. Elgersma Y, Sonzogni M. UBE3A reinstatement as a disease-modifying therapy for Angelman syndrome. Dev Med Child Neurol. 2021;63:802–7. 24. Sonzogni M, Wallaard I, Santos SS, Kingma J, du Mee D, van Woerden GM, et al. A behavioral test battery for mouse models of Angelman syndrome: a powerful tool for testing drugs and novel Ube3a mutants. Mol Autism. 2018;9:47. 25. Kishino T, Lalande M, Wagstaff J. UBE3A/E6-AP mutations cause Angelman syn- drome. Nat Genet. 1997;15:70–3. 26. Matsuura T, Sutcliffe JS, Fang P, Galjaard RJ, Jiang YH, Benton CS, et al. De novo truncating mutations in E6-AP ubiquitin-protein ligase gene (UBE3A) in Angel- man syndrome. Nat Genet. 1997;15:74–7. 27. Buiting K, Williams C, Horsthemke B. Angelman syndrome - insights into a rare neurogenetic disorder. Nat Rev Neurol. 2016;12:584–93. 28. Thibert RL, Larson AM, Hsieh DT, Raby AR, Thiele EA. Neurologic manifestations of Angelman syndrome. Pediatr Neurol. 2013;48:271–9. 29. Bird LM. Angelman syndrome: review of clinical and molecular aspects. Appl Clin Genet. 2014;7:93–104. 30. Willgoss T, Cassater D, Connor S, Krishnan ML, Miller MT, Dias-Barbosa C, et al. Measuring what matters to individuals with Angelman syndrome and their families: development of a patient-centered disease concept model. Child Psy- chiatry Hum Dev. 2021;52:654–68. 31. Rougeulle C, Cardoso C, Fontes M, Colleaux L, Lalande M. An imprinted antisense RNA overlaps UBE3A and a second maternally expressed transcript. Nat Genet. 1998;19:15–6. 32. Yamasaki K, Joh K, Ohta T, Masuzaki H, Ishimaru T, Mukai T, et al. Neurons but not glial cells show reciprocal imprinting of sense and antisense transcripts of Ube3a. Hum Mol Genet. 2003;12:837–47. 33. Judson MC, Sosa-Pagan JO, Del Cid WA, Han JE, Philpot BD. Allelic specificity of Ube3a expression in the mouse brain during postnatal development. J Comp Neurol. 2014;522:1874–96. 34. Jiang YH, Armstrong D, Albrecht U, Atkins CM, Noebels JL, Eichele G, et al. Mutation of the Angelman ubiquitin ligase in mice causes increased cytoplasmic p53 and deficits of contextual learning and long-term potentiation. Neuron. 1998;21:799–811. 35. Huang HS, Allen JA, Mabb AM, King IF, Miriyala J, Taylor-Blake B, et al. Topoisomerase inhibitors unsilence the dormant allele of Ube3a in neurons. Nature. 2011;481:185–9. 36. Meng L, Person RE, Huang W, Zhu PJ, Costa-Mattioli M, Beaudet AL. Truncation of Ube3a-ATS unsilences paternal Ube3a and ameliorates behavioral defects in the Angelman syndrome mouse model. PLoS Genet. 2013;9:e1004039. 37. Meng L, Ward AJ, Chun S, Bennett CF, Beaudet AL, Rigo F. Towards a therapy for Angelman syndrome by targeting a long non-coding RNA. Nature. 2015;518:409–12. 38. Bailus BJ, Pyles B, McAlister MM, O’Geen H, Lockwood SH, Adams AN, et al. Protein delivery of an artificial transcription factor restores widespread Ube3a expression in an Angelman syndrome mouse brain. Mol Ther. 2016;24: 548–55. 39. Lee HM, Clark EP, Kuijer MB, Cushman M, Pommier Y, Philpot BD. Characterization and structure-activity relationships of indenoisoquinoline-derived topoisomerase I inhibitors in unsilencing the dormant Ube3a gene associated with Angelman syndrome. Mol Autism. 2018;9:45. 40. Milazzo C, Mientjes EJ, Wallaard I, Rasmussen SV, Erichsen KD, Kakunuri T, et al. Antisense oligonucleotide treatment rescues UBE3A expression and multiple phenotypes Insight. of 2021;6:e145991. an Angelman syndrome mouse model. JCI Translational Psychiatry (2022) 12:426 J.K. Tanas et al. AUTHOR CONTRIBUTIONS MSS, JKT, and YE contributed to conception and design of the study. IW, JKT, and LW performed behavioral studies. JKT, DDK, MSS, AR, and IW contributed to data analysis and statistical analysis. JKT and MSS wrote the first draft of the manuscript. All authors contributed to manuscript revision, read, and approved the submitted version. 9 FUNDING This work was supported by the Angelman Syndrome Foundation (ASF) and by the District of Columbia Intellectual and Developmental Disabilities Research Center (DC- IDDRC) Award P50HD105328 by NICHD (PI: V. Gallo). Behavioral experiments performed in the Elgersma lab were funded by the ASF, Associazione Angelman, ZonMW (40-46800-98-009), and SFARI (275234). COMPETING INTERESTS The authors declare no competing interests. ADDITIONAL INFORMATION Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41398-022-02206-3. Correspondence and requests for materials should be addressed to Michael S. Sidorov. Reprints and permission information is available at http://www.nature.com/ reprints Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of license, visit http:// creativecommons.org/licenses/by/4.0/. this © The Author(s) 2022 41. Wolter JM, Mao H, Fragola G, Simon JM, Krantz JL, Bazick HO, et al. Cas9 gene therapy for Angelman syndrome traps Ube3a-ATS long non-coding RNA. Nature. 2020;587:281–4. 42. Schmid RS, Deng X, Panikker P, Msackyi M, Breton C, Wilson JM. CRISPR/Cas9 directed to the Ube3a antisense transcript improves Angelman syndrome phe- notype in mice. J Clin Investig. 2021;131:e142574. 43. Copping NA, McTighe SM, Fink KD, Silverman JL. Emerging gene and small molecule therapies for the neurodevelopmental disorder Angelman syndrome. Neurotherapeutics. 2021;18:1535–47. 44. Silva-Santos S, van Woerden GM, Bruinsma CF, Mientjes E, Jolfaei MA, Distel B, et al. Ube3a reinstatement identifies distinct developmental windows in a murine Angelman syndrome model. J Clin Investig. 2015;125:2069–76. 45. Cosgrove JA, Kelly LK, Kiffmeyer EA, Kloth AD. Sex-dependent influence of postweaning environmental enrichment in Angelman syndrome model mice. Brain Behav. 2022;12:e2468. 46. Judson MC, Shyng C, Simon JM, Davis CR, Punt AM, Salmon MT, et al. Dual- isoform hUBE3A gene transfer improves behavioral and seizure outcomes in Angelman syndrome model mice. JCI Insight. 2021;6:e144712. 47. Dodge A, Peters MM, Greene HE, Dietrick C, Botelho R, Chung D, et al. Generation of a novel rat model of Angelman syndrome with a complete Ube3a gene deletion. Autism Res. 2020;13:397–409. 48. Keute M, Miller MT, Krishnan ML, Sadhwani A, Chamberlain S, Thibert RL, et al. Angelman syndrome genotypes manifest varying degrees of clinical severity and developmental impairment. Mol Psychiatry. 2021;26:3625–33. 49. Jiang YH, Pan Y, Zhu L, Landa L, Yoo J, Spencer C, et al. Altered ultrasonic vocalization and impaired learning and memory in Angelman syndrome mouse model with a large maternal deletion from Ube3a to Gabrb3. PLoS ONE. 2010;5:e12278. 50. Koyavski L, Panov J, Simchi L, Rayi PR, Sharvit L, Feuermann Y, et al. Sex- dependent sensory phenotypes and related transcriptomic expression profiles syndrome. Mol Neurobiol. by Angelman are differentially 2019;56:5998–6016. affected 51. Huang HS, Burns AJ, Nonneman RJ, Baker LK, Riddick NV, Nikolova VD, et al. Behavioral deficits in an Angelman syndrome model: effects of genetic back- ground and age. Behav Brain Res. 2013;243:79–90. 52. van Woerden GM, Harris KD, Hojjati MR, Gustin RM, Qiu S, de Avila Freire R, et al. Rescue of neurological deficits in a mouse model for Angelman syndrome by reduction phosphorylation. Nat Neurosci. of 2007;10:280–2. alphaCaMKII inhibitory 53. Mossa A, Manzini MC. Molecular causes of sex-specific deficits in rodent models of neurodevelopmental disorders. J Neurosci Res. 2021;99:37–56. ACKNOWLEDGEMENTS We thank Justin Wolter (UNC) and Juliana Popovitz (Johns Hopkins University) for valuable discussion and guidance in data preparation and analysis. Translational Psychiatry (2022) 12:426
10.3390_economies11010021
Article Do Publicly Listed Insurance Firms in Saudi Arabia Have Strong Corporate Governance? Mamdouh Abdulaziz Saleh Al-Faryan 1,* and Jassem Alokla 2 1 School of Accounting, Economics and Finance, Faculty of Business and Law, University of Portsmouth, Richmond Building, Portland Street, Portsmouth PO1 3DE, UK 2 Department of Accounting and Finance, Sussex Business School, University of Sussex, Brighton BN1 9SN, UK * Correspondence: al-faryan@hotmail.com Abstract: Saudi Arabia has now opened its markets to foreign investors in line with its strategy to diversify its economy. However, investors need to feel confident that Saudi enterprises are being monitored and regulated appropriately. This study identifies the impact of improvements in Saudi corporate governance practices among insurance firms. The effects of corporate governance on the financial performance of 35 insurance firms listed on the Saudi stock market are examined from 2008 to 2014, including Shariah-compliant and life insurance firms. Four different methodologies are used: the generalised least squares random effect, fixed effect models, a difference-in-differences (DID) measurement for comparisons, and the probit model with average marginal effect to address endogeneity. The results indicate that firm performance is affected by information asymmetry. The 2009 exogenous shock from the Saudi regulatory change to board composition and audit committee size shows a positive effect on performance in the DID comparison. However, an increase in independent board and audit committee members has a significant negative effect. Other findings indicate that an increase in CEO (Chief Executive Officer) age has a positive effect on performance, as do three pay variables (director incentives, CEO and top executive pay, and above-the-mean director incentives). However, when CEO and top executive pay increases above the mean, the effect turns negative; this also happens with a change in CEO from poor performance. The results support the importance of Saudi insurance industry corporate governance regulation and reflect the improved governance perspectives of the Saudi Capital Market Authority and Saudi Arabian Monetary Agency. Keywords: corporate governance; Saudi Arabia insurance firms; board composition; audit committee size; managerial pay; shariah compliant; life insurance firms Citation: Al-Faryan, Mamdouh Abdulaziz Saleh, and Jassem Alokla. 2023. Do Publicly Listed Insurance Firms in Saudi Arabia Have Strong Corporate Governance? Economies 11: JEL Classification: G22; G30; G34; G38; J33; K22; L25 21. https://doi.org/10.3390/ economies11010021 Academic Editor: George R.G. Clarke 1. Introduction Received: 14 October 2022 Revised: 1 November 2022 Accepted: 14 November 2022 Published: 10 January 2023 Copyright: © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). In the Kingdom of Saudi Arabia (KSA), issues related to corporate governance have become central in its business environment (Al-Faryan 2020; Bagais and Aljaaidi 2020). Foreign investors can now expand their portfolios by investing in the KSA, as it has now opened its markets and its economy in line with achieving its Vision 2030, a strategy of diversification (Yousaf and Alokla 2022). These recent developments in the country and the role that corporate governance plays in creating confidence in Saudi firms have placed increased emphasis on the importance of effective firm governance (Almaqtari et al. 2020; Hashed and Almaqtari 2021). However, the issue of corporate governance has long been ignored in developing countries, such as the Gulf Cooperation Council economies. This is particularly true in one of the most important sectors in the KSA economy, the insurance industry (Alokla et al. 2022), even though such governance has long been seen as a major factor influencing the industry’s financial performance. To assess this, we investigate governance in the emerging market of the KSA, looking specifically at the insurance industry. Economies 2023, 11, 21. https://doi.org/10.3390/economies11010021 https://www.mdpi.com/journal/economies economies Economies 2023, 11, 21 2 of 41 Effective governance is generally measured by firm performance; however, although the link between corporate governance and firm financial performance has been widely studied in many fields, the academic debate on this continues. The main reason for this is that the relationship between firm financial performance and corporate governance is complex, as many studies have shown (Dalton and Dalton 2011; Fogel and Geier 2007; Hermalin and Weisbach 1991; McGuire et al. 2012). Scholars continue to debate what the optimal size should be for an audit committee, whether there is a link between future returns and stock dividends, and how significant the impact of a high capitalisation ratio is on return on equity. The evidence in extant studies has been contradictory, especially regarding optimal board size, compensation, and the number of outside directors. Although Klein (1998) and Bhagat and Black (2001) appear to report conclusive findings, their studies do not control for risk retention, CEO tenure, and firm age simultaneously. To provide a new perspective unlike that in previous studies (Gugong et al. 2014; Alokla et al. 2022), we add to our models volatility, risk retention, firm size, and firm age. We also consider Shariah compliance, which references the following of Islamic Law or Shariah. From our study perspective, our interest is in whether the performance of Shariah-compliant insurance firms reflects a similar effect in terms of corporate governance as the rest of the industry. As there is no standardised financial measurement of performance, we choose the following variables to measure firm financial performance: return on equity (ROE), return on assets (ROA), and Tobin’s Q. We explore insurance firm corporate governance in terms of board composition, audit committee size, CEO age, CEO turnover, director incentives, and CEO and top executive pay and their impact on insurance firm performance in the KSA. Additionally, as governance is an under-researched issue in this country owing to the lack of data, our study fills a gap in the literature. In the application of the corporate governance code in the KSA, insurance firms differ from other companies, as they are monitored by two official bodies, the Capital Market Authority (CMA) and the Saudi Central Bank or Saudi Arabian Monetary Agency (SAMA). To measure the impact of governance on insurance firm performance in the country, we examine several issues specific to board composition and audit committee size, in particular, the effect of the 2009 financial regulatory reform regarding board composition and audit committee size for publicly listed insurance firms in the KSA. We also address a gap in the literature by considering the endogenous character of board composition, audit committee size, CEO age, CEO turnover, director incentives, and CEO and top executive pay in Saudi insurance companies. Our research questions regarding the relationship between governance and firm performance are as follows. • Do board composition, audit committee size, CEO age, CEO turnover, director in- centives (above the mean), and CEO and top executive pay (above the mean) have a positive relationship with firm performance? • Has the 2009 exogenous shock created by regulatory changes to Saudi boards and audit committees had a positive effect on insurance performance? • Does increasing information asymmetry in KSA insurance firms negatively affect financial performance? • What are the optimal sizes of independent boards and audit committees that have a • • positive effect on performance? Do CEO turnover, above the mean director incentives, and CEO and top executive pay above the mean negatively affect insurance firm performance? Do listed Shariah-compliant insurance firms and life insurance firms follow the same patterns as all listed insurance firms in terms of the effects of board independence and audit committee size on firm performance? Our empirical findings offer five significant contributions to the corporate governance literature. First, to the best of our knowledge, ours is the first study to use Saudi insurance data to examine the endogenous impact of board composition, audit committee size, CEO age, CEO turnover, director incentives, and CEO and top executive pay on insurance firm performance. We do this by investigating 35 Saudi insurance firms between 2008 and Economies 2023, 11, 21 3 of 41 2014, applying four empirical approaches; the generalised least squares (GLS), the random effect (RE) and fixed effect (FE) models, a difference-in-differences (DID) comparison, and the probit model with average marginal effect. Second, we examine the effectiveness of directors when they are challenged by high commission costs or faced with asymmetric information. In addition, we look at the effect of the critical mass of board independence and auditing committee size on firm performance. Third, we investigate the impact of the 2009 regulatory reform of board composition and audit committee size on listed Saudi insurance firms. Fourth, we investigate the effects of board composition and audit committee size (Islamic corporate governance) on the performance of Saudi Shariah-compliant firms and life insurance. Finally, we extend the study of Al-Faryan and Dockery (2021) and find that, overall, the insurance sector has lagged behind in complying with new governance legislation. Namely, not all firms have adhered to various procedures requiring listed companies to improve transparency in corporate financial reporting in the macro insurance sector or in terms of macroeconomics. Our study focuses, in particular, on the link between corporate governance and Saudi insurance firm performance from the perspective of microeconomics. Our results provide useful implications for supervisors, managers, and practitioners regarding the controversial relationship between corporate governance and financial performance, specifically in an emerging market. The rest of the paper is structured as follows. Section 2 presents the literature review and our research hypotheses. Section 3 introduces our data and variables. Section 4 discusses our methodology. Section 5 provides the results and a discussion of the findings, and Section 6 concludes with research and managerial implications. 2. Literature Review and Hypotheses Many studies in different sectors have examined the relationship between corporate governance and firm financial performance (Al-Faryan 2017; Al-Faryan and Dockery 2017; Conyon and Florou 2002; Ghabayen 2012; Gibson 2003; Jenter and Kanaan 2015; Li et al. 2015; Liu et al. 2015; Volpin 2002). The literature on the Saudi market has highlighted several findings. Al-Faryan (2017), Al-Faryan and Dockery (2017), and Ghabayen (2012) have found mixed results. While Ghabayen (2012) results reveal that audit committee size, audit committee composition, and board size have no effect on firm performance in the 102 non-financial firms, the same results were also found by Al-Faryan (2017), who also found two variables to have a significant negative relation: CEO turnover and independent board members with firms’ performance. Al-Faryan and Dockery (2017) found that regulation and instability affect the ownership structure of 169 firms listed from 2008 to 2014. Moreover, evidence from a number of experimental studies has established that there is an inverse and robust statical relationship between corporate governance and a firm’s performance. On a sample of 460 largest UK listed firms from the period 1990–1998, other studies used a sample of 3365 CEO turnovers from 1993 to 2009 and over 1200 firms in eight emerging markets. Conyon and Florou (2002), Jenter and Kanaan (2015), and Gibson (2003), respectively, have found that top executives are fired due to their poor performance. On the other hand, both Li et al. (2015) and Liu et al. (2015) have found a positive relationship between declining ownership concentration and board independence on firms’ performance in China. While Volpin (2002) examined the determinants of executive turnover and firm valuation as a function of ownership and control structure in Italy, and his finding suggests that there is poor governance, as measured by a low sensitivity of turnover to performance. Although few studies have been conducted on the insurance industry in developed economies, extant studies, such as Elamer et al. (2018) in the UK and Eckles et al. (2011) in the US, have examined the relationship between corporate governance and firm per- formance. However, there are several studies on the insurance industry in developing economies; for example, in Nigeria (Ajemunigbohun et al. 2020; Akeem et al. 2014; Azutoru et al. 2017; Fadun 2013; Fodio et al. 2013; Gugong et al. 2014), Ethiopia (Fekadu 2015; Yemane et al. 2015), Kenya (Wanyama and Olweny 2013), Ghana (Tornyeva and Wereko Economies 2023, 11, 21 4 of 41 2012), Nepal (Maharjan 2019), Palestine (Shaheen and Jaradat 2019), India (Chaudhary 2014), Taiwan (Huang et al. 2007), Sri Lanka (Panditharathna 2016), Bangladesh (Datta 2018), Pakistan (Arif 2019), and Bahrain (Najjar 2012). In general, these studies in developing economies are plagued by one of the following issues: (1) Few observations, a short timeframe, and a lack of application of professional software analysis, for example, SPSS; these issues lead to data bias or multicollinearity and omitted variables during regression analysis, which can weaken confidence in the results. (2) Negligence in some assumption tests that would improve robustness, for example, endogeneity, exogenous, and simultaneity problems. (3) Overlooked control variables related to insurance companies that impact corporate governance, as some studies have combined independent variables without adding control variables into the regression mod- els. (4) Overlooked determinants of corporate governance for insurance firm performance (i.e., CEO age, CEO turnover, director incentives, CEO and top executive pay, changes in corporate governance regulations, and the optimal independent board and audit committee size). Our study addresses all of these issues. However, by highlighting different factors, these studies do confirm that insurance company success is linked to corporate governance practices. Thus, an important question remains: which corporate governance (independent) factor (one of the many decisions a company makes) impacts insurance firm performance the most (the dependent variable)? 2.1. Board Composition and Board Independenc Many authors have confirmed that board structure is a fundamental part of agency theory (Bhagat and Bolton 2008; Dalton and Dalton 2011). In a study on Kenya, Wanyama and Olweny (2013) found a positive association between board composition and CEO duality and financial performance. In contrast, in Pakistan, Arif (2019) found that CEO duality is negatively associated with insurance firm performance. The study by Najjar (2012) on insurance companies listed on the Bahrain Stock Exchange between 2005 and 2010 found that board size, firm size, and the number of block holders had a positive significant effect on financial performance. In China, Liu et al. (2015) found that board independence positively affects firm performance, especially in government-controlled firms; the impact of board independence on firm performance increases with ownership concentration. In contrast, in a study on non-financial firms in the Saudi market, Ghabayen (2012) found that board composition had a significant negative relationship with firm performance. Similarly, in Nepal, looking at 2011 to 2018, Maharjan (2019) found that board size and CEO duality were negatively associated with financial performance. Meanwhile, Shaheen and Jaradat (2019) found that, in Palestine, firms that hold more frequent board meetings and have CEO duality achieved better performance. In the case of the KSA market, among listed firms, Al-Faryan (2021) found no sig- nificant association between board composition and firm performance, after examining 169 firms from 2007 to 2014 using RE and FE models. However, his findings indicate that the effects of the regulatory reform in 2009 on board independence are endogenous and positive with a firm’s stock return performance and Tobin’s Q. The implication is that the firms investigated are complying with the country’s corporate governance regulation. In addition, with three to four independent board members, the effectiveness of board independence strongly relates to firm performance (ROE and stock returns). Examining 11 Saudi banks between 2009 and 2012, Al-Sahafi et al. (2015) reported that board independence has a significantly positive relationship with performance. In contrast, looking at the period 2014 to 2017, Almoneef and Samontaray (2019) show that board independence has a negative effect on Saudi bank performance. Prior to this, Alhassan et al. (2015) used ordinary least squares regression to assess the performance of 10 publicly traded Saudi banks between 2007 and 2012. They looked at the number of meetings, board composition, and board size as determinants of corporate governance. The authors found a non-significant relationship between board size and board composition with firm Economies 2023, 11, 21 5 of 41 performance, while board meetings have a significant positive relationship. The timeframe allows them to analyse the impact of the corporate governance code of 2006. However, the above studies overlook insurance firms in the financial sector, even though, as part of that sector, along with banks, such firms are monitored by SAMA and CMA in the KSA (Al-Faryan 2020). Specifically, these studies overlook the 2009 exogenous shock from regulatory changes to board composition, along with endogeneity issues. Looking at the period 2006 to 2010, Fadun (2013) examined 10 insurance firms listed on the Nigerian stock exchange and found that board size positively relates to financial performance. However, for the period 2011 to 2015, Azutoru et al. (2017) argue that board size had a negative association with financial performance. Moreover, Akeem et al. (2014) found no evidence to support the idea that board size or other board composition aspects relate to financial performance. In Ethiopia, Yemane et al. (2015) conducted a study on 10 insurance firms from 2009 to 2013. They concluded that board meetings and board compensation have a statistically significant positive impact on financial performance. However, they found no evidence of the impact of board size. Meanwhile, Fekadu (2015) and Panditharathna (2016) found no evidence of a relationship between corporate gover- nance and firm performance in 56 listed financial firms on the Colombo Stock Exchange in Sri Lanka. Datta (2018) reported a negative relationship between board composition and financial performance in 10 listed insurance companies in Bangladesh. Coles et al. (2008) point out that complex firms, which need more advice than simple ones, have larger boards of directors, with a higher proportion of outside directors. Tobin’s Q and board size have a U-shaped relationship, which implies that either very small or extremely large boards are optimal. However, this relationship occurs as a result of distinc- tions between complex and simple firms. For complex (simple) firms, Tobin’s Q increases (decreases) board size, with this relationship driven by the presence of outside directors. However, other studies have revealed that independent boards do not always enhance value. For example, according to Adams (2012), financial firms in acute trouble during the 2008/2009 financial crisis that required government bailout funding had more independent boards than those that were not in trouble. The author speculates that this was due to the lack of expertise and industry understanding among the majority of the independent directors. According to Yermack (1996) and Eisenberg et al. (1998), board size has a negative impact on firm performance. Namely, larger boards equate to higher monitoring costs, limiting their capacity to effectively monitor and manage. Based on these mixed results, we test the following hypotheses. Hypothesis 1. Board composition/independence is positively associated with financial performance of listed Saudi insurance firms. Hypothesis 2. The exogenous shock from the regulatory changes to Saudi boards are negatively related to the performance of listed Saudi insurance firms. 2.2. Monitoring Costs According to the literature, when monitoring expenses rise, the advantages of having more board independence diminishes (Al-Faryan 2021; Fama and Jensen 1983; Linck et al. 2008). Al-Faryan (2021) contends that external directors incur more costs in comparison with internal ones; the underlying reason for this being the diminished capacity of independent board members to oversee and advise management effectively. There are two major drawbacks to external directors. First, in regard to all firm aspects affected by information asymmetry, internal directors will have more and better information compared with what external directors have access to. Second, external directors may lack sufficient expertise, given that internal managers (directors) have gained company-specific expertise, while external directors are likely to have more general knowledge. As a result, external directors add more costs in terms of the need to gather information on the company and greater expertise in performing their duties effectively (Al-Faryan 2021). Al-Faryan (2021) and Economies 2023, 11, 21 6 of 41 Maug (1997) indicate that, in an environment of high information asymmetry, increasing the number of independent directors lacks beneficial effects and will have a negative effect on performance. Similarly, other studies have shown that an increase in monitoring costs leads to a decrease in the number of external directors (Adams and Ferreira 2007; Raheja 2005). Linck et al. (2008), Duchin et al. (2010), and Al-Faryan (2021) show that, when the cost of acquiring information is lower, independent directors become more valuable. In addition, among Saudi and Chinese firms, board composition seems to have a stronger effect on firm performance (Al-Faryan 2021; Liu et al. 2015). Based on this, we posit the following hypothesis. Hypothesis 3. Information asymmetry (monitoring costs) in Saudi insurance firms will negatively affect firm performance. 2.3. Audit Committee Size Audit committee size is a recognised aspect of the effectiveness of corporate gover- nance and, as such, is expected to play a key role in the financial performance of insurance firms. The audit committee’s role is to enhance firm value. Maharjan (2019), Fadun (2013), and Fekadu (2015) all found a positive relationship between audit committee size and financial performance. In contrast, Datta (2018), Ghabayen (2012), Yemane et al. (2015), and Al-Faryan (2017) found that audit committee size does not affect firm performance. As audit committees have become a fundamental part of a firm’s strategy for generating profits and maximizing shareholder value, according to Braiotta et al. (2010) and Karamanou and Vafeas (2005), large audit committees have been shown to provide superior organisational capacity and authority, as well as a vast knowledge base. However, Karamanou and Vafeas suggest that, when audit committees become too big, they risk becoming ineffective, losing track of procedures and duties, and, eventually, failing to fulfil the duties assigned to them swiftly and properly. Aldamen et al. (2012) discovered a favourable correlation between smaller audit committees with higher expertise and firm performance. However, these extant studies have not examined an exogenous regulatory shock on audit committee size. Thus, we posit the following hypotheses. Hypothesis 4. A larger audit committee is positively related to firm performance of listed Saudi insurance firms. Hypothesis 5. The exogenous shock of the regulatory change to Saudi audit committees have a significant positive effect on the performance of listed Saudi insurance firms. 2.4. CEO Characteristics 2.4.1. CEO Age The influence of a CEO’s age on firm performance has garnered considerable atten- tion in the literature. The presumption is that older CEOs will have a competitive edge over younger ones, who unquestionably have less experience managing a business (Peni 2014). According to Yeoh and Hooy (2022), CEO age has a negative association with firm performance, as risk-taking rises with age; however, beyond a particular age threshold, risk-taking decreases. In contrast, Davis (1979) found no relationship between firm success and CEO age, whereas Amran et al. (2014) reported a negative association between firm performance and CEO age, particularly in terms of ROA. As a result, we posit the following hypothesis. Hypothesis 6. Age of CEO is positively related to firm performance in listed Saudi insurance firms. Economies 2023, 11, 21 7 of 41 2.4.2. CEO Turnover When emerging-market enterprises do poorly, their CEOs are likely to lose their positions (Gibson 2003). Conyon and Florou (2002) show a statistically significant negative relationship between the likelihood of management turnover and firm performance, with top executives dismissed for bad performance. However, reviewing a hand-collected sample of 3365 CEO turnovers from 1993 to 2009, Jenter and Kanaan (2015) demonstrate that CEOs are more likely to be fired after poor industry performance and, to a lesser degree, stock market performance. They indicate that a drop in industry performance from the 90th to the 10th percentile more than doubles the likelihood of a forced CEO change. Al-Faryan (2017) discovered a negative correlation between CEO turnover and performance in all Saudi publicly traded companies. Friedman and Singh (1989) determined that, while company performance is a significant factor affecting CEO turnover, there are additional factors that have been overlooked, such as the CEO’s proximity to retirement age, whether the CEO’s departure was voluntary, and whether a new CEO was found. Thus, we posit the following hypothesis. Hypothesis 7. Firm performance is negatively related to CEO turnover in Saudi insurance firms. 2.5. Managerial Pay 2.5.1. Director Incentives A key role of corporate boards is to ensure that managers undertake their tasks in the best interest of shareholders. Nevertheless, in some cases there is collusion among management. This, therefore, leads to agency problems between shareholders and directors (Fama and Jensen 1983). Compensation policy is considered a key component in a firm’s success (Jensen and Murphy 2010) and a means of addressing board oversight. Murphy (1985) criticises earlier research that focuses only on the visible parts of wages, such as pay and bonuses, while ignoring the most sensitive indicators of firm success, such as stock options and awards. Furthermore, the author contends that, while it is fair to suppose that stock price influences management remuneration, this is not the main predictor of such remuneration. Mehran (1995) studies the relationship between executive pay structure and performance and finds support for incentive pay, claiming that the kind rather than the size of the remuneration encourages managers to improve performance. Contrary to theoretical predictions, in Nigeria, Azutoru et al. (2017) found that executive director incentives do not influence the financial performance of insurance firms. However, according to Al- Faryan (2021), director incentives have a positive relationship with Tobin’s Q in Saudi listed enterprises. Thus, we posit the following hypothesis. Hypothesis 8. Director incentives are positively related to firm performance. 2.5.2. CEO and Top Executive Pay According to Garen (1994), several prior studies have found a positive and significant association between CEO and top executive compensation and financial performance. Top managers who receive greater incentives and stock awards are more likely to make conservative choices that reduce a firm’s profitability (Eckles et al. 2011). In Nigeria, Olaniyi and Olayeni (2020) found a two-way relationship between firm success and CEO salary. They indicate that CEO compensation affects firm performance positively, while firm performance affects CEO compensation adversely. Furthermore, they believe that CEOs should be rewarded for strong performance but not penalised for poor performance; this would mean that CEOs would be well paid even when their firms do poorly. In a study on publicly traded firms in China between 2009 and 2015, Bin et al. (2020) found that there is a positive relationship between firm performance and CEO pay. Al-Faryan (2021) indicates that CEO pay and top executive compensation have a positive association with the performance of Saudi listed enterprises, as well as government listed enterprises, assuming Economies 2023, 11, 21 8 of 41 that the enterprises are competitive and not just controlled by the Saudi government. Thus, we posit the following hypothesis. Hypothesis 9. CEO and top executive pay are positively related to firm performance. Based on the extant research discussed above, our study extends the current literature in the context of listed insurance firms in the following ways. First, we contribute to the literature on audit committee size, CEO top executive pay, CEO age, CEO turnover, board composition, and director incentives by examining data from publicly listed Saudi insurance firms between 2008 and 2014. We believe ours is the first study to review these elements of corporate governance in the context of the Saudi market and Middle East and North Africa economies. Second, we examine the effectiveness of independent directors when they are faced with either high monitoring costs or information asymmetry. Third, we look at the effect of the critical mass of board independence and audit committee size on insurance firm performance. We also evaluate the exogenous impact of changes in board and auditing committee regulation, given that exogenous impacts tend to be rare occurrences in the context of insurance firms. Fourth, we look at Shariah-compliant and life insurance firms in terms of the effects of board composition and audit committee size on firm performance. Finally, we extend the literature that utilises cross-sectional regressions by employing panel data to account for endogeneity and exogenous issues, often ignored in developing economies. 3. Data and Variables To assess whether the listed insurance firms in the KSA have strong corporate gov- ernance, we employ a unique panel dataset consisting of 35 listed Saudi insurance firms in the period 2008 to 2014, including 16 firms that offer life insurance products and seven Shariah-compliant firms that offer insurance products. Data are obtained from the CMA and Mubasher. We begin the study in 2008, since prior to that there are missing observations and variables among the data for Saudi insurance firms. Murphy (1985) argues that cross-section models are inherently flawed due to a lack of theoretical basis for relevant variables and the possibility of omitted variables. Notably, if we assume that the omitted variables will not change over time, we can accurately assess the relationship through time-series regressions. For this reason, we use panel data regressions, employing four different methods: the GLS (Generalised Least Squares), RE (Random Effect), and FE (Fixed Effect) models; a DID (Difference-In-Differences) comparison; and the probit average marginal effect model to account for endogeneity issues (Al-Faryan 2021). There are unique elements in our dataset period. First, the number of firms increased during this timeframe; with 21 listed insurance firms in 2008 and 35 by the end of 2014. Notably, in 2004, there is just one listed insurance firm in Saudi Arabia. Obviously, as the industry expands, the implementation of strong corporate governance grows in importance in this sector. Second, our dataset can be compared with that used by Al-Faryan and Dockery (2021), which addresses the governance code in the Saudi macro insurance sector (in terms of macroeconomics), and covers a similar period. However, herein, we focus on the link between corporate governance and Saudi insurance firm performance from the perspective of microeconomics. Third, during this period, most insurance companies lost more than 20% of their capital, and there was little profitability (Alokla et al. 2022). Thus, despite the recent establishment of these companies, we need to examine their governance during this period. Fourth, the absence of an insurance policy is linked directly to the National Information Center. To avoid discovery, individuals can manipulate their insurance or fail to comply with it. Additionally, there is no fine for individuals who do not follow their insurance policies. Fifth, as many companies had to suspend trading in their stock due to losses during this timeframe, there is a lack of information on many variables Economies 2023, 11, 21 9 of 41 after 2014. For this reason, our sample only goes to 2014, as there would be additional time and effort required, as well as financial restrictions, in compiling data after 2014. 3.1. Dependent Variables We utilise three financial performance variables, ROA, ROE, and Tobin’s Q, to examine whether the findings are sensitive to particular measurements and to ensure consistency. Notably, ROA and ROE are accounting-based and backward-looking performance indi- cators (Al-Faryan 2021; Shan and McIver 2011), whereas Tobin’s Q is a forward-looking, market-based measurement that quantifies the value investors place on future company prospects (Al-Faryan 2021). Tobin’s Q is often troublesome in corporate governance re- search as it is conceived as an indicator for future development potential, describing value as a cause rather than a result of the corporate governance framework (Al-Faryan 2021; Boone et al. 2007; Lehn et al. 2009; Linck et al. 2008). However, Morck et al. (1988) suggest that Tobin’s Q can represent the non-current assets of a firm better than operational perfor- mance can, as operational performance may not convey market expectations arising from reforms, especially when corporate governance structures and rules change. Thus, because of this, Tobin’s Q is included to assist in our analysis and for robustness and comparison with performance indicators specified in the corporate governance literature. 3.2. Independent Variables Our independent variables are board composition, director incentives, independence, board-monitoring costs, audit committee size, regulation (as assessed by DID), CEO and top executive pay, CEO age, and CEO turnover. Saudi board directors are defined as executive directors who work fulltime for the company and are paid a salary; non-executive directors are those who do not work fulltime for the company and do not receive fulltime salaries; independent directors are those who do not work fulltime for the company and do not receive fulltime salaries. Independent directors are entirely impartial to the firm. Non-independent directors meet one of the following criteria: the director holds 5% or more of the company or its subsidiary shares or legally represents another firm or person who owns 5% or more of the company or its subsidiary shares, the director has been a top executive in the firm or one of its subsidiaries in the previous two years or works for another corporation affiliated with the firm, the director has been nominated to a board associated with the holding company’s board of directors, or its management is linked to other board members. According to the 2009 legislation, most directors must be non-executive directors, and the total number of independent directors should be equal to two or to one-third of the total directors, whichever is larger (Al-Faryan 2021; Capital Market Authority CMA 2017). According to the 2009 Saudi legislation, the audit committee must be created by a resolution of the firm’s ordinary general assembly; its members must be shareholders or others, providing that at least one is an independent director and one is a non-executive director. The audit committee cannot have less than three members or more than five; at least one should be a financial and accounting specialist (CMA 2017). 3.3. Control Variables Control variables are identified as the variables that impact and consequently correlate with firm performance. The selection of control variables is critical for the consistent estimation of the analysis. Our choice of control variables follows existing research, and thus, our analyses include control variables adapted from the literature. We believe the addition of control variables improves our estimates while reducing the omitted variable bias reported in earlier research. Further detail on this topic can be found in MacAvoy et al. (1983), Baysinger and Butler (1985), McConnell and Servaes (1990), and Al-Faryan (2021). Our control variables are CEO share %, CEO tenure, Ln(Assets), Ln(gross written), Ln(firm age), volatility, risk retention %, and depth of the insurance market. Economies 2023, 11, 21 10 of 41 We utilise natural logarithms for three reasons: to simplify the interpretation of the results, to simplify the calculation of the results, and to lower the amount of skew in the size distribution of the enterprises, as per Murphy (1985); Demsetz and Lehn (1985) also observed that such data may be transformed from bounded to unbounded. The control variables are explained as follows. 3.3.1. Assets We define the size of the firm as the logarithm of total assets, which is widely employed as a control variable in empirical investigations, including those by Mwangi and Murigu (2015), Alokla et al. (2022), Shaaban and Wahome (2018), Ismail (2013), Eling and Marek (2014), Bubic and Susak (2015), and John et al. (2008). Larger firms require more outside knowledge and benefit further from better monitoring and review; we measure the natural logarithm of assets, as this can address the fact that larger firms are more likely to have more external directors on their boards (Al-Faryan 2021). Firms with higher asset value also enjoy economies of scale and a more diversified insurance portfolio (Alokla et al. 2022). Moreover, firms with higher asset value receive more media attention compared with smaller firms, which may impact investors. 3.3.2. Risk, Volatility, and Market Depth Reinsurance is an important aspect of the insurance industry and plays a critical role in the global industry’s financial stability (International Association of Insurance Supervisors 2012; Alokla et al. 2022). Insurance companies have different financial strengths, which affect their reinsurance arrangements and dictate how much risk they retain. Eling and Marek (2014) include volatility in their model and find that the theoretical volatility of asset returns is lower for non-life insurance firms. We use variables of risk retention, volatility, and market depth since firms differ in terms of market share and growth in market share, which may have a positive impact on financial performance (Alokla et al. 2022; Kozak 2011). 3.3.3. CEO Tenure and Share CEO tenure and CEO share have been used widely in empirical studies as control variables. However, Gibbons and Murphy (1992) provide evidence that CEOs are likely to have different incentives, reputations, and career concerns, depending on their age and their years of tenure. Thus, a CEO who has five years of tenure at age 65 is more likely to have different management qualities than others. These CEOs are likely to have different incentives, reputations, and career concerns (Dikolli et al. 2014). 3.3.4. Firm Age Financial performance and firm age have also appeared in the empirical literature as control variables in studies by Guruswamy and Marew (2017), Charumathi (2012), and Shiu (2004). However, Mwangi and Murigu (2015) found no evidence of the effect of firm age on firm performance. 3.3.5. Gross Written Premiums Economic theory suggests that insurance firms with strong and sustainable business growth are more likely to be financially stable and profitable. Although several studies have widely used gross underwriting premiums as a control variable, the results regarding its relationship with financial performance have been mixed (e.g., Mardi et al. 2017; Mazviona et al. 2017; Alokla et al. 2022; Shaaban and Wahome 2018; Zhang and Nielson 2015). We choose not to use the CEO duality variable, as all Saudi insurance firms follow the corporate governance code (regulation) that does not allow the CEO to be chairperson. In addition, we chose not to use ownership variables. Although many Saudi insurance firms have owners with less than a 5% share, data for these owners are not available since Economies 2023, 11, 21 11 of 41 corporate governance regulation only requires the disclosure of ownership of 5% or more (Al-Faryan 2021; CMA 2017). In addition, we are unable to apply the governance index of 24 rules proposed by Gompers et al. (2003) to proxy the level of shareholder rights to measure corporate gover- nance quality, as the firms listed in the Saudi stock market lack those governance provisions (Al-Faryan 2022). Moreover, we do not differentiate between forced and voluntary CEO turnover (Al-Faryan 2022; Denis et al. 1997), as our source material is each firm’s annual report, wherein forced turnover is not distinguished from voluntary turnover. 4. Methodology We use STATA software (StataCorp LLC, Lakeway Drive College Station, TX, USA) to apply four empirical methods to examine whether the listed Saudi insurance firms have strong corporate governance during our timeframe. As mentioned, we use the RE, FE, and DID models and the probit model with average marginal effect. The RE technique differs from the FE effect technique in that the variation in the error term is believed to be unrelated to the independent variables. Statistically, the FE models are more independent of the observed factors and forecasts. The RE approach takes advantage of partial pooling, while the FE technique does not (Al-Faryan 2021). The Hausman (1978) test can be used to identify which model to use, for example, when there are fewer data points for partial pooling, and the coefficient estimates are influenced by more plentiful data from other groups. In this case, the RE approach may be preferable to putting all the groups together (which may disguise group-wide variation) or calculating effects individually for all groups (which may result in low-sample estimates) (Al-Faryan 2021). The RE technique generalises the fractional pooling method as a statistical model, enabling its application in a variety of scenarios. Our RE and FE models are expressed as follows. PERFit = α + β1%IBit + β2 Ln (BS)it + β3Ln (Assets) it + β4Ln (cid:0)FirmAge(cid:1) it +β7DeInMarketit + β8Ln (gross written)it + εit, + β5VOLit + β6RiskRetit (1) where PERFit, the dependent variable, is either ROA, ROE, or Tobin’s Q. ROA is the firm’s annual net profits over total assets; ROE is the firm’s annual net profit over total equity; Tobin’s Q is the firm’s total market capital over total assets. %IB is the board composition as a percentage of the independent directors on the board in a certain year (the number of independent directors divided by the total number of directors). Ln(BS) is the natural logarithm of the board size (natural logarithm of the number of directors on the board in a certain year); notably, Ln(BS) may be negative since the increase in agency cost caused by having a high number of directors may impair efficiency and lower firm performance (Al-Faryan 2021; Jensen 1993; Lipton and Lorsch 1992; Yermack 1996); Ln(Assets) is the natural logarithm of total firm assets; Ln (FirmAge) is the natural logarithm of firm age since incorporation. VOL is volatility (annualised standard deviation of monthly stock returns). RiskRet is a percentage of risk retention (net premiums written over gross premiums) Alokla et al. (2022). DeInMarket is the depth of the insurance market or insurance penetration (insurance penetration of total gross domestic product (GDP) is defined as gross written premiums divided by total GDP). Ln(gross written) is the natural logarithm of gross premiums written measured by the sum of both direct premiums written and assumed premiums written, before deducting the reinsurance ceded. For the robustness of the results, we use the following equation. PERFit = α + β1 IB (3)it + β2 IB (4)it + β3 IB (5)it + β4 IB (6)it + β5 Ln (BS)it + β6Ln (Assets) it+ β7Ln (cid:0)FirmAge(cid:1) it + β8VOLit + β9RiskRetit + β10DeInMarketit + β11Ln (gross written)it + εit, where IB(3), IB(4), IB(5), and IB(6) are the number of independent members on the board, grouped as three, four, five, and six or more, respectively (dummy variable equals 1 if the firm has up to six or more independent directors, and 0 otherwise). For additional robustness, we examine if monitoring costs affect insurance firm perfor- mance as follows. (2) Economies 2023, 11, 21 12 of 41 PERFit = α + β1%IBit + β2%IB ∗ VOL it + +β3 Ln (BS)it + β4Ln (Assets)it + β5Ln (FirmAge)it + β6VOLit+ β7RiskRetit + β8DeInMarketit + β9Ln (gross written)it + εit, (3) where %IB*VOL is the monitoring cost (the interaction term proxies for information asym- metry or the monitoring cost of independent directors, namely, board composition mul- tiplied by the annual standard deviation of monthly stock returns, following Fama and Jensen (1983), Linck et al. (2008), and Al-Faryan (2021)). To examine the different variables along with audit committee size, we use the follow- ing equation. PERFit = α + β1SACit + β2Ln (BS)it + β3Ln (Assets) it + β4Ln (cid:0)FirmAge(cid:1) it β7DeInMarketit + β8Ln (gross written)it + εit, + β5VOLit + β6RiskRetit+ where SAC is the size of the audit committee (number of audit committee members). To check for robustness, we use the following equation. PERFit = α + β1 Asize (3)it + β2 Asize (4)it + β3 Asize (5)it + β4Ln (BS)it + β5Ln (Assets) it + β6Ln (FirmAge)it+ β7VOLit + β8RiskRetit + β9DeInMarketit + β10Ln (gross written)it + εit, (4) (5) where ASize(3), ASize(4), and ASize(5) are the number of audit committee members: three, four, or five, respectively (dummy variable equals 1 if the firm has up to five, and 0 otherwise). We test CEO characteristics in the combined variable INDVAR (CEOAge, RET Age, CEO turnover) as follows. PERFit = α + β1 I NDVARit + β2Ln (BS)it + β3Ln (Assets) it + β4Ln (cid:0)FirmAge(cid:1) +β7DeInMarketit + β8Ln (gross written)it + εit, + β5VOLit + β6RiskRetit it (6) where CEOAge is a number equal to the CEO’s age; RET Age is the CEO’s retirement age (dummy variable equals 1 if the CEO is age 60 or over, and 0 otherwise); CEO turnover is a dependent dummy variable that equals 1 if the CEO changed, and 0 otherwise. The empirical test for managerial pay using RE and FE is as follows. PERFit = α + β1D − I NCit + β2CEO − Ex − Payit + β3CEO − Sharesit + β4 COE − Tenit + β5Ln (BS)it +β6Ln (Assets) it + β7Ln (cid:0)FirmAge(cid:1) + β8VOLit + β9RiskRetit + β10DeInMarketit + β11Ln (gross written)it + εit, it (7) where D − I NC represents the director incentives (in Saudi riyals (SAR)); these are salary bonuses or attendance allowance for meetings or expense allowance or benefits in kind or a certain percentage of firm profits that may combine two or more of these advantages, expressed annually); CEO − Ex − Pay represents the CEO and top executive pay (in SAR) (this is the sum of the CEO and the senior executive salaries, allowances, in-kind benefits, and end-of-service bonus rewards, expressed annually). The control variables are CEO − Shares, which are the percentage of CEO-owned shares (total CEO shares divided by total firm shares); COE − Ten represents CEO tenure (number of years the CEO has held office). The above-the-mean calculation assesses the impact of the above-the-mean director incentives and above-the-mean CEO and executive pay as follows. PERFit = α + β1 Above mean D − I NCit + β2 Above mean CEO − Ex − Payit + β3 CEO − Sharesit +β4 COE − Tenit + β5Ln (BS)it + β6Ln (Assets) it + β7Ln (cid:0)FirmAge(cid:1) + β8VOLit + β9RiskRetit+ it β10DeInMarketit + β11Ln (gross written)it + εit. (8) 4.1. Difference-in-Differences Approach According to Bertrand and Mullainathan (2003) and Al-Faryan (2021), the DID tech- nique is widely used in the economics and finance literature to address endogeneity. We generate two variables in our DID and DID2 models. The first is a dummy variable that Economies 2023, 11, 21 13 of 41 equals 1 in the post-regulation years from 2009 to 2014, and 0 otherwise. This indicator is referred to as ‘post-regulation’. The treatment variable for DID is also a dummy variable that equals 1 if the percentage of independent members on the board is higher than or equal to one-third, and 0 otherwise. In addition, the treated variable for DID2 is a dummy variable that equals 1 if the audit committee size is greater than or equal to three members, and 0 otherwise. We refer to this variable as ‘treated’ and then multiply post-regulation by treated to create a new variable to represent the influence of the difference between organisations that complied with the corporate governance standards after 2008 and those that did not. We include the variable ‘Treated*Post regulation’ for robustness in the RE and FE models to avoid any bias when comparing the treatment and control groups that may have occurred from variations in groups over time (Al-Faryan 2021). For this comparison, the Saudi insurance firms had to exist before the change in rules (i.e., before the end of 2008). As a result, we exclude 14 of the 35 listed Saudi insurance corporations in our sample according to this criterion. Our study looks at the exogenous shock of insurance regulation and compares the performance of companies complying with the new regulations with those that did not comply. Concerning the built-in endogeneity difficulties, we reduce their presence by introducing the board composition and audit committee changes at Saudi enterprises as a consequence of governance legislation rather than endogenously imposed by insurance industry features (Al-Faryan 2021). In this regard, the change in insurance governance regulations in the KSA allows us to investigate the effects of board composition and audit committee size on performance using the following models. PERFit = α + β1DID it + β2Ln (BS)it + β3Ln (Assets) it + β4Ln (FirmAge)it + β5VOLit + β6RiskRetit +β7DeInMarketit + β8Ln (gross written)it + εit, PERFit = α + β1DID2 it + β2Ln (BS)it + β3Ln (Assets) it + β4Ln (cid:0)FirmAge(cid:1) +β7DeInMarketit + β8Ln (gross written)it + εit, + β5VOLit + β6RiskRetit it (9) (10) where DID represents the DID in board composition, while DID2 represents the DID in the size of the audit committee. DID and DID2 capture the Treated*post-regulation, the treated group multiplied by post-regulation. 4.2. Probit Model Average Marginal Effect The average marginal effect yields a probability effect; namely, a value between 0 and 1. It is the average probability change when x rises by one unit. Because a probit model is nonlinear, the effect will vary from one individual to another. The average marginal effect is calculated by computing it for each individual and then creating an average. We use the following equations for this model. CEO turnoverit = α + β1PERF it + β2Ln (BS)it + β3Ln (Assets) it + β4Ln (FirmAge)it + β5VOLit+ β6RiskRetit + β7DeInMarketit + β8Ln (gross written)it + εit, Above mean director incentive it = α + β1PERF it + β2 Above mean CEO and Top Executive payit + β3CEO− Sharesit + β4 COE − Tenit + β5Ln (BS)it + β6Ln (Assets) it + β7Ln (cid:0)FirmAge(cid:1) + β8VOLit + β9RiskRetit+ β10DeInMarketit + β11Ln (gross written)it + εit, it Above mean CEO and Top Executive payit = α + β1PERF it + β2 Above mean director incentive it + β3CEO − Sharesit + β4 COE − Tenit +β5Ln (BS)it + β6Ln (Assets) it + β7Ln (cid:0)FirmAge(cid:1) +β11Ln (gross written)it + εit, + β8VOLit + β9RiskRetit + β10DeInMarketit it (11) (12) (13) where CEO Turnover, Above mean director incentive, and Above mean CEO and Top Executive Pay are the dependent dummy variables that equal 1 if the CEO changed and the number was above the mean, and 0 otherwise. The independent variable is still PERF where firm performance equates to ROA, ROE, or Tobin’s Q. All control variables are as before and include Above mean director incentive with Above mean CEO and Top Executive Pay. Economies 2023, 11, 21 14 of 41 4.3. Shariah Compliance and Life Insurance Firms To investigate the effects of Shariah compliance and life insurance firms to board composition and the size of the audit committee on performance using the following models: PERFit = α + β1%IB ∗ Shariah − compliantit + β2 Ln (BS)it + β3Ln (Assets)it + β4Ln (FirmAge)it+ β5VOLit + β6RiskRetit + β7DeInMarketit + β8Ln (gross written)it + εit, PERFit = α + β1SAC ∗ Shariah − compliantit + β2 Ln (BS)it + β3Ln (Assets)it + β4Ln (FirmAge)it +β5VOLit + β6RiskRetit + β7DeInMarketit + β8Ln (gross written)it + εit, (14) (15) where Shariah-compliant is a dummy variable that equals 1 if the firm adheres to Islamic Shariah provisions in all of its transactions, insurance operations, and investment activities, and 0 otherwise; %IB*Shariah-compliant is board composition multiplied by the Shariah- compliant variable; SAC* Shariah-compliant is the size of the audit committee multiplied by the Shariah-compliant variable measuring Islamic corporate governance or the corpo- rate governance of Shariah-compliant insurance firms. Other variables are as explained previously. PERFit = α + β1%IB ∗ Li f e insuranceit + β2 Ln (BS)it + β3Ln (Assets)it + β4Ln (FirmAge)it + β5VOLit+ β6RiskRetit + β7DeInMarketit + β8Ln (gross written)it + εit, PERFit = α + β1SAC ∗ Li f e insuranceit + β2 Ln (BS)it + β3Ln (Assets)it + β4Ln (FirmAge)it + β5VOLit+ β6RiskRetit + β7DeInMarketit + β8Ln (gross written)it + εit, (16) (17) Herein, life insurance is a dummy variable that equals 1 if the firm has contracts between insurance policyholders and an insurer, or assurer, in which the insurer guarantees to pay a specified beneficiary a sum of money upon the insured person’s death (often the policyholder)—other occurrences, such as terminal or severe illness, may also trigger payment depending on the contract (typically, the policyholder pays a premium, either on a monthly basis or in one lump sum. Other expenditures may be covered by the benefits of insurance; they may be conventional or takaful (Islamic) life insurance)—and 0 otherwise; %IB* Life insurance is board composition multiplied by the life insurance variable; SAC* Life insurance is the size of the audit committee (those who monitor corporate governance at firms that provide life insurance) multiplied by the life insurance variable. Other variables are as explained previously. 5. Results and Discussion Table 1 provides summary statistics of the key variables in our study. The standard deviations of ROA, ROE, CEO-Ex-Pay, CEO-Shares (%), CEO Turnover, and VOL, six of the 18 variables, are significantly greater than the mean. This suggests that our data are dispersed extensively or that the mean is not accurately reflecting the data. The average board has a mean of 46.43% and a median of 40% independent members. The average board size is 8 to 9 directors, with a minimum of 5 and a maximum of 11. The average size of the audit committee is three, with a maximum of five. The average CEO age is 50 to 51, with a standard deviation of eight years. The average CEO tenure is three years, with a standard deviation of two years. The minimum and maximum CEO tenure are one and nine years, respectively. The dummy variable, CEO Turnover, has a mean value of 0.233 and a standard deviation of 0.424. In addition, there are missing observations in some years for some variables, revealing that some firms had losses that accounted for more than 50% of their capital, leading to the suspension of trading or bankruptcy. Thus, we exclude such data from our study. Economies 2023, 11, 21 15 of 41 Table 1. Summary Statistics. Variable Observations Mean Median Minimum Maximum Standard Deviation ROA ROE Tobin’s Q %IB SAC D-INC (SAR) CEO-Ex-Pay (SAR) CEO-Shares (%) CEO-Ten CEOAge CEO Turnover BS Ln(Assets) FirmAge VOL RiskRet (%) DeInMarket (%) Ln(gross written) 210 210 210 210 210 210 210 210 210 210 210 210 210 210 210 210 210 210 −3.5922 −9.5561 3.7773 46.4327 3.1143 1,004,805 5,146,216 0.5228 3.0667 50.6714 0.2333 8.5571 19.4193 5.7238 16.8407 64.6286 0.8944 23.6997 −0.9512 −2.1808 2.8018 40 3 770,500 4,374,000 0 3 50 0 9 19.2131 5 12.7224 67.5255 0.90 23.6412 −51.8177 −159.0127 0.3335 20 0 0 0 0 1 33 0 5 17.1995 1 3.8282 0 0.62 23.1138 6.7045 30.3397 16.9274 100 5 9,530,000 23,002,000 45.5013 9 77 1 11 23.0224 29 280.1844 99.9934 1.10 24.1404 8.2787 26.9908 3.1330 17.7373 0.6461 1,133,655 3,736,234 4.4401 1.9504 8.4904 0.4240 1.6041 1.1235 4.3031 20.9670 21.9242 0.1590 0.3064 ROA, ROE, and Tobin’s Q are performance measures. %IB is the board composition. SAC is the size of the Audit Committee D-INC is director incentives (in Saudi riyals (SAR)). CEO-Ex-Pay is the CEO and top executive pay (in Saudi riyals (SAR)). CEO-Shares(%) is the percentage of CEO-owned shares. CEO-Ten is CEO tenure. CEOAge is the CEO age. CEO Turnover is a dependent dummy variable that takes 1 if the CEO changed, and 0 otherwise. BS is the board size. Ln(Assets) is the natural logarithm of firm assets. FirmAge is the firm age. VOL is volatility. RiskRe is (risk retention) net premiums written over gross premiums written, and DeInMarket is (depth of insurance market) defined as gross written premiums divided by total GDP. Ln(gross written) is the natural logarithm of gross premiums written, measured by the sum of both direct premiums written and assumed premiums written, before deducting the ceded reinsurance. We do not use the time fixed effects in the models because we do not want to increase the number of predictors with regard to the sample size. As stated previously, the number of firms increase in the timeframe examined (21 listed insurance firms in 2008 and 35 by the end of 2014). Thus, the number of observations is not equal each year; adding time dummies (increasing explanatory variables) would have affected the significance of the estimates and omitted-variable bias in 2013 and 2014. Table 2 provides a correlation matrix of the variables in the analysis. As expected, there is a strong correlation relationship between ROA and ROE (0.7233); therefore, we do not include these in the same empirical model. Additionally, there is a strong negative correlation between Tobin’s Q and Ln(Assets) (−0.6635) in all cases, as Tobin’s Q is calculated by dividing firm market capital by total assets. CEO-Ex-Pay and Ln(Assets) are positively correlated (0.4941). This is as expected, as larger firms are more likely to pay their senior executives more. Expectedly, CEO-Ex-Pay and CEO tenure are also positively correlated (0.4636), as the longer the CEO holds office, the larger the increase in the CEO’s salary. The relationship between CEO Turnover and CEO tenure will always be negative (−05860); if the CEO holds office until retirement, the CEO will not change until after that. Ln(gross written) and Ln(firm age) have a positive correlation (0.5688), which implies that written gross premiums increase over time. The remaining variables do not show a strong correlation. Economies 2023, 11, 21 16 of 41 Table 2. Correlation Matrix. Variable ROA ROE Tobin’s Q %IB SAC D-INC CEO-Ex-Pay CEO-Shares (%) CEO-Ten CEOAge CEO Turnover Ln(BS) Ln(Assets) Ln(cid:0)FirmAge (cid:1) VOL RiskRet DeInMarket Ln(gross written) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 1 1 0.7233 *** (0.0000) −0.2475 *** (0.0003) 0.0511 (0.4615) −0.0379 (0.5849) 0.1526 ** (0.0271) 0.2707 *** (0.0001) 0.1155 * (0.0952) 0.3033 *** (0.0000) 0.1586 ** (0.0215) −0.2045 *** (0.0029) −0.0446 (0.5203) 0.3439 *** (0.0000) 0.3760 *** (0.0000) −0.1022 (0.1399) 0.1827 *** (0.0080) 0.0321 (0.6438) 0.1547 ** (0.0250) 2 1 −0.2906 *** (0.0000) −0.1123 (0.1046) 0.0386 (0.5782) 0.0575 (0.4072) 0.2253 *** (0.0010) 0.1117 (0.1066) 0.1828 *** (0.0079) 0.1080 (0.1186) −0.1928 *** (0.0051) −0.0377 (0.5867) 0.3525 *** (0.0000) 0.1186 * (0.0866) −0.0298 (0.6677) 0.0577 (0.4058) −0.0019 (0.9782) 0.0037 (0.9572) 3 1 −0.0406 (0.5584) 0.0036 (0.9582) −0.0151 (0.8282) −0.1988 *** (0.0038) −0.0770 (0.2669) −0.2042 *** (0.0029) −0.1328 * (0.0546) 0.0846 (0.2222) 0.0450 (0.5163) −0.6635 *** (0.0000) −0.1211 * (0.0799) 0.2069 *** (0.0026) −0.0261 (0.7068) 0.0070 (0.9200) 0.1205 * (0.0814) 4 1 −0.0032 (0.9634) 0.0793 (0.2529) 0.0742 (0.2847) −0.0090 (0.8974) 0.0127 (0.8545) 0.0178 (0.7972) 0.0430 (0.5350) −0.0026 (0.9704) −0.0087 (0.9002) 0.3290 *** (0.0000) −0.0622 (0.3700) 0.1176 * (0.0892) 0.0647 (0.3509) 0.1859 *** (0.0069) 5 1 0.1082 (0.1181) 0.0309 (0.6564) −0.0204 (0.7690) 0.0888 (0.1997) 0.0339 (0.6251) −0.0803 (0.2464) 0.0718 (0.3001) 0.1693 ** (0.0141) 0.1242 * (0.0725) 0.1993 *** (0.0037) 0.0669 (0.3350) 0.0296 (0.6701) 0.0368 (0.5960) 6 7 8 9 10 11 12 13 14 15 16 17 18 VIF 1 0.2239 *** (0.0011) 0.0181 (0.7946) 0.2064 *** (0.0027) 0.2444 *** (0.0004) −0.0713 (0.3036) 0.1796 *** (0.0091) 0.2606 *** (0.0001) 0.2504 *** (0.0002) −0.0592 (0.3933) 0.1482 ** (0.0318) 0.0163 (0.8148) 0.1786 *** (0.0095) 1 0.3053 *** (0.0000) 0.4636 *** (0.0000) 0.1774 *** (0.0100) −0.1160* (0.0936) −0.0221 (0.7498) 0.4941 *** (0.0000) 0.3348 *** (0.0000) −0.1826 *** (0.0080) 0.2497 *** (0.0003) 0.0531 (0.4436) 0.2538 *** (0.0002) 1 0.1516 ** (0.0281) 0.2122 *** (0.0020) −0.0650 (0.3489) 0.0585 (0.3990) 0.1455** (0.0352) 0.0274 (0.6930) −0.0352 (0.6119) 0.0196 (0.7775) −0.0825 (0.2339) 0.0005 (0.9945) 1 0.3911 *** (0.0000) −0.5860 *** (0.0000) −0.1847 *** (0.0073) 0.4062 *** (0.0000) 0.4280 *** (0.0000) −0.1264 * (0.0675) 0.2562 *** (0.0002) 0.1004 (0.1470) 0.2602 *** (0.0001) 1 −0.2152 *** (0.0017) 0.0452 (0.5146) 0.2271 *** (0.0009) 0.1615 ** (0.0192) −0.0919 (0.1844) −0.0889 (0.1995) −0.0526 (0.4487) −0.0914 (0.1872) 1 0.0782 (0.2589) −0.1559 ** (0.0239) −0.0340 (0.6246) 0.1100 (0.1119) −0.0606 (0.3820) −0.0195 (0.7788) −0.0011 (0.9875) 1 −0.1378 ** (0.0461) −0.1389 ** (0.0443) 0.1113 (0.1077) −0.0506 (0.4660) −0.0677 (0.3291) −0.1107 (0.1098) 1 0.3497 *** (0.0000) −0.0994 (0.1512) 0.1331* (0.0542) 0.0433 (0.5324) 0.0669 (0.3349) 1 −0.2509 *** (0.0002) 0.3049 *** (0.0000) 0.1925 *** (0.0051) 0.5688 *** (0.0000) 1 −0.1085 (0.1171) −0.0942 (0.1740) −0.0601 (0.3865) 1 0.1901 *** (0.0057) 0.2493 *** (0.0003) 1 0.4477 *** (0.0000) 1.17 1.13 1.27 1.82 1.17 2.93 1.44 1.79 1.17 1.66 2.45 1.23 1.24 1.33 1 2.13 Notes: *, **, and *** represent significance at 10%, 5%, and 1%, respectively. The p-values are in parentheses. VIF is the variance inflation factor. The mean of VIF is 1.60. Economies 2023, 11, 21 17 of 41 Multicollinearity refers to variables that are highly correlated with each other. The variance inflation factor (VIF) is a measure of multicollinearity. For regression models, the VIF is the ratio of the model variance to the model variance after the introduction of a single independent variable. However, all of our independent variables and the mean values of the VIFs are less than 3, indicating no multicollinearity. It is well established that the higher the VIF, the less accurate the regression findings; VIFs greater than 10 show a high correlation (multicollinearity) and should be concerning. Most studies on board composition utilise ratios or linkage to quantify the degree of board independence. Many governance authorities also stipulate a minimum number of independent directors. As stated, the minimum number of independent directors required in the KSA is two, or one-third of the total number of directors, whichever is larger. In Table 3, in RE Model 1, board composition (%IB) shows a significant negative relationship with ROE and Tobin’s Q, with coefficients of −0.2221 and −0.0176, respectively, at 10% significance. This result is in line with Arif (2019). The result is also in line with the FE model for the %IB, which shows a significantly negative relationship with ROE (−0.3862) at 5% significance. Thus, these findings do not support Hypothesis 1. The %IB with the four binary dummy variables, IB(3), IB(4), IB(5), and IB(6), are shown in Model 2, indicating the presence of three, four, five, or six or more independent directors on a board. These findings also reveal no statistically significant effect between the number of independent directors and ROA or Tobin’s Q, as Akeem et al. (2014) found. In addition, after excluding IB(6) in the RE (−16.5177) and FE (−26.6009) models, there is a significantly negative relationship with ROE at 10% and 5% significance, respectively. Once again, Hypothesis 1 is not supported. When the number of independent board members reaches six or more, the effect is negative on firm performance (ROE). This can be attributed to a lack of knowledge and experience or operational weakness in board efficiency within the insurance company. Our findings support those of Al-Faryan (2017), which show that independent directors are associated with negative firm performance, significant at the 10% level, using the Fama–French three-factor model. In studies of Chinese board independence and firm performance, Liu et al. (2015), looking at ROA and ROE, and Li et al. (2015), looking at ROA and Tobin’s Q in an opposite sign, found a significantly positive relationship between independent directors and firm performance. Liu et al. (2015) and Al-Faryan (2021) also found that boards having at least two and up to five or more independent directors had a significantly positive impact on performance. Broome et al. (2011) argue that independent directors are required for a board to be efficient and achieve critical mass; otherwise, there is a tendency for boards to have no perceptible influence on the firm’s monitoring or decisions. Table 4 shows the results of including an additional variable (%IB*VOL), which mea- sures monitoring costs in the original model (Table 3). The board composition (%IB) coefficient indicates a significantly negative relationship with ROE in the RE (−0.3635) and FE (−0.7037) models, respectively, and with Tobin’s Q in the RE model (−0.0272) at 5%, 1%, and 10% significance, respectively. This result is in line with the results of Al-Faryan (2021) and Maug (1997), except in the FE model for ROA, and, as already stated, rejects Hypothesis 1. The conclusion is that a higher percentage of independent directors on the board contributes to worse performance among Saudi insurance firms. The ROE results show that %IB has a significantly negative relationship at −0.7037 (RE) at 1% significance. Meanwhile, %IB*VOL shows a positive relationship at 1.5229 (RE) at 10% significance, failing to support Hypothesis 3. Insurance firms need to monitor independent board members, which leads to an increase in costs due to the weakness of the expertise of the independent board members or their understanding of the characteristics of the Saudi insurance market, which differ from the West. This then affects the performance of the company and contributes to company losses. Moreover, higher monitoring costs lead to a decrease in the benefits associated with the number of outside directors and board independence (Adams and Ferreira 2007; Al-Faryan 2021; Raheja 2005). Economies 2023, 11, 21 18 of 41 Table 3. The effect of board composition and board independent dummies on insurance performance. ROA ROE Tobin’s Q RE FE RE FE RE FE Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Constant 3.7967 (67.5470) IB(3) IB(4) IB(5) IB(6) %IB Ln(BS) Ln(Assets) Ln(cid:0)FirmAge(cid:1) VOL RiskRet DeInMarket Ln(gross written) Wald Chi2 F Statistic R2 (%) −0.0451 (0.0366) 3.1286 (3.4136) 2.0085 *** (0.7493) 7.1902 *** (1.8539) 0.0310 (0.0250) 0.0189 (0.0275) −0.5144 (3.2055) −2.6890 (2.7749) 53.45 *** −6.3250 (67.0689) 1.5871 (1.9019) −1.8699 (2.2677) −3.3046 (2.5441) −1.4723 (2.5587) 5.8437 (4.0037) 2.0220 *** (0.7525) 7.1526 *** (1.8348) 0.0353 (0.0249) 0.0166 (0.0275) −0.3909 (3.2198) −2.5833 (2.7531) 60.64 *** 61.0747 (119.1293) −0.0505 (0.0427) 6.9750 (4.8228) 3.3110 *** (1.2253) 10.6234 *** (3.4496) 0.0610 ** (0.0277) 0.0008 (0.0315) 1.1928 (3.3645) −6.7638 (5.1649) 27.2824 (120.7192) 1.6737 (2.1202) −1.1789 (2.4949) −3.2452 (2.7010) −1.3378 (2.8331) 10.0082 * (5.1385) 3.1229 ** (1.2654) 9.8178 *** (3.4919) 0.0634 ** (0.0277) −0.0038 (0.0320) 1.1752 (3.3886) −5.4789 (5.2529) −107.7824 (200.2462) −0.2221 * (0.1143) 1.9347 (9.7098) 8.3389 *** (1.9449) 3.5480 (4.9939) 0.0229 (0.0887) 0.0230 (0.0899) −1.1806 (12.3392) −2.6990 (8.2831) 29.20 *** −144.5460 (202.4618) 3.6853 (6.7005) −5.4985 (7.9249) −6.6657 (9.1131) −16.5177 * (8.7572) 13.3729 (12.8572) 7.9351 *** (2.0196) 4.8734 (5.0983) 0.0387 (0.0885) 0.0245 (0.0913) −0.3115 (12.3168) −2.2699 (8.3497) 35.48 *** −37.4082 (469.4880) −0.3862 ** (0.1685) 4.5417 (19.0065) 17.9176 *** (4.8290) 11.8276 (13.5947) 0.0913 (0.1092) 0.0286 (0.1240) 1.6937 (13.2596) −14.1485 (20.3549) −218.9571 (474.3023) 3.5985 (8.3303) −5.3803 (9.8023) −9.8676 (10.6123) −26.6009 ** (11.1314) 22.4153 (20.1892) 15.3649 *** (4.9718) 8.7395 (13.7194) 0.0953 (0.1087) 0.0711 (0.1258) 0.7829 (13.3139) −6.4116 (20.6384) 24.6199 (18.9903) −0.0176 * (0.0106) −0.8803 (0.9607) −2.1879 *** (0.2050) 0.7480 (0.5107) 0.0255 *** (0.0075) 0.0045 (0.0081) −0.3346 (0.9796) 0.9626 (0.7798) 137.57 *** 19.0928 (18.4566) −0.0810 (0.5729) −0.0158 (0.6813) 0.3880 (0.7736) −1.1588 (0.7618) −0.5479 (1.1509) −2.1916 *** (0.1973) 0.6762 (0.4878) 0.0247 *** (0.0075) 0.0056 (0.0081) −0.4015 (1.0059) 1.1465 (0.7575) 150.52 *** 59.6413 * (35.9481) −0.0155 (0.0129) −0.9411 (1.4553) −2.9659 *** (0.3697) 1.4098 (1.0409) 0.0276 *** (0.0084) 0.0055 (0.0095) 0.1062 (1.0153) 0.0557 (1.5586) 60.1045 * (36.5258) −0.4672 (0.6415) −0.4136 (0.7549) 0.0359 (0.8172) −1.3927 (0.8572) −0.5199 (1.5548) −3.0261 *** (0.3829) 1.4832 (1.0565) 0.0268 *** (0.0084) 0.0087 (0.0097) 0.0054 (1.0253) 0.0289 (1.5893) 19.52 22.49 6.42 *** 18.09 5.07 *** 20.20 13.85 16.49 2.94 *** 13.66 2.63 *** 16.06 50.50 51.83 9.35 *** 49.32 7.10 *** 49.94 Notes: *, **, and *** represent significance at 10%, 5%, and 1%, respectively. Standard errors are in parentheses. ROA, ROE, and Tobin’s Q are performance measures. RE is random effects. FE is fixed effects. %IB is the board composition. IB(3) . . . IB(6) are the number of independent members. Ln(BS) is the natural logarithm of the board size. Ln(Assets) is the natural logarithm of firm assets. Ln(cid:0)FirmAge(cid:1) is the natural logarithm of firm age. VOL is volatility. RiskRet is (risk retention) net premiums written over gross premiums written, and DeInMarket is (depth of insurance market) defined as gross written premiums divided by total GDP. Ln(gross written) is the natural logarithm of gross premiums written, measured by the sum of both direct premiums written and assumed premiums written, before deducting the ceded reinsurance. Economies 2023, 11, 21 19 of 41 Table 4. The effect of board composition and monitoring costs on insurance performance. ROA ROE Tobin’s Q RE FE RE FE RE FE 3.1860 (68.1176) −0.0666 (0.0536) 0.1147 (0.2107) 3.3969 (3.4598) 2.0120 *** (0.7570) 7.2352 *** (1.8717) −0.0127 (0.0852) 0.0199 (0.0277) −0.3512 (3.2141) −2.6656 (2.8003) 53.66 *** 19.51 58.1408 (119.5856) −0.0714 (0.0628) 0.1006 (0.2203) 7.2435 (4.8699) 3.2580 *** (1.2337) 10.5759 *** (3.4593) 0.0220 (0.0900) 0.0023 (0.0317) 1.2777 (3.3777) −6.5882 (5.1915) 5.70 *** 18.13 −115.6148 (201.3361) −0.3635 ** (0.1707) 0.8457 (0.7714) 3.0658 (9.8399) 8.4244 *** (1.9663) 3.4240 (5.0430) −0.3032 (0.3114) 0.0297 (0.0905) −0.0556 (12.3495) −2.3550 (8.3258) 30.18 *** 14.13 −81.8409 (467.1969) −0.7037 *** (0.2454) 1.5229 * (0.8607) 8.6085 (19.0257) 17.1137 *** (4.8199) 11.1076 (13.5149) −0.5007 (0.3517) 0.0518 (0.1239) 2.9798 (13.1959) −11.4888 (20.2821) 3.00 *** 14.02 24.0202 (19.0004) −0.0272 * (0.0156) 0.0539 (0.0637) −0.7904 (0.9664) −2.1887 *** (0.2049) 0.7409 (0.5106) 0.0046 (0.0257) 0.0050 (0.0081) −0.2696 (0.9838) 0.9933 (0.7805) 138.29 *** 50.75 58.1242 (36.0419) −0.0263 (0.0189) 0.0520 (0.0664) −0.8022 (1.4677) −2.9934 *** (0.3718) 1.3852 (1.0426) 0.0074 (0.0271) 0.0063 (0.0096) 0.1501 (1.0180) 0.1465 (1.5647) 8.36 *** 49.64 Constant %IB %IB*VOL Ln(BS) Ln(Assets) Ln(FirmAge) VOL RiskRet DeInMarket Ln(gross written) Wald Chi2 F Statistic R2 (%) Notes: *, **, and *** represent significance at 10%, 5%, and 1%, respectively. Standard errors are in parentheses. ROA, ROE, and Tobin’s Q are performance measures. RE is random effects. FE is fixed effects. %IB is the board composition. %IB*VOL is board composition multiplied by volatility. Ln(BS) is the natural logarithm of the board size. Ln(Assets) is the natural logarithm of firm assets. Ln(cid:0)FirmAge(cid:1) is the natural logarithm of firm age. VOL is volatility. RiskRet (risk retention) are net premiums written over gross premiums written, and DeInMarket (depth of insurance market) is defined as gross written premiums divided by total GDP. Ln(gross written) is the natural logarithm of gross premiums written, measured by the sum of both direct premiums written and assumed premiums written, before deducting the ceded reinsurance. We can then ask who can be an effective board member of an insurance firm with sufficient competence and experience to monitor the firm in such a way as to reduce monitoring costs while supporting high performance? The answer may be an individual with practical experience in the market or local actuaries, as opposed to international experts who do not have local market experience or local knowledge. This implies that the increased monitoring costs faced by independent directors will lead to increasing the performance of Saudi insurance firms, a result that contrasts with Liu et al. (2015) and Al-Faryan (2021) regarding ROE (FE). Table 5 shows no effects for the size of the audit committee (SAC) on insurance firm performance in all estimates in Model 1. Thus, Hypothesis 4 is not supported. The auditors play a large representative role, as they are concerned with personal interests rather than those of the company, and this characteristic applies to management as well. Our results support those of Ghabayen (2012) and Al-Faryan (2017), who found that the SAC has a very small insignificant negative coefficient, which means that firm audit committee size has no significant relation with performance at Saudi insurance firms. In Model 2, we replace the SAC with three binary dummy variables: ASize(3), ASize(4), and ASize(5), which indicate the presence of three, four, or five audit committee members, respectively. The results reveal no significant relationship between all ASize and ROE, as also evidenced in Datta (2018), Ghabayen (2012), and Yemane et al. (2015). However, ROA and Tobin’s Q, excluding Tobin’s Q with ASize(4), underestimate RE (1.5592) and relate positively with Tobin’s Q at 10% significance, which supports Hypothesis 4. In contrast, ROA and ASize(5) underestimate RE (−6.3918) and relate negatively at 10% significance, thus rejecting Hypothesis 4. This supports the findings of Maharjan (2019), Fadun (2013), and Fekadu (2015). Economies 2023, 11, 21 20 of 41 Table 5. The effect of audit committee size and audit size dummies on insurance performance. ROA ROE Tobin’s Q RE FE RE FE RE FE Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Constant Asize(3) Asize (4) Asize(5) SAC Ln(BS) Ln(Assets) Ln(cid:0)FirmAge(cid:1) VOL RiskRet DeInMarket Ln(gross written) Wald Chi2 F Statistic R2 (%) −0.7568 (66.2930) −0.9297 (0.8399) 3.4397 (3.3875) 2.1600 *** (0.7299) 6.5863 *** (1.7628) 0.0322 (0.0254) 0.0193 (0.0274) −0.5157 (3.2299) −2.5754 (2.7244) 52.85 *** −17.0516 (65.3392) 3.2457 (2.4301) 0.6149 (2.9607) −6.3918 * (3.7450) 4.0481 (3.2906) 2.1891 *** (0.7029) 6.2249 *** (1.7084) 0.0445 * (0.0250) 0.0169 (0.0265) 0.6701 (3.1885) −2.2153 (2.6978) 68.85 *** 70.3277 (119.3659) −0.2046 (0.9055) 7.3399 (4.8983) 3.5107 *** (1.2260) 10.4731 *** (3.4613) 0.0594 ** (0.0278) −0.0027 (0.0316) 1.3966 (3.3736) −7.4096 (5.1594) 13.4274 (120.9539) 2.3579 (2.5065) 1.7790 (3.1422) −5.0099 (4.1198) 8.0246 (4.8937) 3.4694 *** (1.2091) 8.7540 ** (3.4968) 0.0594 ** (0.0275) 0.0024 (0.0312) 1.3439 (3.3553) −5.0422 (5.2231) −128.2512 (200.3712) −0.9085 (2.9667) 1.7093 (9.7169) 8.8583 *** (1.9291) 0.8718 (4.7850) 0.0221 (0.0920) 0.0172 (0.0901) −1.2593 (12.4940) −2.3583 (8.3071) 25.24 *** −65.3339 (201.7822) 10.7974 (9.3119) −1.5203 (10.8822) 2.7059 (13.6267) 1.7133 (9.6616) 9.3439 *** (1.9202) 2.4242 (4.7664) 0.0386 (0.0935) 0.0080 (0.0896) 3.1063 (12.5623) −6.1528 (8.4508) 31.43 *** 30.6686 (475.7870) 0.7950 (3.6094) 5.3580 (19.5243) 19.1632 *** (4.8867) 10.4817 (13.7965) 0.0739 (0.1110) −0.0072 (0.1259) 3.2645 (13.4468) −18.8359 (20.5650) 37.3265 (488.2025) 8.8437 (10.1168) 0.9108 (12.6827) 4.6029 (16.6285) 2.9538 (19.7521) 19.4286 *** (4.8803) 9.5211 (14.1139) 0.0722 (0.1112) 0.0082 (0.1258) 5.0412 (13.5428) −19.3736 (21.0819) 19.8334 (18.4676) 0.2483 (0.2527) −0.9704 (0.9374) −2.1195 *** (0.1951) 0.4703 (0.4738) 0.0232 *** (0.0077) 0.0029 (0.0080) −0.4051 (1.0010) 1.0762 (0.7599) 143.91 *** 12.5247 (18.3795) 0.8168 (0.7738) 1.5592* (0.9246) 0.2239 (1.1622) −0.7598 (0.9139) −2.1076 *** (0.1876) 0.3711 (0.4583) 0.0254 *** (0.0078) 0.0024 (0.0079) −0.4078 (1.0265) 1.3590* (0.7630) 155.70 *** 20.48 26.66 6.20 *** 17.91 5.68 *** 21.98 12.57 15.05 2.22 ** 12.07 1.95 ** 13.46 50.15 51.22 62.2754 * (36.0115) 0.1118 (0.2732) −0.9753 (1.4778) −2.9256 *** (0.3699) 1.3493 (1.0442) 0.0267 *** (0.0084) 0.0038 (0.0095) 0.1695 (1.0178) −0.1234 (1.5565) 9.12 *** 48.53 56.3933 (37.0938) 0.2252 (0.7687) 0.5412 (0.9636) −0.1812 (1.2634) −0.7952 (1.5008) −2.9266 *** (0.3708) 1.2064 (1.0724) 0.0268 (0.0084) 0.0042 (0.0096) 0.1144 (1.0290) 0.1245 (1.6018) 7.28 *** 49.05 Notes: *, **, and *** represent significance at 10%, 5%, and 1%, respectively. Standard errors are in parentheses. ROA, ROE, and Tobin’s Q are performance measures. RE is random effects, and FE is fixed effects. SAC is the size of the Audit Committee. Asize(3) . . . Asize(5) are the number of Audit Committee members. Ln(BS) is the natural logarithm of the board size. Ln(Assets) is the natural logarithm of firm assets. Ln(cid:0)FirmAge(cid:1) is the natural logarithm of firm age. VOL is volatility. RiskRet is risk retention as net premiums written over gross premiums written, and DeInMarket (depth of insurance market) is defined as gross written premiums divided by total GDP. Ln(gross written) is the natural logarithm of gross premiums written, measured by the sum of both direct premiums written and assumed premiums written, before deducting the ceded reinsurance. Economies 2023, 11, 21 21 of 41 This means that below ASize(4), there is no effect on insurance performance, while there is a positive effect on insurance performance when there are four members. A committee of more than four has a negative effect on insurance performance. Thus, when the SAC is five, it becomes ineffective. One possible reason may be that with five or more, operations and responsibilities may be abandoned, and committees may fail to complete the tasks they are supposed to do quickly and accurately; however, a smaller SAC, such as ASize(4), relates positively to firm performance. Table 6 shows the results of Model 1. The DID for board composition indicates that the performance measures of ROA in the RE (3.8821) and FE (3.0084) models, Tobin’s Q in the RE (0.7917) and FE (0.9705) models, and ROE in the RE (12.8994) model, except for the FE model, are significantly positive with performance at 5% and 10% significance for ROA (RE) and Tobin’s Q (FE) for the listed Saudi insurance firms that implemented the 2009 post-corporate governance regulations (stipulating that at least two or one-third of the directors must be independent). These results support Hypothesis 2. According to our data and the annual reports, all of the listed Saudi insurance firms in our sample, except for four, implemented immediately the SAMA and CMA corporate governance regulations as per the expected date. The insurance firms generally complied with the corporate governance regulations, as fines for non-compliance had doubled. Our results support those of Liu et al. (2015), who found significantly positive relationships with ROA and ROE in the Chinese market; however, there is a non-significantly positive relationship with ROE under the FE model. Al-Faryan (2021) found the opposite sign with ROA, while supporting our result for Tobin’s Q among the listed Saudi firms. Model 2 assesses the effect of the board of directors’ establishment of audit committees per regulations. According to an amendment to the audit committee regulation in 2017, the audit committee must be formed by a decision of the company’s ordinary general assembly; the members of the committee will be shareholders or others, provided that at least one is an independent member and that no member is an executive director. As stated, the SAC must not be less than three or more than five and include one member who is a specialist in financial and accounting matters. In Model 1, similar to the reason for immediately implementing independent board members in the year of regulation, the difference in the result is that insurance firms that applied the 2009 regulation regarding SAC show a positive effect on performance, as shown in Model 2, wherein the measure of ROA in the RE (7.0552) and FE (6.8950) models and of Tobin’s Q in the FE (1.3653) model are positive at 1% significance and the others at 5% significance. These results support Hypothesis 5. After the regulation, three of the listed insurance firms had fewer than three members on their audit committees, and four had fewer than two or one-third independent board members. These firms that were non-compliant with the corporate governance regulations were fined. Table 7 presents the results for the relationship between CEO age and insurance performance in the RE and FE models. We found a significantly positive relationship in the Tobin’s Q models but not in the ROA and ROE models. In Model 1, CEO age has a significantly positive effect on Tobin’s Q in the RE (0.0382) and FE (0.0668) models at 10% and 5% significance, respectively. Model 2 uses a CEO retirement dummy variable equal to 1 if the CEO is aged 60 or above, and 0 otherwise. The results show that the CEO retirement age has a significantly positive effect on Tobin’s Q in the RE (1.2933) and FE (1.5081) models at 1% significance. Thus, both models support Hypothesis 6. This means that in Model 1, firms with CEOs reaching retirement age (60) or above and still in the job show better performance in terms of Tobin’s Q than the other firms. Thus, an increase in CEO age improves firm performance, which could be attributed to a CEO’s cumulative work experience in the insurance sector. In contrast, Peni (2014) finds CEO age to have a non-significant effect on Tobin’s Q while having a positive significant effect on ROA. Economies 2023, 11, 21 22 of 41 Table 6. The effect of board composition and audit committee size regulation (DID-DID 2) on insurance performance. ROA ROE Tobin’s Q RE FE RE FE RE FE Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Constant DID DID 2 Ln(BS) Ln(Assets) Ln(cid:0)FirmAge(cid:1) VOL RiskRet DeInMarket Ln(gross written) Wald Chi2 F Statistic R2 (%) −92.2933 (66.4366) 3.8821 ** (1.5305) 0.1098 (2.9148) 2.3461 *** (0.5579) −0.1509 (1.7505) −0.0868 (0.0575) 0.0064 (0.0258) 0.3865 (3.0514) 1.7765 (2.7892) 58.87 *** −67.9857 (65.4469) 7.0552 *** (1.6080) −1.6913 (2.9098) 2.1267 *** (0.5523) −0.1958 (1.7408) −0.1018 * (0.0537) −0.0090 (0.0254) −3.6936 (3.1105) 1.1826 (2.7246) 75.52 *** 26.9915 (120.9797) 3.0084 * (1.6810) −1.0057 (4.9194) 3.5746 *** (1.1618) 4.6909 (3.9975) −0.0769 (0.0601) 0.0056 (0.0315) 1.7690 (3.2517) −4.5541 (5.0984) −0.9747 (114.7766) 6.8950 *** (1.7449) −4.6586 (4.6522) 3.2747 *** (1.1042) 2.1213 (3.8002) −0.0908 (0.0565) −0.0092 (0.0300) −3.0050 (3.3867) −2.5247 (4.8761) −142.0675 (281.4380) 12.8994 * (7.2036) 1.1069 (11.1196) 9.9616 *** (2.0767) −7.1141 (6.3850) 0.0517 (0.2742) 0.0238 (0.1110) −2.4828 (14.8382) −2.5671 (11.8905) 27.08 *** −100.1930 (283.9532) 17.9074 ** (7.9955) −1.6003 (11.3074) 9.2508 *** (2.0692) −6.2605 (6.3942) −0.0113 (0.2660) −0.0043 (0.1132) −12.0066 (15.6670) −3.2911 (11.8585) 28.56 *** −377.8101 (589.2447) 11.6330 (8.1876) −0.3441 (23.9605) 21.6705 *** (5.6586) −11.9363 (19.4704) 0.0837 (0.2929) 0.0444 (0.1532) −4.2964 (15.8380) −1.7715 (24.8321) −425.8992 (579.2212) 20.1698 ** (8.8054) −12.4160 (23.4774) 20.6508 *** (5.5725) −17.1877 (19.1778) 0.0262 (0.2851) 0.0132 (0.1513) −17.1251 (17.0911) 2.8474 (24.6075) 48.6276 ** (22.0616) 0.7917 * (0.4641) −0.8654 (0.9984) −2.0213 *** (0.1970) 0.8497 (0.6260) 0.0551 *** (0.0173) 0.0124 (0.0082) 0.6236 (0.9138) −0.3520 (0.9233) 141.80 *** 49.4751 ** (22.4167) 0.9933 ** (0.5038) −1.2404 (1.0200) −2.0982 *** (0.2000) 0.8433 (0.6377) 0.0508 *** (0.0168) 0.0109 (0.0083) 0.1067 (0.9715) −0.2723 (0.9334) 139.39 *** 17.5573 (34.7025) 0.9705** (0.4822) −0.5865 (1.4111) −2.7714 *** (0.3333) −0.3550 (1.1467) 0.0546 *** (0.0173) 0.0058 (0.0090) 0.3983 (0.9327) 1.6639 (1.4624) 16.4818 (34.1703) 1.3653 *** (0.5195) −1.4933 (1.3850) −2.8496 *** (0.3287) −0.5640 (1.1314) 0.0496 *** (0.0168) 0.0044 (0.0089) −0.3966 (1.0083) 1.8914 (1.4517) 32.66 36.41 6.12 *** 25.93 8.26 *** 32.02 17.41 18.26 2.15 ** 16.53 2.61 ** 17.23 63.02 62.66 11.75 *** 58.55 12.37 *** 57.75 Notes: *, **, and *** represent significance at 10%, 5%, and 1%, respectively. Standard errors are in parentheses. ROA, ROE, and Tobin’s Q are performance measures. RE is random effects, and FE is fixed effects. DID is difference-in-differences for board composition, and DID 2 is difference-in-differences for audit size of committees, wherein DID and DID 2 are Treated*post regulation (the treated group multiplied by post regulation). Ln(BS) is the natural logarithm of the board size. Ln(Assets) is the natural logarithm of firm assets. Ln(cid:0)FirmAge(cid:1) is the natural logarithm of firm age. VOL is volatility. RiskRet is (risk retention) net premium written over gross premium written, and DeInMarket (depth of insurance market) is defined as gross written premiums divided by the total GDP. Ln(gross written) is the natural logarithm of gross premiums written, measured by sum of both direct premiums written and assumed premiums written, before deducting the ceded reinsurance. Economies 2023, 11, 21 23 of 41 Table 7. The effect of CEO age on insurance performance. ROA ROE Tobin’s Q RE FE RE FE RE FE Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Constant RET Age Ln(BS) Ln(Assets) Ln(cid:0)FirmAge(cid:1) VOL RiskRet DeInMarket Ln(gross written) Wald Chi2 F Statistic R2 (%) −0.7497 (69.0916) 0.0061 (0.0762) 3.1008 (3.4161) 2.1032 *** (0.7463) 6.6178 *** (1.8354) 0.0288 (0.0250) 0.0162 (0.0275) −0.5123 (3.2221) −2.6260 (2.8249) 51.50 *** 0.3980 (67.4842) 0.0056 (1.6876) 3.0894 (3.4138) 2.1075 *** (0.7514) 6.6414 *** (1.8060) 0.0287 (0.0250) 0.0162 (0.0275) −0.5089 (3.2321) −2.6655 (2.7737) 51.49 *** 78.9727 (121.1916) −0.0379 (0.0906) 6.8593 (4.8937) 3.4949 *** (1.2209) 10.6310 *** (3.4844) 0.0596 ** (0.0278) −0.0028 (0.0314) 1.4249 (3.3729) −7.6756 (5.2023) 69.7052 (119.5041) −0.1336 (1.8976) 7.1221 (4.8849) 3.4890 *** (1.2220) 10.4479 *** (3.4630) 0.0590 ** (0.0278) −0.0035 (0.0314) 1.3702 (3.3964) −7.3670 (5.1674) −144.1782 (203.8803) 0.0782 (0.2372) 1.2179 (9.6894) 8.6833 *** (1.9353) 0.3955 (4.8551) 0.0165 (0.0894) 0.0194 (0.0906) −1.4682 (12.4855) −1.7478 (8.4277) 25.56 *** −132.8353 (199.6423) −2.2552 (5.5483) 1.8034 (9.6797) 8.9523 *** (1.9597) 0.8641 (4.7522) 0.0142 (0.0893) 0.0146 (0.0899) −1.7319 (12.5155) −2.3248 (8.2820) 25.77 *** 37.9489 (483.3050) −0.02715 (0.3612) 5.8017 (19.5159) 19.2644 *** (4.8689) 10.6729 (13.8954) 0.0760 (0.1109) −0.0036 (0.1253) 3.2804 (13.4511) −19.1279 (20.7465) 7.2923 (474.5818) −8.3932 (7.5357) 3.0994 (19.3993) 19.4303 *** (4.8528) 10.0254 (13.7523) 0.0742 (0.1103) −0.0019 (0.1247) 1.5406 (13.4878) −17.6283 (20.5213) 16.5741 (19.4919) 0.0382 * (0.0221) −0.8896 (0.9653) −2.1928 *** (0.2062) 0.3798 (0.5075) 0.0249 *** (0.0074) 0.0041 (0.0081) −0.3517 (0.9769) 1.2180 (0.7974) 136.66 *** 23.0009 (18.5765) 1.2933 *** (0.4900) −0.9545 (0.9400) −2.2256 *** (0.2012) 0.4381 (0.4828) 0.0251 *** (0.0074) 0.0039 (0.0079) −0.1599 (0.9774) 1.0421 (0.7638) 147.65 *** 46.7401 (35.9342) 0.0668 ** (0.0269) −0.3361 (1.4510) −2.9275 *** (0.3620) 1.0502 (1.0331) 0.0259 *** (0.0082) 0.0030 (0.0093) 0.1209 (1.0001) 0.3725 (1.5425) 66.7694 * (35.3095) 1.5081 *** (0.5607) −0.3563 (1.4433) −2.9431 *** (0.3611) 1.4524 (1.0232) 0.0272 *** (0.0082) 0.0038 (0.0093) 0.4781 (1.0035) −0.3678 (1.5268) 19.35 19.30 6.22 *** 17.56 6.20 *** 17.70 12.58 12.51 2.22 ** 12.10 2.39 ** 12.11 49.58 51.09 10.20 *** 47.73 10.39 *** 49.14 Notes: *, **, and *** represent significance at 10%, 5%, and 1%, respectively. Standard errors are in parentheses. ROA, ROE, and Tobin’s Q are performance measures. RE is random effects. FE is fixed effects. CEOAge is the CEO age. RET Age is the retirement age of CEOs. Ln(BS) is the natural logarithm of the board size. Ln(Assets) is the natural logarithm of firm assets. Ln(FirmAge) is the natural logarithm of firm age. VOL is volatility. RiskRet (risk retention) is net premiums written over gross premiums written, and DeInMarket (depth of insurance market) is defined as gross written premiums divided by total GDP. Ln(gross written) is the natural logarithm of gross premiums written, measured by the sum of both direct premiums written and assumed premiums written, before deducting the ceded reinsurance. Economies 2023, 11, 21 24 of 41 Table 8 shows significantly negative effects for CEO turnover on ROA and ROE under the RE model only. CEO turnover has a significantly negative effect on ROA (−2.3196) and ROE (−9.0104) at 5% significance in the RE model, while the effect on ROA (−1.8652) decreases in significance to 10% in the FE model. This means that, in the year the CEO is replaced, the firm experiences poor performance, which could be attributed to the new CEO’s lack of knowledge in the first year. This result is consistent with that of Conyon and Florou (2002) and Al-Faryan (2017), who found that CEO changes convey a negative signal to investors and stock market participants. This results in negative equity returns for firms that are already underperforming relative to all listed firms. Other critical criteria that should be examined include whether the CEO is approaching retirement age, whether the CEO’s departure is voluntary, and whether a new CEO has been already identified. Table 7 shows that the retirement age of the CEO has a strong impact on insurance firm performance. These results support Hypothesis 7. Table 8. The effect of CEO turnover on insurance performance. ROA ROE Tobin’s Q RE FE RE FE RE FE Constant CEO Turnover Ln(BS) Ln(Assets) Ln(cid:0)FirmAge(cid:1) VOL RiskRet DeInMarket Ln(gross written) Wald Chi2 F Statistic R2 (%) 0.4078 (65.4366) −2.3196 ** (1.1207) 3.0732 (3.3161) 1.9658 *** (0.7158) 6.4874 *** (1.7346) 0.0306 (0.0249) 0.0172 (0.0271) −0.6451 (3.2121) −2.5134 (2.6886) 56.62 *** 21.54 73.4899 (118.4352) −1.8652 * (1.1298) 7.1657 (4.8010) 3.5251 *** (1.2118) 10.4865 *** (3.4332) 0.0619 ** (0.0276) −0.0018 (0.0311) 1.3658 (3.3469) −7.5518 (5.1186) 6.64 *** 19.05 −117.7657 (191.8118) −9.0104 ** (4.2328) 2.1923 (9.0368) 7.9865 *** (1.7723) 0.9294 (4.4112) 0.0276 (0.0885) 0.0121 (0.0864) −1.7487 (12.5175) −2.1348 (7.9806) 33.00 *** 14.51 44.1000 (472.6367) −6.8692 (4.5089) 6.0125 (19.1594) 19.4013 *** (4.8357) 10.6595 (13.7006) 0.0864 (0.1102) 0.0023 (0.1243) 3.1437 (13.3565) −19.5263 (20.4268) 2.54 ** 13.16 23.0344 (18.9582) −0.3170 (0.3435) −0.8609 (0.9599) −2.1519 *** (0.2032) 0.5375 (0.4930) 0.0254 *** (0.0075) 0.0033 (0.0081) −0.3433 (0.9868) 0.9846 (0.7792) 135.91 *** 49.72 63.0219 * (35.9286) −0.3390 (0.3428) −0.8823 (1.4564) −2.9052 *** (0.3676) 1.3641 (1.0415) 0.0275 *** (0.0084) 0.0045 (0.0094) 0.1632 (1.0153) −0.1653 (1.5528) 9.27 *** 48.40 Notes: *, **, and *** represent significance at 10%, 5%, and 1%, respectively. Standard errors are in parentheses. ROA, ROE, and Tobin’s Q are performance measures. RE is random effects. FE is fixed effects. CEO Turnover is a dependent dummy variable that takes 1 if the CEO changed, and 0 otherwise. Ln(BS) is the natural logarithm of the board size. Ln(Assets) is the natural logarithm of firm assets. Ln(cid:0)FirmAge(cid:1) is the natural logarithm of firm age. VOL is volatility. RiskRet (risk retention) is net premiums written over gross premiums written, and DeInMarket (depth of insurance market) is defined as gross written premiums divided by total GDP. Ln(gross written) is the natural logarithm of gross premiums written, measured by the sum of both direct premiums written and assumed premiums written, before deducting the ceded reinsurance. Looking at the results of Model 1, Table 9 shows that the director incentives and CEO and top executive pay have a significantly positive relationship with Tobin’s Q in the RE model (3.75e-07 and 1.37e-07, respectively) at 5% significance and the CEO and top executive pay in the FE model (1.07e-07), with decreasing significance at 10%. These results support Hypotheses 8 and 9. Our findings also align with those of Al-Faryan (2021), who found that managerial pay has a positive and significant effect on Tobin’s Q. Economies 2023, 11, 21 25 of 41 Table 9. The effect of managerial pay on insurance performance. ROA ROE Tobin’s Q RE FE RE FE RE FE Constant D-INC CEO-Ex-Pay Above mean D-INC Above mean CEO-Ex-Pay CEO-Shares (%) CEO-Ten Ln(BS) Ln(Assets) Ln(cid:0)FirmAge(cid:1) VOL RiskRet DeInMarket Ln(gross written) Wald Chi2 F Statistic R2 (%) Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 −0.9302 (70.2888) −5.97e−08 (5.08e−07) −7.32e−08 (1.86e−07) 0.0587 (0.1161) 0.1132 (0.3387) 3.5406 (3.5847) 2.1615 *** (0.8060) 6.7396 *** (1.8949) 0.0293 (0.0254) 0.0156 (0.0282) −0.4018 (3.2736) −2.7008 (2.8524) 51.44 *** −12.3938 (69.8774) 0.7280 (1.2329) −2.3338 * (1.2980) 0.0572 (0.1131) 0.1833 (0.3304) 2.9395 (3.5002) 2.3894 *** (0.8004) 6.3939 *** (1.8986) 0.0245 (0.0252) 0.0183 (0.0279) −0.7778 (3.2316) −2.3186 (2.8453) 55.79 *** 74.2481 (120.4338) −1.93e−07 (5.41e−07) −1.65e−07 (2.05e−07) 0.0440 (0.1195) −0.1055 (0.3626) 8.0985 (4.9747) 3.6980 *** (1.2499) 11.4721 *** (3.6472) 0.0577 ** (0.0281) 0.0004 (0.0320) 1.3615 (3.4154) −7.8392 (5.2223) 64.0658 (119.1055) 0.3208 (1.3726) −2.8441 ** (1.4166) 0.0413 (0.1167) −0.0459 (0.3534) 8.0019* (4.8587) 3.8714 *** (1.2415) 11.0515 *** (3.5675) 0.0501 * (0.0282) 0.0062 (0.0317) 0.9960 (3.3867) −7.5226 (5.1547) −106.2319 (211.9266) −1.24e−06 (1.78e−06) 3.60e−07 (6.48e−07) 0.2446 (0.4302) 0.3250 (1.1693) 2.3566 (10.2264) 8.2415 *** (2.1975) 0.8064 (4.9248) 0.0243 (0.0907) 0.0062 (0.0924) −0.3084 (12.6859) −3.0357 (8.5717) 26.23 *** −115.4812 (208.8970) 0.9535 (4.1159) −1.1613 (4.5487) 0.3216 (0.4194) 0.5113 (1.1480) 0.4885 (10.0479) 8.4487 *** (2.1427) 0.1848 (4.9017) 0.0150 (0.0904) 0.0082 (0.0918) −0.4782 (12.6820) −2.5968 (8.5204) 26.36 *** 11.2629 (478.7733) −2.04e−06 (2.15e−06) 2.64e−07 (8.13e−07) 0.1722 (0.4749) −1.1042 (1.4413) 7.3184 (19.7764) 19.8919 *** (4.9688) 12.6826 (14.4990) 0.0847 (0.1115) −0.0058 (0.1272) 3.3509 (13.5775) −18.6782 (20.7608) 33.7276 (478.8131) −4.5717 (5.5180) −1.3268 (5.6950) 0.2398 (0.4691) −0.9594 (1.4208) 8.3767 (19.5324) 20.2076 *** (4.9908) 14.4660 (14.3417) 0.0848 (0.1133) 0.0013 (0.1276) 4.1902 (13.6148) −20.0951 (20.7223) 36.9417 ** (18.6380) 3.75e−07 ** (1.48e−07) 1.37e−07 ** (5.41e−08) −0.0076 (0.0348) −0.0082 (0.0983) −1.6436* (0.9380) −2.3514 *** (0.2052) 0.3028 (0.4688) 0.0274 *** (0.0074) −0.0009 (0.0080) 0.0771 (0.9970) 0.5933 (0.7521) 164.53 *** 27.8411 (18.2432) 1.0920 *** (0.3532) 0.2337 (0.3825) 0.0004 (0.0346) 0.0413 (0.0971) −1.2638 (0.9024) −2.1782 *** (0.1967) 0.2588 (0.4546) 0.0245 *** (0.0075) 0.0001 (0.0079) −0.2958 (1.0166) 0.8325 (0.7411) 170.05 *** 59.1962 * (35.6815) 2.61e−07 (1.60e−07) 1.07e−07 * (6.06e−08) −0.0176 (0.0354) 0.0859 (0.1074) −1.5823 (1.4739) −3.0796 *** (0.3703) 0.5365 (1.0806) 0.0276 *** (0.0083) 0.0012 (0.0095) 0.2574 (1.0119) 0.2136 (1.5472) 61.3998 * (35.9382) 0.5355 (0.4142) −0.3697 (0.4275) −0.0151 (0.0352) 0.1586 (0.1066) −0.9798 (1.4660) −2.9776 *** (0.3746) 0.9458 (1.0764) 0.0243 *** (0.0085) 0.0049 (0.0096) 0.0112 (1.0219) −0.0212 (1.5553) 19.51 20.21 4.55 *** 17.01 4.95 *** 17.71 13.34 13.10 1.78 * 12.14 1.75 * 11.53 53.41 53.61 7.45 *** 51.56 7.08 *** 49.57 Notes: *, **, and *** represent significance at 10%, 5%, and 1%, respectively. Standard errors are in parentheses. ROA, ROE, and Tobin’s Q are performance measures. RE is random effects. FE is fixed effects. D-INC is director incentives. CEO-Ex-Pay is the CEO and top executive pay. Above mean D-INC and Above mean CEO-Ex-Pay are dependent dummy variables of director incentives and CEO and top executive pay, respectively, that take 1 if above the mean. CEO-Shares(%) is the percentage of CEO-owned shares. CEO-Ten is CEO tenure. Ln(BS) is the natural logarithm of the board size. Ln(Assets) is the natural logarithm of firm assets. Ln(cid:0)FirmAge(cid:1) is the natural logarithm of firm age. VOL is volatility. RiskRet (risk retention) is net premiums written over gross premiums written, and DeInMarket (depth of insurance market) is defined as gross written premiums divided by total GDP. Ln(gross written) is the natural logarithm of gross premiums written, measured by the sum of both direct premiums written and assumed premiums written, before deducting the ceded reinsurance. Economies 2023, 11, 21 26 of 41 Model 2 uses the above-the-mean director incentives and above-the-mean CEO and top-executive-pay dummy variables equal to 1 if above the mean, and 0 otherwise (see Table 1 regarding the mean). We find that above-the-mean CEO and top executive pay has a significantly negative effect on ROA in the RE (−2.3338) and FE (−2.8441) models at 10% and 5% significance, respectively. Thus, for above the mean, Hypothesis 9 is not supported. However, above-the-mean director incentives have a significantly positive effect on Tobin’s Q in the RE (1.0920) model at 1% significance. Thus, Hypothesis 8 is supported by Model 1. In sum, an increase in director incentives and CEO and top executive pay and above- the-mean director incentives leads to better performance in terms of Tobin’s Q. However, director incentives, CEO and top executive pay, and above-the-mean director incentives do not affect the accounting performance indicators ROA and ROE. As accounting perfor- mance indicators should capture the effects of above-the-mean CEO and top executive pay, the results imply that insurance firms that pay their CEO and top executives above the mean have worse firm performance (ROA) than firms that pay less than the market average. In conclusion, the director incentives have a positive effect on insurance performance, as do CEO and top executives pay up to the mean, after which the performance effect turns negative. One reason insurance firms may lose value is that they are overpaying their CEOs and executives, which results in poor performance as these CEOs and top executives seek self-benefits instead of shareholder benefits. Other reasons can be noted in the standard deviation of CEO and top executive pay. Larger pay indicates that CEO and top executive data are more dispersed, while director incentives show that more director incentive data are clustered about the mean, as seen in Table 1. 5.1. Probit Model Average Marginal Effect Results We also use the probit model average marginal effect, as explained in the methodology section. Table 10 shows the results of the probit model we use to capture the likelihood of CEO turnover due to poor performance. The estimated regressions show a significant and negative relationship between the two accounting operating performance indicators, ROA (−0.0320) and ROE (−0.0074) and CEO turnover at 5% significance and vice versa (Table 8). These results support Hypothesis 7. This means that a firm with better performance is less likely to change CEOs than a firm with bad performance is. Thus, firms with poor performance have a higher probability of changing CEOs. CEO turnover is a critical variable that captures when CEOs are replaced due to poor performance. Many studies have shown an inverse effect between firm performance and CEO turnover (Conyon and Florou 2002; Coughlan and Schmidt 1985; Jenter and Kanaan 2015). Volpin (2002) and Gibson (2003) argue that there is a greater likelihood of CEO turnover at firms with strong governance systems and that the likelihood of CEO turnover increases with poor firm performance. Table 10. The effect of insurance performance on CEO turnover. Performance Ln(BS) Ln(Assets) Ln(FirmAge) VOL RiskRet ROA −0.0320 ** (0.0129) 0.4790 (0.5098) −0.1584 (0.1007) 0.3556 (0.2617) 0.0062 (0.0051) −0.0024 (0.0048) ROE −0.0074 ** (0.0036) 0.4253 (0.5035) −0.1415 (0.1038) 0.1759 (0.2541) 0.0061 (0.0052) −0.0032 (0.0047) Tobin’s Q −0.0318 (0.0432) 0.3715 (0.4948) −0.2746 ** (0.1317) 0.1735 (0.2595) 0.0069 (0.0056) −0.0029 (0.0047) Economies 2023, 11, 21 27 of 41 Table 10. Cont. DeInMarket Ln(gross written) Likelihood ratio ROA −0.0708 (0.7062) −0.0726 (0.4414) 14.55 * ROE −0.0196 (0.6999) −0.0461 (0.4461) 12.36 Tobin’s Q −0.0479 (0.7011) 0.0470 (0.4601) 8.74 Notes: *, **, and *** represent significance at 10%, 5%, and 1%, respectively. Standard errors are in parentheses. This table presents the average marginal effects from the probit estimation, wherein CEO turnover is the dependent variable. ROA, ROE, and Tobin’s Q are performance measures. Ln(BS) is the natural logarithm of the board size. Ln(Assets) is the natural logarithm of firm assets. Ln(cid:0)FirmAge(cid:1) is the natural logarithm of firm age. VOL is volatility. RiskRet (risk retention) is net premiums written over gross premiums written, and DeInMarket (depth of insurance market) is defined as gross written premiums divided by total GDP. Ln(gross written) is the natural logarithm of gross premiums written, measured by the sum of both direct premiums written and assumed premiums written, before deducting the ceded reinsurance. Table 11 presents our findings that show a significantly positive relationship between insurance performance and above-the-mean director incentives. We find that Tobin’s Q has a significantly positive effect on above-the-mean director incentives (0.1867) at 5% significance and vice versa (Table 9). Once again, these findings support Hypothesis 8. Board members can receive incentives of up to 10% of a company’s net profits based on good management performance. This can motivate strong firm performance. However, we do not find a significant relationship between above-the-mean CEO and top executive pay and above-the-mean director incentives. Table 11. The effect of insurance performance on above-the-mean director incentives. Performance Above mean CEO-Ex-Pay CEO-Shares (%) CEO-Ten Ln(BS) Ln(Assets) Ln(FirmAge) VOL RiskRet DeInMarket Ln(gross written) Likelihood ratio ROA 0.0141 (0.0197) −0.0138 (0.3593) 0.0067 (0.0753) 0.0589 (0.0954) 1.9720 * (1.1154) 0.3151 0.2569 0.9510 (0.6359) 0.0073 (0.0075) 0.0108 (0.0084) 0.3329 (0.9247) 0.5871 (0.9539) 39.87 *** ROE −0.0017 (0.0052) −0.0186 (0.3568) 0.0073 (0.0751) 0.0643 (0.0952) 2.0964 * (1.1001) 0.3652 (0.2598) 1.0234 (0.6282) 0.0074 (0.0074) 0.0101 (0.0083) 0.2307 (0.9222) 0.5823 (0.9546) 39.44 *** Tobin’s Q 0.1867 ** (0.0745) −0.0571 (0.3630) 0.0076 (0.0776) 0.0369 (0.0952) 2.2934 ** (1.0835) 0.8320 ** (0.3251) 0.8383 (0.6106) 0.0027 (0.0077) 0.0114 (0.0084) 0.2778 (0.9197) 0.6802 (0.9410) 46.13 *** Notes: *, **, and *** represent significance at 10%, 5%, and 1%, respectively. Standard errors are in parentheses. This table presents the average marginal effects from the probit estimation, wherein above-the-mean director incentives is the dependent dummy variable that takes 1 if above the mean, and 0 otherwise. ROA, ROE, and Tobin’s Q are performance measures, Above mean CEO-Ex-Pay is a dependent dummy variable that takes 1 if above the mean, and 0 otherwise. CEO-Shares (%) is the percentage of CEO-owned shares. CEO-Ten is CEO tenure. Ln(BS) is the natural logarithm of the board size. Ln(Assets) is the natural logarithm of firm assets. Ln(cid:0)FirmAge(cid:1) is the natural logarithm of firm age. VOL is volatility. RiskRet (risk retention) is net premiums written over gross premiums written, and DeInMarket (depth of insurance market) is defined as gross written premiums divided by total GDP. Ln(gross written) is the natural logarithm of gross premiums written, measured by the sum of both direct premiums written and assumed premiums written, before deducting the ceded reinsurance. Economies 2023, 11, 21 28 of 41 Table 12 presents findings that show a significantly negative relationship between poor insurance performance and above-the-mean CEO and top executive pay. We find that ROA has a significantly negative effect on above-the-mean CEO and top executive pay (−0.0359) at 5% significance and vice versa (Table 9), thereby rejecting Hypothesis 9 in terms of above the mean. Poor performance at insurance firms increases when CEOs and top executives are paid above the mean. One explanation for increases in CEO and top executive pay may be due to high insurance administrative costs in addition to the increasing importance of accounting operating performance indicators such as ROA, along with increasing competition, and the desire to raise insurance CEO and top executive pay to be on par with or even higher than that of banks. This issue requires control of technical performance and job attractiveness, along with the monitoring of wages and petty cash, compensation, and competency development. However, CEOs and top executives may exploit internal information to achieve personal gain. There may be a lack of accountability for high managerial pay, even with poor performance. Thus, insurance companies will lose capital due to weak management, weak governance, and gaps in the application of corporate governance regulation and the corporate governance code. Table 12. The effect of insurance performance on above the mean CEO and top executive pay. Performance Above mean D-INC CEO-Shares (%) CEO-Ten Ln(BS) Ln(Assets) Ln(cid:0)FirmAge(cid:1) VOL RiskRet DeInMarket Ln(gross written) Likelihood ratio ROA −0.0359 ** (0.0183) 0.0474 (0.3315) 0.0322 (0.1236) 0.2898 *** (0.1051) 0.0465 (0.9475) 0.7354 *** (0.2251) −0.0063 (0.5243) −0.0253 (0.0169) 0.0078 (0.0080) −1.4224 (0.9562) 0.9580 (0.8161) 52.83 *** ROE −0.0036 (0.0049) −0.0416 (0.3219) 0.0330 0.1181 0.2604 *** (0.1006) −0.0919 (0.9042) 0.6837 *** (0.2187) −0.1993 (0.4854) −0.0244 (0.0166) 0.0066 (0.0077) −1.3156 (0.9359) 1.1193 (0.7734) 49.39 *** Tobin’s Q 0.0387 (0.0755) −0.0764 (0.3254) 0.0322 (0.1166) 0.2540 ** (0.0990) −0.0749 (0.8886) 0.7170 *** (0.2553) −0.2002 (0.4738) −0.0272 0.0169 0.0065 0.0076 −1.2592 0.9267 1.0534 0.7681 49.12 *** Notes: *, **, and *** represent significance at 10%, 5%, and 1%, respectively. Standard errors are in parentheses. This table presents the average marginal effects from the probit estimation, wherein above-the-mean CEO and top executive pay is the dependent dummy variable of the CEO and top executive pay, which takes 1 if above the mean, and 0 otherwise. ROA, ROE, and Tobin’s Q are performance measures, Above mean D-INC is a dependent dummy variable of director incentives, which takes 1 if above the mean, and 0 otherwise. CEO-Shares (%) is the percentage of CEO-owned shares. CEO-Ten is CEO tenure. Ln(BS) is the natural logarithm of the board size. Ln(Assets) is the natural logarithm of firm assets. Ln(cid:0)FirmAge(cid:1) is the natural logarithm of firm age. VOL is volatility. RiskRet (risk retention) is net premiums written over gross premiums written, and DeInMarket (depth of insurance market) is defined as gross written premiums divided by total GDP. Ln(gross written) is the natural logarithm of gross premiums written, measured by the sum of both direct premiums written and assumed premiums written, before deducting the ceded reinsurance. The results shown in Tables 8–12 reflect the fact that Saudi insurance companies have gone bankrupt due to a lack of compliance with regulatory requirements, high administrative costs, and weak management (Alokla et al. 2022). A policy of inheritance has been widespread across high positions in these firms to keep decision-making in the hands of a few and to retain high salaries, even if those holding these high positions lack Economies 2023, 11, 21 29 of 41 expertise. This has led to failure in the management of Saudi insurance companies; these are still suffering from their losses. In one case, the reason for the firm’s loss was because it appointed an ineffective CEO who had been dismissed from another company. CEO turnover such as this can lead to the dismissal of current employees who are then replaced by favourites, with these positions redistributed. The redistribution of power among employees in these companies, for example, through the appointment of a general manager and an executive director or someone in another leadership position, occurs through bias, with the focus on the development of certain employees over others. In these cases, the CEO focuses on one person by offering promotion, workshops, courses, and certificates to prepare them for an administrative position. Poor management or functional exploitation through resignation or dismissal occurs without anyone held accountable or even demanding monies that may have been taken without right. In such cases, the biggest losers are small shareholders who eagerly await the return on their shares and poor employees who have no union to protect their rights. In these firms, there are no profits to distribute or revenue increases, and there may be collusion in administrative affairs. This scenario may be repeated with subsequent CEOs and top executives, and the result is likely the closure of the company. When this occurs, the shares evaporate, and large shareholders escape by selling their shares. The management of the firm’s investments is assigned by the CEO and the financial or investment portfolio official, and the individuals in these roles may not have sufficient experience to accomplish this effectively. 5.2. Shariah Compliance and Life Insurance Firm Results We examined seven publicly listed Shariah-compliant firms in our timeframe, provid- ing 36 observations to investigate the relationships between Shariah-compliant insurance firm performance and board composition and audit committee size. All of the firms follow the regulations of corporate governance, except for one, with one observation of the audit committee and board composition. Because the firms are compliant with Shariah and corporate governance regulations, the exogenous shock of regulation in terms of audit com- mittee and board composition will not have an impact, as they already complied. Although the sample is small, at a rate of 97%, the results imply that Shariah-compliant firms are more committed to implementing corporate governance regulations. These findings are consistent with those in a study by Ezzine (2018). Bhatti and Bhatti (2010) argue that Islamic corporate governance has the same goals as conventional corporate governance but operates within the framework of Islam’s religiously based moral precepts. Hasan (2009) contrasts Western and Islamic corporate governance. To begin, Islamic corporate governance is governed by Islamic law, or Shariah, which encompasses all facets of life. In addition, the impact of Shariah law and certain Islamic economic and financial concepts as they relate to corporate activities and policies should be considered. This means that, in contrast to other systems of corporate governance, Islamic corporate governance’s ultimate purpose is Shariah compliance. Thus, Shariah- compliant insurance firms are concerned with adhering to Islamic Shariah provisions in all transactions, operations, and investment activities. In terms of corporate governance, all factors, such as board composition and audit committee size, must be compatible with Shariah compliance. In this context, we conduct an analysis to determine the governance impact of board composition and audit committee size on Shariah-compliant insurance firm performance in Saudi Arabia. Table 13 shows the results of Model 1. Of the listed Saudi insurance firms that are Shariah compliant, for board composition, the performance measures of ROA in the RE (−0.1305) and FE (−0.3452) models, ROE in the RE (−0.2492) model but not the FE model, and Tobin’s Q (RE and FE) are significantly negative, at 1% significance for ROA and ROE (RE). This result aligns with the results in Table 3, which indicate that board composition is significantly negatively correlated with performance for all insurance firms. Thus, these findings do not support Hypothesis 1. Economies 2023, 11, 21 30 of 41 Table 13. The effect of board composition and audit committee size on performance in Shariah-compliant insurance firms. ROA ROE Tobin’s Q RE FE RE FE RE FE Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Constant %IB * Shariah-compliant SAC * Shariah-compliant Ln(BS) Ln(Assets) Ln(cid:0)FirmAge(cid:1) VOL RiskRet DeInMarket Ln(gross written) Wald Chi2 F Statistic R2 (%) −24.1626 (64.7476) −0.1305 *** (0.0471) 2.7418 (3.2585) 2.0529 *** (0.6991) 5.8467 *** (1.7087) 0.0205 (0.0247) 0.0237 (0.0269) −0.9860 (3.1998) −1.4551 (2.6647) 60.76 *** −15.1048 (64.9679) −1.5274 ** (0.6468) 2.7593 (3.2818) 2.0242 *** (0.7048) 6.0032 *** (1.7198) 0.0223 (0.0248) 0.02333 (0.0270) −0.8470 (3.2122) −1.8358 (2.6730) 58.16 *** 77.2051 (117.3777) −0.3452 *** (0.1427) 7.6204 (4.7611) 3.3900 *** (1.2012) 11.1440 *** (3.4138) 0.0595 ** (0.0273) −0.0056 (0.0309) 1.3130 (3.3166) −7.5831 (5.0718) 98.1568 (120.2730) −3.9037 (2.7331) 6.9556 (4.8131) 3.2670 *** (1.2236) 11.2824 *** (3.4884) 0.0610 ** (0.0276) −0.0018 (0.0312) 1.5695 (3.3558) −8.3519 (5.1723) −186.7492 (195.8050) −0.2492 *** (0.1176) 1.8371 (9.2344) 8.6788 *** (1.8067) −0.8654 (4.5685) −0.0035 (0.0885) 0.0417 (0.0881) −2.1544 (12.4531) 0.3023 (8.1421) 31.95 *** −169.6112 (196.2174) −2.9509 * (1.5868) 1.9703 (9.3146) 8.6608 *** (1.8246) −0.5353 (4.5995) 0.0011 (0.0886) 0.0390 (0.0886) −1.9154 (12.4659) −0.4524 (8.1546) 30.54 *** 47.7447 (473.3621) −0.7841 (0.5754) 7.0495 (19.2004) 19.0395 *** (4.8441) 12.1101 (13.7670) 0.0766 (0.1101) −0.0088 (0.1245) 3.0688 (13.3753) −19.3657 (20.4535) 90.2536 (481.5170) −8.1604 (10.9422) 5.5780 (19.2693) 18.7999 *** (4.8989) 12.2748 (13.9657) 0.0798 (0.1107) −0.0004 (0.1250) 3.6203 (13.4351) −20.9374 (20.7074) 24.5880 (19.1016) 0.0110 (0.0135) −0.8860 (0.9595) −2.1390 *** (0.2026) 0.5805 (0.4965) 0.0253*** (0.0075) 0.0029 (0.0081) −0.3116 (0.9880) 0.9013 (0.7864) 135.70 *** 24.9281 (18.9105) 0.2700 (0.1806) −0.8851 (0.9532) −2.1316 *** (0.2010) 0.6064 (0.4919) 0.0255 *** (0.0075) 0.0023 (0.0081) −0.3067 (0.9851) 0.8773 (0.7783) 139.00 *** 63.4746 * (35.8875) −0.0519 (0.0436) −0.8138 (1.4557) −2.9267 *** (0.3673) 1.4620 (1.0437) 0.0270 *** (0.0083) 0.0039 (0.0094) 0.1562 (1.0140) −0.1649 (1.5507) 58.7432 (36.4777) 0.5092 (0.8289) −0.8539 (1.4598) −2.8836 *** (0.3711) 1.2508 (1.0580) 0.0267 *** (0.0084) 0.0040 (0.0095) 0.1465 (1.0178) −0.0096 (1.5687) 25.46 24.52 7.14 *** 22.54 6.53 *** 22.07 14.62 14.20 2.47 ** 13.53 2.29 ** 13.31 49.96 50.30 9.35 *** 42.52 9.16 *** 49.07 Notes: *, **, and *** represent significance at 10%, 5%, and 1%, respectively. Standard errors are in parentheses. ROA, ROE, and Tobin’s Q are performance measures. RE is random effects. FE is fixed effects. %IB is the board composition. SAC is the size of the audit committee. Shariah-compliant is a dummy variable that equals 1 if the firm is Shariah compliant in all of its transactions, insurance operations, and investment activities, and 0 otherwise. %IB*Shariah-compliant is board composition multiplied by the Shariah-compliant variable. SAC* Shariah-compliant is the size of the audit committee multiplied by Shariah-compliant variable to measure Islamic corporate governance or the corporate governance of Shariah-compliant insurance firms. Ln(BS) is the natural logarithm of the board size. Ln(Assets) is the natural logarithm of firm assets. Ln(cid:0)FirmAge(cid:1) is the natural logarithm of firm age. VOL is volatility. RiskRet is (risk retention) net premiums written over gross premiums written, and DeInMarket (depth of insurance market) is defined as gross written premiums divided by total GDP. Ln(gross written) is the natural logarithm of gross premiums written, measured by the sum of both direct premiums written and assumed premiums written, before deducting the ceded reinsurance. Economies 2023, 11, 21 31 of 41 In Model 2, we find that Shariah compliance and the size of the audit committee (SAC) is significantly negatively correlated with insurance firm performance. The measures of the RE under the ROA (−1.5274) and the ROE (−2.9509) models are significant at 5% and 10%, except under Tobin’s Q models. Thus, Hypothesis 4 is not supported, similar to the results in Table 5. It is clear that firms providing Islamic insurance are similar to all other Saudi insurance firms, which operate for the same reasons as those mentioned in Tables 3 and 4 with a few exceptions. However, independent boards and audit committees have little experience with Is- lamic finance and insurance. The current legislative environment is inconsistent with the logic underlying Islamic insurance companies. This is because of the rapid rise to prominence in the stock market of Islamic insurance. Islamic insurance is based on the principle that those who are insured own the insurance company; those in charge of its management do not own it but instead receive fixed salaries or a percentage of the profits earned by the company in exchange for managing it. The excess of the company’s assets is distributed to the insured. For robustness, we also examine 16 firms providing life insurance between 2008 and 2014, providing 90 observations to investigate the relationship between these firms in terms of board composition and audit committee size on firm performance. All of the firms, with the exception of two, follow corporate governance regulations based on one observation of board composition and two observations of audit committee size. Because life insurance firms already follow corporate governance regulations, the exogenous shock of the regulation of board composition and audit committee size will not have an impact. As before, based on this sample, it appears that firms that provide life insurance are more committed to implementing corporate governance regulations, at a rate of 97.8% for the audit committee and 98.9% for board composition. These findings are consistent with Table 13. Table 14 indicates that life insurance firms are distinguished by board composition and audit committee size, which have distinct characteristics. First, unlike many non-life insurance companies, the independent board of directors and audit committee comprise highly competent financial professionals. Second, the time and effort spent monitoring and disciplining a corporation may benefit the whole customer base. Finally, since, in general, life insurance is a long-term investment, individual policyholders may purchase it only once. As a result, it is necessary to maintain the customer base at all times as this has an impact on firm performance. In this context, we examine whether listed Saudi firms that provide life insurance have similar results to all listed insurance firms. We find that board composition at these life insurance firms is significantly negative for the RE Model 1 under the ROA (−0.0570) and under the ROE (−0.1469), and the FE Model 1 under Tobin’s Q (−0.0312), with significance at 10%. Thus, Hypothesis 1 is not supported, similar to our results in Tables 3 and 13. In Model 2, we find that the SAC is significantly negative with life insurance firm performance at 10% and 5% significance by measures of the RE in the ROA (−0.8617) and the ROE (−2.4960) models but not in the FE models. Thus, Hypothesis 4 is not supported, a result similar to that in Tables 5 and 13. However, the SAC has a significantly positive effect on Tobin’s Q in the RE (0.3137) model at 5% significance, which supports Hypothesis 4. Thus, the findings are consistent with Tobin’s Q with Asize(4) and under the RE (1.5592), as they relate positively with Tobin’s Q at 10% significance, as shown in Table 5. In sum, board composition and audit committee size in life insurance firms reflect similar patterns as all insurance firms, wherein medical and motor insurance make up 84% of the total Saudi insurance market, as these are mandatory. Notably, life insurance is almost non-existent in Saudi Arabia (due to cultural and religious reasons, among other factors, such as the lack of awareness). However, with the implementation of the Saudi Vision 2030, the market will become officially regulated and more developed. As this development occurs, the spread of conventional and takaful life insurance products is expected to rise.1 Economies 2023, 11, 21 32 of 41 Table 14. The effect of board composition and audit committee size on performance of firms providing life insurance. ROA ROE Tobin’s Q RE FE RE FE RE FE Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Constant %IB* Life insurance SAC* Life insurance Ln(BS) Ln(Assets) Ln(cid:0)FirmAge(cid:1) VOL RiskRet DeInMarket Ln(gross written) Wald Chi2 F Statistic R2 (%) −13.5532 (66.6831) −0.0570 * (0.0318) 3.1701 (3.3582) 2.3556 *** (0.7427) 6.7912 *** (1.7699) 0.0285 (0.0249) 0.0135 (0.0273) −0.7381 (3.2096) −2.2330 (2.7315) 55.37 *** −8.5681 (66.4946) −0.8617 * (0.5111) 3.0200 (3.3588) 2.3609 *** (0.7457) 6.7423 *** (1.7694) 0.0266 (0.0249) 0.0088 (0.0276) −0.7244 (3.2125) −2.4142 (2.7278) 54.94 *** 64.7072 (119.4495) −0.0387 (0.0551) 7.4156 (4.8457) 3.5017 *** (1.2198) 10.6575 *** (3.4678) 0.0610 ** (0.0279) −0.0044 (0.0314) 1.2974 (3.3721) −7.1715 (5.1608) 78.0837 (120.7469) 0.5321 (1.2353) 7.0817 (4.8416) 3.5013 *** (1.2212) 10.5179 *** (3.4620) 0.0586 ** (0.0278) 0.0006 (0.0328) 1.5553 (3.3920) −7.7811 (5.2366) −181.8415 (196.9191) −0.1469 * (0.0809) 1.7650 (9.2895) 9.5891 *** (1.8949) 0.9460 (4.5519) 0.0115 (0.0884) −0.0018 (0.0885) −1.9786 (12.4794) −0.6160 (8.1395) 30.47 *** −169.1073 (195.8273) −2.4960 ** (1.2322) 1.7598 (9.3061) 9.7646 *** (1.9066) 1.1247 (4.5659) 0.0083 (0.0883) −0.0151 (0.0893) −1.9069 (12.4428) −1.2495 (8.1079) 31.22 *** 53.2379 (476.1041) 0.1556 (0.2195) 5.0301 (19.3140) 19.1962 *** (4.8618) 9.7379 (13.8219) 0.0673 (0.1112) −0.0003 (0.1250) 3.6639 (13.4406) −19.7904 (20.5700) 85.8705 (480.7785) 3.6147 (4.9185) 5.4324 (19.2777) 19.3601 *** (4.8624) 10.9670 (13.7849) 0.0727 (0.1106) 0.0236 (0.1305) 4.3323 (13.5060) −21.5943 (20.8508) 23.9459 (18.6537) 0.0119 (0.0085) −0.9037 (0.9345) −2.1705 *** (0.2006) 0.4644 (0.4765) 0.0247 *** (0.0075) 0.0044 (0.0080) −0.3319 (0.9956) 0.9538 (0.7642) 143.95 *** 23.9584 (18.3406) 0.3137 ** (0.1315) −0.8892 (0.9219) −2.2180 *** (0.1984) 0.4348 (0.4695) 0.0252 *** (0.0074) 0.0067 (0.0080) −0.3182 (0.9884) 0.9771 (0.7528) 151.43 *** 58.0626 (35.7207) −0.0312 * (0.0165) −0.6826 (1.4491) −2.8998 *** (0.3648) 1.5209 (1.0370) 0.0286 *** (0.0083) 0.0034 (0.0094) 0.0881 (1.0084) 0.0388 (1.5433) 55.9978 (36.3181) −0.4266 (0.3715) −0.8120 (1.4562) −2.9243 *** (0.3673) 1.3091 (1.0413) 0.0273 *** (0.0084) 0.0009 (0.0099) 0.0425 (1.0202) 0.1800 (1.5751) 22.28 22.70 6.27 *** 19.32 6.23 *** 15.81 14.16 14.52 2.29 ** 10.67 2.29 ** 9.65 50.95 52.18 9.74 *** 42.66 9.33 *** 42.70 Notes: *, **, and *** represent significance at 10%, 5%, and 1%, respectively. Standard errors are in parentheses. ROA, ROE, and Tobin’s Q are performance measures. RE is random effects. FE is fixed effects. %IB is board composition. SAC is the size of the audit committee. Life insurance is a dummy variable that equals 1 if the firm has life insurance contracts, and 0 otherwise. %IB* Life insurance is board composition multiplied by the life insurance firm variable. SAC* Life insurance is the size of the audit committee multiplied by the life insurance firm variable. Ln(BS) is the natural logarithm of the board size. Ln(Assets) is the natural logarithm of firm assets. Ln(cid:0)FirmAge(cid:1) is the natural logarithm of firm age. VOL is volatility. RiskRet is (risk retention) net premiums written over gross premiums written, and DeInMarket (depth of insurance market) is defined as gross written premiums divided by total GDP. Ln(gross written) is the natural logarithm of gross premiums written, measured by the sum of both direct premiums written and assumed premiums written, before deducting the ceded reinsurance. Economies 2023, 11, 21 33 of 41 5.3. The Impact of Saudi Inflation and the Consumer Price Index on Insurance Prior studies have noted the importance of the great interplay between fiscal and monetary policy and suggest that both policies behave differently to the economic cycle. However, the impact of further fiscal consolidation efforts on domestic demand was also observed in the KSA and in some countries of the Eurozone (Cecrdlova (2021) and Albassam (2021)). Correspondingly, it is worth mentioning that the discrepancy between fiscal and monetary policy plays a significant role in developing and maintaining the economics. These developments can not be achieved without reforming the existing economic and financial policies (Cecrdlova (2021) and Albassam (2021)). Saudi Arabia’s purchasing power remained strong over our seven-year timeframe. From 2008 to 2014, the purchasing power of the currency improved for two reasons. First, inflation decreased from 9.87% in 2008 to 2.24% in 2014 (see Appendix A). Second, the Saudi riyal is pegged to the US dollar at a rate of 3.75 (Al-Faryan 2021), and insurance policies are indexed to the consumer price index. We investigate whether this impacted the results of the inflation rate on insurance performance. Appendix B shows that the inflation rate (%) has a significant negative effect on insurance performance at 1% for ROA and Tobin’s Q, while ROE has a significant negative effect at 5% significance. This indicates that the higher the inflation rate, the lower the insurance company profits. There are two major interpretations of this result. First, inflation has a negative impact on insurance performance because it leads to a decline in the real income of individuals due to the purchasing power of their money. Demand for insurance policies to insure cars, real estate, and so on is directly proportional to individual income and vice versa. The less real income an individual has, the less he/she spends on insurance. Second, inflation will lead to a rise in the cost of the partial compensation settlement, as that cost rises as a result of the price increase. This problem is addressed in property insurance policies by increasing the amount of insurance in the same inflation line by re-evaluating the value of the property insured. The proportionality clause in compensation is also partially applied in the event that the insured sum is less than the market value of the property. While reinsurance agreements address the impact of inflation on the settlement of claims by setting a cost-of-living index clause or stabilisation clause so that the reinsurer does not bear the increase in the value of compensation resulting from inflation, this condition is usually included in excess loss treaties. In particular, with regard to the insurance of civil liability and motor vehicles, the effect of this condition is that, if the compensation is settled at a time when the value of the cash differs significantly from its value on a specific date (usually the date of commencement of the agreement), the division of compensation between the assigned company and the reinsurer takes place on the basis of the ratios that the value of that compensation would have been divided by had it been settled on that particular date on the basis of the terms of the agreement. 6. Conclusions Our principal objective is to investigate whether corporate governance factors have an effect on financial performance in Saudi insurance firms. Specifically, we examine how board composition, director incentives, board independence, monitoring costs, audit committee size, CEO and top executive pay, CEO age, and CEO turnover influence financial performance by looking at 35 insurance firms listed on the Saudi stock market from 2008 to 2014. Based on the literature, we consider the endogeneity of the statistical approach and discover in the RE and FE models that board composition has a significant negative relationship with ROE and Tobin’s Q, except in the RE model. However, audit committee size does not have a significant relationship with performance, using the DID and DID2 approaches for Saudi insurance companies that choose to comply with regulations. The evidence indicates that firm performance is positively affected by information asymmetry, which increases monitoring costs for independent directors and also reduces the positive effect of independent directors on Saudi insurance firm performance. We focus on critical mass and find that a negative relationship remains when the number of independent Economies 2023, 11, 21 34 of 41 directors increases to six or more, despite the decline in performance. In general, boards with three, four, or five independent members do not show a significant relationship with performance. Similarly, an audit committee of less than four members does not affect performance; however, the size does have a positive effect on insurance performance when the committee has four members, but more than four hurts performance. In general, we find that most firms lack expertise in managing audit committee size and board independence. Moreover, the audit committees do not appear to have any relationship with either internal or external auditors. These findings are not significant based on audit committee size and lack of independent results, which do not play an effective part in listed Saudi insurance companies. We examine the effect of the CEO’s age on Saudi insurance performance and find that an older CEO has better performance. To complement this result, we also analyse the effect of the CEO’s retirement age. For CEOs over 60, we find an increase in significance from 10% and 5% to 1% when the CEO is post-retirement age. When CEOs stay in the position, we find a negative relationship between CEO turnover and firm performance and vice versa. This means that firms with poor performance should change their CEOs, as poor CEO performance has a significant negative effect on performance. Our study also examines the relationship between firm performance and managerial pay using four pay variables, namely; director incentives, CEO and top executive pay, above-the-mean director incentives, and above-the-mean CEO and top executive pay. We find three of these four variables to relate positively to financial performance. These are director incentives, CEO and top executive pay, and above-the-mean director incentives, measuring above-the-mean director incentives using Tobin’s Q and vice versa. However, above-the-mean CEO and top executive pay have the opposite effect on ROA and vice versa. The increased age of a CEO means greater experience and an increase in the CEO’s pay. This leads to better firm performance up to when the mean of the CEO’s pay reaches the average CEO pay in the insurance industry. After this, the effect on performance turns negative. Firms allocate 10% of the profits to board directors as a reward for firm performance. This is an attractive incentive and explains the increase in significance from 5% to 1% for above-the-mean director incentives. In addition, we find that Shariah-compliant firms and life insurance firms have similar results to all listed insurance firms in terms of the effect of board independence and audit committee size on performance. The insurance sector in the KSA has been slow to adhere to new procedures and to comply with newly imposed governance regulations that require listed companies to improve transparency in corporate financial reporting (Al-Faryan and Dockery 2021). Our study reveals that 13 of the insurance companies listed in the Saudi market from 2008 to 2014 were not complying with certain corporate governance regulations in their first year due to a lack of experienced board members and executive management. 6.1. Implications We believe that the Saudi corporate governance codes, along with other longer-term development plans, should be considered successful. Particularly, the 2009 changes made to the board and auditing committee size regulations, have had a positive impact on Saudi listed insurance firm performance. Further, insurance company structure influences firm practices and performance. Nonetheless, there is room for improvement and development, as the positive impact of these issues raises some questions. Our study has implications for both domestic and foreign investors, including institu- tional investors, by highlighting the effective functioning of boards of directors and auditing committees. Domestic investors should see better governance as a positive sign—one they can take advantage of when investing in Saudi listed insurance firms at a time when the Saudi economy is looking to diversify and expand in accordance with the country’s Vision 2030 plans. Foreign investors can diversify their portfolios by investing in the KSA, as it has now opened its markets and its economy in line with achieving this strategy. Economies 2023, 11, 21 35 of 41 The results also demonstrate the importance of the role of corporate governance in Saudi Arabia in relation to the issues studied. Access to foreign savings is important for diversifying the Saudi economy, as well as achieving economic growth. However, foreign investors will only be willing to invest in the KSA if they are confident that their investments are being monitored and protected. In the context of insurance firms, we highlight the growing compliance in the industry, reflecting the improvements in Saudi corporate governance practices. Generally, the most interesting aspect of the findings of this study is that it not only evidenced the impact of corporate governance on financial performance but also, provided further insight on the impact by market share and insurance density. Therefore, the gross premia and insurance penetration are also significantly related to economic growth. (Apergis and Poufinas 2020; Balcilar et al. 2018; Alokla and Daynes 2017; Alokla et al. 2022; Gaganis et al. 2019) 6.2. Limitations Our study is not without limitations. First, the main limitation is, in contrast to the banking sector, in the insurance sector in the KSA, there is a severe lack of available data (Alokla et al. 2022; Yousaf and Alokla 2022; Alokla and Daynes 2017). As a result, data are obtained from the CMA and Mubasher. Second, data availability restricts our analysis to seven years, as there are no data from our unique database available before that time. Financial costs are another constraint that exacerbates our data problem. Future studies should explore whether our results hold over longer periods. Third, corporate governance may vary from one insurance company to another based on many factors, such as the percentage of the government’s share and external shareholders. Future research should consider testing the impact of corporate governance based on qualitative studies. Fourth, there may be other internal factors influencing the underlying relationships that are difficult to model. This may be exacerbated in the KSA by the lack of readily available data. In addition, there are few studies in the literature on the Saudi insurance market or on any developing economies for comparison with our study. However, we find that our results are robust to endogeneity and exogenous issues. Despite limited data and resources, we believe our study contributes by bridging the gap in the literature on corporate governance in the Saudi insurance sector. Further, it can prove useful to managers, investors, market practitioners, and regulators. 6.3. Recommendations For some Saudi insurance firms, their losses have reached up to 50% of their capital, with these companies continuing to work only to increase or decrease capital rather than to find new ways to avoid losses. Working in the same manner in every successive administration only yields the same results, which is the continuation of losses. In addition, a change in departments must be followed by a change in the concept of management, which means making new decisions that change the company’s path to avoid accumulated losses and increase profits. These decisions must be linked to risk management, which includes controlling risks by assessing them and then developing strategies to reduce or avoid them (Alokla and Daynes 2017). Additionally, depending on the risks a company faces, risk assessment may mean the care and regulation of risks and management of crises through financial management, operations, monitoring, and action. To rectify problems caused by a lack of risk management, the board of directors must compensate the shareholders for the company’s accumulated losses above 50%, as the board is responsible for monitoring the company, and its members are nominated by shareholders. In addition, to reduce risks, companies should be transparent, sophisticated, and work to establish modern insurance programs and solutions appropriate for the needs of the stage of the company, such as programs concerned with technical and informational dimensions applied in global markets but not yet in the Saudi market. As a result, policymakers should create a new authority for insurance development and regulations. Economies 2023, 11, 21 36 of 41 Moreover, the industry may benefit from a mandate to merge some insurance com- panies into whatever form is deemed suitable. This will limit the number of firms and, simultaneously, contribute to creating a more organised and dynamic market. Mergers allow companies the ability to change and expand as well as mix experiences and capabili- ties. Finally, companies should adopt modern marketing methods, which rely on online connections and social media and are more flexible in terms of reaching customers, as well as creating new customer segments. Author Contributions: Conceptualization, M.A.S.A.-F.; methodology, M.A.S.A.-F. and J.A.; software, M.A.S.A.-F.; validation, M.A.S.A.-F. and J.A.; formal analysis, M.A.S.A.-F.; investigation, M.A.S.A.-F.; resources, M.A.S.A.-F.; data curation, M.A.S.A.-F.; writing—original draft preparation, M.A.S.A.-F.; writing—review and editing M.A.S.A.-F. and J.A.; visualization, M.A.S.A.-F. and J.A.; supervision, M.A.S.A.-F.; project administration, M.A.S.A.-F.; All authors have read and agreed to the published version of the manuscript. Funding: This research received no external funding. Institutional Review Board Statement: Not applicable. Informed Consent Statement: Not applicable. Acknowledgments: We thank the editor and the reviewers for the helpful comments and suggestions that significantly enhanced this work. We also express our gratitude to the Saudi Arabia Capital Market Authority, Tadawul, and Mubasher for providing the data. The usual disclaimer applies. Conflicts of Interest: The authors declare that they have no conflicts of interest. Appendix A Table A1. Saudi Arabia’s annual inflation rate. Year 2008 2009 2010 2011 2012 Inflation Rate (%) 9.87% 5.06% 5.34% 5.83% 2.87% 2013 3.53% 2014 2.24% Source: WorldData. Appendix B Table A2. The effect of the inflation rate on insurance performance. ROA ROE Tobin’s Q RE FE RE FE RE FE Constant Inflation Rate (%) Ln(BS) Ln(Assets) Ln(cid:0)FirmAge(cid:1) VOL RiskRet DeInMarket 191.6578 ** (94.1971) −1.4913 *** (0.5308) 2.4135 (3.1675) 1.9500 *** (0.6704) 5.7946 *** (1.6346) 0.0162 (0.0247) 0.0098 (0.0268) −5.5036 (3.6416) 206.6571 (127.4094) −1.4104 *** (0.5240) 7.4700 (4.7397) 3.1728 *** (1.2014) 8.5919 ** (3.4583) 0.0496 * (0.0274) −0.0173 (0.0312) −3.6340 (3.7955) 548.3650 (341.5285) −4.9384 ** (2.0455) 0.9498 (9.2462) 8.6261 *** (1.8117) 0.1477 (4.5319) −0.0003 (0.0881) −0.0073 (0.0880) −16.8457 (13.9176) 483.8100 (511.2455) −4.6704 ** (2.1026) 7.0212 (19.0185) 18.2202 *** (4.8206) 4.3745 (13.8770) 0.0445 (0.1100) −0.0499 (0.1250) −13.4009 (15.2297) 117.466 *** (27.3461) −0.7000 *** (0.1523) −0.9651 (0.9417) −2.1955 *** (0.2017) 0.3970 (0.4912) 0.0220 *** (0.0072) −0.0014 (0.0078) −2.5156 ** (1.0465) 133.9358 *** (36.7349) − 0.7387 *** (0.1511) −0.7239 (1.3666) −3.0764 *** (0.3464) 0.3821 (0.9971) 0.0220 *** (0.0079) −0.0030 (0.0090) −2.4665 ** (1.0943) Economies 2023, 11, 21 37 of 41 Table A2. Cont. ROA ROE Tobin’s Q RE FE RE FE RE FE Ln(gross written) Wald Chi2 F Statistic R2 (%) −9.9836 *** (3.7831) 60.74 *** 21.77 −12.2869 ** (5.3688) 7.37 *** 19.36 −29.0856 ** (13.7680) 33.26 *** 14.89 −35.1452 (21.5430) 2.90 *** 13.98 −2.7147 ** (1.0981) 162.14 *** 52.95 −2.7016 * (1.5479) 13.38 *** 51.93 Notes: *, **, and *** represent significance at 10%, 5%, and 1%, respectively. Standard errors are in parentheses. Inflation Rate (%) is the percent inflation rate = Final Consumer Price Index Value − Initial Consumer Price Index Value ×100; Ln(BS) is the natural logarithm of the board size; Ln(Assets) is the natural logarithm of firm assets; Ln(cid:0)FirmAge(cid:1) is the natural logarithm of firm age; VOL is volatility; RiskRet is (risk retention) net premiums written over gross premiums written, and DeInMarket (depth of insurance market) is defined as gross written premiums divided by total GDP. Ln(gross written) is the natural logarithm of gross premiums written, measured by the sum of both direct premiums written and assumed premiums written, before deducting the ceded reinsurance. We use the following equation for this model PERFit = α + β1 In f lation Rate (%)it + β2 Ln (BS)it + β3 Ln (Assets) it + β4 Ln (cid:0)FirmAge(cid:1) + β5VOLit + β6RiskRetit + β7DeInMarketit + β8 Ln (gross written)it + εit. Initial Consumer Price Index Value it Note 1 No insurance company offers only life insurance, but many offer various types of insurance, such as aviation, energy, engineering, marine, property, fire, accident and liability, motor, and health. References Adams, Renée Birgit. 2012. Governance and the financial crisis. International Review of Finance 12: 7–38. [CrossRef] Adams, Renee B., and Daniel Ferreira. 2007. A theory of friendly boards. The Journal of Finance 62: 217–50. [CrossRef] Ajemunigbohun, Sunday, Ayobami Folarin Elegunde, and Fatimo Titilope Azeez. 2020. Evaluation of corporate governance practices on financial performance of selected insurance firms in Nigeria. Acta Universitatis Danubius 16: 86–110. Akeem, Lawal Babatunde, Edwin Terer K, Odelabu Adedire Temitope, and Oluwaremi Feyitimi. 2014. Measuring impact of corporate governance on the performance of the Nigerian insurance company. International journal of Economics, Commerce, and Management 2: 1–17. Albassam, Bassam. 2021. Managing Public Debt: The Case of Saudi Arabia. International Journal of Economic Sciences 10: 20–38. [CrossRef] Aldamen, Husam, Keith Duncan, Simone Kelly, Ray McNamara, and Stephan Nagel. 2012. Audit committee characteristics and firm performance during the global financial crisis. Accounting and Finance 52: 971–1000. [CrossRef] Al-Faryan, Mamdouh Abdulaziz Saleh. 2017. The relationship between corporate governance mechanisms and the performance of Saudi listed firms. Corporate Ownership & Control 14: 338–49. [CrossRef] Al-Faryan, Mamdouh Abdulaziz Saleh. 2020. Corporate governance in Saudi Arabia: An overview of its evolution and recent trends. Risk Governance and Control: Financial Markets & Institutions 10: 23–36. [CrossRef] Al-Faryan, Mamdouh Abdulaziz Saleh. 2021. The effect of board composition and managerial pay on Saudi firm performance. Review of Quantitative Finance and Accounting 57: 693–758. [CrossRef] Al-Faryan, Mamdouh Abdulaziz Saleh. 2022. The Determinants of CEO Turnover for Listed Saudi Firms. Working Paper. Portsmouth: University of Portsmouth. Al-Faryan, Mamdouh Abdulaziz Saleh, and Everton Dockery. 2017. Ownership structure and corporate governance: What does the data reveal about Saudi listed firms? Corporate Ownership & Control 14: 413–24. [CrossRef] Al-Faryan, Mamdouh Abdulaziz Saleh, and Everton Dockery. 2021. Testing for efficiency in the Saudi stock market: Does corporate governance change matter? Review of Quantitative Finance and Accounting 57: 61–90. [CrossRef] Alhassan, Ali Faya, Mohammed Saleh Bajaher, and Amer M. Alshehri. 2015. Corporate governance, firm attributes and financial performance of Saudi listed banks. World Review of Business Research 5: 282–95. Almaqtari, Faozi A., Mohd Shamim, Hamood Mohd Al-Hattami, and Saleem Ahmed Aqlan. 2020. Corporate governance in India and some selected Gulf countries. International Journal of Managerial and Financial Accounting 12: 165–85. [CrossRef] Almoneef, Ahmed, and Durga Prasad Samontaray. 2019. Corporate governance and firm performance in the Saudi banking industry. Banks & Bank Systems 14: 147–58. [CrossRef] Alokla, Jassem, and Arief Daynes. 2017. Risk management in the Takaful industry: An empirical study of the Kuwaiti market. Journal of Islamic Economics, Banking and Finance 113: 1–21. [CrossRef] Alokla, Jassem, Arief Daynes, Paraskevas Pagas, and Panagiotis Tzouvanas. 2022. Solvency determinants: Evidence from the Takaful insurance industry. The Geneva Papers on Risk and Insurance-Issues and Practice. Available online: https://link.springer.com/ article/10.1057/s41288-021-00263-1 (accessed on 1 November 2022). [CrossRef] Economies 2023, 11, 21 38 of 41 Al-Sahafi, Ashraf, Marcus Rodrigs, and Lisa Barnes. 2015. Does corporate governance affect financial performance in the banking sector? Evidence from Saudi Arabia. International Journal of Economics, Commerce and Management 3: 1–26. Amran, Noor Afza, Mohd ‘Atef Md Yusof, Rokiah Ishak, and Norhani Aripin. 2014. Do characteristics of CEO and chairman influence government-linked companies performance? Procedia-Social and Behavioral Sciences 109: 799–803. [CrossRef] Apergis, Nicholas, and Thomas Poufinas. 2020. The role of insurance growth in economic growth: Fresh evidence from a panel of OECD countries. The North American Journal of Economics and Finance 53: 101217. [CrossRef] Arif, Hafiz Muhammad Raheel. 2019. Corporate governance and firm performance: A pragmatic investigation from insurance industry of Pakistan. Net Journal of Business Management 6: 8–16. Azutoru, Ibe Happy Chukwudike, Ugwuanyi Georgina Obinne, and Okanya Ogochukwu Chinelo. 2017. Effect of corporate governance mechanisms on financial performance of insurance companies in Nigeria. Journal of Finance and Accounting 5: 93–103. [CrossRef] Bagais, Omar Ali, and Khaled Salmen Aljaaidi. 2020. Corporate governance attributes and firm performance in Saudi Arabia. Accounting 6: 923–30. [CrossRef] Balcilar, Mehmet, Rangan Gupta, Chien-Chiang Lee, and Chien-Chiang Lee. 2018. The synergistic effect of insurance and banking sector activities on economic growth in Africa. Economic Systems 42: 637–48. [CrossRef] Baysinger, Barry D., and Henry N. Butler. 1985. Corporate governance and the board of directors: Performance effects of changes in board composition. Journal of Law, Economics, & Organization 1: 101–24. Bertrand, Marianne, and Sendhil Mullainathan. 2003. Enjoying the quiet life? Corporate governance and managerial preferences. Journal of Political Economy 111: 1043–75. [CrossRef] Bhagat, Sanjai, and Bernard S. Black. 2001. The non-correlation between board independence and long-term firm performance. SSRN Electron Journal of Corporation Law 27: 231–73. [CrossRef] Bhagat, Sanjai, and Brian Bolton. 2008. Corporate governance and firm performance. Journal of Corporate Finance 14: 257–73. [CrossRef] Bhatti, Maria, and M. Ishaq Bhatti. 2010. Toward understanding Islamic corporate governance issues in Islamic finance. Asian Politics & Policy 2: 25–38. [CrossRef] Bin, Leo, Jianguo Chen, and Anh Xuan Ngo. 2020. Revisiting executive pay, firm performance, and corporate governance in China. Economics, Management and Financial Markets 15: 9–32. [CrossRef] Boone, Audra L., Laura Casares Field, Jonathan M. Karpoff, and Charu G. Raheja. 2007. The determinants of corporate board size and composition: An empirical analysis. Journal of financial Economics 85: 66–101. [CrossRef] Braiotta, Louis, Jr., R. Trent Gazzaway, Robert H. Colson, and Sridhar Ramamoorti. 2010. The Audit Committee Handbook. Hoboken: John Wiley & Sons. Broome, Lissa Lamkin, John M. Conley, and Kimberly D. Krawiec. 2011. Does critical mass matter? views from the boardroom. Seattle University Law Review 34: 1049–80. Bubic, Jasenka, and Toni Susak. 2015. The Impact of Intangible Assets on Financial Performance of Croatian Companies. Paper presented at the Ninth International Scientific Conference on Economic and Social Development, Istanbul, Turkey, April 9–10; pp. 462–68. Available online: https://bib.irb.hr/datoteka/798685.Book_of_Proceedings_Istanbul_2015_Final.pdf#page=470 (accessed on 1 November 2022). Cecrdlova, Andrea. 2021. Comparison of the approach of the Czech National Bank and the European Central Bank to the effects of the global financial crisis. International Journal of Economic Sciences 10: 18–46. [CrossRef] Charumathi, B. 2012. On the Determinants of Profitability of Indian Life Insurers—An Empirical Study. Paper presented at the World Congress on Engineering, London, UK, July 4–6; vol. 1. Available online: http://www.iaeng.org/publication/WCE2012/WCE2 012_pp505-510.pdf (accessed on 6 June 2021). Chaudhary, Priyanka. 2014. Corporate governance in insurance sector. International Journal of Social Science and Economic Research 4: 33–39. Available online: http://www.hindex.org/2014/p618.pdf (accessed on 6 June 2021). Capital Market Authority (CMA). 2017. Corporate Governance Regulations in the Kingdom of Saudi Arabia. Available online: https: //cma.org.sa/en/RulesRegulations/Regulations/Documents/CGRegulations_en.pdf (accessed on 6 June 2021). Coles, Jeffrey L., Naveen D. Daniel, and Naveen D. Daniel. 2008. Boards: Does one size fit all? Journal of Financial Economics 87: 329–56. [CrossRef] Conyon, Martin J., and Annita Florou. 2002. Top executive dismissal, ownership and corporate performance. Accounting and Business Research 32: 209–25. [CrossRef] Coughlan, Anne T., and Ronald M. Schmidt. 1985. Executive compensation, management turnover, and firm performance: An empirical investigation. Journal of Accounting and Economics 7: 43–66. [CrossRef] Dalton, Dan R., and Catherine M. Dalton. 2011. Integration of micro and macro studies in governance research: CEO duality, board composition, and financial performance. Journal of Management 37: 404–11. [CrossRef] Datta, Nibedita. 2018. Impact of corporate governance on financial performance: A study on DSE listed insurance companies in Bangladesh. Global Journal of Management and Business Research 18: 32–39. Available online: https://journalofbusiness.org/index. php/GJMBR/article/view/2463/4-Impact-of-Corporate-Governance_JATS_NLM_xml (accessed on 6 June 2021). Davis, Stanley M. 1979. No connection between executive age and corporate performance. Harvard Business Review 57: 6–8. Demsetz, Harold, and Kenneth Lehn. 1985. The structure of corporate ownership: Causes and consequences. Journal of Political Economy 93: 1155–77. [CrossRef] Economies 2023, 11, 21 39 of 41 Denis, David J., Diane K. Denis, and Atulya Sarin. 1997. Ownership structure and top executive turnover. Journal of Financial Economics 45: 193–221. [CrossRef] Dikolli, Shane S., William J. Mayew, and Dhananjay Nanda. 2014. CEO tenure and the performance-turnover relation. Review of Accounting Studies 19: 281–327. [CrossRef] Duchin, Ran, John G. Matsusaka, and Oguzhan Ozbas. 2010. When are outside directors effective? Journal of Financial Economics 96: 195–214. [CrossRef] Eckles, David L., Martin Halek, Martin Halek, David W. Sommer, and Rongrong Zhang. 2011. Earnings smoothing, executive compensation, and corporate governance: Evidence from the property–liability insurance industry. Journal of Risk and Insurance 78: 761–90. [CrossRef] Eisenberg, Theodore, Stefan Sundgren, and Martin T. Wells. 1998. Larger board size and decreasing firm value in small firms. Journal of Financial Economics 48: 35–54. [CrossRef] Elamer, Ahmed A., Aws AlHares, Collins G. Ntim, and Ismail Benyazid. 2018. The corporate governance–risk-taking nexus: Evidence from insurance companies. International Journal of Ethics and Systems 34: 493–509. [CrossRef] Eling, Martin, and Sebastian D. Marek. 2014. Corporate governance and risk taking: Evidence from the U.K. and German insurance markets. Journal of Risk and Insurance 81: 653–82. [CrossRef] Ezzine, Hanene. 2018. Corporate governance and social norms during financial crisis: Evidence from France and Saudi Arabia. Journal of Management and Governance 22: 707–48. [CrossRef] Fadun, Olajide Solomon. 2013. Corporate governance and insurance firms performance: Empirical study of Nigerian experience. Journal of Insurance Law & Practice 3: 11–31. Fama, Eugene F., and Michael C. Jensen. 1983. Separation of ownership and control. The Journal of Law and Economics 26: 301–25. [CrossRef] Fekadu, Gardachew Worku. 2015. Corporate governance on financial performance of insurance industry. Corporate Ownership & Control 13: 1201–9. [CrossRef] Fodio, Musa Inuwa, Jide Ibikunle, and Victor Chiedu Oba. 2013. Corporate governance mechanisms and reported earnings quality in listed Nigerian insurance firms. International Journal of Finance and Accounting 2: 279–86. Fogel, Eric, and Andrew M. Geier. 2007. Strangers in the House: Rethinking Sarbanes-Oxley and the Independent Board of Directors. Delaware Journal of Corporate Law 32: 33–72. Friedman, Stewart D., and Harbir Singh. 1989. CEO succession and stockholder reaction: The influence of organizational context and event content. Academy of Management Journal 32: 718–44. Available online: https://www.jstor.org/stable/256566 (accessed on 15 November 2016). [CrossRef] Gaganis, Chrysovalantis, Iftekhar Hasan, Panagiota Papadimitri, and Menelaos Tasiou. 2019. National culture and risk-taking: Evidence from the insurance industry. Journal of Business Research 97: 104–16. [CrossRef] Garen, John E. 1994. Executive compensation and principal-agent theory. Journal of Political Economy 102: 1175–99. [CrossRef] Ghabayen, Mohammad Ahid. 2012. Board Characteristics and Firm Performance: Case of Saudi Arabia. International Journal of Accounting and Financial Reporting 2: 168–200. [CrossRef] Gibbons, Robert, and Kevin J. Murphy. 1992. Optimal incentive contracts in the presence of career concerns: Theory and evidence. Journal of Political Economy 100: 468–505. [CrossRef] Gibson, Michael S. 2003. Is corporate governance ineffective in emerging markets? The Journal of Financial and Quantitative Analysis 38: 231–50. [CrossRef] Gompers, Paul, Joy Ishii, and Andrew Metrick. 2003. Corporate governance and equity prices. The Quarterly Journal of Economics 118: 107–56. [CrossRef] Gugong, Benjamin Kumai, Love O. Arugu, and Kabiru Isa Dandago. 2014. The impact of ownership structure on the financial performance of listed insurance firms in Nigeria. International Journal of Academic Research in Accounting, Finance and Management Sciences 4: 560–71. [CrossRef] Guruswamy, D., and Adugnaw Marew. 2017. Determinants of profitability performance of insurance companies: A case study of selected insurance companies in Ethiopia. i-Manager’s Journal on Management 12: 26–44. Hasan, Zulkifli. 2009. Corporate governance: Western and Islamic perspectives. International Review of Business Research Papers 5: 277–93. Hashed, Abdulwahid A., and Faozi A. Almaqtari. 2021. The impact of corporate governance mechanisms and IFRS on earning management in Saudi Arabia. Accounting 7: 207–24. [CrossRef] Hausman, J. A. 1978. Specification tests in econometrics. Econometrica: Journal of the Econometric Society 46: 1251–71. [CrossRef] Hermalin, Benjamin E., and Michael S. Weisbach. 1991. The effects of board composition and direct incentives on firm performance. Financial Management 20: 101–12. Available online: https://www.jstor.org/stable/3665716 (accessed on 15 November 2016). [CrossRef] Huang, Li-Ying, Tzy-yih Hsiaon, and Gene C. Lai. 2007. Does corporate governance and ownership structure influence performance? Evidence from Taiwan life insurance companies. Journal of Insurance Issues 30: 123–51. Available online: https://www.jstor.org/ stable/41946275 (accessed on 6 June 2021). Economies 2023, 11, 21 40 of 41 International Association of Insurance Supervisors. 2012. International Association of Insurance Supervisors Report. Available online: http://www.actuaries.org/CTTEES_INSREG/Documents/2-Year_Roadmap_for_IAIS_Activities_2013-2014.pdf (accessed on 6 June 2021). Ismail, Muhaizam. 2013. Determinants of financial performance: The case of general takaful and insurance companies in Malaysia. International Review of Business Research Papers 9: 111–30. Available online: https://www.bnm.gov.my/documents/20124/761697 /04_RP_Muhaizam.Ismail_Determinants_of_Financial_Performance.pdf (accessed on 6 June 2021). Jensen, Michael C. 1993. The modern Industrial Revolution, exit, and the failure of internal control systems. Journal of Finance 48: 831–80. [CrossRef] Jensen, Michael C., and Kevin J. Murphy. 2010. CEO incentives—It’s not how much you pay, but how. Journal of Applied Corporate Finance 22: 64–76. [CrossRef] Jenter, Dirk, and Fadi Kanaan. 2015. CEO turnover and relative performance evaluation. The Journal of Finance 70: 2155–84. [CrossRef] John, Kose, Lubomir Litov, and Bernard Yeung. 2008. Corporate governance and risk-taking. The journal of Finance 63: 1679–728. [CrossRef] Karamanou, Irene, and Nikos Vafeas. 2005. The association between corporate boards, audit committees, and management earnings forecasts: An empirical analysis. Journal of Accounting Research 43: 453–86. [CrossRef] Klein, April. 1998. Firm performance and board committee structure. The Journal of Law and Economics 41: 275–304. [CrossRef] Kozak, Sylwester. 2011. Integration with the European financial system and changes of the non-life insurance sector concentration in Poland. Acta Scientiarum Polonorum. Oeconomia 10: 41–53. Available online: https://js.wne.sggw.pl/index.php/aspe/article/ view/4009 (accessed on 1 November 2022). Lehn, Kenneth M., Sukesh Patro, and Mengxin Zhao. 2009. Determinants of the size and composition of US corporate boards: 1935-2000. Financial Management 38: 747–80. [CrossRef] Li, Ke, Lei Lu, Usha R. Mittoo, and Zhou Zhang. 2015. Board independence, ownership concentration and corporate performance— Chinese evidence. International Review of Financial Analysis 41: 162–75. [CrossRef] Linck, James S., Jeffry M. Netter, and Tina Yang. 2008. The determinants of board structure. Journal of Financial Economics 87: 308–28. [CrossRef] Lipton, Martin, and Jay W. Lorsch. 1992. A modest proposal for improved corporate governance. The Business Lawyer 48: 59–77. Available online: https://www.jstor.org/stable/40687360 (accessed on 6 June 2021). Liu, Yu, Mihail K. Miletkov, Zuobao Wei, and Tina Yang. 2015. Board independence and firm performance in China. Journal of Corporate Finance 30: 223–44. [CrossRef] MacAvoy, Paul W., Scott Cantor, Jim Dana, and Sarah Peck. 1983. ALI Proposals for Increased Control of the Corporation by the Board of Directors: An Economic Analysis. Statement of the Business Roundtable on the American Law Institute’s Proposed “Principles of Corporate Governance and Structure: Restatement and Recommendation”. New York: Business Roundtable. Maharjan, Rajendra. 2019. Corporate Governance and Financial Performance of Insurance Companies in Nepal. International Research Journal of Management Science 4: 99–117. [CrossRef] Mardi, Markonah, Achmad Sudiro, Surachman, and Mintarti Rahayu. 2017. The effect of good corporate governance and premium growth on the performance of insurance companies. Journal of Finance and Banking Review 2: 1–7. [CrossRef] Maug, Ernst. 1997. Boards of directors and capital structure: Alternative forms of corporate restructuring. Journal of Corporate Finance 3: 113–39. [CrossRef] Mazviona, Batsirai Winmore, Mbakisi Dube, and Tendai Sakahuhwa. 2017. An analysis of factors affecting the performance of insurance companies in Zimbabwe. Journal of Finance and Investment Analysis 6: 1–21. Available online: http://www.scienpress. com/Upload/JFIA%2FVol%206_1_2.pdf (accessed on 6 June 2021). McConnell, John J., and Henri Servaes. 1990. Additional evidence on equity ownership and corporate value. Journal of Financial Economics 27: 595–612. [CrossRef] McGuire, Jean, Sandra Dow, and Bakr Ibrahim. 2012. All in the family? Social performance and corporate governance in the family firm. Journal of Business Research 65: 1643–50. [CrossRef] Mehran, Hamid. 1995. Executive compensation structure, ownership, and firm performance. Journal of Financial Economics 38: 163–84. [CrossRef] Morck, Randall, Andrei Shleifer, and Robert W. Vishny. 1988. Management ownership and market valuation: An empirical analysis. Journal of Financial Economics 20: 293–315. [CrossRef] Murphy, Kevin J. 1985. Corporate performance and managerial remuneration: An empirical analysis. Journal of Accounting and Economics 7: 11–42. [CrossRef] Mwangi, Mirie, and Jane Wanjugu Murigu. 2015. The determinants of financial performance in general insurance companies in Kenya. European Scientific Journal 11: 288–97. Available online: https://eujournal.org/index.php/esj/article/view/4953 (accessed on 6 June 2021). Najjar, Naser Jamil. 2012. The impact of corporate governance on the insurance firm’s performance in Bahrain. International Journal of Learning and Development 2: 1–17. [CrossRef] Olaniyi, Clement Olalekan, and Olaolu Richard Olayeni. 2020. A new perspective into the relationship between CEO pay and firm performance: Evidence from Nigeria’s listed firms. Journal of Social and Economic Development 22: 250–77. [CrossRef] Economies 2023, 11, 21 41 of 41 Panditharathna, K. M. 2016. Corporate governance determinants of firm performance: Empirical evidence from banking, finance and insurance companies in Sri Lanka. Kelaniya Journal of Human Resource Management 11: 1–16. [CrossRef] Peni, Emilia. 2014. CEO and Chairperson characteristics and firm performance. Journal of Management & Governance 18: 185–205. [CrossRef] Raheja, Charu G. 2005. Determinants of board size and composition: A theory of corporate boards. Journal of Financial and Quantitative Analysis 40: 283–306. [CrossRef] Shaaban, Ramadhan, and Michael N. Wahome. 2018. Influence of premium deposits on financial performance of insurance companies listed at Nairobi securities. IJSSIT. Available online: http://repository.seku.ac.ke/handle/123456789/4300 (accessed on 6 June 2021). Shaheen, Yaser, and Nasser Jaradat. 2019. Corporate governance impact on insurance firm’s performance. The case of Palestine. International Journal of Academic Research in Accounting, Finance and Management Sciences 9: 201–10. [CrossRef] [PubMed] Shan, Yuan George, and Ron P. McIver. 2011. Corporate governance mechanisms and financial performance in China: Panel data evidence on listed non financial companies. Asia Pacific Business Review 17: 301–24. [CrossRef] Shiu, Yung-Ming. 2004. Determinants of United Kingdom general insurance company performance. British Actuarial Journal 10: 1079–110. [CrossRef] Tornyeva, Kingsley, and Theophilus Wereko. 2012. “Soft” governance and firm performance: A study of Ghanaian insurance firms. European Journal of Business and Management 4: 86–94. Available online: https://www.iiste.org/Journals/index.php/EJBM/ article/view/2707/2731 (accessed on 8 August 2021). Volpin, Paolo F. 2002. Governance with poor investor protection: Evidence from top executive turnover in Italy. Journal of Financial Economics 64: 61–90. [CrossRef] Wanyama, David Wanyonyi, and Tobias Olweny. 2013. Effects of corporate governance on financial performance of listed insurance firms in Kenya. Public Policy and Administration Research 3: 96–120. Available online: https://www.iiste.org/Journals/index.php/ PPAR/article/view/5340/5446 (accessed on 7 August 2021). Yemane, Abreham Amanuel, M. Lakshmipathi Raju, and R. Madhusudana Raju. 2015. The impact of corporate governance on firm’s performance: Evidence from Ethiopian insurance companies. Research Journal of Finance and Accounting 6: 225–33. Available online: https://www.iiste.org/Journals/index.php/RJFA/article/view/22171/23370 (accessed on 5 August 2021). Yeoh, Siew-Boey, and Chee-Wooi Hooy. 2022. CEO age and risk-taking of family business in Malaysia: The inverse S-curve relationship. Asia Pacific Journal of Management 39: 273–93. [CrossRef] Yermack, David. 1996. Higher market valuation of companies with a small board of directors. Journal of Financial Economics 40: 185–211. [CrossRef] Yousaf, Imran, and Jassem Alokla. 2022. Herding behaviour in the Islamic bank market: Evidence from the Gulf region. Review of Behavioral Finance. [CrossRef] Zhang, Li, and Norma Nielson. 2015. Solvency analysis and prediction in property–casualty insurance: Incorporating economic and market predictors. Journal of Risk and Insurance 82: 97–124. [CrossRef] Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
10.1007_s41109-023-00538-7
Boyd et al. Applied Network Science (2023) 8:15 https://doi.org/10.1007/s41109-023-00538-7 Applied Network Science RESEARCH Open Access The persistent homology of genealogical networks Zachary M. Boyd*, Nick Callor, Taylor Gledhill, Abigail Jenkins, Robert Snellman, Benjamin Webb and Raelynn Wonnacott *Correspondence: zach_boyd@byu.edu Department of Mathematics, Brigham Young University, Provo, UT 84602, USA Abstract Genealogical networks (i.e. family trees) are of growing interest, with the largest known data sets now including well over one billion individuals. Interest in family history also supports an 8.5 billion dollar industry whose size is projected to double within 7 years [FutureWise report HC-1137]. Yet little mathematical attention has been paid to the complex network properties of genealogical networks, especially at large scales. The structure of genealogical networks is of particular interest due to the practice of forming unions, e.g. marriages, that are typically well outside one’s immediate family. In most other networks, including other social networks, no equivalent restriction exists on the distance at which relationships form. To study the effect this has on genealogi- cal networks we use persistent homology to identify and compare the structure of 101 genealogical and 31 other social networks. Specifically, we introduce the notion of a network’s persistence curve, which encodes the network’s set of persistence intervals. We find that the persistence curves of genealogical networks have a distinct structure when compared to other social networks. This difference in structure also extends to subnetworks of genealogical and social networks suggesting that, even with incom- plete data, persistent homology can be used to meaningfully analyze genealogical networks. Here we also describe how concepts from genealogical networks, such as common ancestor cycles, are represented using persistent homology. We expect that persistent homology tools will become increasingly important in genealogical explora- tion as popular interest in ancestry research continues to expand. Keywords: Persistent homology, Genealogical networks, Social networks, Persistence curves, Bottleneck distance Introduction The study of genealogical networks, that is networks relating parents with children and spouses with each other through successive generations is of rapidly growing interest, both because of genealogy’s popular appeal and its applications in genetics (Kaplanis et al. 2018), sociology (Hamberger et al. 2011), population sciences (Rohde et al. 2004), and economics (Greenwood et  al. 2014). Growing data availability of rich, temporally resolved data is also driving interest in genealogy. For example, FamilySearch has con- structed a human family tree with over 1.40 billion individuals, based on 2.21 billion sources, including 4.78 billion images (https:// www. famil ysear ch. org/ en/ newsr oom/ © The Author(s) 2023. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the mate- rial. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/. Boyd et al. Applied Network Science (2023) 8:15 Page 2 of 29 compa ny- facts). Popularization of DNA testing services and increasing availability of audio sources, geographic tags, occupation metadata, and migration records combine to make genealogical networks some of the largest, most richly featured, geospatially embedded temporal networks in existence. Examples of relevant academic studies include methods for automatically constructing networks from documents (Malmi et al. 2018; Bloothooft et  al. 2015), analyzing marriage patterns (Greenwood et  al. 2014), structured population modeling, branching processes (Hey and Machado 2003), and biconnected components (Hamberger et al. 2011; Hage and Harary 1983). Of particular interest to us are works that study distance to recent common ancestors, both theoreti- cally and via simulation (e.g. Chang 1999; Rohde et al. 2004). A growing body of litera- ture also uses genealogical networks for genetic inference, as in (Kaplanis et al. 2018). Related to these genealogical endeavors, a major goal of network science is to describe the structure of such real-world networks. In this paper, we consider persistent homol- ogy as a tool to both analyze and explore the structure of genealogical networks. Per- sistent homology, roughly speaking, is a method of representing voids or gaps in the structure of a network, that distinguishes how significant these voids are to the overall network structure. Persistent homology can be used to compare these voids across two networks without requiring a correspondence between the individual vertices or edges, or even requiring the networks to be the same size. The basic idea involves “filling in” the network with simplices (points, edges, triangles, tetrahedra, etc.) and keeping track of how the network changes as we do so (see "Persistent homology of networks" for details). Some similar applications of persistent homology in the study of networks include Carstens and Horadam (2013), Kannan et  al. (2019); Horak(2009); Petri et  al. (2013). The collaboration networks studied in Carstens and Horadam (2013) are similar to the social networks that we use for comparison in this paper, though our focus is primar- ily on distinguishing these from genealogical networks. Both Kannan et  al. (2019) and Horak (2009) apply persistent homology techniques to general randomized networks of various forms. It is also possible to vary the technique for generating a topological object from a network, as in Petri et  al. (2013) where three methods are compared. We also recommend Aktas et al. (2019) and Otter et al. (2017) as good overviews of the general methods of applying persistent homology. For this paper, our method of constructing a topological representative for each network follows the same general pattern as the work cited above. However, we also acknowledge the wide variety of alternatives for encoding such information. Chazal et al. (2013) and Vandaele et al. (2018) encode their information as point-clouds rather than graphs. A higher-dimensional version of persistent homology is presented in Blumberg and Lesnick (2020), which may permit the inclusion of time-varying networks. Finally, the formulation in Arafat et al. (2020) may allow for better analysis of corrupted or too- large datasets. We also wish to bring attention to four particular applications that demonstrate the versatility of persistent homology. In each of these applications, persistent homology has been used to identify structural voids in data and then to associate these voids to recognizable features in the underlying networks. It is the latter use that we wish to emphasize. Robins et  al. Robins et  al. (2016) have shown that voids found using per- sistent homology correspond to percolating spheres in a porous material. In Lee et  al. Boyd et al. Applied Network Science (2023) 8:15 Page 3 of 29 (2012), structural voids arise when several groups of neurons are strongly connected sequentially, but out-of-sequence pairs are only weakly connected. In these neurologi- cal networks, persistent homology provides a way to identify and classify these differ- ent sequences as well as quantify the strength of these connections. The application in Duman and Pirim (2018) provides a method for extending traditional genetic analysis tools to a parameterized family of datasets by constructing an appropriate topological object. Lastly, Mattia et al. (2016) shows that structural voids or gaps can also represent much more abstract concepts. In this case persistent voids are shown to correspond to the atonality in music compositions. Intuitively, the voids or gaps in genealogical networks should be quite different when compared with other networks, such as social networks, since unions1 (such as mar- riages) in genealogical networks typically form at specific distances, rather than through other mechanisms e.g. triadic closure. That is, distances between individuals who form unions are typically not too small or too large (see "Background: genealogical and social networks"). In contrast, in other social networks, new connections can form at any dis- tance but are often quite small (Sintos and Tsaparas 2014). This difference in network growth between genealogical and other social networks causes differences in network topology that are reflected in the network’s persistent homology. Thus persistent homol- ogy is a useful descriptive tool for exploring and modeling the structure of genealogical networks. Here, we propose a new method for representing persistent homology, which we call a persistence curve (see "Comparing networks using persistent homology"). The persistence curves of many genealogical networks are very similar to each other, and importantly the persistence curves of subsets of genealogical networks, that is, sampled genealogical networks, are also similar to the persistence curves of unsampled genea- logical networks (see "Results"). To give our study of genealogical networks context we also study the persistent homol- ogy of social networks. We find that the same result holds for the social networks we consider, in that the persistence curves of social networks show a common pattern and the persistence curves for social and sampled social networks are similar (see "Results"). We confirm our analysis using another tool for comparing persistent homologies, the bottleneck distance, which is also capable of detecting and differentiating the distinct homology patterns between genealogical and other social networks. In summary, we make the following contributions: • Introduce the notion of a persistence curve and introduce the use of this together with the bottleneck distance as a tool for the analysis of general networks. • Report the distinct persistent homology structure of genealogical networks using both persistence curves and the bottleneck distance. • Link this structure to genealogically relevant concepts. • Similarly, report the distinct persistence homology structure of social networks and compare this to the structure of genealogical networks. 1 In order to be inclusive of various relevant relationships in this paper, we use the word “union” to describe not only legal marriages and common law marriages but also some others, including any relationship that produced children. Boyd et al. Applied Network Science (2023) 8:15 Page 4 of 29 • Report evidence that persistent homology methods work well even in the presence of incomplete data. This is particularly relevant given that genealogical data is often, if not necessarily, incomplete. Throughout the paper, examples from family networks are contrasted with other social networks to highlight the unique features of genealogical networks from a persistent homology point of view. The paper is organized as follows. In "Background: genealogical and social networks" we describe both genealogical and social networks. In "Persistent homology of net- works" we define the persistent homology of a network and introduce the notion of persistence curves. In "Comparing networks using persistent homology" we define the bottleneck distance and show how both this distance and persistence curves can be used to compare networks. In "Results" we describe the genealogical and social data sets we use in our study and give our experimental results in "Results". In "Results" also includes a discussion of how certain structural features of social and genealogical networks are represented using persistent homology. In "Conclusion" we summarize our results and conclude with a discussion regarding the use of persistent homology as a tool for ana- lyzing general network structure and recovering network features. Throughout we give examples of each of the concepts we introduce. Background: genealogical and social networks We represent genealogical networks with a graph G = (V , E) , where V = {1, 2, . . . , n} are the individuals within the network, and E are the (genealogical) relationships. These relationships consist of both parent–child edges and spouse (or more generally union) edges. For the sake of simplicity, these edges are considered to be undirected. We note that the structure of a genealogical network is often thought of as being “tree- like”, since genealogical networks are often constructed from an individual, their parents, their grandparents, and so on, ignoring union edges. The result is a tree, i.e. a connected acyclic graph, if we create only a few generations of the family. However, full genealogi- cal networks are not trees due to the presence, for example, of triangles consisting of two parents and a child (with the two parent–child edges and one union edge). Because of the frequency of such cycles and the fact that they are the smallest possible cycles, we refer to them as trivial cycles. The other typical familial cycle, or cycle found within a family consisting of two parents and some number of children, is a cycle of length four consisting of two parents and two children. Although familial cycles are ubiquitous in genealogical networks, they are not the only cycles that can form. Going far enough through an individual’s ancestors, it is often possible to find a nearest common ancestor, i.e., a common ancestor of one’s father and mother. If such an ancestor exists (and it usually does exist), then the genealogical net- work has a nontrivial cycle. We refer to this as a common ancestor cycle, which consists of only parent–child edges. Other nontrivial cycles are possible in genealogical networks via unions. For instance, a “double cousins” relationship occurs when two siblings from one family form unions with two siblings from another family. The result is a union cycle, or a cycle that contains only union edges and the parent–child edges connecting siblings. Boyd et al. Applied Network Science (2023) 8:15 Page 5 of 29 Fig. 1 Left: The largest connected component of the Tikopia genealogical network consisting of 288 individuals from the island of Tikopia in Polynesia from the year 1930 to 2008, is shown (https:// www. kinso urces. net/ brows er/ datas ets. xhtml). Parent-child edges are shown in blue and union edges are shown in red. Right: The largest connected component of the Residence Hall social network consisting of 217 individuals and their friendships from the Australian National University campus is shown (Residence 2022) Fig. 2 The histogram representing the finite “distance to union” distances is shown where data is collected from 104 genealogical networks from kinsources.net. The height of each bar represents the fraction of unions that form at a specific distance In genealogical networks, union and parent–child edges can combine in any number of ways to create complex non-tree structures (see Fig. 1 left). A feature that is particular to genealogical networks is that union edges typically form at specific distances within these networks. Here the distance d(i, j) between i and j is the shortest path distance between these individuals if such a path exists. Otherwise, it is infinite. In a genealogical network we refer to the distance between two individu- als before they form a union as the couple’s distance to union. For cultural, genetic, and other reasons these distance are typically not small, i.e. usually larger than four. Conse- quently, genealogical networks do not typically have small nonfamilial cycles and often have large extended cycles. This is illustrated in Fig.  2 where distance to union data is collected from 104 publicly available genealogical networks given in Table  2 in the Appendix. Here familial cycles are omitted and the height of each bar represents the Boyd et al. Applied Network Science (2023) 8:15 Page 6 of 29 Fig. 3 Left: Shown in orange is the distribution of the lengths of the cycles forming a basis of the nonfamilial cycle lengths in the San Marino (SM) genealogical network. The analogous distribution of cycle lengths is shown in blue for all cycles in the Deezer Europe (DE) social network. Center: Shown in orange is again the basis cycle length distribution of the San Marino genealogical network. In red is the distribution of the basis cycle lengths averaged over ten realizations of the (loopy, multi-edged) configuration model on the San Marino network. Since the configuration model generates graphs with the same degree distribution as the SM network, this panel indicates that SM’s longer cycles do not arise simply from the degree distribution. Right: Shown in blue is again the basis cycle length distribution of the Deezer social network. In green is the distribution of the basis cycle lengths averaged over ten realizations of the configuration model on the Deezer social network. For this social network, the cycle length distribution can be mostly explained by the degree distribution alone fraction of unions that form at a specific distance. Noticeably, few unions form at dis- tances less than five with the large majority of distance falling between 5 and 10. The observation that genealogical networks have large extended cycles is illustrated in Fig.  3. Shown left in orange is the distribution of cycle lengths of the San Marino genealogical network, a network of the population of the Republic of San Marino from the 15th to the end of the 19th century (https:// www. kinso urces. net/ brows er/ datas ets. xhtml). In this network, which consists of 28,586 individuals, there are 7,146 familial cycles of length three and 8,636 familial cycles of length four. These are omitted in the figure so we can observe the lengths of the cycles forming a basis of nonfamilial cycles in the network. For the sake of contrast, in blue is the distribution of cycle lengths in a basis of the cycles found in the Deezer Europe social network, consisting of 28,281 individuals. Here, similar to genealogical networks, a social network is represented by a graph G = (V , E) where the vertices V also represent individuals. The difference is that in a social network the edges represent some type of social interaction(s). The Deezer network is an online music streaming platform whose social network represents indi- viduals in Europe who use the platform where edges represent mutual user-follower relationships. Noticeably, the San Marino network has relatively few nonfamilial basis cycles under length ten but quite a few cycles with lengths greater than thirty. In contrast, the Deezer social network has a much tighter distribution of basis cycles ranging from roughly five to fifteen in length. To understand the extent to which these cycle distributions are related to the local structure of the associated networks we compare these to the cycle distribution of the associated configuration models of these two networks, respectively. The configuration model is a model for generating random networks with a given degree sequence (New- man 2006). Taking the degree sequences from both the San Marino genealogical and Deezer social network, we create ten versions of these networks each with the same degree sequences. The result of averaging the basis cycle length distributions of these versions of the San Marino and Deezer networks is shown in Fig. 3 (center and right in Boyd et al. Applied Network Science (2023) 8:15 Page 7 of 29 red and green, respectively). While the cycle distribution for the San Marino network is quite different from what the configuration model produces, the Deezer social net- work is quite similar to the distribution predicted by its configuration model. This sug- gests that much of the cycle structure in the Deezer social network is dominated by local interactions, whereas the cycles in the San Marino genealogical network are affected by nonlocal mechanisms that form the network. This includes, presumably, the nonlocal distance to union phenomena described above. The relations we see in Fig. 3 between the cycle length distribution for the San Marino genealogical network and the Deezer social network are typical of the genealogical and social networks we consider in "Data". This suggests that cycle length distribution is a feature that can be used to distinguish genealogical from social networks. Specifically, when we consider two networks with a similar number of cycles, genealogical networks have a much wider distribution of cycle lengths than social networks. However, the method used to calculate the cycle length distribution in Fig.  3 does not provide any further insight into this phenomenon. This limitation motivates us to apply tools from persistent homology which provides ways to describe and measure the relation between any two network cycles. The additional structure that can be obtained by these meth- ods allow us to further distinguish the structure of genealogical and social networks (see "Network comparison using bottleneck distance") and to relate the structural differences demonstrated in Fig.  3 to mechanisms that produce genealogical and social networks, respectively (see "Connections"). Persistent homology of networks Persistent homology provides a method for studying cycles in a network. For the pur- poses of this paper, a brief explanation of persistent homology will be given from the context of simplicial homology. For a more in-depth treatment of simplicial homology, see Chapter  2.1 of Hatcher (2002). For those readers who are either familiar with the basics of persistent homology or who wish to skip the following technical discussion it is possible to proceed to "Data" where we discuss the social and genealogical networks we analyze. For a network given by a graph G = (V , E) we define the distance matrix D(G) = [dij] to have entries dij = d(i, j) , which is the length of the shortest path between individual i and j. For each value δ that appears in the distance matrix D(G), we form a simplicial complex Gδ as follows. The set of 0-simplices is equivalent to the set of vertices of G, where each 0-simplex is identified with a single vertex. Since the distinction between 0-simplices and vertices is purely formal, we will use the terms 0-simplex and vertex interchangeably, and the 0-simplices will be indexed the same way as the vertices. The set of 1-simplices Eδ corresponds to the set of edges {i, j} such that d(i, j) ≤ δ , where the edge {i, j} is identified with the 1-simplex formed by i and j . Again the distinction here is unnecessary for our present discussion, so we will use the same notation for 1-sim- plices and edges. However, the simplicial complex Gδ may also contain objects that do not have equivalent representatives in the graph G, namely the n-simplices for n ≥ 2 . For each integer n ≥ 2 , the set of n-simplices in Gδ consists of all n-simplices [a0 a1 . . . an] Boyd et al. Applied Network Science (2023) 8:15 Page 8 of 29 Fig. 4 The hexagonal network G = G1 in Example 3.1 is filled in as i increases from 0 to 3. This produces the simplicial complexes G0, G1, G2, G3 shown left to right such that d(ai, aj) ≤ δ for 0 ≤ i < j ≤ n . That is, Gδ includes an n-simplex σ if each ver- tex listed in σ is within δ of every vertex listed in σ. In order to simplify our remaining definitions, we extend our definition of Gδ to include all non-negative integers. For i ≥ 0 , let δi be the greatest entry of D(G) such that δi ≤ i . Let Gi = Gδi . This definition together with our construction of Gδ ensures the fol- lowing three important properties are true for all Gi . 1. For i < j , Gi is a subcomplex of Gj , i.e. every simplex of Gi is a simplex of Gj. 2. For i ≥ 1 , there exists a subcomplex of Gi that can be identified with the original graph G. 3. Since G is finite, let M = maxij d(i, j) , then, for all i ≥ M , Gi = GM. Example 3.1 (Hexagonal Network) Consider the hexagonal network G = (V , E) with six vertices, form- ing a single cycle, shown in Fig. 4b. This network has the distance matrix D(G) = 0 1 2 3 2 1 1 0 1 2 3 2 2 1 0 1 2 3 3 2 1 0 1 2 2 3 2 1 0 1 1 2 3 2 1 0 . For the values i = 0 , 1, 2, 3, we form four simplicial complexes, G0 , G1 , G2 , and G3 where we let Gi = (Vi, Ei) . For i = 0 , E0 is empty. Thus, G0 consists of six vertices. For i = 1 the set E1 contains the six edges that form the network’s single cycle, so G1 = G . This graph has no trivial cycles (i.e., triangles), so G1 contains no simplices of dimension greater than 1 (i.e., no n-simplices for n > 1 ). For i = 2 the set E2 gains six additional edges. We also now have eight trivial cycles. Each of these cycles is the boundary of a 2-simplex, so G2 contains these eight 2-simplices as well. However, no subset of these 2-simplices forms the boundary of a 3-simplex, so G2 has no simplices of dimension greater than 2. For i = 3 the set E3 contains all possible edges between the vertices of G, so all possible triv- ial cycles are present. Additionally, all possible 2-simplices, and hence all possible n-sim- plices, are also present in G3 . In particular, G3 is a 6-simplex with its boundary. Since M = 3 is the largest value we see in the distance matrix, then Gi = G3 for i ∈ Z , i > 3. Boyd et al. Applied Network Science (2023) 8:15 Page 9 of 29 The persistent homology of the network G measures how the homology of Gi changes as i increases. If certain features can be identified across multiple values of i, we say they persist. Intuitively, features that arise from the actual network structure should persist for many values of i, while features that arise because of measurement error, ‘noise’, should only appear sporadically. The Stability Theorem (the main theorem of Cohen- Steiner et al. (2007)) states that if the error in measuring a network is bounded by some constant C, then the persistent homology of the true network and the persistent homol- ogy of the noisy network will differ by at most C. We will make this statement more pre- cise in "Persistence diagrams and bottleneck distance". Here we give a formal definition of persistent homology in terms of simplicial homol- ogy, which we will immediately follow this with equivalent definitions in the context of networks. We use Hp(Gi) to denote the dimension-p simplicial homology of the simpli- cial complex Gi with coefficients in Z2 , as Hp(X) is a vector space of Z2. (pth Persistent Homology) For a graph G, and integers i, j with 0 ≤ i ≤ j , Definition 1 let the function φi,j : Hp(Gi) → Hp(Gj) be the linear map induced by the inclusion Gi → Gj . The pth persistent homology of G, PHp(G) is the pair ({Hp(Gi)}i≥0, {φi,j}0≤i<j). Our analysis in the "Comparing networks using persistent homology" and "Results" sections only requires the first few dimensions of persistent homology to distinguish the genealogical and social networks we consider. In order to better understand what persis- tent homology calculates, in what follows we will provide equivalent definitions for PH0 , PH1 , and PH2 using network concepts. We also illustrate how these definitions apply to the hexagonal network in Fig. 4b. (See Examples 3.3, 3.4, and 3.5 for PH0 , PH1 , and PH2 ; respectively.) (Births and Deaths) Let G = (V , E) be a network with simplicial com- Definition 2 plexes G0, G1, G2, · · · . The pth persistent homology of G provides maps φi,j between the pth homology of Gi and the pth homology of Gj . Suppose that basis elements have been chosen for each Hp(Gi) so that if α is a basis element of Hp(Gi) , then φi,j(α) is either trivial in Hp(Gj) or a basis element of Hp(Gj) . The birth of a basis element α ∈ Hp(Gj) is the minimum index i such that α = φi,j( ˆα) for some basis element ˆα ∈ Gi . The death of α is the minimum index k such that φj,k (α) is trivial. Remark 3.2 Those already familiar with persistent homology will find that the preceding definition is somewhat nonstandard, although it is equivalent to the standard definition. We have taken this approach to reduce the notation burden on non-specialist readers. We have done similarly with some of the other persistent homology definitions. We will demonstrate how to choose such representatives for H0 , H1 , and H2 in the fol- lowing definitions. Given such representatives, though, the maps φi,j and φj,k are simply the maps on homology induced by the inclusion maps Gi ⊂ Gj ⊂ Gk . That is, if a repre- sents α ∈ Hp(Gi) , then a also represents φi,j(α) . The Fundamental Theorem of Persistent Boyd et al. Applied Network Science (2023) 8:15 Page 10 of 29 Homology ensures that we can choose a single representative that corresponds to α ∈ Hp(Gj) , ˆα ∈ Hp(Gi) , and φj,k (α) ∈ Hp(Gk ) . The birth of α is then just the first Gi in which the representative exists, and the death of α is the first Gk in which the representa- tive is null-homotopic i.e., homotopic to a trivial cycle. (Representing Persistent Homology: Dimension 0) Let G = (V , E) be a Definition 3 network with vertices V = {1, 2, . . . , n} which form k connected components. Then H0(G0) ∼= Zn 2 , so we can identify the basis for H0(G0) with the set of all n vertices. Like- wise, we may choose k vertices, one from each connected component, to represent the basis for H0(Gi) ∼= Zk 2 for i ≥ 1 . Thus, we will refer to the vertices of G as representa- tives of PH0(G) . (In fact, PH0(G) is a vector space whose basis elements are equivalence classes of formal sums of 0-simplices.) Example 3.3 We now consider PH0(G) for the hexagonal network G in Fig. 4, with G0 , G1 , G2 , and G3 in the same figure. Recall that G has six distinct vertices forming one connected component. If we take any numbering of the vertices, V = {1, 2, 3, 4, 5, 6} , then H0(G0) ∼= Z6 2 , which Z2 with basis V. For i > 0 , H0(Gi) ∼= Z2 , which is is equivalent to the vector space over Z2 with basis {1} . For any v ∈ V , since i = 0 is the first equivalent to the vector space over time we see v, we call this the birth of v. At i = 1 , since we have removed all vertices except 1 from the basis, we say this is the death of those five 0-simplices. Since 1 will always be in the basis for Gi , the death of 1 is said to be ∞. (Representing Persistent Homology: Dimension 1) Let G = (V , E) be a net- Definition 4 work with one connected component. For each i ≥ 0 , we can identify the basis of H1(Gi) with a set Ci of cycles in Gi . The Fundamental Theorem of Persistent Homology allows us to choose these cycles so that if σ is a cycle in Ci , then exactly one of the following is true for any integer j ≥ 0 : 1. σ does not exist in Gj , in which case j < i, 2. σ is trivial or null-homotopic in Gj , in which case i < j, 3. σ is a cycle in Cj. Thus, we will refer to the cycles in (cid:31)i≥0 Ci as the representatives of PH1(G) . (Again, PH1(G) is actually much larger than this. These are actually representatives of equiva- lence classes that form a basis for PH1(G) as a vector space.) We note that C0 is always empty, since there are no edges in G0 . Furthermore, rank(H1(Gi)) = |Ci| for all i ≥ 0 . Because of the construction of the Gi all representa- tives of PH1(G) will be present in G1 . One can think of the representatives of PH1(G) as representing “large” cycles. More specifically, if a cycle σ is contained in (cid:31)s≤i≤t Ci , then it must have a diameter of at least t and at least one pair of consecutive vertices distance s apart. Boyd et al. Applied Network Science (2023) 8:15 Page 11 of 29 2 Fig. 5 A visual depiction of simplices and cycles present in G2 . Left: Four trivial cycles filled by individual 3 1 3] , [3 2-simplices: [1 cycle, 1, 2, 3, 5, 1 filled in by two 2-simplices [1 3] and [1 represented as the faces of a regular octahedron. Right: The closed surface of G2 is filled in by four 3-simplices 4 [1 5] . Center Left: A non-trivial, but null-homotopic 5] . Center Right: All eight 2-simplices 6] , and [ 2 5] , [1 6], [1 6], [3 6], [2 6] 3 3 3 5 4 5 2 5 4 3 Example 3.4 we now consider PH1(G) for the hexagonal network G in Fig. 4b. In both Fig. 4a and 4b we see that G0 has no cycles, G1 has exactly one cycle, and that the cycle in G1 is non-trivial. In Figs. 5a and 5b, we have indicated some of the cycles in G2 , namely the cycles 1,2,3,1; 3,4,5,3; 1,5,6,1; and 1,3,5,1 in Fig.  5a and the cycle 1,2,3,5,1 in Fig.  5b. In fact, Fig.  5c shows us that G2 is an octahedron and therefore every cycle in G2 is either trivial or null- homotopic. Finally, G3 contains even more cycles than G2 , such as 1,3,6,1; but these are all null-homotopic since G3 also contains every possible 2-simplex for six vertices. Therefore, PH1(G) has only one representative, the cycle 1,2,3,4,5,6,1; which appears in G1 , so we say that t = 1 is the birth of the cycle. The cycle is null-homotopic in G2 , so t = 2 is the death of the cycle. We now turn our attention to PH2(G) , but in order to represent PH2(G) we need to introduce some new structure for the induced graphs. A triangle [a b c] in Gi is a set of three vertices, a, b, and c, that form a trivial cycle in Gi . That is, the edges {a, b} , {b, c} , and {a, c} are all present in Gi . A closed surface in Gi is a set of distinct triangles so that for each [a b c] in the set there is exactly one other triangle [a b d] also in the set. A closed surface in Gi is trivial if the corresponding set of 2-simplices is null- homotopic in Gi . That is, the closed surface is “filled in” by some collection of 3-sim- plices in Gi . For example, the octahedron in Fig. 5c is a non-trivial closed surface in G2 because there are no 3-simplices in G2 . In G3 , however, we add edges between ver- tices at distance 3. In turn, we gain several 3-simplices, including [1 2 3 6] , [1 3 5 6] , [3 4 5 6] , and [2 3 4 6] . Figure  5d shows three of these 3-simplices to demonstrate how the closed surface from G2 is filled in by all four. (Representing Persistent Homology: Dimension 2) Let G = (V , E) be a Definition 5 network with one connected component. For each i ≥ 0 , we can identify the basis for H2(Gi) with a set Si of non-trivial closed surfaces in Gi . The Fundamental Theorem of Persistent Homology allows us to choose these representatives so that if σ is a closed surface in Si , then exactly one of the following is true for any integer j ≥ 0 Boyd et al. Applied Network Science (2023) 8:15 Page 12 of 29 1. σ does not exist in Gj , in which case j < i, 2. σ is trivial in Gj , in which case i < j, 3. σ is a cycle in Sj. Thus we will refer to the closed surfaces in (cid:31)i≥0 Si as the representatives of PH2(G). The geometric intuition for PH2(G) is similar to that of PH1(G) in identifying large (cid:31)s≤i≤t Si , then σ is a closed surface with diameter at least t. The value ‘voids’ in G. If σ ∈ of s is harder to describe, but is related to the density of vertices. Example 3.5 We now consider PH2(G) for the hexagonal graph G in Example 3.1. Recall from Exam- ple 3.4 that G0 and G1 have no trivial cycles, and therefore contain no closed surfaces. We can see in Fig. 5 that G2 has exactly one closed surface and it must be non-trivial, since there are no 3-simplices. Finally, G3 has many closed surfaces, but because it contains every possible 3-simplex on six vertices, these are all trivial. Therefore, PH2(G) has only one representative, the octahedral closed surface in G2 . This surface first appears in G2 , so t = 2 is its birth, and the surface is filled by a solid in G3 , so t = 3 is its death. (Persistence Intervals) Recall that the birth of a representative σ ∈ PHp(G) Definition 6 (vertex, cycle, or closed surface) of the persistent homology of a network G is the small- est integer i so that σ ∈ Gi , and the death of σ is the largest integer j so that σ ∈ Gj−1 and σ is trivial in Gk for k ≥ j , if such an integer exists. The persistence interval for σ is [a, b) , where a and b are the birth and death of σ , respectively. This represents the set of all parameter values i for which the equivalence class corresponding to σ is a non-trivial element of Hp(Gi) . The persistence of σ is b − a. Example 3.6 We now finish our consideration of the persistent homology of G from Fig. 4b. Recall from Example 3.3 that PH0(G) has six representatives. These all have birth t = 0 . Five of these have a death of t = 1 , and one of these has a death of ∞ . Therefore the persistence inter- vals for PH0(G) are [0, 1) × 5 and [0, ∞) × 1. From Example 3.4, we know PH1(G) has one representative, with birth t = 1 and death t = 2 . Therefore the corresponding persistence interval is [1, 2) . Note that the diameter of the cycle is 3 and every pair of consecutive vertices is distance 1 apart. This follows the idea mentioned earlier that the representatives of PH1(G) indicate ‘large’ cycles. Specifi- cally, the diameter of σ is at least the death of σ , and the birth of σ is the maximum dis- tance between consecutive vertices. From Example  3.5, PH2(G) has one representative, with birth t = 2 and death t = 3 . Therefore, the persistence interval for that element is [2, 3) . Note that the diameter of the corresponding set of vertices is 3 in G. This also follows the idea mentioned earlier that Boyd et al. Applied Network Science (2023) 8:15 Page 13 of 29 Table 1 The persistence intervals of the Tikopia genealogical network and the hexagon network are shown Dimension Interval Type and Persistence Tikopia Hexagon Dimension 1 Dimension 0 [0, ∞) × 8, [0, 1) × 286 [1, 2) × 16 , [1, 3) × 19 , [1, 4) × 5 , [1, 5) × 3 , [1, 6) × 2 , [1, 7) × 1 [2, 3) × 4 , [3, 4) × 11 , [4, 5) × 12 , [5, 6) × 4 , [6, 7) × 5 , [7, 8) × 1 , [8, 9) × 1 Here the notation [a, b) × k indicates that the network has k persistence intervals [a, b). The corresponding persistence diagrams are shown in Fig. 6 and the corresponding persistence curve for the Tikopia network is shown in Fig. 7 [0, ∞) × 1, [0, 1) × 1 [1, 2) × 1 Dimension 2 [2, 3) × 1 PH2(G) identifies large ‘voids’ in G. Specifically, the death of σ is a lower bound on the diameter of σ. Given the representatives chosen in Definitions 3, 4, 5, and 6, we have the following three observations regarding the persistent homology of a finite, undirected, unweighted graph G: (i) If G has n vertices, then PH0(G) will have exactly n persistence intervals, with exactly one [0, ∞) interval for each connected component and the rest will be [0, 1) intervals. (ii) In dimension 1, PH1(G) describes the number and sizes of the non-trivial cycles in the original network. The persistence intervals will all be of the form [1, b) for some integer b > 1 . The value of b is related to the diameter of the corresponding cycle. In the networks we have studied, we note that a persistence interval [1, b) in PH1(G) corresponds to a simple cycle with between 3b − 2 and 3b vertices, inclu- sive. (iii) In dimension 2, the voids we detect in PH2(G) tell us about the nontrivial intersec- tions of cycles. Such intersections are hard to visualize but, roughly speaking, a representative in PH2(G) can only form if several large cycles intersect each other pairwise. Comparing networks using persistent homology In this section we demonstrate how methods based on persistent homology can be used to compare different networks. The two methods we introduce in this paper are based on using (a) the bottleneck distance and (b) the persistence curves of a given set of networks. Both (a) and (b) rely on first computing persistence intervals then analyzing the differences in these intervals. The two networks we consider throughout this section to demonstrate these meth- ods are the Tikopia genealogical network from Fig.  1 (left) and the hexagonal net- work from Fig.  4. The persistence intervals for these networks are given in Table  1, respectively. Boyd et al. Applied Network Science (2023) 8:15 Page 14 of 29 Persistence diagrams and bottleneck distance One common way to represent persistence intervals is to plot them as points in R × (R ∪ {∞}) , which is typically referred to as a persistence diagram. While this method of visualizing a network’s persistent homology does not indicate how often a given persistence interval occurs, it does provide information on what kind of persis- tence intervals occur for a given network. Definition 7 network G. The persistence diagram for PHp(G) is a multiset of points in defined as follows. (Persistence Diagrams) Let PHp(G) be the pth persistent homology of a R × (R ∪ {∞}) • For each σ ∈ PHp(G) with persistence interval [a, b) , we include one copy of the point (a, b). • For each c ∈ R , we include infinitely many copies of the point (c, c). Note that we include the points (a, a) to represent features in G that are considered trivial in PHp(G) , such as cycles consisting of exactly three vertices. This inclusion is necessary for us to define a meaningful metric on the space of persistence diagrams. The metric we use here is called the bottleneck distance. (Bottleneck Distance) Let S1 and S2 be persistence diagrams for two Definition 8 graphs G and H, respectively. Let η range over the set of bijections from S1 to S2 . Then the bottleneck distance between S1 and S2 is dB(S1, S2) = inf η sup x∈S1 �x − η(x)�∞. The Fundamental Theorem of Persistent Homology (introduced in Zomorodian and Carlsson (2005), explained well in Otter et al. (2017) and Aktas et al. (2019)) ensures that if two graphs are isomorphic, the corresponding persistence diagrams will be equal, and thus the bottleneck distance will be 0. However, it is possible for non-iso- morphic graphs to have identical persistence diagrams. Example 4.1 (Bottleneck Distance Between the Tikopia and Hexagonal Networks) Notice that the per- sistence intervals for the Tikopia genealogical network (see Table 1) include, as a subset, the persistence intervals from the hexagonal network we considered in Example 3.6. We can form a bijection between the persistence diagrams of the Tikopia and hexagonal net- work by identifying the non-trivial intervals from the hexagonal network with those of the Tikopia network. We then map any additional intervals from the Tikopia network of the form [a, b) to the trivial interval [ a+b 2 ) . (The perceptive reader may notice that this is not clearly a bijection, but there is a standard technique from set theory for modifying it to be bijective.) 2 , a+b Boyd et al. Applied Network Science (2023) 8:15 Page 15 of 29 Fig. 6 Left: The persistence diagram of the hexagonal network in Fig. 4b is shown. Center: The persistence diagram of the Tikopia genealogical network in Fig. 1 (left) is shown. Right: A bottleneck bijection between the persistence intervals of the hexagonal and Tikopia family network is shown. Orange lines show which points are matched to points of the form (a, a) where a ∈ R This mapping is shown in Fig.  6 (right). Here, [1, 7) is mapped to [4, 4) . As this pair of points is further apart than any other pair in this bijection, the bottleneck distance for the two networks is at most three, since we take an infimum over all possible bijections. Conversely, there is no interval in the hexagonal persistence diagram that is closer to [1, 7) than 3, so the bottleneck distance is at least three. Thus, the bottleneck distance for these two persistence diagrams is exactly 3. Suppose that two networks, each of which is connected, admit isometric embeddings Rn in . The Stability Theorem (Cohen-Steiner et al. 2007) guarantees that if the Hausdorff distance between the embeddings is δ , then the bottleneck distance for the correspond- ing persistence diagrams is at most δ . For example, if the PH1 persistence diagrams differ by δ , then any attempt to pair up cycles in the networks must include at least one pair of cycles for any isometric embedding that are δ apart in that embedding. In "Network comparison using bottleneck distance" we apply this idea to a large collection of genea- logical and social networks. Persistence curves For the network data we consider, persistence diagrams obfuscate a key difference that we consider important: the number of persistence intervals. For a simple exam- ple of this, consider networks of the form V = {1, 2, . . . , n} with edges of the form {i, i + 1} for 1 ≤ i < n . For n ≥ 2 , any network of this type will have persistence intervals [0, 1) × (n − 1) and [0, ∞) × 1 . However, when plotting the persistence diagram we will only ‘see’ two points: (0, 1) and (0, ∞). To address this limitation, we introduce the notion of a persistence curve as a new way to visualize the persistent homology of a network (see Definition 9). The difference between the persistence curve and the persistence diagram of a network is that the per- sistence curve also includes the number of intervals of a particular type. To create a per- sistence curve we first compute a network’s persistence intervals, then sort the intervals of a given dimension by their persistence into a bar graph. For instance, in dimension 1 the Tikopia genealogical network has thirteen [1, 2) intervals, nineteen [1, 3) intervals, Boyd et al. Applied Network Science (2023) 8:15 Page 16 of 29 Fig. 7 Left: The barcode of the Tikopia genealogical network in dimension 1 is shown. The individual bars are formed from the persistence intervals given in Table 1. Right: The associated persistence curve for the Tikopia network in Fig. 1 is shown etc. which are sequentially stacked as shown in Fig. 7 (left) to create what we will call a barcode. To create the associated persistence curve we connect the endpoints of each subsequent bar as shown in Fig. 7 (right). In dimension-one, the birth times of our intervals will all start at 1, as the networks we consider are unweighted, undirected, and connected. This means that in this dimension the resulting bar graph is also a plot of the death times for each interval. For higher- dimensions, which have varied birth times, we also plot the lengths of the intervals but for simplicity we start at 1 as in dimension-one. A formal definition of a network’s persistence curves is the following. (Persistence Curves) Let G = (V , E) be a network with nonempty vertex Definition 9 and edge sets. Let {[aj, bj)} be the set of all persistence intervals for each σj ∈ PHn(G) where j ∈ N the persistence curve PHn(G) is the linear interpolation of the set of points {(bj − (aj − 1), j)} where bj−1 − (aj−1 − 1) ≤ bj − (aj − 1). . For all n ∈ N Visualizing persistence intervals as a curve allows us to compare the persistent homol- ogy of different networks in a similar fashion to persistence diagrams while retaining different information. In particular, we can see how many intervals there are of a given persistence, whereas the persistence diagram only indicates the presence of such an interval. In what follows we will typically plot the persistence curves of multiple net- works on the same axes to indicate what differences exist in the persistent homology of different networks (cf. "Results"). Data The data we consider in this paper is of two types; genealogical network data and other social network data. The genealogical networks we consider are drawn from ninety-seven genea- logical networks found in (https:// www. kinso urces. net/ brows er/ datas ets. xhtml), which range in size from n = 17 to 5, 016 individuals. The social network data we use is taken from twenty-seven different social networks obtained from (http:// snap. stanf ord. edu/ data/ index. html# socne ts, http:// snap. stanf ord. edu/ data/ index. html# socne ts, http:// netwo rkrep osito ry. Boyd et al. Applied Network Science (2023) 8:15 Page 17 of 29 com/ soc. php, http:// netwo rkrep osito ry. com/ soc. php). These range in size from n = 16 to 2, 539 individuals. (See Table 2 in the Appendix for a full description of this data set.) Although many larger genealogical and social network data sets are available we are limited by both the temporal and spacial complexity of the algorithm used to compute persistence intervals. The program we used, called Ripser (from the python package Ripser) (Ripser 2021), has a computational and spacial complexity of O((n + m)3) where n is the number of individuals and m is the number of edges in a network. The number n + m is the number of simplicies in the network. In the genealogical networks we con- sider there are between n + m = 41 to 15, 735 simplicies and in the social networks we consider between n + m = 41 to 19, 056 simplices. To understand how a network’s persistence intervals are effected by the completeness or incompleteness of data we also consider subnetworks sampled from a few, much larger, genealogical and social networks. These sampled networks are created by randomly selecting an individual with a single neighbor, i.e. a vertex of degree 1, then perform- ing a breadth-first-search starting with this individual to find the η closest individuals in the network to this individual. Because of the spatial and computational limitations of Ripser we choose 600 ≤ η ≤ 3, 000 to ensure we can compute the persistence intervals of these sampled networks. In total we sampled from four different genealogical networks and four different social networks. These are the Advogat, LastFM Asia, Deezer HU and Deezer RO social networks and the genealogical networks 96–99 shown in Table 2, respectively. We sampled from each of these networks five times each to create a total of 20 sampled genealogical networks and 20 sampled social networks. The reason we begin our breadth-first search with a vertex of degree 1 is to ensure that our sampled networks have vertices both on the boundary and the interior of the original network we sampled to better mimic the structure of the original genealogical and social networks. Apart from the (i) genealogical and social networks we consider and (ii) sampled ver- sions of these networks, we also consider what we refer to as (iii) atypical genealogical networks. There are a number of genealogical networks that appear to be created with no attempt to represent all or even a fraction of the familial relationships. For example, the US Presidents network, cited as Atyp. Gen. Network 2 in Table 2, follows the shortest genealogical path between presidents leaving out extraneous relationships. We consider a number these atypical genealogical networks, which form a contrast to the more stand- ard genealogical networks we consider especially in terms of their peristent homology. A description of each of the (i) genealogical, social, (ii) sampled genealogical, sampled social, and (iii) atypical genealogical networks we consider is given at the end of the Appendix. Results Here we compare genealogical and other social networks using the (a) bottleneck dis- tance and the (b) persistence curves defined in "Comparing networks using persistent homology" (see Definitions 8 and 9, respectively). For those who have skipped in "Persis- tent homology of networks" and  “Comparing networks using persistent homology”, the bottleneck distance gives us a distance between two networks based on the differences in their persistent homology. Persistence curves give us a way of visualizing this difference but in greater detail (cf. Figure 7). Boyd et al. Applied Network Science (2023) 8:15 Page 18 of 29 Fig. 8 PCA projections of the bottleneck distances between networks are shown. Left: The bottleneck distance between each of the twenty sampled genealogical and sampled social networks is shown. Center: The bottleneck distances are shown between the genealogical, social, and atypical genealogical networks we consider. Right: The bottleneck distances in the center panel are shown for only the genealogical and social networks we consider Network comparison using bottleneck distance Here we compute the bottleneck distance between every pair from the social and gene- alogical networks we consider. To visualize these results we use principal component analysis to identify the two components that account for the most variance and then plot this data in (see Fig. 8). R2 From each part of Fig. 8 we can see that genealogical networks are generally separated from social networks and form clusters that are easily distinguished. For the sampled networks (shown left), we can easily separate genealogical and social networks, and we can identify at least two distinct subclasses of genealogical networks. However, the bottleneck distance does an inferior job separating the non-sampled genealogical and social networks (shown center and right). The exception are the atypical genealogical networks, whose persistence intervals differ significantly enough from all of the other networks to be distinguishable as a third class of networks (shown center). Comparison of genealogical and social networks using persistence curves Persistence curves give us a new alternative way of comparing networks. The advantage of using these curves compared to the bottleneck distance is that these curves give us a more detailed picture of how the number of persistence intervals varies from network to network. This allows us to better differentiate the structure of genealogical networks from social networks as well as observe the structure common to genealogical networks and those common to social networks, respectively. In Fig.  9 the persistence curves for the unsampled genealogical and unsampled social networks are shown in blue and red, respectively. The atypical genealogical networks are shown in green. The social networks have persistence curves that are quite vertical in both dimension 1 and dimension 2. For dimension 1, this indicates that most cycles in a social network are close to being trivial; either because they have a relatively small circumference or because they can be decomposed into a union of cycles with small circumferences. In particular, most of the social networks have a maximum death time of three (see Defini- tion  2), which corresponds to having a basis of cycles whose maximal circumference is at most nine. In other words, any cycle of circumference ten or more decomposes as the union of smaller cycles. For dimension 2, the steepness of the persistence curves indicate the presence of many distinct, yet similar, paths between certain pairs of vertices. Boyd et al. Applied Network Science (2023) 8:15 Page 19 of 29 Fig. 9 Comparison of persistence curves for full networks vs sampled networks, grouped by dimension and type of network. Upper Row: Sampling social networks typically stretches the persistence curve in only one axis without affecting the other axis. Lower Row: Sampling genealogical networks typically shrink the persistence curve in both axes. Overall the average slope for social networks tends to increase when sampled, while genealogical networks experience a decrease in average slope In contrast, the genealogical networks have persistence curves that have a much more horizontal profile indicating that most cycles are quite long and there are fewer ‘alternate paths’ between pairs of vertices. In the extreme, the atypical genealogical networks are nearly flat in dimension 1, which reflects the fact that these atypical networks were inten- tionally constructed to have very few cycles. In dimension 2, the atypical networks show a similar slope to most of the typical genealogical networks, but the size of the alternative paths in these networks are much larger. This is likely due to the high number of individu- als who were added only to link distant individuals, e.g. presidents. In a typical genea- logical network, the additional relationships between such individuals would allow large cycles to decompose but in the atypical genealogical networks this in not the case. In Fig. 10, we see the persistence curves for the sampled genealogical and sampled social networks shown in blue and red, respectively. The atypical genealogical networks are shown in green. Again the social networks have persistence curves that are quite vertical in both dimensions, although these curves are not as tall as in the case of unsampled social net- works. This indicates that as a social network is sampled it retains a similar proportion of close-to-trivial cycles, but may lose many of the alternative paths between vertices that appear in dimension 2. By contrast, for genealogical networks the persistence curves indi- cate the complete loss of very large cycles in conjunction with a proportional loss of close- to-trivial cycles. In dimension 2, genealogical networks experience a more severe loss of alternative paths than the social networks. As a result, though sampling shrinks the scale of the persistence curves for social and genealogical networks, they remain visually distinct. As in the bottleneck distance plots, genealogical and social networks appear to cluster together in that they have similar types of persistence curve. In fact, this is true whether or not the networks are sampled or unsampled. This suggests that even with incomplete Boyd et al. Applied Network Science (2023) 8:15 Page 20 of 29 Fig. 10 Upper Row: Comparison of persistence curves for full networks by type. Lower Row: Comparison of persistence curves for sampled networks by type, excluding atypical genealogical networks. In each dimension, the average slope for genealogical networks is typically lower than the average slope for a social network. The atypical genealogical networks have the lowest average slope and much greater total length. The behavior for average slopes is more pronounced for sampled networks than for full networks data social network and genealogical networks have a distinguishable persistent homol- ogy, at least at the scales we consider. It is worth mentioning that, while the bottleneck distance plots show us to an extent how different genealogical and social networks are the persistence curves show us what are differences are. The distance plots in Fig. 8 do have the advantage of simplicity, how- ever, and could presumably be used to more quickly identify differences in networks that are not as apparent as those we find between genealogical and social networks. Connections It is also possible to use persistent homology to study properties of a network, such as the number of connected components, the typical size of cycles, or even “missing links” in the data. For genealogical and social networks, we can convert these mathematical concepts into more familiar ideas such as family groups or common ancestors. This also allows us to make conjectures about the persistent homology for such networks by converting standard assumptions about families or social networks into the language of persistence. In dimension 0, the number of connected components determines the number of [0, ∞) intervals, and the total number of distinct vertices is the number of [0, ∞) intervals plus the number of [0, 1) intervals. In the context of a genealogical network, each connected compo- nent represents a family group that is not related to the other family groups by any known connection. Thus, if a given family network is indeed a single “family” of relatives, there should be exactly one [0, ∞) interval. In our Tikopia example we have eight [0, ∞) intervals each of which correspond to exactly one connected component of this genealogical network. (Note that Fig. 1 (left) shows only the largest of these components). In this example, most of the the other ‘family groups’ are actually individuals with no relation edges in the network. Boyd et al. Applied Network Science (2023) 8:15 Page 21 of 29 Fig. 11 Left: A common ancestor cycle. The top most vertex is a common ancestor of the lowest vertex. The horizontal red line is a marriage, all other lines are parent–children edges. Center: A union cycle, specifically the double cousin situation described in "Background: genealogical and social networks". The left-most and right-most vertices are parents of their neighboring vertices. The two horizontal red lines are marriage edges. Right: A θ-cycle formed by a common ancestor cycle with two overlapping hybrid cycles In social networks, the connected components create what could be referred to as friend groups. Unlike genealogical networks, there are usually few restrictions on which edges form in a social network. As such, we do not have a conjecture about the number of [0, ∞) intervals in this setting in general. However, sampling any network as described in "Data" will result in a new network with a single [0, ∞) interval. Moving to dimension 1, persistence intervals in this dimension describe the way that each connected component is internally structured. In sufficiently large genealogical net- works, we will see three kinds of features that we call common ancestors, union cycles, and hybrid cycles. A common ancestor cycle occurs when two descendants of an individual form a union or have a child together. We use the term union cycle to refer to situations where a cycle is formed through union edges and edges connecting two siblings. The final type of cycle of note, the hybrid cycles, are those formed by any other combination of parent–child edges and union edges, which includes everything that is not a strict common ancestor or union cycle. These three types of cycles are illustrated in Fig. 11, where marriage edges are indicated by red edges and parent–child edges are indicated by blue edges. We show a com- mon ancestor in Fig. 11a. Figure 11b is an example of a union cycle in which two siblings in one family form unions with two siblings in another, where only a single parent in each family is shown. In Fig. 11c we give an example of a θ-cycle, which is the union of a com- mon ancestor cycle and two overlapping hybrid cycles. This example comes from siblings of one family marrying cousins from another family. These cycles can be any length theoreti- cally, but cultural norms affect the typical size and number of each type of cycle differently. Recording practices and incomplete data also limit whether these cycles appear in a given dataset. Thus having a description of these cycles together with an understanding of the culture may help identify errors in the recorded data. Conversely, understanding the distri- bution of cycles in high fidelity datasets can help identify the underlying cultural norms and help extrapolate where individuals are missing in incomplete data sets. Since many cultures avoid marrying close relatives, common ancestor cycles tend to have a fairly large circumference. In the Tikopia network (see Fig. 1) we see persistence intervals with death values as high as 7 corresponding to cycles with a circumference Boyd et al. Applied Network Science (2023) 8:15 Page 22 of 29 of at least 21 individuals, which appear to be common ancestor cycles. This partially explains why persistence curves are so flat: there are relatively few minimal common ancestor cycles in a network, but they have very high persistence. More precisely, if the distance to union (the total number of individuals in a common ancestor cycle) is n, then the persistence of that cycle is ⌊n/3⌋ . However, the representatives of persistent homol- ogy only include a basis for these cycles, instead of including every possible distinct cycle. In particular, a large common ancestor cycle will decompose into the union of two hybrid cycles if the hybrid cycles are each shorter than the common ancestor cycle, as shown in Fig. 11c. Persistent homology will reflect the size of the two smaller cycles instead of the larger common ancestor cycle. We note that it is possible to identify the actual cycles chosen for our basis, but the software we used does not provide that infor- mation and size of the networks prohibits us from identifying the cycles manually. In social networks, we see that highly persistence cycles are quite rare. In order to have a cycle of persistence 3, for instance, we need a loop with circumference 9 or higher with no shorter paths between any two vertices in the loops. It may be that phenomena like the small-world effect or, more colloquially, six-degrees of freedom limit the maxi- mal persistence of social networks. We see this reflected in our example data sets with a maximum persistence of 3 for all but one of the social networks. Conclusion In this paper, we explore the persistent homology structure of genealogical networks, motivated by the observation that family links tend to form in a fixed range of intermedi- ate distances, which makes genealogical networks homologically distinct from most other social networks. We also introduce the notion of a persistence curve, which can be used to summarize and compare the persistent homology structure of any network. We also relate specific genealogical structures, such as the common ancestor cycle, to homology objects. We find that, in the presence of incomplete data homology analysis is still genealogi- cally useful. We note missing data due to recording practices and incomplete data (a ubiq- uitous feature of real genealogical networks), limits the kind of cycles that appear in a given dataset. Thus having a description of these cycles together with an understanding of the culture may help identify errors in the recorded data. Conversely, understanding the distribution of cycles in high fidelity datasets can help identify the underlying cultural norms and help extrapolate where individuals are missing in incomplete data sets. There are several interesting directions in which this work could be expanded. For exam- ple, our work has made it clear that there is a real need to analyze the persistent homology of large networks, with at least tens of thousands of nodes, since family formation generally takes place at these scales. The Ripser library we relied on was not able to reach these scales. Additionally, we are very interested in creating random graph models which reflect the actual homology of human family networks—a first attempt at this by our group has been fairly successful at the scale of hundreds of nodes (Flores 2021). More broadly, there is a need to model the ground truth human family network. All the extant data sources represent biased, limited, and noisy subnetworks, while the true interest of the genealogical community is in the ground truth network. Tools for signal denoising, image inpainting, and graph extrapola- tion, for example, could be useful in this context. Finally, an important aspect of genealogical Boyd et al. Applied Network Science (2023) 8:15 Page 23 of 29 Table 2 Social and genealogical network data sets Network data Network type & name Vertices Edges Citation Social networks Friendship & aquaintance Dolphins Zachary karate club Residence hall Highland tribes Seventh graders Physicians Highschool Dutch College Sampson’s monastery Adolescent health Hamsterster friends Social network 1 Social network 2 Social network 5 Social network 7 Social network 8 Social network 9 Firm Hi-Tech Wiki-Vote FB-PAGES-FOOD Advogato LastFM Asia Deezer HU Deezer RO Collaboration & business Social network 4 Social network 6 Social network 11 Social network 12 Disease transmission Taro exchange Information sharing Social network 3 Social network 10 62 34 217 16 29 241 70 32 25 2539 2952 32 32 32 32 32 32 33 889 620 6541 7624 47538 41773 32 32 32 32 22 32 32 159 78 http:// www- perso nal. umich. edu/ mejn/ netda ta/ http:// vlado. fmf. uni- lj. si/ pub/ netwo rks/ data/ ucinet/ ucida ta. htm# zacha ry 2672 http:// konect. cc/ netwo rks/ moreno_ oz/ 58 376 http:// konect. cc/ netwo rks/ ucida ta- gama/ http:// konect. cc/ netwo rks/ moreno_ seven th/ 1098 http:// konect. cc/ netwo rks/ moreno_ innov ation/ 366 354 322 12969 12534 220 191 90 61 79 58 http:// konect. cc/ netwo rks/ moreno_ highs chool/ http:// konect. cc/ netwo rks/ moreno_ vdb/ http:// vlado wiki. fmf. uni- lj. si/ doku. php? id samps on pajek: data: esna3: = http:// konect. cc/ netwo rks/ moreno_ health/ http:// konect. cc/ netwo rks/ petst er- hamst er- friend/ http:// vlado. fmf. uni- lj. si/ pub/ netwo rks/ doc/ ECPR/ assign. 1/ as1. net http:// vlado. fmf. uni- lj. si/ pub/ netwo rks/ doc/ ECPR/ assign. 1/ as2. net http:// vlado. fmf. uni- lj. si/ pub/ netwo rks/ doc/ ECPR/ assign. 1/ as5. net http:// vlado. fmf. uni- lj. si/ pub/ netwo rks/ doc/ ECPR/ assign. 1/ as7. net http:// vlado. fmf. uni- lj. si/ pub/ netwo rks/ doc/ ECPR/ assign. 1/ as8. net http:// vlado. fmf. uni- lj. si/ pub/ netwo rks/ doc/ ECPR/ assign. 1/ as9. net 124.5 https:// netwo rkrep osito ry. com/ soc- firm- hi- tech. php 2.9K 2.1K 51127 27806 https:// netwo rkrep osito ry. com/ soc- wiki- Vote. php https:// netwo rkrep osito ry. com/ fb- pages- food. php http:// konect. cc/ netwo rks/ advog ato/ https:// snap. stanf ord. edu/ data/ feath er- lastfm- social. html 222887 https:// snap. stanf ord. edu/ data/ gemsec- Deezer. html 125826 https:// snap. stanf ord. edu/ data/ gemsec- Deezer. html 218 103 83 65 http:// vlado. fmf. uni- lj. si/ pub/ netwo rks/ doc/ ECPR/ assign. 1/ as4. net http:// vlado. fmf. uni- lj. si/ pub/ netwo rks/ doc/ ECPR/ assign. 1/ as6. net http:// vlado. fmf. uni- lj. si/ pub/ netwo rks/ doc/ ECPR/ assign. 1/ as11. net http:// vlado. fmf. uni- lj. si/ pub/ netwo rks/ doc/ ECPR/ assign. 1/ as12. net 78 http:// konect. cc/ netwo rks/ moreno_ taro/ 119 80 http:// vlado. fmf. uni- lj. si/ pub/ netwo rks/ doc/ ECPR/ assign. 1/ as3. net http:// vlado. fmf. uni- lj. si/ pub/ netwo rks/ doc/ ECPR/ assign. 1/ as10. net Boyd et al. Applied Network Science (2023) 8:15 Page 24 of 29 Table 2 (continued) Network data Network type & name Vertices Edges Citation Genealogical networks Genealogical network 1 Genealogical network 2 Genealogical network 3 Genealogical network 4 Genealogical network 5 Genealogical network 6 Genealogical network 7 Genealogical network 8 Genealogical network 9 Genealogical network 10 Genealogical network 11 Genealogical network 12 310 303 371 795 636 782 128 439 244 410 337 216 Genealogical network 13 77 322 537 718 1387 1151 1366 202 626 481 746 572 378 134 Genealogical network 14 815 1582 Genealogical network 15 20 Genealogical network 16 Genealogical network 17 Genealogical network 18 219 17 168 Genealogical network 19 64 28 371 24 221 109 Genealogical network 20 1423 3211 Genealogical network 21 645 1097 Genealogical network 22 4463 8416 Genealogical network 23 48 86 Genealogical network 24 104 172 Genealogical network 25 1263 2021 Genealogical network 26 80 132 Genealogical network 27 1269 2395 https:// www. kinso urces. net/ kidar ep/ datas et- 209- mowan jum- kalum buru. xhtml https:// www. kinso urces. net/ kidar ep/ datas et-2- mbuti- villa ge- 1957- af03. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 58- ojibwa- 1930- nd07. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 150- achuar- pasta za. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 92- chenc hu- 1940- as02. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 28- trio- 1960s. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 23- shosh one- 1880- nd11. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 70- genes is. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 66- waimi ri- atroa ri. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 240- kodiak. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 51- wilca nia. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 22- ainu- 1880- as01. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 69- slavey- 1911- nd12. xhtml https:// www. kinso urces. net/ kidar ep/ datas et-7- pakaa- nova. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 38- wanin dilja ugwa- 1948- au06. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 171- suya. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 31- family. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 14- labra dor- inuit- 1776- nu02. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 91- takam iut- 1927- 64- nu03. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 258- todas. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 65- iglul igmiut- 1961- nu07. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 115- charl evoix. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 41- vedda- 1905- as04. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 71- iglul igmiut- 1960- 61- nu08. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 223- sambu ru. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 10- apache- 1932- nd01. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 24- ayd- nl- yoruk- 2005. xhtml Genealogical network 28 299 532 https:// www. kinso urces. net/ kidar ep/ datas et- 13- tory. xhtml Boyd et al. Applied Network Science (2023) 8:15 Page 25 of 29 Table 2 (continued) Network data Network type & name Vertices Edges Citation Genealogical network 29 19 Genealogical network 30 Genealogical network 31 399 377 30 592 712 Genealogical network 32 1263 2021 Genealogical network 33 118 Genealogical network 34 98 Genealogical network 35 479 192 161 830 Genealogical network 36 1695 3206 Genealogical network 37 Genealogical network 38 Genealogical network 39 Genealogical network 40 Genealogical network 41 256 798 738 525 619 441 1416 1212 855 1224 Genealogical network 42 3008 6074 Genealogical network 43 Genealogical network 44 Genealogical network 45 Genealogical network 46 Genealogical network 47 Genealogical network 48 Genealogical network 49 Genealogical network 50 Genealogical network 51 Genealogical network 52 Genealogical network 53 Genealogical network 54 Genealogical network 55 Genealogical network 56 Genealogical network 57 278 105 240 4178 216 147 277 330 35 48 105 116 116 657 659 464 172 395 7351 286 242 516 622 53 76 245 220 176 1166 1288 https:// www. kinso urces. net/ kidar ep/ datas et- 21- ngata tjara- 1966- au04. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 204- dogon- konso gu- donyu. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 49- alyaw arra- 1971- au01. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 223- sambu ru. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 39- eyak- 1890. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 75- nunam iut- 1885- nu11. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 19- ojibwa- 1949- nd08. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 103- tikuna- arara. xhtml https:// github. com/ Abiga ilJ32/ The- persi stent- homol ogy- of- genea logic al- netwo rks https:// www. kinso urces. net/ kidar ep/ datas et- 229- nucoo rilma- tingha. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 32- yaral di. xhtml https:// github. com/ Abiga ilJ32/ The- persi stent- homol ogy- of- genea logic al- netwo rks https:// www. kinso urces. net/ kidar ep/ datas et- 251- nuniv ak. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 80- torsh an. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 62- dogrib- 1911- 25- 59- nd04. xhtml https:// www. kinso urces. net/ kidar ep/ datas et-5- konka ma- 1931- 44- 51- eu02. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 158- tikar. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 45- obidos. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 254- port- keats. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 78- pul- eliya- 1954- simpl er- versi on. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 213- sarmi. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 73- parak ana. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 81- gunda ngborn- 1948- au02. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 84- hare- 1956- nd05. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 87- arara. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 89- nunam iut- 1960- nu13. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 226- jie. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 27- nyung ar. xhtml https:// www. kinso urces. net/ kidar ep/ datas et-3- anuta- 1972. xhtmlj Boyd et al. Applied Network Science (2023) 8:15 Page 26 of 29 Table 2 (continued) Genealogical network 58 Genealogical network 59 112 218 Genealogical network 60 90 182 353 119 https:// www. kinso urces. net/ kidar ep/ datas et- 15- oodna datta. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 17- laini ovouma- 1952- eu03. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 12- miwuyt- 1967- au03. xhtml Genealogical network 61 289 477 Genealogical network 62 1463 1969 Genealogical network 63 4109 6517 Genealogical network 64 Genealogical network 65 Genealogical network 66 Genealogical network 67 Genealogical network 68 Genealogical network 69 29 40 294 502 83 95 48 59 441 786 126 157 Genealogical network 70 2588 5651 Genealogical network 71 88 144 Genealogical network 72 1513 2217 Genealogical network 73 3014 5454 Genealogical network 74 139 201 Genealogical network 75 5016 10719 Genealogical network 76 Genealogical network 77 Genealogical network 78 Genealogical network 79 Genealogical network 80 Genealogical network 81 Genealogical network 82 Genealogical network 83 Genealogical network 84 125 272 378 926 706 435 128 169 178 202 445 609 1951 1177 672 114 275 274 Genealogical network 85 Genealogical network 86 87 2049 111 4159 Genealogical network 87 868 980 https:// www. kinso urces. net/ kidar ep/ datas et-9- konka ma- 1951- eu01. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 306- nobles- ile- de- france- 1000- 1440. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 287- duu- rea. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 46- hatfi elds- and- mccoys. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 33- angma gsalik- 1884- nu01. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 18- tikop ia- 1930. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 34- netsi lik- 1922- nu09. xhtml https:// www. kinso urces. net/ kidar ep/ datas et-8- semang- 1924- 50- as03. xhtml https:// www. kinso urces. net/ kidar ep/ datas et-4- shosh one- 1860- nd10. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 61- kelku mmer. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 77- apache- 1935- nd02. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 90- omaha- 1880. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 128- ammon ni. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 79- paiute- 1880- nd09. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 249- baruya. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 242- tling it. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 36- copper- 1922- nu10. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 52- apache- 1936- nd03. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 68- surui. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 60- mbuti- for- est- 1957- af02. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 64- melom bo. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 164- kaing ang. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 11- top- of- the- mount ain. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 37- iglul igmiut- 1921- nu05. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 216- tiwi. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 35- chuuk ese- 1947- 1940. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 20- saudi- royal- genea logy. xhtml Boyd et al. Applied Network Science (2023) 8:15 Page 27 of 29 Table 2 (continued) Genealogical network 88 2821 5079 Genealogical network 89 Genealogical network 90 Genealogical network 91 454 304 367 980 472 671 Genealogical network 92 3151 4289 Genealogical network 93 2975 5107 Genealogical network 94 Genealogical network 95 585 334 1249 530 Genealogical network 96 9595 14988 Genealogical network 97 28586 51446 Genealogical network 98 18645 32439 Genealogical network 99 8809 15643 Atypical Genealogical Networks Atyp. Gen. Network 1 Atyp. Gen. Network 2 429 2477 705 4015 https:// www. kinso urces. net/ kidar ep/ datas et- 30- manus- 1929. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 74- arawe te. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 42- nunam iut- tareu miut- 1900- nu12. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 48- wanin dilja ugwa- 1941- au05. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 54- feist ritz- am- gael- 1990. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 159- cocama- cocam illa. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 44- torres- strait. xhtml https:// www. kinso urces. net/ kidar ep/ datas et-6- iglul igmiut- 1949- nu06. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 93- sainte- cathe rine. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 76- san- marino. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 307- bwa- slam- biogs urvey. xhtml https:// www. kinso urces. net/ kidar ep/ datas et- 194- kel- owey. xhtml (created using FamilySearch.org) https:// www. kinso urces. net/ kidar ep/ datas et- 56- us- presi dents. xhtml networks is the relationship between various supporting documents/metadata and the links that are discoverable through them. For example, one can consider optimal document collec- tion strategies with a limited budget or document collection that is fair in terms of capturing minority information, which is often underrepresented. Appendix Here we indicate both the genealogical and social networks used in our persistent homol- ogy computations (see "Results"). We distinguish the datasets by network type: Friendship/ Acquaintance, Social Media, Collaboration/Business, Disease Transmission, Information Shar- ing, Genealogical, and Atypical Genealogical networks. We also provide the network name, number of vertices and edges in the network, and a citation where the network can be found. Also, a special thanks to Kolton Baldwin for help with numerical simulations on this paper. Acknowledgements We acknowledge helpful conversations with Joseph Price and the FamilySearch Engineering Research team. We also acknowledge Kolton Baldwin for helping to improve our code and simulations. Author contributions Designed the experiments: ZB, NC, BW, RW. Performed the experiments: RF, RW. Wrote the paper: ZB, NC, TG, AJ, RS, BW, RW. All authors read and approved the final manuscript. Funding ZB, BW, and AJ, were supported by a BYU CPMS CHIRP grant. ZB was additionally supported by NFS award #2137511 and Army Research Office grant #W911NF-18-1-0244, and the James S. McDonnell Foundation 21st Century Science Initia- tive-Complex Systems Scholar Award grant #2200203. BW was additionally supported by the Simons Foundation grant #714015. The views and conclusions contained in this document are those of the authors and should not be interpreted Boyd et al. Applied Network Science (2023) 8:15 Page 28 of 29 as representing the official policies, either expressed or implied, of the Army Research Office or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein. Availability of data and materials Links to the datasets generated and/or analysed during the current study can be found in Table 2. Code to replicate and extend this work can be found at https:// github. com/ Abiga ilJ32/ The- persi stent- homol ogy- of- genea logic al- netwo rks. Declarations Competing interests The authors declare that they have no competing interests. Received: 13 July 2022 Accepted: 27 January 2023 References Aktas ME, Akbas E, Fatmaoui AE (2019) Persistence homology of networks: methods and applications. Appl Netw Sci 4:61. https:// doi. org/ 10. 1007/ s41109- 019- 0179-3 Arafat NA, Basu D, Bressan S (2020) ǫ-net Induced Lazy Witness Complexes on Graphs, Preprint arXiv: https:// arxiv. org/ abs/ 2009. 13071 Bloothooft G, Christen P, Mandemakers K, Schraagen M (2015) Population Reconstruction. Springer, Cham Blumberg AJ, Lesnick M (2020) Stability of 2-parameter persistent homology, Preprint arXiv: https:// arxiv. org/ abs/ 2010. 09628 Carstens CJ, Horadam KJ (2013) Persistent homology of collaboration networks, In: Mathematical problems in engineer- ing, vol 2013, Article ID 815035, p 7 Chang JT (1999) Recent common ancestors of all present-day individuals. Adv App Prob 31:1002–1026. https:// doi. org/ 10. 1239/ aap/ 10299 55256 Chazal F, Guibas LJ, Oudot SY, Skraba P (2013) Persistence-based clustering in Riemannian manifolds. J ACM (JACM) 60(6):41 Cohen-Steiner D, Edelsbrunner H, Harer J (2007) Stability of persistence diagrams. Discret Comput Geom 37:103–120 Duman AN, Pirim H (2018) Gene coexpression network comparison via persistent homology. Int J Genom 2018 Flores R (2021) Modeling a human family network. https:// schol arsar chive. byu. edu/ etd/ 9357/ Greenwood J, Guner N, Kocharkov G, Santos C (2014) Marry your like: assortative mating and income inequality. Am Econ Rev 104:348–353 Hage P, Harary F (1983) Structural models in anthropology. Cambridge University Press, Cambridge Hamberger K, Houseman M, White DR (2011) Kinship, class, and community. In: Scott JP, Carrington PJ (eds) The SAGE handbook of social network analysis. Sage Publications Ltd., Thousand Oaks, pp 129–147 Hatcher A (2002) Algebraic topology. Cambridge University Press, Cambridge Hey J, Machado CA (2003) The study of structured populations-new hope for a difficult and divided science. Nat Rev Genet 4:535–543 Horak D et al (2009) Persistent Homology of Complex Networks. J Stat Mech https:// www. kinso urces. net/ brows er/ datas ets. xhtml. Accessed 21 Jun 2022 http:// konect. cc/ netwo rks/. Accessed 10 Jun 2022 http:// snap. stanf ord. edu/ data/ index. html# socne ts. Accessed Aug 2020 http:// netwo rkrep osito ry. com/ soc. php Accessed Aug 2020 http://vladowiki.fmf.uni-lj.si/doku.php?id Kannan H, Saucan E, Roy I et al (2019) Persistent homology of unweighted complex networks via discrete Morse theory. pajek:data:pajek:index. Accessed Aug 2020 = Sci Rep 9(13817):1–18 Kaplanis J, Gordon A, Shor T, Weissbrod O, Geiger D, Wahl M, Gershovits M, Markus B, Sheikh M, Gymrek M, Bhatia G, Mac- Arthur DG, Price AL, Erlich Y (2018) Quantitative analysis of population-scale family trees with millions of relatives, American Association for the Advancement of Science. http:// scien ce. scien cemag. org/ conte nt/ early/ 2018/ 02/ 28/ scien ce. aam93 09 Lee H, Kang H, Chung MK, Kim B, Lee DS (2012) Persistent brain network homology from the perspective of dendrogram. IEEE Trans Med Imaging 31(12):2267–2277 Malmi E, Gionis A, Solin A (2018) Computationally inferred genealogical networks uncover long-term trends in assorta- tive mating. In: Proceedings of the 2018 world wide web conference WWW 2018, Lyon, France, pp 883–892. http:// doi. acm. org/ 10. 1145/ 31788 76. 31861 36 Mattia B, Adriano B, Barbara DF (2016) Towards a Topological Fingerprint of Music. In: Proceedings of the 6th international workshop on computational topology in image context, vol 9667, pp 88–100 Newman MEJ (2006) Modularity and community structure in networks. Proc Natl Acad Sci USA 103(23):8577–8582 Otter N, Porter MA, Tillmann U, et al (2017) A roadmap for the computation of persistent homology, EPJ Data Science Petri G, Scolamiero M, Donato I, Vaccarino F (2013) Networks and cycles: a persistent homology approach to complex networks, In: Gilbert T, Kirkilionis M, Nicolis G (eds) Proceedings of the European conference on complex systems 2012, Springer proceedings in complexity. Springer, Cham. https:// doi. org/ 10. 1007/ 978-3- 319- 00395-5_ 15 Residence hall social network data http:// konect. cc/ netwo rks/ moreno_ oz/. Accessed 20 Jan 2022 Ripser Python package. https:// anaco nda. org/ conda- forge/ ripser. Accessed 4 Oct 2021 Boyd et al. Applied Network Science (2023) 8:15 Page 29 of 29 Robins V, Saadatfar M, Delgado-Friedrichs O, Sheppard AP (2016) Percolating length scales from topological persistence analysis of micro-CT images of porous materials. Water Resour Res 52(1):315–329 Rohde DLT, Olson S, Chang JT (2004) Modelling the recent common ancestry of all living humans. Nature 431:562–566 Sintos S, Tsaparas P (2014) Using strong triadic closure to characterize ties in social networks. IN: Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pp 1466–1475 Vandaele R, De Bie T, Saeys Y (2018) Local topological data analysis to uncover the global structure of data approach- ing graph-structured topologies. In: Joint European conference on machine learning and knowledge discovery in databases. Springer, pp 19–36 Zomorodian A, Carlsson G (2005) Computing persistent homology. Discret Comput Geom 33:249–274 Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
10.1186_s12915-021-01024-1
Avin et al. BMC Biology (2021) 19:97 https://doi.org/10.1186/s12915-021-01024-1 R E S E A R C H A R T I C L E Open Access An agent-based model clarifies the importance of functional and developmental integration in shaping brain evolution Shahar Avin1, Adrian Currie2† and Stephen H. Montgomery3*† Abstract Background: Vertebrate brain structure is characterised not only by relative consistency in scaling between components, but also by many examples of divergence from these general trends.. Alternative hypotheses explain these patterns by emphasising either ‘external’ processes, such as coordinated or divergent selection, or ‘internal’ processes, like developmental coupling among brain regions. Although these hypotheses are not mutually exclusive, there is little agreement over their relative importance across time or how that importance may vary across evolutionary contexts. Results: We introduce an agent-based model to simulate brain evolution in a ‘bare-bones’ system and examine dependencies between variables shaping brain evolution. We show that ‘concerted’ patterns of brain evolution do not, in themselves, provide evidence for developmental coupling, despite these terms often being treated as synonymous in the literature. Instead, concerted evolution can reflect either functional or developmental integration. Our model further allows us to clarify conditions under which such developmental coupling, or uncoupling, is potentially adaptive, revealing support for the maintenance of both mechanisms in neural evolution. Critically, we illustrate how the probability of deviation from concerted evolution depends on the cost/benefit ratio of neural tissue, which increases when overall brain size is itself under constraint. Conclusions: We conclude that both developmentally coupled and uncoupled brain architectures can provide adaptive mechanisms, depending on the distribution of selection across brain structures, life history and costs of neural tissue. However, when constraints also act on overall brain size, heterogeneity in selection across brain structures will favour region specific, or mosaic, evolution. Regardless, the respective advantages of developmentally coupled and uncoupled brain architectures mean that both may persist in fluctuating environments. This implies that developmental coupling is unlikely to be a persistent constraint, but could evolve as an adaptive outcome to selection to maintain functional integration. Keywords: Brain structure, Brain size, Constraint, Concerted evolution, Mosaic evolution, Neuro-evo-devo * Correspondence: s.montgomery@bristol.ac.uk †Adrian Currie and Stephen H. Montgomery contributed equally to this work. 3School of Biological Sciences, University of Bristol, Bristol, UK Full list of author information is available at the end of the article © The Author(s). 2021 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data. Avin et al. BMC Biology (2021) 19:97 Page 2 of 18 Background How are macro-evolutionary patterns in vertebrate brain structure best characterised, and what processes drive those patterns? Answering such questions requires un- derstanding how developmental mechanisms or archi- tectural constraints, as well as selection acting on neural traits, together shape and support behavioural and cog- nitive evolution. Debates over these conflicting pressures on variation have dominated vertebrate evolutionary neurobiology for decades, with no unified theoretical framework in sight. At the heart of this debate are two views of vertebrate brain evolution which, at their most polarised, make seemingly opposing predictions while both appearing logically sound and empirically supported. Under one hypothesis, brain components are developmentally coupled such that the size of each component is largely determined by common developmental mechanisms, such as the schedule, timing and duration of neurogen- esis [1–3]. This would lead to the majority of brain structures evolving in a ‘concerted’ manner, with the size of separate components being closely predicted by over- all brain size [1–4]. Initially, this coupling was discussed as a potential evolutionary constraint, associated with ‘spandrels’ whereby late developing brain regions such as the neocortex, may have expanded disproportionately as a by-product of architectural constraints, before later be- ing co-opted functionally [2]. Proponents of this hypoth- esis now largely argue that developmental coupling is a mechanism that evolves in response to selection favour- ing conservative scaling, and is, as such, a potential adaptive mechanism rather than a constraint per se [1, 3]. However, the view that concertedness in itself indi- cates developmental constraint remains widespread in the literature (e.g. [4–12]). A contrasting hypothesis instead argues that vari- ation in brain components is largely developmen- tally independent of both other brain structures, and of total brain size, allowing them to respond to targeted selection pressures in a ‘mosaic’ way [13– 17]. Mosaic evolution is often discussed as facilitat- ing neural adaptations, reflected in non-allometric changes in brain structure, but it also invokes stabi- lising selection to otherwise maintain scaling rela- tionships functionally In es- interdependent brain components [14, 18]. sence, the mosaic model favours ‘external’ explana- tions that emphasise the role of both divergent and coordinated selection in driving both independent evolution and co-variation of brain phenotypic structures, while ‘concerted’ ‘in- ternal’ mechanisms which emphasise the role of de- velopmental coupling as a route to maintaining scaling relationships [19]. co-evolving, theorists between stress Perhaps confusingly, both hypotheses have at times been supported by analyses of the same volumetric data (e.g. [2, 14, 20]). Proponents of the ‘concerted’ view of brain evolution pointed to consistent allometric scaling between brain components and total brain size as evi- dence of strong developmental integration across brain structures [1, 2, 4]. Proponents of the ‘mosaic’ model in- stead pointed towards co-evolution between brain com- ponents that is independent of total brain size, and evidence for ‘grade shifts’ that indicate deviations in scal- ing between taxonomic groups, as evidence that brain components are caught between distinct selection pres- sures, and constrained from functional interdependence [14]. Distinctions between these hypotheses have be- come more nuanced, with the concerted hypothesis in- corporating brain, accommodating some mosaic change [3, 21]. But, re- gardless, universally satisfactory tests of the generality of these hypotheses have proven elusive, there is frequent confusion in the literature between the distinction be- tween patterns and mechanisms of brain evolution, and little data exists on when one mechanism may be favoured over another. restructuring periodic the of There are two key reasons for this deadlock. First, pro- ponents of concerted and mosaic models are engaged in a ‘relative significance debate’ [22]. Both sides agree that brain evolution exhibits features associated with both concerted and mosaic evolution, but disagree on which pattern dominates across evolutionary time, and why (see for example, [1], p. 299). Relative significance makes hypothesis testing difficult. Neither hypothesis is subject to critical tests, as both accommodate—and even ex- pect—different degrees of departure from the ‘norm’. Al- ternative views of brain evolution therefore run the risk of being too indeterminate for definitive testing. Second, tests of these hypotheses are underdetermined by available evidence. Although proponents of both mechanisms can point to support from developmental data (reviewed in [18, 23])—showing, for example, how concerted patterns of brain evolution can be produced by changes in the regulation of neural progenitor cell proliferation [24–27], or how changes in the allocation, rate or duration of cell division among the cell popula- tions that lead to specific regions can produce mosaic changes in brain structure [28–31]—these data are nat- urally less readily available than volumetric data, and therefore, tests of generalisation are limited. As such, empirical support for concerted or mosaic evolution is most often drawn from comparative analyses of volu- metric brain data. These data reflect the outcome of the interaction between competing adaptive and constrain- ing factors and do not, in themselves, provide evidence of the developmental mechanisms involved [32, 33]. This is a critical point, as ‘concertedness’ has frequently Avin et al. BMC Biology (2021) 19:97 Page 3 of 18 also predicts co-variation become a byword for developmental constraint (e.g. [4– 12, 34]), potentially biasing the interpretation and pres- entation of many studies. However, the mosaic brain hy- between pothesis interdependent brain regions. If the brain is viewed as a interdependent networks, these functional network of constraints could produce consistent scaling relation- ships across brain components — i.e. concerted evolu- tion — without invoking developmental coupling [18]. As inferred through classic evolutionary theory [35], merely recognising a concerted pattern is insufficient evidence to assess alternative mechanisms, or to support the predominance of either hypothesis. If patterns of phenotypic variation alone are unsuitable for identifying the mechanisms that underpin allometric scaling, what evidence could? As noted by previous au- thors, ‘it is not the phenotypic correlation that matters, so much as the genetic correlation’ [36]. Although brain morphology can be highly plastic, responding to effects of the physical or social environment, which may alter the appearance of how brain structures scale within spe- cies (e.g. [37, 38]), the majority of comparative studies interrogating patterns of brain evolution implicitly as- sume these effects are small relative to heritable vari- ation. Quantitative, intra-specific genetic studies provide a test of this assumption and of the relative strength of genetic correlations between brain size and structure. Several recent quantitative genetics studies have found evidence of substantial genetic independence between brain components [5, 8, 39], a central prediction of the mosaic brain hypothesis (reviewed in [18]). However, these studies typically concern standing genetic variation within populations. The developmental coupling hypoth- esis can accommodate this evidence if much of this gen- etic variation is mildly deleterious and is maintained in the population due to negative selection being weaker than drift, for example. If this was the case, selection for changes in brain structure or brain size may more fre- quently act on de novo mutations that are distinct in their developmental effects compared to standing gen- etic variation, and which are generally purged from the population during times of evolutionary stasis in brain structure, perhaps because they have larger fitness ef- fects. If this were the case, intra-specific studies might not reflect the genetic architecture favoured by selection over evolutionary timescales. Currently, we have insuffi- cient evidence either way. At a comparative level, some authors also argue that both concerted and mosaic pat- terns are observed in their data, with pairs of structures evolving in a coordinated, or concerted, manner, poten- tially supported by direct mechanisms linking their de- velopment, while others evolve independently [7, 40– 42]. This would invoke complex patterns of develop- integration that occur after the major brain mental the limited attempts divisions are established [40], rather than the more glo- bal developmental integration suggested by previous au- thors, but this using molecular data do not currently support this idea [43]. Hence, neither phenotypic data nor currently available genetic data are sufficient to unite views on the relative importance of developmental and functional coupling, constraint and adaptive lability in the evolution of brain structure. to test When faced with relative significance and empirical underdetermination, simple mathematical models can help realise basic causal dynamics in a ‘bare-bones’ sys- tem and are a way of examining the dependencies be- tween variables that are thought to be important. We can envision ‘bare-bones’ models as tools that serve to make explicit the assumptions and reasoning involved in otherwise linguistic arguments, sometimes revealing pre- viously hidden assumptions [44, 45]. While they lack empirical data, and must therefore be treated with care, they can be critical for informing future empirical stud- ies and aiding the interpretation of existing literature [46]. This is particularly true for relevant significance de- bates which lack a straightforward way to weigh the im- portance of multiple mechanisms in different contexts using empirical data. Here, a modelling approach can be used to explore how key variables behave, which can dovetail with existing experimental or comparative stud- ies, or prompt new ones. This approach has recently been applied to debates over the socio-ecological selec- tion pressures shaping brain size [47–50], providing a new approach to the field of evolutionary neurobiology. Here, we introduce an agent-based model of brain struc- ture that allows us to explore the interactions between fitness and constraints derived from selection, develop- ment and function (summarised in Additional file 1: Fig- ure S1). In particular, our model allows us to formalise several verbal arguments over the role of developmental coupling in brain evolution; specifically, we ask: 1. Do functional dependencies produce concerted patterns of evolution as well as developmental coupling (e.g. [32, 33])? 2. Can both mechanisms be adaptive (e.g. [1])? 3. Do the costs of neural tissue select against 4. concertedness when selection acts on a specific brain component (e.g. [51])? Is developmental integration evolutionarily labile (e.g. [52]), and do functional dependencies select for developmental coupling (e.g. [32, 52])? 5. Does stabilising selection or constraint on brain size lead to increased frequencies of mosaic evolution (e.g. [51])? Our model allows us to explore these previously verbal arguments and interpretations of volumetric data. We Avin et al. BMC Biology (2021) 19:97 Page 4 of 18 demonstrate that this ‘bare-bones’ model helps clarify current debates surrounding the evolution of brain structure by capturing the basic evolutionary dynamics at play, and hope that it shifts these debates in a pro- ductive theoretical and empirical direction. is by caused can be correlated, Results Do functional dependencies produce concerted patterns of evolution as well as developmental coupling? By varying the degree of both developmental coupling (D) and functional interdependence (F), between brain components, our model suggests that patterns of ‘con- certed’ brain evolution, in which the size of brain com- ponents both developmental and functional coupling (Fig. 1, Add- itional file 1: Figure S2–4). Unsurprisingly, the probabil- ity of patterns of concerted evolution declines with decreasing D (t = 50.330, p < 0.001; Fig. 1d). However, there is a significant interaction between F and D (t = 28.730, p < 0.001) whereby, for a given value of D, where D < 1, the probability of concerted evolution increases with higher values of F (Fig. 1a–d), demonstrating that functional interdependence also promotes concerted- ness. Even when components are completely develop- mentally independent (D is set to 0), high values of F can favour comparatively low levels of mosaicism (Fig. 1c). Mutation size also has an effect on the out- come, with more mosaic brains evolving with larger mu- tation step sizes for a given D value (t = 133.650, p < 0.001; compare Figure S2 and S3). These results illus- trate that macro-evolutionary patterns of allometric scal- in ing consistent with concerted evolution are not, themselves, sufficient to distinguish between alternative mechanistic models of brain evolution. Can both mechanisms be adaptive? By competing alternative D values against one another, we found that both D (t = 21.358, p < 0.001) and F (t = 42.595, p < 0.001) affect survival in a particular environ- ment. At an intermediate benefit to cost ratio (B/C = 1.5), we found that the probability of success of a mo- saic, low D value population increases when F is low, while high values of F result in a higher probability of success for the concerted, high D value population (Fig. 2a–c; Additional file 1: Figure S6-S8). A ‘partially mosaic’ population (D = 0.5) is very rarely favoured (Fig. 2a–c; Additional file 1: Figure S6-S8). These com- parisons indicate that variation in selection across com- ponents alters the outcome of competition between populations with different levels of developmental coupling. Do the costs of neural tissue select against concertedness when selection acts on a specific brain component? In the preceding comparisons, the degree of mosaicism is also associated with the B/C ratio (t = 131.790, p < 0.001; Fig. 1d, Additional file 1: Figure S2), which inter- acts with D (t = − 90.970, p < 0.001) such that the degree of mosaicism tends to increase with B/C (Fig. 1d). We next repeated the simulations described in (ii) while varying the B/C ratio associated with additional brain tissue. The initial models were run with an average B/C ratio of 1.5 (Fig. 1, Additional file 1: Figure S6–8 s row) and were re-run with ratios of 2 and 0.5 (Fig. 3a–c, S6– 8, first and third rows), simulating low and high costs of brain tissue. This revealed that relative tissue costs have a major effect on the success of populations with differ- ent D values (t = 12.116, p < 0.001). B/C interacts with D (t = − 13.885, p < 0.001) such that, for a given combin- ation of F and D, a high B/C (=2) consistently increases Fig. 1 Evolution of ‘mosaicism’ under alternative conditions. a–c Each plot depicts the ‘degree of mosaicism’ (y-axis, defined as the natural log of the ratio between the largest brain component and the smallest brain component in each individual, averaged across the population) as a function of developmental coupling D (x-axis) at the end of a 100-generation simulation run, compiled over 1000 simulation runs, for a population of 100 individuals with an identical D, under different environments, defined by their functional coupling F, with an average benefit to cost ratio (B /C) of 1.5 and a mutation step size of 5%. Each data point is the outcome of one simulation run and the black bar indicates the mean of these runs. d Summary of the effects of varying F and B /C on the degree of mosaicism when D is 0. See Additional file 1: Figure S2-S4 for full results varying B /C and F, iteration numbers and mutation step size Avin et al. BMC Biology (2021) 19:97 Page 5 of 18 Fig. 2 Selected examples of competition between evolving populations with different D values. a–c Each plot depicts the frequency of that D value relative to the total population (y-axis) as a function of developmental coupling D (x-axis) at the end of a 100-generation simulation run, compiled over 1000 simulation runs, under alternative environments defined by their functional coupling F, with an average benefit to cost ratio (B /C) of 1.5 and a mutation step size of 5%. Populations are initialised such that there are 100 fully mosaic individuals (D = 0.0), 100 partially mosaic individuals (D = 0.5) and 100 concerted individuals (D = 1.0). Each data point is the outcome of one simulation run and the black bar indicates the mean of these runs. d–g Selected, representative, individual simulations showing the change in population frequencies over generations for a 5% mutation size and B /C = 1.5. Colours indicate the D value, where yellow is D = 0, where blue is D = 0.5 and where green is D = 1. These show a general pattern of smooth progression from the starting state of equal populations to one D value winning out (d, f), with only a minority of iterations showing signs of populations ‘swapping’ the lead (e, g). This consistency is expected under constant selection regimes. See Additional file 1: Figure S6-S8 for full results varying B /C and F, iteration numbers and mutation step size, and Additional file 1: Figure S9–11 for simulations in tailored environments that illustrate parameter effects the probability of success for the population with a high D value (Fig. 3a–c, Additional file 1: Figure S6-S8). In contrast, for a given combination of F and D, a moderate B/C (=1.5) consistently increases the probability of suc- cess for low D, mosaic models. However, when the B/C ratio is low (=0.5), the non-linear nature of the inter- action between C, F and D is revealed, such that con- low F certedness again becomes successful, even at values (Fig. 3d–f, Additional file 1: Figure S6-S8 third rows). In these competition experiments, mutation size had no effect on the outcome (t = 0.740, p = 0.459; com- pare Additional file 1: Figure S6 and S7). The non-linear success of low D values can be explained if developmen- tal coupling facilitates rapid decreases in all brain regions when costs of brain tissue exceed the benefit, allowing a quick escape from a costly phenotype, or when the fitness of the whole system is dominated by a single component, such that increases in total brain size approximate the fit- ness benefit of increasing specific components. To further illustrate these effects, we specified par- ticular parameter comparisons that show how B/C interacts with changes in selection regimes such that the most successful D value switches based on chan- ging the benefit-cost ratio (Fig. 4, Additional file 1: Figure S10-S12 rows) or variation of selection across components (Fig. 4, Additional file 1: Figure S10-S12 columns). In particular, we note that (i) simulations can maintain multiple populations with different D values where fitness is dominated by the contribu- tion of one component (Fig. 4a); (ii) shifts in the probability of a population of a mosaic, low D popu- lation being successful are otherwise associated with increased variation in the B value among compo- nents (Fig. 4bii); and (iii) the relative success of a mosaic, low D population increases when one or two components provide a net benefit to size and other component(s) provide a net cost (Fig. 4biii), while populations with high D values are favoured when Avin et al. BMC Biology (2021) 19:97 Page 6 of 18 Fig. 3 Selected examples of competition between evolving populations with different D values, showing the effects of varying the average benefit to cost ratio (B /C). a–c Each plot depicts the frequency of that D value relative to the total population (y-axis) as a function of developmental coupling D (x-axis) at the end of a 100-generation simulation run, compiled over 1000 simulation runs, with a mutation step size of 5%, and F = 0 to exemplify the effects of varying the B /C from 0 to 2. Populations are initialised such that there are 100 fully mosaic individuals (D = 0.0), 100 partially mosaic individuals (D = 0.5) and 100 concerted individuals (D = 1.0). Each data point is the outcome of one simulation run and the black bar indicates the mean of these runs. d, e Each plot depicts the average relative frequencies of D = 0 (yellow) and D = 1 (green) at the end of 1000 iterations of a 100-generation simulation, at three B /C ratios, across three F values representing low (d), moderate (e) and high (f) functional coupling. See Additional file 1: Figure S6-S8 for full results varying B /C and F, iteration numbers and mutation step size, and Additional file 1: Figure S9–11 for simulations in tailored environments that illustrate parameter effects all components provide either a net benefit or a net cost (Fig. 4bi). Is developmental integration evolutionarily labile, and do functional dependencies select for developmental coupling? We examined whether populations with alternative D values persist when the selection regime is temporally variable (due to randomised, independent changes in B, C and F). Under initial conditions, where offspring num- ber was set to 1 and maximum age was set to 3, both D = 1 and D = 0 populations persist over 150 generations with roughly equal probabilities (Fig. 5a). Plotting the frequency of D values from individual simulations shows that the success of each population can fluctuate over time (Fig. 5e–g), with multiple populations persisting, on average, file 1: Figure S13A). This is substantially more than is found in simu- lations with fixed environments (Additional file 1: Figure for 62 generations (Additional S5) and is also reflected in the high average generation at extinction for each D value (D = 0, generation 91; D = 0.5, generation 57; D = 1, generation 98; Add- itional file 1: Figure S13). We subsequently varied max- imum age and offspring number to explore how ‘slow’ (long lives, few offspring) or ‘fast’ (short lives, many off- spring) life histories buffer the effects of environmental heterogeneity. This revealed that the main effect cap- tured in the model was that of offspring number inter- acting with D (t = 43,622, p < 0.001), with higher offspring numbers increasing the probability of success of the concerted, high D value populations (Fig. 5a, c, d; Additional file 1: Figure S14-S16 columns). Increasing maximum lifespan had a significant but smaller effect in the opposite direction (t = − 4.349, p < 0.001; Fig. 5a, b, d; Additional file 1: Figure S14-S16 rows). Altering the amplitude of fluctuations in the environment also has a subtle effect on the probability of success of competing populations with different D values (t = − 8.986, p < Avin et al. BMC Biology (2021) 19:97 Page 7 of 18 Fig. 4 Selected examples of competition between evolving populations with different D values, showing the average frequency of each population during the first 50 generations of 1000 simulated runs, with ‘hand-crafted’ environments depicted in the figure to illustrate specific situations of interest. Each plot depicts the average relative size of the 3 components within the artificial brain, with D values colour coded (D = 0 in gold, D = 0.5 in blue, D = 1.0 in green). a When the benefits of each components are very strongly skewed, both D = 0 and D = 1 values can persist as each rapidly adjusts to approximate the optimal condition. b With moderate levels of variation in B across components, a D = 1 value spreads in the population (i); with increased skew in B across components, the probability of spreading switches to D = 0 (ii) unless skew is extreme in which case D = 1 again becomes successful (a). Shifts in the probability of success for D = 1 are also associated with components having opposite signs in the B /C difference across components (ii) 0.001), with more extreme conditions slightly increasing the success of concertedness (Additional file 1: Figure S18). In these competition experiments, mutation size again had no effect on the outcome (t = 1.000, p = 1.000; compare Additional file 1: Figure S14 and S15). Does stabilising selection or constraint on brain size lead to increased frequencies of mosaic evolution? Imposing upper and lower bounds on the total size of the system has notable effects on the probability of mo- saic and concerted evolution. Under these conditions, we again see that patterns of ‘concerted’ brain evolution can be caused by both developmental and functional coupling (Fig. 6a–c; Additional file 1: Figure S20). Mo- saicism is more likely under all scenarios where D < 1 and F < 1 (t = 96.66, p < 0.001; Fig. 6d; Additional file 1: Figure S20). However, for a given D value, the degree of mosaicism is less impacted by variation in B/C than was observed in the boundless base model (Fig. 6d; Add- itional file 1: Figure S21). This can be explained by the dynamic nature of C when imposing upper and lower bounds on brain size. Regardless of the starting value, across iterations, B/C will tend to converge as brain size approaches its limits, which tends to happen well within the 100 generations of the simulations (Additional file 1: Figure S13). As a result, in some simulations, D = 1 pop- the ulations in frequency until increase initially population approaches the upper/lower boundary when the increasingly upscaled C results in the D = 0 popula- tion rising in frequency and becoming dominant (Add- itional file 1: Figure S22). In a competitive scenario, the relative frequency at which brains with low D values are favoured over brains with high D values also increases across the majority of parameter combinations (t = 36.88, p < 0.001; Fig. 6f–k), with a high D population be- ing favoured in the majority of runs when F = 1, or for F = 0.5 when B/C is 0.5 or 2 (Additional file 1: Figure S23). Similar patterns are found in simulations across fluctuating environments with the frequency of D = 0 in- dividuals in the final populations increasing across all runs (t = 24.59, p < 0.001; Fig. 6f–k). Discussion The results of our agent-based simulations have several implications for studies using comparative volumetric data to interrogate mechanistic hypotheses of how brains ‘con- evolve. First, they demonstrate that patterns of certed’ evolution (consistent allometric scaling of all brain components time) should not be taken as evidence for developmental coup- ling, as concertedness can also evolve due to multiple mechanisms, including high functional interdependence in the total absence of developmental coupling. Second, depending on the context, both concerted and mosaic across macro-evolutionary Avin et al. BMC Biology (2021) 19:97 Page 8 of 18 Fig. 5 Selected examples of competition between populations with different D values in a randomly varying environment. a–c Each plot depicts the ratio of each population to the total population (y-axis) as a function of developmental coupling D (x-axis) at the end of a 150-generations, compiled over 1000 simulation iterations; every 2 generations, the environment was replaced, using a uniform random distribution for cost [0.5, 5], max average benefit [0.51, 105] and F functional coupling [0, 1]. Populations are initialized such that there are 100 fully mosaic individuals (D = 0.0), 100 partially mosaic individuals (D = 0.5) and 100 concerted individuals (D = 1.0). Simulations were performed varying two life history conditions: maximum lifespan, or the number of generations an individual persists alive, and offspring number, which are produced once every generation. Three combinations of lifespan and offspring are shown, for full comparisons see S13. d A summary of the ratio of concerted individuals (D = 1) to mosaic individuals (D = 0) at the end of each iteration of the 150 generation simulation, showing the effects of maximum lifespan and offspring number. e–h Selected, representative, individual simulations showing fluctuations in D value frequencies over generations. Colours indicate the D value, or D value, where yellow is D = 0, where blue is D = 0.5 and where green is D = 1. See Additional file 1: Figure S13- S16 for full results varying B /C and F, iteration numbers and mutation step size, and Additional file 1: S17 for effects of increasing the size of environmental fluctuations evolution can be adaptive, and the most probable route to adaptive brain evolution is strongly influenced by whether the change in selection regime is skewed to- wards one brain component or is evenly distributed across the whole brain. Third, probabilities of concerted or mosaic evolution are also dependent on what the relative costs and benefits of increased investment in brain tissue are. Fourth, our model shows that hetero- geneity in selection regimes across time can result in both mechanisms being maintained in a population. Fi- nally, we demonstrate that when upper and lower bounds are placed on brain size, the probability of mo- saic evolution increases under most scenarios. Given these results, pluralism is a reasonable settlement. How- ever, our results suggest ways of going beyond the ap- parent deadlock of empirical underdetermination and relative significance, as discussed in the introduction. It does this by opening and providing an initial exploration of new questions: Under what conditions do develop- mental coupling and uncoupling succeed? What causes switches between mosaic and concerted modes of evolu- tion? And how can we empirically distinguish them? Our results further stress that mosaicism and concerted evolution are not competing models of brain evolution, but instead are reflections of evolutionary mechanisms that are jointly responsible for adaptive patterns of neural evolution. Concertedness is a phenotypic pattern devoid of mechanistic information Our simulations clearly show that concerted brain evolu- tion can occur under any level of developmental coup- ling, or D value (Fig. 1). This is consistent with classic quantitative genetics frameworks, which show that cor- related evolution does not depend on strong genetic cor- relations among traits [35]. However, despite early Avin et al. BMC Biology (2021) 19:97 Page 9 of 18 Fig. 6. (See legend on next page.) Avin et al. BMC Biology (2021) 19:97 Page 10 of 18 (See figure on previous page.) Fig. 6. Mosaic and concerted evolution when brain size has upper and lower limits. a–c Evolution of ‘mosaicism’ under alternative conditions. a–c Each plot depicts the ‘degree of mosaicism’ (y-axis, defined as the natural log of the ratio between the largest brain component and the smallest brain component in each individual, averaged across the population) as a function of developmental coupling D (x-axis) under different environments, defined by their functional coupling F, with an average benefit to cost ratio (B /C) of 1.5 and a mutation step size of 5%. d, e Summary of the effects of varying F and D, and F and B /C, respectively, on the degree of mosaicism. f–h Selected examples of competition between evolving populations with different D values. Each plot depicts the frequency of that D value relative to the total population (y-axis) as a function of developmental coupling D (x-axis) under alternative environments defined by their functional coupling F, with a benefit to cost ratio (B /C) of 1.5 Each data point is the outcome of one simulation run and the black bar indicates the mean of these runs. For comparison, grey bars show the means from the same unbounded simulations in Fig. 2a–c. i–k Each plot depicts the average relative frequencies of D = 0 (yellow) and D = 1 (green) at three B /C ratios, across three F values representing low (d), moderate (e) and high (f) functional coupling. For comparison, results from the same unbounded simulations in Fig. 3d, e are shown in faded colours. See Additional file 1: Figure S20-S23 for further full results varying B /C and F arguments that this is the case [32, 33, 36], the distinc- tion between concerted, or correlated, evolution and de- velopmental constraints is often neglected in the debates surrounding the evolution of brain structure. In our model, the probability of concertedness predictably in- creases with D, but it also increases with F, even when D is 0. Why do high F values result in concerted evolution? If we assume that selection acts on specific brain com- ponents, and excess brain tissue is generally expensive, relationships then selection to maintain functional should be expected and would result in co-evolution among brain components. These formal results are con- sistent with empirical data. For example, across mam- mals, the major components of the visual processing pathway, including the peripheral visual system, lateral geniculate nucleus and visual cortex, tend to co-evolve with one another, as predicted by their functional inte- gration [53–55]. Similarly, major components of the ol- factory pathway, including the olfactory bulbs and olfactory cortex, also co-evolve [54, 56]. However, the visual and olfactory pathways show no consistent pattern of co-evolution between them [54]. Indeed, whether they co-vary negatively, positively or not at all can be ex- plained by how diet and activity pattern interact to shape foraging behaviour [54]. In addition, major brain struc- tures connected by long-range axons, such as the neo- co-evolve cortex independently of total brain size, while also showing evi- dence of temporally transient independent change [57– 60]. These examples illustrate the effects of functional integration on co-evolution among brain components which are consistent with our model outputs. Finally, our model demonstrates an interaction between D and F, which may suggest that a pattern of concertedness driven by F could select for, or against, developmental integration. Regardless, our results demonstrate that concerted patterns of brain evolution provide no real evidence either for or against the prevalence of develop- mental coupling. cerebellum, tend also and to Both concerted and mosaic evolution can be adaptive, depending on the costs and benefits of brain tissue, and the complexity of the mutation landscape Our simulations suggest that the cost of excess brain tis- sue and the distribution of selection across networks of brain components play critical roles in determining how brains evolve. When the strength and direction of selec- tion is skewed towards one brain component, the prob- ability that a population with mosaic brains, or low D values, will be successful is increased. However, when the relative costs of neural tissue are either very low or high, the balance can switch to favour high D values. This likely reflects the ‘speed’ at which the two popula- tions can respond to selection. With only one mutational mechanism, it is potentially more likely for a develop- mentally coupled brain to evolve towards a new ‘adap- tive peak’ than it is for a mosaic brain. This is because landscape for a mosaic brain is more the mutational complex, and the probability of hitting the optimal mu- tational path is reduced by a greater number of potential mutational outcomes. At face value, our simulations therefore support previ- ous arguments that ‘… a coordinated enlargement of many independent components of one functional system without enlargement of the rest of the brain may be more difficult, as its probability would be the vanishingly small product of the probability of each component en- larged individually’ (2, p. 1583). However, if the costs of brain components are unequal, as may be expected if en- ergetic consumption scales with neuron number [61] and neuron density varies between components [62], then this effect would be dampened according to the distribution of costs and benefits across the whole sys- tem. Hence, an uneven distribution of neural costs would likely increase the probability of mosaicism. Evi- dence that this effect occurs in nature may be provided by recent data showing life history traits constrain spe- cific brain regions independently of overall brain size [63]. Avin et al. BMC Biology (2021) 19:97 Page 11 of 18 In addition, when we impose an upper and lower bound on brain size, to reflect scenarios where the total size of the system is under constraint, for example due to brain/body allometry, we further see the importance of considering the costs of neural tissue. Here, the cost of brain expansion increases as the system approaches either boundary, such that the option to inflate brain size as a whole becomes increasingly problematic. Under these conditions, the probability of mosaicism increases as it becomes more energetically conservative, and in some parameter spaces, individuals may have to reallo- cate energetic investment from one brain region to an- other, potentially resulting in investment trade-offs. We also note that, although we present the results of one way in which costs may scale under brain size con- straints, the results of alternative cost scaling relation- ships can be clearly predicted from formula 2; with altered settings for upper and lower bounds, the effects would ‘kick in’ at earlier or later points, and with differ- ent C exponents, the effects would ‘kick in’ with faster or slower rates. While accurately assessing the fitness costs and bene- fits of brain tissue remains challenging, we suggest some tentative predictions based on these interpretations. First, we expect mosaicism to be less likely during transi- tions to high-quality diets, as the cost of neural tissue relative to the overall energy budget may be reduced. Second, under sudden periods of energy resource limita- in a con- tion, brains should shrink, at least initially, certed manner, before approaching a lower bound for some structures when mosaicism will kick in to adap- tively restructure neural investment. Third, mosaicism may be more likely when energy intake is relatively con- stant, but selection favours changes in neural investment which involve energetic trade-offs. This may be common during periods of ecological change that are not associ- ated with changes in body size (and by proxy energetic intake), or in taxa that are size limited. It is tempting to interpret some notable patterns of brain evolution in this context. For example, within hominin evolution, human brains are largely structured in a way that is consistent with neural scaling across hominoids despite its massive increase in size [64, 65]. Given that some authors have argued brain expansion was facilitated by dietary shifts [66–68], this could have provided the conditions for a concerted pattern of brain expansion. In contrast, the brain of the diminutive hominin H. floresiensis is likely to have evolved under considerable size constraints and other authors have suggested cognitive evolution in this lineage was facilitated by a distinct pattern of change in brain morphology, perhaps reflecting a response to se- lection under size constraints [69]. Of course, further ex- perimentation with amenable systems will be necessary to test the predictions of our model. Partially mosaic brains and the maintenance of diversity It may be surprising that partially mosaic individuals, which are seemingly a compromise-state, generally have low probability of success in our simulations. This can again be explained by the ‘cost’ of mutation. In short, partially mosaic individuals incur the highest cost of mu- tational complexity. If the direction of mutation is ran- dom, in a given generation, the partially mosaic brain simply has lower odds of increasing its fitness because mutations affecting the whole brain and region-specific size may often be in conflict. The relative lack of genetic variation simultaneously contributing to both whole brain and region-specific size in natural populations [5, 8, 39] is in keeping with this conclusion. However, how mutational effects interact during brain development re- quires further investigation. However, our simulations across temporally variable selection regimes indicate that populations with alterna- tive D values can co-persist within a single population for many generations. This provides an alternative route to mixed temporal patterns of concerted and mosaic evolution within a single population. If we view D values in our model as alternative genotypes, this implies that genetic variation affecting specific components and gen- etic variation affecting total brain size may both either persist at low levels in natural populations or periodic- ally arise de novo and spread through a population. Re- gardless, selection to fluctuate between favouring mosaic and concerted mechanisms, permitting both adaptive restructuring and adaptive con- servation of brain structure, without invoking genetic mechanisms size and whole brain size. link brain structure this would enable that partly Reconciling mosaic and concerted views Our simulation results suggest a way of reconciling mo- saic and concerted views of brain evolution. Develop- mentally coupled brains evolve in scenarios involving some combination of tissue costs being evenly distrib- uted, and an extreme and variable fitness landscape, while mosaic brains are the result of environmental sta- bility coupled with differentiated selection among com- ponents, and/or strong constraints on brain size. The quick evolutionary response enabled by developmentally coupled brain evolution makes it ideal for circumstances where getting it ‘approximately correct’ quickly is advan- tageous, while mosaic evolution is favoured when accur- acy trumps speed. This is also tied to life history and environmental heterogeneity. For example, large off- spring number, which increases short-term competition, favours the fast response provided by developmental coupling, while lower competition allows mosaic popula- tions to persist, find the optimum brain structure and out-compete developmentally coupled individuals. Avin et al. BMC Biology (2021) 19:97 Page 12 of 18 The contrasting benefits of concerted and mosaic evo- lution bring us back to the initial major division between the two polarised hypotheses [1, 2, 13–15, 17] which continues to be widely reflected in the literature (e.g. [4– 12]); where developmental coupling does occur, is it a constraint, or has it evolved and is therefore evolvable? Our simulations support developmental coupling as a scenario-dependent adaptive mechanism, rather than a constraint. First, our simulations show that the fate of the two opposing mechanisms can vary through time. In nature, this would be reflected in the formation and breaking of genetic correlations between traits. Second, the general absence of genetic correlations observed in quantitative genetics studies is reconciled with concert- edness via developmental coupling by invoking the im- portance of de novo mutation, but our simulations suggest that the probability that these mutations will spread to fixation depends on the environmental con- text. By showing that the distribution and size of costs and benefits are critical in determining the outcome of competing mechanisms, our model supports a view whereby global selection regimes determine the con- straints acting on brain evolution, not the developmental program. Two key conclusions from this work are there- fore (i) patterns of concertedness should not be equated with developmental coupling and (ii) developmental coupling should not be equated to developmental constraint. Conclusion In sum, our agent-based simulations of alternative views of brain evolution provide a number of informative pre- dictions that should refresh our view of this long run- ning debate. First, we show concertedness is an outcome, not a mechanism, and on its own does not provide evidence for developmental coupling or con- straint. Second, we demonstrate that selection regime and structural interdependencies are critical to the out- come of competing mechanisms. Third, we argue that developmentally coupled and uncoupled brain architec- tures can both be adaptive, but we contest the assump- coupling is necessarily a tion that developmental persistent evolutionary constraint. Finally, our model provides a way to integrate patterns of brain evolution, life history and environmental heterogeneity. We of course acknowledge that our model is a simplification of a highly complex biological phenomenon. While the as- sumptions we make in constructing the model are intended to help us clearly explore previous verbal argu- ments, several extensions are possible. These include the addition of heterogeneous distributions of energetic costs, costs of functional integration to mirror energetic costs of long-range axons, and a degree of plasticity in brain structure and size that would reflect the plasticity of neural systems in response to their social and physical environment. Nevertheless, we hope the general ap- proach taken can trigger greater formalisation of evolu- tionary hypotheses in this field, and work that further refines our computational models. Methods To explore the interplay between developmental and functional coupling on the evolution of brain structure, we devised a model that simulates the evolution of a population of individuals in an environment, where fit- ness is determined by the size and costs of brain components. Definitions We employ terminology from the evolutionary neuro- biology literature, but the debate could also be under- stood in terms of principles derived from quantitative genetics. ‘Concerted evolution’ is the result of correlated evolution among all brain components, which can be caused by two main processes. The first, referred to as ‘developmental coupling’, occurs through genetic corre- lations that act through development, such that variation at a particular locus affects the development of multiple, ‘Developmental de-coupling’ or all, brain components. refers to the breakdown, or absence, of these genetic correlations. The second, referred to as ‘functional coup- ling’, instead occurs through correlated selection pres- components arise sures, which contributing to shared behavioural or computational functions, or the environment selecting on multiple as- pects of brain structure (e.g. dark environments selecting for increased olfactory processing and against visual pro- cessing; in this case, the selection correlation is nega- tive). Both correlated selection and genetic correlations can lead to co-evolution between traits [35], and our model is designed to explore this in the context of brain evolution. However, we note it is likely applicable to other anatomies. either due to Base model components is characterised by three components: the The model population of individuals, the environment and an evolu- tionary algorithm. We provide a simplified description of the model in Additional file 1: Figure S1. Each individual is characterised by a number (N) of brain components, each component having a specific size (Si,j), where i denotes the individual and j denotes the jth of N components. For example, among individuals with three brain components (N = 3), the brain of ‘Individual A’ could have a first component of size SA,1 = 5, a second component of size SA,2 = 7 and a third component of size SA,3 = 2, which we denote as SA = (5, 7, 2). A second indi- vidual, ‘Individual B’, might have component brain sizes Avin et al. BMC Biology (2021) 19:97 Page 13 of 18 of 3, 3 and 1, respectively, denoted SB = (3, 3, 1). In this case, the total size of Individual A’s brain would be 14, while the size of Individual B’s brain would be 7. Our model assumes that all brain components have the same fitness cost per unit size (see below), so it is total brain size that determines the evolutionary cost of an individ- ual’s brain, while the size of individual components con- tributes differently to fitness benefits. Whether evolution is mosaic or concerted depends upon the factors influen- cing changes in brain component sizes over time. The sizes of brain components are allowed to vary, through mutation, and this variation is influenced by develop- mental coupling (D), which takes values between 0.0 (no developmental coupling, i.e. a fully ‘mosaic’ brain) and 1.0 (complete developmental coupling, i.e. a brain struc- ture fully determined by total brain size). D is analogous to the strength of genetic correlations between compo- nents. For example, a D of 0.5 would indicate that 50% of the variation in each component is determined by variation in total brain size, and 50% of the variation in each brain component is independent of variation in both total brain size and other components. When a mutation event occurs, the program generates N + 1 ran- dom mutation factors (mj), for example between 0.5 and 1.5 for a 50% mutation step size, where there is one fac- tor for the whole brain (m0) and one for each component (m1 to mN), each brain component is then scaled by these mutation factors, with the variation in mutations affecting particular brain components being flattened depending on D’s value. For example, when D is 0, m0 for the total brain size mutation will be multiplied by 0 but other mutation fac- tors will vary independently according to a scaling factor (1 – D, i.e. unscaled when D is 0), whereas when D is 1 all mu- tation factors for individual brain components will be multi- plied by 0 and only the mutation factor for total brain size will persist. Models with intermediate values of D fall be- tween these extremes. This is determined according to the following formula: Si; jðnewÞ ¼ ½ðD (cid:2) m0Þ þ ðð1−DÞ (cid:2) m jÞ(cid:3) (cid:2) Si; jðoldÞ ; j∈f1; …; Ng benefits B = (1, 2, 3), Individual A, from above, will have total fitness contributions from brain component sizes of 5 × 1 + 7 × 2 + 2 × 3 = 25, while Individual B will have total fit- ness benefits of 3 × 1 + 3 × 2 + 1 × 3 = 12. This is imple- mented by varying Bmax, the highest possible benefit a unit of size could give, with the benefit per component size Bj given as a random fraction (uniformly sampled from the range 0 to 1) of this maximum benefit. Where F = 1, the generated fraction is the same for all components; if F = 0, then the benefit per component is independent (i.e. three generated fractions), and for any intermediate value of F, the benefit provided by each component is determined by con- tributions from both the cross-brain fraction and the per- component fractions, scaled so that the sum of fractions per component never exceed 1. With higher F values, the indi- vidual component benefits are constrained to be more simi- lar. The average benefit per component is always half the maximum benefit permitted, and the degree of functional coupling (F) determines how correlated they are. Third, the environment imposes a fitness cost (C) per size unit of the brain, which is uniform for all brain components based on the assumption that there is a linear ‘per neuron’ energetic cost [61], and that units of ‘size’ in the model are analogous to neuron number. More specifically, the units of ‘size’ in the model are analogous to the ratio between the neuron number of the current organism and the neuron number for the common ancestor, with the common ancestor starting with equal sized brain components, S0 = (1,1,1), in all our simulations. Total fitness for an individual i with brain component sizes Si,j in an environment defined by ðSi; j (cid:2) B jÞ−ðC benefits Bj and cost C is thus given by X X j (cid:2) Si; jÞ . For example, if the environment described j above had a cost C = 1, Individual A will experience a fit- ness cost of 14, giving a total fitness of 25–14 = 11, while Individual B will experience a fitness cost of 7, giving a total fitness of 5. To visualise the effects of varying C and Bmax, we measure the ratio of the average B across com- ponents (annotated B), to C. ð1Þ The Evolutionary Process progresses through the fol- lowing steps: The Environment is characterised by three factors. is characterised by Functional First, an environment Coupling (F), between 0.0 (no functional links between two brain components) and 1.0 (complete functional interdependence between two brain components), which determines how similar brain component benefits are to each other (for F = 1, all benefits are identical) — see below for implementation details. Second, a set of bene- fits (Bj), one for each of the N brain components, which represents the fitness contribution of each size unit an individual gains from that particular brain component, at a particular size. For example, in an environment with 1. Determine the number of ‘offspring’ for each individual in a population, and the age of all individuals, measured in the number of generations (these are identical across the population). Initialise an environment and a population of individuals with identical, uniform brain component sizes (Si,j = 1). 2. 3. For a given number of simulation steps a. Generate offspring for all individuals b. Mutate all offspring Avin et al. BMC Biology (2021) 19:97 Page 14 of 18 Increase the age of all individuals c. d. Remove individuals whose age exceeds the maximum age e. Rank individuals according to their total fitness in the environment (calculated as described above) f. Remove the lowest ranking individuals until the population size returns to the origin size (i.e. population size is stable over time) Given the above model parameters, we can examine the effects of developmental coupling (D) and functional coupling (F) on the evolution of brain component sizes. We can also explore the evolution of brain structure in populations of individuals with an intermediate value of D (unless specified otherwise, we use D = 0.5), which are neither fully concerted nor fully mosaic. We call these ‘partially mosaic’ individuals, where some of the muta- tions affect total brain size, scaling each component equally, and some affect each component independently, as given by Eq. (1) above. We can then assess which mechanism, for example, a fully mosaic brain (D = 0), a fully concerted brain (D = 1) or a partially mosaic brain (D = 0.5), in different scenarios by measuring the frequency of individuals in a population with that D value, as a proportion of total population size, after n generations. The model was initially imple- mented with no upper ceiling on overall brain size, in which case it most accurately simulates periods reflect- ing directional increases/decreases in brain components, and by extension brain size, which is a common but not universal trend [70, 71]. This means that under static se- lection regimes the base model can lead to continuous directional changes in total brain size. We also allow the environment to change randomly over the course of a simulation to examine how temporal heterogeneity in selection regimes affects the long-term success of alter- native brain models. is most successful Introducing constraints on total brain size As described above, the initial model imposes no upper or lower limit on total brain size. While this will reflect periods of directional changes in the brain, or brain component, size [70, 71], the close correlation between brain and body size, which may evolve under contrasting selection pressures [35], may impose limitations on how brains respond to selection that is not captured in the base model. However, these upper and lower boundaries can be envisaged in terms of non-linear relationships be- tween the size of brain components and their costs. As brain size approaches an upper/lower ceiling, the relative cost of increasing/decreasing each component is likely increased, resulting in increased disparity of B/C ratios for each component when selection is favouring in- creases in particular brain regions. To compare how such boundaries may impact the probability of different outcomes for D values, we implemented an extension to the model in which C is scaled according to its distance from the brain size of the common ancestor, with costs becoming increasingly prohibitive near an upper or lower boundary, set as 1.5 and 0.5 times the starting combined size of all three components, respectively. This scaling factor was implemented as: ! 2 Ci ¼ C (cid:2) 0:25 (cid:2) ð2Þ P jS0; jP jSi; j P jSi; jP jS0; j þ where C is the environmental-determined base cost, ∑jSi, j is the total size of all brain components for the current individual and ∑jS0, j is the total brain size of the common ancestor. Under these conditions, the model aims to examine situations where variation in total brain size is under stabilising selection or some form of constraint. Comparing effects of parameter variation on probabilities of mosaicism Using the base model, we first conducted a series of sim- ulations to explore four key questions identified in the introduction: 1. What mechanisms can produce concerted evolution? Here, we fixed C to equal 1 and fix Bmax to be 1, 2 or 4, giving a range of average B/C conditions, while varying D and F. We then ran simulations to examine the degree of mosaicism observed under high, moderate and low levels of F and D. 2. Can both mechanisms be adaptive? Here, we repeat the comparisons above, but in competitive environments to examine the probability of obtaining coordinated changes in brain components under high, moderate and low levels of F. 3. Do the costs of neural tissue select against concertedness when selection acts on specific brain components? i.e. How does variation in fitness contributions from different components affect the way brains respond to selection? This can be addressed in two ways: first, by varying F, which determines how correlated B values of each structure are, or by varying Bmax, to alter the B/C ratio. Here, our aim was to test whether different levels of variation in the fitness contribution of additional brain tissue alter the probability of obtaining a mosaic or concerted brain. Is developmental integration evolutionary labile? i.e. in a fluctuating environment do strong developmental constraints evolve and/or collapse? In this comparison, we took a different approach. 4. Avin et al. BMC Biology (2021) 19:97 Page 15 of 18 We introduced a starting population of brains with a range of component sizes and D. We then allowed C and Bmax to vary randomly every 2 generations to test what combination of factors persist over time. Here, costs were sampled uniformly in the range 0.5–5, and Bmax was sampled uniformly in the range 1–10. As a result, the average benefit (B) is in the range 0.5–5, which is the same range as the cost, but the actual B/C ratio will vary widely between generations. F was also sampled uniformly in the range 0–1. We subsequently explored how varying key life history traits (numbers of offspring and maximum age) might buffer the effects of random environmental fluctuations. 5. How do constraints on brain size interact with the probability of mosaicism? Finally, we subsequently examined how imposing upper and lower bounds on total brain size impact the probability of obtaining a mosaic or concerted brain under each of the conditions described above. experiments total n = 27,000 iterations). We estimated the effects of all parameters and interactions where indi- cated, with a Gaussian distribution when comparing ‘de- grees of mosaicism’ (defined as the natural log of the ratio between the largest brain component and the smal- lest brain component in each individual, averaged across the population) and a quasibinomial distribution when comparing proportional frequencies. The code files in which the model is implemented are openly available for readers to implement additional param- eter settings, or to extend the model, and can be accessed from github.com/shaharavin/BrainEvolutionSimulator [73]. Biplots were made using PlotsOfData [74]. Many simula- tions produce a bimodal distribution of frequencies, we therefore display the mean of these iterations solely to illus- trate the skew in the outcome. Supplementary Information Supplementary information accompanies this paper at https://doi.org/10. 1186/s12915-021-01024-1. results from a also present In all case, the simulations were run over 100 genera- tions, with 1000 iterations, a fixed population size of 300 individuals, initiated with 100 individuals per D value. In the main text, we present results of simulations with a mutation step size of 5%, but runs using the base model were also repeated with a larger mutation step size of 50% to examine how effects were influenced by mutation size (full results are presented in the Supplementary In- formation). To explore the early stages of the simula- tions, we subset of simulations with 10 generations for these base compari- sons. In experiments 2 and 3, simulations were run with an initial population containing equal numbers of indi- viduals with different D values (D = 0, 0.5 or 1), which were then evolved under different environment condi- tions (determined by F and the ratio of benefits to cost). In experiment 4, environmental conditions were ran- domly varied every 2 generations for 150 generations. When imposing the upper and lower bounds on brain size, we repeated experiments 1–4, as described above, with a mutation step size of 5%. The ‘success’ of a D value was determined by the proportion of individuals in a population with that D value at the end of the simula- tion run. The full output of all models are summarised in Additional file 1: Figure S2-S23. The frequency of D values was compared using generalised linear models and the glm() function in R [72] across batches of nine, 1000-iteration simulations where F, D and Bmax and/or C varied (experiments 1–3, see for example Add- itional file 1: Figure S2,S6), or where F, D, maximum lifespan and offspring number varied (experiment 4, see, (all for example Additional 1: Figure S14) file Additional file 1: Figures S1-S23. Figure S1. A simplified, pictorial depiction of the model. Pages 4–5. Figure S2. Evolution of ‘mosaicism’ under alternative conditions, full comparisons, run with 5% mutation size and 100 generations (sister to Fig. 1). Page 6. Figure S3. Evolution of ‘mosaicism’ under alternative conditions, full comparisons, run with 50% mutation size and 100 generations (extension of Fig. 1). Page 7. Figure S4. Evolution of ‘mosaicism’ under alternative conditions, full comparisons, run with 50% mutation size and 10 generations (extension of Fig. 1). Page 8. Figure S5. Generation number at convergence during simulations of competition between evolving populations with different D values under alternative conditions (companion to Fig. 2). Page 9. Figure S6. Competition between evolving populations with different D values under alternative conditions, full comparisons, run with 5% mutation size and 100 generations (sister to Fig. 2). Page 10. Figure S7. Competition between evolving populations with different D values under alternative conditions, full comparisons, run with 50% mutation size and 100 generations (extension of Figs. 2 and 3). Page 11. Figure S8. Competition between evolving populations with different D values under alternative conditions, full comparisons, run with 50% mutation size and 10 generations (extension of Figs. 2 and 3). Page 12. Figure S9. Competition between evolving populations with different D values under tailored environmental conditions, run with 5% mutation size, showing the average size of each brain component and the frequency of competing D values (extension of Figs. 2 and 3). Page 13. Figure S10. Competition between evolving populations with different D values under tailored environmental conditions, full comparisons, run with 5% mutation size and 100 generations (extension of Figs. 2 and 3). Page 14. Figure S11. Competition between evolving populations with different D values under tailored environmental conditions, full comparisons, run with 50% mutation size and 100 generations (extension of Figs. 2 and 3). Page 15. Figure S12. Competition between evolving populations with different D values under tailored environmental conditions, full comparisons, run with 50% mutation size and 10 generations (extension of Figs. 2 and 3). Page 16. Figure S13. Conditions at convergence of simulations of competition between evolving populations with different D values in a randomly varying environment, under different life history conditions (companion to Fig. 2). Page 17. Figure S14. Competition between evolving populations with different D values in a varying environment, under different life history conditions, full comparisons, run with 5% mutation size and 100 generations (sister to Fig. 4). Page 18. Figure S15. Competition between evolving populations with different D values in a varying environment, under different life history conditions, full comparisons, run with 50% mutation size and 100 generations Avin et al. BMC Biology (2021) 19:97 Page 16 of 18 (extension of Fig. 4). Page 19. Figure S16. Competition between evolving populations with different D values in a varying environment, under different life history conditions, full comparisons, run with 50% mutation size and 10 generations (extension of Fig. 4). Page 20. Figure S17. Selected, representative, individual simulations showing fluctuations in population frequencies over 100 generations for a 5% mutation size (A-D) or a 50% mutation size (E-H) (extension of Fig. 4). Page 21. Figure S18. Subtle effects of the size of environmental fluctuations on competition between evolving populations with different D values in a varying environment (extension of Figure S16). Page 22. Figure S19. Relationship between population frequencies between partially mosaic brains (D = 0.5) and fully mosaic (D = 0), or concerted brains (D = 1) from simulations in varying environmental conditions and a 5% mutation size (A) or 50% mutation size (B) (extension of Fig. 4). Page 23. Figure S20. Evolution of ‘mosaicism’ under alternative conditions with upper and lower bounds on brain size, full comparisons, run with 5% mutation size and 100 generations (extension of Fig. 6). Page 24. Figure S21. Competition between evolving populations with different D values under alternative conditions with upper and lower bounds with upper and lower bounds on brain size, full comparisons, run with 5% mutation size and 100 generations (extension of Fig. 6). Page 25. Figure S22. Selected, representative, individual simulations showing fluctuations in population frequencies over 50 generations for a 5% mutation size, with upper and lower bounds with upper and lower bounds on brain size (extension of Fig. 6). Page 26. Figure S23. Conditions at convergence of simulations of competition between evolving populations with different D values in a randomly varying environment, with upper and lower bounds on brain size, under different life history conditions (extension of Fig. 6). Page 27. Acknowledgements We thank Corina Logan, for instigating this collaboration, the Castiglione family for facilitating discussions, and several anonymous reviewers at this and other journals for constructive criticism and advice. Authors’ contributions SA, AC and SHM devised the model, interpreted the output and wrote the manuscript. SA produced the code to implement the model. All authors read and approved the final manuscript. Funding AC was supported by the Templeton World Charity Foundation (TWCF0303). SHM was supported by a NERC Independent Research Fellowship (NE/ N014936/1). Availability of data and materials The code generated during the current study are available at github.com/ shaharavin/BrainEvolutionSimulator [73]. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare they have no competing interests. Author details 1Centre for the Study of Existential Risk, University of Cambridge, Cambridge, UK. 2Department of Sociology, Philosophy and Anthropology, University of Exeter, Exeter, UK. 3School of Biological Sciences, University of Bristol, Bristol, UK. Received: 8 October 2020 Accepted: 13 April 2021 References 1. 2. Finlay BL, Darlington RB, Nicastro N. Developmental structure in brain evolution. Brain Behav Sci. 2001;24:298–304. Finlay B, Darlington R. Linked regularities in the development and evolution of mammalian brains. Science. 1995;268(5217):1578–84. https://doi.org/1 0.1126/science.7777856. 7. 5. 4. 6. 3. Workman AD, Charvet CJ, Clancy B, Darlington RB, Finlay BL. Modeling transformations of neurodevelopmental sequences across mammalian species. J Neurosci. 2013;33(17):7368–83. https://doi.org/10.1523/JNEUROSCI. 5746-12.2013. Yopak KE, Lisney TJ, Darlington RB, Collin SP, Montgomery JC, Finlay BL. A conserved pattern of brain scaling from sharks to primates. Proc Natl Acad Sci. 2010;107(29):12946–51. https://doi.org/10.1073/pnas.1002195107. Hager R, Lu L, Rosen GD, Williams RW. Genetic architecture supports mosaic brain evolution and independent brain–body size regulation. Nat Commun. 2012;3(1):1079. https://doi.org/10.1038/ncomms2086. Powell BJ, Leal M. Brain evolution across the Puerto Rican anole radiation. Brain Behav Evol. 2012;80(3):170–80. https://doi.org/10.1159/000341161. Gutiérrez-Ibáñez C, Iwaniuk AN, Moore BA, Fernández-Juricic E, Corfield JR, Krilow JM, Kolominsky J, Wylie DR. Mosaic and concerted evolution in the visual system of birds. Plos One. 2014;9(3):e90102. https://doi.org/10.1371/ journal.pone.0090102. Noreikiene K, Herczeg G, Gonda A, Balázs G, Husby A, Merilä J. Quantitative genetic analysis of brain size variation in sticklebacks: support for the mosaic model of brain evolution. Proc R Soc B Biol Sci. 2015;282(1810): 20151008. https://doi.org/10.1098/rspb.2015.1008. Carlisle A, Selwood L, Hinds LA, Saunders N, Habgood M, Mardon K, Weisbecker V. Testing hypotheses of developmental constraints on mammalian brain partition evolution, using marsupials. Sci Rep. 2017;7(1): 4241. https://doi.org/10.1038/s41598-017-02726-9. Kotrschal A, Zeng H-L, van der Bijl W, Öhman-Mägi C, Kotrschal K, Pelckmans K, Kolm N. Evolution of brain region volumes during artificial selection for relative brain size. Evolution. 2017;71(12):2942–51. https://doi. org/10.1111/evo.13373. Sukhum KV, Shen J, Carlson BA. Extreme enlargement of the cerebellum in a clade of teleost fishes that evolved a novel active sensory system. Curr Biol. 2018;28:3857–3863.e3. 11. 10. 9. 8. 12. D’Aniello B, Di Cosmo A, Scandurra A, Pinelli C. Mosaic and concerted brain evolution: the contribution of microscopic comparative neuroanatomy in lower vertebrates. Front Neuroanat. 2019;13:86. https://doi.org/10.3389/fna na.2019.00086. 13. Harvey P. Krebs. Comparing brains. Science. 1990;249(4965):140–6. https:// doi.org/10.1126/science.2196673. 14. Barton RA, Harvey PH. Mosaic evolution of brain structure in mammals. 15. Nature. 2000;405:4. Iwaniuk AN, Dean KM, Nelson JE. A mosaic pattern characterizes the evolution of the avian brain. Proc R Soc Lond B Biol Sci. 2004;271(suppl_4): S148–S151. https://doi.org/10.1098/rsbl.2003.0127. 16. Willemet R. Understanding the evolution of mammalian brain structures; the need for a (new) cerebrotype approach. Brain Sci. 2012;2(2):203–24. https://doi.org/10.3390/brainsci2020203. 17. de Winter W, Oxnard CE. Evolutionary radiations and convergences in the structural organization of mammalian brains. Nature. 2001;409(6821):710–4. https://doi.org/10.1038/35055547. 18. Montgomery SH, Mundy NI, Barton RA. Brain evolution and development: adaptation, allometry and constraint. Proc R Soc B Biol Sci. 2016;283(1838): 20160433. https://doi.org/10.1098/rspb.2016.0433. 19. Godfrey-Smith. Complexity and the function of mind in nature. New York: Cambridge University Press; 1996. 20. Brown WM. Natural selection of mammalian brain components. Trends Ecol Evol. 2001;16(9):471–3. https://doi.org/10.1016/S0169-5347(01)02246-7. 21. Charvet CJ, Darlington RB, Finlay BL. Variation in human brains may facilitate evolutionary change toward a limited range of phenotypes. Brain Behav Evol. 2013;81(2):74–85. https://doi.org/10.1159/000345940. 22. Beatty J. Why do biologists argue like they do? Philos Sci. 1997;64:S432–43. https://doi.org/10.1086/392620. 23. Charvet CJ, Striedter GF, Finlay BL. Evo-devo and brain scaling: candidate developmental mechanisms for variation and constancy in vertebrate brain Avin et al. BMC Biology (2021) 19:97 Page 17 of 18 evolution. Brain Behav Evol. 2011;78(3):248–57. https://doi.org/10.1159/ 000329851. 46. Parke EC. Experiments, simulations, and epistemic privilege. Philos Sci. 2014; 81(4):516–36. https://doi.org/10.1086/677956. 24. Chen Y-C, Harrison PW, Kotrschal A, Kolm N, Mank JE, Panula P. Expression change in Angiopoietin-1 underlies change in relative brain size in fish. Proc R Soc B Biol Sci. 2015;282(1810):20150872. https://doi.org/10.1098/rspb.2015.0872. 47. González-Forero M, Gardner A. Inference of ecological and social drivers of human brain-size evolution. Nature. 2018;557(7706):554–7. https://doi.org/1 0.1038/s41586-018-0127-x. 25. Montgomery SH, Capellini I, Venditti C, Barton RA, Mundy NI. Adaptive 48. González-Forero M, Faulwasser T, Lehmann L. A model for brain life history evolution of four microcephaly genes and the evolution of brain size in anthropoid Primates. Mol Biol Evol. 2011;28(1):625–38. https://doi.org/10.1 093/molbev/msq237. evolution. Plos Comput Biol. 2017;13(3):e1005380. https://doi.org/10.1371/ journal.pcbi.1005380. 49. dos Santos M, West SA. The coevolution of cooperation and cognition in 26. Charvet CJ, Striedter GF. Bigger brains cycle faster before neurogenesis begins: a comparison of brain development between chickens and bobwhite quail. Proc R Soc B Biol Sci. 2010;277(1699):3469–75. https://doi. org/10.1098/rspb.2010.0811. 28. 27. Dyer MA, Martins R, da Silva FM, Muniz JAPC, Silveira LCL, Cepko CL, et al. Developmental sources of conservation and variation in the evolution of the primate eye. Proc Natl Acad Sci. 2009;106(22):8963–8. https://doi.org/1 0.1073/pnas.0901484106. Striedter GF, Charvet CJ. Developmental origins of species differences in telencephalon and tectum size: morphometric comparisons between a parakeet (Melopsittacus undulatus ) and a quail (Colinus virgianus). J Comp Neurol. 2008;507(5):1663–75. https://doi.org/10.1002/cne.21640. Sylvester JB, Rich CA, Loh Y-HE, van Staaden MJ, Fraser GJ, Streelman JT. Brain diversity evolves via differences in patterning. Proc Natl Acad Sci. 2010;107(21):9718–23. https://doi.org/10.1073/pnas.1000395107. 29. 30. Charvet CJ, Striedter GF. Developmental basis for telencephalon expansion in waterfowl: enlargement prior to neurogenesis. Proc R Soc B Biol Sci. 2009;276(1672):3421–7. https://doi.org/10.1098/rspb.2009.0888. Striedter GF, Charvet CJ. Telencephalon enlargement by the convergent evolution of expanded subventricular zones. Biol Lett. 2009;5(1):134–7. https://doi.org/10.1098/rsbl.2008.0439. 31. 32. Barton RA. The coordinated structure of mosaic brain evolution. Behav Brain Sci. 2001;24(2):281–2. https://doi.org/10.1017/S0140525X01253953. 33. Deaner RO, van Schaik CP. Flaws in evolutionary theory and interpretation. Behav Brain Sci. 2001;24:279–80. 34. Gómez-Robles A, Hopkins WD, Sherwood CC. Modular structure facilitates mosaic evolution of the brain in chimpanzees and humans. Nat Commun. 2014;5(1):4469. https://doi.org/10.1038/ncomms5469. Lande R. Quantitative genetic analysis of multivariate evolution. Appl Brain. 1979;33:402. 35. 36. Airey DC, Williams RW. Quantitative neurogenetic perspectives. Brain Behav 37. Sci. 2001;24:282–3. Eifert C, Farnworth M, Schulz-Mirbach T, Riesch R, Bierbach D, Klaus S, Wurster A, Tobler M, Streit B, Indy JR, Arias-Rodriguez L, Plath M. Brain size variation in extremophile fish: local adaptation versus phenotypic plasticity: brain size variation in extremophile fish. J Zool. 2015;295(2):143–53. https:// doi.org/10.1111/jzo.12190. 38. Gonda A, Herczeg G, Merilä J. Habitat-dependent and -independent plastic responses to social environment in the nine-spined stickleback (Pungitius pungitius) brain. Proc R Soc B Biol Sci. 2009;276(1664):2085–92. https://doi. org/10.1098/rspb.2009.0026. 39. Henriksen R, Johnsson M, Andersson L, Jensen P, Wright D. The domesticated brain: genetics of brain mass and brain structure in an avian species. Sci Rep. 2016;6(1) https://doi.org/10.1038/srep34031. 40. Herculano-Houzel S, Manger PR, Kaas JH. Brain scaling in mammalian evolution as a consequence of concerted and mosaic changes in numbers of neurons and average neuronal cell size. Front Neuroanat. 2014;8 https:// doi.org/10.3389/fnana.2014.00077. 41. Moore JM, DeVoogd TJ. Concerted and mosaic evolution of functional modules in songbird brains. Proc R Soc B Biol Sci. 2017;284(1854):20170469. https://doi.org/10.1098/rspb.2017.0469. 67. 42. Hoops D, Vidal-García M, Ullmann JFP, Janke AL, Stait-Gardner T, Duchêne DA, Price WS, Whiting MJ, Keogh JS. Evidence for concerted and mosaic brain evolution in dragon lizards. Brain Behav Evol. 2017;90(3):211–23. https://doi.org/10.1159/000478738. 43. Harrison PW, Montgomery SH. Genetics of cerebellar and neocortical expansion in anthropoid primates: a comparative approach. Brain Behav Evol. 2017;89(4):274–85. https://doi.org/10.1159/000477432. 44. Currie A. From Models-as-Fictions to Models-as-Tools. Open Access J Philos. 45. 2017;4(20201214). https://doi.org/10.3998/ergo.12405314.0004.027. Knuuttila T. Imagination extended and embedded: artifactual versus fictional accounts of models. Synthese. 2017; https://doi.org/10.1007/s11229-017-1545-2. humans. Proc R Soc B Biol Sci. 2018;285(1879):20180723. https://doi.org/10.1 098/rspb.2018.0723. 50. Muthukrishna M, Doebeli M, Chudek M, Henrich J. The cultural brain hypothesis: how culture drives brain expansion, sociality, and life history. PLoS Comput Biol. 2018;14(11):e1006504. https://doi.org/10.1371/journal. pcbi.1006504. 51. Dunbar, R.I.M. Confounding explanations. . . . Brain Behav Sci. 2001;24:283. Striedter GF. Brain evolution: how constrained is it? Behav Brain Sci. 24 (2): 52. 296–7. https://doi.org/10.1017/S0140525X01403955. 53. Barton RA. Visual specialization and brain evolution in primates. Proc R Soc Lond B Biol Sci. 1998;265(1409):1933–7. https://doi.org/10.1098/rspb.1998.0523. 54. Barton RA. Evolutionary radiation of visual and olfactory brain systems in primates, bats and insectivores. Philos Trans R Soc Lond Ser B Biol Sci. 1995; 348(1326):381–92. https://doi.org/10.1098/rstb.1995.0076. 55. Barton RA. Binocularity and brain evolution in primates. Proc Natl Acad Sci. 2004;101(27):10113–5. https://doi.org/10.1073/pnas.0401955101. 56. Barton RA. Olfactory evolution and behavioral ecology in primates. Am J Primatol. 2006;68(6):545–58. https://doi.org/10.1002/ajp.20251. 57. Barton RA, Venditti C. Rapid evolution of the cerebellum in humans and other great apes. Curr Biol. 2014;24(20):2440–4. https://doi.org/10.1016/j. cub.2014.08.056. Smaers JB, Steele J, Zilles K. Modeling the evolution of cortico-cerebellar systems in primates: Smaers et al. Ann N Y Acad Sci. 2011;1225(1):176–90. https://doi.org/10.1111/j.1749-6632.2011.06003.x. 58. 59. Herculano-Houzel. Coordinated scaling of cortical and cerebellar numbers of neurons. Front Neuroanat. 2010. doi:https://doi.org/10.3389/fnana.2010.00012. 60. Muller AS, Montgomery SH. Co-evolution of cerebral and cerebellar expansion in cetaceans. J Evol Biol. 2019;32(12):1418–31. https://doi.org/1 0.1111/jeb.13539. 61. Herculano-Houzel S. Scaling of brain metabolism with a fixed energy budget per neuron: implications for neuronal activity, plasticity and evolution. Plos One. 2011;6(3):e17514. https://doi.org/10.1371/journal.pone.0017514. 62. Herculano-Houzel S, Catania K, Manger PR, Kaas JH. Mammalian brains are made of these: a dataset of the numbers and densities of neuronal and nonneuronal cells in the brain of glires, primates, scandentia, eulipotyphlans, afrotherians and artiodactyls, and their relationship with body mass. Brain Behav Evol. 2015;86(3-4):145–63. https://doi.org/10.1159/ 000437413. 63. Powell LE, Barton RA, Street SE. Maternal investment, life histories and the evolution of brain structure in primates. Proc R Soc B Biol Sci. 2019; 286(1911):20191608. https://doi.org/10.1098/rspb.2019.1608. 64. Herculano-Houzel S. The remarkable, yet not extraordinary, human brain as a scaled-up primate brain and its associated cost. Proc Natl Acad Sci. 2012; 109(Supplement_1):10661–8. 65. Barton RA, Venditti C. Human frontal lobes are not relatively large. Proc Natl Acad Sci. 2013;110(22):9001–6. https://doi.org/10.1073/pnas.1215723110. 66. Wrangham R. Control of fire in the Paleolithic: evaluating the cooking hypothesis. Curr Anthropol. 2017;58(S16):S303–13. https://doi.org/10.1086/ 692113. Fonseca-Azevedo K, Herculano-Houzel S. Metabolic constraint imposes tradeoff between body size and number of brain neurons in human evolution. Proc Natl Acad Sci. 2012;109(45):18571–6. https://doi.org/10.1073/ pnas.1206390109. 69. 68. Aiello LC, Wheeler P. The expensive-tissue hypothesis: the brain and the digestive system in human and primate evolution. Curr Anthropol. 1995; 36(2):199–221. https://doi.org/10.1086/204350. Falk D, Hildebolt C, Smith K, Morwood MJ, Sutikna T, Jatmiko, et al. LB1’s virtual endocast, microcephaly, and hominin brain evolution. J Hum Evol 2009;57:597–607, 5, doi: https://doi.org/10.1016/j.jhevol.2008.10.008. Shultz S, Dunbar R. Encephalization is not a universal macroevolutionary phenomenon in mammals but is associated with sociality. Proc Natl Acad Sci. 2010;107(50):21582–6. https://doi.org/10.1073/pnas.1005246107. 70. Avin et al. BMC Biology (2021) 19:97 Page 18 of 18 71. Montgomery SH, Capellini I, Barton RA, Mundy NI. Reconstructing the ups and downs of primate brain evolution: implications for adaptive hypotheses and Homo floresiensis. BMC Biol. 2010;8(1):9. https://doi.org/10.1186/1741- 7007-8-9. 72. R Core Team. R: a language and environment for statistical computing. Vienna: R Foundation for Statistical Computing. Available online at https:// www.Rproject.org/. Accessed 4 Aug 2019. 73. Avin S, Currie A, Montgomery SH. BrainEvolutionSimulator. Github. github. com/shaharavin/BrainEvolutionSimulator. Accessed 27 Feb 2021. 74. Postma M, Goedhart J. PlotsOfData—a web app for visualizing data together with their summaries. Plos Biol. 2019;17(3):e3000202. https://doi. org/10.1371/journal.pbio.3000202. Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
10.1080_19420889.2020.1729601
COMMUNICATIVE & INTEGRATIVE BIOLOGY 2020, VOL. 13, NO. 1, 27–38 https://doi.org/10.1080/19420889.2020.1729601 RESEARCH PAPER Does regeneration recapitulate phylogeny? Planaria as a model of body-axis specification in ancestral eumetazoa Chris Fields a and Michael Levin b aCaunes Minervois, France; bAllen Discovery Center, Tufts University, Medford, MA, USA ABSTRACT Metazoan body plans combine well-defined primary, secondary, and in many bilaterians, tertiary body axes with structural asymmetries at multiple scales. Despite decades of study, how axis- defining symmetries and system-defining asymmetries co-emerge during both evolution and development remain open questions. Regeneration studies in asexual planaria have demonstrated an array of viable forms with symmetrized and, in some cases, duplicated body axes. We suggest that such forms may point toward an ancestral eumetazoan form with characteristics of both cnidarians and placazoa. ARTICLE HISTORY Received 19 December 2019 Revised 7 February 2020 Accepted 9 February 2020 KEYWORDS Bioelectricity; BMP pathway; eumetazoa; nervous system; symmetry breaking; whole-body regeneration; Wnt pathway Introduction What is the connection between spatial symmetry break- ing and multicellularity? To what extent can an ur- metazoan ancestor by envisaged as an initially adventi- tious, spherically symmetric aggregation of ancestral uni- suggested by the cells, e.g. of choanoflagellates as “choanoblastaea” model [1], see also [2,3]? Spatial asym- metries clearly predate multicellularity: the Bacilli and the spiral bacteria are classified by their non-spherically sym- metric shapes. Even E. coli exhibits substrate-dependent chirality at the colony scale [4]. Unicellular eukaryotes exhibit a vast array of internal and external spatial asym- metries. How are such spatial asymmetries translated to scale of a multicellular organism, particularly the a metazoan with well-defined cell layers and multiple distinct organ systems arranged in a specific, population- invariant pattern? The ability to systematically manipulate body-axis asymmetries during whole-body regeneration (WBR) may provide a route toward answering these questions. Organisms capable of WBR are found in all five primary metazoan clades, including the placozoa [5], sponges [6], and ctenophores [7] as well as bilaterians and cnidarians [8]; hence WBR is widely regarded as an ancestral metazoan trait [8–10]. Here we will focus on WBR outcomes in asexual freshwater planaria (Platyhelminthes, Turbellaria, Tricladida), by far the most extensively manipulated WBR model system [11,12], from acoel worms mentioning (Hydrazoa) where (Xenacoelomorpha) available. and Hydra supporting results Distinct body axes, along which differentiated structures can be asymmetrically arranged, provide the basis for Eumetazoan morphologies. With the advent of whole- genome sequencing and transcriptomics, it has become evident that the eumetazoan sister clades of cnidarians and bilaterians employ homologous “developmental toolk- its” for body-axis specification [13–16]. Considerable molecular as well as embryological evidence supports homology between the primary cnidarian aboral – oral (A-O) and bilaterian anterior – posterior (A-P) axes [3,17–19]. While a second, dorsal – ventral (D-V) axis breaking the otherwise cylindrical symmetry around the A-P is a defining bilaterian trait, both molecular and ana- tomical evidence support a secondary (“directive”) axis in at least some cnidarians [20–23]. A third, left – right (L-R) asymmetry appears in some arthropods (e.g. in lobsters) and is ubiquitous in vertebrates [24,25]. We focus here on the early-appearing A-P and D-V axes and their morpho- logical correlates, particularly the gut and central nervous system (CNS) axes. While many treatments are known that specifically disrupt axis specification in multicellular systems (e.g. Wnt, BMP, or bioelectrical pathways for the AP, DV, and LR axes; see below), these processes remain difficult to manipulate arbitrarily and with full control with mole- cular or embryological methods in either cnidarians and bilaterians. It is not, for example, completely clear at the molecular or cellular level how the morphological asym- metries of the CNS or the gut, or the behavioral asymme- try of forward locomotion, are aligned along the A-P axis CONTACT Chris Fields © 2020 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. 23 Rue des Lavandières, Caunes Minervois 11160, France fieldsres@gmail.com 28 C. FIELDS AND M. LEVIN systems without in bilaterals. Nor is it known, outside of planaria and acoels (see below), whether such morphological or beha- vioral asymmetries can be selectively reversed, e.g. to produce A-P symmetric nervous systems or guts. Some putatively basal, acoel bilaterians have rudimentary, net- like nervous evident ganglia or nerve cords, while others have more elaborated structures [26–28], suggesting that the correlation between CNS axis and A-P axis is not universal in bilaterians. The vermi- form myxozoa, e.g. Buddenbrockia [29,30] exhibit for- ward locomotion driven by coordinated, A-P aligned muscle groups, but are radially symmetric cnidarians that altogether lack nervous systems. While such organ- isms are morphological outliers and may exhibit substan- tial derived loss of function, their existence renders reconstruction of ancestral axis-specification mechanisms and, in particular, the morphology and expected beha- vioral repertoire of the common eumetazoan ancestor less than straightforward even given extensive comparative genomics. Here we suggest that WBR [8–10] provides a tractable alternative to embryonic development for asking funda- mental questions about body-axis specification and deep ancestral morphology. We use the term “WBR” to indi- cate regeneration of the whole body from non-germ cells following either natural or laboratory-induced injuries. In the asexual freshwater planaria of primary interest here, specific experimental manipulations of WBR can symme- trize the A-P axis, including the nervous system and gut [11], add ectopic A-P axes [31], or remove the A-P axis altogether to produce outcomes radially symmetric around the remaining D-V axis [32]; see also [12] for review and specific details below. These manipulations support a suggested homology between the ventral nerve cord (VNC) of bilaterians and the circumoral nerve ring of cnidarians. We reconstruct a hypothesized ancestral eumetazoan characterized by a D-V axis, a blind gut, a nerve ring with a surrounding nerve net, and asexual reproduction. We suggest that the primary function of the nervous system is this animal was not locomotion or feeding but the regulation of body size and morphology. Planaria exemplify basal bilaterian morphology and WBR capability While the early phylogeny of the Metazoa remains con- troversial, there is broad agreement across models that the Cnidarians and the Bilaterians are sister clades [15,33]. The early phylogeny of Bilaterians is similarly controver- sial, with numerous models now recognizing the Acoelamorpha as basal bilaterians [34–38]; see [39] for conflicts between molecular and a discussion of phylogenetic developmental-morphological analyses. These animals are characterized by an unsegmented body plan, blind gut, and in some species, an L-R symmetric, multiple-VNC nervous system [18,40], although as noted above, nervous-system morphology is highly variable [26,28]. The development of robust acoel model systems including Isodiametra pulchra [41,42] and Hofstenia miamia [43–45] has allowed the biology of these organisms, including their regenerative capabilities, to be characterized. Srivastava et al. [45] showed that Hofstenia miamia is capable of WBR mediated by the Wnt and BMP pathways as it is in both planaria and Hydra [see also 43]. Regeneration is enabled by somatic stem cells (neoblasts) expressing piwi homologs, as it is in WBR-capable planaria [39]. In contrast to Hofstenia mia- mia, Isodiametra pulchra is capable of posterior regenera- tion, but not WBR [42]. Such variability in WBR capability is also observed in planaria [12]. Despite recent progress with acoels, the asexual planar- ian model systems Dugesia japonica and Schmidtea med- iterranea remain the best-characterized and most extensively manipulated organisms with which to study WBR. While the rhabditophoran Platyhelminths, which include the planarians, are no longer regarded as a basal taxon, they share many of the morphological character- istics of the acoels, including unsegmented body plan, blind gut, and L-R symmetric, two-VNC nervous system [18]. Whether these morphological commonalities are ancestral or derived in either extant acoels or extant planaria remains unknown. Both acoels and planaria exhibit atypical embryonic development [46,47]; whether these characteristics are ancestral or derived also remains unknown. As with morphology, basal bilaterian reproductive strategy remains controversial [48,49]. While sexual reproduction far pre-dates multicellularity, obligate sexu- ality appears to be a multicellular innovation in both animal and plant lineages, consistent with Red Queen type arguments [50]. Demosponges and cnidarians such as Hydra exhibit opportunistic sexuality with budding and WBR [51,52], suggesting that obligate sexuality is derived from this more [9,53]. Characterized acoels include male-female and cross- fertilizing hermaphroditic species as well as asexuals that reproduce by budding or fission [40]. Characterized pla- naria include obligate sexual, opportunistic sexual, or asexual species, with some species alternating between sexual reproduction and parthenogenesis or between sex- ual and vegetative (fission followed by WBR) reproduc- tion [54]. Asexual planaria can be sexualized by feeding them closely related sexual planaria, suggesting that inter- cellular morphogen-based signaling promotes or enforces strategy flexible sexuality [55,56], inducing stem-cell lineages that would otherwise reproduce to replicate themselves instead to undergo a stem – germ – stem lineage cycle [53,57]. Manipulating WBR in planaria Asexual planaria reproduce by fission transverse to the A-P axis followed by WBR of missing anterior or pos- terior structures [58]; fission is a size and environmen- tal conditions dependent biomechanical process [59] regulated in part by Wnt and BMP pathways [60]. Experimental transverse amputation of both head and tail produce trunk fragments that regenerate both ante- rior and posterior structures. While amputation of both head and tail does not occur during reproductive fis- sion in the wild, both it and the other manipulations described below are possible outcomes of predation in the wild and engage the same molecular and bioelectric pathways active in reproductive transverse fission. A large number of molecular, pharmaceutical, and bio- electric manipulations have been shown to disrupt trunk, and smaller fragments WBR in head, [11,61]. It is now well-established that the Wnt pathway implements A-P axis specification [62–65], with either bioelectric asymmetry [66] or morphogen transport to the wound site [65] as initiating events. Elements of the Hedgehog (Hh) pathway regulate Wnt pathway activity in both anterior and posterior compartments [67]. Regeneration of specific anterior structures including brain and eyes also depends on the ERK and FGF pathways [65,68,69]. Molecular manipulations impli- cate the BMP pathway as specifying the D-V axis [11] as in other bilaterians [14]. tail, COMMUNICATIVE & INTEGRATIVE BIOLOGY 29 penetrance dose-dependent Here we are primarily interested in manipulations that symmetrize the A-P axis, i.e. replace the asym- metric A-P axial morphology with a symmetric A-P-A morphology, or introduce one or more ectopic A-P axes, with radially symmetric forms in which the A-P axis appears to have been altogether eliminated as the limiting case. If a morphologically normal worm is cut at 60% and 80% of its length to make a “pre-tail” (PT) fragment and the fragment is allowed to regener- ate, a morphologically normal worm will result. If, however, the PT fragment is treated immediately post- amputation with β-catenin RNAi [70,71], octonol (8OH), a gap-junction blocker [31], or a depolarizing ionophore [66], a two-headed (2H) phenotype results [complementary with manipulations produce two-tailed phenotypes; see 31,61]. Examination of these 2 H worms reveals that the pharynx has also been duplicated, and the ventral cilia are oriented toward the point of duplication, i.e. in the “posterior” direction from each head [31] as shown in Figure 1a. The nervous system is also duplicated, with both copies functional in directing behavior [72]. Crucially, the VNCs are not only duplicated but are continuous across the duplication point, yielding a nervous system with two complete brains connected by two uninterrupted and apparently fully functional VNCs [31,65,72]. Hence, not only has a head grown from the posterior wound, but the entire anatomy anterior to the anterior-facing wound has been dupli- cated from the posterior wound. The A-P axis has, in other words, been symmetrized to an A-P-A axis around a point at roughly 70% of the worm’s length, as shown in Figure 1b. Figure 1. (a) Cutting a PT fragment from a WT worm and treating with 8OH, a depolarizing ionophore, or β-catenin RNAi yields a dose-dependent 2 H phenotype in which all structures anterior to the anterior-facing wound are duplicated. (b) This transforma- tion resymmetrizes the A-P axis around a point at roughly 70% of the worm’s length, equivalent to acting with abstract operators “Copy70(π)((cid:129))” and “Rotate70((cid:129))” in sequence. 30 C. FIELDS AND M. LEVIN The symmetrization of the A-P axis can be repre- sented geometrically as an abstract rotation by π radians (180°) of a copy of the anterior 70% of the anatomy of the animal (Figure 1b). The axis of rotation is the preserved D-V axis. This “rotation” of the A-P axis is implemented by regenerative growth from regenerative the posterior-facing blastema, while growth from the anterior-facing blastema reproduces the original A-P axis [31,65,72]. Symmetrization of the A-P axis does not erase the distinction between anterior and posterior; to produce a bidirectional A-P-A axis with its midpoint at 70% of the original length. The symmetrized animal has dupli- cated anterior and no posterior anatomy. rather duplicates it it Symmetrized 2 H planaria regenerate to produce 2 H progeny for as many generations as have been observed, indicating a stable alteration of morphology. Intriguingly, the morphologically normal outcomes of 8OH treatment under the above conditions are not wild-type, but are rather “cryptic” worms that continue to regenerate 2 H progeny, at the same percentage as in the original experiment, for multiple rounds of regen- eration in plain water with no further perturbations [73]. The production of 2 H progeny can be reversed by ionophore treatment, indicating that the “memory” for the 2 H morphology is bioelectric. Prima facie similar axis duplication results have been obtained in acoels [45,74] and Hydra [75]; however, the axis- neither multi-generation inheritance of rounds of duplicated phenotype across multiple regeneration or any analog of the “cryptic” phenotype has been demonstrated in these systems. Experiments in which the two VNCs are nicked midway through a PT fragment produce symmetric 4 H worms as shown in Figure 2a [31]. Here again, the entire anterior anatomy is regenerated from the two side nicks, producing two symmetrized A-P axes at right angles. As in 2 H animals, the continuity of the VNCs is preserved, with each of the four brains con- nected by VNCs to the two neighboring brains [31]. This outcome can be represented geometrically as a repeated copy-and-rotate operation as shown in Figure 2b. The effective duplication of a symmetrized A-P axis in the cruciform 4 H animals produced by Oviedo et al. [31] suggests that radially symmetric, hypercephalized outcomes such as sketched in Figure 3a could be pro- duced by making multiple “copies” of the A-P axis and “rotating” them around a central D-V axis. From a geometric point of view, making a large number of copies of the anterior morphology and rotating them in such a way that the heads are evenly spaced is equiva- lent to simply deleting the A-P axis to produce a radially symmetric, completely anterior morphology. Every radial direction from the central D-V axis is, in this case, “anterior”; hence, any regenerative mechan- ism that “anteriorized” the animal in a radially sym- metric way could be expected to yield this outcome. Such radially symmetric, hypercephalized outcomes were observed by Iglesias et al. [32] up to 4 weeks Figure 2. (a) Nicking the two VNCs produces symmetric outcomes with two A-P axes. (b) The outcome can be represented by a repeated copy-and-rotate operation. COMMUNICATIVE & INTEGRATIVE BIOLOGY 31 Figure 3. (a) Radially symmetric, hypercephalized outcome of multiple A-P axis duplication and symmetrization as the number n of duplicates becomes large. Such outcomes have been observed following β-catenin RNAi [32]. (b, c) Radially symmetric, hyperce- phalized outcomes, visualized with synapsin staining, obtained by allowing PT fragments from cryptic worms [73] to regenerate in plain water. d) detail of apparently duplicated circumferential VNC in (c), showing nearly continuous clustering of neurons into apparently proto-cephalic structures [cf. 32, Figure 3]. following β-catenin RNAi treatment of amputation fragments. Consistently with the 2 H and cruciform 4 H regenerates discussed above, these outcomes have a continuous, circumferential “VNC” nerve cord [32]. As predicted, the D-V axis remains unaffected, indicat- ing a lack of significant cross-talk between the A-P (Wnt) and D-V (BMP) axis specification systems. Pharyngeal anatomy is, however, disorganized or lost altogether in these radially symmetric animals, in con- trast to its preservation and apparent function in 2 H and 4 H animals, suggesting that radially symmetric anteriorization disorganizes tissue specification near the “origin” of the radial axis. Eyes with optic nerves are present in association with some, but not all, of the apparently proto-cephalic clusters of neurons distribu- ted roughly uniformly along the circumferential “VNC” [32], suggesting some loss of tissue specification distally along the radial axis. Radially symmetric, hypercephalized outcomes with continuous, circumferential nerve cords (Figure 3b–d) have also been observed following regeneration, in plain water without further perturbation, of PT frag- ments of cryptic worms [73]. The VNCs appear to be duplicated in some of these regenerates (Figure 3c,d). Neither pharyngeal structures nor eyes are observed in these preparations. These observations suggest that the bioelectric changes that define the cryptic phenotype can have far-reaching and variable consequences for regenerative morphology that await further, detailed investigation. In summary, the planarian A-P axis appears to be not just symmetrizable, but highly manipulable, in regenera- tion-based assays. Both molecular (e.g. β-catenin RNAi) and bioelectric (8OH, ionophore) manipulations can lead to axis symmetrization and duplication. The cryptic phenotype identified with 8OH treatment is the first known example of reversible, bioelectric, epigenetic inheritance [73]; a similar reversible bioelectric manip- ulation has now also been demonstrated in Hydra [76]. In the absence of rescue manipulations, the altered phe- notypes are stable across multiple generations in viable individual planaria, and may be permanent. These results suggest that while the A-P axis is “primary” in planaria as in other bilaterians, it is in a highly plastic state that may reflect loss of evolved constraints on the standard bilaterian body plan. The planarian A-P axis as a transitional state In cnidarians, Wnt pathway components including the Disheveled (Dsh) receptor and β-catenin effector are expressed in a decreasing gradient from the Oral to the Aboral pole [18,19]. Let us call this the “Wnt – anti- Wnt” axis, where here “anti-Wnt” refers to either a Wnt inhibitor or opposite-pole determinant. The Wnt – anti-Wnt axis is aligned with the gut in cnidar- ians, orthogonal to the circumoral nerve ring, and aligned with the long axis of the nerve net driving whole-body contractions and motility [77]. 32 C. FIELDS AND M. LEVIN In bilaterians, Wnt pathway components are expressed in a decreasing gradient from posterior to anterior. In bilaterians possessing a through-gut, the Wnt – anti-Wnt axis is aligned with the gut and the nerve cord(s) extending from the anterior brain. The anterior, anti-Wnt direction is the direction of both motion and the mouth. The structure of the pharynx and hence the loca- tion and orientation of the combined mouth/anus is variable in both acoels and flatworms. In the pla- naria of interest here, the mouth opens ventrally, aligned with the D-V axis [78], along which the pharynx can also be extended as sketched in Figure 4a. The planarian blind gut has the orientation with respect to the D-V axis that the cnidarian blind gut has with respect to the O-A axis. Symmetrizing the A-P axis in planaria duplicates the mouth and phar- ynx while maintaining their D-V orientation (Figure 4b). shown in Figure 3, the D-V axis has become the “primary” body axis about which anatomical structures are symmetry in this the radial radially symmetric; case is analogous to the radial symmetry of cnidar- ians around the O-A axis. In the radially symmetric forms From this perspective, the idea of a “primary axis” appears somewhat ambiguous in planaria. Known manip- ulations of the D-V axis, moreover, are neither as thor- ough-going or as extensive as the A-P manipulations reviewed here [11]; no D-V analogs of the multiple A-P duplications shown in Figure 2 or 3 are known. Topologically, the planarian A-P axis is analogous to the cnidarian directive (secondary) axis; each specifies two opposing “sides” of a central, invaginated body cavity. Could the plasticity of the planarian A-P axis reflect an ancestral state in which this axis was secondary, as the directive axis is in cnidarians? Reconstructing an ancestral eumetazoan Deep metazoan phylogeny remains highly controver- sial, with active disagreement about whether porifera or ctenophores are more basal and considerable uncer- tainty about the placement of placazoa [79–81]. All empirical phylogeny, however, equally suffers from the problem that only extant (or well-preserved fossil) species are accessible for analysis. An empirically informed theoretical phylogeny may therefore have value in considering questions of eumetazoan ancestry. Standard models of the emergence of animal multicel- lularity are based on the aggregation of closely related cells [82], typically choanoflagellates [1–3]. We have recently proposed an alternative, non-aggregative model in which ancestral, free-living stem cells produce a protective “body” comprising their own reproductively disabled progeny as a means of self-defense in a challenging environment [83]. The principal regulator in this scenario is a “do not pro- liferate” (DNP) signal that the parent stem cell employs to shut down proliferative capability in its progeny, rendering them fully “somatic” cells with no independent genetic interests or fitness. As Wnt-pathway components are already used for proliferation suppression of prestalk cells in Dictyostelium [84,85], it is plausible on phylogenetic grounds that this DNP signal may be a Wnt or Wnt analog. The DNP signal is assumed to be secreted only by prolif- erative stem cells and to be short-range; hence its distribu- tion within a multicellular system will depend on whether the system’s proliferative cells are dispersed or clustered, as sketched in Figure 5a,b. A primitive organism comprising Figure 4. (a) The planarian mouth opening is aligned along the D-V axis, with respect to which the blind gut has the radial symmetry of the blind gut in cnidarians. (b) Symmetrizing the A-P axis duplicates the mouth-opening axis while preserving its D-V orientation. COMMUNICATIVE & INTEGRATIVE BIOLOGY 33 Figure 5. (a) An ancestral proliferative cell produces progeny for protection, employing a short-range “do not proliferate” (DNP) signal to suppress their proliferation. (b) A somatic cell layer enclosing dispersed proliferate cells has uniform [DNP]; if proliferative cells cluster, [DNP] is non-uniform. (c) A primitive organism comprising a cell layer enclosing dispersed proliferative cells is stable; one enclosing clustered proliferative cell has insufficient [DNP] to prevent rogue proliferation at its margins, so is not stable. a cellular envelope enclosing a uniform distribution of dispersed proliferative cells may be expected, assuming cell-cycle synchronization or some other mechanism to coordinate stem-cell proliferation, to have relatively uni- form internal [DNP] and to be stable. However, such an organism with clustered proliferative cells is expected to have non-uniform internal [DNP] and to be unstable due to uncontrolled reproduction by “somatic” cells in which independent proliferation has not been fully suppressed (Figure 5c). Reproductive stability is possible with clustered prolif- erative cells if longer-range communication of a DNP-like signal that suppresses rogue cell division by somatic cells is possible. Neurons provide an ideal solution to this problem, as they allow long-range, error-correcting com- munication between source cells and specific target cells [86]. Neurons likewise provide a means of coordinating the proliferation of dispersed populations of stem cells that are too far apart or too distant within a cell lineage to be reproductively coordinated by other mechanisms. An 34 C. FIELDS AND M. LEVIN Figure 6. (a, b) A reproductively unstable system can achieve stability by employing neurons to transmit a DNP-like signal (green curves) to distant somatic cells in order to suppress rogue cell division. (c) Neurons enable the development of complex anatomies, e.g. invaginated body cavities. organism with neurons can adopt a more complex body plan, e.g. by elongating its periphery into an invagination as sketched in Figure 6. The three extant animal lineages with complex body plans – the ctenophores, cnidarians, and bilaterians – all have neurons. Structural, biochem- ical, and molecular differences between ctenophore neu- rons and those of cnidarians and bilaterians suggest convergent evolution to a common function [87]. We have suggested that the primary ancestral function of neurons in all three lineages is the long-distance control of cell proliferation that enables a stable multicellular morphology even with clustered stem cells [86]. While this hypothesis remains to be tested, manipulations in Xenopus embryos provide initial evidence for CNS regu- lation of distal morphogenesis [88,89]. These theoretical considerations suggest an interpreta- tion of the radially symmetric, hypercephalized regenera- tion outcomes shown in Figure 3 as regressions toward an ancestral state, one that may pre-date not only planaria but eumetazoa in general. Only one extant animal lineage comprises flat, approximately radially symmetric organ- isms: the placozoa [5,90]. Placozoa have no differentiated neurons but have neurotransmitters and behavior. The dispersed, interior “fiber” cells of placozoa are extended and may serve a communicative function, e.g. by para- crine signaling as in sponges [91]. Characterized placozoa do not have differentiated mouths or guts, but appear to digest food externally and absorb nutrients through dis- tributed ventral-surface cells [5]. Placozoa are predomi- nately asexual, reproducing by budding or fission with WBR, but also exhibit opportunistic sexuality. They are often regarded as amorphous but have a primary D-V axis, the axis normal to the substrate, around which they are approximately radially symmetric. Their anatomy resembles the left side of Figure 5c. Do the radially symmetric, gutless, hypercephalized regeneration outcomes in Figure 3b, c resemble placozoa with neurons added? While gross morphology suggests that placozoa may be their closest affinity, further investi- gation of both parties is clearly required to answer this question. If these planarian regenerates indeed resemble placozoa at more than a superficial level, they may be pointing toward an ancestral eumetazoan with radial sym- metry around a primary D-V axis, a blind or undifferen- tiated gut, a rudimentary circumferential nerve cord with radial branching, and asexual reproduction with WBR. Conclusion We have suggested here that WBR provides an alternative to embryology for studying the mechanisms of body-axis specification and their contributions to the evolution of complex morphologies. Asexual planaria appear to be COMMUNICATIVE & INTEGRATIVE BIOLOGY 35 particularly attractive model systems in this regard. The A-P axis of planaria, in particular, is highly malleable using molecular, pharmacological, and bioelectric manipulations. This axis can not only be symmetrized but also duplicated to such an extent that it effectively disappears, leaving a radially symmetric, fully anteriorized form. Whether the A-P axis of acoels, or of other bilaterians, is similarly manipulable remains to be seen; commonalities in axis- specification mechanisms across the bilaterians as well as specific results in acoels [45,74] suggest that they may be. The evolutionary emergence of complex body plans appears intimately connected to the emergence of neu- rons as specialized long-range signaling systems [86]. We suggest that the emergence of neurons in a radially sym- metric, placozoan-like animal may have set the stage for the differentiation of the eumetazoan lineages. Further work is clearly required to elaborate and test these hypotheses. Life-history studies of the symmetrized forms shown in Figure 3b, c have already been initiated; if such forms can be reliably produced and maintained, functional investigation of their nervous and digestive systems will be possible. The results of Müller [75] and Braun and Ori [76] suggest that Hydra may be an attrac- tive Cnidarian model system with which to pursue similar axis-symmetrization studies. Thorough investigation of cell-cell signaling mechanisms in Placozoa would comple- ment these investigations. More broadly, the study of the regulation of morphogenesis outside of the nervous sys- tem per se by neural activity remains in its infancy. Recent as well as classical evidence of regulation of regeneration [92] and of transformation and tumor growth by the nervous system [93–95] suggests that such studies may also have significant clinical relevance. Author Contributions CF and Ml jointly conceived of the ideas and wrote the paper. Acknowledgments We gratefully acknowledge support by an Allen Discovery Center award from the Paul G. Allen Frontiers Group (No. 12171), and the Templeton World Charity Foundation (No. TWCF0089/AB55). Disclosure statement No potential conflicts of interest were disclosed. Funding This work was supported by the Paul G. Allen Frontiers Group [12171];Templeton World Charity Foundation [TWCF0089/AB55]. 36 C. FIELDS AND M. LEVIN ORCID Chris Fields Michael Levin http://orcid.org/0000-0002-4812-0744 http://orcid.org/0000-0001-7292-8084 References [1] Nielsen C. Six major steps in animal evolution: are we derived sponge larvae?. Evol Devel. 2008;10:241–257. [2] Funayama N. The stem cell system in demosponges: insights into the origin of somatic stem cells. Devel Growth Diff. 2010;52:1–14. [3] Arendt D, Benito-Gutierrez E, Brunet T, et al. Gastric pouches and the mucociliary sole: setting the stage for nervous system evolution. Philos Trans Royal Soc B. 2015;370:20150286. [4] Jauffred L, Vejborg RM, Korolev KS, et al. Chirality in microbial biofilms is mediated by close interactions between the cell surface and the substratum. Isme J. 2017;11:1688–1701. [5] Miller DJ, Ball EE. Animal evolution: the enigmatic phy- lum Placozoa revisited. Curr Biol. 2005;15:R26–R28. [6] Ereskovsky AV, Renard E, Borchiellini C. Cellular and molecular processes leading to embryo formation in sponges: evidences for high conservation of processes throughout animal evolution. Devel Genes Evol. 2013;223:5–22. [7] Martindale MQ. The onset of regenerative properties in ctenophores. Curr Opin Genet Devel. 2016;40:113–119. [8] Slack JMW. Animal regeneration: ancestral character or evolutionary novelty? EMBO Rep. 2017;18:1497–1508. [9] Lai AG, Aboobaker AA. EvoRegen in animals: time to uncover deep conservation or convergence of adult stem cell evolution and regenerative processes. Dev Biol. 2018;433:118–131. [10] Tiozzo S, Copley RR. Reconsidering regeneration in metazoans: an evo-devo approach. Front Ecol Evol. 2015;3:67. [11] Elliott SA, Sánchez Alvarado A. The history and endur- ing contributions of planarians to the study of animal regeneration. WIRES Devel Biol. 2012;2:301–326. [12] Rink JC. Stem cells, patterning and regeneration in planarians: self-organization at the organismal scale. In: Rink JC, editor. Planarian regeneration: methods and protocols (Methods in Molecular Biology). Vol. 1774. New York (NY): Springer; 2018. p. 57–172. [13] Adamska M. Developmental signalling and emergence of I, Nedelcu AM, editors. Evolutionary transitions to mul- ticellular life. Dordrecht: Springer; 2015. p. 425–450. DOI:10.1007/978-94-017-9642-2_20 animal multicellularity. Ruiz-Trillo In: [14] Niehrs C. On growth and form: a Cartesian coordinate system of Wnt and BMP signaling specifies bilaterian body axes. Development. 2010;137:845–857. [15] Sebé-Pedrós A, Degnan BM, Ruiz-Trillo I. The origin of Metazoa: a unicellular perspective. Nat Rev Genet. 2017;18:498–512. [16] Tweedt SM, Erwin DH. Origin of metazoan develop- toolkits and their expression in the fossil editors. In: Ruiz-Trillo I, Nedelcu AM, life. mental record. Evolutionary to multicellular transitions Dordrecht: Springer; 2015. p. 47–78. DOI:10.1007/ 978-94-017-9642-2_3 [17] Loh KM, van Amerongen R, Nusse R. Generating cellular diversity and spatial form: Wnt signaling and the evolution of multicellular animals. Devel Cell. 2016;38:643–655. [18] Martindale MQ, Hejnol A. A developmental perspec- tive: changes in the position of the blastopore during bilaterian evolution. Devel Cell. 2009;17:162–174. [19] Petersen CP, Reddien PW. Wnt signaling and the polarity of the primary body axis. Cell. 2009;139:1056–1068. [20] Genikhovich G, Fried P, Prünster MM, et al. Axis patterning by BMPs: cnidarian network reveals evolu- tionary constraints. Cell Rep. 2015;10:1646–1654. [21] He S, Del Viso F, Chen C-Y, et al. An axial Hox code controls tissue segmentation and body patterning in Nematostella vectensis. Science. 2018;361:1377–1380. [22] Layden MJ, Rentzsch F, Röttinger E. The rise of the starlet sea anemone Nematostella vectensis as a model system to investigate development and regeneration. WIRES Devel Biol. 2016;5:408–428. [23] Wijesena N, Simmons DK, Martindale MQ. Antagonistic BMP–cWNT signaling in the cnidarian Nematostella vectensis reveals insight into the evolution of mesoderm. Proc Natl Acad Sci USA. 2017;114:E5608–E5615. [24] Neville AC. Animal asymmetry. London: Edward Arnold; 1976. [25] Palmer AR. Animal asymmetry. Curr Biol. 2009;19: R473–R477. B, [26] Gavilán Perea-Atienza P. Xenacoelomorpha: a case of independent nervous system centralization? B. 2016;371:20150039. E, Martínez Philos Trans Royal Soc [27] Hejnol A, Rentzsch F. Neural nets. Curr Biol. 2015;25: R775–R792. [28] Martínez P, Perea-Atienza E, Gavilán B, et al. The study of xenacoelomorph nervous systems: molecular and mor- phological perspectives. Invert Zool. 2017;14:32–44. [29] Gruhl A, Okamura B. Development and myogenesis of the vermiform Buddenbrockia (Myxozoa) and implications for cnidarian body plan evolution. EvoDevo. 2012;3:10. [30] Jiménez-Guri E, Philippe H, Okamura B, et al. Science. cnidarian worm. a Buddenbrockia is 2007;317:116–118. [31] Oviedo NJ, Morokuma J, Walentek P, et al. Long-range neural and gap junction protein-mediated cues control polarity during planarian regeneration. Dev Biol. 2010;339:188–199. [32] Iglesias M, Gomez-Skarmeta JL, Saló E, et al. Silencing of Smed-βcatenin1 generates radial-like hypercepha- lized planarians. Development. 2008;135:1215–1221. [33] Erwin DH. Early metazoan life: divergence, environment B. ecology. Philos Trans Royal Soc and 2015;370:20150036. [34] Jondelius U, Ruiz-Trillo I, Baguñà J, et al. The Nemertodermatida are basal bilaterians and not members of the Platyhelminthes. Zool Scripta. 2002;31:201–215. [35] Giribet G. Assembling the lophotrochozoan (=spira- life. Philos Trans Royal Soc B. lian) tree of 2008;363:1513–1522. [36] Hejnol A, Obst M, Stamatakis A, et al. Assessing the root of bilaterian animals with scalable phylogenomic methods. Proc R Soc B. 2009;276:4261–4270. [37] Dunn CW, Giribet G, Edgecombe GD, et al. Animal phylogeny and its evolutionary implications. Annu Rev Ecol Evol Syst. 2014;45:371–395. [38] Cannon JT, Vellutini BC, Smith J al. Xenacoelomorpha is the sister group to Nephrozoa. Nature. 2016;530:89–93. III, et [39] Egger B, Steinke D, Tarui H, et al. To be or not to be a flatworm: the Acoel controversy. PLoS ONE. 2009;4(5): e5502. [40] Bouriat SJ, Hejnol A. Acoels. Curr Biol. 2009;19:R279– R280. et [41] De Mulder K, Kuales G, Willems M, al. Characterization of the stem cell system of the acoel Isodiametra pulchra. BMC Devel Biol. 2009;9:69. [42] Perea-Atienza E, Botta M, Salvenmoser W, et al. pulchra Posterior (Acoela, Acoelomorpha). Front Zool. 2013;10:64. [43] Gehrke AR, Neverett E, Luo Y-J, et al. Acoel genome landscape of whole-body reveals regeneration. Science. 2019;363:eaau6173. regeneration Isodiametra regulatory the in [44] Raz AA, Srivastava M, Salvamoser R, et al. Acoel regeneration mechanisms indicate an ancient role for muscle in regenerative patterning. Nat Comms. 2017;8:1260. [45] Srivastava M, Mazza-Curll KL, van Wolfswinkel JC, et al. Whole-body acoel regeneration is controlled by Wnt and Bmp-Admp signaling. Curr Biol. 2014;24:1107–1113. JM, Saló E, A, In: Wanninger [46] Adell T, Martín-Durán et al. Platyhelminthes. editor. Evolutionary developmental biology of invertebrates 2: lophotrochozoa (Spiralia). Wein: Springer; 2015. p. 21–40. DOI:10.1007/978-3-7091-1871-9_3 [47] Hejnol A. Acoelomorpha and Xenoturbellida. In: Wanninger A, editor. Evolutionary developmental biol- introduction, non-bilateria, ogy of acoelomorpha, xenoturbellida, chaetognatha. Wein: Springer; 2015. p. 203–214. DOI:10.1007/978-3-7091- 1862-7_9 invertebrates 1: [48] Petralia RS, Mattson MP, Yao PJ. Aging and longevity in the simplest animals and the quest for immortality. Ageing Res Rev. 2014;16:66–82. [49] Sköld HN, Obst M, Sköld M, et al. Stem cells in asexual reproduction of marine invertebrates. In: Rinkevich B, Matranga V, editors. Stem cells in marine organisms. Berlin: Springer; 2009. p. 105–137. DOI:10.1007/978- 90-481-2767-2_5 [50] Lively CM. A review of Red Queen models for the persistence of obligate sexual reproduction. J Hered. 2010;101:S13–S20. [51] Maldonado M, Riesgo A. Reproduction in the phylum Porifera: a synoptic overview. Treb Soc Catalana Biol. 2008;59:29–49. [52] Watanabe H, Hoang VT, Mättner R, et al. Immortality and the base of multicellular life: lessons from cnidar- ian stem cells. Sem Cell Devel Biol. 2009;20:1114–1125. [53] Solana J. Closing the circle of germline and stem cells: the primordial stem cell hypothesis. EvoDevo. 2013;4:2. [54] Vila-Farré M, Rink JC. The ecology of freshwater pla- narians. In: Rink JC, editor. Planarian regeneration: methods and protocols. New York (NY): Humana; 2018. p. 173–205. DOI:10.1007/978-1-4939-7802-1_3 COMMUNICATIVE & INTEGRATIVE BIOLOGY 37 [55] Hoshi M, Kobayashi K, Arioka S, et al. Switch from asexual to sexual reproduction in the planarian Dugesia ryukyuensis. Integr Comp Biol. 2003;43:242–246. [56] Kobayashi K, Koyanagi R, Matsumoto M, et al. Switching from asexual to sexual reproduction in the planarian Dugesia ryukyuensis: bioassay system and basic description of sexualizing process. Zool Sci. 1999;16:291–298. [57] Fields C, Levin M. Are planaria individuals? What regenerative biology is telling us about the nature of multicellularity. Evol Biol. 2018;45:237–247. [58] Newmark PA, Sánchez Alvarado A. Not your father’s planarian: a classic model enters the era of functional genomics. Nat Rev Genet. 2002;3:210–219. [59] Malinowski PT, Cochet-Escartin O, Kaj KJ, et al. Mechanics dictate where and how freshwater planarians fission. Proc Natl Acad Sci USA. 2017;114:10888–10893. [60] Arnold CP, Bentham-Pyle BW, Lange JJ, et al. Wnt and TGFβ coordinate growth and patterning to regulate size-dependent behaviour. Nature. 2019;572:655–659. [61] Lobo D, Beane WS, Levin M. Modeling planarian regeneration: a primer for reverse-engineering the worm. PLoS Comp Biol. 2012;8:ee1002481. [62] Lobo D, Levin M. Inferring regulatory networks from experimental phenotypes: morphological a computational method reverse-engineers planarian regeneration. PLoS Comp Biol. 2015;11:e1004295. [63] Owlarn S, Bartscherer K. Go ahead, grow a head! regeneration. anterior A planarian’s Regeneration. 2016;3:139–155. guide to [64] Petersen CP, Reddien PW. Polarized notum activation at wounds inhibits Wnt function to promote planarian head regeneration. Science. 2011;332:852–855. [65] Pietak A, Bischof J, LaPalme J, et al. Neural control of body-plan axis in regenerating planaria. PLoS Comp Biol. 2019;15:e1006904. [66] Durant F, Bischof J, Fields C, et al. The role of early bioelectric signals in the regeneration of planarian ante- rior/posterior polarity. Biophys J. 2019;116:948–961. [67] Yazawa S, Umesono Y, Hayashi T, et al. Planarian Hedgehog/Patched establishes anterior-posterior polar- ity by regulating Wnt signaling. Proc Natl Acad Sci USA. 2009;106:22329–22334. [68] Fraguas S, Barberán S, Iglesias M, et al. egr-4, a target of EGFR signaling, is required for the formation of the brain primordia and head regeneration in planarians. Development. 2014;141:1835–1847. [69] Iglesias M, Almuedo-Castillo M, Aboobaker AA, et al. Early planarian brain regeneration is independent of blastema polarity mediated by the Wnt/β-catenin path- way. Devel Biol. 2011;358:68–78. [70] Gurley KA, Rink JC, Sánchez Alvarado A. Beta-catenin defines head versus tail identity during planarian regen- eration and homeostasis. Science. 2008;319:323–327. [71] Petersen CP, Reddien PW. Smed-betacatenin-1 is required for anteroposterior blastema polarity in pla- narian regeneration. Science. 2008;319:327–330. [72] Bischof J, Day ME, Miller KA, et al. Nervous system and tissue polarity dynamically adapt to new morphol- ogies in planaria. Preprint bioRXiv:815688. 2019. DOI:10.1101/815688 38 C. FIELDS AND M. LEVIN [73] Durant F, Morokuma J, Fields C, et al. Long-term, sto- chastic editing of regenerative anatomy via targeting J. endogenous 2017;112:2231–2243. bioelectric gradients. Biophys [74] Sikes JM, Bely AE. Making heads from tails: develop- ment of a reversed anterior-posterior axis during bud- ding in an acoel. Devel Biol. 2010;338:86–97. [75] Müller WA. Head formation at the basal end and mirror-image pattern duplication in Hydra vulgaris. Int J Devel Biol. 1996;40: 1119–1141. PMID: 9032017. [76] Braun E, Ori H. Electric-induced reversal of morpho- genesis in Hydra. Biophys J. 2019;117:1514–1523. [77] Watanable H. Back through time: how cnidarians and basal metazoans shed light on ancient nervous systems. In: Shigeno S, et al., editor. brain evolution by design, diversity and commonality in animals. Tokyo: Springer; 2017. p. 45–75. DOI:10.1007/978-4-431-56469-0_3 [78] Sluys R, Riutort M. Planarian diversity and phylogeny. In: Rink JC, editor. Planarian regeneration: methods and protocols (Methods in Molecular Biology). Vol. 1774. New York (NY): Springer; 2018. p. 1–56. [79] Feuda R, Dohrmann M, Pett W, et al. Improved model- ing of compositional heterogeneity supports sponges as sister to all other animals. Curr Biol. 2017;37:3864–3870. [80] Simion P, Philippe H, Baurain D, et al. A large and consistent phylogenomic dataset supports sponges as the sister group to all other animals. Curr Biol. 2017;27:958–967. [81] Whelan NV, Kocot KM, Moroz TP, et al. Ctenophore relationships and their placement as the sister group to all other animals. Nat Evol Ecol. 2017;1:1737–1746. [82] Hanschen ER, Shelton DE, Michod RE. Evolutionary tran- sitions in individuality and recent models of multicellular- ity. In: Ruiz-Trillo I, Nedelcu AM, editors. Evolutionary transitions to multicellular life. Dordrecht: Springer; 2015. p. 165–188. DOI:10.1007/978-94-017-9642-2_9 [83] Fields C, Levin M. Somatic multicellularity as a satisficing solution to the prediction-error minimiza- tion problem. Commun Integr Biol 2019;12:119–132. [84] Morris HR, Masento MS, Taylor GW, et al. Structure two differentiation inducing factors slime Biochem J. elucidation of (DIF-2 mould Dictyostelium discoideum. 1988;249:903–906. and DIF-3) from the cellular [85] Morris HR, Taylor GW, Masento MS, et al. Chemical structure of the morphogen differentiation inducing from Dictyostelium discoideum. Nature. factor 1987;328:811–814. [86] Fields C, Bischof J, Levin M. Morphological coordina- tion: a common ancestral function unifying neural and non-neural signaling. Physiology. 2020;35:16–30. [87] Moroz LL, Kohn AB. Independent origins of neurons and synapses: insights from ctenophores. Philos Trans R Soc Lond B. 2016;371:20150041. [88] Herrera-Rincon C, Levin M. Booting up the organism during development: prebehavioral functions of the vertebrate brain in guiding body morphogenesis. Commun Integr Biol. 2018;11:e1433440. [89] Herrera-Rincon C, Pai VP, Moran KM, et al. The brain is required for normal muscle and nerve patterning during early Xenopus development. Nat Commun. 2017;8:587. [90] Schierwater B, DeSalle R. Placozoa. Curr Biol. 2017;28: R97–R98. [91] Nickel M. Evolutionary emergence of synaptic nervous systems: what can we learn from the non-synaptic, nerveless Porifera? Invert Biol. 2010;129:1–16. [92] Kumar A, Brockes JP. Nerve dependence in tissue, organ, and appendage regeneration. Trends Neurosci. 2012;35:691–699. [93] Boilly B, Faulkner S, Jobling P, et al. Nerve depen- from regeneration to cancer. Cancer Cell. dence: 2017;31:342–354. [94] Kuol N, Stojanovska L, Apostolopoulos V, et al. Role of the nervous system in cancer metastasis. J Exp Clin Cancer Res. 2018;37:5. [95] Pawlowski A, Weddell G. Induction of tumors in denervated skin. Nature. 1967;213:1234–1236.
10.1111_jopy.12803
Received: 5 June 2022 DOI: 10.1111/jopy.12803 | Revised: 18 November 2022 | Accepted: 17 December 2022 O R I G I N A L A R T I C L E Proximal and distal honor fit and subjective well- being in the Mediterranean region | Felix D. Schönbrodt3 | Ayse K. Uskul1,2 | | Rosa Rodríguez- Bailón4 | Vanessa A. Castillo5,6 | Alexander Kirchner- Häusler1,2 Vivian L. Vignoles2 Susan E. Cross6 Keiko Ishii11 Evangelia Kateri12 | Zafer Özkan16 Jinkyung Na15 Dina Rabie19 | Manuel Teresi17 | Meral Gezici- Yalçın7 | Panagiota Karamaouna12 | Juan Matamoros- Lima4 | Charles Harb8,9 | Shenel Husnu10 | | Konstantinos Kafetsios13 | Rania Miniesy14 | | | Stefano Pagliaro17 | Yukiko Uchida20 | Charis Psaltis18 | 1Department of Psychology, University of Kent, Canterbury, UK 2Department of Psychology, University of Sussex, Brighton, UK 3Department of Psychology, Ludwig- Maximilians- Universität, Munich, Germany 4Department of Psychology, University of Granada, Granada, Spain 5Department of Psychology, Coe College, Cedar Rapids, Iowa, USA 6Department of Psychology, Iowa State University, Ames, Iowa, USA 7Department of Psychology, Bolu Abant İzzet Baysal University, Bolu, Turkey 8Department of Psychology, American University of Beirut, Beirut, Lebanon 9Department of Psychology, Doha Institute, Doha, Qatar 10Department of Psychology, Eastern Mediterranean University, Famagusta, Cyprus 11Department of Cognitive and Psychological Sciences, Nagoya University, Nagoya, Japan 12Department of Psychology, University of Crete, Crete, Greece 13Department of Psychology, Aristotle University of Thessaloniki, Thessaloniki, Greece 14Department of Economics, British University of Egypt, Cairo, Egypt 15Department of Psychology, Sogang University, Seoul, South Korea 16Department of Psychology, Ordu University, Ordu, Turkey 17Department of Neuroscience, Imaging and Clinical Sciences, University di Chieti- Pescara, Chieti, Italy 18Department of Psychology, University of Cyprus, Nicosia, Cyprus 19Nottingham Business School, Nottingham Trent University Business School, Nottingham, UK 20Kokoro Research Center, Kyoto University, Kyoto, Japan Correspondence Alexander Kirchner- Häusler, School of Psychology, University of Sussex, Pevensey 1 Building, Falmer, BN1 9QH, UK. Email: a.kirchnerhausler@sussex.ac.uk Abstract Objective: People's psychological tendencies are attuned to their sociocultural context and culture- specific ways of being, feeling, and thinking are believed to assist individuals in successfully navigating their environment. Supporting this idea, a stronger “fit” with one's cultural environment has often been linked to This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. © 2022 The Authors. Journal of Personality published by Wiley Periodicals LLC. Journal of Personality. 2023;00:1–17. wileyonlinelibrary.com/journal/jopy | 1 14676494, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jopy.12803 by Universidad De Granada, Wiley Online Library on [09/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 2 | Funding information European Research Council positive psychological outcomes. The current research expands the cultural, conceptual, and methodological space of cultural fit research by exploring the link between well- being and honor, a central driver of social behavior in the Mediterranean region. Method: Drawing on a multi- national sample from eight countries circum- Mediterranean (N = 2257), we examined the relationship between cultural fit in honor and well- being at the distal level (fit with one's perceived society) using re- sponse surface analysis (RSA) and at the proximal level (fit with one's university gender group) using profile analysis. Results: We found positive links between fit and well- being in both distal (for some, but not all, honor facets) and proximal fit analyses (across all honor fac- ets). Furthermore, most fit effects in the RSA were complemented with positive level effects of the predictors, with higher average honor levels predicting higher well- being. Conclusions: Our findings highlight the interplay between individual and envi- ronmental factors in honor as well as the important role honor plays in well- being in the Mediterranean region. K E Y W O R D S fit, honor concerns, honor values, Mediterranean region, profile fit, response surface analysis, well- being 1 | INTRODU CT ION Scientific evidence accumulated over the last three decades has clearly demonstrated that individuals' psychologi- cal tendencies are attuned to their sociocultural context. For example, individuals in different cultural groups dif- fer systematically in their models of selfhood (Vignoles et al., 2016), emotional (De Leersnyder et al., 2021), and cognitive processes (Nisbett et al., 2001). Culture- specific ways of being, feeling, and thinking are believed to assist individuals in successfully navigating their sociocultural environment (Kitayama & Uskul,  2011). Following this reasoning, a stronger fit between people's psychological make- up and characteristics of their sociocultural en- vironment (to which we refer as “cultural fit”) is often assumed to be associated with better well- being. Past re- search has supported this assumption by demonstrating positive consequences of cultural fit in different psycho- logical domains, including emotional experience (De Leersnyder, 2017), personality (Fulmer et al., 2010), life- style and social support (Dressler, 2012), and internalized cultural norms (Stephens et al., 2012). The current study contributes to existing research on the psychological con- sequences of cultural fit by expanding its cultural, concep- tual, and methodological space. Using a multi- national sample from the Mediterranean region and adopting a multi- method approach, we examined the relationship between cultural fit and subjective well- being focusing on a cultural construct central to this region: endorsement of honor values and concerns as guiding principles in indi- viduals' social life. 1.1 | Honor Honor has been established as a core value and salient driver of social behavior in different regions of the world including the Mediterranean, the Middle East, Latin America, South Asia, and the Southern U.S. (for reviews see Cross & Uskul,  2022; Uskul & Cross,  2019; Uskul et al.,  2019). Honor has been described as “the value of a person in his own eyes, but also in the eyes of society” (Pitt- Rivers,  1965, p. 21), reflecting the central idea that in cultural groups that emphasize honor an individual's worth is not only self- defined (e.g., to be proud of one's personal accomplishments) or claimed, but also defined in terms of one's reputation and status bestowed by others (e.g., to be known by others as a respectable and moral person) (Cross et al., 2014; Cross & Uskul, 2022). The posi- tive self- view of a person thus combines both intra- and in- terpersonal elements, and this distinct combined focus on both the personal and social image has often been taken as a defining characteristic of so- called “honor cultures” compared with Western and East Asian cultural contexts KIRCHNER-­HÄUSLER­et­al. 14676494, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jopy.12803 by Universidad De Granada, Wiley Online Library on [09/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License (Leung & Cohen, 2011). What further characterizes honor is its relational nature where threats to or enhancements of honor can have direct positive or negative implications for close others or social groups affiliated with the per- son, shaping their honor in their own eyes and the eyes of others (e.g., Korteweg & Yurdakul,  2009; Rodriguez Mosquera et al., 2008; Uskul et al., 2012). Although Pitt- Rivers did not clarify the dimensions upon which individuals base their valuation in his defi- nition of honor, others have identified culturally specific moral codes, gender roles, and economic and social status as the primary sources of these valuations in cultures of honor (Campbell, 1964; Péristiany, 1965). Thus, in order to be considered an honorable person in cultures where honor values are prevalent, an individual has to exhibit certain traits (e.g., morality, virtue, strength) and actively claim one's right to honor, but also has to develop a keen sense of their social reputation and conform to prescribed norms and behaviors (e.g., defending oneself against in- sults, protecting family reputation). Some dimensions of honor may hold equal importance for all members of honor cultures (e.g., family reputation, moral integrity), whereas others often emphasize different expectations for men versus women (e.g., strength and authority within the family for men, sexual purity and loyalty to men and family for women). These norms can restrict behavior in wide- reaching ways: deviations from the prescribed honor code can trigger strong opposition by other group mem- bers, as dishonorable behavior often has dire implications for the reputation of oneself and close others. As such, in- dividuals in honor cultures work toward promoting a pos- itive social image as well as staying vigilant toward honor threats that may stain their own or the honor of close oth- ers, as honor is hard to gain, but easy to lose (for reviews, see Bowman, 2006; Brown, 2016; Nisbett & Cohen, 1996; Uskul et al., 2019). Studies into honor in social psychology, criminology, and law have so far focused primarily on (interpersonal) retaliation following honor threats, highlighting the neg- ative consequences of honor. For example, Nisbett and Cohen  (1996) have demonstrated the interpersonal and institutional emphases on honor in the Southern U.S. through the study of interpersonal aggression, homicide rates, and legal decisions. Similarly, in the decades follow- ing, honor research has focused on the negative aspects of honor dynamics in domains such as intimate partner vio- lence (Baldry et al., 2013), risk- taking (Barnes et al., 2012), (delayed or lack of) health care seeking (Foster et al., 2022) and school shootings (Brown et al., 2009). In comparison, positive aspects of honor dynamics have received less at- tention, with the exception of a few studies that examined the role of politeness (Cohen et al.,  1999), moral behav- ior (Cross et al., 2014) and reciprocity of positive behavior | 3 (e.g., favors, hospitality, Leung & Cohen, 2011). The cur- rent research further contributes to filling in this gap in the literature by focusing on the implications of honor for well- being when there is a fit between individuals' own honor endorsement and honor endorsement by others in one's proximal and distal social environment. 1.2 | Cultural fit Cultures have often been described as “systems of mean- ing”, conceptual systems that are shared between mem- bers of groups and that organize beliefs, values, and practices in a given society (Markus & Kitayama,  2010). Yet, although all individuals exist within a cultural envi- ronment, they vary in the extent to which they endorse the culturally dominant ways of being and acting (Leung & Cohen,  2011). Cultural fit represents this relationship between an individual and their social environment, re- flecting the “process of thinking and acting in ways that are aligned with the thoughts and behavioral expectations of members of a social group” (Mobasseri et al.,  2019, p. 305). As such, it goes beyond the comparison of mere cultural prototypes or averages (see Leung & Cohen, 2011) and can offer an insightful way to examine the psychologi- cal consequences of individual variation within cultural groups. A major idea underlying most cultural fit research is that fitting in relatively more with one's environment is linked with more positive outcomes, as a stronger fit should support individuals in successfully navigating the central demands and tasks of their social environment (Kitayama et al., 2010), provide them with resources to de- code and understand others' and their own behavior better (Edwards & Cable, 2009), and foster feelings of belonging by highlighting similarities between themselves and oth- ers (e.g., Hogg & Terry, 2000). Past research has supported this idea across a wide variety of domains and outcomes. For example, a stronger fit with culturally dominant pat- terns of emotion is associated with better relational well- being (De Leersnyder et al.,  2014), person- culture fit in three personality traits (extraversion, locomotion, and pro- motion focus) is consistently linked to higher self- esteem and better subjective well- being (Fulmer et al., 2010), and endorsing values that fit one's cultural environment is re- lated to better well- being in collectivistic, but not in indi- vidualistic, societies (Li & Hamamura, 2010). Questions of (cultural) fit have been studied using dif- ferent approaches, including difference scores (Edwards, 2002), cross- level in multilevel models (Fulmer et al.,  2010), and correlations between individ- uals and cultural response profiles (Dressler,  2012). In a widely used profile approach, De Leersnyder et al. (2014) interactions KIRCHNER-­HÄUSLER­et­al. 14676494, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jopy.12803 by Universidad De Granada, Wiley Online Library on [09/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 4 | operationalized cultural fit in emotions as the correlation of a person's pattern of responses with the average pat- tern of responses of their cultural group. More recently, response surface approaches, which describe the rela- tionship between two fit variables and the outcome as a three- dimensional surface using polynomial models, have also become increasingly common (see e.g., Humberg et al.,  2019; Schönbrodt,  2016). In the present work, we drew upon profile analysis and response surface analysis in studying the role of honor fit in subjective well- being, both of which we describe in greater detail below. 1.3 | The present study In the present research, we examined the link between individuals' cultural fit in honor values and concerns and their subjective well- being, further expanding the concep- tual space of psychological fit phenomena by including a unique construct central to wide regions of the world (Cross & Uskul, 2022). We conducted our study in com- munities around the Mediterranean, an understudied region in social and cultural psychological research (e.g., Rad et al., 2018; Thalmayer et al., 2021), where honor has been shown to play a central role in coordinating social life (e.g., Gul & Schuster, 2020; Lopez- Zafra et al., 2020; Ramirez Marin et al., 2020; Uskul & Cross, 2020). We meas- ured honor in two different ways: as values and concerns. Honor values (i.e., stable beliefs about what is good, right, and desirable, Schwartz, 1992) tapped into beliefs that in- dividuals should exhibit characteristics or display behav- iors that align with the standards of an honorable person, whereas honor concerns (i.e., appraisals of the relevance of situations to our values, goals, and needs, Rodriguez Mosquera et al., 2002) tapped into how bad one would feel if they behaved in a specific way or had a specific negative reputation that was incongruent with the honor code. We assessed both values and concerns twice: once in terms of how much participants endorsed honor themselves (their own endorsement), and once as participants' perceptions of how much most people in their society endorsed honor (their perceived- societal endorsement). We expected fit with one's environment to play a role in well- being since honor as a cultural construct inher- ently combines elements of both personal (e.g., personal characteristics, moral convictions, and behaviors) and social domains (e.g., social reputation, respect bestowed by others, and normative expectations). As such, to the extent that an individual endorses honor as a moral prin- ciple, the social dynamics surrounding the negotiation of honor may unfold fully only if the environment also em- phasizes honor to a certain degree, is responsive to one's claims to honor, and sanctions dishonorable behaviors in others. Similarly, if an individual does not endorse honor as a moral principle, high- honor environments may high- light conventions and behaviors that are not necessarily aligned with one's own convictions, and thus may restrict the individuals in living their life as they desire and pres- ent costly consequences for their well- being. Based on this reasoning, we examined the following research question concerning honor fit: Do people who show stronger fit be- tween their own honor endorsement and the honor endorse- ment of others in their social environment also show better well- being? We examined the link between cultural fit in honor and subjective well- being at different levels of analysis, fo- cusing on fit with one's perceptions of society as a whole (distal fit) and fit with one's immediate same- gender peer group (proximal fit).1 Specifically, we conceptualized fit with the distal environment as fit with individuals' per- ceptions of honor endorsement in the wider society of their respective country, and fit with the proximal envi- ronment as fit with the average honor endorsement of the matching gender group in the participants' sample (as a highly relevant social category given the gendered nature of honor, Rodriguez Mosquera, 2016). Examining fit sep- arately at distal and proximal levels offers a well- rounded approach to describing fit characteristics and may tap into different processes underlying fit, such as feelings of pro- totypicality or similarity at the perceived- societal level, as well as imminent normative or interpersonal effects at the peer group level. 2 | MATERIALS A ND M ET HODS 2.1 | Participants We recruited 3852 participants from eight communi- ties around the Mediterranean (Cyprus [Greek Cypriot and Turkish Cypriot communities], Egypt, Greece, Italy, Lebanon, Spain, Turkey) primarily via the participant pools of collaborating universities. The data were collected as part of a larger study designed to examine additional research questions concerning cultural group differences in self- construal, social orientation, and cognitive style. In each sample, we aimed for a gender- balanced sample of approximately 200 participants to allow for robust gen- der comparisons, a sample size goal that was guided by sample sizes of past comparative studies in honor contexts that build the foundation of the larger study from which the current data originates (e.g., Salvador et al., 2020; San Martin et al., 2018). To be eligible, participants had to be (a) 18 years or older, (b) born in the country of data collection, and (c) had lived in the country of data collection for more than KIRCHNER-­HÄUSLER­et­al. 14676494, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jopy.12803 by Universidad De Granada, Wiley Online Library on [09/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License | 5 half of their lives. During the data analysis stage, we in- cluded participants who self- identified (d) as a member of the country's majority ethnic group (e.g., White British in the UK), and (e) as male or female (to allow sizable gender groups for comparative purposes). Finally, we ex- cluded from the final sample participants who failed one or more of the four included attention checks embedded within the questionnaire (Table 1). Applying these crite- ria left us with N = 2257 participants in our final samples (Mage = 21.74, SD = 4.78, Min = 18, Max = 71). The sam- ple had a relatively balanced gender distribution (57.02% women), slightly higher than average socio- economic sta- tus (MSES = 6.06, SD = 1.35; on a scale of 0 = Bottom [of society] to 10 = Top [of society]) and included participants who had lived in urban environments for most of their lives (60.08%). 2.3 | Measures Study materials and instructions were compiled in English and then translated into other study languages (Arabic, Spanish, Italian, Turkish, Greek) following a team trans- lation approach (Survey Research Center,  2022). All measures were first translated by native speakers (either by a team member or a professional translator) and then checked by a second team member (fluent in both the English and the local language) to ensure it was under- standable, meaningful, familiar, and appropriate for the respective cultural context. In any given measure, we calculated scale values for each participant only if a par- ticipant answered more than half of the necessary items, and otherwise assigned a missing value to the participants (less than 1.28% for all measures). 2.2 | Procedure 2.3.1 | Honor values Participants completed a study on “Individual Differences in Social and Cognitive Orientation” between December 2019 and February 2021, either on a computer in the lab (10.59%) or on their own devices outside the lab (89.41%). After providing consent, all participants completed a se- ries of measures that were presented in the same order and with items randomized within measures. Depending on the recruitment site, participants received either course credit or monetary compensation, had the possibility to make a financial contribution to a COVID- related charity, or were entered into a raffle for vouchers from local online vendors. We assessed honor values both as the extent to which par- ticipants endorsed honor values (own endorsement) and as participants' perceptions of the extent to which most people in their society endorsed honor values (perceived- societal endorsement). Both sets included the same ten items, with wording adjusted to reflect the own and perceived- societal endorsement focus. We took four items from Yao et al. (2017) (e.g., “People should not allow others to insult their family”) and six items from Smith et al. (2017) (e.g., “People always need to show off their power in front of their competitors”) to increase the conceptual coverage of our honor measure. To reflect the endorsement of values T A B L E 1 Overview of data collection sites and recruitment information Country Women Men Age SES Language Institution Compensation Cyprus (Turkish) 110 45 24.23 (9.03) 6.4 (1.31) Turkish Eastern Mediterranean Course credit, raffle University Cyprus (Greek) Egypt Greece Italy Lebanon Spain Turkey 214 110 196 135 165 116 241 103 95 284 112 20.89 (2.36) 6.04 (1.19) Greek University of Cyprus Course credit, raffle 20.73 (1.56) 6.44 (1.31) Arabic British University of Donation to charity Egypt 23.14 (6.07) 6.04 (1.21) Greek University of Crete 22.76 (4.07) 5.9 (1.39) Italian University of Course credit Course credit Chieti- Pescara 96 19.14 (1.63) 6.7 (1.41) English American University of Course credit Beirut 124 111 22.53 (6.02) 5.72 (1.47) Spanish University of Granada Course credit 20.8 (1.59) 5.64 (1.29) Turkish Bolu Abant Izzet Course credit Total 1287 970 21.74 (4.78) 6.06 (1.35) – – – Baysal University, Ordu University, Zonguldak Bülent Ecevit University KIRCHNER-­HÄUSLER­et­al. 14676494, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jopy.12803 by Universidad De Granada, Wiley Online Library on [09/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 6 | rather than states or behaviors, we rephrased the items to read “People should…” instead of the original wording of “People are…” or “People do…”. Items were rated using a 7- point Likert scale (1 = strongly disagree to 7 = strongly agree). An exploratory factor analysis revealed a two- factor solution for both own (CFI  =  0.977, TLI  =  0.959, SRMR  =  0.026) and perceived- societal endorsement (CFI = 0.991, TLI = 0.984, SRMR = 0.018). In both mea- sures, the two factors that emerged were a factor for Family Reputation (emphasizing the maintenance and defense of family reputation; αOwn = 0.85, αSociety = 0.84) and a factor for a strong self- image (emphasizing the need to project oneself as strong and powerful and to respond decisively to threats to one's honor; αOwn  =  0.65, αSociety  =  0.78). Invariance testing suggested that two items (“People must always be ready to defend their honor” of the family repu- tation factor and “It is important to promote oneself to oth- ers” of the strong self- image factor) loaded inconsistently across the eight samples; we thus excluded these two items from both item sets (see SM for final factor loadings). 2.3.2 | Honor concerns As with honor values, honor concerns were measured with two item sets, one assessing participants' own en- dorsement of honor concerns and one assessing their perceptions of the extent to which most people in their so- ciety endorse honor concerns. We took items from Guerra et al.’ (2013) short version of the Honor Scale (originally developed by Rodriguez Mosquera et al., 2002), designed to assess four honor facets (originally named family honor, feminine honor, masculine honor, integrity honor) with four items in each subscale. Participants rated the extent to which behaving in a specific way or having a specific reputation would make them feel bad about themselves (e.g., “How bad would you feel about yourself if you let other people insult your family?”) using a 7- point Likert scale (1 = Not at all bad to 7 = Very bad).2 An exploratory factor analysis largely supported the original four- factor solution for both own endorsement (CFI = 0.985, TLI = 0.971, SRMR = 0.015) and perceived- societal endorsement (CFI  =  0.991, TLI  =  0.983, SRMR = 0.012). We excluded two items (“…you were known as someone who cannot support a family” and “…you had the reputation of being someone without sexual experience”; both from the masculine honor subscale) from both item sets, as they did not load most strongly on the expected factor, with 14 items retained in the final version. Finally, we renamed the “masculine honor” factor as “family au- thority” to reflect the new focus of the items in this sub- scale, the “feminine honor” factor as “sexual propriety” to more closely follow the conceptual meaning of the items and to reflect that we were collecting data from both men and women, and “family honor” and “integrity honor” to “family reputation” and “integrity”, respectively, to reflect that these dimensions are facets of honor, not different types of honor. All factors showed acceptable reliability (all α > 0.76; see SM for final factor loadings). 2.3.3 | Subjective well- being To assess subjective well- being (SWB), we asked partici- pants to rate their satisfaction in nine domains of their lives (standard of living, health, what one is achieving in life, personal relationships, how safe one feels, feeling part of one's community, future security, amount of time one has to do the things that one likes doing, and the quality of one's local environment [e.g., pollution, green spaces]) using the OECD Guidelines on Measuring Well- Being (OECD, 2013). In addition to the nine domains, we also included an item that asked participants to rate their sat- isfaction with “life as a whole”. All items were rated using a 10- point Likert scale (0 = not at all to 10 = completely satisfied). An exploratory component analysis suggested a single component structure (α = 0.85); we thus averaged items to create one overall subjective well- being score.3 3 | RESULTS 3.1 | Analytic strategy We investigated the link between cultural fit in honor val- ues and concerns and subjective well- being in two ways. First, we conducted a series of Response Surface Analyses (RSA) (Edwards, 2002; Schönbrodt, 2016) to examine fit at the distal level separately for all facets of honor values and concerns. Second, we conducted a profile fit analysis to examine fit at the proximal level across all six facets si- multaneously (De Leersnyder,  2017; McCrae,  2008) (see Table 2 for descriptive statistics).4 Given the exploratory nature of our analyses, following recent recommenda- tions (Benjamin et al.,  2018), we applied a significance level of .005 for greater protection against false- positive re- sults; we thus refer to p- values less than p < .005 as “signif- icant” and to p- values in the range of p < .005 to p < .05 as “suggestive.” For all analyses, we used a Full Information Maximum Likelihood (FIML) approach to impute miss- ing values. All analyses were conducted using R Studio, Version 1.2.5001 (R Core Team, 2020). The data and syn- tax that support the findings of this study and produced this manuscript are openly available in the Open Science Framework at https://osf.io/4tyk5/. KIRCHNER-­HÄUSLER­et­al. 14676494, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jopy.12803 by Universidad De Granada, Wiley Online Library on [09/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License T A B L E 2 Descriptive statistics for study variables Variable Honor values Family reputation (own) Strong self- image (own) Family reputation (society) Strong self- image (society) Honor concerns Family reputation (own) Sexual propriety (own) Family authority (own) Integrity (own) Family reputation (society) Sexual propriety (society) Family authority (society) Integrity (society) SWB SWB (all items) n 2246 2246 2228 2228 2253 2250 2240 2255 2246 2241 2235 2246 2257 M 5.10 3.70 5.83 5.33 5.88 4.59 4.25 6.09 5.74 4.28 4.72 4.56 6.17 SD Min Max SE 1.36 1.39 1.02 1.34 1.20 1.84 1.80 0.98 1.19 1.73 1.64 1.54 1.72 1 1 1 1 1 1 1 1 1 1 1 1 0 7 7 7 7 7 7 7 7 7 7 7 7 10 0.03 0.03 0.02 0.03 0.03 0.04 0.04 0.02 0.03 0.04 0.03 0.03 0.04 | 7 Cronbach's α 0.85 0.65 0.84 0.78 0.76 0.85 0.85 0.76 0.80 0.87 0.85 0.88 0.85 | Distal fit: Own endorsement and 3.2 perceived- societal endorsement of honor To assess the role of distal fit in subjective well- being, we used RSA, an analytical tool designed to test whether the fit (or “congruence”) between two variables (x and y) shows a systematic relationship to a third, dependent vari- able (z) (Schönbrodt, 2016). We examined the congruence between participants' own endorsement of honor values and concerns (x) and their perceived- societal endorse- ment of honor values and concerns (y) to predict their subjective well- being (z). The basic steps of RSA consist of fitting a full polynomial regression model (i.e., linear terms, their interaction, and squared terms for both vari- ables), as well as simpler alternative models, to the data, and then interpreting the resulting coefficients both sta- tistically and graphically. The applied RSA model can be represented as a three- dimensional response surface, which maps pairs of scores on the predictors (x and y axes) against the predicted scores on the outcome variable (z axis; see e.g. Figure 1). Of particular interest to questions of fit are three elements of the response surface. First, the Line of Incongruence (LOIC; shown in blue in the plots), which is the line for which x equals the opposite of y (i.e., x = −y, or the line leading from the front left corner of the coordi- nate cube to the back right corner of the coordinate cube), representing different levels of mismatch between the two predictors. The shape of this line is represented by the model parameters a3 (describing the slope of the LOIC at the midpoint 0,0) and a4 (describing the curvature of the LOIC, i.e., flat, u- shape, or inverted u- shape). Second, the Line of Congruence (LOC; shown in red in the plots, or the line leading from the bottom corner of the coordinate cube to the top corner of the coordinate cube), which is the line for which x equals y, representing different lev- els of matching values of x and y (and thus representing the line where a congruence effect should take place). The shape of this line is represented by the model parameters a1 (describing the slope of the line at the midpoint 0,0) and a2 (describing the curvature of the line). And finally, the First Principal Axis (FPA), which in non- mathematical terms represents the “ridge” of the response surface (or the line following the “bend” of the surface), if the surface is curved. For questions of fit, of particular interest are the parameters p10 and p11, which represent the vertical shift and the rotation, respectively, of the projection of the FPA onto the bottom of the surface cube. Conceptually, these two parameters can give insight into whether the FPA is aligned with the LOC or whether it significantly differs from the LOC. The presence of congruence effects is determined by the joint interpretation of these three elements (and their associated statistical parameters). Humberg et al.  (2019) outline four conditions to conclude a congruence ef- fect in the broadest way: First, the FPA must not deviate KIRCHNER-­HÄUSLER­et­al. 14676494, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jopy.12803 by Universidad De Granada, Wiley Online Library on [09/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 8 | significantly from the LOC. This is reflected statistically in the two conditions that (1) p10 must not be significantly different from 0, and (2) that p11 must not be significantly different from 1, respectively. This condition represents that the highest outcome scores for people are predicted for individuals with congruent predictors. The other condition for a broad congruence effect with a positively scaled outcome variable is that the LOIC must represent an inverted u- shape, with its peak above the midpoint (0,0). This is reflected statistically in the two conditions that (3) a4 must be significantly negative, and (4) that a3 must not be significantly different from 0, respectively. This condition represents that people with increasingly incongruent predictor scores have lower outcome values and that the peak of this inverted u- shape lies over the midpoint (i.e., the LOC). If these four conditions are met, one can conclude that the data support a congruence effect in a broad sense, i.e., a pattern in which congruence has a positive effect on the outcome, but which also allows for main effects of the two predictor variables (e.g., higher av- erage values in x and y are by themselves linked to better outcomes).5 In our analyses, we tested for these four conditions of a broad congruence effect as a statistical representation of our verbal hypotheses. Yet, while we primarily empha- sized a congruence pattern for the various facets of honor, we also neither precluded nor predicted the presence of specific additional, level- related effects. Hence, our testing approach of the RSA surface contains confirmatory (i.e., the four conditions of Humberg et al., 2019) as well as ex- ploratory aspects (i.e., the potential tilt and/or curvature of the LOC, which can be modeled by the broad congru- ence model). We conducted our hypothesis testing in two steps: first, we tested a full polynomial model against vari- ous simpler, more constrained models, and chose the best fitting and most parsimonious model as our final model to interpret the RSA model parameters (for an overview of the different models, please see the supplementary ma- terials). Second, we then checked the conditions needed for a broad congruence effect by examining the respective parameters in the final model. For simpler models, the in- troduced constraints may already fulfill some of the condi- tions needed for a broad congruence effect; we thus focus on the remaining conditions to conclude a broad congru- ence effect depending on the particular final model. All models were run as multilevel structural equation models in the R package lavaan (Rosseel, 2012), nesting participants within countries and including random in- tercepts. To facilitate interpretation, we standardized the predictors around their shared grand mean and grand standard deviation prior to all analyses. We also country- mean centered our predictor variables and entered the country means as separate variables into the model to not confound individual fit with differences in overall levels between our country groups (Enders & Tofighi,  2007). Finally, we examined gender differences by adding gen- der, as well as all interactions of gender with the other predictors, to the final model, and comparing the model fit of this gender- added model. We found no indication of gender differences in any of our final models: adding interactions with gender into any final model did not sig- nificantly increase the model fit as measured by the Chi- Square, and the model without gender interactions also showed a consistently better fit than the one including in- teractions based on the AIC (difference in AIC of at least 2). We thus report results for the pooled sample across both genders only. 3.2.1 | Honor values The analysis of honor values provided at least suggestive support for a broad congruence effect between both di- mensions of honor values and subjective well- being. For family reputation values, model comparisons indicated that a simpler “Rising Ridge” model emerged as the final and most parsimonious model (see Figure  1; a detailed overview of model parameters can be found in the supple- mentary materials). In a “Rising Ridge” model, conditions 1, 2, and 4 for a broad congruence effect are met as a result of the introduced model constraints, leaving a test of con- dition 3 (an inverted u- shape of the LOIC) to conclude a broad congruence effect. The LOIC of the current model indeed showed an inverted u- shape (as indicated by a sig- nificant negative a4  =  −0.35, p < .001, 95% CI  =  [−0.45, −0.24]). This effect was accompanied by a positive linear effect of the levels of predictors, or a positive linear slope of the LOC at the point 0,0 (as indicated by a significant positive a1 =  0.26, p < .001, 95% CI =  [0.15, 0.37]), sug- gesting an additional link between higher predictor values and better well- being. For strong self- image values, model comparisons in- dicated that a simpler “Interaction” model emerged as the final and most parsimonious model (see Figure 1; a detailed overview of model parameters can be found in the supplementary materials). In an “Interaction” model, condition 2 for a broad congruence effect is met as a re- sult of the introduced model constraints, leaving a test of condition 1 (no shift of the FPA), condition 3 (an inverted u- shape of the LOIC), and 4 (slope of the LOIC at 0,0 is 0) to show a broad congruence effect. The current model indeed showed an FPA that was not significantly shifted from the LOC (as indicated by a non- significant p10 = 0.24, p = .7, 95% CI = [−0.99, 1.47]), and a slope of the LOIC not different from 0 at the midpoint 0,0 (as indicated by a non- significant a3 = −0.04, p = .724, 95% CI = [−0.23, 0.16]). In KIRCHNER-­HÄUSLER­et­al. 14676494, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jopy.12803 by Universidad De Granada, Wiley Online Library on [09/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License (cid:11)(cid:68)(cid:12) (cid:37) (cid:58) (cid:54) (cid:28) (cid:27) (cid:26) (cid:25) (cid:24) (cid:23) (cid:23) (cid:21) (cid:19) (cid:51)(cid:72)(cid:85)(cid:70)(cid:72)(cid:76)(cid:89)(cid:72)(cid:71)(cid:237)(cid:54)(cid:82)(cid:70)(cid:76)(cid:72)(cid:87)(cid:68)(cid:79)(cid:3) (cid:40)(cid:81)(cid:71)(cid:82)(cid:85)(cid:86)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87) (cid:237)(cid:21) (cid:237)(cid:23) (cid:237)(cid:23) (cid:23) (cid:21) (cid:19) (cid:237)(cid:21) (cid:50)(cid:90)(cid:81)(cid:3)(cid:40)(cid:81)(cid:71)(cid:82)(cid:85)(cid:86)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87) (cid:11)(cid:69)(cid:12) (cid:28) (cid:27) (cid:37) (cid:58) (cid:54) (cid:26) (cid:25) (cid:24) (cid:23) (cid:20) (cid:22) (cid:21) (cid:51) (cid:72)(cid:85)(cid:70) (cid:40) (cid:72)(cid:76)(cid:89) (cid:81) (cid:71) (cid:72) (cid:82)(cid:85)(cid:86) (cid:71) (cid:237) (cid:54) (cid:72) (cid:80) (cid:82) (cid:72) (cid:81)(cid:87) (cid:19) (cid:237)(cid:20) (cid:237)(cid:21) (cid:70)(cid:76)(cid:72)(cid:87)(cid:68)(cid:79)(cid:3) (cid:237)(cid:22) (cid:237)(cid:22) (cid:237)(cid:21) | 9 (cid:22) (cid:21) (cid:20) (cid:19) (cid:237)(cid:20) (cid:50)(cid:90)(cid:81)(cid:3)(cid:40)(cid:81)(cid:71)(cid:82)(cid:85)(cid:86)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87) F I G U R E 1 Shown are the response surface plot for (a) family reputation values as well as (b) strong self- image values. Please note that the plot for strong self- image values has been rotated to allow for better visualization; however, the coordinate space is the same as for all other plots. Black points represent the (jittered) data points of participants at their predicted level of SWB. The red line marks the line of congruence, the blue line marks the line of incongruence. The two inner circles mark a bagplot, which describes the position of the inner 50% of points (the inner circle) and the outer 50% of points (the outer circle), except outliers. addition, there was suggestive evidence for a LOIC in the shape of an inverted u- shape (as indicated by a suggestive negative a4 = −0.14, p = .005, 95% CI = [−0.25, −0.04]). The data thus provided suggestive support for a broad con- gruence effect for strong self- image values. Furthermore, this congruence effect was combined with both a negative slope of the LOC at the midpoint 0,0 (i.e., a suggestive neg- ative a1 = −0.12, p = .045, 95% CI = [−0.23, −0.002]) and a positive curvature (u- shape) of the LOC (i.e., a significant positive a2 = 0.14, p = .005, 95% CI = [0.04, 0.25], which is constrained to be the opposite of a4 in an “Interaction” model), also suggesting a relationship between general levels of honor endorsement and well- being. Taken together, these results thus show support for a broad congruence effect in both facets of honor values: in- dividuals who showed a match in their own and perceived- societal endorsement of family reputation values or strong self- image values also showed higher well- being than in- dividuals that showed a mismatch, when comparing the same average predictor levels. Furthermore, these congru- ence effects were complemented by links between the gen- eral levels of value endorsement and well- being: given the same degree of (mis)match in their own and perceived- societal value endorsement, individuals at higher levels of honor endorsement showed higher well- being scores than individuals at moderate endorsement levels in both value facets; for strong self- image values individuals at low lev- els of honor endorsement also showed higher well- being scores than individuals at moderate endorsement levels. Notably, for both facets a majority of non- matching cases were located left of the LOC (as shown by the projected black dots on the surface and the bagplot), suggesting that instances of mismatch in which participants rated their society to hold stronger values than they themselves did were more frequent than vice- versa. Our conclusions are therefore more robust for this type of pattern compared to the opposite pattern (in which participants rated them- selves as holding stronger values than their society). 3.2.2 | Honor concerns Next, we applied the same RSA analysis to examine cul- tural fit within honor concerns. An “Interaction” model emerged as the best fitting model for most honor con- cerns (family reputation, sexual propriety, or integrity concerns), whereas for family authority concerns both an “Interaction” model and a “Rising Ridge” model emerged as the best fitting models, showing equal model fit. However, in testing the four conditions for a broad con- gruence effect, we found that the data supported a broad congruence effect only for family authority concerns (with the same conclusions for both the “Interaction” and “Rising Ridge” model), but not for the remaining honor concern facets. For family authority concerns, simpler “Interaction” and “Rising Ridge” models emerged as the final and most parsimonious models (see Figure 2; a detailed overview of KIRCHNER-­HÄUSLER­et­al. 14676494, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jopy.12803 by Universidad De Granada, Wiley Online Library on [09/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 10 | (cid:11)(cid:68)(cid:12) (cid:37) (cid:58) (cid:54) (cid:28) (cid:27) (cid:26) (cid:25) (cid:24) (cid:23) (cid:22) (cid:21) (cid:20) (cid:19) (cid:51)(cid:72)(cid:85)(cid:70)(cid:72)(cid:76)(cid:89)(cid:72)(cid:71)(cid:237)(cid:54)(cid:82)(cid:70)(cid:76)(cid:72)(cid:87)(cid:68)(cid:79)(cid:3) (cid:40)(cid:81)(cid:71)(cid:82)(cid:85)(cid:86)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87) (cid:237)(cid:20) (cid:237)(cid:21) (cid:237)(cid:22) (cid:237)(cid:22) (cid:237)(cid:21) (cid:22) (cid:21) (cid:20) (cid:19) (cid:237)(cid:20) (cid:50)(cid:90)(cid:81)(cid:3)(cid:40)(cid:81)(cid:71)(cid:82)(cid:85)(cid:86)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87) (cid:11)(cid:69)(cid:12) (cid:37) (cid:58) (cid:54) (cid:28) (cid:27) (cid:26) (cid:25) (cid:24) (cid:23) (cid:22) (cid:21) (cid:20) (cid:19) (cid:51)(cid:72)(cid:85)(cid:70)(cid:72)(cid:76)(cid:89)(cid:72)(cid:71)(cid:237)(cid:54)(cid:82)(cid:70)(cid:76)(cid:72)(cid:87)(cid:68)(cid:79)(cid:3) (cid:40)(cid:81)(cid:71)(cid:82)(cid:85)(cid:86)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87) (cid:237)(cid:20) (cid:237)(cid:21) (cid:237)(cid:22) (cid:237)(cid:22) (cid:237)(cid:21) (cid:22) (cid:21) (cid:20) (cid:19) (cid:237)(cid:20) (cid:50)(cid:90)(cid:81)(cid:3)(cid:40)(cid:81)(cid:71)(cid:82)(cid:85)(cid:86)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87) F I G U R E 2 Shown are response surface plots of the two final models for (a) family authority concerns, (a) an interaction model, and (b) a rising ridge model. Black points represent the (jittered) data points of participants at their predicted level of SWB. The red line marks the line of congruence, the blue line marks the line of incongruence. The two inner circles mark a bagplot, which describes the position of the inner 50% of points (the inner circle) and the outer 50% of points (the outer circle), except outliers. model parameters can be found in the supplementary ma- terials). In an “Interaction” model, condition 2 for a broad congruence effect is met as a result of the introduced model constraints, leaving a test of condition 1 (no shift of the FPA), condition 3 (an inverted u- shape of the LOIC), and 4 (slope of the LOIC at 0,0 is 0) to conclude a broad congruence effect. The current “Interaction” model indeed showed no shift of the FPA from the LOC (as indicated by a non- significant p10 = −0.70, p = .457, 95% CI = [−2.53, 1.14]), and a slope of the LOIC not different from 0 at the midpoint 0,0 (as indicated by a non- significant a3 = 0.06, p = .39, 95% CI = [−0.07, 0.18]). Furthermore, there was also a suggestive inverted u- shape of the LOIC (as indi- cated by a suggestive negative a4 = −0.08, p = .025, 95% CI = [−0.15, −0.01]). The “Interaction” model thus sug- gestively supported a broad congruence effect. These conclusions converged with the interpretation of the alternative “Rising Ridge” model: In a “Rising Ridge” model, conditions 1, 2, and 4 for a broad congruence effect are met as a result of the introduced model constraints, leaving a test of condition 3 (an inverted u- shape of the LOIC) to conclude a broad congruence effect. The current “Rising Ridge” model also showed suggestive evidence for an inverted u- shape of the LOIC (as indicated by a sug- gestive negative a4  =  −0.14, p  =  .025, 95% CI  =  [−0.26, −0.02]), thus also suggestively supporting a broad con- gruence effect. Both final models (“Interaction” and “Rising Ridge”) also showed a positive slope of the LOC at the midpoint 0,0 (indicated by a significant positive a1 parameter; “Interaction” Model: a1  =  0.18, p < .001, 95% CI  =  [0.09, 0.28]; “Rising Ridge” Model: a1  =  0.16, p < .001, 95% CI = [0.07, 0.26]), but the Interaction Model also showed a suggestive curvilinear shape (u- shape) of the LOC (a significant positive a2  =  0.08, p  =  .025, 95% CI = [0.01, 0.15]; a2 is constrained to be the opposite of a4 in an “Interaction” model). Taken together, the results from both final models converge in (suggestive) support for a broad congruence effect in family authority concerns: comparing individu- als at the same average level of both honor endorsement predictors, individuals who showed a match in their own and perceived- societal endorsement of family authority concerns showed better well- being compared to individ- uals that showed a mismatch. But, this broad congru- ence effect was complemented by an effect between the levels of honor endorsement and well- being: while the exact relationship differed between the final models, the shared characteristic was that, given the same degree of (mis)match in their ratings, individuals at higher levels of honor endorsement showed higher levels of well- being compared to individuals at low or medium levels. For the remaining honor concern facets of family reputation, sexual propriety, or integrity concerns, an “Interaction” model emerged as the best fitting and most parsimonious model for all three facets (see Figure 3; a detailed overview of model parameters can be found in the supplementary materials). In an Interaction model, condition 2 for a broad congruence effect is met as a re- sult of the introduced model constraints, leaving a test of condition 1 (no shift of the FPA), condition 3 (an inverted u- shape of the LOIC), and 4 (slope of the LOIC at 0,0 is 0) to show a broad congruence effect. None of the three KIRCHNER-­HÄUSLER­et­al. 14676494, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jopy.12803 by Universidad De Granada, Wiley Online Library on [09/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License (cid:11)(cid:68)(cid:12) (cid:37) (cid:58) (cid:54) (cid:28) (cid:27) (cid:26) (cid:25) (cid:24) (cid:23) (cid:11)(cid:69)(cid:12) (cid:37) (cid:58) (cid:54) (cid:28) (cid:27) (cid:26) (cid:25) (cid:24) (cid:23) (cid:11)(cid:70)(cid:12) (cid:20)(cid:19) (cid:37) (cid:58) (cid:54) (cid:27) (cid:25) (cid:23) (cid:21) (cid:22) (cid:20) (cid:21) (cid:51)(cid:72)(cid:85)(cid:70)(cid:72)(cid:76)(cid:89)(cid:72)(cid:71)(cid:237)(cid:54)(cid:82)(cid:70)(cid:76)(cid:72)(cid:87)(cid:68)(cid:79)(cid:3) (cid:40)(cid:81)(cid:71)(cid:82)(cid:85)(cid:86)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87) (cid:19) (cid:237)(cid:20) (cid:237)(cid:21) (cid:237)(cid:22) (cid:237)(cid:20) (cid:19) (cid:20) (cid:21) (cid:22) (cid:50)(cid:90)(cid:81)(cid:3)(cid:40)(cid:81)(cid:71)(cid:82)(cid:85)(cid:86)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87) (cid:237)(cid:21) (cid:237)(cid:22) (cid:22) (cid:19) (cid:21) (cid:20) (cid:51)(cid:72)(cid:85)(cid:70)(cid:72)(cid:76)(cid:89)(cid:72)(cid:71)(cid:237)(cid:54)(cid:82)(cid:70)(cid:76)(cid:72)(cid:87)(cid:68)(cid:79)(cid:3) (cid:40)(cid:81)(cid:71)(cid:82)(cid:85)(cid:86)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87) (cid:237)(cid:21) (cid:237)(cid:20) (cid:237)(cid:22) (cid:237)(cid:22) (cid:22) (cid:21) (cid:20) (cid:19) (cid:237)(cid:21) (cid:237)(cid:20) (cid:50)(cid:90)(cid:81)(cid:3)(cid:40)(cid:81)(cid:71)(cid:82)(cid:85)(cid:86)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87) (cid:22) (cid:20) (cid:21) (cid:51)(cid:72)(cid:85)(cid:70)(cid:72)(cid:76)(cid:89)(cid:72)(cid:71)(cid:237)(cid:54)(cid:82)(cid:70)(cid:76)(cid:72)(cid:87)(cid:68)(cid:79)(cid:3) (cid:40)(cid:81)(cid:71)(cid:82)(cid:85)(cid:86)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87) (cid:19) (cid:237)(cid:20) (cid:237)(cid:21) (cid:237)(cid:22) | 11 (cid:237)(cid:20) (cid:19) (cid:20) (cid:21) (cid:22) (cid:50)(cid:90)(cid:81)(cid:3)(cid:40)(cid:81)(cid:71)(cid:82)(cid:85)(cid:86)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87) (cid:237)(cid:21) (cid:237)(cid:22) F I G U R E 3 Shown are response surface plots for (a) family reputation concerns, (b) sexual propriety concerns, and (c) integrity concerns. Black points represent the (jittered) data points of participants at their predicted level of SWB. The red line marks the line of congruence, the blue line marks the line of incongruence. The two inner circles mark a bagplot, which describes the position of the inner 50% of points (the inner circle) and the outer 50% of points (the outer circle), except outliers. final “Interaction” models met these conditions: all mod- els showed an at least suggestive shift of the FPA from the LOC (as shown by a p10 that was at least suggestively different from 0; family reputation: p10 = −1.62, p = .039, 95% CI  =  [−3.15, −0.08]; sexual propriety: p10  =  −0.93, p = .038, 95% CI = [−1.80, −0.05]; Integrity: p10 = −1.34, p < .001, 95% CI = [−2.01, −0.67]), as well as an at least suggestive positive slope of the LOIC at the midpoint 0,0 (as shown by an at least suggestively positive a3; family reputation: a3 = 0.16, p = .01, 95% CI = [0.04, 0.28]; sexual propriety: a3 = 0.15, p = .019, 95% CI = [0.02, 0.28]; integ- rity: a3 = 0.28, p = .006, 95% CI = [0.08, 0.49]). As such, we did not find support for a broad congruence effect for honor concerns related to family reputation, sexual pro- priety, or integrity. | Proximal fit: Own endorsement and 3.3 group- level endorsement of honor To examine the role of proximal fit in subjective well- being, we inspected the fit between participants and their imme- diate social group (as opposed to their perceived- societal environment in distal fit) by applying a profile fit analy- sis, following steps outlined by De Leersnyder (2017). In profile fit analysis, the individual- environment fit is con- ceptualized as the similarity between a participant's pat- tern of scores (their “profile”) and the pattern of averages for their respective sample group (the “group profile”).6 Following past conventions (De Leersnyder,  2017), for each participant, we conducted profile analysis across the scores of all six honor facets, as profile fit analysis is most suitable when profile elements have distinct conceptual meaning and thus allow for more variation in the profile shape (as opposed to, for example, items within a sub- scale, which co- vary strongly). We therefore calculated profile similarity as the overlap of an individual's scores in the six honor facets with the corresponding six aver- age scores of the individual's matching gender group at their university (but excluding the individual themselves, see De Leersnyder, 2017).7 To illustrate, we calculated the proximal fit index for a Spanish female participant as the similarity of her pattern of honor scores with the average pattern of honor scores of all other female students at her university. We focused on gendered comparison groups as the proximal environment often prescribes different norms and expectations for men and women (Rodriguez Mosquera, 2016) and since from a young age, peer groups can play a critical role in the learning and endorsement of societal conventions (Killen & Stangor, 2001). We used the Intraclass Correlation with Double Entry (ICC- DE) (McCrae, 2008) as the statistical index of profile similarity. The ICC- DE is sensitive to differences in pro- file levels and profile shape and has been shown to per- form generally better than other indices of fit (e.g., simple Pearson Correlations, McCrae,  2008) (see SM for more information on the calculation). We Fisher- transformed the ICC- DE scores (to normalize their distribution, see De Leersnyder,  2017) and included them as predictors in a multilevel structural equation model (with random intercepts between countries), predicting subjective well- being scores. We again examined gender differences by comparing model fit for a model that included interaction of the profile similarity score with gender to one that did not; however, this did not improve fit for our model and we thus report results for our pooled sample across both genders.The profile fit analysis revealed that, on average, participants' scores across the six honor facets showed KIRCHNER-­HÄUSLER­et­al. 14676494, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jopy.12803 by Universidad De Granada, Wiley Online Library on [09/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 12 | overlap with their matching- gender university groups (MICC- DE  =  0.45, SD  =  0.35, Min  =  −0.85, Max  =  0.99). No gender differences in proximal fit levels were found (MMen = 0.43, MWomen = 0.46, F(1, 2255) = 2.57, p = .109). Our regression analyses also provided some support for a link between fit across the six honor facets at the prox- imal level and subjective well- being (see Table  3 for an overview of results). We thus found a suggestive effect for our index of profile similarity, showing that individuals whose profile across the six honor facets was more similar to their same- gender university group also showed better subjective well- being, b = 0.17, p = .025, 95% CI = [0.02, 0.32]. 4 | DISCUSSIO N The current research contributes to the increasing body of work on the role of honor in psychological processes in the Mediterranean region (Uskul & Cross,  2019) and the association between cultural fit and psychologi- cal outcomes (Mobasseri et al.,  2019), by examining the link between individuals' cultural fit in honor values and concerns and their subjective well- being in communities circum Mediterranean. We conceptualized honor as the endorsement of honor at the individual level and consid- ering its multifaceted nature. We also assessed cultural fit both at distal (as congruence with one's perceptions of the wider society) and proximal levels (as profile similarity with one's gender group at university), and operational- ized subjective well- being as satisfaction with one's life across a variety of domains. Finally, we employed a multi- method approach to study the relationship between cul- tural fit and subjective well- being, providing insights from response surface analysis for distal fit (Schönbrodt, 2016) and profile fit analysis for proximal fit (McCrae,  2008). Our findings provide some support for a link between cul- tural fit in honor and subjective well- being: The RSA anal- yses showed that stronger distal fit in honor (i.e., fit with one's perceptions of society) was linked to better well- being for three out of six facets of honor (family reputa- tion values, strong self- image values, and family authority concerns). Furthermore, the profile analyses also showed that stronger proximal fit (i.e., fit with one's same- gender university group) calculated across all honor facets was associated with better well- being. As the current data were cross- sectional, we can only speculate about the exact underpinnings of these fit ef- fects; however, greater environmental fit in honor- related values and concerns may help individuals pay attention to important aspects of (interpersonal) situations, to engage in normative behavior, and to ensure that criti- cal psychological needs are met, supporting their well- being. Similarly, endorsing similar values and concerns as one's (actual or perceived) social environment can also be linked to greater feelings of similarity and be- longing (Hogg & Terry, 2000). Generally, these fit effects between individuals and their environment are consis- tent with the idea that honor is a conceptual construct that includes an interplay of both individual and societal elements (Pitt- Rivers, 1965): Honor is a privilege that an individual has to claim for themselves, but that also has to be “responded” to in the environment. In line with this point, in models that yielded a fit effect, participants with the highest levels of well- being were consistently located in the region with high own endorsement as well as high perceived- societal endorsement. Furthermore, the RSA analyses of distal fit also showed that models that included an interaction term between predictors consistently fit the data better than a “Main Effects” model that did not include such an interaction term (and as such viewed the influence of participants' own and T A B L E 3 Model parameters for regression analyses of subjective well- being on proximal fit Estimate SE z p LL 95%- CI 6.08 0.17 0.14 0.08 44.86 2.25 <0.001* 0.02† 5.81 0.02 Variable Fixed effects Intercept Proximal honor fit Error terms Intercept variance 0.12 0.07 1.79 0.07 −0.01 (Lvl- 2) Residual variance 2.85 0.08 33.53 <0.001* 2.68 (Lvl- 1) UL 6.35 0.32 0.25 3.02 Note: Shown are parameter coefficients for the multi- level regression model for proximal fit across all six honor facets. The fit was computed with the ICC- Double Entry with one's same- gender university group. *p < .005; †p < .05. KIRCHNER-­HÄUSLER­et­al. 14676494, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jopy.12803 by Universidad De Granada, Wiley Online Library on [09/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License perceived- societal honor endorsement as more indepen- dent from each other). Simultaneously, our findings illuminate the impor- tance of considering the multi- faceted nature of honor, rather than treating it as a unitary construct that may ob- scure more nuanced processes and differences (Rodriguez Mosquera,  2016). When using RSA to examine distal fit for specific dimensions of honor, fit effects emerged for three out of six facets of honor: both dimensions of honor values (family reputation values and strong self- image val- ues) and one out of four dimensions of honor concerns (family authority concerns). Our present data do not allow us to draw firm conclusions on why fit effects emerged for these three facets and not for the others. However, both family reputation and strong self- image values are cru- cial components that have long been studied in the in- terpersonal dynamics of honor (for a review see Uskul & Cross,  2019), where fit may be particularly important to coordinate social behavior. Our measure of family au- thority concerns has traditionally been regarded as part of the “masculine” dimension of honor, so it is interest- ing that we found no gender differences for the fit effect (Rodriguez Mosquera, 2016). It is possible that, in our rel- atively young and educated student samples, the family authority may be seen less as an exclusively male territory and that both women and men may be perceived as play- ing an increasingly comparable role in shaping family life, thus higher fit in this facet among both men and women may partly reflect changing perceptions of gender roles (note that we did not assess perceived- societal ratings in a gendered way). Finally, as we found more fit effects in values than concerns, it may also be possible that our way of assessing these facets may have influenced our results. We assessed values as agreement with positively worded beliefs and norms, and concerns as instances of threat or obstruction to relevant honor goals (a more “negative” perspective, Guerra et al., 2013)— possibly a more unusual approach for participants to answer in terms of the per- spective of others. Future research should explore if dif- ferent measurement approaches hold implications for the detection of fit effects. Finally, our results also respond to calls for a shift away from limiting the study of honor to topics of inter- personal retaliation and violence (Uskul & Cross,  2019), and contributes to the scarce but growing evidence on po- tential positive outcomes of honor endorsement (Cohen et al.,  1999; Cross et al.,  2014; Leung & Cohen,  2011): Our findings showed that, within a sample of circum- Mediterranean countries, higher honor fit (both actual and perceived) can be associated with better well- being. Furthermore, a majority of our final RSA models (all but strong self- image values) showed positive linear and cur- vilinear effects of our predictors on well- being, which in | 13 combination suggested that, aside from the fit effects, in- dividuals at high levels of both endorsement predictors also showed higher well- being scores. This may be to some extent a reflection of our sample choice, as we collected individuals from countries in a region in which honor has traditionally been found to be a core guiding value; as such individuals at high levels of own and perceived- societal endorsement of honor may match this cultural environment relatively more than those at medium or low levels and thus allow culturally ingrained dynamics to un- fold in a more fluid way, possibly supporting well- being (or, at least, not obstructing it). Generally, our results align with and support previous findings highlighting the im- portant role that honor occupies in the Mediterranean re- gion, as a central cultural construct that guides cultural expectations about how to live a “good” and appropriate life (Uskul & Cross, 2019). 4.1 | Limitations and future directions Our study comes with several limitations. First, al- though our research expands the existing evidence on cultural fit to an understudied region, future research in cultural fit in honor would benefit from an even more diverse pool of countries and regions. While we particu- larly focused on the Mediterranean region as a regional case- in- point in which honor is endorsed as an impor- tant cultural value, it would be informative to examine whether the current findings hold in other world regions identified as promoting strong honor values and con- cerns (e.g., Latin- America, Southern U.S., parts of South Asia) and which may differ from those here included across a wide variety of other characteristics. Relatedly, it would be beneficial to investigate fit patterns in re- gions where honor endorsement is low, in order to have more fine- grained insight into the implications of a mis- match for well- being (e.g., the well- being of immigrants from the Mediterranean region who reside in Western Europe) (see Gebauer et al., 2017, for a similar approach to religiosity). In addition to greater regional diversity, a stricter test of the generalizability of the current findings would also require future research to include greater di- versity in age, socio- economic status, and other forms of demographic background. This more representative sampling approach would be beneficial for the accuracy of our measures of both distal (allowing for comparisons of perceptions across various societal strata) and proxi- mal fit (examining fit with average scores of representa- tive samples, instead of students). Second, although we showed to some extent converg- ing fit patterns across both distal and proximal levels, our data do not provide any insights into the processes KIRCHNER-­HÄUSLER­et­al. 14676494, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jopy.12803 by Universidad De Granada, Wiley Online Library on [09/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 14 | through which fit in honor is linked to better subjec- tive well- being. While previous work has put forward some explanations as to how the two are linked (van Vianen, 2018), future research is needed to uncover the meaning and potential underlying processes of what it means to “fit in” when it comes to honoring. Finally, we assessed subjective well- being as satisfaction with several socio- economic domains of life and one's life in general. This, however, is in itself a particular perspec- tive on well- being, and other models of well- being may tap into more procedural facets of well- being, such as finding meaning or self- acceptance (Ryff, 2018). Future research could examine how fit effects unfold with other measures of well- being, to uncover which further ben- efits cultural fit might produce for well- being or how these benefits for satisfaction are realized within differ- ent life domains. 5 | CONCLU SIO N In the current research, we examined the link between cultural fit in honor and subjective well- being across eight communities in the Mediterranean region, re- sponding to calls to globalize psychological science (e.g., Thalmayer et al., 2021). Using a multi- faceted and multi- method approach to examine the role of honor fit in well- being, we found that a stronger distal fit (i.e., fit- ting in relatively more with one's perception of society) was linked to higher subjective well- being for some, but not all facets of honor. Furthermore, we also found that stronger proximal fit (i.e., fitting with one's university gender group) calculated across all honor facets was also linked to better subjective well- being. Our findings support previous work by demonstrating honor as an important social construct in the Mediterranean region for “living a good life,” and that a stronger fit with one's cultural environment is associated with positive psycho- logical outcomes. Overall, the current findings expand the cultural, conceptual, and methodological space of cultural fit research and highlight the need to consider the level at which fit occurs. AUTHOR CONTRIBUTIONS Alexander Kirchner- Häusler and Ayse K. Uskul con- ceived of the idea for the paper. Ayse K. Uskul, Alexander Kirchner- Häusler, and Vivian L. Vignoles designed the study. All co- authors contributed to translation of materi- als, conceptual feedback, and data collection. Alexander Kirchner- Häusler and Felix D. Schönbrodt conducted the analyses. Alexander Kirchner- Häusler, Ayse K. Uskul, and Felix D. Schönbrodt took the lead in writing the man- uscript with feedback from all co- authors. ACKNOWLEDGMENT The research was supported by a European Research Council Consolidator Grant (HONORLOGIC, 817577) awarded to Ayse K. Uskul. CONFLICT OF INT EREST The authors declare no conflict of interest. ET HICS STAT EMENT The study received approval from the ethical committees of all involved institutions or national bioethics commit- tees (Greek- Cypriot collection site). ORCID Alexander Kirchner- Häusler org/0000-0002-2406-7635 Felix D. Schönbrodt org/0000-0002-8282-3910 Ayse K. Uskul Vivian L. Vignoles org/0000-0002-7628-6776 Rosa Rodríguez- Bailón org/0000-0002-3489-0107 Vanessa A. Castillo org/0000-0002-8034-7631 Susan E. Cross Meral Gezici- Yalçın org/0000-0002-8751-3428 Charles Harb Shenel Husnu Keiko Ishii Panagiota Karamaouna org/0000-0002-8803-5413 Konstantinos Kafetsios org/0000-0001-5933-4409 Evangelia Kateri org/0000-0002-9366-9648 Juan Matamoros- Lima org/0000-0001-6669-8745 Rania Miniesy Jinkyung Na Zafer Özkan Stefano Pagliaro org/0000-0003-0573-0937 Charis Psaltis Manuel Teresi Yukiko Uchida https://orcid. https://orcid. https://orcid.org/0000-0001-8013-9931 https://orcid. https://orcid. https://orcid. https://orcid.org/0000-0002-9368-4397 https://orcid. https://orcid.org/0000-0003-3958-0092 https://orcid.org/0000-0001-5679-5568 https://orcid.org/0000-0003-0581-8138 https://orcid. https://orcid. https://orcid. https://orcid. https://orcid.org/0000-0001-9176-5989 https://orcid.org/0000-0002-7100-9606 https://orcid.org/0000-0001-7831-2491 https://orcid. https://orcid.org/0000-0001-8724-665X https://orcid.org/0000-0002-1103-7621 https://orcid.org/0000-0002-8336-2423 ENDN OTES 1 We derive these labels from their conceptual representations of fit with the broader society as well as fit with the narrower peer group but are aware that these labels do not perfectly correspond to the actual level of measurement (distal fit being calculated with KIRCHNER-­HÄUSLER­et­al. 14676494, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jopy.12803 by Universidad De Granada, Wiley Online Library on [09/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License people's own perceptions, and proximal fit being calculated based on actual group averages). 2 Due to an error, we mistakenly included the item “…your sister or mother had the reputation of sleeping around” instead of the correct item “…you were unable to defend your family's reputa- tion”, but as both items were part of the original family honor subscale of the original measurement tool and loaded as ex- pected in our exploratory factor analyses we decided to retain the item. 3 In a set of exploratory analyses, we also conducted all of our fit analyses with an SWB score that excluded the one item on the “quality of one's local environment”, including only purely psy- chological elements of well- being. We found no differences in the pattern of results. 4 In drawing upon RSA and profile fit analysis, the focus of the cur- rent paper is the individual level, i.e., how specific individuals fit in their perceived or actual environment and what this may mean for their well- being. While the samples we recruited in different parts of the Mediterranean may not be completely homogenous in their endorsement across all facets of honor, we did not test any hypotheses about the effect of these group differences and con- trolled for them where possible in our analyses. 5 Humberg et al. (2019) also outline the conditions for a strict con- gruence effect, which does not allow for the main effects of the predictors and for which two more conditions (a2 and a1 must not be significantly different from 0) must be met. We tested for strict congruence effects using a “squared difference model” in our model comparison approach (please see the supplementary materials for an overview). 6 We chose to use profile analysis rather than RSA to examine proximal fit since these calculated average group profiles show relatively little variation on the level of groups (in contrast to par- ticipants' perceptions of their society, which varied considerably between individuals). This fact makes RSA relatively less suited as in this case the response surface would reflect the level differ- ences between the groups rather than a fit of individuals (and potentially leading to estimation problems). In contrast, profile fit analysis does not suffer from the same drawbacks since the fit with the group average is calculated separately for each individual participant. 7 We collected data from only one university in all countries ex- cept Turkey, where we recruited participants from three different universities. R E F E R E N C E S Baldry, A. C., Pagliaro, S., & Porcaro, C. (2013). The rule of law at time of masculine honor: Afghan police attitudes and intimate partner violence. Group Processes & Intergroup Relations, 16(3), 363– 374. https://doi.org/10.1177/13684 30212 462492 Barnes, C. D., Brown, R. P., & Tamborski, M. (2012). Living danger- ously: Culture of honor, risk- taking, and the nonrandomness of “accidental” deaths. Social Psychological and Personality Science, 3(1), 100– 107. https://doi.org/10.1177/19485 50611 410440 | 15 L., Camerer, C., Cesarini, D., Chambers, C. D., Clyde, M., Cook, T. D., De Boeck, P., Dienes, Z., Dreber, A., Easwaran, K., Efferson, C., … Johnson, V. E. (2018). Redefine statistical sig- nificance. Nature Human Behaviour, 2(1), 6– 10. https://doi. org/10.1038/s4156 2- 017- 0189- z Bowman, J. (2006). Honor: A history. Encounter Books. Brown, R. P. (2016). Honor bound: How a cultural ideal has shaped the American psyche. Oxford University Press. Brown, R. P., Osterman, L. L., & Barnes, C. D. (2009). School vio- lence and the culture of honor. Psychological Science, 20(11), 1400– 1405. https://doi.org/10.1111/j.1467- 9280.2009.02456.x Campbell, J. (1964). Honour, family and patronage. A study of in- stitutions and moral values in a Greek mountain community. Clarendon Press. Cohen, D., Vandello, J., Puente, S., & Rantilla, A. (1999). “When you call me that, smile!” How norms for politeness, inter- action styles, and aggression work together in southern cul- ture. Social Psychology Quarterly, 62(3), 257. https://doi. org/10.2307/2695863 Cross, S. E., & Uskul, A. K. (2022). The pursuit of honor: Novel contexts, varied approaches, and new developments. In M. J. Gelfand, C. Chiu, & Y. Hong (Eds.), Advances in culture and psychology series. Oxford University Press. Cross, S. E., Uskul, A. K., Gerçek- Swing, B., Sunbay, Z., Alözkan, C., Günsoy, C., Ataca, B., & Karakitapoğlu- Aygün, Z. (2014). Cultural prototypes and dimensions of honor. Personality and Social Psychology Bulletin, 40(2), 232– 249. https://doi. org/10.1177/01461 67213 510323 De Leersnyder, J. (2017). Emotional acculturation: A first re- view. Current Opinion in Psychology, 17, 67– 73. https://doi. org/10.1016/j.copsyc.2017.06.007 De Leersnyder, J., Boiger, M., & Mesquita, B. (2021). What has cul- ture got to do with emotions? (A lot). In M. J. Gelfand (Ed.), Handbook of advances in cultural psychology (Vol. 8, pp. 62– 119). Oxford University Press. De Leersnyder, J., Mesquita, B., Kim, H., Eom, K., & Choi, H. (2014). Emotional fit with culture: A predictor of individual differences in relational well- being. Emotion, 14(2), 241– 245. https://doi. org/10.1037/a0035296 Dressler, W. W. (2012). Cultural consonance: Linking culture, the individual and health. Preventive Medicine, 55(5), 390– 393. https://doi.org/10.1016/j.ypmed.2011.12.022 Edwards, J. R. (2002). Alternatives to difference scores: Polynomial regression analysis and response surface methodology. In The Jossey- Bass Business & Management Series. Measuring and ana- lyzing behavior in organizations: Advances in measurement and data analysis (pp. 350– 400). Jossey- Bass. Edwards, J. R., & Cable, D. M. (2009). The value of value congru- ence. Journal of Applied Psychology, 94(3), 654– 677. https://doi. org/10.1037/a0014891 Enders, C. K., & Tofighi, D. (2007). Centering predictor variables in cross- sectional multilevel models: A new look at an old issue. Psychological Methods, 12(2), 121– 138. https://doi. org/10.1037/1082- 989X.12.2.121 Foster, S., Carvallo, M., & Lee, J. (2022). Feminine honor endorse- ment and young women's STI screenings. Psychology, Health & Medicine, 27, 1450– 1455. https://doi.org/10.1080/13548 506.2021.1883693 Benjamin, D. J., Berger, J. O., Johannesson, M., Nosek, B. A., Wagenmakers, E.- J., Berk, R., Bollen, K. A., Brembs, B., Brown, Fulmer, C. A., Gelfand, M. J., Kruglanski, A. W., Kim- Prieto, C., Diener, E., Pierro, A., & Higgins, E. T. (2010). On “feeling KIRCHNER-­HÄUSLER­et­al. 14676494, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jopy.12803 by Universidad De Granada, Wiley Online Library on [09/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 16 | right” in cultural contexts: How person- culture match affects self- esteem and subjective well- being. Psychological Science, 21(11), https://doi.org/10.1177/09567 97610 384742 1563– 1569. Gebauer, J. E., Sedikides, C., Schönbrodt, F. D., Bleidorn, W., Rentfrow, P. J., Potter, J., & Gosling, S. D. (2017). The religiosity as social value hypothesis: A multi- method replication and ex- tension across 65 countries and three levels of spatial aggrega- tion. Journal of Personality and Social Psychology, 113, e18– e39. https://doi.org/10.1037/pspp0 000104 Guerra, V. M., Giner- Sorolla, R., & Vasiljevic, M. (2013). The im- portance of honor concerns across eight countries. Group Processes & Intergroup Relations, 16(3), 298– 318. https://doi. org/10.1177/13684 30212 463451 Gul, P., & Schuster, I. (2020). Judgments of marital rape as a func- tion of honor culture, masculine reputation threat, and ob- server gender: A cross- cultural comparison between Turkey, Germany, and the UK. Aggressive Behavior, 46(4), 341– 353. https://doi.org/10.1002/ab.21893 Hogg, M., & Terry, D. (2000). Social identity and self- categorization in organizational contexts. The Academy of processes Management Review, 25(1), 121– 140. Humberg, S., Nestler, S., & Back, M. D. (2019). Response surface analysis in personality and social psychology: Checklist and clarifications for the case of congruence hypotheses. Social Psychological and Personality Science, 10(3), 409– 419. https:// doi.org/10.1177/19485 50618 757600 Killen, M., & Stangor, C. (2001). Children's social reasoning about inclusion and exclusion in gender and race peer group contexts. Child Development, 72(1), 174– 186. https://doi. org/10.1111/1467- 8624.00272 Kitayama, S., Karasawa, M., Curhan, K. B., Ryff, C. D., & Markus, H. R. (2010). Independence and interdependence predict health and wellbeing: Divergent patterns in the United States and Japan. Frontiers in Psychology, 1, 1– 10. https://doi.org/10.3389/ fpsyg.2010.00163 Kitayama, S., & Uskul, A. K. (2011). Culture, mind, and the brain: Current evidence and future directions. Annual Review of Psychology, 62(1), 419– 449. https://doi.org/10.1146/annur ev- psych - 12070 9- 145357 Korteweg, A., & Yurdakul, G. (2009). Islam, gender, and immigrant integration: Boundary drawing in discourses on honour killing in The Netherlands and Germany. Ethnic and Racial Studies, 32(2), 218– 238. https://doi.org/10.1080/01419 87080 2065218 Leung, A. K.- Y., & Cohen, D. (2011). Within- and between- culture variation: Individual differences and the cultural logics of honor, face, and dignity cultures. Journal of Personality and Social Psychology, 100(3), 507– 526. https://doi.org/10.1037/ a0022151 Li, L. M. W., & Hamamura, T. (2010). Cultural fit and life satisfac- tion: Endorsement of cultural values predicts life satisfaction only in collectivistic societies. Journal of Psychology in Chinese Societies: Hua Ren Xin Li Xue Bao, 11(2), 109– 122. Lopez- Zafra, E., Rodríguez- Espartal, N., & Ramos- Alvarez, M. M. (2020). Women's and men's role in culture of honor endorse- ment within families. European Journal of Women's Studies, 27(1), 72– 88. https://doi.org/10.1177/13505 06818 824369 Markus, H. R., & Kitayama, S. (2010). Cultures and selves: A cycle of mutual constitution. Perspectives on Psychological Science, 5(4), 420– 430. https://doi.org/10.1177/17456 91610 375557 McCrae, R. R. (2008). A note on some measures of profile agreement. Journal of Personality Assessment, 90(2), 105– 109. https://doi. org/10.1080/00223 89070 1845104 Mobasseri, S., Srivastava, S. B., & Goldberg, A. (2019). What is cul- tural fit?: From cognition to behavior (and back). In Oxford handbooks. The oxford handbook of cognitive sociology. Oxford University Press. Nisbett, R. E., & Cohen, D. (1996). Culture of honor: The psychology of violence in the south. Westview Press. Nisbett, R. E., Peng, K., Choi, I., & Norenzayan, A. (2001). Culture and systems of thought: Holistic versus analytic cognition. Psychological Review, 108(2), 291– 310. OECD. (2013). OECD guidelines on measuring subjective well- being. OECD. https://doi.org/10.1787/97892 64191 655- en Péristiany, J. G. (1965). Honour and shame: The values of Mediterranean society. University Press. Pitt- Rivers, J. (1965). Honor and social status. In Honor and shame: The values of Mediterranean society (pp. 19– 78). Weidenfeld and Nicolson. R Core Team. (2020). R: A language and environment for statistical computing [manual]. R Foundation for Statistical Computing. Rad, M. S., Martingano, A. J., & Ginges, J. (2018). Toward a psy- chology of homo sapiens: Making psychological science more representative of the human population. Proceedings of the National Academy of Sciences, 115(45), 11401– 11405. https:// doi.org/10.1073/pnas.17211 65115 Ramirez Marin, J., Aslani, S., & Brett, J. (2020). Strategic adaptation in intercultural negotiation: Spanish honor and US dignity negoti- ations. Academy of Management Global Proceedings, 2020, 187. Rodriguez Mosquera, P. M. (2016). On the importance of family, mo- rality, masculine, and feminine honor for theory and research: Facets of honor. Social and Personality Psychology Compass, 10(8), 431– 442. https://doi.org/10.1111/spc3.12262 Rodriguez Mosquera, P. M., Fischer, A. H., Manstead, A. S. R., & Zaalberg, R. (2008). Attack, disapproval, or withdrawal? The role of honour in anger and shame responses to being in- sulted. Cognition and Emotion, 22(8), 1471– 1498. https://doi. org/10.1080/02699 93070 1822272 Rodriguez Mosquera, P. M., Manstead, A. S. R., & Fischer, A. H. (2002). The role of honour concerns in emotional reactions to offences. Cognition & Emotion, 16(1), 143– 163. https://doi. org/10.1080/02699 93014 3000167 Rosseel, Y. (2012). Lavaan: An R package for structural equation modeling. Journal of Statistical Software, 48, 1– 36. Ryff, C. D. (2018). Eudaimonic well- being: Highlights from 25 years of inquiry. In Diversity in harmony from psychology: Proceedings of the 31st international congress of psychology (pp. 375– 395). Wiley Online Library. Salvador, C., Carlier, S. I., Ishii, K., Castillo, C. T., Nanakdewa, K., Alvaro, S. M., Savani, K., Kitayama, S. (2020). Emotionally expressive interdependence in Latin America: Triangulating through a comparison of three cultural region. PsyArXiv. https://doi.org/10.31234/ osf.io/pw4yk. San Martin, A., Sinaceur, M., Madi, A., Tompson, S., Maddux, W. W., & Kitayama, S. (2018). Self- assertive interdependence in Arab culture. Nature Human Behaviour, 2(11), 830– 837. https://doi. org/10.1038/s4156 2- 018- 0435- z Schönbrodt, F. D. (2016). Testing fit patterns with polynomial regres- sion models. OSF Preprints, 1– 20. https://doi.org/10.31219/ osf. io/ndggf KIRCHNER-­HÄUSLER­et­al. 14676494, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jopy.12803 by Universidad De Granada, Wiley Online Library on [09/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Schwartz, S. H. (1992). Universals in the content and structure of values: Theoretical advances and empirical tests in 20 coun- tries. In Advances in experimental social psychology (Vol. 25, pp. 1– 65). Elsevier. https://doi.org/10.1016/S0065 - 2601(08)60281 - 6 Smith, P. B., Easterbrook, M. J., Blount, J., Koç, Y., Harb, C., Torres, C. V., Ahmad, A. H., Ping, H., Celikkol, G. C., Loving, R. D., & Rizwan, M. (2017). Culture as perceived context: An explora- tion of the distinction between dignity, face and honor cultures. Acta de Investigación Psicológica, 7(1), 2568– 2576. https://doi. org/10.1016/j.aipprr.2017.03.001 Stephens, N. M., Townsend, S. S. M., Markus, H. R., & Phillips, L. T. (2012). A cultural mismatch: Independent cultural norms pro- duce greater increases in cortisol and more negative emotions among first- generation college students. Journal of Experimental Social Psychology, 48(6), 1389– 1393. https://doi.org/10.1016/j. jesp.2012.07.008 Survey Research Center. (2022). Cross- cultural survey guide- https://ccsg.isr.umich.edu/chapt ers/trans latio n/overv lines. iew/#Team_trans lation Thalmayer, A. G., Toscanelli, C., & Arnett, J. J. (2021). The ne- glected 95% revisited: Is American psychology becoming less American? American Psychologist, 76(1), 116– 129. https://doi. org/10.1037/amp00 00622 Uskul, A. K., & Cross, S. E. (2019). The social and cultural psychol- ogy of honour: What have we learned from researching honour in Turkey? European Review of Social Psychology, 30(1), 39– 73. https://doi.org/10.1080/10463 283.2018.1542903 Uskul, A. K., & Cross, S. E. (2020). Socio- ecological roots of cultures of honor. Current Opinion in Psychology, 32, 177– 180. https:// doi.org/10.1016/j.copsyc.2019.11.001 Uskul, A. K., Cross, S. E., Günsoy, C., & Gul, P. (2019). Cultures of honor. In S. Kitayama & D. Cohen (Eds.), Handbook of cultural psychology (2nd ed., pp. 793– 821). The Guilford Press. Uskul, A. K., Cross, S. E., Sunbay, Z., Gercek- Swing, B., & Ataca, B. (2012). Honor bound: The cultural construction of honor in Turkey and the northern United States. Journal of Cross- Cultural Psychology, 43(7), 1131– 1151. https://doi.org/10.1177/00220 22111 422258 | 17 van Vianen, A. E. M. (2018). Person: A review of its basic tenets. Annual Review of Organizational Psychology and Organizational Behavior, 5(1), 75– 101. https://doi.org/10.1146/annur ev- orgps ych- 03211 7- 104702 Vignoles, V. L., Owe, E., Becker, M., Smith, P. B., Easterbrook, M. J., Brown, R., Gonzalez, R., Didier, N., Carrasco, D., Paz Cadena, M., Lay, S., Schwartz, S. J., Des Rosiers, S. E., Villamar, J. A., Gavreliuc, A., Zinkeng, M., Kreuzbauer, R., Baguma, P., Martin, M., … Bond, M. H. (2016). Beyond the “eastwest” dichotomy: Global variation in cultural models of selfhood. Journal of Experimental Psychology: General, 145(8), 966– 1000. https:// doi.org/10.1037/xge00 00175 Yao, J., Ramirez- Marin, J., Brett, J., Aslani, S., & Semnani- Azad, Z. (2017). A measurement model for dignity, face, and honor cul- tural norms. Management and Organization Review, 13(4), 713– 738. https://doi.org/10.1017/mor.2017.49 SUPPORT IN G INFORMAT ION Additional supporting information can be found online in the Supporting Information section at the end of this article. How to cite this article: Kirchner-Häusler, A., Schönbrodt, F. D., Uskul, A. K., Vignoles, V. L., Rodríguez-Bailón, R., Castillo, V. A., Cross, S. E., Gezici-Yalçın, M., Harb, C., Husnu, S., Ishii, K., Karamaouna, P., Kafetsios, K., Kateri, E., Matamoros-Lima, J., Miniesy, R., Na, J., Özkan, Z., Pagliaro, S., … Uchida, Y. (2023). Proximal and distal honor fit and subjective well-being in the Mediterranean region. Journal of Personality, 00, 1– 17. https://doi.org/10.1111/jopy.12803 KIRCHNER-­HÄUSLER­et­al. 14676494, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jopy.12803 by Universidad De Granada, Wiley Online Library on [09/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
10.1016_j.isci.2023.106601
iScience ll OPEN ACCESS Article The human E3 ligase RNF185 is a regulator of the SARS-CoV-2 envelope protein Charles Zou, Hojong Yoon, Paul M.C. Park, ..., Robert A. Davey, Benjamin L. Ebert, Mikołaj Słabicki benjamin_ebert@dfci.harvard. edu (B.L.E.) slabicki@broadinstitute.org (M.S.) Highlights RNF185, a human E3 ligase, regulates the stability of the SARS-CoV- 2 envelope protein RNF185 and the SARS- CoV-2 envelope protein co-localize at the endoplasmic reticulum Depletion of RNF185 significantly increases SARS-CoV-2 viral titer in a cellular model Zou et al., iScience 26, 106601 May 19, 2023 ª 2023 The Authors. https://doi.org/10.1016/ j.isci.2023.106601 iScience ll OPEN ACCESS Article The human E3 ligase RNF185 is a regulator of the SARS-CoV-2 envelope protein Charles Zou,1 Hojong Yoon,1,2 Paul M.C. Park,1,2 J.J. Patten,3 Jesse Pellman,1,2 Jeannie Carreiro,1,2 Jonathan M. Tsai,1,2 Yen-Der Li,1,2 Shourya S. Roy Burman,4,5 Katherine A. Donovan,4,5 Jessica Gasser,1,2 Adam S. Sperling,1,2,6 Radosław P. Nowak,4,5 Eric S. Fischer,4,5 Robert A. Davey,3 Benjamin L. Ebert,1,2,7,* and Mikołaj Słabicki1,2,8,* SUMMARY Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) hijacks multiple human proteins during infection and viral replication. To examine whether any viral proteins employ human E3 ubiquitin ligases, we evaluated the stability of SARS-CoV-2 proteins with inhibition of the ubiquitin proteasome pathway. Using genetic screens to dissect the molecular machinery involved in the degradation of candidate viral proteins, we identified human E3 ligase RNF185 as a regulator of protein stability for the SARS-CoV-2 envelope protein. We found that RNF185 and the SARS-CoV-2 envelope co-localize to the endoplasmic reticulum (ER). Finally, we demonstrate that the depletion of RNF185 significantly increases SARS-CoV-2 viral titer in a cellular model. Modulation of this interaction could provide opportunities for novel antiviral therapies. INTRODUCTION Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) is responsible for the current devastating global pandemic. While vaccines have achieved tremendous success, morbidity and mortality from SARS-CoV-2 continues to be a public health crisis due to evolution of new viral variants, infections in unvac- cinated or immunocompromised individuals, and breakthrough infections in those who have been vaccinated. Several SARS-CoV-2 antiviral therapies are in clinical use or development, including antibody cocktails that aim to block the receptor-binding domain of the SARS-CoV-2 spike protein1, small molecules that inhibit the host RNA polymerase,2,3 small molecules that incorporate missense mutations into newly synthesized viral RNA,4 and small-molecule inhibitors of the viral main protease (MPro).5 Efforts to understand the viral cycle and map the function of each SARS-CoV-2 protein have identified potential antiviral targets.6,7 Comprehensive CRISPR-Cas9 screens have revealed several mammalian host genes critical for viral entry and replication including the ACE2 receptor,8 HMGB1, the SWI/SNF chromatin remodeling complex,9 and the lysosomal proteins TMEM106B10 and TMEM41B.11 Some viral proteins modulate the ubiquitin-proteasome system (UPS) by redirecting E3 ubiquitin ligases to target and degrade host proteins, augmenting viral infection.12 For example, the human papillomavirus E6 oncoprotein redirects the UBE3A ubiquitin ligase to induce degradation of the tumor suppressor p53.13 Among coronaviruses, SARS-CoV ORF-9b hijacks two ubiquitin ligases, PCBP2 and AIP4, to degrade MAVS, TRAF3, and TRAF6, thereby impairing the host interferon response.14 It was recently reported that SARS-CoV-2 ORF10 interacts with CUL2-ZYG11B,6 though it is not yet clear whether this interaction is required for infection.15 Though less well studied, it is also likely that host E3 ligases regulate viral protein stability, providing a mechanism for viral peptide quality control and clearance. Targeted protein degradation has emerged as a powerful pharmacologic approach to alter the abundance of proteins in cells.16,17 One potential antiviral strategy is to modulate the interaction of a viral protein with a host ubiquitin ligase using a small molecule to promote its removal by the UPS. We used CRISPR screens to identify SARS-CoV-2 proteins targeted for degradation by human E3 ubiquitin ligases. Such E3 ligase–viral 1Department of Medical Oncology, Dana-Farber Cancer Institute, Boston, MA 02215, USA 2Broad Institute of MIT and Harvard, Cambridge, MA, USA 3Department of Microbiology, Boston University School of Medicine and NEIDL, Boston University, Boston, MA 02118, USA 4Department of Cancer Biology, Dana-Farber Cancer Institute, Boston, MA 02215, USA 5Department of Biological Chemistry and Molecular Pharmacology, Harvard Medical School, Boston, MA 02115, USA 6Division of Hematology, Department of Medicine, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, USA 7Howard Hughes Medical Institute, Boston, MA, USA 8Lead contact *Correspondence: benjamin_ebert@dfci. harvard.edu (B.L.E.), slabicki@broadinstitute.org (M.S.) https://doi.org/10.1016/j.isci. 2023.106601 iScience 26, 106601, May 19, 2023 ª 2023 The Authors. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). 1 ll OPEN ACCESS iScience Article A B ] O S M D / t n e m a e r T t [ y r r e h C m / 2 V o C S R A S P F G e 3 2 1 0 SARS-CoV-2 Protein Stability Reporters SARS-CoV-2 EGFP IRES mCherry MG132 Treatment: proteasome inhibitor MLN7243 E1 inhibitor MLN4924 Neddylation inhibitor Flow analysis MLN7243 MG132 MLN4924 No Treatment b 3 f r O 6 f r O 0 1 f r O 8 f r O c 9 f r O l e p o e v n E 7 p s N 6 p s N 6 1 p s N 8 p s N b 7 f r O b 9 f r O a 7 f r O 2 p s N 5 1 p s N 4 1 p s N 0 1 p s N Proteins selected for further characterization i t n e o r p o e c u N l 5 p s N 9 p s N a 3 f r O Figure 1. The stability of SARS-CoV-2 proteins in HEK293T cells (A) Schematic overview of protein stability assay for SARS-CoV-2 proteins. (B) Flow cytometry analysis of SARS-CoV-2 protein stability reporters after 6 h of treatment with 10 mM of MG132, 1 mM MLN7243, 1 mM MLN4924, or DMSO (bars represent mean, n = 3). Black line represents mean eGFP/mCherry expression of non-treated cells; red line represents a 30% increase of expression. protein interacting pairs can serve as starting points for the identification of molecular glues that strengthen such interactions and boost antiviral activity. RESULTS Identification of SARS-CoV-2 proteins degraded by the human ubiquitin-proteasome system We first sought to identify SARS-CoV-2 proteins that are degraded by human E3 ubiquitin ligases. To monitor the stability of SARS-CoV-2 proteins, we generated a set of fluorescent reporter constructs. Each construct contains the sequence of a single, full-length SARS-CoV-2 protein fused in-frame with GFP, eGFP, followed by an internal ribosome entry site and red fluorescent protein, mCherry, for signal normalization (Figure 1A).18 Out of 28 SARS-CoV-2 proteins, we successfully generated 21 reporter cell lines in human HEK293T cells (Table S1). To test whether stability of the viral proteins is influenced by the UPS, we treated reporter-expressing cells with a ubiquitin-activating enzyme inhibitor (MLN7243), a 26S proteasome inhibitor (MG132), or a neddy- lation inhibitor (MLN4924), and monitored eGFP/mCherry expression levels of treated cells using flow cy- tometry normalized to untreated controls (Figure 1B). Inhibitor treatment affected the levels of eGFP- SARS-CoV-2 proteins without affecting the mCherry controls (Figure S1). If the levels of eGFP/mCherry expression increased over 30% upon inhibitor treatment (p value <0.01), we considered the SARS-CoV-2 protein reporter to be destabilized by the human UPS. As expected, all reporters that were stabilized by MLN7243 E1 inhibitor treatment were also stabilized by inhibition of proteasome with MG132. Treatment with MLN4924, however, did not change the stability of any SARS-CoV-2 eGFP fusions, indicating that no SARS-CoV-2 proteins are degraded by cullin ring E3 ligases in the cell line tested.19 By our criteria, ten out of the twenty-one SARS-CoV-2 fusion proteins tested here are degraded by the human UPS (Figure 1B). 2 iScience 26, 106601, May 19, 2023 iScience Article ll OPEN ACCESS Genetic screens for human E3 ligases that destabilize SARS-CoV-2 proteins We next sought to identify the specific E3 ubiquitin ligase components responsible for destabilization of the ten identified SARS-CoV-2 proteins. We transduced the reporter-expressing HEK293T cell line with a single-guide RNA (sgRNA) library targeting (cid:1)700 E3 ligases, E2 conjugating enzymes, and deubiquitinat- ing enzymes (DUBs).20,21 The fluorescent reporter cell lines were sorted into four populations based on eGFP/mCherry ratio ((Gate A) bottom 5%, (Gate B) bottom 5%–10%, (Gate C) top 5%–10%, (Gate D) top 5%)) (Figure 2A). Cells sorted into Gates A and B had reduced eGFP/mCherry expression, indicating rela- tively enhanced degradation of the SARS-CoV-2 eGFP fusion protein, whereas cells sorted into Gates C and D had relatively increased stability of SARS-CoV-2 eGFP fusion protein. After sorting, cell pellets were lysed and sgRNAs were amplified and quantified by next-generation sequencing. We compared sgRNAs read counts from the SARS-CoV-2 most stable population (Gate D) to the most unstable popula- tion (Gate A). For eight out of the ten SARS-CoV-2 protein reporters screened, the genetic screen revealed significant enrichment for at least one E3 ligase (Figures 2B, 2C and S2). Two of the most significantly enriched ligases in the stable population compared to the unstable population were RNF185, in the SARS-CoV-2 envelope- eGFP screen (Figure 2B), and SYVN1, in the SARS-CoV-2 Orf8-eGFP screen (Figure S3A). The E3 ubiquitin- protein ligase UBE3C, which scored in six screens, has been shown to influence the rate of degradation for domains fused to eGFP,22 and was therefore excluded as a potential assay artifact. Validation of E3 ligases that induce degradation of SARS-CoV-2 proteins We next validated the two identified pathogen-host pairs, envelope – RNF185 and Orf8 – SYVN1, in addi- tional cellular contexts. We tested the effect of UPS inhibitors on the stability of envelope-eGFP and Orf8- eGFP fusions in A549 (lung epithelial) cells and K562 (acute myeloid leukemia) cells. We treated the re- porter cell lines with a set of ubiquitin-proteasome inhibitors and monitored the levels of SARS-CoV-2 eGFP fusion. For the SARS-CoV-2 envelope protein, we observed an increase of protein stability in A549 and K562 cells upon the treatment with MLN7243 and MG132 (p.value <0.01), but not MLN4924, consistent with the results for HEK293T cells (Figures S3B and S3C). The SARS-CoV-2 Orf8 protein was stabilized upon MLN7243 and MG132 treatment in HEK293T and K562 cells, but not in the A549 cell line (Figure S3D), sug- gesting that one or more factors were needed to regulate the envelope protein, and that ORF8 is differen- tially expressed in these cell lines. We next sought to validate the E3 ligases that scored in our genetic screens, RNF185 and SYVN1, by tar- geting them with multiple independent sgRNAs in three cell lines. We knocked out RNF185 and SYVN1 with four sgRNAs each and compared the levels of SARS-CoV-2 envelope protein or Orf8 in the eGFP/ mCherry expression system to a non-targeting control guide by flow cytometry. CRISPR-Cas9 knockout of RNF185 efficiently depleted RNF185 protein levels (Figure 2D), which resulted in increased expression of SARS-CoV-2 envelope-eGFP by 2- to 3-fold in all cell lines tested (p.value <0.01) (Figure 2E), validating the involvement of RNF185 in the stability of the envelope protein. The knockout of SYVN1 in HEK293T and K562 cells significantly increased levels of SARS-CoV-2 Orf8-GFP (Figure S3D), but not in A549 cells. RNF185 colocalizes with the SARS-CoV-2 envelope protein Since RNF185 knockout consistently rescued envelope protein degradation across multiple individual sgRNAs and three cell lines, we decided to focus on this E3 ligase. RNF185 has previously been shown to form an endoplasmic reticulum (ER)-associated protein degradation (ERAD) complex involved in quality control of proteins on the ER membrane.23,24 To investigate the cellular localization of the RNF185 and en- velope proteins, HEK293T cells stably expressing the SARS-CoV-2 envelope-eGFP fusion protein cells were transfected with a RNF185-iRFP720 fusion construct and stained with an ER dye. RNF185 and the SARS- CoV-2 envelope protein partially co-localized with the ER marker (Figure 3A). Prior studies have demonstrated that TMEM259 is necessary for ubiquitin ligase RNF185 function in the context of the ERAD complex for the quality control of membrane proteins.23 To examine whether this complex is needed for targeting the viral envelope protein, we used CRISPR-Cas9 to knock out TMEM259 and monitored SARS-CoV-2 envelope protein levels by flow cytometry. Depletion of TMEM259 by sgRNAs increased the stability of SARS-CoV-2 envelope protein (Figure 3B), confirming that an ERAD complex member is involved in modulating SARS-CoV-2 envelope levels. iScience 26, 106601, May 19, 2023 3 A B D ) e u a v l p ( 0 1 g o l 5 4 3 2 1 0 RNF185 -Actin ll OPEN ACCESS iScience Article sgRNA library eGFPSARS-CoV-2 reporter HEK293TCas9 SARS-CoV-2 stable Gate D Gate C - 2 - V o C S R A S P F G e mCherry Gate B FACS SARS-CoV-2 unstable Gate A Next Generation Sequencing Envelope RNF185 Gate A C B D 5 P S U C 3 E B U 1 2 Fold change (stable / unstable) 1000 500 t n u o c d a e r d e z i l a m r o N E * * * * 5 8 1 F N R C n o i t a d a r g e D , y r r e h C m P F G d e z / i l a m r o n [ ] O S M D / t n e m t a e r T 3 2 1 0 MLN7243 MG132 MLN4924 No Treatment Envelope * * * * * * * * * * * * HEK293T A549 Cell Line K562 sgRNF185 A sgRNF185 B sgRNF185 C sgRNF185 D sgNTC 0 1 HEK293T Envelope A549 K562 * * * * * * * * * n s * * * * 2 2 Degradation [normalized GFP/mCherry, sgRNA / sgNTC] 4 0 4 0 3 1 3 2 1 * * * * * * * * 3 4 sgRNF185 A sgNTC sgRNF185 B Figure 2. Functional genomic screens identified RNF185 ligases for SARS-CoV-2 envelope stability (A) Schematic of CRISPR sorting screen of the SARS-CoV-2 protein stability. Reporter cell lines were transduced with Bison CRISPR sgRNA library (713 E1, E2, and E3 ubiquitin ligases, deubiquitinases, and control genes containing 2,852 sgRNAs), cultured for 8 days, then sorted into 4 gates based on GFP/mCherry ratio. (Gate A: bottom 0%–5%; Gate B: bottom 5%–10%; Gate C: top 90%–95%; Gate D: top 95%–100%). (B) CRISPR-Cas9 sorting screen for cells with altered stability of SARS-CoV-2 envelope. Guides were collapsed to gene levels (n = 2, four guides per gene, two-sided empirical ran-sum test-statistic, horizontal dashed line indicates p value = 0.0005, vertical dashed line indicates fold change of 2, vertical dot line indicates fold change of 1.6). Normalized read count for sgRNAs targeting RNF185 is shown. (C) Flow cytometry analysis of degradation of SARS-CoV-2 envelope in a panel of three cell lines after 6 h of treatment with 10 mM of MG132, 1 mM MLN7243, 1 mM MLN4924, or DMSO (** p value <0.01, ns – not significant). (D) Immunoblots of RNF185 levels after infection with sgRNF185 or sgNTC. (E) Flow cytometry analysis of SARS-CoV-2 envelope reporter stability following single-guide knockout of RNF185 across three different cell lines (* p value <0.05, ** p value <0.01). 4 iScience 26, 106601, May 19, 2023 iScience Article ll OPEN ACCESS A B Sars-CoV2 Envelope RNF185 ER GFP iRFP720 DAPI Blue stain Merge 10 µm sgTMEM259 sgRNA A sgTMEM259 sgRNA B sgNTC Envelope n s * sgRNA A sgRNA B sgNTC 0.0 0.5 1.0 1.5 Degradation [normalized GFP/mCherry, sgRNA / sgNTC] Figure 3. Degradation of SARS-CoV-2 envelope protein is mediated by RNF185/TMEM259 complex (A) Fluorescent microscopy of HEK293T cells transiently transfected with iRFP720 tagged RNF185 and GFP tagged envelope proteins. (B) Flow cytometry analysis of CRISPR-Cas9 knocked out TMEM259 compared to a non-targeting control. RNF185 is involved in the degradation of SARS-CoV-2 clinical variants as well as SARS-CoV envelope proteins Since initial isolation of SARS-CoV-2, the envelope protein has been one of the most rapidly evolving pro- teins for subsequent variants of the virus. The most common mutations in the SARS-CoV-2 envelope pro- tein are in the C-terminus of the protein (amino acids (aa) 55–73) and the transmembrane domain (aa 9). To test whether RNF185 is still able to modulate envelope protein levels in these variants, we expressed them as eGFP fusions in HEK293T cells. Depletion of RNF185 increases the levels of SARS-CoV-2 envelope for all variants tested (Figure 4A). SARS-CoV is closely related to SARS-CoV-2 and their envelope proteins share a high degree of protein sequence homology (91% sequence identity), while the MERS envelope protein is more distinct (35% sequence identity) (Figure 4B). Having observed RNF185-mediated degradation of the SARS-CoV-2 envelope protein, we sought to determine whether degradation of the SARS-CoV and MERS envelope proteins is mediated by the same E3 ligase. The SARS-CoV and MERS envelope proteins were fused to eGFP in our reporter vector and expressed in HEK293T cells. We observed that sgRNA targeting RNF185 stabilized protein levels for both the SARS-CoV-2 and the SARS-CoV envelope proteins, albeit to a greater degree for SARS-CoV-2, indicating that RNF185 is involved in the degradation of the envelope protein of both SARS viruses. In contrast, the MERS envelope pro- tein stability was not affected by RNF185 depletion (Figure 4C) consistent with observed differences between the sequences of the envelope protein in MERS compared to that of SARS-CoV and SARS-CoV-2. RNF185 knockout increases SARS-CoV-2 viral titer To examine whether RNF185 is relevant to SARS-CoV-2 virus production, we examined viral titer in host cells with RNF185 inactivation. We knocked out RNF185 in A549-ACE2 cells using CRISPR/Cas9 and in- fected the cells with three SARS-CoV-2 strains: WA, Beta, or Delta.25 Three days following infection, viral titers were measured by a PCR release assay. Knockout of RNF185 resulted in a (cid:1)60% increase in viral titers for all three SARS-CoV-2 strains which was statistically significant for the Beta and Delta variants (Figure 5). DISCUSSION In this study, we systematically examined whether SARS-CoV-2 proteins are destabilized by human E3 ubiq- uitin ligases, and we report that RNF185 induces degradation of the SARS-CoV-2 envelope protein. We iScience 26, 106601, May 19, 2023 5 ll OPEN ACCESS iScience Article * * * * * * * * * * n s * * * * * * * * * * * * * * * * 0 Degradation [normalized GFP/mCherry, sgRNA / sgNTC] 2 1 4 3 Envelope sgRNF185 A sgRNF185 B sgNTC * * * * * * WT T9I S55F V62F S68F R69I P71S L73F A B SARS-CoV-2 SARS-CoV MERS C Envelope SARS2 Envelope SARS Envelope MERS * * * * * * * * sgRNF185 A sgRNF185 B sgNTC 0.0 0.5 1.0 1.5 Degradation [normalized GFP/mCherry, sgRNA / sgNTC] Figure 4. RNF185-mediated degradation of the envelope protein is seen in SARS-CoV-2 wild-type, clinical variants as well as SARS-CoV but not MERS (A) Flow cytometry analysis of degradation of clinical variants of SARS-CoV-2 envelope protein following infection of CRISPR guides targeting RNF185. (B) Sequence alignment of envelope protein for SARS-CoV-2, SARS-CoV, and MERS. (C) Flow cytometry analysis of degradation of SARS-CoV-2, SARS-CoV, and MERS envelope protein following infection of CRISPR guides targeting RNF185. demonstrated that the SARS-CoV-2 envelope colocalizes with RNF185. Furthermore, depletion of TMEM259, a member of the ERAD complex required for RNF185 function, phenocopies the effect of RNF185 knockout. Our results demonstrate that RNF185 modulates envelope protein levels for several SARS-CoV-2 clinical variants and for SARS-CoV, but not for MERS. Finally, in viral titer assays, depletion of RNF185 increased viral titer, demonstrating that the role of RNF185 in regulating envelope protein levels has biological consequences for virus production. 6 iScience 26, 106601, May 19, 2023 iScience Article ll OPEN ACCESS Beta * Delta * 4e+09 3e+09 2e+09 1e+09 0e+00 WA ns 1.5e+09 1.0e+09 5.0e+08 0.0e+00 1.0e+10 7.5e+09 5.0e+09 2.5e+09 0.0e+00 C T N g s 5 8 1 F N R g s C T N g s 5 8 1 F N R g s C T N g s 5 8 1 F N R g s sgRNA sgRNF185 sgNTC Figure 5. SARS-CoV-2 virus titers in A549-ACE2 cell line upon RNF185 depletion Virus titers were measured by detection of genomes in culture supernatants by PCR-based assay (* p value <0.05, ns – not significant). Titers are expressed as genome equivalents per mL of culture supernatant. RNF185 is a non-cullin E3 ubiquitin ligase that is part of the ERAD family of E3 ligases and is important for quality control of ER-synthesized proteins, targeting misfolded or unfolded proteins for degradation to reduce cellular stress.23 RNF185 was previously implicated in cGAS-mediated innate immune response upon HSV-1 infection,22 suggesting that RNF185 may be more broadly involved in viral pathogenic processes. The SARS-CoV-2 envelope protein is synthesized in the ER and subsequently trafficked to the Golgi and ER- Golgi intermediate compartment where it is involved in multiple steps of the viral life cycle: viral assembly, budding, viral release, and inflammasome activation.26 Our results indicate that RNF185 is involved in degradation of the envelope protein. The SARS-CoV-2 envelope protein has been suggested to be a good target for an antiviral therapeutic that induces protein degradation as it would inhibit viral entry, repli- cation, and assembly.27 In multiple cases, a small molecule can augment ubiquitination by a protein’s cognate E3 ligase. For example, small-molecule-induced polymerization of the transcriptional repressor BCL6 leads to its enhanced ubiquitination and degradation by the E3 ligase SIAH1,19 and small-molecule enhancement of the protein-protein interaction between the transcription factor b-catenin and its endogenous E3 ligase b(cid:3)TrCP28 leads to degradation of b-catenin. A small molecule that increases the affinity of the SARS- SCF CoV-2 envelope protein to RNF185 could be an effective antiviral treatment strategy to reduce the effective concentration of viral structural components, lowering the barrier to ubiquitination and clearance. In addi- tion, a small-molecule binder to the envelope protein could increase the retention time of the SARS-CoV-2 envelope protein in the ER, leading to enhanced degradation by the RNF185 E3 ubiquitin ligase. The finding that RNF185 degrades SARS-CoV-2 envelope protein provides a rationale for seeking small mole- cules that strengthen this interaction, thereby enhancing degradation of the SARS-CoV-2 envelope pro- tein, resulting in antiviral activity. Limitations of the study A potential limitation of this study is that we only evaluated the stability of SARS-CoV-2 proteins in human HEK293T cells, which may not fully represent the diverse range of host cells and tissues infected by the virus in vivo. While this study identified a potential target for antiviral therapeutics, additional research is required to explore the biochemical and structural basis for the impact of RNF185 on SARS-CoV-2 enve- lope protein stability and the therapeutic potential of modulating the interaction between RNF185 and the envelope protein. STAR+METHODS Detailed methods are provided in the online version of this paper and include the following: d KEY RESOURCES TABLE d RESOURCE AVAILABILITY B Lead contact iScience 26, 106601, May 19, 2023 7 iScience Article ll OPEN ACCESS B Material availability B Data and code availability d EXPERIMENTAL MODEL AND SUBJECT DETAILS B Cell lines B Generation of stability constructs B Lentivirus production B Lentiviral transduction B SARS-CoV-2 stability assay B BISON SARS-CoV-2 reporter screen in HEK293T cells B Data analysis of CRISPR-Cas9 knockout screens B Generation of CRISPR-Cas9 knock-out cells B Immunoblots B Transient transfection B Fluorescence microscopy B Virus cultivation B Effect of RNF185 KO on SARS-CoV-2 replication SUPPLEMENTAL INFORMATION Supplemental information can be found online at https://doi.org/10.1016/j.isci.2023.106601. ACKNOWLEDGMENTS M.S. is supported by Scientific Projects to Accelerate Research and Collaboration (Broad Institute). H.Y. is supported by NIH grant NCI K00CA253754. J.J.P. and R.A.D. were supported by a grant from MassCPR. A.S.S. was supported by NCI grant K08-CA252174 and DOD grant CA21827. We are grateful to Nathanael S. Gray and all members of the Ebert laboratory for discussion, particularly, Peter G. Miller, Roger Belizaire, Marek Nagiec, and Christopher Hergott. AUTHOR CONTRIBUTIONS M.S., C.Z., B.L.E. conceptualized and initiated the study; M.S. designed, and C.Z. performed experiments with the help of H.Y., P.M.C.P., J.J.P., J.P., J.C., J.M.T., Y.D.L., S.S.R.B., K.A.D., J.G., A.S.S., R.P.N. M.S., B.L.E., R.P.N., E.S.F., R.A.D. supervised the project. M.S., C.Z., B.L.E. wrote the manuscript with input from all authors. DECLARATION OF INTERESTS M.S. has received research funding from Calico Life Sciences LLC. B.L.E. has received research funding from Celgene, Deerfield, Novartis, and Calico Life Sciences LLC and consulting fees from GRAIL. He is a member of the scientific advisory board and shareholder for Neomorph Inc., TenSixteen Bio, Skyhawk Therapeutics, and Exo Therapeutics. E.S.F. is a founder, member of the scientific advisory board (SAB), and equity holder of Civetta Therapeutics, Proximity Therapeutics, and Neomorph Inc (also board of direc- tors), a SAB member and equity holder in Avilar Therapeutics and Photys Therapeutics, equity holder in Lighthorse Therapeutics, and a consultant to Sanofi, Novartis, Deerfield, Odyssey Therapeutics, and EcoR1 capital. The Fischer laboratory receives or has received research funding from Novartis, Deerfield, Ajax, Interline, and Astellas. K.A.D. is a consultant to Kronos Bio and Neomorph Inc. A.S.S. reports consul- ting fees from Adaptive Technologies and Roche. INCLUSION AND DIVERSITY We support inclusive, diverse, and equitable conduct of research. Received: March 31, 2022 Revised: January 31, 2023 Accepted: March 29, 2023 Published: April 8, 2023 8 iScience 26, 106601, May 19, 2023 iScience Article REFERENCES 1. Weinreich, D.M., Sivapalasingam, S., Norton, T., Ali, S., Gao, H., Bhore, R., Musser, B.J., Soo, Y., Rofail, D., Im, J., et al.; Trial Investigators (2021). REGN-COV2, a neutralizing antibody cocktail, in outpatients with Covid-19. N. Engl. J. Med. 384, 238–251. 2. Caly, L., Druce, J.D., Catton, M.G., Jans, D.A., and Wagstaff, K.M. (2020). The FDA- approved drug ivermectin inhibits the replication of SARS-CoV-2 in vitro. Antiviral Res. 178, 104787. 3. Wang, M., Cao, R., Zhang, L., Yang, X., Liu, J., Xu, M., Shi, Z., Hu, Z., Zhong, W., and Xiao, G. (2020). Remdesivir and chloroquine effectively inhibit the recently emerged novel coronavirus (2019-nCoV) in vitro. Cell Res. 30, 269–271. 4. Sheahan, T.P., Sims, A.C., Zhou, S., Graham, R.L., Pruijssers, A.J., Agostini, M.L., Leist, S.R., Scha¨ fer, A., Dinnon, K.H., 3rd, Stevens, L.J., et al. (2020). An orally bioavailable broad- spectrum antiviral inhibits SARS-CoV-2 in human airway epithelial cell cultures and multiple coronaviruses in mice. Sci. Transl. Med. 12, eabb5883. 5. Owen, D.R., Allerton, C.M.N., Anderson, A.S., Aschenbrenner, L., Avery, M., Berritt, S., Boras, B., Cardin, R.D., Carlo, A., Coffman, K.J., et al. (2021). An oral SARS-CoV-2 M(pro) inhibitor clinical candidate for the treatment of COVID-19. Science 374, 1586–1593. 6. Gordon, D.E., Jang, G.M., Bouhaddou, M., Xu, J., Obernier, K., O’Meara, M.J., Guo, J.Z., Swaney, D.L., Tummino, T.A., Huttenhain, R., et al. (2020). A SARS-CoV-2-human protein- protein interaction map reveals drug targets and potential drug-repurposing. Nature 583, 459–468. 7. Gordon, D.E., Jang, G.M., Bouhaddou, M., Xu, J., Obernier, K., White, K.M., O’Meara, M.J., Rezelj, V.V., Guo, J.Z., Swaney, D.L., et al. (2020). A SARS-CoV-2 protein interaction map reveals targets for drug repurposing. Nature 583, 459–468. 8. Zhu, Y., Feng, F., Hu, G., Wang, Y., Yu, Y., Zhu, Y., Xu, W., Cai, X., Sun, Z., Han, W., et al. (2021). A genome-wide CRISPR screen identifies host factors that regulate SARS- CoV-2 entry. Nat. Commun. 12, 961. 9. Wei, J., Alfajaro, M.M., DeWeirdt, P.C., Hanna, R.E., Lu-Culligan, W.J., Cai, W.L., Strine, M.S., Zhang, S.M., Graziano, V.R., Schmitz, C.O., et al. (2021). Genome-wide CRISPR screens reveal host factors critical for SARS-CoV-2 infection. Cell 184, 76–91.e13. 10. Baggen, J., Persoons, L., Vanstreels, E., Jansen, S., Van Looveren, D., Boeckx, B., Geudens, V., De Man, J., Jochmans, D., Wauters, J., et al. (2021). Genome-wide CRISPR screening identifies TMEM106B as a proviral host factor for SARS-CoV-2. Nat. Genet. 53, 435–444. 11. Schneider, W.M., Luna, J.M., Hoffmann, H.H., Sa´ nchez-Rivera, F.J., Leal, A.A., Ashbrook, A.W., Le Pen, J., Ricardo-Lax, I., Michailidis, E., Peace, A., et al. (2021). Genome-scale identification of SARS-CoV-2 and pan- coronavirus host factor networks. Cell 184, 120–132.e14. 12. Mahon, C., Krogan, N.J., Craik, C.S., and Pick, E. (2014). Cullin E3 ligases and their rewiring by viral factors. Biomolecules 4, 897–930. 13. Scheffner, M., Huibregtse, J.M., Vierstra, R.D., and Howley, P.M. (1993). The HPV-16 E6 and E6-AP complex functions as a ubiquitin- protein ligase in the ubiquitination of p53. Cell 75, 495–505. 14. Shi, C.S., Qi, H.Y., Boularan, C., Huang, N.N., Abu-Asab, M., Shelhamer, J.H., and Kehrl, J.H. (2014). SARS-coronavirus open reading frame-9b suppresses innate immunity by targeting mitochondria and the MAVS/ TRAF3/TRAF6 signalosome. J. Immunol. 193, 3080–3089. 15. Mena, E.L., Donahue, C.J., Vaites, L.P., Li, J., Rona, G., O’Leary, C., Lignitto, L., Miwatani- Minter, B., Paulo, J.A., Dhabaria, A., et al. (2021). ORF10-Cullin-2-ZYG11B complex is not required for SARS-CoV-2 infection. Proc. Natl. Acad. Sci. USA 118. e2023157118. 16. Mullard, A. (2021). Targeted protein degraders crowd into the clinic. Nat. Rev. Drug Discov. 20, 247–250. 17. Jan, M., Sperling, A.S., and Ebert, B.L. (2021). Cancer therapies based on targeted protein degradation - lessons learned with lenalidomide. Nat. Rev. Clin. Oncol. 18, 401–417. 18. Yen, H.C.S., Xu, Q., Chou, D.M., Zhao, Z., and Elledge, S.J. (2008). Global protein stability profiling in mammalian cells. Science 322, 918–923. 19. Soucy, T.A., Smith, P.G., Milhollen, M.A., Berger, A.J., Gavin, J.M., Adhikari, S., Brownell, J.E., Burke, K.E., Cardin, D.P., Critchley, S., et al. (2009). An inhibitor of NEDD8-activating enzyme as a new approach to treat cancer. Nature 458, 732–736. 20. Słabicki, M., Kozicka, Z., Petzold, G., Li, Y.D., Manojkumar, M., Bunker, R.D., Donovan, ll OPEN ACCESS K.A., Sievers, Q.L., Koeppel, J., Suchyta, D., et al. (2020). The CDK inhibitor CR8 acts as a molecular glue degrader that depletes cyclin K. Nature 585, 293–297. 21. Słabicki, M., Yoon, H., Koeppel, J., Nitsch, L., Roy Burman, S.S., Di Genua, C., Donovan, K.A., Sperling, A.S., Hunkeler, M., Tsai, J.M., et al. (2020). Small-molecule-induced polymerization triggers degradation of BCL6. Nature 588, 164–168. 22. Chu, B.W., Kovary, K.M., Guillaume, J., Chen, L.C., Teruel, M.N., and Wandless, T.J. (2013). The E3 ubiquitin ligase UBE3C enhances proteasome processivity by ubiquitinating partially proteolyzed substrates. J. Biol. Chem. 288, 34575–34587. 23. El Khouri, E., Le Pavec, G., Toledano, M.B., and Delaunay-Moisan, A. (2013). RNF185 is a novel E3 ligase of endoplasmic reticulum- associated degradation (ERAD) that targets cystic fibrosis transmembrane conductance regulator (CFTR). J. Biol. Chem. 288, 31177– 31191. 24. van de Weijer, M.L., Krshnan, L., Liberatori, S., Guerrero, E.N., Robson-Tull, J., Hahn, L., Lebbink, R.J., Wiertz, E., Fischer, R., Ebner, D., and Carvalho, P. (2020). Quality control of ER membrane proteins by the RNF185/ membralin ubiquitin ligase complex. Mol. Cell 79, 768–781.e67. 25. Patten, J.J., Keiser, P.T., Morselli-Gysi, D., Menichetti, G., Mori, H., Donahue, C.J., Gan, X., Valle, I.D., Geoghegan-Barek, K., Anantpadma, M., et al. (2022). Identification of potent inhibitors of SARS-CoV-2 infection by combined pharmacological evaluation and cellular network prioritization. iScience 25, 104925. 26. Nieto-Torres, J.L., Dediego, M.L., Alvarez, E., Jime´ nez-Guarden˜ o, J.M., Regla-Nava, J.A., Llorente, M., Kremer, L., Shuo, S., and Enjuanes, L. (2011). Subcellular location and topology of severe acute respiratory syndrome coronavirus envelope protein. Virology 415, 69–82. 27. Martinez-Ortiz, W., and Zhou, M.M. (2020). Could PROTACs protect us from COVID-19? Drug Discov. Today 25, 1894–1896. 28. Simonetta, K.R., Taygerly, J., Boyle, K., Basham, S.E., Padovani, C., Lou, Y., Cummins, T.J., Yung, S.L., von Soly, S.K., Kayser, F., et al. (2019). Prospective discovery of small molecule enhancers of an E3 ligase-substrate interaction. Nat. Commun. 10, 1402. iScience 26, 106601, May 19, 2023 9 ll OPEN ACCESS STAR+METHODS KEY RESOURCES TABLE iScience Article REAGENT or RESOURCE SOURCE IDENTIFIER Antibodies Rabbit polyclonal GFP Rabbit RNF185 Polyclonal Antibody Mouse beta-actin Antibody Bacterial and virus strains Stbl3 SARS-CoV-2 WA (hCoV-19/USA-WA1/2020) SARS-CoV-2 Beta (B.1.351, hCoV-19/South Africa/KRISP-K005325/2020) SARS-CoV-2 Delta (B.1.617, hCoV-19/USA/MA-NEIDL-01399/2021) Chemicals, peptides, and recombinant proteins MG132 TAK (MLN7243) Pevonedistat (MLN4924) TRIzol LS Chloroform/Isoamyl Alcohol Isopropanol Ethanol Critical commercial assays Cell Navigator(cid:2) Live Cell Endoplasmic Reticulum (ER) Staining Kit *Blue Fluorescence* Luna Universal Probe One Step RT-qPCR Kit 2019-nCoV RUO Kit TransIT-L-LT1 Transfection Reagent Experimental models: Cell lines Human: HEK293T-cas9 Human: K562-cas9 Human: A549-cas9 Human: A549 African Green Monkey Kidney: VeroE6 Oligonucleotides sgRNA sequence: RNF185_A Forward: CACCGCATCTTACCTGATGTAAACA Reverse: AAACTGTTTACATCAGGTAAGATGC sgRNA sequence: RNF185_B Forward: CACCGCTGAGAACTCCAGTGCAGGG Reverse: AAACCCCTGCACTGGAGTTCTCAGC 10 iScience 26, 106601, May 19, 2023 Cell Signaling Technology ThermoFisher Scientific Cell Signaling ThermoFisher Scientific BEI BEI Cat#2555S Cat#PA5-78615 #3700 Cat#C737303 Cat#NR-52281 Cat#NR-55282 Dr. John H. Connor N/A Selleckchem Selleckchem Selleckchem Ambion Acros Organics Fisher Bioreagents Fisher Bioreagents Cat#S2619 Cat#S8341 Cat#S7109 10296028 327155000 BP2618-1 BP2818-500 AAT Bioquest Cat#22636 NEB IDT Mirus Broad Institute Broad Institute Broad Institute ATCC ATCC IDT IDT Cat#E3006 Cat#10006713 Cat # MIR 2300 https://www.atcc.org/products/crl-3216 https://www.atcc.org/products/ccl-243 https://www.atcc.org/products/ccl-185 Cat#CCL-185 Cat#CRL-1586 N/A N/A (Continued on next page) iScience Article Continued REAGENT or RESOURCE sgRNA sequence: RNF185_C Forward: CACCGGGAGACCAGACCTAACAGAC Reverse: AAACGTCTGTTAGGTCTGGTCTCCC sgRNA sequence: RNF185_D Forward: CACCGGTGGCCACACAGGCTGATGA Reverse: AAACTCATCAGCCTGTGTGGCCACC sgRNA sequence: SYVN1_A Forward: CACCGCCTCCAGAGTGAGAACCCCT Reverse: AAACAGGGGTTCTCACTCTGAGGC sgRNA sequence: SYVN1_B Forward: CACCGCTTGACTCACAAAGTCCACA Reverse: AAACTGTGGACTTTGTGAGTCAAGC sgRNA sequence: SYVN1_C Forward: CACCGGTATGGAAAATGTGGTTGCA Reverse: AAACTGCAACCACATTTTCCATACC sgRNA sequence: TMEM259_A Forward: CACCGACACGCAGGATGCCCTCACG Reverse: AAACCGTGAGGGCATCCTGCGTGTC sgRNA sequence: TMEM259_B Forward: CACCGCAAGCCGCCGAGTAGCACAG Reverse: AAACCTGTGCTACTCGGCGGCTTGC Recombinant DNA Cilantro 2 Bison sgRNA library lentiCas9-Blast lentiGuide-Puro Software and algorithms FlowJo(cid:3)10 ImageJ R R Studio ll OPEN ACCESS SOURCE IDT IDENTIFIER N/A IDT IDT IDT IDT IDT IDT N/A N/A N/A N/A N/A N/A Addgene Addgene Addgene Addgene Plasmid #74450 Cat# 169942 Cat# 52962 Cat# 52963 BD Biosciences N/A N/A https://imagej.nih.gov/ij/ https://www.r-project.org/ https://www.r-project.org/ https://www.rstudio.com/ products/rstudio/download/ https://www.rstudio.com/ products/rstudio/download/ iScience 26, 106601, May 19, 2023 11 iScience Article ll OPEN ACCESS RESOURCE AVAILABILITY Lead contact Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Mikołaj Słabicki (slabicki@broadinstitute.org). Material availability Plasmids and all unique reagents generated in this study are available from the lead contact with a completed material transfer agreement. Data and code availability d Additional Supplemental datasets/h6rn55h86x/1. Items are available from Mendeley Data at https://data.mendeley.com/ d All data reported in this paper will be shared by the lead contact upon request. d This paper does not report original code. d Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. EXPERIMENTAL MODEL AND SUBJECT DETAILS Cell lines HEK293T-Cas9 and Vero E6 cells were cultured in DMEM (Gibco) and A549-Cas9, K562-Cas9 cells were grown in RPMI (Gibco), both supplemented with 10% fetal bovine serum (FBS) (Invitrogen), glutamine (In- vitrogen), and penicillin-streptomycin (Invitrogen) at 37(cid:4)C and 5% CO2. Generation of stability constructs SARS-CoV-2 proteins (Table S1) were synthesized and cloned by Twist Bioscience into ‘Cilantro2’ vector (Addgene 74450, PGK target–eGFP–IRES–mCherry, puromycin resistance). Lentivirus production In a 96-well plate format, 11,000 HEK293T cells were seeded per well in 100 mL medium. The next day, 0.1 mL of TransIT-LT1 (Mirus, MIR2305) was added to 5 mL of OPTI-MEM (Invitrogen), incubated for 5 min at room temperature and combined with a mixture containing 28.9 ng of the SARS-CoV-2 - eGFP fusion protein sta- bility plasmid, 166.7 ng psPAX2 and 16.7 ng pVSV-G in 3.3 mL OPTI-MEM. The solution was incubated for 30 min at room temperature and 10 mL was added to HEK293T cells in a dropwise manner. The lentivirus containing medium was collected two days after transfection and stored at – 80 (cid:4)C. Lentiviral transduction Cells were transduced by spin infection. 30,000 cells per well in 100 mL of culture medium was transferred to a well of a 96-well plate. 20% (v/v) of virus was added, which typically results in infection of 10–20% of cells (MOI 0.1–0.23). The plates were centrifuged for 2 h (2,200 rpm, 37(cid:4)C). Two days post infection, cells were selected with puromycin at a concentration of 2 mg/mL. SARS-CoV-2 stability assay A total of 1.2 3 105 HEK293T-Cas9 cells stably expressing a SARS-CoV-2 stability construct were seeded in 96-well plate. 24 h post seeding, cells were either untreated, treated with 1 mM E1 inhibitor (MLN7243), treated with 10 mM 26S proteasome inhibitor treated (MG132), or treated with 1 mM neddylation inhibitor (MLN4924). 6 h post treatment, cells were detached using Trypsin (Gibco) and analyzed by flow cytometry on the BD LSRFortessa. Mean eGFP/mCherry signal was calculated using FlowJo and subsequently normalized to the DMSO treated sample of each reporter. All degradation assays in HEK293T were per- formed in triplicate, in A549 in four replicates, and in K562 in six replicates. Geometric means of eGFP and mCherry fluorescent signals for live and mCherry-positive cells were exported using flow cytometry analysis software (FlowJo, BD Biosciences). Ratios of eGFP to mCherry were normalized to the average of DMSO-treated controls. 12 iScience 26, 106601, May 19, 2023 iScience Article ll OPEN ACCESS BISON SARS-CoV-2 reporter screen in HEK293T cells The BISON CRISPR library targets 713 E1, E2, and E3 ubiquitin ligases, deubiquitinates, and control genes, contains 2,852 sgRNAs, and was cloned into the pXPR003 vector as previously described.20 The virus for the library was produced in a T-175 flask format, as described above in ‘lentivirus production’ with the following adjustments: 1.8 3 107 HEK293T cells in 25 mL complete DMEM medium, 244 mL TransIT-LT1, 5 mL OPTI- MEM, 32 mg of library, 40 mg psPAX2 and 4 mg pVSV-G in 1 mL OPTI-MEM. Ten per cent (v/v) of BISON CRISPR library was added (MOI (cid:1)0.5–0.7) to a 6-well plate of 2.2 3 106 HEK293T-Cas9 cells in 2 mL of cell culture medium per well, repeated for each stability construct. Cells were spin infected (2200 rpm, 2 h at 37(cid:4)C) and selected with 2 mg/mL puromycin 24 h post transduction. On the eighth day, cells were sorted on a Sony MA900 Multi-Application Cell Sorter using fluorescence-activated cell sorting (FACS). Four populations were collected (top 5%, top 5–10%, bottom 5–10% and bottom 5%) based on the eGFP/mCherry ratio. For each condition, at least 100 3 106 cells were subjected to sorting. Sorted popu- lations were collected by centrifugation (5000 rpm, 5 min) and their cell pellets were flash-frozen in dry ice. Sorted cell pellets were resuspended in 100 mL direct lysis buffer (1 mM CaCl2, 3 mM MgCl2, 1 mM EDTA, 1% Triton X-100, Tris pH 7.5) with freshly supplemented 0.2 mg/mL proteinase K. Lysates were incubated at 65(cid:4)C for 15 min then 95(cid:4)C for 10 min. A 25 mL volume of this mix was used for library amplifications in each sorted sample in a 50 mL reaction volume 0.04U Titanium Taq (Takara Bio 639210), 0.53 Titanium Taq buffer, 800 mM dNTP mix, 200 nM P5-SBS3 forward primer, 200 nM SBS12-pXPR003 reverse primer), 94(cid:4)C for 5 min, 15 cycles of [94(cid:4)C for 30 s, 58(cid:4)C for 15 s, 72(cid:4)C for 30 s], 72(cid:4)C for 2 min. Two microlitres of the first PCR reaction was used as the template for 15 cycles of the second PCR, in which Illumina adapters and barcodes were added (0.04U Titanium Taq (Takara Bio 639210), 13 Titanium Taq buffer, 800 mM dNTP mix, 200 nM SBS3-Stagger-pXPR003 forward primer, 200 nM P7-barcode-SBS12 reverse primer). An equal amount of all samples was pooled and subjected to preparative agarose electrophoresis followed by gel purification (Qiagen). Eluted DNA was further purified by NaOAc and isopropanol precipitation. Amplified sgRNAs were quantified using the Illumina NextSeq platform. Data analysis of CRISPR-Cas9 knockout screens Our data analysis pipeline consisted of the following steps: (1) Normalize each sample to the total number of reads. (2) For each guide, calculate the ratio of reads in the stable versus unstable sorted gate and rank the sgRNAs. (3) Sum the ranks for each guide across all replicates. (4) Determine the gene rank as the median rank of the four guides targeting it. (5) Calculate p values by simulating a distribution with guide RNAs that have randomly assigned ranks over 100 iterations. The R scripts for these steps were previously published.20 Generation of CRISPR-Cas9 knock-out cells sgRNAs targeting genes of interest were cloned into the sgRNA.EFS.tBFP vector using BsmBI digestion as previously described.20 Stability reporter cell lines were transduced in a 96-well plate as described above. Successful guide integration was confirmed by expression of BFP protein by flow cytometry using the BD LSRFortessa. Immunoblots Whole cell protein lysates were mixed with Laemmli (SDS-sample) buffer (reducing, 6X) (Boston BioProducts) and resolved on a polyacrylamide gel along with GFP pulldown lysates described above. The gel was trans- ferred to a membrane and immunoblotted for GFP and HA as previously described.21 Transient transfection 5 mg of iRFP720 tagged RNF185 plasmid was combined with 500 mL of OptiMEM. 15 mL of Mirus TransIT1 is added to the plasmid OptiMEM mix and incubated at room temperature for 30 min. This transfection mix was added to 1.6 3 106 cells in a 6-well plate. Fluorescence signal is observed 48 h post transfection using fluorescence microscopy. Fluorescence microscopy HEK293T cells expressing SARS-CoV-2 Envelope – eGFP were plated into a 96-well plate at 10,000 cells/ well and transiently transfected with RNF185-iRFP720 construct as described above. Two days following transfection, cells were stained with Cell Navigator Live Cell Endoplasmic Reticulum (ER) working solution iScience 26, 106601, May 19, 2023 13 ll OPEN ACCESS iScience Article (blue fluorescence) for 30 min at 37(cid:4)C. Cells were imaged on a confocal microscope with SARS-CoV-2 En- velope in the GFP channel, RNF185 in the far-red channel, and the ER in the DAPI channel. Virus cultivation Three strains of SARS-CoV-2 were used in this study: SARS-CoV-2 WA (2019-nCoV/USA-WA1/2020), B.1.351, designated variant Beta (hCoV-19/South Africa/KRISP-K005325/2020), and B.1.617, designated variant Delta (hCoV-19/USA/MA-NEIDL-01399/2021). WA and Beta were obtained from BEI, and Delta was obtained from Dr. John H. Connor at the NEIDL, Boston University. The viruses were grown by passaging twice on VeroE6 cells (African green monkey kidney cells, known to be naturally permissive to viral growth). VeroE6 cells were infected with an MOI of <0.01 and incubated for around 3 days, when cy- topathology was noted. Samples of each stock were sequenced and showed no evidence of contaminants. Each stock was aliquoted into small tubes and stored at (cid:3)80(cid:4)C until needed. All live virus work was per- formed in the BSL4 laboratory at the NEIDL, Boston University. Effect of RNF185 KO on SARS-CoV-2 replication The evening before the experiment, 7.5 3 104 non-targeting (NT) and RNF185 KO in A549-ACE2 cell line were plated into wells of a 24 well plate in DMEM with 10% fetal bovine serum. Cells were challenged with SARS-CoV-2 WA, Beta, or Delta at an MOI of 0.1 in duplicate. The cells were placed at 37(cid:4)C for 1 h, then the initial inoculum was removed, cells were washed with PBS, and fresh medium was added to the cells. After 3 days, 250 mL of the supernatant was harvested and inactivated in 750 mL of TRIzol LS to be removed from containment. Virus RNA was then extracted. In a fume hood, 200 mL of chloroform/isoamyl alcohol was added to each sample. Tubes were mixed, incubated at room temperature for 5 min, and spun at 11,600 xg for 15 min at 4(cid:4)C. The aqueous phases were transferred to fresh tubes containing 500 mL isopropanol. Samples were incubated for 10 min, then centrifuged as before. The isopropanol was decanted, and RNA pellets were washed in 1 mL of cold 75% ethanol. Samples were centrifuged again at 10,000 xg for 5 min at 4(cid:4)C. The ethanol was removed, RNA pellets were air dried, and then resuspended in 50 mL nuclease-free water. For PCR analysis, the Luna Universal Probe One Step RT-qPCR Kit was used. Each reaction consisted of the following: 10 mL 2X reaction mix (from kit), 1 mL enzyme (from kit), 1 mL primer/probe mix (see below), 1 mL sample, and 7 mL nuclease-free water (from kit). The PCR was performed with the following parameters: 55(cid:4)C/10min >95(cid:4)C/1 min > [95(cid:4)C/10 s > 60(cid:4)C/30 s]/40 cycles. The primer and probe sets were from 2019-nCoV RUO Kit and nCoV_N2 forward and reverse primers. The primer sequences were forward primer 2019-nCoV_N2-F 50-GACCCCAAAATCAGCGAAAT-30, reverse primer 2019-nCoV_N2-R 50-TCTGGTTACT GCCAGTTGAATCTG-30, and probe 2019-nCoV_N2-P 50-FAM-ACC CCG CAT TAC GTT TGG TGG ACC- BHQ1-3’. A standard curve was generated from a series of dilutions with known concentrations of genome copies and was used to equate Cq values to the concentration of genome equivalents in the supernatant. Each sample was run in duplicate and averaged for the analysis. 14 iScience 26, 106601, May 19, 2023
10.1038_s41467-022-29759-7
ARTICLE https://doi.org/10.1038/s41467-022-29759-7 OPEN Singlet and triplet to doublet energy transfer: improving organic light-emitting diodes with radicals 1,2,6, Alexander J. Gillett Feng Li William K. Myers 4, Richard H. Friend 2,6, Qinying Gu2, Junshuai Ding1, Zhangwu Chen1, Timothy J. H. Hele 2,5✉ 2✉ 3, & Emrys W. Evans ; , : ) ( 0 9 8 7 6 5 4 3 2 1 Organic light-emitting diodes (OLEDs) must be engineered to circumvent the efficiency limit imposed by the 3:1 ratio of triplet to singlet exciton formation following electron-hole capture. Here we show the spin nature of luminescent radicals such as TTM-3PCz allows direct energy harvesting from both singlet and triplet excitons through energy transfer, with sub- sequent rapid and efficient light emission from the doublet excitons. This is demonstrated with a model Thermally-Activated Delayed Fluorescence (TADF) organic semiconductor, 4CzIPN, where reverse intersystem crossing from triplets is characteristically slow (50% emission by 1 µs). The radical:TADF combination shows much faster emission via the doublet channel (80% emission by 100 ns) than the comparable TADF-only system, and sustains higher electroluminescent efficiency with increasing current density than a radical-only device. By unlocking energy transfer channels between singlet, triplet and doublet excitons, further technology opportunities are enabled for optoelectronics using organic radicals. 1 State Key Laboratory of Supramolecular Structure and Materials, College of Chemistry, Jilin University, Qianjin Avenue 2699, Changchun 130012, P. R. China. 2 Cavendish Laboratory, University of Cambridge, JJ Thomson Avenue, Cambridge CB3 0HE, UK. 3 Department of Chemistry, University College London, Christopher Ingold Building, London WC1H 0AJ, UK. 4 Centre for Advanced Electron Spin Resonance (CAESR), Department of Chemistry, University of Oxford, Inorganic Chemistry Laboratory, South Parks Road, Oxford OX1 3QR, UK. 5 Department of Chemistry, Swansea University, Singleton Park, Swansea SA2 8PP, UK. 6These authors contributed equally: Feng Li, Alexander J. Gillett. email: rhf10@cam.ac.uk; emrys.evans@swansea.ac.uk ✉ NATURE COMMUNICATIONS | (2022) 13:2744 | https://doi.org/10.1038/s41467-022-29759-7 | www.nature.com/naturecommunications 1 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29759-7 Spin management is an important consideration for organic light-emitting diode (OLED) efficiency in display and lighting technologies. For closed-shell molecules with singlet-spin-0 ground state, spin statistics with electrical exci- tation leads to the formation of 25% singlet (spin-0, S1) and 75% triplet (spin-1, T1) excitons1,2. In first-generation OLEDs, this results in maximum electroluminescence (EL) internal quantum efficiency (EQE) of 25% as singlet emission (fluores- þ hν) triplet emission cence, S1 þ hν) is spin-forbidden. In com- (phosphorescence, T1 mercial applications, triplet–triplet annihilation- and enhanced phosphorescence-based schemes have been used to obtain efficient luminescence from triplet states3–7. Other technologies under development include thermally activated delayed fluor- electron donor–acceptor escence molecular designs promote reduced exchange interaction and minimised S1-T1 energy gap for reverse intersystem is allowed whereas ↛ S0 (i.e., TADF)8–12, where ! S0 crossing (rISC, T1 electroluminescence mechanism is shown in Fig. 1a. ! S1) and delayed S1 emission. The TADF Another possibility to extract emission from the otherwise dark T1 state is to transfer its energy to another energy acceptor molecule, which then emits light. However, if the acceptor is a ground-state singlet, converting the donor triplet to an acceptor excited-state emissive singlet is spin-forbidden: (cid:3) (cid:1) D T1 (cid:1) (cid:3) þ A S0 (cid:1) (cid:3) ↛ D S0 (cid:1) (cid:3) þ A S1 ð1Þ where DðXÞ stands for the energy donor molecule in state X and AðXÞ for the energy acceptor, and !=↛ denotes spin-allowed/ forbidden. It is possible to convert the donor triplet to an acceptor triplet, but emission from this state is spin-forbidden (cid:1) (cid:3) ! D S0 (cid:1) (cid:3) ↛ D S0 (cid:1) (cid:3) þ A S0 (cid:1) (cid:3) þ A S0 (cid:1) þ A T1 (cid:1) D T1 þ hν (cid:3) (cid:3) ð2Þ TADF-only OLED Radical-only OLED (b) y g r e n E (e) (a) (d) y g r e n E rISC ISC y g r e n E Electrical excitation − + hv S1 T1 S0 TADF:radical energy transfer OLED Electrical excitation − + rISCC ISC 2{D(S1)A(D0)} 4{D(T1)A(D0)} FRET 2{D(T1)A(D0)} Dexter 2{D(S0)A(D1)} hv D = Energy donor (4CzIPN, non-radical) A = Energy acceptor (TTM-3PCz, radical) 2{D(S0)A(D0)} TADF:non-radical energy transfer OLED 'hyperfluorescence' (c) Electrical excitation − + Electrical excitation − + Q1 hv y g r e n E D1 D0 rISC ISC 1{D(S1)A(S0)} 3{D(T1)A(S0)} FRET Dexter hv 1{D(S0)A(S1)} 3{D(S0)A(T1)} non-radiative D = Energy donor (non-radical) A = Energy acceptor (non-radical) 1{D(S0)A(S0)} (f) TTM-3PCz Cl Cl Cl 4CzIPN NC N N N CN N Cl Cl Cl Cl N Cl ) 1 – m c 1 – M 4 0 1 ( i t n e c i f f e o c n o i t c n i t x E 4 3 2 1 0 400 4CzIPN TTM-3PCz 1.0 0.8 0.6 0.4 0.2 N o r m a l i s e d P L ( a r b . u n i t s ) 800 0.0 900 500 600 700 Wavelength (nm) Fig. 1 Light emission mechanisms and the radical energy transfer system. Electroluminescence mechanisms for TADF-only, radical-only and energy transfer OLEDs. Spin-allowed radiative transitions from excited to ground states are indicated by blue arrows labelled ‘hv.’ a Scheme for TADF OLED mechanism with emission from singlet S1 exciton, and singlet–triplet intersystem crossing (ISC) and reverse intersystem crossing (rISC) processes with non-emissive triplet T1 exciton. b Scheme for radical OLED mechanism with emission from doublet D1 exciton, formed by direct electrical excitation. Higher energy and non-emissive quartet Q1 exciton state are shown. c Scheme for TADF:non-radical energy transfer OLED mechanism. Electrical excitation generates singlet D(S1) and triplet D(T1) excitons, with FRET singlet-singlet energy transfer to non-radical energy acceptor (A) to form emissive singlet excitons, A(S1). Dexter triplet–triplet energy transfer forms non-emissive triplet excitons, A(T1); non-radiative decay to the ground state is shown by a wavy arrow. ISC and rISC steps between D(T1) and D(S1) are indicated. Spin multiplicity of D and A pairs are denoted by 2 S+1 in 2S+1{D A}. d Scheme for TADF:radical energy transfer OLED mechanism. Electrical excitation generates singlet D(S1) and triplet D(T1) excitons, with singlet–doublet FRET and triplet–doublet Dexter energy transfer to radical energy acceptor (A) to form emissive doublet excitons, A(D1). e Chemical structures for 4CzIPN and TTM-3PCz used to test the mechanism in (d). f Absorption (black) and normalised PL (red) profiles for 4CzIPN (dotted lines) and TTM-3PCz (solid lines). 2 NATURE COMMUNICATIONS | (2022) 13:2744 | https://doi.org/10.1038/s41467-022-29759-7 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29759-7 ARTICLE such that the process converts one dark state to another dark state. A donor singlet can transfer its energy to the acceptor singlet by Förster transfer: (cid:1) (cid:3) (cid:1) (cid:3) ! D S0 þ A S0 (cid:1) (cid:3) ! D S0 (cid:1) (cid:3) þ A S0 (cid:1) (cid:3) þ A S1 (cid:1) (cid:3) D S1 þ hν ð3Þ but since this converts one bright state to another bright state, it does not improve the device efficiency, though could improve other device characteristics such as colour purity. Singlet to singlet energy transfer has been achieved, in previous work13–17, where TADF materials have been used as sensitisers in Förster- type energy transfer of TADF S1 to non-radical fluorescent molecules in a ‘hyperfluorescence’ scheme as depicted in Fig. 1c. In these systems, energy transfer of TADF triplet excitons is indirect and proceeds following reverse intersystem crossing to the TADF S1. However, the undesirable triplet–triplet energy transfer to lower energy triplets on the ‘hyperfluorescent’ mole- cule, as mentioned above, as well as undesirable triplet- annihilation interactions, must therefore be suppressed. ! D0 In contrast to OLED technologies employing electronic excita- tions with paired electrons, efficient radical-based OLEDs offer an alternative route to overcoming the spin-statistics limit using doublet þ hν excitons with spin-allowed doublet emission (D1 fluorescence), since the dark quartet state Q1 lies above the D1 state in energy18–24 (note that Dx denotes doublet electronic states, and D denotes energy donor). The radical OLED photophysical mechan- ism is shown in Fig. 1b. However, despite demonstrating an excel- lent peak EQE at low injection current densities, the ‘roll-off’— decreasing efficiency with increasing current density—is severe in reported radical devices using single-dopant emissive layers where charge trapping directly forms doublet excitons20,25. The role of exciton-exciton and exciton-charge annihilation effects were ruled out by transient PL measurements on electrically-driven OLEDs, leading to the conclusion that the charge-trapping mechanism for EL must be improved to advance the performance of radical-based devices25. Here we consider if the desirable properties of radical emitters could be used to ‘brighten’ otherwise dark (or slowly emissive) triplet states where emission efficiency cannot easily be improved by using a ground-state singlet acceptor. In the SI section 1, we show how, using a ground-state radical acceptor, triplet energy transfer leading to an emissive excited-state doublet can be quantum mechanically spin-allowed by Dexter transfer: (cid:1) þ A D0 (cid:1) (cid:3) ! D S0 (cid:1) (cid:3) ! D S0 (cid:1) þ A D1 (cid:1) þ A D0 (cid:1) D T1 þ hν (cid:3) (cid:3) (cid:3) (cid:3) ð4Þ unlike the case of a ground-state singlet acceptor considered earlier. Energy transfer from an excited-state singlet to a doublet is also allowed via a Förster-type mechanism (cid:1) (cid:3) (cid:1) (cid:3) ! D S0 ! D S0 (cid:1) þ A D1 (cid:1) þ A D0 (cid:1) þ A D0 (cid:1) (cid:3) D S1 þ hν (cid:3) (cid:3) (cid:3) ð5Þ meaning that the radicals’ doublet-spin nature enables energy harvesting of all electronic excitations in standard organic semi- conductors. In addition, rapid EL emission can be enabled in radical energy transfer-based devices, which is desirable: to enhance EL efficiency in OLEDs by outcompeting non-radiative channels, and to avoid building up of high excitation densities at high drive currents that can cause efficiency roll-off. Previously, triplet to doublet energy transfer has been demonstrated in experiments using transient radical acceptors26, but to the best of our knowledge has not been demonstrated using a stable, emissive radical nor in an optoelectronic device. We have combined non-radical organic semiconductors as energy donors with radical emitters as energy acceptors to form light-emitting layers. In principle, the strategy we propose can work with a wide range of standard OLED semiconductors so long as their singlet and triplet states are higher in energy than the doublet exciton in the radical material. It is desirable to choose systems for which the spin-exchange energy is kept low, so that the singlet energy is kept low, and (as in the case of ‘hyperfluorescence’ mentioned earlier) we use here TADF materials that are engineered to reduce the exchange energy to thermally accessible values. A further advantage here is that TADF systems undergo intersystem crossing following photo- excitation, allowing us to follow singlet and triplet dynamics in transient all-optical measurements. Thus our energy donors and acceptors in double-dopant emissive layers were chosen to be the benchmark TADF material, 1,2,3,5-tetrakis(carbazole-9-yl)-4,6- tris(2,4,6-trichlorophenyl) dicyanobenzene methyl-3-substituted-9-phenyl-9H-carbazole (TTM-3PCz) radi- cal from our previous work20. Transient PL (trPL) and absorp- tion (TA) measurements were used to probe the singlet–doublet and triplet–doublet showing rapid energy transfer on picosecond and microsecond timescales from singlet respectively. Magneto- electroluminescence studies support the role of triplet–doublet energy transfer in radical-based OLEDs. The TADF:radical devices show improved device characteristics, with reduced turn- on voltage and roll-off in the EQE, as well as better device sta- bility than single-dopant radical structures. TADF:radical sys- tems extend the spin space of organic optoelectronics, where advantageous ‘hyperfluorescence’ can be retained, dark triplet states removed, and more direct triplet–doublet energy transfer used for efficient radical-based optoelectronics. energy transfer mechanisms, (4CzIPN)8, and triplet excitons, and Results and discussion Radical energy harvesting for doublet emission. Figure 1d shows an energy level diagram for radical-based OLEDs using double-dopant emissive layers containing non-radical organic components (D, energy donor) and radical emitters (A, energy acceptor). General design rules are formulated: singlet (S1) and triplet (T1) excitons of D can transfer energy to the doublet (D1) of A for efficient doublet emission where As energy donors acceptors, Þ > EðA; D1 Þ and EðD; T1 Þ where EðD; S1 1. The singlet and triplet energy levels of the donor are higher Þ > EðA; D1 Þ Þ are Þ is the than the D1 state of the acceptor, i.e., EðD; S1 and EðD; T1 the S1 and T1 exciton energies of D, and EðA; D1 radical A D1 exciton energy; 2. The donor-cation/acceptor-anion, D•+ A•− or donor-anion/ acceptor-cation, D•− A•+ states must be higher energy than the radical D1-exciton, i.e., E(D•+ A•−) > E(A, D1) and E(D•−A•+) > E(A, D1). and 4CzIPN (EðD; HOMOÞ = −5.8 eV; EðD; LUMOÞ = −3.4 eV)27 and TTM-3PCz (EðA; HOMOÞ = −5.8–6 eV; EðA; SOMO reductionÞ = −3.7 eV)20 were chosen, and their molecular structures are given in Fig. 1e. Singlet–doublet transfer (Fig. 1d, dotted arrow) by a dipolar fluorescence resonance energy transfer, FRET, mechanism results in conservation of doublet-spin multiplicity from 2S1 to 2S0. This was promoted by spectral overlap of TTM-3PCz A-absorption and D-fluorescence of 4CzIPN (Fig. 1f), a well-studied TADF emitter with a singlet–triplet exchange energy gap of <50 meV28,29. The small singlet–triplet energy gap also allows substantial spectral overlap of D-phosphorescence and A-absorption, which also leads to a resonant energy condition. This sets up conditions for triplet–doublet energy transfer by electron-exchange Dexter mechanism (Fig. 1d, dotted arrow) from long-lived (>microsecond) 4CzIPN triplet excitons, which can be harvested for light emission. NATURE COMMUNICATIONS | (2022) 13:2744 | https://doi.org/10.1038/s41467-022-29759-7 | www.nature.com/naturecommunications 3 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29759-7 The reverse process—doublet to triplet energy transfer—was previously demonstrated by us and others with TTM-carbazole and anthracene derivatives30. Triplet–doublet energy transfer to form 2S0 is spin-allowed by the 2T1 state, which is mixed with the 4T1 state because of the negligible doublet–quartet 2;4T1 energy difference (estimated to be ~10 µeV from the intermolecular approach with no bond formation where antiferromagnetic coupled doublet is the lowest energy state31, meaning they are effectively degenerate) and spin mixing terms such as the triplet zero-field splitting interaction32. The mixed 2;4T1 states allow unlocked triplet–doublet channels for direct energy transfer with organic radicals. The theoretical considerations for singlet–doublet and triplet–doublet energy transfer by FRET and Dexter mechan- isms are discussed further in Supplementary Information 1. the energy radical transfer studying Energy transfer photophysics with radical emitters. In order to understand the photophysics of combined TADF:radical materi- als we firstly studied films that were radical-only, TADF-only and TADF:radical blends. We used time-resolved optical spectroscopy measurements to probe energy transfer from 4CzIPN to TTM- 3PCz on pico- to microsecond timescales. The film composition for concept was 4CzIPN:TTM-3PCz:CBP (ratio = 0.25:0.03:0.72). Reference films were studied for TTM-3PCz radical only (TTM-3PCz:CBP, 0.03:0.97) and 4CzIPN TADF only (4CzIPN:CBP, 0.25:0.75). The composition is based on the starting point of our previous work on TTM-3PCz OLEDs20, which here allows us to test energy transfer mechanisms in proof-of-principle studies. 4CzIPN and TTM-3PCz were blended in CBP (4,4’bis(N-carbazolyl)-1,1’- biphenyl) to reduce the effects of exciton self-quenching33, and with higher doping of 4CzIPN than the radical to promote charge trapping at the TADF sites and subsequent energy transfer to TTM-3PCz for light emission. TrPL profiles for nano-to-microsecond time ranges (with 355 nm excitation, all fluences = 5 μJ/cm2) of 4CzIPN:TTM- 3PCz:CBP films are found to be superpositions of TTM-3PCz (~700 nm) and 4CzIPN (~530 nm) emission. PL timeslices (2.5 ns) are given in Fig. 2a for 4CzIPN:TTM-3PCz:CBP (red), 4CzIPN:CBP (black) and TTM-3PCz:CBP (blue). In Fig. 2b, normalised PL spectra with respect to radical emission (timeslices from 2.5 to 50 ns) show substantial quenching of 4CzIPN on nanosecond timescales. For OLED applications it is desirable to reduce the overall emission time to minimise exciton quenching mechanisms34, leading us to consider plots of the integrated PL fraction for total emission (Fig. 2c). From this, we observe in 4CzIPN:TTM-3PCz:CBP that 95% of all photons are emitted by 1 μs, and over 80% of emission occurring by 100 ns. This compares favourably to 4CzIPN:CBP where only ~50% of emission happens by 1 μs, such that the donor–acceptor blend shows faster emission than the 4CzIPN-only blend. We have performed TA studies of 4CzIPN:TTM-3PCz:CBP, TTM-3PCz:CBP and 4CzIPN:CBP films in order to elucidate the energy transfer processes from excited-state absorption kinetics. In Fig. 3a, ΔT/T spectral timeslices are presented for short-time TA of 4CzIPN:TTM-3PCz:CBP from 0.2–0.3 ps to 1000–1700 ps. Excitation at 400 nm allowed for the preferential formation of excitons on 4CzIPN, owing to its strong absorption in this region and significantly higher loading fraction. The initial TA spectrum of 4CzIPN:TTM-3PCz:CBP (0.2–0.3 ps) closely resembles that of 4CzIPN:CBP, where we have assigned the 4CzIPN ground-state bleach between 360–460 nm, the 4CzIPN stimulated emission overlaid on a photoinduced absorption (PIA) between 480 and 700 nm, and the primary 4CzIPN S1 PIA at 830 nm (see Supplementary Figs. 1 and 2 for TA of 4CzIPN:CBP films). By 10 ps, we observe new PIA bands that grow in for 4CzIPN:TTM- 3PCz:CBP at 620, 950 and 1650 nm. These features match with the TTM-3PCz D1 spectral profile obtained from studies of TTM- 3PCz:CBP films (Supplementary Figs. 3 and 4), showing energy transfer from TADF singlet to radical doublet. In Fig. 3b, the normalised ΔT/T kinetic profiles for 4CzIPN:TTM-3PCz:CBP in and 4CzIPN S1 TTM-3PCz D1 (800–830 nm, orange line) PIA regions are shown. We highlight an additional quenching of 4CzIPN in 4CzIPN:TTM-3PCz:CBP compared to 4CzIPN:CBP films (black line, Fig. 3b). The quenching of 4CzIPN S1 PIA and the growth of TTM-3PCz D1 PIA on picosecond timescales prior to nanosecond 4CzIPN intersystem crossing is attributed to Förster-type singlet–doublet energy transfer35. As the 4CzIPN S1 PIA lies in a region where there is reduced absorption by the TTM-3PCz D1, we can use the ΔT/T with and without the presence of TTM-3PCz to estimate a lower bound for the fraction of singlet–doublet energy transfer. By 1.7 ns, the 4CzIPN S1 PIA falls to approximately 45% and 60% of the initial signal with (orange) and without (black) TTM-3PCz present, respectively, suggesting that ≥15% of S1 from 4CzIPN have already undergone fluorescence resonance energy transfer (FRET) to TTM-3PCz in 4CzIPN:TTM-3PCz:CBP. With selective excitation of TTM-3PCz at 600 nm (below the 4CzIPN bandgap) (610–630 nm, red line) (a) ) s t i n u . b r a ( L P d e s i l a m r o N 1.0 0.8 0.6 0.4 0.2 0.0 4CzIPN:TTM-3PCz:CBP TTM-3PCz:CBP 4CzIPN:CBP (b) 4CzIPN:TTM-3PCz:CBP 2.5 ns 2.5 ns 5 ns 10 ns 20 ns 50 ns ) s t i n u . b r a ( L P d e s i l a m r o N 1.0 0.8 0.6 0.4 0.2 0.0 (c) n o i t c a r F L P d e t a r g e t n I 1.0 0.8 0.6 0.4 0.2 0.0 450 500 550 600 650 700 750 800 850 450 500 550 600 650 700 750 800 850 101 Wavelength (nm) Wavelength (nm) 4CzIPN (450 - 800 nm) TTM-3PCz (575 - 840 nm) 4CzIPN:TTM-3PCz (650 - 840 nm) 102 Time (ns) 103 104 Fig. 2 Transient photoluminescence studies of 4CzIPN and TTM-3PCz with 355 nm excitation. a PL timeslices at 2.5 ns for 4CzIPN:TTM-3PCz:CBP (ratio = 0.25:0.03:0.72, red line); 4CzIPN:CBP (0.25:0.75, black line); TTM-3PCz:CBP (0.03:0.97, blue line), showing emission from both TADF and radical in the combined film. b PL timeslices for 4CzIPN:TTM-3PCz:CBP at various times from 2.5 to 50 ns, showing the 4CzIPN emission decaying relative to the radical emission at longer times. c Integrated PL fraction time profiles from 2.5 ns to 25 µs for 4CzIPN:TTM-3PCz:CBP in 650–840 nm range (red line); 4CzIPN:CBP in 450–800 nm range (black line); and TTM-3PCz:CBP in 575–840 nm range (blue line), showing faster luminescence for the combined TADF:radical film than the TADF-only film. 4 NATURE COMMUNICATIONS | (2022) 13:2744 | https://doi.org/10.1038/s41467-022-29759-7 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29759-7 (a) 1.0×10-3 5.0×10-4 / T T ∆ 0.0 -5.0×10-4 -1.0×10-3 (d) / T T ∆ d e s i l a m r o N 1.0 0.8 0.6 0.4 0.2 4CzIPN:TTM-3PCz:CBP 0.2 - 0.3 ps 1 - 2 ps 10 - 20 ps 100 - 200 ps 1000 - 1700 ps 400 500 600 700 800 900 1400 1500 1600 Wavelength (nm) Probe range: 610-630 nm 4CzIPN:TTM-3PCz:CBP TTM-3PCz:CBP Bi-exponential Fit τ 1 = 18.8 ns τ 2 = 1.6 μs Mono-exponential Fit τ = 16.8 ns (b) / T T ∆ d e s i l a m r o N 1.0 0.8 0.6 0.4 0.2 0.0 0.1 (e) / T T ∆ d e s i l a m r o N 1.0 0.8 0.6 0.4 0.2 4CzIPN:TTM-3PCz:CBP (c) 0.0 -1.0×10-4 -2.0×10-4 / T T ∆ -3.0×10-4 -4.0×10-4 -5.0×10-4 -6.0×10-4 610 - 630 nm (TTM-3PCz D1 PIA) 800 - 830 nm (4CzIPN S1 PIA) 1500 - 1600 nm (TTM-3PCz D1 PIA) 4CzIPN S1 PIA without TTM-3PCz ARTICLE 4CzIPN:TTM-3PCz:CBP 1 - 2 ns 5 - 6 ns 10 - 15 ns 20 - 30 ns 50 - 60 ns 100 - 200 ns 1000 - 2000 ns 1 10 100 1000 500 600 700 800 900 1000 Time (ps) Wavelength (nm) Probe range: 800-830 nm 4CzIPN:TTM-3PCz:CBP 4CzIP:CBP Bi-exponential Fit τ prompt = 7.8 ns τ delayed = 1.0 μs Bi-exponential Fit τ prompt = 12.1 ns τ delayed = 2.5 μs (f) ) K 3 9 2 = T ( L P d e y a e D l / ) T ( L P d e y a e D l 1 0.8 0.6 0.4 0.2 50 4CzIPN:CBP 4CzIPN:TTM-3PCz:CBP 100 150 200 250 300 Temperature (K) 0.0 100 101 103 102 Time (ns) 104 105 0.0 100 101 103 102 Time (ns) 104 105 Fig. 3 Transient absorption and temperature dependence studies of 4CzIPN and TTM-3PCz. Picosecond to nanosecond (a) timeslices and (b) kinetic profiles from transient absorption studies of 4CzIPN:TTM-3PCz:CBP (ratio = 0.25:0.03:0.72). 400 nm excitation, fluence = 89.1 μJ/cm2. This shows the decay of the singlet PIA around 830 nm and the growth of the radical PIAs around 620 and 1650 nm. c Nanosecond to microsecond timeslices of the 4CzIPN:TTM-3PCz:CBP blend (0.25:0.03:0.72). 355 nm excitation, fluence = 17.0 μJ/cm2. Discontinuities in timeslice spectral profiles for (a) and (c) arise because multiple experiments are used to cover the studied wavelength probe regions. Transient absorption kinetic profiles for photoinduced absorption features of (d) TTM-3PCz (610–630 nm) and (e) 4CzIPN (800–830 nm). d TTM-3PCz excited-state kinetics are shown for 4CzIPN:TTM-3PCz:CBP (0.25:0.03:0.72, red squares); and TTM-3PCz:CBP (0.03:0.97, black circles). This shows delayed radical emission is active in 4CzIPN:TTM-3PCz:CBP (TADF:radical) from triplet–doublet energy transfer. e 4CzIPN excited-state kinetics are shown for 4CzIPN:TTM-3PCz:CBP (red squares); and 4CzIPN:CBP (0.25:0.75, black circles). This shows delayed radical emission in 4CzIPN:TTM-3PCz:CBP (TADF:radical) is more rapid than delayed emission in 4CzIPN:CBP (TADF only). Mono- and bi-exponential fits are indicated by solid lines in (d and e). f Ratio of integrated delayed PL contribution for 4CzIPN:CBP (black circles) and 4CzIPN:TTM-3PCz:CBP (red circles) at different temperatures. Three-point moving average and trends for these profiles are indicated by square and line plots, and show different temperature dependencies. in 4CzIPN:TTM-3PCz:CBP, the resulting TA profiles resemble TTM-3PCz:CBP, showing that the D1 exciton—once formed— does not interact with 4CzIPN by further energy or charge transfer processes (Supplementary Figs. 5 and 6). We have studied energy transfer for timescales beyond 1 ns with long-time TA measurements of 4CzIPN:TTM-3PCz:CBP films (excited at 355 nm). ΔT/T spectral timeslices (1–2 ns to 1000–2000 ns) in Fig. 3c display features at 620, 830 and 1600 nm, which can be attributed to the TTM-3PCz D1 PIA and 4CzIPN S1 PIA from radical-only (Supplementary Figs. 3 and 4) and TADF-only films (Supplementary Figs. 1 and 2). The kinetic decay profile of the TTM-3PCz PIA (600–630 nm) has an extended lifetime in 4CzIPN:TTM-3PCz:CBP films (red squares, Fig. 3d) compared to TTM-3PCz:CBP (black circles). The 4CzIPN:TTM-3PCz:CBP kinetic profile can be fitted to a bi- = 1.6 μs. exponential with time constants of τ The presence of a long-lived D1 state in 4CzIPN:TTM-3PCz:CBP, beyond the D1 excited-state lifetime measured from TTM- 3PCz:CBP (τ = 16.8 ns, Supplementary Fig. 4), suggests energy transfer from 4CzIPN triplet (T1) states. By comparing the kinetic traces of the PIA associated with 4CzIPN from 800 to 830 nm in 4CzIPN:CBP (black circles, Fig. 3e) and 4CzIPN:TTM-3PCz:CBP (red squares), we observed reductions in both the prompt and from 12.1 to 7.8 ns and 2.5 μs to 1.0 μs, delayed lifetimes, respectively, from the presence of TTM-3PCz. This provides further evidence for energy transfer from 4CzIPN T1 (delayed = 18.8 ns and τ 2 1 (cid:3) (cid:3) (cid:3) (cid:3) or transfer, (cid:1) D T1 (cid:1) þ A D0 (cid:1) þ A D1 kinetic), and additionally from 4CzIPN S1 (prompt kinetic), to form TTM-3PCz D1. (cid:1) (cid:3) ! D S0 Triplet–doublet energy transfer from 4CzIPN, a TADF molecule, can be attributed to a hyperfluorescent-type mechan- ism by breakout from S1-T1 ISC and rISC cycles36, (cid:1) (cid:3) ! D S1 (cid:1) ð6Þ þ A D0 i.e., 4CzIPN reverse intersystem crossing, then singlet–doublet triplet–doublet Förster direct Dexter-type mechanism37,38 as given in Eq. (4). Both mechanisms lead to reduced T1 lifetime. In order to distinguish the energy transfer mechanisms, we have performed temperature dependence studies (50–293 K) on trPL of 4CzIPN:CBP (Supplementary Fig. 10) and 4CzIPN:TTM-3PCz:CBP (Supplementary Fig. 11). In both films there is negligible temperature dependence on trPL up to 100 ns, which we define as the prompt emission; we classify light emission from 100 ns onwards as delayed-type. The ratio of integrated delayed emission at different temperatures (T) with respect to the integrated value at 293 K is shown in Fig. 3f (i.e., delayed PL(T)/delayed PL(T = 293 K)). The delayed PL ratio is reduced in 4CzIPN:CBP films compared to 4CzIPN:TTM- 3PCz:CBP, falling to 0.2 and 0.8 at 50 K, respectively. This supports a Dexter-type triplet–doublet energy transfer channel in 4CzIPN:TTM-3PCz:CBP, with lower activation energy than reverse intersystem in 4CzIPN:CBP for thermally activated delayed fluorescence. However, the signal:noise for delayed PL NATURE COMMUNICATIONS | (2022) 13:2744 | https://doi.org/10.1038/s41467-022-29759-7 | www.nature.com/naturecommunications 5 ARTICLE (a) ) V e ( y g r e n E –2 –3 –4 –5 –6 –7 LiF/ Al C P A T P B C ITO/ MoO3 40 nm 30 nm M P M Y P 3 B 60 nm 4CzIPN TTM-3PCz 4CzIPN:TTM-3PCz:CBP TTM-3PCz:CBP 4CzIPN:CBP (d) 25 ) % ( E Q E 20 15 10 5 (b) 103 ) 2 – 102 m c A m ( y t i s n e d t n e r r u C 101 100 10–1 10–2 10–3 10–4 (e) 1.2 L E d e z i l a m r o N 1 0.8 0.6 0.4 0.2 NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29759-7 4CzIPN:TTM-3PCz:CBP TTM-3PCz:CBP 4CzIPN:CBP 5 6 Voltage (V) 7 0 1 2 3 4 3V 3.5V 4V 4.5V 5V 5.5V 6V 7V 8V 9V 10V (c) ) 2 – m 1 – r s W ( e c n a d a R i 103 102 101 100 10–1 10–2 10–3 10–4 10–5 (f) ) % ( L E M 6 5 4 3 2 1 8 9 10 11 12 4CzIPN:TTM-3PCz:CBP TTM-3PCz:CBP 4CzIPN:CBP 0 1 2 3 4 5 6 Voltage (V) 7 8 9 10 11 12 4CzIPN:(TTM-3PCz):CBP (4CzIPN):TTM-3PCz:CBP (4CzIPN):CBP 0 10–3 10–2 10–1 100 101 102 Current density (mA cm–2) 0 400 500 600 700 800 900 Wavelength (nm) 0 –300 –200 0 –100 Magnetic field (mT) 100 200 300 Fig. 4 4CzIPN and TTM-3PCz organic light-emitting diodes. a Device architecture for OLEDs with varying emissive layer: 4CzIPN:TTM-3PCz:CBP, 4CzIPN:CBP; TTM-3PCz:CBP. b–d Current density–voltage (J–V), radiance–voltage, EQE–current density (from 10−3 mA/cm2) curves for OLEDs. e Normalised EL profiles for 4CzIPN:TTM-3PCz:CBP OLEDs with varying voltage, and 4CzIPN and TTM-3PCz emission contributions. f Magneto- electroluminescence (MEL) studies of TTM-3PCz (red squares) and 4CzIPN (red diamonds) emission in 4CzIPN:TTM-3PCz:CBP OLEDs; 4CzIPN emission in 4CzIPN:CBP (black triangles). OLED devices were biased at 8 V. MEL studies show different magnetic field dependencies for 4CzIPN and TTM-3PCz emission from 4CzIPN:TTM-3PCz:CBP devices, which supports Dexter triplet–doublet energy transfer and not the hyperfluorescence mechanism of 4CzIPN triplet exciton energy harvesting. ratio varies in 4CzIPN:TTM-3PCz:CBP with changing tempera- ture, restricting further quantitative analysis. From the film photophysical studies, we have demonstrated efficient singlet–doublet and triplet–doublet energy transfer in 4CzIPN:TTM-3PCz:CBP from picosecond to microsecond time- scales, which we have attributed to Förster and Dexter mechanisms that enable luminescent TADF:radical films with emission from radical D1. Radical OLEDs and magneto-electroluminescence studies. Following our demonstration of singlet–triplet–doublet energy transfer photophysics, we aimed to exploit these processes in more efficient radical-based OLED designs. We fabricated TADF:radical OLEDs using the device structure in Fig. 4a. B3PYMPM (4,6- bis(3,5-di(pyridine-3-yl)phenyl)-2-methylpyrimidine) and TAPC (1,1-bis[(di-4-tolylamino)phenyl]cyclohexane) were used as elec- tron transport and hole transport layers, respectively. The emissive layer (EML) was 4CzIPN:TTM-3PCz:CBP (0.25:0.03:0.72)—the same composition as the photophysics studies. Single-dopant OLEDs were also fabricated where EML was 4CzIPN:CBP (0.25:0.75) for TADF reference devices; and EML was TTM- 3PCz:CBP (0.03:0.97) for radical reference OLEDs. The current density–voltage (J–V), radiance–voltage and EQE plots for the 4CzIPN:TTM-3PCz:CBP (red squares), 4CzIPN:CBP (black triangles) and TTM-3PCz:CBP (blue circles) OLEDs are shown in Fig. 4b–d. We found that the turn-on voltages decrease from 2.9 V (TTM-3PCz:CBP device) to 2.3 V (4CzIPN:TTM- 3PCz:CBP) to 2.2 V (4CzIPN:CBP). Here, we define the turn-on voltage to be that corresponding to current density >0.1 µA/cm2, above the electrical noise level of the devices. The trend in turn- on voltage suggests that the inclusion of the TADF sensitiser leverages more energy-efficient doublet exciton formation in electroluminescence. However, the higher turn-on voltage for TADF:radical OLEDs compared to TADF, and different J–V profiles in Fig. 3b, imply that both CBP and 4CzIPN mediate some electrical excitation of TTM-3PCz in TADF:radical devices. If all doublet electroluminescence originated by energy transfer from TADF sensitisation as in Fig. 1d, the J–V curves and turn-on voltages would be identical for 4CzIPN:CBP and 4CzIPN:TTM- 3PCz:CBP OLEDs. are achievable We note there is a plateau in maximum radiance of ~1 W sr−1 m−2 from 5 V for TTM-3PCz:CBP devices in Fig. 4c; radiance values up to 10 W sr−1 m−2 in 4CzIPN:TTM-3PCz:CBP. At voltages higher than 5 V, there is an increasing component of 4CzIPN emission in the total EL of 4CzIPN:TTM-3PCz:CBP OLEDs. At 10 V the EL from the device contains 89% TTM-3PCz and 11% 4CzIPN contributions. The higher radiance at 10 V for 4CzIPN:TTM-3PCz:CBP (5.0 W sr−1 cm−2) compared to TTM-3PCz:CBP (1.1 W sr−1 cm−2) in Fig. 4c is therefore consistent with increasing energy transfer contribution from electrically excited 4CzIPN. The EL profile at the steady-state PL profile for 10 V in Fig. 4e resembles 4CzIPN:TTM-3PCz blends (Supplementary Fig. 8). Figure 4d shows that there is substantial increase in maximum EQE on going from 4CzIPN:CBP (7.8%) and TTM-3PCz:CBP (10.7%) devices to 4CzIPN:TTM-3PCz:CBP (16.4%) OLEDs. The EQE is evaluated for the total EL output. We note that the 25% wt. 4CzIPN:CBP reference device shown here has lower EQE than previous reports with 3% wt. 4CzIPN concentration due to exciton self-quenching effects8,33. The high 4CzIPN concentration is necessary to promote charge trapping at the TADF component in 4CzIPN:TTM-3PCz:CBP blends. Here the higher EQE on going from 4CzIPN:CBP to 4CzIPN:TTM-3PCz:CBP OLEDs suggests efficient energy transfer from 4CzIPN to TTM-3PCz, leading 6 NATURE COMMUNICATIONS | (2022) 13:2744 | https://doi.org/10.1038/s41467-022-29759-7 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29759-7 ARTICLE is not J0, limited by the EL efficiency of to performance that the 4CzIPN:CBP device. the critical current density that corresponds to the device current at half the maximum EQE, increases from 2.1 mA cm−2 for TTM-3PCz:CBP to 9.5 mA cm−2 for 4CzIPN:TTM-3PCz:CBP. The better roll-off and sustained EL efficiency in 4CzIPN:TTM-3PCz:CBP OLEDs is also attributed to an increasing contribution of 4CzIPN energy transfer to the EL at higher current densities. At lower voltages (<5 V) and current densities (<0.1 mA cm−2), the EL shows TTM-3PCz emission only (Fig. 4e). We performed studies to obtain the device’s half-lifetime, T50 (time for luminance to fall to half of the initial value under a constant current density). The T50 of energy transfer-type 4CzIPN:TTM-3PCz:CBP OLEDs was found to be 42 min at 0.4 mA/cm2 (see Supplementary Fig. 7), indicating some improve- ment over charge-trapping-type devices that we have previously reported for radical OLEDs with TTM-derivative:host EML (10 min at 0.1 mA/cm2)25. Magneto-electroluminescence (MEL) and magnetoconduc- tance (MC) studies have been performed on the 4CzIPN:CBP and 4CzIPN:TTM-3PCz:CBP devices. The devices were biased at 8 V and the data for magneto-EL and magnetoconductance were collected simultaneously. In 4CzIPN:CBP devices, MEL and MC profiles show enhanced EL and current density upon application of magnetic field (Fig. 4f and Supplementary Fig. 9). The profiles are fitted to double Lorentzian functions that capture low (<10 mT) and high (>10 mT) magnetic field effects (MFEs). The low field dependence is characteristic of magnetic field effects on hyperfine-mediate spin mixing of singlet and triplet polaron pair39, the precursors of excitons, which affect the ratio of singlet and triplet exciton formation. High field effects can arise from triplet exciton–polaron quenching and singlet–triplet dephasing effects40,41. MFEs of 4CzIPN:CBP devices are positive and show typical behaviour for MEL and MC from non-radical dopant systems, as previously reported42. that In TADF:radical OLEDs (4CzIPN:TTM-3PCz:CBP) we have studied magnetic field effects on EL from TTM-3PCz (680–800 nm) and 4CzIPN (500–550 nm) emission contributions. We observe positive magnetic field effects for both TTM-3PCz and 4CzIPN contributions, which indicates the main magnetic field sensitivity originates from hyperfine-mediated spin mixing of singlet–triplet polaron pairs, as found in the TADF-only devices. However the size of MEL for 4CzIPN (+4% at 250 mT) and TTM-3PCz emission components are different in TADF:radical OLEDs. We consider that non-identical MEL profiles for 4CzIPN and TTM-3PCz emission in 4CzIPN:TTM-3PCz:CBP devices supports a Dexter triplet–doublet energy transfer mechanism because an identical field sensitivity would be expected for the 4CzIPN and TTM- 3PCz MEL in TADF:radical hyperfluorescent-type devices. (+1% at 250 mT) We have demonstrated efficient energy transfer of 4CzIPN singlet and triplet excitons to obtain emissive doublet excitons of TTM-3PCz. In trPL studies we observed more rapid light emission in 4CzIPN:TTM-3PCz:CBP blends than 4CzIPN:CBP, as up to 95% and 50% of photons are emitted by 1 µs, respectively. TA measurements revealed singlet–doublet and triplet–doublet energy transfer on 10–100 ns and 100 ns–1 µs timescales, though the observed timescale of triplet transfer is limited by the time taken for intersystem crossing to take place on 4CzIPN and, as a spin-allowed process, may be faster than this. layer were OLEDs with 4CzIPN:TTM-3PCz:CBP emissive = 9.5 mA/cm2, demonstrated with max EQE = 16.4% and J0 EQE = 10.7%, which (max outperforms TTM-3PCz:CBP = 2.1 mA/cm2) for the same charge transport layer architec- J0 ture. With also an order of magnitude improvement in device stability, the energy transfer-type radical OLEDs therefore show a substantial improvement in device characteristics compared to previous reports of charge-trapping radical OLEDs. The MEL results allow us to rule out a fully hyperfluorescence-type (Eq. (6)) mechanism for EL, and support Dexter-type T1-D1 energy transfer pathways enabled by organic radicals, here TTM-3PCz. We highlight that Dexter triplet–triplet transfer from energy donor to acceptor is a loss route for light emission with non- radicals, and must be suppressed in energy transfer devices using non-radical fluorescent emitters, for example, hyperfluorescence- type devices15. However fluorescent radical (doublet) emitters can exploit the triplet–doublet energy transfer pathway for radical OLEDs as we have demonstrated here, without a lower-lying radical ‘triplet state’ that must be avoided for emission losses. In future work, our device concepts can be used in improved material combinations for more efficient energy transfer with reduced exciton quenching, and with increased radical lumines- cence for advancing the performance beyond this starting point. By unlocking new energy transfer channels, an optoelectronic design for improved radical-based light-emitting devices is enabled by their unpaired electron spin properties. Methods Materials. TTM-3PCz precursor was synthesised by Suzuki coupling of tris(2,4,6- trichlorophenyl)methane (HTTM) and 4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl- 3PCz20. In this procedure, TTM-3PCz radicals were generated from the precursor by treatment with potassium t-butoxide in tetrahydrofuran, followed by oxidation with p-chloranil. 4CzIPN, TAPC, B3PYMPM, CBP of sublimed grade and other OLED materials were obtained from Ossila, Xi’an Polymer Light and Lumtec. Photophysics. TrPL and TA studies were performed on home-built setups pow- ered by a Ti:sapphire amplifier (Spectra Physics Solstice Ace, 100 fs pulses at 800 nm, 7 W output at 1 kHz). TrPL profiles were recorded using an Andor spectrometer setup with electrically gated intensified CCD camera (Andor SR303i; Andor iStar). Sample excitation with 400 nm pump pulse was provided by frequency-doubled 800 nm pulse from Ti:sapphire amplifier in trPL and short-time (ps–ns) TA studies. Short-time TA studies with 600 nm excitation were achieved from the wavelength tuneable output of TOPAS optical parametric amplifier (Light Conversion), which was pumped by the 800 nm laser pulses from the Ti:sapphire amplifier. Long-time (ns–µs) TA studies were performed with 355 nm pump pulses from an Innolas Picolo 25. Probe pulses for TA were obtained from non-collinear optical parametric amplifier (NOPA) systems for the visible (500–780 nm), near- infrared (830–1000 nm) and infrared (1250–1650 nm) wavelength ranges. The NOPA probe pulses were divided into two identical beams by a 50/50 beamsplitter; this allowed for the use of a second reference beam for improved signal:noise. The probe pulse for the UV (350–500 nm) region was provided by a white light supercontinuum generated in a CaF2 crystal. The probe pulses were detected by Si (Hamamatsu S8381-1024Q) and InGaAs (Hamamatsu G11608-512DA) dual-line array with a custom-built board from Stresing Entwicklungsbüro. Device fabrication and characterisation. Organic semiconductor films and devices were fabricated by vacuum-deposition processing (<6 × 10‒7 torr) using an Angstrom Engineering EvoVac 700 system. Current density, voltage and electro- luminescence characteristics were measured using a Keithley 2400 sourcemeter, Keithley 2000 multimeter and calibrated silicon photodiode. The EL spectra were recorded by an Ocean Optics Flame spectrometer. Magneto-EL measurements were performed with Andor spectrometer (Shamrock 303i and iDus camera) for modulation of EL in presence of magnetic field applied by GMW 3470 electromagnet. Reporting summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article. Data availability The data generated in this study have been deposited in the figshare database under the accession code: https://doi.org/10.6084/m9.figshare.17026607.v1. Code availability Code used to analyse data in this manuscript are available from the corresponding author upon reasonable request. Received: 30 August 2021; Accepted: 2 March 2022; NATURE COMMUNICATIONS | (2022) 13:2744 | https://doi.org/10.1038/s41467-022-29759-7 | www.nature.com/naturecommunications 7 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29759-7 References 1. Tang, C. W. & VanSlyke, S. A. Organic electroluminescent diodes. Appl. Phys. Lett. 51, 913–915 (1987). Forrest, S. R. et al. Excitonic singlet-triplet ratio in a semiconducting organic thin film. Phys. Rev. 60, 14422–14428 (1999). 2. 30. Han, J. et al. Doublet–triplet energy transfer-dominated photon upconversion. J. Phys. Chem. Letts. 8, 5865–5870 (2017). 31. Kawai, A. & Shibuya, K. Charge-transfer controlled exchange interaction in radical-triplet encounter pairs as studied by FT-EPR spectroscopy. J. Phys. Chem. A 111, 4890–4901 (2007). 3. Kido, J. & Iizumi, Y. Fabrication of highly efficient organic electroluminescent 32. Ern, V. & Merrifield, R. E. Magnetic field effect on triplet exciton quenching in devices. Appl. Phys. Lett. 73, 2721–2723 (1998). 4. Di, D. et al. Efficient triplet exciton fusion in molecularly doped polymer light- 5. emitting diodes. Adv. Mater. 29, 1605987 (2017). Baldo, M. A. et al. Highly efficient phosphorescent emission from organic electroluminescent devices. Nature 395, 151–154 (1998). 6. Ma, Y., Houyu, Z., Shen, J. & Che, C. Electroluminescence from triplet metal- ligand charge-transfer excited state of transition metal complexes. Synth. Met. 94, 245–248 (1998). 7. Adachi, C., Baldo, M. A., Thompson, M. E. & Forrest, S. R. Nearly 100% internal phosphorescence efficiency in an organic light-emitting device. J. Appl. Phys. 90, 5048–5051 (2001). 8. Uoyama, H., Goushi, K., Shizu, K., Nomura, H. & Adachi, C. Highly efficient organic light-emitting diodes from delayed fluorescence. Nature 492, 234–240 (2012). Shizu, K. et al. Enhanced electroluminescence from a thermally activated delayed-fluorescence emitter by suppressing nonradiative decay. Phys. Rev. Appl. 3, 014001 (2015). 9. 10. Kim, D. H. et al. High-efficiency electroluminescence and amplified spontaneous emission from a thermally activated delayed fluorescent near- infrared emitter. Nat. Photonics 12, 98–104 (2018). 11. Kotadiya, N. B., Blom, P. W. M. & Wetzelaer, G. J. A. H. Efficient and stable single-layer organic light-emitting diodes based on thermally activated delayed fluorescence. Nat. Photonics 13, 765–769 (2019). 12. Hu, Y. et al. Novel efficient blue and bluish-green light-emitting polymers with delayed fluorescence. J. Mater. Chem. C. 6, 2690–2695 (2018). 13. Nakanotani, H. et al. High-efficiency organic light-emitting diodes with fluorescent emitters. Nat. Commun. 5, 1–7 (2014). 14. Chan, C. Y. et al. Stable pure-blue hyperfluorescence organic light-emitting diodes with high-efficiency and narrow emission. Nat. Photonics 15, 203–207 (2021). 15. Abroshan, H., Coropceanu, V. & Brédas, J. L. Hyperfluorescence-based emission in purely organic materials: suppression of energy-loss mechanisms via alignment of triplet excited states. ACS Mater. Lett. 2, 1412–1418 (2020). 16. Cui, L. S. et al. Fast spin-flip enables efficient and stable organic electroluminescence from charge-transfer states. Nat. Photonics 14, 636–642 (2020). 17. Zhang, D. et al. High-efficiency fluorescent organic light-emitting devices using sensitizing hosts with a small singlet-triplet exchange energy. Adv. Mater. 26, 5050–5055 (2014). 18. Peng, Q., Obolda, A., Zhang, M. & Li, F. Organic light-emitting diodes using a neutral π radical as emitter: the emission from a doublet. Angew. Chem. Int. Ed. 54, 7091–7095 (2015). 19. Neier, E. et al. Solution-processed organic light-emitting diodes with emission from a doublet exciton; using (2,4,6-trichlorophenyl)methyl as emitter. Org. Electron. 44, 126–131 (2017). 20. Ai, X. et al. Efficient radical-based light-emitting diodes with doublet emission. Nature 563, 536–540 (2018). 21. Cui, Z., Abdurahman, A., Ai, X. & Li, F. Stable luminescent radicals and radical-based LEDs with doublet emission. CCS Chem. 2, 1129–1145 (2020). 22. Hudson, J. M., Hele, T. J. H. & Evans, E. W. Efficient light-emitting diodes from organic radicals with doublet emission. J. Appl. Phys. 129, 180901 (2021). 23. Hattori, Y. et al. Luminescent mono-, di-, and triradicals: bridging polychlorinated triarylmethyl radicals by triarylamines and triarylboranes. Chem. - A Eur. J. 25, 15463–15471 (2019). 24. Hattori, Y., Kusamoto, T. & Nishihara, H. Luminescence, stability, and proton response of an open-shell (3,5-dichloro-4-pyridyl)bis(2,4,6-trichlorophenyl) methyl radical. Angew. Chem. Int. Ed. 126, 12039–12042 (2014). 25. Abdurahman, A. et al. Understanding the luminescent nature of organic radicals for efficient doublet emitters and pure-red light-emitting diodes. Nat. Mater. 19, 1224–1229 (2020). 26. Hiratsuka, H., Rajadurai, S., Das, P. K., Hug, G. L. & Fessenden, R. W. Evidence for non-radiative triplet-doublet energy transfer in the naphthalene- benzophenone system in tetrahydrofuran. Chem. Phys. Lett. 137, 255–260 (1987). 27. Nakanotani, H., Masui, K., Nishide, J., Shibata, T. & Adachi, C. Promising operational stability of high-efficiency organic light-emitting diodes based on thermally activated delayed fluorescence. Sci. Rep. 3, 1–6 (2013). 28. Niwa, A. et al. Temperature dependence of photoluminescence properties in a thermally activated delayed fluorescence emitter. Appl. Phys. Lett. 104, 213303 (2014). 29. Menke, S. M. & Holmes, R. J. Exciton transport in an organic semiconductor exhibiting thermally activated delayed fluorescence. J. Phys. Chem. C. 120, 8502–8508 (2016). organic crystals. Phys. Rev. Lett. 21, 609–611 (1968). 33. Kim, H. S., Park, S. R. & Suh, M. C. Concentration quenching behavior of thermally activated delayed fluorescence in a solid film. J. Phys. Chem. C. 121, 13986–13997 (2017). 34. Murawski, C., Leo, K. & Gather, M. C. Efficiency roll-off in organic light- emitting diodes. Adv. Mater. 25, 6801–6827 (2013). 35. Förster, T. 10th spiers memorial lecture. Transfer mechanisms of electronic excitation. in Discussions of the Faraday Society. 7–17 (The Royal Society of Chemistry, 1959). 36. Yurash, B. et al. Photoluminescence quenching probes spin conversion and exciton dynamics in thermally activated delayed fluorescence materials. Adv. Mater. 31, 1804490 (2019). 37. Dexter, D. L. A theory of sensitized luminescence in solids. J. Chem. Phys. 21, 836–850 (1953). 38. Sudha Devi, L. et al. Triplet energy transfer in conjugated polymers. I. Experimental investigation of a weakly disordered compound. Phys. Rev. B - Condens. Matter Mater. Phys. 78, 045210 (2008). 39. Nguyen, T. D. et al. Isotope effect in spin response of π-conjugated polymer films and devices. Nat. Mater. 9, 345–352 (2010). 40. Lawrence, J. E., Lewis, A. M., Manolopoulos, D. E. & Hore, P. J. Magnetoelectroluminescence in organic light-emitting diodes. J. Chem. Phys. 144, 21409 (2016). 41. Keevers, T. L., Baker, W. J. & McCamey, D. R. Theory of exciton-polaron complexes in pulsed electrically detected magnetic resonance. Phys. Rev. B - Condens. Matter Mater. Phys. 91, 205206 (2015). 42. Tanaka, M., Nagata, R., Nakanotani, H. & Adachi, C. Understanding degradation of organic light-emitting diodes from magnetic field effects. Commun. Mater. 1, 1–9 (2020). Acknowledgements J.D., Z.C. and F.L. are grateful for financial support from the National Natural Science Foundation of China (grant no. 51925303). E.W.E. is grateful to the Leverhulme Trust for an Early Career Fellowship; and the Royal Society for a University Research Fellowship (grant no. URF\R1\201300). TJHH thanks the Royal Society for a University Research Fellowship (grant no. URF\R1\201502). WKM and the Centre for Advanced Electron Spin Resonance is supported by EPSRC (EP/L011972/1). F.L. is an academic visitor at the Cavendish Laboratory, Cambridge, and is supported by the Talents Cultivation Pro- gramme (Jilin University, China). A.J.G. and RHF acknowledge support from the Simons Foundation (grant no. 601946) and the EPSRC (EP/M01083X/1 and EP/M005143/1). This project has received funding from the ERC under the European Union’s Horizon 2020 research and innovation programme (grant agreement no. 670405 and 101020167). Author contributions E.W.E. and F.L. fabricated thin films and OLED devices, which were characterised by photoluminescence, J–V-radiance measurements and magnetic field studies. AJG per- formed transient absorption measurements. A.J.G. and E.W.E. carried out the transient PL measurements. Q.G. conducted OLED time dependence studies. J.D. and Z.C. syn- thesised the radical materials. TJHH formulated theory on the photophysical mechan- isms. W.K.M. conducted spin physics studies. EWE, RHF and FL conceived the project and supervised the work. The results were analysed and the manuscript was written with input from all authors. Competing interests The authors declare no competing interests. Additional information Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41467-022-29759-7. Correspondence and requests for materials should be addressed to Richard H. Friend or Emrys W. Evans. Peer review information Nature Communications thanks Nadzeya Kukhta and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available. Reprints and permission information is available at http://www.nature.com/reprints Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. 8 NATURE COMMUNICATIONS | (2022) 13:2744 | https://doi.org/10.1038/s41467-022-29759-7 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-29759-7 ARTICLE Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/ licenses/by/4.0/. © The Author(s) 2022 NATURE COMMUNICATIONS | (2022) 13:2744 | https://doi.org/10.1038/s41467-022-29759-7 | www.nature.com/naturecommunications 9
10.3390_v14112464
Article A Modified PINN Approach for Identifiable Compartmental Models in Epidemiology with Application to COVID-19 Haoran Hu †, Connor M. Kennedy †, Panayotis G. Kevrekidis † and Hong-Kun Zhang *,† Department of Mathematics and Statistics, University of Massachusetts Amherst, Amherst, MA 01003, USA * Correspondence: hongkunz@umass.edu † These authors contributed equally to this work. Abstract: Many approaches using compartmental models have been used to study the COVID-19 pandemic, with machine learning methods applied to these models having particularly notable suc- cess. We consider the Susceptible–Infected–Confirmed–Recovered–Deceased (SICRD) compartmental model, with the goal of estimating the unknown infected compartment I, and several unknown parameters. We apply a variation of a “Physics Informed Neural Network” (PINN), which uses knowledge of the system to aid learning. First, we ensure estimation is possible by verifying the model’s identifiability. Then, we propose a wavelet transform to process data for the network training. Finally, our central result is a novel modification of the PINN’s loss function to reduce the number of simultaneously considered unknowns. We find that our modified network is capable of stable, efficient, and accurate estimation, while the unmodified network consistently yields incorrect values. The modified network is also shown to be efficient enough to be applied to a model with time-varying parameters. We present an application of our model results for ranking states by their estimated relative testing efficiency. Our findings suggest the effectiveness of our modified PINN network, especially in the case of multiple unknown variables. Keywords: network dynamics; COVID-19; PINNs; wavelets 1. Introduction On 31 December 2019, 27 cases of pneumonia were reported in Wuhan City, Hubei Province in China. The cause was identified on 7 January 2020, and was subsequently termed SARS-CoV-2 (the virus) and COVID-19 (the disease) by the World Health Orga- nization (WHO) [1]. The disease subsequently grew to the point that the WHO officially declared it a pandemic on 11 March 2020 [2]. As of 31 May 2022, a cumulative total of 526,558,033 cases and 6,287,117 deaths attributed to the disease had occurred [3]. The sheer scale of the disease’s development prompted research to combat it across the globe. A proper modeling of the disease has been crucial not only in evaluating the behavior of the virus, but also towards policy decisions made in response to it [4,5]. An extensive evalua- tion of social-distancing measures and other nonpharmaceutical interventions found that these approaches were effective. These results supported the adoption of these policies on numerous occasions throughout the pandemic [6–8]. More recently, the role of vaccination has also been considered in the relevant models [9]. One of the greatest issues in studying the disease has been the difficulty in estimating its spread. Even simply getting an accurate estimate of the current number of infected individuals has been extremely difficult [10]. Accurate counts of the infected population are crucial. They serve as a guiding tool on policy decisions from the distribution of testing resources, to the allocation of treatment materials, to the severity of lockdown procedures. Limited testing supplies in the early stages and the frequency of asymptomatic cases have caused major information gathering difficulties [11]. The accurate estimation of certain parameters used in modeling the disease, notably the base reproduction rate and case Citation: Hu, H.; Kennedy, C.M.; Kevrekidis, P.G.; Zhang, H.-K. A Modified PINN Approach for Identifiable Compartmental Models in Epidemiology with Application to COVID-19. Viruses 2022, 14, 2464. https://doi.org/10.3390/v14112464 Academic Editors: Thanasis Fokas, George Kastis and Hernan Garcia-Ruiz Received: 29 August 2022 Accepted: 4 November 2022 Published: 7 November 2022 Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affil- iations. Copyright: © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). Viruses 2022, 14, 2464. https://doi.org/10.3390/v14112464 https://www.mdpi.com/journal/viruses viruses Viruses 2022, 14, 2464 2 of 27 confirmation rate, are also useful in guiding policy. The parameters’ true values, however, are dependent on these unknown infected cases. There are some methods, however, to infer these unknown quantities from limited available data; see, e.g., [12] for a recent discussion. We focus here specifically on the usage of compartmental epidemiological models in conjunction with the usage of the widely applicable and highly successful technology of the so-called physics-informed neural networks (PINNs) (see [13] for a recent review thereof) to estimate these quantities. The classic SIR ODE model is, arguably, the most well-known compartmental model. It separates the population into susceptible, infected, and removed (recovered) groups, with the model’s origins tracing back to the seminal work of [14]. In the almost century that has followed there has been a wide variety of variations of the model to fit the particular features of different diseases [15]. In the case of COVID-19, several different modifications have been considered to more accurately describe its development. The SEAHIR model (involving susceptible, exposed, asymptomatic, hospitalized, infected, and recovered populations) incorporates the effect of social isolation measures [16]. Meanwhile, the SIRSi (susceptible, infected, recovered, sick) approach models temporary immunity that wanes over time [17]. There have been numerous other proposals, involving different numbers of components (and other features such as, e.g., age stratification [18] and spatial distribution [19]). More complex models do run the risk of their unknown quantities being difficult to estimate, or in some cases even impossible [20]. If we want to estimate these unknowns, we need a model that reflects the practical realities of COVID-19, the effects of isolation [6], and the prevalence of asymptomatic cases [10,11], while still allowing estimations. We adopt the usage of the SICRD model implemented in [21]. The susceptible, infected, and recovered groups are similar to the standard SIR setting discussed above. We now also consider two additional compartments, the confirmed and death cases, which we assume to be the only directly known variables. We adopt this model as it incorporates the noteworthy effects of testing and quarantining. It also crucially splits the infected population into unknown/known compartments, as well as includes data on fatalities. The latter is an extremely important piece of accessible information as we note in Section 2.3.2. Indeed without it, it would actually be impossible to estimate the number of infected individuals, or most of the important parameters from confirmed case counts alone. The relevance of the inclusion of the death data (as well as potentially and more recently, the data on hospitalizations) is a feature that some of the present authors have argued in various earlier works [12,18]. The SEAHIR model has similar advantages and much of the analysis below could be adapted to that model as well (an avenue that we do not pursue herein). We propose the usage of a sophisticated neural network approach to estimate the unknown cases and parameters, i.e., a methodology that has proved extremely successful in a variety of applications within scientific computing [13,22]. A key component in the usage of a neural network over traditional regression techniques is in their improved flexibility. Regression methods such as those in [23] may only model the data as exactly following a specified model. A neural network has many more adjustable parameters, allowing for a more flexible modeling of a given problem. Indeed, in this vein, there have been some quite promising recent results in the usage of neural networks in disease modeling. The work of [24] effectively forecasts COVID-19 spread using a combination of an ensemble neural network with a fuzzy logic system. The usage of a graph neural network leveraged human mobility information to improve forecasting of COVID-19 in Germany [25]. Meanwhile, the spread of influenza in the U.S. has been modeled using a recurrent neural network to better reflect the time-sequential nature of the viral spread [26]. The usage of a neural network is also quite natural for us, as our model is particularly well suited for an adaptation of the “Physics Informed Neural Network” (PINN) approach [27]. The principle of PINNs was developed for the estimation of unknown quantities in a system adhering to a physical law, generally a nonlinear PDE (or lattice dynamical equation [28]) that the system is assumed to obey [27]. The effectiveness of this approach has been substantial across a wide variety of disciplines, notably having excellent compati- Viruses 2022, 14, 2464 3 of 27 bility with other techniques such as the previously mentioned graph and recurrent neural networks [22]. We note there have been several other attempts to apply the PINN concept to studying COVID-19. The work of [29] considered the SIR model. They suitably modified the dynamical equations thereof through a nonlinear transformation, aiming to leverage neural networks to identify the (assumed-to-be) linear dependence of the infection rate and subsequently performed short-term predictions. This has some interesting parallels to our proposals and analysis below, although we seek to avoid some of the relevant as- sumptions on the reliability of the total case counts or of the linear time-dependence of the infection rate. Meanwhile, the work of [30] also attempted a similar inference of parameters to our work and considered the more nuanced susceptible–exposed–infected–recovered– susceptible model. This attempt, however, also considered all variables to be known. Yet another effort in this vein of leveraging data-driven techniques in epidemiological problems can be found in the interesting work of [31]. In this paper, the authors sought to discover the associated dynamic models directly from data in the case of childhood diseases such as chickenpox, rubella, and measles, with partial success. The novelty of our work comes from the substantial difficulty of having unknown vari- ables in our system and the modifications of the PINN technique we make to overcome this. The standard PINN directly uses the governing equations to define its loss term, with no further reformulation specific to a given model. We found that in our case, with several unknown variables, the standard PINN consistently converged to incorrect values for all parameters and variables. We address this by introducing a new approach to the training loss. We derive a new set of terms for our loss function in Section 2.5. The new loss terms are structured to reduce the number of unknowns that must be simultaneously considered in a single term. This avoids errors due to too many degrees of freedom in the network’s attempted estimation. The result is a network with rapid and accurate convergence. We achieve very promising results, even with a simple feed-forward network architecture. This suggests the relevance of further research into other architectures, such as recurrent net- works to improve performance, or an application to a graph neural network to incorporate human mobility [19,32]. The exact form of our new loss is specific to the SICRD model, but the principle of deriving new equations for the purpose of improved stability of the training is far broader. The concept could be applied to the training loss of any machine learning system, so long as there exist some governing equations for the considered data. Thus, any such analysis could see improvements in the network’s convergence, stability, and efficiency with appropriately derived loss terms from the governing equations. In sum- mary, we propose the usage of a neural network approach to avoid the more substantial restraints of traditional regression methodologies. We use the PINNs among the former to connect to a widely used tool. Finally, we provide an approach that leads to an im- provement of the training loss of the latter and whose principle can be generalized to other systems within this class (of relevance to epidemiological applications). We begin in Sections 2.1 and 2.2 by explicitly defining the SICRD model, as well as explaining the general concepts of the model and the sources of our data. In Section 2.3, we consider the question of whether the unknown variables and parameters may be estimated from the partially known information. We formally address this by showing our model is identifiable, meaning the unknown parameters and unknown variables can be uniquely determined by the known variables. We thus verify we are in a regime where the estimation goal is possible. It is also shown that several variations of our model are not identifiable, which justifies our particular choice over several considered alternatives. In Section 2.4, we address the issue of potential noise in our collected data with the use of a wavelet transform, which separates the signal into key features and the noisy component caused by smaller scale fluctuations. This filtering also smooths the data, which improves our ability to apply an ODE model to it. With the model defined and the data appropriately processed, Section 2.5 explicitly presents our novel loss function for the network along with a corresponding explanation for its usage. Viruses 2022, 14, 2464 4 of 27 We then demonstrate the neural network’s effectiveness by running estimations on the ideal case, with simulated data that explicitly obey our model in Section 3.1. We briefly review the results of denoising via the wavelet transform in Section 3.2. Then, in Section 3.3 we demonstrate the network’s ability to estimate the unknown infected population and several unknown parameters given real data reported in U.S. states. In Section 3.3.1, we also demonstrate that, due to the high efficiency in the network’s calculations, it may also be used to perform estimates using a model with time-varying parameters as opposed to simply assuming constant parameters, and no particular restrictions on the form of time-dependence are necessary, extending the work of other attempts to allow time-varying parameters; see, e.g., [21] or [29]. Our results suggest that when the parameters are assumed to vary, the variance is substantial enough that over any longer time frame, constant parameter assumptions would be unrealistic. In Section 3.3.2, we propose a ranking of states, according to how well the states are conducting testing by calculating the ratio of the (estimated) infected population over the confirmed population. The intention is to give a more precise estimate on how effectively testing is being conducted in a region, rather than simply looking at the per capita number of infections. This can highlight population centers where infections are low, as the disease has only entered it recently, but testing is low even relative to the number of infections. If a region with poor testing can be recognized early, then the problem can be addressed before it reaches a major outbreak. Finally, in Section 4, we briefly review our results and propose several possible directions of future research. We discuss some of the limitations of the model and network, as well as several methods to expand the work to address those limitations. 2. Materials and Methods 2.1. The SICRD Model In modeling the development of COVID-19, we focused on U.S. states as our pop- ulation centers due to the relative ease of access to epidemiological statistics. In each state, we used a modification of the SIR model [21] by introducing two new population compartments: the death cases and the confirmed cases [21]. This gives the SICRD model, a compartmental model in which the population is divided into susceptible (S), infectious (I), confirmed (C), recovered (R), and dead (D) individuals; see also Figure 1. Note that each compartment refers to the current count and not the total cumulative number of cases. Figure 1. Schematic of the modified-PINN model. We assume here that I and the parameters R0, α, and β are not known. Viruses 2022, 14, 2464 5 of 27 This model reflects the real-life situation in which an infectious person may recover without receiving a formal diagnosis, as well as accounting for the effect of testing for the disease and quarantining. The model is defined explicitly by the following ODE system: I (1) SI N R0 TL SI N − R0 TL α 1 TL TR β(1 − α) TL I − 1 TL C ˙S = − ˙I = ˙C = ˙D = ˙R = C I + β TR (1 − β)(1 − α) TL I + 1 − β TR C Here, N = S + I + C + D + R is the total population size. R0 is the basic reproductive number, which refers to the the average number of cases directly infected by one infectious case in a completely susceptible population. TL is the average number of days from first becoming infectious to confirmation, recovery, or death. TR is the average number of days from being confirmed as infected to recovery or death. α is the proportion of confirmed cases among all cases transferred from the unconfirmed infectious state. Finally, β is the fatality rate. The deaths (D) and recoveries (R) stem from either the infectious compartment (I) or from the confirmed infection one (C). We also note that we assume individuals who have been confirmed to be infectious sufficiently quarantine and do not infect any further individuals. Our goal is to estimate the unknown infected population I(t), along with the unknown parameters α, β, and R0. We wish to perform an estimation using only our accessible data, the count of confirmed cases C(t), deaths D(t), and the more directly estimable time scale parameters TL and TR. The reasons for this precise formulation of the SICRD model are discussed further in Section 2.3. We note that other, much more complex compartmental epidemiological models with more compartments such as those presented in [18] were also considered; here, inspired by the latter work, we present a reduced version of the model for reasons also relating to identifiability, as discussed below. Furthermore, this more straightforward model was chosen to test the capabilities of our neural network approach without added complications, while the testing of more detailed models is left to be considered in future work. SICRD with Time-Dependent Parameters The assumption that the parameters in Equation (1) are constant (i.e., time-independent) is a highly restrictive and unrealistic one for long-enough time scales. In a real-world scenario, many of these parameters are changing as time passes, due to factors such as mutations of the virus and a shifting government policy; the latter is well-known to modify social interactions in the case of nonpharmacological interventions and hence affect factors such as R0 [6,18]. In this case, we update the model in (1) to replace α, R0, and β with α(t), R0(t), and β(t), while TL and TR are left fixed as they are relatively well-known, stable parameters. In the case of mutations, it is not unreasonable to expect variations of these parameters too, but for the cases under consideration, we expect such variations to be small and anyway secondary to the above effects. This gives the alternate model: Viruses 2022, 14, 2464 6 of 27 (2) I I + β(t) TR C − SI N SI N 1 TL R0(t) TL R0(t) TL α(t) TL β(t)(1 − α(t)) TL 1 TR I − C ˙S = − ˙I = ˙C = ˙D = ˙R = (1 − β(t))(1 − α(t)) TL I + 1 − β((t) TR C To perform an estimation on the vector of parameters p(t) = (R0(t), α(t), β(t)), we first estimate a constant value for the parameters over a time interval of length ∆ t. We then perform the estimation of p(∆ t + t) with a simple “rolling window” approach, where we take the constant parameter estimation over the interval [t, t + ∆ t + t). This method allows an estimation of the parameter values which does not presume any particular form for their time dependence. This expands on the interesting, very recent work of [29]. The primary potential issue is the costliness of performing so many estimations, but we demonstrate in Section 3.3 that our network performs efficiently enough to make this scheme feasible. t] to approximate p(∆ 2.2. Data Set All nonsimulated data on COVID-19’s development used in this paper were pulled from the tool developed in [33]. The tool aggregates data from a variety of data sources including the WHO, each state’s individual department of health, and the CDC. The data used included reports on case counts, active cases, and deaths within each individual U.S. state. The usage of these data is expanded upon in Sections 3.2, 3.3.1 and 3.3.2. 2.3. Identifiability 2.3.1. Identifiability Definitions In order to reasonably use Equation (1), we need to verify the identifiability of the model. Identifiability is the ability to uniquely identify the model parameters from the known variables. We know from the results of [18,20,34] that many compartmental epi- demiological models have potential issues with identifiability. To be concrete, we recall the precise definition of the term. Consider an n-dimensional ODE system. We let p ∈ Ω p ⊂ Rnp be the vector of constant parameters for the ODE system where Ω p is our allowable parameter space. We note that the initial values of the variables in the system are also considered parameters. We let m(t, p) ∈ Rnm be all the assumed known (measured) variables of the system and h(t, p) ∈ Rnh be all the assumed unknown (hidden) variables of the ODE system, with nm + nh = n. Then, the ODE system can be represented by [20]: ˙m = f (m, h, p) ˙h = g(m, h, p). (3) Note that the same ODE system may be treated with different separations into m and h depending on what variables are assumed known. Given such a system and a choice of known variables, we say that it is structurally globally identifiable if m(t, p) = m(t, ˆp), ∀t ≥ 0 =⇒ ˆp = p. (4) The interpretation is that the measured variables uniquely determine the values of the constant parameters for the ODE system. In such a case it is not possible for two distinct Viruses 2022, 14, 2464 7 of 27 choices of parameters to give precisely the same values for the measured variables. Note though, from a practical perspective, that this does not disallow the possibility of p and ˆp with very different values, resulting in m(t, p) ≈ m(t, ˆp). Whether a narrow range for the parameter values can be determined from an uncertain value for m is the question of whether the system is practically identifiable or not. Basic global identifiability needs to be verified first though, as the practical identifiability of a system is dependent on the particular values chosen for the parameters and the range of the variables [20]. For an interesting example where the range of variables may play a crucial role in the practical identifiability, see, e.g., the very recent work of [35]. It is also possible for a system to fail to be globally identifiable, but still be “locally identifiable” meaning that no unique choice of parameters may be determined from m(t, p), but that there are finitely many choices of p that can be made. We obtain in Section 2.3.2 that no cases of local but not global identifiability for any parameters are found for our considered systems. For Equation (1), we always assume that C and D are known and thus m(t, p) = (C, D). I is, by definition, not directly known, while S and R both require knowledge of I to be known and thus h(t, p) = (I, S, R). (Reported data on recoveries are available but this information is only on recoveries from confirmed cases and thus is not actually the same quantity as R in our model.) We want to guarantee that our model at least satisfies structural global identifiability, meaning it is possible to estimate the parameters from C and D alone. 2.3.2. Numerical Results on Identifiability The precise calculation of structural identifiability results for a given ODE system is usually infeasible to perform by hand for all but the simplest of systems. There exist many software applications to perform this type of calculation, such as the differential algebra techniques of [34]. We elected to use the SIAN (Structural Identifiability ANalyser) package to test the identifiability of our model and several slight variations of the model [36]. The code runs a Monte Carlo algorithm to verify both local and global identifiability for an ODE system to within a high degree of certainty (>99%). The following models were tested using the code. The first model is the system of equations given in (1) without modifications. The sec- ond model matches (1) for ˙S, ˙I, and ˙C with the following equations for ˙D and ˙R ˙D = ˙R = C β TR (1 − α) TL I + 1 − β TR C (5) It is relevant to note that in (1), we allow deaths to occur from the unknown infected cases I, but that we assume the rate is the same as that for C, as well as allowing these deaths to be known. This is intended to reflect circumstances where a person is diagnosed posthumously, or extremely close to death. Meanwhile, in (5), we assume each fatality to occur in a case where the individual is first diagnosed. While (1) makes some stronger assumptions, it will soon be apparent that these or similar assumptions are necessary for identifiability. We thus need this for the parameters of the model to be capable of estimation from known data. The third case is a slight modification of (1), now allowing C and I to have two distinct death rates, referred to as β1 and β2, respectively. Explicitly, the changed equations are ˙D = ˙R = I + β2(1 − α) TL β1 TR (1 − β2)(1 − α) TL I + 1 − β1 TR C C (6) Viruses 2022, 14, 2464 8 of 27 Finally, we consider the case of (6) but with the death information recorded separately, assigning, D1 to C and D2 to I. Here, we consider both to be known as the case where only D1 is known was found to have only β1 identifiable and the case where only D2 is known is unreasonable, as confirmed-case-induced deaths are naturally expected to be known. ˙D1 = ˙D2 = ˙R = C β1 TR β2(1 − α) TL I (1 − β2)(1 − α) TL I + 1 − β1 TR C (7) Each of the models, with varying assumptions on the assumed known parameters, were run through the SIAN code. In each case there were no instances of parameters which were locally, but not globally identifiable. We also note that in cases where C(t) was the only known variable, α, β, and R0 were all not identifiable. In each case, R(0) was not identifiable, due to the fact that the current value for R did not affect the rate at which any compartment was changing. It can be estimated though, if each of the other variables’ initial value is estimated and the total population is known. We see from Table 1 that under even the worst cases, TL and TR may be estimated, which aligns with the common assumption that they can be more straightforwardly inferred from known incidence data. To estimate any other parameter requires either knowledge of the death rate for I, or for I and C to be assumed to have the same death rate. In the cases where the identifiability of parameters failed, not even local identifiability held, meaning that an infinite number of parameter combinations could give rise to the exact same solution for the known variables. Thus, in order to numerically estimate the parameters we moved forward with model (1), as it had the best conditions for parameter estimation, while still having reasonable underlying assumptions. A similar approach would also work for models with hospitalization data. There, hospitalizations take on a similar role to the deaths in providing indirect information on the unknown case numbers. Table 1. Comparing Model Identifiability. Model Known Variables Known Parameters Globally Identifiable Not Globally Identifiable (5) (5) (1) (6) (6) (6) (7) (7) (7) C(t), D(t) C(t), D(t) C(t), D(t) C(t), D(t) C(t), D(t) C(t), D(t) TL, TR, β none none TL, TR, β1 TL, TR, β2 [C(0), D(0)] [α, R0, S(0), I(0), R(0)] [TL, TR, β, C(0), D(0)] [α, R0, S(0), I(0), R(0)] [TL, TR, α, β, R0, S(0), I(0), C(0), D(0)] R(0) [C(0), D(0)] [I(0), S(0), α, β2, R0, R(0)] [α, β1, R0, S(0), I(0), C(0), D(0)] TL, TR, β1, β2 [α, R0, S(0), I(0), C(0), D(0)] C(t), D1(t), D2(t) C(t), D1(t), D2(t) TL, TR, β1 TL, TR, β2 [C(0), D1(0), D2(0)] [α, β2, R0, S(0), I(0), R(0)] [α, β1, R0, S(0), I(0), C(0), D1(0), D2(0)] C(t), D1(t), D2(t) TL, TR, β1, β2 [α, R0, S(0), I(0), C(0), D1(0), D2(0)] R(0) R(0) R(0) R(0) As an additional remark, the properties of model (5) are worse than those of (1) because for the former the time derivative of D essentially provides information for C (which is known) and hence does not assist towards identifying I. The latter identification Viruses 2022, 14, 2464 9 of 27 is, apparently, possible within (1). In the case of (6), β2 is needed to allow for identifiability (once again to allow for the detection of I), while β1 connected with the confirmed cases does not suffice. It also may not be reasonable to know D1 and D2 separately as generally, only total death counts are in the reported information. The combination of these factors, as well as the knowledge that even with both death counts, the identifiability only holds in this model when β2 is known, which is unreasonable if it is assumed distinct from β1, supports that model (6) is not as useful practically. While having two distinct values of β would be more realistic, we see that it creates fundamental issues with the identifiability of the model. We thus made the assumption of a single β value here, in order to use it as a test case for our network. There may be potential methods to resolve these issues while retaining identifiability, but in the present work our intention was to keep the assumptions of the model relatively simple while testing the novel loss function introduced in Section 2.5. The application of the network to a more complex model is left to future work. 2.4. Data Processing—Denoising of Data Using a Wavelet Transform Before feeding our data to the neural network we constructed, we first processed the data using a wavelet transform to separate the noise of the signal from its primary features. Our usage of this process was motivated by many other works where the denoising of an input signal for a neural network using a wavelet transform produced notable improvements [37–39]. We implemented code from the PyWavelets package to perform our wavelet analysis and denoising on our signal [40]. We considered the following framework for our data: in our model, each piece of measured signal data can be mathematically expressed as f(t) = ftrue(t) + ε, ∀t ∈ [0, T] where ftrue ∈ RT is the true data that would have been obtained in ideal measuring conditions and ε comprises the adverse effects of the local environment or faulty activity for the feature, and is referred to as “noise”. To avoid dealing with the noise in the data set, we applied a wavelet transform to denoise the data. (8) Wavelet theory provides a mathematical tool for hierarchically decomposing signals and, hence, constitutes an elegant technique for representing signals at multiple levels of detail [41]. In general, the wavelet transform is generated by the choice of a single “mother” wavelet ψ(t). In our implementation we used “Symlets 5” as our choice of mother wavelet (details can be found in the code documentation of [40]). Wavelets at different locations and spatial scales are formed by translating and scaling the mother wavelet. The translation and dilation of the mother wavelet are written with the operator U(u, v) acting as follows [42]: Uu,vψ(t) = e−u/2ψ(e−ut − v) Essentially, we can treat the “daughter” wavelets above, generated by rescaling and shifting the “mother” wavelet, as analogous to the sinusoidal functions of the Fourier transform. The advantage is that the rescaling and shifting allow the wavelets to capture more local behavior. This is performed by taking wavelets with small scales u, and shifting across the considered time range using varying values of v. A potentially very robust analysis of signals may be conducted in this way, though for our purposes the goal was to use the wavelets to decompose the signal into high and low frequency components, then discard the especially high frequency components as noise. Now, for a given signal f , the wavelet transform of f , Φψ f (u, v), at scale u and location v, is given by the inner product: Φψ f (u, v) = (cid:90) ∞ −∞ f (t)(Uu,vψ(t))∗ dt =< f , U(u, v)ψ > Viruses 2022, 14, 2464 10 of 27 Here, star represents the complex conjugate. This is known as the continuous wavelet transform or CWT. To ensure that the inverse CWT is well-defined, we need the follow- ing inequality Cψ := (cid:90) ∞ 0 | ˆψ(ξ)|2 ξ dξ < ∞ (9) Here, ˆψ(ξ) is simply the Fourier transform of ψ. This inequality is referred to as the admissibility condition [42]. One interpretation of this is that the choice of a mother wavelet must have no zero frequency component, i.e., no nonzero constant component. The finiteness of this integral guarantees that the result of the CWT always has a finite L2 norm. Generally, once this integral is verified to be finite, the mother wavelet is rescaled by it so that the inverse CWT may simply be defined as f (t) = (cid:90) ∞ (cid:90) ∞ −∞ −∞ Φψ f (u, v)U(u, v)(ψ(t)) dudv. This method of constructing the wavelet transform proceeds by producing the wavelets directly in the signal domain, through scaling and translation. When the signal frequency is higher, the wavelet base with a higher time domain resolution and a lower frequency resolution is used for the analysis. Conversely, when the signal frequency is lower, a lower time domain resolution and higher frequency resolution are used. This adaptive resolution analysis performance of the wavelet transform can effectively distinguish the local characteristics of the signal and the high-frequency noise, and accordingly perform the noise filtering of the signal. The general steps of the wavelet transform threshold filtering method are as follows: (1) Perform the multiscale wavelet decomposition on noisy time series signals; this process can be continued until the “noisy” or detailed component of the signal is sufficiently low in variance (see Section 3.2); (2) Determine a reasonable cutoff threshold and eliminate the high-frequency coefficients at each scale after decomposition; (3) Perform a wavelet signal reconstruction from the wavelet coefficients after the zeroing process to obtain the filtered denoised signal. The original signal f can be decomposed into f = A1 + D1 with A1 being a lower frequency approximation and D1 a high frequency signal. Essentially we are decomposing the original signal into a sum of wavelet terms of the form Uu,vψ(t), with gradually decreasing values of u and suitable values of v. After reaching a cutoff for u, we reassemble the function using the approximation via lower-frequency and larger-scale wavelets, giving A1, and the remaining portion of the function is represented as D1. The process may be repeated, treating A1 as the new “original” signal, to get the decomposition A1 = A2 + D2. The process may be repeated for several iterations. In addition to the removal of noise, filtering also smooths the data out as well. This smoothing does not substantially change the actual numerical values or trends, but it does help achieve a better fit for the model. As an ODE, the SICRD model assumes each of the variables is at least continuously differentiable, thus smoothing the inherently nondiffer- entiable accessible data helps the network perform an analysis on it. This decomposition allows us to obtain a denoised signal, while still retaining information about the exact nature of the filtered noise. With a method to appropriately process a noisy signal and prepare it as a training set, we can now define the network intended to analyze our data. 2.5. Setup of the Neural Network The ability of neural networks to act as universal approximators is well known [43] but the ability to perform accurate estimations in reasonable time frames is a central element of their increasing appeal. A variety of network structures have been developed to suit the features of particular problems. For example, recurrent neural networks, such as the long Viruses 2022, 14, 2464 11 of 27 short-term memory (LSTM) network, have been used to improve pattern recognition in data that have historical dependencies [44]. For our approach, we implemented a new “modified” “Physics Informed Neural Network” (PINN) to learn the values of the parameters and the unknown variable I(t) in our model 1. The PINN approach involves the changing of the loss function for the network, rather than any particular change in the network architecture itself, and in fact is compatible with a wide variety of possible architectures [13,22]. The concept of the PINN is to incorporate some physical law that must be obeyed by the system, and to introduce an extra term into the loss function which becomes smaller the closer the network output adheres to the law [27]. The approach was developed to estimate parameters and do forecasting in cases where data are only available at relatively sparse time intervals. The method is also robust to issues of overfitting, making it particularly useful for studies in limited-information regimes such as disease modeling [27]. While the method has been applied to more complex network architectures [22], we considered only the fully connected feed-forward network in our work as we tested our “proof of concept” modification. Initially, this concept was introduced in the context of nonlinear partial differential equations, though we see that our case of a nonlinear ordinary differential equation still falls into the applicable category (and other studies have also indeed used them in this context [28,45]). The common approach is to simply take the difference of the differential equations’ sides and introduce that as an extra term to minimize in the loss function. This is sufficient for many applications but we found it unable to generate acceptable results in our case. This was largely due to the entirely unknown variables S, I, and R in our model. We also only had a single time series from the system to use in estimating the parameters, whereas in many attempts to estimate system parameters using PINN methods, an entire ensemble of starting conditions and subsequent trajectories is used. For example, the DeepXDE deep learning method was used to attempt to estimate the parameters of a Lorenz system but needed to take a large collection of different starting trajectories to achieve good estimates [46]. Many methods of augmenting the PINN with other approaches exist [22], but generally the core structure of the loss function, the defining feature of the PINN, is not adjusted to improve training. Once a choice of modeling equation is made, it is directly used for the loss terms. The work of [29] is an appealing recent example with a reformulation of the original two-dimensional ODE system with an equivalent one-dimensional second-order ODE. Their goal was to formulate a more straightforward expression for the infection rate parameter, though the inherent effect of the approach on the network training was not considered. To address our case, we developed a new method, changing the structure of the loss function to that in Equation (14). The new terms still constituted the “physical law” for our system, but were of a form where the network’s gradient descent was far more stable. This novel approach, in our view, nontrivially extends PINN methods, and hence that is why we refer to it as a “modified” PINN system. This reveals the importance of tailoring the loss function not only to each system, but to the specific circumstances of accessible variables and parameters for a given problem. The tests on artificially generated data in Section 3.1 were extremely promising, with accurate estimations of the unknown parameters R0 and α. This was performed with only a single time series of C(t) and D(t) being used to perform the estimation. The network architecture was also the simple fully connected feed-forward network, yet it still performed well. This attests to the efficiency of this suitably chosen loss function. Before we precisely define our modified approach, we first review the standard PINN implementation. Generally, although many forms of PINN systems have been used, the loss function is basically of the same form [13,22]. To give a precise definition for the ODE case, consider an n-dimensional ODE system defined by ˙X = f (X, p), (10) Viruses 2022, 14, 2464 12 of 27 where f := f (X, p) is some function of all our variables X, and the given system parameters are p. The loss term Losspinn,1 may then be defined by Losspinn,1 := n ∑ i=1 MSEt( fi(Xest, pest) − ( ˙Xest)i). (11) where (Xest, pest) are the estimated values of the variables and parameters output from the neural network. MSEt is the mean square error up until time t. This loss essentially measures how well the output of the neural network is obeying the physical laws that govern the system. Note that the estimated values of the system parameters appear in f and thus may be learned using this loss term. The PINN loss is then combined with the standard loss function, with Xtest being the given test data: Lossmse,1 = n ∑ i=1 MSEt((Xest)i − (Xtest)i). Thus, we have the total loss: Loss1 = Losspinn,1 + Lossmse,1 (12) (13) Our initial testing essentially implemented the above scheme using the “Disease Informed Neural Network” code as a base [45]. Our version of Lossmse,1 could only include C and D, due to these being (assumed to be) the only known variables, but otherwise the loss function was the same. It was found that the network consistently led to incorrect values for all variables and parameters. The existence of unknown variables created several complications: the network had a fundamental gap in its training data, and the network was trying to estimate the unknown variables (S, I, R) in addition to estimating the unknown parameters. The presence of so much unknown information created substantial problems in the network’s ability to minimize loss for any particular term from fi, without creating significant losses in another element of the equations fi. The standard PINN loss terms also varied quite substantially in magnitude. The size of ˙D was much smaller than ˙C, and thus might impact the network’s learning less substantially, despite how crucial the death data were (as seen in our analysis in Section 2.3). Motivated by these deficiencies, we developed a modified interpretation of the PINN concept in the following way. If our variables for an ODE system are given by the vector X(t), we choose functions gi(Xest, p) (where we recall p is the vector of parameters for the system) which should each be 0 if the time series Xest precisely obeys the ODE system. Concretely, we consider the SICRD model and derive the following functions gi from Equation (1): g1 = (1 − α) ˙C − α ˙D/β + C/TR (14) g2 = αI TL − ˙C − C TR g3 = ˙I + ˙S + ˙C + ˙D/β g4 = R0 IS − NTL ˙S. Notice that these expressions involve suitable modifications of the dynamical equa- tions through (linear) combinations thereof. We can then use these functions to define the term: Losspinn = Then, the total loss function is given by i=4 ∑ i=1 MSE(gi). Losstotal = Lossmse + Losspinn, (15) (16) Viruses 2022, 14, 2464 13 of 27 where Lossmse = MSE(Cest − Ctest) + MSE(Dest − Dtest). (17) Some issues with the performance of the code were found when trying to use ratios or log scales to try to equalize the terms in the MSE loss, but ultimately the performance of the network worked well enough with the standard form. Alternative approaches with these methods of equalizing could be considered though. With these new loss functions, the network converged quickly and accurately when tested on artificially generated data (results in Section 3.1). Part of the motivation for this choice is that it excludes from consideration the re- covered population for which there are always unidentifiable features (such as R(0)) and focuses on the rest of the populations, including an effective rewrite of the conservation law associated with the total population. We also note that in our new equations, we tried to reduce how frequently multiple unknown quantities appeared together. To illustrate this point, g1 involves only known quantities except for α, if we assume β to be known. This allows this equation to be used to learn one parameter. If α is in principle known from g1, then all quantities except I are known in g2, and thus I can be well estimated. Then, all quantities except ˙S are in principle known in g3, letting ˙S and S themselves be estimated. Finally g4 can then estimate the basic reproduction number R0. Practically, this is not precisely what is happening, as each term is being minimized simultaneously by the network. This is just meant to illustrate in principle how this structure reduces the interference of multiple competing unknown quantities (a detail known to cause issues in the training of some networks [47]). We also note that all of the equations are of similar order, whereas in the standard PINN network, the quantity ˙D is substantially smaller than any other derivatives appearing in the system. In that light, the present restructuring of the equations systematically builds the optimization of the system parameters. This approach is in no way unique to our particular choice of system either. It shows that the concept of the PINN can be made broader than the standard choice of how to in- corporate information from the original set of differential equations a system is assumed to obey. Various different ODE or PDE systems could have alternatively derived formulations that can then be used to train a neural network. The most crucial point of our example is that it shows the existence of a case where the standard PINN approach is insufficient, but the modified approach is extremely effective. As far as the structure of the network itself is concerned, we note that it is a simple fully connected neural network, constructed with some minor modifications of the base code of [45]. The hidden dimension and learning rate initially presented in [45] created some issues, due to the original work assuming that all of the population compartments were known. This assumption allowed the original code to still converge reasonably well with a much lower learning rate, as well as a lower hidden dimension. In our implementation, we created 6 layers for our fully connected linear neural network, with hidden dimension 64. The nonlinear activation function in each layer was chosen to be the usual ReLu function, defined as: ReLu(x) = x, if x > 0, and ReLu(x) = 0 otherwise. The learning rate was taken to be lr = 0.0001, while the optimizer used was Adam [48]. 3. Results 3.1. Artificial Data Testing In order to test the validity of our estimation method, we first tested it on a set of artificially generated data. With artificial data, we can know the “hidden” variables and parameters while the network is only fed the corresponding information known in the real case. Then, the network’s resulting estimates can be compared with the true values to evaluate its performance, unlike with real data where this information is unknown. We assumed the real system corresponded at least roughly to the description of (8), behaving on average as the ODE system (1) prescribed, with some random noise. Viruses 2022, 14, 2464 14 of 27 We tested the neural network on artificial data with no noise, i.e., the ideal case, to verify its baseline capabilities. We generated the time series X(t), our vector of variables for the SICRD model, using a standard ODE solver. We then fed only the “known” variables C(t) and D(t) to our network with our new loss function (17), to generate an estimated time series and parameter values. We could then measure MSEt(Iest, Itest), as well as the difference between the estimated parameters and the true values. Small MSE values suggested that so long as our base assumptions about the system were reasonable, our method accurately estimated the infected population and unknown parameters. For our testing, we fixed the known values TL = 8.3, TR = 9.2, and β = 0.05. R0 = 2 and α = 0.8 were chosen for the unknown parameter values. We ran over a time interval of 40 days with 400 data points distributed evenly. The code was run for 50,000 epochs and executed in 3 h. We see in Figure 2 that the correct values for the unknown parameters were converged to quite rapidly and in a stable fashion. The oscillations in the loss function were not unexpected, as we noted that the oscillations were small due to the logarithmic scale. This was essentially just the process where once the loss was small enough, the network could no further diminish the loss without passing through regions where the loss increased temporarily while the Adam optimizer ran. (a) (b) Figure 2. (a) The learned parameters R0 = 2 and α = 0.8 in the artificial data testing. (b) The log-loss curve for the training process as a function of the number of Epochs. 2.2 2.0 1.8 1.6 1.4 1.2 1.0 0.8 0.6 �--------0 10,000 20,000 30,000 Learned parameter R0Learned parameter a40,000 50,000 010,00020,00030,00040,000Epochs1412108642Loss Viruses 2022, 14, 2464 15 of 27 Moreover, the total MSE loss along the time series we obtained was: S total loss : 0.09977; I total loss : 0.00069; D total loss : 0.00262; C total loss : 0.00093 Visually, we can see in Figure 3 how each learned time series (the dotted lines) over- lapped almost precisely with the true time series. Clearly, in this example, the modified PINN performed quite sufficiently in identifying both the parameters and the unavailable time series information. Figure 3. Artificial data time series compared with the network estimated values for each quantity; see the legend for each of the relevant comparisons. Note that the network only had access to the data for C and D. The true curves for I, R, and S were unknown to the network. We also present the results from the standard PINN, i.e., directly using the form from Equation (11) for the PINN loss. Testing was performed under the same conditions as for the modified PINN, with the results presented in Figures 4 and 5. We note with an initial value of α as 0.82, as we did in the modified PINN, α decreased directly to 0.7 during the training, as seen in the first graph. Moreover, the fitted values for C and D were quite poor, as seen in Figure 5 despite these being known quantities. Without modification, the PINN gave results which were not merely poor, or took substantial time to converge but the results also consistently converged to entirely incorrect values. Our modification was crucial towards recovering accurate estimates of I, α, and R0. 3.2. Wavelet Denoising Results With the testing of the network on artificial data concluded, the next step involved the application to real data sets. The given time series, C(t) and D(t), each had noticeable noise. We applied the wavelet decomposition defined in Section 2.4 before applying the network to the filtered signal in Section 3.3. We demonstrate several layers of wavelet decomposition from the original signal, illustrated in Figure 6. Note that the overall shape and scale of the signal was preserved in this smoothing process. Ai for higher values contained less of the precise information, but that information was preserved in the corresponding sequence of Di. A further analysis which incorporated the signal information for Di could have been performed, but we did not pursue that avenue. Viruses 2022, 14, 2464 16 of 27 Figure 4. The learned parameter α (with real value α = 0.8) using the standard PINN loss function. Figure 5. Estimated C and D using the standard PINN loss function. Figure 6. Wavelet decomposition of confirmed case data for Alabama starting 1 May 2020. The origi- nal time series and the subsequent five layers of decomposition into approximations (Ai) on the left and high-frequency signal (Di) on the right are shown. DWT: Frequency and phase change -Confirmed Cases Data 50,000 25,000-l::::=======�===::::::::::::=:::::::::::::::=:::==::========--------J' ' ' ' ' ' ' ' 50,000 25,000 50,000 25,000 50,000 25,000 75,000 50,000 25,000 75,000 50,000 25,000 0 20 40 60 80 100 1,000 0 -1,000 2,5000 -2,500 2,0000 1,000 0 -1,000 2,5000 0 20 40 60 80 100 Viruses 2022, 14, 2464 17 of 27 The A1 approximation was used for the training of our network, as this gave a rea- sonable balance between being smoothed out well enough for ease of training, while still retaining substantial details of the original sequence. 3.3. Real Data Testing After verifying the effectiveness of the network at testing on artificial data and pro- cessing via the wavelet transform, the next step was to perform testing on real data. Only C and D were known variables with the real data though, so it was not possible to evaluate the network based on its ability to estimate the unknown variables. It was still possible to at least measure how well the network was able to fit the available data while generating estimates for the unknown variables and parameters. The network was trained on reported data from Alabama for 90 days starting on 1 May 2020, with the resulting estimation the network produced for C, D, and I given in Figure 7. (a) (b) Figure 7. (a) The estimated value for infected cases, with corresponding fitted values for confirmed cases, compared with the actual confirmed count. (b) The fitted values for death cases, compared with the actual death count. In each case the mean absolute percentage error (MAPE) is included. Viruses2022,1,017of27Vl (I) Vl 60,000 50,000 40,000 Ctl 30,000 u 20,000 10,000 -confirmedConfirmed Estimation--Infected EstimationI Confirmed MAPE = 0.0068 I -Dead1,200 --Dead Estimation1,000 "' 800 (I) Ctl 600 u 400 200 Figure7.(a)Theestimatedvalueforinfectedcases,withcorrespondingfittedvaluesforconfirmedcases,comparedwiththeactualconfirmedcount.(b)Thefittedvaluesfordeathcases,comparedwiththeactualdeathcount.Ineachcasethemeanabsolutepercentageerror(MAPE)isincluded.3.3.RealDataTestingAfterverifyingtheeffectivenessofthenetworkattestingonartificialdataandpro-cessingviathewavelettransform,thenextstepwastoperformtestingonrealdata.OnlyCandDwereknownvariableswiththerealdatathough,soitwasnotpossibletoevaluatethenetworkbasedonitsabilitytoestimatetheunknownvariables.Itwasstillpossibletoatleastmeasurehowwellthenetworkwasabletofittheavailabledatawhilegeneratingestimatesfortheunknownvariablesandparameters.ThenetworkwastrainedonreporteddatafromAlabamafor90daysstartingon1May2020,withtheresultingestimationthenetworkproducedforC,D,andIgiveninFigure7.TheestimatedvaluesfortheassumedunknownparameterswereR0=1.8162,β=0.0098,andα=0.8762.ThevaluesTL=8.3andTR=9.2wereassumedbasedonmedicalrecordresults[21].WeseethatthenetworkwasabletogenerateveryclosefitstotheknownvariablesCandDoverthe90-daytimespan,despitetherelativelynoisydata.Wedidnottrainoverlongerintervalsoftimeastheassumptionofconstantparametervaluesbecamemoreunrealistic.Apotentialapproachusingtime-varyingparametersistestedinSection3.3.1,whereashere,testingwasconstrainedtoconstantparametervaluesinordertoperformourinitialverification.Viruses2022,1,017of27Vl (I) Vl 60,000 50,000 40,000 Ctl 30,000 u 20,000 10,000 -confirmedConfirmed Estimation--Infected EstimationI Confirmed MAPE = 0.0068 I -Dead1,200 --Dead Estimation1,000 "' 800 (I) Ctl 600 u 400 200 Figure7.(a)Theestimatedvalueforinfectedcases,withcorrespondingfittedvaluesforconfirmedcases,comparedwiththeactualconfirmedcount.(b)Thefittedvaluesfordeathcases,comparedwiththeactualdeathcount.Ineachcasethemeanabsolutepercentageerror(MAPE)isincluded.3.3.RealDataTestingAfterverifyingtheeffectivenessofthenetworkattestingonartificialdataandpro-cessingviathewavelettransform,thenextstepwastoperformtestingonrealdata.OnlyCandDwereknownvariableswiththerealdatathough,soitwasnotpossibletoevaluatethenetworkbasedonitsabilitytoestimatetheunknownvariables.Itwasstillpossibletoatleastmeasurehowwellthenetworkwasabletofittheavailabledatawhilegeneratingestimatesfortheunknownvariablesandparameters.ThenetworkwastrainedonreporteddatafromAlabamafor90daysstartingon1May2020,withtheresultingestimationthenetworkproducedforC,D,andIgiveninFigure7.TheestimatedvaluesfortheassumedunknownparameterswereR0=1.8162,β=0.0098,andα=0.8762.ThevaluesTL=8.3andTR=9.2wereassumedbasedonmedicalrecordresults[21].WeseethatthenetworkwasabletogenerateveryclosefitstotheknownvariablesCandDoverthe90-daytimespan,despitetherelativelynoisydata.Wedidnottrainoverlongerintervalsoftimeastheassumptionofconstantparametervaluesbecamemoreunrealistic.Apotentialapproachusingtime-varyingparametersistestedinSection3.3.1,whereashere,testingwasconstrainedtoconstantparametervaluesinordertoperformourinitialverification. Viruses 2022, 14, 2464 18 of 27 The estimated values for the assumed unknown parameters were R0 = 1.8162, β = 0.0098, and α = 0.8762. The values TL = 8.3 and TR = 9.2 were assumed based on medical record results [21]. We see that the network was able to generate very close fits to the known variables C and D over the 90-day time span, despite the relatively noisy data. We did not train over longer intervals of time as the assumption of constant parameter values became more unrealistic. A potential approach using time-varying parameters is tested in Section 3.3.1, whereas here, testing was constrained to constant parameter values in order to perform our initial verification. Thus, so long as the model had a reasonable fit to the assumption of Equation (8) with our SICRD model, the estimates for I(t) and the unknown parameters should be relatively accurate in their determination. Note that the general modification concept could easily be applied to other models as well, though new functions serving the role of the equations in (14) would have to be derived. 3.3.1. Time-Dependent Parameter Estimation Our next step was to perform an estimation using the time-dependent parameter model of Section 2.1. We found that our modified PINN was able to effectively and efficiently minimize loss over a scale of about 60–90 days. We present an example estimation of varying values for the parameters using available data in Alabama from 1 May 2020 to 31 December 2020, with a rolling window of 90 days, using the approach described in Section 2.1. We present our estimates of the time-dependent parameters in Figures 8 and 9. If the true values were constant, or close to constant, we would not expect to see such substantial variation in the graphs, so it seems that realistically the parameters should be assumed to vary for long-enough timescales. The value of R0 seems relatively stable, which may suggest that there were no substantial social distancing efforts implemented. Meanwhile, the variation of the confirmation rate α seems to indicate a difficulty in keeping testing up with infections at first, though some eventual adaptation takes place. The highly “noisy” look of α may be due to artifacts of how confirmed cases are reported. The most substantial variation though is the drop in β, which may reflect improvements in the treatment of severe cases of COVID-19. Care has to be taken in interpreting the values though, as uncertainties in the measured information and accuracy of the model are always present. The results are suggestive of the viability of the use of this network approach on systems with time-varying parameters. In the following section, we restrict ourselves to a 90-day time window with the assumption of fixed parameters to make estimations on all 50 U.S. states more feasible. Figure 8. Cont. Viruses 2022, 14, 2464 19 of 27 Figure 8. Learned values of the time-dependent parameters α and β as a function of time from 1 May 2020 for Alabama. Figure 9. Learned values of the time-dependent parameter R0 as a function of time from 1 May 2020 for Alabama. 3.3.2. Ranking States by Testing Now that we have a network that is able to, with reasonable effectiveness and efficiency, estimate these unknown quantities, we consider a concrete question this information can help us answer: “Which states have the worst relative rates of testing?”. This is an important piece of information for federal policymakers, as it determines which areas would be in the most dire need of additional federal aid in the form of (limited) testing supplies. The sense of “worst relative testing rate” can be quantified by finding which states have the highest ratio of I to C, averaged over a given time interval. Explicitly, we selected our corresponding diagnostic to be: A(t) = 1 t t ∑ k=1 I(k) C(k) (18) The importance of A(t) is that it lets us estimate which areas of the country are strug- gling the most to efficiently test their populations, relative to the prevalence of the disease. The idea of this particular metric is that it gives more specific information than just simply taking the estimation for the unknown infected population. Taking the infected population alone might just give information on locations that happen to have an exceptionally high number of infections occurring. A(t) more directly identifies areas (in our case: states) Viruses 2022, 14, 2464 20 of 27 where information is scarce relative to current levels of infection. This could be vital for identifying population centers that may not have a high number of infections currently, but are highly vulnerable to potential outbreaks due to poor testing infrastructure. The in- formation from our direct estimates for I(t) as well as for A(t) together give a more robust measure of the current testing situation. The importance of early intervention in the spreading of COVID-19 is already known [21], so the potential for this metric to provide an early warning is highly valuable. Even with limitations in estimating I(t), as long as the network is able to distinguish population centers of major need from ones of minor need, then it is still useful as a tool for policy- making. The confirmed case counts over 90 days starting from 1 May 2020, for the states with the top 3 and bottom 3 estimated A(t) values are given in Figure 10. The case counts over 6 months starting from same date are given in Figure 11. The values of A(t) over a 90-day interval starting on 1 May 2020, estimated by our network, are given in Tables 2 and 3, as well as the estimated values for the unknown parameters. We see that the estimated value for A(90) did not correlate in a direct and simple way with the values of the parameters. Thus, the metric did appear to provide extra information that could not be extrapolated simply from the parameter estimates alone. Figure 10. For 90 days starting from 1 May 2020, we plotted the confirmed cases of the bottom 3 estimated A(t) values, corresponding to states PA, NH, and OH, as well as the top 3 estimated A(t) values, corresponding to states ID, AK, and MO. We can see that ID, AK, and MO had a smaller number of cases in May 2020. Their confirmed case counts significantly grew after the middle of June though. Figure 11. Six months of confirmed cases for these 6 states. We see that the trend for ID, AK, and MO continues for months past the time period over which we estimated A(t). QJ-Q. 0 QJ � 60 0 r-f '-G) Q. � 40 tO u > ,a "CJ t 200 ,,., , I I ID AK ri10 FL ---USA Viruses 2022, 14, 2464 21 of 27 Table 2. States organized by ranking values estimated over 90 days starting at 1 May 2020. State Pennsylvania New Hampshire Ohio New York South Dakota Maine Indiana Massachusetts Nebraska New Jersey Connecticut Rhode Island Michigan Delaware A(90) 1.011 1.0296 1.0547 1.0549 1.0656 1.0722 1.0768 1.0771 1.0771 1.0848 1.0895 1.096 1.1105 1.1361 District of Columbia 1.1662 Iowa Wisconsin Minnesota Colorado Maryland North Dakota New Mexico North Carolina Wyoming West Virginia Virginia Illinois 1.1779 1.2021 1.2078 1.2115 1.232 1.2371 1.2607 1.2708 1.2844 1.286 1.3242 1.3434 R0 1.8139 1.8125 1.8093 1.7952 1.8015 1.8028 1.8107 1.8033 1.8159 1.7946 1.7966 1.7964 1.8059 1.8026 1.7992 1.8039 1.8038 1.8166 1.8068 1.8022 1.8085 1.8063 1.8063 1.8081 1.809 1.8127 1.8022 β 0.0375 0.0475 0.0089 0.0032 0.0237 0.0238 0.0212 0.0126 0.0069 0.0139 0.0102 0.0091 0.0188 0.014 0.0077 0.0172 0.0179 0.0417 0.0052 0.006 0.0183 0.0142 0.019 0.0123 0.0133 0.0051 0.014 α 0.8598 0.83 0.9099 0.8898 0.8037 0.8288 0.9268 0.9134 0.8671 0.8585 0.8884 0.9046 0.8416 0.8804 0.9101 0.8872 0.8872 0.841 0.8836 0.8888 0.8802 0.9054 0.9193 0.8875 0.8939 0.8745 0.8022 Viruses 2022, 14, 2464 22 of 27 Table 3. States organized by ranking values estimated over 90 days starting at 1 May 2020. State Utah Louisiana Vermont Kentucky Tennessee Mississippi Alabama Oklahoma Arkansas Kansas Oregon Washington Nevada Missouri Texas Georgia Arizona California A(90) 1.3906 1.3956 1.4032 1.4166 1.4332 1.4338 1.4714 1.4979 1.505 1.5718 1.6298 1.6314 1.6836 1.6889 1.7384 1.7721 1.7958 1.8618 South Carolina 1.8715 Hawaii Puerto Rico Florida Montana Alaska Idaho 2.0421 2.0655 2.1034 2.1703 2.7863 2.8028 R0 1.8144 1.8112 1.8264 1.8184 1.8156 1.814 1.8162 1.8119 1.8058 1.8142 1.8142 1.8081 1.8097 1.8028 1.8037 1.8125 1.8063 1.8618 1.8124 1.8235 1.8159 1.8139 1.8167 1.8165 1.8127 β 0.0072 0.0165 0.0024 0.0066 0.0106 0.0259 0.0098 0.0182 0.0181 0.0017 0.0054 0.0188 0.0043 0.0069 0.0201 0.0077 0.0078 0.0091 0.0118 0.0054 0.0139 0.0375 0.0132 0.0022 0.0036 α 0.8667 0.8968 0.7205 0.873 0.8761 0.9032 0.8762 0.9009 0.9048 0.7393 0.8221 0.8081 0.7946 0.9101 0.9028 0.83 0.8814 0.8898 0.8866 0.7464 0.8037 0.9134 0.8808 0.7136 0.7543 At the highest extreme, Idaho had a roughly 2.8 to 1 ratio for infected to confirmed. We note that while the time frame for this type of statistics was early in the pandemic, a state’s ranking may be suggestive towards subsequent effects. During the tested period, Idaho neglected to implement mask mandates in the summer of 2020 [49], later facing extreme overburden of its medical system by December 2020 [50], and further on having the lowest vaccination level later in September 2021 [51], features that appeared to be in line with our findings (although, of course, a further exploration of such case examples is useful to Viruses 2022, 14, 2464 23 of 27 consider in future efforts). Pennsylvania, on the other hand, had the lowest ranking, with a ratio of 1.011, just barely more cases of unknown infections estimated than confirmed cases. Pennsylvania had already instituted some mask mandates in April 2020 [52], and these mandates were substantially expanded in July 2020 [53]. Universities within the state also began declaring that classes would be moved online for the fall semester [54], with all of these features indicating a higher degree of preparedness to address the pandemic. This once again suggests the relevance of further consideration of this type of diagnostic. Overall, the existence of such a substantial disparity between states supports that this type of analysis may be interesting to consider and explore further. While we do not intend to reach any definitive conclusions in this connection since, realistically, there are many factors that may interfere with an accurate picture of this estimate, the consideration of the case examples that we have examined suggests that a further examination of such ideas and diagnostics would be worthwhile. 4. Discussion In the present work, we chose a modified version of the SIR model for the time evolu- tion of COVID-19, motivated by the features of the disease and the nature of available data. The model allowed the distinguishing of confirmed and unconfirmed cases, as well as a recording of the number of fatalities due to the disease. Upon verification of the identifia- bility of the unknown quantities within the model (based on the practically available data), we were able to use our modified PINN approach to analyze available data, with a special loss function tailored to the lack of available information. The most striking result was the substantial improvement that our modification made, as it was not obvious that a reformulation of the governing equations should have such an impact on the network’s learning. In principle the equations of (14) and (1) with N = S + I + C + D + R contain the same information. The exact nature of this dependence has likely gone unnoticed due to the basic implementation of the PINN being sufficient for most applications. This kind of circumstance though, performing inference when some variables are unknown, is not a rare one. The PINN was originally developed to be effective for inference with sparse data [27] but our work demonstrates that it extends to even incomplete data effectively, when appropriately modified. There are, of course, many further refinements that could be made to our methods. The SICRD model could always be further separated into more distinct population com- partments relevant to the spread of the disease. Compartments for exposed, though not yet infectious individuals, explicitly asymptomatic individuals, hospitalized individuals, and more could be included, and we discussed some principles on the basis of which such considerations could be explored. Additional topics that are quite relevant are the consideration of aspects of age stratification [15,18,55], as well as the incorporation of the role of vaccines and immunity waning (for later time frames than the ones considered herein, during which a vaccine was available) [9,56–58]. We could also consider cases where confirmed individuals may still infect others, albeit at a reduced rate. Our current framework, however, was intended to be used to study the development of the disease on relatively short time frames in the early pandemic stages, mitigating the effect of immunity loss as individuals would most likely still have immunity. Within this time frame, we could also disregard vaccinations, though, as stated earlier, the inclusion of vaccination effects could be incorporated if later times were considered. Instances of confirmed individuals spreading infection were assumed to be an exception as most individuals followed appropriate quarantining procedures on a positive test result. The crucial role of the death data (or hospitalization data in an SEAHIR model [16]) for estimation does raise the important issue of the validity of such data. The existence of underreporting, delays, and misattribution (namely, the important concern about “death from COVID” vs. “death with COVID” [59]) can create potential issues that warrant further examination. Nevertheless, we consider it to be the most relevant starting point for the study conducted herein and certainly a far more reliable one than the total case count, Viruses 2022, 14, 2464 24 of 27 as has been explained also earlier in connection to studies from the CDC; see [18] for a relevant discussion. Indeed, it is also relevant to point out that, as demonstrated in Section 2.3, the un- known infected population is impossible to estimate from recorded case counts alone. Hospitalization case information has also been incorporated into other methods of disease modeling with effective results so there is an existing precedent [16]. Such issues exist in models of waterborne illness, where the presence of difficult-to-track pathogens in water creates similar complications to our asymptomatic cases [34]. Further work can and should include further data to improve the model’s robustness. What the present work does is provide a mathematical basis and a computational implementation such that, even with a more complicated model, the PINN approach will still be reasonably applicable. Before our modification, it was not clear that a PINN could feasibly work even with our simpler model, in the presence of multiple unknown variables. Ultimately the simplifications made for the model, as well as for the architecture of the neural network, were taken so that the evaluation of the novel loss function could be performed with minimal complications. Future work could naturally focus on expanding both the model’s complexity as mentioned above as well as incorporating more nuanced network structures. The work of [25] on graph neural networks gives a very promising direction for future work in both directions, with a natural extension of the model to incorporate human mobility, paired with the natural choice of a graph neural network architecture. Indeed, the study of such metapopulation models of wide appeal within the modeling of COVID-19 [4,21], in conjunction with some of the technical approaches and methodologies presented herein, constitutes a promising direction for future study. The approach could also be suitably adapted to some form of recurrent neural network (RNN) as well. RNNs have already been used in disease modeling [26] and the combination of the LSTM, a specific form of RNN, with the PINN concept has been attempted in other works with notable success [22]. A limitation of our current work is the inability to estimate more of the unknown parameters simultaneously. The modified PINN still struggles to perform well if it must estimate parameters TL and TR in addition to R0, β, and α. There are simply too many simultaneous unknowns for the network to perform estimation efficiently. Indeed, this issue of numerous local minima that perform equally well has been encountered in various other studies; see, e.g., [18]. An RNN approach could be efficient enough to allow this broader parameter estimation. A particularly relevant direction, in terms of how widely it may expand our results, would be a rigorously proven criterion for alternate PINN loss terms. We gave an informed explanation of why the choice of Equation (14) led to better performance, but without a formalized proof. A natural future goal could be to formally prove that an alternate set of loss terms with fewer unknowns generally leads to faster or more accurate convergence. Some analysis on this concept of a network’s difficulty to balance multiple competing objectives does exist [47], so it would be reasonable to expect a form of generalized criteria to hold. If such a proof could be made, it would allow our approach to be applied extremely broadly with a general recipe for how to adapt it to a particular problem. It should also be added the very recently proposed idea of incorporating causality into the loss function of the PINNs [60] may also be quite relevant to consider in the present context. In addition, one could extend the study using graph neural networks [61]. Author Contributions: H.H.: Data curation, software,validation, C.M.K.: Methodology, writing— review and editing, formal analysis, conceptualization, investigation. P.G.K.: Conceptualization, methodology, formal analysis, writing—original draft, supervision. H.-K.Z.: Methodology, software, writing—original draft, project administration. All authors have read and agreed to the published version of the manuscript. Funding: HK Zhang is partially supported by Simons Foundation Collaboration Grants for Mathe- maticians (706383). PGK acknowledges support through the C3.ai Digital Transformation Institute. The authors are grateful to the three reviewers for their insightful comments that have helped improve the manuscript. Viruses 2022, 14, 2464 25 of 27 Institutional Review Board Statement: Not applicable. Informed Consent Statement: Not applicable. Data Availability Statement: All nonsimulated data on COVID-19’s development used in this paper were pulled from the tool developed in [33]. The tool aggregates data from a variety of data sources including the WHO, each state’s individual department of health, and the CDC. The data used included reports on case counts, active cases, and deaths within each individual U.S. state. We also include a direct link to the github for the tool here: https://github.com/CSSEGISandData/COVID-19 (accessed on 29 August 2022). Conflicts of Interest: The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results. References 1. Sohrabi, C.; Alsafi, Z.; O’Neill, N.; Khan, M.; Kerwan, A.; Al-Jabir, A.; Iosifidis, C.; Agha, R. World Health Organization declares global emergency: A review of the 2019 novel coronavirus (COVID-19). Int. J. Surg. 2020, 76, 71–76. [CrossRef] [PubMed] Cucinotta, D.; Vanelli, M. WHO Declares COVID-19 a Pandemic. Acta Biomed. Atenei Parm. 2020, 91, 157–160. [CrossRef] 2. 3. World Health Organization. WHO COVID-19 Dashboard. 2020. Available online: https://covid19.who.int (accessed on 1 April 4. 2021). Calvetti, D.; Hoover, A.P.; Rose, J.; Somersalo, E. Metapopulation Network Models for Understanding, Predicting, and Managing the Coronavirus Disease COVID-19. Front. Phys. 2020, 8, 261. [CrossRef] 5. Holmdahl, I.; Buckee, C. Wrong but Useful—What COVID-19 Epidemiologic Models Can and Cannot Tell Us. N. Engl. J. Med. 6. 2020, 383, 303–305. [CrossRef] [PubMed] Flaxman, S.; Mishra, S.; Gandy, A.; Unwin, H.J.T.; Coupland, H.; Mellan, T.A.; Zhu, H.; Berah, T.; Eaton, J.W.; Guzman, P.N.P.; et al. Estimating the number of infections and the impact of non-pharmaceutical interventions on COVID-19 in European countries: Technical description update. arXiv 2020, arXiv:2004.11342. [PubMed]. 8. 7. Wells, C.R.; Sah, P.; Moghadas, S.M.; Pandey, A.; Shoukat, A.; Wang, Y.; Wang, Z.; Meyers, L.A.; Singer, B.H.; Galvani, A.P. Impact of international travel and border control measures on the global spread of the novel 2019 coronavirus outbreak. Proc. Natl. Acad. Sci. USA 2020, 117, 7504–7509. [CrossRef] Chinazzi, M.; Davis, J.T.; Ajelli, M.; Gioannini, C.; Litvinova, M.; Merler, S.; Y Piontti, A.P.; Mu, K.; Rossi, L.; Sun, K.; et al. The effect of travel restrictions on the spread of the 2019 novel coronavirus (COVID-19) outbreak. Science 2020, 368, 395–400. [CrossRef] Usherwood, T.; LaJoie, Z.; Srivastava, V. A model and predictions for COVID-19 considering population behavior and vaccination. Sci. Rep. 2021, 11, 12051. [CrossRef] 9. 10. Peirlinck, M.; Linka, K.; Sahli Costabal, F.; Bhattacharya, J.; Bendavid, E.; Ioannidis, J.A.P.; Kuhl, E. Visualizing the invisible: The effect of asymptomatic transmission on the outbreak dynamics of COVID-19. Comput. Methods Appl. Mech. Eng. 2020, 372, 113410. [CrossRef] 11. Ma, Q.; Liu, J.; Liu, Q.; Kang, L.; Liu, R.; Jing, W.; Wu, Y.; Liu, M. Global Percentage of Asymptomatic SARS-CoV-2 Infections among the Tested Population and Individuals with Confirmed COVID-19 Diagnosis: A Systematic Review and Meta-analysis. JAMA Netw. Open 2021, 4, e2137257. [CrossRef] 12. Kevrekidis, G.A.; Rapti, Z.; Drossinos, Y.; Kevrekidis, P.G.; Barmann, M.A.; Chen, Q.Y.; Cuevas-Maraver, J. Backcasting COVID-19: A Physics-Informed Estimate for Early Case Incidence. arXiv 2022, arXiv:2202.00507. [PubMed]. 13. Karniadakis, G.E.; Kevrekidis, I.G.; Lu, L.; Perdikaris, P.; Wang, S.; Yang, L. Physics-informed machine learning. Nat. Rev. Phys. 2021, 3, 422–440. [CrossRef] 14. Kermack, W.O.; McKendrick, A.G. A Contribution to the Mathematical Theory of Epidemics. Proc. R. Soc. London Ser. A Contain. Pap. A Math. Phys. Character 1927, 115, 700–721. [CrossRef] 15. Hethcote, H.W. The Mathematics of Infectious Diseases. SIAM Rev. 2000, 42, 599–653. [CrossRef] 16. Leontitsis, A.; Senok, A.; Alsheikh-Ali, A.; Al Nasser, Y.; Loney, T.; Alshamsi, A. SEAHIR: A Specialized Compartmental Model for COVID-19. Int. J. Environ. Res. Public Health 2021, 18, 2667. [CrossRef] 17. Batistela, C.M.; Correa, D.P.; Bueno, Á.M.; Piqueira, J.R.C. SIRSi compartmental model for COVID-19 pandemic with immunity loss. Chaos Solitons Fractals 2021, 142, 110388. [CrossRef] 18. Cuevas-Maraver, J.; Kevrekidis, P.; Chen, Q.; Kevrekidis, G.; Villalobos-Daniel, V.; Rapti, Z.; Drossinos, Y. Lockdown measures and their impact on single- and two-age-structured epidemic model for the COVID-19 outbreak in Mexico. Math. Biosci. 2021, 336, 108590. [CrossRef] 19. Arenas, A.; Cota, W.; Gómez-Gardeñes, J.; Gómez, S.; Granell, C.; Matamalas, J.T.; Soriano-Paños, D.; Steinegger, B. Modeling the Spatiotemporal Epidemic Spreading of COVID-19 and the Impact of Mobility and Social Distancing Interventions. Phys. Rev. X 2020, 10, 041055. [CrossRef] 20. Gallo, L.; Frasca, M.; Latora, V.; Russo, G. Lack of practical identifiability may hamper reliable predictions in COVID-19 epidemic models. Sci. Adv. 2022, 8, eabg5234, [CrossRef] Viruses 2022, 14, 2464 26 of 27 21. Zhang, J.; Dong, L.; Zhang, Y.; Chen, X.; Yao, G.; Han, Z. Investigating time, strength, and duration of measures in controlling the spread of COVID-19 using a networked meta-population model. Nonlinear Dyn. 2020, 101, 1789–1800. [CrossRef] 22. Cuomo, S.; Cola, V.S.D.; Giampaolo, F.; Rozza, G.; Raissi, M.; Piccialli, F. Scientific Machine Learning through Physics-Informed Neural Networks: Where we are and What’s next. arXiv 2022, arXiv:2201.05624. [CrossRef] 23. Mahanty, C.; Kumar, R.; Mishra, B.; Hemanth, D.J.; Gupta, D.; Khanna, A. Prediction of COVID-19 active cases using exponential and non-linear growth models. Expert Syst. 2022, 39, e12648. [CrossRef] 24. Melin, P.; Sánchez, D.; Castro, J.R.; Castillo, O. Design of Type-3 Fuzzy Systems and Ensemble Neural Networks for COVID-19 25. Time Series Prediction Using a Firefly Algorithm. Axioms 2022, 11, 410. [CrossRef] Fritz, C.; Dorigatti, E.; Rügamer, D. Combining Graph Neural Networks and Spatio-temporal Disease Models to Predict COVID-19 Cases in Germany. arXiv 2021, arXiv:2101.00661. [CrossRef] 26. Aiken, E.L.; Nguyen, A.T.; Santillana, M. Towards the Use of Neural Networks for Influenza Prediction at Multiple Spatial Resolutions. arXiv 2019, arXiv:1911.02673. [CrossRef] 27. Raissi, M.; Perdikaris, P.; Karniadakis, G.E. Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations. arXiv 2017, arXiv:1711.10561. [CrossRef] 28. Zhu, W.; Khademi, W.; Charalampidis, E.G.; Kevrekidis, P.G. Neural networks enforcing physical symmetries in nonlinear dynamical lattices: The case example of the Ablowitz–Ladik model. Phys. D Nonlinear Phenom. 2022, 434, 133264. [CrossRef] 29. Barmparis, G.D.; Tsironis, G.P. Physics-informed machine learning for the COVID-19 pandemic: Adherence to social distancing 30. and short-term predictions for eight countries. Quant. Biol. 2022, 10, 139. [CrossRef] Schiassi, E.; De Florio, M.; D’Ambrosio, A.; Mortari, D.; Furfaro, R. Physics-Informed Neural Networks and Functional Interpolation for Data-Driven Parameters Discovery of Epidemiological Compartmental Models. Mathematics 2021, 9, 2069. [CrossRef] 31. Horrocks, J.; Bauch, C. Algorithmic discovery of dynamic models from infectious disease data. Sci. Rep. 2020, 10, 7061. [CrossRef] 32. Balcan, D.; Colizza, V.; Gonçalves, B.; Hu, H.; Ramasco, J.J.; Vespignani, A. Multiscale mobility networks and the spatial spreading of infectious diseases. Proc. Natl. Acad. Sci. USA 2009, 106, 21484–21489. [CrossRef] [PubMed] 33. Dong, E.; Du, H.; Gardner, L. An interactive web-based dashboard to track COVID-19 in real time. Lancet 2020, 20, 533–534. [CrossRef] 34. Eisenberg, M.C.; Robertson, S.L.; Tien, J.H. Identifiability and estimation of multiple transmission pathways in cholera and 35. waterborne disease. J. Theor. Biol. 2013, 324, 84–102. [CrossRef] [PubMed] Sauer, T.; Berry, T.; Ebeigbe, D.; Norton, M.M.; Whalen, A.J.; Schiff, S.J. Identifiability of Infection Model Parameters Early in an Epidemic. SIAM J. Control Optim. 2022, 60, S27–S48, [CrossRef] 36. Hong, H.; Ovchinnikov, A.; Pogudin, G.; Yap, C. SIAN: Software for structural identifiability analysis of ODE models. Bioinfor- matics 2019, 35, 2873–2874. [CrossRef] 37. Liu, K.; Cheng, J.; Yi, J. Copper price forecasted by hybrid neural network with Bayesian Optimization and wavelet transform. Resour. Policy 2022, 75, 102520. [CrossRef] 38. Boto Giralda, D.; Díaz-Pernas, F.; González-Ortega, D.; Díez, J.; Antón-Rodríguez, M.; Martínez Zarzuela, M.; De la Torre Díez, I. Wavelet-Based Denoising for Traffic Volume Time Series Forecasting with Self-Organizing Neural Networks. Comp.-Aided Civ. Infrastruct. Eng. 2010, 25, 530–545. [CrossRef] 39. Campisi-Pinto, S.; Adamowski, J.; Oron, G. Forecasting urban water demand via wavelet-denoising and neural network models. Case study: City of Syracuse, Italy. Water Resour. Manag. 2012, 26, 3539–3558. [CrossRef] 40. Lee, G.R.; Gommers, R.; Waselewski, F.; Wohlfahrt, K.; O’Leary, A. PyWavelets: A Python package for wavelet analysis. J. Open Source Softw. 2019, 4, 1237. [CrossRef] 41. Merry, R. Wavelet Theory and Applications: A Literature Study; DCT Rapporten; Eindhoven University of Technology: Eindhoven, The Netherlands, 2005; DCT 2005.053. 42. Heil, C.E.; Walnut, D.F. Continuous and Discrete Wavelet Transforms. SIAM Rev. 1989, 31, 628–666. [CrossRef] 43. Hornik, K.; Stinchcombe, M.; White, H. Multilayer feedforward networks are universal approximators. Neural Netw. 1989, 2, 359–366. [CrossRef] 44. Hochreiter, S.; Schmidhuber, J. Long Short-term Memory. Neural Comput. 1997, 9, 1735–1780. [CrossRef] [PubMed] 45. Shaier, S. Disease Informed Neural Networks. Available online: https://towardsdatascience.com/disease-informed-neural- networks-aa1f17f598a4 (accessed on 7 January 2022). 46. Lu, L.; Meng, X.; Mao, Z.; Karniadakis, G.E. DeepXDE: A deep learning library for solving differential equations. arXiv 2019, arXiv:1907.04502. [CrossRef] 47. Balduzzi, D.; Racaniere, S.; Martens, J.; Foerster, J.; Tuyls, K.; Graepel, T. The Mechanics of n-Player Differentiable Games. Machine Learning Research. In Proceedings of the 35th International Conference on Machine Learning, Stockholm, Sweden, 10–15 July 2018; Dy, J., Krause, A., Eds.; Volume 80, pp. 354–363. 48. Kingma, D.P.; Ba, J. Adam: A Method for Stochastic Optimization. arXiv 2014, arXiv:1412.6980. [CrossRef] 49. Shanks, A. With COVID-19 Cases Climbing, North Idaho Mayors Decline to Impose Mask Mandates. 2020. Available online: https://www.spokesman.com/stories/2020/jul/21/with-covid-19-cases-climbing-north-idaho-mayors-de/ (accessed on 1 April 2021). Viruses 2022, 14, 2464 27 of 27 50. Ridler, K. Overtaxed Idaho Health Facilities on Brink of Rationing Care. 2020. Available online: https://www.pbs.org/newshour/ health/overtaxed-idaho-health-facilities-on-brink-of-rationing-care (accessed on 1 April 2021). 51. Cohen, R. Idaho Department Of Health And Welfare Activates Crisis Standards In North Idaho. 2020. Available online: https://www.boisestatepublicradio.org/news/2021-09-07/idaho-activates-crisis-standards-in-north-idaho (accessed on 1 April 2021). 52. ABC. Employees, Customers to Be Required to Wear Masks Inside Essential Pennsylvania Businesses, Wolf Administration Orders. 2020. Available online: https://6abc.com/pennsylvania-schools-closed-major-disaster-fema-covid-19-pa/6105247/ (accessed on 1 April 2021). 53. Wolf, T. Secretary of Health Signs Expanded Mask-Wearing Order. 2020 Available online: https://www.governor.pa.gov/ newsroom/gov-wolf-sec-of-health-signs-expanded-mask-wearing-order/ (accessed on 1 April 2021). 54. ABC. Bloomsburg University Going Online Due to Spike in COVID-19 Cases. 2020. Available online: https://6abc.com/ bloomsburg-remote-learning-university-online-virtual-coronavirus-colleges/6392211/ (accessed on 1 April 2021). 55. Ram, V.; Schaposnik, L.P. A modified age-structured SIR model for COVID-19 type viruses. Sci. Rep. 2021, 11, 15194. [CrossRef] [PubMed] 56. Wagner, C.E.; Saad-Roy, C.M.; Grenfell, B.T. Modelling vaccination strategies for COVID-19. Nat. Rev. Immunol. 2022, 22, 139–141. [CrossRef] [PubMed] 57. Marinov, T.T.; Marinova, R.S. Adaptive SIR model with vaccination: Simultaneous identification of rates and functions illustrated with COVID-19. Sci. Rep. 2022, 12, 15688. [CrossRef] 58. Rychtar, J.; Diagne, M.L.; Rwezaura, H.; Tchoumi, S.Y.; Tchuenche, J.M. A Mathematical Model of COVID-19 with Vaccination and Treatment. Comput. Math. Methods Med. 2021, 2021, 1250129. 59. De-Giorgio, F.; Grassi, V.M.; Bergamin, E.; Cina, A.; Del Nonno, F.; Colombo, D.; Nardacci, R.; Falasca, L.; Conte, C.; d’Aloja, E.; et al. Dying “from” or “with” COVID-19 during the Pandemic: Medico-Legal Issues According to a Population Perspective. Int. J. Environ. Res. Public Health 2021, 18, 8851. [CrossRef] 60. Wang, S.; Sankaran, S.; Perdikaris, P. Respecting causality is all you need for training physics-informed neural networks. arXiv 2022, arXiv:2203.07404. [CrossRef] 61. Geng, R.; Gao, Y.; Zhang, H.-K; Zu, J. Analysis of the Spatio-Temporal Dynamics of COVID-19 in Massachusetts via Spectral Graph Wavelet Theory. IEEE Trans. Signal Inf. Process. Over Netw. 2022, 8, 670–683. [CrossRef]
10.1002_advs.202200181
RESEARCH ARTICLE www.advancedscience.com Moiré-Driven Topological Transitions and Extreme Anisotropy in Elastic Metasurfaces Simon Yves, Matheus Inguaggiato Nora Rosa, Yuning Guo, Mohit Gupta, Massimo Ruzzene,* and Andrea Alù* The twist angle between a pair of stacked 2D materials has been recently shown to control remarkable phenomena, including the emergence of flat-band superconductivity in twisted graphene bilayers, of higher-order topological phases in twisted moiré superlattices, and of topological polaritons in twisted hyperbolic metasurfaces. These discoveries, at the foundations of the emergent field of twistronics, have so far been mostly limited to explorations in atomically thin condensed matter and photonic systems, with limitations on the degree of control over geometry and twist angle, and inherent challenges in the fabrication of carefully engineered stacked multilayers. Here, this work extends twistronics to widely reconfigurable macroscopic elastic metasurfaces consisting of LEGO pillar resonators. This work demonstrates highly tailored anisotropy over a single-layer metasurface driven by variations in the twist angle between a pair of interleaved spatially modulated pillar lattices. The resulting quasi-periodic moiré patterns support topological transitions in the isofrequency contours, leading to strong tunability of highly directional waves. The findings illustrate how the rich phenomena enabled by twistronics and moiré physics can be translated over a single-layer metasurface platform, introducing a practical route toward the observation of extreme phenomena in a variety of wave systems, potentially applicable to both quantum and classical settings without multilayered fabrication requirements. S. Yves, A. Alù Photonics Initiative Advanced Science Research Center City University of New York New York, NY 10031, USA E-mail: aalu@gc.cuny.edu M. I. N. Rosa, Y. Guo, M. Gupta, M. Ruzzene Department of Mechanical Engineering University of Colorado Boulder Boulder, CO 80309, USA E-mail: massimo.ruzzene@colorado.edu A. Alù Physics Program Graduate Center City University of New York New York, NY 10026, USA The ORCID identification number(s) for the author(s) of this article can be found under https://doi.org/10.1002/advs.202200181 © 2022 The Authors. Advanced Science published by Wiley-VCH GmbH. This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. DOI: 10.1002/advs.202200181 1. Introduction New discoveries in condensed matter physics have recently shown how a twist in pairs of 2D stacked layers can produce highly unexpected emergent phenomena. Notably, the fine-tuning of such twist allows the emergence of a magic angle at which a plethora of new phenom- ena can be observed, including flat-band superconductivity,[1] the quantum Hall effect,[2] the creation of moiré excitons,[3–8] as well as interlayer magnetism.[9] Based on these concepts, atomic photonic crystals in twisted bilayer graphene have shown the ability to route solitons[10,11] and pro- duce quasi-crystalline phases,[12] higher- topology,[13] non-Abelian gauge order potential,[14] and helical topological state mosaics.[15,16] These phenomena, at the heart of the thriving field of twistronics,[17] arise from the hybridization of the band structures associated with the two isolated monolayers, and the associated formation of moiré superlattices. Macroscopic-scale implementations of these concepts using phononic and photonic metamaterials[18,19] have demonstrated flat bands in macroscopic analogues of bi- layer graphene,[20–23] field localization within moiré lattices,[24–26] the destruction of valley topological protection,[27] artificial gauge fields,[28] and broadband tunable bianisotropy for biosensing applications.[29–31] These concepts have also been recently transposed to optical metamaterials, based on extreme anisotropic responses over hy- perbolic metasurfaces (HMTs).[32] Their iso-frequency contours (IFCs) support an open, hyperbolic topology,[33–37] featuring wave propagation with enhanced local density of states, and enabling subwavelength imaging, as well as negative refraction and canal- ization, inherently broadband in nature. By stacking two hyper- bolic metasurfaces and rotating one with respect to the other, it is possible to largely modify the IFCs, inducing transitions be- tween different topologies, from hyperbolic to elliptical.[38] Such effect is the wave analogue of a Lifshitz transition in electronic band structures,[39] which is known to play a crucial role in the physics of Weyl and Dirac semimetals.[40] These exciting phe- nomena have also been recently demonstrated in polaritonic systems.[41–43] The remarkable features of twisted bilayers exploit the in- terplay between two distinct layers with exotic wave responses, Adv. Sci. 2022, 9, 2200181 2200181 (1 of 8) © 2022 The Authors. Advanced Science published by Wiley-VCH GmbH www.advancedsciencenews.com www.advancedscience.com especially for field localization effects, and generally require a precise control over their coupling, alignment and twist angle. Hence, experimental setups, however reconfigurable, are quite challenging.[44] In an attempt to circumvent such difficulties, a few recent studies have theoretically explored the emergence of analogous responses in single-layer systems, with interac- tions or properties modulated by a second virtual layer. For in- stance, quasi-flat bands, Dirac cones, and quantum anomalous phases have been predicted in modulated optical lattices,[45,46] while topological spectral gaps characterized by second Chern numbers akin to the 4D quantum Hall effect were illustrated in phononic lattices.[47] Lifting the requirement of two stacked layers opens new prospects for the implementation of twistronics across several electronic, photonic, and phononic platforms. Toward this goal, in this Letter we explore the effects of emergent moiré patterns in monolayer pillared metasurfaces formed by the relative twist of two 2D spatial features: the lattice defined by the position of the pillars and the one defined by the anisotropic modula- tion profile of their height, which defines their resonant fea- tures. We first demonstrate that aligning these two lattices, re- sulting in an untwisted metasurface, and controlling their fea- tures, can produce a wide range of elliptical and hyperbolic IFCs. Next, we show that introducing a relative rotation between these two 2D lattices generates quasi-periodic moiré patterns govern- ing topological transitions between open and closed IFCs for spe- cific twist angles. Such transitions inherently occur in a differ- ent way from those emerging in twisted bilayers,[32,41,42,43] for which the interplay between two material hyperbolic surfaces de- fines the transition instead of the emerging moiré patterns.We demonstrate the extreme wave phenomena in such twisted inter- leaved lattices with a highly reconfigurable metasurface formed by LEGO pillar-cone resonators over an elastic plate, which is a 2D extension of previously employed implementations used to study the role of disorder[48] and quasi-periodicity.[49] Our re- sults show the great potential of this platform to study analogues of condensed matter phenomena at the macroscopic scale and in classical settings, and open the door to applications harness- ing both strong anisotropy and moiré physics for enhanced wave manipulation. 2. Results and Discussion Our metasurface consists of a thin elastic plate featuring an ar- ray of pillars in a square lattice of period a (Figure 1a). The pillars can be modeled as mechanical dipolar resonators cou- pled to the transverse motion of the plate, and they are char- acterized by two bending resonant modes (along x and y) of equal frequency due to their symmetric cross section. Pillar-type resonators have been employed in the design of metamateri- als and metasurfaces, notably in the context of bandgaps,[50–54] cloaking,[55] and for seismic mitigation.[56] Nonsymmetric em- bedded resonators have previously been used to implement elas- tic hyperbolic metamaterials,[57–61] whereby large asymmetric couplings within the plate can generate anisotropic effective properties, which can be exploited in the context of waveguiding and subdiffraction imaging. Rather than breaking the resonator symmetry, here we induce strong anisotropy through lattice ef- fects, by spatially modulating the resonant features of the array with a wavelength 𝜆 = Na. This effect introduces a spatial mis- match within the lattice, effectively creating a resonant macrocell including N distinct resonators responsible for asymmetric cou- plings across the metasurface. More specifically, we modify the pillar heights according to the modulation profile S (x, y, 𝜃) = cos [2𝜋/𝜆(cos 𝜃x + sin 𝜃y)], where 𝜃 is a twist angle measured with re- spect to the x axis (Figure 1a). The height hn of each pillar defines the resonant frequency of the dominant mode of interest, and it is assigned by sampling the modulation surface at the lattice sites xn, yn, i.e., hn = h0 [1 + 𝛼S(xn,yn,𝜃)], where h0 is the mean height and 𝛼 is the modulation amplitude. This scheme generates two interleaved spatial features, consisting of the underlying square lattice of period a and of the sampled height distribution at the lattice sites. We begin by highlighting the wave propagation features of the plate in the untwisted (𝜃 = 0◦ , 𝛼 = 0.1) configuration, with pe- riod a along y and Na along x. We consider N = 2, resulting in a diatomic lattice of resonators (Figure 1b), for which the band structure is shown in Figure 1c (the complete band structure can be found in Figure S1, Supporting Information). All band struc- ture computations and response simulations here are calculated with COMSOL Multiphysics, with details provided in the Sec- tion S1 (Supporting Information). The interleaving of the two lattices corresponding to the position of the resonators on the plate and to the resonance modulation, introduces an asymme- try and therefore a mismatch in IFCs along x and y, which re- sults in hyperbolic IFCs around the resonance, two of which are highlighted by black lines in the figure, as well as elliptical ones. The existence of hyperbolic and elliptical IFCs is confirmed by simulating the harmonic response due to a point source exci- tation applied at the center of a finite sample comprising 80 × 80 unit cells. The resulting out-of-plane displacement field, and its Fourier transform (FT) displayed in Figure 1d, illustrate the emergence of hyperbolic and elliptic bands for the three frequen- cies marked in Figure 1c, namely 485 Hz, 625 Hz and 670 Hz. Modifying the height modulation, quantified by the parameter 𝛼, can dramatically change the coupling asymmetry within the sur- face, and correspondingly tailor the IFC shape. Two examples are displayed in Figure 1e at the frequencies of the hyperbolic con- tours in Figure 1c. In the top panel, at 485 Hz, an increase in 𝛼 results in an inversion of IFC curvature, which changes from hyperbolic, to flat, to open-elliptical, to finally close into an ellip- tical shape, demonstrating a topological transition. In the bottom panel, at 625 Hz, another topological transition from hyperbolic to elliptical phases occurs, this time as 𝛼 decreases. In this case, the presence of elliptical IFCs at neighboring frequencies (Fig- ure 1c) facilitates the transition between the two regimes, requir- ing smaller variations of 𝛼 to drive the process. We confirm these phenomena experimentally using our elastic metasurface platform, exploiting the fact that underlying sampling lattice is square. Our metasurface comprises 44 × 44 resonators whose heights are modulated with 𝜆 = 2a (Figure 1f), and we choose the maximum value of 𝛼 allowed by the LEGO pillar geometry, as shown in the inset. The resonances are tuned by sliding the cones along the pillars, following the modulation of hn defined above (Figure 1f, inset). Our LEGO platform provides straightforward tunability and reconfigurability, which we harness to demonstrate extreme wave phenomena and topological transitions. The plate is excited at its center by an Adv. Sci. 2022, 9, 2200181 2200181 (2 of 8) © 2022 The Authors. Advanced Science published by Wiley-VCH GmbH www.advancedsciencenews.com www.advancedscience.com Figure 1. a) Moiré interleaved metasurface: A square lattice of pillars whose heights are modulated according to a rotating profile, here for 𝛼 = 0.1. b) Schematic of the periodic system with 𝜃 = 0◦ , and corresponding unit cell (inset). c) Numerical band structure of (b) with three contours corresponding to hyperbolic (at 485 Hz) along y, hyperbolic along x (at 625 Hz) and elliptical (at 670 Hz) highlighted with black lines. d) Simulated displacement field and corresponding spatial FT zoomed in at the center of the Brillouin zone for the three frequencies highlighted in (c). e) Modification of the IFCs as a function of height modulation at 485 Hz (top) and 625 Hz (bottom). f) Corresponding sample made of LEGO elements with cones at alternating heights (inset). g) Experimentally measured out-of-plane displacement field map and corresponding spatial FT at 345 Hz (left), 470 Hz (center), and 510 Hz (left). electrodynamic shaker, which applies a pseudo-random excita- tion in the 200 − 700 Hz range, and the resulting out-of-plane wave fields are recorded by a scanning Doppler vibrometer (see Figure S2, Supporting Information for the experimental setup). While some hybridization exists between symmetric and asym- metric Lamb waves in the close vicinity of the resonances due to out-of-plane breaking of mirror symmetry, the modes of interest are mainly polarized along the out-of-plane direction (see Figure S3, Supporting Information for the out-of-plane polarization). Figure 1g displays the real and reciprocal space maps of the mea- sured fields at three selected frequencies (345, 470, and 510 Hz): the measured hyperbolic and elliptical propagation are consis- tent with those predicted in simulations, with a small frequency shift attributed to minor differences between experimental and numerical models (see Figure S1, Supporting Information for the simulation of the LEGO lattice band structure). These results clearly show that a spatially anisotropic resonance frequency modulation can generate broadband hyperbolic mechanical Lamb waves, easily implemented over our platform. Moreover, the straightforward tuning of the height modulation amplitude enables a precise control and drastic variations of the supported IFCs. Next, we explore the effect of rotating the interleaved lattices, by twisting the modulation profile relative to the underlying square lattice of resonators (Figure 1a). The misalignment be- tween the lattice and modulation profile produces moiré patterns associated with complex spatial arrangements of the couplings. As illustrated in Figure 2a, 2D modulation patterns with a strong angular dependence appear for 0◦ < 𝜃 < 45◦ (after 45◦, the be- havior is simply inverted because of symmetry). For a generic twist angle, the resulting pattern is quasiperiodic, and the peri- odicity is only restored for specific angles 𝜃 = cos −1(p2/q), where {p2,q} are integers belonging to a Pythagorean triple satisfying p2 1 + p2 2 = q2.[24] These periodic configurations are characterized by unit cells that are typically very large: for instance, the two smallest super-cells are obtained for 𝜃 = cos −1(4/5)≅36.87°, re- sulting in a 5 × 10 super-cell, and 𝜃 = cos −1(12/13) ≅22.62°, re- sulting in a 13 × 26 super-cell. The complexity of the periodic angles and the increasing size of the super-cells makes the analy- sis through Bloch procedures very challenging, if not prohibitive. Adv. Sci. 2022, 9, 2200181 2200181 (3 of 8) © 2022 The Authors. Advanced Science published by Wiley-VCH GmbH www.advancedsciencenews.com www.advancedscience.com Figure 2. a) Pillar height modulation profile as a function of the rotation angle (zoomed detail in inset). b) Simulated out-of-plane displacement field maps (top) and spatial FT (bottom) as a function of the rotation angle for hyperbolicity along the y axis at 485 Hz. c) Same as (b) for the hyperbolicity along the x axis at 625 Hz. Instead, we observe the proposed moiré phenomena by analyz- ing the out-of-plane displacement in real and reciprocal spaces for fixed frequency as a function of the twist angle. Overall, we find evidence of a very rich behavior of the result- ing metasurfaces, whereby different IFC transitions occur at dif- ferent frequencies. We focus on the two hyperbolic regimes pre- sented in Figure 1c. The first example at 485 Hz is illustrated in Figure 2b, at which the wave directionality rotates in the opposite direction compared to the twist angle, until 𝜃 = 30°. The effec- tive wavelength of the guided waves drastically increases in the small angle regime (𝜃 < 10◦), as displayed on Figure 2b, evidence of the progressive emergence of a moiré pattern introducing a super-lattice with long spatial wavelengths (Figure 2a). We note that the metasurface response is strongly affected by the twist an- gle, as long as the wavelength of the moiré pattern is larger than the wavelength of the propagating waves. In reciprocal space, we correspondingly observe the presence of spatial harmonics that move away from the center of the Brillouin zone toward larger wavenumbers as the twist angle increases, in line with a decrease in moiré periodicity. When 𝜃 gets closer to 45◦, an inverted phe- nomenon arises, albeit less noticeable in the 2D modulation pro- file, and some spatial harmonics move closer to the center, caus- ing a distortion of the IFCs. Similar to the case at 𝜃 = 4◦ , this effect hampers surface wave propagation, which is linked to the emergence of partial bandgaps and band flattening caused by the interaction of different spatial harmonics. The rigorous analysis of this phenomenon is inherently complex due to the quasiperi- odic nature of the system and it goes beyond the scope of this work. A different evolution of the supported band structure as a func- tion of the twist angle can be observed in Figure 2c, correspond- ing to excitation at 625 Hz. At 𝜃 = 0◦ the wave propagation is hyperbolic but with opposite orientation. As 𝜃 increases, the field progressively loses directionality, and becomes completely delocalized above 10◦. In reciprocal space (bottom row), this field evolution manifests itself as a topological transition of the asso- ciated IFCs, which evolve from open hyperbolic to closed ellip- tical for increasing twist angle. Similar to Figure 2b, this is ex- plained by the distortion of the original contours due to emerging quasi-periodic modulation and super-lattice patterns. Although the transition here is driven by the twist angle, its emergence is inherently different from the ones observed in previous studies of twisted hyperbolic metasurface bilayers:[32,41,42,43] here the sys- tem consists of a single layer whose interleaved spatial modu- lation lattices and emerging moiré patterns directly control the coupling between resonators, governing the IFC features. Adv. Sci. 2022, 9, 2200181 2200181 (4 of 8) © 2022 The Authors. Advanced Science published by Wiley-VCH GmbH www.advancedsciencenews.com www.advancedscience.com Figure 3. a) Doubling the modulation period enables a better sampling of the resulting modulation profile. b) The modulation profile is better preserved during the twist. c) Pillar height modulation as a function of rotation angle, with a zoom-in inset for four angles. d) Simulated out-of-plane displacement field maps (top) and spatial FT (bottom) as a function of the rotation angle for hyperbolicity along y at 470 Hz. e–g) Same as (d) in the case of topological transitions at 595, 620, and 645 Hz, respectively. Next, we explore the possibility of precisely tuning the dis- persion profile across smoother topological transitions. As noted above, the spatial features emerging from twisting the two inter- leaved lattices are characterized by a large change in their peri- odicities as the twist angle is varied. Increasing the twist angle can quickly degrade the nature of the modulation, as a function of how coarse the height modulation is sampled by the period of the square lattice. The associated angular sensitivity of this phe- nomenon can be expectedly reduced by increasing the periodicity of the modulation profile. For example, Figure 3a considers the untwisted scenario when the modulation period is doubled to 𝜆 = 4a. This change translates into a smoother correlation between twist angle and resulting anisotropic contours, with considerably smaller distortions (Figure 3b,c). As a consequence, the original spatially anisotropic distribution of couplings within the mono- layer, which is responsible for the hyperbolic features, can be bet- ter preserved as the twist angle changes. The resulting wave propagation features are summarized in Figure 3. Figure 3d considers the frequency for which the original untwisted structure supports directional hyperbolic waves ori- ented along the y axis, for excitation at 470 Hz. Although super- lattice phenomena occur at small angles, their impact on the IFCs Adv. Sci. 2022, 9, 2200181 2200181 (5 of 8) © 2022 The Authors. Advanced Science published by Wiley-VCH GmbH www.advancedsciencenews.com www.advancedscience.com Figure 4. a) LEGO metasurfaces as a function of rotation angle for 𝜆 = 4a. b) Experimentally measured field maps (top) and spatial FT (bottom) as a function of the rotation angle in the case of hyperbolicity along the y axis at 311 Hz. c–e) Same as (b) in the case of topological transitions at 430, 452.5, and 462.5 Hz, respectively. is less pronounced now, compared to Figure 2a. Indeed, as the angle increases, the propagation of directional waves smoothly follows the modulation rotation. In reciprocal space, the corre- sponding hyperbolic contours, albeit progressively flatter due to small changes in the couplings induced by moiré effects, undergo a similar rotation, indicating an effective twist of the metasurface properties occurring over a large angular range. Next, we focus on the frequency range associated with hyper- bolicity along the x direction, displayed on Figure 3e–g (for 595, 620, and at 645 Hz, respectively). In the untwisted case (𝜃 = 0◦ ), these frequencies are related to different anisotropic phases: the contour in (e) is an ellipse that progressively opens as the fre- quency is increased to become flat in (f). A further frequency in- crease results in a curvature inversion, leading to a hyperbolic IFC (Figure 3g). As we increase the twist angle, Figure 3e demon- strates an opening of the IFC, and correspondingly a topological transition from elliptical to hyperbolic. The resulting canalized waves follow the rotation of the modulation profile until 𝜃 = 45◦ . In the case of Figure 3f, an overall rotation of the flat contour, as well as its progressive curvature inversion, is observed as a func- tion of the twist angle. Finally, Figure 3g shows a complete topo- logical transition from hyperbolic to elliptical contours, driven by the twist. These findings clearly show that the moiré patterns in- duced by the twist between the interleaved lattices are responsi- ble for topological transitions and canalized waves. The increased modulation wavelength (𝜆 = 4a) results in a better preservation of the untwisted anisotropic coupling distribution. This smoothens the transitions compared to the results of Figure 2b and allows to observe these moiré phenomena over larger angular ranges. These results suggest a straightforward experimental imple- mentation and observation of these phenomena on our reconfig- urable LEGO platform comprising 44 × 44 resonators. We im- plemented several configurations for 𝜃 = 0◦ , 15◦, 30◦, 45◦, as shown in Figure 4a. The sample snapshots illustrate the rotation of the modulation profile, as well as the distortion caused by the sampling as it is twisted relative to the interleaved metasurface lattice (see the pattern formed by the black and blue stripes in the insets). Figure 4b shows the experimentally measured field Adv. Sci. 2022, 9, 2200181 2200181 (6 of 8) © 2022 The Authors. Advanced Science published by Wiley-VCH GmbH www.advancedsciencenews.com www.advancedscience.com profile (top) and corresponding spatial FT (bottom) for hyperbol- icity along y (at 311 Hz) as a function of the twist angle. As 𝜃 in- creases, the wave directionality accordingly rotates, reflected into the corresponding IFCs, which also become flatter and closer to the origin in Fourier space. This behavior, albeit less clear than in Figure 3d because of the smaller size of the sample, follows the trend seen in simulations. Figure 4c–e consider frequencies that support hyperbolic waves along x. Panel (c), for excitation at 430 Hz, starts from delocalized fields for 𝜃 = 0◦ , and the prop- agation becomes strongly canalized as the angle increases, with a directionality following the modulation twist, consistent with Figure 3e. Our measurements confirm a moiré-driven topologi- cal phase transition between closed and open contours. Next, Fig- ure 4d, for excitation at 452.5 Hz, shows a rotation in wave direc- tionality from 𝜃 = 0◦ to 45◦. Although less evident due to the smaller size of the plate, this transition is consistent with the re- sults in Figure 3f. Finally, Figure 4e presents results at 462.5 Hz, at which the waves are canalized and twisted from 𝜃 = 0◦ to 30◦, and then become delocalized at 45◦, experimentally confirming a reverse topological transition, from open to closed contours, as the twist angle increases, similar to Figure 3f. Overall, these ex- perimental results clearly illustrate that tailoring the modulation parameters of twisted interleaved lattices over a metasurface pro- duces topological transitions between delocalized and canaliza- tion regimes, as well as an effective rotation of the guided wave directionality, with the frequency being a key parameter that de- fines the type of observed transition. Overall, these results showcase the rich behavior associated with moiré physics and hyperbolic dispersion within a single- layer metasurface. We note that not all moiré physics associated to bilayer systems can be easily transposed to single layers. No- tably, the interlayer coupling parameter is important for some applications, and it can be challenging to find its equivalent in single-layer systems.[45,46] Moreover, while bilayers may be mod- eled based on the properties of the individual layers,[32,41] addi- tional moiré effects and quasi-periodicity are inevitable in our sys- tem, which makes their modeling more complex. Finally, while inducing dynamical reconfigurability in our single layer moiré system requires a more complex implementation of active de- vices, it does not rely on any physical displacement between the layers, making it more robust and fully controllable. 3. Conclusion In this paper, we have investigated the effect of twisting inter- leaved lattices over a single-layer pillared metasurface. We first explored the case where the two governing spatial features, the position and height modulation profile of the pillars, are aligned but feature a mismatch in their spatial period along one direc- tion. The resulting periodic metasurface supports hyperbolic fea- tures over a broad range of frequencies, whose emergence has been observed both numerically and experimentally over a LEGO platform. Next, we introduced a relative rotation between the in- terleaved lattices, which induces moiré patterns that generates emerging wave phenomena, resulting in drastic modifications of the IFCs that undergo a transition from open to closed contours as the twist angle varies. A coarser sampling of the modulation patterns causes these transitions to be abrupt and to occur over a limited range of twist angles. The effect of sampling is expounded by increasing the modulation wavelength, which results in a bet- ter preservation of the spatial features as the twist angle changes, and produces smoother topological transitions. Such transitions are associated with extreme anisotropic features, inducing wave canalization along specific directions that are controlled by the twist angle within a range of angles and frequencies. In stark contrast to the case of twisted hyperbolic bilayers,[32,41,42,43] these transitions are driven by the moiré patterns emerging within the sample as the rotation angle changes. Moreover, albeit of topo- logical nature, they differ from topological phases in chiral hy- perbolic metamaterials, which are related to pseudo-spin propa- gation at the edge of the system.[62] We have observed these phe- nomena over a simple, practical, and highly reconfigurable LEGO platform, which allowed us to observe with flexibility the various regimes discovered in our study. As such and considering ad- ditional practical implementation challenges, they can be trans- lated over a broad range of physical domains, including quan- tum and nanophotonic systems or microphononics in the con- text of pillared media,[63,64] opening opportunities for twistronic- induced phenomena that do not require multilayered fabrication and careful control alignment, interlayer couplings, and twisting. The reconfigurability of our approach, in contrast with previously investigated twisted bilayer configurations, opens new opportu- nities for single-layer moiré metasurfaces featuring high tunabil- ity of anisotropic responses. Such tunability emerges as a func- tion of the twist angle, which is a single parameter defining the considered modulation. This suggests new opportunities stem- ming from the rich physics of twistronics and moiré phenomena, which may also open the door to dynamically reconfigurable de- vices capable of real-time enhanced wave manipulation. Supporting Information Supporting Information is available from the Wiley Online Library or from the author. Acknowledgements S.Y. and M.I.N.R. contributed equally to this work. S.Y. and A.A. acknowl- edge funding from the Simons Foundation, the National Science Founda- tion EFRI program, and the Air Force Office of Scientific Research MURI program. M. I. N.R. and M.R. gratefully acknowledge the support from the National Science Foundation (NSF) through the EFRI 1741685 grant and from the Army Research Office through grant W911NF-18-1-0036. LEGO is a trademark of the LEGO Group, which does not sponsor, authorize, or endorse this paper. Conflict of Interest The authors declare no conflict of interest. Data Availability Statement The data that support the findings of this study are available from the cor- responding authors upon reasonable request. Keywords hyperbolic, metasurface, moiré materials, phononics, quasi-periodicity, topological transitions, wave steering Adv. Sci. 2022, 9, 2200181 2200181 (7 of 8) © 2022 The Authors. Advanced Science published by Wiley-VCH GmbH www.advancedsciencenews.com www.advancedscience.com Received: January 10, 2022 Revised: February 9, 2022 Published online: March 6, 2022 [30] Y. Zhao, A. N. Askarpour, L. Sun, J. Shi, X. Li, A. Alù, Nat. Commun. 2017, 8, 14180. [31] H. X. Xu, G. Hu, L. Han, M. Jiang, Y. Huang, Y. Li, X. Yang, X. Ling, L. Chen, J. Zhao, C. W. Qiu, Adv. Opt. Mater. 2019, 7, 1801479. [32] G. Hu, A. Krasnok, Y. Mazor, C. W. Qiu, A. Alù, Nano Lett. 2020, 20, 3217. [1] Y. Cao, V. Fatemi, S. Fang, K. Watanabe, T. Taniguchi, E. Kaxiras, P. Jarillo-Herrero, Nature 2018, 556, 43. [2] C. R. Dean, L. Wang, P. Maher, C. Forsythe, F. Ghahari, Y. Gao, J. Ka- toch, M. Ishigami, P. Moon, M. Koshino, T. Taniguchi, Nature 2013, 497, 598. [33] A. Poddubny, I. Iorsh, P. Belov, Y. Kivshar, Nat. Photonics 2013, 7, 948. [34] A. A. High, R. C. Devlin, A. Dibos, M. Polking, D. S. Wild, J. Perczel, [35] N. P. De Leon, M. D. Lukin, H. Park, Nature 2015, 522, 192. J. S. Gomez-Diaz, M. Tymchenko, A. Alù, Phys. Rev. Lett. 2015, 114, 233901. [3] K. Tran, G. Moody, F. Wu, X. Lu, J. Choi, K. Kim, A. Rai, D. A. Sanchez, [36] D. Correas-Serrano, J. S. Gomez-Diaz, A. A. Melcon, A. Alù, J. Opt. J. Quan, A. Singh, J. Embley, Nature 2019, 567, 71. [4] K. L. Seyler, P. Rivera, H. Yu, N. P. Wilson, E. L. Ray, D. G. Mandrus, J. Yan, W. Yao, X. Xu, Nature 2019, 567, 66. [5] C. Jin, E. C. Regan, A. Yan, M. I. B. Utama, D. Wang, S. Zhao, Y. Qin, S. Yang, Z. Zheng, S. Shi, K. Watanabe, Nature 2019, 567, 76. [6] E. M. Alexeev, D. A. Ruiz-Tijerina, M. Danovich, M. J. Hamer, D. J. Terry, P. K. Nayak, S. Ahn, S. Pak, J. Lee, J. I. Sohn, M. R. Molas, Nature 2019, 567, 81. [7] H. Yu, G. B. Liu, J. Tang, X. Xu, W. Yao, Sci. Adv. 2017, 3, e1701696. [8] F. Wu, T. Lovorn, A. H. MacDonald, Phys. Rev. Lett. 2017, 118, 147401. [9] W. Chen, Z. Sun, Z. Wang, L. Gu, X. Xu, S. Wu, C. Gao, Science 2019, 366, 983. 2016, 18, 104006. J. S. Gomez-Diaz, A. Alù, ACS Photonics 2016, 3, 2211. [37] [38] H. N. S. Krishnamoorthy, Z. Jacob, E. Narimanov, I. Kretzschmar, V. M. Menon, Science 2012, 336, 205. I. M. Lifshitz, Sov. Phys. - JETP 1960, 11, 1130. [39] [40] N. P. Armitage, E. J. Mele, A. Vishwanath, Rev. Mod. Phys. 2018, 90, 015001. [41] G. Hu, Q. Ou, G. Si, Y. Wu, J. Wu, Z. Dai, A. Krasnok, Y. Mazor, Q. [42] Zhang, Q. Bao, C. W. Qiu, A. Alù, Nature 2020, 582, 209. J. Duan, N. Capote-Robayna, J. Taboada-Gutiérrez, G. Álvarez-Pérez, I. Prieto, J. Martín-Sánchez, A. Y. Nikitin, P. Alonso-González, Nano Lett. 2020, 20, 5323. [10] L. Jiang, Z. Shi, B. Zeng, S. Wang, J. H. Kang, T. Joshi, C. Jin, L. Ju, J. [43] Z. Zheng, F. Sun, W. Huang, J. Jiang, R. Zhan, Y. Ke, H. Chen, S. Deng, Kim, T. Lyu, Y. R. Shen, Nat. Mater. 2016, 15, 840. Nano Lett. 2020, 20, 5301. [11] S. S. Sunku, G. Ni, B. Y. Jiang, H. Yoo, A. Sternbach, A. S. McLeod, T. Stauber, L. Xiong, T. Taniguchi, K. Watanabe, P. Kim, Science 2018, 362, 1153. [12] S. J. Ahn, P. Moon, T. H. Kim, H. W. Kim, H. C. Shin, E. H. Kim, H. W. Cha, S. J. Kahng, P. Kim, M. Koshino, Y. W. Son, Science 2018, 361, 782. [13] B. Liu, L. Xian, H. Mu, G. Zhao, Z. Liu, A. Rubio, Z. F. Wang, Phys. Rev. Lett. 2021, 126, 066401. [14] P. San-Jose, J. Gonzalez, F. Guinea, Phys. Rev. Lett. 2012, 108, 216802. [15] S. Huang, K. Kim, D. K. Efimkin, T. Lovorn, T. Taniguchi, K. Watanabe, A. H. MacDonald, E. Tutuc, B. J. LeRoy, Phys. Rev. Lett. 2018, 121, 037702. [44] R. Ribeiro-Palau, C. Zhang, K. Watanabe, T. Taniguchi, J. Hone, C. R. Dean, Science 2018, 361, 690. [45] T. Salamon, A. Celi, R. W. Chhajlany, I. Frérot, M. Lewenstein, L. Tar- ruell, D. Rakshit, Phys. Rev. Lett. 2020, 125, 030504. [46] T. Salamon, R. W. Chhajlany, A. Dauphin, M. Lewenstein, D. Rakshit, Phys. Rev B 2020, 102, 235126. [47] M. I. N. Rosa, M. Ruzzene, E. Prodan, Commun. Phys. 2021, 4, 130. [48] P. Celli, B. Yousefzadeh, C. Daraio, S. Gonella, Appl. Phys. Lett. 2019, 114, 091903. [49] M. I. N. Rosa, Y. Guo, M. Ruzzene, Appl. Phys. Lett. 2021, 118, 131901. [50] Y. Achaoui, A. Khelif, S. Benchabane, L. Robert, V. Laude, Phys. Rev. B 2011, 83, 10401. [16] Q. Tong, H. Yu, Q. Zhu, Y. Wang, X. Xu, W. Yao, Nat. Phys. 2017, 13, [51] M. Rupin, F. Lemoult, G. Lerosey, P. Roux, Phys. Rev. Lett. 2014, 112, 356. 234301. [17] S. Carr, D. Massatt, S. Fang, P. Cazeaux, M. Luskin, E. Kaxiras, Phys. Rev. B 2017, 95, 075420. [18] G. Hu, C. W. Qiu, A. Alù, Opt. Mater. Express 2021, 11, 1377. [19] B. Lou, N. Zhao, M. Minkov, C. Guo, M. Orenstein, S. Fan, Phys. Rev. Lett. 2021, 126, 136101. [52] L. Cao, Z, Y., Y. Xu, B. Assouar, Smart Mater. Struct. 2018, 27, 075051. [53] W. Wang, B. Bonello, B. Djafari-Rouhani, Y. Pennec, J. Zhao, Phys. Rev. Appl. 2018, 10, 064011. [54] O. R. Bilal, A. Foehr, C. Daraio, Extreme Mech. Lett. 2017, 15, 103. [55] A. Colombi, P. Roux, S. Guenneau, M. Rupin, J. Acoust. Soc. Am. 2015, [20] M. R. López, F. Peñaranda, J. Christensen, P. San-Jose, Phys. Rev. Lett. 137, 1783. 2020, 125, 214301. [21] Y. Deng, M. Oudich, N. J. Gerard, J. Ji, M. Lu, Y. Jing, Phys. Rev. B 2020, 102, 180304. [22] S. M. Gardezi, H. Pirie, S. Carr, W. Dorrell, J. E. Hoffman, 2D Mater. 2021, 8, 031002. [23] K. Dong, T. Zhang, J. Li, Q. Wang, F. Yang, Y. Rho, D. Wang, C. P. Grigoropoulos, J. Wu, J. Yao, Phys. Rev. Lett. 2021, 126, 223601. [24] P. Wang, Y. Zheng, X. Chen, C. Huang, Y. V. Kartashov, L. Torner, V. V. [56] P. Roux, D. Bindi, T. Boxberger, A. Colombi, F. Cotton, I. Douste- Bacque, S. Garambois, P. Gueguen, G. Hillers, D. Hollis, T. Lecocq, Seismol. Res. Lett. 2018, 89, 582. J. H. Oh, H. M. Seung, Y. Y. Kim, Appl. Phys. Lett. 2014, 104, 073503. J. H. Oh, Y. K. Ahn, Y. Y. Kim, Struct. Multidiscip. Optim. 2015, 52, 1023. [57] [58] [59] H. Lee, J. H. Oh, H. M. Seung, S. H. Cho, Y. Y. Kim, Sci. Rep. 2016, 6, 24026. Konotop, F. Ye, Nature 2020, 577, 42. [60] R. Zhu, Y. Y. Chen, Y. S. Wang, G. K. Hu, G. L. Huang, J. Acoust. Soc. [25] M. Martí-Sabaté, D. Torrent, Phys. Rev. Appl. 2021, 15, L011001. [26] Q. Fu, P. Wang, C. Huang, Y. V. Kartashov, L. Torner, V. V. Konotop, F. Ye, Nat. Photonics 2020, 14, 663. Am. 2016, 139, 3303. [61] H. W. Dong, S. D. Zhao, Y. S. Wang, C. Zhang, Sci. Rep. 2018, 8, 2247. [62] W. Gao, M. Lawrence, B. Yang, F. Liu, F. Fang, B. Béri, J. Li, S. Zhang, [27] Y. Jin, W. Wang, Z. Wen, D. Torrent, B. Djafari-Rouhani, Extreme Mech. Phys. Rev. Lett. 2015, 114, 037402. Lett. 2020, 39, 100777. [63] M. Yan, J. Lu, F. Li, W. Deng, X. Huang, J. Ma, Z. Liu, Nat. Mater. 2018, [28] M. I. Cohen, C. Jörg, Y. Lumer, Y. Plotnik, E. H. Waller, J. Schulz, G. 17, 993. von Freymann, M. Segev, Light: Sci. Appl. 2020, 9, 200. [29] Y. Zhao, M. Belkin, A. Alù, Nat. Commun. 2012, 3, 870. [64] M. Yan, W. Deng, X. Huang, Y. Wu, Y. Yang, J. Lu, F. Li, Z. Liu, Phys. Rev. Lett. 2021, 127, 136401. Adv. Sci. 2022, 9, 2200181 2200181 (8 of 8) © 2022 The Authors. Advanced Science published by Wiley-VCH GmbH
10.1038_s41597-022-01645-3
OPEN DATA DESCRIPTOR Probabilistic atlas for the language network based on precision fMRI data from >800 individuals Benjamin Lipkin1,2 ✉, Greta Tuckute1,2, Josef Affourtit1,2, Hannah Small3, Zachary Mineroff Hope Kean1,2,13, Olessia Jouravlev5,13, Lara Rakocevic1,2,13, Brianna Pritchett1,2,13, Matthew Siegelman6,13, Caitlyn Hoeflin7,13, Alvincé Pongos Melissa Kline Struhl1,13, Anna Ivanova1,2,13, Steven Shannon2, Aalok Sathe Malte Hoffmann 10, Alfonso Nieto-Castañón2,11 & Evelina Fedorenko 8,13, Idan A. Blank9,13, 1,2, 1,2,12 ✉ 4, Two analytic traditions characterize fMRI language research. One relies on averaging activations across individuals. This approach has limitations: because of inter-individual variability in the locations of language areas, any given voxel/vertex in a common brain space is part of the language network in some individuals but in others, may belong to a distinct network. An alternative approach relies on identifying language areas in each individual using a functional ‘localizer’. Because of its greater sensitivity, functional resolution, and interpretability, functional localization is gaining popularity, but it is not always feasible, and cannot be applied retroactively to past studies. To bridge these disjoint approaches, we created a probabilistic functional atlas using fMRI data for an extensively validated language localizer in 806 individuals. This atlas enables estimating the probability that any given location in a common space belongs to the language network, and thus can help interpret group- level activation peaks and lesion locations, or select voxels/electrodes for analysis. More meaningful comparisons of findings across studies should increase robustness and replicability in language research. Background & Summary fMRI is an invaluable non-invasive tool for illuminating the brain’s architecture, especially for human-unique abilities like language. A common analytic approach in fMRI language studies is to average activation maps voxel-wise in a common brain space and perform statistical inference across individuals in each voxel. However, because of the well-established inter-individual variability in the locations of functional areas in the association cortex1,2, activations do not align well across individuals, leading to low sensitivity and functional resolution3. Further, the results of group-averaging analyses are generally interpreted through reverse inference from ana- tomical locations to function4,5, but because of the variability mentioned above, combined with the functional heterogeneity of the association cortex, locations in a common brain space cannot be meaningfully linked to function (see6 for a discussion of this issue for ‘Broca’s area’). 1Department of Brain and cognitive Sciences, Massachusetts institute of technology, cambridge, MA, USA. 2McGovern institute for Brain Research, Massachusetts institute of technology, cambridge, MA, USA. 3Department of cognitive Science, Johns Hopkins University, Baltimore, MD, USA. 4Human-computer Interaction Institute, carnegie Mellon University, Pittsburgh, PA, USA. 5Department of cognitive Science, carleton University, Ottawa, On, canada. 6Department of Psychology, columbia University, new York, nY, USA. 7Harris School of Public Policy, University of chicago, chicago, iL, USA. 8Department of Bioengineering, University of california, Berkeley, cA, USA. 9Department of Psychology, University of california, Los Angeles, cA, USA. 10Athinoula A. Martinos center for Biomedical imaging, Massachusetts General Hospital, cambridge, MA, USA. 11Department of Speech, Language, and Hearing Sciences, Boston University, Boston, MA, USA. 12Department of Speech, Hearing, Bioscience, and technology, Harvard University, cambridge, MA, USA. 13these authors contributed equally: Hope Kean, Olessia Jouravlev, Lara Rakocevic, Brianna Pritchett, Matthew Siegelman, Caitlyn Hoeflin, Alvincé Pongos, Idan A. Blank, Melissa Kline Struh, Anna ivanova. ✉e-mail: lipkinb@mit.edu; evelina9@mit.edu Scientific Data | (2022) 9:529 | https://doi.org/10.1038/s41597-022-01645-3 1 www.nature.com/scientificdata Fig. 1 Language atlas topography. Probabilistic functional atlas for the language > control contrast based on overlaid individual binarized activation maps (where in each map, the top 10% of voxels are selected, as described in the text). (a) SPM-analyzed volume data in the MNI template space (based on 806 individual maps). (b) FreeSurfer-analyzed surface data in the FSaverage template space (based on 804 individual maps). In both figures, the color scale reflects the proportion of participants for whom that voxel/vertex belongs to the top 10% of language > control voxels/vertices (threholded at p = 0.2 for visualization purposes). An alternative analytic approach, which circumvents voxel-wise brain averaging, is known as ‘functional localization’3,7. In this approach, a brain region or network that supports a mental process of interest is identified with a functional contrast in each individual and then its responses to some new critical condition(s) are exam- ined. This approach yields greater sensitivity, functional resolution, and interpretability, and has been successful across many domains of perception and cognition, including language. As a result, many research groups are now moving away from group-averaging analyses toward individual-subject analyses4,8. However, functional localization is not always feasible. Further, although studies that rely on functional local- ization can be straightforwardly compared to each other, it is at present unclear how to relate the results from such studies to group-averaging fMRI studies, or other studies that rely on brain averaging (e.g., studies that use voxel-based morphometry (VBM) or voxel-based lesion-symptom mapping (VLSM) in patient work9). To help bridge the gap between these two analytic traditions in language research, we created a probabilistic functional atlas of the language network (‘Language Atlas’ or LanA) by overlaying 806 individual activation maps for a robust at the individual-subject level and extensively validated language ‘localizer’10,11. The language localizer relies on a contrast between the processing of sentences and a linguistically/acousti- cally degraded control condition and is robust to changes in materials, modality of presentation, and task (see Methods). This localizer identifies the left-lateralized fronto-temporal language network (e.g.12–14) that selec- tively6,15 supports high-level language comprehension and production16–18, including the processing of word meanings and combinatorial syntactic/semantic processing19–21. By design, this contrast excludes lower-level perceptual22–25 and speech-articulatory26–28 processes, as well as discourse-level comprehension29–32. Further, a network that closely corresponds to this functional contrast emerges from task-free resting state data33. (Many researchers have postulated functional dissociations among the different brain regions that comprise the lan- guage network (e.g.12–14,34–36). However, the empirical landscape remains complex and ridden with controversy, and the evidence is now overwhelming that—even if dissociations exist within the network— all the language regions are strongly synchronized in their activity33,37,38, suggesting that they form a functionally integrated system). LanA allows one to estimate for any location in a common brain space the probability that it falls within the language network. In this way, this atlas can provide a common reference frame and help interpret (a) group-level activation peaks from past and future fMRI studies, or results of meta-analyses of such peaks39, (b) lesion locations in individual brains40 or lesion overlap loci in VBM/VLSM analyses, and (c) electrode locations in ECoG/SEEG investigations or locations of source-localized activity in MEG studies. Furthermore, LanA (d) can help select language-selective units (voxels, electrodes, MEG channels, even single cells) for analysis in exist- ing datasets, including in studies that aim to relate human neural representations to those from artificial neural network language models41–46, (e) can be related voxel-by-voxel to any whole-brain data47, including structural data, gene expression data48, or receptor density data49, in order to ask whether/how these features correlate with the language network’s topography, and (f) can help select patches in post-mortem brains for cellular analyses to maximize the chances of examining language cortex. Finally, LanA (g) can help guide/constrain functional mapping during brain surgery when fMRI is not possible, although, of course, no clinical decisions should be made based on LanA alone. We make the atlas available for two most commonly used brain templates (Fig. 1): a volume-based template (MNI IXI549Space; SPM1250) and a surface-based template (fsaverage; FreeSurfer51). The use of these common data formats will allow for easy interfacing with existing open data repositories such Scientific Data | (2022) 9:529 | https://doi.org/10.1038/s41597-022-01645-3 2 www.nature.com/scientificdatawww.nature.com/scientificdata/ Age 30.23 ± 7.08 Gender Handedness Native English Speaker Status 40.57% 59.18% 87.72% 4.71% 1.61% 5.96% 78.04% 21.96% Years Male Female Right-Handed Left-Handed Ambidextrous No Handedness Info Native Speakers of English Proficient Speakers of English Table 1. Participant demographics. Summary demographics of the 806 participants included in the atlas. as NeuroVault52 and ENIGMA53. We emphasize that LanA is not a replacement for localizers: when possible, a language localizer task should be performed54. As we show in SI-1, the effect sizes obtained from group-level regions of interest (ROIs) based on LanA, or from commonly used Glasser parcels55 are underestimated relative to individually defined language functional ROIs. We also release (i) individual activation maps (in the MNI and FSaverage spaces), along with demographic data, and (ii) individual-level neural markers (based on the volumetric analyses), including effect sizes, voxel count (activation extent), and stability of activation across runs. The neural marker data can be used as nor- mative distributions, based on neurotypical relatively young adults, against which any new population (e.g., children or individuals with developmental or acquired brain disorders) can be evaluated. Methods Participants. A total of 806 neurotypical adults (477, ~59%, female), aged 19 to 75 (441, ~55%, aged 19–29; 310, ~38%, aged 30–39; 55, ~7%, aged 40+), participated for payment between September 2007 and June 2021, as summarized in Table 1. All participants had normal or corrected-to-normal vision, and no history of neurolog- ical, developmental, or language impairments. Handedness information was collected for 758 (~94%) of the 806 participants. Of those, 707 participants (~93%) were right-handed, as determined by the Edinburgh handedness inventory56 or self-report, 38 (~5%) were left-handed, and 13 (~2%) – ambidextrous. (The participants for whom handedness is missing in the database are most likely right-handed because most of them were tested during the earlier years of data collection when right-handedness was one of the requirements for participation.) Of the 806 participants, 629 (~78%) were native speakers of English, and the remaining 177 (~22%) – native speakers of another language and proficient speakers of English (see38 for evidence that the topography of language responses for a language that an individual is proficient in is similar to that of their native language, and see SI-2 for a com- parison between the atlas generated using the 629 native English speakers vs. the 177 proficient non-native speak- ers). Given this demographic distribution, this atlas represents certain populations better than others, and these biases should be taken into account when the data are interpreted, including in comparison to other populations. Each participant completed a language localizer task10 as part of one of the studies in the Fedorenko lab. Each scanning session lasted between 1 and 2 hours and included a variety of additional tasks. All participants gave informed written consent in accordance with the requirements of the MIT’s Committee on the Use of Humans as Experimental Subjects (COUHES). Participant and session selection. The 806 scanning sessions above (one session per participant) were selected from a total of 1,065 sessions across 819 participants that were available in the Fedorenko Lab database as of June 2021. The goal was to include as many participants as possible and, for the 163 participants who performed a language localizer in multiple sessions, to select a single session with high-quality data. To assess data quality, we examined the stability of the activation topography for the language localizer contrast (see Language localizer paradigm) across runs. This analysis was performed on the data preprocessed and analyzed in the volume (i.e., SPM-based analyses; see SPM preprocessing and analysis pipeline). For 1,062 of the 1,065 sessions, we calculated voxel-wise spatial correlations in activation magnitudes the language > control contrast (see Language localizer paradigm) between the odd-numbered and even-numbered runs (the three remaining sessions consisted of a single run and were evaluated by visual inspection of the contrast maps). The correlation values were calculated within the language ‘parcels’—masks that denote typical locations of language areas. These masks (available at54 http://evlab.mit.edu/funcloc) were derived from a probabilistic language atlas based on 220 participants (a subset of the participants in the current set of 806) and have been used in much past work57–63. Six masks (three in the frontal cortex and three in the temporal and parietal cortex) were derived from the probabilistic atlas in the left hemisphere and mirror-projected onto the right hemisphere. For each session, the correlation values were aver- aged across the twelve parcels, leading to a single value per session. This spatial correlation measure quantifies the stability of the activation landscape and is an objective proxy for data quality; it is affected by factors like head motion or sleepiness, but does not require subjective visual inspection of contrast maps (see SI-4 for evidence that this measure works similarly when considering all voxels vs. only voxels with positive values for the contrast of interest, suggesting that the values are not driven by the difference between responsive and non-responsive vox- els). Sessions where the spatial correlation value was negative (n = 23; ~2%) were excluded, leaving 1,042 sessions across 806 participants. For the 163 participants with more than one session, we selected the session with the highest spatial correlation value for inclusion in the atlas (see11 for evidence of the stability of spatial correlation Scientific Data | (2022) 9:529 | https://doi.org/10.1038/s41597-022-01645-3 3 www.nature.com/scientificdatawww.nature.com/scientificdata/ Version Number of participants Task type Words/ Nonwords per trial A 624 BP 12 Modality Visual Trial duration (ms) Trial-initial Fixation 6000 100 B 67 MP 8 Visual 4800 C 60 N/MP 12 Visual 6000 D 15 MP 12 Visual 6000 E 17 MP 8 Visual 4800 F 6 BP 12 Visual 4800 G 4 MP 8 Visual 4800 H 8 N I 4 MP J 1 N Variable 8 Variable Auditory Auditory Auditory 18000 6000 12000 300 300 300 300 600 300 Stimulus Button icon/ Memory probe Trial-final Fixation Trials per block Block duration (s) Blocks per condition per run 5400 (450/ word) 2800 (350/ word) 4200 (350/ word) 4200 (350/ word) 2800 (350/ word) 4200 (350/ word) 2800 (350/ word) 400 100 3 18 8 1350 0 or 1000 1000 1350 350 500 or 1500 500 350 5 24 4 3 18 8 3 18 6 5 24 4 5 24 4 1350 350 5 24 8 1 18 8 3300–4300 ≤1000 Until 6000 4 24 4 1 12 16 Conditions Sentences, Nonwords Sentences, Wordlists, Nonwords Sentences, Nonwords Sentences, Wordlists, Nonwords Sentences, Wordlists, Jabberwocky, Nonwords Sentences, Wordlists, Jabberwocky, Nonwords Sentences, Nonwords Intact, Degraded Sentences, Wordlists, Jabberwocky, Nonwords Sentences, Nonwords Fixation block duration (s) 14 Fixation blocks per run Total run time (s) Number of runs 5 358 2 16 3 336 2–5 18 5 378 2 18 4 396 2–3 16 5 464 6–8 25 4 504 2–8 16 5 464 2 14 5 358 2 16 5 464 4 16 5 464 2 Table 2. Language localizer versions. Timing parameters for each version of the language localizer task. Under task type, the options are defined as follows: BP = Button Press, MP = Memory Probe, N = No Task. (For the Memory Probe task, the correct probes were approximately equally likely to come from early, middle, and late parts of the string). values across sessions: i.e., if a participant shows a high spatial correlation in one session, they are likely to show a high spatial correlation in another session; unpublished data replicates this result across a larger population and multiple functionally distinct networks). Following this data selection procedure, the Fisher-transformed spatial correlations of the participants’ language > control contrast were r = 0.98 and r = 0.57 for the left and right hemi- spheres, respectively (see11 for similar values on a subset (n = 150) of these participants). Language localizer paradigm. Across the 806 participants, ten language localizer versions were used, as summarized in Table 2. In each version, a sentence comprehension condition was contrasted with a linguistically or acoustically degraded control condition. Visual (reading) and auditory (listening) contrasts have been pre- viously established to engage the same fronto-temporal language network10,38,64,65. Activity in this network has further been shown to not depend on task or materials10 and to show robust effects across typologically diverse languages38. Furthermore, this network can be recovered from naturalistic task-free (resting state) data based on patterns of BOLD signal fluctuations over time33,37, and corresponds nearly perfectly to the network based on the sentences > nonwords contrast10. As a result, we pooled data from across the different localizer versions in the current study (see SI-2 for evidence that an atlas defined on only Localizer Version A, used in the majority of participants, is nearly identical to an atlas that leverages data from all other versions, and SI-3 for a supplementary analysis showing robust language > control effects across all ten versions). The vast majority of participants (624, ~77.4%) performed Localizer version A – a reading version, where sentences and nonword strings are presented one word/nonword at a time at the rate of 450 ms per word/non- word, in a blocked design (with 3 sentences/nonword strings in each 18 s block). Participants were instructed to read attentively and to press a button at the end of each trial, when a picture of a finger pressing a button appeared on the screen. The experiment consisted of two ~6-minute-long runs, for a total of 16 blocks for each of the two conditions. The presentation script and stimuli for this localizer version can be downloaded at54 http://evlab.mit.edu/funcloc/ (for the stimuli used in the other localizer versions, contact EF). Localizer versions Scientific Data | (2022) 9:529 | https://doi.org/10.1038/s41597-022-01645-3 4 www.nature.com/scientificdatawww.nature.com/scientificdata/ Fig. 2 Data processing flowchart. Overview of the SPM and FreeSurfer preprocessing and analysis pipelines. Raw dicom images were converted to NIfTI format, motion-corrected, mapped to a common space and smoothed during preprocessing. Each session was then modeled, t-maps were extracted and thresholded, and all sessions were aggregated to create the probabilistic atlas. B-G (performed by 169 participants, ~21.0%) also used visual presentation, and Localizer versions H-J (per- formed by 13 participants, ~1.6%) used auditory presentation. Details of the similarities and differences in trial structure, timing, and other experimental parameters across versions are summarized in Table 2. fMRI data acquisition. Structural and functional data were collected on the whole-body, 3 Tesla, Siemens Trio scanner with a 12-channel (G1; n = 18) or 32-channel (G2; n = 788) head coil, at the Athinoula A. Martinos Imaging Center at the McGovern Institute for Brain Research at MIT. T1-weighted structural images were col- lected in 176 sagittal slices with 1 mm isotropic voxels (TR = 2,530 ms, TE = 3.48 ms). Functional, blood oxygen- ation level dependent (BOLD), data were acquired using an EPI sequence (with a 90 degree flip angle and using GRAPPA with an acceleration factor of 2), with the following acquisition parameters: 33 (G1) or 31 (G2) 4 mm thick near-axial slices acquired in the interleaved order (with 10% distance factor), 3.0 mm × 3.0 mm (G1) or 2.1 mm × 2.1 mm (G2) in-plane resolution, FoV in the phase encoding (A ≫ P) direction 192 mm (G1) or 200 mm (G2) and matrix size 64 mm × 64 mm (G1) or 96 mm × 96 mm (G2), TR = 2,000 ms and TE = 30 ms. Prospective acquisition correction66 was used to adjust the positions of the gradients based on the participant’s motion from the previous TR. The first 10 s of each run were excluded to allow for steady state magnetization. SPM preprocessing and analysis pipeline. Preprocessing. For the SPM analyses (Fig. 2 [volume]), fMRI data were analyzed using SPM12 (release 7487), CONN EvLab module (release 19b), and custom MATLAB scripts. Each participant’s functional and structural data were converted from DICOM to NIfTI for- mat. All functional scans were coregistered and resampled using B-spline interpolation to the first scan of the first session (Friston et al.67). Potential outlier scans were identified from the resulting subject-motion estimates as well as from BOLD signal indicators using default thresholds in CONN preprocessing pipeline (5 standard deviations above the mean in global BOLD signal change, or framewise displacement values above 0.9 mm; Nieto-Castañón68). Functional and structural data were independently normalized into a common space (the MontrealNeurological Institute [MNI] template; IXI549Space) using SPM12 unified segmentation and normal- ization procedure (Ashburner & Friston69) with a reference functional image computed as the mean functional data after realignment across all timepoints omitting outlier scans. The output data were resampled to a common bounding box between MNI-space coordinates (−90, −126, −72) and (90, 90, 108), using 2 mm isotropic voxels and 4th order spline interpolation for the functional data, and 1 mm isotropic voxels and trilinear interpolation for the structural data. Last, the functional data were smoothed spatially using spatial convolution with a 4 mm FWHM Gaussian kernel. First-level analysis. Effects were estimated using a General Linear Model (GLM) in which each experimental condition was modeled with a boxcar function convolved with the canonical hemodynamic response function (HRF) (fixation was modeled implicitly, such that all timepoints that did not correspond to one of the conditions Scientific Data | (2022) 9:529 | https://doi.org/10.1038/s41597-022-01645-3 5 www.nature.com/scientificdatawww.nature.com/scientificdata/ were assumed to correspond to a fixation period). Temporal autocorrelations in the BOLD signal timeseries were accounted for by a combination of high-pass filtering with a 128 seconds cutoff, and whitening using an AR(0.2) model (first-order autoregressive model linearized around the coefficient a = 0.2) to approximate the observed covariance of the functional data in the context of Restricted Maximum Likelihood estimation (ReML). In addition to experimental condition effects, the GLM design included first-order temporal derivatives for each condition (included to model variability in the HRF delays), as well as nuisance regressors to control for the effect of slow linear drifts, subject-motion parameters, and potential outlier scans on the BOLD signal. FreeSurfer preprocessing and analysis pipeline. For the FreeSurfer analysis (Fig. 2 [surface]), fMRI data were analyzed using FreeSurfer v6.0.0. Each participant’s functional and structural data were converted from DICOM to NIfTI format using the default unpacksdcmdir parameters. (Two of the 806 participants could not be included in this pipeline because their raw dicom files were lost, leaving 804 participants for this analysis.) The raw data were then sampled onto both hemispheres of the FSaverage surface, motion corrected and registered using the middle time point of each run. The data were then smoothed spatially with a 4 mm FWHM Gaussian filter. For the first-level analyses, effects were estimated using a GLM in which each condition was modeled with a first order polynomial regressor fitting the canonical HRF. The GLM also included nuisance regressors for offline-estimated subject-motion parameters. Atlas creation. SPM. Using custom code (available at OSF70), we computed the overlap of the individual activation maps for the language > control contrast using the 806 participants analyzed in the SPM12 pipeline (see SI-5 for evidence that the atlas reaches stability at sample sizes much smaller than 806, which suggests that the current sample size is sufficient to be generalizable). In particular, we used whole-brain t-maps that were gener- ated by the first-level analysis and that contain a t-value for the relevant contrast in each voxel (a post-hoc analysis compared the whole-brain t-maps to their corresponding unscaled contrast maps and found strong voxel-wise correlations over the set of 806 participants: r = 0.93 ± 0.03; see SI-2 for evidence that atlases generated from t-maps vs. contrast maps are highly similar). In each individual map, we selected the 10% of voxels across the brain with the highest t-values for the language > control contrast (average and median minimum t-values across participants were 1.73 and 1.62, respectively; average and median maximum t-values were 13.8 and 13.7, respec- tively). These maps were then binarized so that the selected voxels got assigned a value of 1 and the remaining voxels—a value of 0. Finally, these values were averaged across participants in each voxel. The resulting atlas contains in each voxel a value between 0 and 1, which corresponds to the proportion of the 806 participants for whom that voxel falls in the 10% of voxels across the brain with the highest t-values for the language > control contrast. In the left hemisphere, these values range from 0 to 0.82, and in the RH—from 0 to 0.64 (the values are lower in the RH presumably because the majority of the selected voxels fall in the LH: across participants, average and median proportions of selected voxels falling in the LH are 58.3% and 57.8%, respectively). For more details on ROI-level probability values, see SI-6. FreeSurfer. Using custom code (available at OSF70), we computed the overlap of the individual activation maps for the language > control contrast, using the 804 participants analyzed in the FreeSurfer pipeline. The procedure was similar to that used for the SPM-based atlas, except that the selection of the highest t-values was performed on the surface vertices. To maintain hemispheric asymmetries, rather than evaluating each hemisphere sepa- rately, as is generally common for FreeSurfer analyses, the top 10% of vertices were selected from the vertices pooled across the LH and RH, as for the SPM-based atlas. For this atlas, in the left hemisphere, the proportion values range from 0 to 0.90, and in the RH—from 0 to 0.80 (these values are expectedly higher than those in the SPM-based atlas given the superiority of surface-based inter-individual alignment71). General. We chose the top 10% approach over an approach where each individual map is thresholded at a fixed t-value (as in10) to account for inter-individual variability in the overall strength of BOLD signal responses due to trait or state factors72–75. However, because differences in the language network size may correspond to differences in linguistic experience or ability76, we additionally provide versions of the atlas that are derived from t-maps that are thresholded at p < 0.001, p < 0.01, or p < 0.05 (https://doi.org/10.17605/OSF.IO/KZWBH70). Atlases based on the fixed t-value thresholding approach yield topographies that are very similar to the one based on the top 10% approach (see SI-2). The critical difference between these versions of the atlas is in the interpretation of the overlap values: whereas, as noted above, in the top 10% approach, the overlap values cor- respond to the proportion of the 806 participants for whom that voxel falls in the 10% of voxels across the brain with the highest t-values for the language > control contrast, in the atlases based on the fixed-thresholding approach, the overlap values correspond to the proportion of the 806 participants for whom that voxel is signif- icant for the language > control contrast at the relevant threshold. Note that in addition to the classic left frontal and left temporal areas (and their right-hemispheric homo- topes), several other areas emerge in the atlas, including in the right cerebellum and in the visual cortex. These less canonical areas have been reported in past language studies (e.g.77,78), but we acknowledge that in general, these have not been as thoroughly functionally characterized as the core frontal and temporal areas, and may in future work be shown not to be selective and/or critical for language function. Finally, one might ask: how similar is the topography of a probabilistic functional atlas to that of a random-effects group map for the same data. Of course, these are expected to be correlated given that voxels which are task-responsive in a greater portion of the participant population (i.e., have higher probability overlap values in the atlas) are likely to yield higher t-values in the voxel-wise t-tests (see SI-7 for this comparison for LanA). The critical advantage of the probabilistic functional atlas like LanA over a random-effects map is the Scientific Data | (2022) 9:529 | https://doi.org/10.1038/s41597-022-01645-3 6 www.nature.com/scientificdatawww.nature.com/scientificdata/ Neural Markers Minimum 0.25 Median 0.75 LH Effect Size RH Effect Size LH Voxel Count RH Voxel Count −0.28 −0.40 0 0 LH Spatial Correlation −0.01 RH Spatial Correlation −0.29 0.88 0.20 1197 196 0.74 0.33 1.28 0.43 1908 495 1.00 0.57 1.65 0.73 2594 1044 1.23 0.78 Max 3.68 2.51 4887 4473 1.90 1.75 Table 3. Neural marker distributions. Summary of the neural markers for the language > control contrast of the 803 participants included in the atlas for whom we have 2 or more runs. Effect sizes reflect the % BOLD signal change for the language > control contrast in the language fROIs (estimated using across-runs cross-validation, as described in the text). Voxel counts reflect the number of significant voxels for the critical language > control contrast at a fixed statistical threshold (p < 0.001 uncorrected) within the language parcel boundaries (see text for details; Neural markers). Spatial correlation is defined as the Fisher-transformed Pearson correlation coefficient between the voxel responses for the language > control contrast across odd- and even-numbered runs within the language parcel boundaries. LH = Left Hemisphere; RH = Right Hemisphere. The columns show the values that correspond to the minimum value, the value at the 25th percentile of the population distribution, the median, the value at the 75th percentile, and the maximum value. straightforward interpretation of the voxel values that it affords, in terms of the probability that the voxel belongs to the relevant functional area/network (the language network in this case). Such information cannot be inferred from t-values in a random-effect map without additional assumptions/mapping functions. Neural markers. In addition to the population-level atlases, we also provide a set of individual-level neural markers (based on the volumetric SPM analyses) for the language network in each participant. These neural markers include: effect size, voxel count, and spatial correlation (additional information on these markers pro- vided below). All of these markers have all been shown to be reliable within individuals over time, including across scanning sessions11. We provide each of these measures for each of the ROIs constrained by the previously defined language ‘parcels’ (available at54 http://evlab.mit.edu/funcloc), which include in each hemisphere three frontal parcels (inferior frontal gyrus [IFG], its orbital portion [IFGorb], and middle frontal gyrus [MFG]) and three temporal/parietal ones (anterior temporal [AntTemp], posterior temporal [PostTemp], and angular gyrus [AngG]), for a total of 12 parcels. Of the 806 participants included in the atlas, only 803 completed the 2 or more runs, as needed to calculate the effect size and spatial correlation markers; for the remaining 3 participants, only voxel count is provided. Effect size was operationalized as the magnitude (% BOLD signal change) of the critical language > control contrast. Within each parcel, we defined—for each participant—a functional ROI (fROI) by selecting 10% of the mask’s total voxels with the highest t-values for the language > control contrast using all but one run of the data. We then extracted from the left-out run the responses to the language and control conditions and computed the lan- guage > control difference. This procedure was repeated across all run partitions. This across-runs cross-validation procedure3 ensures independence between the data used to define the fROIs and estimate their responses79. In the final step, the estimates were averaged across the cross-validation folds to derive a single value per participant per fROI. Voxel count (activation extent) was defined as the number of significant voxels for the critical language > con- trol contrast at a fixed statistical threshold (p < 0.001 uncorrected threshold). Spatial correlation (stability of the activation landscape) was defined—for the voxels falling within the language parcels—as the Fisher-transformed Pearson correlation coefficient between the voxel responses for the language > control contrast across odd- and even-numbered runs. As noted above, for all three measures, we provide 14 values per participant: one for each of the 12 ROIs (6 in each hemisphere), and two additional per-hemisphere values (averaging across the 6 ROIs in each hemisphere). See Table 3 for a summary of these neural markers within the atlas population. Additional measures can be computed based on the measures we provide (e.g., lateralization can be computed from the voxel counts80), and other measures can be extracted from the whole-brain activation maps (see Data records). These different measures can be explored with respect to each other, or to the demographic variables (but see81 for a discussion about the prevalence of underpowered brain-behavior individual differences studies). These measures can also serve as normative distributions against which any new population can be evaluated, including children or individuals with developmental and acquired brain disorders, or otherwise atypical brains82. Data Records The full dataset, including the SPM and FreeSurfer atlases are available for download83 (https://doi.org/10.6084/ m9.figshare.20425209). Along with the atlases, we make available i) individual contrast and significance maps (for both the volume-based SPM and the surface-based FreeSurfer pipelines; because we had not obtained con- sent for raw data release, we cannot make publicly available the raw dicom/NIfTI images), and ii) a dataset of individual neural markers. The complete dataset can additionally be accessed at http://evlabwebapps.mit.edu/langatlas/ via the prepack- aged download links. The ‘Download SPM Atlas’ and ‘Download FS Atlas’ links provide a copy of the language atlas in their respective formats. The SPM atlas is a single volumetric NIfTI file, whereas the FS atlas is com- prised of two overlay NIfTI files, one for each hemisphere. Under ‘Download All SPM Data’ and ‘Download All FS Data’, each of the individual participant’s data can be downloaded. In particular, for each of the 806 Scientific Data | (2022) 9:529 | https://doi.org/10.1038/s41597-022-01645-3 7 www.nature.com/scientificdatawww.nature.com/scientificdata/ participants (804 for FS), we provide a ‘Demograhics_&_Summary.txt’ file, which contains relevant information as in Tables 2 and 3, the individual contrast and significance maps, and a visualization of their individual activa- tion profile in the selected template space. As well as allowing the user to download the data, the LanA website offers opportunities for online explo- ration and the retrieval of relevant subsets of data. In particular, individual activation maps can be explored under the ‘Explore Activation Maps’ tab, and relevant neural markers can be explored under the ‘Explore Neural Markers’ tab. In addition, data can be filtered by demographic variables including, age, gender, handedness, native English speaker status, language network lateralization, etc., and these subsets can be downloaded, or their maps/neural markers can be explored. This flexible tool allows individual users to access relevant data for their needs without the requirement for offline filtering. Finally, at54, we provide a version of the language localizer experiment (Localizer Version A, which is used for the majority of participants) for download. technical Validation The individual participants’ data quality check was performed as described in the Participants and session selec- tion section. Individual localizer versions were evaluated to confirm they each elicited a strong language > control effect, as described in SI-3. The atlas creation process was evaluated with respect to several hyperparameter choices, and the atlas remained robust to each decision, including the inclusion of non-native but proficient English speakers, differ- ent localizer versions, the use of whole-brain maps based on t-values vs. contrast values, and definition of the language system as the top 10% of language > control voxels vs. as language > control voxels that pass a specific significance threshold. We summarize the (minimal) impact of all these choices in SI-2. Finally, in SI-1, we demonstrate that group-level ROIs defined based on the highest-overlap voxels in LanA outperform commonly used Glasser ROIs derived from multi-modal Human Connectome Project data55 in effect size estimation. The latter grossly underestimate effect sizes, especially for the frontal language areas. Of course, as expected3, individual-level language fROIs are still the best for accurately estimating effect sizes, and these outperform the group-based LanA fROIs, but in cases where individual localization may not be possible (e.g., in retroactively re-analyzing past studies), LanA-based group ROIs are recommended, as they fare substan- tially better than Glasser ROIs. Usage Notes The data records presented in this paper, including materials for download and exploration at the http:// evlabwebapps.mit.edu are available for free and fair use to individual and academic researchers, institutions, and entities provided that this work is appropriately referenced. Although this atlas has potential for clinical applications, the authors assume no responsibility for the use or misuse of LanA and associated data records in clinical and other settings. Code availability Code associated with this manuscript can be found at OSF70. Received: 18 March 2022; Accepted: 9 August 2022; Published: xx xx xxxx References 1. Frost, M. A. & Goebel, R. Measuring structural–functional correspondence: Spatial variability of specialised brain regions after macro-anatomical alignment. NeuroImage 59, 1369–1381 (2012). 2. Tahmasebi, A. M. et al. Is the Link between Anatomical Structure and Function Equally Strong at All Cognitive Levels of Processing? Cereb. Cortex 22, 1593–1603 (2012). 3. Nieto-Castañón, A. & Fedorenko, E. Subject-specific functional localizers increase sensitivity and functional resolution of multi- subject analyses. NeuroImage 63, 1646–1669 (2012). 4. Fedorenko, E. The early origins and the growing popularity of the individual-subject analytic approach in human neuroscience. Curr. Opin. Behav. Sci. 40, 105–112 (2021). 5. Poldrack, R. A. Inferring Mental States from Neuroimaging Data: From Reverse Inference to Large-Scale Decoding. Neuron 72, 692–697 (2011). 6. Fedorenko, E. & Blank, I. A. Broca’s Area Is Not a Natural Kind. Trends Cogn. Sci. 24, 270–284 (2020). 7. Saxe, R. Why and how to study Theory of Mind with fMRI. Brain Res. 1079, 57–65 (2006). 8. Gratton, C. & Braga, R. M. Editorial overview: Deep imaging of the individual brain: past, practice, and promise. Curr. Opin. Behav. Sci. 40, iii–vi (2021). 9. Wilson, S. M. Lesion-symptom mapping in the study of spoken language understanding. Lang. Cogn. Neurosci. 32, 891–899 (2017). 10. Fedorenko, E., Hsieh, P.-J., Nieto-Castañón, A., Whitfield-Gabrieli, S. & Kanwisher, N. New Method for fMRI Investigations of Language: Defining ROIs Functionally in Individual Subjects. J. Neurophysiol. 104, 1177–1194 (2010). 11. Mahowald, K. & Fedorenko, E. Reliable individual-level neural markers of high-level language processing: A necessary precursor for relating neural variability to behavioral and genetic variability. NeuroImage 139, 74–93 (2016). 12. Friederici, A. D. The cortical language circuit: from auditory perception to sentence comprehension. Trends Cogn. Sci. 16, 262–268 (2012). 13. Price, C. J. A review and synthesis of the first 20 years of PET and fMRI studies of heard speech, spoken language and reading. Neuroimage 62, 816–847 (2012). 14. Hagoort, P. & Indefrey, P. The Neurobiology of Language Beyond Single Words. Annu. Rev. Neurosci. 37, 347–362 (2014). 15. Fedorenko, E., Behr, M. K. & Kanwisher, N. Functional specificity for high-level linguistic processing in the human brain. Proc. Natl. Acad. Sci. 108, 16428–16433 (2011). 16. Menenti, L., Gierhan, S. M. E., Segaert, K. & Hagoort, P. Shared Language: Overlap and Segregation of the Neuronal Infrastructure for Speaking and Listening Revealed by Functional MRI. Psychol. Sci. 22, 1173–1182 (2011). Scientific Data | (2022) 9:529 | https://doi.org/10.1038/s41597-022-01645-3 8 www.nature.com/scientificdatawww.nature.com/scientificdata/ 17. Silbert, L. J., Honey, C. J., Simony, E., Poeppel, D. & Hasson, U. Coupled neural systems underlie the production and comprehension of naturalistic narrative speech. Proc. Natl. Acad. Sci. 111, E4687–E4696 (2014). 18. Hu, J. et al. The language network supports both lexical access and sentence generation during language production. 2021.09.10.459596 Preprint at https://doi.org/10.1101/2021.09.10.459596 (2021). 19. Bautista, A. & Wilson, S. M. Neural responses to grammatically and lexically degraded speech. Lang. Cogn. Neurosci. 31, 567–574 (2016). 20. Fedorenko, E., Nieto-Castañon, A. & Kanwisher, N. Lexical and syntactic representations in the brain: An fMRI investigation with multi-voxel pattern analyses. Neuropsychologia 50, 499–513 (2012). 21. Fedorenko, E., Blank, I. A., Siegelman, M. & Mineroff, Z. Lack of selectivity for syntax relative to word meanings throughout the language network. Cognition 203, 104348 (2020). 22. Overath, T., McDermott, J. H., Zarate, J. M. & Poeppel, D. The cortical analysis of speech-specific temporal structure revealed by responses to sound quilts. Nat. Neurosci. 18, 903–911 (2015). 23. Norman-Haignere, S., Kanwisher, N. G. & McDermott, J. H. Distinct Cortical Pathways for Music and Speech Revealed by Hypothesis-Free Voxel Decomposition. Neuron 88, 1281–1296 (2015). 24. McCandliss, B. D., Cohen, L. & Dehaene, S. The visual word form area: expertise for reading in the fusiform gyrus. Trends Cogn. Sci. 7, 293–299 (2003). 25. Baker, C. I. et al. Visual word processing and experiential origins of functional selectivity in human extrastriate cortex. Proc. Natl. Acad. Sci. 104, 9087–9092 (2007). 26. Bohland, J. W. & Guenther, F. H. An fMRI investigation of syllable sequence production. NeuroImage 32, 821–841 (2006). 27. Basilakos, A., Smith, K. G., Fillmore, P., Fridriksson, J. & Fedorenko, E. Functional Characterization of the Human Speech Articulation Network. Cereb. Cortex 28, 1816–1830 (2018). 28. Bouchard, K. E., Mesgarani, N., Johnson, K. & Chang, E. F. Functional organization of human sensorimotor cortex for speech articulation. Nature 495, 327–332 (2013). 29. Ferstl, E. C. & von Cramon, D. Y. The role of coherence and cohesion in text comprehension: an event-related fMRI study. Cogn. Brain Res. 11, 325–340 (2001). 30. Lerner, Y., Honey, C. J., Silbert, L. J. & Hasson, U. Topographic Mapping of a Hierarchy of Temporal Receptive Windows Using a Narrated Story. J. Neurosci. 31, 2906–2915 (2011). 31. Blank, I. A. & Fedorenko, E. No evidence for differences among language regions in their temporal receptive windows. NeuroImage 219, 116925 (2020). 32. Jacoby, N. & Fedorenko, E. Discourse-level comprehension engages medial frontal Theory of Mind brain regions even for expository texts. Lang. Cogn. Neurosci. 35, 780–796 (2020). 33. Braga, R. M., DiNicola, L. M., Becker, H. C. & Buckner, R. L. Situating the left-lateralized language network in the broader organization of multiple specialized large-scale distributed networks. J. Neurophysiol. 124, 1415–1448 (2020). 34. Hickok, G. & Poeppel, D. The cortical organization of speech processing. Nat. Rev. Neurosci. 8, 393–402 (2007). 35. Hagoort, P. MUC (Memory, Unification, Control) and beyond. Front. Psychol. 4 (2013). 36. Duffau, H., Moritz-Gasser, S. & Mandonnet, E. A re-examination of neural basis of language processing: Proposal of a dynamic hodotopical model from data provided by brain stimulation mapping during picture naming. Brain Lang. 131, 1–10 (2014). 37. Blank, I., Kanwisher, N. & Fedorenko, E. A functional dissociation between language and multiple-demand systems revealed in patterns of BOLD signal fluctuations. J. Neurophysiol. 112, 1105–1118 (2014). 38. Malik-Moraleda, S. et al. An investigation across 45 languages and 12 language families reveals a universal language network. Nat. Neurosci. 25, 1014–1019 (2022). 39. Hauptman, M., Blank, I. & Fedorenko, E. Non-literal language processing is jointly supported by the language and Theory of Mind networks: Evidence from a novel meta-analytic fMRI approach. (2022). 40. Woolgar, A., Duncan, J., Manes, F. & Fedorenko, E. Fluid intelligence is supported by the multiple-demand system not the language system. Nat. Hum. Behav. 2, 200–204 (2018). 41. Schrimpf, M. et al. The neural architecture of language: Integrative modeling converges on predictive processing. Proc. Natl. Acad. Sci. 118, e2105646118 (2021). 42. Caucheteux, C. & King, J.-R. Language processing in brains and deep neural networks: computational convergence and its limits. https://doi.org/10.1101/2020.07.03.186288 (2020). 43. Millet, J. et al. Toward a realistic model of speech processing in the brain with self-supervised learning. Preprint at http://arxiv.org/ abs/2206.01685 (2022). 44. Toneva, M. & Wehbe, L. Interpreting and improving natural-language processing (in machines) with natural language-processing (in the brain). ArXiv190511833 Cs Q-Bio (2019). 45. Goldstein, A. et al. Shared computational principles for language processing in humans and deep language models. Nat. Neurosci. 25, 369–380 (2022). 46. Jain, S. & Huth, A. G. Incorporating Context into Language Encoding Models for fMRI. http://biorxiv.org/lookup/doi/10.1101/327601 (2018). 47. Markello, R. D. et al. Neuromaps: structural and functional interpretation of brain maps. http://biorxiv.org/lookup/ doi/10.1101/2022.01.06.475081 (2022). 48. Richiardi, J. et al. Correlated gene expression supports synchronous activity in brain networks. Science 348, 1241–1244 (2015). 49. Hansen, J. Y. et al. Mapping gene transcription and neurocognition across human neocortex. Nat. Hum. Behav. 5, 1240–1250 (2021). 50. Statistical parametric mapping: the analyis of funtional brain images. (Elsevier/Academic Press, 2007). 51. Fischl, B., Sereno, M. I., Tootell, R. B. H. & Dale, A. M. High-resolution intersubject averaging and a coordinate system for the cortical surface. Hum. Brain Mapp. 8, 272–284 (1999). 52. Gorgolewski, K. J. et al. NeuroVault.org: a web-based repository for collecting and sharing unthresholded statistical maps of the human brain. Front. Neuroinformatics 9 (2015). 53. ENIGMA. https://enigma.ini.usc.edu/. 54. EvLab Functional Localization. http://evlab.mit.edu/funcloc/. 55. Glasser, M. F. et al. A multi-modal parcellation of human cerebral cortex. Nature 536, 171–178 (2016). 56. Oldfield, R. C. The assessment and analysis of handedness: The Edinburgh inventory. Neuropsychologia 9, 97–113 (1971). 57. Diachek, E., Blank, I., Siegelman, M., Affourtit, J. & Fedorenko, E. The Domain-General Multiple Demand (MD) Network Does Not Support Core Aspects of Language Comprehension: A Large-Scale fMRI Investigation. J. Neurosci. 40, 4536–4550 (2020). 58. Ivanova, A. A. et al. Comprehension of computer code relies primarily on domain-general executive brain regions. eLife 9, e58906 (2020). 59. Jouravlev, O. et al. Speech-accompanying gestures are not processed by the language-processing mechanisms. Neuropsychologia 132, 107132 (2019). 60. Jouravlev, O. et al. Reduced Language Lateralization in Autism and the Broader Autism Phenotype as Assessed with Robust Individual‐Subjects Analyses. Autism Res. 13, 1746–1761 (2020). 61. Mollica, F. et al. Composition is the Core Driver of the Language-selective Network. Neurobiol. Lang. 1, 104–134 (2020). 62. Shain, C., Blank, I. A., Fedorenko, E., Gibson, E. & Schuler, W. Robust effects of working memory demand during naturalistic language comprehension in language-selective cortex. http://biorxiv.org/lookup/doi/10.1101/2021.09.18.460917 (2021). 63. Wehbe, L. et al. Incremental Language Comprehension Difficulty Predicts Activity in the Language Network but Not the Multiple Demand Network. Cereb. Cortex 31, 4006–4023 (2021). Scientific Data | (2022) 9:529 | https://doi.org/10.1038/s41597-022-01645-3 9 www.nature.com/scientificdatawww.nature.com/scientificdata/ 64. Chen, X. et al. The human language system does not support music processing. http://biorxiv.org/lookup/doi/10.1101/2021.06.01.446439 (2021). 65. Scott, T. L., Gallée, J. & Fedorenko, E. A new fun and robust version of an fMRI localizer for the frontotemporal language system. Cogn. Neurosci. 8, 167–176 (2017). 66. Thesen, S., Heid, O., Mueller, E. & Schad, L. R. Prospective acquisition correction for head motion with image-based tracking for real-time fMRI. Magn. Reson. Med. 44, 457–465 (2000). 67. Friston, K. J. et al. Spatial registration and normalization of images. Hum. Brain Mapp. 3, 165–189 (1995). 68. Nieto-Castanon, A. Handbook of functional connectivity magnetic resonance imaging methods in CONN. (Hilbert Press, 2020). 69. Ashburner, J. & Friston, K. J. Unified segmentation. NeuroImage 26, 839–851 (2005). 70. Lipkin, B. & Tuckute, G. LanA (Language Atlas): A probabilistic atlas for the language network based on fMRI data from 800+ individuals, Open Science Framework, https://doi.org/10.17605/OSF.IO/KZWBH (2022). 71. Fischl, B. et al. Cortical Folding Patterns and Predicting Cytoarchitecture. Cereb. Cortex 18, 1973–1980 (2008). 72. Erdoğan, S. B., Tong, Y., Hocke, L. M., Lindsey, K. P. & deB Frederick, B. Correcting for Blood Arrival Time in Global Mean Regression Enhances Functional Connectivity Analysis of Resting State fMRI-BOLD Signals. Front. Hum. Neurosci. 10 (2016). 73. He, H., Shin, D. D. & Liu, T. T. Resting state BOLD fluctuations in large draining veins are highly correlated with the global mean signal. 1 (2010). 74. Power, J. D., Schlaggar, B. L. & Petersen, S. E. Recent progress and outstanding issues in motion correction in resting state fMRI. NeuroImage 105, 536–551 (2015). 75. Wong, C. W., Olafsson, V., Tal, O. & Liu, T. T. The amplitude of the resting-state fMRI global signal is related to EEG vigilance measures. NeuroImage 83, 983–990 (2013). 76. Jouravlev, O., Mineroff, Z., Blank, I. A. & Fedorenko, E. The Small and Efficient Language Network of Polyglots and Hyper-polyglots. Cereb. Cortex 31, 62–76 (2021). 77. Murdoch, B. E. The cerebellum and language: Historical perspective and review. Cortex 46, 858–868 (2010). 78. Seydell-Greenwald, A., Chambers, C. E., Ferrara, K. & Newport, E. L. What you say versus how you say it: Comparing sentence comprehension and emotional prosody processing using fMRI. NeuroImage 209, 116509 (2020). 79. Kriegeskorte, N. Pattern-information analysis: From stimulus decoding to computational-model testing. NeuroImage 56, 411–421 (2011). 80. Binder, J. R., Swanson, S. J., Hammeke, T. A. & Sabsevitz, D. S. A comparison of five fMRI protocols for mapping speech comprehension systems. Epilepsia 49, 1980–1997 (2008). 81. Marek, S. et al. Reproducible brain-wide association studies require thousands of individuals. Nature 603, 654–660 (2022). 82. Tuckute, G. et al. Frontal language areas do not emerge in the absence of temporal language areas: A case study of an individual born without a left temporal lobe. Neuropsychologia 169, 108184 (2022). 83. Lipkin, B. et al. LanA Dataset. figshare https://doi.org/10.6084/m9.figshare.20425209 (2022). acknowledgements EF and some of this work were supported by NIH awards K99R00 HD057522, R01 DC016607 (and supplement 3R01DC016607-04S1), R01 DC016950, and research funds from the Department of Brain and Cognitive Sciences, the McGovern Institute for Brain Research, and the Simons Center for the Social Brain. The authors are grateful to all past EvLab members for help with data collection, especially Dima Ayyash, Jeanne Gallée, Alex Paunov, Nir Jacoby, Jayden Ziegler, Rachel Ryskin, Nafisa Syed, Saima Malik-Moraleda, Yotaro Sueoka, Yev Diachek, Elinor Amit, Tariq Cannonier, Jenn Hu, and Meilin Zhan. EF is also grateful to Nancy Kanwisher for mentorship and support, which laid a foundation for this line of work, and to Ted Gibson for the support over the years. author contributions Conceptualization: E.F. Investigation (data collection; ordered by number of fMRI sessions contributed): E.F., S.S., Z.M., H.K., O.J., B.P., M.S., C.H., A.P., I.B., M.K., J.A., A.I. Data curation: B.L., G.T., J.A., L.R., A.N.C., E.F. Formal analysis: B.L., G.T., J.A., A.N.C. Methodology (here, referring to the design of the language localizer experiments): E.F., Z.M., B.P., M.S. Software: B.L., G.T., J.A., H.S., Z.M., H.K., B.P., M.S., A.P., A.S., M.H., A.N.C. Visualization: B.L., G.T., J.A., M.H. Writing–original draft: B.L., J.A., E.F. Writing–review and editing: G.T., I.B., A.I. Project administration, supervision, resources, and funding acquisition: E.F. Competing interests The authors declare no competing interests. additional information Supplementary information The online version contains supplementary material available at https://doi. org/10.1038/s41597-022-01645-3. Correspondence and requests for materials should be addressed to B.L. or E.F. Reprints and permissions information is available at www.nature.com/reprints. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Cre- ative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not per- mitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/. © The Author(s) 2022 Scientific Data | (2022) 9:529 | https://doi.org/10.1038/s41597-022-01645-3 1 0 www.nature.com/scientificdatawww.nature.com/scientificdata/
10.7554_elife.85542
RESEARCH ARTICLE Pooled genome- wide CRISPR activation screening for rapamycin resistance genes in Drosophila cells Baolong Xia1, Raghuvir Viswanatha1, Yanhui Hu1,2, Stephanie E Mohr1,2, Norbert Perrimon1,2,3* 1Department of Genetics, Blavatnik Institute, Harvard Medical School, Boston, United States; 2Drosophila RNAi Screening Center, Harvard Medical School, Boston, United States; 3Howard Hughes Medical Institute, Boston, United States Abstract Loss- of- function and gain- of- function genetic perturbations provide valuable insights into gene function. In Drosophila cells, while genome- wide loss- of- function screens have been exten- sively used to reveal mechanisms of a variety of biological processes, approaches for performing genome- wide gain- of- function screens are still lacking. Here, we describe a pooled CRISPR activa- tion (CRISPRa) screening platform in Drosophila cells and apply this method to both focused and genome- wide screens to identify rapamycin resistance genes. The screens identified three genes as novel rapamycin resistance genes: a member of the SLC16 family of monocarboxylate transporters (CG8468), a member of the lipocalin protein family (CG5399), and a zinc finger C2H2 transcrip- tion factor (CG9932). Mechanistically, we demonstrate that CG5399 overexpression activates the RTK- Akt- mTOR signaling pathway and that activation of insulin receptor (InR) by CG5399 requires cholesterol and clathrin- coated pits at the cell membrane. This study establishes a novel platform for functional genetic studies in Drosophila cells. Editor's evaluation This work is well- structured, with clear objectives and experiments. The authors successfully demon- strated genome- wide gene activation which overcame previous failed attempts to replicate gene activation that worked well in mammalian systems. The study is detailed and relevant for the appli- cation of CRISPRa in understanding the function of candidate genes. Introduction Although Drosophila is one of the most intensively studied organisms, about half of Drosophila protein- coding genes still lack functional characterization (Ewen- Campen et  al., 2017). Genome- scale loss- of- function (LOF) and gain- of- function (GOF) genetic perturbations facilitate functional genomic studies. In Drosophila, genome- wide LOF screens using RNA interference or CRISPR- Cas9 have helped elucidate the mechanisms of a variety of biological processes (Boutros et  al., 2004; Björklund et al., 2006; D’Ambrosio and Vale, 2010; Bard et al., 2006; Guo et al., 2008; Hao et al., 2008; Song et al., 2022; Viswanatha et al., 2018). However, genome- wide GOF screens have not been feasible, largely due to a lack of the reagents for genome- scale GOF perturbation. Current state- of- the- art for Drosophila GOF studies utilizes cDNA library overexpression (Stapleton et al., 2002; Yu et al., 2011), which only covers a subset of the genome and underrepresents genes with long open reading frames (ORFs). *For correspondence: perrimon@genetics.med. harvard.edu Competing interest: The authors declare that no competing interests exist. Funding: See page 15 Preprinted: 10 December 2022 Received: 13 December 2022 Accepted: 09 April 2023 Published: 20 April 2023 Reviewing Editor: K VijayRaghavan, National Centre for Biological Sciences, Tata Institute of Fundamental Research, India Copyright Xia et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 1 of 18 Research article CRISPRa is a complementary approach for GOF studies. Nuclease- dead Cas9 (dCas9) is fused with different transcriptional activators and targeted to the promoter region by single- guide RNAs (sgRNAs) to facilitate transcription at endogenous loci. Synergistic activation mediator (SAM) complex is one of the most efficient CRISPRa systems (Konermann et al., 2015; Chavez et al., 2016). With the SAM system, synthetic transcriptional activators dCas9- VP64 and MCP- p65- HSF1 are recruited to the promoter region of endogenous genes by MS2 hairpin- containing sgRNAs. Genome- scale pertur- bation with CRISPRa can be achieved by synthesizing a genome- wide sgRNA library, bypassing the need to generate a genome- wide ORF library. CRISPR- based GOF approaches have been used for genome- wide genetic screens in mammalian cells (Konermann et al., 2015; Rodríguez et al., 2022; Zhu et al., 2022; Sofer et al., 2022; Zhang et al., 2022), but the feasibility has not yet been tested in Drosophila cells. To address this gap, we developed a pooled genome- wide CRISPRa screening platform in Drosophila cells using the SAM complex. To demonstrate the feasibility of this platform, we performed focused and genome- wide CRISPRa screens and identified novel candidates conferring rapamycin resistance. Next, we focused on one of the top candidates, a member of the lipocalin protein family (CG5399), and demonstrate that it positively regulates the Receptor Tyrosine Kinase (RTK)- Akt- mTOR signaling pathway by regulating cholesterol and clathrin- coated pits at the cell membrane. Results Inducible CRISPRa using the SAM complex in Drosophila S2R+ cells To establish CRISPRa in Drosophila cells in an inducible manner, the synthetic transcriptional activa- tors of the SAM complex (dCas9- VP64 and MCP- p65- HSF1) were placed under a metallothionein promoter, which can be induced in the presence of copper ions. VP64, p65, and HSF1 are transcrip- tional activators while the MS2 coat protein (MCP) recognizes and binds to MS2 hairpins present in the sgRNAs, recruiting the fused transcriptional activators to gene promoters targeted by sgRNAs. In addition, the MS2 hairpin- containing sgRNA was expressed from a separate plasmid under the control of the U6 promoter (Figure 1A). To determine whether this system can mediate transcriptional activa- tion in Drosophila cells, plasmids with sgRNAs targeting the promoter regions of Jon25Biii or Sdr were transfected into Drosophila S2R+ cells stably expressing the metallothionein promoter- driven SAM complex. Without copper induction, sgRNAs moderately activated the target genes, likely due to leaky expression of the SAM complex from the metallothionein promoter. In the presence of copper, A pMT dCas9 VP64 T2A MCP p65 HSF1 U6 Actin attB attB sgRNA GFP T2A PuroR B l e v e l i n o s s e r p x e e v i t l a e R 25 20 15 10 5 0 Jon25Biii Sdr *** C *** ** * *** ns l e v e l i n o s s e r p x e e v i t l a e R 50 40 30 20 10 0 copper - - + + - - + + empty vector empty vector sgJon25Biii empty vector empty vector sgJon25Biii sgSdr sgSdr *** empty vector sgCG9877 ns CG9877 CG13538 Figure 1. Inducible transcriptional activation by the synergistic activation mediator (SAM) complex in Drosophila cells. (A) Schematic of the SAM complex for inducible transcriptional activation. dCas9- VP64 and MCP- p65- HSF1 were driven by an inducible metallothionein promoter. dCas9- VP64 and MCP- p65- HSF1 were expressed as T2A- containing bicistronic transcript. single- guide RNA (sgRNA) was expressed from pLib8 plasmid, which contains an attB flanking GFP- T2A- PuroR cassette for attP sites recombination. (B) Fold activation of Jon25Biii and Sdr expression measured by qPCR. Three biological replicates are shown as individual circles. (C) Fold activation of CG9877 and CG13538 expression measured by qPCR. Three biological replicates are shown as individual circles. t- test, *p<0.05; **p<0.01; ***p<0.001; ns, not significant. The online version of this article includes the following source data for figure 1: Source data 1. Full source data for Figure 1. Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 2 of 18 Genetics and Genomics Research article sgRNAs robustly upregulated the target genes (Figure 1B), showing that the SAM complex is able to mediate gene activation in Drosophila S2R+ cells. The SAM complex modulates gene expression by recruiting transcriptional activators to gene promoters at endogenous genomic loci. In the Drosophila genome, 32% of genes form divergent gene pairs with transcriptional start sites (TSS) <1 kb apart (Yang and Yu, 2009). Thus, we next assessed the potential for collateral activation by the SAM complex of the closely spaced promoters. CG9877 and CG13538 are a divergent gene pair with transcriptional start sites 908 bp apart, thus this gene pair is a good test case for assessing collateral activation. sgRNAs were designed within a region 300 bp upstream of the TSS of CG9877 and ~600 bp away from the TSS of CG13538. The sgRNAs specifically activated CG9877, but not CG13538 (Figure  1C), suggesting that the SAM complex activates the target gene without affecting nearby genes at least in some divergent gene pairs. Pooled CRISPRa screening with a focused library Next, we used the SAM complex to perform a pooled CRISPRa screen. Unlike in mammalian cells, lentivirus vectors are extremely inefficient in Drosophila cells. To overcome this limitation, we previ- ously established a pooled library delivery method based on site- specific recombination following plasmid transfection of Drosophila cells (Viswanatha et al., 2018). In this method, attB sites flanking sgRNAs are integrated into attP sites flanking landing cassette in the presence of phiC31 integrase. We used the S2R+ PT5 cell line in which the attP cassette is inserted into the Clic locus (Neumüller et al., 2012). To establish pooled library cells for CRISPRa screens, we first generated SAM cells that stably expressed the metallothionein promoter- driven SAM complex. Then, using phiC31- mediated cassette exchange, we integrated a pooled guide RNA library into the landing cassette (Figure 2— figure supplement 1A). To test the pooled screen approach, we performed a CRISPRa screen to identify rapamycin resis- tance genes. Rapamycin is an allosteric inhibitor of the kinase mTOR, a key regulator of cell growth and proliferation. Rapamycin inhibits Drosophila S2R+ cell proliferation in a dose- dependent manner. Cell proliferation was partially inhibited by rapamycin at 0.1 nM and almost completely inhibited at 1 nM (Figure 2—figure supplement 1B). To identify a suitable rapamycin treatment condition for a pooled CRISPRa screen, we performed a pilot- focused screen with different rapamycin concentrations and treatment durations. The focused library consists of 6335 sgRNAs targeting the promoter regions of 652 genes, which include genes involved in the mTOR signaling pathway as well as candidates from our previous CRISPR knockout screen (Viswanatha et al., 2018). Pooled library cells were subjected to passaging in 0.1  nM, 1  nM rapamycin, or DMSO (vehicle) containing culture medium in parallel for 15 days or 30 days. After treatment, we examined sgRNA abundance in each condition by next- generation sequencing. First, we identified genes affecting cell fitness by comparing the final popula- tion after passaging in DMSO with the initial population. Scyl and Cyp12a4 are significantly depleted (false- discovery rate (FDR)<0.05) after passaging in DMSO for 15 days and 30 days (Figure 2—figure supplement 1C and D), suggesting that overexpression of scyl or Cyp12a4 affects cell fitness. Consis- tent with this, scyl is known to inhibit cell proliferation and act as a cell death activator (Reiling and Hafen, 2004; Corradetti et al., 2005; Scuderi et al., 2006). These results demonstrate that pooled CRISPRa screen can be used to identify genes affecting cell fitness. We next identified rapamycin resistance genes by comparing the rapamycin- treated population with the DMSO- treated population, reasoning that sgRNAs that lead to activation of rapamycin resistance genes would be enriched after treatment due to a growth advantage in the presence of rapamycin. The focused screen revealed that one candidate, CG8468, was significantly enriched (FDR<0.05) in the population treated with 1  nM rapamycin. Moreover, with prolonged treatment, CG8468 was further enriched as we observed a higher fold- change value at day 30 than at day 15 (Figure 2—figure supplement 1E and F). However, no gene was significantly enriched in the 0.1 nM rapamycin treatment condition, probably because cells have a higher proliferation rate at 0.1  nM rapamycin concentration compared to 1  nM. Taken together, these data demonstrate that pooled CRISPRa screen using the SAM complex is feasible in Drosophila cells. Genome-wide pooled CRISPRa screen Next, we sought to screen for rapamycin resistance genes at the genome- wide scale. A previous study showed that multiplexed sgRNAs performed better than single sgRNAs for CRISPRi and CRISPRa Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 3 of 18 Genetics and Genomics Research article (Replogle et  al., 2020). Thus, we designed a dual- sgRNA library in which each vector expressed two distinct sgRNAs targeting the promoter region of the same gene within 500  bp upstream of the TSS (Figure  2—figure supplement 2A). The library consists of 84,143 vectors targeting the promoter regions of 13,293 protein- coding genes and 2332 non- coding genes. The dual- sgRNA library was constructed using a three- step pooled cloning strategy (Figure  2—figure supplement 2B and Methods). After library construction, we checked the quality of the dual- sgRNA library by deep sequencing, which revealed that ~98.5% of designed vectors were present in the final library. The difference in representation between the 10th percentile of the final library (19 reads) and the 90th percentile (292 reads) was 15.5- fold (Figure 2—figure supplement 2C). The integrity and distribution of our dual- sgRNA library are comparable to other published genome- wide libraries (Sanjana et al., 2014), indicating the high quality of the library. To establish the pooled library cells, we integrated the genome- wide dual- sgRNA library into SAM cells by phiC31- mediated cassette exchange. The pooled library cells were then passaged in 1 nM rapamycin or DMSO- containing medium for 3 weeks. The abundance of dual- sgRNA vectors in the initial and final cell populations following rapamycin or DMSO treatment was analyzed by next- generation sequencing (Figure 2A). First, we identified genes affecting cell fitness by comparing the final population after passaging in DMSO for 3 weeks with the initial population. Consistent with the results of the focused screen, scyl is also significantly depleted (FDR<0.05) in the genome- wide screen dataset. In addition to the scyl, the screen also identified other genes known to be involved in the suppression of cell proliferation (Figure 2—figure supplement 3A and Table 1). By comparing the rapamycin- treated population with the DMSO- treated population, we identified rapamycin resistance genes. Pka- C3 and Cdc25 were among the top 50 ranked genes identified in both replicates of the genome- wide screen (Table 2). Pka- C3 encodes the catalytic subunit of PKA, and overexpression of the catalytic subunit of PKA or activation of the PKA pathway is known to confer resistance to rapamycin (Cutler et al., 2001; Zurita- Martinez and Cardenas, 2005; Schmelzle et al., 2004). Cdc25 is a tyrosine phosphatase gene that regulates cell cycle progression. Previous studies have indicated that the level of Cdc25 expression is positively correlated with rapamycin resis- tance in cancer cells (Reikvam et al., 2014; Chen et al., 2009). In addition to the known rapamycin resistance genes, novel candidates were also identified in the CRISPRa screen. In particular, three genes, CG8468, CG5399, and CG9932, were significantly enriched (FDR<0.05) in both replicates (Figure  2B). CG8468 is a member of the SLC16 family of monocarboxylate transporters. CG5399 encodes a member of the lipocalin protein family, members of which have been implicated in lipid binding and transport. CG9932 encodes a zinc finger C2H2 transcription factor. Interestingly, CG8468 was also the top hit from our focused library screen, indi- cating the consistency of the pooled CRISPRa screen approach in Drosophila cells (note that the other two genes were not included in the focused library). Validation of the novel rapamycin resistance genes To validate that the novel hits from the CRISPRa screen could indeed confer resistance to rapamycin, we first cloned all dual- sgRNA vectors targeting CG8468, CG5399, or CG9932 that were present in the genome- wide library and established individual stable SAM cell lines for each vector. The target activation efficiency of each vector was evaluated in individual cell lines by qPCR. The dual- sgRNA vectors showed variable activation efficiency, probably due to the complex transcriptional regulation of the target genes or different sgRNA binding efficiencies (Figure  2C, F and I). Interestingly, the enrichment of each dual- sgRNA vector in the screen was highly correlated with its target activation efficiency, as only the vectors that efficiently activate target genes were enriched in the rapamycin- treated samples (Figure 2D, G and J and Figure 2—figure supplement 3B–D). Collectively, these results confirm that the sgRNA vectors enriched in the screen were able to upregulate the target genes. To validate that overexpression of the hits confers a growth advantage in the presence of rapa- mycin, we mixed wild- type SAM cells (GFP negative) and individual dual- sgRNA vector expressing cell lines (GFP positive), then monitored the proportion of GFP- positive cells in the mixed cell popu- lations following 1 nM rapamycin or DMSO treatment for 2 weeks. We reasoned that if a dual- sgRNA vector confers resistance to rapamycin, cells with the vector will proliferate more than wild- type SAM cells in the presence of rapamycin, leading to a higher proportion of GFP- positive cells in the Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 4 of 18 Genetics and Genomics Research article Figure 2. Genome- wide CRISPR activation screen for rapamycin resistance genes. (A) Schematic of CRISPR activation screen (See methods). (B) Two replicates of genome- wide CRISPR activation screen. Data were analyzed by MAGeCK- RRA, a smaller RRA score indicates a stronger selection effect. Each circle represents a gene. Circle size corresponds to the significance (p value) of enrichment. Significantly enriched genes (false- discovery rate (FDR)<0.05) are colored. (C) Fold activation of CG8468 expression measured by qPCR. Three biological replicates are shown as individual circles. Figure 2 continued on next page Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 5 of 18 Genetics and Genomics Research article Figure 2 continued (D) Counts of sgCG8468 vectors from the genome- wide screen. Each dot represents a vector. Vectors targeting intergenic regions are shown in blue. Vectors targeting CG8468 are shown in red and annotated as V1- V6. (E) sgCG8468- expressing cell proliferation in cell mixture following 1 nM rapamycin or DMSO treatment. GFP proportion was measured by flow cytometry. Three biological replicates are shown as individual circles. (F) Fold activation of CG5399 expression measured by qPCR. Three biological replicates are shown as individual circles. (G) Counts of sgCG5399 vectors from the genome- wide screen. Each dot represents a vector. Vectors targeting intergenic regions are shown in blue. Vectors targeting CG5399 are shown in red and annotated as V1- V6. (H) sgCG5399- expressing cell proliferation in cell mixture following 1 nM rapamycin or DMSO treatment. GFP proportion was measured by flow cytometry. Three biological replicates are shown as individual circles. (I) Fold activation of CG9932 expression measured by qPCR. Three biological replicates are shown as individual circles. (J) Counts of sgCG9932 vectors from the genome- wide screen. Each dot represents a vector. Vectors targeting intergenic regions are shown in blue. Vectors targeting CG9932 are shown in red and annotated as V1- V6. (K) sgCG9932- expressing cell proliferation in cell mixture following 1 nM rapamycin or DMSO treatment. GFP proportion was measured by flow cytometry. Three biological replicates are shown as individual circles. t- test, *p<0.05; **p<0.01; ***p<0.001; ns, not significant. The online version of this article includes the following source data and figure supplement(s) for figure 2: Source data 1. Full source data for Figure 2. Figure supplement 1. Pooled CRISPR activation screen with a focused library. Figure supplement 1—source data 1. Full source data for Figure 2—figure supplement 1. Figure supplement 2. Design of genome- wide dual- sgRNA library. Figure supplement 2—source data 1. Full source data for Figure 2—figure supplement 2. Figure supplement 3. Genome- wide cell fitness screen and rapamycin screen. Figure supplement 3—source data 1. Full source data for Figure 2—figure supplement 3. rapamycin- treated sample as compared to the DMSO- treated sample. As expected, we only observed higher proportions of GFP for vectors that efficiently activate target genes, but not for inefficient vectors or an empty vector (Figure 2E, H and K). These results confirmed that overexpression of the hits from the genome- wide screen conferred resistance to rapamycin. Table 1. Significantly depleted genes in genome- wide fitness screen. Gene zld Eaat1 CR44587 Human ortholog Known gene affecting cell fitness Reference ZNF485 SLC1A3 - Lis- 1/Ptp52F* PAFAH1B1/Ptprb scyl αTub85E Poxm Dll LKRSDH DDIT4 TUBA1A Pax9 DLX6 AASS scro NKX2- 1 CG3168 CG2930 SV2A SLC15A1 LIS- 1- overexpressing mitotic cells show a variety of spindle defects PMID: 10722879 Scyl inhibits cell growth by regulating the Tor pathway PMID: 15545626 PAX9 overexpression inhibits cancer cell proliferation PMID: 35628401 Overexpression of Aass suppresses cancer cell proliferation PMID: 31601242 NKX2- 1 suppresses lung cancer progression by dampening ERK activity PMID: 34689179 Overexpression of SV2A inhibits the PI3K signaling pathway PMID: 34277597 *Lis- 1 and Ptp52F form divergent gene pair ~500 bp apart. Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 6 of 18 Genetics and Genomics Research article Table 2. Overlapping genes of top- ranked 50 hits from two genome- wide screen replicates. Rank (Rep 1, Rep 2) 1, 1 2, 3 5, 2 4, 10 22, 13 41, 8 Gene CG8468 CG5399 CG9932 Human ortholog Function SLC16A8 monocarboxylate transporter APOD/LCN2 lipocalin ZFN462/REST transcription factor CG34459 / unknown Pka- C3 Ps PRKX NOVA1 catalytic subunit of PKA RNA splicing 43, 45 CDC25 CDC25A/CDC25B tyrosine phosphatase Known rapamycin resistance gene PMID: 15643061, 14673167, 11739804 PMID: 24383842, 19276368 RTK-Akt-mTOR signaling activation by CG5399 overexpression To characterize the mechanism of rapamycin resistance, we first examined the mTOR activity in SAM cell lines with dual- sgRNA vectors activating the target genes. As ribosomal protein S6 is phosphor- ylated by S6K, which is a downstream target of mTOR, the phosphorylation status of S6 can serve as a readout for mTOR activity. In the presence of rapamycin, S6 phosphorylation was strongly inhibited in wild- type SAM cells and empty vector- expressing SAM cells. In contrast, compared to control cells, the phospho- S6 levels were dramatically elevated in CG5399- overexpressing cells (Figure 3A). Inter- estingly, we did not observe higher phospho- S6 levels in CG8468- overexpressing cells, suggesting that CG8468 acts downstream of mTOR or in a parallel pathway. CG9932- overexpressing cells also displayed higher phospho- S6 levels, possibly reflecting a previous observation that overexpression of the human ortholog of CG9932, REST, activates Akt, which acts upstream of S6 (Dobson et al., 2019). The increase of phospho- S6 levels in CG5399- overexpressing cells in the presence of rapamycin might be explained by different mechanisms: (1) CG5399 might alter the pharmacokinetics of rapa- mycin, decreasing cellular rapamycin concentrations; (2) CG5399 might competitively bind to rapa- mycin, releasing mTOR from inhibition; (3) CG5399 might be a positive regulator of mTOR signaling. To distinguish among these possibilities, phospho- S6 was assessed in CG5399- overexpressing cells without rapamycin treatment. Compared with wild- type SAM cells and empty vector- expressing SAM cells, higher phospho- S6 levels were observed in CG5399- overexpressing cells (Figure  3B), indi- cating that CG5399 is a positive regulator of mTOR. Moreover, in addition to higher phopho- S6 levels, an increase in phospho- Akt was also observed in CG5399- overexpressing cells. Furthermore, knocking down CG5399 mRNA levels using either of two nonoverlapping double- stranded RNAs (dsRNAs) in CG5399- overexpressing cells totally abolished the increase of phospho- Akt and phos- pho- S6 (Figure 3C), excluding the possibility that an off- target effect of the sgRNAs explains these observations. Akt is phosphorylated by PI3K when receptor tyrosine kinases (RTKs) are activated. To test whether PI3K is involved in Akt activation by CG5399 overexpression, we used two nonoverlapping dsRNAs to deplete the catalytic subunit of PI3K, Pi3K92E, in CG5399- overexpressing cells. Knockdown of Pi3K92E abolished Akt activation by CG5399 overexpression, suggesting that CG5399 activates Akt- mTOR through PI3K (Figure 3—figure supplement 1A). As Akt is regulated by both insulin receptor (InR) and PDGF/VEGF receptor (Pvr) (Sopko et  al., 2015), we examined whether InR and Pvr are involved in CG5399 function. Two nonoverlapping dsRNAs targeting InR or Pvr were transfected into CG5399- overexpressing cells. Knockdown of InR or Pvr inhibited upregulation of phospho- Akt and phospho- S6 induced by CG5399 overexpression (Figure 3D, Figure 3—figure supplement 1B), suggesting that CG5399 activates Akt- mTOR via InR and Pvr. As expected, higher phospho- InR levels were also observed in CG5399- overexpressing cells (Figure 3E) in a normal medium without insulin stimulation. Finally, the activation of InR- Akt- mTOR signaling in S2R+ cells was also observed with CG5399 ORF overexpression by co- transfecting pUAS- CG5399 and pActin- Gal4 vectors (Figure 3F), further excluding the possibility that off- targets of sgRNAs contribute to the phenotype. Taken Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 7 of 18 Genetics and Genomics Research article A empty vector wt SAM wt SAM sgCG8468 vector3 sgCG8468 vector4 sgCG5399 vector4 sgCG9932 vector3 sgCG5399 vector6 sgCG9932 vector6 empty vector 1 nM Rapa - - + + + + + + ++ E empty vector sgCG5399 wt SAM ** ** Phospho-InR (Y1545) Phospho-S6 (S233/235) Actin B Phospho-S6 (S233/235) Actin sgCG5399 vector6 sgCG5399 vector4 sgCG5399 vector4 sgCG5399 vector6 empty vector empty vector wt SAM wt SAM 40 35 55 40 kDa 40 35 55 40 n i t c A / 6 S o h p s o h P ) d e t a e r t a p a R M n 1 ( 6 4 2 0 ns sgCG5399 vector4 wt SAM sgCG5399 vector6 empty vector ** *** n i t c A / 6 S o h p s o h P ) d e t a e r t a p a R o / w ( 5 4 3 2 1 0 ns F sgCG5399 vector6 sgCG5399 vector4 wt SAM empty vector untransfected empty vector dsInR-1 wt SAM dsGFP dsInR-2 C untransfected empty vector dsCG5399-1 dsCG5399-2 wt SAM dsGFP D sgCG5399 vector4 sgCG5399 vector4 Phospho-Akt (S505) total Akt Actin Phospho-S6 (S233/235) Actin 100 70 100 70 55 40 35 25 55 40 Phospho-Akt (S505) total Akt Actin Phospho-S6 (S233/235) Actin 100 70 100 70 55 40 35 25 55 40 180 130 100 55 40 100 70 100 70 55 40 40 35 55 40 untransfected GFP CG5399 180 130 100 55 40 100 70 100 70 55 40 40 35 55 40 Actin Phospho-Akt (S505) total Akt Actin Phospho-S6 (S233/235) Actin Phospho-InR (Y1545) Actin Phospho-Akt (S505) total Akt Actin Phospho-S6 (S233/235) Actin Figure 3. CG5399 overexpression activates RTK- Akt- mTOR signaling. (A) Phospho- S6 levels in cells expressing dual- sgRNA vectors in the presence of 1 nM rapamycin. Western blot signals are quantitatively analyzed by ImageJ. Three biological replicates are shown as individual circles. (B) Phospho- S6 levels in cells expressing sgCG5399 vectors without rapamycin treatment. Western blot signals are quantitatively analyzed by ImageJ. Four biological replicates are shown as individual circles. (C) Phospho- Akt and Phospho- S6 in CG5399- overexpressing cells following CG5399 knockdown. Two nonoverlapping double- stranded RNAs (dsRNAs) targeting CG5399 were used. (D) Phospho- Akt and Phospho- S6 in CG5399- overexpressing cells following insulin receptor (InR) knockdown. Two nonoverlapping dsRNAs targeting InR were used. (E) Phospho- InR, phospho- Akt, and phospho- S6 in sgCG5399- expressing synergistic activation mediator (SAM) cells. (F) Phospho- InR, phospho- Akt, and phospho- S6 in CG5399 ORF- overexpressing S2R+ cells using UAS- Gal4. t- test, **<0.01; ***p<0.001; ns, not significant. The online version of this article includes the following source data and figure supplement(s) for figure 3: Source data 1. Full source data for Figure 3. Figure supplement 1. CG5399 overexpression activates Akt- mTOR through RTK/PI3K. Figure supplement 1—source data 1. Full source data for Figure 3—figure supplement 1. Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 8 of 18 Genetics and Genomics Research article together, these data demonstrate that CG5399 overexpression activates RTK- Akt- mTOR signaling in a normal medium without insulin stimulation. Insulin receptors form covalent homodimers at the cell surface. Upon insulin binding, the ecto- domain of InR changes from the inverted U- shape structure to the T- shape structure, facilitating the proximity and autophosphorylation of the cytoplasmic kinase domains (Gutmann et al., 2018; Scapin et al., 2018). Drosophila cells are cultured in Schneider medium supplemented with 10% FBS (fetal bovine serum). The Schneider medium only consists of amino acids and inorganic salts while FBS is a biological product that might contain a trace amount of insulin and insulin- like growth factors (Tu et  al., 2018). As we observed that CG5399 overexpression activates the InR in a normal medium without insulin stimulation, we removed FBS from the culture medium to further exclude the effect of insulin and insulin- like growth factors in FBS. Increase of phospho- InR and phospho- Akt in CG5399- overexpressing cells could still be observed after serum starvation for 2 hr (Figure 3—figure supple- ment 1C), suggesting InR activation by CG5399 overexpression does not require insulin. InR regulation by CG5399 requires cholesterol and clathrin-coated pits InR is embedded in the lipid bilayer environment of the cell membrane. Given that CG5399 encodes a member of the lipocalin protein family, members of which have been implicated in the binding and transport of lipid molecules, we hypothesized that CG5399 might regulate InR by remodeling lipid components at the cell membrane. Structure prediction of CG5399 by AlphaFold revealed a highly conserved barrel structure formed by eight beta- sheets as a putative ligand pocket (Figure 4—figure supplement 1A), similar to the crystal structures of lipocalins in other species (Breustedt et al., 2005; Lakshmi et al., 2015). Molecular docking simulation indicated that cholesterol can be inserted into the barrel structure of CG5399 (Figure 4—figure supplement 1B), suggesting that cholesterol might be a substrate of CG5399. To test whether cholesterol is relevant to CG5399 function, we used methyl- beta- cyclodextrin (MβCD) to deplete cholesterol from cell membranes in CG5399- overexpressing cells. MβCD is a heptasaccharide with a high affinity to cholesterol and has been widely used to manipulate membrane cholesterol content (Zidovetzki and Levitan, 2007). MβCD treatment eliminated the increase of phos- pho- InR, phospho- Akt, and phospho- S6 in a dose- dependent manner in CG5399- overexpressing cells (Figure 4A), indicating that activation of InR- Akt- mTOR by CG5399 overexpression requires choles- terol at the membrane. Moreover, supplementation of cholesterol into the cell membrane can rescue the decrease of phospho- Akt induced by MβCD treatment in CG5399- overexpressing cells, excluding the possibility that an off- target effect of MβCD contributes to the observed effect (Figure 4—figure supplement 1C). To test the possibility that MβCD treatment affected the normal function of InR, S2R+ cells were stimulated with insulin following MβCD treatment. No difference in insulin response was observed in MβCD treated and untreated cells (Figure 4—figure supplement 1D), suggesting that MβCD treatment does not affect InR function. Moreover, direct supplementation of cholesterol to the cell membrane activated InR- Akt- mTOR signaling in wild- type S2R+ cells (Figure 4B), indicating that an increase in the level of cholesterol at the cell membrane was able to activate InR. Cell membranes form distinct microdomains, such as caveolin- coated caveolae, flotillin- coated microdomains, and clathrin- coated pits. Cholesterol is required for the formation of different micro- domains (Lu and Fairn, 2018). To distinguish which microdomain was involved in the CG5399 function, we knocked down flotillins and clathrins in CG5399- overexpressing cells using dsRNAs. Knockdown of flotillin genes (Flo1 and Flo2) did not affect phospho- Akt, whereas the increase of phospho- InR and phospho- Akt normally observed in CG5399- overexpression cells was dampened by knockdown of clathrin heavy chain (Chc) (Figure 4C and D). Moreover, InR activation by cholesterol supplementation was also eliminated by Chc knockdown (Figure 4E). Taken together, these results suggest that InR- Akt- mTOR signaling activation by CG5399 overexpression requires cholesterol and clathrin- coated pits at the cell membrane. Discussion Although genome- wide LOF screens in Drosophila cells have helped elucidate the mechanism of a variety of biological processes, genome- wide GOF screens have not been feasible in this organism. To address this gap, we generated a genome- wide dual- sgRNA library that covers both protein- coding Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 9 of 18 Genetics and Genomics Research article A B empty vector wt SAM sgCG5399 vector4 MβCD (mM) - - - 1.25 2.5 5 cholesterol/ethanol/MβCD cholesterol/ethanol/MβCD ethonal/MβCD - ethanol Phospho-InR (Y1545) Actin Phospho-Akt (S505) total Akt Actin Phospho-S6 (S233/235) Actin C Phospho-Akt (S505) total Akt Actin D Phospho-InR (Y1545) Actin 180 130 100 55 40 100 70 100 70 55 40 40 35 55 40 Phospho-InR (Y1545) Actin Phospho-Akt (S505) total Akt Actin Phospho-S6 (S233/235) Actin 180 130 55 40 100 70 100 70 55 40 40 35 55 40 untransfected empty vector dsFlo1-1 dsGFP wt SAM dsFlo1-2 dsFlo2-1 dsFlo2-2 untransfected empty vector dsChc-1 dsGFP dsChc-2 wt SAM sgCG5399 vector4 sgCG5399 vector4 Phospho-Akt (S505) total Akt Actin 100 70 100 70 55 40 E 100 70 100 70 55 40 untransfected dsChc-1 dsGFP untransfected dsChc-2 dsChc-1 dsGFP dsChc-2 untransfected untransfected dsGFP dsGFP dsChc-1 dsChc-1 dsChc-2 dsChc-2 empty vector sgCG5399 cholesterol - + - + - + - + 180 130 100 55 40 Phospho-InR (Y1545) Actin 180 130 100 55 40 Figure 4. Activation of InR- Akt- mTOR signaling by CG5399 overexpression requires cholesterol and clathrin- coated pits at the membrane. (A) Phospho- InR, phospho- Akt, and phospho- S6 in CG5399- overexpressing cells treated with methyl- beta- cyclodextrin (MβCD) at different concentrations. (B) Phospho- InR, phospho- Akt, and phospho- S6 in S2R+ cells with cholesterol supplementation. Two different cholesterol products from Sigma (C3045 for Lane 3 and C2044 for Lane 4) were used. (C) Phospho- Akt in CG5399- overexpressing cells following Flo1, Flo2, or Chc knockdown. Two nonoverlapping double- stranded RNAs (dsRNAs) targeting each gene were used. (D) Phospho- InR in CG5399- overexpressing cells following Chc knockdown. Two nonoverlapping dsRNAs targeting each gene were used. (E) Phospho- InR in S2R+ cells with cholesterol supplementation following clathrin heavy chain (Chc) knockdown. Two nonoverlapping dsRNAs targeting each gene were used. The online version of this article includes the following source data and figure supplement(s) for figure 4: Source data 1. Full source data for Figure 4. Figure supplement 1. CG5399 interacts with cholesterol by molecular docking. Figure supplement 1—source data 1. Full source data for Figure 4—figure supplement 1. Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 10 of 18 Genetics and Genomics Research article genes and non- coding genes. This library also captures transcriptional complexity by targeting alter- native promoters of the same gene, making it suitable for screens in different contexts in the future. Moreover, in this library, each gene is targeted by 4–6 dual- sgRNA vectors, which helps overcome the inefficient activation by some sgRNAs to some extent. Indeed, we observed that some dual- sgRNA vectors failed to activate the target genes (Figure 2F and I), possibly due to steric hindrance by pre- existing proteins in the promoter regions, or due to different sgRNA binding efficiency, which might be optimized by applying machine learning to large screen datasets in the future. In our method, the synthetic transcriptional activators of the SAM complex (dCas9- VP64 and MCP- p65- HSF1) are driven by the metallothionein promoter. Due to the leaky expression of this promoter, we observed moderate gene activation without copper induction. Availability of a tighter controlled promoter may be required for more sensitive screens. As the Drosophila genome is relatively compact, one concern for CRISPRa is the collateral activation of adjacent genes. Previous work has shown that sgRNAs targeting areas beyond –600 bp from the TSS in Drosophila lose efficiency (Mao et al., 2020), which may alleviate the concern of collateral activation when the sgRNAs target sites are over 600 bp away from the TSS of neighboring genes. In some cases, sgRNAs are designed within both promoters of closely spaced divergent genes. If the hits from the screen form closely spaced divergent gene pairs, e.g., Lis- 1 and Ptp52F in the cell fitness screen (Table 1), more experiments will be required to identify which one contributes to the phenotype. Currently, in our CRISPRa system, spCas9 requires NGG as a protospacer- adjacent motif (PAM) sequence, which limits sgRNA design in the small region upstream of TSS, especially for closely spaced genes. Next- generation CRISPRa system could use PAMless Cas9 variants (Walton et al., 2020) to remove the PAM constraint for sgRNA design, allowing the construc- tion of libraries with more sgRNAs per gene. Our genome- wide genetic screen identified some known rapamycin resistance genes and novel candidates. Overexpression of CG5399, which encodes a lipocalin family protein, confers resistance to rapamycin and activates RTK- Akt- mTOR signaling. The activation of InR by CG5399 requires choles- terol and clathrin- coated pits at the cell membrane (Figure  4—figure supplement 1E). CG5399 is predicted to have a transmembrane helix at the C- terminus by PredictProtein and be located on the cell membrane by DeepLoc. As InR is embedded in the cell membrane and the rearrangement of InR transmembrane domain is crucial for tyrosine kinase domain activation, changing the lipid environ- ment is a reasonable possible mechanism for the regulation of InR activation (Gutmann et al., 2018; Scapin et  al., 2018). A recent study showed that lipid exchange to form ordered domains at cell membranes induces InR autophosphorylation (Suresh et al., 2021). In our experiment, we observed higher InR activity in CG5399- overexpressing cells without insulin stimulation, suggesting that InR activation can be regulated by the lipid environment. The clinical relevance of this InR activation mech- anism needs to be further investigated. In conclusion, we have established a genome- wide CRISPRa platform in Drosophila cells and iden- tified novel rapamycin resistance genes using a genome- wide CRISPRa screen platform. This platform can be applied broadly to help elucidate the cellular mechanisms of a variety of biological processes. Materials and methods Key resources table Reagent type (species) or resource Designation Gene (Drosophila melanogaster) Gene (Drosophila melanogaster) Gene (Drosophila melanogaster) CG8468 CG5399 CG9932 Cell line (D. melanogaster) S2R+ Cell line (D. melanogaster) PT5 Strain, strain background (Escherichia coli) Continued on next page E.cloni10GF’ Electrocompetent Cells Source or reference Identifiers Additional information FlyBase FLYB:FBgn0033913 FlyBase FLYB:FBgn0038353 FlyBase DRSC DRSC Biosearch Technologies FLYB:FBgn0262160 FLYB:FBtc0000150 FLYB:FBtc0000229 60061–2 sgRNA library construction Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 11 of 18 Genetics and Genomics Research article Continued Reagent type (species) or resource Designation Source or reference Identifiers Additional information Strain, strain background (Escherichia coli) One Shot TOP10 Chemically Competent E. coli Invitrogen C404010 Recombinant Anti- Insulin Receptor (phospho Y1185) antibody (Rabbit monoclonal) Abcam ab62321 1:1000 for WB Phospho- Akt (Ser473) (D9E) XP antibody (Rabbit monoclonal) Cell Signaling Technology Akt Rabbit Antibody (Rabbit polyclonal) Cell Signaling Technology StarBright Blue 700 Goat Anti- Rabbit IgG Bio- Rad StarBright Blue 520 Goat Anti- Rabbit IgG Bio- Rad 4060 9272 12004161 12005869 1:1000 for WB 1:1000 for WB 1:2500 for WB 1:2500 for WB Antibody Antibody Antibody Antibody Antibody Antibody hFAB Rhodamine Anti- Actin Primary Antibody (synthesized, monoclonal) Recombinant DNA reagent pMK33- SAM plasmid Recombinant DNA reagent pLib8 plasmid Recombinant DNA reagent pBS130 plasmid Bio- Rad This paper This paper Addgene 12004163 1:2500 for WB Can be obtained from DRSC U6:3- MS2 sgRNA cassette, can be obtained from DRSC 26290 PhiC31 integrase CG5399 ORF vector, cassette, can be obtained from DRSC Recombinant DNA reagent pUAS- CG5399 plasmid This paper Commercial assay or kit Effectene Transfection Reagent Qiagen 301425 Commercial assay or kit CellTiter- Glo Luminescent Cell Viability Assay Commercial assay or kit RNeasy Mini Kit Commercial assay or kit iScript cDNA Synthesis Kit Promega Qiagen Bio- Rad Chemical compound, drug MEGAscript T7 Transcription Kit Invitrogen Chemical compound, drug Methyl-β-cyclodextrin Chemical compound, drug Cholesterol Chemical compound, drug Cholesterol Software, algorithm GraphPad Prism 7 Software, algorithm FlowJo Sigma- Aldrich Sigma- Aldrich Sigma- Aldrich GraphPad FlowJo G7570 74104 1708890 AM1334 C4555 C3045 C2044 Vectors The pMK33- SAM plasmid was generated by transferring the SAM sequence from the flySAM vector (Jia et al., 2018) into the pMK33 plasmid. MS2 hairpin containing sgRNA was expressed from the pLib8 plasmid. pLib8 is derived from pLib6.4 (Viswanatha et al., 2018) by replacing the U6:2- sgRNA cassette to the U6:3- MS2 sgRNA cassette. The PhiC31 integrase expressing pBS130 plasmid was obtained from Addgene (#26290). The full length of CG5399 was cloned from the cDNA of S2R+ cells and inserted into the pWalium10 vector (DGRC, 1470) to construct the pUAS- CG5399 vector. sgRNA sequences used in this study are listed in Supplementary file 1. Antibodies Phospho- InR antibody (Abcam, #ab62321), Phospho- Akt antibody (Cell signaling, # 4060), Akt anti- body (Cell signaling, #9272) were used in this study. Phospho- S6 antibody is a kind gift of Kim and Choi, 2019. Cell culture, transfection, and proliferation assay Drosophila cells were cultured with Schneider medium (Gibco) supplemented with 10% heat- inactivated FBS (Gibco) at 25°C (Viswanatha et  al., 2018) unless otherwise indicated. The Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 12 of 18 Genetics and Genomics Research article wild- type S2R+ cell line and the attP sites containing S2R+ derivative PT5 cell line were obtained from the Drosophila RNAi Screening Center. The DRSC copy of S2R+ was authenticated by the Drosophila Genomics Resource Center (DGRC), as part of their development of a transposable element- based authentication protocol for Drosophila cell lines (Mariyappa et  al., 2022). The presence of the recombination cassette was confirmed by observation of mCherry fluorescence and the successful introduction of sgRNAs via recombination- mediated cassette exchange. Myco- plasma contamination is not an issue for Drosophila cell lines; throughout the study, we monitored and confirmed through careful observation that media and cells were not infected by bacteria or fungi. The cell lines used in this study are not from the list of commonly misidentified cell lines maintained by the International Cell Line Authentication Committee. PT5 cells were transfected with pMK33- SAM plasmid using Effectene (QIAGEN) according to the manufacturer’s protocol. Briefly, 3 × 106 Drosophila cells were seeded into one well of a six- well plate before transfection. 400  ng plasmids were diluted into Buffer EC to a final volume of 100  μL and mixed with 3.2  μL enhancer by vortexing to form a DNA- enhancer mixture. 10  μL Effectene transfection reagents were added into the DNA- enhancer mixture and mixed by vortexing. After incubating at room temperature for 15 min to allow transfection complex formation, the solution was added drop- wise onto Drosophila cells. Transfected cells were passaged in a culture medium containing 200 μg/mL Hygromycin B (Millipore) for 1 month to generate the stable SAM cell line. To induce SAM complex expression, a culture medium containing 100 μM CuSO4 (Sigma) was used. Cell proliferation under different rapamycin concentrations was tested using CellTiter Glo assay (Promega) according to the manufacturer’s protocol. 1 × 104 Drosophila cells were seeded into each well of a 96- well plate. Rapamycin- containing culture medium was added into each well to make the final concentration from 10–4 nM to 10 nM. After culturing for 4 days, a volume of CellTiter Glo reagent was added into each well before cells reached confluence. The luminescence signal was measured by Plate Reader (Molecular Devices). Pooled library design and construction For the focused sgRNA library, sgRNAs were designed within 500 bp upstream of the transcriptional start site (TSS) for each gene. Ten different sgRNAs were selected for each gene unless fewer sgRNA binding sites were found within the window. Constructing the focused library was performed as previously described (Viswanatha et al., 2018). Briefly, Bbs1 sites flanking sgRNA spacer sequences were synthesized as single- stranded DNA oligos (Agilent). DNA oligos were amplified by PCR using Phusion Polymerase (New England Biolabs). Bbs1 restriction enzyme (New England Biolabs) was used to digest amplicon and pLib8 plasmid. The resulting 24- mer fragment was purified from the Bbs1 digested amplicon by running a 20% TBE polyacrylamide gel (Thermo). The purified fragment and plasmid were ligated using T4 ligase (New England Biolabs). The ligation products were transferred into Ecloni 10GF’ electrocompetent cells (Lucigen) using Gene Pulser Xcell Electroporation Systems (Bio- Rad). Transformed bacteria were spread on LB- carbenicillin agar plates. After overnight culture, the bacteria colonies were collected from plates by scraping and amplified in an LB medium with ampicillin. For the genome- wide dual- sgRNA library, sgRNAs were designed within 500 bp upstream of TSS for each gene. sgRNAs were chosen to make six dual- sgRNA combinations for each gene unless fewer sgRNA binding sites were found. To construct the dual- sgRNA library, Bbs1, and BsmB1 sites flanking two sgRNA spacer sequences were synthesized in a custom array as single- stranded DNA oligos (Agilent). DNA oligos were amplified by PCR using Phusion Polymerase (New England Biolabs). DNA amplicons were ligated to Zero Blunt vector (Thermo) using T4 ligase to generate the first library. An amplicon containing the scaffold sequence for the first sgRNA and U6:2 promoter sequence for the second sgRNA was inserted into the BsmB1 site to generate the second library. The second library and pLib8 vector were digested with Bbs1. The resulting sgRNA cassettes from the digested second library were ligated with Bbs1 digested pLib8 vector to generate the final library using T4 ligase. Each library was transferred into Ecloni 10GF’ electrocompetent cells by electro- poration. Transformed bacteria were spread on LB- carbenicillin agar plates. Bacteria colonies were calculated by serial dilution. Each library needs to reach at least 10 times diversity to maintain the integrity of the library. Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 13 of 18 Genetics and Genomics Research article Pooled screening, library sequencing, and data analysis The library was co- transfected with the same amount of phiC31 plasmid into SAM cells at 3 × 106 cells/well of a six- well plate using Effectene. The total cell number used for transfection was calcu- lated to ensure over 1000 cells/sgRNA to maintain the integrity of the library. The transfected cells were passaged in a culture medium containing 5 μg/mL puromycin for 3 weeks to select the sgRNA- integrated cells. The resulting pooled library cells were split into two populations and passaged in rapamycin or DMSO- containing medium supplemented with 100 μM CuSO4 for indicated days. After treatment, the genomic DNA was extracted from the final cell population and the sgRNAs sequences were amplified by PCR. As each Drosophila cell contains ~0.6 pg DNA, the amount of genomic DNA used as a PCR template was calculated to ensure over 1000 cells/sgRNA to maintain the diversity. The library for next- generation sequencing was constructed by adding Illumina adaptors to sgRNA amplicons by PCR. The final PCR products had the following sequence: P5- read1- (N)n- (B)6- sgRNA- P7 (where N stands for any nucleotide, n stands for a variable length of nucleotide from 1 to 10, and (B)6 stands for six nucleotides sample barcode). PCR primers used for NGS library construction are listed in Supplementary file 2. The library was sequenced using the NextSeq500 1 × 75 SE platform (Illumina) in HMS Biopolymers Facility. The sequencing data were de- multiplexed using TagDust. Screen hits were identified using MAGeCK- RRA by comparing the treatment and the control according to the previous report (Li et al., 2014). RNA extraction, reverse transcription, and qPCR RNA was extracted from Drosophila cells using RNA Mini Kit (QIAGEN) according to the manufac- turer’s protocol. Total RNA was reverse transcribed into cDNA using the iScript cDNA Synthesis Kit (Bio- Rad). qPCR was done with SYBR Green Master Mix (Bio- Rad). The housekeeping gene rp49 was used as the reference gene for qPCR. qPCR primers used in this study are listed in Supplementary file 2. The statistical analysis was performed using the GraphPad Prism 7 software. t- tests were performed to test the significance of gene expression data. *p<0.05; **p<0.01; ***p<0.001; ns, not significant. GFP proportion analysis by flow cytometry Cells were transferred into 5  mL FACS tubes (Falcon 352235) and analyzed with a BD LSR II Flow Cytometer in the Department of Immunology Flow Cytometry Facility, Harvard Medical School. Wild- type cells (GFP negative cells) and an empty vector expressing cells (GFP positive cells) were used as a negative and positive control to set the gate in Alexa Fluor 488 channel, respectively. With this gate, the GFP proportion in wild- type cells is 0.031%, and in an empty vector expressing cells is 96.7%. Three biological replicates for each condition were tested. The flow cytometry files were analyzed by FlowJo. The statistical analysis was performed using the GraphPad Prism 7 software. t- tests were performed to test the significance of GFP proportion data. *p<0.05; **p<0.01; ***p<0.001; ns, not significant. dsRNA synthesis and transfection dsRNAs were designed by the Drosophila RNAi Screening Center. dsRNA templates were amplified from genomic DNA using primers with T7 promoter sequence TAAT ACGA CTCA CTAT AGGG at 5' end. dsRNAs were synthesized from the resulting amplicons using MEGAscript T7 Transcription Kit (Invit- rogen). dsRNAs were purified with RNeasy Mini Kit (QIAGEN) before transfection. dsRNA sequences used in this study are listed in Supplementary file 3. 10  μg dsRNA were transfected into 3 × 106 Drosophila cells using Effectene (QIAGEN). Cholesterol depletion and supplementation For cholesterol depletion, methyl- beta- cyclodextrin (sigma) was dissolved in serum- free Schneider medium at the indicated concentration. Cells were incubated with MβCD containing serum- free medium for 1 hr before testing. For cholesterol supplementation, cholesterol (sigma) was dissolved in ethanol. Dissolved cholesterol was added into MβCD containing serum- free Schneider medium to form cholesterol/MβCD complex. Cells were incubated with cholesterol/MβCD complex containing serum- free Schneider medium for 1 hr before testing. Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 14 of 18 Genetics and Genomics Research article Acknowledgements We thank Dr. Jianquan Ni for the flySAM vector and Dr. Ah- Ram Kim for the phospho- S6 antibody. This work is supported by NIH NIGMS P41 GM132087 (NP) and NIH P01CA120964. NP is an investigator at Howard Hughes Medical Institute. This article is subject to HHMI’s Open Access to Publications policy. HHMI lab heads have previously granted a nonexclusive CC BY 4.0 license to the public and a sublicensable license to HHMI in their research articles. Pursuant to those licenses, the author- accepted manuscript of this article can be made freely available under a CC BY 4.0 license immediately upon publication. Additional information Funding Funder Grant reference number Author National Institute of General Medical Sciences GM132087 National Cancer Institute CA120964 Howard Hughes Medical Institute Stephanie E Mohr Norbert Perrimon Norbert Perrimon Norbert Perrimon NIH P01CA120964 Norbert Perrimon The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Author contributions Baolong Xia, Conceptualization, Formal analysis, Investigation, Writing - original draft, Project admin- istration; Raghuvir Viswanatha, Yanhui Hu, Formal analysis; Stephanie E Mohr, Funding acquisition, Writing – review and editing; Norbert Perrimon, Conceptualization, Supervision, Funding acquisition, Writing – review and editing Author ORCIDs Baolong Xia Raghuvir Viswanatha Stephanie E Mohr Norbert Perrimon http://orcid.org/0000-0003-2536-0267 http://orcid.org/0000-0002-9457-6953 http://orcid.org/0000-0001-9639-7708 http://orcid.org/0000-0001-7542-472X Decision letter and Author response Decision letter https://doi.org/10.7554/eLife.85542.sa1 Author response https://doi.org/10.7554/eLife.85542.sa2 Additional files Supplementary files • Supplementary file 1. sgRNA vectors used in this study. • Supplementary file 2. PCR primers used in this study. • Supplementary file 3. dsRNAs used in this study. • MDAR checklist • Source data 1. The original files of the full raw unedited blots and figures with the uncropped blots with relevant bands labeled in this study. Data availability All data generated or analysed during this study are included in the manuscript and source data files. The pMK33- SAM vector, pLib8 vector, and libraries used in this study are available through DRSC/TRiP Functional Genomics Resources. Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 15 of 18 Genetics and Genomics Research article The following dataset was generated: Author(s) Xia B, Viswanatha R, Hu Y, Mohr SE, Perrimon N Year 2023 Dataset title Dataset URL Database and Identifier Data from: Pooled genome- wide CRISPR activation screening for rapamycin resistance genes in Drosophila cells https:// dx. doi. org/ 10. 5061/ dryad. 2547d7ww8 Dryad Digital Repository, 10.5061/dryad.2547d7ww8 References Bard F, Casano L, Mallabiabarrena A, Wallace E, Saito K, Kitayama H, Guizzunti G, Hu Y, Wendler F, Dasgupta R, Perrimon N, Malhotra V. 2006. Functional genomics reveals genes involved in protein secretion and golgi organization. Nature 439:604–607. DOI: https://doi.org/10.1038/nature04377, PMID: 16452979 Björklund M, Taipale M, Varjosalo M, Saharinen J, Lahdenperä J, Taipale J. 2006. Identification of pathways regulating cell size and cell- cycle progression by rnai. Nature 439:1009–1013. DOI: https://doi.org/10.1038/ nature04469, PMID: 16496002 Boutros M, Kiger AA, Armknecht S, Kerr K, Hild M, Koch B, Haas SA, Paro R, Perrimon N, Heidelberg Fly Array Consortium. 2004. Genome- wide rnai analysis of growth and viability in Drosophila cells. Science 303:832–835. DOI: https://doi.org/10.1126/science.1091266, PMID: 14764878 Breustedt DA, Korndörfer IP, Redl B, Skerra A. 2005. The 1.8- A crystal structure of human tear lipocalin reveals an extended branched cavity with capacity for multiple ligands. The Journal of Biological Chemistry 280:484– 493. DOI: https://doi.org/10.1074/jbc.M410466200, PMID: 15489503 Chavez A, Tuttle M, Pruitt BW, Ewen- Campen B, Chari R, Ter- Ovanesyan D, Haque SJ, Cecchi RJ, Kowal EJK, Buchthal J, Housden BE, Perrimon N, Collins JJ, Church G. 2016. Comparison of cas9 activators in multiple species. Nature Methods 13:563–567. DOI: https://doi.org/10.1038/nmeth.3871, PMID: 27214048 Chen RQ, Yang QK, Lu BW, Yi W, Cantin G, Chen YL, Fearns C, Yates JR, Lee JD. 2009. Cdc25B mediates rapamycin- induced oncogenic responses in cancer cells. Cancer Research 69:2663–2668. DOI: https://doi.org/ 10.1158/0008-5472.CAN-08-3222, PMID: 19276368 Corradetti MN, Inoki K, Guan KL. 2005. The stress- inducted proteins RTP801 and RTP801L are negative regulators of the mammalian target of rapamycin pathway. The Journal of Biological Chemistry 280:9769– 9772. DOI: https://doi.org/10.1074/jbc.C400557200, PMID: 15632201 Cutler NS, Pan X, Heitman J, Cardenas ME. 2001. The TOR signal transduction cascade controls cellular differentiation in response to nutrients. Molecular Biology of the Cell 12:4103–4113. DOI: https://doi.org/10. 1091/mbc.12.12.4103, PMID: 11739804 D’Ambrosio MV, Vale RD. 2010. A whole genome rnai screen of Drosophila S2 cell spreading performed using automated computational image analysis. The Journal of Cell Biology 191:471–478. DOI: https://doi.org/10. 1083/jcb.201003135, PMID: 21041442 Dobson THW, Tao R- H, Swaminathan J, Maegawa S, Shaik S, Bravo- Alegria J, Sharma A, Kennis B, Yang Y, Callegari K, Haltom AR, Taylor P, Kogiso M, Qi L, Khatua S, Goldman S, Lulla RR, Fangusaro J, MacDonald TJ, Li X- N, et al. 2019. Transcriptional repressor rest drives lineage stage- specific chromatin compaction at PTCH1 and increases Akt activation in a mouse model of medulloblastoma. Science Signaling 12:565. DOI: https://doi. org/10.1126/scisignal.aan8680, PMID: 30670636 Ewen- Campen B, Mohr SE, Hu Y, Perrimon N. 2017. Accessing the phenotype gap: enabling systematic investigation of paralog functional complexity with CRISPR. Developmental Cell 43:6–9. DOI: https://doi.org/ 10.1016/j.devcel.2017.09.020, PMID: 29017030 Guo Y, Walther TC, Rao M, Stuurman N, Goshima G, Terayama K, Wong JS, Vale RD, Walter P, Farese RV. 2008. Functional genomic screen reveals genes involved in lipid- droplet formation and utilization. Nature 453:657– 661. DOI: https://doi.org/10.1038/nature06928, PMID: 18408709 Gutmann T, Kim KH, Grzybek M, Walz T, Coskun Ü. 2018. Visualization of ligand- induced transmembrane signaling in the full- length human insulin receptor. The Journal of Cell Biology 217:1643–1649. DOI: https:// doi.org/10.1083/jcb.201711047, PMID: 29453311 Hao L, Sakurai A, Watanabe T, Sorensen E, Nidom CA, Newton MA, Ahlquist P, Kawaoka Y. 2008. Drosophila rnai screen identifies host genes important for influenza virus replication. Nature 454:890–893. DOI: https://doi. org/10.1038/nature07151, PMID: 18615016 Jia Y, Xu R- G, Ren X, Ewen- Campen B, Rajakumar R, Zirin J, Yang- Zhou D, Zhu R, Wang F, Mao D, Peng P, Qiao H- H, Wang X, Liu L- P, Xu B, Ji J- Y, Liu Q, Sun J, Perrimon N, Ni J- Q. 2018. Next- generation CRISPR/cas9 transcriptional activation in Drosophila using flysam. PNAS 115:4719–4724. DOI: https://doi.org/10.1073/pnas. 1800677115, PMID: 29666231 Kim AR, Choi KW. 2019. Tric/Cct chaperonins are essential for organ growth by interacting with insulin/TOR signaling in Drosophila. Oncogene 38:4739–4754. DOI: https://doi.org/10.1038/s41388-019-0754-1, PMID: 30792539 Konermann S, Brigham MD, Trevino AE, Joung J, Abudayyeh OO, Barcena C, Hsu PD, Habib N, Gootenberg JS, Nishimasu H, Nureki O, Zhang F. 2015. Genome- scale transcriptional activation by an engineered CRISPR- cas9 complex. Nature 517:583–588. DOI: https://doi.org/10.1038/nature14136, PMID: 25494202 Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 16 of 18 Genetics and Genomics Research article Lakshmi B, Mishra M, Srinivasan N, Archunan G. 2015. Structure- Based phylogenetic analysis of the lipocalin superfamily. PLOS ONE 10:e0135507. DOI: https://doi.org/10.1371/journal.pone.0135507, PMID: 26263546 Li W, Xu H, Xiao T, Cong L, Love MI, Zhang F, Irizarry RA, Liu JS, Brown M, Liu XS. 2014. MAGeCK enables robust identification of essential genes from genome- scale CRISPR/Cas9 knockout screens. Genome Biology 15:554. DOI: https://doi.org/10.1186/s13059-014-0554-4, PMID: 25476604 Lu SM, Fairn GD. 2018. Mesoscale organization of domains in the plasma membrane- beyond the lipid raft. Critical Reviews in Biochemistry and Molecular Biology 53:192–207. DOI: https://doi.org/10.1080/10409238. 2018.1436515, PMID: 29457544 Mao D, Jia Y, Peng P, Shen D, Ren X, Zhu R, Qiu Y, Han Y, Yu J, Che Q, Li Y, Lu X, Liu L- P, Wang Z, Liu Q, Sun J, Ni J- Q. 2020. Enhanced efficiency of flysam by optimization of sgRNA parameters in Drosophila . G3: Genes, Genomes, Genetics 10:4483–4488. DOI: https://doi.org/10.1534/g3.120.401614 Mariyappa D, Rusch DB, Han S, Luhur A, Overton D, Miller DFB, Bergman CM, Zelhof AC. 2022. A novel transposable element- based authentication protocol for Drosophila cell lines. G3: Genes, Genomes, Genetics 12:jkab403. DOI: https://doi.org/10.1093/g3journal/jkab403, PMID: 34849844 Neumüller RA, Wirtz- Peitz F, Lee S, Kwon Y, Buckner M, Hoskins RA, Venken KJT, Bellen HJ, Mohr SE, Perrimon N. 2012. Stringent analysis of gene function and protein- protein interactions using fluorescently tagged genes. Genetics 190:931–940. DOI: https://doi.org/10.1534/genetics.111.136465, PMID: 22174071 Reikvam H, Tamburini J, Skrede S, Holdhus R, Poulain L, Ersvaer E, Hatfield KJ, Bruserud Ø. 2014. Antileukaemic effect of PI3K- mTOR inhibitors in acute myeloid leukaemia- gene expression profiles reveal Cdc25B expression as determinate of pharmacological effect. British Journal of Haematology 164:200–211. DOI: https://doi.org/ 10.1111/bjh.12611, PMID: 24383842 Reiling JH, Hafen E. 2004. The hypoxia- induced paralogs Scylla and Charybdis inhibit growth by down- regulating S6K activity upstream of TSC in Drosophila. Genes & Development 18:2879–2892. DOI: https://doi. org/10.1101/gad.322704, PMID: 15545626 Replogle JM, Norman TM, Xu A, Hussmann JA, Chen J, Cogan JZ, Meer EJ, Terry JM, Riordan DP, Srinivas N, Fiddes IT, Arthur JG, Alvarado LJ, Pfeiffer KA, Mikkelsen TS, Weissman JS, Adamson B. 2020. Combinatorial single- cell CRISPR screens by direct guide RNA capture and targeted sequencing. Nature Biotechnology 38:954–961. DOI: https://doi.org/10.1038/s41587-020-0470-y, PMID: 32231336 Rodríguez Y, Unno K, Truica MI, Chalmers ZR, Yoo YA, Vatapalli R, Sagar V, Yu J, Lysy B, Hussain M, Han H, Abdulkadir SA. 2022. A genome- wide CRISPR activation screen identifies PRRX2 as A regulator of enzalutamide resistance in prostate cancer. Cancer Research 82:2110–2123. DOI: https://doi.org/10.1158/ 0008-5472.CAN-21-3565, PMID: 35405009 Sanjana NE, Shalem O, Zhang F. 2014. Improved vectors and genome- wide libraries for CRISPR screening. Nature Methods 11:783–784. DOI: https://doi.org/10.1038/nmeth.3047, PMID: 25075903 Scapin G, Dandey VP, Zhang Z, Prosise W, Hruza A, Kelly T, Mayhood T, Strickland C, Potter CS, Carragher B. 2018. Structure of the insulin receptor- insulin complex by single- particle cryo- EM analysis. Nature 556:122– 125. DOI: https://doi.org/10.1038/nature26153, PMID: 29512653 Schmelzle T, Beck T, Martin DE, Hall MN. 2004. Activation of the RAS/cyclic AMP pathway suppresses a TOR deficiency in yeast. Molecular and Cellular Biology 24:338–351. DOI: https://doi.org/10.1128/MCB.24.1.338- 351.2004, PMID: 14673167 Scuderi A, Simin K, Kazuko SG, Metherall JE, Letsou A. 2006. Scylla and charybde, homologues of the human apoptotic gene RTP801, are required for head involution in Drosophila. Developmental Biology 291:110–122. DOI: https://doi.org/10.1016/j.ydbio.2005.12.014, PMID: 16423342 Sofer S, Lamkiewicz K, Armoza Eilat S, Partouche S, Marz M, Moskovits N, Stemmer SM, Shlomai A, Sklan EH. 2022. A genome- wide CRISPR activation screen reveals hexokinase 1 as a critical factor in promoting resistance to multi- kinase inhibitors in hepatocellular carcinoma cells. FASEB Journal 36:e22191. DOI: https://doi.org/10. 1096/fj.202101507RR, PMID: 35147243 Song J, Mizrak A, Lee CW, Cicconet M, Lai ZW, Tang WC, Lu CH, Mohr SE, Farese RV, Walther TC. 2022. Identification of two pathways mediating protein targeting from ER to lipid droplets. Nature Cell Biology 24:1364–1377. DOI: https://doi.org/10.1038/s41556-022-00974-0, PMID: 36050470 Sopko R, Lin YB, Makhijani K, Alexander B, Perrimon N, Brückner K. 2015. A systems- level interrogation identifies regulators of Drosophila blood cell number and survival. PLOS Genetics 11:e1005056. DOI: https:// doi.org/10.1371/journal.pgen.1005056, PMID: 25749252 Stapleton M, Carlson J, Brokstein P, Yu C, Champe M, George R, Guarin H, Kronmiller B, Pacleb J, Park S, Wan K, Rubin GM, Celniker SE. 2002. A Drosophila full- length cdna resource. Genome Biology 3:RESEARCH0080. DOI: https://doi.org/10.1186/gb-2002-3-12-research0080, PMID: 12537569 Suresh P, Miller WT, London E. 2021. Phospholipid exchange shows insulin receptor activity is supported by both the propensity to form wide bilayers and ordered raft domains. The Journal of Biological Chemistry 297:101010. DOI: https://doi.org/10.1016/j.jbc.2021.101010, PMID: 34324831 Tu C, Fiandalo MV, Pop E, Stocking JJ, Azabdaftari G, Li J, Wei H, Ma D, Qu J, Mohler JL, Tang L, Wu Y. 2018. Proteomic analysis of charcoal- stripped fetal bovine serum reveals changes in the insulin- like growth factor signaling pathway. Journal of Proteome Research 17:2963–2977. DOI: https://doi.org/10.1021/acs.jproteome. 8b00135 Viswanatha R, Li Z, Hu Y, Perrimon N. 2018. Pooled genome- wide CRISPR screening for basal and context- specific fitness gene essentiality in Drosophila cells. eLife 7:e36333. DOI: https://doi.org/10.7554/eLife.36333, PMID: 30051818 Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 17 of 18 Genetics and Genomics Research article Walton RT, Christie KA, Whittaker MN, Kleinstiver BP. 2020. Unconstrained genome targeting with near- pamless engineered CRISPR- Cas9 variants. Science 368:290–296. DOI: https://doi.org/10.1126/science.aba8853, PMID: 32217751 Yang L, Yu J. 2009. A comparative analysis of divergently- paired genes (DPGs) among Drosophila and vertebrate genomes. BMC Evolutionary Biology 9:55. DOI: https://doi.org/10.1186/1471-2148-9-55, PMID: 19284596 Yu C, Wan KH, Hammonds AS, Stapleton M, Carlson JW, Celniker SE. 2011. Development of expression- ready constructs for generation of proteomic libraries. Methods in Molecular Biology 723:257–272. DOI: https://doi. org/10.1007/978-1-61779-043-0_17, PMID: 21370071 Zhang Q, Wang S, Li W, Yau E, Hui H, Singh PK, Achuthan V, Young Karris MA, Engelman AN, Rana TM. 2022. Genome- wide crispr/cas9 transcriptional activation screen identifies a histone acetyltransferase inhibitor complex as a regulator of hiv- 1 integration. Nucleic Acids Research 50:6687–6701. DOI: https://doi.org/10. 1093/nar/gkac464, PMID: 35713529 Zhu S, Liu Y, Zhou Z, Zhang Z, Xiao X, Liu Z, Chen A, Dong X, Tian F, Chen S, Xu Y, Wang C, Li Q, Niu X, Pan Q, Du S, Xiao J, Wang J, Wei W. 2022. Genome- Wide CRISPR activation screen identifies candidate receptors for SARS- cov- 2 entry. Science China Life Sciences 65:701–717. DOI: https://doi.org/10.1007/s11427-021-1990-5 Zidovetzki R, Levitan I. 2007. Use of cyclodextrins to manipulate plasma membrane cholesterol content: evidence, misconceptions and control strategies. Biochimica et Biophysica Acta 1768:1311–1324. DOI: https:// doi.org/10.1016/j.bbamem.2007.03.026, PMID: 17493580 Zurita- Martinez SA, Cardenas ME. 2005. Tor and cyclic AMP- protein kinase A: two parallel pathways regulating expression of genes required for cell growth. Eukaryotic Cell 4:63–71. DOI: https://doi.org/10.1128/EC.4.1.63- 71.2005, PMID: 15643061 Xia et al. eLife 2023;12:e85542. DOI: https://doi.org/10.7554/eLife.85542 18 of 18 Genetics and Genomics
10.1039_d2sc04160j
Showcasing research from Professors Leman’s and Krishnamurthy’s laboratory, Department of Chemistry, The Scripps Research Institute, California, USA. One-pot chemical pyro- and tri-phosphorylation of peptides by using diamidophosphate in water Pyrophosphopeptides are produced in good yields from diamidophosphate mediated phosphorylation of phosphopeptides followed by hydrolysis of the resulting amidopyrophosphate. The overall reaction proceeds effi ciently in water-ice medium without the need for protecting groups. The simplicity of this process enables a repetitive phosphorylation-hydrolysis sequence of reactions in a single pot that produces triphosphopeptides effi ciently. The potential for continual reiteration of this sequence of reactions suggests that it can become a practical synthetic tool for protein polyphosphorylation. As featured in: See Huacan Lin, Luke J. Leman, Ramanarayanan Krishnamurthy, Chem. Sci., 2022, 13, 13741. rsc.li/chemical-science Registered charity number: 207890 Chemical Science EDGE ARTICLE Cite this: Chem. Sci., 2022, 13, 13741 All publication charges for this article have been paid for by the Royal Society of Chemistry Received 26th July 2022 Accepted 30th October 2022 DOI: 10.1039/d2sc04160j One-pot chemical pyro- and tri-phosphorylation of peptides by using diamidophosphate in water† Huacan Lin, * Luke J. Leman * and Ramanarayanan Krishnamurthy * Protein (pyro)phosphorylation is emerging as a post-translational modification (PTM) in signalling pathways involved in many cellular processes. However, access to synthetic pyrophosphopeptides that can serve as tools for understanding protein pyrophosphorylation is quite limited. Herein, we report a chemical phosphorylation method that enables the synthesis of pyrophosphopeptides in aqueous medium without the need for protecting groups. The strategy employs diamidophosphate (DAP) in a one-pot sequential phosphorylation-hydrolysis of mono-phosphorylated peptide precursors. This operationally simple method exploits the intrinsic nucleophilicity of a phosphate moiety installed on serine, threonine or tyrosine residues in complex peptides with excellent chemoselectivity and good yields under mild conditions. We demonstrate the installation of the pyrophosphate group within a wide range of model peptides and showcase the potential of this methodology by selectively pyrophosphorylating the highly functionalized Nopp140 peptide fragment. The potential to produce higher (poly)phosphorylated peptides was demonstrated as a proof-of-principle experiment where we synthesized the triphosphorylated peptides using this rsc.li/chemical-science one-pot strategy. Introduction Post-translational modications (PTMs) play signicant roles in modulating protein function and activity.1 Among a broad range of PTMs,2–6 protein phosphorylation arguably is the most widely studied PTM and plays a crucial role in regulating many other cellular processes. Since its discovery in 1959,7 phos- phorylation has been closely linked with many cellular func- tions in the cell cycle,8 apoptosis,9 differentiation10 and others. Various techniques for enrichment of phosphorylated proteins as well as their identication and quantitation are accessible today for the analysis of phosphorylation.11 In contrast to protein phosphorylation, protein pyrophosphorylation is a poorly characterized PTM. It is known to be mediated by a group of second messengers termed inositol pyrophosphates (PP-InsPs)12 which are capable of transferring the beta- phosphoryl group from 5PP-InsP5 in the presence of Mg2+ to a phosphorylated serine residue in protein substrates to generate a pyrophosphorylated protein (Fig. 1a).13,14 Some signicant advancements have been made in understanding the important roles that inositol pyrophosphates play in various pathways.15–18 Even so, the biological functions of protein pyrophosphorylation remain largely unknown due to the Department of Chemistry, The Scripps Research Institute, La Jolla, California 92037, USA. E-mail: hulin@scripps.edu; lleman@scripps.edu; rkrishna@scripps.edu † Electronic supplementary information (ESI) available: Experimental procedures, synthetic DOI: schemes, https://doi.org/10.1039/d2sc04160j and MS spectra. traces HPLC See in identifying pyrophosphorylated proteins difficulty in complex cellular contexts and the few efficient methodologies for chemically synthesizing pyrophosphorylated peptides and proteins in vitro. Thus, there is a need to develop new methods and expand the toolbox for accessible peptide and protein pyrophosphorylation. In 2014, Fiedler and co-workers reported a synthetic methodology to access pyrophosphorylated peptides and proteins from phosphorylated precursors (Fig. 1b).19,20 This protocol involves pyrophosphorylation of a phosphoserine residue by using a benzyl phosphorimidazolide reagent in water/dimethylacetamide co-solvent as the rst step, followed by hydrogenolysis of the isolated benzyl-protected interme- diates in water/dimethylformamide co-solvent to provide the pyrophosphopeptides. This protocol also proceeded in water but with slower reaction kinetics. Surprisingly, there are no other methods available that allow easy access to pyrophos- phorylated peptides, despite the considerable number of synthetic approaches small molecules such as nucleosides.21–24 Here, we present a simple one-pot chemical strategy for the facile synthesis of pyro- phosphopeptides. Our approach starts from a peptide precursor bearing a phosphorylated Ser, Thr or Tyr residue, which site-selectively reacts with diamidophosphate (DAP) in water to generate an amidopyrophosphorylated intermediate, followed by nitrous acid-induced hydrolysis to afford the ex- pected pyrophosphopeptide (Fig. 1c). Distinct from the existing method, our approach does not require protecting/ for polyphosphorylation of © 2022 The Author(s). Published by the Royal Society of Chemistry Chem. Sci., 2022, 13, 13741–13747 | 13741 Chemical Science Edge Article Fig. 1 Site-selective peptide pyrophosphorylation via a one-pot sequential amidophosphorylation-hydrolysis methodology. (a) Inositol pyro- phosphate messenger-mediated protein pyrophosphorylation in the presence of magnesium. (b) Chemical pyrophosphorylation of peptides using phosphorimidazolide reagents.20 (c) Diamidophosphate-mediated site-selective pyrophosphorylation of peptides. The use of dia- midophosphate in aqueous medium, presented in this work, offers several advantages. Circles represent amino acids; DMA, dimethylacetamide; DMF, dimethylformamide. deprotecting group chemistry, or isolation of peptide inter- mediates. Furthermore, this higher yielding one-pot trans- formation displays for phosphate moieties over the other nucleophilic peptide side chains, and wide applicability for accessing various synthetic and native pyrophosphopeptides. selectivity excellent Results and discussion ′ ′ -NMPs) and 5 ′ -nucleoside monophosphates (5 Previous work from our group showed that DAP, which was prepared via saponication of phenyl phosphorodiamidate, is a versatile phosphorylating agent.25 Particularly, DAP reacted ′ with 5 -nucleo- ′ ′ -NDPs) to generate the corresponding 5 - side diphosphates (5 ′ amidodiphosphate and 5 -amidotriphosphate derivatives, ′ respectively, and converted 2 -ribonucleotides to the corre- ,3 ′ ′ -cyclic-phosphates.26 Recently, we showed that sponding 2 ′ ′ DAP enables the conversion of 5 -oligonucleotide -NMPs and 5 monophosphates into the corresponding nucleoside triphos- ′ phates (5 -oligonucleotide triphosphates via ami- dophosphates in a one-pot amidophosphorylation-hydrolysis setting in water.27 Inspired by these observations, we reasoned that phosphopeptides might similar amidophosphorylation-hydrolysis to afford the corresponding pyrophosphopeptides in one-pot,28,29 considering the plausible ′ -NTPs) and 5 also undergo ,3 a To 1 mM peptide reactive groups in the side chains (e.g., lysine, aspartate, gluta- mate, cysteine, etc.). this test hypothesis, 1 ( , phosphoserine indicated in bold and red, prepared by the incorporation of commercially available Fmoc-O-benzyl-phosphoserine, ESI Fig. S2†),30 was treated with an excess of DAP (30 equiv.), MgCl2 (10 equiv.) and imidazole (10 equiv.) at pH 5.5 in water at room temperature, 45 °C or −20 °C (in a freezer). The formation of the amido- pyrophosphopeptide 19 in the crude reaction mixture was monitored by reverse-phase high-performance liquid chro- matography (HPLC) and mass spectrometry (MS). Imidazole acts as a catalyst to increase the efficiency of the amidophos- phorylation reaction of the phosphopeptide by forming the amidophosphorimidazolide intermediate in situ.26 Encourag- ingly, we observed almost full consumption of the starting peptide 1 aer 48 hours to produce amidopyrophosphopep- tide 19 with 91% conversion at −20 °C as indicated by HPLC, while signicantly lower conversions of 22% and 20% were obtained at room temperature and 45 °C, respectively (ESI Fig. S23–S25†). The excellent conversion at −20 °C was consistent with our previous results indicating that the eutectic concen- trating environment of water-ice enabled efficient formation of ribonucleoside 2 -cyclophosphate from the corresponding ′ ribonucleoside 3 -NMPs) in the presence ,3 -monophosphates (3 ′ ′ ′ 13742 | Chem. Sci., 2022, 13, 13741–13747 © 2022 The Author(s). Published by the Royal Society of Chemistry Edge Article Chemical Science of DAP, MgCl2 and imidazole.31 We attribute this signicantly higher yield observed at −20 °C to (a) the freeze concentration effect32 in co-existing water-ice phases, an effect seen in other contexts27,33 and (b) a much slower hydrolytic degradation of DAP (when compared to room temperature and 45 °C), which allows the DAP to be available for the phosphorylation reaction over a longer period of time (see ESI Fig. S26–S31† for details). Through further investigations, we found that decreasing the amounts of DAP (5 equiv.), MgCl2 (2 equiv.) and imidazole (2 equiv.) still produced 19 with 86% conversion (Table 1, entry 1, and ESI Fig. S34†). Thus, we adopted these conditions for future amidophosphorylation reactions. For the hydrolysis step, 5 equiv. of sodium nitrite were added to the crude 19 in the same pot and the pH was adjusted to 3.0 and le at −20 °C. Monitoring by HPLC showed complete conversion of amidopyrophosphopeptide 19 into pyrophos- phopeptide 35 aer an additional 20 hours at −20 °C. This resulted in overall 85% conversion over two steps starting from 1, as conrmed by LC-MS (Table 1, entry 1 and ESI Fig. S35†), suggesting that the nitrous acid–induced hydrolysis of the amidopyrophosphate group in 19 was almost quantitative. Having established appropriate conditions, we sought to eval- uate the functional group tolerance by incorporating other amino acids with nucleophiles in their side chains. Peptide 2 , and ESI Fig. S3†) bearing a reactive ( lysine proceeded efficiently and produced the expected amido- pyrophosphopeptide 20 with near quantitative conversion at 96% as conrmed by LC-MS (Table 1, entry 2 and ESI Fig. S36†). Treatment of crude 20 with NaNO2 at pH 3.0 produced pyro- phosphopeptide 36 as the desired product with 95% conversion over two steps (Table 1, entry 2 and ESI Fig. S37†). As a representative example for investigating the effect of the position of the phosphate group on the efficiency and selec- tivity, we chose peptide 3 ( , and ESI Fig. S4†), which has a phosphoserine at the N-terminal residue with a neighboring histidine unit (Table 1, entry 3). Almost full conversion to the amidopyrophosphopeptide 21 (97%) and pyrophosphopeptide 37 (96% over two steps from the starting peptide 3) were observed under the conditions as indicated by HPLC analysis (Fig. 2). No other by-products were detectable during the formation of 21 and 37. Although the retention times for 21 and 37 were close, the 1-unit difference in [M − H] and 0.5-unit difference in [M − 2H]2− between the amidopyr- ophosphopeptide ion and pyrophosphopeptide ion in the MS spectra clearly showed the transformation from –NH2 in the phosphate group to –OH (Fig. 2). − Furthermore, a series of model peptides with a sequence where X = cysteine (5), aspartic similar to acid (7), arginine (8), tyrosine (9), proline (10) or serine (11) were synthesized and investigated for the scope of this chem- istry (ESI Fig. S6, S8–S12†). As outlined by the results in Table 1 (entries 7–11), the reactions from phosphopeptides 7–11 pro- ceeded smoothly to afford the amidopyrophosphopeptide intermediates 23–27 with 93–96% and the corresponding pyrophosphopeptides 39–43 with 86–92% conversion over two steps (ESI Fig. S46–S55†). Despite the presence of other nucleophilic residues, the clean conversion indicated exclusive a phosphorylated serine amidophosphorylation in the phosphoserine residue without any interferences from the existing tryptophan, lysine, histi- dine, aspartic acid, arginine, tyrosine, proline, or (non- phosphorylated) serine. In contrast, reactions of peptides lacking amidopyr- ophosphopeptide products (entries 4 and 6, and ESI Fig. S40 and S45†) even in the presence of a reactive hydroxyl and thiol group, respectively. It is noteworthy that no signicant side reactions were observed during the nitrous acid–induced hydrolysis with a nucleophilic NH2 group in lysine, trypto- phan, histidine, or arginine residues, thus suggesting the high reactivity and chemoselectivity of the amino moiety in the amidophosphate group under our conditions. gave no Previous studies on peptide ligation39 and isocyanate formation40 showed that the optimal pH value for NaNO2– mediated oxidation of hydrazides was 3.0–4.0 where all amines from standard amino acid residues were protonated and thus rendered unreactive as nucleophiles. In our case, the amido- pyrophosphate group remained active due to its low value of pKa z 1.2 25,26 as well as the limited NaNO2 added (equivalent amount to DAP), probably explaining the chemoselective HNO2- induced hydrolysis. In this context, the reaction of peptide 5 containing a cysteine is notable. As expected peptide 5 under- went, in the rst step of amidophosphorylation, almost quan- titative conversion (95%) to amidopyrophosphopeptide 22 (Table 1, entry 5, and ESI Fig. S41†). However, in the hydrolysis step with NaNO2, we observed the formation of the peptide nitrothioite (RSNO) as the major product – as indicated by the new peak in the LC-MS being +29 Da heavier than that of the free thiol peptide. It is known that the thiol group (RSH) could be oxidized using HNO2 to generate nitrothioite (RSNO).41 Nevertheless, the nitrothioite was cleanly reduced back to the original thiol by adding excess 4-mercaptophenylacetic acid (MPAA). By utilizing this protocol, the desired pyrophospho- peptide 38 with native cysteine was obtained with 66% conversion (ESI Fig. S42–S44†). Methionine, the other sulfur- containing residue, was reported to be stable with nitrous acid treatment,39 and is, therefore, not expected to interfere in this protocol as well. To expand the potential application of this methodology further, we subjected phosphothreonine- and phosphotyrosine- containing peptides (12 and 13, and ESI Fig. S13 and S14†) to the one-pot reaction conditions (Table 1, entries 12 and 13). Gratifyingly, both reactions furnished the expected pyrophos- phopeptides 44 and 45 with 93% and 83% conversions via amidopyrophosphopeptides 28 and 29, (ESI Fig. S56–S59†). The reaction of phosphopeptide 14 (ESI Fig. S15†) containing multiple nucleophilic amino acid residues proceeded well to afford the expected pyrophosphopeptide 46 with 94% conversion (Table 1, entry 14, and ESI Fig. S60 and S61†). To demonstrate the practicality of the reactions, we scaled-up the reactions of 3, 9, 11 and 12 and the progress was monitored by LC-MS (ESI Fig. S62–S69†). All reactions pro- ceeded to near complete conversion over the two steps to afford the desired pyrophosphopeptides, each of which were puried and isolated by preparative HPLC in 65%, 63%, 71% and 68% yields, respectively (Table 1, entries 3, 9, 11 and 12). The respectively © 2022 The Author(s). Published by the Royal Society of Chemistry Chem. Sci., 2022, 13, 13741–13747 | 13743 Chemical Science Edge Article Table 1 One-pot pyrophosphorylation of peptides using a sequential two-step protocol Entry Substrate peptide sequencea b b 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Sequence origin Random Random Random Random Random Random Random Random Random Random Random Random Random Random RPA190 34 IC2C35 Gcr1 36,37 EIF2S2 37,38 Amidopyrophosphorylated productc (% conversion) Pyrophosphorylated productc (% conversion over two steps) 19 (86) 20 (96) 21 (97) NR 22 (95) NR 23 (94) 24 (93) 25 (96) 26 (94) 27 (96) 28 (94) 29 (90) 30 (95) 31 (82) 32 (63) 33 (88) 34 (70) 35 (85) 36 (95) 37 (96) (65)d NAf 38 (66)g NAf 39 (92) 40 (86) 41 (90) (63)d (83)e 42 (86) 43 (90) (71)d (87)e 44 (93) (68)d (85)e 45 (83) 46 (94) 47 (79) 48 (60) 49 (86) 50 (68) a Amino acids highlighted in bold and red bear a phosphate group. Substrate numbers are given in bold. b Peptides that do not contain a phosphate group. c % conversions were estimated by reverse phase HPLC analysis based on the area-under-the-curve of the amidopyrophosphopeptide intermediates 19–34 and pyrophosphopeptides 35–50 vs. sum area of all peptidic-peaks. d The isolated yield (aer two steps) of the puried pyrophosphopeptide by reverse phase HPLC purication. e The yield (aer two steps) determined using standard curves of the puried pyrophosphopeptides. f NA = not applicable. g A nitrothioite peptide was observed by LC-MS as the product from oxidation of the thiol group by HNO2. Pyrophosphopeptide 38 (containing the thiol) was formed from reduction of the nitrothioite-intermediate by the addition of excess 4- mercaptophenylacetic acid (MPAA). Reaction conditions: (1) peptide (1 mM, 1.0 equiv.), DAP (5 mM, 5.0 equiv.), MgCl2 (2 mM, 2.0 equiv.), imidazole (2 mM, 2.0 equiv.), H2O, pH = 5.5, −20 °C, 48–72 h; (2) NaNO2 (5 mM, 5.0 equiv.), H2O, pH = 3.0, −20 °C, 20 h. identities of the pyrophosphopeptide products were also established by 31P-NMR (ESI Fig. S73–S78†). Four additional known sequences (15, 16, 17 and 18, and ESI Fig. S16–S19†) containing free N- and C-termini, derived from the DNA- directed RNA polymerase I subunit RPA190 34 (RPA190, Table 1, entry 15), the cytoplasmic dynein 1 intermediate chain 2 35 (IC2C, Table 1, entry 16), the glycolytic gene transcriptional activator36,37 (Gcr1, Table 1, entry 17) and the mammalian protein eukaryotic translation initiation factor 2 37,38 (EIF2S2, Table 1, entry 18) were selected as substrates. The one-pot reactions of all these fully unprotected peptides 15, 16, 17 and 18 proceeded to give desired pyrophosphopeptides 47, 48, 49 and 50 with 79%, 60%, 86% and 68% conversion, respectively (ESI Fig. S79–S86†). The slight drop in yield for 16 and 18 may be a sequence effect and not due to the unprotected N- or C- terminus since the unprotected amino and carboxylic side chains (e.g., in sequences 2 and 7) have not impacted the yields. The reaction displayed excellent chemoselectivity considering the presence of multiple nucleophilic amino acids such as glutamic acid, aspartic acid, serine, and lysine in these peptides. Tandem mass spectrometry (MS/MS) was reported to be used as a fragmentation technique for the identication of serine and threonine pyrophosphorylation.42 To further conrm the constitutional integrity of the pyrophosphopeptide product, peptide 43 was characterized by collision-induced dissociation (CID) MS/MS spectrometry (Fig. 3). The observed fragment ion (M-178), generated from 43 by loss of the pyrophosphate motif, is a clear indicator for the pyrophosphorylated peptide and supports the assignment of phosphorylated serine as the site of pyrophosphorylation.42 The b-ions and y-ions with varying lengths of amino acids retained by the amino and carboxyl- terminal part of M-178 were identied, precisely conrming the sequence of 43 as (ESI Fig. S87 and S88†). The MS/MS analysis revealed no modication of tryptophan, serine, or lysine side chains in 43. Combined with the exclusively single peak observed in HPLC for 43 (ESI Fig. S67†) and 31P NMR (ESI Fig. S75†), this analysis again demonstrated the efficiency and chemoselectivity of the pyro- phosphorylation reaction. Next, we assessed the generality of the one-pot aqueous pyrophosphorylation strategy with a highly functionalized 13744 | Chem. Sci., 2022, 13, 13741–13747 © 2022 The Author(s). Published by the Royal Society of Chemistry Edge Article Chemical Science The efficient one-pot pyrophosphorylation of the highly func- tionalized 52 leading to the formation of 54 once again high- lighted the high intrinsic nucleophilicity and selectivity of the phosphoserine residue towards DAP and the superior reac- tivity of the peptide-amidophosphate towards HNO2 even in a complex sequence. And yields of this one-pot strategy were much better than the previous results (ca. 11%) of the benzyl protected diphosphate derivative reported in the literature.20 as via one-pot To further validate our methodology on producing valuable polyphosphorylated peptides, as a proof-of-principle experi- ment, we selected phosphoserine- and phosphothreonine- containing peptides (3 and 12, respectively) to generate tri- phosphopeptides sequential a amidophosphorylation-hydrolysis scenario in water (Fig. 5). As observed previously, reactions of phosphopeptides 3 and 12 reached completion to quantitatively afford the pyrophospho- peptides 37 and 44 via amidopyrophosphopeptides 21 and 28, respectively (ESI Fig. S92, S93, S97 and S98†). By simply repeating the second amidophosphorylation and subsequent hydrolysis steps in the same pot, satisfactory conversions to nal triphosphopeptides 56 and 58 were achieved in 61% and 70% yield, via amidotriphosphopeptides 55 and 57, respec- tively, as analysed by HPLC and anion exchange chromatog- raphy Successful triphosphorylation of phosphopeptides 3 and 12 could poten- tially broaden the applicability of this method in generating polyphosphorylated peptides. and S99–S101†). S94–S96 (ESI Fig. Fig. 2 Crude LC-MS chromatograms of reaction mixtures confirming the formation of pyrophosphorylated peptide 37 by a one-pot sequential amidophosphorylation-hydrolysis scenario. (i) The starting material 3 and the production of (ii) amidopyrophosphopeptide intermediate 21 and (iii) desired pyrophosphopeptide 37. The conversion in each step, as judged by using the HPLC traces, was nearly quantitative. substrate, the Nopp140 peptide fragment containing amino acids 76–100, which is known to be pyrophosphorylated by inositol pyrophosphate in vitro.14 The Nopp140 peptide frag- ment, a 25-residue peptide bearing a single phosphoserine, seven serines/threonines, ve lysines, and eight glutamic/ aspartic acids, seemed a challenging substrate to explore (Fig. 4a). The substrate 52 was synthesized by post-assembly phosphitylation and oxidation in the solid phase30,43 in high efficiency and characterized by LC-MS (Fig. 4b, and ESI Fig. S1, S20 and S21†). Peptide 52 was reacted with DAP for 20 hours and was efficiently amidophosphorylated to form amidopyr- ophosphopeptide 53 in 75% conversion (Fig. 4a, and ESI Fig. S89†). Remarkably, subsequent treatment of 53 with NaNO2 in the same pot successfully afforded pyrophospho- peptide 54 with overall 70% conversion over two steps (Fig. 4a, and ESI Fig. S90†). The identities of 53 and 54 were conrmed by high-resolution mass spectrometry (HRMS) (Fig. 4c and d). mode, confirming Fig. 3 CID MS/MS analysis of the pyrophosphopeptide 43 in positive as ion . The sequence was identified by matching b-ion (red) and y-ion (blue) fragments. M-178 was generated from 43 losing the pyrophosphate motif. The b-ions were retained by the amino-terminal part of M-178 and the y-ions were retained by the carboxyl-terminal part of M-178. sequence the © 2022 The Author(s). Published by the Royal Society of Chemistry Chem. Sci., 2022, 13, 13741–13747 | 13745 Chemical Science Edge Article Fig. 4 Formation of the pyrophosphopeptide 54 from the highly functionalized Nopp140 peptide fragment. (a) General reaction scheme showing the pyrophosphorylation of the Nopp140 peptide fragment 52 to generate 54 via a one-pot sequential amidophosphorylation- hydrolysis scenario. (b)–(d) High resolution ESI-MS (m/z) spectra in positive ion mode of the starting peptide 52, amidopyrophosphopeptide intermediate 53, and desired pyrophosphopeptide 54, respectively. Conversions were calculated by analytical HPLC based on the area-under- the-curve of 53 and 54 vs. sum area of all peptidic peaks. pyrophosphorylation of complex phosphopeptide targets. Furthermore, triphosphorylation of phosphopeptides was ach- ieved by simply repeating the two steps of amidophosphor- ylation and hydrolysis in the same pot. Given the high yields and lack of byproducts, it is plausible that the thus-generated crude pyrophosphopeptides could be used directly in subse- quent exploratory studies. For example, a repeat of this DAP- mediated phosphorylation has the potential to produce peptide- and protein-polyphosphates44,45 in the same pot. Thus, we anticipate that this method will serve as a practical tool to quickly generate biologically relevant pyrophosphopeptides and explore their biochemistry in the emerging area of protein pyrophosphorylation. Data availability The data supporting this article are available in the ESI.† Author contributions H. Lin, L. J. Leman and R. Krishnamurthy conceived the project. H. Lin performed the experiments and analysed the data. H. Lin, L. J. Leman and R. Krishnamurthy wrote the manuscript and approved the nal version. Fig. 5 Synthesis of triphosphopeptides 56 and 58 from respective phosphopeptides 3 and 12 via a one-pot four step sequential amido- phosphorylation-hydrolysis scenario in water. Conclusions In summary, we have developed a one-pot strategy that enables efficient peptide pyrophosphorylation in water via a sequential amidophosphorylation-hydrolysis scenario. This method is operationally simple and highly chemoselective, obviating the need for orthogonal protecting groups for the canonical amino acid residues within the peptide. The efficacy of the pyrophos- phorylation reaction was demonstrated by the broad scope of phosphopeptide substrates with different amino acid compo- this method allowed the site-selective sitions. Notably, 13746 | Chem. Sci., 2022, 13, 13741–13747 © 2022 The Author(s). Published by the Royal Society of Chemistry Edge Article Conflicts of interest There are no conicts to declare. Acknowledgements This work was supported by NSF and the NASA Astrobiology Program under the Centre for Chemical Evolution (CHE- 150421) and a grant from the Simons Foundation to R. K. (327124FY19). References Chemical Science 19 A. M. Marmelstein, J. A. M. Morgan, M. Penkert, D. T. Rogerson, J. W. Chin, E. Krause and D. Fiedler, Chem. Sci., 2018, 9, 5929–5936. 20 A. M. Marmelstein, L. M. Yates, J. H. Conway and D. Fiedler, J. Am. Chem. Soc., 2014, 136, 108–111. 21 T. Durr-Mayer, D. Qiu, V. B. Eisenbeis, N. Steck, M. Haner, J. S. Siegel, K. K. Baldridge and A. Hofer, A. Mayer, H. J. Jessen, Nat. Commun., 2021, 12, 5368. 22 T. M. Haas, S. Wiesler, T. Durr-Mayer, A. Ripp, P. Fouka, D. Qiu and H. J. Jessen, Angew. Chem., Int. Ed., 2022, 61, e202113231. 23 B. Roy, A. Depaix, C. Perigaud and S. Peyrottes, Chem. Rev., 2016, 116, 7854–7897. 24 S. M. Shepard, H. J. Jessen and C. C. Cummins, J. Am. Chem. 1 Y. L. Deribe, T. Pawson and I. Dikic, Nat. Struct. Mol. Biol., Soc., 2022, 144, 7517–7530. 2010, 17, 666–672. 25 A. Osumah and R. Krishnamurthy, ChemBioChem, 2021, 22, 2 K. K. Biggar and S. S. Li, Nat. Rev. Mol. Cell Biol., 2015, 16, 5– 3001–3009. 17. 3 C. Choudhary and M. Mann, Nat. Rev. Mol. Cell Biol., 2010, 11, 427–439. 4 S. J. Humphrey, D. E. James and M. Mann, Trends Endocrinol. Metab., 2015, 26, 676–687. 26 C. Gibard, S. Bhowmik, M. Karki, E. K. Kim and R. Krishnamurthy, Nat. Chem., 2018, 10, 212–217. I. 27 H. Lin, E. J. T. Arriola, U. F. Muller and R. Krishnamurthy, Angew. Chem., Int. Ed., 2022, 61, e202113625. 28 M. Ikehara, S. Uesugi and T. Fukui, Chem. Pharm. Bull., 1967, Jimenez, 5 K. W. Moremen, M. Tiemeyer and A. V. Nairn, Nat. Rev. Mol. 15, 440–447. Cell Biol., 2012, 13, 448–462. 29 W. P. Jencks and M. Gilchrist, J. Am. Chem. Soc., 1964, 86, 6 E. Verdin and M. Ott, Nat. Rev. Mol. Cell Biol., 2015, 16, 258– 1410–1417. 264. 30 T. J. Attard, N. O'Brien-Simpson and E. C. Reynolds, Int. J. 7 E. H. Fischer, D. J. Graves, E. R. S. Crittenden and E. G. Krebs, Pept. Res. Ther., 2007, 13, 447–468. J. Biol. Chem., 1959, 234, 1698–1704. 8 T. Garcia-Garcia, S. Poncet, A. Derouiche, L. Shi, I. Mijakovic and M. F. Noirot-Gros, Front. Microbiol., 2016, 7, 184. 9 I. Kitazumi and M. Tsukahara, FEBS J., 2011, 278, 427–441. 10 D. V. Hoof, J. Munoz, S. R. Braam, M. W. H. Pinkse, R. Linding, A. J. R. Heck, C. L. Mummery and J. Krijgsveld, Cell Stem Cell, 2009, 5, 214–226. 11 M. Mann, S. E. Ong, M. Gronborg, H. Steen, O. N. Jensen and A. Pandey, Trends Biotechnol., 2002, 20, 261–268. 12 A. Chakraborty, S. Kim and S. H. Snyder, Sci. Signaling, 2011, 4, re1. 31 E. Y. Song, E. I. Jimenez, H. Lin, K. L. Vay, R. Krishnamurthy and H. Mutschler, Angew. Chem., Int. Ed., 2021, 60, 2952– 2957. 32 L. Y. An, Z. Dai, B. Di and L. L. Xu, Molecules, 2021, 26, 750. 33 A. Kanavarioti, P. Monnard and D. W. Deamer, Astrobiology, 2001, 1, 271–281. 34 S. G. Thota, C. P. Unnikannan, S. R. Thampatty, R. Manorama and R. Bhandari, Biochem. J., 2015, 466, 105–114. 35 M. Chanduri, A. Rai, A. B. Malla, M. Wu, D. Fiedler, R. Mallik and R. Bhandari, Biochem. J., 2016, 473, 3031–3047. 36 Z. Szijgyarto, A. Garedew, C. Azevedo and A. Saiardi, Science, 13 A. Saiardi, R. Bhandari, A. C. Resnick, A. M. Snowman and 2011, 334, 802–805. S. H. Snyder, Science, 2004, 306, 2101–2105. 14 R. Bhandari, A. Saiardi, Y. Ahmadibeni, A. M. Snowman, A. C. Resnick, T. Z. Kristiansen, H. Molina, A. Pandey, J. K. Werner, K. R. Juluri, Y. Xu, G. D. Prestwich, K. Parang and S. H. Snyder, Proc. Natl. Acad. Sci. U. S. A., 2007, 104, 15305–15310. 37 L. M. Yates and D. Fiedler, ChemBioChem, 2015, 16, 415–423. 38 V. K. Pathak, P. J. Nielsen, H. Trachsel and J. W. B. Hershey, Cell, 1988, 54, 633–639. 39 G. M. Fang, Y. M. Li, F. Shen, Y. C. Huang, J. B. Li, Y. Lin, H. K. Cui and L. Liu, Angew. Chem., Int. Ed., 2011, 50, 7645–7649. 15 A. Burton, C. Azevedo, C. Andreassi, A. Riccio and A. Saiardi, 40 A. A. Vinogradov, M. D. Simon and B. L. Pentelute, Org. Lett., Proc. Natl. Acad. Sci. U. S. A., 2013, 110, 18970–18975. 2016, 18, 1222–1225. 16 A. Chakraborty, M. A. Koldobskiy, N. T. Bello, M. Maxwell, J. J. Potter, K. R. Juluri, D. Maag, S. Kim, A. S. Huang, M. J. Dailey, M. Saleh, A. M. Snowman, T. H. Moran, E. Mezey and S. H. Snyder, Cell, 2010, 143, 897–910. 17 T. S. Lee, J. Y. Lee, J. W. Kyung, Y. Yang, S. J. Park, S. Lee, I. Pavlovic, B. Kong, Y. S. Jho, H. J. Jessen, D. H. Kweon, Y. K. Shin, S. H. Kim, T. Y. Yoon and S. Kim, Proc. Natl. Acad. Sci. U. S. A., 2016, 113, 8314–8319. 18 J. Worley, X. Luo and A. P. Capaldi, Cell Rep., 2013, 3, 1476– 1482. 41 L. Grossi and P. C. Montevecchi, J. Org. Chem., 2002, 67, 8625–8630. 42 M. Penkert, L. M. Yates, M. Schumann, D. Perlman, D. Fiedler and E. Krause, Anal. Chem., 2017, 89, 3672–3680. 43 D. M. Andrews, J. Kitchin and P. W. Seale, Int. J. Pept. Protein Res., 1991, 38, 469–475. 44 C. Azevedo, T. Livermore and A. Saiardi, Mol. Cell, 2015, 58, 71–82. 45 C. Azevedo, J. Singh, N. Steck, A. Hofer, F. A. Ruiz, T. Singh, H. J. Jessen and A. Saiardi, ACS Chem. Biol., 2018, 13, 1958–1963. © 2022 The Author(s). Published by the Royal Society of Chemistry Chem. Sci., 2022, 13, 13741–13747 | 13747
10.3389_fonc.2022.895544
ORIGINAL RESEARCH published: 13 May 2022 doi: 10.3389/fonc.2022.895544 Post-Radiotherapy PET Image Outcome Prediction by Deep Learning Under Biological Model Guidance: A Feasibility Study of Oropharyngeal Cancer Application Hangjie Ji 1, Kyle Lafata 2,3,4, Yvonne Mowery 2, David Brizel 2, Andrea L. Bertozzi 5,6, Fang-Fang Yin 2 and Chunhao Wang 2* Edited by: Jing Cai, Hong Kong Polytechnic University, Hong Kong SAR, China Reviewed by: Jianping Bi, Hubei Cancer Hospital, China Pei Yang, Central South University, China Sai Kit Lam, Hong Kong Polytechnic University, Hong Kong SAR, China *Correspondence: Chunhao Wang chunhao.wang@duke.edu Specialty section: This article was submitted to Radiation Oncology, a section of the journal Frontiers in Oncology Received: 14 March 2022 Accepted: 11 April 2022 Published: 13 May 2022 Citation: Ji H, Lafata K, Mowery Y, Brizel D, Bertozzi AL, Yin F-F and Wang C (2022) Post-Radiotherapy PET Image Outcome Prediction by Deep Learning Under Biological Model Guidance: A Feasibility Study of Oropharyngeal Cancer Application. Front. Oncol. 12:895544. doi: 10.3389/fonc.2022.895544 1 Department of Mathematics, North Carolina State University, Raleigh, NC, United States, 2 Department of Radiation Oncology, Duke University Medical Center, Durham, NC, United States, 3 Department of Radiology, Duke University Medical Center, Durham, NC, United States, 4 Department of Electrical and Computer Engineering, Duke University, Durham, NC, United States, 5 Mechanical and Aerospace Engineering Department, University of California, Los Angeles, Los Angeles, CA, United States, 6 Department of Mathematics, University of California, Los Angeles, Los Angeles, CA, United States Purpose: To develop a method of biologically guided deep learning for post-radiation 18FDG-PET image outcome prediction based on pre-radiation images and radiotherapy dose information. Methods: Based on the classic reaction–diffusion mechanism, a novel biological model was proposed using a partial differential equation that incorporates spatial radiation dose distribution as a patient-specific treatment information variable. A 7-layer encoder–decoder- based convolutional neural network (CNN) was designed and trained to learn the proposed biological model. As such, the model could generate post-radiation 18FDG-PET image outcome predictions with breakdown biological components for enhanced explainability. The proposed method was developed using 64 oropharyngeal patients with paired 18FDG- PET studies before and after 20-Gy delivery (2 Gy/day fraction) by intensity-modulated radiotherapy (IMRT). In a two-branch deep learning execution, the proposed CNN learns specific terms in the biological model from paired 18FDG-PET images and spatial dose distribution in one branch, and the biological model generates post-20-Gy 18FDG-PET image prediction in the other branch. As in 2D execution, 718/233/230 axial slices from 38/13/13 patients were used for training/validation/independent test. The prediction image results in test cases were compared with the ground-truth results quantitatively. Results: The proposed method successfully generated post-20-Gy 18FDG-PET image outcome prediction with breakdown illustrations of biological model components. Standardized uptake value (SUV) mean values in 18FDG high-uptake regions of predicted images (2.45 ± 0.25) were similar to ground-truth results (2.51 ± 0.33). In 2D-based Gamma analysis, the median/mean Gamma Index (<1) passing rate of test Frontiers in Oncology | www.frontiersin.org 1 May 2022 | Volume 12 | Article 895544 Ji et al. Biologically Guided Deep Learning of PET images was 96.5%/92.8% using the 5%/5 mm criterion; such result was improved to 99.9%/99.6% when 10%/10 mm was adopted. Conclusion: The developed biologically guided deep learning method achieved post-20- Gy 18FDG-PET image outcome predictions in good agreement with ground-truth results. With the breakdown biological modeling components, the outcome image predictions could be used in adaptive radiotherapy decision-making to optimize personalized plans for the best outcome in the future. Keywords: biological modeling, deep learning, image outcome prediction, radiotherapy, 18FDG-PET INTRODUCTION Radiotherapy is a central component of the standard of care for many cancers. In the current era of image-guided radiotherapy (IGRT), medical imaging plays a critical role in radiotherapy practice regarding patient assessment, treatment volume definition, on-board patient positioning, and outcome assessment (1). In particular, imaging-based radiotherapy outcome assessment can capture early therapeutic responses for adaptive therapy to enhance radiotherapy efficacy (2). In addition, long-term therapeutic outcomes from image-based analysis provide useful information in treatment intervention of each patient towards optimized cancer care (3). Thus, medical imaging analysis for radiotherapy outcome assessment has become an irreplaceable component in precision medicine. Technologies of medical imaging analysis have revolutionized image-based radiotherapy outcome reporting. Radiographic assessment of post-radiotherapy tumor morphological changes (i.e., Response Evaluation Criteria in Solid Tumors [RECIST]) was standardized to describe the response to therapy (4). Functional imaging modalities have now shifted outcome analysis from morphological description to physiological characterization. PET tracks the in vivo radioactive tracer distribution, for example, estimating glucose metabolism (18F- FDG) or measuring tissue hypoxia (18FMISO) (5). MR functional imaging, including dynamic contrast-enhanced MRI (DCE- MRI), diffusion-weighted MRI (DWI), and diffusion tensor MRI (DTI), can measure tissue properties such as blood volume/perfusion (6), cellular density (7), and cell movement direction (8). To non-invasively quantify in vivo physiology, functional imaging relies on mathematical models to extract quantitative parameters from phenotype image data. These mathematical models, which are often referred to as mechanism-based models, describe complex physiological processes using basic biological theories and fundamental laws in physical/chemical interactions (9, 10). The derived parameters of mechanism-based models can serve as surrogates of individual physiology functions to facilitate developing a personalized therapeutic approach. Treatment response assessment using functional imaging is often reported as posttreatment changes relative to pretreatment baseline values. Image-based treatment outcome prediction, i.e., forecasting posttreatment image volumes before treatment initiation, has become an emerging topic in clinical oncology (11). The potential clinical application of image-based treatment outcome prediction in radiotherapy is conceptually promising: given an individual’s pre-radiotherapy image, post-radiotherapy image predictions could be available at the treatment planning stage. Guided by these predictions, clinicians could simulate alternative treatment plans, such as target delineation revision and plan parameter tuning (beam angle, energy selection, etc.), for normal tissue sparing and could select a plan that predicts improved response to radiotherapy. This scenario can be applied to adaptive radiotherapy: the predicted intra-treatment images can be used to determine whether a revised radiotherapy plan would be advantageous. Additionally, when new intra-treatment image data are collected, the updated predictions can guide the adaptive planning strategy for optimal radiotherapy outcomes for individual patients (10). Driven by the rapid growth of computation power, deep learning techniques have recently become a practical approach for image-based treatment outcome prediction (12–14). However, few investigators have reported functional image outcome prediction in radiotherapy applications. Aside from the colossal computational workload due to image dimension requirement, the current mechanism- based models focus on spatial decoding of physiology within an image volume; for outcome prediction, a mechanism-based model must incorporate patient-specific treatment information to simulate spatiotemporal physiology evolution during a treatment course. Although pilot studies have reported the feasibility of post-radiotherapy functional image outcome prediction using treatment information (15), the adopted deep learning network ignored the biophysical modeling and generated its prediction via a “black box”; thus, the achieved prediction was reported at a fixed time point without any biological interpretation about how radiation dose affects the outcome. Radiotherapy outcome prediction with breakdowns from biological modeling is an unmet need. In this work, we design a biologically guided deep learning framework for intra-treatment 18FDG-PET image outcome prediction in response to oropharyngeal cancer intensity- modulated radiotherapy (IMRT). Based on the classic reaction–diffusion mechanism in disease progression, we propose a novel partial differential equation (PDE) as a biological model that incorporates spatial radiation dose distribution as a patient-specific treatment information variable. An encoder–decoder-based convolutional neural network (CNN) is designed and trained to learn the proposed Frontiers in Oncology | www.frontiersin.org 2 May 2022 | Volume 12 | Article 895544 Ji et al. Biologically Guided Deep Learning of PET model, which governs the dynamics of tissue response to radiotherapy. Thus, with the explainability of the biological model, the developed deep learning model can generate post- radiotherapy 18FDG-PET image outcome predictions with breakdown biological components. MATERIALS AND METHODS Biological Modeling We hypothesize that the standardized uptake value (SUV) change in 18FDG-PET in response to radiation can be described in a reaction–diffusion system, which represents a family of mathematical models widely used in describing pattern formations and evolving densities in physical, ecological, and biological systems (16). In the context of modeling tumor growth and therapeutic response dynamics, reaction–diffusion models have been applied to both preclinical and clinical works (9, 17, 18). Disease progression, in general, can be summarized by Eq. (1), which describes the malignancy proliferation (reaction) and spread (diffusion) (10): Ut = aDU + bU (1) where U is the spatial distribution of disease (i.e., SUV intensity distribution in this work) and Ut = ∂ U ∂ t is the time derivative term describing the change of U in time. The term aDU = a( ∂2 U ∂ x2 + ∂2 U ∂ y2 ) describes the spreading of abnormal cell activities, where a > 0 is the diffusion coefficient. The linear term bU represents the proliferation of localized malignancy. To incorporate tissue response to radiotherapy in the model in Eq. (2), we propose a new response term for the dose-induced changes of U, Ut = aDU + bU + F DUð F Þ (2) where F(DU) is an N unknown operator that depicts U's local response to radiotherapy. Here we assume that the response term depends on the product of U and the radiotherapy plan’s spatial dose distribution D. We also assume that the operator F depends on DU as the tissue response to cell killing from localized high radiation (10), and we will use a CNN to learn this operator. Thus, Eq. (2) is the core time-dependent PDE that models the post-radiotherapy biological response of abnormal tissue metabolism as SUV intensity (i.e., U) evolves in time. k k , U post Deep Learning Design Formally, our problem is defined as follows: given a set of pre- and post-radiation 18FDG-PET image pairs f(U pre )gk=1,2,…m and the imposed radiation dose distribution images { Dk }k=1,2,…m where m is the total number of image pairs, our goal is to learn the unknown response operator F and coefficients a, b in the model in Eq. (2) with the collected data of the form f(U pre , Dk) gk=1,2, …m. Accordingly the learned model can predict a post- radiation 18FDG-PET image U post given the pre-radiation image U pre and the associated spatial dose distribution Dk. In addition, k since the learned model describes the evolution dynamics of Uk k , U post k k between the two states U pre k and U post disease progression. k frames of Uk between U pre k can be simulated to study the intermediate stages of and U post k While a large body of work has focused on solving reaction– diffusion models like Eq. (2), i.e., finding U based on known coefficients and operators, little research has been devoted to the inverse problem of learning the model’s coefficients and operators from observed U data. The numerical treatments of the inverse problem are typically complicated, as the observed data usually cannot provide sufficient information to determine a unique model, and regularizations are needed to produce meaningful model estimates. As such, we propose a deep neural network framework to learn the model in Eq. (2) from 18FDG-PET images taken before and after radiation. Applying the forward Euler method on the PDE in Eq. (2), we obtain the discretized update rule: U n+1 = U n + haDU n + hbU n + hF DU n ð Þ (3) where h is the time step, Un is the approximate solution of the state U at the time tn=nh, and the Lssssaplacian operator D can be approximated by a discrete operator D2 xy represented by a nine- point refined stencil (19): 0 B B @ 1=4 1=2 1=4 1=2 −3 1=2 1=4 1=2 1=4 1 C C A D2 xy = (4) A deep neural network NF is designed to approximate the response operator F: NF : y ! NF y ;  q ð Þ (5) where q represents the free parameters. For simplicity, we assume that the operator F only depends on, y=DU, the product of the dose distribution D and the 18FDG-PET image state variable U. A diffusion–proliferation operator G is used to combine both the diffusion and proliferation terms with undetermined coefficients a and bs G U n G ð Þ = aD2 xyU n + bU n (6) Given a group of three images consisting of the initial state 18FDG-PET image U 0 at t = 0 prior to radiation, dose distribution map Dk, and the ground-truth final state 18FDG- PET image U post at t = T (post-radiation), from Eqs. (3)–(6), we obtain the intermediate states U n+1 k = U pre by k k k U n+1 k = U n G k + hG U n kð Þ + hNF Dk F ð ∘ U n k ; q Þ (7) ∘ U n for n = 0,1, …, Nt − 1. Here, Dk k represents the element-wise product of the dose distribution map Dk and the 18FDG-PET image U n k at the time step tn, Nt is the total number of steps, and the step size h=T/Nt. As a feasibility study, we consider the final time T = 1 and set the number of steps Nt = 4 in this work. The similarity between the predicted post-radiation 18FDG- is k and the associated ground-truth image U post PET image U Nt k Frontiers in Oncology | www.frontiersin.org 3 May 2022 | Volume 12 | Article 895544 Ji et al. Biologically Guided Deep Learning of PET defined based on the l2 norm loss function: L qð Þ = L 1 m om k=1 ∥ U Nt k − U post k ∥2 2 (8) where m is the number of samples. By minimixing L(q), the deep neural network can learn the weights q that characterize the response operator F and the undetermined coefficients a and b. Figure 1 illustrates the designed deep neural network architecture. The network’s input space is composed of pre- radiation 18FDG-PET image Upre and planned dose distribution map D as a set. The network is split into two branches: one that uses a CNN to learn the response operator NF(DUn) and the other one with only two trainable parameters to apply the diffusion–proliferation operator G [in Eq. (6)] on Un. Specifically, the second branch of the network architecture mimics the traditional finite difference method and applies the discrete Laplacian operator and the linear operator on Un with predicted a and b. Both branches are then merged by the rule in Eq. (7), which feeds the output Un+1 forward to the next cycle. This process is then repeated for Nt time steps to generate a predicted post-radiation 18FDG-PET image, which will be compared against the ground-truth post-radiation 18FDG- PET image. The branch that learns the response operator NF(DUn) consist of a total of 7 convolutional layers and is built upon U- Net’s encoder–decoder architecture (20). The architecture consists of a contracting path that extracts sufficient semantic context from D∘Un and a symmetric expanding path that produces the up-sampled output. The contracting path starts with two applications of 3 × 3 convolutions, each followed by a batch normalization layer and a ReLU operation. Then a 2 × 2 max-pooling operation is performed for down-sampling where the number of feature channels is doubled. Then another two 3 × 3 convolutions operations are applied, each followed by batch normalization and a ReLU activation. The expanding path consists of an up-sampling of the feature map, followed by two 3 × 3 convolutional layers, again with batch normalization and ReLU operations. Finally, a 1 × 1 convolution is applied to map the 16-component feature to a single feature channel that reconstructs the transformed image corresponding to NF(DUn). Patient Data and Network Training In this work, 64 eligible oropharyngeal cancer patients who received curative-intent IMRT in our department were retrospectively studied under an institutional review board (IRB)-approved 18FDG-PET imaging study (21). All patients were prescribed 70 Gy at 2 Gy/day fraction with concurrent chemotherapy. Prior to treatment initiation, each patient underwent an 18FDG-PET/CT scan for target delineation. After 20-Gy delivery, each patient underwent a second 18FDG- PET/CT scan as an intra-treatment evaluation for consideration for adaptive planning. These post-20-Gy 18FDG-PET acquisitions were treated as the post-radiation scans in the modeling. All 18FDG-PET/CT exams were acquired by a PET/CT scanner (Siemens, Erlangen, Germany) in our department. PET acquisitions were performed using 400 × 400 matrix size FIGURE 1 | A partial differential equation (PDE)-informed deep neural network design. Layers are color-coded by operations with associated feature numbers. Frontiers in Oncology | www.frontiersin.org 4 May 2022 | Volume 12 | Article 895544 Ji et al. Biologically Guided Deep Learning of PET ™ ™ in a standard field of view (FOV) of 54 cm, and slice thickness was 2 mm. CT acquisitions were performed using 512 × 512 matrix size in an extended FOV of 65 cm (in-plane resolution = 1.27 mm), and slice thickness was 3 mm. PET images were reconstructed by the ordered subset expectation maximization (OSEM) algorithm with attenuation corrections using the CT acquisition information. The post-20-Gy 18FDG-PET/CT images were registered to the corresponding pre-radiation images using Velocity software (Varian Medical Systems, Palo Alto, CA, USA). Registrations started with rigid bony structure alignment, and a multi-pass deformable registration algorithm was adopted to improve soft tissue alignment near the anterior body surface. In the process of IMRT planning, all treatment plans were optimized and calculated using the Eclipse treatment planning system (Varian Medical Systems, Palo Alto, CA, USA) with a 2.5-mm dose calculation grid size. All 18FDG- PET images and spatial dose distribution maps of 20-Gy treatment were resampled to the CT simulation image grid size. Of all 2D 18FDG-PET axial images, those with sufficient 18FDG uptake in the pre-radiation acquisition were selected by SUVmax > 1.5 excluding brain regions (22). Overall, 718 axial slices collected from 38 patients were used for neural network training, 233 axial slices from 13 patients were used for validation, and 230 axial slices from 13 patients were used for independent tests. During the neural network training, the loss function was defined as based on the l2 norm in Eq. (8). Gradient updates were computed using batch sizes of 10 samples, and batch normalization was performed after each convolutional layer. The training utilized the Adam optimizer for up to 400 epochs, while an early stopping strategy on the loss function evaluated on the validation samples was adopted with a patience of 100 epochs. The overall training time was about 15 min in a TensorFlow environment using an NVIDIA TITAN Xp graphic card. ™ Evaluation The accuracy of post-20-Gy 18FDG-PET image prediction was evaluated using 230 axial slices’ results from 13 test patients. The prediction results were first visually inspected as qualitative evaluation. SUV mean values in high-uptake regions determined by Otsu’s method (23) were quantitatively compared with the ground-truth results. Pixel-to-pixel SUV numerical differences were evaluated by Gamma tests within the body region (24). Multiple Gamma tests with different SUV difference tolerances and distance-to-agreement (DTA) tolerances were performed. While Gamma Index <1 was considered as acceptable pixel-wise results, Gamma Index passing rates, i.e., the percentage of pixels with Gamma Index <1, were reported as summarizing metrics. RESULTS Figure 2 shows an example case of post-20-Gy 18FDG-PET image outcome prediction. As seen in the pre-radiation 18FDG- PET image, SUV hotspots with clear edges were found on the patient’s right side. After 20-Gy delivery shown by the bilateral side dose distribution in D, the ground-truth post-20-Gy18FDG- PET image results demonstrated good therapy response with reduced hotspot sizes and decreased SUV intensities. The predicted 18FDG-PET image captured the overall appearance in the ground-truth results without noticeable artifact marks. Two hotspots corresponding with the nodal disease were found in the prediction image at the same locations. The hotspots’ sizes and SUV intensities were comparable, though the anterior hotspot intensity appeared to be lower than the ground-truth result. In the breakdown illustration of biological model terms in Eq. (3), the diffusion term demonstrated overall uniform intensity distribution around 0 except in hotspot regions; the core regions in hotspots had negative diffusion intensities, which suggested a spatial retraction of abnormal metabolism. The proliferation term had a similar appearance to the pre- radiation 18FDG-PET image. The dose-response term indicated an elevated intensity region that corresponds to the anterior SUV hotspot; this suggests that the anterior SUV hotspot had a better response to 20-Gy than the posterior SUV hotspot, which had limited intensity in the dose-response map. The other areas within the body had close-to-zero dose-response intensity, while low intensities were found near the body surface. The Gamma Index map showed a good quantitative pixel-to-pixel SUV comparison between ground-truth and predicted post-20- Gy 18FDG-PET images using the 5%/10 mm criterion. In the test patient cohort, the SUV mean value of high-uptake regions in post-20-Gy predicted images was 2.45 ± 0.25, which was slightly lower than ground-truth results (2.51 ± 0.33, p = 0.015); the dice coefficient results of the segmented high-uptake regions were 0.89 ± 0.12. Gamma Index passing rate results of all testing axial slices are summarized in Figure 3. When the 5%/5 mm Gamma criterion was adopted, the median 2D Gamma passing rate was 96.5%. With the use of looser Gamma criteria, the passing rate results improved (5%/10 mm, median 99.2%, average 97.6%; 10%/5 mm, median 99.5%, average 98.6%). The highest median passing rate was 99.9% (average = 99.6%) when 10%/10 mm was used. DISCUSSION In this work, we successfully demonstrated the design of a biological model-guided deep learning framework for post-20- Gy 18FDG-PET image outcome prediction in a unique cohort of patients undergoing IMRT for oropharyngeal cancer. For the first time, we demonstrated 3 breakdown biological components of oropharyngeal cancer response to radiation. One of the key innovations in this work is the biological model in Eq. (2), which was hypothesized as the mathematical equation that governs the post-radiation SUV change. The model was derived from the classic reaction–diffusion system, which has been utilized in many works of tumor growth and disease progression modeling (25–27). Although applying reaction–diffusion models to 18FDG-PET image analysis (particularly to head and neck cancer) is less reported, some exploratory studies have demonstrated the validation of reaction–diffusion-type models Frontiers in Oncology | www.frontiersin.org 5 May 2022 | Volume 12 | Article 895544 Ji et al. Biologically Guided Deep Learning of PET FIGURE 2 | An example of post-radiotherapy 18FDG-PET image outcome with given pre-radiation 18FDG-PET image and dose distribution map D, with a breakdown of predicted biological effects (diffusion, proliferation, and dose response in absolute value) in Eq. (2). The 2D Gamma (G) test result was obtained through acceptance criteria of 5%/10 mm. in intracranial PET image modeling (28). Compared to the original reaction–diffusion models, the newly introduced dose- response term in Eq. (2) was hypothesized as a semantic component of dose-induced SUV image state changes. Adding additional terms in reaction–diffusion family models to account for therapeutic effect has been reported before in breast, lung, and pancreatic cancer studies (29–31); nevertheless, our approach of using spatial dose distribution in biological modeling is a novel design. Compared to the use of prescription dose levels for outcome assessment/prediction in many studies, the adoption of spatial dose distribution maintained heterogeneous radiation deposition information at the pixel level, which may be a more accurate approach for image-based outcome prediction with explainability from existing biology domain knowledge. Nevertheless, the designed deep learning model relies on the reaction–diffusion system hypothesis, which has yet to be widely acknowledged as general domain knowledge of tissue radiation response. In addition to the image result supports, the reaction–diffusion system hypothesis can be studied via in vivo functional imaging (such as diffusion-weighted MRI) and in vitro cell study to establish the benchmark evidence for oropharyngeal cancer applications. As a deep learning approach, a CNN was designed to learn the proposed biological model. PDEs with known coefficients and operators can be solved by various numerical methods such as finite difference methods, finite element methods, and spectral methods. In the scientific computing field, solving differential FIGURE 3 | Gamma Index passing rate summary with different gamma criteria. Green line positions represent median value, and box represents 25%/75% percentile with whiskers indicating 5%/95% percentile. Frontiers in Oncology | www.frontiersin.org 6 May 2022 | Volume 12 | Article 895544 Ji et al. Biologically Guided Deep Learning of PET equations using CNN in complex systems has become popular for efficiency and accuracy (32). Additionally, differential equations specified by CNN can parameterize the continuous state transition with non-uniform sampling step sizes (33); that is, one may use images from different patients with different acquisition time points. The applied analysis of stochastic differential equations has demonstrated value for recent radiomic applications (34, 35); deep learning-based data assimilation may improve the performance of these techniques by providing a more accurate estimation of model hyper- parameters and coefficients. The use of CNN is necessary to learn the dose-response term F in Eq. (2), which is an unknown operator that is assumed to be related to the product of spatial dose distribution and 18FDG-PET image variable (DU); without it is difficult to approximate the an analytical expression, operator F by classic numerical treatments of inverse problems. The proposed CNN in Figure 1 revealed the dose- response term F(DU) as a whole, while the detailed mechanism of DU‘s contribution of 18FDG-PET image prediction remains unclear. Inspired by the classic encoder–decoder U-net implementation, the CNN architecture in Figure 1 was dedicated to the problem in Eqs. (3)–(7); with the loss function defined in Eq. (8), the training process had a fast convergence (Supplementary Figure 1). It would be of interest to further study the operator F for its analytical expression and possible biological explanations. Such works require more advanced mathematical theories supported by experimental data, preferably as in vitro implementations, to validate analytical designs as a biological model calibration process (10). Based on the Gamma test results in Figure 3, the achieved 18FDG-PET image predictions showed good agreement with ground-truth images. As a common quality assurance method in radiotherapy practice, Gamma analysis accounts for both intensity differences and systematic shifts in image prediction error. The Gamma test criteria need to consider multiple uncertainty sources in data processing and clinical preferences. For instance, the dose-response term results in Figure 2 indicated very small but non-zero intensity values near the body surface, especially in anterior skin regions. While other normal tissues demonstrated very limited dose response, the observed skin regions’ dose response may be noisy results related to deformable image registration uncertainties, which was mainly determined by patient weight loss during the radiotherapy course (36). Radiotherapy margin formulism that models treatment margin statistics should also be weighted in image prediction evaluation (37). In addition to these two potential factors, the adopted Gamma test criteria have including SUV’s intrinsic incorporated many other factors, uncertainty, PET image acquisition resolution, PET-CT QA protocol, and SUV-based metabolic volume definition. The current results demonstrated accurate image outcome prediction at the time point of post-20-Gy radiotherapy. The actual physiological change during the 20-Gy radiotherapy course is a continuous process, which is an inherent feature in the proposed model in Eq. (2); in other words, in addition to post-20-Gy 18FDG-PET image outcomes at t = 1, our model can predict intermediate stage image outcomes between t = 0 and t = 1. To demonstrate this merit, Figure 4 shows a simulation of intermediate stage 18FDG-PET image outcome predictions as biological model solutions from the pre-radiation result at t = 0 to post-20-Gy prediction at t = 1. In general, the four predicted 18FDG-PET images demonstrated a reasonable image state transition from t = 0 to t = 1 without abrupt changes. While the majority of normal tissue maintained steady SUV intensities during the presented time evolution, the SUV hotspot corresponding to the primary oropharyngeal tumor had shrinkage at its posterior boundary with slightly reduced intensity. Compared to the ground-truth post-20-Gy 18FDG- PET image, the prediction image at t = 1 captured the SUV hotspot’s morphological features, particularly at its posterior boundary. However, this simulation result cannot be validated by current clinical results because of the lack of longitudinal 18FDG-PET scans during a radiotherapy course, which is mainly limited by ionizing radiation risk and potential high financial cost. On the other hand, longitudinal MRI exams are commonly utilized for cranial radiotherapy follow-up as standard care, and the image series can be used to validate the cranial model continuity in future works. The current biological model was implemented in a 2D fashion on axial images. For each test patient, the post- radiation 18FDG-PET image predictions were generated slice- by-slice to approximate volumetric rendering. In theory, the biological model in Eq. (2) and the demonstrated deep neural network could be implemented as 3D in the spatial domain; however, the computation workload for 3D implementation, especially for a generative task with complex nature, requires a large data sample size with curated data collection. In this work, 64 patients were collected with paired 18FDG-PET exams in a clinical trial setup, and 1,181 high 18FDG uptake axial slices were collected and were assigned to neural network training/ validation/tests following a 60%/20%/20% ratio. Given the fact that 1) image slice thickness (3 mm) is larger than in-plane resolution (1.27 mm) and 2) paired image acquisitions were performed with a 2-week time interval, the model was confined for locoregional computation with a small 3 × 3 in-plane kernel size, and thus the information extraction was within an axial “slab” and did not involve information exchange in other slices. This underlying design made all 2D slices eligible as independent samples for deep learning training, and the current results from 2D implementation demonstrated good image prediction accuracy and established the technical feasibility of the proposed biological model-guided deep learning. 3D-based modeling would be ideal for brainstorming experiments, but this data cohort would be a very limited data sample size for generative deep learning tasks. Future studies using a larger patient cohort, potentially in a multi-institution collaboration, are planned to further investigate the proposed framework based on 3D implementation. Additionally, experiments using small animals are also planned for future developments of deep learning in image outcome prediction. Further investigation of the biological interpretation of the learned dose-response term may also lead to improved mathematical modeling for this problem. Frontiers in Oncology | www.frontiersin.org 7 May 2022 | Volume 12 | Article 895544 Ji et al. Biologically Guided Deep Learning of PET FIGURE 4 | A simulation of 18FDG-PET image outcome transition based on a four-step execution with step size h = 0.25 showing the predicted transition from pre- radiation U at t = 0 to the predicted post-radiotherapy U at t = 1. The ground-truth post-20-Gy 18FDG-PET image is included for comparison at t = 1. As a feasibility study, the current results showed that the achieved post-20-Gy 18FDG-PET image outcome prediction had good agreement with ground-truth results. Post-20-Gy 18FDG- PET has been demonstrated as informing surrogates of recurrence-free survival and overall survival of human papillomavirus (HPV)-related oropharyngeal cancer (38). In a potential clinical application scenario, the current framework would allow a physician to determine if an 18FDG-PET scan after 20-Gy radiation would facilitate improved adaptive radiotherapy clinical decision making. The impact of image prediction accuracy on clinical decision making was not rendered by the current results of technical development work; future works, preferably in a prospective fashion, are planned to investigate such clinical impacts from physicians’ perspectives in clinical practice. Another crucial step toward this clinical application scenario is to verify the models’ responses to different radiation therapy strategies. The current patient cohort from a clinical study received a uniform treatment regimen; thus, the developed model may not capture certain individual reactions after a drastically different radiotherapy approach. For deep learning developments, it would be ethically challenging to collect patient data with intentional treatment variations. Following the small animal experiments discussed above, with dedicated imaging platforms and radiotherapy machines, one can generate post- radiation samples with heterogeneous treatment strategies in multiple imaging sessions. Such experiments may provide valuable opportunities for studying biological models for improved deep learning intelligibility. CONCLUSION In this work, we developed a biological model-guided deep learning method for post-radiation 18FDG-PET image outcome prediction. The proposed biological model incorporates spatial radiation dose distribution as a patient-specific variable, and a novel CNN architecture was implemented to predict post- radiotherapy 18FDG-PET images from pre-radiation results. Current results demonstrate good agreements between post- 20-Gy predictions and ground-truth results in a cohort of patients with oropharyngeal cancer. Future developments of the current methodology design will enhance the applicability of image outcome prediction in clinical practice. DATA AVAILABILITY STATEMENT The datasets studied for this study is collected from a clinical trial; due to PHI protection, the original image data cannot be published in the public domain. The studied deep learning design model will be available upon direct request to the corresponding author. Requests to access the datasets should be directed to chunhao.wang@duke.edu. Frontiers in Oncology | www.frontiersin.org 8 May 2022 | Volume 12 | Article 895544 Ji et al. Biologically Guided Deep Learning of PET ETHICS STATEMENT FUNDING The studies involving human participants were reviewed and approved by Duke University. The patients/participants provided their written informed consent to participate in this study. HJ and AB were supported by the Simons Foundation Math+X investigator award number 510776 and the National Science Foundation under grant NSF DMS-1952339 during this work. AUTHOR CONTRIBUTIONS SUPPLEMENTARY MATERIAL All authors participated in the study design. KL and CW collected image data. HJ completed computation works. All authors participated in writing and approved the final version. The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fonc.2022.895544/ full#supplementary-material REFERENCES 1. Jaffray DA. Image-Guided Radiotherapy: From Current Concept to Future Perspectives. Nat Rev Clin Oncol (2012) 9(12):688. doi: 10.1038/ nrclinonc.2012.194 2. Yan D. Adaptive Radiotherapy: Merging Principle Into Clinical Practice. Semin Radiat Oncol (2010) 20(2):79–83. doi: 10.1016/j.semradonc.2009.11.001 3. Chang Z, Wang C. Treatment Assessment of Radiotherapy Using MR Functional Quantitative Imaging. World J Radiol (2015) 7(1):1. doi: 10.4329/wjr.v7.i1.1 4. Therasse P, Arbuck SG, Eisenhauer EA, Wanders J, Kaplan RS, Rubinstein L, Verweij J, et al. New Guidelines to Evaluate the Response to Treatment in Solid Tumors. J Natl Cancer Instit (2000) 92(3):205–16. doi: 10.1093/jnci/92.3.205 5. Cherry SR, Sorenson JA, Phelps ME. Physics in Nuclear Medicine E-Book. Elsevier Health Sciences (2012). 6. d’Arcy JA, Collins DJ, Padhani AR, Walker-Samuel S, Suckling J, Leach MO. Magnetic Resonance Imaging Workbench: Analysis and Visualization of Dynamic Contrast-Enhanced MR Imaging Data. Radiographics (2006) 26 (2):621–32. doi: 10.1148/rg.262045187 7. Koh D-M, Collins DJ. Diffusion-Weighted MRI in the Body: Applications and Challenges in Oncology. Am J Roentgenol (2007) 188(6):1622–35. doi: 10.2214/AJR.06.1403 8. Le Bihan D, Mangin JF, Poupon C, Clark CA, Pappata S, Molko N, et al. Diffusion Tensor Imaging: Concepts and Applications. J Magnet Res Imaging Off J Int Soc Magnet Res Med (2001) 13(4):534–46. doi: 10.1002/jmri.1076 9. Hormuth DA, Jarrett AM, Yankeelov TE. Forecasting Tumor and Vasculature Response Dynamics to Radiation Therapy Via Image Based Mathematical Modeling. Radiat Oncol (2020) 15(1):4. doi: 10.1186/s13014-019-1446-2 10. Hormuth DA, Jarrett AM, Lima EA, McKenna MT, Fuentes DT, Yankeelov TE. Mechanism-Based Modeling of Tumor Growth and Treatment Response Constrained by Multiparametric Imaging Data. JCO Clin Cancer Informat (2019) 3:1–10. doi: 10.1200/CCI.18.00055 16. Britton NF. Reaction-Diffusion Equations and Their Applications to Biology. Cambridge, MA: Academic Press (1986). 17. Hormuth DA, Eldridge SL, Weis JA, Miga MI, Yankeelov TE. Mechanically Coupled Reaction-Diffusion Model to Predict Glioma Growth: Methodological Details. In: Cancer Systems Biology. Humana Press: New York, NY (2018). pp 225–41. 18. Jarrett AM, Hormuth DA, Barnes SL, Feng X, Huang W, Yankeelov TE. Incorporating Drug Delivery Into an Imaging-Driven, Mechanics-Coupled Reaction Diffusion Model for Predicting the Response of Breast Cancer to Neoadjuvant Chemotherapy: Theory and Preliminary Clinical Results. Phys Med Biol (2018) 63(10):105015. doi: 10.1088/1361-6560/aac040 19. Provatas N, Elder K. Phase-Field Methods in Materials Science and Engineering. Weinheim, Germany: John Wiley & Sons (2011). 20. Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation, in: International Conference on Medical Image Computing and Computer-Assisted Intervention 2015. Cham, Switzerland: Springer: (2015). pp. 234–41. 21. Mowery YM, Vergalasova I, Rushing CN, Choudhury KR, Niedzwiecki D, Wu Q, et al. Early 18F-FDG-PET Response During Radiation Therapy for HPV-Related Oropharyngeal Cancer May Predict Disease Recurrence. Int J Radiat Oncol Biol Phys (2020) 108(4):969–76. doi: 10.1016/j.ijrobp. 2020.08.029 22. Simon E, Fox TH, Lee D, Waller AF, Pantalone P, Jani AB. PET Lesion Segmentation Using Automated Iso-Intensity Contouring in Head and Neck Cancer. Technol Cancer Res Treat (2009) 8(4):249–55. doi: 10.1177/ 153303460900800401 23. Otsu N. A Threshold Selection Method From Gray-Level Histograms. IEEE Trans Syst Man Cybernet (1979) 9(1):62–6. doi: 10.1109/TSMC.1979.4310076 24. Low DA, Mutic S, Dempsey JF, Gerber RL., Bosch WR., Perez CA., et al. Quantitative Dosimetric Verification of an IMRT Planning and Delivery System. Radiother Oncol (1998) 49(3):305–16. doi: 10.1016/S0167-8140(98) 00125-X 11. Yankeelov TE, Quaranta V, Evans KJ, Rericha EC. Toward a Science of Tumor Forecasting for Clinical Oncology. Cancer Res (2015) 75(6):918–23. doi: 10.1158/0008-5472.CAN-14-2233 25. Chen X, Summers RM, Yao J. Kidney Tumor Growth Prediction by Coupling Reaction–Diffusion and Biomechanical Model. IEEE Trans Biomed Eng (2012) 60(1):169–73.doi: 10.1109/TBME.2012.2222027 12. Bychkov D, Turkki R, Haglund C, Linder N, Lundin J. Deep Learning for Tissue Microarray Image-Based Outcome Prediction in Patients with Colorectal Cancer. Med Imaging Dig Pathol (2016) 9791:298–303. doi: 10.1117/12.2217072 13. Chang Y, Lafata K, Liu C, Wang C., Cui Y., Ren L., et al. An Encoder-Decoder Based Convolutional Neural Network (ED-CNN) for PET Image Response Prediction Using Pre-RT Information: A Feasibility of Oropharynx Cancer vol. 46(6);pp. E283–E283: 111 RIVER ST, IMRT. In. Medical physics HOBOKEN 07030-5774, NJ USA: WILEY. (2019). 14. Hilbert A, Ramos LA, van Os HJA, Olabarriaga SD, Tolhuisen ML, Wermer MJH, et al. Data-Efficient Deep Learning of Radiological Image Data for Outcome Prediction After Endovascular Treatment of Patients With Acute Ischemic Stroke. Comput Biol Med (2019) 115:103516. doi: 10.1016/ j.compbiomed.2019.103516 15. Wang C, Liu C, Chang Y, Lafata K, Cui Y, Zhang J, et al. Dose-Distribution- Driven PET Image-Based Outcome Prediction (DDD-PIOP): A Deep Learning Study for Oropharyngeal Cancer IMRT Application. Front Oncol (2020) 1592. doi: 10.3389/fonc.2020.01592 26. Hormuth DAII, Weis JA, Barnes SL, Miga MI, Quaranta V, Yankeelov TE. Biophysical Modeling of In Vivo Glioma Response After Whole-Brain Radiation Therapy in a Murine Model of Brain Cancer. Int J Radiat Oncol Biol Phys (2018) 100(5):1270–9. doi: 10.1016/j.ijrobp.2017.12.004 27. Swan A, Hillen T, Bowman JC, Murtha AD. A Patient-Specific Anisotropic Diffusion Model for Brain Tumour Spread. Bull Math Biol (2018) 80(5):1259– 91. doi: 10.1007/s11538-017-0271-8 28. Rockne RC, Trister AD, Jacobs J, Hawkins-Daarud AJ, Neal ML, Hendrickson K, et al. A Patient-Specific Computational Model of Hypoxia-Modulated Radiation Resistance in Glioblastoma Using 18F-FMISO-PET. J R Soc Interface. (2015) 12(103):20141174. doi: 10.1098/rsif.2014.1174 29. Liu Y, Sadowski SM, Weisbrod AB, Kebebew E, Summers RM, Yao J. Patient Specific Tumor Growth Prediction Using Multimodal Images. Med Imag Anal (2014) 18(3):555–66. doi: 10.1016/j.media.2014.02.005 30. Mi H, Petitjean C, Dubray B, Vera P, Ruan S. Prediction of Lung Tumor Evolution During Radiotherapy in Individual Patients With PET. IEEE Trans Med Imag (2014) 33(4):995–1003. doi: 10.1109/TMI.2014.2301892 Frontiers in Oncology | www.frontiersin.org 9 May 2022 | Volume 12 | Article 895544 Ji et al. Biologically Guided Deep Learning of PET 31. Weis JA, Miga MI, Yankeelov TE. Three-Dimensional Image-Based Mechanical Modeling for Predicting the Response of Breast Cancer to Neoadjuvant Therapy. Comput Methods Appl Mechanic Eng (2017) 314:494–512. doi: 10.1016/j.cma.2016.08.024 32. Li A, Chen R, Farimani AB, Zhang YJ. Reaction Diffusion System Prediction Based on Convolutional Neural Network. Sci Rep (2020) 10(1):1–9. doi: 10.1038/s41598-020-60853-2 33. Chen RT, Rubanova Y, Bettencourt J, Duvenaud D. Neural Ordinary Differential Equations. ArXiv Preprint ArXiv (2018) 1806.07366. 34. Lafata KJ, Corradetti MN, Gao J, Jacobs CD, Weng J, Chang Y, et al. Radiogenomic Analysis of Locally Advanced Lung Cancer Based on CT Imaging and Intratreatment Changes in Cell-Free Dna. Radiol: Imaging Cancer (2021) 3(4):e200157. doi: 10.1148/rycan.2021200157 35. Lafata K, Zhou Z, Liu J-G, Yin F-F. Data Clustering Based on Langevin Annealing With a Self-Consistent Potential. Q Appl Math (2019) 77(3):591– 613. doi: 10.1090/qam/1521 36. Paganelli C, Meschini G, Molinelli S, Riboldi M, Baroni G. Patient-Specific Validation of Deformable Image Registration in Radiation Therapy: Overview and Caveats. Med Phys (2018) 45(10):e908–22. doi: 10.1002/mp.13162 37. Bortfeld T, van Herk M, Jiang SB. When Should Systematic Patient Positioning Errors in Radiotherapy Be Corrected? Phys Med Biol (2002) 47 (23):N297. doi: 10.1088/0031-9155/47/23/401 38. Lafata KJ, Chang Y, Wang C, Mowery YM, Vergalasova I, Niedzwiecki D, et al. Intrinsic Radiomic Expression Patterns After 20 Gy Demonstrate Early Metabolic Response of Oropharyngeal Cancers. Med Phys (2021) 48(7):3767– 77. doi: 10.1002/mp.14926 Conflict of Interest: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. Publisher’s Note: All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher. Copyright © 2022 Ji, Lafata, Mowery, Brizel, Bertozzi, Yin and Wang. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms. Frontiers in Oncology | www.frontiersin.org 10 May 2022 | Volume 12 | Article 895544
10.1016_j.jbc.2022.102361
RESEARCH ARTICLE Autoinhibition of the GEF activity of cytoskeletal regulatory protein Trio is disrupted in neurodevelopmental disorder-related genetic variants Received for publication, January 11, 2022, and in revised form, August 4, 2022 Published, Papers in Press, August 10, 2022, https://doi.org/10.1016/j.jbc.2022.102361 Josie E. Bircher1,‡ From the 1Department of Molecular Biophysics and Biochemistry, and 2Keck MS & Proteomics Resource, Yale University, New Haven, Connecticut, USA; 3Department of Pharmaceutical Chemistry, University of California at San Francisco, San Francisco, California, USA; 4Department of Neuroscience, Yale University, New Haven, Connecticut, USA , Michael J. Trnka3, and Anthony J. Koleske1,4,* , Ellen E. Corcoran1,‡ , TuKiet T. Lam1,2 Edited by Kirill Martemyanov TRIO encodes a cytoskeletal regulatory protein with three catalytic domains—two guanine exchange factor (GEF) do- mains, GEF1 and GEF2, and a kinase domain—as well as several accessory domains that have not been extensively studied. Function-damaging variants in the TRIO gene are known to be enriched in individuals with neurodevelopmental disorders (NDDs). Disease variants in the GEF1 domain or the nine adjacent spectrin repeats (SRs) are enriched in NDDs, suggesting that dysregulated GEF1 activity is linked to these disorders. We provide evidence here that the Trio SRs interact intramolecularly with the GEF1 domain to inhibit its enzymatic activity. We demonstrate that SRs 6-9 decrease GEF1 catalytic activity both in vitro and in cells and show that NDD- associated variants in the SR8 and GEF1 domains relieve this autoinhibitory constraint. Our results from chemical cross- linking and bio-layer interferometry indicate that the SRs pri- marily contact the pleckstrin homology region of the GEF1 domain, reducing GEF1 binding to the small GTPase Rac1. Together, our findings reveal a key regulatory mechanism that is commonly disrupted in multiple NDDs and may offer a new target for therapeutic intervention for TRIO-associated NDDs. The TRIO gene encodes a large (>300 kDa) multidomain protein with three catalytic domains (hence the name, Trio): two guanine nucleotide exchange factor (GEF) domains, each composed of Dbl homology (DH) and pleckstrin homology (PH) regions, and a putative serine/threonine kinase domain. The two GEF domains exhibit distinct substrate specificities: the more N-terminal GEF domain (GEF1) promotes GTP loading onto Rac1 and RhoG GTPases (1–3), while the more C-terminal GEF domain (GEF2) activates RhoA (1, 4, 5). Trio also contains an N-terminal lipid-binding Sec14 domain, nine spectrin repeat (SR) domains, and Src homology 3 and immunoglobulin-like domains (1, 6–9). Beyond the potential for protein–lipid and protein–protein interactions, the func- tions of these accessory domains remain poorly understood. ‡ These authors contributed equally to this work. * For correspondence: Anthony J. Koleske, anthony.koleske@yale.edu. De novo mutations and ultra-rare variants in TRIO are enriched in neurodevelopmental disorders (NDDs) (10–14) and the pattern of these variants differs in different disorders. For example, de novo missense and rare damaging variants in the GEF1 domain and adjacent regulatory SRs are enriched in autism, intellectual disability, and developmental delay, sug- gesting that dysregulated GEF1 activity contributes to the pathophysiology of these disorders. Indeed, our lab and others have shown that some of these variants disrupt the ability of GEF1 to catalyze Rac1 activation (12–15). Clusters of variants in the SR8 and GEF1 domains impacted cellular Rac1 activity in different ways and were associated with distinct endophenotypes in heterozygous carriers: SR8 domain variants were linked to developmental delay, mac- rocephaly, and hyperactive Rac1 activity in cells, whereas GEF1 domain variants were linked to mild intellectual disability, microcephaly, and reduced Rac1 activity in cells (15). However, the role of the SRs in Trio function and the mechanism of SR8 variant-mediated increase in Rac1 activity are unclear. Previous studies demonstrated that expression of Trio GEF1 increased Rac1 activity in cells and resulted in dominant gain-of-function pathfinding defects in fly retinal axons (16, 17). Appending additional regions of Trio, including the SRs, to GEF1 attenuated both Trio GEF1-dependent processes. These observations strongly suggest that the SRs reduce GEF1 activity in Trio. However, it remains unknown whether the SRs autoinhibit GEF1 activity directly or via the recruitment of cellular cofactor(s). It is also unclear how variants in the SRs would impact this regulatory mechanism in vitro and in cells. We provide evidence here that SRs 6-9 directly inhibit Trio GEF1 activity in vitro and in cells. Using a GDP-fluorescein (FL)-BODIPY nucleotide exchange assay (18), we show that inclusion of SRs 6-9 is sufficient to inhibit GEF1 activity in vitro, suggesting an autoinhibitory mechanism. We then find that NDD-associated variants in the SR8 and GEF1 domains increase GEF1 activity by relieving autoinhibition, whereas an NDD-associated variant in SR6 reinforces autoinhibition. Using interferometry, we chemical cross-linking and bio-layer J. Biol. Chem. (2022) 298(9) 102361 1 © 2022 THE AUTHORS. Published by Elsevier Inc on behalf of American Society for Biochemistry and Molecular Biology. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). Trio GEF autoinhibition by spectrin repeats demonstrate that the SRs make contact with the PH region of the GEF1 domain and reduce the affinity of GEF1 for Rac1. Together, our findings provide a novel RhoGEF regulatory mechanism by which SRs disrupt Trio GEF1 activation by reducing the interaction of Trio GEF1 with Rac1 and impairing catalytic efficiency. This mechanism appears to be commonly disrupted by NDD-associated variants in TRIO, making it a potential target for therapeutic intervention. Results Inclusion of SRs 6-9 reduces Trio GEF1 activity Genetic variants in SRs 6-9 are associated with NDDs (15), some of which were previously shown to affect Trio-mediated Rac1 activation in cells. To measure the impact of the SRs on GEF1 activity in vitro, we generated and purified Trio GEF1 alone (42 kDa) and a Trio fragment containing SRs 6-9 appended to the GEF1 domain (SR6-GEF1, 99 kDa) (Fig. 1A). Both proteins were monodisperse upon size-exclusion chro- matography and eluted at a position consistent with being monomers (estimated Stokes radius was 3.8 nm for GEF1, 5.6 nm for SR6-GEF1) (Fig. 1B). Using a fluorescence-based guanine nucleotide exchange assay, we measured the cata- lytic activity of GEF1 and SR6-GEF1. Purified 100 nM GEF1 efficiently catalyzed exchange of BODIPY-FL-GDP for GTP on Rac1, with a first-order dissociation rate constant kobs = 2.4 ± −1 (Fig. 1, C and D). Measurement of the rate 0.6 × 10 constant, kobs, as a function of GEF1 concentration yielded a −1 (Fig. 1, E and F). SR6-GEF1 kcat/KM = 1.9 × 104 M similarly promoted GTP exchange onto Rac1 but with a significantly reduced ((cid:1)20 fold and 6-fold, respectively) kobs = −1 (Fig. 1, C, D −1 and kcat/KM = 3.1 × 103 M 1.2 ± 1.8 ×10 and F). These data indicate that inclusion of SRs 6-9 inhibits Trio GEF1 activity for Rac1 in vitro. −1 s −3 s −1 s −4 s Figure 1. Inclusion of SRs 6-9 reduces Trio GEF1 activity on Rac1. A, schematic of Trio proteins: full-length Trio, SR6-GEF1, and GEF1. B, Trio SR6-GEF1 and GEF1 were purified and size-exclusion chromatography was performed to verify that proteins were monodisperse. Dotted lines indicate peak elution volume, which is used to calculate Stokes radii. Samples (approximately 5 μg) of purified components were analyzed by SDS-PAGE and stained with Coomassie Blue to assess purity. Gel images were spliced from separate lanes of the same gel, original gel shown in Figure 3B. C, 100 nM of Trio GEF proteins were incubated with 12.8 μM Rac1 preloaded with 3.2 μM BODIPY-FL-GDP, and nucleotide exchange was tracked via the decrease in fluorescence over time. Representative trace is shown here; traces in color, exponential fits overlaid in black. D, Trio SR6-GEF1 had approximately 20-fold lower exchange activity, kobs, than GEF1 alone. N = 21 independent kobs measurements for overall quantification of rates per group. Bars represent average ± SD; ****p ≤ 0.0001 in a two-tailed t test. E, GEF1 catalytic efficiency was determined by measuring the kobs of GEF1 at multiple concentrations (top) and extracting a linear fit from the plot of kobs versus GEF concentration. Sample traces shown with exponential fits overlaid in black. F, the catalytic efficiency of SR6-GEF1 was 6-fold lower than GEF1 (n = 4). Bars represent average ± SD of four experimental replicates; **p ≤ 0.005 in a two-tailed t test. DH1, Dbl homology domain; FL, fluorescein; GEF, guanine exchange factor; Ig, Ig-like domains; PH1, pleckstrin homology domain; SH3-1, Src homology 3 domain; SR, spectrin repeat. 2 J. Biol. Chem. (2022) 298(9) 102361 Trio GEF autoinhibition by spectrin repeats NDD-associated variants in SR8 increase Trio GEF1 activity in the context of SR6-GEF1 GEF1 variant D1368V increases GEF activity only in the context of SR6-GEF1 We generated and purified SR6-GEF1 expression constructs containing single NDD-associated variants in SR8 and measured their ability to catalyze nucleotide exchange on Rac1 (Fig. 2, A and B). When tested at 100 nM, all SR8 variants, except N1080I, increased the kobs by 4 to 8 fold over that of WT SR6-GEF1 (Fig. 2, C and D). In agreement with these findings, one representative SR8 variant, SR6-GEF1R1078Q, −1, which had a significantly increased kobs = 1.0 ± 0.5 × 10 −1, a 1.5-fold increase in cat- −1 s had a kcat/KM = 4.7 × 103 M alytic efficiency over WT SR6-GEF1 (Fig. 2E). These findings indicate that NDD-associated variants in SR8 are sufficient to relieve SR autoinhibition. −3 s NDD-associated variants in SR6 decrease GEF1 activity in the context of SR6-GEF1 We also generated two SR6-GEF1 constructs harboring individual disease variants in the SR6 domain. While the rate constant (kobs) values obtained for each construct did not significantly decrease compared to WT SR6-GEF1, measure- ment of catalytic efficiency, kcat/KM, of both WT SR6-GEF1 and SR6-GEF1E883D revealed that SR6-GEF1E883D had a significantly decreased catalytic efficiency of a kcat/KM = 1.7 × −1, 1.8-fold lower than WT SR6-GEF1 (Fig. 2E). This 103 M suggests that NDD-associated variants in SR6 decrease GEF1 activity. −1 s Hypothesizing that the SRs might contact GEF1 to impact catalytic activity, we searched for GEF1 domain variants that might impact potential autoinhibition of GEF1 activity by SRs. lie in the GEF1:Rac1 Unlike GEF1 disease variants that interface and decrease GEF1 activity (12–14), D1368V lies in the DH domain but is distal to the GEF1:Rac1 interface, so its impact is less well understood (Fig. 3A). However, introduc- tion of the D1368V variant greatly potentiates the ability of the Trio9 splice isoform, which contains all of the SRs, to increase activity of a Rac1 reporter in cells (14). We intro- duced D1368V into SR6-GEF1 and found that it significantly −1 increased catalytic activity, with a kobs = 1.4 ± 0.3 × 10 −1 (Fig. 3, B–E), a 1.5-fold in- −1 s and kcat/KM = 4.8 × 103 M crease over the kcat/KM for WT SR6-GEF1. In contrast, introducing D1368V into GEF1 alone did not impact its ac- tivity compared to GEF1 (Fig. 3, B–E), indicating that the activating effects of D1368V require SRs 6-9. Together with data reported above, these are consistent with a model in which NDD-associated variants in SR8 and GEF1 relieve in- hibition of GEF1 activity by the SRs. −3 s The SRs and GEF1 form distinct stable interacting domains We used AlphaFold (19, 20) to model human Trio SR6- GEF1 (Fig. 4, A and B). Strikingly, this model suggests that SRs interact with the GEF1 domain, with SR8 closely apposed Figure 2. Mutations in SR6 and SR8 differentially impact GEF1 activity. A, schematic of disease associated mutations in the SRs used in this study. B, mutants were generated in the context of SR6-GEF1 and purified. C, sample GEF assay traces of SR6-GEF1E883D and SR6-GEF1R1078Q. Traces in color, exponential fits overlaid in black. D, SR8 variants in SR6-GEF1 have significantly enhanced catalytic rates, kobs, at equal molar amounts (100 nM) (except N1080I). **p ≤ 0.005; ***p ≤ 0.001; ****p ≤ 0.001 for a significant difference compared to SR6-GEF1 in a one-way ANOVA adjusted for multiple comparisons (n ≥ 9). E, catalytic efficiency (kcat/KM) of representative SR6/8 mutants was determined by measuring the kobs values at different concentrations of GEF, as shown in Figure 1D. The catalytic efficiency of SR6-GEF1R1078Q is (cid:1)1.5-fold greater than that of SR6-GEF1, while the catalytic efficiency of SR6-GEF1E883D is (cid:1)1.8-fold slower (n = 3). Data for GEF1 and SR6-GEF1 from Figure 1 are shown again for reference, and all are reported as an average ± SD of three or more experimental replicates. * = significantly different from SR6-GEF1, p ≤ 0.05 in a one-way ANOVA adjusted for multiple comparisons. GEF, guanine exchange factor; SR, spectrin repeat. J. Biol. Chem. (2022) 298(9) 102361 3 Trio GEF autoinhibition by spectrin repeats Figure 3. GEF1 variant D1368V increases GEF1 activity in the context of SR6-GEF1. A, crystal structure of Trio GEF1 (light and dark blue) and Rac1 (gray), accessed in PDB, ID = 2NZ8 (5). D1368, identified in the box, is distal to the Rac1-binding interface. B, samples (approximately 5 μg) of purified components were analyzed by SDS-PAGE and stained with Coomassie Blue R250 to assess purity. Gel bands for WT SR6-GEF1 and WT GEF1 are the same as shown spliced in Figure 1B. C, sample GEF assay traces of D1368V in the context of SR6-GEF1 and GEF1. Traces in color, exponential fits overlaid in black. D, D1368V in SR6-GEF1 increases catalytic rate, kobs, at equal molar amounts of GEF but has no impact when inserted into GEF1 alone (****p ≤ 0.0001, unpaired t test for mutant versus WT in respective GEF1 or SR6-GEF1, n = 3). E, catalytic efficiency (kcat/KM) of SR6-GEF1D1368V was determined by measuring the kobs values at different concentrations of GEF, as in Figure 1D. Data for GEF1 and SR6-GEF1 shown again for reference. The catalytic efficiency, kcat/KM, of SR6- GEF1D1368V is (cid:1)1.5-fold greater than that of SR6-GEF1 (n = 3). * = significantly different from SR6-GEF1, p ≤ 0.05 in a one-way ANOVA adjusted for mul- tiple comparisons. GEF, guanine exchange factor; SR, spectrin repeat. to GEF1 and the NDD-associated mutations concentrated at this SR8:GEF1 interface. This model of SR6-GEF1 and addi- tional analysis using DISOPRED predicted the existence of an unstructured loop between SR9 and GEF1, suggesting this flexible region may connect the SRs and GEF1 domain (Fig. 4C) (21). We used limited proteolysis to probe for the presence of a flexible linker between SR9 and the GEF1 domain that might be susceptible to partial proteolysis. Treatment of SR6-GEF1 at intermediate levels of trypsin yielded two major bands, identified by mass spectrometry as composed of SRs 6-9 and GEF1, respectively. This observation indicates that SRs 6-9 and the GEF1 domain each make up distinct folding units with increased relative resistance to protease (Fig. 4D). Together, these findings support a model in which the SRs make contact with GEF1. To test directly for possible interactions between the SRs and GEF1 domain, we incubated SR6-GEF1 with an 11.4 Å spacer lysine cross-linker, BS3 (bis(sulfosuccinimidyl)suberate), and analyzed cross-linked peptides via mass spectrometry to identify sites in close enough proximity to cross-link. Several long-distance cross-links were observed between the SRs and the GEF1 domain (Fig. 5A). Specifically, the SR:GEF1 interface includes a peptide in DH domain which is directly at the Rac1 binding interface (1429–1438, green in Fig. 5A) and a peptide in the PH domain important for stabilizing the Rac1 interaction (1529–1537, orange in Fig. 5A) (Fig. 5A) (5). Multiple regions originating in SR6-9 contact these peptides in the GEF domain. This suggests that SR6-GEF1 may be dynamic, with multiple conformational states captured by cross-linking. We hypothe- size that these SR:GEF1 contacts likely disrupt Rac1 binding to GEF1 We also performed chemical cross-linking on three variants in SR6-GEF1 to understand how intramolecular contacts may change in the variants. The SR6-GEF1 variants that display activated GEF activity, R1078Q and D1368V, both exhibited a loss of contact between SR6, 7, 9, and the GEF1 domain (Fig. 5B). In addition, R1078Q, but not D1368V, also reduced SR8:GEF1 contacts (Table S2). In contrast, the SR6 variant, E883D, which reduced GEF activity, did not reduce intra- molecular contacts with GEF1; in fact, new contacts appeared (SR7 and SR9 contacts, blue and purple arrowheads, Fig. 5B), suggesting this variant may reinforce intramolecular SR:GEF contacts (Fig. 5B). These data are consistent with a model in which specific intramolecular contacts between the SRs and GEF1 are altered in genetic variants with increased GEF1 activity. 4 J. Biol. Chem. (2022) 298(9) 102361 Trio GEF autoinhibition by spectrin repeats Figure 4. AlphaFold predicts an interaction between the SRs and GEF1, which form independent folding units. A, AlphaFold model of human Trio SR6-GEF1. SR6, 8 in light pink, SR7, 9 in dark pink, linker region in gray, and GEF1 in blue. Sites of mutations used in this study are modeled as black spheres, with amino acids labeled. This model predicts an interaction between SR8 and GEF1. B, SR6-GEF1 from AlphaFold model, rotated to view flexible linker region between GEF1 and SR9. C, probability of disorder was predicted using DISOPRED. The region between SR9 and DH1 has a high probability of being disordered (cutoff > 0.5). D, limited proteolysis of SR6-GEF1. His-SR6-GEF1 was incubated with increasing concentrations of trypsin and select bands were identified using mass spectrometry. Relative abundance of identified peptides was plotted to determine composition of each band. The y-axis displays relative abundance of peptides and x-axis is ‘amino acid position’, which refers to the location in SR6-GEF1 that the peptide covers (with SR6-GEF1 diagram below). Band 1 (pink box around gel band at (cid:1)60 kDa) comprises SR6-9 and Band 2 (blue box around band at (cid:1)40 kDa) comprises GEF1. Therefore, SR6-9 and GEF1 form distinct stable domains. DH1, Dbl homology domain; GEF, guanine exchange factor; SR, spectrin repeat. The SRs reduce GEF1 binding to Rac1 Based on our cross-linking data, we hypothesized that an interaction between SRs 6-9 and PH1 may impair the ability of GEF1 to bind Rac1. We used bio-layer interferometry to measure the association of nucleotide-free Rac1 with His- GEF1 or His-SR6-GEF1 immobilized on a nitrilotriacetic acid (Ni-NTA) affinity chip. GEF1 bound to Rac1 with a Kd = 151 ± 49 nM in nucleotide-free conditions (Fig. 6, A–C). SR6- GEF1 had a reduced affinity for Rac1, with a Kd = 316 ± 87 nM (Fig. 6, A–C). Taken together with the cross-linking data, this supports a model where the SRs contact the PH domain to impair GEF1 binding to Rac1, which likely contributes to the reduction in observed GEF1 activity. SRs 6-9 inhibit GEF1-induced cell spreading Trio GEF1 activates Rac1 and RhoG to coordinate down- stream cytoskeletal changes and mediate changes in cell morphology (1–3, 22). We first expressed Trio GEF1-GFP in HEK293 cells and quantified its impact on cell morphology (Fig. 7, A–C). When matched for GFP expression levels, GEF1 expressing cells had significantly increased cell area compared to GFP controls (Fig. 7, A–C). Cells expressing GEF1 appeared to be more spread with round lamellipodia encompassing the cell edge, a common result of Rac1 activation (23) (Fig. 7B). The area of cells expressing a catalytic-dead mutant of GEF1, GEF1 ND/AA (N1465A/D1466A), were similar to GFP con- trols, indicating a key role for GEF1 catalytic activity in this morphological change (24). In contrast to GEF1, SR6-GEF1 expressing cells had no measurable effect on cell area, but the SR8 mutant, SR6-GEF1R1078Q, increased cell area over that of GFP and SR6-GEF1 WT (Fig. 7, B and C). Cells expressing SR6-GEF1R1078Q also appeared qualitatively in morphology to those cells expressing GEF1 alone, with more full, rounded edges (Fig. 7B). Therefore, inclusion of SRs 6-9 inhibits Trio GEF1-dependent changes in cell morphology, and disease-associated variants can disrupt this inhibitory regulation. similar We then expressed GFP-Trio9s, a predominant neuronal isoform throughout neurodevelopment, in HEK293 cells and quantified its impact on cell morphology (25) (Fig. 7, A, D and E). Interestingly, when matched for GFP expression levels, GFP-Trio9s expressing cells had significantly decreased cell area compared to GFP controls. Expressing two variants of Trio9s, the most activated SR8 mutant, GFP-Trio9sR1078Q, and a catalytic-dead mutant of GEF1, GFP-Trio9s ND/AA (N1465A/D1466A), decreased cell area compared to GFP alone (Fig. 7, D and E). Cells expressing any variant of J. Biol. Chem. (2022) 298(9) 102361 5 Trio GEF autoinhibition by spectrin repeats Figure 5. The SRs interact with GEF1. A, SR6-GEF1 was incubated with lysine cross-linker BS3 and cross-linked peptides were identified using mass spectrometry. Crystal structure of GEF1 alone (gray, left panel) and with Rac1 (black, right panel) (from PDB, ID = 2NZ8 (5)) with cross-linked peptides between SR6-9 and GEF1 (in WT case) shown in green (1429–1438), pink (1503–1506), orange (1529–1537), purple (1562–1588), and light blue (1574–1588). SR6-9 contacts the DH domain at a peptide that likely interferes with Rac1 binding (1429–1438) and a region in the PH domain critical for stabilizing the Rac1 interaction (1529–1537) (5). B, representative activating mutants (R1078Q and D1368V) display fewer contacts between SR6-9 and GEF1 (lost contacts shown with dotted lines). Representative inactivating mutant (E883D) displays increased contacts between SR6-9 and GEF1 (New contacts shown with blue or purple arrows). Cross-links were categorized based on their N-terminal cross-link site (in SR6, 7, or 9) and their C-terminal GEF1 contacts were visualized. For the activating mutants, the peptides that were mutually lost for both activating mutants were visualized here. For table of all mutant cross-links between SR6-9 and GEF1, see Table S2. BS3, bis(sulfosuccinimidyl)suberate; DH1, Dbl homology domain; GEF, guanine exchange factor; PH1, pleckstrin homology domain; SR, spectrin repeat. GFP-Trio9s appeared very round, completely lacking lamelli- podia or cell edge protrusions (Fig. 7, D and E). We speculate that activity of the Trio GEF2 domain, which targets RhoA to promote cytoskeleton contractility (26), may dominate in this context, making it difficult to discern specific effects on GEF1 activity. interferometry, we show that the SRs contact regions of GEF1 important for Rac1 binding and that inclusion of the SRs is associated with reduced binding affinity for Rac1 in vitro. We present a model for how Trio GEF1 activity is regulated, and how this regulation is disrupted by disorder-associated variants. Discussion Inclusion of Trio SRs autoinhibits GEF1 activity in vitro We provide evidence here that the Trio SRs 6-9 directly inhibit GEF1 activity via intramolecular interactions in vitro and in cells. We demonstrate that NDD-associated variants in the SR8 and GEF1 domains release this autoinhibitory constraint, strongly suggesting that disruption of this GEF1 regulatory mechanism contributes to the pathophysiology of these disorders. Using chemical cross-linking and bio-layer Previous cell-based studies have shown that removing the SRs is associated with increased downstream Rac1 activity and Trio gain-of-function phenotypes in vivo, suggesting that the Trio SRs function to inhibit GEF1 activity (16, 17, 27). This hypothesis is supported by evidence that other RhoGEFs, like Tiam1, contain autoinhibitory N-terminally adjacent accessory domains (8, 28, 29). In most cases, how inhibition occurs and 6 J. Biol. Chem. (2022) 298(9) 102361 Trio GEF autoinhibition by spectrin repeats Figure 6. Inclusion of SRs 6-9 reduce binding to Rac1. A, His-GEF1 or His-SR6-GEF1 were immobilized on an Ni-NTA biosensor and the association of different concentrations of Rac1 was measured. Representative traces shown, with data in color and one phase exponential fits in black. Full concentration gradients (4–5 Rac1 concentrations) were performed at least three independent times. B, kobs values were extracted from each association curve and plotted against Rac1 concentration to calculate a Kd of GEF1 or SR6-GEF1 binding to Rac1. C, SR6-GEF1 has a 2-fold weaker affinity for Rac1 than GEF1 (*p ≤ 0.05, unpaired t test). GEF, guanine exchange factor; Ni-NTA, nitrilotriacetic acid; SR, spectrin repeat. how it is released to activate GEF activity is unknown. Our results show that SR6-GEF1 is monomeric in solution and that inclusion of SRs significantly decreases GEF1 catalytic activity in vitro. Collectively, these observations suggest that the SRs are sufficient to inhibit GEF1 activity via intramolecular in- teractions in cis. SRs make direct contact with GEF1 and impair interactions with Rac1 Within GEF1, the DH1 domain catalyzes GTP exchange onto Rac1 and serves as the main Rac1-binding interface. The PH domain plays a regulatory role in catalysis but also serves to stabilize the Rac1:DH1 interaction (30, 31). Using chemical cross-linking, we demonstrate that SRs 6-9 make extensive contacts with the GEF1 domain, including at sites critical for Rac1 binding, suggesting that SR6-9 sterically blocks contact with Rac1. In addition, NDD-associated variants that activate GEF1 exhibit reduced contacts between the SRs and GEF1 and those that impair GEF1 activity exhibit increased contacts. Hence, altering the interaction between the SRs and GEF1 impacts catalytic activity (5). We found that inclusion of SRs 6-9 reduces the affinity of GEF1 for Rac1 by 2-fold, compared to GEF1 alone. Whereas our catalytic rate measurements suggest the presence of SRs 6-9 results in a 6-fold decrease in activity, the reduction in affinity that we observed was smaller in magnitude. It is likely that engagement of the SRs with GEF1 impairs other steps in the catalytic cycle, as demonstrated by our catalytic efficiency data, in addition to impacting Rac1-binding affinity. Future studies will elucidate whether other components of the nucleotide exchange process are impacted by the SRs. NDD-associated mutations in SR8 and GEF1 disrupt SR-mediated GEF1 inhibition Two rare variant clusters in TRIO, one in SR8 (Fig. 2A) and one in GEF1, have been linked to distinct endophenotypes in individuals with NDDs (15). For example, TRIO SR8 variants are linked to developmental delay and macrocephaly in humans and cause increased Rac1 (GEF1) activity in cells, whereas most mutations in the GEF1 domain are linked to mild intellectual disability, microcephaly, and reduced Rac1 activity in cells. However, how SR8 variants increased Rac1 activity was completely unknown. We hypothesized that the increased Rac1 activity associated with SR8 domain variants resulted from disruption of SR-mediated GEF1 inhibition. We generated mutant SR6-GEF1 constructs harboring distinct disorder-associated variants and found that nearly all SR8 mutants increased SR6-GEF1 catalytic activity 4 to 8 fold. Interestingly, the one exception, N1080I, disrupts binding to neuroligin-1 and blocks neuroligin-1–mediated synapto- genesis (32). We hypothesize that other sites, including N1080I, in the SRs serve as convergence points for upstream activators to regulate GEF1 activity and discuss this in a following section. Together, these data demonstrate that many J. Biol. Chem. (2022) 298(9) 102361 7 Trio GEF autoinhibition by spectrin repeats Figure 7. SRs 6-9 reduce the impact of GEF1 on cell spreading. A, schematic of constructs used, with mutants shown below. B, constructs in (A) were transfected into HEK293 cells and plated on fibronectin. Cells were fixed and stained using anti-GFP to visualize GFP expression and cell morphology. Cells expressing GEF1 and SR6-GEF1R1078Q appeared to have more rounded edges and circular shapes. The scale bar represents 10 μm. Contrast was adjusted between images shown to best visualize cell edge; cell edge is outlined with a white dashed line. C, cell area, normalized to protein expression on a cell-by- cell basis, was quantified. Cell area increased upon expression of GEF1 and SR6-GEF1R1078Q, while expression of a catalytic-dead GEF1 mutant (ND/AA) or SR6-GEF1 had no effect compared to GFP alone. D, cells visualized and analyzed as in (B). The scale bar represents 10 μm. Cells expressing Trio9s constructs all appeared rounder and lacked cell edge protrusions. E, cell area quantified as in (C). Cell area was decreased upon expression of all GFP-Trio9s constructs compared to GFP alone. Trio9sR1078Q did not increase cell area to levels seen with GFP alone. Two biological replicates were performed for each set of constructs, with 25 to 40 cells analyzed per group per replicate (*p ≤ 0.05, ****p ≤ 0.0001, one-way ANOVA between GFP control and each group and adjusted for multiple comparisons). GEF, guanine exchange factor; SR, spectrin repeat. NDD variants in SR8 are sufficient to relieve SR-mediated GEF1 inhibition. We also found that a GEF1 domain variant associated with Rac1 activation in cells likely impacts SR-mediated GEF1 in- hibition. Unlike GEF1 disease variants that lie at the Rac1- binding interface and decrease GEF1 activity, this variant, D1368V, is distal to the Rac1 interface and hyperactivates Rac1 activity in cells when introduced in the Trio9 splice isoform (12–14, 32). Our results indicate that D1368V significantly increases GEF1 activity in the context of SR6-GEF1 but has no effect on GEF1 alone. We propose that D1368V enhances SR6- GEF1 activity by disrupting SR autoinhibition. Indeed, our cross-linking data suggests that contacts between the SRs and GEF1 are reduced for the D1368V variant. NDD-associated variants in SR6 may reinforce SR-mediated GEF1 inhibition We also generated two SR6-GEF1 constructs harboring individual disease variants in the SR6 domain, whose impact on Trio function remains completely unknown. The catalytic efficiency (kcat/KM) of SR6-GEF1E883D was significantly slower than SR6-GEF1, suggesting that SR6 mutants decrease SR6-GEF1 catalytic activity. While the mechanism for this is unclear, one possibility is that SR6 acts as a hinge region allosterically governing the flexibility of the helices surround- ing SR8 and that SR6 variants may decrease the ability for the SRs to release their inhibitory lock on the GEF1 domain. Indeed, we observed more contacts between SR7 and SR9 and the GEF1 domain in SR6-GEF1E883D, suggesting that the intramolecular contacts are more stable or extensive in the variant case. This observation underscores the importance of understanding how dysregulation of Trio GEF1 activity con- tributes to NDDs. The SRs may serve as a target for activators of Trio GEF1 activity We demonstrated that the SRs inhibit Trio GEF1 activity, but it is unclear how inhibition may be released in a cellular context. SR domains are widely accepted as scaffolding pro- teins that coordinate cytoskeletal interactions with high spatial precision. Considering that Trio is known to act downstream of cell surface receptors to coordinate cytoskeletal rearrange- ments, we anticipate that the Trio SRs serve as a target of interaction partners to engage and activate Trio GEF1 activity in cells. Trio SRs interact with diverse cellular partners, including synaptic scaffolding proteins (Piccolo and Bassoon) (33), cell-adhesion molecules (VE-cadherin and Intercellular Adhesion Molecule 1 (ICAM1)) (34, 35), and membrane 8 J. Biol. Chem. (2022) 298(9) 102361 trafficking proteins (RABIN8) (36). These SR-binding partners may engage Trio to coordinate GEF1 activation and/or deac- tivation in a spatiotemporal manner. Indeed, several studies have shown that Trio interactions with binding partners im- pacts Rac1 activity in cells (32, 34, 35, 37, 38). For example, VE-cadherin binds Trio SR5 and SR6, and this interaction locally increases Rac1 activity in cells (34). Similarly, the ICAM1 intracellular tail binds Trio GEF1, and the Trio/ ICAM1 interaction potentiates ICAM1 clustering at adhesion sites, promoting Rac1 activation in cells (35). Finally, the in- tegral membrane protein Kidins220 regulates Rac1-dependent neurite outgrowth via interactions with the Trio SRs (37). While these studies suggest that the Trio signaling partners may engage and activate Trio GEF1 activity, the specific interaction interfaces and binding stoichiometry that mediates GEF1 activation and how they are impacted by disorder- associated variants is presently unknown. Based on our evi- dence that SR8 variants relieve autoinhibitory constraint, we anticipate that SR8 may be a convergence point for upstream activators and coordinated regulation of GEF1 activity. Conclusions TRIO has emerged as a significant risk gene for NDDs. Using biochemical and genetic tools, we identified a novel regulatory mechanism by which Trio SRs inhibit GEF1 activity and showed that disorder-associated variants are sufficient to relieve this autoinhibitory constraint. This discovery will serve as a model to understand how Trio GEF1 is regulated by physiological signals and how its disruption leads to NDDs. This mechanism may also offer a new target for therapeutic interventions for TRIO-associated NDDs. Experimental procedures Expression construct cloning and protein purification Human Trio SR6-GEF1 was PCR amplified and inserted into the pFastBac1 HTa vector (Invitrogen). Site-directed mutagenesis was used to insert point mutations into pFast- Bac1-Hta-SR6-GEF1 construct and confirmed by DNA sequencing. Primers used for cloning are included in Table S1. Recombinant baculoviruses were generated using Sf9 cells (Bac-to-Bac expression system, Thermo Fisher Scientific). Baculoviruses were used to infect Hi5 cells at an estimated multiplicity of infection = 1 for 48 h before lysis in lysis buffer (20 mM Hepes pH 7.25, 500 mM KCl, 5 mM β-mercaptoe- thanol, 5% glycerol, 1% TritonX-100, 20 mM imidazole, 1 mM DTT, 1 mM PMSF, 1× Roche cOmplete protease inhibitors EDTA free) for 20 min at 4 (cid:3)C. Lysates were affinity purified using Ni-NTA resin (Qiagen) and eluted with 250 mM imid- azole. Elution fractions were further purified over an Sephadex 200 (S200) Increase 10/300 GL column into assay buffer (20 mM Hepes pH 7.25, 150 mM KCl, 5% glycerol, 0.01% TritonX-100, 1 mM DTT), aliquoted, and flash frozen for long-term storage. Human Trio GEF1 and Rac1 were generated and affinity purified from bacterial cells as described in Blaise et al. (18). Point mutants were generated using site-directed mutagenesis. Trio GEF autoinhibition by spectrin repeats Following affinity purification, eluted protein was further pu- rified over an S200 Increase column into assay buffer, ali- quoted, and flash frozen for long-term storage. Stokes radii of proteins were estimated based on the elution volume from the S200 Increase column, calculated based on a standard curve generated by running protein standards (Pro- tein Standard Mix 15–600 kDa, Supelco). BODIPY-FL-GDP nucleotide exchange assays 12.8 μM Rac1 was loaded with 3.2 μM BODIPY-FL-GDP (Invitrogen) in 1× assay buffer (20 mM Hepes pH 7.25, 150 mM KCl, 5% glycerol, 1 mM DTT, 0.01% TritonX-100) plus 2 mM EDTA to a total volume of 25 μl per reaction, then incubated for 1 h at room temperature. BODIPY-FL-GDP loading onto Rac1 was halted by the addition of 5 μl of MgCl2, for a total reaction volume of 30 μl with a final MgCl2 con- centration of 5 mM. Prior to initiating the reaction with 100 nM Trio GEF, 30 μl of GTPase (12.8 μM) plus MgCl2 (5 mM) mix or blank (3.2 μM BODIPY-FL-GDP, 2 mM EDTA, and 1× assay buffer) was added to appropriate wells. During the BODIPY-FL-GDP loading incubation period, GEF1- containing proteins were prepared in 1× assay buffer, 4 mM GTP, and 2 mM MgCl2. Exchange reactions were initiated by adding 10 μl of 100 nM Trio GEF mixture (as stated above) to each well, for a total reaction volume of 40 μl. Real-time fluorescence data was measured every 10 s for 30 min moni- toring BODIPY-FL fluorescence by excitation at 488 nm and emission at 535 nm, as per Blaise et al. (18). All kobs measurements of GEF1 activity represent at least three experimental replicates with three technical replicates per experiment. Results are shown as the mean ± SD from multiple experiments. A one-way ANOVA was used to determine statistical significance between SR6-GEF1 and all other variants (two-tailed p-value < 0.05) and adjusted using Dunnett’s multiple comparisons test. Catalytic efficiencies (kcat/KM) of selected SR6-GEF1 constructs were extracted −1) versus GEF1 con- from a linear fit of catalytic rate (kobs, s centration (nM). Three experimental replicates were per- formed for each SR6-GEF1 construct, and the catalytic efficiency values were averaged. Results are shown as the mean ± SD. A one-way ANOVA was used to determine sta- tistical significance between SR6-GEF1 and all other variants (two-tailed p-value < 0.05) and adjusted using Dunnett’s multiple comparisons test. Protein structure predictions AlphaFold was used to access the predicted structure of human Trio spectrin repeats 1-GEF1 (amino acids 201–1600), entry number AF-O75962-F2 (19, 20). Swiss pdb Viewer was used to model SR6-GEF1, amino acids 788 to 1599 (39). DISOPRED was used to predict the probability of disorder of Trio SR6-GEF1, amino acids 788 to 1599 (21). Limited proteolysis SR6-GEF1 in assay buffer plus 10 mM CaCl2 was diluted to 0.4 mg/ml and incubated with increasing concentrations of J. Biol. Chem. (2022) 298(9) 102361 9 Trio GEF autoinhibition by spectrin repeats trypsin (0.001 mg/ml–0.11 mg/ml) for 1 h at room tempera- ture in a 25 μl total reaction volume. Reactions were quenched with 8 μl quench buffer (50 mM Tris–HCl pH 6.8, 4% SDS, 10% glycerol, 0.1% bromophenol blue, 5% β-mercaptoethanol, 1 mM PMSF, 4 mM EGTA, 4 mM EDTA) and immediately boiled for 10 min. Samples were immediately run on a 12% SDS-PAGE gel, and proteins were visualized by Coomassie R250 staining. Major gel bands were excised and washed with 50:50 ace- tonitrile:water buffer containing 100 mM ammonium bicar- bonate. Proteins in the gel were reduced with 4.5 mM DTT at 37 (cid:3)C for 20 min and alkylated with 10 mM iodoacetamide at room temperature for 20 min in the dark. Gel bands were washed twice with 50:50 acetonitrile:water containing 100 mM bicarbonate and dried for 10 min in a SpeedVac. Trypsin digestion was carried out (1:100 M ratio of trypsin to protein) by incubation with the gel piece at 37 (cid:3)C overnight. The digest samples were analyzed by LC–MS/MS using a Q-Exactive Plus mass spectrometer equipped with a Waters nanoACQUITY ultra-performance liquid chromatography system using a Waters Symmetry C18 180 μm by 20 mm trap column and a 1.7 μm (75 μm inner diameter by 250 mm) nanoACQUITY ultra-performance liquid chromatography column (35 (cid:3)C) for peptide separation. Trapping was done at 15 μl/min with 99% buffer A (100% water, 0.1% formic acid) for 1 min. Peptide separation was performed at 300 nl/min with buffer A and buffer B (100% acetonitrile, 0.1% formic acid) over a linear gradient. High-Energy collisional dissociation was utilized to fragment peptide ions via data-dependent acquisition. Mass spectral data were processed with Proteome Discoverer (v. 2.3) and protein database search was carried out in Mascot search engine (Matrix Science, LLC; v. 2.6.0). Protein searches were conducted against the Trichoplusia ni protein database and the human Trio SR6-GEF1 sequence. Mascot search parameters included the following: parent peptide ion tolerance of 10.0 ppm; peptide fragment ion mass tolerance of 0.020 Da; strict trypsin fragments (enzyme cleavage after the C terminus of K or R, but not if it is followed by P); fixed modification of carbamidomethyl (C); and variable modification of phospho (S, T, Y), oxidation (M), and propioamidation (C), and dea- midation (NQ). Peptide identification confidence was set at 95% confidence probability based on Mascot MOWSE score. Results were transferred to Scaffold software (Proteome Soft- ware; v. 4) for further data analysis to look at peptide abun- dances in reference to their start position. These were utilized to plot in a frequency distribution to determine band identity. Cross-linking mass spectrometry Cross-linking experiments were performed as in Sanchez et al. (40) with deviations noted below. Twenty five micro- grams of protein was incubated in assay buffer with 100 μM BS3 (Thermo Fisher) for 30 min on ice. The reaction was quenched by adding Tris pH 7.25 to 10 mM final concentra- tion. Protein was then acetone precipitated and the pellet was alkylated with iodoacetamide and digested with trypsin. Pep- tides were desalted on a 100 μl Omix C18 tip (Agilent), dried, 10 J. Biol. Chem. (2022) 298(9) 102361 and reconstituted in 100 μl of 0.1% formic acid. Mass spec- trometry was performed on an Orbitrap Exploris 480 equipped with an EasySpray nanoESI source, an EasySpray 75 μm × 15 cm C18 column, and a FAIMS Pro ion mobility interface coupled with an UltiMate 3000 RSLCnano system (Thermo Scientific). Each sample was analyzed at four different FAIMS compensation voltages (CV = −40 V, −50 V, −60 V, −70 V) to provide gas-phase enrichment/fractionation of cross-linked peptide ions (41). Each analysis was a separate injection (2.5 μl sample). The sample was loaded at 2% B at 600 nl/min for 35 min followed by a multisegment elution gradient to 35% B at 200 nl/min over 70 min with the remaining time used for column washing and reequilibration (buffer A: 0.1% formic acid (aq); buffer B: 0.1% formic acid in acetonitrile). Precursor ions were acquired at 120,000 resolving power, and ions with charges 3 to 8+ were isolated in the quadrupole using a 1.6 m/z unit window and dissociated by HCD at 30% NCE. Product ions were measured at 30,000 resolving power. Peak lists were generated using PAVA (in house Python app), searched with Protein Prospector v6.3.23 (42), and classified as unique res- idue pairs using Touchstone (an in-house R library) at SVM.score ≥1.5 corresponding to a residue pair level FDR < 0.1% and then further summarized and presented as domain-domain pairs using Touchstone. A custom database consisting of the human Trio construct and a 10× longer decoy database (11 sequences total) was used in the Prospector search, using tryptic specificity with 2-missed cleavages and tolerance of 10/25 ppm (precursor/product). DSS/BS3 cross- linking was specified. Bio-layer interferometry Kinetic binding assays were performed using a ForteBio BLItz instrument. Ni-NTA biosensors were prehydrated in assay buffer for 10 min prior to the experiment. Biosensors were first measured for a baseline signal for 30 s before loading His-GEF1 (0.5 μM) or SR6-GEF1 (2 μM) in assay buffer for 5 min (con- centrations were optimized for reproducible biosensor loading and signal change). Biosensors were then re-equilibrated in assay buffer for 30 s before introducing varying concentrations of Rac1 (at least four concentrations per experiment) in assay buffer for 5 min to measure association. Association curves were fit to a one phase exponential curve to obtain a kobs value and these values were plotted against Rac1 concentration to calculate a Kd from the linear fit of this line, where the y-intercept = koff and slope = kon (Kd = koff/kon). Concentration gradients were replicated at least three times independently, and the Kd measurements of each interaction were compared using an unpaired t test. Reported values are mean ± SD. Measurement of GEF and SR6-GEF1 impact on cell morphology PEI was used to transfect HEK293 cells with 0.5 to 4 μg of DNA in 6-well dishes at a density of 3 × 105 cells per well. Twenty four hours after transfection, cells were trypsinized and replated at a density of 2.5 × 104 cells per coverslip on fibronectin-coated coverslips (10 μg/ml fibronectin). Twenty four hours post plating, cells were fixed and stained as in Lim et al. (43). Cells were fixed for 5 min in 2% paraformaldehyde in cytoskeleton buffer (10 mM MES pH 6.8, 138 mM KCl, 3 mM MgCl2, 2 mM EGTA, 320 mM sucrose). Cells were rinsed three times in Tris Buffered Saline (TBS) (20 mM Tris pH 7.4, 150 mM NaCl) and incubated with 5 μg/ml Alexa Fluor Wheat Germ Agglutinin 555 in TBS (Thermo Fisher) for 10 min to visualize the cell membrane when imaging. Cells were washed another three times in TBS, then permeabilized for 10 min in 0.3% TritonX-100/TBS and washed another three times in 0.1% TritonX-100/TBS. Cells were blocked for 30 min in antibody dilution buffer (ADB) (0.1% TritonX-100, 2% bovine serum albumin, 0.1% NaN3, 10% fetal bovine serum, TBS) and incu- bated with primary antibody (ADB containing a 1:2000 dilution of Goat Anti-GFP, Rockland) at 4 (cid:3)C overnight. The next morning, cells were washed in 0.1% TritonX-100/TBS three times and incubated in secondary antibody for 1 h at room temperature (in ADB, 1:2000 Alexa Fluor 488 Donkey Anti- Goat, Abcam). Cells were washed once in 0.1% TritonX-100/ TBS, once in TBS, and then mounted onto glass slides using AquaMount (Lerner Laboratories). After drying, coverslips were sealed using clear nail polish and imaged using a 40× objective on a spinning disk confocal microscope (UltraVIEW VoX spinning disk confocal (PerkinElmer) Nikon Ti-E-Eclipse), collecting a full z-stack of images for each cell. Identical mi- croscope settings were used between imaging samples. After imaging cells, images were processed using Fiji/ImageJ (44) to generate a sum projection of the GFP channel for quantifying fluorescence as a proxy for total protein expres- sion. Images were then analyzed using CellProfiler to semi- automatically detect cell edges and compute cell area (45). Cell area was normalized for protein expression on a single cell basis by dividing the total area of the cell by the total GFP fluorescence of the cell (a proxy for total protein expression). Two biological replicates were performed, with 25 to 40 cells quantified per group per replicate. Statistical significance of differences in the normalized cell area was determined using a one-way ANOVA between the GFP control and all other groups (two-tailed p-value < 0.05) and adjusted using Dun- nett’s multiple comparisons test. Data availability Data available upon request. Contact anthony.koleske@yale. edu for more information. The limited proteolysis mass spectrometry data have been deposited to the ProteomeXchange Consortium via the PRIDE (46) partner repository with the dataset identifier PXD034393 (http://www.ebi.ac.uk/pride). The cross-linking raw mass spectrometry data and peak lists are available in the massIVE repository (https://massive.ucsd. edu) with accession number: MSV000089621 Annotated spectra supporting the cross-linked identifica- tions are published on MS-Viewer (https://msviewer.ucsf.edu/ cgi-bin/msform.cgi?form=msviewer) with the following search keys: Trio SR6-GEF1-WT: l4abvtas5a Trio GEF autoinhibition by spectrin repeats Trio SR6-GEF1-E883D: mmmpkfzwvo Trio SR6-GEF1-R1078Q: paout3qryt Trio SR6-GEF1-D1368V: 7xhepmd94b Supporting information—This article contains supporting informa- tion (18). Institutes of Health Acknowledgments—The mass spectrometers and the accompany biotechnology tools within the MS & Proteomics Resource at Yale University (used for limited proteolysis experiments) were funded in part by the Yale School of Medicine and by the Office of The Di- rector, National (S10OD02365101A1, S10OD019967, and S10OD018034). Crosslinking Mass spectrom- etry experiments were supported by the Adelson Medical Research Foundation and the University of California, San Francisco Program for Breakthrough Biomedical Research. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the article. We thank Daisy Duan, Amanda Jeng, and Wanqing Lyu for helpful comments on the article and Titus Boggon and Kimmie Vish for helpful insights on structure modeling and protein purification. We also thank Florine Collin and Jean Kanyo for help with mass spectrometry sample preparation and data collection, respectively. Author contributions—J. E. B., E. E. C., and A. J. K. conceptualiza- tion; J. E. B. and E. E. C. methodology; J. E. B., E. E. C., T. T. L., and M. J. T. formal analysis; J. E. B., E. E. C., T. T. L., and M. J. T. investigation; J. E. B., E. E. C., and A. J. K. writing–original draft; J. E. B. and E. E. C. visualization; J. E. B., E. E. C., and A. J. K. funding acquisition; J. E. B., E. E. C., and A. J. K. project administration; T. T. L. and M. J. T. writing–review and editing; A. J. K. supervision. Funding and additional information—This work was supported by the National Institute of Health (NIH) grants R56MH122449, R01 MH115939, and R01 NS105640 to A. J. K., F31MH127891-01 to E. E. C., and F31 NS113511-03 to J. E. B. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. Conflict of interest—The authors declare no competing financial conflicts of interest. Abbreviations—The abbreviations used are: ADB, antibody dilution buffer; BS3, bis(sulfosuccinimidyl)suberate; DH1, Dbl homology domain; FL, fluorescein; GEF, guanine exchange factor; NDD, neurodevelopmental disorder; Ni-NTA, nitrilotriacetic acid; PH1, pleckstrin homology domain; SR, spectrin repeat; TBS, Tris Buff- ered Saline. References 1. Debant, A., Serra-Pages, C., Seipel, K., O’Brien, S., Tang, M., Park, S. H., et al. (1996) The multidomain protein Trio binds the LAR trans- membrane tyrosine phosphatase, contains a protein kinase domain, and has separate rac-specific and rho-specific guanine nucleotide exchange factor domains. Proc. Natl. Acad. Sci. U. S. A. 93, 5466–5471 2. Steven, R., Kubiseski, T. J., Zheng, H., Kulkarni, S., Mancillas, J., Ruiz Morales, A., et al. (1998) UNC-73 activates the Rac GTPase and is required for cell and growth cone migrations in C. elegans. Cell 92, 785–795 3. Penzes, P., Johnson, R. C., Kambampati, V., Mains, R. E., and Eipper, B. A. (2001) Distinct roles for the two Rho GDP/GTP exchange factor domains J. Biol. Chem. (2022) 298(9) 102361 11 Trio GEF autoinhibition by spectrin repeats of kalirin in regulation of neurite growth and neuronal morphology. J. Neurosci. 21, 8426–8434 4. Bellanger, J. M., Lazaro, J. B., Diriong, S., Fernandez, A., Lamb, N., and Debant, A. (1998) The two guanine nucleotide exchange factor domains of Trio link the Rac1 and the RhoA pathways in vivo. Oncogene 16, 147–152 5. Chhatriwala, M. K., Betts, L., Worthylake, D. K., and Sondek, J. (2007) The DH and PH domains of Trio coordinately engage Rho GTPases for their efficient activation. J. Mol. Biol. 368, 1307–1320 6. Rabiner, C. A., Mains, R. E., and Eipper, B. A. (2005) Kalirin: a dual rho guanine nucleotide exchange factor that is so much more than the sum of its many parts. Neuroscientist 11, 148–160 7. Saito, K., Tautz, L., and Mustelin, T. (2007) The lipid-binding SEC14 domain. Biochim. Biophys. Acta 1771, 719–726 8. Schiller, M. R., Ferraro, F., Wang, Y., Ma, X. M., McPherson, C. E., Sobota, J. A., et al. (2008) Autonomous functions for the Sec14p/spectrin- repeat region of Kalirin. Exp. Cell Res. 314, 2674–2691 9. Vishwanatha, K. S., Wang, Y. P., Keutmann, H. T., Mains, R. E., and Eipper, B. A. (2012) Structural organization of the nine spectrin repeats of Kalirin. Biochemistry 51, 5663–5673 10. Bircher, J. E., and Koleske, A. J. (2021) Trio family proteins as regulators of cell migration and morphogenesis in development and disease - mechanisms and cellular contexts. J. Cell Sci. 134, jcs248393 11. Paskus, J. D., Herring, B. E., and Roche, K. W. (2020) Kalirin and trio: RhoGEFs in synaptic transmission, plasticity, and complex brain disor- ders. Trends Neurosci. 43, 505–518 12. Katrancha, S. M., Wu, Y., Zhu, M., Eipper, B. A., Koleske, A. J., and Mains, R. E. (2017) Neurodevelopmental disease-associated de novo mutations and rare sequence variants affect TRIO GDP/GTP exchange factor activity. Hum. Mol. Genet. 26, 4728–4740 13. Pengelly, R. J., Greville-Heygate, S., Schmidt, S., Seaby, E. G., Jabalameli, M. R., Mehta, S. G., et al. (2016) Mutations specific to the Rac-GEF domain of TRIO cause intellectual disability and microcephaly. J. Med. Genet. 53, 735–742 14. Sadybekov, A., Tian, C., Arnesano, C., Katritch, V., and Herring, B. E. (2017) An autism spectrum disorder-related de novo mutation hotspot discovered in the GEF1 domain of Trio. Nat. Commun. 8, 601 24. Debreceni, B., Gao, Y., Guo, F., Zhu, K., Jia, B., and Zheng, Y. (2004) Mechanisms of guanine nucleotide exchange and Rac-mediated signaling J. Biol. Chem. 279, revealed by a dominant negative trio mutant. 3777–3786 25. McPherson, C. E., Eipper, B. A., and Mains, R. E. (2005) Multiple novel isoforms of Trio are expressed in the developing rat brain. Gene 347, 125–135 26. van Rijssel, J., Hoogenboezem, M., Wester, L., Hordijk, P. L., and Van Buul, J. D. (2012) The N-terminal DH-PH domain of Trio induces cell spreading and migration by regulating lamellipodia dynamics in a Rac1- dependent fashion. PLoS One 7, e29912 27. Schiller, M. R., Chakrabarti, K., King, G. F., Schiller, N. I., Eipper, B. A., and Maciejewski, M. W. (2006) Regulation of RhoGEF activity by intra- molecular and intermolecular SH3 domain interactions. J. Biol. Chem. 281, 18774–18786 28. Schmidt, A., and Hall, A. (2002) Guanine nucleotide exchange factors for rho GTPases: turning on the switch. Genes Dev. 16, 1587–1609 29. Xu, Z., Gakhar, L., Bain, F. E., Spies, M., and Fuentes, E. J. (2017) The Tiam1 guanine nucleotide exchange factor is auto-inhibited by its pleckstrin homology coiled-coil extension domain. J. Biol. Chem. 292, 17777–17793 30. Bellanger, J. M., Estrach, S., Schmidt, S., Briancon-Marjollet, A., Zugasti, O., Fromont, S., et al. (2003) Different regulation of the Trio Dbl- homology domains by their associated PH domains. Biol. Cell 95, 625–634 31. Kubiseski, T. J., Culotti, J., and Pawson, T. (2003) Functional analysis of the Caenorhabditis elegans UNC-73B PH domain demonstrates a role in activation of the Rac GTPase in vitro and axon guidance in vivo. Mol. Cell. Biol. 23, 6823–6835 32. Tian, C., Paskus, J. D., Fingleton, E., Roche, K. W., and Herring, B. E. (2021) Autism spectrum disorder/intellectual disability-associated muta- tions in trio disrupt neuroligin 1-mediated synaptogenesis. J. Neurosci. 41, 7768–7778 33. Terry-Lorenzo, R. T., Torres, V. I., Wagh, D., Galaz, J., Swanson, S. K., Florens, L., et al. (2016) Trio, a rho family GEF, interacts with the pre- synaptic active zone proteins Piccolo and Bassoon. PLoS One 11, e0167535 15. Barbosa, S., Debant, A., Schmidt, S., and Baralle, D. (2020) Opposite modulation of RAC1 by mutations in TRIO is associated with distinct, domain-specific neurodevelopmental disorders. Am. J. Hum. Genet. 106, 338–355 34. Timmerman, I., Heemskerk, N., Kroon, J., Schaefer, A., van Rijssel, J., Hoogenboezem, M., et al. (2015) A local VE-cadherin and Trio-based signaling complex stabilizes endothelial junctions through Rac1. J. Cell Sci. 128, 3514 16. Newsome, T. P., Schmidt, S., Dietzl, G., Keleman, K., Asling, B., Debant, A., et al. (2000) Trio combines with dock to regulate Pak ac- tivity during photoreceptor axon pathfinding in Drosophila. Cell 101, 283–294 35. van Rijssel, J., Kroon, J., Hoogenboezem, M., van Alphen, F. P., de Jong, R. J., Kostadinova, E., et al. (2012) The Rho-guanine nucleotide exchange factor Trio controls leukocyte transendothelial migration by promoting docking structure formation. Mol. Biol. Cell 23, 2831–2844 17. Estrach, S., Schmidt, S., Diriong, S., Penna, A., Blangy, A., Fort, P., et al. (2002) The human Rho-GEF trio and its target GTPase RhoG are involved in the NGF pathway, leading to neurite outgrowth. Curr. Biol. 12, 307–312 18. Blaise, A. M., Corcoran, E. E., Wattenberg, E. S., Zhang, Y. L., Cottrell, J. R., and Koleske, A. J. (2022) In vitro fluorescence assay to measure GDP/ GTP exchange of guanine nucleotide exchange factors of Rho family GTPases. Biol. Methods Protoc. 7, bpab024 19. Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., et al. (2021) Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 20. Varadi, M., Anyango, S., Deshpande, M., Nair, S., Natassia, C., Yorda- nova, G., et al. (2021) AlphaFold protein structure database: massively expanding the structural coverage of protein-sequence space with high- accuracy models. Nucleic Acids Res. 50, D439–D444 21. Ward, J. J., McGuffin, L. J., Bryson, K., Buxton, B. F., and Jones, D. T. (2004) The DISOPRED server for the prediction of protein disorder. Bioinformatics 20, 2138–2139 22. Hall, A. (1998) Rho GTPases and the actin cytoskeleton. Science 279, 509–514 36. Tao, T., Sun, J., Peng, Y., Li, Y., Wang, P., Chen, X., et al. (2019) Golgi- resident TRIO regulates membrane trafficking during neurite outgrowth. J. Biol. Chem. 294, 10954–10968 37. Neubrand, V. E., Thomas, C., Schmidt, S., Debant, A., and Schiavo, G. (2010) Kidins220/ARMS regulates Rac1-dependent neurite outgrowth by direct interaction with the RhoGEF Trio. J. Cell Sci. 123, 2111 38. Son, K., Smith, T. C., and Luna, E. J. (2015) Supervillin binds the rac/rho- GEF trio and increases trio-mediated Rac1 activation. Cytoskeleton (Hoboken) 72, 47–64 39. Guex, N., and Peitsch, M. C. (1997) Swiss-Model and the Swiss- pdbViewer: an environment for comparative protein modeling. Electro- phoresis 18, 2714–2723 40. Sanchez, N. A., Kallweit, L. M., Trnka, M. J., Clemmer, C. L., and Al- Sady, B. (2021) Heterodimerization of H3K9 histone methyltransferases G9a and GLP activates methyl reading and writing capabilities. J. Biol. Chem. 297, 101276 41. Schnirch, L., Nadler-Holly, M., Siao, S. W., Frese, C. K., Viner, R., and Liu, F. (2020) Expanding the depth and sensitivity of cross-link identifi- cation by differential ion mobility using high-field asymmetric waveform ion mobility spectrometry. Anal. Chem. 92, 10495–10503 23. Wells, C. M., Walmsley, M., Ooi, S., Tybulewicz, V., and Ridley, A. J. (2004) Rac1-deficient macrophages exhibit defects in cell spreading and membrane ruffling but not migration. J. Cell Sci. 117, 1259–1268 42. Trnka, M. J., Baker, P. R., Robinson, P. J., Burlingame, A. L., and Chalkley, R. J. (2014) Matching cross-linked peptide spectra: only as good as the worse identification. Mol. Cell. Proteomics 13, 420–434 12 J. Biol. Chem. (2022) 298(9) 102361 Trio GEF autoinhibition by spectrin repeats 43. Lim, C., Berk, J. M., Blaise, A., Bircher, J., Koleske, A. J., Hochstrasser, M., et al. (2020) Crystal structure of a guanine nucleotide exchange factor encoded by the scrub typhus pathogen Orientia tsutsugamushi. Proc. Natl. Acad. Sci. U. S. A. 117, 30380–30390 44. Schindelin, J., Arganda-Carreras, I., Frise, E., Kaynig, V., Longair, M., Pietzsch, T., et al. (2012) Fiji: an open-source platform for biological- image analysis. Nat. Methods 9, 676–682 45. McQuin, C., Goodman, A., Chernyshev, V., Kamentsky, L., Cimini, B. A., Karhohs, K. W., et al. (2018) CellProfiler 3.0: next-generation image processing for biology. PLoS Biol. 16, e2005970 46. Perez-Riverol, Y., Bai, J., Bandla, C., Hewapathirana, S., García-Seisdedos, D., Kamatchinathan, S., et al. (2022) The PRIDE database resources in 2022: a hub for mass spectrometry-based proteomics evidences. Nucleic Acids Res. 50, D543–D552 J. Biol. Chem. (2022) 298(9) 102361 13